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> 53e22bee78STejun Heo 54ea138446STejun Heo #include "workqueue_internal.h" 551da177e4SLinus Torvalds 56c8e55f36STejun Heo enum { 57bc2ae0f5STejun Heo /* 5824647570STejun Heo * worker_pool flags 59bc2ae0f5STejun Heo * 6024647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 61bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 62bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 63bc2ae0f5STejun Heo * is in effect. 64bc2ae0f5STejun Heo * 65bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 66bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6724647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 68bc2ae0f5STejun Heo * 69bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 701258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 714736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 72bc2ae0f5STejun Heo */ 73692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7424647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 75db7bccf4STejun Heo 76c8e55f36STejun Heo /* worker flags */ 77c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 78c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 79e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 80fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 81f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 82a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 83e22bee78STejun Heo 84a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 85a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 86db7bccf4STejun Heo 87e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 884ce62e9eSTejun Heo 8929c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 90c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 91db7bccf4STejun Heo 92e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 93e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 94e22bee78STejun Heo 953233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 963233cdbdSTejun Heo /* call for help after 10ms 973233cdbdSTejun Heo (min two ticks) */ 98e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 99e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1001da177e4SLinus Torvalds 1011da177e4SLinus Torvalds /* 102e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1038698a745SDongsheng Yang * all cpus. Give MIN_NICE. 104e22bee78STejun Heo */ 1058698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1068698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 107ecf6881fSTejun Heo 108ecf6881fSTejun Heo WQ_NAME_LEN = 24, 109c8e55f36STejun Heo }; 110c8e55f36STejun Heo 1111da177e4SLinus Torvalds /* 1124690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1134690c4abSTejun Heo * 114e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 115e41e704bSTejun Heo * everyone else. 1164690c4abSTejun Heo * 117e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 118e22bee78STejun Heo * only be modified and accessed from the local cpu. 119e22bee78STejun Heo * 120d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1214690c4abSTejun Heo * 122d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 123d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 124d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 125d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 126e22bee78STejun Heo * 1271258fae7STejun Heo * A: wq_pool_attach_mutex protected. 128822d8405STejun Heo * 12968e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13076af4d93STejun Heo * 13124acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1325bcab335STejun Heo * 1335b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1345b95e1afSLai Jiangshan * 1355b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 13624acfb71SThomas Gleixner * RCU for reads. 1375b95e1afSLai Jiangshan * 1383c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1393c25a55dSLai Jiangshan * 14024acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1412e109a28STejun Heo * 1422e109a28STejun Heo * MD: wq_mayday_lock protected. 1434690c4abSTejun Heo */ 1444690c4abSTejun Heo 1452eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 146c34056a3STejun Heo 147bd7bdd43STejun Heo struct worker_pool { 148a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 149d84ff051STejun Heo int cpu; /* I: the associated cpu */ 150f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1519daf9e67STejun Heo int id; /* I: pool ID */ 15211ebea50STejun Heo unsigned int flags; /* X: flags */ 153bd7bdd43STejun Heo 15482607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15582607adcSTejun Heo 156bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 157ea1abd61SLai Jiangshan 1585826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1595826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 160bd7bdd43STejun Heo 161bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 162bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 163bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 164bd7bdd43STejun Heo 165c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 166c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 167c9e7cf27STejun Heo /* L: hash of busy workers */ 168c9e7cf27STejun Heo 1692607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17092f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17160f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 172e19e397aSTejun Heo 1737cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 174e19e397aSTejun Heo 1757a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17668e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17768e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1787a4e344cSTejun Heo 179e19e397aSTejun Heo /* 180e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 181e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 182e19e397aSTejun Heo * cacheline. 183e19e397aSTejun Heo */ 184e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18529c91e99STejun Heo 18629c91e99STejun Heo /* 18724acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 18829c91e99STejun Heo * from get_work_pool(). 18929c91e99STejun Heo */ 19029c91e99STejun Heo struct rcu_head rcu; 1918b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1928b03ae3cSTejun Heo 1938b03ae3cSTejun Heo /* 194112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 195112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 196112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 197112202d9STejun Heo * number of flag bits. 1981da177e4SLinus Torvalds */ 199112202d9STejun Heo struct pool_workqueue { 200bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2014690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20273f53c4aSTejun Heo int work_color; /* L: current color */ 20373f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2048864b4e5STejun Heo int refcnt; /* L: reference count */ 20573f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20673f53c4aSTejun Heo /* L: nr of in_flight works */ 2071e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 208a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2091e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2103c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2112e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2128864b4e5STejun Heo 2138864b4e5STejun Heo /* 2148864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2158864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 21624acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 217b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2188864b4e5STejun Heo */ 2198864b4e5STejun Heo struct work_struct unbound_release_work; 2208864b4e5STejun Heo struct rcu_head rcu; 221e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2221da177e4SLinus Torvalds 2231da177e4SLinus Torvalds /* 22473f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22573f53c4aSTejun Heo */ 22673f53c4aSTejun Heo struct wq_flusher { 2273c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2283c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22973f53c4aSTejun Heo struct completion done; /* flush completion */ 23073f53c4aSTejun Heo }; 2311da177e4SLinus Torvalds 232226223abSTejun Heo struct wq_device; 233226223abSTejun Heo 23473f53c4aSTejun Heo /* 235c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 236c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2371da177e4SLinus Torvalds */ 2381da177e4SLinus Torvalds struct workqueue_struct { 2393c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 240e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 24173f53c4aSTejun Heo 2423c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2433c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2443c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 245112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2463c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2473c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2483c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24973f53c4aSTejun Heo 2502e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 25130ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 252e22bee78STejun Heo 25387fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 254a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 255226223abSTejun Heo 2565b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2575b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2586029a918STejun Heo 259226223abSTejun Heo #ifdef CONFIG_SYSFS 260226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 261226223abSTejun Heo #endif 2624e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 263669de8bdSBart Van Assche char *lock_name; 264669de8bdSBart Van Assche struct lock_class_key key; 2654e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2664e6045f1SJohannes Berg #endif 267ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2682728fd2fSTejun Heo 269e2dca7adSTejun Heo /* 27024acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 27124acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 272e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 273e2dca7adSTejun Heo */ 274e2dca7adSTejun Heo struct rcu_head rcu; 275e2dca7adSTejun Heo 2762728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2772728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2782728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2795b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2801da177e4SLinus Torvalds }; 2811da177e4SLinus Torvalds 282e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 283e904e6c2STejun Heo 284bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 285bce90380STejun Heo /* possible CPUs of each node */ 286bce90380STejun Heo 287d55262c4STejun Heo static bool wq_disable_numa; 288d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 289d55262c4STejun Heo 290cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 291552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 292cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 293cee22a15SViresh Kumar 294863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 2953347fa09STejun Heo 296bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 297bce90380STejun Heo 2984c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2994c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3004c16bd32STejun Heo 30168e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3021258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 303a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 304d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 305d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3065bcab335STejun Heo 307e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30868e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3097d19c5ceSTejun Heo 310ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 311ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 312ef557180SMike Galbraith 313ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 314ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 315b05a7928SFrederic Weisbecker 316f303fccbSTejun Heo /* 317f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 318f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 319f303fccbSTejun Heo * to uncover usages which depend on it. 320f303fccbSTejun Heo */ 321f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 322f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 323f303fccbSTejun Heo #else 324f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 325f303fccbSTejun Heo #endif 326f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 327f303fccbSTejun Heo 3287d19c5ceSTejun Heo /* the per-cpu worker pools */ 32925528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3307d19c5ceSTejun Heo 33168e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3327d19c5ceSTejun Heo 33368e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 33429c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 33529c91e99STejun Heo 336c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33729c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33829c91e99STejun Heo 3398a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3408a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3418a2b7538STejun Heo 342d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 343ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 344044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3451aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 346044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 347d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 348044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 349f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 350044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 35124d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3520668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3530668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3540668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3550668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 356d320c038STejun Heo 3577d19c5ceSTejun Heo static int worker_thread(void *__worker); 3586ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 359c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 3607d19c5ceSTejun Heo 36197bd2347STejun Heo #define CREATE_TRACE_POINTS 36297bd2347STejun Heo #include <trace/events/workqueue.h> 36397bd2347STejun Heo 36468e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 36524acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 366f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36724acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 3685bcab335STejun Heo 3695b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 37024acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 371f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 372f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 37324acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 3745b95e1afSLai Jiangshan 375f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 376f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 377f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3787a62c2c8STejun Heo (pool)++) 3794ce62e9eSTejun Heo 38049e3cf44STejun Heo /** 38117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38217116969STejun Heo * @pool: iteration cursor 383611c92a0STejun Heo * @pi: integer used for iteration 384fa1b54e6STejun Heo * 38524acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 38668e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38768e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 388fa1b54e6STejun Heo * 389fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 390fa1b54e6STejun Heo * ignored. 39117116969STejun Heo */ 392611c92a0STejun Heo #define for_each_pool(pool, pi) \ 393611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39468e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 395fa1b54e6STejun Heo else 39617116969STejun Heo 39717116969STejun Heo /** 398822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 399822d8405STejun Heo * @worker: iteration cursor 400822d8405STejun Heo * @pool: worker_pool to iterate workers of 401822d8405STejun Heo * 4021258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 403822d8405STejun Heo * 404822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 405822d8405STejun Heo * ignored. 406822d8405STejun Heo */ 407da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 408da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4091258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 410822d8405STejun Heo else 411822d8405STejun Heo 412822d8405STejun Heo /** 41349e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41449e3cf44STejun Heo * @pwq: iteration cursor 41549e3cf44STejun Heo * @wq: the target workqueue 41676af4d93STejun Heo * 41724acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 418794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 419794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 42076af4d93STejun Heo * 42176af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42276af4d93STejun Heo * ignored. 42349e3cf44STejun Heo */ 42449e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42549e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4265a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 427f3421797STejun Heo 428dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 429dc186ad7SThomas Gleixner 430f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 431dc186ad7SThomas Gleixner 43299777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43399777288SStanislaw Gruszka { 43499777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43599777288SStanislaw Gruszka } 43699777288SStanislaw Gruszka 437b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 438b9fdac7fSDu, Changbin { 439b9fdac7fSDu, Changbin struct work_struct *work = addr; 440b9fdac7fSDu, Changbin 441b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 442b9fdac7fSDu, Changbin } 443b9fdac7fSDu, Changbin 444dc186ad7SThomas Gleixner /* 445dc186ad7SThomas Gleixner * fixup_init is called when: 446dc186ad7SThomas Gleixner * - an active object is initialized 447dc186ad7SThomas Gleixner */ 44802a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 449dc186ad7SThomas Gleixner { 450dc186ad7SThomas Gleixner struct work_struct *work = addr; 451dc186ad7SThomas Gleixner 452dc186ad7SThomas Gleixner switch (state) { 453dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 454dc186ad7SThomas Gleixner cancel_work_sync(work); 455dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45602a982a6SDu, Changbin return true; 457dc186ad7SThomas Gleixner default: 45802a982a6SDu, Changbin return false; 459dc186ad7SThomas Gleixner } 460dc186ad7SThomas Gleixner } 461dc186ad7SThomas Gleixner 462dc186ad7SThomas Gleixner /* 463dc186ad7SThomas Gleixner * fixup_free is called when: 464dc186ad7SThomas Gleixner * - an active object is freed 465dc186ad7SThomas Gleixner */ 46602a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 467dc186ad7SThomas Gleixner { 468dc186ad7SThomas Gleixner struct work_struct *work = addr; 469dc186ad7SThomas Gleixner 470dc186ad7SThomas Gleixner switch (state) { 471dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 472dc186ad7SThomas Gleixner cancel_work_sync(work); 473dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47402a982a6SDu, Changbin return true; 475dc186ad7SThomas Gleixner default: 47602a982a6SDu, Changbin return false; 477dc186ad7SThomas Gleixner } 478dc186ad7SThomas Gleixner } 479dc186ad7SThomas Gleixner 480f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 481dc186ad7SThomas Gleixner .name = "work_struct", 48299777288SStanislaw Gruszka .debug_hint = work_debug_hint, 483b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 484dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 485dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 486dc186ad7SThomas Gleixner }; 487dc186ad7SThomas Gleixner 488dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 489dc186ad7SThomas Gleixner { 490dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 491dc186ad7SThomas Gleixner } 492dc186ad7SThomas Gleixner 493dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 494dc186ad7SThomas Gleixner { 495dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 496dc186ad7SThomas Gleixner } 497dc186ad7SThomas Gleixner 498dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 499dc186ad7SThomas Gleixner { 500dc186ad7SThomas Gleixner if (onstack) 501dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 502dc186ad7SThomas Gleixner else 503dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 504dc186ad7SThomas Gleixner } 505dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 506dc186ad7SThomas Gleixner 507dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 508dc186ad7SThomas Gleixner { 509dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 510dc186ad7SThomas Gleixner } 511dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 512dc186ad7SThomas Gleixner 513ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 514ea2e64f2SThomas Gleixner { 515ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 516ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 517ea2e64f2SThomas Gleixner } 518ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 519ea2e64f2SThomas Gleixner 520dc186ad7SThomas Gleixner #else 521dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 522dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 523dc186ad7SThomas Gleixner #endif 524dc186ad7SThomas Gleixner 5254e8b22bdSLi Bin /** 5264e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5274e8b22bdSLi Bin * @pool: the pool pointer of interest 5284e8b22bdSLi Bin * 5294e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5304e8b22bdSLi Bin * successfully, -errno on failure. 5314e8b22bdSLi Bin */ 5329daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5339daf9e67STejun Heo { 5349daf9e67STejun Heo int ret; 5359daf9e67STejun Heo 53668e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5375bcab335STejun Heo 5384e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5394e8b22bdSLi Bin GFP_KERNEL); 540229641a6STejun Heo if (ret >= 0) { 541e68035fbSTejun Heo pool->id = ret; 542229641a6STejun Heo return 0; 543229641a6STejun Heo } 5449daf9e67STejun Heo return ret; 5459daf9e67STejun Heo } 5469daf9e67STejun Heo 54776af4d93STejun Heo /** 548df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 549df2d5ae4STejun Heo * @wq: the target workqueue 550df2d5ae4STejun Heo * @node: the node ID 551df2d5ae4STejun Heo * 55224acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 5535b95e1afSLai Jiangshan * read locked. 554df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 555df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 556d185af30SYacine Belkadi * 557d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 558df2d5ae4STejun Heo */ 559df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 560df2d5ae4STejun Heo int node) 561df2d5ae4STejun Heo { 5625b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 563d6e022f1STejun Heo 564d6e022f1STejun Heo /* 565d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 566d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 567d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 568d6e022f1STejun Heo * happens, this workaround can be removed. 569d6e022f1STejun Heo */ 570d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 571d6e022f1STejun Heo return wq->dfl_pwq; 572d6e022f1STejun Heo 573df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 574df2d5ae4STejun Heo } 575df2d5ae4STejun Heo 57673f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57773f53c4aSTejun Heo { 57873f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 57973f53c4aSTejun Heo } 58073f53c4aSTejun Heo 58173f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58273f53c4aSTejun Heo { 58373f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58473f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58573f53c4aSTejun Heo } 58673f53c4aSTejun Heo 58773f53c4aSTejun Heo static int work_next_color(int color) 58873f53c4aSTejun Heo { 58973f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 590a848e3b6SOleg Nesterov } 591a848e3b6SOleg Nesterov 5924594bf15SDavid Howells /* 593112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 594112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5957c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5967a22ad75STejun Heo * 597112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 598112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 599bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 600bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6017a22ad75STejun Heo * 602112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6037c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 604112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6057c3eed5cSTejun Heo * available only while the work item is queued. 606bbb68dfaSTejun Heo * 607bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 608bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 609bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 610bbb68dfaSTejun Heo * try to steal the PENDING bit. 6114594bf15SDavid Howells */ 6127a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6137a22ad75STejun Heo unsigned long flags) 6147a22ad75STejun Heo { 6156183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6167a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6177a22ad75STejun Heo } 6187a22ad75STejun Heo 619112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6204690c4abSTejun Heo unsigned long extra_flags) 621365970a1SDavid Howells { 622112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 623112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 624365970a1SDavid Howells } 625365970a1SDavid Howells 6264468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6274468a00fSLai Jiangshan int pool_id) 6284468a00fSLai Jiangshan { 6294468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6304468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6314468a00fSLai Jiangshan } 6324468a00fSLai Jiangshan 6337c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6347c3eed5cSTejun Heo int pool_id) 6354d707b9fSOleg Nesterov { 63623657bb1STejun Heo /* 63723657bb1STejun Heo * The following wmb is paired with the implied mb in 63823657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 63923657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64023657bb1STejun Heo * owner. 64123657bb1STejun Heo */ 64223657bb1STejun Heo smp_wmb(); 6437c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 644346c09f8SRoman Pen /* 645346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 646346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 647346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6488bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 649346c09f8SRoman Pen * the same @work. E.g. consider this case: 650346c09f8SRoman Pen * 651346c09f8SRoman Pen * CPU#0 CPU#1 652346c09f8SRoman Pen * ---------------------------- -------------------------------- 653346c09f8SRoman Pen * 654346c09f8SRoman Pen * 1 STORE event_indicated 655346c09f8SRoman Pen * 2 queue_work_on() { 656346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 657346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 658346c09f8SRoman Pen * 5 set_work_data() # clear bit 659346c09f8SRoman Pen * 6 smp_mb() 660346c09f8SRoman Pen * 7 work->current_func() { 661346c09f8SRoman Pen * 8 LOAD event_indicated 662346c09f8SRoman Pen * } 663346c09f8SRoman Pen * 664346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 665346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 666346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 667346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 668346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 669346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 670346c09f8SRoman Pen * before actual STORE. 671346c09f8SRoman Pen */ 672346c09f8SRoman Pen smp_mb(); 6734d707b9fSOleg Nesterov } 6744d707b9fSOleg Nesterov 6757a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 676365970a1SDavid Howells { 6777c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6787c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6797a22ad75STejun Heo } 6807a22ad75STejun Heo 681112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6827a22ad75STejun Heo { 683e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6847a22ad75STejun Heo 685112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 686e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 687e120153dSTejun Heo else 688e120153dSTejun Heo return NULL; 6897a22ad75STejun Heo } 6907a22ad75STejun Heo 6917c3eed5cSTejun Heo /** 6927c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6937c3eed5cSTejun Heo * @work: the work item of interest 6947c3eed5cSTejun Heo * 69568e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69624acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 69724acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 698fa1b54e6STejun Heo * 699fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 700fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 701fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 702fa1b54e6STejun Heo * returned pool is and stays online. 703d185af30SYacine Belkadi * 704d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7057c3eed5cSTejun Heo */ 7067c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7077a22ad75STejun Heo { 708e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7097c3eed5cSTejun Heo int pool_id; 7107a22ad75STejun Heo 71168e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 712fa1b54e6STejun Heo 713112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 714112202d9STejun Heo return ((struct pool_workqueue *) 7157c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7167a22ad75STejun Heo 7177c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7187c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7197a22ad75STejun Heo return NULL; 7207a22ad75STejun Heo 721fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7227c3eed5cSTejun Heo } 7237c3eed5cSTejun Heo 7247c3eed5cSTejun Heo /** 7257c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7267c3eed5cSTejun Heo * @work: the work item of interest 7277c3eed5cSTejun Heo * 728d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7297c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7307c3eed5cSTejun Heo */ 7317c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7327c3eed5cSTejun Heo { 73354d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7347c3eed5cSTejun Heo 735112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 736112202d9STejun Heo return ((struct pool_workqueue *) 73754d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 73854d5b7d0SLai Jiangshan 73954d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7407c3eed5cSTejun Heo } 7417c3eed5cSTejun Heo 742bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 743bbb68dfaSTejun Heo { 7447c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 745bbb68dfaSTejun Heo 7467c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7477c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 748bbb68dfaSTejun Heo } 749bbb68dfaSTejun Heo 750bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 751bbb68dfaSTejun Heo { 752bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 753bbb68dfaSTejun Heo 754112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 755bbb68dfaSTejun Heo } 756bbb68dfaSTejun Heo 757e22bee78STejun Heo /* 7583270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7593270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 760d565ed63STejun Heo * they're being called with pool->lock held. 761e22bee78STejun Heo */ 762e22bee78STejun Heo 76363d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 764649027d7STejun Heo { 765e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 766649027d7STejun Heo } 767649027d7STejun Heo 768e22bee78STejun Heo /* 769e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 770e22bee78STejun Heo * running workers. 771974271c4STejun Heo * 772974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 773706026c2STejun Heo * function will always return %true for unbound pools as long as the 774974271c4STejun Heo * worklist isn't empty. 775e22bee78STejun Heo */ 77663d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 777e22bee78STejun Heo { 77863d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 779e22bee78STejun Heo } 780e22bee78STejun Heo 781e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78263d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 783e22bee78STejun Heo { 78463d95a91STejun Heo return pool->nr_idle; 785e22bee78STejun Heo } 786e22bee78STejun Heo 787e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 78863d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 789e22bee78STejun Heo { 790e19e397aSTejun Heo return !list_empty(&pool->worklist) && 791e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 792e22bee78STejun Heo } 793e22bee78STejun Heo 794e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79563d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 796e22bee78STejun Heo { 79763d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 798e22bee78STejun Heo } 799e22bee78STejun Heo 800e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80163d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 802e22bee78STejun Heo { 803692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 80463d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80563d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 806e22bee78STejun Heo 807e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 808e22bee78STejun Heo } 809e22bee78STejun Heo 810e22bee78STejun Heo /* 811e22bee78STejun Heo * Wake up functions. 