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 143633e3f0a3SRichard Clark /* 143733e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 143833e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 143933e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 144033e3f0a3SRichard Clark */ 144133e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 144233e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1443e41e704bSTejun Heo return; 144424acfb71SThomas Gleixner rcu_read_lock(); 14459e8cd2f5STejun Heo retry: 1446aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1447aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1448df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1449ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1450df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1451aa202f1fSHillf Danton } else { 1452aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1453aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1454aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1455aa202f1fSHillf Danton } 1456f3421797STejun Heo 145718aa9effSTejun Heo /* 1458c9178087STejun Heo * If @work was previously on a different pool, it might still be 1459c9178087STejun Heo * running there, in which case the work needs to be queued on that 1460c9178087STejun Heo * pool to guarantee non-reentrancy. 146118aa9effSTejun Heo */ 1462c9e7cf27STejun Heo last_pool = get_work_pool(work); 1463112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 146418aa9effSTejun Heo struct worker *worker; 146518aa9effSTejun Heo 1466a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 146718aa9effSTejun Heo 1468c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 146918aa9effSTejun Heo 1470112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1471c9178087STejun Heo pwq = worker->current_pwq; 14728594fadeSLai Jiangshan } else { 147318aa9effSTejun Heo /* meh... not running there, queue here */ 1474a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1475a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 147618aa9effSTejun Heo } 14778930cabaSTejun Heo } else { 1478a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 14798930cabaSTejun Heo } 1480502ca9d8STejun Heo 14819e8cd2f5STejun Heo /* 14829e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14839e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14849e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1485df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1486df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14879e8cd2f5STejun Heo * make forward-progress. 14889e8cd2f5STejun Heo */ 14899e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14909e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1491a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 14929e8cd2f5STejun Heo cpu_relax(); 14939e8cd2f5STejun Heo goto retry; 14949e8cd2f5STejun Heo } 14959e8cd2f5STejun Heo /* oops */ 14969e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14979e8cd2f5STejun Heo wq->name, cpu); 14989e8cd2f5STejun Heo } 14999e8cd2f5STejun Heo 1500112202d9STejun Heo /* pwq determined, queue */ 1501112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1502502ca9d8STejun Heo 150324acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 150424acfb71SThomas Gleixner goto out; 15051e19ffc6STejun Heo 1506112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1507112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 15081e19ffc6STejun Heo 1509112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1510cdadf009STejun Heo trace_workqueue_activate_work(work); 1511112202d9STejun Heo pwq->nr_active++; 1512112202d9STejun Heo worklist = &pwq->pool->worklist; 151382607adcSTejun Heo if (list_empty(worklist)) 151482607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 15158a2e8e5dSTejun Heo } else { 1516f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1517f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 15188a2e8e5dSTejun Heo } 15191e19ffc6STejun Heo 15200687c66bSZqiang debug_work_activate(work); 1521112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 15221e19ffc6STejun Heo 152324acfb71SThomas Gleixner out: 1524a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 152524acfb71SThomas Gleixner rcu_read_unlock(); 15261da177e4SLinus Torvalds } 15271da177e4SLinus Torvalds 15280fcb78c2SRolf Eike Beer /** 1529c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1530c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1531c1a220e7SZhang Rui * @wq: workqueue to use 1532c1a220e7SZhang Rui * @work: work to queue 1533c1a220e7SZhang Rui * 1534c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1535443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1536443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1537d185af30SYacine Belkadi * 1538d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1539c1a220e7SZhang Rui */ 1540d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1541d4283e93STejun Heo struct work_struct *work) 1542c1a220e7SZhang Rui { 1543d4283e93STejun Heo bool ret = false; 15448930cabaSTejun Heo unsigned long flags; 15458930cabaSTejun Heo 15468930cabaSTejun Heo local_irq_save(flags); 1547c1a220e7SZhang Rui 154822df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15494690c4abSTejun Heo __queue_work(cpu, wq, work); 1550d4283e93STejun Heo ret = true; 1551c1a220e7SZhang Rui } 15528930cabaSTejun Heo 15538930cabaSTejun Heo local_irq_restore(flags); 1554c1a220e7SZhang Rui return ret; 1555c1a220e7SZhang Rui } 1556ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1557c1a220e7SZhang Rui 15588204e0c1SAlexander Duyck /** 15598204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15608204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15618204e0c1SAlexander Duyck * 15628204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15638204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15648204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15658204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15668204e0c1SAlexander Duyck */ 15678204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15688204e0c1SAlexander Duyck { 15698204e0c1SAlexander Duyck int cpu; 15708204e0c1SAlexander Duyck 15718204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15728204e0c1SAlexander Duyck if (!wq_numa_enabled) 15738204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15748204e0c1SAlexander Duyck 15758204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15768204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15778204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15788204e0c1SAlexander Duyck 15798204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15808204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15818204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15828204e0c1SAlexander Duyck return cpu; 15838204e0c1SAlexander Duyck 15848204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15858204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15868204e0c1SAlexander Duyck 15878204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15888204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15898204e0c1SAlexander Duyck } 15908204e0c1SAlexander Duyck 15918204e0c1SAlexander Duyck /** 15928204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15938204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15948204e0c1SAlexander Duyck * @wq: workqueue to use 15958204e0c1SAlexander Duyck * @work: work to queue 15968204e0c1SAlexander Duyck * 15978204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 15988204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 15998204e0c1SAlexander Duyck * NUMA node. 16008204e0c1SAlexander Duyck * 16018204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 16028204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 16038204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 16048204e0c1SAlexander Duyck * 16058204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 16068204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 16078204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 16088204e0c1SAlexander Duyck * 16098204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 16108204e0c1SAlexander Duyck */ 16118204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 16128204e0c1SAlexander Duyck struct work_struct *work) 16138204e0c1SAlexander Duyck { 16148204e0c1SAlexander Duyck unsigned long flags; 16158204e0c1SAlexander Duyck bool ret = false; 16168204e0c1SAlexander Duyck 16178204e0c1SAlexander Duyck /* 16188204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 16198204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 16208204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 16218204e0c1SAlexander Duyck * 16228204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 16238204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 16248204e0c1SAlexander Duyck * some round robin type logic. 16258204e0c1SAlexander Duyck */ 16268204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 16278204e0c1SAlexander Duyck 16288204e0c1SAlexander Duyck local_irq_save(flags); 16298204e0c1SAlexander Duyck 16308204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16318204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16328204e0c1SAlexander Duyck 16338204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16348204e0c1SAlexander Duyck ret = true; 16358204e0c1SAlexander Duyck } 16368204e0c1SAlexander Duyck 16378204e0c1SAlexander Duyck local_irq_restore(flags); 16388204e0c1SAlexander Duyck return ret; 16398204e0c1SAlexander Duyck } 16408204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16418204e0c1SAlexander Duyck 16428c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16431da177e4SLinus Torvalds { 16448c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16451da177e4SLinus Torvalds 1646e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 164760c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16481da177e4SLinus Torvalds } 16491438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16501da177e4SLinus Torvalds 16517beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 165252bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16531da177e4SLinus Torvalds { 16547beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16557beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16561da177e4SLinus Torvalds 1657637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 16584b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1659fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1660fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16617beb2edfSTejun Heo 16628852aac2STejun Heo /* 16638852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16648852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16658852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16668852aac2STejun Heo * on that there's no such delay when @delay is 0. 16678852aac2STejun Heo */ 16688852aac2STejun Heo if (!delay) { 16698852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16708852aac2STejun Heo return; 16718852aac2STejun Heo } 16728852aac2STejun Heo 167360c057bcSLai Jiangshan dwork->wq = wq; 16741265057fSTejun Heo dwork->cpu = cpu; 16757beb2edfSTejun Heo timer->expires = jiffies + delay; 16767beb2edfSTejun Heo 1677041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16787beb2edfSTejun Heo add_timer_on(timer, cpu); 1679041bd12eSTejun Heo else 1680041bd12eSTejun Heo add_timer(timer); 16817beb2edfSTejun Heo } 16821da177e4SLinus Torvalds 16830fcb78c2SRolf Eike Beer /** 16840fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16850fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16860fcb78c2SRolf Eike Beer * @wq: workqueue to use 1687af9997e4SRandy Dunlap * @dwork: work to queue 16880fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16890fcb78c2SRolf Eike Beer * 1690d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1691715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1692715f1300STejun Heo * execution. 16930fcb78c2SRolf Eike Beer */ 1694d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 169552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16967a6bc1cdSVenkatesh Pallipadi { 169752bad64dSDavid Howells struct work_struct *work = &dwork->work; 1698d4283e93STejun Heo bool ret = false; 16998930cabaSTejun Heo unsigned long flags; 17008930cabaSTejun Heo 17018930cabaSTejun Heo /* read the comment in __queue_work() */ 17028930cabaSTejun Heo local_irq_save(flags); 17037a6bc1cdSVenkatesh Pallipadi 170422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17057beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1706d4283e93STejun Heo ret = true; 17077a6bc1cdSVenkatesh Pallipadi } 17088930cabaSTejun Heo 17098930cabaSTejun Heo local_irq_restore(flags); 17107a6bc1cdSVenkatesh Pallipadi return ret; 17117a6bc1cdSVenkatesh Pallipadi } 1712ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 17131da177e4SLinus Torvalds 1714c8e55f36STejun Heo /** 17158376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 17168376fe22STejun Heo * @cpu: CPU number to execute work on 17178376fe22STejun Heo * @wq: workqueue to use 17188376fe22STejun Heo * @dwork: work to queue 17198376fe22STejun Heo * @delay: number of jiffies to wait before queueing 17208376fe22STejun Heo * 17218376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 17228376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 17238376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 17248376fe22STejun Heo * current state. 17258376fe22STejun Heo * 1726d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 17278376fe22STejun Heo * pending and its timer was modified. 17288376fe22STejun Heo * 1729e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17308376fe22STejun Heo * See try_to_grab_pending() for details. 17318376fe22STejun Heo */ 17328376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17338376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17348376fe22STejun Heo { 17358376fe22STejun Heo unsigned long flags; 17368376fe22STejun Heo int ret; 17378376fe22STejun Heo 17388376fe22STejun Heo do { 17398376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17408376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17418376fe22STejun Heo 17428376fe22STejun Heo if (likely(ret >= 0)) { 17438376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17448376fe22STejun Heo local_irq_restore(flags); 17458376fe22STejun Heo } 17468376fe22STejun Heo 17478376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17488376fe22STejun Heo return ret; 17498376fe22STejun Heo } 17508376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17518376fe22STejun Heo 175205f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 175305f0fe6bSTejun Heo { 175405f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 175505f0fe6bSTejun Heo 175605f0fe6bSTejun Heo /* read the comment in __queue_work() */ 175705f0fe6bSTejun Heo local_irq_disable(); 175805f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 175905f0fe6bSTejun Heo local_irq_enable(); 176005f0fe6bSTejun Heo } 176105f0fe6bSTejun Heo 176205f0fe6bSTejun Heo /** 176305f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 176405f0fe6bSTejun Heo * @wq: workqueue to use 176505f0fe6bSTejun Heo * @rwork: work to queue 176605f0fe6bSTejun Heo * 176705f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 176805f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1769bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 177005f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 177105f0fe6bSTejun Heo */ 177205f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 177305f0fe6bSTejun Heo { 177405f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 177505f0fe6bSTejun Heo 177605f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 177705f0fe6bSTejun Heo rwork->wq = wq; 1778a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 177905f0fe6bSTejun Heo return true; 178005f0fe6bSTejun Heo } 178105f0fe6bSTejun Heo 178205f0fe6bSTejun Heo return false; 178305f0fe6bSTejun Heo } 178405f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 178505f0fe6bSTejun Heo 17868376fe22STejun Heo /** 1787c8e55f36STejun Heo * worker_enter_idle - enter idle state 1788c8e55f36STejun Heo * @worker: worker which is entering idle state 1789c8e55f36STejun Heo * 1790c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1791c8e55f36STejun Heo * necessary. 1792c8e55f36STejun Heo * 1793c8e55f36STejun Heo * LOCKING: 1794a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1795c8e55f36STejun Heo */ 1796c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 17971da177e4SLinus Torvalds { 1798bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1799c8e55f36STejun Heo 18006183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 18016183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 18026183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 18036183c009STejun Heo return; 1804c8e55f36STejun Heo 1805051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1806cb444766STejun Heo worker->flags |= WORKER_IDLE; 1807bd7bdd43STejun Heo pool->nr_idle++; 1808e22bee78STejun Heo worker->last_active = jiffies; 1809c8e55f36STejun Heo 1810c8e55f36STejun Heo /* idle_list is LIFO */ 1811bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1812db7bccf4STejun Heo 181363d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1814628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1815cb444766STejun Heo 1816989442d7SLai Jiangshan /* Sanity check nr_running. */ 1817bc35f7efSLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 1818c8e55f36STejun Heo } 1819c8e55f36STejun Heo 1820c8e55f36STejun Heo /** 1821c8e55f36STejun Heo * worker_leave_idle - leave idle state 1822c8e55f36STejun Heo * @worker: worker which is leaving idle state 1823c8e55f36STejun Heo * 1824c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1825c8e55f36STejun Heo * 1826c8e55f36STejun Heo * LOCKING: 1827a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1828c8e55f36STejun Heo */ 1829c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1830c8e55f36STejun Heo { 1831bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1832c8e55f36STejun Heo 18336183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18346183c009STejun Heo return; 1835d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1836bd7bdd43STejun Heo pool->nr_idle--; 1837c8e55f36STejun Heo list_del_init(&worker->entry); 1838c8e55f36STejun Heo } 1839c8e55f36STejun Heo 1840f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1841c34056a3STejun Heo { 1842c34056a3STejun Heo struct worker *worker; 1843c34056a3STejun Heo 1844f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1845c8e55f36STejun Heo if (worker) { 1846c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1847affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1848da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1849e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1850e22bee78STejun Heo worker->flags = WORKER_PREP; 1851c8e55f36STejun Heo } 1852c34056a3STejun Heo return worker; 1853c34056a3STejun Heo } 1854c34056a3STejun Heo 1855c34056a3STejun Heo /** 18564736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18574736cbf7SLai Jiangshan * @worker: worker to be attached 18584736cbf7SLai Jiangshan * @pool: the target pool 18594736cbf7SLai Jiangshan * 18604736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18614736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18624736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18634736cbf7SLai Jiangshan */ 18644736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18654736cbf7SLai Jiangshan struct worker_pool *pool) 18664736cbf7SLai Jiangshan { 18671258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18684736cbf7SLai Jiangshan 18694736cbf7SLai Jiangshan /* 18701258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18711258fae7STejun Heo * stable across this function. See the comments above the flag 18721258fae7STejun Heo * definition for details. 18734736cbf7SLai Jiangshan */ 18744736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 18754736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 18765c25b5ffSPeter Zijlstra else 18775c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 18784736cbf7SLai Jiangshan 1879640f17c8SPeter Zijlstra if (worker->rescue_wq) 1880640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 1881640f17c8SPeter Zijlstra 18824736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1883a2d812a2STejun Heo worker->pool = pool; 18844736cbf7SLai Jiangshan 18851258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 18864736cbf7SLai Jiangshan } 18874736cbf7SLai Jiangshan 18884736cbf7SLai Jiangshan /** 188960f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 189060f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 189160f5a4bcSLai Jiangshan * 18924736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 18934736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 18944736cbf7SLai Jiangshan * other reference to the pool. 189560f5a4bcSLai Jiangshan */ 1896a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 189760f5a4bcSLai Jiangshan { 1898a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 189960f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 190060f5a4bcSLai Jiangshan 19011258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1902a2d812a2STejun Heo 19035c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 1904da028469SLai Jiangshan list_del(&worker->node); 1905a2d812a2STejun Heo worker->pool = NULL; 1906a2d812a2STejun Heo 1907da028469SLai Jiangshan if (list_empty(&pool->workers)) 190860f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 19091258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 191060f5a4bcSLai Jiangshan 1911b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1912b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1913b62c0751SLai Jiangshan 191460f5a4bcSLai Jiangshan if (detach_completion) 191560f5a4bcSLai Jiangshan complete(detach_completion); 191660f5a4bcSLai Jiangshan } 191760f5a4bcSLai Jiangshan 191860f5a4bcSLai Jiangshan /** 1919c34056a3STejun Heo * create_worker - create a new workqueue worker 192063d95a91STejun Heo * @pool: pool the new worker will belong to 1921c34056a3STejun Heo * 1922051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1923c34056a3STejun Heo * 1924c34056a3STejun Heo * CONTEXT: 1925c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1926c34056a3STejun Heo * 1927d185af30SYacine Belkadi * Return: 1928c34056a3STejun Heo * Pointer to the newly created worker. 1929c34056a3STejun Heo */ 1930bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1931c34056a3STejun Heo { 1932e441b56fSZhen Lei struct worker *worker; 1933e441b56fSZhen Lei int id; 1934e3c916a4STejun Heo char id_buf[16]; 1935c34056a3STejun Heo 19367cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 1937e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 1938822d8405STejun Heo if (id < 0) 1939e441b56fSZhen Lei return NULL; 1940c34056a3STejun Heo 1941f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1942c34056a3STejun Heo if (!worker) 1943c34056a3STejun Heo goto fail; 1944c34056a3STejun Heo 1945c34056a3STejun Heo worker->id = id; 1946c34056a3STejun Heo 194729c91e99STejun Heo if (pool->cpu >= 0) 1948e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1949e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1950f3421797STejun Heo else 1951e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1952e3c916a4STejun Heo 1953f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1954e3c916a4STejun Heo "kworker/%s", id_buf); 1955c34056a3STejun Heo if (IS_ERR(worker->task)) 1956c34056a3STejun Heo goto fail; 1957c34056a3STejun Heo 195891151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 195925834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 196091151228SOleg Nesterov 1961da028469SLai Jiangshan /* successful, attach the worker to the pool */ 19624736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1963822d8405STejun Heo 1964051e1850SLai Jiangshan /* start the newly created worker */ 1965a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1966051e1850SLai Jiangshan worker->pool->nr_workers++; 1967051e1850SLai Jiangshan worker_enter_idle(worker); 1968051e1850SLai Jiangshan wake_up_process(worker->task); 1969a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1970051e1850SLai Jiangshan 1971c34056a3STejun Heo return worker; 1972822d8405STejun Heo 1973c34056a3STejun Heo fail: 1974e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 1975c34056a3STejun Heo kfree(worker); 1976c34056a3STejun Heo return NULL; 1977c34056a3STejun Heo } 1978c34056a3STejun Heo 1979c34056a3STejun Heo /** 1980c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1981c34056a3STejun Heo * @worker: worker to be destroyed 1982c34056a3STejun Heo * 198373eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 198473eb7fe7SLai Jiangshan * be idle. 1985c8e55f36STejun Heo * 1986c8e55f36STejun Heo * CONTEXT: 1987a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1988c34056a3STejun Heo */ 1989c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1990c34056a3STejun Heo { 1991bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1992c34056a3STejun Heo 1993cd549687STejun Heo lockdep_assert_held(&pool->lock); 1994cd549687STejun Heo 1995c34056a3STejun Heo /* sanity check frenzy */ 19966183c009STejun Heo if (WARN_ON(worker->current_work) || 199773eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 199873eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 19996183c009STejun Heo return; 2000c34056a3STejun Heo 2001bd7bdd43STejun Heo pool->nr_workers--; 2002bd7bdd43STejun Heo pool->nr_idle--; 2003c8e55f36STejun Heo 2004c8e55f36STejun Heo list_del_init(&worker->entry); 2005cb444766STejun Heo worker->flags |= WORKER_DIE; 200660f5a4bcSLai Jiangshan wake_up_process(worker->task); 2007c34056a3STejun Heo } 2008c34056a3STejun Heo 200932a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2010e22bee78STejun Heo { 201132a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 2012e22bee78STejun Heo 2013a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2014e22bee78STejun Heo 20153347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2016e22bee78STejun Heo struct worker *worker; 2017e22bee78STejun Heo unsigned long expires; 2018e22bee78STejun Heo 2019e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 202063d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2021e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2022e22bee78STejun Heo 20233347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 202463d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 20253347fc9fSLai Jiangshan break; 2026e22bee78STejun Heo } 20273347fc9fSLai Jiangshan 20283347fc9fSLai Jiangshan destroy_worker(worker); 2029e22bee78STejun Heo } 2030e22bee78STejun Heo 2031a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2032e22bee78STejun Heo } 2033e22bee78STejun Heo 2034493a1724STejun Heo static void send_mayday(struct work_struct *work) 2035e22bee78STejun Heo { 2036112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2037112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2038493a1724STejun Heo 20392e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2040e22bee78STejun Heo 2041493008a8STejun Heo if (!wq->rescuer) 2042493a1724STejun Heo return; 2043e22bee78STejun Heo 2044e22bee78STejun Heo /* mayday mayday mayday */ 2045493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 204677668c8bSLai Jiangshan /* 204777668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 204877668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 204977668c8bSLai Jiangshan * rescuer is done with it. 205077668c8bSLai Jiangshan */ 205177668c8bSLai Jiangshan get_pwq(pwq); 2052493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2053e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2054493a1724STejun Heo } 2055e22bee78STejun Heo } 2056e22bee78STejun Heo 205732a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2058e22bee78STejun Heo { 205932a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2060e22bee78STejun Heo struct work_struct *work; 2061e22bee78STejun Heo 2062a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2063a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2064e22bee78STejun Heo 206563d95a91STejun Heo if (need_to_create_worker(pool)) { 2066e22bee78STejun Heo /* 2067e22bee78STejun Heo * We've been trying to create a new worker but 2068e22bee78STejun Heo * haven't been successful. We might be hitting an 2069e22bee78STejun Heo * allocation deadlock. Send distress signals to 2070e22bee78STejun Heo * rescuers. 