11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 20b11895c4SLibin * normal work items and the other for high priority ones) and some extra 21b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 22b11895c4SLibin * number of these backing pools is dynamic. 23c54fce6eSTejun Heo * 24c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 251da177e4SLinus Torvalds */ 261da177e4SLinus Torvalds 279984de1aSPaul Gortmaker #include <linux/export.h> 281da177e4SLinus Torvalds #include <linux/kernel.h> 291da177e4SLinus Torvalds #include <linux/sched.h> 301da177e4SLinus Torvalds #include <linux/init.h> 311da177e4SLinus Torvalds #include <linux/signal.h> 321da177e4SLinus Torvalds #include <linux/completion.h> 331da177e4SLinus Torvalds #include <linux/workqueue.h> 341da177e4SLinus Torvalds #include <linux/slab.h> 351da177e4SLinus Torvalds #include <linux/cpu.h> 361da177e4SLinus Torvalds #include <linux/notifier.h> 371da177e4SLinus Torvalds #include <linux/kthread.h> 381fa44ecaSJames Bottomley #include <linux/hardirq.h> 3946934023SChristoph Lameter #include <linux/mempolicy.h> 40341a5958SRafael J. Wysocki #include <linux/freezer.h> 41d5abe669SPeter Zijlstra #include <linux/kallsyms.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> 51e22bee78STejun Heo 52ea138446STejun Heo #include "workqueue_internal.h" 531da177e4SLinus Torvalds 54c8e55f36STejun Heo enum { 55bc2ae0f5STejun Heo /* 5624647570STejun Heo * worker_pool flags 57bc2ae0f5STejun Heo * 5824647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 59bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 60bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 61bc2ae0f5STejun Heo * is in effect. 62bc2ae0f5STejun Heo * 63bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 64bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6524647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 66bc2ae0f5STejun Heo * 67bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 6892f9c5c4SLai Jiangshan * attach_mutex to avoid changing binding state while 694736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 70bc2ae0f5STejun Heo */ 7124647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 72db7bccf4STejun Heo 73c8e55f36STejun Heo /* worker flags */ 74c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 75c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 76e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 77fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 78f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 79a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 80e22bee78STejun Heo 81a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 82a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 83db7bccf4STejun Heo 84e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 854ce62e9eSTejun Heo 8629c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 87c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 88db7bccf4STejun Heo 89e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 90e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 91e22bee78STejun Heo 923233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 933233cdbdSTejun Heo /* call for help after 10ms 943233cdbdSTejun Heo (min two ticks) */ 95e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 96e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds /* 99e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1008698a745SDongsheng Yang * all cpus. Give MIN_NICE. 101e22bee78STejun Heo */ 1028698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1038698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 104ecf6881fSTejun Heo 105ecf6881fSTejun Heo WQ_NAME_LEN = 24, 106c8e55f36STejun Heo }; 107c8e55f36STejun Heo 1081da177e4SLinus Torvalds /* 1094690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1104690c4abSTejun Heo * 111e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 112e41e704bSTejun Heo * everyone else. 1134690c4abSTejun Heo * 114e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 115e22bee78STejun Heo * only be modified and accessed from the local cpu. 116e22bee78STejun Heo * 117d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1184690c4abSTejun Heo * 119d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 120d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 121d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 122d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 123e22bee78STejun Heo * 12492f9c5c4SLai Jiangshan * A: pool->attach_mutex protected. 125822d8405STejun Heo * 12668e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12776af4d93STejun Heo * 12868e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1295bcab335STejun Heo * 1303c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1313c25a55dSLai Jiangshan * 132b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1332e109a28STejun Heo * 1342e109a28STejun Heo * MD: wq_mayday_lock protected. 1354690c4abSTejun Heo */ 1364690c4abSTejun Heo 1372eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 138c34056a3STejun Heo 139bd7bdd43STejun Heo struct worker_pool { 140d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 141d84ff051STejun Heo int cpu; /* I: the associated cpu */ 142f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1439daf9e67STejun Heo int id; /* I: pool ID */ 14411ebea50STejun Heo unsigned int flags; /* X: flags */ 145bd7bdd43STejun Heo 146bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 147bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 148ea1abd61SLai Jiangshan 149ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 150bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 151bd7bdd43STejun Heo 152bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 153bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 154bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 155bd7bdd43STejun Heo 156c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 157c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 158c9e7cf27STejun Heo /* L: hash of busy workers */ 159c9e7cf27STejun Heo 160bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16134a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 1622607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 16392f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 16492f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 16560f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 166e19e397aSTejun Heo 1677cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 168e19e397aSTejun Heo 1697a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17068e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17168e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1727a4e344cSTejun Heo 173e19e397aSTejun Heo /* 174e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 175e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 176e19e397aSTejun Heo * cacheline. 177e19e397aSTejun Heo */ 178e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17929c91e99STejun Heo 18029c91e99STejun Heo /* 18129c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18229c91e99STejun Heo * from get_work_pool(). 18329c91e99STejun Heo */ 18429c91e99STejun Heo struct rcu_head rcu; 1858b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1868b03ae3cSTejun Heo 1878b03ae3cSTejun Heo /* 188112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 189112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 190112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 191112202d9STejun Heo * number of flag bits. 1921da177e4SLinus Torvalds */ 193112202d9STejun Heo struct pool_workqueue { 194bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1954690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19673f53c4aSTejun Heo int work_color; /* L: current color */ 19773f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1988864b4e5STejun Heo int refcnt; /* L: reference count */ 19973f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20073f53c4aSTejun Heo /* L: nr of in_flight works */ 2011e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 202a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2031e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2043c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2052e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2068864b4e5STejun Heo 2078864b4e5STejun Heo /* 2088864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2098864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2108864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 211b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2128864b4e5STejun Heo */ 2138864b4e5STejun Heo struct work_struct unbound_release_work; 2148864b4e5STejun Heo struct rcu_head rcu; 215e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* 21873f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21973f53c4aSTejun Heo */ 22073f53c4aSTejun Heo struct wq_flusher { 2213c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2223c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22373f53c4aSTejun Heo struct completion done; /* flush completion */ 22473f53c4aSTejun Heo }; 2251da177e4SLinus Torvalds 226226223abSTejun Heo struct wq_device; 227226223abSTejun Heo 22873f53c4aSTejun Heo /* 229c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 230c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2311da177e4SLinus Torvalds */ 2321da177e4SLinus Torvalds struct workqueue_struct { 2333c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 234e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 23573f53c4aSTejun Heo 2363c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2373c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2383c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 239112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2403c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2413c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2423c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24373f53c4aSTejun Heo 2442e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 245e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 246e22bee78STejun Heo 24787fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 248a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 249226223abSTejun Heo 2506029a918STejun Heo struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ 2514c16bd32STejun Heo struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ 2526029a918STejun Heo 253226223abSTejun Heo #ifdef CONFIG_SYSFS 254226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 255226223abSTejun Heo #endif 2564e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2574e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2584e6045f1SJohannes Berg #endif 259ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2602728fd2fSTejun Heo 261e2dca7adSTejun Heo /* 262e2dca7adSTejun Heo * Destruction of workqueue_struct is sched-RCU protected to allow 263e2dca7adSTejun Heo * walking the workqueues list without grabbing wq_pool_mutex. 264e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 265e2dca7adSTejun Heo */ 266e2dca7adSTejun Heo struct rcu_head rcu; 267e2dca7adSTejun Heo 2682728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2692728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2702728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 271df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2721da177e4SLinus Torvalds }; 2731da177e4SLinus Torvalds 274e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 275e904e6c2STejun Heo 276bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 277bce90380STejun Heo /* possible CPUs of each node */ 278bce90380STejun Heo 279d55262c4STejun Heo static bool wq_disable_numa; 280d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 281d55262c4STejun Heo 282cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 283cee22a15SViresh Kumar #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT 284cee22a15SViresh Kumar static bool wq_power_efficient = true; 285cee22a15SViresh Kumar #else 286cee22a15SViresh Kumar static bool wq_power_efficient; 287cee22a15SViresh Kumar #endif 288cee22a15SViresh Kumar 289cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 290cee22a15SViresh Kumar 291bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 292bce90380STejun Heo 2934c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2944c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2954c16bd32STejun Heo 29668e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2972e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2985bcab335STejun Heo 299e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30068e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3017d19c5ceSTejun Heo 3027d19c5ceSTejun Heo /* the per-cpu worker pools */ 3037d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 3047d19c5ceSTejun Heo cpu_worker_pools); 3057d19c5ceSTejun Heo 30668e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3077d19c5ceSTejun Heo 30868e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 30929c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 31029c91e99STejun Heo 311c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 31229c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 31329c91e99STejun Heo 3148a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3158a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3168a2b7538STejun Heo 317d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 318ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 319044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3201aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 321044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 322d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 323044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 324f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 325044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 32624d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3270668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3280668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3290668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3300668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 331d320c038STejun Heo 3327d19c5ceSTejun Heo static int worker_thread(void *__worker); 3337d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3347d19c5ceSTejun Heo const struct workqueue_attrs *from); 3356ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3367d19c5ceSTejun Heo 33797bd2347STejun Heo #define CREATE_TRACE_POINTS 33897bd2347STejun Heo #include <trace/events/workqueue.h> 33997bd2347STejun Heo 34068e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3415bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 34268e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 34368e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3445bcab335STejun Heo 345b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 34676af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 347b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 348b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 34976af4d93STejun Heo 350f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 351f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 352f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3537a62c2c8STejun Heo (pool)++) 3544ce62e9eSTejun Heo 35549e3cf44STejun Heo /** 35617116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 35717116969STejun Heo * @pool: iteration cursor 358611c92a0STejun Heo * @pi: integer used for iteration 359fa1b54e6STejun Heo * 36068e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 36168e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 36268e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 363fa1b54e6STejun Heo * 364fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 365fa1b54e6STejun Heo * ignored. 36617116969STejun Heo */ 367611c92a0STejun Heo #define for_each_pool(pool, pi) \ 368611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 36968e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 370fa1b54e6STejun Heo else 37117116969STejun Heo 37217116969STejun Heo /** 373822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 374822d8405STejun Heo * @worker: iteration cursor 375822d8405STejun Heo * @pool: worker_pool to iterate workers of 376822d8405STejun Heo * 37792f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 378822d8405STejun Heo * 379822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 380822d8405STejun Heo * ignored. 381822d8405STejun Heo */ 382da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 383da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 38492f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 385822d8405STejun Heo else 386822d8405STejun Heo 387822d8405STejun Heo /** 38849e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 38949e3cf44STejun Heo * @pwq: iteration cursor 39049e3cf44STejun Heo * @wq: the target workqueue 39176af4d93STejun Heo * 392b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 393794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 394794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 39576af4d93STejun Heo * 39676af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 39776af4d93STejun Heo * ignored. 39849e3cf44STejun Heo */ 39949e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 40076af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 401b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 40276af4d93STejun Heo else 403f3421797STejun Heo 404dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 405dc186ad7SThomas Gleixner 406dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 407dc186ad7SThomas Gleixner 40899777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 40999777288SStanislaw Gruszka { 41099777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 41199777288SStanislaw Gruszka } 41299777288SStanislaw Gruszka 413dc186ad7SThomas Gleixner /* 414dc186ad7SThomas Gleixner * fixup_init is called when: 415dc186ad7SThomas Gleixner * - an active object is initialized 416dc186ad7SThomas Gleixner */ 417dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 418dc186ad7SThomas Gleixner { 419dc186ad7SThomas Gleixner struct work_struct *work = addr; 420dc186ad7SThomas Gleixner 421dc186ad7SThomas Gleixner switch (state) { 422dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 423dc186ad7SThomas Gleixner cancel_work_sync(work); 424dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 425dc186ad7SThomas Gleixner return 1; 426dc186ad7SThomas Gleixner default: 427dc186ad7SThomas Gleixner return 0; 428dc186ad7SThomas Gleixner } 429dc186ad7SThomas Gleixner } 430dc186ad7SThomas Gleixner 431dc186ad7SThomas Gleixner /* 432dc186ad7SThomas Gleixner * fixup_activate is called when: 433dc186ad7SThomas Gleixner * - an active object is activated 434dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 435dc186ad7SThomas Gleixner */ 436dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 437dc186ad7SThomas Gleixner { 438dc186ad7SThomas Gleixner struct work_struct *work = addr; 439dc186ad7SThomas Gleixner 440dc186ad7SThomas Gleixner switch (state) { 441dc186ad7SThomas Gleixner 442dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 443dc186ad7SThomas Gleixner /* 444dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 445dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 446dc186ad7SThomas Gleixner * is tracked in the object tracker. 447dc186ad7SThomas Gleixner */ 44822df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 449dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 450dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 451dc186ad7SThomas Gleixner return 0; 452dc186ad7SThomas Gleixner } 453dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 454dc186ad7SThomas Gleixner return 0; 455dc186ad7SThomas Gleixner 456dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 457dc186ad7SThomas Gleixner WARN_ON(1); 458dc186ad7SThomas Gleixner 459dc186ad7SThomas Gleixner default: 460dc186ad7SThomas Gleixner return 0; 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner } 463dc186ad7SThomas Gleixner 464dc186ad7SThomas Gleixner /* 465dc186ad7SThomas Gleixner * fixup_free is called when: 466dc186ad7SThomas Gleixner * - an active object is freed 467dc186ad7SThomas Gleixner */ 468dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 469dc186ad7SThomas Gleixner { 470dc186ad7SThomas Gleixner struct work_struct *work = addr; 471dc186ad7SThomas Gleixner 472dc186ad7SThomas Gleixner switch (state) { 473dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 474dc186ad7SThomas Gleixner cancel_work_sync(work); 475dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 476dc186ad7SThomas Gleixner return 1; 477dc186ad7SThomas Gleixner default: 478dc186ad7SThomas Gleixner return 0; 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner } 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 483dc186ad7SThomas Gleixner .name = "work_struct", 48499777288SStanislaw Gruszka .debug_hint = work_debug_hint, 485dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 486dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 487dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 488dc186ad7SThomas Gleixner }; 489dc186ad7SThomas Gleixner 490dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 491dc186ad7SThomas Gleixner { 492dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 493dc186ad7SThomas Gleixner } 494dc186ad7SThomas Gleixner 495dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 496dc186ad7SThomas Gleixner { 497dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 498dc186ad7SThomas Gleixner } 499dc186ad7SThomas Gleixner 500dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 501dc186ad7SThomas Gleixner { 502dc186ad7SThomas Gleixner if (onstack) 503dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 504dc186ad7SThomas Gleixner else 505dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 506dc186ad7SThomas Gleixner } 507dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 508dc186ad7SThomas Gleixner 509dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 510dc186ad7SThomas Gleixner { 511dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 512dc186ad7SThomas Gleixner } 513dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 514dc186ad7SThomas Gleixner 515ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 516ea2e64f2SThomas Gleixner { 517ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 518ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 519ea2e64f2SThomas Gleixner } 520ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 521ea2e64f2SThomas Gleixner 522dc186ad7SThomas Gleixner #else 523dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 524dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 525dc186ad7SThomas Gleixner #endif 526dc186ad7SThomas Gleixner 5274e8b22bdSLi Bin /** 5284e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5294e8b22bdSLi Bin * @pool: the pool pointer of interest 5304e8b22bdSLi Bin * 5314e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5324e8b22bdSLi Bin * successfully, -errno on failure. 5334e8b22bdSLi Bin */ 5349daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5359daf9e67STejun Heo { 5369daf9e67STejun Heo int ret; 5379daf9e67STejun Heo 53868e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5395bcab335STejun Heo 5404e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5414e8b22bdSLi Bin GFP_KERNEL); 542229641a6STejun Heo if (ret >= 0) { 543e68035fbSTejun Heo pool->id = ret; 544229641a6STejun Heo return 0; 545229641a6STejun Heo } 5469daf9e67STejun Heo return ret; 5479daf9e67STejun Heo } 5489daf9e67STejun Heo 54976af4d93STejun Heo /** 550df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 551df2d5ae4STejun Heo * @wq: the target workqueue 552df2d5ae4STejun Heo * @node: the node ID 553df2d5ae4STejun Heo * 554df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 555df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 556df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 557d185af30SYacine Belkadi * 558d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 559df2d5ae4STejun Heo */ 560df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 561df2d5ae4STejun Heo int node) 562df2d5ae4STejun Heo { 563df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 564df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 565df2d5ae4STejun Heo } 566df2d5ae4STejun Heo 56773f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 56873f53c4aSTejun Heo { 56973f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 57073f53c4aSTejun Heo } 57173f53c4aSTejun Heo 57273f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 57373f53c4aSTejun Heo { 57473f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 57573f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 57673f53c4aSTejun Heo } 57773f53c4aSTejun Heo 57873f53c4aSTejun Heo static int work_next_color(int color) 57973f53c4aSTejun Heo { 58073f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 581a848e3b6SOleg Nesterov } 582a848e3b6SOleg Nesterov 5834594bf15SDavid Howells /* 584112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 585112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5867c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5877a22ad75STejun Heo * 588112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 589112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 590bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 591bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5927a22ad75STejun Heo * 593112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5947c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 595112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5967c3eed5cSTejun Heo * available only while the work item is queued. 597bbb68dfaSTejun Heo * 598bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 599bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 600bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 601bbb68dfaSTejun Heo * try to steal the PENDING bit. 6024594bf15SDavid Howells */ 6037a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6047a22ad75STejun Heo unsigned long flags) 6057a22ad75STejun Heo { 6066183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6077a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6087a22ad75STejun Heo } 6097a22ad75STejun Heo 610112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6114690c4abSTejun Heo unsigned long extra_flags) 612365970a1SDavid Howells { 613112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 614112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 615365970a1SDavid Howells } 616365970a1SDavid Howells 6174468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6184468a00fSLai Jiangshan int pool_id) 6194468a00fSLai Jiangshan { 6204468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6214468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6224468a00fSLai Jiangshan } 6234468a00fSLai Jiangshan 6247c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6257c3eed5cSTejun Heo int pool_id) 6264d707b9fSOleg Nesterov { 62723657bb1STejun Heo /* 62823657bb1STejun Heo * The following wmb is paired with the implied mb in 62923657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 63023657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 63123657bb1STejun Heo * owner. 