812e22bee78STejun Heo */ 813e22bee78STejun Heo 8141037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8151037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8167e11629dSTejun Heo { 81763d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8187e11629dSTejun Heo return NULL; 8197e11629dSTejun Heo 82063d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8217e11629dSTejun Heo } 8227e11629dSTejun Heo 8237e11629dSTejun Heo /** 8247e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82563d95a91STejun Heo * @pool: worker pool to wake worker from 8267e11629dSTejun Heo * 82763d95a91STejun Heo * Wake up the first idle worker of @pool. 8287e11629dSTejun Heo * 8297e11629dSTejun Heo * CONTEXT: 830a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8317e11629dSTejun Heo */ 83263d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8337e11629dSTejun Heo { 8341037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8357e11629dSTejun Heo 8367e11629dSTejun Heo if (likely(worker)) 8377e11629dSTejun Heo wake_up_process(worker->task); 8387e11629dSTejun Heo } 8397e11629dSTejun Heo 8404690c4abSTejun Heo /** 8416d25be57SThomas Gleixner * wq_worker_running - a worker is running again 842e22bee78STejun Heo * @task: task waking up 843e22bee78STejun Heo * 8446d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 845e22bee78STejun Heo */ 8466d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 847e22bee78STejun Heo { 848e22bee78STejun Heo struct worker *worker = kthread_data(task); 849e22bee78STejun Heo 8506d25be57SThomas Gleixner if (!worker->sleeping) 8516d25be57SThomas Gleixner return; 8526d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 853e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 8546d25be57SThomas Gleixner worker->sleeping = 0; 85536576000SJoonsoo Kim } 856e22bee78STejun Heo 857e22bee78STejun Heo /** 858e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 859e22bee78STejun Heo * @task: task going to sleep 860e22bee78STejun Heo * 8616d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 86262849a96SSebastian Andrzej Siewior * going to sleep. Preemption needs to be disabled to protect ->sleeping 86362849a96SSebastian Andrzej Siewior * assignment. 864e22bee78STejun Heo */ 8656d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 866e22bee78STejun Heo { 8676d25be57SThomas Gleixner struct worker *next, *worker = kthread_data(task); 868111c225aSTejun Heo struct worker_pool *pool; 869e22bee78STejun Heo 870111c225aSTejun Heo /* 871111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 872111c225aSTejun Heo * workers, also reach here, let's not access anything before 873111c225aSTejun Heo * checking NOT_RUNNING. 874111c225aSTejun Heo */ 8752d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 8766d25be57SThomas Gleixner return; 877e22bee78STejun Heo 878111c225aSTejun Heo pool = worker->pool; 879111c225aSTejun Heo 88062849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 88162849a96SSebastian Andrzej Siewior if (worker->sleeping) 8826d25be57SThomas Gleixner return; 8836d25be57SThomas Gleixner 8846d25be57SThomas Gleixner worker->sleeping = 1; 885a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 886e22bee78STejun Heo 887e22bee78STejun Heo /* 888e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 889e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 890e22bee78STejun Heo * Please read comment there. 891e22bee78STejun Heo * 892628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 893628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 894628c78e7STejun Heo * disabled, which in turn means that none else could be 895d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 896628c78e7STejun Heo * lock is safe. 897e22bee78STejun Heo */ 898e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 8996d25be57SThomas Gleixner !list_empty(&pool->worklist)) { 9006d25be57SThomas Gleixner next = first_idle_worker(pool); 9016d25be57SThomas Gleixner if (next) 9026d25be57SThomas Gleixner wake_up_process(next->task); 9036d25be57SThomas Gleixner } 904a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 905e22bee78STejun Heo } 906e22bee78STejun Heo 907e22bee78STejun Heo /** 9081b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 9098194fe94SBart Van Assche * @task: Task to retrieve last work function of. 9101b69ac6bSJohannes Weiner * 9111b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 9121b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 9131b69ac6bSJohannes Weiner * 9141b69ac6bSJohannes Weiner * CONTEXT: 915a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 9161b69ac6bSJohannes Weiner * 9174b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 9184b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 9194b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 9204b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 9214b047002SJohannes Weiner * 9224b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 9234b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 9244b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 9254b047002SJohannes Weiner * is guaranteed to not be processing any works. 9264b047002SJohannes Weiner * 9271b69ac6bSJohannes Weiner * Return: 9281b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 9291b69ac6bSJohannes Weiner * hasn't executed any work yet. 9301b69ac6bSJohannes Weiner */ 9311b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 9321b69ac6bSJohannes Weiner { 9331b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 9341b69ac6bSJohannes Weiner 9351b69ac6bSJohannes Weiner return worker->last_func; 9361b69ac6bSJohannes Weiner } 9371b69ac6bSJohannes Weiner 9381b69ac6bSJohannes Weiner /** 939e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 940cb444766STejun Heo * @worker: self 941d302f017STejun Heo * @flags: flags to set 942d302f017STejun Heo * 943228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 944d302f017STejun Heo * 945cb444766STejun Heo * CONTEXT: 946a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 947d302f017STejun Heo */ 948228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 949d302f017STejun Heo { 950bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 951e22bee78STejun Heo 952cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 953cb444766STejun Heo 954228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 955e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 956e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 957e19e397aSTejun Heo atomic_dec(&pool->nr_running); 958e22bee78STejun Heo } 959e22bee78STejun Heo 960d302f017STejun Heo worker->flags |= flags; 961d302f017STejun Heo } 962d302f017STejun Heo 963d302f017STejun Heo /** 964e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 965cb444766STejun Heo * @worker: self 966d302f017STejun Heo * @flags: flags to clear 967d302f017STejun Heo * 968e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 969d302f017STejun Heo * 970cb444766STejun Heo * CONTEXT: 971a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 972d302f017STejun Heo */ 973d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 974d302f017STejun Heo { 97563d95a91STejun Heo struct worker_pool *pool = worker->pool; 976e22bee78STejun Heo unsigned int oflags = worker->flags; 977e22bee78STejun Heo 978cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 979cb444766STejun Heo 980d302f017STejun Heo worker->flags &= ~flags; 981e22bee78STejun Heo 98242c025f3STejun Heo /* 98342c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 98442c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 98542c025f3STejun Heo * of multiple flags, not a single flag. 98642c025f3STejun Heo */ 987e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 988e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 989e19e397aSTejun Heo atomic_inc(&pool->nr_running); 990d302f017STejun Heo } 991d302f017STejun Heo 992d302f017STejun Heo /** 9938cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 994c9e7cf27STejun Heo * @pool: pool of interest 9958cca0eeaSTejun Heo * @work: work to find worker for 9968cca0eeaSTejun Heo * 997c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 998c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 999a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1000a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1001a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1002a2c1c57bSTejun Heo * being executed. 1003a2c1c57bSTejun Heo * 1004a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1005a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1006a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1007a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1008a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1009a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1010a2c1c57bSTejun Heo * 1011c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1012c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1013c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1014c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1015c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1016c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10178cca0eeaSTejun Heo * 10188cca0eeaSTejun Heo * CONTEXT: 1019a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 10208cca0eeaSTejun Heo * 1021d185af30SYacine Belkadi * Return: 1022d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10238cca0eeaSTejun Heo * otherwise. 10248cca0eeaSTejun Heo */ 1025c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10268cca0eeaSTejun Heo struct work_struct *work) 10278cca0eeaSTejun Heo { 102842f8570fSSasha Levin struct worker *worker; 102942f8570fSSasha Levin 1030b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1031a2c1c57bSTejun Heo (unsigned long)work) 1032a2c1c57bSTejun Heo if (worker->current_work == work && 1033a2c1c57bSTejun Heo worker->current_func == work->func) 103442f8570fSSasha Levin return worker; 103542f8570fSSasha Levin 103642f8570fSSasha Levin return NULL; 10378cca0eeaSTejun Heo } 10388cca0eeaSTejun Heo 10398cca0eeaSTejun Heo /** 1040bf4ede01STejun Heo * move_linked_works - move linked works to a list 1041bf4ede01STejun Heo * @work: start of series of works to be scheduled 1042bf4ede01STejun Heo * @head: target list to append @work to 1043402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1044bf4ede01STejun Heo * 1045bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1046bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1047bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1048bf4ede01STejun Heo * 1049bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1050bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1051bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1052bf4ede01STejun Heo * 1053bf4ede01STejun Heo * CONTEXT: 1054a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1055bf4ede01STejun Heo */ 1056bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1057bf4ede01STejun Heo struct work_struct **nextp) 1058bf4ede01STejun Heo { 1059bf4ede01STejun Heo struct work_struct *n; 1060bf4ede01STejun Heo 1061bf4ede01STejun Heo /* 1062bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1063bf4ede01STejun Heo * use NULL for list head. 1064bf4ede01STejun Heo */ 1065bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1066bf4ede01STejun Heo list_move_tail(&work->entry, head); 1067bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1068bf4ede01STejun Heo break; 1069bf4ede01STejun Heo } 1070bf4ede01STejun Heo 1071bf4ede01STejun Heo /* 1072bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1073bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1074bf4ede01STejun Heo * needs to be updated. 1075bf4ede01STejun Heo */ 1076bf4ede01STejun Heo if (nextp) 1077bf4ede01STejun Heo *nextp = n; 1078bf4ede01STejun Heo } 1079bf4ede01STejun Heo 10808864b4e5STejun Heo /** 10818864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10828864b4e5STejun Heo * @pwq: pool_workqueue to get 10838864b4e5STejun Heo * 10848864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10858864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10868864b4e5STejun Heo */ 10878864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10888864b4e5STejun Heo { 10898864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10908864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10918864b4e5STejun Heo pwq->refcnt++; 10928864b4e5STejun Heo } 10938864b4e5STejun Heo 10948864b4e5STejun Heo /** 10958864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10968864b4e5STejun Heo * @pwq: pool_workqueue to put 10978864b4e5STejun Heo * 10988864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10998864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11008864b4e5STejun Heo */ 11018864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11028864b4e5STejun Heo { 11038864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11048864b4e5STejun Heo if (likely(--pwq->refcnt)) 11058864b4e5STejun Heo return; 11068864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11078864b4e5STejun Heo return; 11088864b4e5STejun Heo /* 11098864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11108864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11118864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11128864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11138864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11148864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11158864b4e5STejun Heo */ 11168864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11178864b4e5STejun Heo } 11188864b4e5STejun Heo 1119dce90d47STejun Heo /** 1120dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1121dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1122dce90d47STejun Heo * 1123dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1124dce90d47STejun Heo */ 1125dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1126dce90d47STejun Heo { 1127dce90d47STejun Heo if (pwq) { 1128dce90d47STejun Heo /* 112924acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1130dce90d47STejun Heo * following lock operations are safe. 1131dce90d47STejun Heo */ 1132a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1133dce90d47STejun Heo put_pwq(pwq); 1134a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1135dce90d47STejun Heo } 1136dce90d47STejun Heo } 1137dce90d47STejun Heo 1138112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1139bf4ede01STejun Heo { 1140112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1141bf4ede01STejun Heo 1142bf4ede01STejun Heo trace_workqueue_activate_work(work); 114382607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 114482607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1145112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1146bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1147112202d9STejun Heo pwq->nr_active++; 1148bf4ede01STejun Heo } 1149bf4ede01STejun Heo 1150112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11513aa62497SLai Jiangshan { 1152112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11533aa62497SLai Jiangshan struct work_struct, entry); 11543aa62497SLai Jiangshan 1155112202d9STejun Heo pwq_activate_delayed_work(work); 11563aa62497SLai Jiangshan } 11573aa62497SLai Jiangshan 1158bf4ede01STejun Heo /** 1159112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1160112202d9STejun Heo * @pwq: pwq of interest 1161bf4ede01STejun Heo * @color: color of work which left the queue 1162bf4ede01STejun Heo * 1163bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1164112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1165bf4ede01STejun Heo * 1166bf4ede01STejun Heo * CONTEXT: 1167a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1168bf4ede01STejun Heo */ 1169112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1170bf4ede01STejun Heo { 11718864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1172bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11738864b4e5STejun Heo goto out_put; 1174bf4ede01STejun Heo 1175112202d9STejun Heo pwq->nr_in_flight[color]--; 1176bf4ede01STejun Heo 1177112202d9STejun Heo pwq->nr_active--; 1178112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1179bf4ede01STejun Heo /* one down, submit a delayed one */ 1180112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1181112202d9STejun Heo pwq_activate_first_delayed(pwq); 1182bf4ede01STejun Heo } 1183bf4ede01STejun Heo 1184bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1185112202d9STejun Heo if (likely(pwq->flush_color != color)) 11868864b4e5STejun Heo goto out_put; 1187bf4ede01STejun Heo 1188bf4ede01STejun Heo /* are there still in-flight works? */ 1189112202d9STejun Heo if (pwq->nr_in_flight[color]) 11908864b4e5STejun Heo goto out_put; 1191bf4ede01STejun Heo 1192112202d9STejun Heo /* this pwq is done, clear flush_color */ 1193112202d9STejun Heo pwq->flush_color = -1; 1194bf4ede01STejun Heo 1195bf4ede01STejun Heo /* 1196112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1197bf4ede01STejun Heo * will handle the rest. 1198bf4ede01STejun Heo */ 1199112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1200112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12018864b4e5STejun Heo out_put: 12028864b4e5STejun Heo put_pwq(pwq); 1203bf4ede01STejun Heo } 1204bf4ede01STejun Heo 120536e227d2STejun Heo /** 1206bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 120736e227d2STejun Heo * @work: work item to steal 120836e227d2STejun Heo * @is_dwork: @work is a delayed_work 1209bbb68dfaSTejun Heo * @flags: place to store irq state 121036e227d2STejun Heo * 121136e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1212d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 121336e227d2STejun Heo * 1214d185af30SYacine Belkadi * Return: 12153eb6b31bSMauro Carvalho Chehab * 12163eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 121736e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 121836e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 121936e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1220bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1221bbb68dfaSTejun Heo * for arbitrarily long 12223eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 122336e227d2STejun Heo * 1224d185af30SYacine Belkadi * Note: 1225bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1226e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1227e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1228e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1229bbb68dfaSTejun Heo * 1230bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1231bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1232bbb68dfaSTejun Heo * 1233e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1234bf4ede01STejun Heo */ 1235bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1236bbb68dfaSTejun Heo unsigned long *flags) 1237bf4ede01STejun Heo { 1238d565ed63STejun Heo struct worker_pool *pool; 1239112202d9STejun Heo struct pool_workqueue *pwq; 1240bf4ede01STejun Heo 1241bbb68dfaSTejun Heo local_irq_save(*flags); 1242bbb68dfaSTejun Heo 124336e227d2STejun Heo /* try to steal the timer if it exists */ 124436e227d2STejun Heo if (is_dwork) { 124536e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 124636e227d2STejun Heo 1247e0aecdd8STejun Heo /* 1248e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1249e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1250e0aecdd8STejun Heo * running on the local CPU. 1251e0aecdd8STejun Heo */ 125236e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 125336e227d2STejun Heo return 1; 125436e227d2STejun Heo } 125536e227d2STejun Heo 125636e227d2STejun Heo /* try to claim PENDING the normal way */ 1257bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1258bf4ede01STejun Heo return 0; 1259bf4ede01STejun Heo 126024acfb71SThomas Gleixner rcu_read_lock(); 1261bf4ede01STejun Heo /* 1262bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1263bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1264bf4ede01STejun Heo */ 1265d565ed63STejun Heo pool = get_work_pool(work); 1266d565ed63STejun Heo if (!pool) 1267bbb68dfaSTejun Heo goto fail; 1268bf4ede01STejun Heo 1269a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1270bf4ede01STejun Heo /* 1271112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1272112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1273112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1274112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1275112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12760b3dae68SLai Jiangshan * item is currently queued on that pool. 1277bf4ede01STejun Heo */ 1278112202d9STejun Heo pwq = get_work_pwq(work); 1279112202d9STejun Heo if (pwq && pwq->pool == pool) { 1280bf4ede01STejun Heo debug_work_deactivate(work); 12813aa62497SLai Jiangshan 12823aa62497SLai Jiangshan /* 128316062836STejun Heo * A delayed work item cannot be grabbed directly because 128416062836STejun Heo * it might have linked NO_COLOR work items which, if left 1285112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 128616062836STejun Heo * management later on and cause stall. Make sure the work 128716062836STejun Heo * item is activated before grabbing. 12883aa62497SLai Jiangshan */ 12893aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1290112202d9STejun Heo pwq_activate_delayed_work(work); 12913aa62497SLai Jiangshan 1292bf4ede01STejun Heo list_del_init(&work->entry); 12939c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 129436e227d2STejun Heo 1295112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12964468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12974468a00fSLai Jiangshan 1298a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 129924acfb71SThomas Gleixner rcu_read_unlock(); 130036e227d2STejun Heo return 1; 1301bf4ede01STejun Heo } 1302a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1303bbb68dfaSTejun Heo fail: 130424acfb71SThomas Gleixner rcu_read_unlock(); 1305bbb68dfaSTejun Heo local_irq_restore(*flags); 1306bbb68dfaSTejun Heo if (work_is_canceling(work)) 1307bbb68dfaSTejun Heo return -ENOENT; 1308bbb68dfaSTejun Heo cpu_relax(); 130936e227d2STejun Heo return -EAGAIN; 1310bf4ede01STejun Heo } 1311bf4ede01STejun Heo 1312bf4ede01STejun Heo /** 1313706026c2STejun Heo * insert_work - insert a work into a pool 1314112202d9STejun Heo * @pwq: pwq @work belongs to 13154690c4abSTejun Heo * @work: work to insert 13164690c4abSTejun Heo * @head: insertion point 13174690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13184690c4abSTejun Heo * 1319112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1320706026c2STejun Heo * work_struct flags. 13214690c4abSTejun Heo * 13224690c4abSTejun Heo * CONTEXT: 1323a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1324365970a1SDavid Howells */ 1325112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1326112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1327b89deed3SOleg Nesterov { 1328112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1329e1d8aa9fSFrederic Weisbecker 13304690c4abSTejun Heo /* we own @work, set data and link */ 1331112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13321a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13338864b4e5STejun Heo get_pwq(pwq); 1334e22bee78STejun Heo 1335e22bee78STejun Heo /* 1336c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1337c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1338c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1339e22bee78STejun Heo */ 1340e22bee78STejun Heo smp_mb(); 1341e22bee78STejun Heo 134263d95a91STejun Heo if (__need_more_worker(pool)) 134363d95a91STejun Heo wake_up_worker(pool); 1344b89deed3SOleg Nesterov } 1345b89deed3SOleg Nesterov 1346c8efcc25STejun Heo /* 1347c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13488d03ecfeSTejun Heo * same workqueue. 1349c8efcc25STejun Heo */ 1350c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1351c8efcc25STejun Heo { 1352c8efcc25STejun Heo struct worker *worker; 1353c8efcc25STejun Heo 13548d03ecfeSTejun Heo worker = current_wq_worker(); 1355c8efcc25STejun Heo /* 1356bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 13578d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1358c8efcc25STejun Heo */ 1359112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1360c8efcc25STejun Heo } 1361c8efcc25STejun Heo 1362ef557180SMike Galbraith /* 1363ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1364ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1365ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1366ef557180SMike Galbraith */ 1367ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1368ef557180SMike Galbraith { 1369f303fccbSTejun Heo static bool printed_dbg_warning; 1370ef557180SMike Galbraith int new_cpu; 1371ef557180SMike Galbraith 1372f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1373ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1374ef557180SMike Galbraith return cpu; 1375f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1376f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1377f303fccbSTejun Heo printed_dbg_warning = true; 1378f303fccbSTejun Heo } 1379f303fccbSTejun Heo 1380ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1381ef557180SMike Galbraith return cpu; 1382ef557180SMike Galbraith 1383ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1384ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1385ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1386ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1387ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1388ef557180SMike Galbraith return cpu; 1389ef557180SMike Galbraith } 1390ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1391ef557180SMike Galbraith 1392ef557180SMike Galbraith return new_cpu; 1393ef557180SMike Galbraith } 1394ef557180SMike Galbraith 1395d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13961da177e4SLinus Torvalds struct work_struct *work) 13971da177e4SLinus Torvalds { 1398112202d9STejun Heo struct pool_workqueue *pwq; 1399c9178087STejun Heo struct worker_pool *last_pool; 14001e19ffc6STejun Heo struct list_head *worklist; 14018a2e8e5dSTejun Heo unsigned int work_flags; 1402b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 14038930cabaSTejun Heo 14048930cabaSTejun Heo /* 14058930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 14068930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 14078930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 14088930cabaSTejun Heo * happen with IRQ disabled. 14098930cabaSTejun Heo */ 14108e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 14111da177e4SLinus Torvalds 1412dc186ad7SThomas Gleixner debug_work_activate(work); 14131e19ffc6STejun Heo 14149ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1415618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1416c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1417e41e704bSTejun Heo return; 141824acfb71SThomas Gleixner rcu_read_lock(); 14199e8cd2f5STejun Heo retry: 1420aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1421aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1422df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1423ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1424df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1425aa202f1fSHillf Danton } else { 1426aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1427aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1428aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1429aa202f1fSHillf Danton } 1430f3421797STejun Heo 143118aa9effSTejun Heo /* 1432c9178087STejun Heo * If @work was previously on a different pool, it might still be 1433c9178087STejun Heo * running there, in which case the work needs to be queued on that 1434c9178087STejun Heo * pool to guarantee non-reentrancy. 143518aa9effSTejun Heo */ 1436c9e7cf27STejun Heo last_pool = get_work_pool(work); 1437112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 143818aa9effSTejun Heo struct worker *worker; 143918aa9effSTejun Heo 1440a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 144118aa9effSTejun Heo 1442c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 144318aa9effSTejun Heo 1444112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1445c9178087STejun Heo pwq = worker->current_pwq; 14468594fadeSLai Jiangshan } else { 144718aa9effSTejun Heo /* meh... not running there, queue here */ 1448a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1449a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 145018aa9effSTejun Heo } 14518930cabaSTejun Heo } else { 1452a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 14538930cabaSTejun Heo } 1454502ca9d8STejun Heo 14559e8cd2f5STejun Heo /* 14569e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14579e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14589e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1459df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1460df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14619e8cd2f5STejun Heo * make forward-progress. 14629e8cd2f5STejun Heo */ 14639e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14649e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1465a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 14669e8cd2f5STejun Heo cpu_relax(); 14679e8cd2f5STejun Heo goto retry; 14689e8cd2f5STejun Heo } 14699e8cd2f5STejun Heo /* oops */ 14709e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14719e8cd2f5STejun Heo wq->name, cpu); 14729e8cd2f5STejun Heo } 14739e8cd2f5STejun Heo 1474112202d9STejun Heo /* pwq determined, queue */ 1475112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1476502ca9d8STejun Heo 147724acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 147824acfb71SThomas Gleixner goto out; 14791e19ffc6STejun Heo 1480112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1481112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14821e19ffc6STejun Heo 1483112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1484cdadf009STejun Heo trace_workqueue_activate_work(work); 1485112202d9STejun Heo pwq->nr_active++; 1486112202d9STejun Heo worklist = &pwq->pool->worklist; 148782607adcSTejun Heo if (list_empty(worklist)) 148882607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14898a2e8e5dSTejun Heo } else { 14908a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1491112202d9STejun Heo worklist = &pwq->delayed_works; 14928a2e8e5dSTejun Heo } 14931e19ffc6STejun Heo 1494112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14951e19ffc6STejun Heo 149624acfb71SThomas Gleixner out: 1497a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 149824acfb71SThomas Gleixner rcu_read_unlock(); 14991da177e4SLinus Torvalds } 15001da177e4SLinus Torvalds 15010fcb78c2SRolf Eike Beer /** 1502c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1503c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1504c1a220e7SZhang Rui * @wq: workqueue to use 1505c1a220e7SZhang Rui * @work: work to queue 1506c1a220e7SZhang Rui * 1507c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1508c1a220e7SZhang Rui * can't go away. 1509d185af30SYacine Belkadi * 1510d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1511c1a220e7SZhang Rui */ 1512d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1513d4283e93STejun Heo struct work_struct *work) 1514c1a220e7SZhang Rui { 1515d4283e93STejun Heo bool ret = false; 15168930cabaSTejun Heo unsigned long flags; 15178930cabaSTejun Heo 15188930cabaSTejun Heo local_irq_save(flags); 1519c1a220e7SZhang Rui 152022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15214690c4abSTejun Heo __queue_work(cpu, wq, work); 1522d4283e93STejun Heo ret = true; 1523c1a220e7SZhang Rui } 15248930cabaSTejun Heo 15258930cabaSTejun Heo local_irq_restore(flags); 1526c1a220e7SZhang Rui return ret; 1527c1a220e7SZhang Rui } 1528ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1529c1a220e7SZhang Rui 15308204e0c1SAlexander Duyck /** 15318204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15328204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15338204e0c1SAlexander Duyck * 15348204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15358204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15368204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15378204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15388204e0c1SAlexander Duyck */ 15398204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15408204e0c1SAlexander Duyck { 15418204e0c1SAlexander Duyck int cpu; 15428204e0c1SAlexander Duyck 15438204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15448204e0c1SAlexander Duyck if (!wq_numa_enabled) 15458204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15468204e0c1SAlexander Duyck 15478204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15488204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15498204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15508204e0c1SAlexander Duyck 15518204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15528204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15538204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15548204e0c1SAlexander Duyck return cpu; 15558204e0c1SAlexander Duyck 15568204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15578204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15588204e0c1SAlexander Duyck 15598204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15608204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15618204e0c1SAlexander Duyck } 15628204e0c1SAlexander Duyck 15638204e0c1SAlexander Duyck /** 15648204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15658204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15668204e0c1SAlexander Duyck * @wq: workqueue to use 15678204e0c1SAlexander Duyck * @work: work to queue 15688204e0c1SAlexander Duyck * 15698204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 15708204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 15718204e0c1SAlexander Duyck * NUMA node. 15728204e0c1SAlexander Duyck * 15738204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 15748204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 15758204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 15768204e0c1SAlexander Duyck * 15778204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 15788204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 15798204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 15808204e0c1SAlexander Duyck * 15818204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 15828204e0c1SAlexander Duyck */ 15838204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 15848204e0c1SAlexander Duyck struct work_struct *work) 15858204e0c1SAlexander Duyck { 15868204e0c1SAlexander Duyck unsigned long flags; 15878204e0c1SAlexander Duyck bool ret = false; 15888204e0c1SAlexander Duyck 15898204e0c1SAlexander Duyck /* 15908204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 15918204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 15928204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 15938204e0c1SAlexander Duyck * 15948204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 15958204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 15968204e0c1SAlexander Duyck * some round robin type logic. 