1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 3c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 41da177e4SLinus Torvalds * 5c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 81da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 9e1f8e874SFrancois Cami * Andrew Morton 101da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 111da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1289ada679SChristoph Lameter * 13cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 14c54fce6eSTejun Heo * 15c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 16c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 17c54fce6eSTejun Heo * 18c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 19c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 20b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 21b11895c4SLibin * normal work items and the other for high priority ones) and some extra 22b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 23b11895c4SLibin * number of these backing pools is dynamic. 24c54fce6eSTejun Heo * 259a261491SBenjamin Peterson * Please read Documentation/core-api/workqueue.rst for details. 261da177e4SLinus Torvalds */ 271da177e4SLinus Torvalds 289984de1aSPaul Gortmaker #include <linux/export.h> 291da177e4SLinus Torvalds #include <linux/kernel.h> 301da177e4SLinus Torvalds #include <linux/sched.h> 311da177e4SLinus Torvalds #include <linux/init.h> 321da177e4SLinus Torvalds #include <linux/signal.h> 331da177e4SLinus Torvalds #include <linux/completion.h> 341da177e4SLinus Torvalds #include <linux/workqueue.h> 351da177e4SLinus Torvalds #include <linux/slab.h> 361da177e4SLinus Torvalds #include <linux/cpu.h> 371da177e4SLinus Torvalds #include <linux/notifier.h> 381da177e4SLinus Torvalds #include <linux/kthread.h> 391fa44ecaSJames Bottomley #include <linux/hardirq.h> 4046934023SChristoph Lameter #include <linux/mempolicy.h> 41341a5958SRafael J. Wysocki #include <linux/freezer.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51c98a9805STal Shorer #include <linux/sched/isolation.h> 5262635ea8SSergey Senozhatsky #include <linux/nmi.h> 53940d71c6SSergey Senozhatsky #include <linux/kvm_para.h> 54e22bee78STejun Heo 55ea138446STejun Heo #include "workqueue_internal.h" 561da177e4SLinus Torvalds 57c8e55f36STejun Heo enum { 58bc2ae0f5STejun Heo /* 5924647570STejun Heo * worker_pool flags 60bc2ae0f5STejun Heo * 6124647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 62bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 63bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 64bc2ae0f5STejun Heo * is in effect. 65bc2ae0f5STejun Heo * 66bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 67bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6824647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 69bc2ae0f5STejun Heo * 70bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 711258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 724736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 73bc2ae0f5STejun Heo */ 74692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7524647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 76db7bccf4STejun Heo 77c8e55f36STejun Heo /* worker flags */ 78c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 79c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 80e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 81fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 82f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 83a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 84e22bee78STejun Heo 85a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 86a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 87db7bccf4STejun Heo 88e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 894ce62e9eSTejun Heo 9029c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 91c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 92db7bccf4STejun Heo 93e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 94e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 95e22bee78STejun Heo 963233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 973233cdbdSTejun Heo /* call for help after 10ms 983233cdbdSTejun Heo (min two ticks) */ 99e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 100e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1011da177e4SLinus Torvalds 1021da177e4SLinus Torvalds /* 103e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1048698a745SDongsheng Yang * all cpus. Give MIN_NICE. 105e22bee78STejun Heo */ 1068698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1078698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 108ecf6881fSTejun Heo 109ecf6881fSTejun Heo WQ_NAME_LEN = 24, 110c8e55f36STejun Heo }; 111c8e55f36STejun Heo 1121da177e4SLinus Torvalds /* 1134690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1144690c4abSTejun Heo * 115e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 116e41e704bSTejun Heo * everyone else. 1174690c4abSTejun Heo * 118e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 119e22bee78STejun Heo * only be modified and accessed from the local cpu. 120e22bee78STejun Heo * 121d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1224690c4abSTejun Heo * 123d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 124d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 125d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 126d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 127e22bee78STejun Heo * 1281258fae7STejun Heo * A: wq_pool_attach_mutex protected. 129822d8405STejun Heo * 13068e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13176af4d93STejun Heo * 13224acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1335bcab335STejun Heo * 1345b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1355b95e1afSLai Jiangshan * 1365b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 13724acfb71SThomas Gleixner * RCU for reads. 1385b95e1afSLai Jiangshan * 1393c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1403c25a55dSLai Jiangshan * 14124acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1422e109a28STejun Heo * 1432e109a28STejun Heo * MD: wq_mayday_lock protected. 1444690c4abSTejun Heo */ 1454690c4abSTejun Heo 1462eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 147c34056a3STejun Heo 148bd7bdd43STejun Heo struct worker_pool { 149a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 150d84ff051STejun Heo int cpu; /* I: the associated cpu */ 151f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1529daf9e67STejun Heo int id; /* I: pool ID */ 15311ebea50STejun Heo unsigned int flags; /* X: flags */ 154bd7bdd43STejun Heo 15582607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15682607adcSTejun Heo 157bc35f7efSLai Jiangshan /* 158bc35f7efSLai Jiangshan * The counter is incremented in a process context on the associated CPU 159bc35f7efSLai Jiangshan * w/ preemption disabled, and decremented or reset in the same context 160bc35f7efSLai Jiangshan * but w/ pool->lock held. The readers grab pool->lock and are 161bc35f7efSLai Jiangshan * guaranteed to see if the counter reached zero. 162bc35f7efSLai Jiangshan */ 163bc35f7efSLai Jiangshan int nr_running; 16484f91c62SLai Jiangshan 165bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 166ea1abd61SLai Jiangshan 1675826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1685826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 169bd7bdd43STejun Heo 1702c1f1a91SLai Jiangshan struct list_head idle_list; /* L: list of idle workers */ 171bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 172bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 173bd7bdd43STejun Heo 174c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 175c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 176c9e7cf27STejun Heo /* L: hash of busy workers */ 177c9e7cf27STejun Heo 1782607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17992f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 18060f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 181e19e397aSTejun Heo 1827cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 183e19e397aSTejun Heo 1847a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 18568e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 18668e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1877a4e344cSTejun Heo 188e19e397aSTejun Heo /* 18924acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 19029c91e99STejun Heo * from get_work_pool(). 19129c91e99STejun Heo */ 19229c91e99STejun Heo struct rcu_head rcu; 19384f91c62SLai Jiangshan }; 1948b03ae3cSTejun Heo 1958b03ae3cSTejun Heo /* 196112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 197112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 198112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 199112202d9STejun Heo * number of flag bits. 2001da177e4SLinus Torvalds */ 201112202d9STejun Heo struct pool_workqueue { 202bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2034690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20473f53c4aSTejun Heo int work_color; /* L: current color */ 20573f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2068864b4e5STejun Heo int refcnt; /* L: reference count */ 20773f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20873f53c4aSTejun Heo /* L: nr of in_flight works */ 209018f3a13SLai Jiangshan 210018f3a13SLai Jiangshan /* 211018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 212018f3a13SLai Jiangshan * 213018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 214018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 215018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 216018f3a13SLai Jiangshan * 217018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 218018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 219018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 220018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 221018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 222018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 223018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 224018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 225018f3a13SLai Jiangshan */ 2261e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 227a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 228f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2293c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2302e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2318864b4e5STejun Heo 2328864b4e5STejun Heo /* 2338864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2348864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 23524acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 236b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2378864b4e5STejun Heo */ 2388864b4e5STejun Heo struct work_struct unbound_release_work; 2398864b4e5STejun Heo struct rcu_head rcu; 240e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2411da177e4SLinus Torvalds 2421da177e4SLinus Torvalds /* 24373f53c4aSTejun Heo * Structure used to wait for workqueue flush. 24473f53c4aSTejun Heo */ 24573f53c4aSTejun Heo struct wq_flusher { 2463c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2473c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 24873f53c4aSTejun Heo struct completion done; /* flush completion */ 24973f53c4aSTejun Heo }; 2501da177e4SLinus Torvalds 251226223abSTejun Heo struct wq_device; 252226223abSTejun Heo 25373f53c4aSTejun Heo /* 254c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 255c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2561da177e4SLinus Torvalds */ 2571da177e4SLinus Torvalds struct workqueue_struct { 2583c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 259e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 26073f53c4aSTejun Heo 2613c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2623c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2633c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 264112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2653c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2663c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2673c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 26873f53c4aSTejun Heo 2692e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 27030ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 271e22bee78STejun Heo 27287fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 273a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 274226223abSTejun Heo 2755b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2765b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2776029a918STejun Heo 278226223abSTejun Heo #ifdef CONFIG_SYSFS 279226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 280226223abSTejun Heo #endif 2814e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 282669de8bdSBart Van Assche char *lock_name; 283669de8bdSBart Van Assche struct lock_class_key key; 2844e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2854e6045f1SJohannes Berg #endif 286ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2872728fd2fSTejun Heo 288e2dca7adSTejun Heo /* 28924acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 29024acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 291e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 292e2dca7adSTejun Heo */ 293e2dca7adSTejun Heo struct rcu_head rcu; 294e2dca7adSTejun Heo 2952728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2962728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2972728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2985b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2991da177e4SLinus Torvalds }; 3001da177e4SLinus Torvalds 301e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 302e904e6c2STejun Heo 303bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 304bce90380STejun Heo /* possible CPUs of each node */ 305bce90380STejun Heo 306d55262c4STejun Heo static bool wq_disable_numa; 307d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 308d55262c4STejun Heo 309cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 310552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 311cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 312cee22a15SViresh Kumar 313863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3143347fa09STejun Heo 315bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 316bce90380STejun Heo 3174c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3184c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3194c16bd32STejun Heo 32068e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3211258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 322a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 323d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 324d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3255bcab335STejun Heo 326e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 32768e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3287d19c5ceSTejun Heo 329ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 330ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 331ef557180SMike Galbraith 332ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 333ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 334b05a7928SFrederic Weisbecker 335f303fccbSTejun Heo /* 336f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 337f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 338f303fccbSTejun Heo * to uncover usages which depend on it. 339f303fccbSTejun Heo */ 340f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 341f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 342f303fccbSTejun Heo #else 343f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 344f303fccbSTejun Heo #endif 345f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 346f303fccbSTejun Heo 3477d19c5ceSTejun Heo /* the per-cpu worker pools */ 34825528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3497d19c5ceSTejun Heo 35068e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3517d19c5ceSTejun Heo 35268e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 35329c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 35429c91e99STejun Heo 355c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 35629c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 35729c91e99STejun Heo 3588a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3598a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3608a2b7538STejun Heo 361d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 362ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 363044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3641aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 365044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 366d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 367044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 368f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 369044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 37024d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3710668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3720668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3730668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3740668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 375d320c038STejun Heo 3767d19c5ceSTejun Heo static int worker_thread(void *__worker); 3776ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 378c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 37955df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 3807d19c5ceSTejun Heo 38197bd2347STejun Heo #define CREATE_TRACE_POINTS 38297bd2347STejun Heo #include <trace/events/workqueue.h> 38397bd2347STejun Heo 38468e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 38524acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 386f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 38724acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 3885bcab335STejun Heo 3895b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 39024acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 391f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 392f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 39324acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 3945b95e1afSLai Jiangshan 395f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 396f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 397f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3987a62c2c8STejun Heo (pool)++) 3994ce62e9eSTejun Heo 40049e3cf44STejun Heo /** 40117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 40217116969STejun Heo * @pool: iteration cursor 403611c92a0STejun Heo * @pi: integer used for iteration 404fa1b54e6STejun Heo * 40524acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 40668e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 40768e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 408fa1b54e6STejun Heo * 409fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 410fa1b54e6STejun Heo * ignored. 41117116969STejun Heo */ 412611c92a0STejun Heo #define for_each_pool(pool, pi) \ 413611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 41468e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 415fa1b54e6STejun Heo else 41617116969STejun Heo 41717116969STejun Heo /** 418822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 419822d8405STejun Heo * @worker: iteration cursor 420822d8405STejun Heo * @pool: worker_pool to iterate workers of 421822d8405STejun Heo * 4221258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 423822d8405STejun Heo * 424822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 425822d8405STejun Heo * ignored. 426822d8405STejun Heo */ 427da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 428da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4291258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 430822d8405STejun Heo else 431822d8405STejun Heo 432822d8405STejun Heo /** 43349e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 43449e3cf44STejun Heo * @pwq: iteration cursor 43549e3cf44STejun Heo * @wq: the target workqueue 43676af4d93STejun Heo * 43724acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 438794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 439794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 44076af4d93STejun Heo * 44176af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 44276af4d93STejun Heo * ignored. 44349e3cf44STejun Heo */ 44449e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 44549e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4465a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 447f3421797STejun Heo 448dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 449dc186ad7SThomas Gleixner 450f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 451dc186ad7SThomas Gleixner 45299777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 45399777288SStanislaw Gruszka { 45499777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 45599777288SStanislaw Gruszka } 45699777288SStanislaw Gruszka 457b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 458b9fdac7fSDu, Changbin { 459b9fdac7fSDu, Changbin struct work_struct *work = addr; 460b9fdac7fSDu, Changbin 461b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 462b9fdac7fSDu, Changbin } 463b9fdac7fSDu, Changbin 464dc186ad7SThomas Gleixner /* 465dc186ad7SThomas Gleixner * fixup_init is called when: 466dc186ad7SThomas Gleixner * - an active object is initialized 467dc186ad7SThomas Gleixner */ 46802a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 469dc186ad7SThomas Gleixner { 470dc186ad7SThomas Gleixner struct work_struct *work = addr; 471dc186ad7SThomas Gleixner 472dc186ad7SThomas Gleixner switch (state) { 473dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 474dc186ad7SThomas Gleixner cancel_work_sync(work); 475dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 47602a982a6SDu, Changbin return true; 477dc186ad7SThomas Gleixner default: 47802a982a6SDu, Changbin return false; 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner } 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner /* 483dc186ad7SThomas Gleixner * fixup_free is called when: 484dc186ad7SThomas Gleixner * - an active object is freed 485dc186ad7SThomas Gleixner */ 48602a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 487dc186ad7SThomas Gleixner { 488dc186ad7SThomas Gleixner struct work_struct *work = addr; 489dc186ad7SThomas Gleixner 490dc186ad7SThomas Gleixner switch (state) { 491dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 492dc186ad7SThomas Gleixner cancel_work_sync(work); 493dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 49402a982a6SDu, Changbin return true; 495dc186ad7SThomas Gleixner default: 49602a982a6SDu, Changbin return false; 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner } 499dc186ad7SThomas Gleixner 500f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 501dc186ad7SThomas Gleixner .name = "work_struct", 50299777288SStanislaw Gruszka .debug_hint = work_debug_hint, 503b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 504dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 505dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 506dc186ad7SThomas Gleixner }; 507dc186ad7SThomas Gleixner 508dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 511dc186ad7SThomas Gleixner } 512dc186ad7SThomas Gleixner 513dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 514dc186ad7SThomas Gleixner { 515dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 516dc186ad7SThomas Gleixner } 517dc186ad7SThomas Gleixner 518dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 519dc186ad7SThomas Gleixner { 520dc186ad7SThomas Gleixner if (onstack) 521dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 522dc186ad7SThomas Gleixner else 523dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 524dc186ad7SThomas Gleixner } 525dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 526dc186ad7SThomas Gleixner 527dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 528dc186ad7SThomas Gleixner { 529dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 530dc186ad7SThomas Gleixner } 531dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 532dc186ad7SThomas Gleixner 533ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 534ea2e64f2SThomas Gleixner { 535ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 536ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 537ea2e64f2SThomas Gleixner } 538ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 539ea2e64f2SThomas Gleixner 540dc186ad7SThomas Gleixner #else 541dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 542dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 543dc186ad7SThomas Gleixner #endif 544dc186ad7SThomas Gleixner 5454e8b22bdSLi Bin /** 54667dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5474e8b22bdSLi Bin * @pool: the pool pointer of interest 5484e8b22bdSLi Bin * 5494e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5504e8b22bdSLi Bin * successfully, -errno on failure. 5514e8b22bdSLi Bin */ 5529daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5539daf9e67STejun Heo { 5549daf9e67STejun Heo int ret; 5559daf9e67STejun Heo 55668e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5575bcab335STejun Heo 5584e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5594e8b22bdSLi Bin GFP_KERNEL); 560229641a6STejun Heo if (ret >= 0) { 561e68035fbSTejun Heo pool->id = ret; 562229641a6STejun Heo return 0; 563229641a6STejun Heo } 5649daf9e67STejun Heo return ret; 5659daf9e67STejun Heo } 5669daf9e67STejun Heo 56776af4d93STejun Heo /** 568df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 569df2d5ae4STejun Heo * @wq: the target workqueue 570df2d5ae4STejun Heo * @node: the node ID 571df2d5ae4STejun Heo * 57224acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 5735b95e1afSLai Jiangshan * read locked. 574df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 575df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 576d185af30SYacine Belkadi * 577d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 578df2d5ae4STejun Heo */ 579df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 580df2d5ae4STejun Heo int node) 581df2d5ae4STejun Heo { 5825b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 583d6e022f1STejun Heo 584d6e022f1STejun Heo /* 585d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 586d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 587d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 588d6e022f1STejun Heo * happens, this workaround can be removed. 589d6e022f1STejun Heo */ 590d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 591d6e022f1STejun Heo return wq->dfl_pwq; 592d6e022f1STejun Heo 593df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 594df2d5ae4STejun Heo } 595df2d5ae4STejun Heo 59673f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 59773f53c4aSTejun Heo { 59873f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 59973f53c4aSTejun Heo } 60073f53c4aSTejun Heo 601c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 60273f53c4aSTejun Heo { 603c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 60473f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 60573f53c4aSTejun Heo } 60673f53c4aSTejun Heo 60773f53c4aSTejun Heo static int work_next_color(int color) 60873f53c4aSTejun Heo { 60973f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 610a848e3b6SOleg Nesterov } 611a848e3b6SOleg Nesterov 6124594bf15SDavid Howells /* 613112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 614112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6157c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6167a22ad75STejun Heo * 617112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 618112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 619bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 620bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6217a22ad75STejun Heo * 622112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6237c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 624112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6257c3eed5cSTejun Heo * available only while the work item is queued. 