2071e22bee78STejun Heo */ 207263d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2073e22bee78STejun Heo send_mayday(work); 2074e22bee78STejun Heo } 2075e22bee78STejun Heo 2076a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2077a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2078e22bee78STejun Heo 207963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2080e22bee78STejun Heo } 2081e22bee78STejun Heo 2082e22bee78STejun Heo /** 2083e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 208463d95a91STejun Heo * @pool: pool to create a new worker for 2085e22bee78STejun Heo * 208663d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2087e22bee78STejun Heo * have at least one idle worker on return from this function. If 2088e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 208963d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2090e22bee78STejun Heo * possible allocation deadlock. 2091e22bee78STejun Heo * 2092c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2093c5aa87bbSTejun Heo * may_start_working() %true. 2094e22bee78STejun Heo * 2095e22bee78STejun Heo * LOCKING: 2096a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2097e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2098e22bee78STejun Heo * manager. 2099e22bee78STejun Heo */ 210029187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2101d565ed63STejun Heo __releases(&pool->lock) 2102d565ed63STejun Heo __acquires(&pool->lock) 2103e22bee78STejun Heo { 2104e22bee78STejun Heo restart: 2105a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 21069f9c2364STejun Heo 2107e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 210863d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2109e22bee78STejun Heo 2110e22bee78STejun Heo while (true) { 2111051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2112e22bee78STejun Heo break; 2113e22bee78STejun Heo 2114e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 21159f9c2364STejun Heo 211663d95a91STejun Heo if (!need_to_create_worker(pool)) 2117e22bee78STejun Heo break; 2118e22bee78STejun Heo } 2119e22bee78STejun Heo 212063d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2121a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2122051e1850SLai Jiangshan /* 2123051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2124051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2125051e1850SLai Jiangshan * already become busy. 2126051e1850SLai Jiangshan */ 212763d95a91STejun Heo if (need_to_create_worker(pool)) 2128e22bee78STejun Heo goto restart; 2129e22bee78STejun Heo } 2130e22bee78STejun Heo 2131e22bee78STejun Heo /** 2132e22bee78STejun Heo * manage_workers - manage worker pool 2133e22bee78STejun Heo * @worker: self 2134e22bee78STejun Heo * 2135706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2136e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2137706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2138e22bee78STejun Heo * 2139e22bee78STejun Heo * The caller can safely start processing works on false return. On 2140e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2141e22bee78STejun Heo * and may_start_working() is true. 2142e22bee78STejun Heo * 2143e22bee78STejun Heo * CONTEXT: 2144a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2145e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2146e22bee78STejun Heo * 2147d185af30SYacine Belkadi * Return: 214829187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 214929187a9eSTejun Heo * start processing works, %true if management function was performed and 215029187a9eSTejun Heo * the conditions that the caller verified before calling the function may 215129187a9eSTejun Heo * no longer be true. 2152e22bee78STejun Heo */ 2153e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2154e22bee78STejun Heo { 215563d95a91STejun Heo struct worker_pool *pool = worker->pool; 2156e22bee78STejun Heo 2157692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 215829187a9eSTejun Heo return false; 2159692b4825STejun Heo 2160692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 21612607d7a6STejun Heo pool->manager = worker; 2162e22bee78STejun Heo 216329187a9eSTejun Heo maybe_create_worker(pool); 2164e22bee78STejun Heo 21652607d7a6STejun Heo pool->manager = NULL; 2166692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2167d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 216829187a9eSTejun Heo return true; 2169e22bee78STejun Heo } 2170e22bee78STejun Heo 2171a62428c0STejun Heo /** 2172a62428c0STejun Heo * process_one_work - process single work 2173c34056a3STejun Heo * @worker: self 2174a62428c0STejun Heo * @work: work to process 2175a62428c0STejun Heo * 2176a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2177a62428c0STejun Heo * process a single work including synchronization against and 2178a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2179a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2180a62428c0STejun Heo * call this function to process a work. 2181a62428c0STejun Heo * 2182a62428c0STejun Heo * CONTEXT: 2183a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2184a62428c0STejun Heo */ 2185c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2186d565ed63STejun Heo __releases(&pool->lock) 2187d565ed63STejun Heo __acquires(&pool->lock) 21881da177e4SLinus Torvalds { 2189112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2190bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2191112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 2192c4560c2cSLai Jiangshan unsigned long work_data; 21937e11629dSTejun Heo struct worker *collision; 21944e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21954e6045f1SJohannes Berg /* 2196a62428c0STejun Heo * It is permissible to free the struct work_struct from 2197a62428c0STejun Heo * inside the function that is called from it, this we need to 2198a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2199a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2200a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 22014e6045f1SJohannes Berg */ 22024d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 22034d82a1deSPeter Zijlstra 22044d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 22054e6045f1SJohannes Berg #endif 2206807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 220785327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2208ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 220925511a47STejun Heo 22107e11629dSTejun Heo /* 22117e11629dSTejun Heo * A single work shouldn't be executed concurrently by 22127e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 22137e11629dSTejun Heo * already processing the work. If so, defer the work to the 22147e11629dSTejun Heo * currently executing one. 22157e11629dSTejun Heo */ 2216c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 22177e11629dSTejun Heo if (unlikely(collision)) { 22187e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 22197e11629dSTejun Heo return; 22207e11629dSTejun Heo } 22211da177e4SLinus Torvalds 22228930cabaSTejun Heo /* claim and dequeue */ 22231da177e4SLinus Torvalds debug_work_deactivate(work); 2224c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2225c34056a3STejun Heo worker->current_work = work; 2226a2c1c57bSTejun Heo worker->current_func = work->func; 2227112202d9STejun Heo worker->current_pwq = pwq; 2228c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2229d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 22307a22ad75STejun Heo 22318bf89593STejun Heo /* 22328bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 22338bf89593STejun Heo * overridden through set_worker_desc(). 22348bf89593STejun Heo */ 22358bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 22368bf89593STejun Heo 2237a62428c0STejun Heo list_del_init(&work->entry); 2238a62428c0STejun Heo 2239649027d7STejun Heo /* 2240228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2241228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2242228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2243228f1d00SLai Jiangshan * execution of the pending work items. 2244fb0e7bebSTejun Heo */ 2245fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2246228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2247fb0e7bebSTejun Heo 2248974271c4STejun Heo /* 2249a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2250a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2251a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2252a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2253228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2254974271c4STejun Heo */ 2255a489a03eSLai Jiangshan if (need_more_worker(pool)) 225663d95a91STejun Heo wake_up_worker(pool); 2257974271c4STejun Heo 22588930cabaSTejun Heo /* 22597c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2260d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 226123657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 226223657bb1STejun Heo * disabled. 22638930cabaSTejun Heo */ 22647c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22651da177e4SLinus Torvalds 2266a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2267365970a1SDavid Howells 2268a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 22693295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2270e6f3faa7SPeter Zijlstra /* 2271f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2272f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2273e6f3faa7SPeter Zijlstra * 2274e6f3faa7SPeter Zijlstra * However, that would result in: 2275e6f3faa7SPeter Zijlstra * 2276e6f3faa7SPeter Zijlstra * A(W1) 2277e6f3faa7SPeter Zijlstra * WFC(C) 2278e6f3faa7SPeter Zijlstra * A(W1) 2279e6f3faa7SPeter Zijlstra * C(C) 2280e6f3faa7SPeter Zijlstra * 2281e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2282e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2283e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2284f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2285e6f3faa7SPeter Zijlstra * these locks. 2286e6f3faa7SPeter Zijlstra * 2287e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2288e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2289e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2290e6f3faa7SPeter Zijlstra */ 2291f52be570SPeter Zijlstra lockdep_invariant_state(true); 2292e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2293a2c1c57bSTejun Heo worker->current_func(work); 2294e36c886aSArjan van de Ven /* 2295e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2296e36c886aSArjan van de Ven * point will only record its address. 2297e36c886aSArjan van de Ven */ 22981c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 22993295f0efSIngo Molnar lock_map_release(&lockdep_map); 2300112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 23011da177e4SLinus Torvalds 2302d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2303044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2304d75f773cSSakari Ailus " last function: %ps\n", 2305a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2306a2c1c57bSTejun Heo worker->current_func); 2307d5abe669SPeter Zijlstra debug_show_held_locks(current); 2308d5abe669SPeter Zijlstra dump_stack(); 2309d5abe669SPeter Zijlstra } 2310d5abe669SPeter Zijlstra 2311b22ce278STejun Heo /* 2312025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2313b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2314b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2315b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2316789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2317789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2318b22ce278STejun Heo */ 2319a7e6425eSPaul E. McKenney cond_resched(); 2320b22ce278STejun Heo 2321a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2322a62428c0STejun Heo 2323fb0e7bebSTejun Heo /* clear cpu intensive status */ 2324fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2325fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2326fb0e7bebSTejun Heo 23271b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 23281b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 23291b69ac6bSJohannes Weiner 2330a62428c0STejun Heo /* we're done with it, release */ 233142f8570fSSasha Levin hash_del(&worker->hentry); 2332c34056a3STejun Heo worker->current_work = NULL; 2333a2c1c57bSTejun Heo worker->current_func = NULL; 2334112202d9STejun Heo worker->current_pwq = NULL; 2335d812796eSLai Jiangshan worker->current_color = INT_MAX; 2336c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 23371da177e4SLinus Torvalds } 23381da177e4SLinus Torvalds 2339affee4b2STejun Heo /** 2340affee4b2STejun Heo * process_scheduled_works - process scheduled works 2341affee4b2STejun Heo * @worker: self 2342affee4b2STejun Heo * 2343affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2344affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2345affee4b2STejun Heo * fetches a work from the top and executes it. 2346affee4b2STejun Heo * 2347affee4b2STejun Heo * CONTEXT: 2348a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2349affee4b2STejun Heo * multiple times. 2350affee4b2STejun Heo */ 2351affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 23521da177e4SLinus Torvalds { 2353affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2354affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2355a62428c0STejun Heo struct work_struct, entry); 2356c34056a3STejun Heo process_one_work(worker, work); 2357a62428c0STejun Heo } 23581da177e4SLinus Torvalds } 23591da177e4SLinus Torvalds 2360197f6accSTejun Heo static void set_pf_worker(bool val) 2361197f6accSTejun Heo { 2362197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2363197f6accSTejun Heo if (val) 2364197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2365197f6accSTejun Heo else 2366197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2367197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2368197f6accSTejun Heo } 2369197f6accSTejun Heo 23704690c4abSTejun Heo /** 23714690c4abSTejun Heo * worker_thread - the worker thread function 2372c34056a3STejun Heo * @__worker: self 23734690c4abSTejun Heo * 2374c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2375c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2376c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2377c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2378c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2379d185af30SYacine Belkadi * 2380d185af30SYacine Belkadi * Return: 0 23814690c4abSTejun Heo */ 2382c34056a3STejun Heo static int worker_thread(void *__worker) 23831da177e4SLinus Torvalds { 2384c34056a3STejun Heo struct worker *worker = __worker; 2385bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 23861da177e4SLinus Torvalds 2387e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2388197f6accSTejun Heo set_pf_worker(true); 2389c8e55f36STejun Heo woke_up: 2390a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2391affee4b2STejun Heo 2392a9ab775bSTejun Heo /* am I supposed to die? */ 2393a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2394a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2395a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2396197f6accSTejun Heo set_pf_worker(false); 239760f5a4bcSLai Jiangshan 239860f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2399e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2400a2d812a2STejun Heo worker_detach_from_pool(worker); 240160f5a4bcSLai Jiangshan kfree(worker); 2402c8e55f36STejun Heo return 0; 2403c8e55f36STejun Heo } 2404c8e55f36STejun Heo 2405c8e55f36STejun Heo worker_leave_idle(worker); 2406db7bccf4STejun Heo recheck: 2407e22bee78STejun Heo /* no more worker necessary? */ 240863d95a91STejun Heo if (!need_more_worker(pool)) 2409e22bee78STejun Heo goto sleep; 2410e22bee78STejun Heo 2411e22bee78STejun Heo /* do we need to manage? */ 241263d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2413e22bee78STejun Heo goto recheck; 2414e22bee78STejun Heo 2415c8e55f36STejun Heo /* 2416c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2417c8e55f36STejun Heo * preparing to process a work or actually processing it. 2418c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2419c8e55f36STejun Heo */ 24206183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2421c8e55f36STejun Heo 2422e22bee78STejun Heo /* 2423a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2424a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2425a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2426a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2427a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2428e22bee78STejun Heo */ 2429a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2430e22bee78STejun Heo 2431e22bee78STejun Heo do { 2432affee4b2STejun Heo struct work_struct *work = 2433bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2434affee4b2STejun Heo struct work_struct, entry); 2435affee4b2STejun Heo 243682607adcSTejun Heo pool->watchdog_ts = jiffies; 243782607adcSTejun Heo 2438c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2439affee4b2STejun Heo /* optimization path, not strictly necessary */ 2440affee4b2STejun Heo process_one_work(worker, work); 2441affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2442affee4b2STejun Heo process_scheduled_works(worker); 2443affee4b2STejun Heo } else { 2444c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2445affee4b2STejun Heo process_scheduled_works(worker); 2446affee4b2STejun Heo } 244763d95a91STejun Heo } while (keep_working(pool)); 2448affee4b2STejun Heo 2449228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2450d313dd85STejun Heo sleep: 2451c8e55f36STejun Heo /* 2452d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2453d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2454d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2455d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2456d565ed63STejun Heo * event. 2457c8e55f36STejun Heo */ 2458c8e55f36STejun Heo worker_enter_idle(worker); 2459c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2460a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24611da177e4SLinus Torvalds schedule(); 2462c8e55f36STejun Heo goto woke_up; 24631da177e4SLinus Torvalds } 24641da177e4SLinus Torvalds 2465e22bee78STejun Heo /** 2466e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2467111c225aSTejun Heo * @__rescuer: self 2468e22bee78STejun Heo * 2469e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2470493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2471e22bee78STejun Heo * 2472706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2473e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2474e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2475e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2476e22bee78STejun Heo * the problem rescuer solves. 2477e22bee78STejun Heo * 2478706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2479706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2480e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2481e22bee78STejun Heo * 2482e22bee78STejun Heo * This should happen rarely. 2483d185af30SYacine Belkadi * 2484d185af30SYacine Belkadi * Return: 0 2485e22bee78STejun Heo */ 2486111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2487e22bee78STejun Heo { 2488111c225aSTejun Heo struct worker *rescuer = __rescuer; 2489111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2490e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 24914d595b86SLai Jiangshan bool should_stop; 2492e22bee78STejun Heo 2493e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2494111c225aSTejun Heo 2495111c225aSTejun Heo /* 2496111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2497111c225aSTejun Heo * doesn't participate in concurrency management. 2498111c225aSTejun Heo */ 2499197f6accSTejun Heo set_pf_worker(true); 2500e22bee78STejun Heo repeat: 2501c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 25021da177e4SLinus Torvalds 25034d595b86SLai Jiangshan /* 25044d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 25054d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 25064d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 25074d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 25084d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 25094d595b86SLai Jiangshan * list is always empty on exit. 25104d595b86SLai Jiangshan */ 25114d595b86SLai Jiangshan should_stop = kthread_should_stop(); 25121da177e4SLinus Torvalds 2513493a1724STejun Heo /* see whether any pwq is asking for help */ 2514a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2515493a1724STejun Heo 2516493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2517493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2518493a1724STejun Heo struct pool_workqueue, mayday_node); 2519112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2520e22bee78STejun Heo struct work_struct *work, *n; 252182607adcSTejun Heo bool first = true; 2522e22bee78STejun Heo 2523e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2524493a1724STejun Heo list_del_init(&pwq->mayday_node); 2525493a1724STejun Heo 2526a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2527e22bee78STejun Heo 252851697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 252951697d39SLai Jiangshan 2530a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2531e22bee78STejun Heo 2532e22bee78STejun Heo /* 2533e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2534e22bee78STejun Heo * process'em. 2535e22bee78STejun Heo */ 25360479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 253782607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 253882607adcSTejun Heo if (get_work_pwq(work) == pwq) { 253982607adcSTejun Heo if (first) 254082607adcSTejun Heo pool->watchdog_ts = jiffies; 2541e22bee78STejun Heo move_linked_works(work, scheduled, &n); 254282607adcSTejun Heo } 254382607adcSTejun Heo first = false; 254482607adcSTejun Heo } 2545e22bee78STejun Heo 2546008847f6SNeilBrown if (!list_empty(scheduled)) { 2547e22bee78STejun Heo process_scheduled_works(rescuer); 25487576958aSTejun Heo 25497576958aSTejun Heo /* 2550008847f6SNeilBrown * The above execution of rescued work items could 2551008847f6SNeilBrown * have created more to rescue through 2552f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2553008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2554008847f6SNeilBrown * that such back-to-back work items, which may be 2555008847f6SNeilBrown * being used to relieve memory pressure, don't 2556008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2557008847f6SNeilBrown */ 25584f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2559a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2560e66b39afSTejun Heo /* 2561e66b39afSTejun Heo * Queue iff we aren't racing destruction 2562e66b39afSTejun Heo * and somebody else hasn't queued it already. 2563e66b39afSTejun Heo */ 2564e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2565008847f6SNeilBrown get_pwq(pwq); 2566e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2567e66b39afSTejun Heo } 2568a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2569008847f6SNeilBrown } 2570008847f6SNeilBrown } 2571008847f6SNeilBrown 2572008847f6SNeilBrown /* 257377668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 257413b1d625SLai Jiangshan * go away while we're still attached to it. 257577668c8bSLai Jiangshan */ 257677668c8bSLai Jiangshan put_pwq(pwq); 257777668c8bSLai Jiangshan 257877668c8bSLai Jiangshan /* 2579d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 25807576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 25817576958aSTejun Heo * and stalling the execution. 25827576958aSTejun Heo */ 2583d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 258463d95a91STejun Heo wake_up_worker(pool); 25857576958aSTejun Heo 2586a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 258713b1d625SLai Jiangshan 2588a2d812a2STejun Heo worker_detach_from_pool(rescuer); 258913b1d625SLai Jiangshan 2590a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 25911da177e4SLinus Torvalds } 25921da177e4SLinus Torvalds 2593a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2594493a1724STejun Heo 25954d595b86SLai Jiangshan if (should_stop) { 25964d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2597197f6accSTejun Heo set_pf_worker(false); 25984d595b86SLai Jiangshan return 0; 25994d595b86SLai Jiangshan } 26004d595b86SLai Jiangshan 2601111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2602111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2603e22bee78STejun Heo schedule(); 2604e22bee78STejun Heo goto repeat; 26051da177e4SLinus Torvalds } 26061da177e4SLinus Torvalds 2607fca839c0STejun Heo /** 2608fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2609fca839c0STejun Heo * @target_wq: workqueue being flushed 2610fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2611fca839c0STejun Heo * 2612fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2613fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2614fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2615fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2616fca839c0STejun Heo * a deadlock. 