15978204e0c1SAlexander Duyck */ 15988204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 15998204e0c1SAlexander Duyck 16008204e0c1SAlexander Duyck local_irq_save(flags); 16018204e0c1SAlexander Duyck 16028204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16038204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16048204e0c1SAlexander Duyck 16058204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16068204e0c1SAlexander Duyck ret = true; 16078204e0c1SAlexander Duyck } 16088204e0c1SAlexander Duyck 16098204e0c1SAlexander Duyck local_irq_restore(flags); 16108204e0c1SAlexander Duyck return ret; 16118204e0c1SAlexander Duyck } 16128204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16138204e0c1SAlexander Duyck 16148c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16151da177e4SLinus Torvalds { 16168c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16171da177e4SLinus Torvalds 1618e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 161960c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16201da177e4SLinus Torvalds } 16211438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16221da177e4SLinus Torvalds 16237beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 162452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16251da177e4SLinus Torvalds { 16267beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16277beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16281da177e4SLinus Torvalds 1629637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 1630841b86f3SKees Cook WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1631fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1632fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16337beb2edfSTejun Heo 16348852aac2STejun Heo /* 16358852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16368852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16378852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16388852aac2STejun Heo * on that there's no such delay when @delay is 0. 16398852aac2STejun Heo */ 16408852aac2STejun Heo if (!delay) { 16418852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16428852aac2STejun Heo return; 16438852aac2STejun Heo } 16448852aac2STejun Heo 164560c057bcSLai Jiangshan dwork->wq = wq; 16461265057fSTejun Heo dwork->cpu = cpu; 16477beb2edfSTejun Heo timer->expires = jiffies + delay; 16487beb2edfSTejun Heo 1649041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16507beb2edfSTejun Heo add_timer_on(timer, cpu); 1651041bd12eSTejun Heo else 1652041bd12eSTejun Heo add_timer(timer); 16537beb2edfSTejun Heo } 16541da177e4SLinus Torvalds 16550fcb78c2SRolf Eike Beer /** 16560fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16570fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16580fcb78c2SRolf Eike Beer * @wq: workqueue to use 1659af9997e4SRandy Dunlap * @dwork: work to queue 16600fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16610fcb78c2SRolf Eike Beer * 1662d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1663715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1664715f1300STejun Heo * execution. 16650fcb78c2SRolf Eike Beer */ 1666d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 166752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16687a6bc1cdSVenkatesh Pallipadi { 166952bad64dSDavid Howells struct work_struct *work = &dwork->work; 1670d4283e93STejun Heo bool ret = false; 16718930cabaSTejun Heo unsigned long flags; 16728930cabaSTejun Heo 16738930cabaSTejun Heo /* read the comment in __queue_work() */ 16748930cabaSTejun Heo local_irq_save(flags); 16757a6bc1cdSVenkatesh Pallipadi 167622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16777beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1678d4283e93STejun Heo ret = true; 16797a6bc1cdSVenkatesh Pallipadi } 16808930cabaSTejun Heo 16818930cabaSTejun Heo local_irq_restore(flags); 16827a6bc1cdSVenkatesh Pallipadi return ret; 16837a6bc1cdSVenkatesh Pallipadi } 1684ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 16851da177e4SLinus Torvalds 1686c8e55f36STejun Heo /** 16878376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 16888376fe22STejun Heo * @cpu: CPU number to execute work on 16898376fe22STejun Heo * @wq: workqueue to use 16908376fe22STejun Heo * @dwork: work to queue 16918376fe22STejun Heo * @delay: number of jiffies to wait before queueing 16928376fe22STejun Heo * 16938376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 16948376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 16958376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 16968376fe22STejun Heo * current state. 16978376fe22STejun Heo * 1698d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 16998376fe22STejun Heo * pending and its timer was modified. 17008376fe22STejun Heo * 1701e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17028376fe22STejun Heo * See try_to_grab_pending() for details. 17038376fe22STejun Heo */ 17048376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17058376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17068376fe22STejun Heo { 17078376fe22STejun Heo unsigned long flags; 17088376fe22STejun Heo int ret; 17098376fe22STejun Heo 17108376fe22STejun Heo do { 17118376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17128376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17138376fe22STejun Heo 17148376fe22STejun Heo if (likely(ret >= 0)) { 17158376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17168376fe22STejun Heo local_irq_restore(flags); 17178376fe22STejun Heo } 17188376fe22STejun Heo 17198376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17208376fe22STejun Heo return ret; 17218376fe22STejun Heo } 17228376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17238376fe22STejun Heo 172405f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 172505f0fe6bSTejun Heo { 172605f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 172705f0fe6bSTejun Heo 172805f0fe6bSTejun Heo /* read the comment in __queue_work() */ 172905f0fe6bSTejun Heo local_irq_disable(); 173005f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 173105f0fe6bSTejun Heo local_irq_enable(); 173205f0fe6bSTejun Heo } 173305f0fe6bSTejun Heo 173405f0fe6bSTejun Heo /** 173505f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 173605f0fe6bSTejun Heo * @wq: workqueue to use 173705f0fe6bSTejun Heo * @rwork: work to queue 173805f0fe6bSTejun Heo * 173905f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 174005f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1741bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 174205f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 174305f0fe6bSTejun Heo */ 174405f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 174505f0fe6bSTejun Heo { 174605f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 174705f0fe6bSTejun Heo 174805f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 174905f0fe6bSTejun Heo rwork->wq = wq; 175005f0fe6bSTejun Heo call_rcu(&rwork->rcu, rcu_work_rcufn); 175105f0fe6bSTejun Heo return true; 175205f0fe6bSTejun Heo } 175305f0fe6bSTejun Heo 175405f0fe6bSTejun Heo return false; 175505f0fe6bSTejun Heo } 175605f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 175705f0fe6bSTejun Heo 17588376fe22STejun Heo /** 1759c8e55f36STejun Heo * worker_enter_idle - enter idle state 1760c8e55f36STejun Heo * @worker: worker which is entering idle state 1761c8e55f36STejun Heo * 1762c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1763c8e55f36STejun Heo * necessary. 1764c8e55f36STejun Heo * 1765c8e55f36STejun Heo * LOCKING: 1766a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1767c8e55f36STejun Heo */ 1768c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 17691da177e4SLinus Torvalds { 1770bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1771c8e55f36STejun Heo 17726183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 17736183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 17746183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 17756183c009STejun Heo return; 1776c8e55f36STejun Heo 1777051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1778cb444766STejun Heo worker->flags |= WORKER_IDLE; 1779bd7bdd43STejun Heo pool->nr_idle++; 1780e22bee78STejun Heo worker->last_active = jiffies; 1781c8e55f36STejun Heo 1782c8e55f36STejun Heo /* idle_list is LIFO */ 1783bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1784db7bccf4STejun Heo 178563d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1786628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1787cb444766STejun Heo 1788544ecf31STejun Heo /* 1789e8b3f8dbSLai Jiangshan * Sanity check nr_running. Because unbind_workers() releases 1790d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1791628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1792628c78e7STejun Heo * unbind is not in progress. 1793544ecf31STejun Heo */ 179424647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1795bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1796e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1797c8e55f36STejun Heo } 1798c8e55f36STejun Heo 1799c8e55f36STejun Heo /** 1800c8e55f36STejun Heo * worker_leave_idle - leave idle state 1801c8e55f36STejun Heo * @worker: worker which is leaving idle state 1802c8e55f36STejun Heo * 1803c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1804c8e55f36STejun Heo * 1805c8e55f36STejun Heo * LOCKING: 1806a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1807c8e55f36STejun Heo */ 1808c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1809c8e55f36STejun Heo { 1810bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1811c8e55f36STejun Heo 18126183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18136183c009STejun Heo return; 1814d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1815bd7bdd43STejun Heo pool->nr_idle--; 1816c8e55f36STejun Heo list_del_init(&worker->entry); 1817c8e55f36STejun Heo } 1818c8e55f36STejun Heo 1819f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1820c34056a3STejun Heo { 1821c34056a3STejun Heo struct worker *worker; 1822c34056a3STejun Heo 1823f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1824c8e55f36STejun Heo if (worker) { 1825c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1826affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1827da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1828e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1829e22bee78STejun Heo worker->flags = WORKER_PREP; 1830c8e55f36STejun Heo } 1831c34056a3STejun Heo return worker; 1832c34056a3STejun Heo } 1833c34056a3STejun Heo 1834c34056a3STejun Heo /** 18354736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18364736cbf7SLai Jiangshan * @worker: worker to be attached 18374736cbf7SLai Jiangshan * @pool: the target pool 18384736cbf7SLai Jiangshan * 18394736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18404736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18414736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18424736cbf7SLai Jiangshan */ 18434736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18444736cbf7SLai Jiangshan struct worker_pool *pool) 18454736cbf7SLai Jiangshan { 18461258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18474736cbf7SLai Jiangshan 18484736cbf7SLai Jiangshan /* 18494736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 18504736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 18514736cbf7SLai Jiangshan */ 18524736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 18534736cbf7SLai Jiangshan 18544736cbf7SLai Jiangshan /* 18551258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18561258fae7STejun Heo * stable across this function. See the comments above the flag 18571258fae7STejun Heo * definition for details. 18584736cbf7SLai Jiangshan */ 18594736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 18604736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 18614736cbf7SLai Jiangshan 18624736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1863a2d812a2STejun Heo worker->pool = pool; 18644736cbf7SLai Jiangshan 18651258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 18664736cbf7SLai Jiangshan } 18674736cbf7SLai Jiangshan 18684736cbf7SLai Jiangshan /** 186960f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 187060f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 187160f5a4bcSLai Jiangshan * 18724736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 18734736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 18744736cbf7SLai Jiangshan * other reference to the pool. 187560f5a4bcSLai Jiangshan */ 1876a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 187760f5a4bcSLai Jiangshan { 1878a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 187960f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 188060f5a4bcSLai Jiangshan 18811258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1882a2d812a2STejun Heo 1883da028469SLai Jiangshan list_del(&worker->node); 1884a2d812a2STejun Heo worker->pool = NULL; 1885a2d812a2STejun Heo 1886da028469SLai Jiangshan if (list_empty(&pool->workers)) 188760f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 18881258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 188960f5a4bcSLai Jiangshan 1890b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1891b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1892b62c0751SLai Jiangshan 189360f5a4bcSLai Jiangshan if (detach_completion) 189460f5a4bcSLai Jiangshan complete(detach_completion); 189560f5a4bcSLai Jiangshan } 189660f5a4bcSLai Jiangshan 189760f5a4bcSLai Jiangshan /** 1898c34056a3STejun Heo * create_worker - create a new workqueue worker 189963d95a91STejun Heo * @pool: pool the new worker will belong to 1900c34056a3STejun Heo * 1901051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1902c34056a3STejun Heo * 1903c34056a3STejun Heo * CONTEXT: 1904c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1905c34056a3STejun Heo * 1906d185af30SYacine Belkadi * Return: 1907c34056a3STejun Heo * Pointer to the newly created worker. 1908c34056a3STejun Heo */ 1909bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1910c34056a3STejun Heo { 1911c34056a3STejun Heo struct worker *worker = NULL; 1912f3421797STejun Heo int id = -1; 1913e3c916a4STejun Heo char id_buf[16]; 1914c34056a3STejun Heo 19157cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 19167cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1917822d8405STejun Heo if (id < 0) 1918c34056a3STejun Heo goto fail; 1919c34056a3STejun Heo 1920f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1921c34056a3STejun Heo if (!worker) 1922c34056a3STejun Heo goto fail; 1923c34056a3STejun Heo 1924c34056a3STejun Heo worker->id = id; 1925c34056a3STejun Heo 192629c91e99STejun Heo if (pool->cpu >= 0) 1927e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1928e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1929f3421797STejun Heo else 1930e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1931e3c916a4STejun Heo 1932f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1933e3c916a4STejun Heo "kworker/%s", id_buf); 1934c34056a3STejun Heo if (IS_ERR(worker->task)) 1935c34056a3STejun Heo goto fail; 1936c34056a3STejun Heo 193791151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 193825834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 193991151228SOleg Nesterov 1940da028469SLai Jiangshan /* successful, attach the worker to the pool */ 19414736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1942822d8405STejun Heo 1943051e1850SLai Jiangshan /* start the newly created worker */ 1944a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1945051e1850SLai Jiangshan worker->pool->nr_workers++; 1946051e1850SLai Jiangshan worker_enter_idle(worker); 1947051e1850SLai Jiangshan wake_up_process(worker->task); 1948a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1949051e1850SLai Jiangshan 1950c34056a3STejun Heo return worker; 1951822d8405STejun Heo 1952c34056a3STejun Heo fail: 19539625ab17SLai Jiangshan if (id >= 0) 19547cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1955c34056a3STejun Heo kfree(worker); 1956c34056a3STejun Heo return NULL; 1957c34056a3STejun Heo } 1958c34056a3STejun Heo 1959c34056a3STejun Heo /** 1960c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1961c34056a3STejun Heo * @worker: worker to be destroyed 1962c34056a3STejun Heo * 196373eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 196473eb7fe7SLai Jiangshan * be idle. 1965c8e55f36STejun Heo * 1966c8e55f36STejun Heo * CONTEXT: 1967a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1968c34056a3STejun Heo */ 1969c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1970c34056a3STejun Heo { 1971bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1972c34056a3STejun Heo 1973cd549687STejun Heo lockdep_assert_held(&pool->lock); 1974cd549687STejun Heo 1975c34056a3STejun Heo /* sanity check frenzy */ 19766183c009STejun Heo if (WARN_ON(worker->current_work) || 197773eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 197873eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 19796183c009STejun Heo return; 1980c34056a3STejun Heo 1981bd7bdd43STejun Heo pool->nr_workers--; 1982bd7bdd43STejun Heo pool->nr_idle--; 1983c8e55f36STejun Heo 1984c8e55f36STejun Heo list_del_init(&worker->entry); 1985cb444766STejun Heo worker->flags |= WORKER_DIE; 198660f5a4bcSLai Jiangshan wake_up_process(worker->task); 1987c34056a3STejun Heo } 1988c34056a3STejun Heo 198932a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 1990e22bee78STejun Heo { 199132a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 1992e22bee78STejun Heo 1993a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1994e22bee78STejun Heo 19953347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1996e22bee78STejun Heo struct worker *worker; 1997e22bee78STejun Heo unsigned long expires; 1998e22bee78STejun Heo 1999e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 200063d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2001e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2002e22bee78STejun Heo 20033347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 200463d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 20053347fc9fSLai Jiangshan break; 2006e22bee78STejun Heo } 20073347fc9fSLai Jiangshan 20083347fc9fSLai Jiangshan destroy_worker(worker); 2009e22bee78STejun Heo } 2010e22bee78STejun Heo 2011a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2012e22bee78STejun Heo } 2013e22bee78STejun Heo 2014493a1724STejun Heo static void send_mayday(struct work_struct *work) 2015e22bee78STejun Heo { 2016112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2017112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2018493a1724STejun Heo 20192e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2020e22bee78STejun Heo 2021493008a8STejun Heo if (!wq->rescuer) 2022493a1724STejun Heo return; 2023e22bee78STejun Heo 2024e22bee78STejun Heo /* mayday mayday mayday */ 2025493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 202677668c8bSLai Jiangshan /* 202777668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 202877668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 202977668c8bSLai Jiangshan * rescuer is done with it. 203077668c8bSLai Jiangshan */ 203177668c8bSLai Jiangshan get_pwq(pwq); 2032493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2033e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2034493a1724STejun Heo } 2035e22bee78STejun Heo } 2036e22bee78STejun Heo 203732a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2038e22bee78STejun Heo { 203932a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2040e22bee78STejun Heo struct work_struct *work; 2041e22bee78STejun Heo 2042a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2043a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2044e22bee78STejun Heo 204563d95a91STejun Heo if (need_to_create_worker(pool)) { 2046e22bee78STejun Heo /* 2047e22bee78STejun Heo * We've been trying to create a new worker but 2048e22bee78STejun Heo * haven't been successful. We might be hitting an 2049e22bee78STejun Heo * allocation deadlock. Send distress signals to 2050e22bee78STejun Heo * rescuers. 2051e22bee78STejun Heo */ 205263d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2053e22bee78STejun Heo send_mayday(work); 2054e22bee78STejun Heo } 2055e22bee78STejun Heo 2056a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2057a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2058e22bee78STejun Heo 205963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2060e22bee78STejun Heo } 2061e22bee78STejun Heo 2062e22bee78STejun Heo /** 2063e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 206463d95a91STejun Heo * @pool: pool to create a new worker for 2065e22bee78STejun Heo * 206663d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2067e22bee78STejun Heo * have at least one idle worker on return from this function. If 2068e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 206963d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2070e22bee78STejun Heo * possible allocation deadlock. 2071e22bee78STejun Heo * 2072c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2073c5aa87bbSTejun Heo * may_start_working() %true. 2074e22bee78STejun Heo * 2075e22bee78STejun Heo * LOCKING: 2076a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2077e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2078e22bee78STejun Heo * manager. 2079e22bee78STejun Heo */ 208029187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2081d565ed63STejun Heo __releases(&pool->lock) 2082d565ed63STejun Heo __acquires(&pool->lock) 2083e22bee78STejun Heo { 2084e22bee78STejun Heo restart: 2085a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 20869f9c2364STejun Heo 2087e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 208863d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2089e22bee78STejun Heo 2090e22bee78STejun Heo while (true) { 2091051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2092e22bee78STejun Heo break; 2093e22bee78STejun Heo 2094e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 20959f9c2364STejun Heo 209663d95a91STejun Heo if (!need_to_create_worker(pool)) 2097e22bee78STejun Heo break; 2098e22bee78STejun Heo } 2099e22bee78STejun Heo 210063d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2101a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2102051e1850SLai Jiangshan /* 2103051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2104051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2105051e1850SLai Jiangshan * already become busy. 2106051e1850SLai Jiangshan */ 210763d95a91STejun Heo if (need_to_create_worker(pool)) 2108e22bee78STejun Heo goto restart; 2109e22bee78STejun Heo } 2110e22bee78STejun Heo 2111e22bee78STejun Heo /** 2112e22bee78STejun Heo * manage_workers - manage worker pool 2113e22bee78STejun Heo * @worker: self 2114e22bee78STejun Heo * 2115706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2116e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2117706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2118e22bee78STejun Heo * 2119e22bee78STejun Heo * The caller can safely start processing works on false return. On 2120e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2121e22bee78STejun Heo * and may_start_working() is true. 2122e22bee78STejun Heo * 2123e22bee78STejun Heo * CONTEXT: 2124a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2125e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2126e22bee78STejun Heo * 2127d185af30SYacine Belkadi * Return: 212829187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 212929187a9eSTejun Heo * start processing works, %true if management function was performed and 213029187a9eSTejun Heo * the conditions that the caller verified before calling the function may 213129187a9eSTejun Heo * no longer be true. 2132e22bee78STejun Heo */ 2133e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2134e22bee78STejun Heo { 213563d95a91STejun Heo struct worker_pool *pool = worker->pool; 2136e22bee78STejun Heo 2137692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 213829187a9eSTejun Heo return false; 2139692b4825STejun Heo 2140692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 21412607d7a6STejun Heo pool->manager = worker; 2142e22bee78STejun Heo 214329187a9eSTejun Heo maybe_create_worker(pool); 2144e22bee78STejun Heo 21452607d7a6STejun Heo pool->manager = NULL; 2146692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2147d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 214829187a9eSTejun Heo return true; 2149e22bee78STejun Heo } 2150e22bee78STejun Heo 2151a62428c0STejun Heo /** 2152a62428c0STejun Heo * process_one_work - process single work 2153c34056a3STejun Heo * @worker: self 2154a62428c0STejun Heo * @work: work to process 2155a62428c0STejun Heo * 2156a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2157a62428c0STejun Heo * process a single work including synchronization against and 2158a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2159a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2160a62428c0STejun Heo * call this function to process a work. 2161a62428c0STejun Heo * 2162a62428c0STejun Heo * CONTEXT: 2163a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2164a62428c0STejun Heo */ 2165c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2166d565ed63STejun Heo __releases(&pool->lock) 2167d565ed63STejun Heo __acquires(&pool->lock) 21681da177e4SLinus Torvalds { 2169112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2170bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2171112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 217273f53c4aSTejun Heo int work_color; 21737e11629dSTejun Heo struct worker *collision; 21744e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21754e6045f1SJohannes Berg /* 2176a62428c0STejun Heo * It is permissible to free the struct work_struct from 2177a62428c0STejun Heo * inside the function that is called from it, this we need to 2178a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2179a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2180a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21814e6045f1SJohannes Berg */ 21824d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21834d82a1deSPeter Zijlstra 21844d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21854e6045f1SJohannes Berg #endif 2186807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 218785327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2188ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 218925511a47STejun Heo 21907e11629dSTejun Heo /* 21917e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21927e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21937e11629dSTejun Heo * already processing the work. If so, defer the work to the 21947e11629dSTejun Heo * currently executing one. 21957e11629dSTejun Heo */ 2196c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21977e11629dSTejun Heo if (unlikely(collision)) { 21987e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21997e11629dSTejun Heo return; 22007e11629dSTejun Heo } 22011da177e4SLinus Torvalds 22028930cabaSTejun Heo /* claim and dequeue */ 22031da177e4SLinus Torvalds debug_work_deactivate(work); 2204c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2205c34056a3STejun Heo worker->current_work = work; 2206a2c1c57bSTejun Heo worker->current_func = work->func; 2207112202d9STejun Heo worker->current_pwq = pwq; 220873f53c4aSTejun Heo work_color = get_work_color(work); 22097a22ad75STejun Heo 22108bf89593STejun Heo /* 22118bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 22128bf89593STejun Heo * overridden through set_worker_desc(). 22138bf89593STejun Heo */ 22148bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 22158bf89593STejun Heo 2216a62428c0STejun Heo list_del_init(&work->entry); 2217a62428c0STejun Heo 2218649027d7STejun Heo /* 2219228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2220228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2221228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2222228f1d00SLai Jiangshan * execution of the pending work items. 2223fb0e7bebSTejun Heo */ 2224fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2225228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2226fb0e7bebSTejun Heo 2227974271c4STejun Heo /* 2228a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2229a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2230a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2231a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2232228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2233974271c4STejun Heo */ 2234a489a03eSLai Jiangshan if (need_more_worker(pool)) 223563d95a91STejun Heo wake_up_worker(pool); 2236974271c4STejun Heo 22378930cabaSTejun Heo /* 22387c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2239d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 224023657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 224123657bb1STejun Heo * disabled. 22428930cabaSTejun Heo */ 22437c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22441da177e4SLinus Torvalds 2245a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2246365970a1SDavid Howells 2247a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 22483295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2249e6f3faa7SPeter Zijlstra /* 2250f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2251f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2252e6f3faa7SPeter Zijlstra * 2253e6f3faa7SPeter Zijlstra * However, that would result in: 2254e6f3faa7SPeter Zijlstra * 2255e6f3faa7SPeter Zijlstra * A(W1) 2256e6f3faa7SPeter Zijlstra * WFC(C) 2257e6f3faa7SPeter Zijlstra * A(W1) 2258e6f3faa7SPeter Zijlstra * C(C) 2259e6f3faa7SPeter Zijlstra * 2260e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2261e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2262e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2263f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2264e6f3faa7SPeter Zijlstra * these locks. 2265e6f3faa7SPeter Zijlstra * 2266e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2267e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2268e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2269e6f3faa7SPeter Zijlstra */ 2270f52be570SPeter Zijlstra lockdep_invariant_state(true); 2271e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2272a2c1c57bSTejun Heo worker->current_func(work); 2273e36c886aSArjan van de Ven /* 2274e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2275e36c886aSArjan van de Ven * point will only record its address. 