63223657bb1STejun Heo */ 63323657bb1STejun Heo smp_wmb(); 6347c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6354d707b9fSOleg Nesterov } 6364d707b9fSOleg Nesterov 6377a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 638365970a1SDavid Howells { 6397c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6407c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6417a22ad75STejun Heo } 6427a22ad75STejun Heo 643112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6447a22ad75STejun Heo { 645e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6467a22ad75STejun Heo 647112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 648e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 649e120153dSTejun Heo else 650e120153dSTejun Heo return NULL; 6517a22ad75STejun Heo } 6527a22ad75STejun Heo 6537c3eed5cSTejun Heo /** 6547c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6557c3eed5cSTejun Heo * @work: the work item of interest 6567c3eed5cSTejun Heo * 65768e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 65868e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 65968e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 660fa1b54e6STejun Heo * 661fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 662fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 663fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 664fa1b54e6STejun Heo * returned pool is and stays online. 665d185af30SYacine Belkadi * 666d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6677c3eed5cSTejun Heo */ 6687c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6697a22ad75STejun Heo { 670e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6717c3eed5cSTejun Heo int pool_id; 6727a22ad75STejun Heo 67368e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 674fa1b54e6STejun Heo 675112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 676112202d9STejun Heo return ((struct pool_workqueue *) 6777c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6787a22ad75STejun Heo 6797c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6807c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6817a22ad75STejun Heo return NULL; 6827a22ad75STejun Heo 683fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6847c3eed5cSTejun Heo } 6857c3eed5cSTejun Heo 6867c3eed5cSTejun Heo /** 6877c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6887c3eed5cSTejun Heo * @work: the work item of interest 6897c3eed5cSTejun Heo * 690d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6917c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6927c3eed5cSTejun Heo */ 6937c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6947c3eed5cSTejun Heo { 69554d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6967c3eed5cSTejun Heo 697112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 698112202d9STejun Heo return ((struct pool_workqueue *) 69954d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 70054d5b7d0SLai Jiangshan 70154d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7027c3eed5cSTejun Heo } 7037c3eed5cSTejun Heo 704bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 705bbb68dfaSTejun Heo { 7067c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 707bbb68dfaSTejun Heo 7087c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7097c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 710bbb68dfaSTejun Heo } 711bbb68dfaSTejun Heo 712bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 713bbb68dfaSTejun Heo { 714bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 715bbb68dfaSTejun Heo 716112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 717bbb68dfaSTejun Heo } 718bbb68dfaSTejun Heo 719e22bee78STejun Heo /* 7203270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7213270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 722d565ed63STejun Heo * they're being called with pool->lock held. 723e22bee78STejun Heo */ 724e22bee78STejun Heo 72563d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 726649027d7STejun Heo { 727e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 728649027d7STejun Heo } 729649027d7STejun Heo 730e22bee78STejun Heo /* 731e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 732e22bee78STejun Heo * running workers. 733974271c4STejun Heo * 734974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 735706026c2STejun Heo * function will always return %true for unbound pools as long as the 736974271c4STejun Heo * worklist isn't empty. 737e22bee78STejun Heo */ 73863d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 739e22bee78STejun Heo { 74063d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 741e22bee78STejun Heo } 742e22bee78STejun Heo 743e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 74463d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 745e22bee78STejun Heo { 74663d95a91STejun Heo return pool->nr_idle; 747e22bee78STejun Heo } 748e22bee78STejun Heo 749e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 75063d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 751e22bee78STejun Heo { 752e19e397aSTejun Heo return !list_empty(&pool->worklist) && 753e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 754e22bee78STejun Heo } 755e22bee78STejun Heo 756e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 75763d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 758e22bee78STejun Heo { 75963d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 760e22bee78STejun Heo } 761e22bee78STejun Heo 762e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 76363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 764e22bee78STejun Heo { 76534a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 76663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 76763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 768e22bee78STejun Heo 769e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 770e22bee78STejun Heo } 771e22bee78STejun Heo 772e22bee78STejun Heo /* 773e22bee78STejun Heo * Wake up functions. 774e22bee78STejun Heo */ 775e22bee78STejun Heo 7761037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 7771037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 7787e11629dSTejun Heo { 77963d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7807e11629dSTejun Heo return NULL; 7817e11629dSTejun Heo 78263d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7837e11629dSTejun Heo } 7847e11629dSTejun Heo 7857e11629dSTejun Heo /** 7867e11629dSTejun Heo * wake_up_worker - wake up an idle worker 78763d95a91STejun Heo * @pool: worker pool to wake worker from 7887e11629dSTejun Heo * 78963d95a91STejun Heo * Wake up the first idle worker of @pool. 7907e11629dSTejun Heo * 7917e11629dSTejun Heo * CONTEXT: 792d565ed63STejun Heo * spin_lock_irq(pool->lock). 7937e11629dSTejun Heo */ 79463d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7957e11629dSTejun Heo { 7961037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 7977e11629dSTejun Heo 7987e11629dSTejun Heo if (likely(worker)) 7997e11629dSTejun Heo wake_up_process(worker->task); 8007e11629dSTejun Heo } 8017e11629dSTejun Heo 8024690c4abSTejun Heo /** 803e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 804e22bee78STejun Heo * @task: task waking up 805e22bee78STejun Heo * @cpu: CPU @task is waking up to 806e22bee78STejun Heo * 807e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 808e22bee78STejun Heo * being awoken. 809e22bee78STejun Heo * 810e22bee78STejun Heo * CONTEXT: 811e22bee78STejun Heo * spin_lock_irq(rq->lock) 812e22bee78STejun Heo */ 813d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 814e22bee78STejun Heo { 815e22bee78STejun Heo struct worker *worker = kthread_data(task); 816e22bee78STejun Heo 81736576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 818ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 819e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 820e22bee78STejun Heo } 82136576000SJoonsoo Kim } 822e22bee78STejun Heo 823e22bee78STejun Heo /** 824e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 825e22bee78STejun Heo * @task: task going to sleep 826e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 827e22bee78STejun Heo * 828e22bee78STejun Heo * This function is called during schedule() when a busy worker is 829e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 830e22bee78STejun Heo * returning pointer to its task. 831e22bee78STejun Heo * 832e22bee78STejun Heo * CONTEXT: 833e22bee78STejun Heo * spin_lock_irq(rq->lock) 834e22bee78STejun Heo * 835d185af30SYacine Belkadi * Return: 836e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 837e22bee78STejun Heo */ 838d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 839e22bee78STejun Heo { 840e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 841111c225aSTejun Heo struct worker_pool *pool; 842e22bee78STejun Heo 843111c225aSTejun Heo /* 844111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 845111c225aSTejun Heo * workers, also reach here, let's not access anything before 846111c225aSTejun Heo * checking NOT_RUNNING. 847111c225aSTejun Heo */ 8482d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 849e22bee78STejun Heo return NULL; 850e22bee78STejun Heo 851111c225aSTejun Heo pool = worker->pool; 852111c225aSTejun Heo 853e22bee78STejun Heo /* this can only happen on the local cpu */ 85492b69f50SLai Jiangshan if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) 8556183c009STejun Heo return NULL; 856e22bee78STejun Heo 857e22bee78STejun Heo /* 858e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 859e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 860e22bee78STejun Heo * Please read comment there. 861e22bee78STejun Heo * 862628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 863628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 864628c78e7STejun Heo * disabled, which in turn means that none else could be 865d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 866628c78e7STejun Heo * lock is safe. 867e22bee78STejun Heo */ 868e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 869e19e397aSTejun Heo !list_empty(&pool->worklist)) 8701037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 871e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 872e22bee78STejun Heo } 873e22bee78STejun Heo 874e22bee78STejun Heo /** 875e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 876cb444766STejun Heo * @worker: self 877d302f017STejun Heo * @flags: flags to set 878d302f017STejun Heo * 879228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 880d302f017STejun Heo * 881cb444766STejun Heo * CONTEXT: 882d565ed63STejun Heo * spin_lock_irq(pool->lock) 883d302f017STejun Heo */ 884228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 885d302f017STejun Heo { 886bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 887e22bee78STejun Heo 888cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 889cb444766STejun Heo 890228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 891e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 892e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 893e19e397aSTejun Heo atomic_dec(&pool->nr_running); 894e22bee78STejun Heo } 895e22bee78STejun Heo 896d302f017STejun Heo worker->flags |= flags; 897d302f017STejun Heo } 898d302f017STejun Heo 899d302f017STejun Heo /** 900e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 901cb444766STejun Heo * @worker: self 902d302f017STejun Heo * @flags: flags to clear 903d302f017STejun Heo * 904e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 905d302f017STejun Heo * 906cb444766STejun Heo * CONTEXT: 907d565ed63STejun Heo * spin_lock_irq(pool->lock) 908d302f017STejun Heo */ 909d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 910d302f017STejun Heo { 91163d95a91STejun Heo struct worker_pool *pool = worker->pool; 912e22bee78STejun Heo unsigned int oflags = worker->flags; 913e22bee78STejun Heo 914cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 915cb444766STejun Heo 916d302f017STejun Heo worker->flags &= ~flags; 917e22bee78STejun Heo 91842c025f3STejun Heo /* 91942c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 92042c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 92142c025f3STejun Heo * of multiple flags, not a single flag. 92242c025f3STejun Heo */ 923e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 924e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 925e19e397aSTejun Heo atomic_inc(&pool->nr_running); 926d302f017STejun Heo } 927d302f017STejun Heo 928d302f017STejun Heo /** 9298cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 930c9e7cf27STejun Heo * @pool: pool of interest 9318cca0eeaSTejun Heo * @work: work to find worker for 9328cca0eeaSTejun Heo * 933c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 934c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 935a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 936a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 937a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 938a2c1c57bSTejun Heo * being executed. 939a2c1c57bSTejun Heo * 940a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 941a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 942a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 943a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 944a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 945a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 946a2c1c57bSTejun Heo * 947c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 948c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 949c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 950c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 951c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 952c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9538cca0eeaSTejun Heo * 9548cca0eeaSTejun Heo * CONTEXT: 955d565ed63STejun Heo * spin_lock_irq(pool->lock). 9568cca0eeaSTejun Heo * 957d185af30SYacine Belkadi * Return: 958d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9598cca0eeaSTejun Heo * otherwise. 9608cca0eeaSTejun Heo */ 961c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9628cca0eeaSTejun Heo struct work_struct *work) 9638cca0eeaSTejun Heo { 96442f8570fSSasha Levin struct worker *worker; 96542f8570fSSasha Levin 966b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 967a2c1c57bSTejun Heo (unsigned long)work) 968a2c1c57bSTejun Heo if (worker->current_work == work && 969a2c1c57bSTejun Heo worker->current_func == work->func) 97042f8570fSSasha Levin return worker; 97142f8570fSSasha Levin 97242f8570fSSasha Levin return NULL; 9738cca0eeaSTejun Heo } 9748cca0eeaSTejun Heo 9758cca0eeaSTejun Heo /** 976bf4ede01STejun Heo * move_linked_works - move linked works to a list 977bf4ede01STejun Heo * @work: start of series of works to be scheduled 978bf4ede01STejun Heo * @head: target list to append @work to 979bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 980bf4ede01STejun Heo * 981bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 982bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 983bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 984bf4ede01STejun Heo * 985bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 986bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 987bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 988bf4ede01STejun Heo * 989bf4ede01STejun Heo * CONTEXT: 990d565ed63STejun Heo * spin_lock_irq(pool->lock). 991bf4ede01STejun Heo */ 992bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 993bf4ede01STejun Heo struct work_struct **nextp) 994bf4ede01STejun Heo { 995bf4ede01STejun Heo struct work_struct *n; 996bf4ede01STejun Heo 997bf4ede01STejun Heo /* 998bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 999bf4ede01STejun Heo * use NULL for list head. 1000bf4ede01STejun Heo */ 1001bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1002bf4ede01STejun Heo list_move_tail(&work->entry, head); 1003bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1004bf4ede01STejun Heo break; 1005bf4ede01STejun Heo } 1006bf4ede01STejun Heo 1007bf4ede01STejun Heo /* 1008bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1009bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1010bf4ede01STejun Heo * needs to be updated. 1011bf4ede01STejun Heo */ 1012bf4ede01STejun Heo if (nextp) 1013bf4ede01STejun Heo *nextp = n; 1014bf4ede01STejun Heo } 1015bf4ede01STejun Heo 10168864b4e5STejun Heo /** 10178864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10188864b4e5STejun Heo * @pwq: pool_workqueue to get 10198864b4e5STejun Heo * 10208864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10218864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10228864b4e5STejun Heo */ 10238864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10248864b4e5STejun Heo { 10258864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10268864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10278864b4e5STejun Heo pwq->refcnt++; 10288864b4e5STejun Heo } 10298864b4e5STejun Heo 10308864b4e5STejun Heo /** 10318864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10328864b4e5STejun Heo * @pwq: pool_workqueue to put 10338864b4e5STejun Heo * 10348864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10358864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10368864b4e5STejun Heo */ 10378864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10388864b4e5STejun Heo { 10398864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10408864b4e5STejun Heo if (likely(--pwq->refcnt)) 10418864b4e5STejun Heo return; 10428864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10438864b4e5STejun Heo return; 10448864b4e5STejun Heo /* 10458864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10468864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10478864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10488864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10498864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10508864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10518864b4e5STejun Heo */ 10528864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10538864b4e5STejun Heo } 10548864b4e5STejun Heo 1055dce90d47STejun Heo /** 1056dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1057dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1058dce90d47STejun Heo * 1059dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1060dce90d47STejun Heo */ 1061dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1062dce90d47STejun Heo { 1063dce90d47STejun Heo if (pwq) { 1064dce90d47STejun Heo /* 1065dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1066dce90d47STejun Heo * following lock operations are safe. 1067dce90d47STejun Heo */ 1068dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1069dce90d47STejun Heo put_pwq(pwq); 1070dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1071dce90d47STejun Heo } 1072dce90d47STejun Heo } 1073dce90d47STejun Heo 1074112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1075bf4ede01STejun Heo { 1076112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1077bf4ede01STejun Heo 1078bf4ede01STejun Heo trace_workqueue_activate_work(work); 1079112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1080bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1081112202d9STejun Heo pwq->nr_active++; 1082bf4ede01STejun Heo } 1083bf4ede01STejun Heo 1084112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10853aa62497SLai Jiangshan { 1086112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10873aa62497SLai Jiangshan struct work_struct, entry); 10883aa62497SLai Jiangshan 1089112202d9STejun Heo pwq_activate_delayed_work(work); 10903aa62497SLai Jiangshan } 10913aa62497SLai Jiangshan 1092bf4ede01STejun Heo /** 1093112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1094112202d9STejun Heo * @pwq: pwq of interest 1095bf4ede01STejun Heo * @color: color of work which left the queue 1096bf4ede01STejun Heo * 1097bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1098112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1099bf4ede01STejun Heo * 1100bf4ede01STejun Heo * CONTEXT: 1101d565ed63STejun Heo * spin_lock_irq(pool->lock). 1102bf4ede01STejun Heo */ 1103112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1104bf4ede01STejun Heo { 11058864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1106bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11078864b4e5STejun Heo goto out_put; 1108bf4ede01STejun Heo 1109112202d9STejun Heo pwq->nr_in_flight[color]--; 1110bf4ede01STejun Heo 1111112202d9STejun Heo pwq->nr_active--; 1112112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1113bf4ede01STejun Heo /* one down, submit a delayed one */ 1114112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1115112202d9STejun Heo pwq_activate_first_delayed(pwq); 1116bf4ede01STejun Heo } 1117bf4ede01STejun Heo 1118bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1119112202d9STejun Heo if (likely(pwq->flush_color != color)) 11208864b4e5STejun Heo goto out_put; 1121bf4ede01STejun Heo 1122bf4ede01STejun Heo /* are there still in-flight works? */ 1123112202d9STejun Heo if (pwq->nr_in_flight[color]) 11248864b4e5STejun Heo goto out_put; 1125bf4ede01STejun Heo 1126112202d9STejun Heo /* this pwq is done, clear flush_color */ 1127112202d9STejun Heo pwq->flush_color = -1; 1128bf4ede01STejun Heo 1129bf4ede01STejun Heo /* 1130112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1131bf4ede01STejun Heo * will handle the rest. 1132bf4ede01STejun Heo */ 1133112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1134112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11358864b4e5STejun Heo out_put: 11368864b4e5STejun Heo put_pwq(pwq); 1137bf4ede01STejun Heo } 1138bf4ede01STejun Heo 113936e227d2STejun Heo /** 1140bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 114136e227d2STejun Heo * @work: work item to steal 114236e227d2STejun Heo * @is_dwork: @work is a delayed_work 1143bbb68dfaSTejun Heo * @flags: place to store irq state 114436e227d2STejun Heo * 114536e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1146d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 114736e227d2STejun Heo * 1148d185af30SYacine Belkadi * Return: 114936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 115036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 115136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1152bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1153bbb68dfaSTejun Heo * for arbitrarily long 115436e227d2STejun Heo * 1155d185af30SYacine Belkadi * Note: 1156bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1157e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1158e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1159e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1160bbb68dfaSTejun Heo * 1161bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1162bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1163bbb68dfaSTejun Heo * 1164e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1165bf4ede01STejun Heo */ 1166bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1167bbb68dfaSTejun Heo unsigned long *flags) 1168bf4ede01STejun Heo { 1169d565ed63STejun Heo struct worker_pool *pool; 1170112202d9STejun Heo struct pool_workqueue *pwq; 1171bf4ede01STejun Heo 1172bbb68dfaSTejun Heo local_irq_save(*flags); 1173bbb68dfaSTejun Heo 117436e227d2STejun Heo /* try to steal the timer if it exists */ 117536e227d2STejun Heo if (is_dwork) { 117636e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 117736e227d2STejun Heo 1178e0aecdd8STejun Heo /* 1179e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1180e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1181e0aecdd8STejun Heo * running on the local CPU. 1182e0aecdd8STejun Heo */ 118336e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 118436e227d2STejun Heo return 1; 118536e227d2STejun Heo } 118636e227d2STejun Heo 118736e227d2STejun Heo /* try to claim PENDING the normal way */ 1188bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1189bf4ede01STejun Heo return 0; 1190bf4ede01STejun Heo 1191bf4ede01STejun Heo /* 1192bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1193bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1194bf4ede01STejun Heo */ 1195d565ed63STejun Heo pool = get_work_pool(work); 1196d565ed63STejun Heo if (!pool) 1197bbb68dfaSTejun Heo goto fail; 1198bf4ede01STejun Heo 1199d565ed63STejun Heo spin_lock(&pool->lock); 1200bf4ede01STejun Heo /* 1201112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1202112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1203112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1204112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1205112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12060b3dae68SLai Jiangshan * item is currently queued on that pool. 1207bf4ede01STejun Heo */ 1208112202d9STejun Heo pwq = get_work_pwq(work); 1209112202d9STejun Heo if (pwq && pwq->pool == pool) { 1210bf4ede01STejun Heo debug_work_deactivate(work); 12113aa62497SLai Jiangshan 12123aa62497SLai Jiangshan /* 121316062836STejun Heo * A delayed work item cannot be grabbed directly because 121416062836STejun Heo * it might have linked NO_COLOR work items which, if left 1215112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 121616062836STejun Heo * management later on and cause stall. Make sure the work 121716062836STejun Heo * item is activated before grabbing. 12183aa62497SLai Jiangshan */ 12193aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1220112202d9STejun Heo pwq_activate_delayed_work(work); 12213aa62497SLai Jiangshan 1222bf4ede01STejun Heo list_del_init(&work->entry); 12239c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 122436e227d2STejun Heo 1225112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12264468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12274468a00fSLai Jiangshan 1228d565ed63STejun Heo spin_unlock(&pool->lock); 122936e227d2STejun Heo return 1; 1230bf4ede01STejun Heo } 1231d565ed63STejun Heo spin_unlock(&pool->lock); 1232bbb68dfaSTejun Heo fail: 1233bbb68dfaSTejun Heo local_irq_restore(*flags); 1234bbb68dfaSTejun Heo if (work_is_canceling(work)) 1235bbb68dfaSTejun Heo return -ENOENT; 1236bbb68dfaSTejun Heo cpu_relax(); 123736e227d2STejun Heo return -EAGAIN; 1238bf4ede01STejun Heo } 1239bf4ede01STejun Heo 1240bf4ede01STejun Heo /** 1241706026c2STejun Heo * insert_work - insert a work into a pool 1242112202d9STejun Heo * @pwq: pwq @work belongs to 12434690c4abSTejun Heo * @work: work to insert 12444690c4abSTejun Heo * @head: insertion point 12454690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12464690c4abSTejun Heo * 1247112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1248706026c2STejun Heo * work_struct flags. 12494690c4abSTejun Heo * 12504690c4abSTejun Heo * CONTEXT: 1251d565ed63STejun Heo * spin_lock_irq(pool->lock). 1252365970a1SDavid Howells */ 1253112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1254112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1255b89deed3SOleg Nesterov { 1256112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1257e1d8aa9fSFrederic Weisbecker 12584690c4abSTejun Heo /* we own @work, set data and link */ 1259112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12601a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12618864b4e5STejun Heo get_pwq(pwq); 1262e22bee78STejun Heo 1263e22bee78STejun Heo /* 1264c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1265c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1266c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1267e22bee78STejun Heo */ 1268e22bee78STejun Heo smp_mb(); 1269e22bee78STejun Heo 127063d95a91STejun Heo if (__need_more_worker(pool)) 127163d95a91STejun Heo wake_up_worker(pool); 1272b89deed3SOleg Nesterov } 1273b89deed3SOleg Nesterov 1274c8efcc25STejun Heo /* 1275c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12768d03ecfeSTejun Heo * same workqueue. 