2617fca839c0STejun Heo */ 2618fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2619fca839c0STejun Heo struct work_struct *target_work) 2620fca839c0STejun Heo { 2621fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2622fca839c0STejun Heo struct worker *worker; 2623fca839c0STejun Heo 2624fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2625fca839c0STejun Heo return; 2626fca839c0STejun Heo 2627fca839c0STejun Heo worker = current_wq_worker(); 2628fca839c0STejun Heo 2629fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2630d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2631fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 263223d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 263323d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2634d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2635fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2636fca839c0STejun Heo target_wq->name, target_func); 2637fca839c0STejun Heo } 2638fca839c0STejun Heo 2639fc2e4d70SOleg Nesterov struct wq_barrier { 2640fc2e4d70SOleg Nesterov struct work_struct work; 2641fc2e4d70SOleg Nesterov struct completion done; 26422607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2643fc2e4d70SOleg Nesterov }; 2644fc2e4d70SOleg Nesterov 2645fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2646fc2e4d70SOleg Nesterov { 2647fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2648fc2e4d70SOleg Nesterov complete(&barr->done); 2649fc2e4d70SOleg Nesterov } 2650fc2e4d70SOleg Nesterov 26514690c4abSTejun Heo /** 26524690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2653112202d9STejun Heo * @pwq: pwq to insert barrier into 26544690c4abSTejun Heo * @barr: wq_barrier to insert 2655affee4b2STejun Heo * @target: target work to attach @barr to 2656affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 26574690c4abSTejun Heo * 2658affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2659affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2660affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2661affee4b2STejun Heo * cpu. 2662affee4b2STejun Heo * 2663affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2664affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2665affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2666affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2667affee4b2STejun Heo * after a work with LINKED flag set. 2668affee4b2STejun Heo * 2669affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2670112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 26714690c4abSTejun Heo * 26724690c4abSTejun Heo * CONTEXT: 2673a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 26744690c4abSTejun Heo */ 2675112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2676affee4b2STejun Heo struct wq_barrier *barr, 2677affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2678fc2e4d70SOleg Nesterov { 2679d812796eSLai Jiangshan unsigned int work_flags = 0; 2680d812796eSLai Jiangshan unsigned int work_color; 2681affee4b2STejun Heo struct list_head *head; 2682affee4b2STejun Heo 2683dc186ad7SThomas Gleixner /* 2684d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2685dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2686dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2687dc186ad7SThomas Gleixner * might deadlock. 2688dc186ad7SThomas Gleixner */ 2689ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 269022df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 269152fa5bc5SBoqun Feng 2692fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2693fd1a5b04SByungchul Park 26942607d7a6STejun Heo barr->task = current; 269583c22520SOleg Nesterov 2696018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2697018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2698018f3a13SLai Jiangshan 2699affee4b2STejun Heo /* 2700affee4b2STejun Heo * If @target is currently being executed, schedule the 2701affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2702affee4b2STejun Heo */ 2703d812796eSLai Jiangshan if (worker) { 2704affee4b2STejun Heo head = worker->scheduled.next; 2705d812796eSLai Jiangshan work_color = worker->current_color; 2706d812796eSLai Jiangshan } else { 2707affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2708affee4b2STejun Heo 2709affee4b2STejun Heo head = target->entry.next; 2710affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2711d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 2712d812796eSLai Jiangshan work_color = get_work_color(*bits); 2713affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2714affee4b2STejun Heo } 2715affee4b2STejun Heo 2716d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 2717d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 2718d812796eSLai Jiangshan 2719dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2720d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 2721fc2e4d70SOleg Nesterov } 2722fc2e4d70SOleg Nesterov 272373f53c4aSTejun Heo /** 2724112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 272573f53c4aSTejun Heo * @wq: workqueue being flushed 272673f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 272773f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 272873f53c4aSTejun Heo * 2729112202d9STejun Heo * Prepare pwqs for workqueue flushing. 273073f53c4aSTejun Heo * 2731112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2732112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2733112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2734112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2735112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 273673f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 273773f53c4aSTejun Heo * 273873f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 273973f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 274073f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 274173f53c4aSTejun Heo * is returned. 274273f53c4aSTejun Heo * 2743112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 274473f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 274573f53c4aSTejun Heo * advanced to @work_color. 274673f53c4aSTejun Heo * 274773f53c4aSTejun Heo * CONTEXT: 27483c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 274973f53c4aSTejun Heo * 2750d185af30SYacine Belkadi * Return: 275173f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 275273f53c4aSTejun Heo * otherwise. 275373f53c4aSTejun Heo */ 2754112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 275573f53c4aSTejun Heo int flush_color, int work_color) 27561da177e4SLinus Torvalds { 275773f53c4aSTejun Heo bool wait = false; 275849e3cf44STejun Heo struct pool_workqueue *pwq; 27591da177e4SLinus Torvalds 276073f53c4aSTejun Heo if (flush_color >= 0) { 27616183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2762112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2763dc186ad7SThomas Gleixner } 276414441960SOleg Nesterov 276549e3cf44STejun Heo for_each_pwq(pwq, wq) { 2766112202d9STejun Heo struct worker_pool *pool = pwq->pool; 27671da177e4SLinus Torvalds 2768a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 276973f53c4aSTejun Heo 277073f53c4aSTejun Heo if (flush_color >= 0) { 27716183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 277273f53c4aSTejun Heo 2773112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2774112202d9STejun Heo pwq->flush_color = flush_color; 2775112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 277673f53c4aSTejun Heo wait = true; 27771da177e4SLinus Torvalds } 277873f53c4aSTejun Heo } 277973f53c4aSTejun Heo 278073f53c4aSTejun Heo if (work_color >= 0) { 27816183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2782112202d9STejun Heo pwq->work_color = work_color; 278373f53c4aSTejun Heo } 278473f53c4aSTejun Heo 2785a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27861da177e4SLinus Torvalds } 27871da177e4SLinus Torvalds 2788112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 278973f53c4aSTejun Heo complete(&wq->first_flusher->done); 279073f53c4aSTejun Heo 279173f53c4aSTejun Heo return wait; 279283c22520SOleg Nesterov } 27931da177e4SLinus Torvalds 27940fcb78c2SRolf Eike Beer /** 2795c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 27960fcb78c2SRolf Eike Beer * @wq: workqueue to flush 27971da177e4SLinus Torvalds * 2798c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2799c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 28001da177e4SLinus Torvalds */ 2801c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 28021da177e4SLinus Torvalds { 280373f53c4aSTejun Heo struct wq_flusher this_flusher = { 280473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 280573f53c4aSTejun Heo .flush_color = -1, 2806fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 280773f53c4aSTejun Heo }; 280873f53c4aSTejun Heo int next_color; 2809b1f4ec17SOleg Nesterov 28103347fa09STejun Heo if (WARN_ON(!wq_online)) 28113347fa09STejun Heo return; 28123347fa09STejun Heo 281387915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 281487915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 281587915adcSJohannes Berg 28163c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 281773f53c4aSTejun Heo 281873f53c4aSTejun Heo /* 281973f53c4aSTejun Heo * Start-to-wait phase 282073f53c4aSTejun Heo */ 282173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 282273f53c4aSTejun Heo 282373f53c4aSTejun Heo if (next_color != wq->flush_color) { 282473f53c4aSTejun Heo /* 282573f53c4aSTejun Heo * Color space is not full. The current work_color 282673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 282773f53c4aSTejun Heo * by one. 282873f53c4aSTejun Heo */ 28296183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 283073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 283173f53c4aSTejun Heo wq->work_color = next_color; 283273f53c4aSTejun Heo 283373f53c4aSTejun Heo if (!wq->first_flusher) { 283473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 28356183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 283673f53c4aSTejun Heo 283773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 283873f53c4aSTejun Heo 2839112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 284073f53c4aSTejun Heo wq->work_color)) { 284173f53c4aSTejun Heo /* nothing to flush, done */ 284273f53c4aSTejun Heo wq->flush_color = next_color; 284373f53c4aSTejun Heo wq->first_flusher = NULL; 284473f53c4aSTejun Heo goto out_unlock; 284573f53c4aSTejun Heo } 284673f53c4aSTejun Heo } else { 284773f53c4aSTejun Heo /* wait in queue */ 28486183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 284973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2850112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 285173f53c4aSTejun Heo } 285273f53c4aSTejun Heo } else { 285373f53c4aSTejun Heo /* 285473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 285573f53c4aSTejun Heo * The next flush completion will assign us 285673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 285773f53c4aSTejun Heo */ 285873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 285973f53c4aSTejun Heo } 286073f53c4aSTejun Heo 2861fca839c0STejun Heo check_flush_dependency(wq, NULL); 2862fca839c0STejun Heo 28633c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 286473f53c4aSTejun Heo 286573f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 286673f53c4aSTejun Heo 286773f53c4aSTejun Heo /* 286873f53c4aSTejun Heo * Wake-up-and-cascade phase 286973f53c4aSTejun Heo * 287073f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 287173f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 287273f53c4aSTejun Heo */ 287300d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 287473f53c4aSTejun Heo return; 287573f53c4aSTejun Heo 28763c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 287773f53c4aSTejun Heo 28784ce48b37STejun Heo /* we might have raced, check again with mutex held */ 28794ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 28804ce48b37STejun Heo goto out_unlock; 28814ce48b37STejun Heo 288200d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 288373f53c4aSTejun Heo 28846183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 28856183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 288673f53c4aSTejun Heo 288773f53c4aSTejun Heo while (true) { 288873f53c4aSTejun Heo struct wq_flusher *next, *tmp; 288973f53c4aSTejun Heo 289073f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 289173f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 289273f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 289373f53c4aSTejun Heo break; 289473f53c4aSTejun Heo list_del_init(&next->list); 289573f53c4aSTejun Heo complete(&next->done); 289673f53c4aSTejun Heo } 289773f53c4aSTejun Heo 28986183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 289973f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 290073f53c4aSTejun Heo 290173f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 290273f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 290373f53c4aSTejun Heo 290473f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 290573f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 290673f53c4aSTejun Heo /* 290773f53c4aSTejun Heo * Assign the same color to all overflowed 290873f53c4aSTejun Heo * flushers, advance work_color and append to 290973f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 291073f53c4aSTejun Heo * phase for these overflowed flushers. 291173f53c4aSTejun Heo */ 291273f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 291373f53c4aSTejun Heo tmp->flush_color = wq->work_color; 291473f53c4aSTejun Heo 291573f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 291673f53c4aSTejun Heo 291773f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 291873f53c4aSTejun Heo &wq->flusher_queue); 2919112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 292073f53c4aSTejun Heo } 292173f53c4aSTejun Heo 292273f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 29236183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 292473f53c4aSTejun Heo break; 292573f53c4aSTejun Heo } 292673f53c4aSTejun Heo 292773f53c4aSTejun Heo /* 292873f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2929112202d9STejun Heo * the new first flusher and arm pwqs. 293073f53c4aSTejun Heo */ 29316183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 29326183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 293373f53c4aSTejun Heo 293473f53c4aSTejun Heo list_del_init(&next->list); 293573f53c4aSTejun Heo wq->first_flusher = next; 293673f53c4aSTejun Heo 2937112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 293873f53c4aSTejun Heo break; 293973f53c4aSTejun Heo 294073f53c4aSTejun Heo /* 294173f53c4aSTejun Heo * Meh... this color is already done, clear first 294273f53c4aSTejun Heo * flusher and repeat cascading. 294373f53c4aSTejun Heo */ 294473f53c4aSTejun Heo wq->first_flusher = NULL; 294573f53c4aSTejun Heo } 294673f53c4aSTejun Heo 294773f53c4aSTejun Heo out_unlock: 29483c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 29491da177e4SLinus Torvalds } 2950c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 29511da177e4SLinus Torvalds 29529c5a2ba7STejun Heo /** 29539c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 29549c5a2ba7STejun Heo * @wq: workqueue to drain 29559c5a2ba7STejun Heo * 29569c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 29579c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 29589c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2959b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 29609c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 29619c5a2ba7STejun Heo * takes too long. 29629c5a2ba7STejun Heo */ 29639c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 29649c5a2ba7STejun Heo { 29659c5a2ba7STejun Heo unsigned int flush_cnt = 0; 296649e3cf44STejun Heo struct pool_workqueue *pwq; 29679c5a2ba7STejun Heo 29689c5a2ba7STejun Heo /* 29699c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 29709c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2971618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 29729c5a2ba7STejun Heo */ 297387fc741eSLai Jiangshan mutex_lock(&wq->mutex); 29749c5a2ba7STejun Heo if (!wq->nr_drainers++) 2975618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 297687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29779c5a2ba7STejun Heo reflush: 2978c4f135d6STetsuo Handa __flush_workqueue(wq); 29799c5a2ba7STejun Heo 2980b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 298176af4d93STejun Heo 298249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2983fa2563e4SThomas Tuttle bool drained; 29849c5a2ba7STejun Heo 2985a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 2986f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 2987a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 2988fa2563e4SThomas Tuttle 2989fa2563e4SThomas Tuttle if (drained) 29909c5a2ba7STejun Heo continue; 29919c5a2ba7STejun Heo 29929c5a2ba7STejun Heo if (++flush_cnt == 10 || 29939c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2994e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 2995e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 299676af4d93STejun Heo 2997b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 29989c5a2ba7STejun Heo goto reflush; 29999c5a2ba7STejun Heo } 30009c5a2ba7STejun Heo 30019c5a2ba7STejun Heo if (!--wq->nr_drainers) 3002618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 300387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 30049c5a2ba7STejun Heo } 30059c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 30069c5a2ba7STejun Heo 3007d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3008d6e89786SJohannes Berg bool from_cancel) 3009baf59022STejun Heo { 3010baf59022STejun Heo struct worker *worker = NULL; 3011c9e7cf27STejun Heo struct worker_pool *pool; 3012112202d9STejun Heo struct pool_workqueue *pwq; 3013baf59022STejun Heo 3014baf59022STejun Heo might_sleep(); 3015baf59022STejun Heo 301624acfb71SThomas Gleixner rcu_read_lock(); 3017fa1b54e6STejun Heo pool = get_work_pool(work); 3018fa1b54e6STejun Heo if (!pool) { 301924acfb71SThomas Gleixner rcu_read_unlock(); 3020fa1b54e6STejun Heo return false; 3021fa1b54e6STejun Heo } 3022fa1b54e6STejun Heo 3023a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 30240b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3025112202d9STejun Heo pwq = get_work_pwq(work); 3026112202d9STejun Heo if (pwq) { 3027112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3028baf59022STejun Heo goto already_gone; 3029606a5020STejun Heo } else { 3030c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3031baf59022STejun Heo if (!worker) 3032baf59022STejun Heo goto already_gone; 3033112202d9STejun Heo pwq = worker->current_pwq; 3034606a5020STejun Heo } 3035baf59022STejun Heo 3036fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3037fca839c0STejun Heo 3038112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3039a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3040baf59022STejun Heo 3041e159489bSTejun Heo /* 3042a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3043a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3044a1d14934SPeter Zijlstra * 3045a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3046a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3047a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3048a1d14934SPeter Zijlstra * forward progress. 3049e159489bSTejun Heo */ 3050d6e89786SJohannes Berg if (!from_cancel && 3051d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3052112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3053112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3054a1d14934SPeter Zijlstra } 305524acfb71SThomas Gleixner rcu_read_unlock(); 3056baf59022STejun Heo return true; 3057baf59022STejun Heo already_gone: 3058a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 305924acfb71SThomas Gleixner rcu_read_unlock(); 3060baf59022STejun Heo return false; 3061baf59022STejun Heo } 3062baf59022STejun Heo 3063d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3064d6e89786SJohannes Berg { 3065d6e89786SJohannes Berg struct wq_barrier barr; 3066d6e89786SJohannes Berg 3067d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3068d6e89786SJohannes Berg return false; 3069d6e89786SJohannes Berg 30704d43d395STetsuo Handa if (WARN_ON(!work->func)) 30714d43d395STetsuo Handa return false; 30724d43d395STetsuo Handa 307387915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 307487915adcSJohannes Berg lock_map_release(&work->lockdep_map); 307587915adcSJohannes Berg 3076d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3077d6e89786SJohannes Berg wait_for_completion(&barr.done); 3078d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3079d6e89786SJohannes Berg return true; 3080d6e89786SJohannes Berg } else { 3081d6e89786SJohannes Berg return false; 3082d6e89786SJohannes Berg } 3083d6e89786SJohannes Berg } 3084d6e89786SJohannes Berg 3085db700897SOleg Nesterov /** 3086401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3087401a8d04STejun Heo * @work: the work to flush 3088db700897SOleg Nesterov * 3089606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3090606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3091401a8d04STejun Heo * 3092d185af30SYacine Belkadi * Return: 3093401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3094401a8d04STejun Heo * %false if it was already idle. 3095db700897SOleg Nesterov */ 3096401a8d04STejun Heo bool flush_work(struct work_struct *work) 3097db700897SOleg Nesterov { 3098d6e89786SJohannes Berg return __flush_work(work, false); 3099606a5020STejun Heo } 3100db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3101db700897SOleg Nesterov 31028603e1b3STejun Heo struct cwt_wait { 3103ac6424b9SIngo Molnar wait_queue_entry_t wait; 31048603e1b3STejun Heo struct work_struct *work; 31058603e1b3STejun Heo }; 31068603e1b3STejun Heo 3107ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 31088603e1b3STejun Heo { 31098603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 31108603e1b3STejun Heo 31118603e1b3STejun Heo if (cwait->work != key) 31128603e1b3STejun Heo return 0; 31138603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 31148603e1b3STejun Heo } 31158603e1b3STejun Heo 311636e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3117401a8d04STejun Heo { 31188603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3119bbb68dfaSTejun Heo unsigned long flags; 31201f1f642eSOleg Nesterov int ret; 31211f1f642eSOleg Nesterov 31221f1f642eSOleg Nesterov do { 3123bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3124bbb68dfaSTejun Heo /* 31258603e1b3STejun Heo * If someone else is already canceling, wait for it to 31268603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 31278603e1b3STejun Heo * because we may get scheduled between @work's completion 31288603e1b3STejun Heo * and the other canceling task resuming and clearing 31298603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 31308603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 31318603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 31328603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 31338603e1b3STejun Heo * we're hogging the CPU. 31348603e1b3STejun Heo * 31358603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 31368603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 31378603e1b3STejun Heo * wake function which matches @work along with exclusive 31388603e1b3STejun Heo * wait and wakeup. 3139bbb68dfaSTejun Heo */ 31408603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 31418603e1b3STejun Heo struct cwt_wait cwait; 31428603e1b3STejun Heo 31438603e1b3STejun Heo init_wait(&cwait.wait); 31448603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 31458603e1b3STejun Heo cwait.work = work; 31468603e1b3STejun Heo 31478603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 31488603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 31498603e1b3STejun Heo if (work_is_canceling(work)) 31508603e1b3STejun Heo schedule(); 31518603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 31528603e1b3STejun Heo } 31531f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 31541f1f642eSOleg Nesterov 3155bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3156bbb68dfaSTejun Heo mark_work_canceling(work); 3157bbb68dfaSTejun Heo local_irq_restore(flags); 3158bbb68dfaSTejun Heo 31593347fa09STejun Heo /* 31603347fa09STejun Heo * This allows canceling during early boot. We know that @work 31613347fa09STejun Heo * isn't executing. 31623347fa09STejun Heo */ 31633347fa09STejun Heo if (wq_online) 3164d6e89786SJohannes Berg __flush_work(work, true); 31653347fa09STejun Heo 31667a22ad75STejun Heo clear_work_data(work); 31678603e1b3STejun Heo 31688603e1b3STejun Heo /* 31698603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 31708603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 31718603e1b3STejun Heo * visible there. 31728603e1b3STejun Heo */ 31738603e1b3STejun Heo smp_mb(); 31748603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 31758603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 31768603e1b3STejun Heo 31771f1f642eSOleg Nesterov return ret; 31781f1f642eSOleg Nesterov } 31791f1f642eSOleg Nesterov 31806e84d644SOleg Nesterov /** 3181401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3182401a8d04STejun Heo * @work: the work to cancel 31836e84d644SOleg Nesterov * 3184401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3185401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3186401a8d04STejun Heo * another workqueue. On return from this function, @work is 3187401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 31881f1f642eSOleg Nesterov * 3189401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3190401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 31916e84d644SOleg Nesterov * 3192401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 31936e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3194401a8d04STejun Heo * 3195d185af30SYacine Belkadi * Return: 3196401a8d04STejun Heo * %true if @work was pending, %false otherwise. 