2276e36c886aSArjan van de Ven */ 22771c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 22783295f0efSIngo Molnar lock_map_release(&lockdep_map); 2279112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22801da177e4SLinus Torvalds 2281d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2282044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2283d75f773cSSakari Ailus " last function: %ps\n", 2284a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2285a2c1c57bSTejun Heo worker->current_func); 2286d5abe669SPeter Zijlstra debug_show_held_locks(current); 2287d5abe669SPeter Zijlstra dump_stack(); 2288d5abe669SPeter Zijlstra } 2289d5abe669SPeter Zijlstra 2290b22ce278STejun Heo /* 2291025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2292b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2293b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2294b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2295789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2296789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2297b22ce278STejun Heo */ 2298a7e6425eSPaul E. McKenney cond_resched(); 2299b22ce278STejun Heo 2300a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2301a62428c0STejun Heo 2302fb0e7bebSTejun Heo /* clear cpu intensive status */ 2303fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2304fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2305fb0e7bebSTejun Heo 23061b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 23071b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 23081b69ac6bSJohannes Weiner 2309a62428c0STejun Heo /* we're done with it, release */ 231042f8570fSSasha Levin hash_del(&worker->hentry); 2311c34056a3STejun Heo worker->current_work = NULL; 2312a2c1c57bSTejun Heo worker->current_func = NULL; 2313112202d9STejun Heo worker->current_pwq = NULL; 2314112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 23151da177e4SLinus Torvalds } 23161da177e4SLinus Torvalds 2317affee4b2STejun Heo /** 2318affee4b2STejun Heo * process_scheduled_works - process scheduled works 2319affee4b2STejun Heo * @worker: self 2320affee4b2STejun Heo * 2321affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2322affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2323affee4b2STejun Heo * fetches a work from the top and executes it. 2324affee4b2STejun Heo * 2325affee4b2STejun Heo * CONTEXT: 2326a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2327affee4b2STejun Heo * multiple times. 2328affee4b2STejun Heo */ 2329affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 23301da177e4SLinus Torvalds { 2331affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2332affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2333a62428c0STejun Heo struct work_struct, entry); 2334c34056a3STejun Heo process_one_work(worker, work); 2335a62428c0STejun Heo } 23361da177e4SLinus Torvalds } 23371da177e4SLinus Torvalds 2338197f6accSTejun Heo static void set_pf_worker(bool val) 2339197f6accSTejun Heo { 2340197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2341197f6accSTejun Heo if (val) 2342197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2343197f6accSTejun Heo else 2344197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2345197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2346197f6accSTejun Heo } 2347197f6accSTejun Heo 23484690c4abSTejun Heo /** 23494690c4abSTejun Heo * worker_thread - the worker thread function 2350c34056a3STejun Heo * @__worker: self 23514690c4abSTejun Heo * 2352c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2353c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2354c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2355c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2356c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2357d185af30SYacine Belkadi * 2358d185af30SYacine Belkadi * Return: 0 23594690c4abSTejun Heo */ 2360c34056a3STejun Heo static int worker_thread(void *__worker) 23611da177e4SLinus Torvalds { 2362c34056a3STejun Heo struct worker *worker = __worker; 2363bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 23641da177e4SLinus Torvalds 2365e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2366197f6accSTejun Heo set_pf_worker(true); 2367c8e55f36STejun Heo woke_up: 2368a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2369affee4b2STejun Heo 2370a9ab775bSTejun Heo /* am I supposed to die? */ 2371a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2372a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2373a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2374197f6accSTejun Heo set_pf_worker(false); 237560f5a4bcSLai Jiangshan 237660f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 23777cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 2378a2d812a2STejun Heo worker_detach_from_pool(worker); 237960f5a4bcSLai Jiangshan kfree(worker); 2380c8e55f36STejun Heo return 0; 2381c8e55f36STejun Heo } 2382c8e55f36STejun Heo 2383c8e55f36STejun Heo worker_leave_idle(worker); 2384db7bccf4STejun Heo recheck: 2385e22bee78STejun Heo /* no more worker necessary? */ 238663d95a91STejun Heo if (!need_more_worker(pool)) 2387e22bee78STejun Heo goto sleep; 2388e22bee78STejun Heo 2389e22bee78STejun Heo /* do we need to manage? */ 239063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2391e22bee78STejun Heo goto recheck; 2392e22bee78STejun Heo 2393c8e55f36STejun Heo /* 2394c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2395c8e55f36STejun Heo * preparing to process a work or actually processing it. 2396c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2397c8e55f36STejun Heo */ 23986183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2399c8e55f36STejun Heo 2400e22bee78STejun Heo /* 2401a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2402a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2403a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2404a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2405a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2406e22bee78STejun Heo */ 2407a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2408e22bee78STejun Heo 2409e22bee78STejun Heo do { 2410affee4b2STejun Heo struct work_struct *work = 2411bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2412affee4b2STejun Heo struct work_struct, entry); 2413affee4b2STejun Heo 241482607adcSTejun Heo pool->watchdog_ts = jiffies; 241582607adcSTejun Heo 2416c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2417affee4b2STejun Heo /* optimization path, not strictly necessary */ 2418affee4b2STejun Heo process_one_work(worker, work); 2419affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2420affee4b2STejun Heo process_scheduled_works(worker); 2421affee4b2STejun Heo } else { 2422c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2423affee4b2STejun Heo process_scheduled_works(worker); 2424affee4b2STejun Heo } 242563d95a91STejun Heo } while (keep_working(pool)); 2426affee4b2STejun Heo 2427228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2428d313dd85STejun Heo sleep: 2429c8e55f36STejun Heo /* 2430d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2431d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2432d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2433d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2434d565ed63STejun Heo * event. 2435c8e55f36STejun Heo */ 2436c8e55f36STejun Heo worker_enter_idle(worker); 2437c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2438a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24391da177e4SLinus Torvalds schedule(); 2440c8e55f36STejun Heo goto woke_up; 24411da177e4SLinus Torvalds } 24421da177e4SLinus Torvalds 2443e22bee78STejun Heo /** 2444e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2445111c225aSTejun Heo * @__rescuer: self 2446e22bee78STejun Heo * 2447e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2448493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2449e22bee78STejun Heo * 2450706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2451e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2452e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2453e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2454e22bee78STejun Heo * the problem rescuer solves. 2455e22bee78STejun Heo * 2456706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2457706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2458e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2459e22bee78STejun Heo * 2460e22bee78STejun Heo * This should happen rarely. 2461d185af30SYacine Belkadi * 2462d185af30SYacine Belkadi * Return: 0 2463e22bee78STejun Heo */ 2464111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2465e22bee78STejun Heo { 2466111c225aSTejun Heo struct worker *rescuer = __rescuer; 2467111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2468e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 24694d595b86SLai Jiangshan bool should_stop; 2470e22bee78STejun Heo 2471e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2472111c225aSTejun Heo 2473111c225aSTejun Heo /* 2474111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2475111c225aSTejun Heo * doesn't participate in concurrency management. 2476111c225aSTejun Heo */ 2477197f6accSTejun Heo set_pf_worker(true); 2478e22bee78STejun Heo repeat: 2479c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 24801da177e4SLinus Torvalds 24814d595b86SLai Jiangshan /* 24824d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 24834d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 24844d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 24854d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 24864d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 24874d595b86SLai Jiangshan * list is always empty on exit. 24884d595b86SLai Jiangshan */ 24894d595b86SLai Jiangshan should_stop = kthread_should_stop(); 24901da177e4SLinus Torvalds 2491493a1724STejun Heo /* see whether any pwq is asking for help */ 2492a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2493493a1724STejun Heo 2494493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2495493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2496493a1724STejun Heo struct pool_workqueue, mayday_node); 2497112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2498e22bee78STejun Heo struct work_struct *work, *n; 249982607adcSTejun Heo bool first = true; 2500e22bee78STejun Heo 2501e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2502493a1724STejun Heo list_del_init(&pwq->mayday_node); 2503493a1724STejun Heo 2504a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2505e22bee78STejun Heo 250651697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 250751697d39SLai Jiangshan 2508a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2509e22bee78STejun Heo 2510e22bee78STejun Heo /* 2511e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2512e22bee78STejun Heo * process'em. 2513e22bee78STejun Heo */ 25140479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 251582607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 251682607adcSTejun Heo if (get_work_pwq(work) == pwq) { 251782607adcSTejun Heo if (first) 251882607adcSTejun Heo pool->watchdog_ts = jiffies; 2519e22bee78STejun Heo move_linked_works(work, scheduled, &n); 252082607adcSTejun Heo } 252182607adcSTejun Heo first = false; 252282607adcSTejun Heo } 2523e22bee78STejun Heo 2524008847f6SNeilBrown if (!list_empty(scheduled)) { 2525e22bee78STejun Heo process_scheduled_works(rescuer); 25267576958aSTejun Heo 25277576958aSTejun Heo /* 2528008847f6SNeilBrown * The above execution of rescued work items could 2529008847f6SNeilBrown * have created more to rescue through 2530008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2531008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2532008847f6SNeilBrown * that such back-to-back work items, which may be 2533008847f6SNeilBrown * being used to relieve memory pressure, don't 2534008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2535008847f6SNeilBrown */ 25364f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2537a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2538e66b39afSTejun Heo /* 2539e66b39afSTejun Heo * Queue iff we aren't racing destruction 2540e66b39afSTejun Heo * and somebody else hasn't queued it already. 2541e66b39afSTejun Heo */ 2542e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2543008847f6SNeilBrown get_pwq(pwq); 2544e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2545e66b39afSTejun Heo } 2546a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2547008847f6SNeilBrown } 2548008847f6SNeilBrown } 2549008847f6SNeilBrown 2550008847f6SNeilBrown /* 255177668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 255213b1d625SLai Jiangshan * go away while we're still attached to it. 255377668c8bSLai Jiangshan */ 255477668c8bSLai Jiangshan put_pwq(pwq); 255577668c8bSLai Jiangshan 255677668c8bSLai Jiangshan /* 2557d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 25587576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 25597576958aSTejun Heo * and stalling the execution. 25607576958aSTejun Heo */ 2561d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 256263d95a91STejun Heo wake_up_worker(pool); 25637576958aSTejun Heo 2564a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 256513b1d625SLai Jiangshan 2566a2d812a2STejun Heo worker_detach_from_pool(rescuer); 256713b1d625SLai Jiangshan 2568a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 25691da177e4SLinus Torvalds } 25701da177e4SLinus Torvalds 2571a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2572493a1724STejun Heo 25734d595b86SLai Jiangshan if (should_stop) { 25744d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2575197f6accSTejun Heo set_pf_worker(false); 25764d595b86SLai Jiangshan return 0; 25774d595b86SLai Jiangshan } 25784d595b86SLai Jiangshan 2579111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2580111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2581e22bee78STejun Heo schedule(); 2582e22bee78STejun Heo goto repeat; 25831da177e4SLinus Torvalds } 25841da177e4SLinus Torvalds 2585fca839c0STejun Heo /** 2586fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2587fca839c0STejun Heo * @target_wq: workqueue being flushed 2588fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2589fca839c0STejun Heo * 2590fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2591fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2592fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2593fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2594fca839c0STejun Heo * a deadlock. 2595fca839c0STejun Heo */ 2596fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2597fca839c0STejun Heo struct work_struct *target_work) 2598fca839c0STejun Heo { 2599fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2600fca839c0STejun Heo struct worker *worker; 2601fca839c0STejun Heo 2602fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2603fca839c0STejun Heo return; 2604fca839c0STejun Heo 2605fca839c0STejun Heo worker = current_wq_worker(); 2606fca839c0STejun Heo 2607fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2608d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2609fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 261023d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 261123d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2612d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2613fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2614fca839c0STejun Heo target_wq->name, target_func); 2615fca839c0STejun Heo } 2616fca839c0STejun Heo 2617fc2e4d70SOleg Nesterov struct wq_barrier { 2618fc2e4d70SOleg Nesterov struct work_struct work; 2619fc2e4d70SOleg Nesterov struct completion done; 26202607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2621fc2e4d70SOleg Nesterov }; 2622fc2e4d70SOleg Nesterov 2623fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2624fc2e4d70SOleg Nesterov { 2625fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2626fc2e4d70SOleg Nesterov complete(&barr->done); 2627fc2e4d70SOleg Nesterov } 2628fc2e4d70SOleg Nesterov 26294690c4abSTejun Heo /** 26304690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2631112202d9STejun Heo * @pwq: pwq to insert barrier into 26324690c4abSTejun Heo * @barr: wq_barrier to insert 2633affee4b2STejun Heo * @target: target work to attach @barr to 2634affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 26354690c4abSTejun Heo * 2636affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2637affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2638affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2639affee4b2STejun Heo * cpu. 2640affee4b2STejun Heo * 2641affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2642affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2643affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2644affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2645affee4b2STejun Heo * after a work with LINKED flag set. 2646affee4b2STejun Heo * 2647affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2648112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 26494690c4abSTejun Heo * 26504690c4abSTejun Heo * CONTEXT: 2651a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 26524690c4abSTejun Heo */ 2653112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2654affee4b2STejun Heo struct wq_barrier *barr, 2655affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2656fc2e4d70SOleg Nesterov { 2657affee4b2STejun Heo struct list_head *head; 2658affee4b2STejun Heo unsigned int linked = 0; 2659affee4b2STejun Heo 2660dc186ad7SThomas Gleixner /* 2661d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2662dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2663dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2664dc186ad7SThomas Gleixner * might deadlock. 2665dc186ad7SThomas Gleixner */ 2666ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 266722df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 266852fa5bc5SBoqun Feng 2669fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2670fd1a5b04SByungchul Park 26712607d7a6STejun Heo barr->task = current; 267283c22520SOleg Nesterov 2673affee4b2STejun Heo /* 2674affee4b2STejun Heo * If @target is currently being executed, schedule the 2675affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2676affee4b2STejun Heo */ 2677affee4b2STejun Heo if (worker) 2678affee4b2STejun Heo head = worker->scheduled.next; 2679affee4b2STejun Heo else { 2680affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2681affee4b2STejun Heo 2682affee4b2STejun Heo head = target->entry.next; 2683affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2684affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2685affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2686affee4b2STejun Heo } 2687affee4b2STejun Heo 2688dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2689112202d9STejun Heo insert_work(pwq, &barr->work, head, 2690affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2691fc2e4d70SOleg Nesterov } 2692fc2e4d70SOleg Nesterov 269373f53c4aSTejun Heo /** 2694112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 269573f53c4aSTejun Heo * @wq: workqueue being flushed 269673f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 269773f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 269873f53c4aSTejun Heo * 2699112202d9STejun Heo * Prepare pwqs for workqueue flushing. 270073f53c4aSTejun Heo * 2701112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2702112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2703112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2704112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2705112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 270673f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 270773f53c4aSTejun Heo * 270873f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 270973f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 271073f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 271173f53c4aSTejun Heo * is returned. 271273f53c4aSTejun Heo * 2713112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 271473f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 271573f53c4aSTejun Heo * advanced to @work_color. 271673f53c4aSTejun Heo * 271773f53c4aSTejun Heo * CONTEXT: 27183c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 271973f53c4aSTejun Heo * 2720d185af30SYacine Belkadi * Return: 272173f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 272273f53c4aSTejun Heo * otherwise. 272373f53c4aSTejun Heo */ 2724112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 272573f53c4aSTejun Heo int flush_color, int work_color) 27261da177e4SLinus Torvalds { 272773f53c4aSTejun Heo bool wait = false; 272849e3cf44STejun Heo struct pool_workqueue *pwq; 27291da177e4SLinus Torvalds 273073f53c4aSTejun Heo if (flush_color >= 0) { 27316183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2732112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2733dc186ad7SThomas Gleixner } 273414441960SOleg Nesterov 273549e3cf44STejun Heo for_each_pwq(pwq, wq) { 2736112202d9STejun Heo struct worker_pool *pool = pwq->pool; 27371da177e4SLinus Torvalds 2738a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 273973f53c4aSTejun Heo 274073f53c4aSTejun Heo if (flush_color >= 0) { 27416183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 274273f53c4aSTejun Heo 2743112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2744112202d9STejun Heo pwq->flush_color = flush_color; 2745112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 274673f53c4aSTejun Heo wait = true; 27471da177e4SLinus Torvalds } 274873f53c4aSTejun Heo } 274973f53c4aSTejun Heo 275073f53c4aSTejun Heo if (work_color >= 0) { 27516183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2752112202d9STejun Heo pwq->work_color = work_color; 275373f53c4aSTejun Heo } 275473f53c4aSTejun Heo 2755a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27561da177e4SLinus Torvalds } 27571da177e4SLinus Torvalds 2758112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 275973f53c4aSTejun Heo complete(&wq->first_flusher->done); 276073f53c4aSTejun Heo 276173f53c4aSTejun Heo return wait; 276283c22520SOleg Nesterov } 27631da177e4SLinus Torvalds 27640fcb78c2SRolf Eike Beer /** 27651da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 27660fcb78c2SRolf Eike Beer * @wq: workqueue to flush 27671da177e4SLinus Torvalds * 2768c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2769c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 27701da177e4SLinus Torvalds */ 27717ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 27721da177e4SLinus Torvalds { 277373f53c4aSTejun Heo struct wq_flusher this_flusher = { 277473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 277573f53c4aSTejun Heo .flush_color = -1, 2776fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 277773f53c4aSTejun Heo }; 277873f53c4aSTejun Heo int next_color; 2779b1f4ec17SOleg Nesterov 27803347fa09STejun Heo if (WARN_ON(!wq_online)) 27813347fa09STejun Heo return; 27823347fa09STejun Heo 278387915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 278487915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 278587915adcSJohannes Berg 27863c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 278773f53c4aSTejun Heo 278873f53c4aSTejun Heo /* 278973f53c4aSTejun Heo * Start-to-wait phase 279073f53c4aSTejun Heo */ 279173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 279273f53c4aSTejun Heo 279373f53c4aSTejun Heo if (next_color != wq->flush_color) { 279473f53c4aSTejun Heo /* 279573f53c4aSTejun Heo * Color space is not full. The current work_color 279673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 279773f53c4aSTejun Heo * by one. 279873f53c4aSTejun Heo */ 27996183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 280073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 280173f53c4aSTejun Heo wq->work_color = next_color; 280273f53c4aSTejun Heo 280373f53c4aSTejun Heo if (!wq->first_flusher) { 280473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 28056183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 280673f53c4aSTejun Heo 280773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 280873f53c4aSTejun Heo 2809112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 281073f53c4aSTejun Heo wq->work_color)) { 281173f53c4aSTejun Heo /* nothing to flush, done */ 281273f53c4aSTejun Heo wq->flush_color = next_color; 281373f53c4aSTejun Heo wq->first_flusher = NULL; 281473f53c4aSTejun Heo goto out_unlock; 281573f53c4aSTejun Heo } 281673f53c4aSTejun Heo } else { 281773f53c4aSTejun Heo /* wait in queue */ 28186183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 281973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2820112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 282173f53c4aSTejun Heo } 282273f53c4aSTejun Heo } else { 282373f53c4aSTejun Heo /* 282473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 282573f53c4aSTejun Heo * The next flush completion will assign us 282673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 282773f53c4aSTejun Heo */ 282873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 282973f53c4aSTejun Heo } 283073f53c4aSTejun Heo 2831fca839c0STejun Heo check_flush_dependency(wq, NULL); 2832fca839c0STejun Heo 28333c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 283473f53c4aSTejun Heo 283573f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 283673f53c4aSTejun Heo 283773f53c4aSTejun Heo /* 283873f53c4aSTejun Heo * Wake-up-and-cascade phase 283973f53c4aSTejun Heo * 284073f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 284173f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 284273f53c4aSTejun Heo */ 284300d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 284473f53c4aSTejun Heo return; 284573f53c4aSTejun Heo 28463c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 284773f53c4aSTejun Heo 28484ce48b37STejun Heo /* we might have raced, check again with mutex held */ 28494ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 28504ce48b37STejun Heo goto out_unlock; 28514ce48b37STejun Heo 285200d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 285373f53c4aSTejun Heo 28546183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 28556183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 285673f53c4aSTejun Heo 285773f53c4aSTejun Heo while (true) { 285873f53c4aSTejun Heo struct wq_flusher *next, *tmp; 285973f53c4aSTejun Heo 286073f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 286173f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 286273f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 286373f53c4aSTejun Heo break; 286473f53c4aSTejun Heo list_del_init(&next->list); 286573f53c4aSTejun Heo complete(&next->done); 286673f53c4aSTejun Heo } 286773f53c4aSTejun Heo 28686183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 286973f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 287073f53c4aSTejun Heo 287173f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 287273f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 287373f53c4aSTejun Heo 287473f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 287573f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 287673f53c4aSTejun Heo /* 287773f53c4aSTejun Heo * Assign the same color to all overflowed 287873f53c4aSTejun Heo * flushers, advance work_color and append to 287973f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 288073f53c4aSTejun Heo * phase for these overflowed flushers. 288173f53c4aSTejun Heo */ 288273f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 288373f53c4aSTejun Heo tmp->flush_color = wq->work_color; 288473f53c4aSTejun Heo 288573f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 288673f53c4aSTejun Heo 288773f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 288873f53c4aSTejun Heo &wq->flusher_queue); 2889112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 289073f53c4aSTejun Heo } 289173f53c4aSTejun Heo 289273f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 28936183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 289473f53c4aSTejun Heo break; 289573f53c4aSTejun Heo } 289673f53c4aSTejun Heo 289773f53c4aSTejun Heo /* 289873f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2899112202d9STejun Heo * the new first flusher and arm pwqs. 290073f53c4aSTejun Heo */ 29016183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 29026183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 290373f53c4aSTejun Heo 290473f53c4aSTejun Heo list_del_init(&next->list); 290573f53c4aSTejun Heo wq->first_flusher = next; 290673f53c4aSTejun Heo 2907112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 290873f53c4aSTejun Heo break; 290973f53c4aSTejun Heo 291073f53c4aSTejun Heo /* 291173f53c4aSTejun Heo * Meh... this color is already done, clear first 291273f53c4aSTejun Heo * flusher and repeat cascading. 291373f53c4aSTejun Heo */ 291473f53c4aSTejun Heo wq->first_flusher = NULL; 291573f53c4aSTejun Heo } 291673f53c4aSTejun Heo 291773f53c4aSTejun Heo out_unlock: 29183c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 29191da177e4SLinus Torvalds } 29201dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 29211da177e4SLinus Torvalds 29229c5a2ba7STejun Heo /** 29239c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 29249c5a2ba7STejun Heo * @wq: workqueue to drain 29259c5a2ba7STejun Heo * 29269c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 29279c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 29289c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2929b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 29309c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 29319c5a2ba7STejun Heo * takes too long. 29329c5a2ba7STejun Heo */ 29339c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 29349c5a2ba7STejun Heo { 29359c5a2ba7STejun Heo unsigned int flush_cnt = 0; 293649e3cf44STejun Heo struct pool_workqueue *pwq; 29379c5a2ba7STejun Heo 29389c5a2ba7STejun Heo /* 29399c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 29409c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2941618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 29429c5a2ba7STejun Heo */ 294387fc741eSLai Jiangshan mutex_lock(&wq->mutex); 29449c5a2ba7STejun Heo if (!wq->nr_drainers++) 2945618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 294687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29479c5a2ba7STejun Heo reflush: 29489c5a2ba7STejun Heo flush_workqueue(wq); 29499c5a2ba7STejun Heo 2950b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 295176af4d93STejun Heo 295249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2953fa2563e4SThomas Tuttle bool drained; 29549c5a2ba7STejun Heo 2955a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 2956112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2957a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 2958fa2563e4SThomas Tuttle 2959fa2563e4SThomas Tuttle if (drained) 29609c5a2ba7STejun Heo continue; 29619c5a2ba7STejun Heo 29629c5a2ba7STejun Heo if (++flush_cnt == 10 || 29639c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2964c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 29659c5a2ba7STejun Heo wq->name, flush_cnt); 296676af4d93STejun Heo 2967b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 29689c5a2ba7STejun Heo goto reflush; 29699c5a2ba7STejun Heo } 29709c5a2ba7STejun Heo 29719c5a2ba7STejun Heo if (!--wq->nr_drainers) 2972618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 297387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29749c5a2ba7STejun Heo } 29759c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 29769c5a2ba7STejun Heo 2977d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2978d6e89786SJohannes Berg bool from_cancel) 2979baf59022STejun Heo { 2980baf59022STejun Heo struct worker *worker = NULL; 2981c9e7cf27STejun Heo struct worker_pool *pool; 2982112202d9STejun Heo struct pool_workqueue *pwq; 2983baf59022STejun Heo 2984baf59022STejun Heo might_sleep(); 2985baf59022STejun Heo 298624acfb71SThomas Gleixner rcu_read_lock(); 2987fa1b54e6STejun Heo pool = get_work_pool(work); 2988fa1b54e6STejun Heo if (!pool) { 298924acfb71SThomas Gleixner rcu_read_unlock(); 2990fa1b54e6STejun Heo return false; 2991fa1b54e6STejun Heo } 2992fa1b54e6STejun Heo 2993a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 29940b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2995112202d9STejun Heo pwq = get_work_pwq(work); 2996112202d9STejun Heo if (pwq) { 2997112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2998baf59022STejun Heo goto already_gone; 2999606a5020STejun Heo } else { 3000c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3001baf59022STejun Heo if (!worker) 3002baf59022STejun Heo goto already_gone; 3003112202d9STejun Heo pwq = worker->current_pwq; 3004606a5020STejun Heo } 3005baf59022STejun Heo 3006fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3007fca839c0STejun Heo 3008112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3009a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3010baf59022STejun Heo 3011e159489bSTejun Heo /* 3012a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3013a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3014a1d14934SPeter Zijlstra * 3015a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3016a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3017a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3018a1d14934SPeter Zijlstra * forward progress. 