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 15784f91c62SLai Jiangshan /* The current concurrency level. */ 15884f91c62SLai Jiangshan atomic_t nr_running; 15984f91c62SLai Jiangshan 160bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 161ea1abd61SLai Jiangshan 1625826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1635826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 164bd7bdd43STejun Heo 165*2c1f1a91SLai Jiangshan struct list_head idle_list; /* L: list of idle workers */ 166bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 167bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 168bd7bdd43STejun Heo 169c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 170c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 171c9e7cf27STejun Heo /* L: hash of busy workers */ 172c9e7cf27STejun Heo 1732607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17492f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17560f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 176e19e397aSTejun Heo 1777cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 178e19e397aSTejun Heo 1797a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 18068e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 18168e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1827a4e344cSTejun Heo 183e19e397aSTejun Heo /* 18424acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 18529c91e99STejun Heo * from get_work_pool(). 18629c91e99STejun Heo */ 18729c91e99STejun Heo struct rcu_head rcu; 18884f91c62SLai Jiangshan }; 1898b03ae3cSTejun Heo 1908b03ae3cSTejun Heo /* 191112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 192112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 193112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 194112202d9STejun Heo * number of flag bits. 1951da177e4SLinus Torvalds */ 196112202d9STejun Heo struct pool_workqueue { 197bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1984690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19973f53c4aSTejun Heo int work_color; /* L: current color */ 20073f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2018864b4e5STejun Heo int refcnt; /* L: reference count */ 20273f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20373f53c4aSTejun Heo /* L: nr of in_flight works */ 204018f3a13SLai Jiangshan 205018f3a13SLai Jiangshan /* 206018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 207018f3a13SLai Jiangshan * 208018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 209018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 210018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 211018f3a13SLai Jiangshan * 212018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 213018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 214018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 215018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 216018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 217018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 218018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 219018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 220018f3a13SLai Jiangshan */ 2211e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 222a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 223f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2243c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2252e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2268864b4e5STejun Heo 2278864b4e5STejun Heo /* 2288864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2298864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 23024acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 231b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2328864b4e5STejun Heo */ 2338864b4e5STejun Heo struct work_struct unbound_release_work; 2348864b4e5STejun Heo struct rcu_head rcu; 235e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2361da177e4SLinus Torvalds 2371da177e4SLinus Torvalds /* 23873f53c4aSTejun Heo * Structure used to wait for workqueue flush. 23973f53c4aSTejun Heo */ 24073f53c4aSTejun Heo struct wq_flusher { 2413c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2423c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 24373f53c4aSTejun Heo struct completion done; /* flush completion */ 24473f53c4aSTejun Heo }; 2451da177e4SLinus Torvalds 246226223abSTejun Heo struct wq_device; 247226223abSTejun Heo 24873f53c4aSTejun Heo /* 249c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 250c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2511da177e4SLinus Torvalds */ 2521da177e4SLinus Torvalds struct workqueue_struct { 2533c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 254e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 25573f53c4aSTejun Heo 2563c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2573c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2583c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 259112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2603c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2613c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2623c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 26373f53c4aSTejun Heo 2642e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 26530ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 266e22bee78STejun Heo 26787fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 268a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 269226223abSTejun Heo 2705b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2715b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2726029a918STejun Heo 273226223abSTejun Heo #ifdef CONFIG_SYSFS 274226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 275226223abSTejun Heo #endif 2764e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 277669de8bdSBart Van Assche char *lock_name; 278669de8bdSBart Van Assche struct lock_class_key key; 2794e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2804e6045f1SJohannes Berg #endif 281ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2822728fd2fSTejun Heo 283e2dca7adSTejun Heo /* 28424acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 28524acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 286e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 287e2dca7adSTejun Heo */ 288e2dca7adSTejun Heo struct rcu_head rcu; 289e2dca7adSTejun Heo 2902728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2912728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2922728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2935b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2941da177e4SLinus Torvalds }; 2951da177e4SLinus Torvalds 296e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 297e904e6c2STejun Heo 298bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 299bce90380STejun Heo /* possible CPUs of each node */ 300bce90380STejun Heo 301d55262c4STejun Heo static bool wq_disable_numa; 302d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 303d55262c4STejun Heo 304cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 305552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 306cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 307cee22a15SViresh Kumar 308863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3093347fa09STejun Heo 310bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 311bce90380STejun Heo 3124c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3134c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3144c16bd32STejun Heo 31568e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3161258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 317a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 318d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 319d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3205bcab335STejun Heo 321e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 32268e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3237d19c5ceSTejun Heo 324ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 325ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 326ef557180SMike Galbraith 327ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 328ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 329b05a7928SFrederic Weisbecker 330f303fccbSTejun Heo /* 331f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 332f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 333f303fccbSTejun Heo * to uncover usages which depend on it. 334f303fccbSTejun Heo */ 335f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 336f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 337f303fccbSTejun Heo #else 338f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 339f303fccbSTejun Heo #endif 340f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 341f303fccbSTejun Heo 3427d19c5ceSTejun Heo /* the per-cpu worker pools */ 34325528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3447d19c5ceSTejun Heo 34568e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3467d19c5ceSTejun Heo 34768e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 34829c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 34929c91e99STejun Heo 350c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 35129c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 35229c91e99STejun Heo 3538a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3548a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3558a2b7538STejun Heo 356d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 357ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 358044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3591aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 360044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 361d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 362044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 363f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 364044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 36524d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3660668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3670668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3680668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3690668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 370d320c038STejun Heo 3717d19c5ceSTejun Heo static int worker_thread(void *__worker); 3726ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 373c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 37455df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 3757d19c5ceSTejun Heo 37697bd2347STejun Heo #define CREATE_TRACE_POINTS 37797bd2347STejun Heo #include <trace/events/workqueue.h> 37897bd2347STejun Heo 37968e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 38024acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 381f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 38224acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 3835bcab335STejun Heo 3845b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 38524acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 386f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 387f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 38824acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 3895b95e1afSLai Jiangshan 390f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 391f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 392f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3937a62c2c8STejun Heo (pool)++) 3944ce62e9eSTejun Heo 39549e3cf44STejun Heo /** 39617116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 39717116969STejun Heo * @pool: iteration cursor 398611c92a0STejun Heo * @pi: integer used for iteration 399fa1b54e6STejun Heo * 40024acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 40168e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 40268e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 403fa1b54e6STejun Heo * 404fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 405fa1b54e6STejun Heo * ignored. 40617116969STejun Heo */ 407611c92a0STejun Heo #define for_each_pool(pool, pi) \ 408611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 40968e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 410fa1b54e6STejun Heo else 41117116969STejun Heo 41217116969STejun Heo /** 413822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 414822d8405STejun Heo * @worker: iteration cursor 415822d8405STejun Heo * @pool: worker_pool to iterate workers of 416822d8405STejun Heo * 4171258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 418822d8405STejun Heo * 419822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 420822d8405STejun Heo * ignored. 421822d8405STejun Heo */ 422da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 423da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4241258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 425822d8405STejun Heo else 426822d8405STejun Heo 427822d8405STejun Heo /** 42849e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 42949e3cf44STejun Heo * @pwq: iteration cursor 43049e3cf44STejun Heo * @wq: the target workqueue 43176af4d93STejun Heo * 43224acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 433794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 434794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 43576af4d93STejun Heo * 43676af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 43776af4d93STejun Heo * ignored. 43849e3cf44STejun Heo */ 43949e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 44049e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4415a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 442f3421797STejun Heo 443dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 444dc186ad7SThomas Gleixner 445f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 446dc186ad7SThomas Gleixner 44799777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 44899777288SStanislaw Gruszka { 44999777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 45099777288SStanislaw Gruszka } 45199777288SStanislaw Gruszka 452b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 453b9fdac7fSDu, Changbin { 454b9fdac7fSDu, Changbin struct work_struct *work = addr; 455b9fdac7fSDu, Changbin 456b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 457b9fdac7fSDu, Changbin } 458b9fdac7fSDu, Changbin 459dc186ad7SThomas Gleixner /* 460dc186ad7SThomas Gleixner * fixup_init is called when: 461dc186ad7SThomas Gleixner * - an active object is initialized 462dc186ad7SThomas Gleixner */ 46302a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 464dc186ad7SThomas Gleixner { 465dc186ad7SThomas Gleixner struct work_struct *work = addr; 466dc186ad7SThomas Gleixner 467dc186ad7SThomas Gleixner switch (state) { 468dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 469dc186ad7SThomas Gleixner cancel_work_sync(work); 470dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 47102a982a6SDu, Changbin return true; 472dc186ad7SThomas Gleixner default: 47302a982a6SDu, Changbin return false; 474dc186ad7SThomas Gleixner } 475dc186ad7SThomas Gleixner } 476dc186ad7SThomas Gleixner 477dc186ad7SThomas Gleixner /* 478dc186ad7SThomas Gleixner * fixup_free is called when: 479dc186ad7SThomas Gleixner * - an active object is freed 480dc186ad7SThomas Gleixner */ 48102a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 482dc186ad7SThomas Gleixner { 483dc186ad7SThomas Gleixner struct work_struct *work = addr; 484dc186ad7SThomas Gleixner 485dc186ad7SThomas Gleixner switch (state) { 486dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 487dc186ad7SThomas Gleixner cancel_work_sync(work); 488dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 48902a982a6SDu, Changbin return true; 490dc186ad7SThomas Gleixner default: 49102a982a6SDu, Changbin return false; 492dc186ad7SThomas Gleixner } 493dc186ad7SThomas Gleixner } 494dc186ad7SThomas Gleixner 495f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 496dc186ad7SThomas Gleixner .name = "work_struct", 49799777288SStanislaw Gruszka .debug_hint = work_debug_hint, 498b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 499dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 500dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 501dc186ad7SThomas Gleixner }; 502dc186ad7SThomas Gleixner 503dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 504dc186ad7SThomas Gleixner { 505dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 506dc186ad7SThomas Gleixner } 507dc186ad7SThomas Gleixner 508dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 511dc186ad7SThomas Gleixner } 512dc186ad7SThomas Gleixner 513dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 514dc186ad7SThomas Gleixner { 515dc186ad7SThomas Gleixner if (onstack) 516dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 517dc186ad7SThomas Gleixner else 518dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 519dc186ad7SThomas Gleixner } 520dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 521dc186ad7SThomas Gleixner 522dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 523dc186ad7SThomas Gleixner { 524dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 525dc186ad7SThomas Gleixner } 526dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 527dc186ad7SThomas Gleixner 528ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 529ea2e64f2SThomas Gleixner { 530ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 531ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 532ea2e64f2SThomas Gleixner } 533ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 534ea2e64f2SThomas Gleixner 535dc186ad7SThomas Gleixner #else 536dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 537dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 538dc186ad7SThomas Gleixner #endif 539dc186ad7SThomas Gleixner 5404e8b22bdSLi Bin /** 54167dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5424e8b22bdSLi Bin * @pool: the pool pointer of interest 5434e8b22bdSLi Bin * 5444e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5454e8b22bdSLi Bin * successfully, -errno on failure. 5464e8b22bdSLi Bin */ 5479daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5489daf9e67STejun Heo { 5499daf9e67STejun Heo int ret; 5509daf9e67STejun Heo 55168e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5525bcab335STejun Heo 5534e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5544e8b22bdSLi Bin GFP_KERNEL); 555229641a6STejun Heo if (ret >= 0) { 556e68035fbSTejun Heo pool->id = ret; 557229641a6STejun Heo return 0; 558229641a6STejun Heo } 5599daf9e67STejun Heo return ret; 5609daf9e67STejun Heo } 5619daf9e67STejun Heo 56276af4d93STejun Heo /** 563df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 564df2d5ae4STejun Heo * @wq: the target workqueue 565df2d5ae4STejun Heo * @node: the node ID 566df2d5ae4STejun Heo * 56724acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 5685b95e1afSLai Jiangshan * read locked. 569df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 570df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 571d185af30SYacine Belkadi * 572d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 573df2d5ae4STejun Heo */ 574df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 575df2d5ae4STejun Heo int node) 576df2d5ae4STejun Heo { 5775b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 578d6e022f1STejun Heo 579d6e022f1STejun Heo /* 580d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 581d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 582d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 583d6e022f1STejun Heo * happens, this workaround can be removed. 584d6e022f1STejun Heo */ 585d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 586d6e022f1STejun Heo return wq->dfl_pwq; 587d6e022f1STejun Heo 588df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 589df2d5ae4STejun Heo } 590df2d5ae4STejun Heo 59173f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 59273f53c4aSTejun Heo { 59373f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 59473f53c4aSTejun Heo } 59573f53c4aSTejun Heo 596c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 59773f53c4aSTejun Heo { 598c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 59973f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 60073f53c4aSTejun Heo } 60173f53c4aSTejun Heo 60273f53c4aSTejun Heo static int work_next_color(int color) 60373f53c4aSTejun Heo { 60473f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 605a848e3b6SOleg Nesterov } 606a848e3b6SOleg Nesterov 6074594bf15SDavid Howells /* 608112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 609112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6107c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6117a22ad75STejun Heo * 612112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 613112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 614bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 615bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6167a22ad75STejun Heo * 617112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6187c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 619112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6207c3eed5cSTejun Heo * available only while the work item is queued. 621bbb68dfaSTejun Heo * 622bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 623bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 624bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 625bbb68dfaSTejun Heo * try to steal the PENDING bit. 6264594bf15SDavid Howells */ 6277a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6287a22ad75STejun Heo unsigned long flags) 6297a22ad75STejun Heo { 6306183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6317a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6327a22ad75STejun Heo } 6337a22ad75STejun Heo 634112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6354690c4abSTejun Heo unsigned long extra_flags) 636365970a1SDavid Howells { 637112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 638112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 639365970a1SDavid Howells } 640365970a1SDavid Howells 6414468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6424468a00fSLai Jiangshan int pool_id) 6434468a00fSLai Jiangshan { 6444468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6454468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6464468a00fSLai Jiangshan } 6474468a00fSLai Jiangshan 6487c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6497c3eed5cSTejun Heo int pool_id) 6504d707b9fSOleg Nesterov { 65123657bb1STejun Heo /* 65223657bb1STejun Heo * The following wmb is paired with the implied mb in 65323657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 65423657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 65523657bb1STejun Heo * owner. 65623657bb1STejun Heo */ 65723657bb1STejun Heo smp_wmb(); 6587c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 659346c09f8SRoman Pen /* 660346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 661346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 662346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6638bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 664346c09f8SRoman Pen * the same @work. E.g. consider this case: 665346c09f8SRoman Pen * 666346c09f8SRoman Pen * CPU#0 CPU#1 667346c09f8SRoman Pen * ---------------------------- -------------------------------- 668346c09f8SRoman Pen * 669346c09f8SRoman Pen * 1 STORE event_indicated 670346c09f8SRoman Pen * 2 queue_work_on() { 671346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 672346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 673346c09f8SRoman Pen * 5 set_work_data() # clear bit 674346c09f8SRoman Pen * 6 smp_mb() 675346c09f8SRoman Pen * 7 work->current_func() { 676346c09f8SRoman Pen * 8 LOAD event_indicated 677346c09f8SRoman Pen * } 678346c09f8SRoman Pen * 679346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 680346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 681346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 682346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 683346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 684346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 685346c09f8SRoman Pen * before actual STORE. 686346c09f8SRoman Pen */ 687346c09f8SRoman Pen smp_mb(); 6884d707b9fSOleg Nesterov } 6894d707b9fSOleg Nesterov 6907a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 691365970a1SDavid Howells { 6927c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6937c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6947a22ad75STejun Heo } 6957a22ad75STejun Heo 696112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6977a22ad75STejun Heo { 698e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6997a22ad75STejun Heo 700112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 701e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 702e120153dSTejun Heo else 703e120153dSTejun Heo return NULL; 7047a22ad75STejun Heo } 7057a22ad75STejun Heo 7067c3eed5cSTejun Heo /** 7077c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7087c3eed5cSTejun Heo * @work: the work item of interest 7097c3eed5cSTejun Heo * 71068e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 71124acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 71224acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 713fa1b54e6STejun Heo * 714fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 715fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 716fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 717fa1b54e6STejun Heo * returned pool is and stays online. 718d185af30SYacine Belkadi * 719d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7207c3eed5cSTejun Heo */ 7217c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7227a22ad75STejun Heo { 723e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7247c3eed5cSTejun Heo int pool_id; 7257a22ad75STejun Heo 72668e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 727fa1b54e6STejun Heo 728112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 729112202d9STejun Heo return ((struct pool_workqueue *) 7307c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7317a22ad75STejun Heo 7327c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7337c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7347a22ad75STejun Heo return NULL; 7357a22ad75STejun Heo 736fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7377c3eed5cSTejun Heo } 7387c3eed5cSTejun Heo 7397c3eed5cSTejun Heo /** 7407c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7417c3eed5cSTejun Heo * @work: the work item of interest 7427c3eed5cSTejun Heo * 743d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7447c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7457c3eed5cSTejun Heo */ 7467c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7477c3eed5cSTejun Heo { 74854d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7497c3eed5cSTejun Heo 750112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 751112202d9STejun Heo return ((struct pool_workqueue *) 75254d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 75354d5b7d0SLai Jiangshan 75454d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7557c3eed5cSTejun Heo } 7567c3eed5cSTejun Heo 757bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 758bbb68dfaSTejun Heo { 7597c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 760bbb68dfaSTejun Heo 7617c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7627c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 763bbb68dfaSTejun Heo } 764bbb68dfaSTejun Heo 765bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 766bbb68dfaSTejun Heo { 767bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 768bbb68dfaSTejun Heo 769112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 770bbb68dfaSTejun Heo } 771bbb68dfaSTejun Heo 772e22bee78STejun Heo /* 7733270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7743270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 775d565ed63STejun Heo * they're being called with pool->lock held. 776e22bee78STejun Heo */ 777e22bee78STejun Heo 77863d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 779649027d7STejun Heo { 780e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 781649027d7STejun Heo } 782649027d7STejun Heo 783e22bee78STejun Heo /* 784e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 785e22bee78STejun Heo * running workers. 786974271c4STejun Heo * 787974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 788706026c2STejun Heo * function will always return %true for unbound pools as long as the 789974271c4STejun Heo * worklist isn't empty. 790e22bee78STejun Heo */ 79163d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 792e22bee78STejun Heo { 79363d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 794e22bee78STejun Heo } 795e22bee78STejun Heo 796e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 79763d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 798e22bee78STejun Heo { 79963d95a91STejun Heo return pool->nr_idle; 800e22bee78STejun Heo } 801e22bee78STejun Heo 802e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 80363d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 804e22bee78STejun Heo { 805e19e397aSTejun Heo return !list_empty(&pool->worklist) && 806e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 807e22bee78STejun Heo } 808e22bee78STejun Heo 809e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 81063d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 811e22bee78STejun Heo { 81263d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 813e22bee78STejun Heo } 814e22bee78STejun Heo 815e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 81663d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 817e22bee78STejun Heo { 818692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 81963d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 82063d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 821e22bee78STejun Heo 822e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 823e22bee78STejun Heo } 824e22bee78STejun Heo 825e22bee78STejun Heo /* 826e22bee78STejun Heo * Wake up functions. 827e22bee78STejun Heo */ 828e22bee78STejun Heo 829*2c1f1a91SLai Jiangshan /* Return the first idle worker. Called with pool->lock held. */ 8301037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8317e11629dSTejun Heo { 83263d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8337e11629dSTejun Heo return NULL; 8347e11629dSTejun Heo 83563d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8367e11629dSTejun Heo } 8377e11629dSTejun Heo 8387e11629dSTejun Heo /** 8397e11629dSTejun Heo * wake_up_worker - wake up an idle worker 84063d95a91STejun Heo * @pool: worker pool to wake worker from 8417e11629dSTejun Heo * 84263d95a91STejun Heo * Wake up the first idle worker of @pool. 8437e11629dSTejun Heo * 8447e11629dSTejun Heo * CONTEXT: 845a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8467e11629dSTejun Heo */ 84763d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8487e11629dSTejun Heo { 8491037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8507e11629dSTejun Heo 8517e11629dSTejun Heo if (likely(worker)) 8527e11629dSTejun Heo wake_up_process(worker->task); 8537e11629dSTejun Heo } 8547e11629dSTejun Heo 8554690c4abSTejun Heo /** 8566d25be57SThomas Gleixner * wq_worker_running - a worker is running again 857e22bee78STejun Heo * @task: task waking up 858e22bee78STejun Heo * 8596d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 860e22bee78STejun Heo */ 8616d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 862e22bee78STejun Heo { 863e22bee78STejun Heo struct worker *worker = kthread_data(task); 864e22bee78STejun Heo 8656d25be57SThomas Gleixner if (!worker->sleeping) 8666d25be57SThomas Gleixner return; 86707edfeceSFrederic Weisbecker 86807edfeceSFrederic Weisbecker /* 86907edfeceSFrederic Weisbecker * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 87007edfeceSFrederic Weisbecker * and the nr_running increment below, we may ruin the nr_running reset 87107edfeceSFrederic Weisbecker * and leave with an unexpected pool->nr_running == 1 on the newly unbound 87207edfeceSFrederic Weisbecker * pool. Protect against such race. 87307edfeceSFrederic Weisbecker */ 87407edfeceSFrederic Weisbecker preempt_disable(); 8756d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 876e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 87707edfeceSFrederic Weisbecker preempt_enable(); 8786d25be57SThomas Gleixner worker->sleeping = 0; 87936576000SJoonsoo Kim } 880e22bee78STejun Heo 881e22bee78STejun Heo /** 882e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 883e22bee78STejun Heo * @task: task going to sleep 884e22bee78STejun Heo * 8856d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 886ccf45156SLai Jiangshan * going to sleep. 887e22bee78STejun Heo */ 8886d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 889e22bee78STejun Heo { 8906d25be57SThomas Gleixner struct worker *next, *worker = kthread_data(task); 891111c225aSTejun Heo struct worker_pool *pool; 892e22bee78STejun Heo 893111c225aSTejun Heo /* 894111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 895111c225aSTejun Heo * workers, also reach here, let's not access anything before 896111c225aSTejun Heo * checking NOT_RUNNING. 