1277c8efcc25STejun Heo */ 1278c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1279c8efcc25STejun Heo { 1280c8efcc25STejun Heo struct worker *worker; 1281c8efcc25STejun Heo 12828d03ecfeSTejun Heo worker = current_wq_worker(); 1283c8efcc25STejun Heo /* 12848d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12858d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1286c8efcc25STejun Heo */ 1287112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1288c8efcc25STejun Heo } 1289c8efcc25STejun Heo 1290d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12911da177e4SLinus Torvalds struct work_struct *work) 12921da177e4SLinus Torvalds { 1293112202d9STejun Heo struct pool_workqueue *pwq; 1294c9178087STejun Heo struct worker_pool *last_pool; 12951e19ffc6STejun Heo struct list_head *worklist; 12968a2e8e5dSTejun Heo unsigned int work_flags; 1297b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12988930cabaSTejun Heo 12998930cabaSTejun Heo /* 13008930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13018930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13028930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13038930cabaSTejun Heo * happen with IRQ disabled. 13048930cabaSTejun Heo */ 13058930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13061da177e4SLinus Torvalds 1307dc186ad7SThomas Gleixner debug_work_activate(work); 13081e19ffc6STejun Heo 13099ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1310618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1311c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1312e41e704bSTejun Heo return; 13139e8cd2f5STejun Heo retry: 1314df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1315f3421797STejun Heo cpu = raw_smp_processor_id(); 1316df2d5ae4STejun Heo 1317df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1318df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1319c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1320df2d5ae4STejun Heo else 1321df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1322f3421797STejun Heo 132318aa9effSTejun Heo /* 1324c9178087STejun Heo * If @work was previously on a different pool, it might still be 1325c9178087STejun Heo * running there, in which case the work needs to be queued on that 1326c9178087STejun Heo * pool to guarantee non-reentrancy. 132718aa9effSTejun Heo */ 1328c9e7cf27STejun Heo last_pool = get_work_pool(work); 1329112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 133018aa9effSTejun Heo struct worker *worker; 133118aa9effSTejun Heo 1332d565ed63STejun Heo spin_lock(&last_pool->lock); 133318aa9effSTejun Heo 1334c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 133518aa9effSTejun Heo 1336112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1337c9178087STejun Heo pwq = worker->current_pwq; 13388594fadeSLai Jiangshan } else { 133918aa9effSTejun Heo /* meh... not running there, queue here */ 1340d565ed63STejun Heo spin_unlock(&last_pool->lock); 1341112202d9STejun Heo spin_lock(&pwq->pool->lock); 134218aa9effSTejun Heo } 13438930cabaSTejun Heo } else { 1344112202d9STejun Heo spin_lock(&pwq->pool->lock); 13458930cabaSTejun Heo } 1346502ca9d8STejun Heo 13479e8cd2f5STejun Heo /* 13489e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13499e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13509e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1351df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1352df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13539e8cd2f5STejun Heo * make forward-progress. 13549e8cd2f5STejun Heo */ 13559e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13569e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13579e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13589e8cd2f5STejun Heo cpu_relax(); 13599e8cd2f5STejun Heo goto retry; 13609e8cd2f5STejun Heo } 13619e8cd2f5STejun Heo /* oops */ 13629e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13639e8cd2f5STejun Heo wq->name, cpu); 13649e8cd2f5STejun Heo } 13659e8cd2f5STejun Heo 1366112202d9STejun Heo /* pwq determined, queue */ 1367112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1368502ca9d8STejun Heo 1369f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1370112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1371f5b2552bSDan Carpenter return; 1372f5b2552bSDan Carpenter } 13731e19ffc6STejun Heo 1374112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1375112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13761e19ffc6STejun Heo 1377112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1378cdadf009STejun Heo trace_workqueue_activate_work(work); 1379112202d9STejun Heo pwq->nr_active++; 1380112202d9STejun Heo worklist = &pwq->pool->worklist; 13818a2e8e5dSTejun Heo } else { 13828a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1383112202d9STejun Heo worklist = &pwq->delayed_works; 13848a2e8e5dSTejun Heo } 13851e19ffc6STejun Heo 1386112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13871e19ffc6STejun Heo 1388112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13891da177e4SLinus Torvalds } 13901da177e4SLinus Torvalds 13910fcb78c2SRolf Eike Beer /** 1392c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1393c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1394c1a220e7SZhang Rui * @wq: workqueue to use 1395c1a220e7SZhang Rui * @work: work to queue 1396c1a220e7SZhang Rui * 1397c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1398c1a220e7SZhang Rui * can't go away. 1399d185af30SYacine Belkadi * 1400d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1401c1a220e7SZhang Rui */ 1402d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1403d4283e93STejun Heo struct work_struct *work) 1404c1a220e7SZhang Rui { 1405d4283e93STejun Heo bool ret = false; 14068930cabaSTejun Heo unsigned long flags; 14078930cabaSTejun Heo 14088930cabaSTejun Heo local_irq_save(flags); 1409c1a220e7SZhang Rui 141022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14114690c4abSTejun Heo __queue_work(cpu, wq, work); 1412d4283e93STejun Heo ret = true; 1413c1a220e7SZhang Rui } 14148930cabaSTejun Heo 14158930cabaSTejun Heo local_irq_restore(flags); 1416c1a220e7SZhang Rui return ret; 1417c1a220e7SZhang Rui } 1418ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1419c1a220e7SZhang Rui 1420d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14211da177e4SLinus Torvalds { 142252bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14231da177e4SLinus Torvalds 1424e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 142560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14261da177e4SLinus Torvalds } 14271438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14281da177e4SLinus Torvalds 14297beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 143052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14311da177e4SLinus Torvalds { 14327beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14337beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14341da177e4SLinus Torvalds 14357beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14367beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1437fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1438fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14397beb2edfSTejun Heo 14408852aac2STejun Heo /* 14418852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14428852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14438852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14448852aac2STejun Heo * on that there's no such delay when @delay is 0. 14458852aac2STejun Heo */ 14468852aac2STejun Heo if (!delay) { 14478852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14488852aac2STejun Heo return; 14498852aac2STejun Heo } 14508852aac2STejun Heo 14517beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14527beb2edfSTejun Heo 145360c057bcSLai Jiangshan dwork->wq = wq; 14541265057fSTejun Heo dwork->cpu = cpu; 14557beb2edfSTejun Heo timer->expires = jiffies + delay; 14567beb2edfSTejun Heo 14577beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14587beb2edfSTejun Heo add_timer_on(timer, cpu); 14597beb2edfSTejun Heo else 14607beb2edfSTejun Heo add_timer(timer); 14617beb2edfSTejun Heo } 14621da177e4SLinus Torvalds 14630fcb78c2SRolf Eike Beer /** 14640fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14650fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14660fcb78c2SRolf Eike Beer * @wq: workqueue to use 1467af9997e4SRandy Dunlap * @dwork: work to queue 14680fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14690fcb78c2SRolf Eike Beer * 1470d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1471715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1472715f1300STejun Heo * execution. 14730fcb78c2SRolf Eike Beer */ 1474d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 147552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14767a6bc1cdSVenkatesh Pallipadi { 147752bad64dSDavid Howells struct work_struct *work = &dwork->work; 1478d4283e93STejun Heo bool ret = false; 14798930cabaSTejun Heo unsigned long flags; 14808930cabaSTejun Heo 14818930cabaSTejun Heo /* read the comment in __queue_work() */ 14828930cabaSTejun Heo local_irq_save(flags); 14837a6bc1cdSVenkatesh Pallipadi 148422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14857beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1486d4283e93STejun Heo ret = true; 14877a6bc1cdSVenkatesh Pallipadi } 14888930cabaSTejun Heo 14898930cabaSTejun Heo local_irq_restore(flags); 14907a6bc1cdSVenkatesh Pallipadi return ret; 14917a6bc1cdSVenkatesh Pallipadi } 1492ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 14931da177e4SLinus Torvalds 1494c8e55f36STejun Heo /** 14958376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14968376fe22STejun Heo * @cpu: CPU number to execute work on 14978376fe22STejun Heo * @wq: workqueue to use 14988376fe22STejun Heo * @dwork: work to queue 14998376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15008376fe22STejun Heo * 15018376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15028376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15038376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15048376fe22STejun Heo * current state. 15058376fe22STejun Heo * 1506d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15078376fe22STejun Heo * pending and its timer was modified. 15088376fe22STejun Heo * 1509e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15108376fe22STejun Heo * See try_to_grab_pending() for details. 15118376fe22STejun Heo */ 15128376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15138376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15148376fe22STejun Heo { 15158376fe22STejun Heo unsigned long flags; 15168376fe22STejun Heo int ret; 15178376fe22STejun Heo 15188376fe22STejun Heo do { 15198376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15208376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15218376fe22STejun Heo 15228376fe22STejun Heo if (likely(ret >= 0)) { 15238376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15248376fe22STejun Heo local_irq_restore(flags); 15258376fe22STejun Heo } 15268376fe22STejun Heo 15278376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15288376fe22STejun Heo return ret; 15298376fe22STejun Heo } 15308376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15318376fe22STejun Heo 15328376fe22STejun Heo /** 1533c8e55f36STejun Heo * worker_enter_idle - enter idle state 1534c8e55f36STejun Heo * @worker: worker which is entering idle state 1535c8e55f36STejun Heo * 1536c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1537c8e55f36STejun Heo * necessary. 1538c8e55f36STejun Heo * 1539c8e55f36STejun Heo * LOCKING: 1540d565ed63STejun Heo * spin_lock_irq(pool->lock). 1541c8e55f36STejun Heo */ 1542c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15431da177e4SLinus Torvalds { 1544bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1545c8e55f36STejun Heo 15466183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15476183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15486183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15496183c009STejun Heo return; 1550c8e55f36STejun Heo 1551051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1552cb444766STejun Heo worker->flags |= WORKER_IDLE; 1553bd7bdd43STejun Heo pool->nr_idle++; 1554e22bee78STejun Heo worker->last_active = jiffies; 1555c8e55f36STejun Heo 1556c8e55f36STejun Heo /* idle_list is LIFO */ 1557bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1558db7bccf4STejun Heo 155963d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1560628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1561cb444766STejun Heo 1562544ecf31STejun Heo /* 1563706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1564d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1565628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1566628c78e7STejun Heo * unbind is not in progress. 1567544ecf31STejun Heo */ 156824647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1569bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1570e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1571c8e55f36STejun Heo } 1572c8e55f36STejun Heo 1573c8e55f36STejun Heo /** 1574c8e55f36STejun Heo * worker_leave_idle - leave idle state 1575c8e55f36STejun Heo * @worker: worker which is leaving idle state 1576c8e55f36STejun Heo * 1577c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1578c8e55f36STejun Heo * 1579c8e55f36STejun Heo * LOCKING: 1580d565ed63STejun Heo * spin_lock_irq(pool->lock). 1581c8e55f36STejun Heo */ 1582c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1583c8e55f36STejun Heo { 1584bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1585c8e55f36STejun Heo 15866183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15876183c009STejun Heo return; 1588d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1589bd7bdd43STejun Heo pool->nr_idle--; 1590c8e55f36STejun Heo list_del_init(&worker->entry); 1591c8e55f36STejun Heo } 1592c8e55f36STejun Heo 1593f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1594c34056a3STejun Heo { 1595c34056a3STejun Heo struct worker *worker; 1596c34056a3STejun Heo 1597f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1598c8e55f36STejun Heo if (worker) { 1599c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1600affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1601da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1602e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1603e22bee78STejun Heo worker->flags = WORKER_PREP; 1604c8e55f36STejun Heo } 1605c34056a3STejun Heo return worker; 1606c34056a3STejun Heo } 1607c34056a3STejun Heo 1608c34056a3STejun Heo /** 16094736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16104736cbf7SLai Jiangshan * @worker: worker to be attached 16114736cbf7SLai Jiangshan * @pool: the target pool 16124736cbf7SLai Jiangshan * 16134736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16144736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16154736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16164736cbf7SLai Jiangshan */ 16174736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16184736cbf7SLai Jiangshan struct worker_pool *pool) 16194736cbf7SLai Jiangshan { 16204736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16214736cbf7SLai Jiangshan 16224736cbf7SLai Jiangshan /* 16234736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16244736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16254736cbf7SLai Jiangshan */ 16264736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16274736cbf7SLai Jiangshan 16284736cbf7SLai Jiangshan /* 16294736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 16304736cbf7SLai Jiangshan * stable across this function. See the comments above the 16314736cbf7SLai Jiangshan * flag definition for details. 16324736cbf7SLai Jiangshan */ 16334736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 16344736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 16354736cbf7SLai Jiangshan 16364736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 16374736cbf7SLai Jiangshan 16384736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 16394736cbf7SLai Jiangshan } 16404736cbf7SLai Jiangshan 16414736cbf7SLai Jiangshan /** 164260f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 164360f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 164460f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 164560f5a4bcSLai Jiangshan * 16464736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 16474736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 16484736cbf7SLai Jiangshan * other reference to the pool. 164960f5a4bcSLai Jiangshan */ 165060f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 165160f5a4bcSLai Jiangshan struct worker_pool *pool) 165260f5a4bcSLai Jiangshan { 165360f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 165460f5a4bcSLai Jiangshan 165592f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1656da028469SLai Jiangshan list_del(&worker->node); 1657da028469SLai Jiangshan if (list_empty(&pool->workers)) 165860f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 165992f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 166060f5a4bcSLai Jiangshan 1661b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1662b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1663b62c0751SLai Jiangshan 166460f5a4bcSLai Jiangshan if (detach_completion) 166560f5a4bcSLai Jiangshan complete(detach_completion); 166660f5a4bcSLai Jiangshan } 166760f5a4bcSLai Jiangshan 166860f5a4bcSLai Jiangshan /** 1669c34056a3STejun Heo * create_worker - create a new workqueue worker 167063d95a91STejun Heo * @pool: pool the new worker will belong to 1671c34056a3STejun Heo * 1672051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1673c34056a3STejun Heo * 1674c34056a3STejun Heo * CONTEXT: 1675c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1676c34056a3STejun Heo * 1677d185af30SYacine Belkadi * Return: 1678c34056a3STejun Heo * Pointer to the newly created worker. 1679c34056a3STejun Heo */ 1680bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1681c34056a3STejun Heo { 1682c34056a3STejun Heo struct worker *worker = NULL; 1683f3421797STejun Heo int id = -1; 1684e3c916a4STejun Heo char id_buf[16]; 1685c34056a3STejun Heo 16867cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 16877cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1688822d8405STejun Heo if (id < 0) 1689c34056a3STejun Heo goto fail; 1690c34056a3STejun Heo 1691f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1692c34056a3STejun Heo if (!worker) 1693c34056a3STejun Heo goto fail; 1694c34056a3STejun Heo 1695bd7bdd43STejun Heo worker->pool = pool; 1696c34056a3STejun Heo worker->id = id; 1697c34056a3STejun Heo 169829c91e99STejun Heo if (pool->cpu >= 0) 1699e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1700e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1701f3421797STejun Heo else 1702e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1703e3c916a4STejun Heo 1704f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1705e3c916a4STejun Heo "kworker/%s", id_buf); 1706c34056a3STejun Heo if (IS_ERR(worker->task)) 1707c34056a3STejun Heo goto fail; 1708c34056a3STejun Heo 170991151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 171091151228SOleg Nesterov 171191151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 171291151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 171391151228SOleg Nesterov 1714da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17154736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1716822d8405STejun Heo 1717051e1850SLai Jiangshan /* start the newly created worker */ 1718051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1719051e1850SLai Jiangshan worker->pool->nr_workers++; 1720051e1850SLai Jiangshan worker_enter_idle(worker); 1721051e1850SLai Jiangshan wake_up_process(worker->task); 1722051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1723051e1850SLai Jiangshan 1724c34056a3STejun Heo return worker; 1725822d8405STejun Heo 1726c34056a3STejun Heo fail: 17279625ab17SLai Jiangshan if (id >= 0) 17287cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1729c34056a3STejun Heo kfree(worker); 1730c34056a3STejun Heo return NULL; 1731c34056a3STejun Heo } 1732c34056a3STejun Heo 1733c34056a3STejun Heo /** 1734c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1735c34056a3STejun Heo * @worker: worker to be destroyed 1736c34056a3STejun Heo * 173773eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 173873eb7fe7SLai Jiangshan * be idle. 1739c8e55f36STejun Heo * 1740c8e55f36STejun Heo * CONTEXT: 174160f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1742c34056a3STejun Heo */ 1743c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1744c34056a3STejun Heo { 1745bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1746c34056a3STejun Heo 1747cd549687STejun Heo lockdep_assert_held(&pool->lock); 1748cd549687STejun Heo 1749c34056a3STejun Heo /* sanity check frenzy */ 17506183c009STejun Heo if (WARN_ON(worker->current_work) || 175173eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 175273eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 17536183c009STejun Heo return; 1754c34056a3STejun Heo 1755bd7bdd43STejun Heo pool->nr_workers--; 1756bd7bdd43STejun Heo pool->nr_idle--; 1757c8e55f36STejun Heo 1758c8e55f36STejun Heo list_del_init(&worker->entry); 1759cb444766STejun Heo worker->flags |= WORKER_DIE; 176060f5a4bcSLai Jiangshan wake_up_process(worker->task); 1761c34056a3STejun Heo } 1762c34056a3STejun Heo 176363d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1764e22bee78STejun Heo { 176563d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1766e22bee78STejun Heo 1767d565ed63STejun Heo spin_lock_irq(&pool->lock); 1768e22bee78STejun Heo 17693347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1770e22bee78STejun Heo struct worker *worker; 1771e22bee78STejun Heo unsigned long expires; 1772e22bee78STejun Heo 1773e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 177463d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1775e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1776e22bee78STejun Heo 17773347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 177863d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 17793347fc9fSLai Jiangshan break; 1780e22bee78STejun Heo } 17813347fc9fSLai Jiangshan 17823347fc9fSLai Jiangshan destroy_worker(worker); 1783e22bee78STejun Heo } 1784e22bee78STejun Heo 1785d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1786e22bee78STejun Heo } 1787e22bee78STejun Heo 1788493a1724STejun Heo static void send_mayday(struct work_struct *work) 1789e22bee78STejun Heo { 1790112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1791112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1792493a1724STejun Heo 17932e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1794e22bee78STejun Heo 1795493008a8STejun Heo if (!wq->rescuer) 1796493a1724STejun Heo return; 1797e22bee78STejun Heo 1798e22bee78STejun Heo /* mayday mayday mayday */ 1799493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 180077668c8bSLai Jiangshan /* 180177668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 180277668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 180377668c8bSLai Jiangshan * rescuer is done with it. 180477668c8bSLai Jiangshan */ 180577668c8bSLai Jiangshan get_pwq(pwq); 1806493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1807e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1808493a1724STejun Heo } 1809e22bee78STejun Heo } 1810e22bee78STejun Heo 1811706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1812e22bee78STejun Heo { 181363d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1814e22bee78STejun Heo struct work_struct *work; 1815e22bee78STejun Heo 1816b2d82909STejun Heo spin_lock_irq(&pool->lock); 1817b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1818e22bee78STejun Heo 181963d95a91STejun Heo if (need_to_create_worker(pool)) { 1820e22bee78STejun Heo /* 1821e22bee78STejun Heo * We've been trying to create a new worker but 1822e22bee78STejun Heo * haven't been successful. We might be hitting an 1823e22bee78STejun Heo * allocation deadlock. Send distress signals to 1824e22bee78STejun Heo * rescuers. 1825e22bee78STejun Heo */ 182663d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1827e22bee78STejun Heo send_mayday(work); 1828e22bee78STejun Heo } 1829e22bee78STejun Heo 1830b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1831b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1832e22bee78STejun Heo 183363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1834e22bee78STejun Heo } 1835e22bee78STejun Heo 1836e22bee78STejun Heo /** 1837e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 183863d95a91STejun Heo * @pool: pool to create a new worker for 1839e22bee78STejun Heo * 184063d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1841e22bee78STejun Heo * have at least one idle worker on return from this function. If 1842e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 184363d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1844e22bee78STejun Heo * possible allocation deadlock. 1845e22bee78STejun Heo * 1846c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1847c5aa87bbSTejun Heo * may_start_working() %true. 1848e22bee78STejun Heo * 1849e22bee78STejun Heo * LOCKING: 1850d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1851e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1852e22bee78STejun Heo * manager. 