3019e159489bSTejun Heo */ 3020d6e89786SJohannes Berg if (!from_cancel && 3021d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3022112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3023112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3024a1d14934SPeter Zijlstra } 302524acfb71SThomas Gleixner rcu_read_unlock(); 3026baf59022STejun Heo return true; 3027baf59022STejun Heo already_gone: 3028a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 302924acfb71SThomas Gleixner rcu_read_unlock(); 3030baf59022STejun Heo return false; 3031baf59022STejun Heo } 3032baf59022STejun Heo 3033d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3034d6e89786SJohannes Berg { 3035d6e89786SJohannes Berg struct wq_barrier barr; 3036d6e89786SJohannes Berg 3037d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3038d6e89786SJohannes Berg return false; 3039d6e89786SJohannes Berg 30404d43d395STetsuo Handa if (WARN_ON(!work->func)) 30414d43d395STetsuo Handa return false; 30424d43d395STetsuo Handa 304387915adcSJohannes Berg if (!from_cancel) { 304487915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 304587915adcSJohannes Berg lock_map_release(&work->lockdep_map); 304687915adcSJohannes Berg } 304787915adcSJohannes Berg 3048d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3049d6e89786SJohannes Berg wait_for_completion(&barr.done); 3050d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3051d6e89786SJohannes Berg return true; 3052d6e89786SJohannes Berg } else { 3053d6e89786SJohannes Berg return false; 3054d6e89786SJohannes Berg } 3055d6e89786SJohannes Berg } 3056d6e89786SJohannes Berg 3057db700897SOleg Nesterov /** 3058401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3059401a8d04STejun Heo * @work: the work to flush 3060db700897SOleg Nesterov * 3061606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3062606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3063401a8d04STejun Heo * 3064d185af30SYacine Belkadi * Return: 3065401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3066401a8d04STejun Heo * %false if it was already idle. 3067db700897SOleg Nesterov */ 3068401a8d04STejun Heo bool flush_work(struct work_struct *work) 3069db700897SOleg Nesterov { 3070d6e89786SJohannes Berg return __flush_work(work, false); 3071606a5020STejun Heo } 3072db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3073db700897SOleg Nesterov 30748603e1b3STejun Heo struct cwt_wait { 3075ac6424b9SIngo Molnar wait_queue_entry_t wait; 30768603e1b3STejun Heo struct work_struct *work; 30778603e1b3STejun Heo }; 30788603e1b3STejun Heo 3079ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 30808603e1b3STejun Heo { 30818603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 30828603e1b3STejun Heo 30838603e1b3STejun Heo if (cwait->work != key) 30848603e1b3STejun Heo return 0; 30858603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 30868603e1b3STejun Heo } 30878603e1b3STejun Heo 308836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3089401a8d04STejun Heo { 30908603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3091bbb68dfaSTejun Heo unsigned long flags; 30921f1f642eSOleg Nesterov int ret; 30931f1f642eSOleg Nesterov 30941f1f642eSOleg Nesterov do { 3095bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3096bbb68dfaSTejun Heo /* 30978603e1b3STejun Heo * If someone else is already canceling, wait for it to 30988603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 30998603e1b3STejun Heo * because we may get scheduled between @work's completion 31008603e1b3STejun Heo * and the other canceling task resuming and clearing 31018603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 31028603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 31038603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 31048603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 31058603e1b3STejun Heo * we're hogging the CPU. 31068603e1b3STejun Heo * 31078603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 31088603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 31098603e1b3STejun Heo * wake function which matches @work along with exclusive 31108603e1b3STejun Heo * wait and wakeup. 3111bbb68dfaSTejun Heo */ 31128603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 31138603e1b3STejun Heo struct cwt_wait cwait; 31148603e1b3STejun Heo 31158603e1b3STejun Heo init_wait(&cwait.wait); 31168603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 31178603e1b3STejun Heo cwait.work = work; 31188603e1b3STejun Heo 31198603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 31208603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 31218603e1b3STejun Heo if (work_is_canceling(work)) 31228603e1b3STejun Heo schedule(); 31238603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 31248603e1b3STejun Heo } 31251f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 31261f1f642eSOleg Nesterov 3127bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3128bbb68dfaSTejun Heo mark_work_canceling(work); 3129bbb68dfaSTejun Heo local_irq_restore(flags); 3130bbb68dfaSTejun Heo 31313347fa09STejun Heo /* 31323347fa09STejun Heo * This allows canceling during early boot. We know that @work 31333347fa09STejun Heo * isn't executing. 31343347fa09STejun Heo */ 31353347fa09STejun Heo if (wq_online) 3136d6e89786SJohannes Berg __flush_work(work, true); 31373347fa09STejun Heo 31387a22ad75STejun Heo clear_work_data(work); 31398603e1b3STejun Heo 31408603e1b3STejun Heo /* 31418603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 31428603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 31438603e1b3STejun Heo * visible there. 31448603e1b3STejun Heo */ 31458603e1b3STejun Heo smp_mb(); 31468603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 31478603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 31488603e1b3STejun Heo 31491f1f642eSOleg Nesterov return ret; 31501f1f642eSOleg Nesterov } 31511f1f642eSOleg Nesterov 31526e84d644SOleg Nesterov /** 3153401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3154401a8d04STejun Heo * @work: the work to cancel 31556e84d644SOleg Nesterov * 3156401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3157401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3158401a8d04STejun Heo * another workqueue. On return from this function, @work is 3159401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 31601f1f642eSOleg Nesterov * 3161401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3162401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 31636e84d644SOleg Nesterov * 3164401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 31656e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3166401a8d04STejun Heo * 3167d185af30SYacine Belkadi * Return: 3168401a8d04STejun Heo * %true if @work was pending, %false otherwise. 31696e84d644SOleg Nesterov */ 3170401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 31716e84d644SOleg Nesterov { 317236e227d2STejun Heo return __cancel_work_timer(work, false); 3173b89deed3SOleg Nesterov } 317428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3175b89deed3SOleg Nesterov 31766e84d644SOleg Nesterov /** 3177401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3178401a8d04STejun Heo * @dwork: the delayed work to flush 31796e84d644SOleg Nesterov * 3180401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3181401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3182401a8d04STejun Heo * considers the last queueing instance of @dwork. 31831f1f642eSOleg Nesterov * 3184d185af30SYacine Belkadi * Return: 3185401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3186401a8d04STejun Heo * %false if it was already idle. 31876e84d644SOleg Nesterov */ 3188401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3189401a8d04STejun Heo { 31908930cabaSTejun Heo local_irq_disable(); 3191401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 319260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 31938930cabaSTejun Heo local_irq_enable(); 3194401a8d04STejun Heo return flush_work(&dwork->work); 3195401a8d04STejun Heo } 3196401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3197401a8d04STejun Heo 319805f0fe6bSTejun Heo /** 319905f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 320005f0fe6bSTejun Heo * @rwork: the rcu work to flush 320105f0fe6bSTejun Heo * 320205f0fe6bSTejun Heo * Return: 320305f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 320405f0fe6bSTejun Heo * %false if it was already idle. 320505f0fe6bSTejun Heo */ 320605f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 320705f0fe6bSTejun Heo { 320805f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 320905f0fe6bSTejun Heo rcu_barrier(); 321005f0fe6bSTejun Heo flush_work(&rwork->work); 321105f0fe6bSTejun Heo return true; 321205f0fe6bSTejun Heo } else { 321305f0fe6bSTejun Heo return flush_work(&rwork->work); 321405f0fe6bSTejun Heo } 321505f0fe6bSTejun Heo } 321605f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 321705f0fe6bSTejun Heo 3218f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3219f72b8792SJens Axboe { 3220f72b8792SJens Axboe unsigned long flags; 3221f72b8792SJens Axboe int ret; 3222f72b8792SJens Axboe 3223f72b8792SJens Axboe do { 3224f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3225f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3226f72b8792SJens Axboe 3227f72b8792SJens Axboe if (unlikely(ret < 0)) 3228f72b8792SJens Axboe return false; 3229f72b8792SJens Axboe 3230f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3231f72b8792SJens Axboe local_irq_restore(flags); 3232f72b8792SJens Axboe return ret; 3233f72b8792SJens Axboe } 3234f72b8792SJens Axboe 3235401a8d04STejun Heo /** 323657b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 323757b30ae7STejun Heo * @dwork: delayed_work to cancel 323809383498STejun Heo * 3239d185af30SYacine Belkadi * Kill off a pending delayed_work. 3240d185af30SYacine Belkadi * 3241d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3242d185af30SYacine Belkadi * pending. 3243d185af30SYacine Belkadi * 3244d185af30SYacine Belkadi * Note: 3245d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3246d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3247d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 324809383498STejun Heo * 324957b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 325009383498STejun Heo */ 325157b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 325209383498STejun Heo { 3253f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 325409383498STejun Heo } 325557b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 325609383498STejun Heo 325709383498STejun Heo /** 3258401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3259401a8d04STejun Heo * @dwork: the delayed work cancel 3260401a8d04STejun Heo * 3261401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3262401a8d04STejun Heo * 3263d185af30SYacine Belkadi * Return: 3264401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3265401a8d04STejun Heo */ 3266401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 32676e84d644SOleg Nesterov { 326836e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 32696e84d644SOleg Nesterov } 3270f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 32711da177e4SLinus Torvalds 32720fcb78c2SRolf Eike Beer /** 327331ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3274b6136773SAndrew Morton * @func: the function to call 3275b6136773SAndrew Morton * 327631ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 327731ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3278b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 327931ddd871STejun Heo * 3280d185af30SYacine Belkadi * Return: 328131ddd871STejun Heo * 0 on success, -errno on failure. 3282b6136773SAndrew Morton */ 328365f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 328415316ba8SChristoph Lameter { 328515316ba8SChristoph Lameter int cpu; 328638f51568SNamhyung Kim struct work_struct __percpu *works; 328715316ba8SChristoph Lameter 3288b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3289b6136773SAndrew Morton if (!works) 329015316ba8SChristoph Lameter return -ENOMEM; 3291b6136773SAndrew Morton 329295402b38SGautham R Shenoy get_online_cpus(); 329393981800STejun Heo 329415316ba8SChristoph Lameter for_each_online_cpu(cpu) { 32959bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 32969bfb1839SIngo Molnar 32979bfb1839SIngo Molnar INIT_WORK(work, func); 32988de6d308SOleg Nesterov schedule_work_on(cpu, work); 329915316ba8SChristoph Lameter } 330093981800STejun Heo 330193981800STejun Heo for_each_online_cpu(cpu) 33028616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 330393981800STejun Heo 330495402b38SGautham R Shenoy put_online_cpus(); 3305b6136773SAndrew Morton free_percpu(works); 330615316ba8SChristoph Lameter return 0; 330715316ba8SChristoph Lameter } 330815316ba8SChristoph Lameter 3309eef6a7d5SAlan Stern /** 33101fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 33111fa44ecaSJames Bottomley * @fn: the function to execute 33121fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 33131fa44ecaSJames Bottomley * be available when the work executes) 33141fa44ecaSJames Bottomley * 33151fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 33161fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 33171fa44ecaSJames Bottomley * 3318d185af30SYacine Belkadi * Return: 0 - function was executed 33191fa44ecaSJames Bottomley * 1 - function was scheduled for execution 33201fa44ecaSJames Bottomley */ 332165f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 33221fa44ecaSJames Bottomley { 33231fa44ecaSJames Bottomley if (!in_interrupt()) { 332465f27f38SDavid Howells fn(&ew->work); 33251fa44ecaSJames Bottomley return 0; 33261fa44ecaSJames Bottomley } 33271fa44ecaSJames Bottomley 332865f27f38SDavid Howells INIT_WORK(&ew->work, fn); 33291fa44ecaSJames Bottomley schedule_work(&ew->work); 33301fa44ecaSJames Bottomley 33311fa44ecaSJames Bottomley return 1; 33321fa44ecaSJames Bottomley } 33331fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 33341fa44ecaSJames Bottomley 33357a4e344cSTejun Heo /** 33367a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33377a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33387a4e344cSTejun Heo * 33397a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33407a4e344cSTejun Heo */ 3341513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 33427a4e344cSTejun Heo { 33437a4e344cSTejun Heo if (attrs) { 33447a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33457a4e344cSTejun Heo kfree(attrs); 33467a4e344cSTejun Heo } 33477a4e344cSTejun Heo } 33487a4e344cSTejun Heo 33497a4e344cSTejun Heo /** 33507a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33517a4e344cSTejun Heo * 33527a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3353d185af30SYacine Belkadi * return it. 3354d185af30SYacine Belkadi * 3355d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33567a4e344cSTejun Heo */ 3357513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 33587a4e344cSTejun Heo { 33597a4e344cSTejun Heo struct workqueue_attrs *attrs; 33607a4e344cSTejun Heo 3361be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 33627a4e344cSTejun Heo if (!attrs) 33637a4e344cSTejun Heo goto fail; 3364be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 33657a4e344cSTejun Heo goto fail; 33667a4e344cSTejun Heo 336713e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33687a4e344cSTejun Heo return attrs; 33697a4e344cSTejun Heo fail: 33707a4e344cSTejun Heo free_workqueue_attrs(attrs); 33717a4e344cSTejun Heo return NULL; 33727a4e344cSTejun Heo } 33737a4e344cSTejun Heo 337429c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 337529c91e99STejun Heo const struct workqueue_attrs *from) 337629c91e99STejun Heo { 337729c91e99STejun Heo to->nice = from->nice; 337829c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 33792865a8fbSShaohua Li /* 33802865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 33812865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 33822865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 33832865a8fbSShaohua Li */ 33842865a8fbSShaohua Li to->no_numa = from->no_numa; 338529c91e99STejun Heo } 338629c91e99STejun Heo 338729c91e99STejun Heo /* hash value of the content of @attr */ 338829c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 338929c91e99STejun Heo { 339029c91e99STejun Heo u32 hash = 0; 339129c91e99STejun Heo 339229c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 339313e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 339413e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 339529c91e99STejun Heo return hash; 339629c91e99STejun Heo } 339729c91e99STejun Heo 339829c91e99STejun Heo /* content equality test */ 339929c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 340029c91e99STejun Heo const struct workqueue_attrs *b) 340129c91e99STejun Heo { 340229c91e99STejun Heo if (a->nice != b->nice) 340329c91e99STejun Heo return false; 340429c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 340529c91e99STejun Heo return false; 340629c91e99STejun Heo return true; 340729c91e99STejun Heo } 340829c91e99STejun Heo 34097a4e344cSTejun Heo /** 34107a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34117a4e344cSTejun Heo * @pool: worker_pool to initialize 34127a4e344cSTejun Heo * 3413402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3414d185af30SYacine Belkadi * 3415d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 341629c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 341729c91e99STejun Heo * on @pool safely to release it. 34187a4e344cSTejun Heo */ 34197a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34204e1a1f9aSTejun Heo { 3421a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 342229c91e99STejun Heo pool->id = -1; 342329c91e99STejun Heo pool->cpu = -1; 3424f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34254e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 342682607adcSTejun Heo pool->watchdog_ts = jiffies; 34274e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34284e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34294e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34304e1a1f9aSTejun Heo 343132a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 34324e1a1f9aSTejun Heo 343332a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 34344e1a1f9aSTejun Heo 3435da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34367a4e344cSTejun Heo 34377cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 343829c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 343929c91e99STejun Heo pool->refcnt = 1; 344029c91e99STejun Heo 344129c91e99STejun Heo /* shouldn't fail above this point */ 3442be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 34437a4e344cSTejun Heo if (!pool->attrs) 34447a4e344cSTejun Heo return -ENOMEM; 34457a4e344cSTejun Heo return 0; 34464e1a1f9aSTejun Heo } 34474e1a1f9aSTejun Heo 3448669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3449669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3450669de8bdSBart Van Assche { 3451669de8bdSBart Van Assche char *lock_name; 3452669de8bdSBart Van Assche 3453669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3454669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3455669de8bdSBart Van Assche if (!lock_name) 3456669de8bdSBart Van Assche lock_name = wq->name; 345769a106c0SQian Cai 345869a106c0SQian Cai wq->lock_name = lock_name; 3459669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3460669de8bdSBart Van Assche } 3461669de8bdSBart Van Assche 3462669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3463669de8bdSBart Van Assche { 3464669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3465669de8bdSBart Van Assche } 3466669de8bdSBart Van Assche 3467669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3468669de8bdSBart Van Assche { 3469669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3470669de8bdSBart Van Assche kfree(wq->lock_name); 3471669de8bdSBart Van Assche } 3472669de8bdSBart Van Assche #else 3473669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3474669de8bdSBart Van Assche { 3475669de8bdSBart Van Assche } 3476669de8bdSBart Van Assche 3477669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3478669de8bdSBart Van Assche { 3479669de8bdSBart Van Assche } 3480669de8bdSBart Van Assche 3481669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3482669de8bdSBart Van Assche { 3483669de8bdSBart Van Assche } 3484669de8bdSBart Van Assche #endif 3485669de8bdSBart Van Assche 3486e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3487e2dca7adSTejun Heo { 3488e2dca7adSTejun Heo struct workqueue_struct *wq = 3489e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3490e2dca7adSTejun Heo 3491669de8bdSBart Van Assche wq_free_lockdep(wq); 3492669de8bdSBart Van Assche 3493e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3494e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3495e2dca7adSTejun Heo else 3496e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3497e2dca7adSTejun Heo 3498e2dca7adSTejun Heo kfree(wq); 3499e2dca7adSTejun Heo } 3500e2dca7adSTejun Heo 350129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 350229c91e99STejun Heo { 350329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 350429c91e99STejun Heo 35057cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 350629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 350729c91e99STejun Heo kfree(pool); 350829c91e99STejun Heo } 350929c91e99STejun Heo 3510d8bb65abSSebastian Andrzej Siewior /* This returns with the lock held on success (pool manager is inactive). */ 3511d8bb65abSSebastian Andrzej Siewior static bool wq_manager_inactive(struct worker_pool *pool) 3512d8bb65abSSebastian Andrzej Siewior { 3513a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 3514d8bb65abSSebastian Andrzej Siewior 3515d8bb65abSSebastian Andrzej Siewior if (pool->flags & POOL_MANAGER_ACTIVE) { 3516a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3517d8bb65abSSebastian Andrzej Siewior return false; 3518d8bb65abSSebastian Andrzej Siewior } 3519d8bb65abSSebastian Andrzej Siewior return true; 3520d8bb65abSSebastian Andrzej Siewior } 3521d8bb65abSSebastian Andrzej Siewior 352229c91e99STejun Heo /** 352329c91e99STejun Heo * put_unbound_pool - put a worker_pool 352429c91e99STejun Heo * @pool: worker_pool to put 352529c91e99STejun Heo * 352624acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3527c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3528c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3529c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3530a892caccSTejun Heo * 3531a892caccSTejun Heo * Should be called with wq_pool_mutex held. 353229c91e99STejun Heo */ 353329c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 353429c91e99STejun Heo { 353560f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 353629c91e99STejun Heo struct worker *worker; 353729c91e99STejun Heo 3538a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3539a892caccSTejun Heo 3540a892caccSTejun Heo if (--pool->refcnt) 354129c91e99STejun Heo return; 354229c91e99STejun Heo 354329c91e99STejun Heo /* sanity checks */ 354461d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3545a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 354629c91e99STejun Heo return; 354729c91e99STejun Heo 354829c91e99STejun Heo /* release id and unhash */ 354929c91e99STejun Heo if (pool->id >= 0) 355029c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 355129c91e99STejun Heo hash_del(&pool->hash_node); 355229c91e99STejun Heo 3553c5aa87bbSTejun Heo /* 3554692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3555692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3556692b4825STejun Heo * manager and @pool gets freed with the flag set. 3557d8bb65abSSebastian Andrzej Siewior * Because of how wq_manager_inactive() works, we will hold the 3558d8bb65abSSebastian Andrzej Siewior * spinlock after a successful wait. 3559c5aa87bbSTejun Heo */ 3560d8bb65abSSebastian Andrzej Siewior rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool), 3561d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3562692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3563692b4825STejun Heo 35641037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 356529c91e99STejun Heo destroy_worker(worker); 356629c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3567a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 356860f5a4bcSLai Jiangshan 35691258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3570da028469SLai Jiangshan if (!list_empty(&pool->workers)) 357160f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 35721258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 357360f5a4bcSLai Jiangshan 357460f5a4bcSLai Jiangshan if (pool->detach_completion) 357560f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 357660f5a4bcSLai Jiangshan 357729c91e99STejun Heo /* shut down the timers */ 357829c91e99STejun Heo del_timer_sync(&pool->idle_timer); 357929c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 358029c91e99STejun Heo 358124acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 358225b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 358329c91e99STejun Heo } 358429c91e99STejun Heo 358529c91e99STejun Heo /** 358629c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 358729c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 358829c91e99STejun Heo * 358929c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 359029c91e99STejun Heo * reference count and return it. If there already is a matching 359129c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3592d185af30SYacine Belkadi * create a new one. 3593a892caccSTejun Heo * 3594a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3595d185af30SYacine Belkadi * 3596d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3597d185af30SYacine Belkadi * On failure, %NULL. 359829c91e99STejun Heo */ 359929c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 360029c91e99STejun Heo { 360129c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 360229c91e99STejun Heo struct worker_pool *pool; 3603f3f90ad4STejun Heo int node; 3604e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 360529c91e99STejun Heo 3606a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 360729c91e99STejun Heo 360829c91e99STejun Heo /* do we already have a matching pool? */ 360929c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 361029c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 361129c91e99STejun Heo pool->refcnt++; 36123fb1823cSLai Jiangshan return pool; 361329c91e99STejun Heo } 361429c91e99STejun Heo } 361529c91e99STejun Heo 3616e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3617e2273584SXunlei Pang if (wq_numa_enabled) { 3618e2273584SXunlei Pang for_each_node(node) { 3619e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3620e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3621e2273584SXunlei Pang target_node = node; 3622e2273584SXunlei Pang break; 3623e2273584SXunlei Pang } 3624e2273584SXunlei Pang } 3625e2273584SXunlei Pang } 3626e2273584SXunlei Pang 362729c91e99STejun Heo /* nope, create a new one */ 3628e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 362929c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 363029c91e99STejun Heo goto fail; 363129c91e99STejun Heo 36328864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 363329c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3634e2273584SXunlei Pang pool->node = target_node; 363529c91e99STejun Heo 36362865a8fbSShaohua Li /* 36372865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36382865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36392865a8fbSShaohua Li */ 36402865a8fbSShaohua Li pool->attrs->no_numa = false; 36412865a8fbSShaohua Li 364229c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 364329c91e99STejun Heo goto fail; 364429c91e99STejun Heo 364529c91e99STejun Heo /* create and start the initial worker */ 36463347fa09STejun Heo if (wq_online && !create_worker(pool)) 364729c91e99STejun Heo goto fail; 364829c91e99STejun Heo 364929c91e99STejun Heo /* install */ 365029c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 36513fb1823cSLai Jiangshan 365229c91e99STejun Heo return pool; 365329c91e99STejun Heo fail: 365429c91e99STejun Heo if (pool) 365529c91e99STejun Heo put_unbound_pool(pool); 365629c91e99STejun Heo return NULL; 365729c91e99STejun Heo } 365829c91e99STejun Heo 36598864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36608864b4e5STejun Heo { 36618864b4e5STejun Heo kmem_cache_free(pwq_cache, 36628864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36638864b4e5STejun Heo } 36648864b4e5STejun Heo 36658864b4e5STejun Heo /* 36668864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36678864b4e5STejun Heo * and needs to be destroyed. 36688864b4e5STejun Heo */ 36698864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36708864b4e5STejun Heo { 36718864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36728864b4e5STejun Heo unbound_release_work); 36738864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36748864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3675bc0caf09STejun Heo bool is_last; 36768864b4e5STejun Heo 36778864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36788864b4e5STejun Heo return; 36798864b4e5STejun Heo 36803c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 36818864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3682bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 36833c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 36848864b4e5STejun Heo 3685a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 36868864b4e5STejun Heo put_unbound_pool(pool); 3687a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3688a892caccSTejun Heo 368925b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 36908864b4e5STejun Heo 36918864b4e5STejun Heo /* 36928864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3693e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 36948864b4e5STejun Heo */ 3695669de8bdSBart Van Assche if (is_last) { 3696669de8bdSBart Van Assche wq_unregister_lockdep(wq); 369725b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 36986029a918STejun Heo } 3699669de8bdSBart Van Assche } 37008864b4e5STejun Heo 37010fbd95aaSTejun Heo /** 3702699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37030fbd95aaSTejun Heo * @pwq: target pool_workqueue 37040fbd95aaSTejun Heo * 3705699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3706699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3707699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37080fbd95aaSTejun Heo */ 3709699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37100fbd95aaSTejun Heo { 3711699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3712699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 37133347fa09STejun Heo unsigned long flags; 3714699ce097STejun Heo 3715699ce097STejun Heo /* for @wq->saved_max_active */ 3716a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3717699ce097STejun Heo 3718699ce097STejun Heo /* fast exit for non-freezable wqs */ 3719699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3720699ce097STejun Heo return; 3721699ce097STejun Heo 37223347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3723a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3724699ce097STejun Heo 372574b414eaSLai Jiangshan /* 372674b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 372774b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 372874b414eaSLai Jiangshan * is updated and visible. 372974b414eaSLai Jiangshan */ 373074b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3731699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37320fbd95aaSTejun Heo 37330fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 37340fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 37350fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3736951a078aSLai Jiangshan 3737951a078aSLai Jiangshan /* 3738951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3739951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3740951a078aSLai Jiangshan */ 3741951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3742699ce097STejun Heo } else { 3743699ce097STejun Heo pwq->max_active = 0; 3744699ce097STejun Heo } 3745699ce097STejun Heo 3746a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 37470fbd95aaSTejun Heo } 37480fbd95aaSTejun Heo 3749e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3750f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3751f147f29eSTejun Heo struct worker_pool *pool) 3752d2c1d404STejun Heo { 3753d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3754d2c1d404STejun Heo 3755e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3756e50aba9aSTejun Heo 3757d2c1d404STejun Heo pwq->pool = pool; 3758d2c1d404STejun Heo pwq->wq = wq; 3759d2c1d404STejun Heo pwq->flush_color = -1; 37608864b4e5STejun Heo pwq->refcnt = 1; 3761d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 37621befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3763d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37648864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3765f147f29eSTejun Heo } 3766d2c1d404STejun Heo 3767f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37681befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3769f147f29eSTejun Heo { 3770f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3771f147f29eSTejun Heo 3772f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 377375ccf595STejun Heo 37741befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37751befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37761befcf30STejun Heo return; 37771befcf30STejun Heo 377829b1cb41SLai Jiangshan /* set the matching work_color */ 377975ccf595STejun Heo pwq->work_color = wq->work_color; 3780983ca25eSTejun Heo 3781983ca25eSTejun Heo /* sync max_active to the current setting */ 3782983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3783983ca25eSTejun Heo 3784983ca25eSTejun Heo /* link in @pwq */ 37859e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3786df2d5ae4STejun Heo } 37876029a918STejun Heo 3788f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3789f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3790f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3791f147f29eSTejun Heo { 3792f147f29eSTejun Heo struct worker_pool *pool; 3793f147f29eSTejun Heo struct pool_workqueue *pwq; 3794f147f29eSTejun Heo 3795f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3796f147f29eSTejun Heo 3797f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3798f147f29eSTejun Heo if (!pool) 3799f147f29eSTejun Heo return NULL; 3800f147f29eSTejun Heo 3801e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3802f147f29eSTejun Heo if (!pwq) { 3803f147f29eSTejun Heo put_unbound_pool(pool); 3804f147f29eSTejun Heo return NULL; 3805f147f29eSTejun Heo } 3806f147f29eSTejun Heo 3807f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3808f147f29eSTejun Heo return pwq; 3809d2c1d404STejun Heo } 3810d2c1d404STejun Heo 38114c16bd32STejun Heo /** 381230186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3813042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 38144c16bd32STejun Heo * @node: the target NUMA node 38154c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38164c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38174c16bd32STejun Heo * 38184c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38194c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3820d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38214c16bd32STejun Heo * 38224c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38234c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38244c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38254c16bd32STejun Heo * @attrs->cpumask. 