897111c225aSTejun Heo */ 8982d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 8996d25be57SThomas Gleixner return; 900e22bee78STejun Heo 901111c225aSTejun Heo pool = worker->pool; 902111c225aSTejun Heo 90362849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 90462849a96SSebastian Andrzej Siewior if (worker->sleeping) 9056d25be57SThomas Gleixner return; 9066d25be57SThomas Gleixner 9076d25be57SThomas Gleixner worker->sleeping = 1; 908a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 909e22bee78STejun Heo 910e22bee78STejun Heo /* 91145c753f5SFrederic Weisbecker * Recheck in case unbind_workers() preempted us. We don't 91245c753f5SFrederic Weisbecker * want to decrement nr_running after the worker is unbound 91345c753f5SFrederic Weisbecker * and nr_running has been reset. 91445c753f5SFrederic Weisbecker */ 91545c753f5SFrederic Weisbecker if (worker->flags & WORKER_NOT_RUNNING) { 91645c753f5SFrederic Weisbecker raw_spin_unlock_irq(&pool->lock); 91745c753f5SFrederic Weisbecker return; 91845c753f5SFrederic Weisbecker } 91945c753f5SFrederic Weisbecker 920e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 9216d25be57SThomas Gleixner !list_empty(&pool->worklist)) { 9226d25be57SThomas Gleixner next = first_idle_worker(pool); 9236d25be57SThomas Gleixner if (next) 9246d25be57SThomas Gleixner wake_up_process(next->task); 9256d25be57SThomas Gleixner } 926a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 927e22bee78STejun Heo } 928e22bee78STejun Heo 929e22bee78STejun Heo /** 9301b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 9318194fe94SBart Van Assche * @task: Task to retrieve last work function of. 9321b69ac6bSJohannes Weiner * 9331b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 9341b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 9351b69ac6bSJohannes Weiner * 9361b69ac6bSJohannes Weiner * CONTEXT: 937a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 9381b69ac6bSJohannes Weiner * 9394b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 9404b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 9414b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 9424b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 9434b047002SJohannes Weiner * 9444b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 9454b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 9464b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 9474b047002SJohannes Weiner * is guaranteed to not be processing any works. 9484b047002SJohannes Weiner * 9491b69ac6bSJohannes Weiner * Return: 9501b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 9511b69ac6bSJohannes Weiner * hasn't executed any work yet. 9521b69ac6bSJohannes Weiner */ 9531b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 9541b69ac6bSJohannes Weiner { 9551b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 9561b69ac6bSJohannes Weiner 9571b69ac6bSJohannes Weiner return worker->last_func; 9581b69ac6bSJohannes Weiner } 9591b69ac6bSJohannes Weiner 9601b69ac6bSJohannes Weiner /** 961e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 962cb444766STejun Heo * @worker: self 963d302f017STejun Heo * @flags: flags to set 964d302f017STejun Heo * 965228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 966d302f017STejun Heo * 967cb444766STejun Heo * CONTEXT: 968a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 969d302f017STejun Heo */ 970228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 971d302f017STejun Heo { 972bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 973e22bee78STejun Heo 974cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 975cb444766STejun Heo 976228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 977e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 978e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 979e19e397aSTejun Heo atomic_dec(&pool->nr_running); 980e22bee78STejun Heo } 981e22bee78STejun Heo 982d302f017STejun Heo worker->flags |= flags; 983d302f017STejun Heo } 984d302f017STejun Heo 985d302f017STejun Heo /** 986e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 987cb444766STejun Heo * @worker: self 988d302f017STejun Heo * @flags: flags to clear 989d302f017STejun Heo * 990e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 991d302f017STejun Heo * 992cb444766STejun Heo * CONTEXT: 993a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 994d302f017STejun Heo */ 995d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 996d302f017STejun Heo { 99763d95a91STejun Heo struct worker_pool *pool = worker->pool; 998e22bee78STejun Heo unsigned int oflags = worker->flags; 999e22bee78STejun Heo 1000cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 1001cb444766STejun Heo 1002d302f017STejun Heo worker->flags &= ~flags; 1003e22bee78STejun Heo 100442c025f3STejun Heo /* 100542c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 100642c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 100742c025f3STejun Heo * of multiple flags, not a single flag. 100842c025f3STejun Heo */ 1009e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 1010e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 1011e19e397aSTejun Heo atomic_inc(&pool->nr_running); 1012d302f017STejun Heo } 1013d302f017STejun Heo 1014d302f017STejun Heo /** 10158cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1016c9e7cf27STejun Heo * @pool: pool of interest 10178cca0eeaSTejun Heo * @work: work to find worker for 10188cca0eeaSTejun Heo * 1019c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1020c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1021a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1022a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1023a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1024a2c1c57bSTejun Heo * being executed. 1025a2c1c57bSTejun Heo * 1026a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1027a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1028a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1029a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1030a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1031a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1032a2c1c57bSTejun Heo * 1033c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1034c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1035c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1036c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1037c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1038c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10398cca0eeaSTejun Heo * 10408cca0eeaSTejun Heo * CONTEXT: 1041a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 10428cca0eeaSTejun Heo * 1043d185af30SYacine Belkadi * Return: 1044d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10458cca0eeaSTejun Heo * otherwise. 10468cca0eeaSTejun Heo */ 1047c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10488cca0eeaSTejun Heo struct work_struct *work) 10498cca0eeaSTejun Heo { 105042f8570fSSasha Levin struct worker *worker; 105142f8570fSSasha Levin 1052b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1053a2c1c57bSTejun Heo (unsigned long)work) 1054a2c1c57bSTejun Heo if (worker->current_work == work && 1055a2c1c57bSTejun Heo worker->current_func == work->func) 105642f8570fSSasha Levin return worker; 105742f8570fSSasha Levin 105842f8570fSSasha Levin return NULL; 10598cca0eeaSTejun Heo } 10608cca0eeaSTejun Heo 10618cca0eeaSTejun Heo /** 1062bf4ede01STejun Heo * move_linked_works - move linked works to a list 1063bf4ede01STejun Heo * @work: start of series of works to be scheduled 1064bf4ede01STejun Heo * @head: target list to append @work to 1065402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1066bf4ede01STejun Heo * 1067bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1068bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1069bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1070bf4ede01STejun Heo * 1071bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1072bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1073bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1074bf4ede01STejun Heo * 1075bf4ede01STejun Heo * CONTEXT: 1076a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1077bf4ede01STejun Heo */ 1078bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1079bf4ede01STejun Heo struct work_struct **nextp) 1080bf4ede01STejun Heo { 1081bf4ede01STejun Heo struct work_struct *n; 1082bf4ede01STejun Heo 1083bf4ede01STejun Heo /* 1084bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1085bf4ede01STejun Heo * use NULL for list head. 1086bf4ede01STejun Heo */ 1087bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1088bf4ede01STejun Heo list_move_tail(&work->entry, head); 1089bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1090bf4ede01STejun Heo break; 1091bf4ede01STejun Heo } 1092bf4ede01STejun Heo 1093bf4ede01STejun Heo /* 1094bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1095bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1096bf4ede01STejun Heo * needs to be updated. 1097bf4ede01STejun Heo */ 1098bf4ede01STejun Heo if (nextp) 1099bf4ede01STejun Heo *nextp = n; 1100bf4ede01STejun Heo } 1101bf4ede01STejun Heo 11028864b4e5STejun Heo /** 11038864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 11048864b4e5STejun Heo * @pwq: pool_workqueue to get 11058864b4e5STejun Heo * 11068864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 11078864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 11088864b4e5STejun Heo */ 11098864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 11108864b4e5STejun Heo { 11118864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11128864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 11138864b4e5STejun Heo pwq->refcnt++; 11148864b4e5STejun Heo } 11158864b4e5STejun Heo 11168864b4e5STejun Heo /** 11178864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 11188864b4e5STejun Heo * @pwq: pool_workqueue to put 11198864b4e5STejun Heo * 11208864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 11218864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11228864b4e5STejun Heo */ 11238864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11248864b4e5STejun Heo { 11258864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11268864b4e5STejun Heo if (likely(--pwq->refcnt)) 11278864b4e5STejun Heo return; 11288864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11298864b4e5STejun Heo return; 11308864b4e5STejun Heo /* 11318864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11328864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11338864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11348864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11358864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11368864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11378864b4e5STejun Heo */ 11388864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11398864b4e5STejun Heo } 11408864b4e5STejun Heo 1141dce90d47STejun Heo /** 1142dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1143dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1144dce90d47STejun Heo * 1145dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1146dce90d47STejun Heo */ 1147dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1148dce90d47STejun Heo { 1149dce90d47STejun Heo if (pwq) { 1150dce90d47STejun Heo /* 115124acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1152dce90d47STejun Heo * following lock operations are safe. 1153dce90d47STejun Heo */ 1154a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1155dce90d47STejun Heo put_pwq(pwq); 1156a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1157dce90d47STejun Heo } 1158dce90d47STejun Heo } 1159dce90d47STejun Heo 1160f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1161bf4ede01STejun Heo { 1162112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1163bf4ede01STejun Heo 1164bf4ede01STejun Heo trace_workqueue_activate_work(work); 116582607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 116682607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1167112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1168f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1169112202d9STejun Heo pwq->nr_active++; 1170bf4ede01STejun Heo } 1171bf4ede01STejun Heo 1172f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 11733aa62497SLai Jiangshan { 1174f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 11753aa62497SLai Jiangshan struct work_struct, entry); 11763aa62497SLai Jiangshan 1177f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 11783aa62497SLai Jiangshan } 11793aa62497SLai Jiangshan 1180bf4ede01STejun Heo /** 1181112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1182112202d9STejun Heo * @pwq: pwq of interest 1183c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1184bf4ede01STejun Heo * 1185bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1186112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1187bf4ede01STejun Heo * 1188bf4ede01STejun Heo * CONTEXT: 1189a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1190bf4ede01STejun Heo */ 1191c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1192bf4ede01STejun Heo { 1193c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1194c4560c2cSLai Jiangshan 1195018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1196112202d9STejun Heo pwq->nr_active--; 1197f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1198f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1199112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1200f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1201bf4ede01STejun Heo } 1202018f3a13SLai Jiangshan } 1203018f3a13SLai Jiangshan 1204018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1205bf4ede01STejun Heo 1206bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1207112202d9STejun Heo if (likely(pwq->flush_color != color)) 12088864b4e5STejun Heo goto out_put; 1209bf4ede01STejun Heo 1210bf4ede01STejun Heo /* are there still in-flight works? */ 1211112202d9STejun Heo if (pwq->nr_in_flight[color]) 12128864b4e5STejun Heo goto out_put; 1213bf4ede01STejun Heo 1214112202d9STejun Heo /* this pwq is done, clear flush_color */ 1215112202d9STejun Heo pwq->flush_color = -1; 1216bf4ede01STejun Heo 1217bf4ede01STejun Heo /* 1218112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1219bf4ede01STejun Heo * will handle the rest. 1220bf4ede01STejun Heo */ 1221112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1222112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12238864b4e5STejun Heo out_put: 12248864b4e5STejun Heo put_pwq(pwq); 1225bf4ede01STejun Heo } 1226bf4ede01STejun Heo 122736e227d2STejun Heo /** 1228bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 122936e227d2STejun Heo * @work: work item to steal 123036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1231bbb68dfaSTejun Heo * @flags: place to store irq state 123236e227d2STejun Heo * 123336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1234d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 123536e227d2STejun Heo * 1236d185af30SYacine Belkadi * Return: 12373eb6b31bSMauro Carvalho Chehab * 12383eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 123936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 124036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 124136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1242bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1243bbb68dfaSTejun Heo * for arbitrarily long 12443eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 124536e227d2STejun Heo * 1246d185af30SYacine Belkadi * Note: 1247bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1248e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1249e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1250e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1251bbb68dfaSTejun Heo * 1252bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1253bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1254bbb68dfaSTejun Heo * 1255e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1256bf4ede01STejun Heo */ 1257bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1258bbb68dfaSTejun Heo unsigned long *flags) 1259bf4ede01STejun Heo { 1260d565ed63STejun Heo struct worker_pool *pool; 1261112202d9STejun Heo struct pool_workqueue *pwq; 1262bf4ede01STejun Heo 1263bbb68dfaSTejun Heo local_irq_save(*flags); 1264bbb68dfaSTejun Heo 126536e227d2STejun Heo /* try to steal the timer if it exists */ 126636e227d2STejun Heo if (is_dwork) { 126736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 126836e227d2STejun Heo 1269e0aecdd8STejun Heo /* 1270e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1271e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1272e0aecdd8STejun Heo * running on the local CPU. 1273e0aecdd8STejun Heo */ 127436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 127536e227d2STejun Heo return 1; 127636e227d2STejun Heo } 127736e227d2STejun Heo 127836e227d2STejun Heo /* try to claim PENDING the normal way */ 1279bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1280bf4ede01STejun Heo return 0; 1281bf4ede01STejun Heo 128224acfb71SThomas Gleixner rcu_read_lock(); 1283bf4ede01STejun Heo /* 1284bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1285bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1286bf4ede01STejun Heo */ 1287d565ed63STejun Heo pool = get_work_pool(work); 1288d565ed63STejun Heo if (!pool) 1289bbb68dfaSTejun Heo goto fail; 1290bf4ede01STejun Heo 1291a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1292bf4ede01STejun Heo /* 1293112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1294112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1295112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1296112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1297112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12980b3dae68SLai Jiangshan * item is currently queued on that pool. 1299bf4ede01STejun Heo */ 1300112202d9STejun Heo pwq = get_work_pwq(work); 1301112202d9STejun Heo if (pwq && pwq->pool == pool) { 1302bf4ede01STejun Heo debug_work_deactivate(work); 13033aa62497SLai Jiangshan 13043aa62497SLai Jiangshan /* 1305018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1306018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1307018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1308018f3a13SLai Jiangshan * 1309f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1310d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1311f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 131216062836STejun Heo * management later on and cause stall. Make sure the work 131316062836STejun Heo * item is activated before grabbing. 13143aa62497SLai Jiangshan */ 1315f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1316f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13173aa62497SLai Jiangshan 1318bf4ede01STejun Heo list_del_init(&work->entry); 1319c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 132036e227d2STejun Heo 1321112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 13224468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 13234468a00fSLai Jiangshan 1324a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 132524acfb71SThomas Gleixner rcu_read_unlock(); 132636e227d2STejun Heo return 1; 1327bf4ede01STejun Heo } 1328a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1329bbb68dfaSTejun Heo fail: 133024acfb71SThomas Gleixner rcu_read_unlock(); 1331bbb68dfaSTejun Heo local_irq_restore(*flags); 1332bbb68dfaSTejun Heo if (work_is_canceling(work)) 1333bbb68dfaSTejun Heo return -ENOENT; 1334bbb68dfaSTejun Heo cpu_relax(); 133536e227d2STejun Heo return -EAGAIN; 1336bf4ede01STejun Heo } 1337bf4ede01STejun Heo 1338bf4ede01STejun Heo /** 1339706026c2STejun Heo * insert_work - insert a work into a pool 1340112202d9STejun Heo * @pwq: pwq @work belongs to 13414690c4abSTejun Heo * @work: work to insert 13424690c4abSTejun Heo * @head: insertion point 13434690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13444690c4abSTejun Heo * 1345112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1346706026c2STejun Heo * work_struct flags. 13474690c4abSTejun Heo * 13484690c4abSTejun Heo * CONTEXT: 1349a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1350365970a1SDavid Howells */ 1351112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1352112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1353b89deed3SOleg Nesterov { 1354112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1355e1d8aa9fSFrederic Weisbecker 1356e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1357f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1358e89a85d6SWalter Wu 13594690c4abSTejun Heo /* we own @work, set data and link */ 1360112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13611a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13628864b4e5STejun Heo get_pwq(pwq); 1363e22bee78STejun Heo 136463d95a91STejun Heo if (__need_more_worker(pool)) 136563d95a91STejun Heo wake_up_worker(pool); 1366b89deed3SOleg Nesterov } 1367b89deed3SOleg Nesterov 1368c8efcc25STejun Heo /* 1369c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13708d03ecfeSTejun Heo * same workqueue. 1371c8efcc25STejun Heo */ 1372c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1373c8efcc25STejun Heo { 1374c8efcc25STejun Heo struct worker *worker; 1375c8efcc25STejun Heo 13768d03ecfeSTejun Heo worker = current_wq_worker(); 1377c8efcc25STejun Heo /* 1378bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 13798d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1380c8efcc25STejun Heo */ 1381112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1382c8efcc25STejun Heo } 1383c8efcc25STejun Heo 1384ef557180SMike Galbraith /* 1385ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1386ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1387ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1388ef557180SMike Galbraith */ 1389ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1390ef557180SMike Galbraith { 1391f303fccbSTejun Heo static bool printed_dbg_warning; 1392ef557180SMike Galbraith int new_cpu; 1393ef557180SMike Galbraith 1394f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1395ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1396ef557180SMike Galbraith return cpu; 1397f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1398f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1399f303fccbSTejun Heo printed_dbg_warning = true; 1400f303fccbSTejun Heo } 1401f303fccbSTejun Heo 1402ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1403ef557180SMike Galbraith return cpu; 1404ef557180SMike Galbraith 1405ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1406ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1407ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1408ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1409ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1410ef557180SMike Galbraith return cpu; 1411ef557180SMike Galbraith } 1412ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1413ef557180SMike Galbraith 1414ef557180SMike Galbraith return new_cpu; 1415ef557180SMike Galbraith } 1416ef557180SMike Galbraith 1417d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 14181da177e4SLinus Torvalds struct work_struct *work) 14191da177e4SLinus Torvalds { 1420112202d9STejun Heo struct pool_workqueue *pwq; 1421c9178087STejun Heo struct worker_pool *last_pool; 14221e19ffc6STejun Heo struct list_head *worklist; 14238a2e8e5dSTejun Heo unsigned int work_flags; 1424b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 14258930cabaSTejun Heo 14268930cabaSTejun Heo /* 14278930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 14288930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 14298930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 14308930cabaSTejun Heo * happen with IRQ disabled. 14318930cabaSTejun Heo */ 14328e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 14331da177e4SLinus Torvalds 14341e19ffc6STejun Heo 14359ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1436618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1437c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1438e41e704bSTejun Heo return; 143924acfb71SThomas Gleixner rcu_read_lock(); 14409e8cd2f5STejun Heo retry: 1441aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1442aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1443df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1444ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1445df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1446aa202f1fSHillf Danton } else { 1447aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1448aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1449aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1450aa202f1fSHillf Danton } 1451f3421797STejun Heo 145218aa9effSTejun Heo /* 1453c9178087STejun Heo * If @work was previously on a different pool, it might still be 1454c9178087STejun Heo * running there, in which case the work needs to be queued on that 1455c9178087STejun Heo * pool to guarantee non-reentrancy. 145618aa9effSTejun Heo */ 1457c9e7cf27STejun Heo last_pool = get_work_pool(work); 1458112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 145918aa9effSTejun Heo struct worker *worker; 146018aa9effSTejun Heo 1461a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 146218aa9effSTejun Heo 1463c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 146418aa9effSTejun Heo 1465112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1466c9178087STejun Heo pwq = worker->current_pwq; 14678594fadeSLai Jiangshan } else { 146818aa9effSTejun Heo /* meh... not running there, queue here */ 1469a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1470a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 147118aa9effSTejun Heo } 14728930cabaSTejun Heo } else { 1473a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 14748930cabaSTejun Heo } 1475502ca9d8STejun Heo 14769e8cd2f5STejun Heo /* 14779e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14789e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14799e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1480df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1481df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14829e8cd2f5STejun Heo * make forward-progress. 14839e8cd2f5STejun Heo */ 14849e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14859e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1486a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 14879e8cd2f5STejun Heo cpu_relax(); 14889e8cd2f5STejun Heo goto retry; 14899e8cd2f5STejun Heo } 14909e8cd2f5STejun Heo /* oops */ 14919e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14929e8cd2f5STejun Heo wq->name, cpu); 14939e8cd2f5STejun Heo } 14949e8cd2f5STejun Heo 1495112202d9STejun Heo /* pwq determined, queue */ 1496112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1497502ca9d8STejun Heo 149824acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 149924acfb71SThomas Gleixner goto out; 15001e19ffc6STejun Heo 1501112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1502112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 15031e19ffc6STejun Heo 1504112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1505cdadf009STejun Heo trace_workqueue_activate_work(work); 1506112202d9STejun Heo pwq->nr_active++; 1507112202d9STejun Heo worklist = &pwq->pool->worklist; 150882607adcSTejun Heo if (list_empty(worklist)) 150982607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 15108a2e8e5dSTejun Heo } else { 1511f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1512f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 15138a2e8e5dSTejun Heo } 15141e19ffc6STejun Heo 15150687c66bSZqiang debug_work_activate(work); 1516112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 15171e19ffc6STejun Heo 151824acfb71SThomas Gleixner out: 1519a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 152024acfb71SThomas Gleixner rcu_read_unlock(); 15211da177e4SLinus Torvalds } 15221da177e4SLinus Torvalds 15230fcb78c2SRolf Eike Beer /** 1524c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1525c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1526c1a220e7SZhang Rui * @wq: workqueue to use 1527c1a220e7SZhang Rui * @work: work to queue 1528c1a220e7SZhang Rui * 1529c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1530443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1531443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1532d185af30SYacine Belkadi * 1533d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1534c1a220e7SZhang Rui */ 1535d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1536d4283e93STejun Heo struct work_struct *work) 1537c1a220e7SZhang Rui { 1538d4283e93STejun Heo bool ret = false; 15398930cabaSTejun Heo unsigned long flags; 15408930cabaSTejun Heo 15418930cabaSTejun Heo local_irq_save(flags); 1542c1a220e7SZhang Rui 154322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15444690c4abSTejun Heo __queue_work(cpu, wq, work); 1545d4283e93STejun Heo ret = true; 1546c1a220e7SZhang Rui } 15478930cabaSTejun Heo 15488930cabaSTejun Heo local_irq_restore(flags); 1549c1a220e7SZhang Rui return ret; 1550c1a220e7SZhang Rui } 1551ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1552c1a220e7SZhang Rui 15538204e0c1SAlexander Duyck /** 15548204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15558204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15568204e0c1SAlexander Duyck * 15578204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15588204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15598204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15608204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15618204e0c1SAlexander Duyck */ 15628204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15638204e0c1SAlexander Duyck { 15648204e0c1SAlexander Duyck int cpu; 15658204e0c1SAlexander Duyck 15668204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15678204e0c1SAlexander Duyck if (!wq_numa_enabled) 15688204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15698204e0c1SAlexander Duyck 15708204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15718204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15728204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15738204e0c1SAlexander Duyck 15748204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15758204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15768204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15778204e0c1SAlexander Duyck return cpu; 15788204e0c1SAlexander Duyck 15798204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15808204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15818204e0c1SAlexander Duyck 15828204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15838204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15848204e0c1SAlexander Duyck } 15858204e0c1SAlexander Duyck 15868204e0c1SAlexander Duyck /** 15878204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15888204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15898204e0c1SAlexander Duyck * @wq: workqueue to use 15908204e0c1SAlexander Duyck * @work: work to queue 15918204e0c1SAlexander Duyck * 15928204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 15938204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 15948204e0c1SAlexander Duyck * NUMA node. 15958204e0c1SAlexander Duyck * 15968204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 15978204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 15988204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 15998204e0c1SAlexander Duyck * 16008204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 16018204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 16028204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 16038204e0c1SAlexander Duyck * 16048204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 16058204e0c1SAlexander Duyck */ 16068204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 16078204e0c1SAlexander Duyck struct work_struct *work) 16088204e0c1SAlexander Duyck { 16098204e0c1SAlexander Duyck unsigned long flags; 16108204e0c1SAlexander Duyck bool ret = false; 16118204e0c1SAlexander Duyck 16128204e0c1SAlexander Duyck /* 16138204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 16148204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 16158204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 16168204e0c1SAlexander Duyck * 16178204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 16188204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 16198204e0c1SAlexander Duyck * some round robin type logic. 