626bbb68dfaSTejun Heo * 627bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 628bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 629bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 630bbb68dfaSTejun Heo * try to steal the PENDING bit. 6314594bf15SDavid Howells */ 6327a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6337a22ad75STejun Heo unsigned long flags) 6347a22ad75STejun Heo { 6356183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6367a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6377a22ad75STejun Heo } 6387a22ad75STejun Heo 639112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6404690c4abSTejun Heo unsigned long extra_flags) 641365970a1SDavid Howells { 642112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 643112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 644365970a1SDavid Howells } 645365970a1SDavid Howells 6464468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6474468a00fSLai Jiangshan int pool_id) 6484468a00fSLai Jiangshan { 6494468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6504468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6514468a00fSLai Jiangshan } 6524468a00fSLai Jiangshan 6537c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6547c3eed5cSTejun Heo int pool_id) 6554d707b9fSOleg Nesterov { 65623657bb1STejun Heo /* 65723657bb1STejun Heo * The following wmb is paired with the implied mb in 65823657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 65923657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 66023657bb1STejun Heo * owner. 66123657bb1STejun Heo */ 66223657bb1STejun Heo smp_wmb(); 6637c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 664346c09f8SRoman Pen /* 665346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 666346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 667346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6688bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 669346c09f8SRoman Pen * the same @work. E.g. consider this case: 670346c09f8SRoman Pen * 671346c09f8SRoman Pen * CPU#0 CPU#1 672346c09f8SRoman Pen * ---------------------------- -------------------------------- 673346c09f8SRoman Pen * 674346c09f8SRoman Pen * 1 STORE event_indicated 675346c09f8SRoman Pen * 2 queue_work_on() { 676346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 677346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 678346c09f8SRoman Pen * 5 set_work_data() # clear bit 679346c09f8SRoman Pen * 6 smp_mb() 680346c09f8SRoman Pen * 7 work->current_func() { 681346c09f8SRoman Pen * 8 LOAD event_indicated 682346c09f8SRoman Pen * } 683346c09f8SRoman Pen * 684346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 685346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 686346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 687346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 688346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 689346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 690346c09f8SRoman Pen * before actual STORE. 691346c09f8SRoman Pen */ 692346c09f8SRoman Pen smp_mb(); 6934d707b9fSOleg Nesterov } 6944d707b9fSOleg Nesterov 6957a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 696365970a1SDavid Howells { 6977c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6987c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6997a22ad75STejun Heo } 7007a22ad75STejun Heo 701112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7027a22ad75STejun Heo { 703e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7047a22ad75STejun Heo 705112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 706e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 707e120153dSTejun Heo else 708e120153dSTejun Heo return NULL; 7097a22ad75STejun Heo } 7107a22ad75STejun Heo 7117c3eed5cSTejun Heo /** 7127c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7137c3eed5cSTejun Heo * @work: the work item of interest 7147c3eed5cSTejun Heo * 71568e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 71624acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 71724acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 718fa1b54e6STejun Heo * 719fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 720fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 721fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 722fa1b54e6STejun Heo * returned pool is and stays online. 723d185af30SYacine Belkadi * 724d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7257c3eed5cSTejun Heo */ 7267c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7277a22ad75STejun Heo { 728e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7297c3eed5cSTejun Heo int pool_id; 7307a22ad75STejun Heo 73168e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 732fa1b54e6STejun Heo 733112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 734112202d9STejun Heo return ((struct pool_workqueue *) 7357c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7367a22ad75STejun Heo 7377c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7387c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7397a22ad75STejun Heo return NULL; 7407a22ad75STejun Heo 741fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7427c3eed5cSTejun Heo } 7437c3eed5cSTejun Heo 7447c3eed5cSTejun Heo /** 7457c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7467c3eed5cSTejun Heo * @work: the work item of interest 7477c3eed5cSTejun Heo * 748d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7497c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7507c3eed5cSTejun Heo */ 7517c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7527c3eed5cSTejun Heo { 75354d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7547c3eed5cSTejun Heo 755112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 756112202d9STejun Heo return ((struct pool_workqueue *) 75754d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 75854d5b7d0SLai Jiangshan 75954d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7607c3eed5cSTejun Heo } 7617c3eed5cSTejun Heo 762bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 763bbb68dfaSTejun Heo { 7647c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 765bbb68dfaSTejun Heo 7667c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7677c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 768bbb68dfaSTejun Heo } 769bbb68dfaSTejun Heo 770bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 771bbb68dfaSTejun Heo { 772bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 773bbb68dfaSTejun Heo 774112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 775bbb68dfaSTejun Heo } 776bbb68dfaSTejun Heo 777e22bee78STejun Heo /* 7783270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7793270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 780d565ed63STejun Heo * they're being called with pool->lock held. 781e22bee78STejun Heo */ 782e22bee78STejun Heo 78363d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 784649027d7STejun Heo { 785bc35f7efSLai Jiangshan return !pool->nr_running; 786649027d7STejun Heo } 787649027d7STejun Heo 788e22bee78STejun Heo /* 789e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 790e22bee78STejun Heo * running workers. 791974271c4STejun Heo * 792974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 793706026c2STejun Heo * function will always return %true for unbound pools as long as the 794974271c4STejun Heo * worklist isn't empty. 795e22bee78STejun Heo */ 79663d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 797e22bee78STejun Heo { 79863d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 799e22bee78STejun Heo } 800e22bee78STejun Heo 801e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 80263d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 803e22bee78STejun Heo { 80463d95a91STejun Heo return pool->nr_idle; 805e22bee78STejun Heo } 806e22bee78STejun Heo 807e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 80863d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 809e22bee78STejun Heo { 810bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 811e22bee78STejun Heo } 812e22bee78STejun Heo 813e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 81463d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 815e22bee78STejun Heo { 81663d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 817e22bee78STejun Heo } 818e22bee78STejun Heo 819e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 82063d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 821e22bee78STejun Heo { 822692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 82363d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 82463d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 825e22bee78STejun Heo 826e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 827e22bee78STejun Heo } 828e22bee78STejun Heo 829e22bee78STejun Heo /* 830e22bee78STejun Heo * Wake up functions. 831e22bee78STejun Heo */ 832e22bee78STejun Heo 8332c1f1a91SLai Jiangshan /* Return the first idle worker. Called with pool->lock held. */ 8341037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8357e11629dSTejun Heo { 83663d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8377e11629dSTejun Heo return NULL; 8387e11629dSTejun Heo 83963d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8407e11629dSTejun Heo } 8417e11629dSTejun Heo 8427e11629dSTejun Heo /** 8437e11629dSTejun Heo * wake_up_worker - wake up an idle worker 84463d95a91STejun Heo * @pool: worker pool to wake worker from 8457e11629dSTejun Heo * 84663d95a91STejun Heo * Wake up the first idle worker of @pool. 8477e11629dSTejun Heo * 8487e11629dSTejun Heo * CONTEXT: 849a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8507e11629dSTejun Heo */ 85163d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8527e11629dSTejun Heo { 8531037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8547e11629dSTejun Heo 8557e11629dSTejun Heo if (likely(worker)) 8567e11629dSTejun Heo wake_up_process(worker->task); 8577e11629dSTejun Heo } 8587e11629dSTejun Heo 8594690c4abSTejun Heo /** 8606d25be57SThomas Gleixner * wq_worker_running - a worker is running again 861e22bee78STejun Heo * @task: task waking up 862e22bee78STejun Heo * 8636d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 864e22bee78STejun Heo */ 8656d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 866e22bee78STejun Heo { 867e22bee78STejun Heo struct worker *worker = kthread_data(task); 868e22bee78STejun Heo 8696d25be57SThomas Gleixner if (!worker->sleeping) 8706d25be57SThomas Gleixner return; 87107edfeceSFrederic Weisbecker 87207edfeceSFrederic Weisbecker /* 87307edfeceSFrederic Weisbecker * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 87407edfeceSFrederic Weisbecker * and the nr_running increment below, we may ruin the nr_running reset 87507edfeceSFrederic Weisbecker * and leave with an unexpected pool->nr_running == 1 on the newly unbound 87607edfeceSFrederic Weisbecker * pool. Protect against such race. 87707edfeceSFrederic Weisbecker */ 87807edfeceSFrederic Weisbecker preempt_disable(); 8796d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 880bc35f7efSLai Jiangshan worker->pool->nr_running++; 88107edfeceSFrederic Weisbecker preempt_enable(); 8826d25be57SThomas Gleixner worker->sleeping = 0; 88336576000SJoonsoo Kim } 884e22bee78STejun Heo 885e22bee78STejun Heo /** 886e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 887e22bee78STejun Heo * @task: task going to sleep 888e22bee78STejun Heo * 8896d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 890ccf45156SLai Jiangshan * going to sleep. 891e22bee78STejun Heo */ 8926d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 893e22bee78STejun Heo { 894cc5bff38SLai Jiangshan struct worker *worker = kthread_data(task); 895111c225aSTejun Heo struct worker_pool *pool; 896e22bee78STejun Heo 897111c225aSTejun Heo /* 898111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 899111c225aSTejun Heo * workers, also reach here, let's not access anything before 900111c225aSTejun Heo * checking NOT_RUNNING. 901111c225aSTejun Heo */ 9022d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 9036d25be57SThomas Gleixner return; 904e22bee78STejun Heo 905111c225aSTejun Heo pool = worker->pool; 906111c225aSTejun Heo 90762849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 90862849a96SSebastian Andrzej Siewior if (worker->sleeping) 9096d25be57SThomas Gleixner return; 9106d25be57SThomas Gleixner 9116d25be57SThomas Gleixner worker->sleeping = 1; 912a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 913e22bee78STejun Heo 914e22bee78STejun Heo /* 91545c753f5SFrederic Weisbecker * Recheck in case unbind_workers() preempted us. We don't 91645c753f5SFrederic Weisbecker * want to decrement nr_running after the worker is unbound 91745c753f5SFrederic Weisbecker * and nr_running has been reset. 91845c753f5SFrederic Weisbecker */ 91945c753f5SFrederic Weisbecker if (worker->flags & WORKER_NOT_RUNNING) { 92045c753f5SFrederic Weisbecker raw_spin_unlock_irq(&pool->lock); 92145c753f5SFrederic Weisbecker return; 92245c753f5SFrederic Weisbecker } 92345c753f5SFrederic Weisbecker 924bc35f7efSLai Jiangshan pool->nr_running--; 925bc35f7efSLai Jiangshan if (need_more_worker(pool)) 926cc5bff38SLai Jiangshan wake_up_worker(pool); 927a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 928e22bee78STejun Heo } 929e22bee78STejun Heo 930e22bee78STejun Heo /** 9311b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 9328194fe94SBart Van Assche * @task: Task to retrieve last work function of. 9331b69ac6bSJohannes Weiner * 9341b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 9351b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 9361b69ac6bSJohannes Weiner * 9371b69ac6bSJohannes Weiner * CONTEXT: 938a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 9391b69ac6bSJohannes Weiner * 9404b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 9414b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 9424b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 9434b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 9444b047002SJohannes Weiner * 9454b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 9464b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 9474b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 9484b047002SJohannes Weiner * is guaranteed to not be processing any works. 9494b047002SJohannes Weiner * 9501b69ac6bSJohannes Weiner * Return: 9511b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 9521b69ac6bSJohannes Weiner * hasn't executed any work yet. 9531b69ac6bSJohannes Weiner */ 9541b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 9551b69ac6bSJohannes Weiner { 9561b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 9571b69ac6bSJohannes Weiner 9581b69ac6bSJohannes Weiner return worker->last_func; 9591b69ac6bSJohannes Weiner } 9601b69ac6bSJohannes Weiner 9611b69ac6bSJohannes Weiner /** 962e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 963cb444766STejun Heo * @worker: self 964d302f017STejun Heo * @flags: flags to set 965d302f017STejun Heo * 966228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 967d302f017STejun Heo * 968cb444766STejun Heo * CONTEXT: 969a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 970d302f017STejun Heo */ 971228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 972d302f017STejun Heo { 973bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 974e22bee78STejun Heo 975cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 976cb444766STejun Heo 977228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 978e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 979e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 980bc35f7efSLai Jiangshan pool->nr_running--; 981e22bee78STejun Heo } 982e22bee78STejun Heo 983d302f017STejun Heo worker->flags |= flags; 984d302f017STejun Heo } 985d302f017STejun Heo 986d302f017STejun Heo /** 987e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 988cb444766STejun Heo * @worker: self 989d302f017STejun Heo * @flags: flags to clear 990d302f017STejun Heo * 991e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 992d302f017STejun Heo * 993cb444766STejun Heo * CONTEXT: 994a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 995d302f017STejun Heo */ 996d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 997d302f017STejun Heo { 99863d95a91STejun Heo struct worker_pool *pool = worker->pool; 999e22bee78STejun Heo unsigned int oflags = worker->flags; 1000e22bee78STejun Heo 1001cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 1002cb444766STejun Heo 1003d302f017STejun Heo worker->flags &= ~flags; 1004e22bee78STejun Heo 100542c025f3STejun Heo /* 100642c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 100742c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 100842c025f3STejun Heo * of multiple flags, not a single flag. 100942c025f3STejun Heo */ 1010e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 1011e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 1012bc35f7efSLai Jiangshan pool->nr_running++; 1013d302f017STejun Heo } 1014d302f017STejun Heo 1015d302f017STejun Heo /** 10168cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1017c9e7cf27STejun Heo * @pool: pool of interest 10188cca0eeaSTejun Heo * @work: work to find worker for 10198cca0eeaSTejun Heo * 1020c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1021c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1022a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1023a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1024a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1025a2c1c57bSTejun Heo * being executed. 1026a2c1c57bSTejun Heo * 1027a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1028a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1029a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1030a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1031a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1032a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1033a2c1c57bSTejun Heo * 1034c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1035c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1036c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1037c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1038c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1039c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10408cca0eeaSTejun Heo * 10418cca0eeaSTejun Heo * CONTEXT: 1042a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 10438cca0eeaSTejun Heo * 1044d185af30SYacine Belkadi * Return: 1045d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10468cca0eeaSTejun Heo * otherwise. 10478cca0eeaSTejun Heo */ 1048c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10498cca0eeaSTejun Heo struct work_struct *work) 10508cca0eeaSTejun Heo { 105142f8570fSSasha Levin struct worker *worker; 105242f8570fSSasha Levin 1053b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1054a2c1c57bSTejun Heo (unsigned long)work) 1055a2c1c57bSTejun Heo if (worker->current_work == work && 1056a2c1c57bSTejun Heo worker->current_func == work->func) 105742f8570fSSasha Levin return worker; 105842f8570fSSasha Levin 105942f8570fSSasha Levin return NULL; 10608cca0eeaSTejun Heo } 10618cca0eeaSTejun Heo 10628cca0eeaSTejun Heo /** 1063bf4ede01STejun Heo * move_linked_works - move linked works to a list 1064bf4ede01STejun Heo * @work: start of series of works to be scheduled 1065bf4ede01STejun Heo * @head: target list to append @work to 1066402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1067bf4ede01STejun Heo * 1068bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1069bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1070bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1071bf4ede01STejun Heo * 1072bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1073bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1074bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1075bf4ede01STejun Heo * 1076bf4ede01STejun Heo * CONTEXT: 1077a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1078bf4ede01STejun Heo */ 1079bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1080bf4ede01STejun Heo struct work_struct **nextp) 1081bf4ede01STejun Heo { 1082bf4ede01STejun Heo struct work_struct *n; 1083bf4ede01STejun Heo 1084bf4ede01STejun Heo /* 1085bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1086bf4ede01STejun Heo * use NULL for list head. 1087bf4ede01STejun Heo */ 1088bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1089bf4ede01STejun Heo list_move_tail(&work->entry, head); 1090bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1091bf4ede01STejun Heo break; 1092bf4ede01STejun Heo } 1093bf4ede01STejun Heo 1094bf4ede01STejun Heo /* 1095bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1096bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1097bf4ede01STejun Heo * needs to be updated. 1098bf4ede01STejun Heo */ 1099bf4ede01STejun Heo if (nextp) 1100bf4ede01STejun Heo *nextp = n; 1101bf4ede01STejun Heo } 1102bf4ede01STejun Heo 11038864b4e5STejun Heo /** 11048864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 11058864b4e5STejun Heo * @pwq: pool_workqueue to get 11068864b4e5STejun Heo * 11078864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 11088864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 11098864b4e5STejun Heo */ 11108864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 11118864b4e5STejun Heo { 11128864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11138864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 11148864b4e5STejun Heo pwq->refcnt++; 11158864b4e5STejun Heo } 11168864b4e5STejun Heo 11178864b4e5STejun Heo /** 11188864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 11198864b4e5STejun Heo * @pwq: pool_workqueue to put 11208864b4e5STejun Heo * 11218864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 11228864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11238864b4e5STejun Heo */ 11248864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11258864b4e5STejun Heo { 11268864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11278864b4e5STejun Heo if (likely(--pwq->refcnt)) 11288864b4e5STejun Heo return; 11298864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11308864b4e5STejun Heo return; 11318864b4e5STejun Heo /* 11328864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11338864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11348864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11358864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11368864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11378864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11388864b4e5STejun Heo */ 11398864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11408864b4e5STejun Heo } 11418864b4e5STejun Heo 1142dce90d47STejun Heo /** 1143dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1144dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1145dce90d47STejun Heo * 1146dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1147dce90d47STejun Heo */ 1148dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1149dce90d47STejun Heo { 1150dce90d47STejun Heo if (pwq) { 1151dce90d47STejun Heo /* 115224acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1153dce90d47STejun Heo * following lock operations are safe. 1154dce90d47STejun Heo */ 1155a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1156dce90d47STejun Heo put_pwq(pwq); 1157a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1158dce90d47STejun Heo } 1159dce90d47STejun Heo } 1160dce90d47STejun Heo 1161f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1162bf4ede01STejun Heo { 1163112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1164bf4ede01STejun Heo 1165bf4ede01STejun Heo trace_workqueue_activate_work(work); 116682607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 116782607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1168112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1169f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1170112202d9STejun Heo pwq->nr_active++; 1171bf4ede01STejun Heo } 1172bf4ede01STejun Heo 1173f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 11743aa62497SLai Jiangshan { 1175f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 11763aa62497SLai Jiangshan struct work_struct, entry); 11773aa62497SLai Jiangshan 1178f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 11793aa62497SLai Jiangshan } 11803aa62497SLai Jiangshan 1181bf4ede01STejun Heo /** 1182112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1183112202d9STejun Heo * @pwq: pwq of interest 1184c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1185bf4ede01STejun Heo * 1186bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1187112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1188bf4ede01STejun Heo * 1189bf4ede01STejun Heo * CONTEXT: 1190a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1191bf4ede01STejun Heo */ 1192c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1193bf4ede01STejun Heo { 1194c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1195c4560c2cSLai Jiangshan 1196018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1197112202d9STejun Heo pwq->nr_active--; 1198f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1199f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1200112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1201f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1202bf4ede01STejun Heo } 1203018f3a13SLai Jiangshan } 1204018f3a13SLai Jiangshan 1205018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1206bf4ede01STejun Heo 1207bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1208112202d9STejun Heo if (likely(pwq->flush_color != color)) 12098864b4e5STejun Heo goto out_put; 1210bf4ede01STejun Heo 1211bf4ede01STejun Heo /* are there still in-flight works? */ 1212112202d9STejun Heo if (pwq->nr_in_flight[color]) 12138864b4e5STejun Heo goto out_put; 1214bf4ede01STejun Heo 1215112202d9STejun Heo /* this pwq is done, clear flush_color */ 1216112202d9STejun Heo pwq->flush_color = -1; 1217bf4ede01STejun Heo 1218bf4ede01STejun Heo /* 1219112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1220bf4ede01STejun Heo * will handle the rest. 1221bf4ede01STejun Heo */ 1222112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1223112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12248864b4e5STejun Heo out_put: 12258864b4e5STejun Heo put_pwq(pwq); 1226bf4ede01STejun Heo } 1227bf4ede01STejun Heo 122836e227d2STejun Heo /** 1229bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 123036e227d2STejun Heo * @work: work item to steal 123136e227d2STejun Heo * @is_dwork: @work is a delayed_work 1232bbb68dfaSTejun Heo * @flags: place to store irq state 123336e227d2STejun Heo * 123436e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1235d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 123636e227d2STejun Heo * 1237d185af30SYacine Belkadi * Return: 12383eb6b31bSMauro Carvalho Chehab * 12393eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 124036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 124136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 124236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1243bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1244bbb68dfaSTejun Heo * for arbitrarily long 12453eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 124636e227d2STejun Heo * 1247d185af30SYacine Belkadi * Note: 1248bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1249e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1250e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1251e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1252bbb68dfaSTejun Heo * 1253bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1254bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1255bbb68dfaSTejun Heo * 1256e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1257bf4ede01STejun Heo */ 1258bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1259bbb68dfaSTejun Heo unsigned long *flags) 1260bf4ede01STejun Heo { 1261d565ed63STejun Heo struct worker_pool *pool; 1262112202d9STejun Heo struct pool_workqueue *pwq; 1263bf4ede01STejun Heo 1264bbb68dfaSTejun Heo local_irq_save(*flags); 1265bbb68dfaSTejun Heo 126636e227d2STejun Heo /* try to steal the timer if it exists */ 126736e227d2STejun Heo if (is_dwork) { 126836e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 126936e227d2STejun Heo 1270e0aecdd8STejun Heo /* 1271e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1272e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1273e0aecdd8STejun Heo * running on the local CPU. 1274e0aecdd8STejun Heo */ 127536e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 127636e227d2STejun Heo return 1; 127736e227d2STejun Heo } 127836e227d2STejun Heo 127936e227d2STejun Heo /* try to claim PENDING the normal way */ 1280bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1281bf4ede01STejun Heo return 0; 1282bf4ede01STejun Heo 128324acfb71SThomas Gleixner rcu_read_lock(); 1284bf4ede01STejun Heo /* 1285bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1286bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1287bf4ede01STejun Heo */ 1288d565ed63STejun Heo pool = get_work_pool(work); 1289d565ed63STejun Heo if (!pool) 1290bbb68dfaSTejun Heo goto fail; 1291bf4ede01STejun Heo 1292a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1293bf4ede01STejun Heo /* 1294112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1295112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1296112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1297112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1298112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12990b3dae68SLai Jiangshan * item is currently queued on that pool. 1300bf4ede01STejun Heo */ 1301112202d9STejun Heo pwq = get_work_pwq(work); 1302112202d9STejun Heo if (pwq && pwq->pool == pool) { 1303bf4ede01STejun Heo debug_work_deactivate(work); 13043aa62497SLai Jiangshan 13053aa62497SLai Jiangshan /* 1306018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1307018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1308018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1309018f3a13SLai Jiangshan * 1310f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1311d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1312f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 131316062836STejun Heo * management later on and cause stall. Make sure the work 131416062836STejun Heo * item is activated before grabbing. 