38264c16bd32STejun Heo * 38274c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38284c16bd32STejun Heo * stable. 3829d185af30SYacine Belkadi * 3830d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3831d185af30SYacine Belkadi * %false if equal. 38324c16bd32STejun Heo */ 38334c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38344c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38354c16bd32STejun Heo { 3836d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38374c16bd32STejun Heo goto use_dfl; 38384c16bd32STejun Heo 38394c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38404c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38414c16bd32STejun Heo if (cpu_going_down >= 0) 38424c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38434c16bd32STejun Heo 38444c16bd32STejun Heo if (cpumask_empty(cpumask)) 38454c16bd32STejun Heo goto use_dfl; 38464c16bd32STejun Heo 38474c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38484c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38491ad0f0a7SMichael Bringmann 38501ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 38511ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 38521ad0f0a7SMichael Bringmann "possible intersect\n"); 38531ad0f0a7SMichael Bringmann return false; 38541ad0f0a7SMichael Bringmann } 38551ad0f0a7SMichael Bringmann 38564c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 38574c16bd32STejun Heo 38584c16bd32STejun Heo use_dfl: 38594c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38604c16bd32STejun Heo return false; 38614c16bd32STejun Heo } 38624c16bd32STejun Heo 38631befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38641befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38651befcf30STejun Heo int node, 38661befcf30STejun Heo struct pool_workqueue *pwq) 38671befcf30STejun Heo { 38681befcf30STejun Heo struct pool_workqueue *old_pwq; 38691befcf30STejun Heo 38705b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 38711befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38721befcf30STejun Heo 38731befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 38741befcf30STejun Heo link_pwq(pwq); 38751befcf30STejun Heo 38761befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 38771befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 38781befcf30STejun Heo return old_pwq; 38791befcf30STejun Heo } 38801befcf30STejun Heo 38812d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 38822d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 38832d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 38842d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3885042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 38862d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 38872d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 38882d5f0764SLai Jiangshan }; 38892d5f0764SLai Jiangshan 38902d5f0764SLai Jiangshan /* free the resources after success or abort */ 38912d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 38922d5f0764SLai Jiangshan { 38932d5f0764SLai Jiangshan if (ctx) { 38942d5f0764SLai Jiangshan int node; 38952d5f0764SLai Jiangshan 38962d5f0764SLai Jiangshan for_each_node(node) 38972d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 38982d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 38992d5f0764SLai Jiangshan 39002d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 39012d5f0764SLai Jiangshan 39022d5f0764SLai Jiangshan kfree(ctx); 39032d5f0764SLai Jiangshan } 39042d5f0764SLai Jiangshan } 39052d5f0764SLai Jiangshan 39062d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 39072d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 39082d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 39092d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 39102d5f0764SLai Jiangshan { 39112d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 39122d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 39132d5f0764SLai Jiangshan int node; 39142d5f0764SLai Jiangshan 39152d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39162d5f0764SLai Jiangshan 3917acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 39182d5f0764SLai Jiangshan 3919be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 3920be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 39212d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 39222d5f0764SLai Jiangshan goto out_free; 39232d5f0764SLai Jiangshan 3924042f7df1SLai Jiangshan /* 3925042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3926042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3927042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3928042f7df1SLai Jiangshan */ 39292d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3930b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3931042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3932042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 39332d5f0764SLai Jiangshan 39342d5f0764SLai Jiangshan /* 39352d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 39362d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 39372d5f0764SLai Jiangshan * pools. 39382d5f0764SLai Jiangshan */ 39392d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 39402d5f0764SLai Jiangshan 39412d5f0764SLai Jiangshan /* 39422d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 39432d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 39442d5f0764SLai Jiangshan * it even if we don't use it immediately. 39452d5f0764SLai Jiangshan */ 39462d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39472d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 39482d5f0764SLai Jiangshan goto out_free; 39492d5f0764SLai Jiangshan 39502d5f0764SLai Jiangshan for_each_node(node) { 3951042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 39522d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39532d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 39542d5f0764SLai Jiangshan goto out_free; 39552d5f0764SLai Jiangshan } else { 39562d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 39572d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 39582d5f0764SLai Jiangshan } 39592d5f0764SLai Jiangshan } 39602d5f0764SLai Jiangshan 3961042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3962042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3963042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 39642d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3965042f7df1SLai Jiangshan 39662d5f0764SLai Jiangshan ctx->wq = wq; 39672d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 39682d5f0764SLai Jiangshan return ctx; 39692d5f0764SLai Jiangshan 39702d5f0764SLai Jiangshan out_free: 39712d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 39722d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 39732d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 39742d5f0764SLai Jiangshan return NULL; 39752d5f0764SLai Jiangshan } 39762d5f0764SLai Jiangshan 39772d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 39782d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 39792d5f0764SLai Jiangshan { 39802d5f0764SLai Jiangshan int node; 39812d5f0764SLai Jiangshan 39822d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 39832d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 39842d5f0764SLai Jiangshan 39852d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 39862d5f0764SLai Jiangshan 39872d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 39882d5f0764SLai Jiangshan for_each_node(node) 39892d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 39902d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 39912d5f0764SLai Jiangshan 39922d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 39932d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 39942d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 39952d5f0764SLai Jiangshan 39962d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 39972d5f0764SLai Jiangshan } 39982d5f0764SLai Jiangshan 3999a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4000a0111cf6SLai Jiangshan { 4001a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4002a0111cf6SLai Jiangshan get_online_cpus(); 4003a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4004a0111cf6SLai Jiangshan } 4005a0111cf6SLai Jiangshan 4006a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4007a0111cf6SLai Jiangshan { 4008a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4009a0111cf6SLai Jiangshan put_online_cpus(); 4010a0111cf6SLai Jiangshan } 4011a0111cf6SLai Jiangshan 4012a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4013a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4014a0111cf6SLai Jiangshan { 4015a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4016a0111cf6SLai Jiangshan 4017a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4018a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4019a0111cf6SLai Jiangshan return -EINVAL; 4020a0111cf6SLai Jiangshan 4021a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 40220a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 40230a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4024a0111cf6SLai Jiangshan return -EINVAL; 4025a0111cf6SLai Jiangshan 40260a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 40270a94efb5STejun Heo } 40280a94efb5STejun Heo 4029a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 40306201171eSwanghaibin if (!ctx) 40316201171eSwanghaibin return -ENOMEM; 4032a0111cf6SLai Jiangshan 4033a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4034a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4035a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4036a0111cf6SLai Jiangshan 40376201171eSwanghaibin return 0; 4038a0111cf6SLai Jiangshan } 4039a0111cf6SLai Jiangshan 40409e8cd2f5STejun Heo /** 40419e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 40429e8cd2f5STejun Heo * @wq: the target workqueue 40439e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 40449e8cd2f5STejun Heo * 40454c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 40464c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 40474c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 40484c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 40494c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 40504c16bd32STejun Heo * back-to-back will stay on its current pwq. 40519e8cd2f5STejun Heo * 4052d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4053d185af30SYacine Belkadi * 4054509b3204SDaniel Jordan * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus(). 4055509b3204SDaniel Jordan * 4056d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 40579e8cd2f5STejun Heo */ 4058513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 40599e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 40609e8cd2f5STejun Heo { 4061a0111cf6SLai Jiangshan int ret; 40629e8cd2f5STejun Heo 4063509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4064509b3204SDaniel Jordan 4065509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4066a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4067509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 40682d5f0764SLai Jiangshan 40692d5f0764SLai Jiangshan return ret; 40709e8cd2f5STejun Heo } 40719e8cd2f5STejun Heo 40724c16bd32STejun Heo /** 40734c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 40744c16bd32STejun Heo * @wq: the target workqueue 40754c16bd32STejun Heo * @cpu: the CPU coming up or going down 40764c16bd32STejun Heo * @online: whether @cpu is coming up or going down 40774c16bd32STejun Heo * 40784c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 40794c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 40804c16bd32STejun Heo * @wq accordingly. 40814c16bd32STejun Heo * 40824c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 40834c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 40844c16bd32STejun Heo * correct. 40854c16bd32STejun Heo * 40864c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 40874c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 40884c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 40894c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 40904c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 40914c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 40924c16bd32STejun Heo * CPU_DOWN_PREPARE. 40934c16bd32STejun Heo */ 40944c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 40954c16bd32STejun Heo bool online) 40964c16bd32STejun Heo { 40974c16bd32STejun Heo int node = cpu_to_node(cpu); 40984c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 40994c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 41004c16bd32STejun Heo struct workqueue_attrs *target_attrs; 41014c16bd32STejun Heo cpumask_t *cpumask; 41024c16bd32STejun Heo 41034c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 41044c16bd32STejun Heo 4105f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4106f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 41074c16bd32STejun Heo return; 41084c16bd32STejun Heo 41094c16bd32STejun Heo /* 41104c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 41114c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 41124c16bd32STejun Heo * CPU hotplug exclusion. 41134c16bd32STejun Heo */ 41144c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 41154c16bd32STejun Heo cpumask = target_attrs->cpumask; 41164c16bd32STejun Heo 41174c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 41184c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 41194c16bd32STejun Heo 41204c16bd32STejun Heo /* 41214c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4122042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4123042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4124042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 41254c16bd32STejun Heo */ 4126042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 41274c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4128f7142ed4SLai Jiangshan return; 41294c16bd32STejun Heo } else { 41304c16bd32STejun Heo goto use_dfl_pwq; 41314c16bd32STejun Heo } 41324c16bd32STejun Heo 41334c16bd32STejun Heo /* create a new pwq */ 41344c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41354c16bd32STejun Heo if (!pwq) { 41362d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41374c16bd32STejun Heo wq->name); 413877f300b1SDaeseok Youn goto use_dfl_pwq; 41394c16bd32STejun Heo } 41404c16bd32STejun Heo 4141f7142ed4SLai Jiangshan /* Install the new pwq. */ 41424c16bd32STejun Heo mutex_lock(&wq->mutex); 41434c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41444c16bd32STejun Heo goto out_unlock; 41454c16bd32STejun Heo 41464c16bd32STejun Heo use_dfl_pwq: 4147f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4148a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 41494c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4150a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 41514c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 41524c16bd32STejun Heo out_unlock: 41534c16bd32STejun Heo mutex_unlock(&wq->mutex); 41544c16bd32STejun Heo put_pwq_unlocked(old_pwq); 41554c16bd32STejun Heo } 41564c16bd32STejun Heo 415730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 41581da177e4SLinus Torvalds { 415949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 41608a2b7538STejun Heo int cpu, ret; 4161e1d8aa9fSFrederic Weisbecker 416230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4163420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4164420c0ddbSTejun Heo if (!wq->cpu_pwqs) 416530cdf249STejun Heo return -ENOMEM; 416630cdf249STejun Heo 416730cdf249STejun Heo for_each_possible_cpu(cpu) { 41687fb98ea7STejun Heo struct pool_workqueue *pwq = 41697fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 41707a62c2c8STejun Heo struct worker_pool *cpu_pools = 4171f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 417230cdf249STejun Heo 4173f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4174f147f29eSTejun Heo 4175f147f29eSTejun Heo mutex_lock(&wq->mutex); 41761befcf30STejun Heo link_pwq(pwq); 4177f147f29eSTejun Heo mutex_unlock(&wq->mutex); 417830cdf249STejun Heo } 417930cdf249STejun Heo return 0; 4180509b3204SDaniel Jordan } 4181509b3204SDaniel Jordan 4182509b3204SDaniel Jordan get_online_cpus(); 4183509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 41848a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 41858a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 41868a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 41878a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 41888a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 41899e8cd2f5STejun Heo } else { 4190509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 41919e8cd2f5STejun Heo } 4192509b3204SDaniel Jordan put_online_cpus(); 4193509b3204SDaniel Jordan 4194509b3204SDaniel Jordan return ret; 41950f900049STejun Heo } 41960f900049STejun Heo 4197f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4198f3421797STejun Heo const char *name) 4199b71ab8c2STejun Heo { 4200f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4201f3421797STejun Heo 4202f3421797STejun Heo if (max_active < 1 || max_active > lim) 4203044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4204f3421797STejun Heo max_active, name, 1, lim); 4205b71ab8c2STejun Heo 4206f3421797STejun Heo return clamp_val(max_active, 1, lim); 4207b71ab8c2STejun Heo } 4208b71ab8c2STejun Heo 4209983c7515STejun Heo /* 4210983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4211983c7515STejun Heo * to guarantee forward progress. 4212983c7515STejun Heo */ 4213983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4214983c7515STejun Heo { 4215983c7515STejun Heo struct worker *rescuer; 4216b92b36eaSDan Carpenter int ret; 4217983c7515STejun Heo 4218983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4219983c7515STejun Heo return 0; 4220983c7515STejun Heo 4221983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4222983c7515STejun Heo if (!rescuer) 4223983c7515STejun Heo return -ENOMEM; 4224983c7515STejun Heo 4225983c7515STejun Heo rescuer->rescue_wq = wq; 4226983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4227f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4228b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 4229983c7515STejun Heo kfree(rescuer); 4230b92b36eaSDan Carpenter return ret; 4231983c7515STejun Heo } 4232983c7515STejun Heo 4233983c7515STejun Heo wq->rescuer = rescuer; 4234983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4235983c7515STejun Heo wake_up_process(rescuer->task); 4236983c7515STejun Heo 4237983c7515STejun Heo return 0; 4238983c7515STejun Heo } 4239983c7515STejun Heo 4240a2775bbcSMathieu Malaterre __printf(1, 4) 4241669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 424297e37d7bSTejun Heo unsigned int flags, 4243669de8bdSBart Van Assche int max_active, ...) 42443af24433SOleg Nesterov { 4245df2d5ae4STejun Heo size_t tbl_size = 0; 4246ecf6881fSTejun Heo va_list args; 42473af24433SOleg Nesterov struct workqueue_struct *wq; 424849e3cf44STejun Heo struct pool_workqueue *pwq; 4249b196be89STejun Heo 42505c0338c6STejun Heo /* 42515c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 42525c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 42535c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 42545c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 42555c0338c6STejun Heo * on NUMA. 42565c0338c6STejun Heo */ 42575c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 42585c0338c6STejun Heo flags |= __WQ_ORDERED; 42595c0338c6STejun Heo 4260cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4261cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4262cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4263cee22a15SViresh Kumar 4264ecf6881fSTejun Heo /* allocate wq and format name */ 4265df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4266ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4267df2d5ae4STejun Heo 4268df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4269b196be89STejun Heo if (!wq) 4270d2c1d404STejun Heo return NULL; 4271b196be89STejun Heo 42726029a918STejun Heo if (flags & WQ_UNBOUND) { 4273be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 42746029a918STejun Heo if (!wq->unbound_attrs) 42756029a918STejun Heo goto err_free_wq; 42766029a918STejun Heo } 42776029a918STejun Heo 4278669de8bdSBart Van Assche va_start(args, max_active); 4279ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4280b196be89STejun Heo va_end(args); 42813af24433SOleg Nesterov 4282d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4283b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 42843af24433SOleg Nesterov 4285b196be89STejun Heo /* init wq */ 428697e37d7bSTejun Heo wq->flags = flags; 4287a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 42883c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4289112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 429030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 429173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 429273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4293493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 42943af24433SOleg Nesterov 4295669de8bdSBart Van Assche wq_init_lockdep(wq); 4296cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 42973af24433SOleg Nesterov 429830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 429982efcab3SBart Van Assche goto err_unreg_lockdep; 43001537663fSTejun Heo 430140c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4302d2c1d404STejun Heo goto err_destroy; 4303e22bee78STejun Heo 4304226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4305226223abSTejun Heo goto err_destroy; 4306226223abSTejun Heo 43076af8bf3dSOleg Nesterov /* 430868e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 430968e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 431068e13a67SLai Jiangshan * list. 43116af8bf3dSOleg Nesterov */ 431268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4313a0a1a5fdSTejun Heo 4314a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 431549e3cf44STejun Heo for_each_pwq(pwq, wq) 4316699ce097STejun Heo pwq_adjust_max_active(pwq); 4317a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4318a0a1a5fdSTejun Heo 4319e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4320a0a1a5fdSTejun Heo 432168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43223af24433SOleg Nesterov 43233af24433SOleg Nesterov return wq; 4324d2c1d404STejun Heo 432582efcab3SBart Van Assche err_unreg_lockdep: 4326009bb421SBart Van Assche wq_unregister_lockdep(wq); 4327009bb421SBart Van Assche wq_free_lockdep(wq); 432882efcab3SBart Van Assche err_free_wq: 43296029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 43304690c4abSTejun Heo kfree(wq); 4331d2c1d404STejun Heo return NULL; 4332d2c1d404STejun Heo err_destroy: 4333d2c1d404STejun Heo destroy_workqueue(wq); 43344690c4abSTejun Heo return NULL; 43351da177e4SLinus Torvalds } 4336669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 43371da177e4SLinus Torvalds 4338c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4339c29eb853STejun Heo { 4340c29eb853STejun Heo int i; 4341c29eb853STejun Heo 4342c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4343c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4344c29eb853STejun Heo return true; 4345c29eb853STejun Heo 4346c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4347c29eb853STejun Heo return true; 4348c29eb853STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 4349c29eb853STejun Heo return true; 4350c29eb853STejun Heo 4351c29eb853STejun Heo return false; 4352c29eb853STejun Heo } 4353c29eb853STejun Heo 43543af24433SOleg Nesterov /** 43553af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 43563af24433SOleg Nesterov * @wq: target workqueue 43573af24433SOleg Nesterov * 43583af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 43593af24433SOleg Nesterov */ 43603af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 43613af24433SOleg Nesterov { 436249e3cf44STejun Heo struct pool_workqueue *pwq; 43634c16bd32STejun Heo int node; 43643af24433SOleg Nesterov 4365def98c84STejun Heo /* 4366def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4367def98c84STejun Heo * lead to sysfs name conflicts. 4368def98c84STejun Heo */ 4369def98c84STejun Heo workqueue_sysfs_unregister(wq); 4370def98c84STejun Heo 43719c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 43729c5a2ba7STejun Heo drain_workqueue(wq); 4373c8efcc25STejun Heo 4374def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4375def98c84STejun Heo if (wq->rescuer) { 4376def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4377def98c84STejun Heo 4378def98c84STejun Heo /* this prevents new queueing */ 4379a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4380def98c84STejun Heo wq->rescuer = NULL; 4381a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4382def98c84STejun Heo 4383def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4384def98c84STejun Heo kthread_stop(rescuer->task); 43858efe1223STejun Heo kfree(rescuer); 4386def98c84STejun Heo } 4387def98c84STejun Heo 4388c29eb853STejun Heo /* 4389c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4390c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4391c29eb853STejun Heo */ 4392c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4393b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 439449e3cf44STejun Heo for_each_pwq(pwq, wq) { 4395a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4396c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 43971d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4398e66b39afSTejun Heo __func__, wq->name); 4399c29eb853STejun Heo show_pwq(pwq); 4400a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4401b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4402c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 4403fa07fb6aSTejun Heo show_workqueue_state(); 44046183c009STejun Heo return; 440576af4d93STejun Heo } 4406a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 440776af4d93STejun Heo } 4408b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 44096183c009STejun Heo 4410a0a1a5fdSTejun Heo /* 4411a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4412a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4413a0a1a5fdSTejun Heo */ 4414e2dca7adSTejun Heo list_del_rcu(&wq->list); 441568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44163af24433SOleg Nesterov 44178864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4418669de8bdSBart Van Assche wq_unregister_lockdep(wq); 441929c91e99STejun Heo /* 44208864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4421e2dca7adSTejun Heo * schedule RCU free. 442229c91e99STejun Heo */ 442325b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 44248864b4e5STejun Heo } else { 44258864b4e5STejun Heo /* 44268864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 44274c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 44284c16bd32STejun Heo * @wq will be freed when the last pwq is released. 44298864b4e5STejun Heo */ 44304c16bd32STejun Heo for_each_node(node) { 44314c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 44324c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 44334c16bd32STejun Heo put_pwq_unlocked(pwq); 44344c16bd32STejun Heo } 44354c16bd32STejun Heo 44364c16bd32STejun Heo /* 44374c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 44384c16bd32STejun Heo * put. Don't access it afterwards. 44394c16bd32STejun Heo */ 44404c16bd32STejun Heo pwq = wq->dfl_pwq; 44414c16bd32STejun Heo wq->dfl_pwq = NULL; 4442dce90d47STejun Heo put_pwq_unlocked(pwq); 444329c91e99STejun Heo } 44443af24433SOleg Nesterov } 44453af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 44463af24433SOleg Nesterov 4447dcd989cbSTejun Heo /** 4448dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4449dcd989cbSTejun Heo * @wq: target workqueue 4450dcd989cbSTejun Heo * @max_active: new max_active value. 4451dcd989cbSTejun Heo * 4452dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4453dcd989cbSTejun Heo * 4454dcd989cbSTejun Heo * CONTEXT: 4455dcd989cbSTejun Heo * Don't call from IRQ context. 4456dcd989cbSTejun Heo */ 4457dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4458dcd989cbSTejun Heo { 445949e3cf44STejun Heo struct pool_workqueue *pwq; 4460dcd989cbSTejun Heo 44618719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 44620a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 44638719dceaSTejun Heo return; 44648719dceaSTejun Heo 4465f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4466dcd989cbSTejun Heo 4467a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4468dcd989cbSTejun Heo 44690a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4470dcd989cbSTejun Heo wq->saved_max_active = max_active; 4471dcd989cbSTejun Heo 4472699ce097STejun Heo for_each_pwq(pwq, wq) 4473699ce097STejun Heo pwq_adjust_max_active(pwq); 4474dcd989cbSTejun Heo 4475a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4476dcd989cbSTejun Heo } 4477dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4478dcd989cbSTejun Heo 4479dcd989cbSTejun Heo /** 448027d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 448127d4ee03SLukas Wunner * 448227d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 448327d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 448427d4ee03SLukas Wunner * 448527d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 448627d4ee03SLukas Wunner */ 448727d4ee03SLukas Wunner struct work_struct *current_work(void) 448827d4ee03SLukas Wunner { 448927d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 449027d4ee03SLukas Wunner 449127d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 449227d4ee03SLukas Wunner } 449327d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 449427d4ee03SLukas Wunner 449527d4ee03SLukas Wunner /** 4496e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4497e6267616STejun Heo * 4498e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4499e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4500d185af30SYacine Belkadi * 4501d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4502e6267616STejun Heo */ 4503e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4504e6267616STejun Heo { 4505e6267616STejun Heo struct worker *worker = current_wq_worker(); 4506e6267616STejun Heo 45076a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4508e6267616STejun Heo } 4509e6267616STejun Heo 4510e6267616STejun Heo /** 4511dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4512dcd989cbSTejun Heo * @cpu: CPU in question 4513dcd989cbSTejun Heo * @wq: target workqueue 4514dcd989cbSTejun Heo * 4515dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4516dcd989cbSTejun Heo * no synchronization around this function and the test result is 4517dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4518dcd989cbSTejun Heo * 4519d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4520d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4521d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4522d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4523d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4524d3251859STejun Heo * 4525d185af30SYacine Belkadi * Return: 4526dcd989cbSTejun Heo * %true if congested, %false otherwise. 