16208204e0c1SAlexander Duyck */ 16218204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 16228204e0c1SAlexander Duyck 16238204e0c1SAlexander Duyck local_irq_save(flags); 16248204e0c1SAlexander Duyck 16258204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16268204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16278204e0c1SAlexander Duyck 16288204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16298204e0c1SAlexander Duyck ret = true; 16308204e0c1SAlexander Duyck } 16318204e0c1SAlexander Duyck 16328204e0c1SAlexander Duyck local_irq_restore(flags); 16338204e0c1SAlexander Duyck return ret; 16348204e0c1SAlexander Duyck } 16358204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16368204e0c1SAlexander Duyck 16378c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16381da177e4SLinus Torvalds { 16398c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16401da177e4SLinus Torvalds 1641e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 164260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16431da177e4SLinus Torvalds } 16441438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16451da177e4SLinus Torvalds 16467beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 164752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16481da177e4SLinus Torvalds { 16497beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16507beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16511da177e4SLinus Torvalds 1652637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 165398173112SSami Tolvanen WARN_ON_FUNCTION_MISMATCH(timer->function, delayed_work_timer_fn); 1654fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1655fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16567beb2edfSTejun Heo 16578852aac2STejun Heo /* 16588852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16598852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16608852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16618852aac2STejun Heo * on that there's no such delay when @delay is 0. 16628852aac2STejun Heo */ 16638852aac2STejun Heo if (!delay) { 16648852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16658852aac2STejun Heo return; 16668852aac2STejun Heo } 16678852aac2STejun Heo 166860c057bcSLai Jiangshan dwork->wq = wq; 16691265057fSTejun Heo dwork->cpu = cpu; 16707beb2edfSTejun Heo timer->expires = jiffies + delay; 16717beb2edfSTejun Heo 1672041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16737beb2edfSTejun Heo add_timer_on(timer, cpu); 1674041bd12eSTejun Heo else 1675041bd12eSTejun Heo add_timer(timer); 16767beb2edfSTejun Heo } 16771da177e4SLinus Torvalds 16780fcb78c2SRolf Eike Beer /** 16790fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16800fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16810fcb78c2SRolf Eike Beer * @wq: workqueue to use 1682af9997e4SRandy Dunlap * @dwork: work to queue 16830fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16840fcb78c2SRolf Eike Beer * 1685d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1686715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1687715f1300STejun Heo * execution. 16880fcb78c2SRolf Eike Beer */ 1689d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 169052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16917a6bc1cdSVenkatesh Pallipadi { 169252bad64dSDavid Howells struct work_struct *work = &dwork->work; 1693d4283e93STejun Heo bool ret = false; 16948930cabaSTejun Heo unsigned long flags; 16958930cabaSTejun Heo 16968930cabaSTejun Heo /* read the comment in __queue_work() */ 16978930cabaSTejun Heo local_irq_save(flags); 16987a6bc1cdSVenkatesh Pallipadi 169922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17007beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1701d4283e93STejun Heo ret = true; 17027a6bc1cdSVenkatesh Pallipadi } 17038930cabaSTejun Heo 17048930cabaSTejun Heo local_irq_restore(flags); 17057a6bc1cdSVenkatesh Pallipadi return ret; 17067a6bc1cdSVenkatesh Pallipadi } 1707ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 17081da177e4SLinus Torvalds 1709c8e55f36STejun Heo /** 17108376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 17118376fe22STejun Heo * @cpu: CPU number to execute work on 17128376fe22STejun Heo * @wq: workqueue to use 17138376fe22STejun Heo * @dwork: work to queue 17148376fe22STejun Heo * @delay: number of jiffies to wait before queueing 17158376fe22STejun Heo * 17168376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 17178376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 17188376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 17198376fe22STejun Heo * current state. 17208376fe22STejun Heo * 1721d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 17228376fe22STejun Heo * pending and its timer was modified. 17238376fe22STejun Heo * 1724e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17258376fe22STejun Heo * See try_to_grab_pending() for details. 17268376fe22STejun Heo */ 17278376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17288376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17298376fe22STejun Heo { 17308376fe22STejun Heo unsigned long flags; 17318376fe22STejun Heo int ret; 17328376fe22STejun Heo 17338376fe22STejun Heo do { 17348376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17358376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17368376fe22STejun Heo 17378376fe22STejun Heo if (likely(ret >= 0)) { 17388376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17398376fe22STejun Heo local_irq_restore(flags); 17408376fe22STejun Heo } 17418376fe22STejun Heo 17428376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17438376fe22STejun Heo return ret; 17448376fe22STejun Heo } 17458376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17468376fe22STejun Heo 174705f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 174805f0fe6bSTejun Heo { 174905f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 175005f0fe6bSTejun Heo 175105f0fe6bSTejun Heo /* read the comment in __queue_work() */ 175205f0fe6bSTejun Heo local_irq_disable(); 175305f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 175405f0fe6bSTejun Heo local_irq_enable(); 175505f0fe6bSTejun Heo } 175605f0fe6bSTejun Heo 175705f0fe6bSTejun Heo /** 175805f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 175905f0fe6bSTejun Heo * @wq: workqueue to use 176005f0fe6bSTejun Heo * @rwork: work to queue 176105f0fe6bSTejun Heo * 176205f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 176305f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1764bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 176505f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 176605f0fe6bSTejun Heo */ 176705f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 176805f0fe6bSTejun Heo { 176905f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 177005f0fe6bSTejun Heo 177105f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 177205f0fe6bSTejun Heo rwork->wq = wq; 177305f0fe6bSTejun Heo call_rcu(&rwork->rcu, rcu_work_rcufn); 177405f0fe6bSTejun Heo return true; 177505f0fe6bSTejun Heo } 177605f0fe6bSTejun Heo 177705f0fe6bSTejun Heo return false; 177805f0fe6bSTejun Heo } 177905f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 178005f0fe6bSTejun Heo 17818376fe22STejun Heo /** 1782c8e55f36STejun Heo * worker_enter_idle - enter idle state 1783c8e55f36STejun Heo * @worker: worker which is entering idle state 1784c8e55f36STejun Heo * 1785c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1786c8e55f36STejun Heo * necessary. 1787c8e55f36STejun Heo * 1788c8e55f36STejun Heo * LOCKING: 1789a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1790c8e55f36STejun Heo */ 1791c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 17921da177e4SLinus Torvalds { 1793bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1794c8e55f36STejun Heo 17956183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 17966183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 17976183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 17986183c009STejun Heo return; 1799c8e55f36STejun Heo 1800051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1801cb444766STejun Heo worker->flags |= WORKER_IDLE; 1802bd7bdd43STejun Heo pool->nr_idle++; 1803e22bee78STejun Heo worker->last_active = jiffies; 1804c8e55f36STejun Heo 1805c8e55f36STejun Heo /* idle_list is LIFO */ 1806bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1807db7bccf4STejun Heo 180863d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1809628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1810cb444766STejun Heo 1811989442d7SLai Jiangshan /* Sanity check nr_running. */ 1812989442d7SLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && 1813e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1814c8e55f36STejun Heo } 1815c8e55f36STejun Heo 1816c8e55f36STejun Heo /** 1817c8e55f36STejun Heo * worker_leave_idle - leave idle state 1818c8e55f36STejun Heo * @worker: worker which is leaving idle state 1819c8e55f36STejun Heo * 1820c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1821c8e55f36STejun Heo * 1822c8e55f36STejun Heo * LOCKING: 1823a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1824c8e55f36STejun Heo */ 1825c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1826c8e55f36STejun Heo { 1827bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1828c8e55f36STejun Heo 18296183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18306183c009STejun Heo return; 1831d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1832bd7bdd43STejun Heo pool->nr_idle--; 1833c8e55f36STejun Heo list_del_init(&worker->entry); 1834c8e55f36STejun Heo } 1835c8e55f36STejun Heo 1836f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1837c34056a3STejun Heo { 1838c34056a3STejun Heo struct worker *worker; 1839c34056a3STejun Heo 1840f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1841c8e55f36STejun Heo if (worker) { 1842c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1843affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1844da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1845e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1846e22bee78STejun Heo worker->flags = WORKER_PREP; 1847c8e55f36STejun Heo } 1848c34056a3STejun Heo return worker; 1849c34056a3STejun Heo } 1850c34056a3STejun Heo 1851c34056a3STejun Heo /** 18524736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18534736cbf7SLai Jiangshan * @worker: worker to be attached 18544736cbf7SLai Jiangshan * @pool: the target pool 18554736cbf7SLai Jiangshan * 18564736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18574736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18584736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18594736cbf7SLai Jiangshan */ 18604736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18614736cbf7SLai Jiangshan struct worker_pool *pool) 18624736cbf7SLai Jiangshan { 18631258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18644736cbf7SLai Jiangshan 18654736cbf7SLai Jiangshan /* 18661258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18671258fae7STejun Heo * stable across this function. See the comments above the flag 18681258fae7STejun Heo * definition for details. 18694736cbf7SLai Jiangshan */ 18704736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 18714736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 18725c25b5ffSPeter Zijlstra else 18735c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 18744736cbf7SLai Jiangshan 1875640f17c8SPeter Zijlstra if (worker->rescue_wq) 1876640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 1877640f17c8SPeter Zijlstra 18784736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1879a2d812a2STejun Heo worker->pool = pool; 18804736cbf7SLai Jiangshan 18811258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 18824736cbf7SLai Jiangshan } 18834736cbf7SLai Jiangshan 18844736cbf7SLai Jiangshan /** 188560f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 188660f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 188760f5a4bcSLai Jiangshan * 18884736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 18894736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 18904736cbf7SLai Jiangshan * other reference to the pool. 189160f5a4bcSLai Jiangshan */ 1892a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 189360f5a4bcSLai Jiangshan { 1894a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 189560f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 189660f5a4bcSLai Jiangshan 18971258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1898a2d812a2STejun Heo 18995c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 1900da028469SLai Jiangshan list_del(&worker->node); 1901a2d812a2STejun Heo worker->pool = NULL; 1902a2d812a2STejun Heo 1903da028469SLai Jiangshan if (list_empty(&pool->workers)) 190460f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 19051258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 190660f5a4bcSLai Jiangshan 1907b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1908b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1909b62c0751SLai Jiangshan 191060f5a4bcSLai Jiangshan if (detach_completion) 191160f5a4bcSLai Jiangshan complete(detach_completion); 191260f5a4bcSLai Jiangshan } 191360f5a4bcSLai Jiangshan 191460f5a4bcSLai Jiangshan /** 1915c34056a3STejun Heo * create_worker - create a new workqueue worker 191663d95a91STejun Heo * @pool: pool the new worker will belong to 1917c34056a3STejun Heo * 1918051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1919c34056a3STejun Heo * 1920c34056a3STejun Heo * CONTEXT: 1921c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1922c34056a3STejun Heo * 1923d185af30SYacine Belkadi * Return: 1924c34056a3STejun Heo * Pointer to the newly created worker. 1925c34056a3STejun Heo */ 1926bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1927c34056a3STejun Heo { 1928e441b56fSZhen Lei struct worker *worker; 1929e441b56fSZhen Lei int id; 1930e3c916a4STejun Heo char id_buf[16]; 1931c34056a3STejun Heo 19327cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 1933e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 1934822d8405STejun Heo if (id < 0) 1935e441b56fSZhen Lei return NULL; 1936c34056a3STejun Heo 1937f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1938c34056a3STejun Heo if (!worker) 1939c34056a3STejun Heo goto fail; 1940c34056a3STejun Heo 1941c34056a3STejun Heo worker->id = id; 1942c34056a3STejun Heo 194329c91e99STejun Heo if (pool->cpu >= 0) 1944e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1945e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1946f3421797STejun Heo else 1947e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1948e3c916a4STejun Heo 1949f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1950e3c916a4STejun Heo "kworker/%s", id_buf); 1951c34056a3STejun Heo if (IS_ERR(worker->task)) 1952c34056a3STejun Heo goto fail; 1953c34056a3STejun Heo 195491151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 195525834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 195691151228SOleg Nesterov 1957da028469SLai Jiangshan /* successful, attach the worker to the pool */ 19584736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1959822d8405STejun Heo 1960051e1850SLai Jiangshan /* start the newly created worker */ 1961a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1962051e1850SLai Jiangshan worker->pool->nr_workers++; 1963051e1850SLai Jiangshan worker_enter_idle(worker); 1964051e1850SLai Jiangshan wake_up_process(worker->task); 1965a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1966051e1850SLai Jiangshan 1967c34056a3STejun Heo return worker; 1968822d8405STejun Heo 1969c34056a3STejun Heo fail: 1970e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 1971c34056a3STejun Heo kfree(worker); 1972c34056a3STejun Heo return NULL; 1973c34056a3STejun Heo } 1974c34056a3STejun Heo 1975c34056a3STejun Heo /** 1976c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1977c34056a3STejun Heo * @worker: worker to be destroyed 1978c34056a3STejun Heo * 197973eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 198073eb7fe7SLai Jiangshan * be idle. 1981c8e55f36STejun Heo * 1982c8e55f36STejun Heo * CONTEXT: 1983a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1984c34056a3STejun Heo */ 1985c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1986c34056a3STejun Heo { 1987bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1988c34056a3STejun Heo 1989cd549687STejun Heo lockdep_assert_held(&pool->lock); 1990cd549687STejun Heo 1991c34056a3STejun Heo /* sanity check frenzy */ 19926183c009STejun Heo if (WARN_ON(worker->current_work) || 199373eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 199473eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 19956183c009STejun Heo return; 1996c34056a3STejun Heo 1997bd7bdd43STejun Heo pool->nr_workers--; 1998bd7bdd43STejun Heo pool->nr_idle--; 1999c8e55f36STejun Heo 2000c8e55f36STejun Heo list_del_init(&worker->entry); 2001cb444766STejun Heo worker->flags |= WORKER_DIE; 200260f5a4bcSLai Jiangshan wake_up_process(worker->task); 2003c34056a3STejun Heo } 2004c34056a3STejun Heo 200532a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2006e22bee78STejun Heo { 200732a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 2008e22bee78STejun Heo 2009a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2010e22bee78STejun Heo 20113347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2012e22bee78STejun Heo struct worker *worker; 2013e22bee78STejun Heo unsigned long expires; 2014e22bee78STejun Heo 2015e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 201663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2017e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2018e22bee78STejun Heo 20193347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 202063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 20213347fc9fSLai Jiangshan break; 2022e22bee78STejun Heo } 20233347fc9fSLai Jiangshan 20243347fc9fSLai Jiangshan destroy_worker(worker); 2025e22bee78STejun Heo } 2026e22bee78STejun Heo 2027a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2028e22bee78STejun Heo } 2029e22bee78STejun Heo 2030493a1724STejun Heo static void send_mayday(struct work_struct *work) 2031e22bee78STejun Heo { 2032112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2033112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2034493a1724STejun Heo 20352e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2036e22bee78STejun Heo 2037493008a8STejun Heo if (!wq->rescuer) 2038493a1724STejun Heo return; 2039e22bee78STejun Heo 2040e22bee78STejun Heo /* mayday mayday mayday */ 2041493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 204277668c8bSLai Jiangshan /* 204377668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 204477668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 204577668c8bSLai Jiangshan * rescuer is done with it. 204677668c8bSLai Jiangshan */ 204777668c8bSLai Jiangshan get_pwq(pwq); 2048493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2049e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2050493a1724STejun Heo } 2051e22bee78STejun Heo } 2052e22bee78STejun Heo 205332a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2054e22bee78STejun Heo { 205532a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2056e22bee78STejun Heo struct work_struct *work; 2057e22bee78STejun Heo 2058a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2059a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2060e22bee78STejun Heo 206163d95a91STejun Heo if (need_to_create_worker(pool)) { 2062e22bee78STejun Heo /* 2063e22bee78STejun Heo * We've been trying to create a new worker but 2064e22bee78STejun Heo * haven't been successful. We might be hitting an 2065e22bee78STejun Heo * allocation deadlock. Send distress signals to 2066e22bee78STejun Heo * rescuers. 2067e22bee78STejun Heo */ 206863d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2069e22bee78STejun Heo send_mayday(work); 2070e22bee78STejun Heo } 2071e22bee78STejun Heo 2072a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2073a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2074e22bee78STejun Heo 207563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2076e22bee78STejun Heo } 2077e22bee78STejun Heo 2078e22bee78STejun Heo /** 2079e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 208063d95a91STejun Heo * @pool: pool to create a new worker for 2081e22bee78STejun Heo * 208263d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2083e22bee78STejun Heo * have at least one idle worker on return from this function. If 2084e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 208563d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2086e22bee78STejun Heo * possible allocation deadlock. 2087e22bee78STejun Heo * 2088c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2089c5aa87bbSTejun Heo * may_start_working() %true. 2090e22bee78STejun Heo * 2091e22bee78STejun Heo * LOCKING: 2092a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2093e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2094e22bee78STejun Heo * manager. 2095e22bee78STejun Heo */ 209629187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2097d565ed63STejun Heo __releases(&pool->lock) 2098d565ed63STejun Heo __acquires(&pool->lock) 2099e22bee78STejun Heo { 2100e22bee78STejun Heo restart: 2101a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 21029f9c2364STejun Heo 2103e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 210463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2105e22bee78STejun Heo 2106e22bee78STejun Heo while (true) { 2107051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2108e22bee78STejun Heo break; 2109e22bee78STejun Heo 2110e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 21119f9c2364STejun Heo 211263d95a91STejun Heo if (!need_to_create_worker(pool)) 2113e22bee78STejun Heo break; 2114e22bee78STejun Heo } 2115e22bee78STejun Heo 211663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2117a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2118051e1850SLai Jiangshan /* 2119051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2120051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2121051e1850SLai Jiangshan * already become busy. 2122051e1850SLai Jiangshan */ 212363d95a91STejun Heo if (need_to_create_worker(pool)) 2124e22bee78STejun Heo goto restart; 2125e22bee78STejun Heo } 2126e22bee78STejun Heo 2127e22bee78STejun Heo /** 2128e22bee78STejun Heo * manage_workers - manage worker pool 2129e22bee78STejun Heo * @worker: self 2130e22bee78STejun Heo * 2131706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2132e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2133706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2134e22bee78STejun Heo * 2135e22bee78STejun Heo * The caller can safely start processing works on false return. On 2136e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2137e22bee78STejun Heo * and may_start_working() is true. 2138e22bee78STejun Heo * 2139e22bee78STejun Heo * CONTEXT: 2140a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2141e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2142e22bee78STejun Heo * 2143d185af30SYacine Belkadi * Return: 214429187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 214529187a9eSTejun Heo * start processing works, %true if management function was performed and 214629187a9eSTejun Heo * the conditions that the caller verified before calling the function may 214729187a9eSTejun Heo * no longer be true. 2148e22bee78STejun Heo */ 2149e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2150e22bee78STejun Heo { 215163d95a91STejun Heo struct worker_pool *pool = worker->pool; 2152e22bee78STejun Heo 2153692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 215429187a9eSTejun Heo return false; 2155692b4825STejun Heo 2156692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 21572607d7a6STejun Heo pool->manager = worker; 2158e22bee78STejun Heo 215929187a9eSTejun Heo maybe_create_worker(pool); 2160e22bee78STejun Heo 21612607d7a6STejun Heo pool->manager = NULL; 2162692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2163d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 216429187a9eSTejun Heo return true; 2165e22bee78STejun Heo } 2166e22bee78STejun Heo 2167a62428c0STejun Heo /** 2168a62428c0STejun Heo * process_one_work - process single work 2169c34056a3STejun Heo * @worker: self 2170a62428c0STejun Heo * @work: work to process 2171a62428c0STejun Heo * 2172a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2173a62428c0STejun Heo * process a single work including synchronization against and 2174a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2175a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2176a62428c0STejun Heo * call this function to process a work. 2177a62428c0STejun Heo * 2178a62428c0STejun Heo * CONTEXT: 2179a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2180a62428c0STejun Heo */ 2181c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2182d565ed63STejun Heo __releases(&pool->lock) 2183d565ed63STejun Heo __acquires(&pool->lock) 21841da177e4SLinus Torvalds { 2185112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2186bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2187112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 2188c4560c2cSLai Jiangshan unsigned long work_data; 21897e11629dSTejun Heo struct worker *collision; 21904e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21914e6045f1SJohannes Berg /* 2192a62428c0STejun Heo * It is permissible to free the struct work_struct from 2193a62428c0STejun Heo * inside the function that is called from it, this we need to 2194a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2195a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2196a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21974e6045f1SJohannes Berg */ 21984d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21994d82a1deSPeter Zijlstra 22004d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 22014e6045f1SJohannes Berg #endif 2202807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 220385327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2204ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 220525511a47STejun Heo 22067e11629dSTejun Heo /* 22077e11629dSTejun Heo * A single work shouldn't be executed concurrently by 22087e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 22097e11629dSTejun Heo * already processing the work. If so, defer the work to the 22107e11629dSTejun Heo * currently executing one. 22117e11629dSTejun Heo */ 2212c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 22137e11629dSTejun Heo if (unlikely(collision)) { 22147e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 22157e11629dSTejun Heo return; 22167e11629dSTejun Heo } 22171da177e4SLinus Torvalds 22188930cabaSTejun Heo /* claim and dequeue */ 22191da177e4SLinus Torvalds debug_work_deactivate(work); 2220c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2221c34056a3STejun Heo worker->current_work = work; 2222a2c1c57bSTejun Heo worker->current_func = work->func; 2223112202d9STejun Heo worker->current_pwq = pwq; 2224c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2225d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 22267a22ad75STejun Heo 22278bf89593STejun Heo /* 22288bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 22298bf89593STejun Heo * overridden through set_worker_desc(). 22308bf89593STejun Heo */ 22318bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 22328bf89593STejun Heo 2233a62428c0STejun Heo list_del_init(&work->entry); 2234a62428c0STejun Heo 2235649027d7STejun Heo /* 2236228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2237228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2238228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2239228f1d00SLai Jiangshan * execution of the pending work items. 2240fb0e7bebSTejun Heo */ 2241fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2242228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2243fb0e7bebSTejun Heo 2244974271c4STejun Heo /* 2245a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2246a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2247a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2248a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2249228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2250974271c4STejun Heo */ 2251a489a03eSLai Jiangshan if (need_more_worker(pool)) 225263d95a91STejun Heo wake_up_worker(pool); 2253974271c4STejun Heo 22548930cabaSTejun Heo /* 22557c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2256d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 225723657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 225823657bb1STejun Heo * disabled. 22598930cabaSTejun Heo */ 22607c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22611da177e4SLinus Torvalds 2262a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2263365970a1SDavid Howells 2264a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 22653295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2266e6f3faa7SPeter Zijlstra /* 2267f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2268f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2269e6f3faa7SPeter Zijlstra * 2270e6f3faa7SPeter Zijlstra * However, that would result in: 2271e6f3faa7SPeter Zijlstra * 2272e6f3faa7SPeter Zijlstra * A(W1) 2273e6f3faa7SPeter Zijlstra * WFC(C) 2274e6f3faa7SPeter Zijlstra * A(W1) 2275e6f3faa7SPeter Zijlstra * C(C) 2276e6f3faa7SPeter Zijlstra * 2277e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2278e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2279e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2280f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2281e6f3faa7SPeter Zijlstra * these locks. 2282e6f3faa7SPeter Zijlstra * 2283e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2284e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2285e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2286e6f3faa7SPeter Zijlstra */ 2287f52be570SPeter Zijlstra lockdep_invariant_state(true); 2288e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2289a2c1c57bSTejun Heo worker->current_func(work); 2290e36c886aSArjan van de Ven /* 2291e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2292e36c886aSArjan van de Ven * point will only record its address. 2293e36c886aSArjan van de Ven */ 22941c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 22953295f0efSIngo Molnar lock_map_release(&lockdep_map); 2296112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22971da177e4SLinus Torvalds 2298d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2299044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2300d75f773cSSakari Ailus " last function: %ps\n", 2301a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2302a2c1c57bSTejun Heo worker->current_func); 2303d5abe669SPeter Zijlstra debug_show_held_locks(current); 2304d5abe669SPeter Zijlstra dump_stack(); 2305d5abe669SPeter Zijlstra } 2306d5abe669SPeter Zijlstra 2307b22ce278STejun Heo /* 2308025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2309b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2310b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2311b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2312789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2313789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2314b22ce278STejun Heo */ 2315a7e6425eSPaul E. McKenney cond_resched(); 2316b22ce278STejun Heo 2317a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2318a62428c0STejun Heo 2319fb0e7bebSTejun Heo /* clear cpu intensive status */ 2320fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2321fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2322fb0e7bebSTejun Heo 23231b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 23241b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 23251b69ac6bSJohannes Weiner 2326a62428c0STejun Heo /* we're done with it, release */ 232742f8570fSSasha Levin hash_del(&worker->hentry); 2328c34056a3STejun Heo worker->current_work = NULL; 2329a2c1c57bSTejun Heo worker->current_func = NULL; 2330112202d9STejun Heo worker->current_pwq = NULL; 2331d812796eSLai Jiangshan worker->current_color = INT_MAX; 2332c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 23331da177e4SLinus Torvalds } 23341da177e4SLinus Torvalds 2335affee4b2STejun Heo /** 2336affee4b2STejun Heo * process_scheduled_works - process scheduled works 2337affee4b2STejun Heo * @worker: self 2338affee4b2STejun Heo * 2339affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2340affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2341affee4b2STejun Heo * fetches a work from the top and executes it. 2342affee4b2STejun Heo * 2343affee4b2STejun Heo * CONTEXT: 2344a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2345affee4b2STejun Heo * multiple times. 