31976e84d644SOleg Nesterov */ 3198401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 31996e84d644SOleg Nesterov { 320036e227d2STejun Heo return __cancel_work_timer(work, false); 3201b89deed3SOleg Nesterov } 320228e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3203b89deed3SOleg Nesterov 32046e84d644SOleg Nesterov /** 3205401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3206401a8d04STejun Heo * @dwork: the delayed work to flush 32076e84d644SOleg Nesterov * 3208401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3209401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3210401a8d04STejun Heo * considers the last queueing instance of @dwork. 32111f1f642eSOleg Nesterov * 3212d185af30SYacine Belkadi * Return: 3213401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3214401a8d04STejun Heo * %false if it was already idle. 32156e84d644SOleg Nesterov */ 3216401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3217401a8d04STejun Heo { 32188930cabaSTejun Heo local_irq_disable(); 3219401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 322060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 32218930cabaSTejun Heo local_irq_enable(); 3222401a8d04STejun Heo return flush_work(&dwork->work); 3223401a8d04STejun Heo } 3224401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3225401a8d04STejun Heo 322605f0fe6bSTejun Heo /** 322705f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 322805f0fe6bSTejun Heo * @rwork: the rcu work to flush 322905f0fe6bSTejun Heo * 323005f0fe6bSTejun Heo * Return: 323105f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 323205f0fe6bSTejun Heo * %false if it was already idle. 323305f0fe6bSTejun Heo */ 323405f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 323505f0fe6bSTejun Heo { 323605f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 323705f0fe6bSTejun Heo rcu_barrier(); 323805f0fe6bSTejun Heo flush_work(&rwork->work); 323905f0fe6bSTejun Heo return true; 324005f0fe6bSTejun Heo } else { 324105f0fe6bSTejun Heo return flush_work(&rwork->work); 324205f0fe6bSTejun Heo } 324305f0fe6bSTejun Heo } 324405f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 324505f0fe6bSTejun Heo 3246f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3247f72b8792SJens Axboe { 3248f72b8792SJens Axboe unsigned long flags; 3249f72b8792SJens Axboe int ret; 3250f72b8792SJens Axboe 3251f72b8792SJens Axboe do { 3252f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3253f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3254f72b8792SJens Axboe 3255f72b8792SJens Axboe if (unlikely(ret < 0)) 3256f72b8792SJens Axboe return false; 3257f72b8792SJens Axboe 3258f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3259f72b8792SJens Axboe local_irq_restore(flags); 3260f72b8792SJens Axboe return ret; 3261f72b8792SJens Axboe } 3262f72b8792SJens Axboe 326373b4b532SAndrey Grodzovsky /* 326473b4b532SAndrey Grodzovsky * See cancel_delayed_work() 326573b4b532SAndrey Grodzovsky */ 326673b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 326773b4b532SAndrey Grodzovsky { 326873b4b532SAndrey Grodzovsky return __cancel_work(work, false); 326973b4b532SAndrey Grodzovsky } 327073b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 327173b4b532SAndrey Grodzovsky 3272401a8d04STejun Heo /** 327357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 327457b30ae7STejun Heo * @dwork: delayed_work to cancel 327509383498STejun Heo * 3276d185af30SYacine Belkadi * Kill off a pending delayed_work. 3277d185af30SYacine Belkadi * 3278d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3279d185af30SYacine Belkadi * pending. 3280d185af30SYacine Belkadi * 3281d185af30SYacine Belkadi * Note: 3282d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3283d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3284d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 328509383498STejun Heo * 328657b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 328709383498STejun Heo */ 328857b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 328909383498STejun Heo { 3290f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 329109383498STejun Heo } 329257b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 329309383498STejun Heo 329409383498STejun Heo /** 3295401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3296401a8d04STejun Heo * @dwork: the delayed work cancel 3297401a8d04STejun Heo * 3298401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3299401a8d04STejun Heo * 3300d185af30SYacine Belkadi * Return: 3301401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3302401a8d04STejun Heo */ 3303401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 33046e84d644SOleg Nesterov { 330536e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 33066e84d644SOleg Nesterov } 3307f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 33081da177e4SLinus Torvalds 33090fcb78c2SRolf Eike Beer /** 331031ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3311b6136773SAndrew Morton * @func: the function to call 3312b6136773SAndrew Morton * 331331ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 331431ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3315b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 331631ddd871STejun Heo * 3317d185af30SYacine Belkadi * Return: 331831ddd871STejun Heo * 0 on success, -errno on failure. 3319b6136773SAndrew Morton */ 332065f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 332115316ba8SChristoph Lameter { 332215316ba8SChristoph Lameter int cpu; 332338f51568SNamhyung Kim struct work_struct __percpu *works; 332415316ba8SChristoph Lameter 3325b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3326b6136773SAndrew Morton if (!works) 332715316ba8SChristoph Lameter return -ENOMEM; 3328b6136773SAndrew Morton 3329ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 333093981800STejun Heo 333115316ba8SChristoph Lameter for_each_online_cpu(cpu) { 33329bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 33339bfb1839SIngo Molnar 33349bfb1839SIngo Molnar INIT_WORK(work, func); 33358de6d308SOleg Nesterov schedule_work_on(cpu, work); 333615316ba8SChristoph Lameter } 333793981800STejun Heo 333893981800STejun Heo for_each_online_cpu(cpu) 33398616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 334093981800STejun Heo 3341ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3342b6136773SAndrew Morton free_percpu(works); 334315316ba8SChristoph Lameter return 0; 334415316ba8SChristoph Lameter } 334515316ba8SChristoph Lameter 3346eef6a7d5SAlan Stern /** 33471fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 33481fa44ecaSJames Bottomley * @fn: the function to execute 33491fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 33501fa44ecaSJames Bottomley * be available when the work executes) 33511fa44ecaSJames Bottomley * 33521fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 33531fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 33541fa44ecaSJames Bottomley * 3355d185af30SYacine Belkadi * Return: 0 - function was executed 33561fa44ecaSJames Bottomley * 1 - function was scheduled for execution 33571fa44ecaSJames Bottomley */ 335865f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 33591fa44ecaSJames Bottomley { 33601fa44ecaSJames Bottomley if (!in_interrupt()) { 336165f27f38SDavid Howells fn(&ew->work); 33621fa44ecaSJames Bottomley return 0; 33631fa44ecaSJames Bottomley } 33641fa44ecaSJames Bottomley 336565f27f38SDavid Howells INIT_WORK(&ew->work, fn); 33661fa44ecaSJames Bottomley schedule_work(&ew->work); 33671fa44ecaSJames Bottomley 33681fa44ecaSJames Bottomley return 1; 33691fa44ecaSJames Bottomley } 33701fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 33711fa44ecaSJames Bottomley 33727a4e344cSTejun Heo /** 33737a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33747a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33757a4e344cSTejun Heo * 33767a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33777a4e344cSTejun Heo */ 3378513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 33797a4e344cSTejun Heo { 33807a4e344cSTejun Heo if (attrs) { 33817a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33827a4e344cSTejun Heo kfree(attrs); 33837a4e344cSTejun Heo } 33847a4e344cSTejun Heo } 33857a4e344cSTejun Heo 33867a4e344cSTejun Heo /** 33877a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33887a4e344cSTejun Heo * 33897a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3390d185af30SYacine Belkadi * return it. 3391d185af30SYacine Belkadi * 3392d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33937a4e344cSTejun Heo */ 3394513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 33957a4e344cSTejun Heo { 33967a4e344cSTejun Heo struct workqueue_attrs *attrs; 33977a4e344cSTejun Heo 3398be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 33997a4e344cSTejun Heo if (!attrs) 34007a4e344cSTejun Heo goto fail; 3401be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 34027a4e344cSTejun Heo goto fail; 34037a4e344cSTejun Heo 340413e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 34057a4e344cSTejun Heo return attrs; 34067a4e344cSTejun Heo fail: 34077a4e344cSTejun Heo free_workqueue_attrs(attrs); 34087a4e344cSTejun Heo return NULL; 34097a4e344cSTejun Heo } 34107a4e344cSTejun Heo 341129c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 341229c91e99STejun Heo const struct workqueue_attrs *from) 341329c91e99STejun Heo { 341429c91e99STejun Heo to->nice = from->nice; 341529c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 34162865a8fbSShaohua Li /* 34172865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 34182865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 34192865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 34202865a8fbSShaohua Li */ 34212865a8fbSShaohua Li to->no_numa = from->no_numa; 342229c91e99STejun Heo } 342329c91e99STejun Heo 342429c91e99STejun Heo /* hash value of the content of @attr */ 342529c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 342629c91e99STejun Heo { 342729c91e99STejun Heo u32 hash = 0; 342829c91e99STejun Heo 342929c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 343013e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 343113e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 343229c91e99STejun Heo return hash; 343329c91e99STejun Heo } 343429c91e99STejun Heo 343529c91e99STejun Heo /* content equality test */ 343629c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 343729c91e99STejun Heo const struct workqueue_attrs *b) 343829c91e99STejun Heo { 343929c91e99STejun Heo if (a->nice != b->nice) 344029c91e99STejun Heo return false; 344129c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 344229c91e99STejun Heo return false; 344329c91e99STejun Heo return true; 344429c91e99STejun Heo } 344529c91e99STejun Heo 34467a4e344cSTejun Heo /** 34477a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34487a4e344cSTejun Heo * @pool: worker_pool to initialize 34497a4e344cSTejun Heo * 3450402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3451d185af30SYacine Belkadi * 3452d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 345329c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 345429c91e99STejun Heo * on @pool safely to release it. 34557a4e344cSTejun Heo */ 34567a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34574e1a1f9aSTejun Heo { 3458a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 345929c91e99STejun Heo pool->id = -1; 346029c91e99STejun Heo pool->cpu = -1; 3461f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34624e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 346382607adcSTejun Heo pool->watchdog_ts = jiffies; 34644e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34654e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34664e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34674e1a1f9aSTejun Heo 346832a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 34694e1a1f9aSTejun Heo 347032a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 34714e1a1f9aSTejun Heo 3472da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34737a4e344cSTejun Heo 34747cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 347529c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 347629c91e99STejun Heo pool->refcnt = 1; 347729c91e99STejun Heo 347829c91e99STejun Heo /* shouldn't fail above this point */ 3479be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 34807a4e344cSTejun Heo if (!pool->attrs) 34817a4e344cSTejun Heo return -ENOMEM; 34827a4e344cSTejun Heo return 0; 34834e1a1f9aSTejun Heo } 34844e1a1f9aSTejun Heo 3485669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3486669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3487669de8bdSBart Van Assche { 3488669de8bdSBart Van Assche char *lock_name; 3489669de8bdSBart Van Assche 3490669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3491669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3492669de8bdSBart Van Assche if (!lock_name) 3493669de8bdSBart Van Assche lock_name = wq->name; 349469a106c0SQian Cai 349569a106c0SQian Cai wq->lock_name = lock_name; 3496669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3497669de8bdSBart Van Assche } 3498669de8bdSBart Van Assche 3499669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3500669de8bdSBart Van Assche { 3501669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3502669de8bdSBart Van Assche } 3503669de8bdSBart Van Assche 3504669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3505669de8bdSBart Van Assche { 3506669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3507669de8bdSBart Van Assche kfree(wq->lock_name); 3508669de8bdSBart Van Assche } 3509669de8bdSBart Van Assche #else 3510669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3511669de8bdSBart Van Assche { 3512669de8bdSBart Van Assche } 3513669de8bdSBart Van Assche 3514669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3515669de8bdSBart Van Assche { 3516669de8bdSBart Van Assche } 3517669de8bdSBart Van Assche 3518669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3519669de8bdSBart Van Assche { 3520669de8bdSBart Van Assche } 3521669de8bdSBart Van Assche #endif 3522669de8bdSBart Van Assche 3523e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3524e2dca7adSTejun Heo { 3525e2dca7adSTejun Heo struct workqueue_struct *wq = 3526e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3527e2dca7adSTejun Heo 3528669de8bdSBart Van Assche wq_free_lockdep(wq); 3529669de8bdSBart Van Assche 3530e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3531e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3532e2dca7adSTejun Heo else 3533e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3534e2dca7adSTejun Heo 3535e2dca7adSTejun Heo kfree(wq); 3536e2dca7adSTejun Heo } 3537e2dca7adSTejun Heo 353829c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 353929c91e99STejun Heo { 354029c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 354129c91e99STejun Heo 35427cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 354329c91e99STejun Heo free_workqueue_attrs(pool->attrs); 354429c91e99STejun Heo kfree(pool); 354529c91e99STejun Heo } 354629c91e99STejun Heo 3547d8bb65abSSebastian Andrzej Siewior /* This returns with the lock held on success (pool manager is inactive). */ 3548d8bb65abSSebastian Andrzej Siewior static bool wq_manager_inactive(struct worker_pool *pool) 3549d8bb65abSSebastian Andrzej Siewior { 3550a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 3551d8bb65abSSebastian Andrzej Siewior 3552d8bb65abSSebastian Andrzej Siewior if (pool->flags & POOL_MANAGER_ACTIVE) { 3553a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3554d8bb65abSSebastian Andrzej Siewior return false; 3555d8bb65abSSebastian Andrzej Siewior } 3556d8bb65abSSebastian Andrzej Siewior return true; 3557d8bb65abSSebastian Andrzej Siewior } 3558d8bb65abSSebastian Andrzej Siewior 355929c91e99STejun Heo /** 356029c91e99STejun Heo * put_unbound_pool - put a worker_pool 356129c91e99STejun Heo * @pool: worker_pool to put 356229c91e99STejun Heo * 356324acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3564c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3565c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3566c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3567a892caccSTejun Heo * 3568a892caccSTejun Heo * Should be called with wq_pool_mutex held. 356929c91e99STejun Heo */ 357029c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 357129c91e99STejun Heo { 357260f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 357329c91e99STejun Heo struct worker *worker; 357429c91e99STejun Heo 3575a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3576a892caccSTejun Heo 3577a892caccSTejun Heo if (--pool->refcnt) 357829c91e99STejun Heo return; 357929c91e99STejun Heo 358029c91e99STejun Heo /* sanity checks */ 358161d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3582a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 358329c91e99STejun Heo return; 358429c91e99STejun Heo 358529c91e99STejun Heo /* release id and unhash */ 358629c91e99STejun Heo if (pool->id >= 0) 358729c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 358829c91e99STejun Heo hash_del(&pool->hash_node); 358929c91e99STejun Heo 3590c5aa87bbSTejun Heo /* 3591692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3592692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3593692b4825STejun Heo * manager and @pool gets freed with the flag set. 3594d8bb65abSSebastian Andrzej Siewior * Because of how wq_manager_inactive() works, we will hold the 3595d8bb65abSSebastian Andrzej Siewior * spinlock after a successful wait. 3596c5aa87bbSTejun Heo */ 3597d8bb65abSSebastian Andrzej Siewior rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool), 3598d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3599692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3600692b4825STejun Heo 36011037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 360229c91e99STejun Heo destroy_worker(worker); 360329c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3604a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 360560f5a4bcSLai Jiangshan 36061258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3607da028469SLai Jiangshan if (!list_empty(&pool->workers)) 360860f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 36091258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 361060f5a4bcSLai Jiangshan 361160f5a4bcSLai Jiangshan if (pool->detach_completion) 361260f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 361360f5a4bcSLai Jiangshan 361429c91e99STejun Heo /* shut down the timers */ 361529c91e99STejun Heo del_timer_sync(&pool->idle_timer); 361629c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 361729c91e99STejun Heo 361824acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 361925b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 362029c91e99STejun Heo } 362129c91e99STejun Heo 362229c91e99STejun Heo /** 362329c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 362429c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 362529c91e99STejun Heo * 362629c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 362729c91e99STejun Heo * reference count and return it. If there already is a matching 362829c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3629d185af30SYacine Belkadi * create a new one. 3630a892caccSTejun Heo * 3631a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3632d185af30SYacine Belkadi * 3633d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3634d185af30SYacine Belkadi * On failure, %NULL. 363529c91e99STejun Heo */ 363629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 363729c91e99STejun Heo { 363829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 363929c91e99STejun Heo struct worker_pool *pool; 3640f3f90ad4STejun Heo int node; 3641e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 364229c91e99STejun Heo 3643a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 364429c91e99STejun Heo 364529c91e99STejun Heo /* do we already have a matching pool? */ 364629c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 364729c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 364829c91e99STejun Heo pool->refcnt++; 36493fb1823cSLai Jiangshan return pool; 365029c91e99STejun Heo } 365129c91e99STejun Heo } 365229c91e99STejun Heo 3653e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3654e2273584SXunlei Pang if (wq_numa_enabled) { 3655e2273584SXunlei Pang for_each_node(node) { 3656e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3657e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3658e2273584SXunlei Pang target_node = node; 3659e2273584SXunlei Pang break; 3660e2273584SXunlei Pang } 3661e2273584SXunlei Pang } 3662e2273584SXunlei Pang } 3663e2273584SXunlei Pang 366429c91e99STejun Heo /* nope, create a new one */ 3665e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 366629c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 366729c91e99STejun Heo goto fail; 366829c91e99STejun Heo 36698864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 367029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3671e2273584SXunlei Pang pool->node = target_node; 367229c91e99STejun Heo 36732865a8fbSShaohua Li /* 36742865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36752865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36762865a8fbSShaohua Li */ 36772865a8fbSShaohua Li pool->attrs->no_numa = false; 36782865a8fbSShaohua Li 367929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 368029c91e99STejun Heo goto fail; 368129c91e99STejun Heo 368229c91e99STejun Heo /* create and start the initial worker */ 36833347fa09STejun Heo if (wq_online && !create_worker(pool)) 368429c91e99STejun Heo goto fail; 368529c91e99STejun Heo 368629c91e99STejun Heo /* install */ 368729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 36883fb1823cSLai Jiangshan 368929c91e99STejun Heo return pool; 369029c91e99STejun Heo fail: 369129c91e99STejun Heo if (pool) 369229c91e99STejun Heo put_unbound_pool(pool); 369329c91e99STejun Heo return NULL; 369429c91e99STejun Heo } 369529c91e99STejun Heo 36968864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36978864b4e5STejun Heo { 36988864b4e5STejun Heo kmem_cache_free(pwq_cache, 36998864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 37008864b4e5STejun Heo } 37018864b4e5STejun Heo 37028864b4e5STejun Heo /* 37038864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 37048864b4e5STejun Heo * and needs to be destroyed. 37058864b4e5STejun Heo */ 37068864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 37078864b4e5STejun Heo { 37088864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 37098864b4e5STejun Heo unbound_release_work); 37108864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 37118864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3712b42b0bddSYang Yingliang bool is_last = false; 37138864b4e5STejun Heo 3714b42b0bddSYang Yingliang /* 3715b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 3716b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3717b42b0bddSYang Yingliang */ 3718b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 37198864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 37208864b4e5STejun Heo return; 37218864b4e5STejun Heo 37223c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 37238864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3724bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 37253c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3726b42b0bddSYang Yingliang } 37278864b4e5STejun Heo 3728a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 37298864b4e5STejun Heo put_unbound_pool(pool); 3730a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3731a892caccSTejun Heo 373225b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 37338864b4e5STejun Heo 37348864b4e5STejun Heo /* 37358864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3736e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 37378864b4e5STejun Heo */ 3738669de8bdSBart Van Assche if (is_last) { 3739669de8bdSBart Van Assche wq_unregister_lockdep(wq); 374025b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 37416029a918STejun Heo } 3742669de8bdSBart Van Assche } 37438864b4e5STejun Heo 37440fbd95aaSTejun Heo /** 3745699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37460fbd95aaSTejun Heo * @pwq: target pool_workqueue 37470fbd95aaSTejun Heo * 3748699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3749f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 3750699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37510fbd95aaSTejun Heo */ 3752699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37530fbd95aaSTejun Heo { 3754699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3755699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 37563347fa09STejun Heo unsigned long flags; 3757699ce097STejun Heo 3758699ce097STejun Heo /* for @wq->saved_max_active */ 3759a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3760699ce097STejun Heo 3761699ce097STejun Heo /* fast exit for non-freezable wqs */ 3762699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3763699ce097STejun Heo return; 3764699ce097STejun Heo 37653347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3766a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3767699ce097STejun Heo 376874b414eaSLai Jiangshan /* 376974b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 377074b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 377174b414eaSLai Jiangshan * is updated and visible. 377274b414eaSLai Jiangshan */ 377374b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 377401341fbdSYunfeng Ye bool kick = false; 377501341fbdSYunfeng Ye 3776699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37770fbd95aaSTejun Heo 3778f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 377901341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 3780f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 378101341fbdSYunfeng Ye kick = true; 378201341fbdSYunfeng Ye } 3783951a078aSLai Jiangshan 3784951a078aSLai Jiangshan /* 3785951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 378601341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 378701341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 378801341fbdSYunfeng Ye * let's kick iff work items were activated. 