1853e22bee78STejun Heo */ 185429187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1855d565ed63STejun Heo __releases(&pool->lock) 1856d565ed63STejun Heo __acquires(&pool->lock) 1857e22bee78STejun Heo { 1858e22bee78STejun Heo restart: 1859d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18609f9c2364STejun Heo 1861e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 186263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1863e22bee78STejun Heo 1864e22bee78STejun Heo while (true) { 1865051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1866e22bee78STejun Heo break; 1867e22bee78STejun Heo 1868e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 18699f9c2364STejun Heo 187063d95a91STejun Heo if (!need_to_create_worker(pool)) 1871e22bee78STejun Heo break; 1872e22bee78STejun Heo } 1873e22bee78STejun Heo 187463d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1875d565ed63STejun Heo spin_lock_irq(&pool->lock); 1876051e1850SLai Jiangshan /* 1877051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1878051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1879051e1850SLai Jiangshan * already become busy. 1880051e1850SLai Jiangshan */ 188163d95a91STejun Heo if (need_to_create_worker(pool)) 1882e22bee78STejun Heo goto restart; 1883e22bee78STejun Heo } 1884e22bee78STejun Heo 1885e22bee78STejun Heo /** 1886e22bee78STejun Heo * manage_workers - manage worker pool 1887e22bee78STejun Heo * @worker: self 1888e22bee78STejun Heo * 1889706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1890e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1891706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1892e22bee78STejun Heo * 1893e22bee78STejun Heo * The caller can safely start processing works on false return. On 1894e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1895e22bee78STejun Heo * and may_start_working() is true. 1896e22bee78STejun Heo * 1897e22bee78STejun Heo * CONTEXT: 1898d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1899e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1900e22bee78STejun Heo * 1901d185af30SYacine Belkadi * Return: 190229187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 190329187a9eSTejun Heo * start processing works, %true if management function was performed and 190429187a9eSTejun Heo * the conditions that the caller verified before calling the function may 190529187a9eSTejun Heo * no longer be true. 1906e22bee78STejun Heo */ 1907e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1908e22bee78STejun Heo { 190963d95a91STejun Heo struct worker_pool *pool = worker->pool; 1910e22bee78STejun Heo 1911bc3a1afcSTejun Heo /* 1912bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1913bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1914bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1915bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1916bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1917bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1918bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1919bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1920bc3a1afcSTejun Heo */ 192134a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 192229187a9eSTejun Heo return false; 19232607d7a6STejun Heo pool->manager = worker; 1924e22bee78STejun Heo 192529187a9eSTejun Heo maybe_create_worker(pool); 1926e22bee78STejun Heo 19272607d7a6STejun Heo pool->manager = NULL; 192834a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 192929187a9eSTejun Heo return true; 1930e22bee78STejun Heo } 1931e22bee78STejun Heo 1932a62428c0STejun Heo /** 1933a62428c0STejun Heo * process_one_work - process single work 1934c34056a3STejun Heo * @worker: self 1935a62428c0STejun Heo * @work: work to process 1936a62428c0STejun Heo * 1937a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1938a62428c0STejun Heo * process a single work including synchronization against and 1939a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1940a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1941a62428c0STejun Heo * call this function to process a work. 1942a62428c0STejun Heo * 1943a62428c0STejun Heo * CONTEXT: 1944d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1945a62428c0STejun Heo */ 1946c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 1947d565ed63STejun Heo __releases(&pool->lock) 1948d565ed63STejun Heo __acquires(&pool->lock) 19491da177e4SLinus Torvalds { 1950112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1951bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1952112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 195373f53c4aSTejun Heo int work_color; 19547e11629dSTejun Heo struct worker *collision; 19554e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 19564e6045f1SJohannes Berg /* 1957a62428c0STejun Heo * It is permissible to free the struct work_struct from 1958a62428c0STejun Heo * inside the function that is called from it, this we need to 1959a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1960a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1961a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 19624e6045f1SJohannes Berg */ 19634d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 19644d82a1deSPeter Zijlstra 19654d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 19664e6045f1SJohannes Berg #endif 1967807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 196885327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1969ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 197025511a47STejun Heo 19717e11629dSTejun Heo /* 19727e11629dSTejun Heo * A single work shouldn't be executed concurrently by 19737e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 19747e11629dSTejun Heo * already processing the work. If so, defer the work to the 19757e11629dSTejun Heo * currently executing one. 19767e11629dSTejun Heo */ 1977c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 19787e11629dSTejun Heo if (unlikely(collision)) { 19797e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 19807e11629dSTejun Heo return; 19817e11629dSTejun Heo } 19821da177e4SLinus Torvalds 19838930cabaSTejun Heo /* claim and dequeue */ 19841da177e4SLinus Torvalds debug_work_deactivate(work); 1985c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 1986c34056a3STejun Heo worker->current_work = work; 1987a2c1c57bSTejun Heo worker->current_func = work->func; 1988112202d9STejun Heo worker->current_pwq = pwq; 198973f53c4aSTejun Heo work_color = get_work_color(work); 19907a22ad75STejun Heo 1991a62428c0STejun Heo list_del_init(&work->entry); 1992a62428c0STejun Heo 1993649027d7STejun Heo /* 1994228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 1995228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 1996228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 1997228f1d00SLai Jiangshan * execution of the pending work items. 1998fb0e7bebSTejun Heo */ 1999fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2000228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2001fb0e7bebSTejun Heo 2002974271c4STejun Heo /* 2003a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2004a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2005a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2006a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2007228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2008974271c4STejun Heo */ 2009a489a03eSLai Jiangshan if (need_more_worker(pool)) 201063d95a91STejun Heo wake_up_worker(pool); 2011974271c4STejun Heo 20128930cabaSTejun Heo /* 20137c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2014d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 201523657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 201623657bb1STejun Heo * disabled. 20178930cabaSTejun Heo */ 20187c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20191da177e4SLinus Torvalds 2020d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2021365970a1SDavid Howells 2022112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20233295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2024e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2025a2c1c57bSTejun Heo worker->current_func(work); 2026e36c886aSArjan van de Ven /* 2027e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2028e36c886aSArjan van de Ven * point will only record its address. 2029e36c886aSArjan van de Ven */ 2030e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20313295f0efSIngo Molnar lock_map_release(&lockdep_map); 2032112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 20331da177e4SLinus Torvalds 2034d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2035044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2036044c782cSValentin Ilie " last function: %pf\n", 2037a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2038a2c1c57bSTejun Heo worker->current_func); 2039d5abe669SPeter Zijlstra debug_show_held_locks(current); 2040d5abe669SPeter Zijlstra dump_stack(); 2041d5abe669SPeter Zijlstra } 2042d5abe669SPeter Zijlstra 2043b22ce278STejun Heo /* 2044b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2045b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2046b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2047b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2048789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2049789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2050b22ce278STejun Heo */ 20513e28e377SJoe Lawrence cond_resched_rcu_qs(); 2052b22ce278STejun Heo 2053d565ed63STejun Heo spin_lock_irq(&pool->lock); 2054a62428c0STejun Heo 2055fb0e7bebSTejun Heo /* clear cpu intensive status */ 2056fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2057fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2058fb0e7bebSTejun Heo 2059a62428c0STejun Heo /* we're done with it, release */ 206042f8570fSSasha Levin hash_del(&worker->hentry); 2061c34056a3STejun Heo worker->current_work = NULL; 2062a2c1c57bSTejun Heo worker->current_func = NULL; 2063112202d9STejun Heo worker->current_pwq = NULL; 20643d1cb205STejun Heo worker->desc_valid = false; 2065112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 20661da177e4SLinus Torvalds } 20671da177e4SLinus Torvalds 2068affee4b2STejun Heo /** 2069affee4b2STejun Heo * process_scheduled_works - process scheduled works 2070affee4b2STejun Heo * @worker: self 2071affee4b2STejun Heo * 2072affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2073affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2074affee4b2STejun Heo * fetches a work from the top and executes it. 2075affee4b2STejun Heo * 2076affee4b2STejun Heo * CONTEXT: 2077d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2078affee4b2STejun Heo * multiple times. 2079affee4b2STejun Heo */ 2080affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 20811da177e4SLinus Torvalds { 2082affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2083affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2084a62428c0STejun Heo struct work_struct, entry); 2085c34056a3STejun Heo process_one_work(worker, work); 2086a62428c0STejun Heo } 20871da177e4SLinus Torvalds } 20881da177e4SLinus Torvalds 20894690c4abSTejun Heo /** 20904690c4abSTejun Heo * worker_thread - the worker thread function 2091c34056a3STejun Heo * @__worker: self 20924690c4abSTejun Heo * 2093c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2094c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2095c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2096c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2097c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2098d185af30SYacine Belkadi * 2099d185af30SYacine Belkadi * Return: 0 21004690c4abSTejun Heo */ 2101c34056a3STejun Heo static int worker_thread(void *__worker) 21021da177e4SLinus Torvalds { 2103c34056a3STejun Heo struct worker *worker = __worker; 2104bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21051da177e4SLinus Torvalds 2106e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2107e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2108c8e55f36STejun Heo woke_up: 2109d565ed63STejun Heo spin_lock_irq(&pool->lock); 2110affee4b2STejun Heo 2111a9ab775bSTejun Heo /* am I supposed to die? */ 2112a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2113d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2114a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2115e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 211660f5a4bcSLai Jiangshan 211760f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21187cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 211960f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 212060f5a4bcSLai Jiangshan kfree(worker); 2121c8e55f36STejun Heo return 0; 2122c8e55f36STejun Heo } 2123c8e55f36STejun Heo 2124c8e55f36STejun Heo worker_leave_idle(worker); 2125db7bccf4STejun Heo recheck: 2126e22bee78STejun Heo /* no more worker necessary? */ 212763d95a91STejun Heo if (!need_more_worker(pool)) 2128e22bee78STejun Heo goto sleep; 2129e22bee78STejun Heo 2130e22bee78STejun Heo /* do we need to manage? */ 213163d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2132e22bee78STejun Heo goto recheck; 2133e22bee78STejun Heo 2134c8e55f36STejun Heo /* 2135c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2136c8e55f36STejun Heo * preparing to process a work or actually processing it. 2137c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2138c8e55f36STejun Heo */ 21396183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2140c8e55f36STejun Heo 2141e22bee78STejun Heo /* 2142a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2143a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2144a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2145a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2146a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2147e22bee78STejun Heo */ 2148a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2149e22bee78STejun Heo 2150e22bee78STejun Heo do { 2151affee4b2STejun Heo struct work_struct *work = 2152bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2153affee4b2STejun Heo struct work_struct, entry); 2154affee4b2STejun Heo 2155c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2156affee4b2STejun Heo /* optimization path, not strictly necessary */ 2157affee4b2STejun Heo process_one_work(worker, work); 2158affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2159affee4b2STejun Heo process_scheduled_works(worker); 2160affee4b2STejun Heo } else { 2161c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2162affee4b2STejun Heo process_scheduled_works(worker); 2163affee4b2STejun Heo } 216463d95a91STejun Heo } while (keep_working(pool)); 2165affee4b2STejun Heo 2166228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2167d313dd85STejun Heo sleep: 2168c8e55f36STejun Heo /* 2169d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2170d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2171d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2172d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2173d565ed63STejun Heo * event. 2174c8e55f36STejun Heo */ 2175c8e55f36STejun Heo worker_enter_idle(worker); 2176c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2177d565ed63STejun Heo spin_unlock_irq(&pool->lock); 21781da177e4SLinus Torvalds schedule(); 2179c8e55f36STejun Heo goto woke_up; 21801da177e4SLinus Torvalds } 21811da177e4SLinus Torvalds 2182e22bee78STejun Heo /** 2183e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2184111c225aSTejun Heo * @__rescuer: self 2185e22bee78STejun Heo * 2186e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2187493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2188e22bee78STejun Heo * 2189706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2190e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2191e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2192e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2193e22bee78STejun Heo * the problem rescuer solves. 2194e22bee78STejun Heo * 2195706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2196706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2197e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2198e22bee78STejun Heo * 2199e22bee78STejun Heo * This should happen rarely. 2200d185af30SYacine Belkadi * 2201d185af30SYacine Belkadi * Return: 0 2202e22bee78STejun Heo */ 2203111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2204e22bee78STejun Heo { 2205111c225aSTejun Heo struct worker *rescuer = __rescuer; 2206111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2207e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22084d595b86SLai Jiangshan bool should_stop; 2209e22bee78STejun Heo 2210e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2211111c225aSTejun Heo 2212111c225aSTejun Heo /* 2213111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2214111c225aSTejun Heo * doesn't participate in concurrency management. 2215111c225aSTejun Heo */ 2216111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2217e22bee78STejun Heo repeat: 2218e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22191da177e4SLinus Torvalds 22204d595b86SLai Jiangshan /* 22214d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22224d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22234d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22244d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22254d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22264d595b86SLai Jiangshan * list is always empty on exit. 22274d595b86SLai Jiangshan */ 22284d595b86SLai Jiangshan should_stop = kthread_should_stop(); 22291da177e4SLinus Torvalds 2230493a1724STejun Heo /* see whether any pwq is asking for help */ 22312e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2232493a1724STejun Heo 2233493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2234493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2235493a1724STejun Heo struct pool_workqueue, mayday_node); 2236112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2237e22bee78STejun Heo struct work_struct *work, *n; 2238e22bee78STejun Heo 2239e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2240493a1724STejun Heo list_del_init(&pwq->mayday_node); 2241493a1724STejun Heo 22422e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2243e22bee78STejun Heo 224451697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 224551697d39SLai Jiangshan 224651697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2247b3104104SLai Jiangshan rescuer->pool = pool; 2248e22bee78STejun Heo 2249e22bee78STejun Heo /* 2250e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2251e22bee78STejun Heo * process'em. 2252e22bee78STejun Heo */ 22530479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 2254bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2255112202d9STejun Heo if (get_work_pwq(work) == pwq) 2256e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2257e22bee78STejun Heo 2258008847f6SNeilBrown if (!list_empty(scheduled)) { 2259e22bee78STejun Heo process_scheduled_works(rescuer); 22607576958aSTejun Heo 22617576958aSTejun Heo /* 2262008847f6SNeilBrown * The above execution of rescued work items could 2263008847f6SNeilBrown * have created more to rescue through 2264008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2265008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2266008847f6SNeilBrown * that such back-to-back work items, which may be 2267008847f6SNeilBrown * being used to relieve memory pressure, don't 2268008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2269008847f6SNeilBrown */ 2270008847f6SNeilBrown if (need_to_create_worker(pool)) { 2271008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2272008847f6SNeilBrown get_pwq(pwq); 2273008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2274008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2275008847f6SNeilBrown } 2276008847f6SNeilBrown } 2277008847f6SNeilBrown 2278008847f6SNeilBrown /* 227977668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 228013b1d625SLai Jiangshan * go away while we're still attached to it. 228177668c8bSLai Jiangshan */ 228277668c8bSLai Jiangshan put_pwq(pwq); 228377668c8bSLai Jiangshan 228477668c8bSLai Jiangshan /* 2285d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 22867576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 22877576958aSTejun Heo * and stalling the execution. 22887576958aSTejun Heo */ 2289d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 229063d95a91STejun Heo wake_up_worker(pool); 22917576958aSTejun Heo 2292b3104104SLai Jiangshan rescuer->pool = NULL; 229313b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 229413b1d625SLai Jiangshan 229513b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 229613b1d625SLai Jiangshan 229713b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 22981da177e4SLinus Torvalds } 22991da177e4SLinus Torvalds 23002e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2301493a1724STejun Heo 23024d595b86SLai Jiangshan if (should_stop) { 23034d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23044d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23054d595b86SLai Jiangshan return 0; 23064d595b86SLai Jiangshan } 23074d595b86SLai Jiangshan 2308111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2309111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2310e22bee78STejun Heo schedule(); 2311e22bee78STejun Heo goto repeat; 23121da177e4SLinus Torvalds } 23131da177e4SLinus Torvalds 2314fc2e4d70SOleg Nesterov struct wq_barrier { 2315fc2e4d70SOleg Nesterov struct work_struct work; 2316fc2e4d70SOleg Nesterov struct completion done; 23172607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2318fc2e4d70SOleg Nesterov }; 2319fc2e4d70SOleg Nesterov 2320fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2321fc2e4d70SOleg Nesterov { 2322fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2323fc2e4d70SOleg Nesterov complete(&barr->done); 2324fc2e4d70SOleg Nesterov } 2325fc2e4d70SOleg Nesterov 23264690c4abSTejun Heo /** 23274690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2328112202d9STejun Heo * @pwq: pwq to insert barrier into 23294690c4abSTejun Heo * @barr: wq_barrier to insert 2330affee4b2STejun Heo * @target: target work to attach @barr to 2331affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23324690c4abSTejun Heo * 2333affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2334affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2335affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2336affee4b2STejun Heo * cpu. 2337affee4b2STejun Heo * 2338affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2339affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2340affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2341affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2342affee4b2STejun Heo * after a work with LINKED flag set. 2343affee4b2STejun Heo * 2344affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2345112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23464690c4abSTejun Heo * 23474690c4abSTejun Heo * CONTEXT: 2348d565ed63STejun Heo * spin_lock_irq(pool->lock). 23494690c4abSTejun Heo */ 2350112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2351affee4b2STejun Heo struct wq_barrier *barr, 2352affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2353fc2e4d70SOleg Nesterov { 2354affee4b2STejun Heo struct list_head *head; 2355affee4b2STejun Heo unsigned int linked = 0; 2356affee4b2STejun Heo 2357dc186ad7SThomas Gleixner /* 2358d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2359dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2360dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2361dc186ad7SThomas Gleixner * might deadlock. 2362dc186ad7SThomas Gleixner */ 2363ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 236422df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2365fc2e4d70SOleg Nesterov init_completion(&barr->done); 23662607d7a6STejun Heo barr->task = current; 236783c22520SOleg Nesterov 2368affee4b2STejun Heo /* 2369affee4b2STejun Heo * If @target is currently being executed, schedule the 2370affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2371affee4b2STejun Heo */ 2372affee4b2STejun Heo if (worker) 2373affee4b2STejun Heo head = worker->scheduled.next; 2374affee4b2STejun Heo else { 2375affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2376affee4b2STejun Heo 2377affee4b2STejun Heo head = target->entry.next; 2378affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2379affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2380affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2381affee4b2STejun Heo } 2382affee4b2STejun Heo 2383dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2384112202d9STejun Heo insert_work(pwq, &barr->work, head, 2385affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2386fc2e4d70SOleg Nesterov } 2387fc2e4d70SOleg Nesterov 238873f53c4aSTejun Heo /** 2389112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 239073f53c4aSTejun Heo * @wq: workqueue being flushed 239173f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 239273f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 239373f53c4aSTejun Heo * 2394112202d9STejun Heo * Prepare pwqs for workqueue flushing. 239573f53c4aSTejun Heo * 2396112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2397112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2398112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2399112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2400112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 240173f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 240273f53c4aSTejun Heo * 240373f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 240473f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 240573f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 240673f53c4aSTejun Heo * is returned. 240773f53c4aSTejun Heo * 2408112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 240973f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 241073f53c4aSTejun Heo * advanced to @work_color. 241173f53c4aSTejun Heo * 241273f53c4aSTejun Heo * CONTEXT: 24133c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 241473f53c4aSTejun Heo * 2415d185af30SYacine Belkadi * Return: 241673f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 241773f53c4aSTejun Heo * otherwise. 241873f53c4aSTejun Heo */ 2419112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 242073f53c4aSTejun Heo int flush_color, int work_color) 24211da177e4SLinus Torvalds { 242273f53c4aSTejun Heo bool wait = false; 242349e3cf44STejun Heo struct pool_workqueue *pwq; 24241da177e4SLinus Torvalds 242573f53c4aSTejun Heo if (flush_color >= 0) { 24266183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2427112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2428dc186ad7SThomas Gleixner } 242914441960SOleg Nesterov 243049e3cf44STejun Heo for_each_pwq(pwq, wq) { 2431112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24321da177e4SLinus Torvalds 2433b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 243473f53c4aSTejun Heo 243573f53c4aSTejun Heo if (flush_color >= 0) { 24366183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 243773f53c4aSTejun Heo 2438112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2439112202d9STejun Heo pwq->flush_color = flush_color; 2440112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 244173f53c4aSTejun Heo wait = true; 24421da177e4SLinus Torvalds } 244373f53c4aSTejun Heo } 244473f53c4aSTejun Heo 244573f53c4aSTejun Heo if (work_color >= 0) { 24466183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2447112202d9STejun Heo pwq->work_color = work_color; 244873f53c4aSTejun Heo } 244973f53c4aSTejun Heo 2450b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 24511da177e4SLinus Torvalds } 24521da177e4SLinus Torvalds 2453112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 245473f53c4aSTejun Heo complete(&wq->first_flusher->done); 245573f53c4aSTejun Heo 245673f53c4aSTejun Heo return wait; 245783c22520SOleg Nesterov } 24581da177e4SLinus Torvalds 24590fcb78c2SRolf Eike Beer /** 24601da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 24610fcb78c2SRolf Eike Beer * @wq: workqueue to flush 24621da177e4SLinus Torvalds * 2463c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2464c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 24651da177e4SLinus Torvalds */ 24667ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 24671da177e4SLinus Torvalds { 246873f53c4aSTejun Heo struct wq_flusher this_flusher = { 246973f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 247073f53c4aSTejun Heo .flush_color = -1, 247173f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 247273f53c4aSTejun Heo }; 247373f53c4aSTejun Heo int next_color; 2474b1f4ec17SOleg Nesterov 24753295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 24763295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 247773f53c4aSTejun Heo 24783c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 247973f53c4aSTejun Heo 248073f53c4aSTejun Heo /* 248173f53c4aSTejun Heo * Start-to-wait phase 248273f53c4aSTejun Heo */ 248373f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 248473f53c4aSTejun Heo 248573f53c4aSTejun Heo if (next_color != wq->flush_color) { 248673f53c4aSTejun Heo /* 248773f53c4aSTejun Heo * Color space is not full. The current work_color 248873f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 248973f53c4aSTejun Heo * by one. 249073f53c4aSTejun Heo */ 24916183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 249273f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 249373f53c4aSTejun Heo wq->work_color = next_color; 249473f53c4aSTejun Heo 249573f53c4aSTejun Heo if (!wq->first_flusher) { 249673f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 24976183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 249873f53c4aSTejun Heo 249973f53c4aSTejun Heo wq->first_flusher = &this_flusher; 250073f53c4aSTejun Heo 2501112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 250273f53c4aSTejun Heo wq->work_color)) { 250373f53c4aSTejun Heo /* nothing to flush, done */ 250473f53c4aSTejun Heo wq->flush_color = next_color; 250573f53c4aSTejun Heo wq->first_flusher = NULL; 250673f53c4aSTejun Heo goto out_unlock; 250773f53c4aSTejun Heo } 250873f53c4aSTejun Heo } else { 250973f53c4aSTejun Heo /* wait in queue */ 25106183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 251173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2512112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 251373f53c4aSTejun Heo } 251473f53c4aSTejun Heo } else { 251573f53c4aSTejun Heo /* 251673f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 251773f53c4aSTejun Heo * The next flush completion will assign us 251873f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 251973f53c4aSTejun Heo */ 252073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 252173f53c4aSTejun Heo } 252273f53c4aSTejun Heo 25233c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 252473f53c4aSTejun Heo 252573f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 252673f53c4aSTejun Heo 252773f53c4aSTejun Heo /* 252873f53c4aSTejun Heo * Wake-up-and-cascade phase 252973f53c4aSTejun Heo * 253073f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 253173f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 253273f53c4aSTejun Heo */ 253373f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 253473f53c4aSTejun Heo return; 253573f53c4aSTejun Heo 25363c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 253773f53c4aSTejun Heo 25384ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25394ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25404ce48b37STejun Heo goto out_unlock; 25414ce48b37STejun Heo 254273f53c4aSTejun Heo wq->first_flusher = NULL; 254373f53c4aSTejun Heo 25446183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25456183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 254673f53c4aSTejun Heo 254773f53c4aSTejun Heo while (true) { 254873f53c4aSTejun Heo struct wq_flusher *next, *tmp; 254973f53c4aSTejun Heo 255073f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 255173f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 255273f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 255373f53c4aSTejun Heo break; 255473f53c4aSTejun Heo list_del_init(&next->list); 255573f53c4aSTejun Heo complete(&next->done); 255673f53c4aSTejun Heo } 255773f53c4aSTejun Heo 25586183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 255973f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 256073f53c4aSTejun Heo 256173f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 256273f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 256373f53c4aSTejun Heo 256473f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 256573f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 256673f53c4aSTejun Heo /* 256773f53c4aSTejun Heo * Assign the same color to all overflowed 256873f53c4aSTejun Heo * flushers, advance work_color and append to 256973f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 257073f53c4aSTejun Heo * phase for these overflowed flushers. 