2346affee4b2STejun Heo */ 2347affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 23481da177e4SLinus Torvalds { 2349affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2350affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2351a62428c0STejun Heo struct work_struct, entry); 2352c34056a3STejun Heo process_one_work(worker, work); 2353a62428c0STejun Heo } 23541da177e4SLinus Torvalds } 23551da177e4SLinus Torvalds 2356197f6accSTejun Heo static void set_pf_worker(bool val) 2357197f6accSTejun Heo { 2358197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2359197f6accSTejun Heo if (val) 2360197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2361197f6accSTejun Heo else 2362197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2363197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2364197f6accSTejun Heo } 2365197f6accSTejun Heo 23664690c4abSTejun Heo /** 23674690c4abSTejun Heo * worker_thread - the worker thread function 2368c34056a3STejun Heo * @__worker: self 23694690c4abSTejun Heo * 2370c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2371c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2372c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2373c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2374c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2375d185af30SYacine Belkadi * 2376d185af30SYacine Belkadi * Return: 0 23774690c4abSTejun Heo */ 2378c34056a3STejun Heo static int worker_thread(void *__worker) 23791da177e4SLinus Torvalds { 2380c34056a3STejun Heo struct worker *worker = __worker; 2381bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 23821da177e4SLinus Torvalds 2383e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2384197f6accSTejun Heo set_pf_worker(true); 2385c8e55f36STejun Heo woke_up: 2386a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2387affee4b2STejun Heo 2388a9ab775bSTejun Heo /* am I supposed to die? */ 2389a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2390a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2391a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2392197f6accSTejun Heo set_pf_worker(false); 239360f5a4bcSLai Jiangshan 239460f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2395e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2396a2d812a2STejun Heo worker_detach_from_pool(worker); 239760f5a4bcSLai Jiangshan kfree(worker); 2398c8e55f36STejun Heo return 0; 2399c8e55f36STejun Heo } 2400c8e55f36STejun Heo 2401c8e55f36STejun Heo worker_leave_idle(worker); 2402db7bccf4STejun Heo recheck: 2403e22bee78STejun Heo /* no more worker necessary? */ 240463d95a91STejun Heo if (!need_more_worker(pool)) 2405e22bee78STejun Heo goto sleep; 2406e22bee78STejun Heo 2407e22bee78STejun Heo /* do we need to manage? */ 240863d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2409e22bee78STejun Heo goto recheck; 2410e22bee78STejun Heo 2411c8e55f36STejun Heo /* 2412c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2413c8e55f36STejun Heo * preparing to process a work or actually processing it. 2414c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2415c8e55f36STejun Heo */ 24166183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2417c8e55f36STejun Heo 2418e22bee78STejun Heo /* 2419a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2420a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2421a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2422a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2423a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2424e22bee78STejun Heo */ 2425a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2426e22bee78STejun Heo 2427e22bee78STejun Heo do { 2428affee4b2STejun Heo struct work_struct *work = 2429bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2430affee4b2STejun Heo struct work_struct, entry); 2431affee4b2STejun Heo 243282607adcSTejun Heo pool->watchdog_ts = jiffies; 243382607adcSTejun Heo 2434c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2435affee4b2STejun Heo /* optimization path, not strictly necessary */ 2436affee4b2STejun Heo process_one_work(worker, work); 2437affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2438affee4b2STejun Heo process_scheduled_works(worker); 2439affee4b2STejun Heo } else { 2440c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2441affee4b2STejun Heo process_scheduled_works(worker); 2442affee4b2STejun Heo } 244363d95a91STejun Heo } while (keep_working(pool)); 2444affee4b2STejun Heo 2445228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2446d313dd85STejun Heo sleep: 2447c8e55f36STejun Heo /* 2448d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2449d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2450d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2451d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2452d565ed63STejun Heo * event. 2453c8e55f36STejun Heo */ 2454c8e55f36STejun Heo worker_enter_idle(worker); 2455c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2456a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24571da177e4SLinus Torvalds schedule(); 2458c8e55f36STejun Heo goto woke_up; 24591da177e4SLinus Torvalds } 24601da177e4SLinus Torvalds 2461e22bee78STejun Heo /** 2462e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2463111c225aSTejun Heo * @__rescuer: self 2464e22bee78STejun Heo * 2465e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2466493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2467e22bee78STejun Heo * 2468706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2469e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2470e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2471e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2472e22bee78STejun Heo * the problem rescuer solves. 2473e22bee78STejun Heo * 2474706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2475706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2476e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2477e22bee78STejun Heo * 2478e22bee78STejun Heo * This should happen rarely. 2479d185af30SYacine Belkadi * 2480d185af30SYacine Belkadi * Return: 0 2481e22bee78STejun Heo */ 2482111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2483e22bee78STejun Heo { 2484111c225aSTejun Heo struct worker *rescuer = __rescuer; 2485111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2486e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 24874d595b86SLai Jiangshan bool should_stop; 2488e22bee78STejun Heo 2489e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2490111c225aSTejun Heo 2491111c225aSTejun Heo /* 2492111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2493111c225aSTejun Heo * doesn't participate in concurrency management. 2494111c225aSTejun Heo */ 2495197f6accSTejun Heo set_pf_worker(true); 2496e22bee78STejun Heo repeat: 2497c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 24981da177e4SLinus Torvalds 24994d595b86SLai Jiangshan /* 25004d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 25014d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 25024d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 25034d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 25044d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 25054d595b86SLai Jiangshan * list is always empty on exit. 25064d595b86SLai Jiangshan */ 25074d595b86SLai Jiangshan should_stop = kthread_should_stop(); 25081da177e4SLinus Torvalds 2509493a1724STejun Heo /* see whether any pwq is asking for help */ 2510a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2511493a1724STejun Heo 2512493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2513493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2514493a1724STejun Heo struct pool_workqueue, mayday_node); 2515112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2516e22bee78STejun Heo struct work_struct *work, *n; 251782607adcSTejun Heo bool first = true; 2518e22bee78STejun Heo 2519e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2520493a1724STejun Heo list_del_init(&pwq->mayday_node); 2521493a1724STejun Heo 2522a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2523e22bee78STejun Heo 252451697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 252551697d39SLai Jiangshan 2526a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2527e22bee78STejun Heo 2528e22bee78STejun Heo /* 2529e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2530e22bee78STejun Heo * process'em. 2531e22bee78STejun Heo */ 25320479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 253382607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 253482607adcSTejun Heo if (get_work_pwq(work) == pwq) { 253582607adcSTejun Heo if (first) 253682607adcSTejun Heo pool->watchdog_ts = jiffies; 2537e22bee78STejun Heo move_linked_works(work, scheduled, &n); 253882607adcSTejun Heo } 253982607adcSTejun Heo first = false; 254082607adcSTejun Heo } 2541e22bee78STejun Heo 2542008847f6SNeilBrown if (!list_empty(scheduled)) { 2543e22bee78STejun Heo process_scheduled_works(rescuer); 25447576958aSTejun Heo 25457576958aSTejun Heo /* 2546008847f6SNeilBrown * The above execution of rescued work items could 2547008847f6SNeilBrown * have created more to rescue through 2548f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2549008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2550008847f6SNeilBrown * that such back-to-back work items, which may be 2551008847f6SNeilBrown * being used to relieve memory pressure, don't 2552008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2553008847f6SNeilBrown */ 25544f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2555a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2556e66b39afSTejun Heo /* 2557e66b39afSTejun Heo * Queue iff we aren't racing destruction 2558e66b39afSTejun Heo * and somebody else hasn't queued it already. 2559e66b39afSTejun Heo */ 2560e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2561008847f6SNeilBrown get_pwq(pwq); 2562e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2563e66b39afSTejun Heo } 2564a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2565008847f6SNeilBrown } 2566008847f6SNeilBrown } 2567008847f6SNeilBrown 2568008847f6SNeilBrown /* 256977668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 257013b1d625SLai Jiangshan * go away while we're still attached to it. 257177668c8bSLai Jiangshan */ 257277668c8bSLai Jiangshan put_pwq(pwq); 257377668c8bSLai Jiangshan 257477668c8bSLai Jiangshan /* 2575d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 25767576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 25777576958aSTejun Heo * and stalling the execution. 25787576958aSTejun Heo */ 2579d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 258063d95a91STejun Heo wake_up_worker(pool); 25817576958aSTejun Heo 2582a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 258313b1d625SLai Jiangshan 2584a2d812a2STejun Heo worker_detach_from_pool(rescuer); 258513b1d625SLai Jiangshan 2586a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 25871da177e4SLinus Torvalds } 25881da177e4SLinus Torvalds 2589a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2590493a1724STejun Heo 25914d595b86SLai Jiangshan if (should_stop) { 25924d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2593197f6accSTejun Heo set_pf_worker(false); 25944d595b86SLai Jiangshan return 0; 25954d595b86SLai Jiangshan } 25964d595b86SLai Jiangshan 2597111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2598111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2599e22bee78STejun Heo schedule(); 2600e22bee78STejun Heo goto repeat; 26011da177e4SLinus Torvalds } 26021da177e4SLinus Torvalds 2603fca839c0STejun Heo /** 2604fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2605fca839c0STejun Heo * @target_wq: workqueue being flushed 2606fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2607fca839c0STejun Heo * 2608fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2609fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2610fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2611fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2612fca839c0STejun Heo * a deadlock. 2613fca839c0STejun Heo */ 2614fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2615fca839c0STejun Heo struct work_struct *target_work) 2616fca839c0STejun Heo { 2617fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2618fca839c0STejun Heo struct worker *worker; 2619fca839c0STejun Heo 2620fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2621fca839c0STejun Heo return; 2622fca839c0STejun Heo 2623fca839c0STejun Heo worker = current_wq_worker(); 2624fca839c0STejun Heo 2625fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2626d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2627fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 262823d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 262923d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2630d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2631fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2632fca839c0STejun Heo target_wq->name, target_func); 2633fca839c0STejun Heo } 2634fca839c0STejun Heo 2635fc2e4d70SOleg Nesterov struct wq_barrier { 2636fc2e4d70SOleg Nesterov struct work_struct work; 2637fc2e4d70SOleg Nesterov struct completion done; 26382607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2639fc2e4d70SOleg Nesterov }; 2640fc2e4d70SOleg Nesterov 2641fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2642fc2e4d70SOleg Nesterov { 2643fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2644fc2e4d70SOleg Nesterov complete(&barr->done); 2645fc2e4d70SOleg Nesterov } 2646fc2e4d70SOleg Nesterov 26474690c4abSTejun Heo /** 26484690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2649112202d9STejun Heo * @pwq: pwq to insert barrier into 26504690c4abSTejun Heo * @barr: wq_barrier to insert 2651affee4b2STejun Heo * @target: target work to attach @barr to 2652affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 26534690c4abSTejun Heo * 2654affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2655affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2656affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2657affee4b2STejun Heo * cpu. 2658affee4b2STejun Heo * 2659affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2660affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2661affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2662affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2663affee4b2STejun Heo * after a work with LINKED flag set. 2664affee4b2STejun Heo * 2665affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2666112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 26674690c4abSTejun Heo * 26684690c4abSTejun Heo * CONTEXT: 2669a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 26704690c4abSTejun Heo */ 2671112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2672affee4b2STejun Heo struct wq_barrier *barr, 2673affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2674fc2e4d70SOleg Nesterov { 2675d812796eSLai Jiangshan unsigned int work_flags = 0; 2676d812796eSLai Jiangshan unsigned int work_color; 2677affee4b2STejun Heo struct list_head *head; 2678affee4b2STejun Heo 2679dc186ad7SThomas Gleixner /* 2680d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2681dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2682dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2683dc186ad7SThomas Gleixner * might deadlock. 2684dc186ad7SThomas Gleixner */ 2685ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 268622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 268752fa5bc5SBoqun Feng 2688fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2689fd1a5b04SByungchul Park 26902607d7a6STejun Heo barr->task = current; 269183c22520SOleg Nesterov 2692018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2693018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2694018f3a13SLai Jiangshan 2695affee4b2STejun Heo /* 2696affee4b2STejun Heo * If @target is currently being executed, schedule the 2697affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2698affee4b2STejun Heo */ 2699d812796eSLai Jiangshan if (worker) { 2700affee4b2STejun Heo head = worker->scheduled.next; 2701d812796eSLai Jiangshan work_color = worker->current_color; 2702d812796eSLai Jiangshan } else { 2703affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2704affee4b2STejun Heo 2705affee4b2STejun Heo head = target->entry.next; 2706affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2707d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 2708d812796eSLai Jiangshan work_color = get_work_color(*bits); 2709affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2710affee4b2STejun Heo } 2711affee4b2STejun Heo 2712d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 2713d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 2714d812796eSLai Jiangshan 2715dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2716d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 2717fc2e4d70SOleg Nesterov } 2718fc2e4d70SOleg Nesterov 271973f53c4aSTejun Heo /** 2720112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 272173f53c4aSTejun Heo * @wq: workqueue being flushed 272273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 272373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 272473f53c4aSTejun Heo * 2725112202d9STejun Heo * Prepare pwqs for workqueue flushing. 272673f53c4aSTejun Heo * 2727112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2728112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2729112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2730112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2731112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 273273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 273373f53c4aSTejun Heo * 273473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 273573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 273673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 273773f53c4aSTejun Heo * is returned. 273873f53c4aSTejun Heo * 2739112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 274073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 274173f53c4aSTejun Heo * advanced to @work_color. 274273f53c4aSTejun Heo * 274373f53c4aSTejun Heo * CONTEXT: 27443c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 274573f53c4aSTejun Heo * 2746d185af30SYacine Belkadi * Return: 274773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 274873f53c4aSTejun Heo * otherwise. 274973f53c4aSTejun Heo */ 2750112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 275173f53c4aSTejun Heo int flush_color, int work_color) 27521da177e4SLinus Torvalds { 275373f53c4aSTejun Heo bool wait = false; 275449e3cf44STejun Heo struct pool_workqueue *pwq; 27551da177e4SLinus Torvalds 275673f53c4aSTejun Heo if (flush_color >= 0) { 27576183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2758112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2759dc186ad7SThomas Gleixner } 276014441960SOleg Nesterov 276149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2762112202d9STejun Heo struct worker_pool *pool = pwq->pool; 27631da177e4SLinus Torvalds 2764a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 276573f53c4aSTejun Heo 276673f53c4aSTejun Heo if (flush_color >= 0) { 27676183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 276873f53c4aSTejun Heo 2769112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2770112202d9STejun Heo pwq->flush_color = flush_color; 2771112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 277273f53c4aSTejun Heo wait = true; 27731da177e4SLinus Torvalds } 277473f53c4aSTejun Heo } 277573f53c4aSTejun Heo 277673f53c4aSTejun Heo if (work_color >= 0) { 27776183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2778112202d9STejun Heo pwq->work_color = work_color; 277973f53c4aSTejun Heo } 278073f53c4aSTejun Heo 2781a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27821da177e4SLinus Torvalds } 27831da177e4SLinus Torvalds 2784112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 278573f53c4aSTejun Heo complete(&wq->first_flusher->done); 278673f53c4aSTejun Heo 278773f53c4aSTejun Heo return wait; 278883c22520SOleg Nesterov } 27891da177e4SLinus Torvalds 27900fcb78c2SRolf Eike Beer /** 27911da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 27920fcb78c2SRolf Eike Beer * @wq: workqueue to flush 27931da177e4SLinus Torvalds * 2794c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2795c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 27961da177e4SLinus Torvalds */ 27977ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 27981da177e4SLinus Torvalds { 279973f53c4aSTejun Heo struct wq_flusher this_flusher = { 280073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 280173f53c4aSTejun Heo .flush_color = -1, 2802fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 280373f53c4aSTejun Heo }; 280473f53c4aSTejun Heo int next_color; 2805b1f4ec17SOleg Nesterov 28063347fa09STejun Heo if (WARN_ON(!wq_online)) 28073347fa09STejun Heo return; 28083347fa09STejun Heo 280987915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 281087915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 281187915adcSJohannes Berg 28123c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 281373f53c4aSTejun Heo 281473f53c4aSTejun Heo /* 281573f53c4aSTejun Heo * Start-to-wait phase 281673f53c4aSTejun Heo */ 281773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 281873f53c4aSTejun Heo 281973f53c4aSTejun Heo if (next_color != wq->flush_color) { 282073f53c4aSTejun Heo /* 282173f53c4aSTejun Heo * Color space is not full. The current work_color 282273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 282373f53c4aSTejun Heo * by one. 282473f53c4aSTejun Heo */ 28256183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 282673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 282773f53c4aSTejun Heo wq->work_color = next_color; 282873f53c4aSTejun Heo 282973f53c4aSTejun Heo if (!wq->first_flusher) { 283073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 28316183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 283273f53c4aSTejun Heo 283373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 283473f53c4aSTejun Heo 2835112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 283673f53c4aSTejun Heo wq->work_color)) { 283773f53c4aSTejun Heo /* nothing to flush, done */ 283873f53c4aSTejun Heo wq->flush_color = next_color; 283973f53c4aSTejun Heo wq->first_flusher = NULL; 284073f53c4aSTejun Heo goto out_unlock; 284173f53c4aSTejun Heo } 284273f53c4aSTejun Heo } else { 284373f53c4aSTejun Heo /* wait in queue */ 28446183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 284573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2846112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 284773f53c4aSTejun Heo } 284873f53c4aSTejun Heo } else { 284973f53c4aSTejun Heo /* 285073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 285173f53c4aSTejun Heo * The next flush completion will assign us 285273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 285373f53c4aSTejun Heo */ 285473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 285573f53c4aSTejun Heo } 285673f53c4aSTejun Heo 2857fca839c0STejun Heo check_flush_dependency(wq, NULL); 2858fca839c0STejun Heo 28593c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 286073f53c4aSTejun Heo 286173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 286273f53c4aSTejun Heo 286373f53c4aSTejun Heo /* 286473f53c4aSTejun Heo * Wake-up-and-cascade phase 286573f53c4aSTejun Heo * 286673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 286773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 286873f53c4aSTejun Heo */ 286900d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 287073f53c4aSTejun Heo return; 287173f53c4aSTejun Heo 28723c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 287373f53c4aSTejun Heo 28744ce48b37STejun Heo /* we might have raced, check again with mutex held */ 28754ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 28764ce48b37STejun Heo goto out_unlock; 28774ce48b37STejun Heo 287800d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 287973f53c4aSTejun Heo 28806183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 28816183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 288273f53c4aSTejun Heo 288373f53c4aSTejun Heo while (true) { 288473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 288573f53c4aSTejun Heo 288673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 288773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 288873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 288973f53c4aSTejun Heo break; 289073f53c4aSTejun Heo list_del_init(&next->list); 289173f53c4aSTejun Heo complete(&next->done); 289273f53c4aSTejun Heo } 289373f53c4aSTejun Heo 28946183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 289573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 289673f53c4aSTejun Heo 289773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 289873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 289973f53c4aSTejun Heo 290073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 290173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 290273f53c4aSTejun Heo /* 290373f53c4aSTejun Heo * Assign the same color to all overflowed 290473f53c4aSTejun Heo * flushers, advance work_color and append to 290573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 290673f53c4aSTejun Heo * phase for these overflowed flushers. 290773f53c4aSTejun Heo */ 290873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 290973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 291073f53c4aSTejun Heo 291173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 291273f53c4aSTejun Heo 291373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 291473f53c4aSTejun Heo &wq->flusher_queue); 2915112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 291673f53c4aSTejun Heo } 291773f53c4aSTejun Heo 291873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 29196183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 292073f53c4aSTejun Heo break; 292173f53c4aSTejun Heo } 292273f53c4aSTejun Heo 292373f53c4aSTejun Heo /* 292473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2925112202d9STejun Heo * the new first flusher and arm pwqs. 292673f53c4aSTejun Heo */ 29276183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 29286183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 292973f53c4aSTejun Heo 293073f53c4aSTejun Heo list_del_init(&next->list); 293173f53c4aSTejun Heo wq->first_flusher = next; 293273f53c4aSTejun Heo 2933112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 293473f53c4aSTejun Heo break; 293573f53c4aSTejun Heo 293673f53c4aSTejun Heo /* 293773f53c4aSTejun Heo * Meh... this color is already done, clear first 293873f53c4aSTejun Heo * flusher and repeat cascading. 293973f53c4aSTejun Heo */ 294073f53c4aSTejun Heo wq->first_flusher = NULL; 294173f53c4aSTejun Heo } 294273f53c4aSTejun Heo 294373f53c4aSTejun Heo out_unlock: 29443c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 29451da177e4SLinus Torvalds } 29461dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 29471da177e4SLinus Torvalds 29489c5a2ba7STejun Heo /** 29499c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 29509c5a2ba7STejun Heo * @wq: workqueue to drain 29519c5a2ba7STejun Heo * 29529c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 29539c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 29549c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2955b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 29569c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 29579c5a2ba7STejun Heo * takes too long. 29589c5a2ba7STejun Heo */ 29599c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 29609c5a2ba7STejun Heo { 29619c5a2ba7STejun Heo unsigned int flush_cnt = 0; 296249e3cf44STejun Heo struct pool_workqueue *pwq; 29639c5a2ba7STejun Heo 29649c5a2ba7STejun Heo /* 29659c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 29669c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2967618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 29689c5a2ba7STejun Heo */ 296987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 29709c5a2ba7STejun Heo if (!wq->nr_drainers++) 2971618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 297287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29739c5a2ba7STejun Heo reflush: 29749c5a2ba7STejun Heo flush_workqueue(wq); 29759c5a2ba7STejun Heo 2976b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 297776af4d93STejun Heo 297849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2979fa2563e4SThomas Tuttle bool drained; 29809c5a2ba7STejun Heo 2981a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 2982f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 2983a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 2984fa2563e4SThomas Tuttle 2985fa2563e4SThomas Tuttle if (drained) 29869c5a2ba7STejun Heo continue; 29879c5a2ba7STejun Heo 29889c5a2ba7STejun Heo if (++flush_cnt == 10 || 29899c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2990e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 2991e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 299276af4d93STejun Heo 2993b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 29949c5a2ba7STejun Heo goto reflush; 29959c5a2ba7STejun Heo } 29969c5a2ba7STejun Heo 29979c5a2ba7STejun Heo if (!--wq->nr_drainers) 2998618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 299987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 30009c5a2ba7STejun Heo } 30019c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 30029c5a2ba7STejun Heo 3003d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3004d6e89786SJohannes Berg bool from_cancel) 3005baf59022STejun Heo { 3006baf59022STejun Heo struct worker *worker = NULL; 3007c9e7cf27STejun Heo struct worker_pool *pool; 3008112202d9STejun Heo struct pool_workqueue *pwq; 3009baf59022STejun Heo 3010baf59022STejun Heo might_sleep(); 3011baf59022STejun Heo 301224acfb71SThomas Gleixner rcu_read_lock(); 3013fa1b54e6STejun Heo pool = get_work_pool(work); 3014fa1b54e6STejun Heo if (!pool) { 301524acfb71SThomas Gleixner rcu_read_unlock(); 3016fa1b54e6STejun Heo return false; 3017fa1b54e6STejun Heo } 3018fa1b54e6STejun Heo 3019a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 30200b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3021112202d9STejun Heo pwq = get_work_pwq(work); 3022112202d9STejun Heo if (pwq) { 3023112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3024baf59022STejun Heo goto already_gone; 3025606a5020STejun Heo } else { 3026c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3027baf59022STejun Heo if (!worker) 3028baf59022STejun Heo goto already_gone; 3029112202d9STejun Heo pwq = worker->current_pwq; 3030606a5020STejun Heo } 3031baf59022STejun Heo 3032fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3033fca839c0STejun Heo 3034112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3035a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3036baf59022STejun Heo 3037e159489bSTejun Heo /* 3038a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3039a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3040a1d14934SPeter Zijlstra * 3041a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3042a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3043a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3044a1d14934SPeter Zijlstra * forward progress. 3045e159489bSTejun Heo */ 3046d6e89786SJohannes Berg if (!from_cancel && 3047d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3048112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3049112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3050a1d14934SPeter Zijlstra } 305124acfb71SThomas Gleixner rcu_read_unlock(); 3052baf59022STejun Heo return true; 3053baf59022STejun Heo already_gone: 3054a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 305524acfb71SThomas Gleixner rcu_read_unlock(); 3056baf59022STejun Heo return false; 3057baf59022STejun Heo } 3058baf59022STejun Heo 3059d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3060d6e89786SJohannes Berg { 3061d6e89786SJohannes Berg struct wq_barrier barr; 3062d6e89786SJohannes Berg 3063d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3064d6e89786SJohannes Berg return false; 3065d6e89786SJohannes Berg 30664d43d395STetsuo Handa if (WARN_ON(!work->func)) 30674d43d395STetsuo Handa return false; 30684d43d395STetsuo Handa 306987915adcSJohannes Berg if (!from_cancel) { 307087915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 307187915adcSJohannes Berg lock_map_release(&work->lockdep_map); 307287915adcSJohannes Berg } 307387915adcSJohannes Berg 3074d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3075d6e89786SJohannes Berg wait_for_completion(&barr.done); 3076d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3077d6e89786SJohannes Berg return true; 3078d6e89786SJohannes Berg } else { 3079d6e89786SJohannes Berg return false; 3080d6e89786SJohannes Berg } 3081d6e89786SJohannes Berg } 3082d6e89786SJohannes Berg 3083db700897SOleg Nesterov /** 3084401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3085401a8d04STejun Heo * @work: the work to flush 3086db700897SOleg Nesterov * 3087606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3088606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3089401a8d04STejun Heo * 3090d185af30SYacine Belkadi * Return: 3091401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3092401a8d04STejun Heo * %false if it was already idle. 3093db700897SOleg Nesterov */ 3094401a8d04STejun Heo bool flush_work(struct work_struct *work) 3095db700897SOleg Nesterov { 3096d6e89786SJohannes Berg return __flush_work(work, false); 3097606a5020STejun Heo } 3098db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3099db700897SOleg Nesterov 31008603e1b3STejun Heo struct cwt_wait { 3101ac6424b9SIngo Molnar wait_queue_entry_t wait; 31028603e1b3STejun Heo struct work_struct *work; 31038603e1b3STejun Heo }; 31048603e1b3STejun Heo 3105ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 31068603e1b3STejun Heo { 31078603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 31088603e1b3STejun Heo 31098603e1b3STejun Heo if (cwait->work != key) 31108603e1b3STejun Heo return 0; 31118603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 31128603e1b3STejun Heo } 31138603e1b3STejun Heo 311436e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3115401a8d04STejun Heo { 31168603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3117bbb68dfaSTejun Heo unsigned long flags; 31181f1f642eSOleg Nesterov int ret; 31191f1f642eSOleg Nesterov 31201f1f642eSOleg Nesterov do { 3121bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3122bbb68dfaSTejun Heo /* 31238603e1b3STejun Heo * If someone else is already canceling, wait for it to 31248603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 31258603e1b3STejun Heo * because we may get scheduled between @work's completion 31268603e1b3STejun Heo * and the other canceling task resuming and clearing 31278603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 31288603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 31298603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 31308603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 31318603e1b3STejun Heo * we're hogging the CPU. 31328603e1b3STejun Heo * 31338603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 31348603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 31358603e1b3STejun Heo * wake function which matches @work along with exclusive 31368603e1b3STejun Heo * wait and wakeup. 