4527dcd989cbSTejun Heo */ 4528d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4529dcd989cbSTejun Heo { 45307fb98ea7STejun Heo struct pool_workqueue *pwq; 453176af4d93STejun Heo bool ret; 453276af4d93STejun Heo 453324acfb71SThomas Gleixner rcu_read_lock(); 453424acfb71SThomas Gleixner preempt_disable(); 45357fb98ea7STejun Heo 4536d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4537d3251859STejun Heo cpu = smp_processor_id(); 4538d3251859STejun Heo 45397fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 45407fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 45417fb98ea7STejun Heo else 4542df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4543dcd989cbSTejun Heo 454476af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 454524acfb71SThomas Gleixner preempt_enable(); 454624acfb71SThomas Gleixner rcu_read_unlock(); 454776af4d93STejun Heo 454876af4d93STejun Heo return ret; 4549dcd989cbSTejun Heo } 4550dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4551dcd989cbSTejun Heo 4552dcd989cbSTejun Heo /** 4553dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4554dcd989cbSTejun Heo * @work: the work to be tested 4555dcd989cbSTejun Heo * 4556dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4557dcd989cbSTejun Heo * synchronization around this function and the test result is 4558dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4559dcd989cbSTejun Heo * 4560d185af30SYacine Belkadi * Return: 4561dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4562dcd989cbSTejun Heo */ 4563dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4564dcd989cbSTejun Heo { 4565fa1b54e6STejun Heo struct worker_pool *pool; 4566dcd989cbSTejun Heo unsigned long flags; 4567dcd989cbSTejun Heo unsigned int ret = 0; 4568dcd989cbSTejun Heo 4569dcd989cbSTejun Heo if (work_pending(work)) 4570dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4571038366c5SLai Jiangshan 457224acfb71SThomas Gleixner rcu_read_lock(); 4573fa1b54e6STejun Heo pool = get_work_pool(work); 4574038366c5SLai Jiangshan if (pool) { 4575a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4576c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4577dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4578a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4579038366c5SLai Jiangshan } 458024acfb71SThomas Gleixner rcu_read_unlock(); 4581dcd989cbSTejun Heo 4582dcd989cbSTejun Heo return ret; 4583dcd989cbSTejun Heo } 4584dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4585dcd989cbSTejun Heo 45863d1cb205STejun Heo /** 45873d1cb205STejun Heo * set_worker_desc - set description for the current work item 45883d1cb205STejun Heo * @fmt: printf-style format string 45893d1cb205STejun Heo * @...: arguments for the format string 45903d1cb205STejun Heo * 45913d1cb205STejun Heo * This function can be called by a running work function to describe what 45923d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 45933d1cb205STejun Heo * information will be printed out together to help debugging. The 45943d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 45953d1cb205STejun Heo */ 45963d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 45973d1cb205STejun Heo { 45983d1cb205STejun Heo struct worker *worker = current_wq_worker(); 45993d1cb205STejun Heo va_list args; 46003d1cb205STejun Heo 46013d1cb205STejun Heo if (worker) { 46023d1cb205STejun Heo va_start(args, fmt); 46033d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 46043d1cb205STejun Heo va_end(args); 46053d1cb205STejun Heo } 46063d1cb205STejun Heo } 46075c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 46083d1cb205STejun Heo 46093d1cb205STejun Heo /** 46103d1cb205STejun Heo * print_worker_info - print out worker information and description 46113d1cb205STejun Heo * @log_lvl: the log level to use when printing 46123d1cb205STejun Heo * @task: target task 46133d1cb205STejun Heo * 46143d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 46153d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 46163d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 46173d1cb205STejun Heo * 46183d1cb205STejun Heo * This function can be safely called on any task as long as the 46193d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 46203d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 46213d1cb205STejun Heo */ 46223d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 46233d1cb205STejun Heo { 46243d1cb205STejun Heo work_func_t *fn = NULL; 46253d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 46263d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 46273d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 46283d1cb205STejun Heo struct workqueue_struct *wq = NULL; 46293d1cb205STejun Heo struct worker *worker; 46303d1cb205STejun Heo 46313d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 46323d1cb205STejun Heo return; 46333d1cb205STejun Heo 46343d1cb205STejun Heo /* 46353d1cb205STejun Heo * This function is called without any synchronization and @task 46363d1cb205STejun Heo * could be in any state. Be careful with dereferences. 46373d1cb205STejun Heo */ 4638e700591aSPetr Mladek worker = kthread_probe_data(task); 46393d1cb205STejun Heo 46403d1cb205STejun Heo /* 46418bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 46428bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 46433d1cb205STejun Heo */ 4644fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4645fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4646fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4647fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4648fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 46493d1cb205STejun Heo 46503d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4651d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 46528bf89593STejun Heo if (strcmp(name, desc)) 46533d1cb205STejun Heo pr_cont(" (%s)", desc); 46543d1cb205STejun Heo pr_cont("\n"); 46553d1cb205STejun Heo } 46563d1cb205STejun Heo } 46573d1cb205STejun Heo 46583494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 46593494fc30STejun Heo { 46603494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 46613494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 46623494fc30STejun Heo pr_cont(" node=%d", pool->node); 46633494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 46643494fc30STejun Heo } 46653494fc30STejun Heo 46663494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 46673494fc30STejun Heo { 46683494fc30STejun Heo if (work->func == wq_barrier_func) { 46693494fc30STejun Heo struct wq_barrier *barr; 46703494fc30STejun Heo 46713494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 46723494fc30STejun Heo 46733494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 46743494fc30STejun Heo task_pid_nr(barr->task)); 46753494fc30STejun Heo } else { 4676d75f773cSSakari Ailus pr_cont("%s %ps", comma ? "," : "", work->func); 46773494fc30STejun Heo } 46783494fc30STejun Heo } 46793494fc30STejun Heo 46803494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 46813494fc30STejun Heo { 46823494fc30STejun Heo struct worker_pool *pool = pwq->pool; 46833494fc30STejun Heo struct work_struct *work; 46843494fc30STejun Heo struct worker *worker; 46853494fc30STejun Heo bool has_in_flight = false, has_pending = false; 46863494fc30STejun Heo int bkt; 46873494fc30STejun Heo 46883494fc30STejun Heo pr_info(" pwq %d:", pool->id); 46893494fc30STejun Heo pr_cont_pool_info(pool); 46903494fc30STejun Heo 4691e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4692e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 46933494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 46943494fc30STejun Heo 46953494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 46963494fc30STejun Heo if (worker->current_pwq == pwq) { 46973494fc30STejun Heo has_in_flight = true; 46983494fc30STejun Heo break; 46993494fc30STejun Heo } 47003494fc30STejun Heo } 47013494fc30STejun Heo if (has_in_flight) { 47023494fc30STejun Heo bool comma = false; 47033494fc30STejun Heo 47043494fc30STejun Heo pr_info(" in-flight:"); 47053494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47063494fc30STejun Heo if (worker->current_pwq != pwq) 47073494fc30STejun Heo continue; 47083494fc30STejun Heo 4709d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 47103494fc30STejun Heo task_pid_nr(worker->task), 471130ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 47123494fc30STejun Heo worker->current_func); 47133494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 47143494fc30STejun Heo pr_cont_work(false, work); 47153494fc30STejun Heo comma = true; 47163494fc30STejun Heo } 47173494fc30STejun Heo pr_cont("\n"); 47183494fc30STejun Heo } 47193494fc30STejun Heo 47203494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47213494fc30STejun Heo if (get_work_pwq(work) == pwq) { 47223494fc30STejun Heo has_pending = true; 47233494fc30STejun Heo break; 47243494fc30STejun Heo } 47253494fc30STejun Heo } 47263494fc30STejun Heo if (has_pending) { 47273494fc30STejun Heo bool comma = false; 47283494fc30STejun Heo 47293494fc30STejun Heo pr_info(" pending:"); 47303494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47313494fc30STejun Heo if (get_work_pwq(work) != pwq) 47323494fc30STejun Heo continue; 47333494fc30STejun Heo 47343494fc30STejun Heo pr_cont_work(comma, work); 47353494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47363494fc30STejun Heo } 47373494fc30STejun Heo pr_cont("\n"); 47383494fc30STejun Heo } 47393494fc30STejun Heo 47403494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 47413494fc30STejun Heo bool comma = false; 47423494fc30STejun Heo 47433494fc30STejun Heo pr_info(" delayed:"); 47443494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 47453494fc30STejun Heo pr_cont_work(comma, work); 47463494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47473494fc30STejun Heo } 47483494fc30STejun Heo pr_cont("\n"); 47493494fc30STejun Heo } 47503494fc30STejun Heo } 47513494fc30STejun Heo 47523494fc30STejun Heo /** 47533494fc30STejun Heo * show_workqueue_state - dump workqueue state 47543494fc30STejun Heo * 47557b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 47567b776af6SRoger Lu * all busy workqueues and pools. 47573494fc30STejun Heo */ 47583494fc30STejun Heo void show_workqueue_state(void) 47593494fc30STejun Heo { 47603494fc30STejun Heo struct workqueue_struct *wq; 47613494fc30STejun Heo struct worker_pool *pool; 47623494fc30STejun Heo unsigned long flags; 47633494fc30STejun Heo int pi; 47643494fc30STejun Heo 476524acfb71SThomas Gleixner rcu_read_lock(); 47663494fc30STejun Heo 47673494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 47683494fc30STejun Heo 47693494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 47703494fc30STejun Heo struct pool_workqueue *pwq; 47713494fc30STejun Heo bool idle = true; 47723494fc30STejun Heo 47733494fc30STejun Heo for_each_pwq(pwq, wq) { 47743494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 47753494fc30STejun Heo idle = false; 47763494fc30STejun Heo break; 47773494fc30STejun Heo } 47783494fc30STejun Heo } 47793494fc30STejun Heo if (idle) 47803494fc30STejun Heo continue; 47813494fc30STejun Heo 47823494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 47833494fc30STejun Heo 47843494fc30STejun Heo for_each_pwq(pwq, wq) { 4785a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 47863494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 47873494fc30STejun Heo show_pwq(pwq); 4788a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 478962635ea8SSergey Senozhatsky /* 479062635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 479162635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 479262635ea8SSergey Senozhatsky * hard lockup. 479362635ea8SSergey Senozhatsky */ 479462635ea8SSergey Senozhatsky touch_nmi_watchdog(); 47953494fc30STejun Heo } 47963494fc30STejun Heo } 47973494fc30STejun Heo 47983494fc30STejun Heo for_each_pool(pool, pi) { 47993494fc30STejun Heo struct worker *worker; 48003494fc30STejun Heo bool first = true; 48013494fc30STejun Heo 4802a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 48033494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 48043494fc30STejun Heo goto next_pool; 48053494fc30STejun Heo 48063494fc30STejun Heo pr_info("pool %d:", pool->id); 48073494fc30STejun Heo pr_cont_pool_info(pool); 480882607adcSTejun Heo pr_cont(" hung=%us workers=%d", 480982607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 481082607adcSTejun Heo pool->nr_workers); 48113494fc30STejun Heo if (pool->manager) 48123494fc30STejun Heo pr_cont(" manager: %d", 48133494fc30STejun Heo task_pid_nr(pool->manager->task)); 48143494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 48153494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 48163494fc30STejun Heo task_pid_nr(worker->task)); 48173494fc30STejun Heo first = false; 48183494fc30STejun Heo } 48193494fc30STejun Heo pr_cont("\n"); 48203494fc30STejun Heo next_pool: 4821a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 482262635ea8SSergey Senozhatsky /* 482362635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 482462635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 482562635ea8SSergey Senozhatsky * hard lockup. 482662635ea8SSergey Senozhatsky */ 482762635ea8SSergey Senozhatsky touch_nmi_watchdog(); 48283494fc30STejun Heo } 48293494fc30STejun Heo 483024acfb71SThomas Gleixner rcu_read_unlock(); 48313494fc30STejun Heo } 48323494fc30STejun Heo 48336b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 48346b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 48356b59808bSTejun Heo { 48366b59808bSTejun Heo int off; 48376b59808bSTejun Heo 48386b59808bSTejun Heo /* always show the actual comm */ 48396b59808bSTejun Heo off = strscpy(buf, task->comm, size); 48406b59808bSTejun Heo if (off < 0) 48416b59808bSTejun Heo return; 48426b59808bSTejun Heo 4843197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 48446b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 48456b59808bSTejun Heo 4846197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4847197f6accSTejun Heo struct worker *worker = kthread_data(task); 4848197f6accSTejun Heo struct worker_pool *pool = worker->pool; 48496b59808bSTejun Heo 48506b59808bSTejun Heo if (pool) { 4851a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 48526b59808bSTejun Heo /* 4853197f6accSTejun Heo * ->desc tracks information (wq name or 4854197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4855197f6accSTejun Heo * current, prepend '+', otherwise '-'. 48566b59808bSTejun Heo */ 48576b59808bSTejun Heo if (worker->desc[0] != '\0') { 48586b59808bSTejun Heo if (worker->current_work) 48596b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 48606b59808bSTejun Heo worker->desc); 48616b59808bSTejun Heo else 48626b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 48636b59808bSTejun Heo worker->desc); 48646b59808bSTejun Heo } 4865a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 48666b59808bSTejun Heo } 4867197f6accSTejun Heo } 48686b59808bSTejun Heo 48696b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 48706b59808bSTejun Heo } 48716b59808bSTejun Heo 487266448bc2SMathieu Malaterre #ifdef CONFIG_SMP 487366448bc2SMathieu Malaterre 4874db7bccf4STejun Heo /* 4875db7bccf4STejun Heo * CPU hotplug. 4876db7bccf4STejun Heo * 4877e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4878112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4879706026c2STejun Heo * pool which make migrating pending and scheduled works very 4880e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 488194cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4882e22bee78STejun Heo * blocked draining impractical. 4883e22bee78STejun Heo * 488424647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4885628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4886628c78e7STejun Heo * cpu comes back online. 4887db7bccf4STejun Heo */ 4888db7bccf4STejun Heo 4889e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4890db7bccf4STejun Heo { 48914ce62e9eSTejun Heo struct worker_pool *pool; 4892db7bccf4STejun Heo struct worker *worker; 4893db7bccf4STejun Heo 4894f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 48951258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 4896a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4897e22bee78STejun Heo 4898f2d5a0eeSTejun Heo /* 489992f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 490094cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 490194cf58bbSTejun Heo * except for the ones which are still executing works from 490294cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 490394cf58bbSTejun Heo * this, they may become diasporas. 4904f2d5a0eeSTejun Heo */ 4905da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4906403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4907db7bccf4STejun Heo 490824647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4909f2d5a0eeSTejun Heo 4910a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 49111258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 4912e22bee78STejun Heo 4913e22bee78STejun Heo /* 4914eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4915eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4916eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4917eb283428SLai Jiangshan * from other cpus. 4918628c78e7STejun Heo */ 4919628c78e7STejun Heo schedule(); 4920628c78e7STejun Heo 4921628c78e7STejun Heo /* 4922eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4923eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4924eb283428SLai Jiangshan * and keep_working() are always true as long as the 4925eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4926eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4927eb283428SLai Jiangshan * are served by workers tied to the pool. 4928e22bee78STejun Heo */ 4929e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4930eb283428SLai Jiangshan 4931eb283428SLai Jiangshan /* 4932eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4933eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4934eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4935eb283428SLai Jiangshan */ 4936a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4937eb283428SLai Jiangshan wake_up_worker(pool); 4938a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 4939eb283428SLai Jiangshan } 4940db7bccf4STejun Heo } 4941db7bccf4STejun Heo 4942bd7c089eSTejun Heo /** 4943bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4944bd7c089eSTejun Heo * @pool: pool of interest 4945bd7c089eSTejun Heo * 4946a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4947bd7c089eSTejun Heo */ 4948bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4949bd7c089eSTejun Heo { 4950a9ab775bSTejun Heo struct worker *worker; 4951bd7c089eSTejun Heo 49521258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 4953bd7c089eSTejun Heo 4954bd7c089eSTejun Heo /* 4955a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4956a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4957402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4958a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4959a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4960bd7c089eSTejun Heo */ 4961da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4962a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4963a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4964a9ab775bSTejun Heo 4965a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4966f7c17d26SWanpeng Li 49673de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4968a9ab775bSTejun Heo 4969da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4970a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4971a9ab775bSTejun Heo 4972a9ab775bSTejun Heo /* 4973a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4974a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4975a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4976a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4977a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4978a9ab775bSTejun Heo * be bound before @pool->lock is released. 4979a9ab775bSTejun Heo */ 4980a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4981bd7c089eSTejun Heo wake_up_process(worker->task); 4982bd7c089eSTejun Heo 4983bd7c089eSTejun Heo /* 4984a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4985a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4986a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4987a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4988a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4989a9ab775bSTejun Heo * concurrency management. Note that when or whether 4990a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4991a9ab775bSTejun Heo * 4992c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 4993a9ab775bSTejun Heo * tested without holding any lock in 49946d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 4995a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4996a9ab775bSTejun Heo * management operations. 4997bd7c089eSTejun Heo */ 4998a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4999a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5000a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5001c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5002bd7c089eSTejun Heo } 5003a9ab775bSTejun Heo 5004a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5005bd7c089eSTejun Heo } 5006bd7c089eSTejun Heo 50077dbc725eSTejun Heo /** 50087dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 50097dbc725eSTejun Heo * @pool: unbound pool of interest 50107dbc725eSTejun Heo * @cpu: the CPU which is coming up 50117dbc725eSTejun Heo * 50127dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 50137dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 50147dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 50157dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 50167dbc725eSTejun Heo */ 50177dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 50187dbc725eSTejun Heo { 50197dbc725eSTejun Heo static cpumask_t cpumask; 50207dbc725eSTejun Heo struct worker *worker; 50217dbc725eSTejun Heo 50221258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 50237dbc725eSTejun Heo 50247dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 50257dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 50267dbc725eSTejun Heo return; 50277dbc725eSTejun Heo 50287dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 50297dbc725eSTejun Heo 50307dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5031da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5032d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 50337dbc725eSTejun Heo } 50347dbc725eSTejun Heo 50357ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 50361da177e4SLinus Torvalds { 50374ce62e9eSTejun Heo struct worker_pool *pool; 50381da177e4SLinus Torvalds 5039f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 50403ce63377STejun Heo if (pool->nr_workers) 50413ce63377STejun Heo continue; 5042051e1850SLai Jiangshan if (!create_worker(pool)) 50437ee681b2SThomas Gleixner return -ENOMEM; 50443af24433SOleg Nesterov } 50457ee681b2SThomas Gleixner return 0; 50467ee681b2SThomas Gleixner } 50471da177e4SLinus Torvalds 50487ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 50497ee681b2SThomas Gleixner { 50507ee681b2SThomas Gleixner struct worker_pool *pool; 50517ee681b2SThomas Gleixner struct workqueue_struct *wq; 50527ee681b2SThomas Gleixner int pi; 50537ee681b2SThomas Gleixner 505468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 50557dbc725eSTejun Heo 50567dbc725eSTejun Heo for_each_pool(pool, pi) { 50571258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 505894cf58bbSTejun Heo 5059f05b558dSLai Jiangshan if (pool->cpu == cpu) 506094cf58bbSTejun Heo rebind_workers(pool); 5061f05b558dSLai Jiangshan else if (pool->cpu < 0) 50627dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 506394cf58bbSTejun Heo 50641258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 506594cf58bbSTejun Heo } 50667dbc725eSTejun Heo 50674c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 50684c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 50694c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 50704c16bd32STejun Heo 507168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 50727ee681b2SThomas Gleixner return 0; 507365758202STejun Heo } 507465758202STejun Heo 50757ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 507665758202STejun Heo { 50774c16bd32STejun Heo struct workqueue_struct *wq; 50788db25e78STejun Heo 50794c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5080e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5081e8b3f8dbSLai Jiangshan return -1; 5082e8b3f8dbSLai Jiangshan 5083e8b3f8dbSLai Jiangshan unbind_workers(cpu); 50844c16bd32STejun Heo 50854c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 50864c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 50874c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 50884c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 50894c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 50904c16bd32STejun Heo 50917ee681b2SThomas Gleixner return 0; 509265758202STejun Heo } 509365758202STejun Heo 50942d3854a3SRusty Russell struct work_for_cpu { 5095ed48ece2STejun Heo struct work_struct work; 50962d3854a3SRusty Russell long (*fn)(void *); 50972d3854a3SRusty Russell void *arg; 50982d3854a3SRusty Russell long ret; 50992d3854a3SRusty Russell }; 51002d3854a3SRusty Russell 5101ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 51022d3854a3SRusty Russell { 5103ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5104ed48ece2STejun Heo 51052d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 51062d3854a3SRusty Russell } 51072d3854a3SRusty Russell 51082d3854a3SRusty Russell /** 510922aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 51102d3854a3SRusty Russell * @cpu: the cpu to run on 51112d3854a3SRusty Russell * @fn: the function to run 51122d3854a3SRusty Russell * @arg: the function arg 51132d3854a3SRusty Russell * 511431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 51156b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5116d185af30SYacine Belkadi * 5117d185af30SYacine Belkadi * Return: The value @fn returns. 51182d3854a3SRusty Russell */ 5119d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 51202d3854a3SRusty Russell { 5121ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 51222d3854a3SRusty Russell 5123ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5124ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 512512997d1aSBjorn Helgaas flush_work(&wfc.work); 5126440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 51272d3854a3SRusty Russell return wfc.ret; 51282d3854a3SRusty Russell } 51292d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 51300e8d6a93SThomas Gleixner 51310e8d6a93SThomas Gleixner /** 51320e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 51330e8d6a93SThomas Gleixner * @cpu: the cpu to run on 51340e8d6a93SThomas Gleixner * @fn: the function to run 51350e8d6a93SThomas Gleixner * @arg: the function argument 51360e8d6a93SThomas Gleixner * 51370e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 51380e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 51390e8d6a93SThomas Gleixner * 51400e8d6a93SThomas Gleixner * Return: The value @fn returns. 51410e8d6a93SThomas Gleixner */ 51420e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 51430e8d6a93SThomas Gleixner { 51440e8d6a93SThomas Gleixner long ret = -ENODEV; 51450e8d6a93SThomas Gleixner 51460e8d6a93SThomas Gleixner get_online_cpus(); 51470e8d6a93SThomas Gleixner if (cpu_online(cpu)) 51480e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 51490e8d6a93SThomas Gleixner put_online_cpus(); 51500e8d6a93SThomas Gleixner return ret; 51510e8d6a93SThomas Gleixner } 51520e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 51532d3854a3SRusty Russell #endif /* CONFIG_SMP */ 51542d3854a3SRusty Russell 5155a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5156e7577c50SRusty Russell 5157a0a1a5fdSTejun Heo /** 5158a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5159a0a1a5fdSTejun Heo * 516058a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5161c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 5162706026c2STejun Heo * pool->worklist. 5163a0a1a5fdSTejun Heo * 5164a0a1a5fdSTejun Heo * CONTEXT: 5165a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5166a0a1a5fdSTejun Heo */ 5167a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5168a0a1a5fdSTejun Heo { 516924b8a847STejun Heo struct workqueue_struct *wq; 517024b8a847STejun Heo struct pool_workqueue *pwq; 5171a0a1a5fdSTejun Heo 517268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5173a0a1a5fdSTejun Heo 51746183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5175a0a1a5fdSTejun Heo workqueue_freezing = true; 5176a0a1a5fdSTejun Heo 517724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5178a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5179699ce097STejun Heo for_each_pwq(pwq, wq) 5180699ce097STejun Heo pwq_adjust_max_active(pwq); 5181a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5182a1056305STejun Heo } 51835bcab335STejun Heo 518468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5185a0a1a5fdSTejun Heo } 5186a0a1a5fdSTejun Heo 5187a0a1a5fdSTejun Heo /** 518858a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5189a0a1a5fdSTejun Heo * 5190a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5191a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5192a0a1a5fdSTejun Heo * 5193a0a1a5fdSTejun Heo * CONTEXT: 519468e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5195a0a1a5fdSTejun Heo * 5196d185af30SYacine Belkadi * Return: 519758a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 519858a69cb4STejun Heo * is complete. 5199a0a1a5fdSTejun Heo */ 5200a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5201a0a1a5fdSTejun Heo { 5202a0a1a5fdSTejun Heo bool busy = false; 520324b8a847STejun Heo struct workqueue_struct *wq; 520424b8a847STejun Heo struct pool_workqueue *pwq; 5205a0a1a5fdSTejun Heo 520668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5207a0a1a5fdSTejun Heo 52086183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5209a0a1a5fdSTejun Heo 521024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 521124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 521224b8a847STejun Heo continue; 5213a0a1a5fdSTejun Heo /* 5214a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5215a0a1a5fdSTejun Heo * to peek without lock. 5216a0a1a5fdSTejun Heo */ 521724acfb71SThomas Gleixner rcu_read_lock(); 521824b8a847STejun Heo for_each_pwq(pwq, wq) { 52196183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5220112202d9STejun Heo if (pwq->nr_active) { 5221a0a1a5fdSTejun Heo busy = true; 522224acfb71SThomas Gleixner rcu_read_unlock(); 5223a0a1a5fdSTejun Heo goto out_unlock; 5224a0a1a5fdSTejun Heo } 5225a0a1a5fdSTejun Heo } 522624acfb71SThomas Gleixner rcu_read_unlock(); 5227a0a1a5fdSTejun Heo } 5228a0a1a5fdSTejun Heo out_unlock: 522968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5230a0a1a5fdSTejun Heo return busy; 5231a0a1a5fdSTejun Heo } 5232a0a1a5fdSTejun Heo 5233a0a1a5fdSTejun Heo /** 5234a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5235a0a1a5fdSTejun Heo * 5236a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5237706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5238a0a1a5fdSTejun Heo * 5239a0a1a5fdSTejun Heo * CONTEXT: 5240a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5241a0a1a5fdSTejun Heo */ 5242a0a1a5fdSTejun Heo void thaw_workqueues(void) 5243a0a1a5fdSTejun Heo { 524424b8a847STejun Heo struct workqueue_struct *wq; 524524b8a847STejun Heo struct pool_workqueue *pwq; 5246a0a1a5fdSTejun Heo 524768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5248a0a1a5fdSTejun Heo 5249a0a1a5fdSTejun Heo if (!workqueue_freezing) 5250a0a1a5fdSTejun Heo goto out_unlock; 5251a0a1a5fdSTejun Heo 525274b414eaSLai Jiangshan workqueue_freezing = false; 525324b8a847STejun Heo 525424b8a847STejun Heo /* restore max_active and repopulate worklist */ 525524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5256a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5257699ce097STejun Heo for_each_pwq(pwq, wq) 5258699ce097STejun Heo pwq_adjust_max_active(pwq); 5259a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 526024b8a847STejun Heo } 526124b8a847STejun Heo 5262a0a1a5fdSTejun Heo out_unlock: 526368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5264a0a1a5fdSTejun Heo } 5265a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5266a0a1a5fdSTejun Heo 5267042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5268042f7df1SLai Jiangshan { 5269042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5270042f7df1SLai Jiangshan int ret = 0; 5271042f7df1SLai Jiangshan struct workqueue_struct *wq; 5272042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5273042f7df1SLai Jiangshan 5274042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5275042f7df1SLai Jiangshan 5276042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5277042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5278042f7df1SLai Jiangshan continue; 5279042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5280042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5281042f7df1SLai Jiangshan continue; 5282042f7df1SLai Jiangshan 5283042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5284042f7df1SLai Jiangshan if (!ctx) { 5285042f7df1SLai Jiangshan ret = -ENOMEM; 5286042f7df1SLai Jiangshan break; 5287042f7df1SLai Jiangshan } 5288042f7df1SLai Jiangshan 5289042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5290042f7df1SLai Jiangshan } 5291042f7df1SLai Jiangshan 5292042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5293042f7df1SLai Jiangshan if (!ret) 5294042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5295042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5296042f7df1SLai Jiangshan } 5297042f7df1SLai Jiangshan 5298042f7df1SLai Jiangshan return ret; 5299042f7df1SLai Jiangshan } 5300042f7df1SLai Jiangshan 5301042f7df1SLai Jiangshan /** 5302042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5303042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5304042f7df1SLai Jiangshan * 5305042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5306042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5307042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5308042f7df1SLai Jiangshan * 5309042f7df1SLai Jiangshan * Retun: 0 - Success 5310042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5311042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5312042f7df1SLai Jiangshan */ 5313042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5314042f7df1SLai Jiangshan { 5315042f7df1SLai Jiangshan int ret = -EINVAL; 5316042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5317042f7df1SLai Jiangshan 5318042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 5319042f7df1SLai Jiangshan return -ENOMEM; 5320042f7df1SLai Jiangshan 5321c98a9805STal Shorer /* 5322c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5323c98a9805STal Shorer * If the user wishes to include them, we allow that. 5324c98a9805STal Shorer */ 5325042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5326042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5327a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5328042f7df1SLai Jiangshan 5329042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5330042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5331042f7df1SLai Jiangshan 5332042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5333042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5334042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5335042f7df1SLai Jiangshan 5336042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5337042f7df1SLai Jiangshan if (ret < 0) 5338042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5339042f7df1SLai Jiangshan 5340a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5341042f7df1SLai Jiangshan } 5342042f7df1SLai Jiangshan 5343042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 5344042f7df1SLai Jiangshan return ret; 5345042f7df1SLai Jiangshan } 5346042f7df1SLai Jiangshan 53476ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 53486ba94429SFrederic Weisbecker /* 53496ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 53506ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 53516ba94429SFrederic Weisbecker * following attributes. 