257173f53c4aSTejun Heo */ 257273f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 257373f53c4aSTejun Heo tmp->flush_color = wq->work_color; 257473f53c4aSTejun Heo 257573f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 257673f53c4aSTejun Heo 257773f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 257873f53c4aSTejun Heo &wq->flusher_queue); 2579112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 258073f53c4aSTejun Heo } 258173f53c4aSTejun Heo 258273f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 25836183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 258473f53c4aSTejun Heo break; 258573f53c4aSTejun Heo } 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo /* 258873f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2589112202d9STejun Heo * the new first flusher and arm pwqs. 259073f53c4aSTejun Heo */ 25916183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 25926183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 259373f53c4aSTejun Heo 259473f53c4aSTejun Heo list_del_init(&next->list); 259573f53c4aSTejun Heo wq->first_flusher = next; 259673f53c4aSTejun Heo 2597112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 259873f53c4aSTejun Heo break; 259973f53c4aSTejun Heo 260073f53c4aSTejun Heo /* 260173f53c4aSTejun Heo * Meh... this color is already done, clear first 260273f53c4aSTejun Heo * flusher and repeat cascading. 260373f53c4aSTejun Heo */ 260473f53c4aSTejun Heo wq->first_flusher = NULL; 260573f53c4aSTejun Heo } 260673f53c4aSTejun Heo 260773f53c4aSTejun Heo out_unlock: 26083c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26091da177e4SLinus Torvalds } 2610ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26111da177e4SLinus Torvalds 26129c5a2ba7STejun Heo /** 26139c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26149c5a2ba7STejun Heo * @wq: workqueue to drain 26159c5a2ba7STejun Heo * 26169c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26179c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26189c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26199c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26209c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26219c5a2ba7STejun Heo * takes too long. 26229c5a2ba7STejun Heo */ 26239c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26249c5a2ba7STejun Heo { 26259c5a2ba7STejun Heo unsigned int flush_cnt = 0; 262649e3cf44STejun Heo struct pool_workqueue *pwq; 26279c5a2ba7STejun Heo 26289c5a2ba7STejun Heo /* 26299c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26309c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2631618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26329c5a2ba7STejun Heo */ 263387fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26349c5a2ba7STejun Heo if (!wq->nr_drainers++) 2635618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 263687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26379c5a2ba7STejun Heo reflush: 26389c5a2ba7STejun Heo flush_workqueue(wq); 26399c5a2ba7STejun Heo 2640b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 264176af4d93STejun Heo 264249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2643fa2563e4SThomas Tuttle bool drained; 26449c5a2ba7STejun Heo 2645b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2646112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2647b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2648fa2563e4SThomas Tuttle 2649fa2563e4SThomas Tuttle if (drained) 26509c5a2ba7STejun Heo continue; 26519c5a2ba7STejun Heo 26529c5a2ba7STejun Heo if (++flush_cnt == 10 || 26539c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2654c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 26559c5a2ba7STejun Heo wq->name, flush_cnt); 265676af4d93STejun Heo 2657b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 26589c5a2ba7STejun Heo goto reflush; 26599c5a2ba7STejun Heo } 26609c5a2ba7STejun Heo 26619c5a2ba7STejun Heo if (!--wq->nr_drainers) 2662618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 266387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26649c5a2ba7STejun Heo } 26659c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 26669c5a2ba7STejun Heo 2667606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2668baf59022STejun Heo { 2669baf59022STejun Heo struct worker *worker = NULL; 2670c9e7cf27STejun Heo struct worker_pool *pool; 2671112202d9STejun Heo struct pool_workqueue *pwq; 2672baf59022STejun Heo 2673baf59022STejun Heo might_sleep(); 2674baf59022STejun Heo 2675fa1b54e6STejun Heo local_irq_disable(); 2676fa1b54e6STejun Heo pool = get_work_pool(work); 2677fa1b54e6STejun Heo if (!pool) { 2678fa1b54e6STejun Heo local_irq_enable(); 2679fa1b54e6STejun Heo return false; 2680fa1b54e6STejun Heo } 2681fa1b54e6STejun Heo 2682fa1b54e6STejun Heo spin_lock(&pool->lock); 26830b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2684112202d9STejun Heo pwq = get_work_pwq(work); 2685112202d9STejun Heo if (pwq) { 2686112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2687baf59022STejun Heo goto already_gone; 2688606a5020STejun Heo } else { 2689c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2690baf59022STejun Heo if (!worker) 2691baf59022STejun Heo goto already_gone; 2692112202d9STejun Heo pwq = worker->current_pwq; 2693606a5020STejun Heo } 2694baf59022STejun Heo 2695112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2696d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2697baf59022STejun Heo 2698e159489bSTejun Heo /* 2699e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2700e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2701e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2702e159489bSTejun Heo * access. 2703e159489bSTejun Heo */ 2704493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2705112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2706e159489bSTejun Heo else 2707112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2708112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2709e159489bSTejun Heo 2710baf59022STejun Heo return true; 2711baf59022STejun Heo already_gone: 2712d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2713baf59022STejun Heo return false; 2714baf59022STejun Heo } 2715baf59022STejun Heo 2716db700897SOleg Nesterov /** 2717401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2718401a8d04STejun Heo * @work: the work to flush 2719db700897SOleg Nesterov * 2720606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2721606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2722401a8d04STejun Heo * 2723d185af30SYacine Belkadi * Return: 2724401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2725401a8d04STejun Heo * %false if it was already idle. 2726db700897SOleg Nesterov */ 2727401a8d04STejun Heo bool flush_work(struct work_struct *work) 2728db700897SOleg Nesterov { 272912997d1aSBjorn Helgaas struct wq_barrier barr; 273012997d1aSBjorn Helgaas 27310976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27320976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27330976dfc1SStephen Boyd 273412997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 273512997d1aSBjorn Helgaas wait_for_completion(&barr.done); 273612997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 273712997d1aSBjorn Helgaas return true; 273812997d1aSBjorn Helgaas } else { 273912997d1aSBjorn Helgaas return false; 274012997d1aSBjorn Helgaas } 2741606a5020STejun Heo } 2742db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2743db700897SOleg Nesterov 27448603e1b3STejun Heo struct cwt_wait { 27458603e1b3STejun Heo wait_queue_t wait; 27468603e1b3STejun Heo struct work_struct *work; 27478603e1b3STejun Heo }; 27488603e1b3STejun Heo 27498603e1b3STejun Heo static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) 27508603e1b3STejun Heo { 27518603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 27528603e1b3STejun Heo 27538603e1b3STejun Heo if (cwait->work != key) 27548603e1b3STejun Heo return 0; 27558603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 27568603e1b3STejun Heo } 27578603e1b3STejun Heo 275836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2759401a8d04STejun Heo { 27608603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2761bbb68dfaSTejun Heo unsigned long flags; 27621f1f642eSOleg Nesterov int ret; 27631f1f642eSOleg Nesterov 27641f1f642eSOleg Nesterov do { 2765bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2766bbb68dfaSTejun Heo /* 27678603e1b3STejun Heo * If someone else is already canceling, wait for it to 27688603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 27698603e1b3STejun Heo * because we may get scheduled between @work's completion 27708603e1b3STejun Heo * and the other canceling task resuming and clearing 27718603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 27728603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 27738603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 27748603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 27758603e1b3STejun Heo * we're hogging the CPU. 27768603e1b3STejun Heo * 27778603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 27788603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 27798603e1b3STejun Heo * wake function which matches @work along with exclusive 27808603e1b3STejun Heo * wait and wakeup. 2781bbb68dfaSTejun Heo */ 27828603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 27838603e1b3STejun Heo struct cwt_wait cwait; 27848603e1b3STejun Heo 27858603e1b3STejun Heo init_wait(&cwait.wait); 27868603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 27878603e1b3STejun Heo cwait.work = work; 27888603e1b3STejun Heo 27898603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 27908603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 27918603e1b3STejun Heo if (work_is_canceling(work)) 27928603e1b3STejun Heo schedule(); 27938603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 27948603e1b3STejun Heo } 27951f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 27961f1f642eSOleg Nesterov 2797bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2798bbb68dfaSTejun Heo mark_work_canceling(work); 2799bbb68dfaSTejun Heo local_irq_restore(flags); 2800bbb68dfaSTejun Heo 2801606a5020STejun Heo flush_work(work); 28027a22ad75STejun Heo clear_work_data(work); 28038603e1b3STejun Heo 28048603e1b3STejun Heo /* 28058603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 28068603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 28078603e1b3STejun Heo * visible there. 28088603e1b3STejun Heo */ 28098603e1b3STejun Heo smp_mb(); 28108603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 28118603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 28128603e1b3STejun Heo 28131f1f642eSOleg Nesterov return ret; 28141f1f642eSOleg Nesterov } 28151f1f642eSOleg Nesterov 28166e84d644SOleg Nesterov /** 2817401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2818401a8d04STejun Heo * @work: the work to cancel 28196e84d644SOleg Nesterov * 2820401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2821401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2822401a8d04STejun Heo * another workqueue. On return from this function, @work is 2823401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28241f1f642eSOleg Nesterov * 2825401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2826401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28276e84d644SOleg Nesterov * 2828401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28296e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2830401a8d04STejun Heo * 2831d185af30SYacine Belkadi * Return: 2832401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28336e84d644SOleg Nesterov */ 2834401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28356e84d644SOleg Nesterov { 283636e227d2STejun Heo return __cancel_work_timer(work, false); 2837b89deed3SOleg Nesterov } 283828e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2839b89deed3SOleg Nesterov 28406e84d644SOleg Nesterov /** 2841401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2842401a8d04STejun Heo * @dwork: the delayed work to flush 28436e84d644SOleg Nesterov * 2844401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2845401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2846401a8d04STejun Heo * considers the last queueing instance of @dwork. 28471f1f642eSOleg Nesterov * 2848d185af30SYacine Belkadi * Return: 2849401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2850401a8d04STejun Heo * %false if it was already idle. 28516e84d644SOleg Nesterov */ 2852401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2853401a8d04STejun Heo { 28548930cabaSTejun Heo local_irq_disable(); 2855401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 285660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28578930cabaSTejun Heo local_irq_enable(); 2858401a8d04STejun Heo return flush_work(&dwork->work); 2859401a8d04STejun Heo } 2860401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2861401a8d04STejun Heo 2862401a8d04STejun Heo /** 286357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 286457b30ae7STejun Heo * @dwork: delayed_work to cancel 286509383498STejun Heo * 2866d185af30SYacine Belkadi * Kill off a pending delayed_work. 2867d185af30SYacine Belkadi * 2868d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2869d185af30SYacine Belkadi * pending. 2870d185af30SYacine Belkadi * 2871d185af30SYacine Belkadi * Note: 2872d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2873d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2874d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 287509383498STejun Heo * 287657b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 287709383498STejun Heo */ 287857b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 287909383498STejun Heo { 288057b30ae7STejun Heo unsigned long flags; 288157b30ae7STejun Heo int ret; 288257b30ae7STejun Heo 288357b30ae7STejun Heo do { 288457b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 288557b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 288657b30ae7STejun Heo 288757b30ae7STejun Heo if (unlikely(ret < 0)) 288857b30ae7STejun Heo return false; 288957b30ae7STejun Heo 28907c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28917c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 289257b30ae7STejun Heo local_irq_restore(flags); 2893c0158ca6SDan Magenheimer return ret; 289409383498STejun Heo } 289557b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 289609383498STejun Heo 289709383498STejun Heo /** 2898401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2899401a8d04STejun Heo * @dwork: the delayed work cancel 2900401a8d04STejun Heo * 2901401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2902401a8d04STejun Heo * 2903d185af30SYacine Belkadi * Return: 2904401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2905401a8d04STejun Heo */ 2906401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29076e84d644SOleg Nesterov { 290836e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29096e84d644SOleg Nesterov } 2910f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29111da177e4SLinus Torvalds 29120fcb78c2SRolf Eike Beer /** 291331ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2914b6136773SAndrew Morton * @func: the function to call 2915b6136773SAndrew Morton * 291631ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 291731ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2918b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 291931ddd871STejun Heo * 2920d185af30SYacine Belkadi * Return: 292131ddd871STejun Heo * 0 on success, -errno on failure. 2922b6136773SAndrew Morton */ 292365f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 292415316ba8SChristoph Lameter { 292515316ba8SChristoph Lameter int cpu; 292638f51568SNamhyung Kim struct work_struct __percpu *works; 292715316ba8SChristoph Lameter 2928b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2929b6136773SAndrew Morton if (!works) 293015316ba8SChristoph Lameter return -ENOMEM; 2931b6136773SAndrew Morton 293295402b38SGautham R Shenoy get_online_cpus(); 293393981800STejun Heo 293415316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29359bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29369bfb1839SIngo Molnar 29379bfb1839SIngo Molnar INIT_WORK(work, func); 29388de6d308SOleg Nesterov schedule_work_on(cpu, work); 293915316ba8SChristoph Lameter } 294093981800STejun Heo 294193981800STejun Heo for_each_online_cpu(cpu) 29428616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 294393981800STejun Heo 294495402b38SGautham R Shenoy put_online_cpus(); 2945b6136773SAndrew Morton free_percpu(works); 294615316ba8SChristoph Lameter return 0; 294715316ba8SChristoph Lameter } 294815316ba8SChristoph Lameter 2949eef6a7d5SAlan Stern /** 2950eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2951eef6a7d5SAlan Stern * 2952eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2953eef6a7d5SAlan Stern * completion. 2954eef6a7d5SAlan Stern * 2955eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2956eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2957eef6a7d5SAlan Stern * will lead to deadlock: 2958eef6a7d5SAlan Stern * 2959eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2960eef6a7d5SAlan Stern * a lock held by your code or its caller. 2961eef6a7d5SAlan Stern * 2962eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2963eef6a7d5SAlan Stern * 2964eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2965eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2966eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2967eef6a7d5SAlan Stern * 2968eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2969eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2970eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2971eef6a7d5SAlan Stern * cancel_work_sync() instead. 2972eef6a7d5SAlan Stern */ 29731da177e4SLinus Torvalds void flush_scheduled_work(void) 29741da177e4SLinus Torvalds { 2975d320c038STejun Heo flush_workqueue(system_wq); 29761da177e4SLinus Torvalds } 2977ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29781da177e4SLinus Torvalds 29791da177e4SLinus Torvalds /** 29801fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29811fa44ecaSJames Bottomley * @fn: the function to execute 29821fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29831fa44ecaSJames Bottomley * be available when the work executes) 29841fa44ecaSJames Bottomley * 29851fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29861fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29871fa44ecaSJames Bottomley * 2988d185af30SYacine Belkadi * Return: 0 - function was executed 29891fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29901fa44ecaSJames Bottomley */ 299165f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29921fa44ecaSJames Bottomley { 29931fa44ecaSJames Bottomley if (!in_interrupt()) { 299465f27f38SDavid Howells fn(&ew->work); 29951fa44ecaSJames Bottomley return 0; 29961fa44ecaSJames Bottomley } 29971fa44ecaSJames Bottomley 299865f27f38SDavid Howells INIT_WORK(&ew->work, fn); 29991fa44ecaSJames Bottomley schedule_work(&ew->work); 30001fa44ecaSJames Bottomley 30011fa44ecaSJames Bottomley return 1; 30021fa44ecaSJames Bottomley } 30031fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30041fa44ecaSJames Bottomley 30057a4e344cSTejun Heo /** 30067a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 30077a4e344cSTejun Heo * @attrs: workqueue_attrs to free 30087a4e344cSTejun Heo * 30097a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 30107a4e344cSTejun Heo */ 30117a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 30127a4e344cSTejun Heo { 30137a4e344cSTejun Heo if (attrs) { 30147a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 30157a4e344cSTejun Heo kfree(attrs); 30167a4e344cSTejun Heo } 30177a4e344cSTejun Heo } 30187a4e344cSTejun Heo 30197a4e344cSTejun Heo /** 30207a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 30217a4e344cSTejun Heo * @gfp_mask: allocation mask to use 30227a4e344cSTejun Heo * 30237a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3024d185af30SYacine Belkadi * return it. 3025d185af30SYacine Belkadi * 3026d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 30277a4e344cSTejun Heo */ 30287a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 30297a4e344cSTejun Heo { 30307a4e344cSTejun Heo struct workqueue_attrs *attrs; 30317a4e344cSTejun Heo 30327a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 30337a4e344cSTejun Heo if (!attrs) 30347a4e344cSTejun Heo goto fail; 30357a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 30367a4e344cSTejun Heo goto fail; 30377a4e344cSTejun Heo 303813e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 30397a4e344cSTejun Heo return attrs; 30407a4e344cSTejun Heo fail: 30417a4e344cSTejun Heo free_workqueue_attrs(attrs); 30427a4e344cSTejun Heo return NULL; 30437a4e344cSTejun Heo } 30447a4e344cSTejun Heo 304529c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 304629c91e99STejun Heo const struct workqueue_attrs *from) 304729c91e99STejun Heo { 304829c91e99STejun Heo to->nice = from->nice; 304929c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 30502865a8fbSShaohua Li /* 30512865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 30522865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 30532865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 30542865a8fbSShaohua Li */ 30552865a8fbSShaohua Li to->no_numa = from->no_numa; 305629c91e99STejun Heo } 305729c91e99STejun Heo 305829c91e99STejun Heo /* hash value of the content of @attr */ 305929c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 306029c91e99STejun Heo { 306129c91e99STejun Heo u32 hash = 0; 306229c91e99STejun Heo 306329c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 306413e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 306513e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 306629c91e99STejun Heo return hash; 306729c91e99STejun Heo } 306829c91e99STejun Heo 306929c91e99STejun Heo /* content equality test */ 307029c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 307129c91e99STejun Heo const struct workqueue_attrs *b) 307229c91e99STejun Heo { 307329c91e99STejun Heo if (a->nice != b->nice) 307429c91e99STejun Heo return false; 307529c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 307629c91e99STejun Heo return false; 307729c91e99STejun Heo return true; 307829c91e99STejun Heo } 307929c91e99STejun Heo 30807a4e344cSTejun Heo /** 30817a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 30827a4e344cSTejun Heo * @pool: worker_pool to initialize 30837a4e344cSTejun Heo * 30847a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 3085d185af30SYacine Belkadi * 3086d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 308729c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 308829c91e99STejun Heo * on @pool safely to release it. 30897a4e344cSTejun Heo */ 30907a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 30914e1a1f9aSTejun Heo { 30924e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 309329c91e99STejun Heo pool->id = -1; 309429c91e99STejun Heo pool->cpu = -1; 3095f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 30964e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 30974e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 30984e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 30994e1a1f9aSTejun Heo hash_init(pool->busy_hash); 31004e1a1f9aSTejun Heo 31014e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 31024e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 31034e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 31044e1a1f9aSTejun Heo 31054e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 31064e1a1f9aSTejun Heo (unsigned long)pool); 31074e1a1f9aSTejun Heo 31084e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 310992f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3110da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 31117a4e344cSTejun Heo 31127cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 311329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 311429c91e99STejun Heo pool->refcnt = 1; 311529c91e99STejun Heo 311629c91e99STejun Heo /* shouldn't fail above this point */ 31177a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 31187a4e344cSTejun Heo if (!pool->attrs) 31197a4e344cSTejun Heo return -ENOMEM; 31207a4e344cSTejun Heo return 0; 31214e1a1f9aSTejun Heo } 31224e1a1f9aSTejun Heo 3123e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3124e2dca7adSTejun Heo { 3125e2dca7adSTejun Heo struct workqueue_struct *wq = 3126e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3127e2dca7adSTejun Heo 3128e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3129e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3130e2dca7adSTejun Heo else 3131e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3132e2dca7adSTejun Heo 3133e2dca7adSTejun Heo kfree(wq->rescuer); 3134e2dca7adSTejun Heo kfree(wq); 3135e2dca7adSTejun Heo } 3136e2dca7adSTejun Heo 313729c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 313829c91e99STejun Heo { 313929c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 314029c91e99STejun Heo 31417cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 314229c91e99STejun Heo free_workqueue_attrs(pool->attrs); 314329c91e99STejun Heo kfree(pool); 314429c91e99STejun Heo } 314529c91e99STejun Heo 314629c91e99STejun Heo /** 314729c91e99STejun Heo * put_unbound_pool - put a worker_pool 314829c91e99STejun Heo * @pool: worker_pool to put 314929c91e99STejun Heo * 315029c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3151c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3152c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3153c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3154a892caccSTejun Heo * 3155a892caccSTejun Heo * Should be called with wq_pool_mutex held. 315629c91e99STejun Heo */ 315729c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 315829c91e99STejun Heo { 315960f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 316029c91e99STejun Heo struct worker *worker; 316129c91e99STejun Heo 3162a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3163a892caccSTejun Heo 3164a892caccSTejun Heo if (--pool->refcnt) 316529c91e99STejun Heo return; 316629c91e99STejun Heo 316729c91e99STejun Heo /* sanity checks */ 316861d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3169a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 317029c91e99STejun Heo return; 317129c91e99STejun Heo 317229c91e99STejun Heo /* release id and unhash */ 317329c91e99STejun Heo if (pool->id >= 0) 317429c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 317529c91e99STejun Heo hash_del(&pool->hash_node); 317629c91e99STejun Heo 3177c5aa87bbSTejun Heo /* 3178c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3179c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 318092f9c5c4SLai Jiangshan * attach_mutex. 