13153aa62497SLai Jiangshan */ 1316f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1317f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13183aa62497SLai Jiangshan 1319bf4ede01STejun Heo list_del_init(&work->entry); 1320c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 132136e227d2STejun Heo 1322112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 13234468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 13244468a00fSLai Jiangshan 1325a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 132624acfb71SThomas Gleixner rcu_read_unlock(); 132736e227d2STejun Heo return 1; 1328bf4ede01STejun Heo } 1329a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1330bbb68dfaSTejun Heo fail: 133124acfb71SThomas Gleixner rcu_read_unlock(); 1332bbb68dfaSTejun Heo local_irq_restore(*flags); 1333bbb68dfaSTejun Heo if (work_is_canceling(work)) 1334bbb68dfaSTejun Heo return -ENOENT; 1335bbb68dfaSTejun Heo cpu_relax(); 133636e227d2STejun Heo return -EAGAIN; 1337bf4ede01STejun Heo } 1338bf4ede01STejun Heo 1339bf4ede01STejun Heo /** 1340706026c2STejun Heo * insert_work - insert a work into a pool 1341112202d9STejun Heo * @pwq: pwq @work belongs to 13424690c4abSTejun Heo * @work: work to insert 13434690c4abSTejun Heo * @head: insertion point 13444690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13454690c4abSTejun Heo * 1346112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1347706026c2STejun Heo * work_struct flags. 13484690c4abSTejun Heo * 13494690c4abSTejun Heo * CONTEXT: 1350a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1351365970a1SDavid Howells */ 1352112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1353112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1354b89deed3SOleg Nesterov { 1355112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1356e1d8aa9fSFrederic Weisbecker 1357e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1358f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1359e89a85d6SWalter Wu 13604690c4abSTejun Heo /* we own @work, set data and link */ 1361112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13621a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13638864b4e5STejun Heo get_pwq(pwq); 1364e22bee78STejun Heo 136563d95a91STejun Heo if (__need_more_worker(pool)) 136663d95a91STejun Heo wake_up_worker(pool); 1367b89deed3SOleg Nesterov } 1368b89deed3SOleg Nesterov 1369c8efcc25STejun Heo /* 1370c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13718d03ecfeSTejun Heo * same workqueue. 1372c8efcc25STejun Heo */ 1373c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1374c8efcc25STejun Heo { 1375c8efcc25STejun Heo struct worker *worker; 1376c8efcc25STejun Heo 13778d03ecfeSTejun Heo worker = current_wq_worker(); 1378c8efcc25STejun Heo /* 1379bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 13808d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1381c8efcc25STejun Heo */ 1382112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1383c8efcc25STejun Heo } 1384c8efcc25STejun Heo 1385ef557180SMike Galbraith /* 1386ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1387ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1388ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1389ef557180SMike Galbraith */ 1390ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1391ef557180SMike Galbraith { 1392f303fccbSTejun Heo static bool printed_dbg_warning; 1393ef557180SMike Galbraith int new_cpu; 1394ef557180SMike Galbraith 1395f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1396ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1397ef557180SMike Galbraith return cpu; 1398f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1399f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1400f303fccbSTejun Heo printed_dbg_warning = true; 1401f303fccbSTejun Heo } 1402f303fccbSTejun Heo 1403ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1404ef557180SMike Galbraith return cpu; 1405ef557180SMike Galbraith 1406ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1407ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1408ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1409ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1410ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1411ef557180SMike Galbraith return cpu; 1412ef557180SMike Galbraith } 1413ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1414ef557180SMike Galbraith 1415ef557180SMike Galbraith return new_cpu; 1416ef557180SMike Galbraith } 1417ef557180SMike Galbraith 1418d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 14191da177e4SLinus Torvalds struct work_struct *work) 14201da177e4SLinus Torvalds { 1421112202d9STejun Heo struct pool_workqueue *pwq; 1422c9178087STejun Heo struct worker_pool *last_pool; 14231e19ffc6STejun Heo struct list_head *worklist; 14248a2e8e5dSTejun Heo unsigned int work_flags; 1425b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 14268930cabaSTejun Heo 14278930cabaSTejun Heo /* 14288930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 14298930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 14308930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 14318930cabaSTejun Heo * happen with IRQ disabled. 14328930cabaSTejun Heo */ 14338e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 14341da177e4SLinus Torvalds 14351e19ffc6STejun Heo 14369ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1437618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1438c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1439e41e704bSTejun Heo return; 144024acfb71SThomas Gleixner rcu_read_lock(); 14419e8cd2f5STejun Heo retry: 1442aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1443aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1444df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1445ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1446df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1447aa202f1fSHillf Danton } else { 1448aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1449aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1450aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1451aa202f1fSHillf Danton } 1452f3421797STejun Heo 145318aa9effSTejun Heo /* 1454c9178087STejun Heo * If @work was previously on a different pool, it might still be 1455c9178087STejun Heo * running there, in which case the work needs to be queued on that 1456c9178087STejun Heo * pool to guarantee non-reentrancy. 145718aa9effSTejun Heo */ 1458c9e7cf27STejun Heo last_pool = get_work_pool(work); 1459112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 146018aa9effSTejun Heo struct worker *worker; 146118aa9effSTejun Heo 1462a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 146318aa9effSTejun Heo 1464c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 146518aa9effSTejun Heo 1466112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1467c9178087STejun Heo pwq = worker->current_pwq; 14688594fadeSLai Jiangshan } else { 146918aa9effSTejun Heo /* meh... not running there, queue here */ 1470a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1471a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 147218aa9effSTejun Heo } 14738930cabaSTejun Heo } else { 1474a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 14758930cabaSTejun Heo } 1476502ca9d8STejun Heo 14779e8cd2f5STejun Heo /* 14789e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14799e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14809e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1481df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1482df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14839e8cd2f5STejun Heo * make forward-progress. 14849e8cd2f5STejun Heo */ 14859e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14869e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1487a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 14889e8cd2f5STejun Heo cpu_relax(); 14899e8cd2f5STejun Heo goto retry; 14909e8cd2f5STejun Heo } 14919e8cd2f5STejun Heo /* oops */ 14929e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14939e8cd2f5STejun Heo wq->name, cpu); 14949e8cd2f5STejun Heo } 14959e8cd2f5STejun Heo 1496112202d9STejun Heo /* pwq determined, queue */ 1497112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1498502ca9d8STejun Heo 149924acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 150024acfb71SThomas Gleixner goto out; 15011e19ffc6STejun Heo 1502112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1503112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 15041e19ffc6STejun Heo 1505112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1506cdadf009STejun Heo trace_workqueue_activate_work(work); 1507112202d9STejun Heo pwq->nr_active++; 1508112202d9STejun Heo worklist = &pwq->pool->worklist; 150982607adcSTejun Heo if (list_empty(worklist)) 151082607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 15118a2e8e5dSTejun Heo } else { 1512f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1513f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 15148a2e8e5dSTejun Heo } 15151e19ffc6STejun Heo 15160687c66bSZqiang debug_work_activate(work); 1517112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 15181e19ffc6STejun Heo 151924acfb71SThomas Gleixner out: 1520a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 152124acfb71SThomas Gleixner rcu_read_unlock(); 15221da177e4SLinus Torvalds } 15231da177e4SLinus Torvalds 15240fcb78c2SRolf Eike Beer /** 1525c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1526c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1527c1a220e7SZhang Rui * @wq: workqueue to use 1528c1a220e7SZhang Rui * @work: work to queue 1529c1a220e7SZhang Rui * 1530c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1531443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1532443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1533d185af30SYacine Belkadi * 1534d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1535c1a220e7SZhang Rui */ 1536d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1537d4283e93STejun Heo struct work_struct *work) 1538c1a220e7SZhang Rui { 1539d4283e93STejun Heo bool ret = false; 15408930cabaSTejun Heo unsigned long flags; 15418930cabaSTejun Heo 15428930cabaSTejun Heo local_irq_save(flags); 1543c1a220e7SZhang Rui 154422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15454690c4abSTejun Heo __queue_work(cpu, wq, work); 1546d4283e93STejun Heo ret = true; 1547c1a220e7SZhang Rui } 15488930cabaSTejun Heo 15498930cabaSTejun Heo local_irq_restore(flags); 1550c1a220e7SZhang Rui return ret; 1551c1a220e7SZhang Rui } 1552ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1553c1a220e7SZhang Rui 15548204e0c1SAlexander Duyck /** 15558204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15568204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15578204e0c1SAlexander Duyck * 15588204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15598204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15608204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15618204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15628204e0c1SAlexander Duyck */ 15638204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15648204e0c1SAlexander Duyck { 15658204e0c1SAlexander Duyck int cpu; 15668204e0c1SAlexander Duyck 15678204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15688204e0c1SAlexander Duyck if (!wq_numa_enabled) 15698204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15708204e0c1SAlexander Duyck 15718204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15728204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15738204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15748204e0c1SAlexander Duyck 15758204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15768204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15778204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15788204e0c1SAlexander Duyck return cpu; 15798204e0c1SAlexander Duyck 15808204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15818204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15828204e0c1SAlexander Duyck 15838204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15848204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15858204e0c1SAlexander Duyck } 15868204e0c1SAlexander Duyck 15878204e0c1SAlexander Duyck /** 15888204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15898204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15908204e0c1SAlexander Duyck * @wq: workqueue to use 15918204e0c1SAlexander Duyck * @work: work to queue 15928204e0c1SAlexander Duyck * 15938204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 15948204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 15958204e0c1SAlexander Duyck * NUMA node. 15968204e0c1SAlexander Duyck * 15978204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 15988204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 15998204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 16008204e0c1SAlexander Duyck * 16018204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 16028204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 16038204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 16048204e0c1SAlexander Duyck * 16058204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 16068204e0c1SAlexander Duyck */ 16078204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 16088204e0c1SAlexander Duyck struct work_struct *work) 16098204e0c1SAlexander Duyck { 16108204e0c1SAlexander Duyck unsigned long flags; 16118204e0c1SAlexander Duyck bool ret = false; 16128204e0c1SAlexander Duyck 16138204e0c1SAlexander Duyck /* 16148204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 16158204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 16168204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 16178204e0c1SAlexander Duyck * 16188204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 16198204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 16208204e0c1SAlexander Duyck * some round robin type logic. 16218204e0c1SAlexander Duyck */ 16228204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 16238204e0c1SAlexander Duyck 16248204e0c1SAlexander Duyck local_irq_save(flags); 16258204e0c1SAlexander Duyck 16268204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16278204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16288204e0c1SAlexander Duyck 16298204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16308204e0c1SAlexander Duyck ret = true; 16318204e0c1SAlexander Duyck } 16328204e0c1SAlexander Duyck 16338204e0c1SAlexander Duyck local_irq_restore(flags); 16348204e0c1SAlexander Duyck return ret; 16358204e0c1SAlexander Duyck } 16368204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16378204e0c1SAlexander Duyck 16388c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16391da177e4SLinus Torvalds { 16408c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16411da177e4SLinus Torvalds 1642e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 164360c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16441da177e4SLinus Torvalds } 16451438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16461da177e4SLinus Torvalds 16477beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 164852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16491da177e4SLinus Torvalds { 16507beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16517beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16521da177e4SLinus Torvalds 1653637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 16544b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1655fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1656fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16577beb2edfSTejun Heo 16588852aac2STejun Heo /* 16598852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16608852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16618852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16628852aac2STejun Heo * on that there's no such delay when @delay is 0. 16638852aac2STejun Heo */ 16648852aac2STejun Heo if (!delay) { 16658852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16668852aac2STejun Heo return; 16678852aac2STejun Heo } 16688852aac2STejun Heo 166960c057bcSLai Jiangshan dwork->wq = wq; 16701265057fSTejun Heo dwork->cpu = cpu; 16717beb2edfSTejun Heo timer->expires = jiffies + delay; 16727beb2edfSTejun Heo 1673041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16747beb2edfSTejun Heo add_timer_on(timer, cpu); 1675041bd12eSTejun Heo else 1676041bd12eSTejun Heo add_timer(timer); 16777beb2edfSTejun Heo } 16781da177e4SLinus Torvalds 16790fcb78c2SRolf Eike Beer /** 16800fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16810fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16820fcb78c2SRolf Eike Beer * @wq: workqueue to use 1683af9997e4SRandy Dunlap * @dwork: work to queue 16840fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16850fcb78c2SRolf Eike Beer * 1686d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1687715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1688715f1300STejun Heo * execution. 16890fcb78c2SRolf Eike Beer */ 1690d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 169152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16927a6bc1cdSVenkatesh Pallipadi { 169352bad64dSDavid Howells struct work_struct *work = &dwork->work; 1694d4283e93STejun Heo bool ret = false; 16958930cabaSTejun Heo unsigned long flags; 16968930cabaSTejun Heo 16978930cabaSTejun Heo /* read the comment in __queue_work() */ 16988930cabaSTejun Heo local_irq_save(flags); 16997a6bc1cdSVenkatesh Pallipadi 170022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17017beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1702d4283e93STejun Heo ret = true; 17037a6bc1cdSVenkatesh Pallipadi } 17048930cabaSTejun Heo 17058930cabaSTejun Heo local_irq_restore(flags); 17067a6bc1cdSVenkatesh Pallipadi return ret; 17077a6bc1cdSVenkatesh Pallipadi } 1708ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 17091da177e4SLinus Torvalds 1710c8e55f36STejun Heo /** 17118376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 17128376fe22STejun Heo * @cpu: CPU number to execute work on 17138376fe22STejun Heo * @wq: workqueue to use 17148376fe22STejun Heo * @dwork: work to queue 17158376fe22STejun Heo * @delay: number of jiffies to wait before queueing 17168376fe22STejun Heo * 17178376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 17188376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 17198376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 17208376fe22STejun Heo * current state. 17218376fe22STejun Heo * 1722d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 17238376fe22STejun Heo * pending and its timer was modified. 17248376fe22STejun Heo * 1725e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17268376fe22STejun Heo * See try_to_grab_pending() for details. 17278376fe22STejun Heo */ 17288376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17298376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17308376fe22STejun Heo { 17318376fe22STejun Heo unsigned long flags; 17328376fe22STejun Heo int ret; 17338376fe22STejun Heo 17348376fe22STejun Heo do { 17358376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17368376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17378376fe22STejun Heo 17388376fe22STejun Heo if (likely(ret >= 0)) { 17398376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17408376fe22STejun Heo local_irq_restore(flags); 17418376fe22STejun Heo } 17428376fe22STejun Heo 17438376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17448376fe22STejun Heo return ret; 17458376fe22STejun Heo } 17468376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17478376fe22STejun Heo 174805f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 174905f0fe6bSTejun Heo { 175005f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 175105f0fe6bSTejun Heo 175205f0fe6bSTejun Heo /* read the comment in __queue_work() */ 175305f0fe6bSTejun Heo local_irq_disable(); 175405f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 175505f0fe6bSTejun Heo local_irq_enable(); 175605f0fe6bSTejun Heo } 175705f0fe6bSTejun Heo 175805f0fe6bSTejun Heo /** 175905f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 176005f0fe6bSTejun Heo * @wq: workqueue to use 176105f0fe6bSTejun Heo * @rwork: work to queue 176205f0fe6bSTejun Heo * 176305f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 176405f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1765bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 176605f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 176705f0fe6bSTejun Heo */ 176805f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 176905f0fe6bSTejun Heo { 177005f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 177105f0fe6bSTejun Heo 177205f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 177305f0fe6bSTejun Heo rwork->wq = wq; 1774*a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 177505f0fe6bSTejun Heo return true; 177605f0fe6bSTejun Heo } 177705f0fe6bSTejun Heo 177805f0fe6bSTejun Heo return false; 177905f0fe6bSTejun Heo } 178005f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 178105f0fe6bSTejun Heo 17828376fe22STejun Heo /** 1783c8e55f36STejun Heo * worker_enter_idle - enter idle state 1784c8e55f36STejun Heo * @worker: worker which is entering idle state 1785c8e55f36STejun Heo * 1786c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1787c8e55f36STejun Heo * necessary. 