3789951a078aSLai Jiangshan */ 379001341fbdSYunfeng Ye if (kick) 3791951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3792699ce097STejun Heo } else { 3793699ce097STejun Heo pwq->max_active = 0; 3794699ce097STejun Heo } 3795699ce097STejun Heo 3796a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 37970fbd95aaSTejun Heo } 37980fbd95aaSTejun Heo 379967dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 3800f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3801f147f29eSTejun Heo struct worker_pool *pool) 3802d2c1d404STejun Heo { 3803d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3804d2c1d404STejun Heo 3805e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3806e50aba9aSTejun Heo 3807d2c1d404STejun Heo pwq->pool = pool; 3808d2c1d404STejun Heo pwq->wq = wq; 3809d2c1d404STejun Heo pwq->flush_color = -1; 38108864b4e5STejun Heo pwq->refcnt = 1; 3811f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 38121befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3813d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 38148864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3815f147f29eSTejun Heo } 3816d2c1d404STejun Heo 3817f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 38181befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3819f147f29eSTejun Heo { 3820f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3821f147f29eSTejun Heo 3822f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 382375ccf595STejun Heo 38241befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 38251befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 38261befcf30STejun Heo return; 38271befcf30STejun Heo 382829b1cb41SLai Jiangshan /* set the matching work_color */ 382975ccf595STejun Heo pwq->work_color = wq->work_color; 3830983ca25eSTejun Heo 3831983ca25eSTejun Heo /* sync max_active to the current setting */ 3832983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3833983ca25eSTejun Heo 3834983ca25eSTejun Heo /* link in @pwq */ 38359e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3836df2d5ae4STejun Heo } 38376029a918STejun Heo 3838f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3839f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3840f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3841f147f29eSTejun Heo { 3842f147f29eSTejun Heo struct worker_pool *pool; 3843f147f29eSTejun Heo struct pool_workqueue *pwq; 3844f147f29eSTejun Heo 3845f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3846f147f29eSTejun Heo 3847f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3848f147f29eSTejun Heo if (!pool) 3849f147f29eSTejun Heo return NULL; 3850f147f29eSTejun Heo 3851e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3852f147f29eSTejun Heo if (!pwq) { 3853f147f29eSTejun Heo put_unbound_pool(pool); 3854f147f29eSTejun Heo return NULL; 3855f147f29eSTejun Heo } 3856f147f29eSTejun Heo 3857f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3858f147f29eSTejun Heo return pwq; 3859d2c1d404STejun Heo } 3860d2c1d404STejun Heo 38614c16bd32STejun Heo /** 386230186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3863042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 38644c16bd32STejun Heo * @node: the target NUMA node 38654c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38664c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38674c16bd32STejun Heo * 38684c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38694c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3870d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38714c16bd32STejun Heo * 38724c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38734c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38744c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38754c16bd32STejun Heo * @attrs->cpumask. 38764c16bd32STejun Heo * 38774c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38784c16bd32STejun Heo * stable. 3879d185af30SYacine Belkadi * 3880d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3881d185af30SYacine Belkadi * %false if equal. 38824c16bd32STejun Heo */ 38834c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38844c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38854c16bd32STejun Heo { 3886d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38874c16bd32STejun Heo goto use_dfl; 38884c16bd32STejun Heo 38894c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38904c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38914c16bd32STejun Heo if (cpu_going_down >= 0) 38924c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38934c16bd32STejun Heo 38944c16bd32STejun Heo if (cpumask_empty(cpumask)) 38954c16bd32STejun Heo goto use_dfl; 38964c16bd32STejun Heo 38974c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38984c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38991ad0f0a7SMichael Bringmann 39001ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 39011ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 39021ad0f0a7SMichael Bringmann "possible intersect\n"); 39031ad0f0a7SMichael Bringmann return false; 39041ad0f0a7SMichael Bringmann } 39051ad0f0a7SMichael Bringmann 39064c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 39074c16bd32STejun Heo 39084c16bd32STejun Heo use_dfl: 39094c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 39104c16bd32STejun Heo return false; 39114c16bd32STejun Heo } 39124c16bd32STejun Heo 39131befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 39141befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 39151befcf30STejun Heo int node, 39161befcf30STejun Heo struct pool_workqueue *pwq) 39171befcf30STejun Heo { 39181befcf30STejun Heo struct pool_workqueue *old_pwq; 39191befcf30STejun Heo 39205b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39211befcf30STejun Heo lockdep_assert_held(&wq->mutex); 39221befcf30STejun Heo 39231befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 39241befcf30STejun Heo link_pwq(pwq); 39251befcf30STejun Heo 39261befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 39271befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 39281befcf30STejun Heo return old_pwq; 39291befcf30STejun Heo } 39301befcf30STejun Heo 39312d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 39322d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 39332d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 39342d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3935042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 39362d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 39372d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 39382d5f0764SLai Jiangshan }; 39392d5f0764SLai Jiangshan 39402d5f0764SLai Jiangshan /* free the resources after success or abort */ 39412d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 39422d5f0764SLai Jiangshan { 39432d5f0764SLai Jiangshan if (ctx) { 39442d5f0764SLai Jiangshan int node; 39452d5f0764SLai Jiangshan 39462d5f0764SLai Jiangshan for_each_node(node) 39472d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 39482d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 39492d5f0764SLai Jiangshan 39502d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 39512d5f0764SLai Jiangshan 39522d5f0764SLai Jiangshan kfree(ctx); 39532d5f0764SLai Jiangshan } 39542d5f0764SLai Jiangshan } 39552d5f0764SLai Jiangshan 39562d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 39572d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 39582d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 39592d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 39602d5f0764SLai Jiangshan { 39612d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 39622d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 39632d5f0764SLai Jiangshan int node; 39642d5f0764SLai Jiangshan 39652d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39662d5f0764SLai Jiangshan 3967acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 39682d5f0764SLai Jiangshan 3969be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 3970be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 39712d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 39722d5f0764SLai Jiangshan goto out_free; 39732d5f0764SLai Jiangshan 3974042f7df1SLai Jiangshan /* 3975042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3976042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3977042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3978042f7df1SLai Jiangshan */ 39792d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3980b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3981042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3982042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 39832d5f0764SLai Jiangshan 39842d5f0764SLai Jiangshan /* 39852d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 39862d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 39872d5f0764SLai Jiangshan * pools. 39882d5f0764SLai Jiangshan */ 39892d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 39902d5f0764SLai Jiangshan 39912d5f0764SLai Jiangshan /* 39922d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 39932d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 39942d5f0764SLai Jiangshan * it even if we don't use it immediately. 39952d5f0764SLai Jiangshan */ 39962d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39972d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 39982d5f0764SLai Jiangshan goto out_free; 39992d5f0764SLai Jiangshan 40002d5f0764SLai Jiangshan for_each_node(node) { 4001042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 40022d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 40032d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 40042d5f0764SLai Jiangshan goto out_free; 40052d5f0764SLai Jiangshan } else { 40062d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 40072d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 40082d5f0764SLai Jiangshan } 40092d5f0764SLai Jiangshan } 40102d5f0764SLai Jiangshan 4011042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4012042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4013042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 40142d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4015042f7df1SLai Jiangshan 40162d5f0764SLai Jiangshan ctx->wq = wq; 40172d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40182d5f0764SLai Jiangshan return ctx; 40192d5f0764SLai Jiangshan 40202d5f0764SLai Jiangshan out_free: 40212d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40222d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 40232d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 40242d5f0764SLai Jiangshan return NULL; 40252d5f0764SLai Jiangshan } 40262d5f0764SLai Jiangshan 40272d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 40282d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 40292d5f0764SLai Jiangshan { 40302d5f0764SLai Jiangshan int node; 40312d5f0764SLai Jiangshan 40322d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 40332d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 40342d5f0764SLai Jiangshan 40352d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 40362d5f0764SLai Jiangshan 40372d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 40382d5f0764SLai Jiangshan for_each_node(node) 40392d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 40402d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 40412d5f0764SLai Jiangshan 40422d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 40432d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 40442d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 40452d5f0764SLai Jiangshan 40462d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 40472d5f0764SLai Jiangshan } 40482d5f0764SLai Jiangshan 4049a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4050a0111cf6SLai Jiangshan { 4051a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4052ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4053a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4054a0111cf6SLai Jiangshan } 4055a0111cf6SLai Jiangshan 4056a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4057a0111cf6SLai Jiangshan { 4058a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4059ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4060a0111cf6SLai Jiangshan } 4061a0111cf6SLai Jiangshan 4062a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4063a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4064a0111cf6SLai Jiangshan { 4065a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4066a0111cf6SLai Jiangshan 4067a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4068a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4069a0111cf6SLai Jiangshan return -EINVAL; 4070a0111cf6SLai Jiangshan 4071a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 40720a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 40730a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4074a0111cf6SLai Jiangshan return -EINVAL; 4075a0111cf6SLai Jiangshan 40760a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 40770a94efb5STejun Heo } 40780a94efb5STejun Heo 4079a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 40806201171eSwanghaibin if (!ctx) 40816201171eSwanghaibin return -ENOMEM; 4082a0111cf6SLai Jiangshan 4083a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4084a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4085a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4086a0111cf6SLai Jiangshan 40876201171eSwanghaibin return 0; 4088a0111cf6SLai Jiangshan } 4089a0111cf6SLai Jiangshan 40909e8cd2f5STejun Heo /** 40919e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 40929e8cd2f5STejun Heo * @wq: the target workqueue 40939e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 40949e8cd2f5STejun Heo * 40954c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 40964c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 40974c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 40984c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 40994c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 41004c16bd32STejun Heo * back-to-back will stay on its current pwq. 41019e8cd2f5STejun Heo * 4102d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4103d185af30SYacine Belkadi * 4104ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4105509b3204SDaniel Jordan * 4106d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 41079e8cd2f5STejun Heo */ 4108513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 41099e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 41109e8cd2f5STejun Heo { 4111a0111cf6SLai Jiangshan int ret; 41129e8cd2f5STejun Heo 4113509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4114509b3204SDaniel Jordan 4115509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4116a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4117509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 41182d5f0764SLai Jiangshan 41192d5f0764SLai Jiangshan return ret; 41209e8cd2f5STejun Heo } 41219e8cd2f5STejun Heo 41224c16bd32STejun Heo /** 41234c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 41244c16bd32STejun Heo * @wq: the target workqueue 41254c16bd32STejun Heo * @cpu: the CPU coming up or going down 41264c16bd32STejun Heo * @online: whether @cpu is coming up or going down 41274c16bd32STejun Heo * 41284c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 41294c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 41304c16bd32STejun Heo * @wq accordingly. 41314c16bd32STejun Heo * 41324c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 41334c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 41344c16bd32STejun Heo * correct. 41354c16bd32STejun Heo * 41364c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 41374c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 41384c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 41394c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 41404c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 41414c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 41424c16bd32STejun Heo * CPU_DOWN_PREPARE. 41434c16bd32STejun Heo */ 41444c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 41454c16bd32STejun Heo bool online) 41464c16bd32STejun Heo { 41474c16bd32STejun Heo int node = cpu_to_node(cpu); 41484c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 41494c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 41504c16bd32STejun Heo struct workqueue_attrs *target_attrs; 41514c16bd32STejun Heo cpumask_t *cpumask; 41524c16bd32STejun Heo 41534c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 41544c16bd32STejun Heo 4155f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4156f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 41574c16bd32STejun Heo return; 41584c16bd32STejun Heo 41594c16bd32STejun Heo /* 41604c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 41614c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 41624c16bd32STejun Heo * CPU hotplug exclusion. 41634c16bd32STejun Heo */ 41644c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 41654c16bd32STejun Heo cpumask = target_attrs->cpumask; 41664c16bd32STejun Heo 41674c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 41684c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 41694c16bd32STejun Heo 41704c16bd32STejun Heo /* 41714c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4172042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4173042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4174042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 41754c16bd32STejun Heo */ 4176042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 41774c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4178f7142ed4SLai Jiangshan return; 41794c16bd32STejun Heo } else { 41804c16bd32STejun Heo goto use_dfl_pwq; 41814c16bd32STejun Heo } 41824c16bd32STejun Heo 41834c16bd32STejun Heo /* create a new pwq */ 41844c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41854c16bd32STejun Heo if (!pwq) { 41862d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41874c16bd32STejun Heo wq->name); 418877f300b1SDaeseok Youn goto use_dfl_pwq; 41894c16bd32STejun Heo } 41904c16bd32STejun Heo 4191f7142ed4SLai Jiangshan /* Install the new pwq. */ 41924c16bd32STejun Heo mutex_lock(&wq->mutex); 41934c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41944c16bd32STejun Heo goto out_unlock; 41954c16bd32STejun Heo 41964c16bd32STejun Heo use_dfl_pwq: 4197f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4198a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 41994c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4200a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 42014c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 42024c16bd32STejun Heo out_unlock: 42034c16bd32STejun Heo mutex_unlock(&wq->mutex); 42044c16bd32STejun Heo put_pwq_unlocked(old_pwq); 42054c16bd32STejun Heo } 42064c16bd32STejun Heo 420730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 42081da177e4SLinus Torvalds { 420949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 42108a2b7538STejun Heo int cpu, ret; 4211e1d8aa9fSFrederic Weisbecker 421230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4213420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4214420c0ddbSTejun Heo if (!wq->cpu_pwqs) 421530cdf249STejun Heo return -ENOMEM; 421630cdf249STejun Heo 421730cdf249STejun Heo for_each_possible_cpu(cpu) { 42187fb98ea7STejun Heo struct pool_workqueue *pwq = 42197fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 42207a62c2c8STejun Heo struct worker_pool *cpu_pools = 4221f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 422230cdf249STejun Heo 4223f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4224f147f29eSTejun Heo 4225f147f29eSTejun Heo mutex_lock(&wq->mutex); 42261befcf30STejun Heo link_pwq(pwq); 4227f147f29eSTejun Heo mutex_unlock(&wq->mutex); 422830cdf249STejun Heo } 422930cdf249STejun Heo return 0; 4230509b3204SDaniel Jordan } 4231509b3204SDaniel Jordan 4232ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4233509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 42348a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 42358a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 42368a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 42378a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 42388a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 42399e8cd2f5STejun Heo } else { 4240509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 42419e8cd2f5STejun Heo } 4242ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4243509b3204SDaniel Jordan 4244509b3204SDaniel Jordan return ret; 42450f900049STejun Heo } 42460f900049STejun Heo 4247f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4248f3421797STejun Heo const char *name) 4249b71ab8c2STejun Heo { 4250f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4251f3421797STejun Heo 4252f3421797STejun Heo if (max_active < 1 || max_active > lim) 4253044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4254f3421797STejun Heo max_active, name, 1, lim); 4255b71ab8c2STejun Heo 4256f3421797STejun Heo return clamp_val(max_active, 1, lim); 4257b71ab8c2STejun Heo } 4258b71ab8c2STejun Heo 4259983c7515STejun Heo /* 4260983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4261983c7515STejun Heo * to guarantee forward progress. 4262983c7515STejun Heo */ 4263983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4264983c7515STejun Heo { 4265983c7515STejun Heo struct worker *rescuer; 4266b92b36eaSDan Carpenter int ret; 4267983c7515STejun Heo 4268983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4269983c7515STejun Heo return 0; 4270983c7515STejun Heo 4271983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4272983c7515STejun Heo if (!rescuer) 4273983c7515STejun Heo return -ENOMEM; 4274983c7515STejun Heo 4275983c7515STejun Heo rescuer->rescue_wq = wq; 4276983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4277f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4278b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 4279983c7515STejun Heo kfree(rescuer); 4280b92b36eaSDan Carpenter return ret; 4281983c7515STejun Heo } 4282983c7515STejun Heo 4283983c7515STejun Heo wq->rescuer = rescuer; 4284983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4285983c7515STejun Heo wake_up_process(rescuer->task); 4286983c7515STejun Heo 4287983c7515STejun Heo return 0; 4288983c7515STejun Heo } 4289983c7515STejun Heo 4290a2775bbcSMathieu Malaterre __printf(1, 4) 4291669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 429297e37d7bSTejun Heo unsigned int flags, 4293669de8bdSBart Van Assche int max_active, ...) 42943af24433SOleg Nesterov { 4295df2d5ae4STejun Heo size_t tbl_size = 0; 4296ecf6881fSTejun Heo va_list args; 42973af24433SOleg Nesterov struct workqueue_struct *wq; 429849e3cf44STejun Heo struct pool_workqueue *pwq; 4299b196be89STejun Heo 43005c0338c6STejun Heo /* 43015c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 43025c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 43035c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 43045c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 43055c0338c6STejun Heo * on NUMA. 43065c0338c6STejun Heo */ 43075c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 43085c0338c6STejun Heo flags |= __WQ_ORDERED; 43095c0338c6STejun Heo 4310cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4311cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4312cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4313cee22a15SViresh Kumar 4314ecf6881fSTejun Heo /* allocate wq and format name */ 4315df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4316ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4317df2d5ae4STejun Heo 4318df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4319b196be89STejun Heo if (!wq) 4320d2c1d404STejun Heo return NULL; 4321b196be89STejun Heo 43226029a918STejun Heo if (flags & WQ_UNBOUND) { 4323be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 43246029a918STejun Heo if (!wq->unbound_attrs) 43256029a918STejun Heo goto err_free_wq; 43266029a918STejun Heo } 43276029a918STejun Heo 4328669de8bdSBart Van Assche va_start(args, max_active); 4329ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4330b196be89STejun Heo va_end(args); 43313af24433SOleg Nesterov 4332d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4333b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 43343af24433SOleg Nesterov 4335b196be89STejun Heo /* init wq */ 433697e37d7bSTejun Heo wq->flags = flags; 4337a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 43383c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4339112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 434030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 434173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 434273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4343493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 43443af24433SOleg Nesterov 4345669de8bdSBart Van Assche wq_init_lockdep(wq); 4346cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 43473af24433SOleg Nesterov 434830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 434982efcab3SBart Van Assche goto err_unreg_lockdep; 43501537663fSTejun Heo 435140c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4352d2c1d404STejun Heo goto err_destroy; 4353e22bee78STejun Heo 4354226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4355226223abSTejun Heo goto err_destroy; 4356226223abSTejun Heo 43576af8bf3dSOleg Nesterov /* 435868e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 435968e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 436068e13a67SLai Jiangshan * list. 43616af8bf3dSOleg Nesterov */ 436268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4363a0a1a5fdSTejun Heo 4364a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 436549e3cf44STejun Heo for_each_pwq(pwq, wq) 4366699ce097STejun Heo pwq_adjust_max_active(pwq); 4367a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4368a0a1a5fdSTejun Heo 4369e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4370a0a1a5fdSTejun Heo 437168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43723af24433SOleg Nesterov 43733af24433SOleg Nesterov return wq; 4374d2c1d404STejun Heo 437582efcab3SBart Van Assche err_unreg_lockdep: 4376009bb421SBart Van Assche wq_unregister_lockdep(wq); 4377009bb421SBart Van Assche wq_free_lockdep(wq); 437882efcab3SBart Van Assche err_free_wq: 43796029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 43804690c4abSTejun Heo kfree(wq); 4381d2c1d404STejun Heo return NULL; 4382d2c1d404STejun Heo err_destroy: 4383d2c1d404STejun Heo destroy_workqueue(wq); 43844690c4abSTejun Heo return NULL; 43851da177e4SLinus Torvalds } 4386669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 43871da177e4SLinus Torvalds 4388c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4389c29eb853STejun Heo { 4390c29eb853STejun Heo int i; 4391c29eb853STejun Heo 4392c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4393c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4394c29eb853STejun Heo return true; 4395c29eb853STejun Heo 4396c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4397c29eb853STejun Heo return true; 4398f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4399c29eb853STejun Heo return true; 4400c29eb853STejun Heo 4401c29eb853STejun Heo return false; 4402c29eb853STejun Heo } 4403c29eb853STejun Heo 44043af24433SOleg Nesterov /** 44053af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 44063af24433SOleg Nesterov * @wq: target workqueue 44073af24433SOleg Nesterov * 44083af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 44093af24433SOleg Nesterov */ 44103af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 44113af24433SOleg Nesterov { 441249e3cf44STejun Heo struct pool_workqueue *pwq; 44134c16bd32STejun Heo int node; 44143af24433SOleg Nesterov 4415def98c84STejun Heo /* 4416def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4417def98c84STejun Heo * lead to sysfs name conflicts. 