3181c5aa87bbSTejun Heo */ 318229c91e99STejun Heo mutex_lock(&pool->manager_arb); 318329c91e99STejun Heo 318460f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 31851037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 318629c91e99STejun Heo destroy_worker(worker); 318729c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 318829c91e99STejun Heo spin_unlock_irq(&pool->lock); 318960f5a4bcSLai Jiangshan 319092f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3191da028469SLai Jiangshan if (!list_empty(&pool->workers)) 319260f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 319392f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 319460f5a4bcSLai Jiangshan 319560f5a4bcSLai Jiangshan if (pool->detach_completion) 319660f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 319760f5a4bcSLai Jiangshan 319829c91e99STejun Heo mutex_unlock(&pool->manager_arb); 319929c91e99STejun Heo 320029c91e99STejun Heo /* shut down the timers */ 320129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 320229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 320329c91e99STejun Heo 320429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 320529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 320629c91e99STejun Heo } 320729c91e99STejun Heo 320829c91e99STejun Heo /** 320929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 321029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 321129c91e99STejun Heo * 321229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 321329c91e99STejun Heo * reference count and return it. If there already is a matching 321429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3215d185af30SYacine Belkadi * create a new one. 3216a892caccSTejun Heo * 3217a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3218d185af30SYacine Belkadi * 3219d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3220d185af30SYacine Belkadi * On failure, %NULL. 322129c91e99STejun Heo */ 322229c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 322329c91e99STejun Heo { 322429c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 322529c91e99STejun Heo struct worker_pool *pool; 3226f3f90ad4STejun Heo int node; 322729c91e99STejun Heo 3228a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 322929c91e99STejun Heo 323029c91e99STejun Heo /* do we already have a matching pool? */ 323129c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 323229c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 323329c91e99STejun Heo pool->refcnt++; 32343fb1823cSLai Jiangshan return pool; 323529c91e99STejun Heo } 323629c91e99STejun Heo } 323729c91e99STejun Heo 323829c91e99STejun Heo /* nope, create a new one */ 323929c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 324029c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 324129c91e99STejun Heo goto fail; 324229c91e99STejun Heo 32438864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 324429c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 324529c91e99STejun Heo 32462865a8fbSShaohua Li /* 32472865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 32482865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 32492865a8fbSShaohua Li */ 32502865a8fbSShaohua Li pool->attrs->no_numa = false; 32512865a8fbSShaohua Li 3252f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3253f3f90ad4STejun Heo if (wq_numa_enabled) { 3254f3f90ad4STejun Heo for_each_node(node) { 3255f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3256f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3257f3f90ad4STejun Heo pool->node = node; 3258f3f90ad4STejun Heo break; 3259f3f90ad4STejun Heo } 3260f3f90ad4STejun Heo } 3261f3f90ad4STejun Heo } 3262f3f90ad4STejun Heo 326329c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 326429c91e99STejun Heo goto fail; 326529c91e99STejun Heo 326629c91e99STejun Heo /* create and start the initial worker */ 3267051e1850SLai Jiangshan if (!create_worker(pool)) 326829c91e99STejun Heo goto fail; 326929c91e99STejun Heo 327029c91e99STejun Heo /* install */ 327129c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 32723fb1823cSLai Jiangshan 327329c91e99STejun Heo return pool; 327429c91e99STejun Heo fail: 327529c91e99STejun Heo if (pool) 327629c91e99STejun Heo put_unbound_pool(pool); 327729c91e99STejun Heo return NULL; 327829c91e99STejun Heo } 327929c91e99STejun Heo 32808864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 32818864b4e5STejun Heo { 32828864b4e5STejun Heo kmem_cache_free(pwq_cache, 32838864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 32848864b4e5STejun Heo } 32858864b4e5STejun Heo 32868864b4e5STejun Heo /* 32878864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 32888864b4e5STejun Heo * and needs to be destroyed. 32898864b4e5STejun Heo */ 32908864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 32918864b4e5STejun Heo { 32928864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 32938864b4e5STejun Heo unbound_release_work); 32948864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 32958864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3296bc0caf09STejun Heo bool is_last; 32978864b4e5STejun Heo 32988864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 32998864b4e5STejun Heo return; 33008864b4e5STejun Heo 33013c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 33028864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3303bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 33043c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 33058864b4e5STejun Heo 3306a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 33078864b4e5STejun Heo put_unbound_pool(pool); 3308a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3309a892caccSTejun Heo 33108864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 33118864b4e5STejun Heo 33128864b4e5STejun Heo /* 33138864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3314e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 33158864b4e5STejun Heo */ 3316e2dca7adSTejun Heo if (is_last) 3317e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 33186029a918STejun Heo } 33198864b4e5STejun Heo 33200fbd95aaSTejun Heo /** 3321699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 33220fbd95aaSTejun Heo * @pwq: target pool_workqueue 33230fbd95aaSTejun Heo * 3324699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3325699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3326699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 33270fbd95aaSTejun Heo */ 3328699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 33290fbd95aaSTejun Heo { 3330699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3331699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3332699ce097STejun Heo 3333699ce097STejun Heo /* for @wq->saved_max_active */ 3334a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3335699ce097STejun Heo 3336699ce097STejun Heo /* fast exit for non-freezable wqs */ 3337699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3338699ce097STejun Heo return; 3339699ce097STejun Heo 3340a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3341699ce097STejun Heo 334274b414eaSLai Jiangshan /* 334374b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 334474b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 334574b414eaSLai Jiangshan * is updated and visible. 334674b414eaSLai Jiangshan */ 334774b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3348699ce097STejun Heo pwq->max_active = wq->saved_max_active; 33490fbd95aaSTejun Heo 33500fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 33510fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 33520fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3353951a078aSLai Jiangshan 3354951a078aSLai Jiangshan /* 3355951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3356951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3357951a078aSLai Jiangshan */ 3358951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3359699ce097STejun Heo } else { 3360699ce097STejun Heo pwq->max_active = 0; 3361699ce097STejun Heo } 3362699ce097STejun Heo 3363a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 33640fbd95aaSTejun Heo } 33650fbd95aaSTejun Heo 3366e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3367f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3368f147f29eSTejun Heo struct worker_pool *pool) 3369d2c1d404STejun Heo { 3370d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3371d2c1d404STejun Heo 3372e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3373e50aba9aSTejun Heo 3374d2c1d404STejun Heo pwq->pool = pool; 3375d2c1d404STejun Heo pwq->wq = wq; 3376d2c1d404STejun Heo pwq->flush_color = -1; 33778864b4e5STejun Heo pwq->refcnt = 1; 3378d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 33791befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3380d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 33818864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3382f147f29eSTejun Heo } 3383d2c1d404STejun Heo 3384f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 33851befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3386f147f29eSTejun Heo { 3387f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3388f147f29eSTejun Heo 3389f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 339075ccf595STejun Heo 33911befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 33921befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 33931befcf30STejun Heo return; 33941befcf30STejun Heo 339529b1cb41SLai Jiangshan /* set the matching work_color */ 339675ccf595STejun Heo pwq->work_color = wq->work_color; 3397983ca25eSTejun Heo 3398983ca25eSTejun Heo /* sync max_active to the current setting */ 3399983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3400983ca25eSTejun Heo 3401983ca25eSTejun Heo /* link in @pwq */ 34029e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3403df2d5ae4STejun Heo } 34046029a918STejun Heo 3405f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3406f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3407f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3408f147f29eSTejun Heo { 3409f147f29eSTejun Heo struct worker_pool *pool; 3410f147f29eSTejun Heo struct pool_workqueue *pwq; 3411f147f29eSTejun Heo 3412f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3413f147f29eSTejun Heo 3414f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3415f147f29eSTejun Heo if (!pool) 3416f147f29eSTejun Heo return NULL; 3417f147f29eSTejun Heo 3418e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3419f147f29eSTejun Heo if (!pwq) { 3420f147f29eSTejun Heo put_unbound_pool(pool); 3421f147f29eSTejun Heo return NULL; 3422f147f29eSTejun Heo } 3423f147f29eSTejun Heo 3424f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3425f147f29eSTejun Heo return pwq; 3426d2c1d404STejun Heo } 3427d2c1d404STejun Heo 34284c16bd32STejun Heo /** 34294c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 34304c16bd32STejun Heo * @attrs: the wq_attrs of interest 34314c16bd32STejun Heo * @node: the target NUMA node 34324c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 34334c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 34344c16bd32STejun Heo * 34354c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 34364c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3437d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 34384c16bd32STejun Heo * 34394c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 34404c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 34414c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 34424c16bd32STejun Heo * @attrs->cpumask. 34434c16bd32STejun Heo * 34444c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 34454c16bd32STejun Heo * stable. 3446d185af30SYacine Belkadi * 3447d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3448d185af30SYacine Belkadi * %false if equal. 34494c16bd32STejun Heo */ 34504c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 34514c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 34524c16bd32STejun Heo { 3453d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 34544c16bd32STejun Heo goto use_dfl; 34554c16bd32STejun Heo 34564c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 34574c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 34584c16bd32STejun Heo if (cpu_going_down >= 0) 34594c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 34604c16bd32STejun Heo 34614c16bd32STejun Heo if (cpumask_empty(cpumask)) 34624c16bd32STejun Heo goto use_dfl; 34634c16bd32STejun Heo 34644c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 34654c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 34664c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 34674c16bd32STejun Heo 34684c16bd32STejun Heo use_dfl: 34694c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 34704c16bd32STejun Heo return false; 34714c16bd32STejun Heo } 34724c16bd32STejun Heo 34731befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 34741befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 34751befcf30STejun Heo int node, 34761befcf30STejun Heo struct pool_workqueue *pwq) 34771befcf30STejun Heo { 34781befcf30STejun Heo struct pool_workqueue *old_pwq; 34791befcf30STejun Heo 34801befcf30STejun Heo lockdep_assert_held(&wq->mutex); 34811befcf30STejun Heo 34821befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 34831befcf30STejun Heo link_pwq(pwq); 34841befcf30STejun Heo 34851befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 34861befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 34871befcf30STejun Heo return old_pwq; 34881befcf30STejun Heo } 34891befcf30STejun Heo 34902d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 34912d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 34922d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 34932d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 34942d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 34952d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 34962d5f0764SLai Jiangshan }; 34972d5f0764SLai Jiangshan 34982d5f0764SLai Jiangshan /* free the resources after success or abort */ 34992d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 35002d5f0764SLai Jiangshan { 35012d5f0764SLai Jiangshan if (ctx) { 35022d5f0764SLai Jiangshan int node; 35032d5f0764SLai Jiangshan 35042d5f0764SLai Jiangshan for_each_node(node) 35052d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 35062d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 35072d5f0764SLai Jiangshan 35082d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 35092d5f0764SLai Jiangshan 35102d5f0764SLai Jiangshan kfree(ctx); 35112d5f0764SLai Jiangshan } 35122d5f0764SLai Jiangshan } 35132d5f0764SLai Jiangshan 35142d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 35152d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 35162d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 35172d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 35182d5f0764SLai Jiangshan { 35192d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 35202d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 35212d5f0764SLai Jiangshan int node; 35222d5f0764SLai Jiangshan 35232d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 35242d5f0764SLai Jiangshan 35252d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 35262d5f0764SLai Jiangshan GFP_KERNEL); 35272d5f0764SLai Jiangshan 35282d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 35292d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 35302d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 35312d5f0764SLai Jiangshan goto out_free; 35322d5f0764SLai Jiangshan 35332d5f0764SLai Jiangshan /* make a copy of @attrs and sanitize it */ 35342d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 35352d5f0764SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 35362d5f0764SLai Jiangshan 35372d5f0764SLai Jiangshan /* 35382d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 35392d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 35402d5f0764SLai Jiangshan * pools. 35412d5f0764SLai Jiangshan */ 35422d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 35432d5f0764SLai Jiangshan 35442d5f0764SLai Jiangshan /* 35452d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 35462d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 35472d5f0764SLai Jiangshan * it even if we don't use it immediately. 35482d5f0764SLai Jiangshan */ 35492d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 35502d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 35512d5f0764SLai Jiangshan goto out_free; 35522d5f0764SLai Jiangshan 35532d5f0764SLai Jiangshan for_each_node(node) { 35542d5f0764SLai Jiangshan if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 35552d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 35562d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 35572d5f0764SLai Jiangshan goto out_free; 35582d5f0764SLai Jiangshan } else { 35592d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 35602d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 35612d5f0764SLai Jiangshan } 35622d5f0764SLai Jiangshan } 35632d5f0764SLai Jiangshan 35642d5f0764SLai Jiangshan ctx->attrs = new_attrs; 35652d5f0764SLai Jiangshan ctx->wq = wq; 35662d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 35672d5f0764SLai Jiangshan return ctx; 35682d5f0764SLai Jiangshan 35692d5f0764SLai Jiangshan out_free: 35702d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 35712d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 35722d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 35732d5f0764SLai Jiangshan return NULL; 35742d5f0764SLai Jiangshan } 35752d5f0764SLai Jiangshan 35762d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 35772d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 35782d5f0764SLai Jiangshan { 35792d5f0764SLai Jiangshan int node; 35802d5f0764SLai Jiangshan 35812d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 35822d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 35832d5f0764SLai Jiangshan 35842d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 35852d5f0764SLai Jiangshan 35862d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 35872d5f0764SLai Jiangshan for_each_node(node) 35882d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 35892d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 35902d5f0764SLai Jiangshan 35912d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 35922d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 35932d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 35942d5f0764SLai Jiangshan 35952d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 35962d5f0764SLai Jiangshan } 35972d5f0764SLai Jiangshan 35989e8cd2f5STejun Heo /** 35999e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36009e8cd2f5STejun Heo * @wq: the target workqueue 36019e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36029e8cd2f5STejun Heo * 36034c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 36044c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 36054c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 36064c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 36074c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 36084c16bd32STejun Heo * back-to-back will stay on its current pwq. 36099e8cd2f5STejun Heo * 3610d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3611d185af30SYacine Belkadi * 3612d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 36139e8cd2f5STejun Heo */ 36149e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36159e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36169e8cd2f5STejun Heo { 36172d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 36182d5f0764SLai Jiangshan int ret = -ENOMEM; 36199e8cd2f5STejun Heo 36208719dceaSTejun Heo /* only unbound workqueues can change attributes */ 36219e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 36229e8cd2f5STejun Heo return -EINVAL; 36239e8cd2f5STejun Heo 36248719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 36258719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 36268719dceaSTejun Heo return -EINVAL; 36278719dceaSTejun Heo 36284c16bd32STejun Heo /* 36294c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 36304c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 36314c16bd32STejun Heo * pwqs accordingly. 36324c16bd32STejun Heo */ 36334c16bd32STejun Heo get_online_cpus(); 36344c16bd32STejun Heo 36354c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 36362d5f0764SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 36374c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 36384c16bd32STejun Heo 36392d5f0764SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 36402d5f0764SLai Jiangshan if (ctx) { 36412d5f0764SLai Jiangshan apply_wqattrs_commit(ctx); 36424862125bSTejun Heo ret = 0; 36432d5f0764SLai Jiangshan } 364413e2e556STejun Heo 36454c16bd32STejun Heo put_online_cpus(); 36462d5f0764SLai Jiangshan 36472d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 36482d5f0764SLai Jiangshan 36492d5f0764SLai Jiangshan return ret; 36509e8cd2f5STejun Heo } 36519e8cd2f5STejun Heo 36524c16bd32STejun Heo /** 36534c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 36544c16bd32STejun Heo * @wq: the target workqueue 36554c16bd32STejun Heo * @cpu: the CPU coming up or going down 36564c16bd32STejun Heo * @online: whether @cpu is coming up or going down 36574c16bd32STejun Heo * 36584c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 36594c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 36604c16bd32STejun Heo * @wq accordingly. 36614c16bd32STejun Heo * 36624c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 36634c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 36644c16bd32STejun Heo * correct. 36654c16bd32STejun Heo * 36664c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 36674c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 36684c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 36694c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 36704c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 36714c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 36724c16bd32STejun Heo * CPU_DOWN_PREPARE. 36734c16bd32STejun Heo */ 36744c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 36754c16bd32STejun Heo bool online) 36764c16bd32STejun Heo { 36774c16bd32STejun Heo int node = cpu_to_node(cpu); 36784c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 36794c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 36804c16bd32STejun Heo struct workqueue_attrs *target_attrs; 36814c16bd32STejun Heo cpumask_t *cpumask; 36824c16bd32STejun Heo 36834c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 36844c16bd32STejun Heo 36854c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 36864c16bd32STejun Heo return; 36874c16bd32STejun Heo 36884c16bd32STejun Heo /* 36894c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 36904c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 36914c16bd32STejun Heo * CPU hotplug exclusion. 36924c16bd32STejun Heo */ 36934c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 36944c16bd32STejun Heo cpumask = target_attrs->cpumask; 36954c16bd32STejun Heo 36964c16bd32STejun Heo mutex_lock(&wq->mutex); 3697d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 3698d55262c4STejun Heo goto out_unlock; 36994c16bd32STejun Heo 37004c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 37014c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 37024c16bd32STejun Heo 37034c16bd32STejun Heo /* 37044c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 37054c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 37064c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 3707534a3fbbSDaeseok Youn * wq's, the default pwq should be used. 37084c16bd32STejun Heo */ 37094c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 37104c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 37114c16bd32STejun Heo goto out_unlock; 37124c16bd32STejun Heo } else { 37134c16bd32STejun Heo goto use_dfl_pwq; 37144c16bd32STejun Heo } 37154c16bd32STejun Heo 37164c16bd32STejun Heo mutex_unlock(&wq->mutex); 37174c16bd32STejun Heo 37184c16bd32STejun Heo /* create a new pwq */ 37194c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 37204c16bd32STejun Heo if (!pwq) { 37212d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 37224c16bd32STejun Heo wq->name); 372377f300b1SDaeseok Youn mutex_lock(&wq->mutex); 372477f300b1SDaeseok Youn goto use_dfl_pwq; 37254c16bd32STejun Heo } 37264c16bd32STejun Heo 37274c16bd32STejun Heo /* 37284c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 37294c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 37304c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 37314c16bd32STejun Heo * inbetween. 