1788c8e55f36STejun Heo * 1789c8e55f36STejun Heo * LOCKING: 1790a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1791c8e55f36STejun Heo */ 1792c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 17931da177e4SLinus Torvalds { 1794bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1795c8e55f36STejun Heo 17966183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 17976183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 17986183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 17996183c009STejun Heo return; 1800c8e55f36STejun Heo 1801051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1802cb444766STejun Heo worker->flags |= WORKER_IDLE; 1803bd7bdd43STejun Heo pool->nr_idle++; 1804e22bee78STejun Heo worker->last_active = jiffies; 1805c8e55f36STejun Heo 1806c8e55f36STejun Heo /* idle_list is LIFO */ 1807bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1808db7bccf4STejun Heo 180963d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1810628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1811cb444766STejun Heo 1812989442d7SLai Jiangshan /* Sanity check nr_running. */ 1813bc35f7efSLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && 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 /** 2791c4f135d6STetsuo Handa * __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 */ 2797c4f135d6STetsuo Handa 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 } 2946c4f135d6STetsuo Handa 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: 2974c4f135d6STetsuo Handa __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 lock_map_acquire(&work->lockdep_map); 307087915adcSJohannes Berg lock_map_release(&work->lockdep_map); 307187915adcSJohannes Berg 3072d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3073d6e89786SJohannes Berg wait_for_completion(&barr.done); 3074d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3075d6e89786SJohannes Berg return true; 3076d6e89786SJohannes Berg } else { 3077d6e89786SJohannes Berg return false; 3078d6e89786SJohannes Berg } 3079d6e89786SJohannes Berg } 3080d6e89786SJohannes Berg 3081db700897SOleg Nesterov /** 3082401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3083401a8d04STejun Heo * @work: the work to flush 3084db700897SOleg Nesterov * 3085606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3086606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3087401a8d04STejun Heo * 3088d185af30SYacine Belkadi * Return: 3089401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3090401a8d04STejun Heo * %false if it was already idle. 3091db700897SOleg Nesterov */ 3092401a8d04STejun Heo bool flush_work(struct work_struct *work) 3093db700897SOleg Nesterov { 3094d6e89786SJohannes Berg return __flush_work(work, false); 3095606a5020STejun Heo } 3096db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3097db700897SOleg Nesterov 30988603e1b3STejun Heo struct cwt_wait { 3099ac6424b9SIngo Molnar wait_queue_entry_t wait; 31008603e1b3STejun Heo struct work_struct *work; 31018603e1b3STejun Heo }; 31028603e1b3STejun Heo 3103ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 31048603e1b3STejun Heo { 31058603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 31068603e1b3STejun Heo 31078603e1b3STejun Heo if (cwait->work != key) 31088603e1b3STejun Heo return 0; 31098603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 31108603e1b3STejun Heo } 31118603e1b3STejun Heo 311236e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3113401a8d04STejun Heo { 31148603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3115bbb68dfaSTejun Heo unsigned long flags; 31161f1f642eSOleg Nesterov int ret; 31171f1f642eSOleg Nesterov 31181f1f642eSOleg Nesterov do { 3119bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3120bbb68dfaSTejun Heo /* 31218603e1b3STejun Heo * If someone else is already canceling, wait for it to 31228603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 31238603e1b3STejun Heo * because we may get scheduled between @work's completion 31248603e1b3STejun Heo * and the other canceling task resuming and clearing 31258603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 31268603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 31278603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 31288603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 31298603e1b3STejun Heo * we're hogging the CPU. 31308603e1b3STejun Heo * 31318603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 31328603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 31338603e1b3STejun Heo * wake function which matches @work along with exclusive 31348603e1b3STejun Heo * wait and wakeup. 3135bbb68dfaSTejun Heo */ 31368603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 31378603e1b3STejun Heo struct cwt_wait cwait; 31388603e1b3STejun Heo 31398603e1b3STejun Heo init_wait(&cwait.wait); 31408603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 31418603e1b3STejun Heo cwait.work = work; 31428603e1b3STejun Heo 31438603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 31448603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 31458603e1b3STejun Heo if (work_is_canceling(work)) 31468603e1b3STejun Heo schedule(); 31478603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 31488603e1b3STejun Heo } 31491f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 31501f1f642eSOleg Nesterov 3151bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3152bbb68dfaSTejun Heo mark_work_canceling(work); 3153bbb68dfaSTejun Heo local_irq_restore(flags); 3154bbb68dfaSTejun Heo 31553347fa09STejun Heo /* 31563347fa09STejun Heo * This allows canceling during early boot. We know that @work 31573347fa09STejun Heo * isn't executing. 31583347fa09STejun Heo */ 31593347fa09STejun Heo if (wq_online) 3160d6e89786SJohannes Berg __flush_work(work, true); 31613347fa09STejun Heo 31627a22ad75STejun Heo clear_work_data(work); 31638603e1b3STejun Heo 31648603e1b3STejun Heo /* 31658603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 31668603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 31678603e1b3STejun Heo * visible there. 31688603e1b3STejun Heo */ 31698603e1b3STejun Heo smp_mb(); 31708603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 31718603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 31728603e1b3STejun Heo 31731f1f642eSOleg Nesterov return ret; 31741f1f642eSOleg Nesterov } 31751f1f642eSOleg Nesterov 31766e84d644SOleg Nesterov /** 3177401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3178401a8d04STejun Heo * @work: the work to cancel 31796e84d644SOleg Nesterov * 3180401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3181401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3182401a8d04STejun Heo * another workqueue. On return from this function, @work is 3183401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 31841f1f642eSOleg Nesterov * 3185401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3186401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 31876e84d644SOleg Nesterov * 3188401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 31896e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3190401a8d04STejun Heo * 3191d185af30SYacine Belkadi * Return: 3192401a8d04STejun Heo * %true if @work was pending, %false otherwise. 31936e84d644SOleg Nesterov */ 3194401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 31956e84d644SOleg Nesterov { 319636e227d2STejun Heo return __cancel_work_timer(work, false); 3197b89deed3SOleg Nesterov } 319828e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3199b89deed3SOleg Nesterov 32006e84d644SOleg Nesterov /** 3201401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3202401a8d04STejun Heo * @dwork: the delayed work to flush 32036e84d644SOleg Nesterov * 3204401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3205401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3206401a8d04STejun Heo * considers the last queueing instance of @dwork. 32071f1f642eSOleg Nesterov * 3208d185af30SYacine Belkadi * Return: 3209401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3210401a8d04STejun Heo * %false if it was already idle. 32116e84d644SOleg Nesterov */ 3212401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3213401a8d04STejun Heo { 32148930cabaSTejun Heo local_irq_disable(); 3215401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 321660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 32178930cabaSTejun Heo local_irq_enable(); 3218401a8d04STejun Heo return flush_work(&dwork->work); 3219401a8d04STejun Heo } 3220401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3221401a8d04STejun Heo 322205f0fe6bSTejun Heo /** 322305f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 322405f0fe6bSTejun Heo * @rwork: the rcu work to flush 322505f0fe6bSTejun Heo * 322605f0fe6bSTejun Heo * Return: 322705f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 322805f0fe6bSTejun Heo * %false if it was already idle. 322905f0fe6bSTejun Heo */ 323005f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 323105f0fe6bSTejun Heo { 323205f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 323305f0fe6bSTejun Heo rcu_barrier(); 323405f0fe6bSTejun Heo flush_work(&rwork->work); 323505f0fe6bSTejun Heo return true; 323605f0fe6bSTejun Heo } else { 323705f0fe6bSTejun Heo return flush_work(&rwork->work); 323805f0fe6bSTejun Heo } 323905f0fe6bSTejun Heo } 324005f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 324105f0fe6bSTejun Heo 3242f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3243f72b8792SJens Axboe { 3244f72b8792SJens Axboe unsigned long flags; 3245f72b8792SJens Axboe int ret; 3246f72b8792SJens Axboe 3247f72b8792SJens Axboe do { 3248f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3249f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3250f72b8792SJens Axboe 3251f72b8792SJens Axboe if (unlikely(ret < 0)) 3252f72b8792SJens Axboe return false; 3253f72b8792SJens Axboe 3254f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3255f72b8792SJens Axboe local_irq_restore(flags); 3256f72b8792SJens Axboe return ret; 3257f72b8792SJens Axboe } 3258f72b8792SJens Axboe 325973b4b532SAndrey Grodzovsky /* 326073b4b532SAndrey Grodzovsky * See cancel_delayed_work() 326173b4b532SAndrey Grodzovsky */ 326273b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 326373b4b532SAndrey Grodzovsky { 326473b4b532SAndrey Grodzovsky return __cancel_work(work, false); 326573b4b532SAndrey Grodzovsky } 326673b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 326773b4b532SAndrey Grodzovsky 3268401a8d04STejun Heo /** 326957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 327057b30ae7STejun Heo * @dwork: delayed_work to cancel 327109383498STejun Heo * 3272d185af30SYacine Belkadi * Kill off a pending delayed_work. 3273d185af30SYacine Belkadi * 3274d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3275d185af30SYacine Belkadi * pending. 3276d185af30SYacine Belkadi * 3277d185af30SYacine Belkadi * Note: 3278d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3279d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3280d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 328109383498STejun Heo * 328257b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 328309383498STejun Heo */ 328457b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 328509383498STejun Heo { 3286f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 328709383498STejun Heo } 328857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 328909383498STejun Heo 329009383498STejun Heo /** 3291401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3292401a8d04STejun Heo * @dwork: the delayed work cancel 3293401a8d04STejun Heo * 3294401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3295401a8d04STejun Heo * 3296d185af30SYacine Belkadi * Return: 3297401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3298401a8d04STejun Heo */ 3299401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 33006e84d644SOleg Nesterov { 330136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 33026e84d644SOleg Nesterov } 3303f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 33041da177e4SLinus Torvalds 33050fcb78c2SRolf Eike Beer /** 330631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3307b6136773SAndrew Morton * @func: the function to call 3308b6136773SAndrew Morton * 330931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 331031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3311b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 331231ddd871STejun Heo * 3313d185af30SYacine Belkadi * Return: 331431ddd871STejun Heo * 0 on success, -errno on failure. 3315b6136773SAndrew Morton */ 331665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 331715316ba8SChristoph Lameter { 331815316ba8SChristoph Lameter int cpu; 331938f51568SNamhyung Kim struct work_struct __percpu *works; 332015316ba8SChristoph Lameter 3321b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3322b6136773SAndrew Morton if (!works) 332315316ba8SChristoph Lameter return -ENOMEM; 3324b6136773SAndrew Morton 3325ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 332693981800STejun Heo 332715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 33289bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 33299bfb1839SIngo Molnar 33309bfb1839SIngo Molnar INIT_WORK(work, func); 33318de6d308SOleg Nesterov schedule_work_on(cpu, work); 333215316ba8SChristoph Lameter } 333393981800STejun Heo 333493981800STejun Heo for_each_online_cpu(cpu) 33358616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 333693981800STejun Heo 3337ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3338b6136773SAndrew Morton free_percpu(works); 333915316ba8SChristoph Lameter return 0; 334015316ba8SChristoph Lameter } 334115316ba8SChristoph Lameter 3342eef6a7d5SAlan Stern /** 33431fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 33441fa44ecaSJames Bottomley * @fn: the function to execute 33451fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 33461fa44ecaSJames Bottomley * be available when the work executes) 33471fa44ecaSJames Bottomley * 33481fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 33491fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 33501fa44ecaSJames Bottomley * 3351d185af30SYacine Belkadi * Return: 0 - function was executed 33521fa44ecaSJames Bottomley * 1 - function was scheduled for execution 33531fa44ecaSJames Bottomley */ 335465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 33551fa44ecaSJames Bottomley { 33561fa44ecaSJames Bottomley if (!in_interrupt()) { 335765f27f38SDavid Howells fn(&ew->work); 33581fa44ecaSJames Bottomley return 0; 33591fa44ecaSJames Bottomley } 33601fa44ecaSJames Bottomley 336165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 33621fa44ecaSJames Bottomley schedule_work(&ew->work); 33631fa44ecaSJames Bottomley 33641fa44ecaSJames Bottomley return 1; 33651fa44ecaSJames Bottomley } 33661fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 33671fa44ecaSJames Bottomley 33687a4e344cSTejun Heo /** 33697a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33707a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33717a4e344cSTejun Heo * 33727a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33737a4e344cSTejun Heo */ 3374513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 33757a4e344cSTejun Heo { 33767a4e344cSTejun Heo if (attrs) { 33777a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33787a4e344cSTejun Heo kfree(attrs); 33797a4e344cSTejun Heo } 33807a4e344cSTejun Heo } 33817a4e344cSTejun Heo 33827a4e344cSTejun Heo /** 33837a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33847a4e344cSTejun Heo * 33857a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3386d185af30SYacine Belkadi * return it. 3387d185af30SYacine Belkadi * 3388d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33897a4e344cSTejun Heo */ 3390513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 33917a4e344cSTejun Heo { 33927a4e344cSTejun Heo struct workqueue_attrs *attrs; 33937a4e344cSTejun Heo 3394be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 33957a4e344cSTejun Heo if (!attrs) 33967a4e344cSTejun Heo goto fail; 3397be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 33987a4e344cSTejun Heo goto fail; 33997a4e344cSTejun Heo 340013e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 34017a4e344cSTejun Heo return attrs; 34027a4e344cSTejun Heo fail: 34037a4e344cSTejun Heo free_workqueue_attrs(attrs); 34047a4e344cSTejun Heo return NULL; 34057a4e344cSTejun Heo } 34067a4e344cSTejun Heo 340729c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 340829c91e99STejun Heo const struct workqueue_attrs *from) 340929c91e99STejun Heo { 341029c91e99STejun Heo to->nice = from->nice; 341129c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 34122865a8fbSShaohua Li /* 34132865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 34142865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 34152865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 34162865a8fbSShaohua Li */ 34172865a8fbSShaohua Li to->no_numa = from->no_numa; 341829c91e99STejun Heo } 341929c91e99STejun Heo 342029c91e99STejun Heo /* hash value of the content of @attr */ 342129c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 342229c91e99STejun Heo { 342329c91e99STejun Heo u32 hash = 0; 342429c91e99STejun Heo 342529c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 342613e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 342713e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 342829c91e99STejun Heo return hash; 342929c91e99STejun Heo } 343029c91e99STejun Heo 343129c91e99STejun Heo /* content equality test */ 343229c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 343329c91e99STejun Heo const struct workqueue_attrs *b) 343429c91e99STejun Heo { 343529c91e99STejun Heo if (a->nice != b->nice) 343629c91e99STejun Heo return false; 343729c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 343829c91e99STejun Heo return false; 343929c91e99STejun Heo return true; 344029c91e99STejun Heo } 344129c91e99STejun Heo 34427a4e344cSTejun Heo /** 34437a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34447a4e344cSTejun Heo * @pool: worker_pool to initialize 34457a4e344cSTejun Heo * 3446402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3447d185af30SYacine Belkadi * 3448d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 344929c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 345029c91e99STejun Heo * on @pool safely to release it. 34517a4e344cSTejun Heo */ 34527a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34534e1a1f9aSTejun Heo { 3454a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 345529c91e99STejun Heo pool->id = -1; 345629c91e99STejun Heo pool->cpu = -1; 3457f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34584e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 345982607adcSTejun Heo pool->watchdog_ts = jiffies; 34604e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34614e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34624e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34634e1a1f9aSTejun Heo 346432a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 34654e1a1f9aSTejun Heo 346632a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 34674e1a1f9aSTejun Heo 3468da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34697a4e344cSTejun Heo 34707cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 347129c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 347229c91e99STejun Heo pool->refcnt = 1; 347329c91e99STejun Heo 347429c91e99STejun Heo /* shouldn't fail above this point */ 3475be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 34767a4e344cSTejun Heo if (!pool->attrs) 34777a4e344cSTejun Heo return -ENOMEM; 34787a4e344cSTejun Heo return 0; 34794e1a1f9aSTejun Heo } 34804e1a1f9aSTejun Heo 3481669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3482669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3483669de8bdSBart Van Assche { 3484669de8bdSBart Van Assche char *lock_name; 3485669de8bdSBart Van Assche 3486669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3487669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3488669de8bdSBart Van Assche if (!lock_name) 3489669de8bdSBart Van Assche lock_name = wq->name; 349069a106c0SQian Cai 349169a106c0SQian Cai wq->lock_name = lock_name; 3492669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3493669de8bdSBart Van Assche } 3494669de8bdSBart Van Assche 3495669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3496669de8bdSBart Van Assche { 3497669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3498669de8bdSBart Van Assche } 3499669de8bdSBart Van Assche 3500669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3501669de8bdSBart Van Assche { 3502669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3503669de8bdSBart Van Assche kfree(wq->lock_name); 3504669de8bdSBart Van Assche } 3505669de8bdSBart Van Assche #else 3506669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3507669de8bdSBart Van Assche { 3508669de8bdSBart Van Assche } 3509669de8bdSBart Van Assche 3510669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3511669de8bdSBart Van Assche { 3512669de8bdSBart Van Assche } 3513669de8bdSBart Van Assche 3514669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3515669de8bdSBart Van Assche { 3516669de8bdSBart Van Assche } 3517669de8bdSBart Van Assche #endif 3518669de8bdSBart Van Assche 3519e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3520e2dca7adSTejun Heo { 3521e2dca7adSTejun Heo struct workqueue_struct *wq = 3522e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3523e2dca7adSTejun Heo 3524669de8bdSBart Van Assche wq_free_lockdep(wq); 3525669de8bdSBart Van Assche 3526e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3527e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3528e2dca7adSTejun Heo else 3529e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3530e2dca7adSTejun Heo 3531e2dca7adSTejun Heo kfree(wq); 3532e2dca7adSTejun Heo } 3533e2dca7adSTejun Heo 353429c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 353529c91e99STejun Heo { 353629c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 353729c91e99STejun Heo 35387cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 353929c91e99STejun Heo free_workqueue_attrs(pool->attrs); 354029c91e99STejun Heo kfree(pool); 354129c91e99STejun Heo } 354229c91e99STejun Heo 3543d8bb65abSSebastian Andrzej Siewior /* This returns with the lock held on success (pool manager is inactive). */ 3544d8bb65abSSebastian Andrzej Siewior static bool wq_manager_inactive(struct worker_pool *pool) 3545d8bb65abSSebastian Andrzej Siewior { 3546a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 3547d8bb65abSSebastian Andrzej Siewior 3548d8bb65abSSebastian Andrzej Siewior if (pool->flags & POOL_MANAGER_ACTIVE) { 3549a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3550d8bb65abSSebastian Andrzej Siewior return false; 3551d8bb65abSSebastian Andrzej Siewior } 3552d8bb65abSSebastian Andrzej Siewior return true; 3553d8bb65abSSebastian Andrzej Siewior } 3554d8bb65abSSebastian Andrzej Siewior 355529c91e99STejun Heo /** 355629c91e99STejun Heo * put_unbound_pool - put a worker_pool 355729c91e99STejun Heo * @pool: worker_pool to put 355829c91e99STejun Heo * 355924acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3560c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3561c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3562c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3563a892caccSTejun Heo * 3564a892caccSTejun Heo * Should be called with wq_pool_mutex held. 356529c91e99STejun Heo */ 356629c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 356729c91e99STejun Heo { 356860f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 356929c91e99STejun Heo struct worker *worker; 357029c91e99STejun Heo 3571a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3572a892caccSTejun Heo 3573a892caccSTejun Heo if (--pool->refcnt) 357429c91e99STejun Heo return; 357529c91e99STejun Heo 357629c91e99STejun Heo /* sanity checks */ 357761d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3578a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 357929c91e99STejun Heo return; 358029c91e99STejun Heo 358129c91e99STejun Heo /* release id and unhash */ 358229c91e99STejun Heo if (pool->id >= 0) 358329c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 358429c91e99STejun Heo hash_del(&pool->hash_node); 358529c91e99STejun Heo 3586c5aa87bbSTejun Heo /* 3587692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3588692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3589692b4825STejun Heo * manager and @pool gets freed with the flag set. 3590d8bb65abSSebastian Andrzej Siewior * Because of how wq_manager_inactive() works, we will hold the 3591d8bb65abSSebastian Andrzej Siewior * spinlock after a successful wait. 3592c5aa87bbSTejun Heo */ 3593d8bb65abSSebastian Andrzej Siewior rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool), 3594d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3595692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3596692b4825STejun Heo 35971037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 359829c91e99STejun Heo destroy_worker(worker); 359929c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3600a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 360160f5a4bcSLai Jiangshan 36021258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3603da028469SLai Jiangshan if (!list_empty(&pool->workers)) 360460f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 36051258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 360660f5a4bcSLai Jiangshan 360760f5a4bcSLai Jiangshan if (pool->detach_completion) 360860f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 360960f5a4bcSLai Jiangshan 361029c91e99STejun Heo /* shut down the timers */ 361129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 361229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 361329c91e99STejun Heo 361424acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 361525b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 361629c91e99STejun Heo } 361729c91e99STejun Heo 361829c91e99STejun Heo /** 361929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 362029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 362129c91e99STejun Heo * 362229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 362329c91e99STejun Heo * reference count and return it. If there already is a matching 362429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3625d185af30SYacine Belkadi * create a new one. 3626a892caccSTejun Heo * 3627a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3628d185af30SYacine Belkadi * 3629d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3630d185af30SYacine Belkadi * On failure, %NULL. 363129c91e99STejun Heo */ 363229c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 363329c91e99STejun Heo { 363429c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 363529c91e99STejun Heo struct worker_pool *pool; 3636f3f90ad4STejun Heo int node; 3637e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 363829c91e99STejun Heo 3639a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 364029c91e99STejun Heo 364129c91e99STejun Heo /* do we already have a matching pool? */ 364229c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 364329c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 364429c91e99STejun Heo pool->refcnt++; 36453fb1823cSLai Jiangshan return pool; 364629c91e99STejun Heo } 364729c91e99STejun Heo } 364829c91e99STejun Heo 3649e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3650e2273584SXunlei Pang if (wq_numa_enabled) { 3651e2273584SXunlei Pang for_each_node(node) { 3652e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3653e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3654e2273584SXunlei Pang target_node = node; 3655e2273584SXunlei Pang break; 3656e2273584SXunlei Pang } 3657e2273584SXunlei Pang } 3658e2273584SXunlei Pang } 3659e2273584SXunlei Pang 366029c91e99STejun Heo /* nope, create a new one */ 3661e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 366229c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 366329c91e99STejun Heo goto fail; 366429c91e99STejun Heo 36658864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 366629c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3667e2273584SXunlei Pang pool->node = target_node; 366829c91e99STejun Heo 36692865a8fbSShaohua Li /* 36702865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36712865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36722865a8fbSShaohua Li */ 36732865a8fbSShaohua Li pool->attrs->no_numa = false; 36742865a8fbSShaohua Li 367529c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 367629c91e99STejun Heo goto fail; 367729c91e99STejun Heo 367829c91e99STejun Heo /* create and start the initial worker */ 36793347fa09STejun Heo if (wq_online && !create_worker(pool)) 368029c91e99STejun Heo goto fail; 368129c91e99STejun Heo 368229c91e99STejun Heo /* install */ 368329c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 36843fb1823cSLai Jiangshan 368529c91e99STejun Heo return pool; 368629c91e99STejun Heo fail: 368729c91e99STejun Heo if (pool) 368829c91e99STejun Heo put_unbound_pool(pool); 368929c91e99STejun Heo return NULL; 369029c91e99STejun Heo } 369129c91e99STejun Heo 36928864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36938864b4e5STejun Heo { 36948864b4e5STejun Heo kmem_cache_free(pwq_cache, 36958864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36968864b4e5STejun Heo } 36978864b4e5STejun Heo 36988864b4e5STejun Heo /* 36998864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 37008864b4e5STejun Heo * and needs to be destroyed. 