4418def98c84STejun Heo */ 4419def98c84STejun Heo workqueue_sysfs_unregister(wq); 4420def98c84STejun Heo 442133e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 442233e3f0a3SRichard Clark mutex_lock(&wq->mutex); 442333e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 442433e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 442533e3f0a3SRichard Clark 44269c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 44279c5a2ba7STejun Heo drain_workqueue(wq); 4428c8efcc25STejun Heo 4429def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4430def98c84STejun Heo if (wq->rescuer) { 4431def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4432def98c84STejun Heo 4433def98c84STejun Heo /* this prevents new queueing */ 4434a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4435def98c84STejun Heo wq->rescuer = NULL; 4436a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4437def98c84STejun Heo 4438def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4439def98c84STejun Heo kthread_stop(rescuer->task); 44408efe1223STejun Heo kfree(rescuer); 4441def98c84STejun Heo } 4442def98c84STejun Heo 4443c29eb853STejun Heo /* 4444c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4445c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4446c29eb853STejun Heo */ 4447c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4448b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 444949e3cf44STejun Heo for_each_pwq(pwq, wq) { 4450a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4451c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 44521d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4453e66b39afSTejun Heo __func__, wq->name); 4454c29eb853STejun Heo show_pwq(pwq); 4455a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4456b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4457c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 445855df0933SImran Khan show_one_workqueue(wq); 44596183c009STejun Heo return; 446076af4d93STejun Heo } 4461a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 446276af4d93STejun Heo } 4463b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 44646183c009STejun Heo 4465a0a1a5fdSTejun Heo /* 4466a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4467a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4468a0a1a5fdSTejun Heo */ 4469e2dca7adSTejun Heo list_del_rcu(&wq->list); 447068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44713af24433SOleg Nesterov 44728864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4473669de8bdSBart Van Assche wq_unregister_lockdep(wq); 447429c91e99STejun Heo /* 44758864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4476e2dca7adSTejun Heo * schedule RCU free. 447729c91e99STejun Heo */ 447825b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 44798864b4e5STejun Heo } else { 44808864b4e5STejun Heo /* 44818864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 44824c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 44834c16bd32STejun Heo * @wq will be freed when the last pwq is released. 44848864b4e5STejun Heo */ 44854c16bd32STejun Heo for_each_node(node) { 44864c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 44874c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 44884c16bd32STejun Heo put_pwq_unlocked(pwq); 44894c16bd32STejun Heo } 44904c16bd32STejun Heo 44914c16bd32STejun Heo /* 44924c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 44934c16bd32STejun Heo * put. Don't access it afterwards. 44944c16bd32STejun Heo */ 44954c16bd32STejun Heo pwq = wq->dfl_pwq; 44964c16bd32STejun Heo wq->dfl_pwq = NULL; 4497dce90d47STejun Heo put_pwq_unlocked(pwq); 449829c91e99STejun Heo } 44993af24433SOleg Nesterov } 45003af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 45013af24433SOleg Nesterov 4502dcd989cbSTejun Heo /** 4503dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4504dcd989cbSTejun Heo * @wq: target workqueue 4505dcd989cbSTejun Heo * @max_active: new max_active value. 4506dcd989cbSTejun Heo * 4507dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4508dcd989cbSTejun Heo * 4509dcd989cbSTejun Heo * CONTEXT: 4510dcd989cbSTejun Heo * Don't call from IRQ context. 4511dcd989cbSTejun Heo */ 4512dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4513dcd989cbSTejun Heo { 451449e3cf44STejun Heo struct pool_workqueue *pwq; 4515dcd989cbSTejun Heo 45168719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 45170a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 45188719dceaSTejun Heo return; 45198719dceaSTejun Heo 4520f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4521dcd989cbSTejun Heo 4522a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4523dcd989cbSTejun Heo 45240a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4525dcd989cbSTejun Heo wq->saved_max_active = max_active; 4526dcd989cbSTejun Heo 4527699ce097STejun Heo for_each_pwq(pwq, wq) 4528699ce097STejun Heo pwq_adjust_max_active(pwq); 4529dcd989cbSTejun Heo 4530a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4531dcd989cbSTejun Heo } 4532dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4533dcd989cbSTejun Heo 4534dcd989cbSTejun Heo /** 453527d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 453627d4ee03SLukas Wunner * 453727d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 453827d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 453927d4ee03SLukas Wunner * 454027d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 454127d4ee03SLukas Wunner */ 454227d4ee03SLukas Wunner struct work_struct *current_work(void) 454327d4ee03SLukas Wunner { 454427d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 454527d4ee03SLukas Wunner 454627d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 454727d4ee03SLukas Wunner } 454827d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 454927d4ee03SLukas Wunner 455027d4ee03SLukas Wunner /** 4551e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4552e6267616STejun Heo * 4553e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4554e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4555d185af30SYacine Belkadi * 4556d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4557e6267616STejun Heo */ 4558e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4559e6267616STejun Heo { 4560e6267616STejun Heo struct worker *worker = current_wq_worker(); 4561e6267616STejun Heo 45626a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4563e6267616STejun Heo } 4564e6267616STejun Heo 4565e6267616STejun Heo /** 4566dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4567dcd989cbSTejun Heo * @cpu: CPU in question 4568dcd989cbSTejun Heo * @wq: target workqueue 4569dcd989cbSTejun Heo * 4570dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4571dcd989cbSTejun Heo * no synchronization around this function and the test result is 4572dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4573dcd989cbSTejun Heo * 4574d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4575d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4576d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4577d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4578d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4579d3251859STejun Heo * 4580d185af30SYacine Belkadi * Return: 4581dcd989cbSTejun Heo * %true if congested, %false otherwise. 4582dcd989cbSTejun Heo */ 4583d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4584dcd989cbSTejun Heo { 45857fb98ea7STejun Heo struct pool_workqueue *pwq; 458676af4d93STejun Heo bool ret; 458776af4d93STejun Heo 458824acfb71SThomas Gleixner rcu_read_lock(); 458924acfb71SThomas Gleixner preempt_disable(); 45907fb98ea7STejun Heo 4591d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4592d3251859STejun Heo cpu = smp_processor_id(); 4593d3251859STejun Heo 45947fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 45957fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 45967fb98ea7STejun Heo else 4597df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4598dcd989cbSTejun Heo 4599f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 460024acfb71SThomas Gleixner preempt_enable(); 460124acfb71SThomas Gleixner rcu_read_unlock(); 460276af4d93STejun Heo 460376af4d93STejun Heo return ret; 4604dcd989cbSTejun Heo } 4605dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4606dcd989cbSTejun Heo 4607dcd989cbSTejun Heo /** 4608dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4609dcd989cbSTejun Heo * @work: the work to be tested 4610dcd989cbSTejun Heo * 4611dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4612dcd989cbSTejun Heo * synchronization around this function and the test result is 4613dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4614dcd989cbSTejun Heo * 4615d185af30SYacine Belkadi * Return: 4616dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4617dcd989cbSTejun Heo */ 4618dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4619dcd989cbSTejun Heo { 4620fa1b54e6STejun Heo struct worker_pool *pool; 4621dcd989cbSTejun Heo unsigned long flags; 4622dcd989cbSTejun Heo unsigned int ret = 0; 4623dcd989cbSTejun Heo 4624dcd989cbSTejun Heo if (work_pending(work)) 4625dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4626038366c5SLai Jiangshan 462724acfb71SThomas Gleixner rcu_read_lock(); 4628fa1b54e6STejun Heo pool = get_work_pool(work); 4629038366c5SLai Jiangshan if (pool) { 4630a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4631c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4632dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4633a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4634038366c5SLai Jiangshan } 463524acfb71SThomas Gleixner rcu_read_unlock(); 4636dcd989cbSTejun Heo 4637dcd989cbSTejun Heo return ret; 4638dcd989cbSTejun Heo } 4639dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4640dcd989cbSTejun Heo 46413d1cb205STejun Heo /** 46423d1cb205STejun Heo * set_worker_desc - set description for the current work item 46433d1cb205STejun Heo * @fmt: printf-style format string 46443d1cb205STejun Heo * @...: arguments for the format string 46453d1cb205STejun Heo * 46463d1cb205STejun Heo * This function can be called by a running work function to describe what 46473d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 46483d1cb205STejun Heo * information will be printed out together to help debugging. The 46493d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 46503d1cb205STejun Heo */ 46513d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 46523d1cb205STejun Heo { 46533d1cb205STejun Heo struct worker *worker = current_wq_worker(); 46543d1cb205STejun Heo va_list args; 46553d1cb205STejun Heo 46563d1cb205STejun Heo if (worker) { 46573d1cb205STejun Heo va_start(args, fmt); 46583d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 46593d1cb205STejun Heo va_end(args); 46603d1cb205STejun Heo } 46613d1cb205STejun Heo } 46625c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 46633d1cb205STejun Heo 46643d1cb205STejun Heo /** 46653d1cb205STejun Heo * print_worker_info - print out worker information and description 46663d1cb205STejun Heo * @log_lvl: the log level to use when printing 46673d1cb205STejun Heo * @task: target task 46683d1cb205STejun Heo * 46693d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 46703d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 46713d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 46723d1cb205STejun Heo * 46733d1cb205STejun Heo * This function can be safely called on any task as long as the 46743d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 46753d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 46763d1cb205STejun Heo */ 46773d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 46783d1cb205STejun Heo { 46793d1cb205STejun Heo work_func_t *fn = NULL; 46803d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 46813d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 46823d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 46833d1cb205STejun Heo struct workqueue_struct *wq = NULL; 46843d1cb205STejun Heo struct worker *worker; 46853d1cb205STejun Heo 46863d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 46873d1cb205STejun Heo return; 46883d1cb205STejun Heo 46893d1cb205STejun Heo /* 46903d1cb205STejun Heo * This function is called without any synchronization and @task 46913d1cb205STejun Heo * could be in any state. Be careful with dereferences. 46923d1cb205STejun Heo */ 4693e700591aSPetr Mladek worker = kthread_probe_data(task); 46943d1cb205STejun Heo 46953d1cb205STejun Heo /* 46968bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 46978bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 46983d1cb205STejun Heo */ 4699fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4700fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4701fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4702fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4703fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 47043d1cb205STejun Heo 47053d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4706d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 47078bf89593STejun Heo if (strcmp(name, desc)) 47083d1cb205STejun Heo pr_cont(" (%s)", desc); 47093d1cb205STejun Heo pr_cont("\n"); 47103d1cb205STejun Heo } 47113d1cb205STejun Heo } 47123d1cb205STejun Heo 47133494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 47143494fc30STejun Heo { 47153494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 47163494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 47173494fc30STejun Heo pr_cont(" node=%d", pool->node); 47183494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 47193494fc30STejun Heo } 47203494fc30STejun Heo 4721*c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 4722*c76feb0dSPaul E. McKenney bool comma; 4723*c76feb0dSPaul E. McKenney work_func_t func; 4724*c76feb0dSPaul E. McKenney long ctr; 4725*c76feb0dSPaul E. McKenney }; 4726*c76feb0dSPaul E. McKenney 4727*c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 4728*c76feb0dSPaul E. McKenney { 4729*c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 4730*c76feb0dSPaul E. McKenney goto out_record; 4731*c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 4732*c76feb0dSPaul E. McKenney pcwsp->ctr++; 4733*c76feb0dSPaul E. McKenney return; 4734*c76feb0dSPaul E. McKenney } 4735*c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 4736*c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 4737*c76feb0dSPaul E. McKenney else 4738*c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 4739*c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 4740*c76feb0dSPaul E. McKenney out_record: 4741*c76feb0dSPaul E. McKenney if ((long)func == -1L) 4742*c76feb0dSPaul E. McKenney return; 4743*c76feb0dSPaul E. McKenney pcwsp->comma = comma; 4744*c76feb0dSPaul E. McKenney pcwsp->func = func; 4745*c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 4746*c76feb0dSPaul E. McKenney } 4747*c76feb0dSPaul E. McKenney 4748*c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 47493494fc30STejun Heo { 47503494fc30STejun Heo if (work->func == wq_barrier_func) { 47513494fc30STejun Heo struct wq_barrier *barr; 47523494fc30STejun Heo 47533494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 47543494fc30STejun Heo 4755*c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 47563494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 47573494fc30STejun Heo task_pid_nr(barr->task)); 47583494fc30STejun Heo } else { 4759*c76feb0dSPaul E. McKenney if (!comma) 4760*c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 4761*c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 47623494fc30STejun Heo } 47633494fc30STejun Heo } 47643494fc30STejun Heo 47653494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 47663494fc30STejun Heo { 4767*c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 47683494fc30STejun Heo struct worker_pool *pool = pwq->pool; 47693494fc30STejun Heo struct work_struct *work; 47703494fc30STejun Heo struct worker *worker; 47713494fc30STejun Heo bool has_in_flight = false, has_pending = false; 47723494fc30STejun Heo int bkt; 47733494fc30STejun Heo 47743494fc30STejun Heo pr_info(" pwq %d:", pool->id); 47753494fc30STejun Heo pr_cont_pool_info(pool); 47763494fc30STejun Heo 4777e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4778e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 47793494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 47803494fc30STejun Heo 47813494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47823494fc30STejun Heo if (worker->current_pwq == pwq) { 47833494fc30STejun Heo has_in_flight = true; 47843494fc30STejun Heo break; 47853494fc30STejun Heo } 47863494fc30STejun Heo } 47873494fc30STejun Heo if (has_in_flight) { 47883494fc30STejun Heo bool comma = false; 47893494fc30STejun Heo 47903494fc30STejun Heo pr_info(" in-flight:"); 47913494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47923494fc30STejun Heo if (worker->current_pwq != pwq) 47933494fc30STejun Heo continue; 47943494fc30STejun Heo 4795d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 47963494fc30STejun Heo task_pid_nr(worker->task), 479730ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 47983494fc30STejun Heo worker->current_func); 47993494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 4800*c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 4801*c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 48023494fc30STejun Heo comma = true; 48033494fc30STejun Heo } 48043494fc30STejun Heo pr_cont("\n"); 48053494fc30STejun Heo } 48063494fc30STejun Heo 48073494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 48083494fc30STejun Heo if (get_work_pwq(work) == pwq) { 48093494fc30STejun Heo has_pending = true; 48103494fc30STejun Heo break; 48113494fc30STejun Heo } 48123494fc30STejun Heo } 48133494fc30STejun Heo if (has_pending) { 48143494fc30STejun Heo bool comma = false; 48153494fc30STejun Heo 48163494fc30STejun Heo pr_info(" pending:"); 48173494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 48183494fc30STejun Heo if (get_work_pwq(work) != pwq) 48193494fc30STejun Heo continue; 48203494fc30STejun Heo 4821*c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 48223494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 48233494fc30STejun Heo } 4824*c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 48253494fc30STejun Heo pr_cont("\n"); 48263494fc30STejun Heo } 48273494fc30STejun Heo 4828f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 48293494fc30STejun Heo bool comma = false; 48303494fc30STejun Heo 4831f97a4a1aSLai Jiangshan pr_info(" inactive:"); 4832f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 4833*c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 48343494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 48353494fc30STejun Heo } 4836*c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 48373494fc30STejun Heo pr_cont("\n"); 48383494fc30STejun Heo } 48393494fc30STejun Heo } 48403494fc30STejun Heo 48413494fc30STejun Heo /** 484255df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 484355df0933SImran Khan * @wq: workqueue whose state will be printed 48443494fc30STejun Heo */ 484555df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 48463494fc30STejun Heo { 48473494fc30STejun Heo struct pool_workqueue *pwq; 48483494fc30STejun Heo bool idle = true; 484955df0933SImran Khan unsigned long flags; 48503494fc30STejun Heo 48513494fc30STejun Heo for_each_pwq(pwq, wq) { 4852f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 48533494fc30STejun Heo idle = false; 48543494fc30STejun Heo break; 48553494fc30STejun Heo } 48563494fc30STejun Heo } 485755df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 485855df0933SImran Khan return; 48593494fc30STejun Heo 48603494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 48613494fc30STejun Heo 48623494fc30STejun Heo for_each_pwq(pwq, wq) { 4863a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 486457116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 486557116ce1SJohan Hovold /* 486657116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 486757116ce1SJohan Hovold * drivers that queue work while holding locks 486857116ce1SJohan Hovold * also taken in their write paths. 486957116ce1SJohan Hovold */ 487057116ce1SJohan Hovold printk_deferred_enter(); 48713494fc30STejun Heo show_pwq(pwq); 487257116ce1SJohan Hovold printk_deferred_exit(); 487357116ce1SJohan Hovold } 4874a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 487562635ea8SSergey Senozhatsky /* 487662635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 487755df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 487862635ea8SSergey Senozhatsky * hard lockup. 487962635ea8SSergey Senozhatsky */ 488062635ea8SSergey Senozhatsky touch_nmi_watchdog(); 48813494fc30STejun Heo } 488255df0933SImran Khan 48833494fc30STejun Heo } 48843494fc30STejun Heo 488555df0933SImran Khan /** 488655df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 488755df0933SImran Khan * @pool: worker pool whose state will be printed 488855df0933SImran Khan */ 488955df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 489055df0933SImran Khan { 48913494fc30STejun Heo struct worker *worker; 48923494fc30STejun Heo bool first = true; 489355df0933SImran Khan unsigned long flags; 48943494fc30STejun Heo 4895a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 48963494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 48973494fc30STejun Heo goto next_pool; 489857116ce1SJohan Hovold /* 489957116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 490057116ce1SJohan Hovold * queue work while holding locks also taken in their write 490157116ce1SJohan Hovold * paths. 490257116ce1SJohan Hovold */ 490357116ce1SJohan Hovold printk_deferred_enter(); 49043494fc30STejun Heo pr_info("pool %d:", pool->id); 49053494fc30STejun Heo pr_cont_pool_info(pool); 490682607adcSTejun Heo pr_cont(" hung=%us workers=%d", 490782607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 490882607adcSTejun Heo pool->nr_workers); 49093494fc30STejun Heo if (pool->manager) 49103494fc30STejun Heo pr_cont(" manager: %d", 49113494fc30STejun Heo task_pid_nr(pool->manager->task)); 49123494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 49133494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 49143494fc30STejun Heo task_pid_nr(worker->task)); 49153494fc30STejun Heo first = false; 49163494fc30STejun Heo } 49173494fc30STejun Heo pr_cont("\n"); 491857116ce1SJohan Hovold printk_deferred_exit(); 49193494fc30STejun Heo next_pool: 4920a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 492162635ea8SSergey Senozhatsky /* 492262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 492355df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 492462635ea8SSergey Senozhatsky * hard lockup. 492562635ea8SSergey Senozhatsky */ 492662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 492755df0933SImran Khan 49283494fc30STejun Heo } 49293494fc30STejun Heo 493055df0933SImran Khan /** 493155df0933SImran Khan * show_all_workqueues - dump workqueue state 493255df0933SImran Khan * 493355df0933SImran Khan * Called from a sysrq handler or try_to_freeze_tasks() and prints out 493455df0933SImran Khan * all busy workqueues and pools. 493555df0933SImran Khan */ 493655df0933SImran Khan void show_all_workqueues(void) 493755df0933SImran Khan { 493855df0933SImran Khan struct workqueue_struct *wq; 493955df0933SImran Khan struct worker_pool *pool; 494055df0933SImran Khan int pi; 494155df0933SImran Khan 494255df0933SImran Khan rcu_read_lock(); 494355df0933SImran Khan 494455df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 494555df0933SImran Khan 494655df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 494755df0933SImran Khan show_one_workqueue(wq); 494855df0933SImran Khan 494955df0933SImran Khan for_each_pool(pool, pi) 495055df0933SImran Khan show_one_worker_pool(pool); 495155df0933SImran Khan 495224acfb71SThomas Gleixner rcu_read_unlock(); 49533494fc30STejun Heo } 49543494fc30STejun Heo 49556b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 49566b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 49576b59808bSTejun Heo { 49586b59808bSTejun Heo int off; 49596b59808bSTejun Heo 49606b59808bSTejun Heo /* always show the actual comm */ 49616b59808bSTejun Heo off = strscpy(buf, task->comm, size); 49626b59808bSTejun Heo if (off < 0) 49636b59808bSTejun Heo return; 49646b59808bSTejun Heo 4965197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 49666b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 49676b59808bSTejun Heo 4968197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4969197f6accSTejun Heo struct worker *worker = kthread_data(task); 4970197f6accSTejun Heo struct worker_pool *pool = worker->pool; 49716b59808bSTejun Heo 49726b59808bSTejun Heo if (pool) { 4973a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 49746b59808bSTejun Heo /* 4975197f6accSTejun Heo * ->desc tracks information (wq name or 4976197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4977197f6accSTejun Heo * current, prepend '+', otherwise '-'. 49786b59808bSTejun Heo */ 49796b59808bSTejun Heo if (worker->desc[0] != '\0') { 49806b59808bSTejun Heo if (worker->current_work) 49816b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 49826b59808bSTejun Heo worker->desc); 49836b59808bSTejun Heo else 49846b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 49856b59808bSTejun Heo worker->desc); 49866b59808bSTejun Heo } 4987a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 49886b59808bSTejun Heo } 4989197f6accSTejun Heo } 49906b59808bSTejun Heo 49916b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 49926b59808bSTejun Heo } 49936b59808bSTejun Heo 499466448bc2SMathieu Malaterre #ifdef CONFIG_SMP 499566448bc2SMathieu Malaterre 4996db7bccf4STejun Heo /* 4997db7bccf4STejun Heo * CPU hotplug. 4998db7bccf4STejun Heo * 4999e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5000112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5001706026c2STejun Heo * pool which make migrating pending and scheduled works very 5002e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 500394cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5004e22bee78STejun Heo * blocked draining impractical. 