53526ba94429SFrederic Weisbecker * 53536ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 53546ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 53556ba94429SFrederic Weisbecker * 53566ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 53576ba94429SFrederic Weisbecker * 53589a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 53596ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 53606ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 53619a19b463SWang Long * numa RW bool : whether enable NUMA affinity 53626ba94429SFrederic Weisbecker */ 53636ba94429SFrederic Weisbecker struct wq_device { 53646ba94429SFrederic Weisbecker struct workqueue_struct *wq; 53656ba94429SFrederic Weisbecker struct device dev; 53666ba94429SFrederic Weisbecker }; 53676ba94429SFrederic Weisbecker 53686ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 53696ba94429SFrederic Weisbecker { 53706ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 53716ba94429SFrederic Weisbecker 53726ba94429SFrederic Weisbecker return wq_dev->wq; 53736ba94429SFrederic Weisbecker } 53746ba94429SFrederic Weisbecker 53756ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 53766ba94429SFrederic Weisbecker char *buf) 53776ba94429SFrederic Weisbecker { 53786ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53796ba94429SFrederic Weisbecker 53806ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 53816ba94429SFrederic Weisbecker } 53826ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 53836ba94429SFrederic Weisbecker 53846ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 53856ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 53866ba94429SFrederic Weisbecker { 53876ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53886ba94429SFrederic Weisbecker 53896ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 53906ba94429SFrederic Weisbecker } 53916ba94429SFrederic Weisbecker 53926ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 53936ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 53946ba94429SFrederic Weisbecker size_t count) 53956ba94429SFrederic Weisbecker { 53966ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53976ba94429SFrederic Weisbecker int val; 53986ba94429SFrederic Weisbecker 53996ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 54006ba94429SFrederic Weisbecker return -EINVAL; 54016ba94429SFrederic Weisbecker 54026ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 54036ba94429SFrederic Weisbecker return count; 54046ba94429SFrederic Weisbecker } 54056ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 54066ba94429SFrederic Weisbecker 54076ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 54086ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 54096ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 54106ba94429SFrederic Weisbecker NULL, 54116ba94429SFrederic Weisbecker }; 54126ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 54136ba94429SFrederic Weisbecker 54146ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 54156ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54166ba94429SFrederic Weisbecker { 54176ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54186ba94429SFrederic Weisbecker const char *delim = ""; 54196ba94429SFrederic Weisbecker int node, written = 0; 54206ba94429SFrederic Weisbecker 542124acfb71SThomas Gleixner get_online_cpus(); 542224acfb71SThomas Gleixner rcu_read_lock(); 54236ba94429SFrederic Weisbecker for_each_node(node) { 54246ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 54256ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 54266ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 54276ba94429SFrederic Weisbecker delim = " "; 54286ba94429SFrederic Weisbecker } 54296ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 543024acfb71SThomas Gleixner rcu_read_unlock(); 543124acfb71SThomas Gleixner put_online_cpus(); 54326ba94429SFrederic Weisbecker 54336ba94429SFrederic Weisbecker return written; 54346ba94429SFrederic Weisbecker } 54356ba94429SFrederic Weisbecker 54366ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 54376ba94429SFrederic Weisbecker char *buf) 54386ba94429SFrederic Weisbecker { 54396ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54406ba94429SFrederic Weisbecker int written; 54416ba94429SFrederic Weisbecker 54426ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 54436ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 54446ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 54456ba94429SFrederic Weisbecker 54466ba94429SFrederic Weisbecker return written; 54476ba94429SFrederic Weisbecker } 54486ba94429SFrederic Weisbecker 54496ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 54506ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 54516ba94429SFrederic Weisbecker { 54526ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 54536ba94429SFrederic Weisbecker 5454899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5455899a94feSLai Jiangshan 5456be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 54576ba94429SFrederic Weisbecker if (!attrs) 54586ba94429SFrederic Weisbecker return NULL; 54596ba94429SFrederic Weisbecker 54606ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 54616ba94429SFrederic Weisbecker return attrs; 54626ba94429SFrederic Weisbecker } 54636ba94429SFrederic Weisbecker 54646ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 54656ba94429SFrederic Weisbecker const char *buf, size_t count) 54666ba94429SFrederic Weisbecker { 54676ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54686ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5469d4d3e257SLai Jiangshan int ret = -ENOMEM; 5470d4d3e257SLai Jiangshan 5471d4d3e257SLai Jiangshan apply_wqattrs_lock(); 54726ba94429SFrederic Weisbecker 54736ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 54746ba94429SFrederic Weisbecker if (!attrs) 5475d4d3e257SLai Jiangshan goto out_unlock; 54766ba94429SFrederic Weisbecker 54776ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 54786ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5479d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 54806ba94429SFrederic Weisbecker else 54816ba94429SFrederic Weisbecker ret = -EINVAL; 54826ba94429SFrederic Weisbecker 5483d4d3e257SLai Jiangshan out_unlock: 5484d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 54856ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 54866ba94429SFrederic Weisbecker return ret ?: count; 54876ba94429SFrederic Weisbecker } 54886ba94429SFrederic Weisbecker 54896ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 54906ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54916ba94429SFrederic Weisbecker { 54926ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54936ba94429SFrederic Weisbecker int written; 54946ba94429SFrederic Weisbecker 54956ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 54966ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 54976ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 54986ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 54996ba94429SFrederic Weisbecker return written; 55006ba94429SFrederic Weisbecker } 55016ba94429SFrederic Weisbecker 55026ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 55036ba94429SFrederic Weisbecker struct device_attribute *attr, 55046ba94429SFrederic Weisbecker const char *buf, size_t count) 55056ba94429SFrederic Weisbecker { 55066ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55076ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5508d4d3e257SLai Jiangshan int ret = -ENOMEM; 5509d4d3e257SLai Jiangshan 5510d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55116ba94429SFrederic Weisbecker 55126ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55136ba94429SFrederic Weisbecker if (!attrs) 5514d4d3e257SLai Jiangshan goto out_unlock; 55156ba94429SFrederic Weisbecker 55166ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 55176ba94429SFrederic Weisbecker if (!ret) 5518d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55196ba94429SFrederic Weisbecker 5520d4d3e257SLai Jiangshan out_unlock: 5521d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55226ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55236ba94429SFrederic Weisbecker return ret ?: count; 55246ba94429SFrederic Weisbecker } 55256ba94429SFrederic Weisbecker 55266ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 55276ba94429SFrederic Weisbecker char *buf) 55286ba94429SFrederic Weisbecker { 55296ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55306ba94429SFrederic Weisbecker int written; 55316ba94429SFrederic Weisbecker 55326ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55336ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 55346ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 55356ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55366ba94429SFrederic Weisbecker 55376ba94429SFrederic Weisbecker return written; 55386ba94429SFrederic Weisbecker } 55396ba94429SFrederic Weisbecker 55406ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 55416ba94429SFrederic Weisbecker const char *buf, size_t count) 55426ba94429SFrederic Weisbecker { 55436ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55446ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5545d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5546d4d3e257SLai Jiangshan 5547d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55486ba94429SFrederic Weisbecker 55496ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55506ba94429SFrederic Weisbecker if (!attrs) 5551d4d3e257SLai Jiangshan goto out_unlock; 55526ba94429SFrederic Weisbecker 55536ba94429SFrederic Weisbecker ret = -EINVAL; 55546ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 55556ba94429SFrederic Weisbecker attrs->no_numa = !v; 5556d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55576ba94429SFrederic Weisbecker } 55586ba94429SFrederic Weisbecker 5559d4d3e257SLai Jiangshan out_unlock: 5560d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55616ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55626ba94429SFrederic Weisbecker return ret ?: count; 55636ba94429SFrederic Weisbecker } 55646ba94429SFrederic Weisbecker 55656ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 55666ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 55676ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 55686ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 55696ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 55706ba94429SFrederic Weisbecker __ATTR_NULL, 55716ba94429SFrederic Weisbecker }; 55726ba94429SFrederic Weisbecker 55736ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 55746ba94429SFrederic Weisbecker .name = "workqueue", 55756ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 55766ba94429SFrederic Weisbecker }; 55776ba94429SFrederic Weisbecker 5578b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5579b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5580b05a7928SFrederic Weisbecker { 5581b05a7928SFrederic Weisbecker int written; 5582b05a7928SFrederic Weisbecker 5583042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5584b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5585b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5586042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5587b05a7928SFrederic Weisbecker 5588b05a7928SFrederic Weisbecker return written; 5589b05a7928SFrederic Weisbecker } 5590b05a7928SFrederic Weisbecker 5591042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5592042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5593042f7df1SLai Jiangshan { 5594042f7df1SLai Jiangshan cpumask_var_t cpumask; 5595042f7df1SLai Jiangshan int ret; 5596042f7df1SLai Jiangshan 5597042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5598042f7df1SLai Jiangshan return -ENOMEM; 5599042f7df1SLai Jiangshan 5600042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5601042f7df1SLai Jiangshan if (!ret) 5602042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5603042f7df1SLai Jiangshan 5604042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5605042f7df1SLai Jiangshan return ret ? ret : count; 5606042f7df1SLai Jiangshan } 5607042f7df1SLai Jiangshan 5608b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5609042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5610042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5611b05a7928SFrederic Weisbecker 56126ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 56136ba94429SFrederic Weisbecker { 5614b05a7928SFrederic Weisbecker int err; 5615b05a7928SFrederic Weisbecker 5616b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5617b05a7928SFrederic Weisbecker if (err) 5618b05a7928SFrederic Weisbecker return err; 5619b05a7928SFrederic Weisbecker 5620b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 56216ba94429SFrederic Weisbecker } 56226ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 56236ba94429SFrederic Weisbecker 56246ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 56256ba94429SFrederic Weisbecker { 56266ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 56276ba94429SFrederic Weisbecker 56286ba94429SFrederic Weisbecker kfree(wq_dev); 56296ba94429SFrederic Weisbecker } 56306ba94429SFrederic Weisbecker 56316ba94429SFrederic Weisbecker /** 56326ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 56336ba94429SFrederic Weisbecker * @wq: the workqueue to register 56346ba94429SFrederic Weisbecker * 56356ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 56366ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 56376ba94429SFrederic Weisbecker * which is the preferred method. 56386ba94429SFrederic Weisbecker * 56396ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 56406ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 56416ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 56426ba94429SFrederic Weisbecker * attributes. 56436ba94429SFrederic Weisbecker * 56446ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 56456ba94429SFrederic Weisbecker */ 56466ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 56476ba94429SFrederic Weisbecker { 56486ba94429SFrederic Weisbecker struct wq_device *wq_dev; 56496ba94429SFrederic Weisbecker int ret; 56506ba94429SFrederic Weisbecker 56516ba94429SFrederic Weisbecker /* 5652402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 56536ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 56546ba94429SFrederic Weisbecker * workqueues. 56556ba94429SFrederic Weisbecker */ 56560a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 56576ba94429SFrederic Weisbecker return -EINVAL; 56586ba94429SFrederic Weisbecker 56596ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 56606ba94429SFrederic Weisbecker if (!wq_dev) 56616ba94429SFrederic Weisbecker return -ENOMEM; 56626ba94429SFrederic Weisbecker 56636ba94429SFrederic Weisbecker wq_dev->wq = wq; 56646ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 56656ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 566623217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 56676ba94429SFrederic Weisbecker 56686ba94429SFrederic Weisbecker /* 56696ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 56706ba94429SFrederic Weisbecker * everything is ready. 56716ba94429SFrederic Weisbecker */ 56726ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 56736ba94429SFrederic Weisbecker 56746ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 56756ba94429SFrederic Weisbecker if (ret) { 5676537f4146SArvind Yadav put_device(&wq_dev->dev); 56776ba94429SFrederic Weisbecker wq->wq_dev = NULL; 56786ba94429SFrederic Weisbecker return ret; 56796ba94429SFrederic Weisbecker } 56806ba94429SFrederic Weisbecker 56816ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 56826ba94429SFrederic Weisbecker struct device_attribute *attr; 56836ba94429SFrederic Weisbecker 56846ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 56856ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 56866ba94429SFrederic Weisbecker if (ret) { 56876ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 56886ba94429SFrederic Weisbecker wq->wq_dev = NULL; 56896ba94429SFrederic Weisbecker return ret; 56906ba94429SFrederic Weisbecker } 56916ba94429SFrederic Weisbecker } 56926ba94429SFrederic Weisbecker } 56936ba94429SFrederic Weisbecker 56946ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 56956ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 56966ba94429SFrederic Weisbecker return 0; 56976ba94429SFrederic Weisbecker } 56986ba94429SFrederic Weisbecker 56996ba94429SFrederic Weisbecker /** 57006ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 57016ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 57026ba94429SFrederic Weisbecker * 57036ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 57046ba94429SFrederic Weisbecker */ 57056ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 57066ba94429SFrederic Weisbecker { 57076ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 57086ba94429SFrederic Weisbecker 57096ba94429SFrederic Weisbecker if (!wq->wq_dev) 57106ba94429SFrederic Weisbecker return; 57116ba94429SFrederic Weisbecker 57126ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57136ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57146ba94429SFrederic Weisbecker } 57156ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 57166ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 57176ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 57186ba94429SFrederic Weisbecker 571982607adcSTejun Heo /* 572082607adcSTejun Heo * Workqueue watchdog. 572182607adcSTejun Heo * 572282607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 572382607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 572482607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 572582607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 572682607adcSTejun Heo * largely opaque. 572782607adcSTejun Heo * 572882607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 572982607adcSTejun Heo * state if some pools failed to make forward progress for a while where 573082607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 573182607adcSTejun Heo * 573282607adcSTejun Heo * This mechanism is controlled through the kernel parameter 573382607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 573482607adcSTejun Heo * corresponding sysfs parameter file. 573582607adcSTejun Heo */ 573682607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 573782607adcSTejun Heo 573882607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 57395cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 574082607adcSTejun Heo 574182607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 574282607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 574382607adcSTejun Heo 574482607adcSTejun Heo static void wq_watchdog_reset_touched(void) 574582607adcSTejun Heo { 574682607adcSTejun Heo int cpu; 574782607adcSTejun Heo 574882607adcSTejun Heo wq_watchdog_touched = jiffies; 574982607adcSTejun Heo for_each_possible_cpu(cpu) 575082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 575182607adcSTejun Heo } 575282607adcSTejun Heo 57535cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 575482607adcSTejun Heo { 575582607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 575682607adcSTejun Heo bool lockup_detected = false; 575782607adcSTejun Heo struct worker_pool *pool; 575882607adcSTejun Heo int pi; 575982607adcSTejun Heo 576082607adcSTejun Heo if (!thresh) 576182607adcSTejun Heo return; 576282607adcSTejun Heo 576382607adcSTejun Heo rcu_read_lock(); 576482607adcSTejun Heo 576582607adcSTejun Heo for_each_pool(pool, pi) { 576682607adcSTejun Heo unsigned long pool_ts, touched, ts; 576782607adcSTejun Heo 576882607adcSTejun Heo if (list_empty(&pool->worklist)) 576982607adcSTejun Heo continue; 577082607adcSTejun Heo 577182607adcSTejun Heo /* get the latest of pool and touched timestamps */ 577282607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 577382607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 577482607adcSTejun Heo 577582607adcSTejun Heo if (time_after(pool_ts, touched)) 577682607adcSTejun Heo ts = pool_ts; 577782607adcSTejun Heo else 577882607adcSTejun Heo ts = touched; 577982607adcSTejun Heo 578082607adcSTejun Heo if (pool->cpu >= 0) { 578182607adcSTejun Heo unsigned long cpu_touched = 578282607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 578382607adcSTejun Heo pool->cpu)); 578482607adcSTejun Heo if (time_after(cpu_touched, ts)) 578582607adcSTejun Heo ts = cpu_touched; 578682607adcSTejun Heo } 578782607adcSTejun Heo 578882607adcSTejun Heo /* did we stall? */ 578982607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 579082607adcSTejun Heo lockup_detected = true; 579182607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 579282607adcSTejun Heo pr_cont_pool_info(pool); 579382607adcSTejun Heo pr_cont(" stuck for %us!\n", 579482607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 579582607adcSTejun Heo } 579682607adcSTejun Heo } 579782607adcSTejun Heo 579882607adcSTejun Heo rcu_read_unlock(); 579982607adcSTejun Heo 580082607adcSTejun Heo if (lockup_detected) 580182607adcSTejun Heo show_workqueue_state(); 580282607adcSTejun Heo 580382607adcSTejun Heo wq_watchdog_reset_touched(); 580482607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 580582607adcSTejun Heo } 580682607adcSTejun Heo 5807cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 580882607adcSTejun Heo { 580982607adcSTejun Heo if (cpu >= 0) 581082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 581182607adcSTejun Heo else 581282607adcSTejun Heo wq_watchdog_touched = jiffies; 581382607adcSTejun Heo } 581482607adcSTejun Heo 581582607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 581682607adcSTejun Heo { 581782607adcSTejun Heo wq_watchdog_thresh = 0; 581882607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 581982607adcSTejun Heo 582082607adcSTejun Heo if (thresh) { 582182607adcSTejun Heo wq_watchdog_thresh = thresh; 582282607adcSTejun Heo wq_watchdog_reset_touched(); 582382607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 582482607adcSTejun Heo } 582582607adcSTejun Heo } 582682607adcSTejun Heo 582782607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 582882607adcSTejun Heo const struct kernel_param *kp) 582982607adcSTejun Heo { 583082607adcSTejun Heo unsigned long thresh; 583182607adcSTejun Heo int ret; 583282607adcSTejun Heo 583382607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 583482607adcSTejun Heo if (ret) 583582607adcSTejun Heo return ret; 583682607adcSTejun Heo 583782607adcSTejun Heo if (system_wq) 583882607adcSTejun Heo wq_watchdog_set_thresh(thresh); 583982607adcSTejun Heo else 584082607adcSTejun Heo wq_watchdog_thresh = thresh; 584182607adcSTejun Heo 584282607adcSTejun Heo return 0; 584382607adcSTejun Heo } 584482607adcSTejun Heo 584582607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 584682607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 584782607adcSTejun Heo .get = param_get_ulong, 584882607adcSTejun Heo }; 584982607adcSTejun Heo 585082607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 585182607adcSTejun Heo 0644); 585282607adcSTejun Heo 585382607adcSTejun Heo static void wq_watchdog_init(void) 585482607adcSTejun Heo { 58555cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 585682607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 585782607adcSTejun Heo } 585882607adcSTejun Heo 585982607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 586082607adcSTejun Heo 586182607adcSTejun Heo static inline void wq_watchdog_init(void) { } 586282607adcSTejun Heo 586382607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 586482607adcSTejun Heo 5865bce90380STejun Heo static void __init wq_numa_init(void) 5866bce90380STejun Heo { 5867bce90380STejun Heo cpumask_var_t *tbl; 5868bce90380STejun Heo int node, cpu; 5869bce90380STejun Heo 5870bce90380STejun Heo if (num_possible_nodes() <= 1) 5871bce90380STejun Heo return; 5872bce90380STejun Heo 5873d55262c4STejun Heo if (wq_disable_numa) { 5874d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5875d55262c4STejun Heo return; 5876d55262c4STejun Heo } 5877d55262c4STejun Heo 5878be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 58794c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 58804c16bd32STejun Heo 5881bce90380STejun Heo /* 5882bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5883bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5884bce90380STejun Heo * fully initialized by now. 5885bce90380STejun Heo */ 58866396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5887bce90380STejun Heo BUG_ON(!tbl); 5888bce90380STejun Heo 5889bce90380STejun Heo for_each_node(node) 58905a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 58911be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5892bce90380STejun Heo 5893bce90380STejun Heo for_each_possible_cpu(cpu) { 5894bce90380STejun Heo node = cpu_to_node(cpu); 5895bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5896bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5897bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5898bce90380STejun Heo return; 5899bce90380STejun Heo } 5900bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5901bce90380STejun Heo } 5902bce90380STejun Heo 5903bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5904bce90380STejun Heo wq_numa_enabled = true; 5905bce90380STejun Heo } 5906bce90380STejun Heo 59073347fa09STejun Heo /** 59083347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 59093347fa09STejun Heo * 59103347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 59113347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 59123347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 59133347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 59143347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 59153347fa09STejun Heo * created and scheduled right before early initcalls. 59163347fa09STejun Heo */ 59172333e829SYu Chen void __init workqueue_init_early(void) 59181da177e4SLinus Torvalds { 59197a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 59201bda3f80SFrederic Weisbecker int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 59217a4e344cSTejun Heo int i, cpu; 5922c34056a3STejun Heo 592310cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5924e904e6c2STejun Heo 5925b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 59261bda3f80SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 5927b05a7928SFrederic Weisbecker 5928e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5929e904e6c2STejun Heo 5930706026c2STejun Heo /* initialize CPU pools */ 593129c91e99STejun Heo for_each_possible_cpu(cpu) { 59324ce62e9eSTejun Heo struct worker_pool *pool; 59338b03ae3cSTejun Heo 59347a4e344cSTejun Heo i = 0; 5935f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 59367a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5937ec22ca5eSTejun Heo pool->cpu = cpu; 59387a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 59397a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5940f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 59417a4e344cSTejun Heo 59429daf9e67STejun Heo /* alloc pool ID */ 594368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 59449daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 594568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 59464ce62e9eSTejun Heo } 59478b03ae3cSTejun Heo } 59488b03ae3cSTejun Heo 59498a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 595029c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 595129c91e99STejun Heo struct workqueue_attrs *attrs; 595229c91e99STejun Heo 5953be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 595429c91e99STejun Heo attrs->nice = std_nice[i]; 595529c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 59568a2b7538STejun Heo 59578a2b7538STejun Heo /* 59588a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 59598a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 59608a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 59618a2b7538STejun Heo */ 5962be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 59638a2b7538STejun Heo attrs->nice = std_nice[i]; 59648a2b7538STejun Heo attrs->no_numa = true; 59658a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 596629c91e99STejun Heo } 596729c91e99STejun Heo 5968d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 59691aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5970d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5971f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5972f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 597324d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 597424d51addSTejun Heo WQ_FREEZABLE, 0); 59750668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 59760668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 59770668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 59780668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 59790668106cSViresh Kumar 0); 59801aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 59810668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 59820668106cSViresh Kumar !system_power_efficient_wq || 59830668106cSViresh Kumar !system_freezable_power_efficient_wq); 59843347fa09STejun Heo } 59853347fa09STejun Heo 59863347fa09STejun Heo /** 59873347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 59883347fa09STejun Heo * 59893347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 59903347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 59913347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 59923347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 59933347fa09STejun Heo * with the initial workers and enable future kworker creations. 59943347fa09STejun Heo */ 59952333e829SYu Chen void __init workqueue_init(void) 59963347fa09STejun Heo { 59972186d9f9STejun Heo struct workqueue_struct *wq; 59983347fa09STejun Heo struct worker_pool *pool; 59993347fa09STejun Heo int cpu, bkt; 60003347fa09STejun Heo 60012186d9f9STejun Heo /* 60022186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 60032186d9f9STejun Heo * CPU to node mapping may not be available that early on some 60042186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 60052186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 60062186d9f9STejun Heo * affinity, fix them up. 600740c17f75STejun Heo * 600840c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 60092186d9f9STejun Heo */ 60102186d9f9STejun Heo wq_numa_init(); 60112186d9f9STejun Heo 60122186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 60132186d9f9STejun Heo 60142186d9f9STejun Heo for_each_possible_cpu(cpu) { 60152186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 60162186d9f9STejun Heo pool->node = cpu_to_node(cpu); 60172186d9f9STejun Heo } 60182186d9f9STejun Heo } 60192186d9f9STejun Heo 602040c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 60212186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 602240c17f75STejun Heo WARN(init_rescuer(wq), 602340c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 602440c17f75STejun Heo wq->name); 602540c17f75STejun Heo } 60262186d9f9STejun Heo 60272186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 60282186d9f9STejun Heo 60293347fa09STejun Heo /* create the initial workers */ 60303347fa09STejun Heo for_each_online_cpu(cpu) { 60313347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 60323347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 60333347fa09STejun Heo BUG_ON(!create_worker(pool)); 60343347fa09STejun Heo } 60353347fa09STejun Heo } 60363347fa09STejun Heo 60373347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 60383347fa09STejun Heo BUG_ON(!create_worker(pool)); 60393347fa09STejun Heo 60403347fa09STejun Heo wq_online = true; 604182607adcSTejun Heo wq_watchdog_init(); 60421da177e4SLinus Torvalds } 6043