3137bbb68dfaSTejun Heo */ 31388603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 31398603e1b3STejun Heo struct cwt_wait cwait; 31408603e1b3STejun Heo 31418603e1b3STejun Heo init_wait(&cwait.wait); 31428603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 31438603e1b3STejun Heo cwait.work = work; 31448603e1b3STejun Heo 31458603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 31468603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 31478603e1b3STejun Heo if (work_is_canceling(work)) 31488603e1b3STejun Heo schedule(); 31498603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 31508603e1b3STejun Heo } 31511f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 31521f1f642eSOleg Nesterov 3153bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3154bbb68dfaSTejun Heo mark_work_canceling(work); 3155bbb68dfaSTejun Heo local_irq_restore(flags); 3156bbb68dfaSTejun Heo 31573347fa09STejun Heo /* 31583347fa09STejun Heo * This allows canceling during early boot. We know that @work 31593347fa09STejun Heo * isn't executing. 31603347fa09STejun Heo */ 31613347fa09STejun Heo if (wq_online) 3162d6e89786SJohannes Berg __flush_work(work, true); 31633347fa09STejun Heo 31647a22ad75STejun Heo clear_work_data(work); 31658603e1b3STejun Heo 31668603e1b3STejun Heo /* 31678603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 31688603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 31698603e1b3STejun Heo * visible there. 31708603e1b3STejun Heo */ 31718603e1b3STejun Heo smp_mb(); 31728603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 31738603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 31748603e1b3STejun Heo 31751f1f642eSOleg Nesterov return ret; 31761f1f642eSOleg Nesterov } 31771f1f642eSOleg Nesterov 31786e84d644SOleg Nesterov /** 3179401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3180401a8d04STejun Heo * @work: the work to cancel 31816e84d644SOleg Nesterov * 3182401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3183401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3184401a8d04STejun Heo * another workqueue. On return from this function, @work is 3185401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 31861f1f642eSOleg Nesterov * 3187401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3188401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 31896e84d644SOleg Nesterov * 3190401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 31916e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3192401a8d04STejun Heo * 3193d185af30SYacine Belkadi * Return: 3194401a8d04STejun Heo * %true if @work was pending, %false otherwise. 31956e84d644SOleg Nesterov */ 3196401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 31976e84d644SOleg Nesterov { 319836e227d2STejun Heo return __cancel_work_timer(work, false); 3199b89deed3SOleg Nesterov } 320028e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3201b89deed3SOleg Nesterov 32026e84d644SOleg Nesterov /** 3203401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3204401a8d04STejun Heo * @dwork: the delayed work to flush 32056e84d644SOleg Nesterov * 3206401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3207401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3208401a8d04STejun Heo * considers the last queueing instance of @dwork. 32091f1f642eSOleg Nesterov * 3210d185af30SYacine Belkadi * Return: 3211401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3212401a8d04STejun Heo * %false if it was already idle. 32136e84d644SOleg Nesterov */ 3214401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3215401a8d04STejun Heo { 32168930cabaSTejun Heo local_irq_disable(); 3217401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 321860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 32198930cabaSTejun Heo local_irq_enable(); 3220401a8d04STejun Heo return flush_work(&dwork->work); 3221401a8d04STejun Heo } 3222401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3223401a8d04STejun Heo 322405f0fe6bSTejun Heo /** 322505f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 322605f0fe6bSTejun Heo * @rwork: the rcu work to flush 322705f0fe6bSTejun Heo * 322805f0fe6bSTejun Heo * Return: 322905f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 323005f0fe6bSTejun Heo * %false if it was already idle. 323105f0fe6bSTejun Heo */ 323205f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 323305f0fe6bSTejun Heo { 323405f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 323505f0fe6bSTejun Heo rcu_barrier(); 323605f0fe6bSTejun Heo flush_work(&rwork->work); 323705f0fe6bSTejun Heo return true; 323805f0fe6bSTejun Heo } else { 323905f0fe6bSTejun Heo return flush_work(&rwork->work); 324005f0fe6bSTejun Heo } 324105f0fe6bSTejun Heo } 324205f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 324305f0fe6bSTejun Heo 3244f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3245f72b8792SJens Axboe { 3246f72b8792SJens Axboe unsigned long flags; 3247f72b8792SJens Axboe int ret; 3248f72b8792SJens Axboe 3249f72b8792SJens Axboe do { 3250f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3251f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3252f72b8792SJens Axboe 3253f72b8792SJens Axboe if (unlikely(ret < 0)) 3254f72b8792SJens Axboe return false; 3255f72b8792SJens Axboe 3256f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3257f72b8792SJens Axboe local_irq_restore(flags); 3258f72b8792SJens Axboe return ret; 3259f72b8792SJens Axboe } 3260f72b8792SJens Axboe 3261401a8d04STejun Heo /** 326257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 326357b30ae7STejun Heo * @dwork: delayed_work to cancel 326409383498STejun Heo * 3265d185af30SYacine Belkadi * Kill off a pending delayed_work. 3266d185af30SYacine Belkadi * 3267d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3268d185af30SYacine Belkadi * pending. 3269d185af30SYacine Belkadi * 3270d185af30SYacine Belkadi * Note: 3271d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3272d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3273d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 327409383498STejun Heo * 327557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 327609383498STejun Heo */ 327757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 327809383498STejun Heo { 3279f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 328009383498STejun Heo } 328157b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 328209383498STejun Heo 328309383498STejun Heo /** 3284401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3285401a8d04STejun Heo * @dwork: the delayed work cancel 3286401a8d04STejun Heo * 3287401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3288401a8d04STejun Heo * 3289d185af30SYacine Belkadi * Return: 3290401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3291401a8d04STejun Heo */ 3292401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 32936e84d644SOleg Nesterov { 329436e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 32956e84d644SOleg Nesterov } 3296f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 32971da177e4SLinus Torvalds 32980fcb78c2SRolf Eike Beer /** 329931ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3300b6136773SAndrew Morton * @func: the function to call 3301b6136773SAndrew Morton * 330231ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 330331ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3304b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 330531ddd871STejun Heo * 3306d185af30SYacine Belkadi * Return: 330731ddd871STejun Heo * 0 on success, -errno on failure. 3308b6136773SAndrew Morton */ 330965f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 331015316ba8SChristoph Lameter { 331115316ba8SChristoph Lameter int cpu; 331238f51568SNamhyung Kim struct work_struct __percpu *works; 331315316ba8SChristoph Lameter 3314b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3315b6136773SAndrew Morton if (!works) 331615316ba8SChristoph Lameter return -ENOMEM; 3317b6136773SAndrew Morton 3318ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 331993981800STejun Heo 332015316ba8SChristoph Lameter for_each_online_cpu(cpu) { 33219bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 33229bfb1839SIngo Molnar 33239bfb1839SIngo Molnar INIT_WORK(work, func); 33248de6d308SOleg Nesterov schedule_work_on(cpu, work); 332515316ba8SChristoph Lameter } 332693981800STejun Heo 332793981800STejun Heo for_each_online_cpu(cpu) 33288616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 332993981800STejun Heo 3330ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3331b6136773SAndrew Morton free_percpu(works); 333215316ba8SChristoph Lameter return 0; 333315316ba8SChristoph Lameter } 333415316ba8SChristoph Lameter 3335eef6a7d5SAlan Stern /** 33361fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 33371fa44ecaSJames Bottomley * @fn: the function to execute 33381fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 33391fa44ecaSJames Bottomley * be available when the work executes) 33401fa44ecaSJames Bottomley * 33411fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 33421fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 33431fa44ecaSJames Bottomley * 3344d185af30SYacine Belkadi * Return: 0 - function was executed 33451fa44ecaSJames Bottomley * 1 - function was scheduled for execution 33461fa44ecaSJames Bottomley */ 334765f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 33481fa44ecaSJames Bottomley { 33491fa44ecaSJames Bottomley if (!in_interrupt()) { 335065f27f38SDavid Howells fn(&ew->work); 33511fa44ecaSJames Bottomley return 0; 33521fa44ecaSJames Bottomley } 33531fa44ecaSJames Bottomley 335465f27f38SDavid Howells INIT_WORK(&ew->work, fn); 33551fa44ecaSJames Bottomley schedule_work(&ew->work); 33561fa44ecaSJames Bottomley 33571fa44ecaSJames Bottomley return 1; 33581fa44ecaSJames Bottomley } 33591fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 33601fa44ecaSJames Bottomley 33617a4e344cSTejun Heo /** 33627a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33637a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33647a4e344cSTejun Heo * 33657a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33667a4e344cSTejun Heo */ 3367513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 33687a4e344cSTejun Heo { 33697a4e344cSTejun Heo if (attrs) { 33707a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33717a4e344cSTejun Heo kfree(attrs); 33727a4e344cSTejun Heo } 33737a4e344cSTejun Heo } 33747a4e344cSTejun Heo 33757a4e344cSTejun Heo /** 33767a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33777a4e344cSTejun Heo * 33787a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3379d185af30SYacine Belkadi * return it. 3380d185af30SYacine Belkadi * 3381d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33827a4e344cSTejun Heo */ 3383513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 33847a4e344cSTejun Heo { 33857a4e344cSTejun Heo struct workqueue_attrs *attrs; 33867a4e344cSTejun Heo 3387be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 33887a4e344cSTejun Heo if (!attrs) 33897a4e344cSTejun Heo goto fail; 3390be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 33917a4e344cSTejun Heo goto fail; 33927a4e344cSTejun Heo 339313e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33947a4e344cSTejun Heo return attrs; 33957a4e344cSTejun Heo fail: 33967a4e344cSTejun Heo free_workqueue_attrs(attrs); 33977a4e344cSTejun Heo return NULL; 33987a4e344cSTejun Heo } 33997a4e344cSTejun Heo 340029c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 340129c91e99STejun Heo const struct workqueue_attrs *from) 340229c91e99STejun Heo { 340329c91e99STejun Heo to->nice = from->nice; 340429c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 34052865a8fbSShaohua Li /* 34062865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 34072865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 34082865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 34092865a8fbSShaohua Li */ 34102865a8fbSShaohua Li to->no_numa = from->no_numa; 341129c91e99STejun Heo } 341229c91e99STejun Heo 341329c91e99STejun Heo /* hash value of the content of @attr */ 341429c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 341529c91e99STejun Heo { 341629c91e99STejun Heo u32 hash = 0; 341729c91e99STejun Heo 341829c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 341913e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 342013e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 342129c91e99STejun Heo return hash; 342229c91e99STejun Heo } 342329c91e99STejun Heo 342429c91e99STejun Heo /* content equality test */ 342529c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 342629c91e99STejun Heo const struct workqueue_attrs *b) 342729c91e99STejun Heo { 342829c91e99STejun Heo if (a->nice != b->nice) 342929c91e99STejun Heo return false; 343029c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 343129c91e99STejun Heo return false; 343229c91e99STejun Heo return true; 343329c91e99STejun Heo } 343429c91e99STejun Heo 34357a4e344cSTejun Heo /** 34367a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34377a4e344cSTejun Heo * @pool: worker_pool to initialize 34387a4e344cSTejun Heo * 3439402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3440d185af30SYacine Belkadi * 3441d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 344229c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 344329c91e99STejun Heo * on @pool safely to release it. 34447a4e344cSTejun Heo */ 34457a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34464e1a1f9aSTejun Heo { 3447a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 344829c91e99STejun Heo pool->id = -1; 344929c91e99STejun Heo pool->cpu = -1; 3450f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34514e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 345282607adcSTejun Heo pool->watchdog_ts = jiffies; 34534e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34544e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34554e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34564e1a1f9aSTejun Heo 345732a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 34584e1a1f9aSTejun Heo 345932a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 34604e1a1f9aSTejun Heo 3461da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34627a4e344cSTejun Heo 34637cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 346429c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 346529c91e99STejun Heo pool->refcnt = 1; 346629c91e99STejun Heo 346729c91e99STejun Heo /* shouldn't fail above this point */ 3468be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 34697a4e344cSTejun Heo if (!pool->attrs) 34707a4e344cSTejun Heo return -ENOMEM; 34717a4e344cSTejun Heo return 0; 34724e1a1f9aSTejun Heo } 34734e1a1f9aSTejun Heo 3474669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3475669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3476669de8bdSBart Van Assche { 3477669de8bdSBart Van Assche char *lock_name; 3478669de8bdSBart Van Assche 3479669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3480669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3481669de8bdSBart Van Assche if (!lock_name) 3482669de8bdSBart Van Assche lock_name = wq->name; 348369a106c0SQian Cai 348469a106c0SQian Cai wq->lock_name = lock_name; 3485669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3486669de8bdSBart Van Assche } 3487669de8bdSBart Van Assche 3488669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3489669de8bdSBart Van Assche { 3490669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3491669de8bdSBart Van Assche } 3492669de8bdSBart Van Assche 3493669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3494669de8bdSBart Van Assche { 3495669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3496669de8bdSBart Van Assche kfree(wq->lock_name); 3497669de8bdSBart Van Assche } 3498669de8bdSBart Van Assche #else 3499669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3500669de8bdSBart Van Assche { 3501669de8bdSBart Van Assche } 3502669de8bdSBart Van Assche 3503669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3504669de8bdSBart Van Assche { 3505669de8bdSBart Van Assche } 3506669de8bdSBart Van Assche 3507669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3508669de8bdSBart Van Assche { 3509669de8bdSBart Van Assche } 3510669de8bdSBart Van Assche #endif 3511669de8bdSBart Van Assche 3512e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3513e2dca7adSTejun Heo { 3514e2dca7adSTejun Heo struct workqueue_struct *wq = 3515e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3516e2dca7adSTejun Heo 3517669de8bdSBart Van Assche wq_free_lockdep(wq); 3518669de8bdSBart Van Assche 3519e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3520e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3521e2dca7adSTejun Heo else 3522e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3523e2dca7adSTejun Heo 3524e2dca7adSTejun Heo kfree(wq); 3525e2dca7adSTejun Heo } 3526e2dca7adSTejun Heo 352729c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 352829c91e99STejun Heo { 352929c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 353029c91e99STejun Heo 35317cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 353229c91e99STejun Heo free_workqueue_attrs(pool->attrs); 353329c91e99STejun Heo kfree(pool); 353429c91e99STejun Heo } 353529c91e99STejun Heo 3536d8bb65abSSebastian Andrzej Siewior /* This returns with the lock held on success (pool manager is inactive). */ 3537d8bb65abSSebastian Andrzej Siewior static bool wq_manager_inactive(struct worker_pool *pool) 3538d8bb65abSSebastian Andrzej Siewior { 3539a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 3540d8bb65abSSebastian Andrzej Siewior 3541d8bb65abSSebastian Andrzej Siewior if (pool->flags & POOL_MANAGER_ACTIVE) { 3542a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3543d8bb65abSSebastian Andrzej Siewior return false; 3544d8bb65abSSebastian Andrzej Siewior } 3545d8bb65abSSebastian Andrzej Siewior return true; 3546d8bb65abSSebastian Andrzej Siewior } 3547d8bb65abSSebastian Andrzej Siewior 354829c91e99STejun Heo /** 354929c91e99STejun Heo * put_unbound_pool - put a worker_pool 355029c91e99STejun Heo * @pool: worker_pool to put 355129c91e99STejun Heo * 355224acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3553c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3554c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3555c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3556a892caccSTejun Heo * 3557a892caccSTejun Heo * Should be called with wq_pool_mutex held. 355829c91e99STejun Heo */ 355929c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 356029c91e99STejun Heo { 356160f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 356229c91e99STejun Heo struct worker *worker; 356329c91e99STejun Heo 3564a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3565a892caccSTejun Heo 3566a892caccSTejun Heo if (--pool->refcnt) 356729c91e99STejun Heo return; 356829c91e99STejun Heo 356929c91e99STejun Heo /* sanity checks */ 357061d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3571a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 357229c91e99STejun Heo return; 357329c91e99STejun Heo 357429c91e99STejun Heo /* release id and unhash */ 357529c91e99STejun Heo if (pool->id >= 0) 357629c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 357729c91e99STejun Heo hash_del(&pool->hash_node); 357829c91e99STejun Heo 3579c5aa87bbSTejun Heo /* 3580692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3581692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3582692b4825STejun Heo * manager and @pool gets freed with the flag set. 3583d8bb65abSSebastian Andrzej Siewior * Because of how wq_manager_inactive() works, we will hold the 3584d8bb65abSSebastian Andrzej Siewior * spinlock after a successful wait. 3585c5aa87bbSTejun Heo */ 3586d8bb65abSSebastian Andrzej Siewior rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool), 3587d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3588692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3589692b4825STejun Heo 35901037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 359129c91e99STejun Heo destroy_worker(worker); 359229c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3593a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 359460f5a4bcSLai Jiangshan 35951258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3596da028469SLai Jiangshan if (!list_empty(&pool->workers)) 359760f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 35981258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 359960f5a4bcSLai Jiangshan 360060f5a4bcSLai Jiangshan if (pool->detach_completion) 360160f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 360260f5a4bcSLai Jiangshan 360329c91e99STejun Heo /* shut down the timers */ 360429c91e99STejun Heo del_timer_sync(&pool->idle_timer); 360529c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 360629c91e99STejun Heo 360724acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 360825b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 360929c91e99STejun Heo } 361029c91e99STejun Heo 361129c91e99STejun Heo /** 361229c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 361329c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 361429c91e99STejun Heo * 361529c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 361629c91e99STejun Heo * reference count and return it. If there already is a matching 361729c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3618d185af30SYacine Belkadi * create a new one. 3619a892caccSTejun Heo * 3620a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3621d185af30SYacine Belkadi * 3622d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3623d185af30SYacine Belkadi * On failure, %NULL. 362429c91e99STejun Heo */ 362529c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 362629c91e99STejun Heo { 362729c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 362829c91e99STejun Heo struct worker_pool *pool; 3629f3f90ad4STejun Heo int node; 3630e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 363129c91e99STejun Heo 3632a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 363329c91e99STejun Heo 363429c91e99STejun Heo /* do we already have a matching pool? */ 363529c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 363629c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 363729c91e99STejun Heo pool->refcnt++; 36383fb1823cSLai Jiangshan return pool; 363929c91e99STejun Heo } 364029c91e99STejun Heo } 364129c91e99STejun Heo 3642e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3643e2273584SXunlei Pang if (wq_numa_enabled) { 3644e2273584SXunlei Pang for_each_node(node) { 3645e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3646e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3647e2273584SXunlei Pang target_node = node; 3648e2273584SXunlei Pang break; 3649e2273584SXunlei Pang } 3650e2273584SXunlei Pang } 3651e2273584SXunlei Pang } 3652e2273584SXunlei Pang 365329c91e99STejun Heo /* nope, create a new one */ 3654e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 365529c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 365629c91e99STejun Heo goto fail; 365729c91e99STejun Heo 36588864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 365929c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3660e2273584SXunlei Pang pool->node = target_node; 366129c91e99STejun Heo 36622865a8fbSShaohua Li /* 36632865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36642865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36652865a8fbSShaohua Li */ 36662865a8fbSShaohua Li pool->attrs->no_numa = false; 36672865a8fbSShaohua Li 366829c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 366929c91e99STejun Heo goto fail; 367029c91e99STejun Heo 367129c91e99STejun Heo /* create and start the initial worker */ 36723347fa09STejun Heo if (wq_online && !create_worker(pool)) 367329c91e99STejun Heo goto fail; 367429c91e99STejun Heo 367529c91e99STejun Heo /* install */ 367629c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 36773fb1823cSLai Jiangshan 367829c91e99STejun Heo return pool; 367929c91e99STejun Heo fail: 368029c91e99STejun Heo if (pool) 368129c91e99STejun Heo put_unbound_pool(pool); 368229c91e99STejun Heo return NULL; 368329c91e99STejun Heo } 368429c91e99STejun Heo 36858864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36868864b4e5STejun Heo { 36878864b4e5STejun Heo kmem_cache_free(pwq_cache, 36888864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36898864b4e5STejun Heo } 36908864b4e5STejun Heo 36918864b4e5STejun Heo /* 36928864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36938864b4e5STejun Heo * and needs to be destroyed. 36948864b4e5STejun Heo */ 36958864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36968864b4e5STejun Heo { 36978864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36988864b4e5STejun Heo unbound_release_work); 36998864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 37008864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3701b42b0bddSYang Yingliang bool is_last = false; 37028864b4e5STejun Heo 3703b42b0bddSYang Yingliang /* 3704b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 3705b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3706b42b0bddSYang Yingliang */ 3707b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 37088864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 37098864b4e5STejun Heo return; 37108864b4e5STejun Heo 37113c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 37128864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3713bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 37143c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3715b42b0bddSYang Yingliang } 37168864b4e5STejun Heo 3717a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 37188864b4e5STejun Heo put_unbound_pool(pool); 3719a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3720a892caccSTejun Heo 372125b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 37228864b4e5STejun Heo 37238864b4e5STejun Heo /* 37248864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3725e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 37268864b4e5STejun Heo */ 3727669de8bdSBart Van Assche if (is_last) { 3728669de8bdSBart Van Assche wq_unregister_lockdep(wq); 372925b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 37306029a918STejun Heo } 3731669de8bdSBart Van Assche } 37328864b4e5STejun Heo 37330fbd95aaSTejun Heo /** 3734699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37350fbd95aaSTejun Heo * @pwq: target pool_workqueue 37360fbd95aaSTejun Heo * 3737699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3738f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 3739699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37400fbd95aaSTejun Heo */ 3741699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37420fbd95aaSTejun Heo { 3743699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3744699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 37453347fa09STejun Heo unsigned long flags; 3746699ce097STejun Heo 3747699ce097STejun Heo /* for @wq->saved_max_active */ 3748a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3749699ce097STejun Heo 3750699ce097STejun Heo /* fast exit for non-freezable wqs */ 3751699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3752699ce097STejun Heo return; 3753699ce097STejun Heo 37543347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3755a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3756699ce097STejun Heo 375774b414eaSLai Jiangshan /* 375874b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 375974b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 376074b414eaSLai Jiangshan * is updated and visible. 376174b414eaSLai Jiangshan */ 376274b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 376301341fbdSYunfeng Ye bool kick = false; 376401341fbdSYunfeng Ye 3765699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37660fbd95aaSTejun Heo 3767f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 376801341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 3769f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 377001341fbdSYunfeng Ye kick = true; 377101341fbdSYunfeng Ye } 3772951a078aSLai Jiangshan 3773951a078aSLai Jiangshan /* 3774951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 377501341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 377601341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 377701341fbdSYunfeng Ye * let's kick iff work items were activated. 3778951a078aSLai Jiangshan */ 377901341fbdSYunfeng Ye if (kick) 3780951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3781699ce097STejun Heo } else { 3782699ce097STejun Heo pwq->max_active = 0; 3783699ce097STejun Heo } 3784699ce097STejun Heo 3785a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 37860fbd95aaSTejun Heo } 37870fbd95aaSTejun Heo 378867dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 3789f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3790f147f29eSTejun Heo struct worker_pool *pool) 3791d2c1d404STejun Heo { 3792d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3793d2c1d404STejun Heo 3794e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3795e50aba9aSTejun Heo 3796d2c1d404STejun Heo pwq->pool = pool; 3797d2c1d404STejun Heo pwq->wq = wq; 3798d2c1d404STejun Heo pwq->flush_color = -1; 37998864b4e5STejun Heo pwq->refcnt = 1; 3800f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 38011befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3802d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 38038864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3804f147f29eSTejun Heo } 3805d2c1d404STejun Heo 3806f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 38071befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3808f147f29eSTejun Heo { 3809f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3810f147f29eSTejun Heo 3811f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 381275ccf595STejun Heo 38131befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 38141befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 38151befcf30STejun Heo return; 38161befcf30STejun Heo 381729b1cb41SLai Jiangshan /* set the matching work_color */ 381875ccf595STejun Heo pwq->work_color = wq->work_color; 3819983ca25eSTejun Heo 3820983ca25eSTejun Heo /* sync max_active to the current setting */ 3821983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3822983ca25eSTejun Heo 3823983ca25eSTejun Heo /* link in @pwq */ 38249e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3825df2d5ae4STejun Heo } 38266029a918STejun Heo 3827f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3828f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3829f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3830f147f29eSTejun Heo { 3831f147f29eSTejun Heo struct worker_pool *pool; 3832f147f29eSTejun Heo struct pool_workqueue *pwq; 3833f147f29eSTejun Heo 3834f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3835f147f29eSTejun Heo 3836f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3837f147f29eSTejun Heo if (!pool) 3838f147f29eSTejun Heo return NULL; 3839f147f29eSTejun Heo 3840e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3841f147f29eSTejun Heo if (!pwq) { 3842f147f29eSTejun Heo put_unbound_pool(pool); 3843f147f29eSTejun Heo return NULL; 3844f147f29eSTejun Heo } 3845f147f29eSTejun Heo 3846f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3847f147f29eSTejun Heo return pwq; 3848d2c1d404STejun Heo } 3849d2c1d404STejun Heo 38504c16bd32STejun Heo /** 385130186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3852042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 38534c16bd32STejun Heo * @node: the target NUMA node 38544c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38554c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38564c16bd32STejun Heo * 38574c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38584c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3859d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38604c16bd32STejun Heo * 38614c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38624c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38634c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38644c16bd32STejun Heo * @attrs->cpumask. 38654c16bd32STejun Heo * 38664c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38674c16bd32STejun Heo * stable. 3868d185af30SYacine Belkadi * 3869d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3870d185af30SYacine Belkadi * %false if equal. 