37018864b4e5STejun Heo */ 37028864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 37038864b4e5STejun Heo { 37048864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 37058864b4e5STejun Heo unbound_release_work); 37068864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 37078864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3708b42b0bddSYang Yingliang bool is_last = false; 37098864b4e5STejun Heo 3710b42b0bddSYang Yingliang /* 3711b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 3712b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3713b42b0bddSYang Yingliang */ 3714b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 37158864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 37168864b4e5STejun Heo return; 37178864b4e5STejun Heo 37183c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 37198864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3720bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 37213c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3722b42b0bddSYang Yingliang } 37238864b4e5STejun Heo 3724a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 37258864b4e5STejun Heo put_unbound_pool(pool); 3726a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3727a892caccSTejun Heo 372825b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 37298864b4e5STejun Heo 37308864b4e5STejun Heo /* 37318864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3732e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 37338864b4e5STejun Heo */ 3734669de8bdSBart Van Assche if (is_last) { 3735669de8bdSBart Van Assche wq_unregister_lockdep(wq); 373625b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 37376029a918STejun Heo } 3738669de8bdSBart Van Assche } 37398864b4e5STejun Heo 37400fbd95aaSTejun Heo /** 3741699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37420fbd95aaSTejun Heo * @pwq: target pool_workqueue 37430fbd95aaSTejun Heo * 3744699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3745f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 3746699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37470fbd95aaSTejun Heo */ 3748699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37490fbd95aaSTejun Heo { 3750699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3751699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 37523347fa09STejun Heo unsigned long flags; 3753699ce097STejun Heo 3754699ce097STejun Heo /* for @wq->saved_max_active */ 3755a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3756699ce097STejun Heo 3757699ce097STejun Heo /* fast exit for non-freezable wqs */ 3758699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3759699ce097STejun Heo return; 3760699ce097STejun Heo 37613347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3762a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3763699ce097STejun Heo 376474b414eaSLai Jiangshan /* 376574b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 376674b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 376774b414eaSLai Jiangshan * is updated and visible. 376874b414eaSLai Jiangshan */ 376974b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 377001341fbdSYunfeng Ye bool kick = false; 377101341fbdSYunfeng Ye 3772699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37730fbd95aaSTejun Heo 3774f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 377501341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 3776f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 377701341fbdSYunfeng Ye kick = true; 377801341fbdSYunfeng Ye } 3779951a078aSLai Jiangshan 3780951a078aSLai Jiangshan /* 3781951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 378201341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 378301341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 378401341fbdSYunfeng Ye * let's kick iff work items were activated. 3785951a078aSLai Jiangshan */ 378601341fbdSYunfeng Ye if (kick) 3787951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3788699ce097STejun Heo } else { 3789699ce097STejun Heo pwq->max_active = 0; 3790699ce097STejun Heo } 3791699ce097STejun Heo 3792a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 37930fbd95aaSTejun Heo } 37940fbd95aaSTejun Heo 379567dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 3796f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3797f147f29eSTejun Heo struct worker_pool *pool) 3798d2c1d404STejun Heo { 3799d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3800d2c1d404STejun Heo 3801e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3802e50aba9aSTejun Heo 3803d2c1d404STejun Heo pwq->pool = pool; 3804d2c1d404STejun Heo pwq->wq = wq; 3805d2c1d404STejun Heo pwq->flush_color = -1; 38068864b4e5STejun Heo pwq->refcnt = 1; 3807f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 38081befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3809d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 38108864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3811f147f29eSTejun Heo } 3812d2c1d404STejun Heo 3813f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 38141befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3815f147f29eSTejun Heo { 3816f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3817f147f29eSTejun Heo 3818f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 381975ccf595STejun Heo 38201befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 38211befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 38221befcf30STejun Heo return; 38231befcf30STejun Heo 382429b1cb41SLai Jiangshan /* set the matching work_color */ 382575ccf595STejun Heo pwq->work_color = wq->work_color; 3826983ca25eSTejun Heo 3827983ca25eSTejun Heo /* sync max_active to the current setting */ 3828983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3829983ca25eSTejun Heo 3830983ca25eSTejun Heo /* link in @pwq */ 38319e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3832df2d5ae4STejun Heo } 38336029a918STejun Heo 3834f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3835f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3836f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3837f147f29eSTejun Heo { 3838f147f29eSTejun Heo struct worker_pool *pool; 3839f147f29eSTejun Heo struct pool_workqueue *pwq; 3840f147f29eSTejun Heo 3841f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3842f147f29eSTejun Heo 3843f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3844f147f29eSTejun Heo if (!pool) 3845f147f29eSTejun Heo return NULL; 3846f147f29eSTejun Heo 3847e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3848f147f29eSTejun Heo if (!pwq) { 3849f147f29eSTejun Heo put_unbound_pool(pool); 3850f147f29eSTejun Heo return NULL; 3851f147f29eSTejun Heo } 3852f147f29eSTejun Heo 3853f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3854f147f29eSTejun Heo return pwq; 3855d2c1d404STejun Heo } 3856d2c1d404STejun Heo 38574c16bd32STejun Heo /** 385830186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3859042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 38604c16bd32STejun Heo * @node: the target NUMA node 38614c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38624c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38634c16bd32STejun Heo * 38644c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38654c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3866d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38674c16bd32STejun Heo * 38684c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38694c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38704c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38714c16bd32STejun Heo * @attrs->cpumask. 38724c16bd32STejun Heo * 38734c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38744c16bd32STejun Heo * stable. 3875d185af30SYacine Belkadi * 3876d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3877d185af30SYacine Belkadi * %false if equal. 38784c16bd32STejun Heo */ 38794c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38804c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38814c16bd32STejun Heo { 3882d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38834c16bd32STejun Heo goto use_dfl; 38844c16bd32STejun Heo 38854c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38864c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38874c16bd32STejun Heo if (cpu_going_down >= 0) 38884c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38894c16bd32STejun Heo 38904c16bd32STejun Heo if (cpumask_empty(cpumask)) 38914c16bd32STejun Heo goto use_dfl; 38924c16bd32STejun Heo 38934c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38944c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38951ad0f0a7SMichael Bringmann 38961ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 38971ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 38981ad0f0a7SMichael Bringmann "possible intersect\n"); 38991ad0f0a7SMichael Bringmann return false; 39001ad0f0a7SMichael Bringmann } 39011ad0f0a7SMichael Bringmann 39024c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 39034c16bd32STejun Heo 39044c16bd32STejun Heo use_dfl: 39054c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 39064c16bd32STejun Heo return false; 39074c16bd32STejun Heo } 39084c16bd32STejun Heo 39091befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 39101befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 39111befcf30STejun Heo int node, 39121befcf30STejun Heo struct pool_workqueue *pwq) 39131befcf30STejun Heo { 39141befcf30STejun Heo struct pool_workqueue *old_pwq; 39151befcf30STejun Heo 39165b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39171befcf30STejun Heo lockdep_assert_held(&wq->mutex); 39181befcf30STejun Heo 39191befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 39201befcf30STejun Heo link_pwq(pwq); 39211befcf30STejun Heo 39221befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 39231befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 39241befcf30STejun Heo return old_pwq; 39251befcf30STejun Heo } 39261befcf30STejun Heo 39272d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 39282d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 39292d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 39302d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3931042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 39322d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 39332d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 39342d5f0764SLai Jiangshan }; 39352d5f0764SLai Jiangshan 39362d5f0764SLai Jiangshan /* free the resources after success or abort */ 39372d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 39382d5f0764SLai Jiangshan { 39392d5f0764SLai Jiangshan if (ctx) { 39402d5f0764SLai Jiangshan int node; 39412d5f0764SLai Jiangshan 39422d5f0764SLai Jiangshan for_each_node(node) 39432d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 39442d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 39452d5f0764SLai Jiangshan 39462d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 39472d5f0764SLai Jiangshan 39482d5f0764SLai Jiangshan kfree(ctx); 39492d5f0764SLai Jiangshan } 39502d5f0764SLai Jiangshan } 39512d5f0764SLai Jiangshan 39522d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 39532d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 39542d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 39552d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 39562d5f0764SLai Jiangshan { 39572d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 39582d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 39592d5f0764SLai Jiangshan int node; 39602d5f0764SLai Jiangshan 39612d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39622d5f0764SLai Jiangshan 3963acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 39642d5f0764SLai Jiangshan 3965be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 3966be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 39672d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 39682d5f0764SLai Jiangshan goto out_free; 39692d5f0764SLai Jiangshan 3970042f7df1SLai Jiangshan /* 3971042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3972042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3973042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3974042f7df1SLai Jiangshan */ 39752d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3976b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3977042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3978042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 39792d5f0764SLai Jiangshan 39802d5f0764SLai Jiangshan /* 39812d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 39822d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 39832d5f0764SLai Jiangshan * pools. 39842d5f0764SLai Jiangshan */ 39852d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 39862d5f0764SLai Jiangshan 39872d5f0764SLai Jiangshan /* 39882d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 39892d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 39902d5f0764SLai Jiangshan * it even if we don't use it immediately. 39912d5f0764SLai Jiangshan */ 39922d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39932d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 39942d5f0764SLai Jiangshan goto out_free; 39952d5f0764SLai Jiangshan 39962d5f0764SLai Jiangshan for_each_node(node) { 3997042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 39982d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39992d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 40002d5f0764SLai Jiangshan goto out_free; 40012d5f0764SLai Jiangshan } else { 40022d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 40032d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 40042d5f0764SLai Jiangshan } 40052d5f0764SLai Jiangshan } 40062d5f0764SLai Jiangshan 4007042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4008042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4009042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 40102d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4011042f7df1SLai Jiangshan 40122d5f0764SLai Jiangshan ctx->wq = wq; 40132d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40142d5f0764SLai Jiangshan return ctx; 40152d5f0764SLai Jiangshan 40162d5f0764SLai Jiangshan out_free: 40172d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40182d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 40192d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 40202d5f0764SLai Jiangshan return NULL; 40212d5f0764SLai Jiangshan } 40222d5f0764SLai Jiangshan 40232d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 40242d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 40252d5f0764SLai Jiangshan { 40262d5f0764SLai Jiangshan int node; 40272d5f0764SLai Jiangshan 40282d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 40292d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 40302d5f0764SLai Jiangshan 40312d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 40322d5f0764SLai Jiangshan 40332d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 40342d5f0764SLai Jiangshan for_each_node(node) 40352d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 40362d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 40372d5f0764SLai Jiangshan 40382d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 40392d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 40402d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 40412d5f0764SLai Jiangshan 40422d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 40432d5f0764SLai Jiangshan } 40442d5f0764SLai Jiangshan 4045a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4046a0111cf6SLai Jiangshan { 4047a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4048ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4049a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4050a0111cf6SLai Jiangshan } 4051a0111cf6SLai Jiangshan 4052a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4053a0111cf6SLai Jiangshan { 4054a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4055ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4056a0111cf6SLai Jiangshan } 4057a0111cf6SLai Jiangshan 4058a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4059a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4060a0111cf6SLai Jiangshan { 4061a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4062a0111cf6SLai Jiangshan 4063a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4064a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4065a0111cf6SLai Jiangshan return -EINVAL; 4066a0111cf6SLai Jiangshan 4067a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 40680a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 40690a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4070a0111cf6SLai Jiangshan return -EINVAL; 4071a0111cf6SLai Jiangshan 40720a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 40730a94efb5STejun Heo } 40740a94efb5STejun Heo 4075a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 40766201171eSwanghaibin if (!ctx) 40776201171eSwanghaibin return -ENOMEM; 4078a0111cf6SLai Jiangshan 4079a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4080a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4081a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4082a0111cf6SLai Jiangshan 40836201171eSwanghaibin return 0; 4084a0111cf6SLai Jiangshan } 4085a0111cf6SLai Jiangshan 40869e8cd2f5STejun Heo /** 40879e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 40889e8cd2f5STejun Heo * @wq: the target workqueue 40899e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 40909e8cd2f5STejun Heo * 40914c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 40924c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 40934c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 40944c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 40954c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 40964c16bd32STejun Heo * back-to-back will stay on its current pwq. 40979e8cd2f5STejun Heo * 4098d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4099d185af30SYacine Belkadi * 4100ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4101509b3204SDaniel Jordan * 4102d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 41039e8cd2f5STejun Heo */ 4104513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 41059e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 41069e8cd2f5STejun Heo { 4107a0111cf6SLai Jiangshan int ret; 41089e8cd2f5STejun Heo 4109509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4110509b3204SDaniel Jordan 4111509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4112a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4113509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 41142d5f0764SLai Jiangshan 41152d5f0764SLai Jiangshan return ret; 41169e8cd2f5STejun Heo } 41179e8cd2f5STejun Heo 41184c16bd32STejun Heo /** 41194c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 41204c16bd32STejun Heo * @wq: the target workqueue 41214c16bd32STejun Heo * @cpu: the CPU coming up or going down 41224c16bd32STejun Heo * @online: whether @cpu is coming up or going down 41234c16bd32STejun Heo * 41244c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 41254c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 41264c16bd32STejun Heo * @wq accordingly. 41274c16bd32STejun Heo * 41284c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 41294c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 41304c16bd32STejun Heo * correct. 41314c16bd32STejun Heo * 41324c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 41334c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 41344c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 41354c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 41364c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 41374c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 41384c16bd32STejun Heo * CPU_DOWN_PREPARE. 41394c16bd32STejun Heo */ 41404c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 41414c16bd32STejun Heo bool online) 41424c16bd32STejun Heo { 41434c16bd32STejun Heo int node = cpu_to_node(cpu); 41444c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 41454c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 41464c16bd32STejun Heo struct workqueue_attrs *target_attrs; 41474c16bd32STejun Heo cpumask_t *cpumask; 41484c16bd32STejun Heo 41494c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 41504c16bd32STejun Heo 4151f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4152f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 41534c16bd32STejun Heo return; 41544c16bd32STejun Heo 41554c16bd32STejun Heo /* 41564c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 41574c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 41584c16bd32STejun Heo * CPU hotplug exclusion. 41594c16bd32STejun Heo */ 41604c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 41614c16bd32STejun Heo cpumask = target_attrs->cpumask; 41624c16bd32STejun Heo 41634c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 41644c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 41654c16bd32STejun Heo 41664c16bd32STejun Heo /* 41674c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4168042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4169042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4170042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 41714c16bd32STejun Heo */ 4172042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 41734c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4174f7142ed4SLai Jiangshan return; 41754c16bd32STejun Heo } else { 41764c16bd32STejun Heo goto use_dfl_pwq; 41774c16bd32STejun Heo } 41784c16bd32STejun Heo 41794c16bd32STejun Heo /* create a new pwq */ 41804c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41814c16bd32STejun Heo if (!pwq) { 41822d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41834c16bd32STejun Heo wq->name); 418477f300b1SDaeseok Youn goto use_dfl_pwq; 41854c16bd32STejun Heo } 41864c16bd32STejun Heo 4187f7142ed4SLai Jiangshan /* Install the new pwq. */ 41884c16bd32STejun Heo mutex_lock(&wq->mutex); 41894c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41904c16bd32STejun Heo goto out_unlock; 41914c16bd32STejun Heo 41924c16bd32STejun Heo use_dfl_pwq: 4193f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4194a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 41954c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4196a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 41974c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 41984c16bd32STejun Heo out_unlock: 41994c16bd32STejun Heo mutex_unlock(&wq->mutex); 42004c16bd32STejun Heo put_pwq_unlocked(old_pwq); 42014c16bd32STejun Heo } 42024c16bd32STejun Heo 420330cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 42041da177e4SLinus Torvalds { 420549e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 42068a2b7538STejun Heo int cpu, ret; 4207e1d8aa9fSFrederic Weisbecker 420830cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4209420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4210420c0ddbSTejun Heo if (!wq->cpu_pwqs) 421130cdf249STejun Heo return -ENOMEM; 421230cdf249STejun Heo 421330cdf249STejun Heo for_each_possible_cpu(cpu) { 42147fb98ea7STejun Heo struct pool_workqueue *pwq = 42157fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 42167a62c2c8STejun Heo struct worker_pool *cpu_pools = 4217f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 421830cdf249STejun Heo 4219f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4220f147f29eSTejun Heo 4221f147f29eSTejun Heo mutex_lock(&wq->mutex); 42221befcf30STejun Heo link_pwq(pwq); 4223f147f29eSTejun Heo mutex_unlock(&wq->mutex); 422430cdf249STejun Heo } 422530cdf249STejun Heo return 0; 4226509b3204SDaniel Jordan } 4227509b3204SDaniel Jordan 4228ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4229509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 42308a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 42318a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 42328a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 42338a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 42348a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 42359e8cd2f5STejun Heo } else { 4236509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 42379e8cd2f5STejun Heo } 4238ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4239509b3204SDaniel Jordan 4240509b3204SDaniel Jordan return ret; 42410f900049STejun Heo } 42420f900049STejun Heo 4243f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4244f3421797STejun Heo const char *name) 4245b71ab8c2STejun Heo { 4246f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4247f3421797STejun Heo 4248f3421797STejun Heo if (max_active < 1 || max_active > lim) 4249044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4250f3421797STejun Heo max_active, name, 1, lim); 4251b71ab8c2STejun Heo 4252f3421797STejun Heo return clamp_val(max_active, 1, lim); 4253b71ab8c2STejun Heo } 4254b71ab8c2STejun Heo 4255983c7515STejun Heo /* 4256983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4257983c7515STejun Heo * to guarantee forward progress. 4258983c7515STejun Heo */ 4259983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4260983c7515STejun Heo { 4261983c7515STejun Heo struct worker *rescuer; 4262b92b36eaSDan Carpenter int ret; 4263983c7515STejun Heo 4264983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4265983c7515STejun Heo return 0; 4266983c7515STejun Heo 4267983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4268983c7515STejun Heo if (!rescuer) 4269983c7515STejun Heo return -ENOMEM; 4270983c7515STejun Heo 4271983c7515STejun Heo rescuer->rescue_wq = wq; 4272983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4273f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4274b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 4275983c7515STejun Heo kfree(rescuer); 4276b92b36eaSDan Carpenter return ret; 4277983c7515STejun Heo } 4278983c7515STejun Heo 4279983c7515STejun Heo wq->rescuer = rescuer; 4280983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4281983c7515STejun Heo wake_up_process(rescuer->task); 4282983c7515STejun Heo 4283983c7515STejun Heo return 0; 4284983c7515STejun Heo } 4285983c7515STejun Heo 4286a2775bbcSMathieu Malaterre __printf(1, 4) 4287669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 428897e37d7bSTejun Heo unsigned int flags, 4289669de8bdSBart Van Assche int max_active, ...) 