37324c16bd32STejun Heo */ 37334c16bd32STejun Heo mutex_lock(&wq->mutex); 37344c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 37354c16bd32STejun Heo goto out_unlock; 37364c16bd32STejun Heo 37374c16bd32STejun Heo use_dfl_pwq: 37384c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 37394c16bd32STejun Heo get_pwq(wq->dfl_pwq); 37404c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 37414c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 37424c16bd32STejun Heo out_unlock: 37434c16bd32STejun Heo mutex_unlock(&wq->mutex); 37444c16bd32STejun Heo put_pwq_unlocked(old_pwq); 37454c16bd32STejun Heo } 37464c16bd32STejun Heo 374730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37481da177e4SLinus Torvalds { 374949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 37508a2b7538STejun Heo int cpu, ret; 3751e1d8aa9fSFrederic Weisbecker 375230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3753420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3754420c0ddbSTejun Heo if (!wq->cpu_pwqs) 375530cdf249STejun Heo return -ENOMEM; 375630cdf249STejun Heo 375730cdf249STejun Heo for_each_possible_cpu(cpu) { 37587fb98ea7STejun Heo struct pool_workqueue *pwq = 37597fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37607a62c2c8STejun Heo struct worker_pool *cpu_pools = 3761f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 376230cdf249STejun Heo 3763f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3764f147f29eSTejun Heo 3765f147f29eSTejun Heo mutex_lock(&wq->mutex); 37661befcf30STejun Heo link_pwq(pwq); 3767f147f29eSTejun Heo mutex_unlock(&wq->mutex); 376830cdf249STejun Heo } 376930cdf249STejun Heo return 0; 37708a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 37718a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 37728a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 37738a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 37748a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 37758a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 37768a2b7538STejun Heo return ret; 37779e8cd2f5STejun Heo } else { 37789e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 37799e8cd2f5STejun Heo } 37800f900049STejun Heo } 37810f900049STejun Heo 3782f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3783f3421797STejun Heo const char *name) 3784b71ab8c2STejun Heo { 3785f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3786f3421797STejun Heo 3787f3421797STejun Heo if (max_active < 1 || max_active > lim) 3788044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3789f3421797STejun Heo max_active, name, 1, lim); 3790b71ab8c2STejun Heo 3791f3421797STejun Heo return clamp_val(max_active, 1, lim); 3792b71ab8c2STejun Heo } 3793b71ab8c2STejun Heo 3794b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 379597e37d7bSTejun Heo unsigned int flags, 37961e19ffc6STejun Heo int max_active, 3797eb13ba87SJohannes Berg struct lock_class_key *key, 3798b196be89STejun Heo const char *lock_name, ...) 37993af24433SOleg Nesterov { 3800df2d5ae4STejun Heo size_t tbl_size = 0; 3801ecf6881fSTejun Heo va_list args; 38023af24433SOleg Nesterov struct workqueue_struct *wq; 380349e3cf44STejun Heo struct pool_workqueue *pwq; 3804b196be89STejun Heo 3805cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 3806cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 3807cee22a15SViresh Kumar flags |= WQ_UNBOUND; 3808cee22a15SViresh Kumar 3809ecf6881fSTejun Heo /* allocate wq and format name */ 3810df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 3811ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 3812df2d5ae4STejun Heo 3813df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 3814b196be89STejun Heo if (!wq) 3815d2c1d404STejun Heo return NULL; 3816b196be89STejun Heo 38176029a918STejun Heo if (flags & WQ_UNBOUND) { 38186029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38196029a918STejun Heo if (!wq->unbound_attrs) 38206029a918STejun Heo goto err_free_wq; 38216029a918STejun Heo } 38226029a918STejun Heo 3823ecf6881fSTejun Heo va_start(args, lock_name); 3824ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3825b196be89STejun Heo va_end(args); 38263af24433SOleg Nesterov 3827d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3828b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 38293af24433SOleg Nesterov 3830b196be89STejun Heo /* init wq */ 383197e37d7bSTejun Heo wq->flags = flags; 3832a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 38333c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3834112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 383530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 383673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 383773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3838493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 38393af24433SOleg Nesterov 3840eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3841cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 38423af24433SOleg Nesterov 384330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3844d2c1d404STejun Heo goto err_free_wq; 38451537663fSTejun Heo 3846493008a8STejun Heo /* 3847493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3848493008a8STejun Heo * have a rescuer to guarantee forward progress. 3849493008a8STejun Heo */ 3850493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3851e22bee78STejun Heo struct worker *rescuer; 3852e22bee78STejun Heo 3853f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 3854e22bee78STejun Heo if (!rescuer) 3855d2c1d404STejun Heo goto err_destroy; 3856e22bee78STejun Heo 3857111c225aSTejun Heo rescuer->rescue_wq = wq; 3858111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3859b196be89STejun Heo wq->name); 3860d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3861d2c1d404STejun Heo kfree(rescuer); 3862d2c1d404STejun Heo goto err_destroy; 3863d2c1d404STejun Heo } 3864e22bee78STejun Heo 3865d2c1d404STejun Heo wq->rescuer = rescuer; 386614a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3867e22bee78STejun Heo wake_up_process(rescuer->task); 38683af24433SOleg Nesterov } 38691537663fSTejun Heo 3870226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3871226223abSTejun Heo goto err_destroy; 3872226223abSTejun Heo 38733af24433SOleg Nesterov /* 387468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 387568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 387668e13a67SLai Jiangshan * list. 38773af24433SOleg Nesterov */ 387868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3879a0a1a5fdSTejun Heo 3880a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 388149e3cf44STejun Heo for_each_pwq(pwq, wq) 3882699ce097STejun Heo pwq_adjust_max_active(pwq); 3883a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3884a0a1a5fdSTejun Heo 3885e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 3886a0a1a5fdSTejun Heo 388768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38883af24433SOleg Nesterov 38893af24433SOleg Nesterov return wq; 3890d2c1d404STejun Heo 3891d2c1d404STejun Heo err_free_wq: 38926029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 38934690c4abSTejun Heo kfree(wq); 3894d2c1d404STejun Heo return NULL; 3895d2c1d404STejun Heo err_destroy: 3896d2c1d404STejun Heo destroy_workqueue(wq); 38974690c4abSTejun Heo return NULL; 38981da177e4SLinus Torvalds } 3899d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 39001da177e4SLinus Torvalds 39013af24433SOleg Nesterov /** 39023af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 39033af24433SOleg Nesterov * @wq: target workqueue 39043af24433SOleg Nesterov * 39053af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 39063af24433SOleg Nesterov */ 39073af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 39083af24433SOleg Nesterov { 390949e3cf44STejun Heo struct pool_workqueue *pwq; 39104c16bd32STejun Heo int node; 39113af24433SOleg Nesterov 39129c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 39139c5a2ba7STejun Heo drain_workqueue(wq); 3914c8efcc25STejun Heo 39156183c009STejun Heo /* sanity checks */ 3916b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 391749e3cf44STejun Heo for_each_pwq(pwq, wq) { 39186183c009STejun Heo int i; 39196183c009STejun Heo 392076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 392176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3922b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39236183c009STejun Heo return; 392476af4d93STejun Heo } 392576af4d93STejun Heo } 392676af4d93STejun Heo 39275c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 39288864b4e5STejun Heo WARN_ON(pwq->nr_active) || 392976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3930b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39316183c009STejun Heo return; 39326183c009STejun Heo } 393376af4d93STejun Heo } 3934b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39356183c009STejun Heo 3936a0a1a5fdSTejun Heo /* 3937a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3938a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3939a0a1a5fdSTejun Heo */ 394068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3941e2dca7adSTejun Heo list_del_rcu(&wq->list); 394268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 39433af24433SOleg Nesterov 3944226223abSTejun Heo workqueue_sysfs_unregister(wq); 3945226223abSTejun Heo 3946e2dca7adSTejun Heo if (wq->rescuer) 3947e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3948e22bee78STejun Heo 39498864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 395029c91e99STejun Heo /* 39518864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 3952e2dca7adSTejun Heo * schedule RCU free. 395329c91e99STejun Heo */ 3954e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 39558864b4e5STejun Heo } else { 39568864b4e5STejun Heo /* 39578864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 39584c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 39594c16bd32STejun Heo * @wq will be freed when the last pwq is released. 39608864b4e5STejun Heo */ 39614c16bd32STejun Heo for_each_node(node) { 39624c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 39634c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 39644c16bd32STejun Heo put_pwq_unlocked(pwq); 39654c16bd32STejun Heo } 39664c16bd32STejun Heo 39674c16bd32STejun Heo /* 39684c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 39694c16bd32STejun Heo * put. Don't access it afterwards. 39704c16bd32STejun Heo */ 39714c16bd32STejun Heo pwq = wq->dfl_pwq; 39724c16bd32STejun Heo wq->dfl_pwq = NULL; 3973dce90d47STejun Heo put_pwq_unlocked(pwq); 397429c91e99STejun Heo } 39753af24433SOleg Nesterov } 39763af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 39773af24433SOleg Nesterov 3978dcd989cbSTejun Heo /** 3979dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3980dcd989cbSTejun Heo * @wq: target workqueue 3981dcd989cbSTejun Heo * @max_active: new max_active value. 3982dcd989cbSTejun Heo * 3983dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3984dcd989cbSTejun Heo * 3985dcd989cbSTejun Heo * CONTEXT: 3986dcd989cbSTejun Heo * Don't call from IRQ context. 3987dcd989cbSTejun Heo */ 3988dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3989dcd989cbSTejun Heo { 399049e3cf44STejun Heo struct pool_workqueue *pwq; 3991dcd989cbSTejun Heo 39928719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 39938719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 39948719dceaSTejun Heo return; 39958719dceaSTejun Heo 3996f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3997dcd989cbSTejun Heo 3998a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 3999dcd989cbSTejun Heo 4000dcd989cbSTejun Heo wq->saved_max_active = max_active; 4001dcd989cbSTejun Heo 4002699ce097STejun Heo for_each_pwq(pwq, wq) 4003699ce097STejun Heo pwq_adjust_max_active(pwq); 4004dcd989cbSTejun Heo 4005a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4006dcd989cbSTejun Heo } 4007dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4008dcd989cbSTejun Heo 4009dcd989cbSTejun Heo /** 4010e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4011e6267616STejun Heo * 4012e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4013e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4014d185af30SYacine Belkadi * 4015d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4016e6267616STejun Heo */ 4017e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4018e6267616STejun Heo { 4019e6267616STejun Heo struct worker *worker = current_wq_worker(); 4020e6267616STejun Heo 40216a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4022e6267616STejun Heo } 4023e6267616STejun Heo 4024e6267616STejun Heo /** 4025dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4026dcd989cbSTejun Heo * @cpu: CPU in question 4027dcd989cbSTejun Heo * @wq: target workqueue 4028dcd989cbSTejun Heo * 4029dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4030dcd989cbSTejun Heo * no synchronization around this function and the test result is 4031dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4032dcd989cbSTejun Heo * 4033d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4034d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4035d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4036d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4037d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4038d3251859STejun Heo * 4039d185af30SYacine Belkadi * Return: 4040dcd989cbSTejun Heo * %true if congested, %false otherwise. 4041dcd989cbSTejun Heo */ 4042d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4043dcd989cbSTejun Heo { 40447fb98ea7STejun Heo struct pool_workqueue *pwq; 404576af4d93STejun Heo bool ret; 404676af4d93STejun Heo 404788109453SLai Jiangshan rcu_read_lock_sched(); 40487fb98ea7STejun Heo 4049d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4050d3251859STejun Heo cpu = smp_processor_id(); 4051d3251859STejun Heo 40527fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 40537fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 40547fb98ea7STejun Heo else 4055df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4056dcd989cbSTejun Heo 405776af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 405888109453SLai Jiangshan rcu_read_unlock_sched(); 405976af4d93STejun Heo 406076af4d93STejun Heo return ret; 4061dcd989cbSTejun Heo } 4062dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4063dcd989cbSTejun Heo 4064dcd989cbSTejun Heo /** 4065dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4066dcd989cbSTejun Heo * @work: the work to be tested 4067dcd989cbSTejun Heo * 4068dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4069dcd989cbSTejun Heo * synchronization around this function and the test result is 4070dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4071dcd989cbSTejun Heo * 4072d185af30SYacine Belkadi * Return: 4073dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4074dcd989cbSTejun Heo */ 4075dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4076dcd989cbSTejun Heo { 4077fa1b54e6STejun Heo struct worker_pool *pool; 4078dcd989cbSTejun Heo unsigned long flags; 4079dcd989cbSTejun Heo unsigned int ret = 0; 4080dcd989cbSTejun Heo 4081dcd989cbSTejun Heo if (work_pending(work)) 4082dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4083038366c5SLai Jiangshan 4084fa1b54e6STejun Heo local_irq_save(flags); 4085fa1b54e6STejun Heo pool = get_work_pool(work); 4086038366c5SLai Jiangshan if (pool) { 4087fa1b54e6STejun Heo spin_lock(&pool->lock); 4088c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4089dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4090fa1b54e6STejun Heo spin_unlock(&pool->lock); 4091038366c5SLai Jiangshan } 4092fa1b54e6STejun Heo local_irq_restore(flags); 4093dcd989cbSTejun Heo 4094dcd989cbSTejun Heo return ret; 4095dcd989cbSTejun Heo } 4096dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4097dcd989cbSTejun Heo 40983d1cb205STejun Heo /** 40993d1cb205STejun Heo * set_worker_desc - set description for the current work item 41003d1cb205STejun Heo * @fmt: printf-style format string 41013d1cb205STejun Heo * @...: arguments for the format string 41023d1cb205STejun Heo * 41033d1cb205STejun Heo * This function can be called by a running work function to describe what 41043d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 41053d1cb205STejun Heo * information will be printed out together to help debugging. The 41063d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 41073d1cb205STejun Heo */ 41083d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 41093d1cb205STejun Heo { 41103d1cb205STejun Heo struct worker *worker = current_wq_worker(); 41113d1cb205STejun Heo va_list args; 41123d1cb205STejun Heo 41133d1cb205STejun Heo if (worker) { 41143d1cb205STejun Heo va_start(args, fmt); 41153d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 41163d1cb205STejun Heo va_end(args); 41173d1cb205STejun Heo worker->desc_valid = true; 41183d1cb205STejun Heo } 41193d1cb205STejun Heo } 41203d1cb205STejun Heo 41213d1cb205STejun Heo /** 41223d1cb205STejun Heo * print_worker_info - print out worker information and description 41233d1cb205STejun Heo * @log_lvl: the log level to use when printing 41243d1cb205STejun Heo * @task: target task 41253d1cb205STejun Heo * 41263d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 41273d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 41283d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 41293d1cb205STejun Heo * 41303d1cb205STejun Heo * This function can be safely called on any task as long as the 41313d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 41323d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 41333d1cb205STejun Heo */ 41343d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 41353d1cb205STejun Heo { 41363d1cb205STejun Heo work_func_t *fn = NULL; 41373d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 41383d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 41393d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 41403d1cb205STejun Heo struct workqueue_struct *wq = NULL; 41413d1cb205STejun Heo bool desc_valid = false; 41423d1cb205STejun Heo struct worker *worker; 41433d1cb205STejun Heo 41443d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 41453d1cb205STejun Heo return; 41463d1cb205STejun Heo 41473d1cb205STejun Heo /* 41483d1cb205STejun Heo * This function is called without any synchronization and @task 41493d1cb205STejun Heo * could be in any state. Be careful with dereferences. 41503d1cb205STejun Heo */ 41513d1cb205STejun Heo worker = probe_kthread_data(task); 41523d1cb205STejun Heo 41533d1cb205STejun Heo /* 41543d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 41553d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 41563d1cb205STejun Heo */ 41573d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 41583d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 41593d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 41603d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 41613d1cb205STejun Heo 41623d1cb205STejun Heo /* copy worker description */ 41633d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 41643d1cb205STejun Heo if (desc_valid) 41653d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 41663d1cb205STejun Heo 41673d1cb205STejun Heo if (fn || name[0] || desc[0]) { 41683d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 41693d1cb205STejun Heo if (desc[0]) 41703d1cb205STejun Heo pr_cont(" (%s)", desc); 41713d1cb205STejun Heo pr_cont("\n"); 41723d1cb205STejun Heo } 41733d1cb205STejun Heo } 41743d1cb205STejun Heo 41753494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 41763494fc30STejun Heo { 41773494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 41783494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 41793494fc30STejun Heo pr_cont(" node=%d", pool->node); 41803494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 41813494fc30STejun Heo } 41823494fc30STejun Heo 41833494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 41843494fc30STejun Heo { 41853494fc30STejun Heo if (work->func == wq_barrier_func) { 41863494fc30STejun Heo struct wq_barrier *barr; 41873494fc30STejun Heo 41883494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 41893494fc30STejun Heo 41903494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 41913494fc30STejun Heo task_pid_nr(barr->task)); 41923494fc30STejun Heo } else { 41933494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 41943494fc30STejun Heo } 41953494fc30STejun Heo } 41963494fc30STejun Heo 41973494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 41983494fc30STejun Heo { 41993494fc30STejun Heo struct worker_pool *pool = pwq->pool; 42003494fc30STejun Heo struct work_struct *work; 42013494fc30STejun Heo struct worker *worker; 42023494fc30STejun Heo bool has_in_flight = false, has_pending = false; 42033494fc30STejun Heo int bkt; 42043494fc30STejun Heo 42053494fc30STejun Heo pr_info(" pwq %d:", pool->id); 42063494fc30STejun Heo pr_cont_pool_info(pool); 42073494fc30STejun Heo 42083494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 42093494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 42103494fc30STejun Heo 42113494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 42123494fc30STejun Heo if (worker->current_pwq == pwq) { 42133494fc30STejun Heo has_in_flight = true; 42143494fc30STejun Heo break; 42153494fc30STejun Heo } 42163494fc30STejun Heo } 42173494fc30STejun Heo if (has_in_flight) { 42183494fc30STejun Heo bool comma = false; 42193494fc30STejun Heo 42203494fc30STejun Heo pr_info(" in-flight:"); 42213494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 42223494fc30STejun Heo if (worker->current_pwq != pwq) 42233494fc30STejun Heo continue; 42243494fc30STejun Heo 42253494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 42263494fc30STejun Heo task_pid_nr(worker->task), 42273494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 42283494fc30STejun Heo worker->current_func); 42293494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 42303494fc30STejun Heo pr_cont_work(false, work); 42313494fc30STejun Heo comma = true; 42323494fc30STejun Heo } 42333494fc30STejun Heo pr_cont("\n"); 42343494fc30STejun Heo } 42353494fc30STejun Heo 42363494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 42373494fc30STejun Heo if (get_work_pwq(work) == pwq) { 42383494fc30STejun Heo has_pending = true; 42393494fc30STejun Heo break; 42403494fc30STejun Heo } 42413494fc30STejun Heo } 42423494fc30STejun Heo if (has_pending) { 42433494fc30STejun Heo bool comma = false; 42443494fc30STejun Heo 42453494fc30STejun Heo pr_info(" pending:"); 42463494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 42473494fc30STejun Heo if (get_work_pwq(work) != pwq) 42483494fc30STejun Heo continue; 42493494fc30STejun Heo 42503494fc30STejun Heo pr_cont_work(comma, work); 42513494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 42523494fc30STejun Heo } 42533494fc30STejun Heo pr_cont("\n"); 42543494fc30STejun Heo } 42553494fc30STejun Heo 42563494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 42573494fc30STejun Heo bool comma = false; 42583494fc30STejun Heo 42593494fc30STejun Heo pr_info(" delayed:"); 42603494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 42613494fc30STejun Heo pr_cont_work(comma, work); 42623494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 42633494fc30STejun Heo } 42643494fc30STejun Heo pr_cont("\n"); 42653494fc30STejun Heo } 42663494fc30STejun Heo } 42673494fc30STejun Heo 42683494fc30STejun Heo /** 42693494fc30STejun Heo * show_workqueue_state - dump workqueue state 42703494fc30STejun Heo * 42713494fc30STejun Heo * Called from a sysrq handler and prints out all busy workqueues and 42723494fc30STejun Heo * pools. 42733494fc30STejun Heo */ 42743494fc30STejun Heo void show_workqueue_state(void) 42753494fc30STejun Heo { 42763494fc30STejun Heo struct workqueue_struct *wq; 42773494fc30STejun Heo struct worker_pool *pool; 42783494fc30STejun Heo unsigned long flags; 42793494fc30STejun Heo int pi; 42803494fc30STejun Heo 42813494fc30STejun Heo rcu_read_lock_sched(); 42823494fc30STejun Heo 42833494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 42843494fc30STejun Heo 42853494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 42863494fc30STejun Heo struct pool_workqueue *pwq; 42873494fc30STejun Heo bool idle = true; 42883494fc30STejun Heo 42893494fc30STejun Heo for_each_pwq(pwq, wq) { 42903494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 42913494fc30STejun Heo idle = false; 42923494fc30STejun Heo break; 42933494fc30STejun Heo } 42943494fc30STejun Heo } 42953494fc30STejun Heo if (idle) 42963494fc30STejun Heo continue; 42973494fc30STejun Heo 42983494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 42993494fc30STejun Heo 43003494fc30STejun Heo for_each_pwq(pwq, wq) { 43013494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 43023494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 43033494fc30STejun Heo show_pwq(pwq); 43043494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 43053494fc30STejun Heo } 43063494fc30STejun Heo } 43073494fc30STejun Heo 43083494fc30STejun Heo for_each_pool(pool, pi) { 43093494fc30STejun Heo struct worker *worker; 43103494fc30STejun Heo bool first = true; 43113494fc30STejun Heo 43123494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 43133494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 43143494fc30STejun Heo goto next_pool; 43153494fc30STejun Heo 43163494fc30STejun Heo pr_info("pool %d:", pool->id); 43173494fc30STejun Heo pr_cont_pool_info(pool); 43183494fc30STejun Heo pr_cont(" workers=%d", pool->nr_workers); 43193494fc30STejun Heo if (pool->manager) 43203494fc30STejun Heo pr_cont(" manager: %d", 43213494fc30STejun Heo task_pid_nr(pool->manager->task)); 43223494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 43233494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 43243494fc30STejun Heo task_pid_nr(worker->task)); 43253494fc30STejun Heo first = false; 43263494fc30STejun Heo } 43273494fc30STejun Heo pr_cont("\n"); 43283494fc30STejun Heo next_pool: 43293494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 43303494fc30STejun Heo } 43313494fc30STejun Heo 43323494fc30STejun Heo rcu_read_unlock_sched(); 43333494fc30STejun Heo } 43343494fc30STejun Heo 4335db7bccf4STejun Heo /* 4336db7bccf4STejun Heo * CPU hotplug. 4337db7bccf4STejun Heo * 4338e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4339112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4340706026c2STejun Heo * pool which make migrating pending and scheduled works very 4341e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 434294cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4343e22bee78STejun Heo * blocked draining impractical. 4344e22bee78STejun Heo * 434524647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4346628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4347628c78e7STejun Heo * cpu comes back online. 4348db7bccf4STejun Heo */ 4349db7bccf4STejun Heo 4350706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4351db7bccf4STejun Heo { 435238db41d9STejun Heo int cpu = smp_processor_id(); 43534ce62e9eSTejun Heo struct worker_pool *pool; 4354db7bccf4STejun Heo struct worker *worker; 4355db7bccf4STejun Heo 4356f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 435792f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 435894cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4359e22bee78STejun Heo 4360f2d5a0eeSTejun Heo /* 436192f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 436294cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 436394cf58bbSTejun Heo * except for the ones which are still executing works from 436494cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 436594cf58bbSTejun Heo * this, they may become diasporas. 