5005e22bee78STejun Heo * 500624647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5007628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5008628c78e7STejun Heo * cpu comes back online. 5009db7bccf4STejun Heo */ 5010db7bccf4STejun Heo 5011e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5012db7bccf4STejun Heo { 50134ce62e9eSTejun Heo struct worker_pool *pool; 5014db7bccf4STejun Heo struct worker *worker; 5015db7bccf4STejun Heo 5016f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 50171258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5018a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5019e22bee78STejun Heo 5020f2d5a0eeSTejun Heo /* 502192f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 502294cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 502311b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5024b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5025b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5026b4ac9384SLai Jiangshan * is on the same cpu. 5027f2d5a0eeSTejun Heo */ 5028da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5029403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5030db7bccf4STejun Heo 503124647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5032f2d5a0eeSTejun Heo 5033e22bee78STejun Heo /* 5034989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5035989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5036989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5037989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5038989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5039eb283428SLai Jiangshan * are served by workers tied to the pool. 5040e22bee78STejun Heo */ 5041bc35f7efSLai Jiangshan pool->nr_running = 0; 5042eb283428SLai Jiangshan 5043eb283428SLai Jiangshan /* 5044eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5045eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5046eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5047eb283428SLai Jiangshan */ 5048eb283428SLai Jiangshan wake_up_worker(pool); 5049989442d7SLai Jiangshan 5050a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5051989442d7SLai Jiangshan 5052989442d7SLai Jiangshan for_each_pool_worker(worker, pool) { 5053989442d7SLai Jiangshan kthread_set_per_cpu(worker->task, -1); 505446a4d679SLai Jiangshan if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 505510a5a651SZqiang WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 505646a4d679SLai Jiangshan else 505746a4d679SLai Jiangshan WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 5058989442d7SLai Jiangshan } 5059989442d7SLai Jiangshan 5060989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5061eb283428SLai Jiangshan } 5062db7bccf4STejun Heo } 5063db7bccf4STejun Heo 5064bd7c089eSTejun Heo /** 5065bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5066bd7c089eSTejun Heo * @pool: pool of interest 5067bd7c089eSTejun Heo * 5068a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5069bd7c089eSTejun Heo */ 5070bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5071bd7c089eSTejun Heo { 5072a9ab775bSTejun Heo struct worker *worker; 5073bd7c089eSTejun Heo 50741258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5075bd7c089eSTejun Heo 5076bd7c089eSTejun Heo /* 5077a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5078a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5079402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5080a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5081a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5082bd7c089eSTejun Heo */ 50835c25b5ffSPeter Zijlstra for_each_pool_worker(worker, pool) { 50845c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 5085a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5086a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 50875c25b5ffSPeter Zijlstra } 5088a9ab775bSTejun Heo 5089a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5090f7c17d26SWanpeng Li 50913de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5092a9ab775bSTejun Heo 5093da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5094a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5095a9ab775bSTejun Heo 5096a9ab775bSTejun Heo /* 5097a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5098a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5099a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5100a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5101a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5102a9ab775bSTejun Heo * concurrency management. Note that when or whether 5103a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5104a9ab775bSTejun Heo * 5105c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5106a9ab775bSTejun Heo * tested without holding any lock in 51076d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5108a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5109a9ab775bSTejun Heo * management operations. 5110bd7c089eSTejun Heo */ 5111a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5112a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5113a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5114c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5115bd7c089eSTejun Heo } 5116a9ab775bSTejun Heo 5117a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5118bd7c089eSTejun Heo } 5119bd7c089eSTejun Heo 51207dbc725eSTejun Heo /** 51217dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 51227dbc725eSTejun Heo * @pool: unbound pool of interest 51237dbc725eSTejun Heo * @cpu: the CPU which is coming up 51247dbc725eSTejun Heo * 51257dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 51267dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 51277dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 51287dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 51297dbc725eSTejun Heo */ 51307dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 51317dbc725eSTejun Heo { 51327dbc725eSTejun Heo static cpumask_t cpumask; 51337dbc725eSTejun Heo struct worker *worker; 51347dbc725eSTejun Heo 51351258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 51367dbc725eSTejun Heo 51377dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 51387dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 51397dbc725eSTejun Heo return; 51407dbc725eSTejun Heo 51417dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 51427dbc725eSTejun Heo 51437dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5144da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5145d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 51467dbc725eSTejun Heo } 51477dbc725eSTejun Heo 51487ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 51491da177e4SLinus Torvalds { 51504ce62e9eSTejun Heo struct worker_pool *pool; 51511da177e4SLinus Torvalds 5152f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 51533ce63377STejun Heo if (pool->nr_workers) 51543ce63377STejun Heo continue; 5155051e1850SLai Jiangshan if (!create_worker(pool)) 51567ee681b2SThomas Gleixner return -ENOMEM; 51573af24433SOleg Nesterov } 51587ee681b2SThomas Gleixner return 0; 51597ee681b2SThomas Gleixner } 51601da177e4SLinus Torvalds 51617ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 51627ee681b2SThomas Gleixner { 51637ee681b2SThomas Gleixner struct worker_pool *pool; 51647ee681b2SThomas Gleixner struct workqueue_struct *wq; 51657ee681b2SThomas Gleixner int pi; 51667ee681b2SThomas Gleixner 516768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 51687dbc725eSTejun Heo 51697dbc725eSTejun Heo for_each_pool(pool, pi) { 51701258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 517194cf58bbSTejun Heo 5172f05b558dSLai Jiangshan if (pool->cpu == cpu) 517394cf58bbSTejun Heo rebind_workers(pool); 5174f05b558dSLai Jiangshan else if (pool->cpu < 0) 51757dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 517694cf58bbSTejun Heo 51771258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 517894cf58bbSTejun Heo } 51797dbc725eSTejun Heo 51804c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51814c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51824c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 51834c16bd32STejun Heo 518468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 51857ee681b2SThomas Gleixner return 0; 518665758202STejun Heo } 518765758202STejun Heo 51887ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 518965758202STejun Heo { 51904c16bd32STejun Heo struct workqueue_struct *wq; 51918db25e78STejun Heo 51924c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5193e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5194e8b3f8dbSLai Jiangshan return -1; 5195e8b3f8dbSLai Jiangshan 5196e8b3f8dbSLai Jiangshan unbind_workers(cpu); 51974c16bd32STejun Heo 51984c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51994c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 52004c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 52014c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 52024c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 52034c16bd32STejun Heo 52047ee681b2SThomas Gleixner return 0; 520565758202STejun Heo } 520665758202STejun Heo 52072d3854a3SRusty Russell struct work_for_cpu { 5208ed48ece2STejun Heo struct work_struct work; 52092d3854a3SRusty Russell long (*fn)(void *); 52102d3854a3SRusty Russell void *arg; 52112d3854a3SRusty Russell long ret; 52122d3854a3SRusty Russell }; 52132d3854a3SRusty Russell 5214ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 52152d3854a3SRusty Russell { 5216ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5217ed48ece2STejun Heo 52182d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 52192d3854a3SRusty Russell } 52202d3854a3SRusty Russell 52212d3854a3SRusty Russell /** 522222aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 52232d3854a3SRusty Russell * @cpu: the cpu to run on 52242d3854a3SRusty Russell * @fn: the function to run 52252d3854a3SRusty Russell * @arg: the function arg 52262d3854a3SRusty Russell * 522731ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 52286b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5229d185af30SYacine Belkadi * 5230d185af30SYacine Belkadi * Return: The value @fn returns. 52312d3854a3SRusty Russell */ 5232d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 52332d3854a3SRusty Russell { 5234ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 52352d3854a3SRusty Russell 5236ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5237ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 523812997d1aSBjorn Helgaas flush_work(&wfc.work); 5239440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 52402d3854a3SRusty Russell return wfc.ret; 52412d3854a3SRusty Russell } 52422d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 52430e8d6a93SThomas Gleixner 52440e8d6a93SThomas Gleixner /** 52450e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 52460e8d6a93SThomas Gleixner * @cpu: the cpu to run on 52470e8d6a93SThomas Gleixner * @fn: the function to run 52480e8d6a93SThomas Gleixner * @arg: the function argument 52490e8d6a93SThomas Gleixner * 52500e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 52510e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 52520e8d6a93SThomas Gleixner * 52530e8d6a93SThomas Gleixner * Return: The value @fn returns. 52540e8d6a93SThomas Gleixner */ 52550e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 52560e8d6a93SThomas Gleixner { 52570e8d6a93SThomas Gleixner long ret = -ENODEV; 52580e8d6a93SThomas Gleixner 5259ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 52600e8d6a93SThomas Gleixner if (cpu_online(cpu)) 52610e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5262ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 52630e8d6a93SThomas Gleixner return ret; 52640e8d6a93SThomas Gleixner } 52650e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 52662d3854a3SRusty Russell #endif /* CONFIG_SMP */ 52672d3854a3SRusty Russell 5268a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5269e7577c50SRusty Russell 5270a0a1a5fdSTejun Heo /** 5271a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5272a0a1a5fdSTejun Heo * 527358a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5274f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5275706026c2STejun Heo * pool->worklist. 5276a0a1a5fdSTejun Heo * 5277a0a1a5fdSTejun Heo * CONTEXT: 5278a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5279a0a1a5fdSTejun Heo */ 5280a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5281a0a1a5fdSTejun Heo { 528224b8a847STejun Heo struct workqueue_struct *wq; 528324b8a847STejun Heo struct pool_workqueue *pwq; 5284a0a1a5fdSTejun Heo 528568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5286a0a1a5fdSTejun Heo 52876183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5288a0a1a5fdSTejun Heo workqueue_freezing = true; 5289a0a1a5fdSTejun Heo 529024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5291a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5292699ce097STejun Heo for_each_pwq(pwq, wq) 5293699ce097STejun Heo pwq_adjust_max_active(pwq); 5294a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5295a1056305STejun Heo } 52965bcab335STejun Heo 529768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5298a0a1a5fdSTejun Heo } 5299a0a1a5fdSTejun Heo 5300a0a1a5fdSTejun Heo /** 530158a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5302a0a1a5fdSTejun Heo * 5303a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5304a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5305a0a1a5fdSTejun Heo * 5306a0a1a5fdSTejun Heo * CONTEXT: 530768e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5308a0a1a5fdSTejun Heo * 5309d185af30SYacine Belkadi * Return: 531058a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 531158a69cb4STejun Heo * is complete. 5312a0a1a5fdSTejun Heo */ 5313a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5314a0a1a5fdSTejun Heo { 5315a0a1a5fdSTejun Heo bool busy = false; 531624b8a847STejun Heo struct workqueue_struct *wq; 531724b8a847STejun Heo struct pool_workqueue *pwq; 5318a0a1a5fdSTejun Heo 531968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5320a0a1a5fdSTejun Heo 53216183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5322a0a1a5fdSTejun Heo 532324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 532424b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 532524b8a847STejun Heo continue; 5326a0a1a5fdSTejun Heo /* 5327a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5328a0a1a5fdSTejun Heo * to peek without lock. 5329a0a1a5fdSTejun Heo */ 533024acfb71SThomas Gleixner rcu_read_lock(); 533124b8a847STejun Heo for_each_pwq(pwq, wq) { 53326183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5333112202d9STejun Heo if (pwq->nr_active) { 5334a0a1a5fdSTejun Heo busy = true; 533524acfb71SThomas Gleixner rcu_read_unlock(); 5336a0a1a5fdSTejun Heo goto out_unlock; 5337a0a1a5fdSTejun Heo } 5338a0a1a5fdSTejun Heo } 533924acfb71SThomas Gleixner rcu_read_unlock(); 5340a0a1a5fdSTejun Heo } 5341a0a1a5fdSTejun Heo out_unlock: 534268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5343a0a1a5fdSTejun Heo return busy; 5344a0a1a5fdSTejun Heo } 5345a0a1a5fdSTejun Heo 5346a0a1a5fdSTejun Heo /** 5347a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5348a0a1a5fdSTejun Heo * 5349a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5350706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5351a0a1a5fdSTejun Heo * 5352a0a1a5fdSTejun Heo * CONTEXT: 5353a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5354a0a1a5fdSTejun Heo */ 5355a0a1a5fdSTejun Heo void thaw_workqueues(void) 5356a0a1a5fdSTejun Heo { 535724b8a847STejun Heo struct workqueue_struct *wq; 535824b8a847STejun Heo struct pool_workqueue *pwq; 5359a0a1a5fdSTejun Heo 536068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5361a0a1a5fdSTejun Heo 5362a0a1a5fdSTejun Heo if (!workqueue_freezing) 5363a0a1a5fdSTejun Heo goto out_unlock; 5364a0a1a5fdSTejun Heo 536574b414eaSLai Jiangshan workqueue_freezing = false; 536624b8a847STejun Heo 536724b8a847STejun Heo /* restore max_active and repopulate worklist */ 536824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5369a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5370699ce097STejun Heo for_each_pwq(pwq, wq) 5371699ce097STejun Heo pwq_adjust_max_active(pwq); 5372a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 537324b8a847STejun Heo } 537424b8a847STejun Heo 5375a0a1a5fdSTejun Heo out_unlock: 537668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5377a0a1a5fdSTejun Heo } 5378a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5379a0a1a5fdSTejun Heo 5380042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5381042f7df1SLai Jiangshan { 5382042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5383042f7df1SLai Jiangshan int ret = 0; 5384042f7df1SLai Jiangshan struct workqueue_struct *wq; 5385042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5386042f7df1SLai Jiangshan 5387042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5388042f7df1SLai Jiangshan 5389042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5390042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5391042f7df1SLai Jiangshan continue; 5392042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5393042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5394042f7df1SLai Jiangshan continue; 5395042f7df1SLai Jiangshan 5396042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5397042f7df1SLai Jiangshan if (!ctx) { 5398042f7df1SLai Jiangshan ret = -ENOMEM; 5399042f7df1SLai Jiangshan break; 5400042f7df1SLai Jiangshan } 5401042f7df1SLai Jiangshan 5402042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5403042f7df1SLai Jiangshan } 5404042f7df1SLai Jiangshan 5405042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5406042f7df1SLai Jiangshan if (!ret) 5407042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5408042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5409042f7df1SLai Jiangshan } 5410042f7df1SLai Jiangshan 5411042f7df1SLai Jiangshan return ret; 5412042f7df1SLai Jiangshan } 5413042f7df1SLai Jiangshan 5414042f7df1SLai Jiangshan /** 5415042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5416042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5417042f7df1SLai Jiangshan * 5418042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5419042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5420042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5421042f7df1SLai Jiangshan * 542267dc8325SCai Huoqing * Return: 0 - Success 5423042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5424042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5425042f7df1SLai Jiangshan */ 5426042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5427042f7df1SLai Jiangshan { 5428042f7df1SLai Jiangshan int ret = -EINVAL; 5429042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5430042f7df1SLai Jiangshan 5431c98a9805STal Shorer /* 5432c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5433c98a9805STal Shorer * If the user wishes to include them, we allow that. 5434c98a9805STal Shorer */ 5435042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5436042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5437a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5438d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5439d25302e4SMenglong Dong ret = 0; 5440d25302e4SMenglong Dong goto out_unlock; 5441d25302e4SMenglong Dong } 5442d25302e4SMenglong Dong 5443d25302e4SMenglong Dong if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) { 5444d25302e4SMenglong Dong ret = -ENOMEM; 5445d25302e4SMenglong Dong goto out_unlock; 5446d25302e4SMenglong Dong } 5447042f7df1SLai Jiangshan 5448042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5449042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5450042f7df1SLai Jiangshan 5451042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5452042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5453042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5454042f7df1SLai Jiangshan 5455042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5456042f7df1SLai Jiangshan if (ret < 0) 5457042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5458042f7df1SLai Jiangshan 5459d25302e4SMenglong Dong free_cpumask_var(saved_cpumask); 5460d25302e4SMenglong Dong out_unlock: 5461a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5462042f7df1SLai Jiangshan } 5463042f7df1SLai Jiangshan 5464042f7df1SLai Jiangshan return ret; 5465042f7df1SLai Jiangshan } 5466042f7df1SLai Jiangshan 54676ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 54686ba94429SFrederic Weisbecker /* 54696ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 54706ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 54716ba94429SFrederic Weisbecker * following attributes. 54726ba94429SFrederic Weisbecker * 54736ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 54746ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 54756ba94429SFrederic Weisbecker * 54766ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 54776ba94429SFrederic Weisbecker * 54789a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 54796ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 54806ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 54819a19b463SWang Long * numa RW bool : whether enable NUMA affinity 54826ba94429SFrederic Weisbecker */ 54836ba94429SFrederic Weisbecker struct wq_device { 54846ba94429SFrederic Weisbecker struct workqueue_struct *wq; 54856ba94429SFrederic Weisbecker struct device dev; 54866ba94429SFrederic Weisbecker }; 54876ba94429SFrederic Weisbecker 54886ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 54896ba94429SFrederic Weisbecker { 54906ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 54916ba94429SFrederic Weisbecker 54926ba94429SFrederic Weisbecker return wq_dev->wq; 54936ba94429SFrederic Weisbecker } 54946ba94429SFrederic Weisbecker 54956ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 54966ba94429SFrederic Weisbecker char *buf) 54976ba94429SFrederic Weisbecker { 54986ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54996ba94429SFrederic Weisbecker 55006ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 55016ba94429SFrederic Weisbecker } 55026ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 55036ba94429SFrederic Weisbecker 55046ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 55056ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 55066ba94429SFrederic Weisbecker { 55076ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55086ba94429SFrederic Weisbecker 55096ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 55106ba94429SFrederic Weisbecker } 55116ba94429SFrederic Weisbecker 55126ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 55136ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 55146ba94429SFrederic Weisbecker size_t count) 55156ba94429SFrederic Weisbecker { 55166ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55176ba94429SFrederic Weisbecker int val; 55186ba94429SFrederic Weisbecker 55196ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 55206ba94429SFrederic Weisbecker return -EINVAL; 55216ba94429SFrederic Weisbecker 55226ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 55236ba94429SFrederic Weisbecker return count; 55246ba94429SFrederic Weisbecker } 55256ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 55266ba94429SFrederic Weisbecker 55276ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 55286ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 55296ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 55306ba94429SFrederic Weisbecker NULL, 55316ba94429SFrederic Weisbecker }; 55326ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 55336ba94429SFrederic Weisbecker 55346ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 55356ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 55366ba94429SFrederic Weisbecker { 55376ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55386ba94429SFrederic Weisbecker const char *delim = ""; 55396ba94429SFrederic Weisbecker int node, written = 0; 55406ba94429SFrederic Weisbecker 5541ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 554224acfb71SThomas Gleixner rcu_read_lock(); 55436ba94429SFrederic Weisbecker for_each_node(node) { 55446ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 55456ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 55466ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 55476ba94429SFrederic Weisbecker delim = " "; 55486ba94429SFrederic Weisbecker } 55496ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 555024acfb71SThomas Gleixner rcu_read_unlock(); 5551ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 55526ba94429SFrederic Weisbecker 55536ba94429SFrederic Weisbecker return written; 55546ba94429SFrederic Weisbecker } 55556ba94429SFrederic Weisbecker 55566ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 55576ba94429SFrederic Weisbecker char *buf) 55586ba94429SFrederic Weisbecker { 55596ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55606ba94429SFrederic Weisbecker int written; 55616ba94429SFrederic Weisbecker 55626ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55636ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 55646ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55656ba94429SFrederic Weisbecker 55666ba94429SFrederic Weisbecker return written; 55676ba94429SFrederic Weisbecker } 55686ba94429SFrederic Weisbecker 55696ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 55706ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 55716ba94429SFrederic Weisbecker { 55726ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 55736ba94429SFrederic Weisbecker 5574899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5575899a94feSLai Jiangshan 5576be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 55776ba94429SFrederic Weisbecker if (!attrs) 55786ba94429SFrederic Weisbecker return NULL; 55796ba94429SFrederic Weisbecker 55806ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 55816ba94429SFrederic Weisbecker return attrs; 55826ba94429SFrederic Weisbecker } 55836ba94429SFrederic Weisbecker 55846ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 55856ba94429SFrederic Weisbecker const char *buf, size_t count) 55866ba94429SFrederic Weisbecker { 55876ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55886ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5589d4d3e257SLai Jiangshan int ret = -ENOMEM; 5590d4d3e257SLai Jiangshan 5591d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55926ba94429SFrederic