38714c16bd32STejun Heo */ 38724c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38734c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38744c16bd32STejun Heo { 3875d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38764c16bd32STejun Heo goto use_dfl; 38774c16bd32STejun Heo 38784c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38794c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38804c16bd32STejun Heo if (cpu_going_down >= 0) 38814c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38824c16bd32STejun Heo 38834c16bd32STejun Heo if (cpumask_empty(cpumask)) 38844c16bd32STejun Heo goto use_dfl; 38854c16bd32STejun Heo 38864c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38874c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38881ad0f0a7SMichael Bringmann 38891ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 38901ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 38911ad0f0a7SMichael Bringmann "possible intersect\n"); 38921ad0f0a7SMichael Bringmann return false; 38931ad0f0a7SMichael Bringmann } 38941ad0f0a7SMichael Bringmann 38954c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 38964c16bd32STejun Heo 38974c16bd32STejun Heo use_dfl: 38984c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38994c16bd32STejun Heo return false; 39004c16bd32STejun Heo } 39014c16bd32STejun Heo 39021befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 39031befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 39041befcf30STejun Heo int node, 39051befcf30STejun Heo struct pool_workqueue *pwq) 39061befcf30STejun Heo { 39071befcf30STejun Heo struct pool_workqueue *old_pwq; 39081befcf30STejun Heo 39095b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39101befcf30STejun Heo lockdep_assert_held(&wq->mutex); 39111befcf30STejun Heo 39121befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 39131befcf30STejun Heo link_pwq(pwq); 39141befcf30STejun Heo 39151befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 39161befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 39171befcf30STejun Heo return old_pwq; 39181befcf30STejun Heo } 39191befcf30STejun Heo 39202d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 39212d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 39222d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 39232d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3924042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 39252d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 39262d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 39272d5f0764SLai Jiangshan }; 39282d5f0764SLai Jiangshan 39292d5f0764SLai Jiangshan /* free the resources after success or abort */ 39302d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 39312d5f0764SLai Jiangshan { 39322d5f0764SLai Jiangshan if (ctx) { 39332d5f0764SLai Jiangshan int node; 39342d5f0764SLai Jiangshan 39352d5f0764SLai Jiangshan for_each_node(node) 39362d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 39372d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 39382d5f0764SLai Jiangshan 39392d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 39402d5f0764SLai Jiangshan 39412d5f0764SLai Jiangshan kfree(ctx); 39422d5f0764SLai Jiangshan } 39432d5f0764SLai Jiangshan } 39442d5f0764SLai Jiangshan 39452d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 39462d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 39472d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 39482d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 39492d5f0764SLai Jiangshan { 39502d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 39512d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 39522d5f0764SLai Jiangshan int node; 39532d5f0764SLai Jiangshan 39542d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39552d5f0764SLai Jiangshan 3956acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 39572d5f0764SLai Jiangshan 3958be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 3959be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 39602d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 39612d5f0764SLai Jiangshan goto out_free; 39622d5f0764SLai Jiangshan 3963042f7df1SLai Jiangshan /* 3964042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3965042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3966042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3967042f7df1SLai Jiangshan */ 39682d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3969b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3970042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3971042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 39722d5f0764SLai Jiangshan 39732d5f0764SLai Jiangshan /* 39742d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 39752d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 39762d5f0764SLai Jiangshan * pools. 39772d5f0764SLai Jiangshan */ 39782d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 39792d5f0764SLai Jiangshan 39802d5f0764SLai Jiangshan /* 39812d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 39822d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 39832d5f0764SLai Jiangshan * it even if we don't use it immediately. 39842d5f0764SLai Jiangshan */ 39852d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39862d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 39872d5f0764SLai Jiangshan goto out_free; 39882d5f0764SLai Jiangshan 39892d5f0764SLai Jiangshan for_each_node(node) { 3990042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 39912d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39922d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 39932d5f0764SLai Jiangshan goto out_free; 39942d5f0764SLai Jiangshan } else { 39952d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 39962d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 39972d5f0764SLai Jiangshan } 39982d5f0764SLai Jiangshan } 39992d5f0764SLai Jiangshan 4000042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4001042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4002042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 40032d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4004042f7df1SLai Jiangshan 40052d5f0764SLai Jiangshan ctx->wq = wq; 40062d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40072d5f0764SLai Jiangshan return ctx; 40082d5f0764SLai Jiangshan 40092d5f0764SLai Jiangshan out_free: 40102d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40112d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 40122d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 40132d5f0764SLai Jiangshan return NULL; 40142d5f0764SLai Jiangshan } 40152d5f0764SLai Jiangshan 40162d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 40172d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 40182d5f0764SLai Jiangshan { 40192d5f0764SLai Jiangshan int node; 40202d5f0764SLai Jiangshan 40212d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 40222d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 40232d5f0764SLai Jiangshan 40242d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 40252d5f0764SLai Jiangshan 40262d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 40272d5f0764SLai Jiangshan for_each_node(node) 40282d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 40292d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 40302d5f0764SLai Jiangshan 40312d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 40322d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 40332d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 40342d5f0764SLai Jiangshan 40352d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 40362d5f0764SLai Jiangshan } 40372d5f0764SLai Jiangshan 4038a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4039a0111cf6SLai Jiangshan { 4040a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4041ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4042a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4043a0111cf6SLai Jiangshan } 4044a0111cf6SLai Jiangshan 4045a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4046a0111cf6SLai Jiangshan { 4047a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4048ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4049a0111cf6SLai Jiangshan } 4050a0111cf6SLai Jiangshan 4051a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4052a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4053a0111cf6SLai Jiangshan { 4054a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4055a0111cf6SLai Jiangshan 4056a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4057a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4058a0111cf6SLai Jiangshan return -EINVAL; 4059a0111cf6SLai Jiangshan 4060a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 40610a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 40620a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4063a0111cf6SLai Jiangshan return -EINVAL; 4064a0111cf6SLai Jiangshan 40650a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 40660a94efb5STejun Heo } 40670a94efb5STejun Heo 4068a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 40696201171eSwanghaibin if (!ctx) 40706201171eSwanghaibin return -ENOMEM; 4071a0111cf6SLai Jiangshan 4072a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4073a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4074a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4075a0111cf6SLai Jiangshan 40766201171eSwanghaibin return 0; 4077a0111cf6SLai Jiangshan } 4078a0111cf6SLai Jiangshan 40799e8cd2f5STejun Heo /** 40809e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 40819e8cd2f5STejun Heo * @wq: the target workqueue 40829e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 40839e8cd2f5STejun Heo * 40844c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 40854c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 40864c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 40874c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 40884c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 40894c16bd32STejun Heo * back-to-back will stay on its current pwq. 40909e8cd2f5STejun Heo * 4091d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4092d185af30SYacine Belkadi * 4093ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4094509b3204SDaniel Jordan * 4095d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 40969e8cd2f5STejun Heo */ 4097513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 40989e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 40999e8cd2f5STejun Heo { 4100a0111cf6SLai Jiangshan int ret; 41019e8cd2f5STejun Heo 4102509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4103509b3204SDaniel Jordan 4104509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4105a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4106509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 41072d5f0764SLai Jiangshan 41082d5f0764SLai Jiangshan return ret; 41099e8cd2f5STejun Heo } 41109e8cd2f5STejun Heo 41114c16bd32STejun Heo /** 41124c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 41134c16bd32STejun Heo * @wq: the target workqueue 41144c16bd32STejun Heo * @cpu: the CPU coming up or going down 41154c16bd32STejun Heo * @online: whether @cpu is coming up or going down 41164c16bd32STejun Heo * 41174c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 41184c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 41194c16bd32STejun Heo * @wq accordingly. 41204c16bd32STejun Heo * 41214c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 41224c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 41234c16bd32STejun Heo * correct. 41244c16bd32STejun Heo * 41254c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 41264c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 41274c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 41284c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 41294c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 41304c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 41314c16bd32STejun Heo * CPU_DOWN_PREPARE. 41324c16bd32STejun Heo */ 41334c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 41344c16bd32STejun Heo bool online) 41354c16bd32STejun Heo { 41364c16bd32STejun Heo int node = cpu_to_node(cpu); 41374c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 41384c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 41394c16bd32STejun Heo struct workqueue_attrs *target_attrs; 41404c16bd32STejun Heo cpumask_t *cpumask; 41414c16bd32STejun Heo 41424c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 41434c16bd32STejun Heo 4144f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4145f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 41464c16bd32STejun Heo return; 41474c16bd32STejun Heo 41484c16bd32STejun Heo /* 41494c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 41504c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 41514c16bd32STejun Heo * CPU hotplug exclusion. 41524c16bd32STejun Heo */ 41534c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 41544c16bd32STejun Heo cpumask = target_attrs->cpumask; 41554c16bd32STejun Heo 41564c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 41574c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 41584c16bd32STejun Heo 41594c16bd32STejun Heo /* 41604c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4161042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4162042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4163042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 41644c16bd32STejun Heo */ 4165042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 41664c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4167f7142ed4SLai Jiangshan return; 41684c16bd32STejun Heo } else { 41694c16bd32STejun Heo goto use_dfl_pwq; 41704c16bd32STejun Heo } 41714c16bd32STejun Heo 41724c16bd32STejun Heo /* create a new pwq */ 41734c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41744c16bd32STejun Heo if (!pwq) { 41752d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41764c16bd32STejun Heo wq->name); 417777f300b1SDaeseok Youn goto use_dfl_pwq; 41784c16bd32STejun Heo } 41794c16bd32STejun Heo 4180f7142ed4SLai Jiangshan /* Install the new pwq. */ 41814c16bd32STejun Heo mutex_lock(&wq->mutex); 41824c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41834c16bd32STejun Heo goto out_unlock; 41844c16bd32STejun Heo 41854c16bd32STejun Heo use_dfl_pwq: 4186f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4187a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 41884c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4189a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 41904c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 41914c16bd32STejun Heo out_unlock: 41924c16bd32STejun Heo mutex_unlock(&wq->mutex); 41934c16bd32STejun Heo put_pwq_unlocked(old_pwq); 41944c16bd32STejun Heo } 41954c16bd32STejun Heo 419630cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 41971da177e4SLinus Torvalds { 419849e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 41998a2b7538STejun Heo int cpu, ret; 4200e1d8aa9fSFrederic Weisbecker 420130cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4202420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4203420c0ddbSTejun Heo if (!wq->cpu_pwqs) 420430cdf249STejun Heo return -ENOMEM; 420530cdf249STejun Heo 420630cdf249STejun Heo for_each_possible_cpu(cpu) { 42077fb98ea7STejun Heo struct pool_workqueue *pwq = 42087fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 42097a62c2c8STejun Heo struct worker_pool *cpu_pools = 4210f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 421130cdf249STejun Heo 4212f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4213f147f29eSTejun Heo 4214f147f29eSTejun Heo mutex_lock(&wq->mutex); 42151befcf30STejun Heo link_pwq(pwq); 4216f147f29eSTejun Heo mutex_unlock(&wq->mutex); 421730cdf249STejun Heo } 421830cdf249STejun Heo return 0; 4219509b3204SDaniel Jordan } 4220509b3204SDaniel Jordan 4221ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4222509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 42238a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 42248a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 42258a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 42268a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 42278a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 42289e8cd2f5STejun Heo } else { 4229509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 42309e8cd2f5STejun Heo } 4231ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4232509b3204SDaniel Jordan 4233509b3204SDaniel Jordan return ret; 42340f900049STejun Heo } 42350f900049STejun Heo 4236f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4237f3421797STejun Heo const char *name) 4238b71ab8c2STejun Heo { 4239f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4240f3421797STejun Heo 4241f3421797STejun Heo if (max_active < 1 || max_active > lim) 4242044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4243f3421797STejun Heo max_active, name, 1, lim); 4244b71ab8c2STejun Heo 4245f3421797STejun Heo return clamp_val(max_active, 1, lim); 4246b71ab8c2STejun Heo } 4247b71ab8c2STejun Heo 4248983c7515STejun Heo /* 4249983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4250983c7515STejun Heo * to guarantee forward progress. 4251983c7515STejun Heo */ 4252983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4253983c7515STejun Heo { 4254983c7515STejun Heo struct worker *rescuer; 4255b92b36eaSDan Carpenter int ret; 4256983c7515STejun Heo 4257983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4258983c7515STejun Heo return 0; 4259983c7515STejun Heo 4260983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4261983c7515STejun Heo if (!rescuer) 4262983c7515STejun Heo return -ENOMEM; 4263983c7515STejun Heo 4264983c7515STejun Heo rescuer->rescue_wq = wq; 4265983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4266f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4267b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 4268983c7515STejun Heo kfree(rescuer); 4269b92b36eaSDan Carpenter return ret; 4270983c7515STejun Heo } 4271983c7515STejun Heo 4272983c7515STejun Heo wq->rescuer = rescuer; 4273983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4274983c7515STejun Heo wake_up_process(rescuer->task); 4275983c7515STejun Heo 4276983c7515STejun Heo return 0; 4277983c7515STejun Heo } 4278983c7515STejun Heo 4279a2775bbcSMathieu Malaterre __printf(1, 4) 4280669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 428197e37d7bSTejun Heo unsigned int flags, 4282669de8bdSBart Van Assche int max_active, ...) 42833af24433SOleg Nesterov { 4284df2d5ae4STejun Heo size_t tbl_size = 0; 4285ecf6881fSTejun Heo va_list args; 42863af24433SOleg Nesterov struct workqueue_struct *wq; 428749e3cf44STejun Heo struct pool_workqueue *pwq; 4288b196be89STejun Heo 42895c0338c6STejun Heo /* 42905c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 42915c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 42925c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 42935c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 42945c0338c6STejun Heo * on NUMA. 42955c0338c6STejun Heo */ 42965c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 42975c0338c6STejun Heo flags |= __WQ_ORDERED; 42985c0338c6STejun Heo 4299cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4300cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4301cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4302cee22a15SViresh Kumar 4303ecf6881fSTejun Heo /* allocate wq and format name */ 4304df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4305ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4306df2d5ae4STejun Heo 4307df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4308b196be89STejun Heo if (!wq) 4309d2c1d404STejun Heo return NULL; 4310b196be89STejun Heo 43116029a918STejun Heo if (flags & WQ_UNBOUND) { 4312be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 43136029a918STejun Heo if (!wq->unbound_attrs) 43146029a918STejun Heo goto err_free_wq; 43156029a918STejun Heo } 43166029a918STejun Heo 4317669de8bdSBart Van Assche va_start(args, max_active); 4318ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4319b196be89STejun Heo va_end(args); 43203af24433SOleg Nesterov 4321d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4322b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 43233af24433SOleg Nesterov 4324b196be89STejun Heo /* init wq */ 432597e37d7bSTejun Heo wq->flags = flags; 4326a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 43273c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4328112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 432930cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 433073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 433173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4332493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 43333af24433SOleg Nesterov 4334669de8bdSBart Van Assche wq_init_lockdep(wq); 4335cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 43363af24433SOleg Nesterov 433730cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 433882efcab3SBart Van Assche goto err_unreg_lockdep; 43391537663fSTejun Heo 434040c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4341d2c1d404STejun Heo goto err_destroy; 4342e22bee78STejun Heo 4343226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4344226223abSTejun Heo goto err_destroy; 4345226223abSTejun Heo 43466af8bf3dSOleg Nesterov /* 434768e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 434868e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 434968e13a67SLai Jiangshan * list. 43506af8bf3dSOleg Nesterov */ 435168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4352a0a1a5fdSTejun Heo 4353a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 435449e3cf44STejun Heo for_each_pwq(pwq, wq) 4355699ce097STejun Heo pwq_adjust_max_active(pwq); 4356a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4357a0a1a5fdSTejun Heo 4358e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4359a0a1a5fdSTejun Heo 436068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43613af24433SOleg Nesterov 43623af24433SOleg Nesterov return wq; 4363d2c1d404STejun Heo 436482efcab3SBart Van Assche err_unreg_lockdep: 4365009bb421SBart Van Assche wq_unregister_lockdep(wq); 4366009bb421SBart Van Assche wq_free_lockdep(wq); 436782efcab3SBart Van Assche err_free_wq: 43686029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 43694690c4abSTejun Heo kfree(wq); 4370d2c1d404STejun Heo return NULL; 4371d2c1d404STejun Heo err_destroy: 4372d2c1d404STejun Heo destroy_workqueue(wq); 43734690c4abSTejun Heo return NULL; 43741da177e4SLinus Torvalds } 4375669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 43761da177e4SLinus Torvalds 4377c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4378c29eb853STejun Heo { 4379c29eb853STejun Heo int i; 4380c29eb853STejun Heo 4381c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4382c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4383c29eb853STejun Heo return true; 4384c29eb853STejun Heo 4385c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4386c29eb853STejun Heo return true; 4387f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4388c29eb853STejun Heo return true; 4389c29eb853STejun Heo 4390c29eb853STejun Heo return false; 4391c29eb853STejun Heo } 4392c29eb853STejun Heo 43933af24433SOleg Nesterov /** 43943af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 43953af24433SOleg Nesterov * @wq: target workqueue 43963af24433SOleg Nesterov * 43973af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 43983af24433SOleg Nesterov */ 43993af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 44003af24433SOleg Nesterov { 440149e3cf44STejun Heo struct pool_workqueue *pwq; 44024c16bd32STejun Heo int node; 44033af24433SOleg Nesterov 4404def98c84STejun Heo /* 4405def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4406def98c84STejun Heo * lead to sysfs name conflicts. 4407def98c84STejun Heo */ 4408def98c84STejun Heo workqueue_sysfs_unregister(wq); 4409def98c84STejun Heo 44109c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 44119c5a2ba7STejun Heo drain_workqueue(wq); 4412c8efcc25STejun Heo 4413def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4414def98c84STejun Heo if (wq->rescuer) { 4415def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4416def98c84STejun Heo 4417def98c84STejun Heo /* this prevents new queueing */ 4418a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4419def98c84STejun Heo wq->rescuer = NULL; 4420a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4421def98c84STejun Heo 4422def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4423def98c84STejun Heo kthread_stop(rescuer->task); 44248efe1223STejun Heo kfree(rescuer); 4425def98c84STejun Heo } 4426def98c84STejun Heo 4427c29eb853STejun Heo /* 4428c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4429c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4430c29eb853STejun Heo */ 4431c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4432b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 443349e3cf44STejun Heo for_each_pwq(pwq, wq) { 4434a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4435c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 44361d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4437e66b39afSTejun Heo __func__, wq->name); 4438c29eb853STejun Heo show_pwq(pwq); 4439a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4440b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4441c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 444255df0933SImran Khan show_one_workqueue(wq); 44436183c009STejun Heo return; 444476af4d93STejun Heo } 4445a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 444676af4d93STejun Heo } 4447b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 44486183c009STejun Heo 4449a0a1a5fdSTejun Heo /* 4450a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4451a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4452a0a1a5fdSTejun Heo */ 4453e2dca7adSTejun Heo list_del_rcu(&wq->list); 445468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44553af24433SOleg Nesterov 44568864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4457669de8bdSBart Van Assche wq_unregister_lockdep(wq); 445829c91e99STejun Heo /* 44598864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4460e2dca7adSTejun Heo * schedule RCU free. 446129c91e99STejun Heo */ 446225b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 44638864b4e5STejun Heo } else { 44648864b4e5STejun Heo /* 44658864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 44664c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 44674c16bd32STejun Heo * @wq will be freed when the last pwq is released. 44688864b4e5STejun Heo */ 44694c16bd32STejun Heo for_each_node(node) { 44704c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 44714c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 44724c16bd32STejun Heo put_pwq_unlocked(pwq); 44734c16bd32STejun Heo } 44744c16bd32STejun Heo 44754c16bd32STejun Heo /* 44764c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 44774c16bd32STejun Heo * put. Don't access it afterwards. 44784c16bd32STejun Heo */ 44794c16bd32STejun Heo pwq = wq->dfl_pwq; 44804c16bd32STejun Heo wq->dfl_pwq = NULL; 4481dce90d47STejun Heo put_pwq_unlocked(pwq); 448229c91e99STejun Heo } 44833af24433SOleg Nesterov } 44843af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 44853af24433SOleg Nesterov 4486dcd989cbSTejun Heo /** 4487dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4488dcd989cbSTejun Heo * @wq: target workqueue 4489dcd989cbSTejun Heo * @max_active: new max_active value. 4490dcd989cbSTejun Heo * 4491dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4492dcd989cbSTejun Heo * 4493dcd989cbSTejun Heo * CONTEXT: 4494dcd989cbSTejun Heo * Don't call from IRQ context. 4495dcd989cbSTejun Heo */ 4496dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4497dcd989cbSTejun Heo { 449849e3cf44STejun Heo struct pool_workqueue *pwq; 4499dcd989cbSTejun Heo 45008719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 45010a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 45028719dceaSTejun Heo return; 45038719dceaSTejun Heo 4504f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4505dcd989cbSTejun Heo 4506a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4507dcd989cbSTejun Heo 45080a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4509dcd989cbSTejun Heo wq->saved_max_active = max_active; 4510dcd989cbSTejun Heo 4511699ce097STejun Heo for_each_pwq(pwq, wq) 4512699ce097STejun Heo pwq_adjust_max_active(pwq); 4513dcd989cbSTejun Heo 4514a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4515dcd989cbSTejun Heo } 4516dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4517dcd989cbSTejun Heo 4518dcd989cbSTejun Heo /** 451927d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 452027d4ee03SLukas Wunner * 452127d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 452227d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 452327d4ee03SLukas Wunner * 452427d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 452527d4ee03SLukas Wunner */ 452627d4ee03SLukas Wunner struct work_struct *current_work(void) 452727d4ee03SLukas Wunner { 452827d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 452927d4ee03SLukas Wunner 453027d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 453127d4ee03SLukas Wunner } 453227d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 453327d4ee03SLukas Wunner 453427d4ee03SLukas Wunner /** 4535e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4536e6267616STejun Heo * 4537e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4538e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4539d185af30SYacine Belkadi * 4540d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4541e6267616STejun Heo */ 4542e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4543e6267616STejun Heo { 4544e6267616STejun Heo struct worker *worker = current_wq_worker(); 4545e6267616STejun Heo 45466a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4547e6267616STejun Heo } 4548e6267616STejun Heo 4549e6267616STejun Heo /** 4550dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4551dcd989cbSTejun Heo * @cpu: CPU in question 4552dcd989cbSTejun Heo * @wq: target workqueue 4553dcd989cbSTejun Heo * 4554dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4555dcd989cbSTejun Heo * no synchronization around this function and the test result is 4556dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4557dcd989cbSTejun Heo * 4558d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4559d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4560d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4561d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4562d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4563d3251859STejun Heo * 4564d185af30SYacine Belkadi * Return: 4565dcd989cbSTejun Heo * %true if congested, %false otherwise. 4566dcd989cbSTejun Heo */ 4567d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4568dcd989cbSTejun Heo { 45697fb98ea7STejun Heo struct pool_workqueue *pwq; 457076af4d93STejun Heo bool ret; 457176af4d93STejun Heo 457224acfb71SThomas Gleixner rcu_read_lock(); 457324acfb71SThomas Gleixner preempt_disable(); 45747fb98ea7STejun Heo 4575d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4576d3251859STejun Heo cpu = smp_processor_id(); 4577d3251859STejun Heo 45787fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 45797fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 45807fb98ea7STejun Heo else 4581df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4582dcd989cbSTejun Heo 4583f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 458424acfb71SThomas Gleixner preempt_enable(); 458524acfb71SThomas Gleixner rcu_read_unlock(); 458676af4d93STejun Heo 458776af4d93STejun Heo return ret; 4588dcd989cbSTejun Heo } 4589dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4590dcd989cbSTejun Heo 4591dcd989cbSTejun Heo /** 4592dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4593dcd989cbSTejun Heo * @work: the work to be tested 4594dcd989cbSTejun Heo * 4595dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4596dcd989cbSTejun Heo * synchronization around this function and the test result is 4597dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4598dcd989cbSTejun Heo * 4599d185af30SYacine Belkadi * Return: 4600dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4601dcd989cbSTejun Heo */ 4602dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4603dcd989cbSTejun Heo { 4604fa1b54e6STejun Heo struct worker_pool *pool; 4605dcd989cbSTejun Heo unsigned long flags; 4606dcd989cbSTejun Heo unsigned int ret = 0; 4607dcd989cbSTejun Heo 4608dcd989cbSTejun Heo if (work_pending(work)) 4609dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4610038366c5SLai Jiangshan 461124acfb71SThomas Gleixner rcu_read_lock(); 4612fa1b54e6STejun Heo pool = get_work_pool(work); 4613038366c5SLai Jiangshan if (pool) { 4614a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4615c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4616dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4617a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4618038366c5SLai Jiangshan } 461924acfb71SThomas Gleixner rcu_read_unlock(); 4620dcd989cbSTejun Heo 4621dcd989cbSTejun Heo return ret; 4622dcd989cbSTejun Heo } 4623dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4624dcd989cbSTejun Heo 46253d1cb205STejun Heo /** 46263d1cb205STejun Heo * set_worker_desc - set description for the current work item 46273d1cb205STejun Heo * @fmt: printf-style format string 46283d1cb205STejun Heo * @...: arguments for the format string 46293d1cb205STejun Heo * 46303d1cb205STejun Heo * This function can be called by a running work function to describe what 46313d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 46323d1cb205STejun Heo * information will be printed out together to help debugging. The 46333d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 46343d1cb205STejun Heo */ 46353d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 46363d1cb205STejun Heo { 46373d1cb205STejun Heo struct worker *worker = current_wq_worker(); 46383d1cb205STejun Heo va_list args; 46393d1cb205STejun Heo 46403d1cb205STejun Heo if (worker) { 46413d1cb205STejun Heo va_start(args, fmt); 46423d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 46433d1cb205STejun Heo va_end(args); 46443d1cb205STejun Heo } 46453d1cb205STejun Heo } 46465c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 46473d1cb205STejun Heo 46483d1cb205STejun Heo /** 46493d1cb205STejun Heo * print_worker_info - print out worker information and description 46503d1cb205STejun Heo * @log_lvl: the log level to use when printing 46513d1cb205STejun Heo * @task: target task 46523d1cb205STejun Heo * 46533d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 46543d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 46553d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 46563d1cb205STejun Heo * 46573d1cb205STejun Heo * This function can be safely called on any task as long as the 46583d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 46593d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 46603d1cb205STejun Heo */ 46613d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 46623d1cb205STejun Heo { 46633d1cb205STejun Heo work_func_t *fn = NULL; 46643d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 46653d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 46663d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 46673d1cb205STejun Heo struct workqueue_struct *wq = NULL; 46683d1cb205STejun Heo struct worker *worker; 46693d1cb205STejun Heo 46703d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 46713d1cb205STejun Heo return; 46723d1cb205STejun Heo 46733d1cb205STejun Heo /* 46743d1cb205STejun Heo * This function is called without any synchronization and @task 46753d1cb205STejun Heo * could be in any state. Be careful with dereferences. 