42903af24433SOleg Nesterov { 4291df2d5ae4STejun Heo size_t tbl_size = 0; 4292ecf6881fSTejun Heo va_list args; 42933af24433SOleg Nesterov struct workqueue_struct *wq; 429449e3cf44STejun Heo struct pool_workqueue *pwq; 4295b196be89STejun Heo 42965c0338c6STejun Heo /* 42975c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 42985c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 42995c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 43005c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 43015c0338c6STejun Heo * on NUMA. 43025c0338c6STejun Heo */ 43035c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 43045c0338c6STejun Heo flags |= __WQ_ORDERED; 43055c0338c6STejun Heo 4306cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4307cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4308cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4309cee22a15SViresh Kumar 4310ecf6881fSTejun Heo /* allocate wq and format name */ 4311df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4312ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4313df2d5ae4STejun Heo 4314df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4315b196be89STejun Heo if (!wq) 4316d2c1d404STejun Heo return NULL; 4317b196be89STejun Heo 43186029a918STejun Heo if (flags & WQ_UNBOUND) { 4319be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 43206029a918STejun Heo if (!wq->unbound_attrs) 43216029a918STejun Heo goto err_free_wq; 43226029a918STejun Heo } 43236029a918STejun Heo 4324669de8bdSBart Van Assche va_start(args, max_active); 4325ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4326b196be89STejun Heo va_end(args); 43273af24433SOleg Nesterov 4328d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4329b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 43303af24433SOleg Nesterov 4331b196be89STejun Heo /* init wq */ 433297e37d7bSTejun Heo wq->flags = flags; 4333a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 43343c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4335112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 433630cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 433773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 433873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4339493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 43403af24433SOleg Nesterov 4341669de8bdSBart Van Assche wq_init_lockdep(wq); 4342cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 43433af24433SOleg Nesterov 434430cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 434582efcab3SBart Van Assche goto err_unreg_lockdep; 43461537663fSTejun Heo 434740c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4348d2c1d404STejun Heo goto err_destroy; 4349e22bee78STejun Heo 4350226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4351226223abSTejun Heo goto err_destroy; 4352226223abSTejun Heo 43536af8bf3dSOleg Nesterov /* 435468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 435568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 435668e13a67SLai Jiangshan * list. 43576af8bf3dSOleg Nesterov */ 435868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4359a0a1a5fdSTejun Heo 4360a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 436149e3cf44STejun Heo for_each_pwq(pwq, wq) 4362699ce097STejun Heo pwq_adjust_max_active(pwq); 4363a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4364a0a1a5fdSTejun Heo 4365e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4366a0a1a5fdSTejun Heo 436768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43683af24433SOleg Nesterov 43693af24433SOleg Nesterov return wq; 4370d2c1d404STejun Heo 437182efcab3SBart Van Assche err_unreg_lockdep: 4372009bb421SBart Van Assche wq_unregister_lockdep(wq); 4373009bb421SBart Van Assche wq_free_lockdep(wq); 437482efcab3SBart Van Assche err_free_wq: 43756029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 43764690c4abSTejun Heo kfree(wq); 4377d2c1d404STejun Heo return NULL; 4378d2c1d404STejun Heo err_destroy: 4379d2c1d404STejun Heo destroy_workqueue(wq); 43804690c4abSTejun Heo return NULL; 43811da177e4SLinus Torvalds } 4382669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 43831da177e4SLinus Torvalds 4384c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4385c29eb853STejun Heo { 4386c29eb853STejun Heo int i; 4387c29eb853STejun Heo 4388c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4389c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4390c29eb853STejun Heo return true; 4391c29eb853STejun Heo 4392c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4393c29eb853STejun Heo return true; 4394f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4395c29eb853STejun Heo return true; 4396c29eb853STejun Heo 4397c29eb853STejun Heo return false; 4398c29eb853STejun Heo } 4399c29eb853STejun Heo 44003af24433SOleg Nesterov /** 44013af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 44023af24433SOleg Nesterov * @wq: target workqueue 44033af24433SOleg Nesterov * 44043af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 44053af24433SOleg Nesterov */ 44063af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 44073af24433SOleg Nesterov { 440849e3cf44STejun Heo struct pool_workqueue *pwq; 44094c16bd32STejun Heo int node; 44103af24433SOleg Nesterov 4411def98c84STejun Heo /* 4412def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4413def98c84STejun Heo * lead to sysfs name conflicts. 4414def98c84STejun Heo */ 4415def98c84STejun Heo workqueue_sysfs_unregister(wq); 4416def98c84STejun Heo 44179c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 44189c5a2ba7STejun Heo drain_workqueue(wq); 4419c8efcc25STejun Heo 4420def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4421def98c84STejun Heo if (wq->rescuer) { 4422def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4423def98c84STejun Heo 4424def98c84STejun Heo /* this prevents new queueing */ 4425a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4426def98c84STejun Heo wq->rescuer = NULL; 4427a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4428def98c84STejun Heo 4429def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4430def98c84STejun Heo kthread_stop(rescuer->task); 44318efe1223STejun Heo kfree(rescuer); 4432def98c84STejun Heo } 4433def98c84STejun Heo 4434c29eb853STejun Heo /* 4435c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4436c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4437c29eb853STejun Heo */ 4438c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4439b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 444049e3cf44STejun Heo for_each_pwq(pwq, wq) { 4441a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4442c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 44431d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4444e66b39afSTejun Heo __func__, wq->name); 4445c29eb853STejun Heo show_pwq(pwq); 4446a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4447b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4448c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 444955df0933SImran Khan show_one_workqueue(wq); 44506183c009STejun Heo return; 445176af4d93STejun Heo } 4452a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 445376af4d93STejun Heo } 4454b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 44556183c009STejun Heo 4456a0a1a5fdSTejun Heo /* 4457a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4458a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4459a0a1a5fdSTejun Heo */ 4460e2dca7adSTejun Heo list_del_rcu(&wq->list); 446168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44623af24433SOleg Nesterov 44638864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4464669de8bdSBart Van Assche wq_unregister_lockdep(wq); 446529c91e99STejun Heo /* 44668864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4467e2dca7adSTejun Heo * schedule RCU free. 446829c91e99STejun Heo */ 446925b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 44708864b4e5STejun Heo } else { 44718864b4e5STejun Heo /* 44728864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 44734c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 44744c16bd32STejun Heo * @wq will be freed when the last pwq is released. 44758864b4e5STejun Heo */ 44764c16bd32STejun Heo for_each_node(node) { 44774c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 44784c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 44794c16bd32STejun Heo put_pwq_unlocked(pwq); 44804c16bd32STejun Heo } 44814c16bd32STejun Heo 44824c16bd32STejun Heo /* 44834c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 44844c16bd32STejun Heo * put. Don't access it afterwards. 44854c16bd32STejun Heo */ 44864c16bd32STejun Heo pwq = wq->dfl_pwq; 44874c16bd32STejun Heo wq->dfl_pwq = NULL; 4488dce90d47STejun Heo put_pwq_unlocked(pwq); 448929c91e99STejun Heo } 44903af24433SOleg Nesterov } 44913af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 44923af24433SOleg Nesterov 4493dcd989cbSTejun Heo /** 4494dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4495dcd989cbSTejun Heo * @wq: target workqueue 4496dcd989cbSTejun Heo * @max_active: new max_active value. 4497dcd989cbSTejun Heo * 4498dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4499dcd989cbSTejun Heo * 4500dcd989cbSTejun Heo * CONTEXT: 4501dcd989cbSTejun Heo * Don't call from IRQ context. 4502dcd989cbSTejun Heo */ 4503dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4504dcd989cbSTejun Heo { 450549e3cf44STejun Heo struct pool_workqueue *pwq; 4506dcd989cbSTejun Heo 45078719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 45080a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 45098719dceaSTejun Heo return; 45108719dceaSTejun Heo 4511f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4512dcd989cbSTejun Heo 4513a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4514dcd989cbSTejun Heo 45150a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4516dcd989cbSTejun Heo wq->saved_max_active = max_active; 4517dcd989cbSTejun Heo 4518699ce097STejun Heo for_each_pwq(pwq, wq) 4519699ce097STejun Heo pwq_adjust_max_active(pwq); 4520dcd989cbSTejun Heo 4521a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4522dcd989cbSTejun Heo } 4523dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4524dcd989cbSTejun Heo 4525dcd989cbSTejun Heo /** 452627d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 452727d4ee03SLukas Wunner * 452827d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 452927d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 453027d4ee03SLukas Wunner * 453127d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 453227d4ee03SLukas Wunner */ 453327d4ee03SLukas Wunner struct work_struct *current_work(void) 453427d4ee03SLukas Wunner { 453527d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 453627d4ee03SLukas Wunner 453727d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 453827d4ee03SLukas Wunner } 453927d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 454027d4ee03SLukas Wunner 454127d4ee03SLukas Wunner /** 4542e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4543e6267616STejun Heo * 4544e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4545e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4546d185af30SYacine Belkadi * 4547d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4548e6267616STejun Heo */ 4549e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4550e6267616STejun Heo { 4551e6267616STejun Heo struct worker *worker = current_wq_worker(); 4552e6267616STejun Heo 45536a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4554e6267616STejun Heo } 4555e6267616STejun Heo 4556e6267616STejun Heo /** 4557dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4558dcd989cbSTejun Heo * @cpu: CPU in question 4559dcd989cbSTejun Heo * @wq: target workqueue 4560dcd989cbSTejun Heo * 4561dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4562dcd989cbSTejun Heo * no synchronization around this function and the test result is 4563dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4564dcd989cbSTejun Heo * 4565d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4566d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4567d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4568d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4569d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4570d3251859STejun Heo * 4571d185af30SYacine Belkadi * Return: 4572dcd989cbSTejun Heo * %true if congested, %false otherwise. 4573dcd989cbSTejun Heo */ 4574d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4575dcd989cbSTejun Heo { 45767fb98ea7STejun Heo struct pool_workqueue *pwq; 457776af4d93STejun Heo bool ret; 457876af4d93STejun Heo 457924acfb71SThomas Gleixner rcu_read_lock(); 458024acfb71SThomas Gleixner preempt_disable(); 45817fb98ea7STejun Heo 4582d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4583d3251859STejun Heo cpu = smp_processor_id(); 4584d3251859STejun Heo 45857fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 45867fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 45877fb98ea7STejun Heo else 4588df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4589dcd989cbSTejun Heo 4590f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 459124acfb71SThomas Gleixner preempt_enable(); 459224acfb71SThomas Gleixner rcu_read_unlock(); 459376af4d93STejun Heo 459476af4d93STejun Heo return ret; 4595dcd989cbSTejun Heo } 4596dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4597dcd989cbSTejun Heo 4598dcd989cbSTejun Heo /** 4599dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4600dcd989cbSTejun Heo * @work: the work to be tested 4601dcd989cbSTejun Heo * 4602dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4603dcd989cbSTejun Heo * synchronization around this function and the test result is 4604dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4605dcd989cbSTejun Heo * 4606d185af30SYacine Belkadi * Return: 4607dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4608dcd989cbSTejun Heo */ 4609dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4610dcd989cbSTejun Heo { 4611fa1b54e6STejun Heo struct worker_pool *pool; 4612dcd989cbSTejun Heo unsigned long flags; 4613dcd989cbSTejun Heo unsigned int ret = 0; 4614dcd989cbSTejun Heo 4615dcd989cbSTejun Heo if (work_pending(work)) 4616dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4617038366c5SLai Jiangshan 461824acfb71SThomas Gleixner rcu_read_lock(); 4619fa1b54e6STejun Heo pool = get_work_pool(work); 4620038366c5SLai Jiangshan if (pool) { 4621a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4622c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4623dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4624a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4625038366c5SLai Jiangshan } 462624acfb71SThomas Gleixner rcu_read_unlock(); 4627dcd989cbSTejun Heo 4628dcd989cbSTejun Heo return ret; 4629dcd989cbSTejun Heo } 4630dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4631dcd989cbSTejun Heo 46323d1cb205STejun Heo /** 46333d1cb205STejun Heo * set_worker_desc - set description for the current work item 46343d1cb205STejun Heo * @fmt: printf-style format string 46353d1cb205STejun Heo * @...: arguments for the format string 46363d1cb205STejun Heo * 46373d1cb205STejun Heo * This function can be called by a running work function to describe what 46383d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 46393d1cb205STejun Heo * information will be printed out together to help debugging. The 46403d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 46413d1cb205STejun Heo */ 46423d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 46433d1cb205STejun Heo { 46443d1cb205STejun Heo struct worker *worker = current_wq_worker(); 46453d1cb205STejun Heo va_list args; 46463d1cb205STejun Heo 46473d1cb205STejun Heo if (worker) { 46483d1cb205STejun Heo va_start(args, fmt); 46493d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 46503d1cb205STejun Heo va_end(args); 46513d1cb205STejun Heo } 46523d1cb205STejun Heo } 46535c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 46543d1cb205STejun Heo 46553d1cb205STejun Heo /** 46563d1cb205STejun Heo * print_worker_info - print out worker information and description 46573d1cb205STejun Heo * @log_lvl: the log level to use when printing 46583d1cb205STejun Heo * @task: target task 46593d1cb205STejun Heo * 46603d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 46613d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 46623d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 46633d1cb205STejun Heo * 46643d1cb205STejun Heo * This function can be safely called on any task as long as the 46653d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 46663d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 46673d1cb205STejun Heo */ 46683d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 46693d1cb205STejun Heo { 46703d1cb205STejun Heo work_func_t *fn = NULL; 46713d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 46723d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 46733d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 46743d1cb205STejun Heo struct workqueue_struct *wq = NULL; 46753d1cb205STejun Heo struct worker *worker; 46763d1cb205STejun Heo 46773d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 46783d1cb205STejun Heo return; 46793d1cb205STejun Heo 46803d1cb205STejun Heo /* 46813d1cb205STejun Heo * This function is called without any synchronization and @task 46823d1cb205STejun Heo * could be in any state. Be careful with dereferences. 46833d1cb205STejun Heo */ 4684e700591aSPetr Mladek worker = kthread_probe_data(task); 46853d1cb205STejun Heo 46863d1cb205STejun Heo /* 46878bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 46888bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 46893d1cb205STejun Heo */ 4690fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4691fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4692fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4693fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4694fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 46953d1cb205STejun Heo 46963d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4697d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 46988bf89593STejun Heo if (strcmp(name, desc)) 46993d1cb205STejun Heo pr_cont(" (%s)", desc); 47003d1cb205STejun Heo pr_cont("\n"); 47013d1cb205STejun Heo } 47023d1cb205STejun Heo } 47033d1cb205STejun Heo 47043494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 47053494fc30STejun Heo { 47063494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 47073494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 47083494fc30STejun Heo pr_cont(" node=%d", pool->node); 47093494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 47103494fc30STejun Heo } 47113494fc30STejun Heo 47123494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 47133494fc30STejun Heo { 47143494fc30STejun Heo if (work->func == wq_barrier_func) { 47153494fc30STejun Heo struct wq_barrier *barr; 47163494fc30STejun Heo 47173494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 47183494fc30STejun Heo 47193494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 47203494fc30STejun Heo task_pid_nr(barr->task)); 47213494fc30STejun Heo } else { 4722d75f773cSSakari Ailus pr_cont("%s %ps", comma ? "," : "", work->func); 47233494fc30STejun Heo } 47243494fc30STejun Heo } 47253494fc30STejun Heo 47263494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 47273494fc30STejun Heo { 47283494fc30STejun Heo struct worker_pool *pool = pwq->pool; 47293494fc30STejun Heo struct work_struct *work; 47303494fc30STejun Heo struct worker *worker; 47313494fc30STejun Heo bool has_in_flight = false, has_pending = false; 47323494fc30STejun Heo int bkt; 47333494fc30STejun Heo 47343494fc30STejun Heo pr_info(" pwq %d:", pool->id); 47353494fc30STejun Heo pr_cont_pool_info(pool); 47363494fc30STejun Heo 4737e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4738e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 47393494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 47403494fc30STejun Heo 47413494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47423494fc30STejun Heo if (worker->current_pwq == pwq) { 47433494fc30STejun Heo has_in_flight = true; 47443494fc30STejun Heo break; 47453494fc30STejun Heo } 47463494fc30STejun Heo } 47473494fc30STejun Heo if (has_in_flight) { 47483494fc30STejun Heo bool comma = false; 47493494fc30STejun Heo 47503494fc30STejun Heo pr_info(" in-flight:"); 47513494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47523494fc30STejun Heo if (worker->current_pwq != pwq) 47533494fc30STejun Heo continue; 47543494fc30STejun Heo 4755d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 47563494fc30STejun Heo task_pid_nr(worker->task), 475730ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 47583494fc30STejun Heo worker->current_func); 47593494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 47603494fc30STejun Heo pr_cont_work(false, work); 47613494fc30STejun Heo comma = true; 47623494fc30STejun Heo } 47633494fc30STejun Heo pr_cont("\n"); 47643494fc30STejun Heo } 47653494fc30STejun Heo 47663494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47673494fc30STejun Heo if (get_work_pwq(work) == pwq) { 47683494fc30STejun Heo has_pending = true; 47693494fc30STejun Heo break; 47703494fc30STejun Heo } 47713494fc30STejun Heo } 47723494fc30STejun Heo if (has_pending) { 47733494fc30STejun Heo bool comma = false; 47743494fc30STejun Heo 47753494fc30STejun Heo pr_info(" pending:"); 47763494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47773494fc30STejun Heo if (get_work_pwq(work) != pwq) 47783494fc30STejun Heo continue; 47793494fc30STejun Heo 47803494fc30STejun Heo pr_cont_work(comma, work); 47813494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47823494fc30STejun Heo } 47833494fc30STejun Heo pr_cont("\n"); 47843494fc30STejun Heo } 47853494fc30STejun Heo 4786f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 47873494fc30STejun Heo bool comma = false; 47883494fc30STejun Heo 4789f97a4a1aSLai Jiangshan pr_info(" inactive:"); 4790f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 47913494fc30STejun Heo pr_cont_work(comma, work); 47923494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47933494fc30STejun Heo } 47943494fc30STejun Heo pr_cont("\n"); 47953494fc30STejun Heo } 47963494fc30STejun Heo } 47973494fc30STejun Heo 47983494fc30STejun Heo /** 479955df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 480055df0933SImran Khan * @wq: workqueue whose state will be printed 48013494fc30STejun Heo */ 480255df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 48033494fc30STejun Heo { 48043494fc30STejun Heo struct pool_workqueue *pwq; 48053494fc30STejun Heo bool idle = true; 480655df0933SImran Khan unsigned long flags; 48073494fc30STejun Heo 48083494fc30STejun Heo for_each_pwq(pwq, wq) { 4809f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 48103494fc30STejun Heo idle = false; 48113494fc30STejun Heo break; 48123494fc30STejun Heo } 48133494fc30STejun Heo } 481455df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 481555df0933SImran Khan return; 48163494fc30STejun Heo 48173494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 48183494fc30STejun Heo 48193494fc30STejun Heo for_each_pwq(pwq, wq) { 4820a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 482157116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 482257116ce1SJohan Hovold /* 482357116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 482457116ce1SJohan Hovold * drivers that queue work while holding locks 482557116ce1SJohan Hovold * also taken in their write paths. 482657116ce1SJohan Hovold */ 482757116ce1SJohan Hovold printk_deferred_enter(); 48283494fc30STejun Heo show_pwq(pwq); 482957116ce1SJohan Hovold printk_deferred_exit(); 483057116ce1SJohan Hovold } 4831a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 483262635ea8SSergey Senozhatsky /* 483362635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 483455df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 483562635ea8SSergey Senozhatsky * hard lockup. 483662635ea8SSergey Senozhatsky */ 483762635ea8SSergey Senozhatsky touch_nmi_watchdog(); 48383494fc30STejun Heo } 483955df0933SImran Khan 48403494fc30STejun Heo } 48413494fc30STejun Heo 484255df0933SImran Khan /** 484355df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 484455df0933SImran Khan * @pool: worker pool whose state will be printed 484555df0933SImran Khan */ 484655df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 484755df0933SImran Khan { 48483494fc30STejun Heo struct worker *worker; 48493494fc30STejun Heo bool first = true; 485055df0933SImran Khan unsigned long flags; 48513494fc30STejun Heo 4852a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 48533494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 48543494fc30STejun Heo goto next_pool; 485557116ce1SJohan Hovold /* 485657116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 485757116ce1SJohan Hovold * queue work while holding locks also taken in their write 485857116ce1SJohan Hovold * paths. 485957116ce1SJohan Hovold */ 486057116ce1SJohan Hovold printk_deferred_enter(); 48613494fc30STejun Heo pr_info("pool %d:", pool->id); 48623494fc30STejun Heo pr_cont_pool_info(pool); 486382607adcSTejun Heo pr_cont(" hung=%us workers=%d", 486482607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 486582607adcSTejun Heo pool->nr_workers); 48663494fc30STejun Heo if (pool->manager) 48673494fc30STejun Heo pr_cont(" manager: %d", 48683494fc30STejun Heo task_pid_nr(pool->manager->task)); 48693494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 48703494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 48713494fc30STejun Heo task_pid_nr(worker->task)); 48723494fc30STejun Heo first = false; 48733494fc30STejun Heo } 48743494fc30STejun Heo pr_cont("\n"); 487557116ce1SJohan Hovold printk_deferred_exit(); 48763494fc30STejun Heo next_pool: 4877a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 487862635ea8SSergey Senozhatsky /* 487962635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 488055df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 488162635ea8SSergey Senozhatsky * hard lockup. 488262635ea8SSergey Senozhatsky */ 488362635ea8SSergey Senozhatsky touch_nmi_watchdog(); 488455df0933SImran Khan 48853494fc30STejun Heo } 48863494fc30STejun Heo 488755df0933SImran Khan /** 488855df0933SImran Khan * show_all_workqueues - dump workqueue state 488955df0933SImran Khan * 489055df0933SImran Khan * Called from a sysrq handler or try_to_freeze_tasks() and prints out 489155df0933SImran Khan * all busy workqueues and pools. 489255df0933SImran Khan */ 489355df0933SImran Khan void show_all_workqueues(void) 489455df0933SImran Khan { 489555df0933SImran Khan struct workqueue_struct *wq; 489655df0933SImran Khan struct worker_pool *pool; 489755df0933SImran Khan int pi; 489855df0933SImran Khan 489955df0933SImran Khan rcu_read_lock(); 490055df0933SImran Khan 490155df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 490255df0933SImran Khan 490355df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 490455df0933SImran Khan show_one_workqueue(wq); 490555df0933SImran Khan 490655df0933SImran Khan for_each_pool(pool, pi) 490755df0933SImran Khan show_one_worker_pool(pool); 490855df0933SImran Khan 490924acfb71SThomas Gleixner rcu_read_unlock(); 49103494fc30STejun Heo } 49113494fc30STejun Heo 49126b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 49136b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 49146b59808bSTejun Heo { 49156b59808bSTejun Heo int off; 49166b59808bSTejun Heo 49176b59808bSTejun Heo /* always show the actual comm */ 49186b59808bSTejun Heo off = strscpy(buf, task->comm, size); 49196b59808bSTejun Heo if (off < 0) 49206b59808bSTejun Heo return; 49216b59808bSTejun Heo 4922197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 49236b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 49246b59808bSTejun Heo 4925197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4926197f6accSTejun Heo struct worker *worker = kthread_data(task); 4927197f6accSTejun Heo struct worker_pool *pool = worker->pool; 49286b59808bSTejun Heo 49296b59808bSTejun Heo if (pool) { 4930a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 49316b59808bSTejun Heo /* 4932197f6accSTejun Heo * ->desc tracks information (wq name or 4933197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4934197f6accSTejun Heo * current, prepend '+', otherwise '-'. 