Weisbecker 55936ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55946ba94429SFrederic Weisbecker if (!attrs) 5595d4d3e257SLai Jiangshan goto out_unlock; 55966ba94429SFrederic Weisbecker 55976ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 55986ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5599d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56006ba94429SFrederic Weisbecker else 56016ba94429SFrederic Weisbecker ret = -EINVAL; 56026ba94429SFrederic Weisbecker 5603d4d3e257SLai Jiangshan out_unlock: 5604d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 56056ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56066ba94429SFrederic Weisbecker return ret ?: count; 56076ba94429SFrederic Weisbecker } 56086ba94429SFrederic Weisbecker 56096ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 56106ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 56116ba94429SFrederic Weisbecker { 56126ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56136ba94429SFrederic Weisbecker int written; 56146ba94429SFrederic Weisbecker 56156ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 56166ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 56176ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 56186ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 56196ba94429SFrederic Weisbecker return written; 56206ba94429SFrederic Weisbecker } 56216ba94429SFrederic Weisbecker 56226ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 56236ba94429SFrederic Weisbecker struct device_attribute *attr, 56246ba94429SFrederic Weisbecker const char *buf, size_t count) 56256ba94429SFrederic Weisbecker { 56266ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56276ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5628d4d3e257SLai Jiangshan int ret = -ENOMEM; 5629d4d3e257SLai Jiangshan 5630d4d3e257SLai Jiangshan apply_wqattrs_lock(); 56316ba94429SFrederic Weisbecker 56326ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 56336ba94429SFrederic Weisbecker if (!attrs) 5634d4d3e257SLai Jiangshan goto out_unlock; 56356ba94429SFrederic Weisbecker 56366ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 56376ba94429SFrederic Weisbecker if (!ret) 5638d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56396ba94429SFrederic Weisbecker 5640d4d3e257SLai Jiangshan out_unlock: 5641d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 56426ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56436ba94429SFrederic Weisbecker return ret ?: count; 56446ba94429SFrederic Weisbecker } 56456ba94429SFrederic Weisbecker 56466ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 56476ba94429SFrederic Weisbecker char *buf) 56486ba94429SFrederic Weisbecker { 56496ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56506ba94429SFrederic Weisbecker int written; 56516ba94429SFrederic Weisbecker 56526ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 56536ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 56546ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 56556ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 56566ba94429SFrederic Weisbecker 56576ba94429SFrederic Weisbecker return written; 56586ba94429SFrederic Weisbecker } 56596ba94429SFrederic Weisbecker 56606ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 56616ba94429SFrederic Weisbecker const char *buf, size_t count) 56626ba94429SFrederic Weisbecker { 56636ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56646ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5665d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5666d4d3e257SLai Jiangshan 5667d4d3e257SLai Jiangshan apply_wqattrs_lock(); 56686ba94429SFrederic Weisbecker 56696ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 56706ba94429SFrederic Weisbecker if (!attrs) 5671d4d3e257SLai Jiangshan goto out_unlock; 56726ba94429SFrederic Weisbecker 56736ba94429SFrederic Weisbecker ret = -EINVAL; 56746ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 56756ba94429SFrederic Weisbecker attrs->no_numa = !v; 5676d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56776ba94429SFrederic Weisbecker } 56786ba94429SFrederic Weisbecker 5679d4d3e257SLai Jiangshan out_unlock: 5680d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 56816ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56826ba94429SFrederic Weisbecker return ret ?: count; 56836ba94429SFrederic Weisbecker } 56846ba94429SFrederic Weisbecker 56856ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 56866ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 56876ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 56886ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 56896ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 56906ba94429SFrederic Weisbecker __ATTR_NULL, 56916ba94429SFrederic Weisbecker }; 56926ba94429SFrederic Weisbecker 56936ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 56946ba94429SFrederic Weisbecker .name = "workqueue", 56956ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 56966ba94429SFrederic Weisbecker }; 56976ba94429SFrederic Weisbecker 5698b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5699b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5700b05a7928SFrederic Weisbecker { 5701b05a7928SFrederic Weisbecker int written; 5702b05a7928SFrederic Weisbecker 5703042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5704b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5705b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5706042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5707b05a7928SFrederic Weisbecker 5708b05a7928SFrederic Weisbecker return written; 5709b05a7928SFrederic Weisbecker } 5710b05a7928SFrederic Weisbecker 5711042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5712042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5713042f7df1SLai Jiangshan { 5714042f7df1SLai Jiangshan cpumask_var_t cpumask; 5715042f7df1SLai Jiangshan int ret; 5716042f7df1SLai Jiangshan 5717042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5718042f7df1SLai Jiangshan return -ENOMEM; 5719042f7df1SLai Jiangshan 5720042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5721042f7df1SLai Jiangshan if (!ret) 5722042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5723042f7df1SLai Jiangshan 5724042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5725042f7df1SLai Jiangshan return ret ? ret : count; 5726042f7df1SLai Jiangshan } 5727042f7df1SLai Jiangshan 5728b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5729042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5730042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5731b05a7928SFrederic Weisbecker 57326ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 57336ba94429SFrederic Weisbecker { 5734b05a7928SFrederic Weisbecker int err; 5735b05a7928SFrederic Weisbecker 5736b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5737b05a7928SFrederic Weisbecker if (err) 5738b05a7928SFrederic Weisbecker return err; 5739b05a7928SFrederic Weisbecker 5740b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 57416ba94429SFrederic Weisbecker } 57426ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 57436ba94429SFrederic Weisbecker 57446ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 57456ba94429SFrederic Weisbecker { 57466ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 57476ba94429SFrederic Weisbecker 57486ba94429SFrederic Weisbecker kfree(wq_dev); 57496ba94429SFrederic Weisbecker } 57506ba94429SFrederic Weisbecker 57516ba94429SFrederic Weisbecker /** 57526ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 57536ba94429SFrederic Weisbecker * @wq: the workqueue to register 57546ba94429SFrederic Weisbecker * 57556ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 57566ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 57576ba94429SFrederic Weisbecker * which is the preferred method. 57586ba94429SFrederic Weisbecker * 57596ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 57606ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 57616ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 57626ba94429SFrederic Weisbecker * attributes. 57636ba94429SFrederic Weisbecker * 57646ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 57656ba94429SFrederic Weisbecker */ 57666ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 57676ba94429SFrederic Weisbecker { 57686ba94429SFrederic Weisbecker struct wq_device *wq_dev; 57696ba94429SFrederic Weisbecker int ret; 57706ba94429SFrederic Weisbecker 57716ba94429SFrederic Weisbecker /* 5772402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 57736ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 57746ba94429SFrederic Weisbecker * workqueues. 57756ba94429SFrederic Weisbecker */ 57760a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 57776ba94429SFrederic Weisbecker return -EINVAL; 57786ba94429SFrederic Weisbecker 57796ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 57806ba94429SFrederic Weisbecker if (!wq_dev) 57816ba94429SFrederic Weisbecker return -ENOMEM; 57826ba94429SFrederic Weisbecker 57836ba94429SFrederic Weisbecker wq_dev->wq = wq; 57846ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 57856ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 578623217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 57876ba94429SFrederic Weisbecker 57886ba94429SFrederic Weisbecker /* 57896ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 57906ba94429SFrederic Weisbecker * everything is ready. 57916ba94429SFrederic Weisbecker */ 57926ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 57936ba94429SFrederic Weisbecker 57946ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 57956ba94429SFrederic Weisbecker if (ret) { 5796537f4146SArvind Yadav put_device(&wq_dev->dev); 57976ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57986ba94429SFrederic Weisbecker return ret; 57996ba94429SFrederic Weisbecker } 58006ba94429SFrederic Weisbecker 58016ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 58026ba94429SFrederic Weisbecker struct device_attribute *attr; 58036ba94429SFrederic Weisbecker 58046ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 58056ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 58066ba94429SFrederic Weisbecker if (ret) { 58076ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 58086ba94429SFrederic Weisbecker wq->wq_dev = NULL; 58096ba94429SFrederic Weisbecker return ret; 58106ba94429SFrederic Weisbecker } 58116ba94429SFrederic Weisbecker } 58126ba94429SFrederic Weisbecker } 58136ba94429SFrederic Weisbecker 58146ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 58156ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 58166ba94429SFrederic Weisbecker return 0; 58176ba94429SFrederic Weisbecker } 58186ba94429SFrederic Weisbecker 58196ba94429SFrederic Weisbecker /** 58206ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 58216ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 58226ba94429SFrederic Weisbecker * 58236ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 58246ba94429SFrederic Weisbecker */ 58256ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 58266ba94429SFrederic Weisbecker { 58276ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 58286ba94429SFrederic Weisbecker 58296ba94429SFrederic Weisbecker if (!wq->wq_dev) 58306ba94429SFrederic Weisbecker return; 58316ba94429SFrederic Weisbecker 58326ba94429SFrederic Weisbecker wq->wq_dev = NULL; 58336ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 58346ba94429SFrederic Weisbecker } 58356ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 58366ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 58376ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 58386ba94429SFrederic Weisbecker 583982607adcSTejun Heo /* 584082607adcSTejun Heo * Workqueue watchdog. 584182607adcSTejun Heo * 584282607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 584382607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 584482607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 584582607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 584682607adcSTejun Heo * largely opaque. 584782607adcSTejun Heo * 584882607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 584982607adcSTejun Heo * state if some pools failed to make forward progress for a while where 585082607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 585182607adcSTejun Heo * 585282607adcSTejun Heo * This mechanism is controlled through the kernel parameter 585382607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 585482607adcSTejun Heo * corresponding sysfs parameter file. 585582607adcSTejun Heo */ 585682607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 585782607adcSTejun Heo 585882607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 58595cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 586082607adcSTejun Heo 586182607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 586282607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 586382607adcSTejun Heo 586482607adcSTejun Heo static void wq_watchdog_reset_touched(void) 586582607adcSTejun Heo { 586682607adcSTejun Heo int cpu; 586782607adcSTejun Heo 586882607adcSTejun Heo wq_watchdog_touched = jiffies; 586982607adcSTejun Heo for_each_possible_cpu(cpu) 587082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 587182607adcSTejun Heo } 587282607adcSTejun Heo 58735cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 587482607adcSTejun Heo { 587582607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 587682607adcSTejun Heo bool lockup_detected = false; 5877940d71c6SSergey Senozhatsky unsigned long now = jiffies; 587882607adcSTejun Heo struct worker_pool *pool; 587982607adcSTejun Heo int pi; 588082607adcSTejun Heo 588182607adcSTejun Heo if (!thresh) 588282607adcSTejun Heo return; 588382607adcSTejun Heo 588482607adcSTejun Heo rcu_read_lock(); 588582607adcSTejun Heo 588682607adcSTejun Heo for_each_pool(pool, pi) { 588782607adcSTejun Heo unsigned long pool_ts, touched, ts; 588882607adcSTejun Heo 588982607adcSTejun Heo if (list_empty(&pool->worklist)) 589082607adcSTejun Heo continue; 589182607adcSTejun Heo 5892940d71c6SSergey Senozhatsky /* 5893940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 5894940d71c6SSergey Senozhatsky * the watchdog like a stall. 5895940d71c6SSergey Senozhatsky */ 5896940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 5897940d71c6SSergey Senozhatsky 589882607adcSTejun Heo /* get the latest of pool and touched timestamps */ 589989e28ce6SWang Qing if (pool->cpu >= 0) 590089e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 590189e28ce6SWang Qing else 590282607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 590389e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 590482607adcSTejun Heo 590582607adcSTejun Heo if (time_after(pool_ts, touched)) 590682607adcSTejun Heo ts = pool_ts; 590782607adcSTejun Heo else 590882607adcSTejun Heo ts = touched; 590982607adcSTejun Heo 591082607adcSTejun Heo /* did we stall? */ 5911940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 591282607adcSTejun Heo lockup_detected = true; 591382607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 591482607adcSTejun Heo pr_cont_pool_info(pool); 591582607adcSTejun Heo pr_cont(" stuck for %us!\n", 5916940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 591782607adcSTejun Heo } 591882607adcSTejun Heo } 591982607adcSTejun Heo 592082607adcSTejun Heo rcu_read_unlock(); 592182607adcSTejun Heo 592282607adcSTejun Heo if (lockup_detected) 592355df0933SImran Khan show_all_workqueues(); 592482607adcSTejun Heo 592582607adcSTejun Heo wq_watchdog_reset_touched(); 592682607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 592782607adcSTejun Heo } 592882607adcSTejun Heo 5929cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 593082607adcSTejun Heo { 593182607adcSTejun Heo if (cpu >= 0) 593282607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 593389e28ce6SWang Qing 593482607adcSTejun Heo wq_watchdog_touched = jiffies; 593582607adcSTejun Heo } 593682607adcSTejun Heo 593782607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 593882607adcSTejun Heo { 593982607adcSTejun Heo wq_watchdog_thresh = 0; 594082607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 594182607adcSTejun Heo 594282607adcSTejun Heo if (thresh) { 594382607adcSTejun Heo wq_watchdog_thresh = thresh; 594482607adcSTejun Heo wq_watchdog_reset_touched(); 594582607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 594682607adcSTejun Heo } 594782607adcSTejun Heo } 594882607adcSTejun Heo 594982607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 595082607adcSTejun Heo const struct kernel_param *kp) 595182607adcSTejun Heo { 595282607adcSTejun Heo unsigned long thresh; 595382607adcSTejun Heo int ret; 595482607adcSTejun Heo 595582607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 595682607adcSTejun Heo if (ret) 595782607adcSTejun Heo return ret; 595882607adcSTejun Heo 595982607adcSTejun Heo if (system_wq) 596082607adcSTejun Heo wq_watchdog_set_thresh(thresh); 596182607adcSTejun Heo else 596282607adcSTejun Heo wq_watchdog_thresh = thresh; 596382607adcSTejun Heo 596482607adcSTejun Heo return 0; 596582607adcSTejun Heo } 596682607adcSTejun Heo 596782607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 596882607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 596982607adcSTejun Heo .get = param_get_ulong, 597082607adcSTejun Heo }; 597182607adcSTejun Heo 597282607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 597382607adcSTejun Heo 0644); 597482607adcSTejun Heo 597582607adcSTejun Heo static void wq_watchdog_init(void) 597682607adcSTejun Heo { 59775cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 597882607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 597982607adcSTejun Heo } 598082607adcSTejun Heo 598182607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 598282607adcSTejun Heo 598382607adcSTejun Heo static inline void wq_watchdog_init(void) { } 598482607adcSTejun Heo 598582607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 598682607adcSTejun Heo 5987bce90380STejun Heo static void __init wq_numa_init(void) 5988bce90380STejun Heo { 5989bce90380STejun Heo cpumask_var_t *tbl; 5990bce90380STejun Heo int node, cpu; 5991bce90380STejun Heo 5992bce90380STejun Heo if (num_possible_nodes() <= 1) 5993bce90380STejun Heo return; 5994bce90380STejun Heo 5995d55262c4STejun Heo if (wq_disable_numa) { 5996d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5997d55262c4STejun Heo return; 5998d55262c4STejun Heo } 5999d55262c4STejun Heo 6000f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 6001f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6002f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6003f728c4a9SZhen Lei return; 6004f728c4a9SZhen Lei } 6005f728c4a9SZhen Lei } 6006f728c4a9SZhen Lei 6007be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 60084c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 60094c16bd32STejun Heo 6010bce90380STejun Heo /* 6011bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 6012bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 6013bce90380STejun Heo * fully initialized by now. 6014bce90380STejun Heo */ 60156396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6016bce90380STejun Heo BUG_ON(!tbl); 6017bce90380STejun Heo 6018bce90380STejun Heo for_each_node(node) 60195a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 60201be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6021bce90380STejun Heo 6022bce90380STejun Heo for_each_possible_cpu(cpu) { 6023bce90380STejun Heo node = cpu_to_node(cpu); 6024bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6025bce90380STejun Heo } 6026bce90380STejun Heo 6027bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6028bce90380STejun Heo wq_numa_enabled = true; 6029bce90380STejun Heo } 6030bce90380STejun Heo 60313347fa09STejun Heo /** 60323347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 60333347fa09STejun Heo * 60343347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 60353347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 60363347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 60373347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 60383347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 60393347fa09STejun Heo * created and scheduled right before early initcalls. 60403347fa09STejun Heo */ 60412333e829SYu Chen void __init workqueue_init_early(void) 60421da177e4SLinus Torvalds { 60437a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 60447a4e344cSTejun Heo int i, cpu; 6045c34056a3STejun Heo 604610cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6047e904e6c2STejun Heo 6048b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 604904d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 605004d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6051b05a7928SFrederic Weisbecker 6052e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6053e904e6c2STejun Heo 6054706026c2STejun Heo /* initialize CPU pools */ 605529c91e99STejun Heo for_each_possible_cpu(cpu) { 60564ce62e9eSTejun Heo struct worker_pool *pool; 60578b03ae3cSTejun Heo 60587a4e344cSTejun Heo i = 0; 6059f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 60607a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6061ec22ca5eSTejun Heo pool->cpu = cpu; 60627a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 60637a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6064f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 60657a4e344cSTejun Heo 60669daf9e67STejun Heo /* alloc pool ID */ 606768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 60689daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 606968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 60704ce62e9eSTejun Heo } 60718b03ae3cSTejun Heo } 60728b03ae3cSTejun Heo 60738a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 607429c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 607529c91e99STejun Heo struct workqueue_attrs *attrs; 607629c91e99STejun Heo 6077be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 607829c91e99STejun Heo attrs->nice = std_nice[i]; 607929c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 60808a2b7538STejun Heo 60818a2b7538STejun Heo /* 60828a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 60838a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 60848a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 60858a2b7538STejun Heo */ 6086be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 60878a2b7538STejun Heo attrs->nice = std_nice[i]; 60888a2b7538STejun Heo attrs->no_numa = true; 60898a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 609029c91e99STejun Heo } 609129c91e99STejun Heo 6092d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 60931aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6094d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6095f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6096f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 609724d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 609824d51addSTejun Heo WQ_FREEZABLE, 0); 60990668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 61000668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 61010668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 61020668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 61030668106cSViresh Kumar 0); 61041aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 61050668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 61060668106cSViresh Kumar !system_power_efficient_wq || 61070668106cSViresh Kumar !system_freezable_power_efficient_wq); 61083347fa09STejun Heo } 61093347fa09STejun Heo 61103347fa09STejun Heo /** 61113347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 61123347fa09STejun Heo * 61133347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 61143347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 61153347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 61163347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 61173347fa09STejun Heo * with the initial workers and enable future kworker creations. 61183347fa09STejun Heo */ 61192333e829SYu Chen void __init workqueue_init(void) 61203347fa09STejun Heo { 61212186d9f9STejun Heo struct workqueue_struct *wq; 61223347fa09STejun Heo struct worker_pool *pool; 61233347fa09STejun Heo int cpu, bkt; 61243347fa09STejun Heo 61252186d9f9STejun Heo /* 61262186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 61272186d9f9STejun Heo * CPU to node mapping may not be available that early on some 61282186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 61292186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 61302186d9f9STejun Heo * affinity, fix them up. 613140c17f75STejun Heo * 613240c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 61332186d9f9STejun Heo */ 61342186d9f9STejun Heo wq_numa_init(); 61352186d9f9STejun Heo 61362186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 61372186d9f9STejun Heo 61382186d9f9STejun Heo for_each_possible_cpu(cpu) { 61392186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 61402186d9f9STejun Heo pool->node = cpu_to_node(cpu); 61412186d9f9STejun Heo } 61422186d9f9STejun Heo } 61432186d9f9STejun Heo 614440c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 61452186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 614640c17f75STejun Heo WARN(init_rescuer(wq), 614740c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 614840c17f75STejun Heo wq->name); 614940c17f75STejun Heo } 61502186d9f9STejun Heo 61512186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 61522186d9f9STejun Heo 61533347fa09STejun Heo /* create the initial workers */ 61543347fa09STejun Heo for_each_online_cpu(cpu) { 61553347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 61563347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 61573347fa09STejun Heo BUG_ON(!create_worker(pool)); 61583347fa09STejun Heo } 61593347fa09STejun Heo } 61603347fa09STejun Heo 61613347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 61623347fa09STejun Heo BUG_ON(!create_worker(pool)); 61633347fa09STejun Heo 61643347fa09STejun Heo wq_online = true; 616582607adcSTejun Heo wq_watchdog_init(); 61661da177e4SLinus Torvalds } 6167c4f135d6STetsuo Handa 6168c4f135d6STetsuo Handa /* 6169c4f135d6STetsuo Handa * Despite the naming, this is a no-op function which is here only for avoiding 6170c4f135d6STetsuo Handa * link error. Since compile-time warning may fail to catch, we will need to 6171c4f135d6STetsuo Handa * emit run-time warning from __flush_workqueue(). 6172c4f135d6STetsuo Handa */ 6173c4f135d6STetsuo Handa void __warn_flushing_systemwide_wq(void) { } 6174c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6175