46763d1cb205STejun Heo */ 4677e700591aSPetr Mladek worker = kthread_probe_data(task); 46783d1cb205STejun Heo 46793d1cb205STejun Heo /* 46808bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 46818bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 46823d1cb205STejun Heo */ 4683fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4684fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4685fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4686fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4687fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 46883d1cb205STejun Heo 46893d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4690d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 46918bf89593STejun Heo if (strcmp(name, desc)) 46923d1cb205STejun Heo pr_cont(" (%s)", desc); 46933d1cb205STejun Heo pr_cont("\n"); 46943d1cb205STejun Heo } 46953d1cb205STejun Heo } 46963d1cb205STejun Heo 46973494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 46983494fc30STejun Heo { 46993494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 47003494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 47013494fc30STejun Heo pr_cont(" node=%d", pool->node); 47023494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 47033494fc30STejun Heo } 47043494fc30STejun Heo 47053494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 47063494fc30STejun Heo { 47073494fc30STejun Heo if (work->func == wq_barrier_func) { 47083494fc30STejun Heo struct wq_barrier *barr; 47093494fc30STejun Heo 47103494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 47113494fc30STejun Heo 47123494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 47133494fc30STejun Heo task_pid_nr(barr->task)); 47143494fc30STejun Heo } else { 4715d75f773cSSakari Ailus pr_cont("%s %ps", comma ? "," : "", work->func); 47163494fc30STejun Heo } 47173494fc30STejun Heo } 47183494fc30STejun Heo 47193494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 47203494fc30STejun Heo { 47213494fc30STejun Heo struct worker_pool *pool = pwq->pool; 47223494fc30STejun Heo struct work_struct *work; 47233494fc30STejun Heo struct worker *worker; 47243494fc30STejun Heo bool has_in_flight = false, has_pending = false; 47253494fc30STejun Heo int bkt; 47263494fc30STejun Heo 47273494fc30STejun Heo pr_info(" pwq %d:", pool->id); 47283494fc30STejun Heo pr_cont_pool_info(pool); 47293494fc30STejun Heo 4730e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4731e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 47323494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 47333494fc30STejun Heo 47343494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47353494fc30STejun Heo if (worker->current_pwq == pwq) { 47363494fc30STejun Heo has_in_flight = true; 47373494fc30STejun Heo break; 47383494fc30STejun Heo } 47393494fc30STejun Heo } 47403494fc30STejun Heo if (has_in_flight) { 47413494fc30STejun Heo bool comma = false; 47423494fc30STejun Heo 47433494fc30STejun Heo pr_info(" in-flight:"); 47443494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47453494fc30STejun Heo if (worker->current_pwq != pwq) 47463494fc30STejun Heo continue; 47473494fc30STejun Heo 4748d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 47493494fc30STejun Heo task_pid_nr(worker->task), 475030ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 47513494fc30STejun Heo worker->current_func); 47523494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 47533494fc30STejun Heo pr_cont_work(false, work); 47543494fc30STejun Heo comma = true; 47553494fc30STejun Heo } 47563494fc30STejun Heo pr_cont("\n"); 47573494fc30STejun Heo } 47583494fc30STejun Heo 47593494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47603494fc30STejun Heo if (get_work_pwq(work) == pwq) { 47613494fc30STejun Heo has_pending = true; 47623494fc30STejun Heo break; 47633494fc30STejun Heo } 47643494fc30STejun Heo } 47653494fc30STejun Heo if (has_pending) { 47663494fc30STejun Heo bool comma = false; 47673494fc30STejun Heo 47683494fc30STejun Heo pr_info(" pending:"); 47693494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47703494fc30STejun Heo if (get_work_pwq(work) != pwq) 47713494fc30STejun Heo continue; 47723494fc30STejun Heo 47733494fc30STejun Heo pr_cont_work(comma, work); 47743494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47753494fc30STejun Heo } 47763494fc30STejun Heo pr_cont("\n"); 47773494fc30STejun Heo } 47783494fc30STejun Heo 4779f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 47803494fc30STejun Heo bool comma = false; 47813494fc30STejun Heo 4782f97a4a1aSLai Jiangshan pr_info(" inactive:"); 4783f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 47843494fc30STejun Heo pr_cont_work(comma, work); 47853494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47863494fc30STejun Heo } 47873494fc30STejun Heo pr_cont("\n"); 47883494fc30STejun Heo } 47893494fc30STejun Heo } 47903494fc30STejun Heo 47913494fc30STejun Heo /** 479255df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 479355df0933SImran Khan * @wq: workqueue whose state will be printed 47943494fc30STejun Heo */ 479555df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 47963494fc30STejun Heo { 47973494fc30STejun Heo struct pool_workqueue *pwq; 47983494fc30STejun Heo bool idle = true; 479955df0933SImran Khan unsigned long flags; 48003494fc30STejun Heo 48013494fc30STejun Heo for_each_pwq(pwq, wq) { 4802f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 48033494fc30STejun Heo idle = false; 48043494fc30STejun Heo break; 48053494fc30STejun Heo } 48063494fc30STejun Heo } 480755df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 480855df0933SImran Khan return; 48093494fc30STejun Heo 48103494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 48113494fc30STejun Heo 48123494fc30STejun Heo for_each_pwq(pwq, wq) { 4813a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 481457116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 481557116ce1SJohan Hovold /* 481657116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 481757116ce1SJohan Hovold * drivers that queue work while holding locks 481857116ce1SJohan Hovold * also taken in their write paths. 481957116ce1SJohan Hovold */ 482057116ce1SJohan Hovold printk_deferred_enter(); 48213494fc30STejun Heo show_pwq(pwq); 482257116ce1SJohan Hovold printk_deferred_exit(); 482357116ce1SJohan Hovold } 4824a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 482562635ea8SSergey Senozhatsky /* 482662635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 482755df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 482862635ea8SSergey Senozhatsky * hard lockup. 482962635ea8SSergey Senozhatsky */ 483062635ea8SSergey Senozhatsky touch_nmi_watchdog(); 48313494fc30STejun Heo } 483255df0933SImran Khan 48333494fc30STejun Heo } 48343494fc30STejun Heo 483555df0933SImran Khan /** 483655df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 483755df0933SImran Khan * @pool: worker pool whose state will be printed 483855df0933SImran Khan */ 483955df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 484055df0933SImran Khan { 48413494fc30STejun Heo struct worker *worker; 48423494fc30STejun Heo bool first = true; 484355df0933SImran Khan unsigned long flags; 48443494fc30STejun Heo 4845a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 48463494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 48473494fc30STejun Heo goto next_pool; 484857116ce1SJohan Hovold /* 484957116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 485057116ce1SJohan Hovold * queue work while holding locks also taken in their write 485157116ce1SJohan Hovold * paths. 485257116ce1SJohan Hovold */ 485357116ce1SJohan Hovold printk_deferred_enter(); 48543494fc30STejun Heo pr_info("pool %d:", pool->id); 48553494fc30STejun Heo pr_cont_pool_info(pool); 485682607adcSTejun Heo pr_cont(" hung=%us workers=%d", 485782607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 485882607adcSTejun Heo pool->nr_workers); 48593494fc30STejun Heo if (pool->manager) 48603494fc30STejun Heo pr_cont(" manager: %d", 48613494fc30STejun Heo task_pid_nr(pool->manager->task)); 48623494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 48633494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 48643494fc30STejun Heo task_pid_nr(worker->task)); 48653494fc30STejun Heo first = false; 48663494fc30STejun Heo } 48673494fc30STejun Heo pr_cont("\n"); 486857116ce1SJohan Hovold printk_deferred_exit(); 48693494fc30STejun Heo next_pool: 4870a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 487162635ea8SSergey Senozhatsky /* 487262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 487355df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 487462635ea8SSergey Senozhatsky * hard lockup. 487562635ea8SSergey Senozhatsky */ 487662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 487755df0933SImran Khan 48783494fc30STejun Heo } 48793494fc30STejun Heo 488055df0933SImran Khan /** 488155df0933SImran Khan * show_all_workqueues - dump workqueue state 488255df0933SImran Khan * 488355df0933SImran Khan * Called from a sysrq handler or try_to_freeze_tasks() and prints out 488455df0933SImran Khan * all busy workqueues and pools. 488555df0933SImran Khan */ 488655df0933SImran Khan void show_all_workqueues(void) 488755df0933SImran Khan { 488855df0933SImran Khan struct workqueue_struct *wq; 488955df0933SImran Khan struct worker_pool *pool; 489055df0933SImran Khan int pi; 489155df0933SImran Khan 489255df0933SImran Khan rcu_read_lock(); 489355df0933SImran Khan 489455df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 489555df0933SImran Khan 489655df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 489755df0933SImran Khan show_one_workqueue(wq); 489855df0933SImran Khan 489955df0933SImran Khan for_each_pool(pool, pi) 490055df0933SImran Khan show_one_worker_pool(pool); 490155df0933SImran Khan 490224acfb71SThomas Gleixner rcu_read_unlock(); 49033494fc30STejun Heo } 49043494fc30STejun Heo 49056b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 49066b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 49076b59808bSTejun Heo { 49086b59808bSTejun Heo int off; 49096b59808bSTejun Heo 49106b59808bSTejun Heo /* always show the actual comm */ 49116b59808bSTejun Heo off = strscpy(buf, task->comm, size); 49126b59808bSTejun Heo if (off < 0) 49136b59808bSTejun Heo return; 49146b59808bSTejun Heo 4915197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 49166b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 49176b59808bSTejun Heo 4918197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4919197f6accSTejun Heo struct worker *worker = kthread_data(task); 4920197f6accSTejun Heo struct worker_pool *pool = worker->pool; 49216b59808bSTejun Heo 49226b59808bSTejun Heo if (pool) { 4923a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 49246b59808bSTejun Heo /* 4925197f6accSTejun Heo * ->desc tracks information (wq name or 4926197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4927197f6accSTejun Heo * current, prepend '+', otherwise '-'. 49286b59808bSTejun Heo */ 49296b59808bSTejun Heo if (worker->desc[0] != '\0') { 49306b59808bSTejun Heo if (worker->current_work) 49316b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 49326b59808bSTejun Heo worker->desc); 49336b59808bSTejun Heo else 49346b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 49356b59808bSTejun Heo worker->desc); 49366b59808bSTejun Heo } 4937a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 49386b59808bSTejun Heo } 4939197f6accSTejun Heo } 49406b59808bSTejun Heo 49416b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 49426b59808bSTejun Heo } 49436b59808bSTejun Heo 494466448bc2SMathieu Malaterre #ifdef CONFIG_SMP 494566448bc2SMathieu Malaterre 4946db7bccf4STejun Heo /* 4947db7bccf4STejun Heo * CPU hotplug. 4948db7bccf4STejun Heo * 4949e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4950112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4951706026c2STejun Heo * pool which make migrating pending and scheduled works very 4952e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 495394cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4954e22bee78STejun Heo * blocked draining impractical. 4955e22bee78STejun Heo * 495624647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4957628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4958628c78e7STejun Heo * cpu comes back online. 4959db7bccf4STejun Heo */ 4960db7bccf4STejun Heo 4961e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4962db7bccf4STejun Heo { 49634ce62e9eSTejun Heo struct worker_pool *pool; 4964db7bccf4STejun Heo struct worker *worker; 4965db7bccf4STejun Heo 4966f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49671258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 4968a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4969e22bee78STejun Heo 4970f2d5a0eeSTejun Heo /* 497192f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 497294cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 497311b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 4974b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 4975b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 4976b4ac9384SLai Jiangshan * is on the same cpu. 4977f2d5a0eeSTejun Heo */ 4978da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4979403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4980db7bccf4STejun Heo 498124647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4982f2d5a0eeSTejun Heo 4983e22bee78STejun Heo /* 4984989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 4985989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 4986989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 4987989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 4988989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 4989eb283428SLai Jiangshan * are served by workers tied to the pool. 4990e22bee78STejun Heo */ 4991e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4992eb283428SLai Jiangshan 4993eb283428SLai Jiangshan /* 4994eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4995eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4996eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4997eb283428SLai Jiangshan */ 4998eb283428SLai Jiangshan wake_up_worker(pool); 4999989442d7SLai Jiangshan 5000a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5001989442d7SLai Jiangshan 5002989442d7SLai Jiangshan for_each_pool_worker(worker, pool) { 5003989442d7SLai Jiangshan kthread_set_per_cpu(worker->task, -1); 5004989442d7SLai Jiangshan WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 5005989442d7SLai Jiangshan } 5006989442d7SLai Jiangshan 5007989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5008eb283428SLai Jiangshan } 5009db7bccf4STejun Heo } 5010db7bccf4STejun Heo 5011bd7c089eSTejun Heo /** 5012bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5013bd7c089eSTejun Heo * @pool: pool of interest 5014bd7c089eSTejun Heo * 5015a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5016bd7c089eSTejun Heo */ 5017bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5018bd7c089eSTejun Heo { 5019a9ab775bSTejun Heo struct worker *worker; 5020bd7c089eSTejun Heo 50211258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5022bd7c089eSTejun Heo 5023bd7c089eSTejun Heo /* 5024a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5025a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5026402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5027a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5028a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5029bd7c089eSTejun Heo */ 50305c25b5ffSPeter Zijlstra for_each_pool_worker(worker, pool) { 50315c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 5032a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5033a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 50345c25b5ffSPeter Zijlstra } 5035a9ab775bSTejun Heo 5036a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5037f7c17d26SWanpeng Li 50383de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5039a9ab775bSTejun Heo 5040da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5041a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5042a9ab775bSTejun Heo 5043a9ab775bSTejun Heo /* 5044a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5045a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5046a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5047a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5048a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5049a9ab775bSTejun Heo * concurrency management. Note that when or whether 5050a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5051a9ab775bSTejun Heo * 5052c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5053a9ab775bSTejun Heo * tested without holding any lock in 50546d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5055a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5056a9ab775bSTejun Heo * management operations. 5057bd7c089eSTejun Heo */ 5058a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5059a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5060a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5061c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5062bd7c089eSTejun Heo } 5063a9ab775bSTejun Heo 5064a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5065bd7c089eSTejun Heo } 5066bd7c089eSTejun Heo 50677dbc725eSTejun Heo /** 50687dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 50697dbc725eSTejun Heo * @pool: unbound pool of interest 50707dbc725eSTejun Heo * @cpu: the CPU which is coming up 50717dbc725eSTejun Heo * 50727dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 50737dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 50747dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 50757dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 50767dbc725eSTejun Heo */ 50777dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 50787dbc725eSTejun Heo { 50797dbc725eSTejun Heo static cpumask_t cpumask; 50807dbc725eSTejun Heo struct worker *worker; 50817dbc725eSTejun Heo 50821258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 50837dbc725eSTejun Heo 50847dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 50857dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 50867dbc725eSTejun Heo return; 50877dbc725eSTejun Heo 50887dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 50897dbc725eSTejun Heo 50907dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5091da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5092d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 50937dbc725eSTejun Heo } 50947dbc725eSTejun Heo 50957ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 50961da177e4SLinus Torvalds { 50974ce62e9eSTejun Heo struct worker_pool *pool; 50981da177e4SLinus Torvalds 5099f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 51003ce63377STejun Heo if (pool->nr_workers) 51013ce63377STejun Heo continue; 5102051e1850SLai Jiangshan if (!create_worker(pool)) 51037ee681b2SThomas Gleixner return -ENOMEM; 51043af24433SOleg Nesterov } 51057ee681b2SThomas Gleixner return 0; 51067ee681b2SThomas Gleixner } 51071da177e4SLinus Torvalds 51087ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 51097ee681b2SThomas Gleixner { 51107ee681b2SThomas Gleixner struct worker_pool *pool; 51117ee681b2SThomas Gleixner struct workqueue_struct *wq; 51127ee681b2SThomas Gleixner int pi; 51137ee681b2SThomas Gleixner 511468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 51157dbc725eSTejun Heo 51167dbc725eSTejun Heo for_each_pool(pool, pi) { 51171258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 511894cf58bbSTejun Heo 5119f05b558dSLai Jiangshan if (pool->cpu == cpu) 512094cf58bbSTejun Heo rebind_workers(pool); 5121f05b558dSLai Jiangshan else if (pool->cpu < 0) 51227dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 512394cf58bbSTejun Heo 51241258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 512594cf58bbSTejun Heo } 51267dbc725eSTejun Heo 51274c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51284c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51294c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 51304c16bd32STejun Heo 513168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 51327ee681b2SThomas Gleixner return 0; 513365758202STejun Heo } 513465758202STejun Heo 51357ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 513665758202STejun Heo { 51374c16bd32STejun Heo struct workqueue_struct *wq; 51388db25e78STejun Heo 51394c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5140e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5141e8b3f8dbSLai Jiangshan return -1; 5142e8b3f8dbSLai Jiangshan 5143e8b3f8dbSLai Jiangshan unbind_workers(cpu); 51444c16bd32STejun Heo 51454c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51464c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 51474c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51484c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 51494c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 51504c16bd32STejun Heo 51517ee681b2SThomas Gleixner return 0; 515265758202STejun Heo } 515365758202STejun Heo 51542d3854a3SRusty Russell struct work_for_cpu { 5155ed48ece2STejun Heo struct work_struct work; 51562d3854a3SRusty Russell long (*fn)(void *); 51572d3854a3SRusty Russell void *arg; 51582d3854a3SRusty Russell long ret; 51592d3854a3SRusty Russell }; 51602d3854a3SRusty Russell 5161ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 51622d3854a3SRusty Russell { 5163ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5164ed48ece2STejun Heo 51652d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 51662d3854a3SRusty Russell } 51672d3854a3SRusty Russell 51682d3854a3SRusty Russell /** 516922aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 51702d3854a3SRusty Russell * @cpu: the cpu to run on 51712d3854a3SRusty Russell * @fn: the function to run 51722d3854a3SRusty Russell * @arg: the function arg 51732d3854a3SRusty Russell * 517431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 51756b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5176d185af30SYacine Belkadi * 5177d185af30SYacine Belkadi * Return: The value @fn returns. 51782d3854a3SRusty Russell */ 5179d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 51802d3854a3SRusty Russell { 5181ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 51822d3854a3SRusty Russell 5183ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5184ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 518512997d1aSBjorn Helgaas flush_work(&wfc.work); 5186440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 51872d3854a3SRusty Russell return wfc.ret; 51882d3854a3SRusty Russell } 51892d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 51900e8d6a93SThomas Gleixner 51910e8d6a93SThomas Gleixner /** 51920e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 51930e8d6a93SThomas Gleixner * @cpu: the cpu to run on 51940e8d6a93SThomas Gleixner * @fn: the function to run 51950e8d6a93SThomas Gleixner * @arg: the function argument 51960e8d6a93SThomas Gleixner * 51970e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 51980e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 51990e8d6a93SThomas Gleixner * 52000e8d6a93SThomas Gleixner * Return: The value @fn returns. 52010e8d6a93SThomas Gleixner */ 52020e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 52030e8d6a93SThomas Gleixner { 52040e8d6a93SThomas Gleixner long ret = -ENODEV; 52050e8d6a93SThomas Gleixner 5206ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 52070e8d6a93SThomas Gleixner if (cpu_online(cpu)) 52080e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5209ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 52100e8d6a93SThomas Gleixner return ret; 52110e8d6a93SThomas Gleixner } 52120e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 52132d3854a3SRusty Russell #endif /* CONFIG_SMP */ 52142d3854a3SRusty Russell 5215a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5216e7577c50SRusty Russell 5217a0a1a5fdSTejun Heo /** 5218a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5219a0a1a5fdSTejun Heo * 522058a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5221f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5222706026c2STejun Heo * pool->worklist. 5223a0a1a5fdSTejun Heo * 5224a0a1a5fdSTejun Heo * CONTEXT: 5225a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5226a0a1a5fdSTejun Heo */ 5227a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5228a0a1a5fdSTejun Heo { 522924b8a847STejun Heo struct workqueue_struct *wq; 523024b8a847STejun Heo struct pool_workqueue *pwq; 5231a0a1a5fdSTejun Heo 523268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5233a0a1a5fdSTejun Heo 52346183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5235a0a1a5fdSTejun Heo workqueue_freezing = true; 5236a0a1a5fdSTejun Heo 523724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5238a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5239699ce097STejun Heo for_each_pwq(pwq, wq) 5240699ce097STejun Heo pwq_adjust_max_active(pwq); 5241a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5242a1056305STejun Heo } 52435bcab335STejun Heo 524468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5245a0a1a5fdSTejun Heo } 5246a0a1a5fdSTejun Heo 5247a0a1a5fdSTejun Heo /** 524858a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5249a0a1a5fdSTejun Heo * 5250a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5251a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5252a0a1a5fdSTejun Heo * 5253a0a1a5fdSTejun Heo * CONTEXT: 525468e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5255a0a1a5fdSTejun Heo * 5256d185af30SYacine Belkadi * Return: 525758a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 525858a69cb4STejun Heo * is complete. 5259a0a1a5fdSTejun Heo */ 5260a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5261a0a1a5fdSTejun Heo { 5262a0a1a5fdSTejun Heo bool busy = false; 526324b8a847STejun Heo struct workqueue_struct *wq; 526424b8a847STejun Heo struct pool_workqueue *pwq; 5265a0a1a5fdSTejun Heo 526668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5267a0a1a5fdSTejun Heo 52686183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5269a0a1a5fdSTejun Heo 527024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 527124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 527224b8a847STejun Heo continue; 5273a0a1a5fdSTejun Heo /* 5274a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5275a0a1a5fdSTejun Heo * to peek without lock. 5276a0a1a5fdSTejun Heo */ 527724acfb71SThomas Gleixner rcu_read_lock(); 527824b8a847STejun Heo for_each_pwq(pwq, wq) { 52796183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5280112202d9STejun Heo if (pwq->nr_active) { 5281a0a1a5fdSTejun Heo busy = true; 528224acfb71SThomas Gleixner rcu_read_unlock(); 5283a0a1a5fdSTejun Heo goto out_unlock; 5284a0a1a5fdSTejun Heo } 5285a0a1a5fdSTejun Heo } 528624acfb71SThomas Gleixner rcu_read_unlock(); 5287a0a1a5fdSTejun Heo } 5288a0a1a5fdSTejun Heo out_unlock: 528968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5290a0a1a5fdSTejun Heo return busy; 5291a0a1a5fdSTejun Heo } 5292a0a1a5fdSTejun Heo 5293a0a1a5fdSTejun Heo /** 5294a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5295a0a1a5fdSTejun Heo * 5296a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5297706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5298a0a1a5fdSTejun Heo * 5299a0a1a5fdSTejun Heo * CONTEXT: 5300a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5301a0a1a5fdSTejun Heo */ 5302a0a1a5fdSTejun Heo void thaw_workqueues(void) 5303a0a1a5fdSTejun Heo { 530424b8a847STejun Heo struct workqueue_struct *wq; 530524b8a847STejun Heo struct pool_workqueue *pwq; 5306a0a1a5fdSTejun Heo 530768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5308a0a1a5fdSTejun Heo 5309a0a1a5fdSTejun Heo if (!workqueue_freezing) 5310a0a1a5fdSTejun Heo goto out_unlock; 5311a0a1a5fdSTejun Heo 531274b414eaSLai Jiangshan workqueue_freezing = false; 531324b8a847STejun Heo 531424b8a847STejun Heo /* restore max_active and repopulate worklist */ 531524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5316a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5317699ce097STejun Heo for_each_pwq(pwq, wq) 5318699ce097STejun Heo pwq_adjust_max_active(pwq); 5319a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 532024b8a847STejun Heo } 532124b8a847STejun Heo 5322a0a1a5fdSTejun Heo out_unlock: 532368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5324a0a1a5fdSTejun Heo } 5325a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5326a0a1a5fdSTejun Heo 5327042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5328042f7df1SLai Jiangshan { 5329042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5330042f7df1SLai Jiangshan int ret = 0; 5331042f7df1SLai Jiangshan struct workqueue_struct *wq; 5332042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5333042f7df1SLai Jiangshan 5334042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5335042f7df1SLai Jiangshan 5336042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5337042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5338042f7df1SLai Jiangshan continue; 5339042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5340042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5341042f7df1SLai Jiangshan continue; 5342042f7df1SLai Jiangshan 5343042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5344042f7df1SLai Jiangshan if (!ctx) { 5345042f7df1SLai Jiangshan ret = -ENOMEM; 5346042f7df1SLai Jiangshan break; 5347042f7df1SLai Jiangshan } 5348042f7df1SLai Jiangshan 5349042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5350042f7df1SLai Jiangshan } 5351042f7df1SLai Jiangshan 5352042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5353042f7df1SLai Jiangshan if (!ret) 5354042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5355042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5356042f7df1SLai Jiangshan } 5357042f7df1SLai Jiangshan 5358042f7df1SLai Jiangshan return ret; 5359042f7df1SLai Jiangshan } 5360042f7df1SLai Jiangshan 5361042f7df1SLai Jiangshan /** 5362042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5363042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5364042f7df1SLai Jiangshan * 5365042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5366042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5367042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5368042f7df1SLai Jiangshan * 536967dc8325SCai Huoqing * Return: 0 - Success 5370042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5371042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5372042f7df1SLai Jiangshan */ 5373042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5374042f7df1SLai Jiangshan { 5375042f7df1SLai Jiangshan int ret = -EINVAL; 5376042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5377042f7df1SLai Jiangshan 5378c98a9805STal Shorer /* 5379c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5380c98a9805STal Shorer * If the user wishes to include them, we allow that. 5381c98a9805STal Shorer */ 5382042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5383042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5384a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5385d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5386d25302e4SMenglong Dong ret = 0; 5387d25302e4SMenglong Dong goto out_unlock; 5388d25302e4SMenglong Dong } 5389d25302e4SMenglong Dong 5390d25302e4SMenglong Dong if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) { 5391d25302e4SMenglong Dong ret = -ENOMEM; 5392d25302e4SMenglong Dong goto out_unlock; 5393d25302e4SMenglong Dong } 5394042f7df1SLai Jiangshan 5395042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5396042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5397042f7df1SLai Jiangshan 5398042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5399042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5400042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5401042f7df1SLai Jiangshan 5402042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5403042f7df1SLai Jiangshan if (ret < 0) 5404042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5405042f7df1SLai Jiangshan 5406d25302e4SMenglong Dong free_cpumask_var(saved_cpumask); 5407d25302e4SMenglong Dong out_unlock: 5408a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5409042f7df1SLai Jiangshan } 5410042f7df1SLai Jiangshan 5411042f7df1SLai Jiangshan return ret; 5412042f7df1SLai Jiangshan } 5413042f7df1SLai Jiangshan 54146ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 54156ba94429SFrederic Weisbecker /* 54166ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 54176ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 54186ba94429SFrederic Weisbecker * following attributes. 