49356b59808bSTejun Heo */ 49366b59808bSTejun Heo if (worker->desc[0] != '\0') { 49376b59808bSTejun Heo if (worker->current_work) 49386b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 49396b59808bSTejun Heo worker->desc); 49406b59808bSTejun Heo else 49416b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 49426b59808bSTejun Heo worker->desc); 49436b59808bSTejun Heo } 4944a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 49456b59808bSTejun Heo } 4946197f6accSTejun Heo } 49476b59808bSTejun Heo 49486b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 49496b59808bSTejun Heo } 49506b59808bSTejun Heo 495166448bc2SMathieu Malaterre #ifdef CONFIG_SMP 495266448bc2SMathieu Malaterre 4953db7bccf4STejun Heo /* 4954db7bccf4STejun Heo * CPU hotplug. 4955db7bccf4STejun Heo * 4956e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4957112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4958706026c2STejun Heo * pool which make migrating pending and scheduled works very 4959e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 496094cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4961e22bee78STejun Heo * blocked draining impractical. 4962e22bee78STejun Heo * 496324647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4964628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4965628c78e7STejun Heo * cpu comes back online. 4966db7bccf4STejun Heo */ 4967db7bccf4STejun Heo 4968e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4969db7bccf4STejun Heo { 49704ce62e9eSTejun Heo struct worker_pool *pool; 4971db7bccf4STejun Heo struct worker *worker; 4972db7bccf4STejun Heo 4973f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49741258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 4975a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4976e22bee78STejun Heo 4977f2d5a0eeSTejun Heo /* 497892f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 497994cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 498011b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 4981b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 4982b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 4983b4ac9384SLai Jiangshan * is on the same cpu. 4984f2d5a0eeSTejun Heo */ 4985da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4986403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4987db7bccf4STejun Heo 498824647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4989f2d5a0eeSTejun Heo 4990e22bee78STejun Heo /* 4991989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 4992989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 4993989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 4994989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 4995989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 4996eb283428SLai Jiangshan * are served by workers tied to the pool. 4997e22bee78STejun Heo */ 4998bc35f7efSLai Jiangshan pool->nr_running = 0; 4999eb283428SLai Jiangshan 5000eb283428SLai Jiangshan /* 5001eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5002eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5003eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5004eb283428SLai Jiangshan */ 5005eb283428SLai Jiangshan wake_up_worker(pool); 5006989442d7SLai Jiangshan 5007a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5008989442d7SLai Jiangshan 5009989442d7SLai Jiangshan for_each_pool_worker(worker, pool) { 5010989442d7SLai Jiangshan kthread_set_per_cpu(worker->task, -1); 501146a4d679SLai Jiangshan if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 501210a5a651SZqiang WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 501346a4d679SLai Jiangshan else 501446a4d679SLai Jiangshan WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 5015989442d7SLai Jiangshan } 5016989442d7SLai Jiangshan 5017989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5018eb283428SLai Jiangshan } 5019db7bccf4STejun Heo } 5020db7bccf4STejun Heo 5021bd7c089eSTejun Heo /** 5022bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5023bd7c089eSTejun Heo * @pool: pool of interest 5024bd7c089eSTejun Heo * 5025a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5026bd7c089eSTejun Heo */ 5027bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5028bd7c089eSTejun Heo { 5029a9ab775bSTejun Heo struct worker *worker; 5030bd7c089eSTejun Heo 50311258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5032bd7c089eSTejun Heo 5033bd7c089eSTejun Heo /* 5034a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5035a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5036402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5037a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5038a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5039bd7c089eSTejun Heo */ 50405c25b5ffSPeter Zijlstra for_each_pool_worker(worker, pool) { 50415c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 5042a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5043a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 50445c25b5ffSPeter Zijlstra } 5045a9ab775bSTejun Heo 5046a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5047f7c17d26SWanpeng Li 50483de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5049a9ab775bSTejun Heo 5050da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5051a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5052a9ab775bSTejun Heo 5053a9ab775bSTejun Heo /* 5054a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5055a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5056a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5057a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5058a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5059a9ab775bSTejun Heo * concurrency management. Note that when or whether 5060a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5061a9ab775bSTejun Heo * 5062c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5063a9ab775bSTejun Heo * tested without holding any lock in 50646d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5065a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5066a9ab775bSTejun Heo * management operations. 5067bd7c089eSTejun Heo */ 5068a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5069a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5070a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5071c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5072bd7c089eSTejun Heo } 5073a9ab775bSTejun Heo 5074a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5075bd7c089eSTejun Heo } 5076bd7c089eSTejun Heo 50777dbc725eSTejun Heo /** 50787dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 50797dbc725eSTejun Heo * @pool: unbound pool of interest 50807dbc725eSTejun Heo * @cpu: the CPU which is coming up 50817dbc725eSTejun Heo * 50827dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 50837dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 50847dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 50857dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 50867dbc725eSTejun Heo */ 50877dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 50887dbc725eSTejun Heo { 50897dbc725eSTejun Heo static cpumask_t cpumask; 50907dbc725eSTejun Heo struct worker *worker; 50917dbc725eSTejun Heo 50921258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 50937dbc725eSTejun Heo 50947dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 50957dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 50967dbc725eSTejun Heo return; 50977dbc725eSTejun Heo 50987dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 50997dbc725eSTejun Heo 51007dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5101da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5102d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 51037dbc725eSTejun Heo } 51047dbc725eSTejun Heo 51057ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 51061da177e4SLinus Torvalds { 51074ce62e9eSTejun Heo struct worker_pool *pool; 51081da177e4SLinus Torvalds 5109f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 51103ce63377STejun Heo if (pool->nr_workers) 51113ce63377STejun Heo continue; 5112051e1850SLai Jiangshan if (!create_worker(pool)) 51137ee681b2SThomas Gleixner return -ENOMEM; 51143af24433SOleg Nesterov } 51157ee681b2SThomas Gleixner return 0; 51167ee681b2SThomas Gleixner } 51171da177e4SLinus Torvalds 51187ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 51197ee681b2SThomas Gleixner { 51207ee681b2SThomas Gleixner struct worker_pool *pool; 51217ee681b2SThomas Gleixner struct workqueue_struct *wq; 51227ee681b2SThomas Gleixner int pi; 51237ee681b2SThomas Gleixner 512468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 51257dbc725eSTejun Heo 51267dbc725eSTejun Heo for_each_pool(pool, pi) { 51271258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 512894cf58bbSTejun Heo 5129f05b558dSLai Jiangshan if (pool->cpu == cpu) 513094cf58bbSTejun Heo rebind_workers(pool); 5131f05b558dSLai Jiangshan else if (pool->cpu < 0) 51327dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 513394cf58bbSTejun Heo 51341258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 513594cf58bbSTejun Heo } 51367dbc725eSTejun Heo 51374c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51384c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51394c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 51404c16bd32STejun Heo 514168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 51427ee681b2SThomas Gleixner return 0; 514365758202STejun Heo } 514465758202STejun Heo 51457ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 514665758202STejun Heo { 51474c16bd32STejun Heo struct workqueue_struct *wq; 51488db25e78STejun Heo 51494c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5150e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5151e8b3f8dbSLai Jiangshan return -1; 5152e8b3f8dbSLai Jiangshan 5153e8b3f8dbSLai Jiangshan unbind_workers(cpu); 51544c16bd32STejun Heo 51554c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51564c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 51574c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51584c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 51594c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 51604c16bd32STejun Heo 51617ee681b2SThomas Gleixner return 0; 516265758202STejun Heo } 516365758202STejun Heo 51642d3854a3SRusty Russell struct work_for_cpu { 5165ed48ece2STejun Heo struct work_struct work; 51662d3854a3SRusty Russell long (*fn)(void *); 51672d3854a3SRusty Russell void *arg; 51682d3854a3SRusty Russell long ret; 51692d3854a3SRusty Russell }; 51702d3854a3SRusty Russell 5171ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 51722d3854a3SRusty Russell { 5173ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5174ed48ece2STejun Heo 51752d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 51762d3854a3SRusty Russell } 51772d3854a3SRusty Russell 51782d3854a3SRusty Russell /** 517922aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 51802d3854a3SRusty Russell * @cpu: the cpu to run on 51812d3854a3SRusty Russell * @fn: the function to run 51822d3854a3SRusty Russell * @arg: the function arg 51832d3854a3SRusty Russell * 518431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 51856b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5186d185af30SYacine Belkadi * 5187d185af30SYacine Belkadi * Return: The value @fn returns. 51882d3854a3SRusty Russell */ 5189d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 51902d3854a3SRusty Russell { 5191ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 51922d3854a3SRusty Russell 5193ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5194ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 519512997d1aSBjorn Helgaas flush_work(&wfc.work); 5196440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 51972d3854a3SRusty Russell return wfc.ret; 51982d3854a3SRusty Russell } 51992d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 52000e8d6a93SThomas Gleixner 52010e8d6a93SThomas Gleixner /** 52020e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 52030e8d6a93SThomas Gleixner * @cpu: the cpu to run on 52040e8d6a93SThomas Gleixner * @fn: the function to run 52050e8d6a93SThomas Gleixner * @arg: the function argument 52060e8d6a93SThomas Gleixner * 52070e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 52080e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 52090e8d6a93SThomas Gleixner * 52100e8d6a93SThomas Gleixner * Return: The value @fn returns. 52110e8d6a93SThomas Gleixner */ 52120e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 52130e8d6a93SThomas Gleixner { 52140e8d6a93SThomas Gleixner long ret = -ENODEV; 52150e8d6a93SThomas Gleixner 5216ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 52170e8d6a93SThomas Gleixner if (cpu_online(cpu)) 52180e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5219ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 52200e8d6a93SThomas Gleixner return ret; 52210e8d6a93SThomas Gleixner } 52220e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 52232d3854a3SRusty Russell #endif /* CONFIG_SMP */ 52242d3854a3SRusty Russell 5225a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5226e7577c50SRusty Russell 5227a0a1a5fdSTejun Heo /** 5228a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5229a0a1a5fdSTejun Heo * 523058a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5231f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5232706026c2STejun Heo * pool->worklist. 5233a0a1a5fdSTejun Heo * 5234a0a1a5fdSTejun Heo * CONTEXT: 5235a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5236a0a1a5fdSTejun Heo */ 5237a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5238a0a1a5fdSTejun Heo { 523924b8a847STejun Heo struct workqueue_struct *wq; 524024b8a847STejun Heo struct pool_workqueue *pwq; 5241a0a1a5fdSTejun Heo 524268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5243a0a1a5fdSTejun Heo 52446183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5245a0a1a5fdSTejun Heo workqueue_freezing = true; 5246a0a1a5fdSTejun Heo 524724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5248a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5249699ce097STejun Heo for_each_pwq(pwq, wq) 5250699ce097STejun Heo pwq_adjust_max_active(pwq); 5251a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5252a1056305STejun Heo } 52535bcab335STejun Heo 525468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5255a0a1a5fdSTejun Heo } 5256a0a1a5fdSTejun Heo 5257a0a1a5fdSTejun Heo /** 525858a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5259a0a1a5fdSTejun Heo * 5260a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5261a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5262a0a1a5fdSTejun Heo * 5263a0a1a5fdSTejun Heo * CONTEXT: 526468e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5265a0a1a5fdSTejun Heo * 5266d185af30SYacine Belkadi * Return: 526758a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 526858a69cb4STejun Heo * is complete. 5269a0a1a5fdSTejun Heo */ 5270a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5271a0a1a5fdSTejun Heo { 5272a0a1a5fdSTejun Heo bool busy = false; 527324b8a847STejun Heo struct workqueue_struct *wq; 527424b8a847STejun Heo struct pool_workqueue *pwq; 5275a0a1a5fdSTejun Heo 527668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5277a0a1a5fdSTejun Heo 52786183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5279a0a1a5fdSTejun Heo 528024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 528124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 528224b8a847STejun Heo continue; 5283a0a1a5fdSTejun Heo /* 5284a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5285a0a1a5fdSTejun Heo * to peek without lock. 5286a0a1a5fdSTejun Heo */ 528724acfb71SThomas Gleixner rcu_read_lock(); 528824b8a847STejun Heo for_each_pwq(pwq, wq) { 52896183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5290112202d9STejun Heo if (pwq->nr_active) { 5291a0a1a5fdSTejun Heo busy = true; 529224acfb71SThomas Gleixner rcu_read_unlock(); 5293a0a1a5fdSTejun Heo goto out_unlock; 5294a0a1a5fdSTejun Heo } 5295a0a1a5fdSTejun Heo } 529624acfb71SThomas Gleixner rcu_read_unlock(); 5297a0a1a5fdSTejun Heo } 5298a0a1a5fdSTejun Heo out_unlock: 529968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5300a0a1a5fdSTejun Heo return busy; 5301a0a1a5fdSTejun Heo } 5302a0a1a5fdSTejun Heo 5303a0a1a5fdSTejun Heo /** 5304a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5305a0a1a5fdSTejun Heo * 5306a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5307706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5308a0a1a5fdSTejun Heo * 5309a0a1a5fdSTejun Heo * CONTEXT: 5310a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5311a0a1a5fdSTejun Heo */ 5312a0a1a5fdSTejun Heo void thaw_workqueues(void) 5313a0a1a5fdSTejun Heo { 531424b8a847STejun Heo struct workqueue_struct *wq; 531524b8a847STejun Heo struct pool_workqueue *pwq; 5316a0a1a5fdSTejun Heo 531768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5318a0a1a5fdSTejun Heo 5319a0a1a5fdSTejun Heo if (!workqueue_freezing) 5320a0a1a5fdSTejun Heo goto out_unlock; 5321a0a1a5fdSTejun Heo 532274b414eaSLai Jiangshan workqueue_freezing = false; 532324b8a847STejun Heo 532424b8a847STejun Heo /* restore max_active and repopulate worklist */ 532524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5326a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5327699ce097STejun Heo for_each_pwq(pwq, wq) 5328699ce097STejun Heo pwq_adjust_max_active(pwq); 5329a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 533024b8a847STejun Heo } 533124b8a847STejun Heo 5332a0a1a5fdSTejun Heo out_unlock: 533368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5334a0a1a5fdSTejun Heo } 5335a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5336a0a1a5fdSTejun Heo 5337042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5338042f7df1SLai Jiangshan { 5339042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5340042f7df1SLai Jiangshan int ret = 0; 5341042f7df1SLai Jiangshan struct workqueue_struct *wq; 5342042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5343042f7df1SLai Jiangshan 5344042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5345042f7df1SLai Jiangshan 5346042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5347042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5348042f7df1SLai Jiangshan continue; 5349042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5350042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5351042f7df1SLai Jiangshan continue; 5352042f7df1SLai Jiangshan 5353042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5354042f7df1SLai Jiangshan if (!ctx) { 5355042f7df1SLai Jiangshan ret = -ENOMEM; 5356042f7df1SLai Jiangshan break; 5357042f7df1SLai Jiangshan } 5358042f7df1SLai Jiangshan 5359042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5360042f7df1SLai Jiangshan } 5361042f7df1SLai Jiangshan 5362042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5363042f7df1SLai Jiangshan if (!ret) 5364042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5365042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5366042f7df1SLai Jiangshan } 5367042f7df1SLai Jiangshan 5368042f7df1SLai Jiangshan return ret; 5369042f7df1SLai Jiangshan } 5370042f7df1SLai Jiangshan 5371042f7df1SLai Jiangshan /** 5372042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5373042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5374042f7df1SLai Jiangshan * 5375042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5376042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5377042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5378042f7df1SLai Jiangshan * 537967dc8325SCai Huoqing * Return: 0 - Success 5380042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5381042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5382042f7df1SLai Jiangshan */ 5383042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5384042f7df1SLai Jiangshan { 5385042f7df1SLai Jiangshan int ret = -EINVAL; 5386042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5387042f7df1SLai Jiangshan 5388c98a9805STal Shorer /* 5389c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5390c98a9805STal Shorer * If the user wishes to include them, we allow that. 5391c98a9805STal Shorer */ 5392042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5393042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5394a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5395d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5396d25302e4SMenglong Dong ret = 0; 5397d25302e4SMenglong Dong goto out_unlock; 5398d25302e4SMenglong Dong } 5399d25302e4SMenglong Dong 5400d25302e4SMenglong Dong if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) { 5401d25302e4SMenglong Dong ret = -ENOMEM; 5402d25302e4SMenglong Dong goto out_unlock; 5403d25302e4SMenglong Dong } 5404042f7df1SLai Jiangshan 5405042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5406042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5407042f7df1SLai Jiangshan 5408042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5409042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5410042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5411042f7df1SLai Jiangshan 5412042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5413042f7df1SLai Jiangshan if (ret < 0) 5414042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5415042f7df1SLai Jiangshan 5416d25302e4SMenglong Dong free_cpumask_var(saved_cpumask); 5417d25302e4SMenglong Dong out_unlock: 5418a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5419042f7df1SLai Jiangshan } 5420042f7df1SLai Jiangshan 5421042f7df1SLai Jiangshan return ret; 5422042f7df1SLai Jiangshan } 5423042f7df1SLai Jiangshan 54246ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 54256ba94429SFrederic Weisbecker /* 54266ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 54276ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 54286ba94429SFrederic Weisbecker * following attributes. 54296ba94429SFrederic Weisbecker * 54306ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 54316ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 54326ba94429SFrederic Weisbecker * 54336ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 54346ba94429SFrederic Weisbecker * 54359a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 54366ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 54376ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 54389a19b463SWang Long * numa RW bool : whether enable NUMA affinity 54396ba94429SFrederic Weisbecker */ 54406ba94429SFrederic Weisbecker struct wq_device { 54416ba94429SFrederic Weisbecker struct workqueue_struct *wq; 54426ba94429SFrederic Weisbecker struct device dev; 54436ba94429SFrederic Weisbecker }; 54446ba94429SFrederic Weisbecker 54456ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 54466ba94429SFrederic Weisbecker { 54476ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 54486ba94429SFrederic Weisbecker 54496ba94429SFrederic Weisbecker return wq_dev->wq; 54506ba94429SFrederic Weisbecker } 54516ba94429SFrederic Weisbecker 54526ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 54536ba94429SFrederic Weisbecker char *buf) 54546ba94429SFrederic Weisbecker { 54556ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54566ba94429SFrederic Weisbecker 54576ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 54586ba94429SFrederic Weisbecker } 54596ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 54606ba94429SFrederic Weisbecker 54616ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 54626ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54636ba94429SFrederic Weisbecker { 54646ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54656ba94429SFrederic Weisbecker 54666ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 54676ba94429SFrederic Weisbecker } 54686ba94429SFrederic Weisbecker 54696ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 54706ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 54716ba94429SFrederic Weisbecker size_t count) 54726ba94429SFrederic Weisbecker { 54736ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54746ba94429SFrederic Weisbecker int val; 54756ba94429SFrederic Weisbecker 54766ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 54776ba94429SFrederic Weisbecker return -EINVAL; 54786ba94429SFrederic Weisbecker 54796ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 54806ba94429SFrederic Weisbecker return count; 54816ba94429SFrederic Weisbecker } 54826ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 54836ba94429SFrederic Weisbecker 54846ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 54856ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 54866ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 54876ba94429SFrederic Weisbecker NULL, 54886ba94429SFrederic Weisbecker }; 54896ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 54906ba94429SFrederic Weisbecker 54916ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 54926ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54936ba94429SFrederic Weisbecker { 54946ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54956ba94429SFrederic Weisbecker const char *delim = ""; 54966ba94429SFrederic Weisbecker int node, written = 0; 54976ba94429SFrederic Weisbecker 5498ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 549924acfb71SThomas Gleixner rcu_read_lock(); 55006ba94429SFrederic Weisbecker for_each_node(node) { 55016ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 55026ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 55036ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 55046ba94429SFrederic Weisbecker delim = " "; 55056ba94429SFrederic Weisbecker } 55066ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 550724acfb71SThomas Gleixner rcu_read_unlock(); 5508ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 55096ba94429SFrederic Weisbecker 55106ba94429SFrederic Weisbecker return written; 55116ba94429SFrederic Weisbecker } 55126ba94429SFrederic Weisbecker 55136ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 55146ba94429SFrederic Weisbecker char *buf) 55156ba94429SFrederic Weisbecker { 55166ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55176ba94429SFrederic Weisbecker int written; 55186ba94429SFrederic Weisbecker 55196ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55206ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 55216ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55226ba94429SFrederic Weisbecker 55236ba94429SFrederic Weisbecker return written; 55246ba94429SFrederic Weisbecker } 55256ba94429SFrederic Weisbecker 55266ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 55276ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 55286ba94429SFrederic Weisbecker { 55296ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 55306ba94429SFrederic Weisbecker 5531899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5532899a94feSLai Jiangshan 5533be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 55346ba94429SFrederic Weisbecker if (!attrs) 55356ba94429SFrederic Weisbecker return NULL; 55366ba94429SFrederic Weisbecker 55376ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 55386ba94429SFrederic Weisbecker return attrs; 55396ba94429SFrederic Weisbecker } 55406ba94429SFrederic Weisbecker 55416ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 55426ba94429SFrederic Weisbecker const char *buf, size_t count) 55436ba94429SFrederic Weisbecker { 55446ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55456ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5546d4d3e257SLai Jiangshan int ret = -ENOMEM; 5547d4d3e257SLai Jiangshan 5548d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55496ba94429SFrederic Weisbecker 55506ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55516ba94429SFrederic Weisbecker if (!attrs) 5552d4d3e257SLai Jiangshan goto out_unlock; 55536ba94429SFrederic Weisbecker 55546ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 55556ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5556d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55576ba94429SFrederic Weisbecker else 55586ba94429SFrederic Weisbecker ret = -EINVAL; 55596ba94429SFrederic Weisbecker 5560d4d3e257SLai Jiangshan out_unlock: 5561d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55626ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55636ba94429SFrederic Weisbecker return ret ?: count; 55646ba94429SFrederic Weisbecker } 55656ba94429SFrederic Weisbecker 55666ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 55676ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 55686ba94429SFrederic Weisbecker { 55696ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55706ba94429SFrederic Weisbecker int written; 55716ba94429SFrederic Weisbecker 55726ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55736ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 55746ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 55756ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55766ba94429SFrederic Weisbecker return written; 55776ba94429SFrederic Weisbecker } 55786ba94429SFrederic Weisbecker 55796ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 55806ba94429SFrederic Weisbecker struct device_attribute *attr, 55816ba94429SFrederic Weisbecker const char *buf, size_t count) 55826ba94429SFrederic Weisbecker { 55836ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55846ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5585d4d3e257SLai Jiangshan int ret = -ENOMEM; 5586d4d3e257SLai Jiangshan 5587d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55886ba94429SFrederic Weisbecker 55896ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55906ba94429SFrederic Weisbecker if (!attrs) 5591d4d3e257SLai Jiangshan goto out_unlock; 55926ba94429SFrederic Weisbecker 55936ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 55946ba94429SFrederic Weisbecker if (!ret) 5595d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55966ba94429SFrederic Weisbecker 5597d4d3e257SLai Jiangshan out_unlock: 5598d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55996ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56006ba94429SFrederic Weisbecker return ret ?: count; 56016ba94429SFrederic Weisbecker } 56026ba94429SFrederic Weisbecker 56036ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 56046ba94429SFrederic Weisbecker char *buf) 56056ba94429SFrederic Weisbecker { 56066ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56076ba94429SFrederic Weisbecker int written; 56086ba94429SFrederic Weisbecker 56096ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 56106ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 56116ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 56126ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 56136ba94429SFrederic Weisbecker 56146ba94429SFrederic Weisbecker return written; 56156ba94429SFrederic Weisbecker } 56166ba94429SFrederic Weisbecker 56176ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 56186ba94429SFrederic Weisbecker const char *buf, size_t count) 56196ba94429SFrederic Weisbecker { 56206ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56216ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5622d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5623d4d3e257SLai Jiangshan 5624d4d3e257SLai Jiangshan apply_wqattrs_lock(); 56256ba94429SFrederic Weisbecker 56266ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 56276ba94429SFrederic Weisbecker if (!attrs) 5628d4d3e257SLai Jiangshan goto out_unlock; 56296ba94429SFrederic Weisbecker 56306ba94429SFrederic Weisbecker ret = -EINVAL; 56316ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 56326ba94429SFrederic Weisbecker attrs->no_numa = !v; 5633d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56346ba94429SFrederic Weisbecker } 56356ba94429SFrederic Weisbecker 5636d4d3e257SLai Jiangshan out_unlock: 5637d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 56386ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56396ba94429SFrederic Weisbecker return ret ?: count; 56406ba94429SFrederic Weisbecker } 56416ba94429SFrederic Weisbecker 56426ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 56436ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 56446ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 56456ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 56466ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 56476ba94429SFrederic Weisbecker __ATTR_NULL, 56486ba94429SFrederic Weisbecker }; 56496ba94429SFrederic Weisbecker 56506ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 56516ba94429SFrederic Weisbecker .name = "workqueue", 56526ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 56536ba94429SFrederic Weisbecker }; 56546ba94429SFrederic Weisbecker 5655b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5656b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5657b05a7928SFrederic Weisbecker { 5658b05a7928SFrederic Weisbecker int written; 5659b05a7928SFrederic Weisbecker 5660042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5661b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5662b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5663042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5664b05a7928SFrederic Weisbecker 5665b05a7928SFrederic Weisbecker return written; 5666b05a7928SFrederic Weisbecker } 5667b05a7928SFrederic Weisbecker 5668042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5669042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5670042f7df1SLai Jiangshan { 5671042f7df1SLai Jiangshan cpumask_var_t cpumask; 5672042f7df1SLai Jiangshan int ret; 5673042f7df1SLai Jiangshan 5674042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5675042f7df1SLai Jiangshan return -ENOMEM; 5676042f7df1SLai Jiangshan 5677042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5678042f7df1SLai Jiangshan if (!ret) 5679042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5680042f7df1SLai Jiangshan 5681042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5682042f7df1SLai Jiangshan return ret ? ret : count; 5683042f7df1SLai Jiangshan } 5684042f7df1SLai Jiangshan 5685b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5686042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5687042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5688b05a7928SFrederic Weisbecker 56896ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 56906ba94429SFrederic Weisbecker { 5691b05a7928SFrederic Weisbecker int err; 5692b05a7928SFrederic Weisbecker 5693b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5694b05a7928SFrederic Weisbecker if (err) 5695b05a7928SFrederic Weisbecker return err; 5696b05a7928SFrederic Weisbecker 5697b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 56986ba94429SFrederic Weisbecker } 56996ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 57006ba94429SFrederic Weisbecker 57016ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 57026ba94429SFrederic Weisbecker { 57036ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 57046ba94429SFrederic Weisbecker 57056ba94429SFrederic Weisbecker kfree(wq_dev); 57066ba94429SFrederic Weisbecker } 57076ba94429SFrederic Weisbecker 57086ba94429SFrederic Weisbecker /** 57096ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 57106ba94429SFrederic Weisbecker * @wq: the workqueue to register 57116ba94429SFrederic Weisbecker * 57126ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 57136ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 57146ba94429SFrederic Weisbecker * which is the preferred method. 57156ba94429SFrederic Weisbecker * 57166ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 57176ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 57186ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 57196ba94429SFrederic Weisbecker * attributes. 57206ba94429SFrederic Weisbecker * 57216ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 57226ba94429SFrederic Weisbecker */ 57236ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 57246ba94429SFrederic Weisbecker { 57256ba94429SFrederic Weisbecker struct wq_device *wq_dev; 57266ba94429SFrederic Weisbecker int ret; 57276ba94429SFrederic Weisbecker 57286ba94429SFrederic Weisbecker /* 5729402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 57306ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 57316ba94429SFrederic Weisbecker * workqueues. 57326ba94429SFrederic Weisbecker */ 57330a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 57346ba94429SFrederic Weisbecker return -EINVAL; 57356ba94429SFrederic Weisbecker 57366ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 57376ba94429SFrederic Weisbecker if (!wq_dev) 57386ba94429SFrederic Weisbecker return -ENOMEM; 57396ba94429SFrederic Weisbecker 57406ba94429SFrederic Weisbecker wq_dev->wq = wq; 57416ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 57426ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 574323217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 57446ba94429SFrederic Weisbecker 57456ba94429SFrederic Weisbecker /* 57466ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 57476ba94429SFrederic Weisbecker * everything is ready. 57486ba94429SFrederic Weisbecker */ 57496ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 57506ba94429SFrederic Weisbecker 57516ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 57526ba94429SFrederic Weisbecker if (ret) { 5753537f4146SArvind Yadav put_device(&wq_dev->dev); 57546ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57556ba94429SFrederic Weisbecker return ret; 57566ba94429SFrederic Weisbecker } 57576ba94429SFrederic Weisbecker 57586ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 57596ba94429SFrederic Weisbecker struct device_attribute *attr; 57606ba94429SFrederic Weisbecker 57616ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 57626ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 57636ba94429SFrederic Weisbecker if (ret) { 57646ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57656ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57666ba94429SFrederic Weisbecker return ret; 57676ba94429SFrederic Weisbecker } 57686ba94429SFrederic Weisbecker } 57696ba94429SFrederic Weisbecker } 57706ba94429SFrederic Weisbecker 57716ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 57726ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 57736ba94429SFrederic Weisbecker return 0; 57746ba94429SFrederic Weisbecker } 57756ba94429SFrederic Weisbecker 57766ba94429SFrederic Weisbecker /** 57776ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 57786ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 57796ba94429SFrederic Weisbecker * 57806ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 57816ba94429SFrederic Weisbecker */ 57826ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 57836ba94429SFrederic Weisbecker { 57846ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 57856ba94429SFrederic Weisbecker 57866ba94429SFrederic Weisbecker if (!wq->wq_dev) 57876ba94429SFrederic Weisbecker return; 57886ba94429SFrederic Weisbecker 57896ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57906ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57916ba94429SFrederic Weisbecker } 57926ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 57936ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 57946ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 57956ba94429SFrederic Weisbecker 579682607adcSTejun Heo /* 579782607adcSTejun Heo * Workqueue watchdog. 579882607adcSTejun Heo * 579982607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 580082607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 580182607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 580282607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 580382607adcSTejun Heo * largely opaque. 580482607adcSTejun Heo * 580582607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 580682607adcSTejun Heo * state if some pools failed to make forward progress for a while where 580782607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 580882607adcSTejun Heo * 580982607adcSTejun Heo * This mechanism is controlled through the kernel parameter 581082607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 581182607adcSTejun Heo * corresponding sysfs parameter file. 581282607adcSTejun Heo */ 581382607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 581482607adcSTejun Heo 581582607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 58165cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 581782607adcSTejun Heo 581882607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 581982607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 582082607adcSTejun Heo 582182607adcSTejun Heo static void wq_watchdog_reset_touched(void) 582282607adcSTejun Heo { 582382607adcSTejun Heo int cpu; 582482607adcSTejun Heo 582582607adcSTejun Heo wq_watchdog_touched = jiffies; 582682607adcSTejun Heo for_each_possible_cpu(cpu) 582782607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 582882607adcSTejun Heo } 582982607adcSTejun Heo 58305cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 583182607adcSTejun Heo { 583282607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 583382607adcSTejun Heo bool lockup_detected = false; 5834940d71c6SSergey Senozhatsky unsigned long now = jiffies; 583582607adcSTejun Heo struct worker_pool *pool; 583682607adcSTejun Heo int pi; 583782607adcSTejun Heo 583882607adcSTejun Heo if (!thresh) 583982607adcSTejun Heo return; 584082607adcSTejun Heo 584182607adcSTejun Heo rcu_read_lock(); 584282607adcSTejun Heo 584382607adcSTejun Heo for_each_pool(pool, pi) { 584482607adcSTejun Heo unsigned long pool_ts, touched, ts; 584582607adcSTejun Heo 584682607adcSTejun Heo if (list_empty(&pool->worklist)) 584782607adcSTejun Heo continue; 584882607adcSTejun Heo 5849940d71c6SSergey Senozhatsky /* 5850940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 5851940d71c6SSergey Senozhatsky * the watchdog like a stall. 5852940d71c6SSergey Senozhatsky */ 5853940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 5854940d71c6SSergey Senozhatsky 585582607adcSTejun Heo /* get the latest of pool and touched timestamps */ 585689e28ce6SWang Qing if (pool->cpu >= 0) 585789e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 585889e28ce6SWang Qing else 585982607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 586089e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 586182607adcSTejun Heo 586282607adcSTejun Heo if (time_after(pool_ts, touched)) 586382607adcSTejun Heo ts = pool_ts; 586482607adcSTejun Heo else 586582607adcSTejun Heo ts = touched; 586682607adcSTejun Heo 586782607adcSTejun Heo /* did we stall? */ 5868940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 586982607adcSTejun Heo lockup_detected = true; 587082607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 587182607adcSTejun Heo pr_cont_pool_info(pool); 587282607adcSTejun Heo pr_cont(" stuck for %us!\n", 5873940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 587482607adcSTejun Heo } 587582607adcSTejun Heo } 587682607adcSTejun Heo 587782607adcSTejun Heo rcu_read_unlock(); 587882607adcSTejun Heo 587982607adcSTejun Heo if (lockup_detected) 588055df0933SImran Khan show_all_workqueues(); 588182607adcSTejun Heo 588282607adcSTejun Heo wq_watchdog_reset_touched(); 588382607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 588482607adcSTejun Heo } 588582607adcSTejun Heo 5886cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 588782607adcSTejun Heo { 588882607adcSTejun Heo if (cpu >= 0) 588982607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 589089e28ce6SWang Qing 589182607adcSTejun Heo wq_watchdog_touched = jiffies; 589282607adcSTejun Heo } 589382607adcSTejun Heo 589482607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 589582607adcSTejun Heo { 589682607adcSTejun Heo wq_watchdog_thresh = 0; 589782607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 589882607adcSTejun Heo 589982607adcSTejun Heo if (thresh) { 590082607adcSTejun Heo wq_watchdog_thresh = thresh; 590182607adcSTejun Heo wq_watchdog_reset_touched(); 590282607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 590382607adcSTejun Heo } 590482607adcSTejun Heo } 590582607adcSTejun Heo 590682607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 590782607adcSTejun Heo const struct kernel_param *kp) 590882607adcSTejun Heo { 590982607adcSTejun Heo unsigned long thresh; 591082607adcSTejun Heo int ret; 591182607adcSTejun Heo 591282607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 591382607adcSTejun Heo if (ret) 591482607adcSTejun Heo return ret; 591582607adcSTejun Heo 591682607adcSTejun Heo if (system_wq) 591782607adcSTejun Heo wq_watchdog_set_thresh(thresh); 591882607adcSTejun Heo else 591982607adcSTejun Heo wq_watchdog_thresh = thresh; 592082607adcSTejun Heo 592182607adcSTejun Heo return 0; 592282607adcSTejun Heo } 592382607adcSTejun Heo 592482607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 592582607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 592682607adcSTejun Heo .get = param_get_ulong, 592782607adcSTejun Heo }; 592882607adcSTejun Heo 592982607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 593082607adcSTejun Heo 0644); 593182607adcSTejun Heo 593282607adcSTejun Heo static void wq_watchdog_init(void) 593382607adcSTejun Heo { 59345cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 593582607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 593682607adcSTejun Heo } 593782607adcSTejun Heo 593882607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 593982607adcSTejun Heo 594082607adcSTejun Heo static inline void wq_watchdog_init(void) { } 594182607adcSTejun Heo 594282607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 594382607adcSTejun Heo 5944bce90380STejun Heo static void __init wq_numa_init(void) 5945bce90380STejun Heo { 5946bce90380STejun Heo cpumask_var_t *tbl; 5947bce90380STejun Heo int node, cpu; 5948bce90380STejun Heo 5949bce90380STejun Heo if (num_possible_nodes() <= 1) 5950bce90380STejun Heo return; 5951bce90380STejun Heo 5952d55262c4STejun Heo if (wq_disable_numa) { 5953d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5954d55262c4STejun Heo return; 5955d55262c4STejun Heo } 5956d55262c4STejun Heo 5957f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 5958f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 5959f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5960f728c4a9SZhen Lei return; 5961f728c4a9SZhen Lei } 5962f728c4a9SZhen Lei } 5963f728c4a9SZhen Lei 5964be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 59654c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 59664c16bd32STejun Heo 5967bce90380STejun Heo /* 5968bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5969bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5970bce90380STejun Heo * fully initialized by now. 5971bce90380STejun Heo */ 59726396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5973bce90380STejun Heo BUG_ON(!tbl); 5974bce90380STejun Heo 5975bce90380STejun Heo for_each_node(node) 59765a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 59771be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5978bce90380STejun Heo 5979bce90380STejun Heo for_each_possible_cpu(cpu) { 5980bce90380STejun Heo node = cpu_to_node(cpu); 5981bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5982bce90380STejun Heo } 5983bce90380STejun Heo 5984bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5985bce90380STejun Heo wq_numa_enabled = true; 5986bce90380STejun Heo } 5987bce90380STejun Heo 59883347fa09STejun Heo /** 59893347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 59903347fa09STejun Heo * 59913347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 59923347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 59933347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 59943347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 59953347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 59963347fa09STejun Heo * created and scheduled right before early initcalls. 59973347fa09STejun Heo */ 59982333e829SYu Chen void __init workqueue_init_early(void) 59991da177e4SLinus Torvalds { 60007a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 60017a4e344cSTejun Heo int i, cpu; 6002c34056a3STejun Heo 600310cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6004e904e6c2STejun Heo 6005b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 600604d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 600704d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6008b05a7928SFrederic Weisbecker 6009e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6010e904e6c2STejun Heo 6011706026c2STejun Heo /* initialize CPU pools */ 601229c91e99STejun Heo for_each_possible_cpu(cpu) { 60134ce62e9eSTejun Heo struct worker_pool *pool; 60148b03ae3cSTejun Heo 60157a4e344cSTejun Heo i = 0; 6016f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 60177a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6018ec22ca5eSTejun Heo pool->cpu = cpu; 60197a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 60207a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6021f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 60227a4e344cSTejun Heo 60239daf9e67STejun Heo /* alloc pool ID */ 602468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 60259daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 602668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 60274ce62e9eSTejun Heo } 60288b03ae3cSTejun Heo } 60298b03ae3cSTejun Heo 60308a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 603129c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 603229c91e99STejun Heo struct workqueue_attrs *attrs; 603329c91e99STejun Heo 6034be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 603529c91e99STejun Heo attrs->nice = std_nice[i]; 603629c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 60378a2b7538STejun Heo 60388a2b7538STejun Heo /* 60398a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 60408a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 60418a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 60428a2b7538STejun Heo */ 6043be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 60448a2b7538STejun Heo attrs->nice = std_nice[i]; 60458a2b7538STejun Heo attrs->no_numa = true; 60468a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 604729c91e99STejun Heo } 604829c91e99STejun Heo 6049d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 60501aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6051d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6052f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6053f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 605424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 605524d51addSTejun Heo WQ_FREEZABLE, 0); 60560668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 60570668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 60580668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 60590668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 60600668106cSViresh Kumar 0); 60611aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 60620668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 60630668106cSViresh Kumar !system_power_efficient_wq || 60640668106cSViresh Kumar !system_freezable_power_efficient_wq); 60653347fa09STejun Heo } 60663347fa09STejun Heo 60673347fa09STejun Heo /** 60683347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 60693347fa09STejun Heo * 60703347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 60713347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 60723347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 60733347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 60743347fa09STejun Heo * with the initial workers and enable future kworker creations. 60753347fa09STejun Heo */ 60762333e829SYu Chen void __init workqueue_init(void) 60773347fa09STejun Heo { 60782186d9f9STejun Heo struct workqueue_struct *wq; 60793347fa09STejun Heo struct worker_pool *pool; 60803347fa09STejun Heo int cpu, bkt; 60813347fa09STejun Heo 60822186d9f9STejun Heo /* 60832186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 60842186d9f9STejun Heo * CPU to node mapping may not be available that early on some 60852186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 60862186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 60872186d9f9STejun Heo * affinity, fix them up. 608840c17f75STejun Heo * 608940c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 60902186d9f9STejun Heo */ 60912186d9f9STejun Heo wq_numa_init(); 60922186d9f9STejun Heo 60932186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 60942186d9f9STejun Heo 60952186d9f9STejun Heo for_each_possible_cpu(cpu) { 60962186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 60972186d9f9STejun Heo pool->node = cpu_to_node(cpu); 60982186d9f9STejun Heo } 60992186d9f9STejun Heo } 61002186d9f9STejun Heo 610140c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 61022186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 610340c17f75STejun Heo WARN(init_rescuer(wq), 610440c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 610540c17f75STejun Heo wq->name); 610640c17f75STejun Heo } 61072186d9f9STejun Heo 61082186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 61092186d9f9STejun Heo 61103347fa09STejun Heo /* create the initial workers */ 61113347fa09STejun Heo for_each_online_cpu(cpu) { 61123347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 61133347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 61143347fa09STejun Heo BUG_ON(!create_worker(pool)); 61153347fa09STejun Heo } 61163347fa09STejun Heo } 61173347fa09STejun Heo 61183347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 61193347fa09STejun Heo BUG_ON(!create_worker(pool)); 61203347fa09STejun Heo 61213347fa09STejun Heo wq_online = true; 612282607adcSTejun Heo wq_watchdog_init(); 61231da177e4SLinus Torvalds } 6124c4f135d6STetsuo Handa 6125c4f135d6STetsuo Handa /* 6126c4f135d6STetsuo Handa * Despite the naming, this is a no-op function which is here only for avoiding 6127c4f135d6STetsuo Handa * link error. Since compile-time warning may fail to catch, we will need to 6128c4f135d6STetsuo Handa * emit run-time warning from __flush_workqueue(). 6129c4f135d6STetsuo Handa */ 6130c4f135d6STetsuo Handa void __warn_flushing_systemwide_wq(void) { } 6131c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6132