54196ba94429SFrederic Weisbecker * 54206ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 54216ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 54226ba94429SFrederic Weisbecker * 54236ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 54246ba94429SFrederic Weisbecker * 54259a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 54266ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 54276ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 54289a19b463SWang Long * numa RW bool : whether enable NUMA affinity 54296ba94429SFrederic Weisbecker */ 54306ba94429SFrederic Weisbecker struct wq_device { 54316ba94429SFrederic Weisbecker struct workqueue_struct *wq; 54326ba94429SFrederic Weisbecker struct device dev; 54336ba94429SFrederic Weisbecker }; 54346ba94429SFrederic Weisbecker 54356ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 54366ba94429SFrederic Weisbecker { 54376ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 54386ba94429SFrederic Weisbecker 54396ba94429SFrederic Weisbecker return wq_dev->wq; 54406ba94429SFrederic Weisbecker } 54416ba94429SFrederic Weisbecker 54426ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 54436ba94429SFrederic Weisbecker char *buf) 54446ba94429SFrederic Weisbecker { 54456ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54466ba94429SFrederic Weisbecker 54476ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 54486ba94429SFrederic Weisbecker } 54496ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 54506ba94429SFrederic Weisbecker 54516ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 54526ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54536ba94429SFrederic Weisbecker { 54546ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54556ba94429SFrederic Weisbecker 54566ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 54576ba94429SFrederic Weisbecker } 54586ba94429SFrederic Weisbecker 54596ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 54606ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 54616ba94429SFrederic Weisbecker size_t count) 54626ba94429SFrederic Weisbecker { 54636ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54646ba94429SFrederic Weisbecker int val; 54656ba94429SFrederic Weisbecker 54666ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 54676ba94429SFrederic Weisbecker return -EINVAL; 54686ba94429SFrederic Weisbecker 54696ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 54706ba94429SFrederic Weisbecker return count; 54716ba94429SFrederic Weisbecker } 54726ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 54736ba94429SFrederic Weisbecker 54746ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 54756ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 54766ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 54776ba94429SFrederic Weisbecker NULL, 54786ba94429SFrederic Weisbecker }; 54796ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 54806ba94429SFrederic Weisbecker 54816ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_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 const char *delim = ""; 54866ba94429SFrederic Weisbecker int node, written = 0; 54876ba94429SFrederic Weisbecker 5488ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 548924acfb71SThomas Gleixner rcu_read_lock(); 54906ba94429SFrederic Weisbecker for_each_node(node) { 54916ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 54926ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 54936ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 54946ba94429SFrederic Weisbecker delim = " "; 54956ba94429SFrederic Weisbecker } 54966ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 549724acfb71SThomas Gleixner rcu_read_unlock(); 5498ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 54996ba94429SFrederic Weisbecker 55006ba94429SFrederic Weisbecker return written; 55016ba94429SFrederic Weisbecker } 55026ba94429SFrederic Weisbecker 55036ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 55046ba94429SFrederic Weisbecker char *buf) 55056ba94429SFrederic Weisbecker { 55066ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55076ba94429SFrederic Weisbecker int written; 55086ba94429SFrederic Weisbecker 55096ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55106ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 55116ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55126ba94429SFrederic Weisbecker 55136ba94429SFrederic Weisbecker return written; 55146ba94429SFrederic Weisbecker } 55156ba94429SFrederic Weisbecker 55166ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 55176ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 55186ba94429SFrederic Weisbecker { 55196ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 55206ba94429SFrederic Weisbecker 5521899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5522899a94feSLai Jiangshan 5523be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 55246ba94429SFrederic Weisbecker if (!attrs) 55256ba94429SFrederic Weisbecker return NULL; 55266ba94429SFrederic Weisbecker 55276ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 55286ba94429SFrederic Weisbecker return attrs; 55296ba94429SFrederic Weisbecker } 55306ba94429SFrederic Weisbecker 55316ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 55326ba94429SFrederic Weisbecker const char *buf, size_t count) 55336ba94429SFrederic Weisbecker { 55346ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55356ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5536d4d3e257SLai Jiangshan int ret = -ENOMEM; 5537d4d3e257SLai Jiangshan 5538d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55396ba94429SFrederic Weisbecker 55406ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55416ba94429SFrederic Weisbecker if (!attrs) 5542d4d3e257SLai Jiangshan goto out_unlock; 55436ba94429SFrederic Weisbecker 55446ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 55456ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5546d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55476ba94429SFrederic Weisbecker else 55486ba94429SFrederic Weisbecker ret = -EINVAL; 55496ba94429SFrederic Weisbecker 5550d4d3e257SLai Jiangshan out_unlock: 5551d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55526ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55536ba94429SFrederic Weisbecker return ret ?: count; 55546ba94429SFrederic Weisbecker } 55556ba94429SFrederic Weisbecker 55566ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 55576ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 55586ba94429SFrederic Weisbecker { 55596ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55606ba94429SFrederic Weisbecker int written; 55616ba94429SFrederic Weisbecker 55626ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55636ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 55646ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 55656ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55666ba94429SFrederic Weisbecker return written; 55676ba94429SFrederic Weisbecker } 55686ba94429SFrederic Weisbecker 55696ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 55706ba94429SFrederic Weisbecker struct device_attribute *attr, 55716ba94429SFrederic Weisbecker const char *buf, size_t count) 55726ba94429SFrederic Weisbecker { 55736ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55746ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5575d4d3e257SLai Jiangshan int ret = -ENOMEM; 5576d4d3e257SLai Jiangshan 5577d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55786ba94429SFrederic Weisbecker 55796ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55806ba94429SFrederic Weisbecker if (!attrs) 5581d4d3e257SLai Jiangshan goto out_unlock; 55826ba94429SFrederic Weisbecker 55836ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 55846ba94429SFrederic Weisbecker if (!ret) 5585d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55866ba94429SFrederic Weisbecker 5587d4d3e257SLai Jiangshan out_unlock: 5588d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55896ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55906ba94429SFrederic Weisbecker return ret ?: count; 55916ba94429SFrederic Weisbecker } 55926ba94429SFrederic Weisbecker 55936ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 55946ba94429SFrederic Weisbecker char *buf) 55956ba94429SFrederic Weisbecker { 55966ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55976ba94429SFrederic Weisbecker int written; 55986ba94429SFrederic Weisbecker 55996ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 56006ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 56016ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 56026ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 56036ba94429SFrederic Weisbecker 56046ba94429SFrederic Weisbecker return written; 56056ba94429SFrederic Weisbecker } 56066ba94429SFrederic Weisbecker 56076ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 56086ba94429SFrederic Weisbecker const char *buf, size_t count) 56096ba94429SFrederic Weisbecker { 56106ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56116ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5612d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5613d4d3e257SLai Jiangshan 5614d4d3e257SLai Jiangshan apply_wqattrs_lock(); 56156ba94429SFrederic Weisbecker 56166ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 56176ba94429SFrederic Weisbecker if (!attrs) 5618d4d3e257SLai Jiangshan goto out_unlock; 56196ba94429SFrederic Weisbecker 56206ba94429SFrederic Weisbecker ret = -EINVAL; 56216ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 56226ba94429SFrederic Weisbecker attrs->no_numa = !v; 5623d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56246ba94429SFrederic Weisbecker } 56256ba94429SFrederic Weisbecker 5626d4d3e257SLai Jiangshan out_unlock: 5627d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 56286ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56296ba94429SFrederic Weisbecker return ret ?: count; 56306ba94429SFrederic Weisbecker } 56316ba94429SFrederic Weisbecker 56326ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 56336ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 56346ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 56356ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 56366ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 56376ba94429SFrederic Weisbecker __ATTR_NULL, 56386ba94429SFrederic Weisbecker }; 56396ba94429SFrederic Weisbecker 56406ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 56416ba94429SFrederic Weisbecker .name = "workqueue", 56426ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 56436ba94429SFrederic Weisbecker }; 56446ba94429SFrederic Weisbecker 5645b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5646b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5647b05a7928SFrederic Weisbecker { 5648b05a7928SFrederic Weisbecker int written; 5649b05a7928SFrederic Weisbecker 5650042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5651b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5652b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5653042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5654b05a7928SFrederic Weisbecker 5655b05a7928SFrederic Weisbecker return written; 5656b05a7928SFrederic Weisbecker } 5657b05a7928SFrederic Weisbecker 5658042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5659042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5660042f7df1SLai Jiangshan { 5661042f7df1SLai Jiangshan cpumask_var_t cpumask; 5662042f7df1SLai Jiangshan int ret; 5663042f7df1SLai Jiangshan 5664042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5665042f7df1SLai Jiangshan return -ENOMEM; 5666042f7df1SLai Jiangshan 5667042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5668042f7df1SLai Jiangshan if (!ret) 5669042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5670042f7df1SLai Jiangshan 5671042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5672042f7df1SLai Jiangshan return ret ? ret : count; 5673042f7df1SLai Jiangshan } 5674042f7df1SLai Jiangshan 5675b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5676042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5677042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5678b05a7928SFrederic Weisbecker 56796ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 56806ba94429SFrederic Weisbecker { 5681b05a7928SFrederic Weisbecker int err; 5682b05a7928SFrederic Weisbecker 5683b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5684b05a7928SFrederic Weisbecker if (err) 5685b05a7928SFrederic Weisbecker return err; 5686b05a7928SFrederic Weisbecker 5687b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 56886ba94429SFrederic Weisbecker } 56896ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 56906ba94429SFrederic Weisbecker 56916ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 56926ba94429SFrederic Weisbecker { 56936ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 56946ba94429SFrederic Weisbecker 56956ba94429SFrederic Weisbecker kfree(wq_dev); 56966ba94429SFrederic Weisbecker } 56976ba94429SFrederic Weisbecker 56986ba94429SFrederic Weisbecker /** 56996ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 57006ba94429SFrederic Weisbecker * @wq: the workqueue to register 57016ba94429SFrederic Weisbecker * 57026ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 57036ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 57046ba94429SFrederic Weisbecker * which is the preferred method. 57056ba94429SFrederic Weisbecker * 57066ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 57076ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 57086ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 57096ba94429SFrederic Weisbecker * attributes. 57106ba94429SFrederic Weisbecker * 57116ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 57126ba94429SFrederic Weisbecker */ 57136ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 57146ba94429SFrederic Weisbecker { 57156ba94429SFrederic Weisbecker struct wq_device *wq_dev; 57166ba94429SFrederic Weisbecker int ret; 57176ba94429SFrederic Weisbecker 57186ba94429SFrederic Weisbecker /* 5719402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 57206ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 57216ba94429SFrederic Weisbecker * workqueues. 57226ba94429SFrederic Weisbecker */ 57230a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 57246ba94429SFrederic Weisbecker return -EINVAL; 57256ba94429SFrederic Weisbecker 57266ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 57276ba94429SFrederic Weisbecker if (!wq_dev) 57286ba94429SFrederic Weisbecker return -ENOMEM; 57296ba94429SFrederic Weisbecker 57306ba94429SFrederic Weisbecker wq_dev->wq = wq; 57316ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 57326ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 573323217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 57346ba94429SFrederic Weisbecker 57356ba94429SFrederic Weisbecker /* 57366ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 57376ba94429SFrederic Weisbecker * everything is ready. 57386ba94429SFrederic Weisbecker */ 57396ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 57406ba94429SFrederic Weisbecker 57416ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 57426ba94429SFrederic Weisbecker if (ret) { 5743537f4146SArvind Yadav put_device(&wq_dev->dev); 57446ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57456ba94429SFrederic Weisbecker return ret; 57466ba94429SFrederic Weisbecker } 57476ba94429SFrederic Weisbecker 57486ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 57496ba94429SFrederic Weisbecker struct device_attribute *attr; 57506ba94429SFrederic Weisbecker 57516ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 57526ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 57536ba94429SFrederic Weisbecker if (ret) { 57546ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57556ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57566ba94429SFrederic Weisbecker return ret; 57576ba94429SFrederic Weisbecker } 57586ba94429SFrederic Weisbecker } 57596ba94429SFrederic Weisbecker } 57606ba94429SFrederic Weisbecker 57616ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 57626ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 57636ba94429SFrederic Weisbecker return 0; 57646ba94429SFrederic Weisbecker } 57656ba94429SFrederic Weisbecker 57666ba94429SFrederic Weisbecker /** 57676ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 57686ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 57696ba94429SFrederic Weisbecker * 57706ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 57716ba94429SFrederic Weisbecker */ 57726ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 57736ba94429SFrederic Weisbecker { 57746ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 57756ba94429SFrederic Weisbecker 57766ba94429SFrederic Weisbecker if (!wq->wq_dev) 57776ba94429SFrederic Weisbecker return; 57786ba94429SFrederic Weisbecker 57796ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57806ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57816ba94429SFrederic Weisbecker } 57826ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 57836ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 57846ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 57856ba94429SFrederic Weisbecker 578682607adcSTejun Heo /* 578782607adcSTejun Heo * Workqueue watchdog. 578882607adcSTejun Heo * 578982607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 579082607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 579182607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 579282607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 579382607adcSTejun Heo * largely opaque. 579482607adcSTejun Heo * 579582607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 579682607adcSTejun Heo * state if some pools failed to make forward progress for a while where 579782607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 579882607adcSTejun Heo * 579982607adcSTejun Heo * This mechanism is controlled through the kernel parameter 580082607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 580182607adcSTejun Heo * corresponding sysfs parameter file. 580282607adcSTejun Heo */ 580382607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 580482607adcSTejun Heo 580582607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 58065cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 580782607adcSTejun Heo 580882607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 580982607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 581082607adcSTejun Heo 581182607adcSTejun Heo static void wq_watchdog_reset_touched(void) 581282607adcSTejun Heo { 581382607adcSTejun Heo int cpu; 581482607adcSTejun Heo 581582607adcSTejun Heo wq_watchdog_touched = jiffies; 581682607adcSTejun Heo for_each_possible_cpu(cpu) 581782607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 581882607adcSTejun Heo } 581982607adcSTejun Heo 58205cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 582182607adcSTejun Heo { 582282607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 582382607adcSTejun Heo bool lockup_detected = false; 5824940d71c6SSergey Senozhatsky unsigned long now = jiffies; 582582607adcSTejun Heo struct worker_pool *pool; 582682607adcSTejun Heo int pi; 582782607adcSTejun Heo 582882607adcSTejun Heo if (!thresh) 582982607adcSTejun Heo return; 583082607adcSTejun Heo 583182607adcSTejun Heo rcu_read_lock(); 583282607adcSTejun Heo 583382607adcSTejun Heo for_each_pool(pool, pi) { 583482607adcSTejun Heo unsigned long pool_ts, touched, ts; 583582607adcSTejun Heo 583682607adcSTejun Heo if (list_empty(&pool->worklist)) 583782607adcSTejun Heo continue; 583882607adcSTejun Heo 5839940d71c6SSergey Senozhatsky /* 5840940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 5841940d71c6SSergey Senozhatsky * the watchdog like a stall. 5842940d71c6SSergey Senozhatsky */ 5843940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 5844940d71c6SSergey Senozhatsky 584582607adcSTejun Heo /* get the latest of pool and touched timestamps */ 584689e28ce6SWang Qing if (pool->cpu >= 0) 584789e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 584889e28ce6SWang Qing else 584982607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 585089e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 585182607adcSTejun Heo 585282607adcSTejun Heo if (time_after(pool_ts, touched)) 585382607adcSTejun Heo ts = pool_ts; 585482607adcSTejun Heo else 585582607adcSTejun Heo ts = touched; 585682607adcSTejun Heo 585782607adcSTejun Heo /* did we stall? */ 5858940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 585982607adcSTejun Heo lockup_detected = true; 586082607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 586182607adcSTejun Heo pr_cont_pool_info(pool); 586282607adcSTejun Heo pr_cont(" stuck for %us!\n", 5863940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 586482607adcSTejun Heo } 586582607adcSTejun Heo } 586682607adcSTejun Heo 586782607adcSTejun Heo rcu_read_unlock(); 586882607adcSTejun Heo 586982607adcSTejun Heo if (lockup_detected) 587055df0933SImran Khan show_all_workqueues(); 587182607adcSTejun Heo 587282607adcSTejun Heo wq_watchdog_reset_touched(); 587382607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 587482607adcSTejun Heo } 587582607adcSTejun Heo 5876cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 587782607adcSTejun Heo { 587882607adcSTejun Heo if (cpu >= 0) 587982607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 588089e28ce6SWang Qing 588182607adcSTejun Heo wq_watchdog_touched = jiffies; 588282607adcSTejun Heo } 588382607adcSTejun Heo 588482607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 588582607adcSTejun Heo { 588682607adcSTejun Heo wq_watchdog_thresh = 0; 588782607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 588882607adcSTejun Heo 588982607adcSTejun Heo if (thresh) { 589082607adcSTejun Heo wq_watchdog_thresh = thresh; 589182607adcSTejun Heo wq_watchdog_reset_touched(); 589282607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 589382607adcSTejun Heo } 589482607adcSTejun Heo } 589582607adcSTejun Heo 589682607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 589782607adcSTejun Heo const struct kernel_param *kp) 589882607adcSTejun Heo { 589982607adcSTejun Heo unsigned long thresh; 590082607adcSTejun Heo int ret; 590182607adcSTejun Heo 590282607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 590382607adcSTejun Heo if (ret) 590482607adcSTejun Heo return ret; 590582607adcSTejun Heo 590682607adcSTejun Heo if (system_wq) 590782607adcSTejun Heo wq_watchdog_set_thresh(thresh); 590882607adcSTejun Heo else 590982607adcSTejun Heo wq_watchdog_thresh = thresh; 591082607adcSTejun Heo 591182607adcSTejun Heo return 0; 591282607adcSTejun Heo } 591382607adcSTejun Heo 591482607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 591582607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 591682607adcSTejun Heo .get = param_get_ulong, 591782607adcSTejun Heo }; 591882607adcSTejun Heo 591982607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 592082607adcSTejun Heo 0644); 592182607adcSTejun Heo 592282607adcSTejun Heo static void wq_watchdog_init(void) 592382607adcSTejun Heo { 59245cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 592582607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 592682607adcSTejun Heo } 592782607adcSTejun Heo 592882607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 592982607adcSTejun Heo 593082607adcSTejun Heo static inline void wq_watchdog_init(void) { } 593182607adcSTejun Heo 593282607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 593382607adcSTejun Heo 5934bce90380STejun Heo static void __init wq_numa_init(void) 5935bce90380STejun Heo { 5936bce90380STejun Heo cpumask_var_t *tbl; 5937bce90380STejun Heo int node, cpu; 5938bce90380STejun Heo 5939bce90380STejun Heo if (num_possible_nodes() <= 1) 5940bce90380STejun Heo return; 5941bce90380STejun Heo 5942d55262c4STejun Heo if (wq_disable_numa) { 5943d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5944d55262c4STejun Heo return; 5945d55262c4STejun Heo } 5946d55262c4STejun Heo 5947f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 5948f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 5949f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5950f728c4a9SZhen Lei return; 5951f728c4a9SZhen Lei } 5952f728c4a9SZhen Lei } 5953f728c4a9SZhen Lei 5954be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 59554c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 59564c16bd32STejun Heo 5957bce90380STejun Heo /* 5958bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5959bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5960bce90380STejun Heo * fully initialized by now. 5961bce90380STejun Heo */ 59626396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5963bce90380STejun Heo BUG_ON(!tbl); 5964bce90380STejun Heo 5965bce90380STejun Heo for_each_node(node) 59665a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 59671be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5968bce90380STejun Heo 5969bce90380STejun Heo for_each_possible_cpu(cpu) { 5970bce90380STejun Heo node = cpu_to_node(cpu); 5971bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5972bce90380STejun Heo } 5973bce90380STejun Heo 5974bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5975bce90380STejun Heo wq_numa_enabled = true; 5976bce90380STejun Heo } 5977bce90380STejun Heo 59783347fa09STejun Heo /** 59793347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 59803347fa09STejun Heo * 59813347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 59823347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 59833347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 59843347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 59853347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 59863347fa09STejun Heo * created and scheduled right before early initcalls. 59873347fa09STejun Heo */ 59882333e829SYu Chen void __init workqueue_init_early(void) 59891da177e4SLinus Torvalds { 59907a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 59911bda3f80SFrederic Weisbecker int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 59927a4e344cSTejun Heo int i, cpu; 5993c34056a3STejun Heo 599410cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5995e904e6c2STejun Heo 5996b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 59971bda3f80SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 5998b05a7928SFrederic Weisbecker 5999e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6000e904e6c2STejun Heo 6001706026c2STejun Heo /* initialize CPU pools */ 600229c91e99STejun Heo for_each_possible_cpu(cpu) { 60034ce62e9eSTejun Heo struct worker_pool *pool; 60048b03ae3cSTejun Heo 60057a4e344cSTejun Heo i = 0; 6006f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 60077a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6008ec22ca5eSTejun Heo pool->cpu = cpu; 60097a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 60107a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6011f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 60127a4e344cSTejun Heo 60139daf9e67STejun Heo /* alloc pool ID */ 601468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 60159daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 601668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 60174ce62e9eSTejun Heo } 60188b03ae3cSTejun Heo } 60198b03ae3cSTejun Heo 60208a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 602129c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 602229c91e99STejun Heo struct workqueue_attrs *attrs; 602329c91e99STejun Heo 6024be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 602529c91e99STejun Heo attrs->nice = std_nice[i]; 602629c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 60278a2b7538STejun Heo 60288a2b7538STejun Heo /* 60298a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 60308a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 60318a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 60328a2b7538STejun Heo */ 6033be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 60348a2b7538STejun Heo attrs->nice = std_nice[i]; 60358a2b7538STejun Heo attrs->no_numa = true; 60368a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 603729c91e99STejun Heo } 603829c91e99STejun Heo 6039d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 60401aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6041d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6042f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6043f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 604424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 604524d51addSTejun Heo WQ_FREEZABLE, 0); 60460668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 60470668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 60480668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 60490668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 60500668106cSViresh Kumar 0); 60511aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 60520668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 60530668106cSViresh Kumar !system_power_efficient_wq || 60540668106cSViresh Kumar !system_freezable_power_efficient_wq); 60553347fa09STejun Heo } 60563347fa09STejun Heo 60573347fa09STejun Heo /** 60583347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 60593347fa09STejun Heo * 60603347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 60613347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 60623347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 60633347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 60643347fa09STejun Heo * with the initial workers and enable future kworker creations. 60653347fa09STejun Heo */ 60662333e829SYu Chen void __init workqueue_init(void) 60673347fa09STejun Heo { 60682186d9f9STejun Heo struct workqueue_struct *wq; 60693347fa09STejun Heo struct worker_pool *pool; 60703347fa09STejun Heo int cpu, bkt; 60713347fa09STejun Heo 60722186d9f9STejun Heo /* 60732186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 60742186d9f9STejun Heo * CPU to node mapping may not be available that early on some 60752186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 60762186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 60772186d9f9STejun Heo * affinity, fix them up. 607840c17f75STejun Heo * 607940c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 60802186d9f9STejun Heo */ 60812186d9f9STejun Heo wq_numa_init(); 60822186d9f9STejun Heo 60832186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 60842186d9f9STejun Heo 60852186d9f9STejun Heo for_each_possible_cpu(cpu) { 60862186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 60872186d9f9STejun Heo pool->node = cpu_to_node(cpu); 60882186d9f9STejun Heo } 60892186d9f9STejun Heo } 60902186d9f9STejun Heo 609140c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 60922186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 609340c17f75STejun Heo WARN(init_rescuer(wq), 609440c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 609540c17f75STejun Heo wq->name); 609640c17f75STejun Heo } 60972186d9f9STejun Heo 60982186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 60992186d9f9STejun Heo 61003347fa09STejun Heo /* create the initial workers */ 61013347fa09STejun Heo for_each_online_cpu(cpu) { 61023347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 61033347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 61043347fa09STejun Heo BUG_ON(!create_worker(pool)); 61053347fa09STejun Heo } 61063347fa09STejun Heo } 61073347fa09STejun Heo 61083347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 61093347fa09STejun Heo BUG_ON(!create_worker(pool)); 61103347fa09STejun Heo 61113347fa09STejun Heo wq_online = true; 611282607adcSTejun Heo wq_watchdog_init(); 61131da177e4SLinus Torvalds } 6114