4366f2d5a0eeSTejun Heo */ 4367da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4368403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4369db7bccf4STejun Heo 437024647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4371f2d5a0eeSTejun Heo 437294cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 437392f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4374e22bee78STejun Heo 4375e22bee78STejun Heo /* 4376eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4377eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4378eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4379eb283428SLai Jiangshan * from other cpus. 4380628c78e7STejun Heo */ 4381628c78e7STejun Heo schedule(); 4382628c78e7STejun Heo 4383628c78e7STejun Heo /* 4384eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4385eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4386eb283428SLai Jiangshan * and keep_working() are always true as long as the 4387eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4388eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4389eb283428SLai Jiangshan * are served by workers tied to the pool. 4390e22bee78STejun Heo */ 4391e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4392eb283428SLai Jiangshan 4393eb283428SLai Jiangshan /* 4394eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4395eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4396eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4397eb283428SLai Jiangshan */ 4398eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4399eb283428SLai Jiangshan wake_up_worker(pool); 4400eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4401eb283428SLai Jiangshan } 4402db7bccf4STejun Heo } 4403db7bccf4STejun Heo 4404bd7c089eSTejun Heo /** 4405bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4406bd7c089eSTejun Heo * @pool: pool of interest 4407bd7c089eSTejun Heo * 4408a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4409bd7c089eSTejun Heo */ 4410bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4411bd7c089eSTejun Heo { 4412a9ab775bSTejun Heo struct worker *worker; 4413bd7c089eSTejun Heo 441492f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4415bd7c089eSTejun Heo 4416bd7c089eSTejun Heo /* 4417a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4418a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4419a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4420a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4421a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4422bd7c089eSTejun Heo */ 4423da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4424a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4425a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4426a9ab775bSTejun Heo 4427a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 44283de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4429a9ab775bSTejun Heo 4430da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4431a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4432a9ab775bSTejun Heo 4433a9ab775bSTejun Heo /* 4434a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4435a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4436a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4437a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4438a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4439a9ab775bSTejun Heo * be bound before @pool->lock is released. 4440a9ab775bSTejun Heo */ 4441a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4442bd7c089eSTejun Heo wake_up_process(worker->task); 4443bd7c089eSTejun Heo 4444bd7c089eSTejun Heo /* 4445a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4446a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4447a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4448a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4449a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4450a9ab775bSTejun Heo * concurrency management. Note that when or whether 4451a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4452a9ab775bSTejun Heo * 4453a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4454a9ab775bSTejun Heo * tested without holding any lock in 4455a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4456a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4457a9ab775bSTejun Heo * management operations. 4458bd7c089eSTejun Heo */ 4459a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4460a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4461a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4462a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4463bd7c089eSTejun Heo } 4464a9ab775bSTejun Heo 4465a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4466bd7c089eSTejun Heo } 4467bd7c089eSTejun Heo 44687dbc725eSTejun Heo /** 44697dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 44707dbc725eSTejun Heo * @pool: unbound pool of interest 44717dbc725eSTejun Heo * @cpu: the CPU which is coming up 44727dbc725eSTejun Heo * 44737dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 44747dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 44757dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 44767dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 44777dbc725eSTejun Heo */ 44787dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 44797dbc725eSTejun Heo { 44807dbc725eSTejun Heo static cpumask_t cpumask; 44817dbc725eSTejun Heo struct worker *worker; 44827dbc725eSTejun Heo 448392f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 44847dbc725eSTejun Heo 44857dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 44867dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 44877dbc725eSTejun Heo return; 44887dbc725eSTejun Heo 44897dbc725eSTejun Heo /* is @cpu the only online CPU? */ 44907dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 44917dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 44927dbc725eSTejun Heo return; 44937dbc725eSTejun Heo 44947dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4495da028469SLai Jiangshan for_each_pool_worker(worker, pool) 44967dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 44977dbc725eSTejun Heo pool->attrs->cpumask) < 0); 44987dbc725eSTejun Heo } 44997dbc725eSTejun Heo 45008db25e78STejun Heo /* 45018db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 45028db25e78STejun Heo * This will be registered high priority CPU notifier. 45038db25e78STejun Heo */ 45040db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 45051da177e4SLinus Torvalds unsigned long action, 45061da177e4SLinus Torvalds void *hcpu) 45071da177e4SLinus Torvalds { 4508d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45094ce62e9eSTejun Heo struct worker_pool *pool; 45104c16bd32STejun Heo struct workqueue_struct *wq; 45117dbc725eSTejun Heo int pi; 45121da177e4SLinus Torvalds 45138db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 45143af24433SOleg Nesterov case CPU_UP_PREPARE: 4515f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45163ce63377STejun Heo if (pool->nr_workers) 45173ce63377STejun Heo continue; 4518051e1850SLai Jiangshan if (!create_worker(pool)) 45193ce63377STejun Heo return NOTIFY_BAD; 45203af24433SOleg Nesterov } 45211da177e4SLinus Torvalds break; 45221da177e4SLinus Torvalds 452365758202STejun Heo case CPU_DOWN_FAILED: 452465758202STejun Heo case CPU_ONLINE: 452568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45267dbc725eSTejun Heo 45277dbc725eSTejun Heo for_each_pool(pool, pi) { 452892f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 452994cf58bbSTejun Heo 4530f05b558dSLai Jiangshan if (pool->cpu == cpu) 453194cf58bbSTejun Heo rebind_workers(pool); 4532f05b558dSLai Jiangshan else if (pool->cpu < 0) 45337dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 453494cf58bbSTejun Heo 453592f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 453694cf58bbSTejun Heo } 45377dbc725eSTejun Heo 45384c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45394c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45404c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 45414c16bd32STejun Heo 454268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45438db25e78STejun Heo break; 454465758202STejun Heo } 454565758202STejun Heo return NOTIFY_OK; 454665758202STejun Heo } 454765758202STejun Heo 454865758202STejun Heo /* 454965758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 455065758202STejun Heo * This will be registered as low priority CPU notifier. 455165758202STejun Heo */ 45520db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 455365758202STejun Heo unsigned long action, 455465758202STejun Heo void *hcpu) 455565758202STejun Heo { 4556d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45578db25e78STejun Heo struct work_struct unbind_work; 45584c16bd32STejun Heo struct workqueue_struct *wq; 45598db25e78STejun Heo 456065758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 456165758202STejun Heo case CPU_DOWN_PREPARE: 45624c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4563706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 45647635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 45654c16bd32STejun Heo 45664c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45674c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 45684c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45694c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 45704c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 45714c16bd32STejun Heo 45724c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 45738db25e78STejun Heo flush_work(&unbind_work); 4574440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 45758db25e78STejun Heo break; 457665758202STejun Heo } 457765758202STejun Heo return NOTIFY_OK; 457865758202STejun Heo } 457965758202STejun Heo 45802d3854a3SRusty Russell #ifdef CONFIG_SMP 45818ccad40dSRusty Russell 45822d3854a3SRusty Russell struct work_for_cpu { 4583ed48ece2STejun Heo struct work_struct work; 45842d3854a3SRusty Russell long (*fn)(void *); 45852d3854a3SRusty Russell void *arg; 45862d3854a3SRusty Russell long ret; 45872d3854a3SRusty Russell }; 45882d3854a3SRusty Russell 4589ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 45902d3854a3SRusty Russell { 4591ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4592ed48ece2STejun Heo 45932d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 45942d3854a3SRusty Russell } 45952d3854a3SRusty Russell 45962d3854a3SRusty Russell /** 45972d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 45982d3854a3SRusty Russell * @cpu: the cpu to run on 45992d3854a3SRusty Russell * @fn: the function to run 46002d3854a3SRusty Russell * @arg: the function arg 46012d3854a3SRusty Russell * 460231ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 46036b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4604d185af30SYacine Belkadi * 4605d185af30SYacine Belkadi * Return: The value @fn returns. 46062d3854a3SRusty Russell */ 4607d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 46082d3854a3SRusty Russell { 4609ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 46102d3854a3SRusty Russell 4611ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4612ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 461312997d1aSBjorn Helgaas flush_work(&wfc.work); 4614440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 46152d3854a3SRusty Russell return wfc.ret; 46162d3854a3SRusty Russell } 46172d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 46182d3854a3SRusty Russell #endif /* CONFIG_SMP */ 46192d3854a3SRusty Russell 4620a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4621e7577c50SRusty Russell 4622a0a1a5fdSTejun Heo /** 4623a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4624a0a1a5fdSTejun Heo * 462558a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4626c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4627706026c2STejun Heo * pool->worklist. 4628a0a1a5fdSTejun Heo * 4629a0a1a5fdSTejun Heo * CONTEXT: 4630a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4631a0a1a5fdSTejun Heo */ 4632a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4633a0a1a5fdSTejun Heo { 463424b8a847STejun Heo struct workqueue_struct *wq; 463524b8a847STejun Heo struct pool_workqueue *pwq; 4636a0a1a5fdSTejun Heo 463768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4638a0a1a5fdSTejun Heo 46396183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4640a0a1a5fdSTejun Heo workqueue_freezing = true; 4641a0a1a5fdSTejun Heo 464224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4643a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4644699ce097STejun Heo for_each_pwq(pwq, wq) 4645699ce097STejun Heo pwq_adjust_max_active(pwq); 4646a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4647a1056305STejun Heo } 46485bcab335STejun Heo 464968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4650a0a1a5fdSTejun Heo } 4651a0a1a5fdSTejun Heo 4652a0a1a5fdSTejun Heo /** 465358a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4654a0a1a5fdSTejun Heo * 4655a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4656a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4657a0a1a5fdSTejun Heo * 4658a0a1a5fdSTejun Heo * CONTEXT: 465968e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4660a0a1a5fdSTejun Heo * 4661d185af30SYacine Belkadi * Return: 466258a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 466358a69cb4STejun Heo * is complete. 4664a0a1a5fdSTejun Heo */ 4665a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4666a0a1a5fdSTejun Heo { 4667a0a1a5fdSTejun Heo bool busy = false; 466824b8a847STejun Heo struct workqueue_struct *wq; 466924b8a847STejun Heo struct pool_workqueue *pwq; 4670a0a1a5fdSTejun Heo 467168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4672a0a1a5fdSTejun Heo 46736183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4674a0a1a5fdSTejun Heo 467524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 467624b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 467724b8a847STejun Heo continue; 4678a0a1a5fdSTejun Heo /* 4679a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4680a0a1a5fdSTejun Heo * to peek without lock. 4681a0a1a5fdSTejun Heo */ 468288109453SLai Jiangshan rcu_read_lock_sched(); 468324b8a847STejun Heo for_each_pwq(pwq, wq) { 46846183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4685112202d9STejun Heo if (pwq->nr_active) { 4686a0a1a5fdSTejun Heo busy = true; 468788109453SLai Jiangshan rcu_read_unlock_sched(); 4688a0a1a5fdSTejun Heo goto out_unlock; 4689a0a1a5fdSTejun Heo } 4690a0a1a5fdSTejun Heo } 469188109453SLai Jiangshan rcu_read_unlock_sched(); 4692a0a1a5fdSTejun Heo } 4693a0a1a5fdSTejun Heo out_unlock: 469468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4695a0a1a5fdSTejun Heo return busy; 4696a0a1a5fdSTejun Heo } 4697a0a1a5fdSTejun Heo 4698a0a1a5fdSTejun Heo /** 4699a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4700a0a1a5fdSTejun Heo * 4701a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4702706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4703a0a1a5fdSTejun Heo * 4704a0a1a5fdSTejun Heo * CONTEXT: 4705a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4706a0a1a5fdSTejun Heo */ 4707a0a1a5fdSTejun Heo void thaw_workqueues(void) 4708a0a1a5fdSTejun Heo { 470924b8a847STejun Heo struct workqueue_struct *wq; 471024b8a847STejun Heo struct pool_workqueue *pwq; 4711a0a1a5fdSTejun Heo 471268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4713a0a1a5fdSTejun Heo 4714a0a1a5fdSTejun Heo if (!workqueue_freezing) 4715a0a1a5fdSTejun Heo goto out_unlock; 4716a0a1a5fdSTejun Heo 471774b414eaSLai Jiangshan workqueue_freezing = false; 471824b8a847STejun Heo 471924b8a847STejun Heo /* restore max_active and repopulate worklist */ 472024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4721a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4722699ce097STejun Heo for_each_pwq(pwq, wq) 4723699ce097STejun Heo pwq_adjust_max_active(pwq); 4724a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 472524b8a847STejun Heo } 472624b8a847STejun Heo 4727a0a1a5fdSTejun Heo out_unlock: 472868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4729a0a1a5fdSTejun Heo } 4730a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4731a0a1a5fdSTejun Heo 47326ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 47336ba94429SFrederic Weisbecker /* 47346ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 47356ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 47366ba94429SFrederic Weisbecker * following attributes. 47376ba94429SFrederic Weisbecker * 47386ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 47396ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 47406ba94429SFrederic Weisbecker * 47416ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 47426ba94429SFrederic Weisbecker * 47436ba94429SFrederic Weisbecker * id RO int : the associated pool ID 47446ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 47456ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 47466ba94429SFrederic Weisbecker */ 47476ba94429SFrederic Weisbecker struct wq_device { 47486ba94429SFrederic Weisbecker struct workqueue_struct *wq; 47496ba94429SFrederic Weisbecker struct device dev; 47506ba94429SFrederic Weisbecker }; 47516ba94429SFrederic Weisbecker 47526ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 47536ba94429SFrederic Weisbecker { 47546ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 47556ba94429SFrederic Weisbecker 47566ba94429SFrederic Weisbecker return wq_dev->wq; 47576ba94429SFrederic Weisbecker } 47586ba94429SFrederic Weisbecker 47596ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 47606ba94429SFrederic Weisbecker char *buf) 47616ba94429SFrederic Weisbecker { 47626ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 47636ba94429SFrederic Weisbecker 47646ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 47656ba94429SFrederic Weisbecker } 47666ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 47676ba94429SFrederic Weisbecker 47686ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 47696ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 47706ba94429SFrederic Weisbecker { 47716ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 47726ba94429SFrederic Weisbecker 47736ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 47746ba94429SFrederic Weisbecker } 47756ba94429SFrederic Weisbecker 47766ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 47776ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 47786ba94429SFrederic Weisbecker size_t count) 47796ba94429SFrederic Weisbecker { 47806ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 47816ba94429SFrederic Weisbecker int val; 47826ba94429SFrederic Weisbecker 47836ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 47846ba94429SFrederic Weisbecker return -EINVAL; 47856ba94429SFrederic Weisbecker 47866ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 47876ba94429SFrederic Weisbecker return count; 47886ba94429SFrederic Weisbecker } 47896ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 47906ba94429SFrederic Weisbecker 47916ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 47926ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 47936ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 47946ba94429SFrederic Weisbecker NULL, 47956ba94429SFrederic Weisbecker }; 47966ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 47976ba94429SFrederic Weisbecker 47986ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 47996ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 48006ba94429SFrederic Weisbecker { 48016ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48026ba94429SFrederic Weisbecker const char *delim = ""; 48036ba94429SFrederic Weisbecker int node, written = 0; 48046ba94429SFrederic Weisbecker 48056ba94429SFrederic Weisbecker rcu_read_lock_sched(); 48066ba94429SFrederic Weisbecker for_each_node(node) { 48076ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 48086ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 48096ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 48106ba94429SFrederic Weisbecker delim = " "; 48116ba94429SFrederic Weisbecker } 48126ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 48136ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 48146ba94429SFrederic Weisbecker 48156ba94429SFrederic Weisbecker return written; 48166ba94429SFrederic Weisbecker } 48176ba94429SFrederic Weisbecker 48186ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 48196ba94429SFrederic Weisbecker char *buf) 48206ba94429SFrederic Weisbecker { 48216ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48226ba94429SFrederic Weisbecker int written; 48236ba94429SFrederic Weisbecker 48246ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 48256ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 48266ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 48276ba94429SFrederic Weisbecker 48286ba94429SFrederic Weisbecker return written; 48296ba94429SFrederic Weisbecker } 48306ba94429SFrederic Weisbecker 48316ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 48326ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 48336ba94429SFrederic Weisbecker { 48346ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 48356ba94429SFrederic Weisbecker 48366ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 48376ba94429SFrederic Weisbecker if (!attrs) 48386ba94429SFrederic Weisbecker return NULL; 48396ba94429SFrederic Weisbecker 48406ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 48416ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 48426ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 48436ba94429SFrederic Weisbecker return attrs; 48446ba94429SFrederic Weisbecker } 48456ba94429SFrederic Weisbecker 48466ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 48476ba94429SFrederic Weisbecker const char *buf, size_t count) 48486ba94429SFrederic Weisbecker { 48496ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48506ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 48516ba94429SFrederic Weisbecker int ret; 48526ba94429SFrederic Weisbecker 48536ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 48546ba94429SFrederic Weisbecker if (!attrs) 48556ba94429SFrederic Weisbecker return -ENOMEM; 48566ba94429SFrederic Weisbecker 48576ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 48586ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 48596ba94429SFrederic Weisbecker ret = apply_workqueue_attrs(wq, attrs); 48606ba94429SFrederic Weisbecker else 48616ba94429SFrederic Weisbecker ret = -EINVAL; 48626ba94429SFrederic Weisbecker 48636ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 48646ba94429SFrederic Weisbecker return ret ?: count; 48656ba94429SFrederic Weisbecker } 48666ba94429SFrederic Weisbecker 48676ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 48686ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 48696ba94429SFrederic Weisbecker { 48706ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48716ba94429SFrederic Weisbecker int written; 48726ba94429SFrederic Weisbecker 48736ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 48746ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 48756ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 48766ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 48776ba94429SFrederic Weisbecker return written; 48786ba94429SFrederic Weisbecker } 48796ba94429SFrederic Weisbecker 48806ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 48816ba94429SFrederic Weisbecker struct device_attribute *attr, 48826ba94429SFrederic Weisbecker const char *buf, size_t count) 48836ba94429SFrederic Weisbecker { 48846ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48856ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 48866ba94429SFrederic Weisbecker int ret; 48876ba94429SFrederic Weisbecker 48886ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 48896ba94429SFrederic Weisbecker if (!attrs) 48906ba94429SFrederic Weisbecker return -ENOMEM; 48916ba94429SFrederic Weisbecker 48926ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 48936ba94429SFrederic Weisbecker if (!ret) 48946ba94429SFrederic Weisbecker ret = apply_workqueue_attrs(wq, attrs); 48956ba94429SFrederic Weisbecker 48966ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 48976ba94429SFrederic Weisbecker return ret ?: count; 48986ba94429SFrederic Weisbecker } 48996ba94429SFrederic Weisbecker 49006ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 49016ba94429SFrederic Weisbecker char *buf) 49026ba94429SFrederic Weisbecker { 49036ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49046ba94429SFrederic Weisbecker int written; 49056ba94429SFrederic Weisbecker 49066ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 49076ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 49086ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 49096ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 49106ba94429SFrederic Weisbecker 49116ba94429SFrederic Weisbecker return written; 49126ba94429SFrederic Weisbecker } 49136ba94429SFrederic Weisbecker 49146ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 49156ba94429SFrederic Weisbecker const char *buf, size_t count) 49166ba94429SFrederic Weisbecker { 49176ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49186ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 49196ba94429SFrederic Weisbecker int v, ret; 49206ba94429SFrederic Weisbecker 49216ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 49226ba94429SFrederic Weisbecker if (!attrs) 49236ba94429SFrederic Weisbecker return -ENOMEM; 49246ba94429SFrederic Weisbecker 49256ba94429SFrederic Weisbecker ret = -EINVAL; 49266ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 49276ba94429SFrederic Weisbecker attrs->no_numa = !v; 49286ba94429SFrederic Weisbecker ret = apply_workqueue_attrs(wq, attrs); 49296ba94429SFrederic Weisbecker } 49306ba94429SFrederic Weisbecker 49316ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 49326ba94429SFrederic Weisbecker return ret ?: count; 49336ba94429SFrederic Weisbecker } 49346ba94429SFrederic Weisbecker 49356ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 49366ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 49376ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 49386ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 49396ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 49406ba94429SFrederic Weisbecker __ATTR_NULL, 49416ba94429SFrederic Weisbecker }; 49426ba94429SFrederic Weisbecker 49436ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 49446ba94429SFrederic Weisbecker .name = "workqueue", 49456ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 49466ba94429SFrederic Weisbecker }; 49476ba94429SFrederic Weisbecker 49486ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 49496ba94429SFrederic Weisbecker { 49506ba94429SFrederic Weisbecker return subsys_virtual_register(&wq_subsys, NULL); 49516ba94429SFrederic Weisbecker } 49526ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 49536ba94429SFrederic Weisbecker 49546ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 49556ba94429SFrederic Weisbecker { 49566ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 49576ba94429SFrederic Weisbecker 49586ba94429SFrederic Weisbecker kfree(wq_dev); 49596ba94429SFrederic Weisbecker } 49606ba94429SFrederic Weisbecker 49616ba94429SFrederic Weisbecker /** 49626ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 49636ba94429SFrederic Weisbecker * @wq: the workqueue to register 49646ba94429SFrederic Weisbecker * 49656ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 49666ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 49676ba94429SFrederic Weisbecker * which is the preferred method. 49686ba94429SFrederic Weisbecker * 49696ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 49706ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 49716ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 49726ba94429SFrederic Weisbecker * attributes. 49736ba94429SFrederic Weisbecker * 49746ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 49756ba94429SFrederic Weisbecker */ 49766ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 49776ba94429SFrederic Weisbecker { 49786ba94429SFrederic Weisbecker struct wq_device *wq_dev; 49796ba94429SFrederic Weisbecker int ret; 49806ba94429SFrederic Weisbecker 49816ba94429SFrederic Weisbecker /* 49826ba94429SFrederic Weisbecker * Adjusting max_active or creating new pwqs by applyting 49836ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 49846ba94429SFrederic Weisbecker * workqueues. 49856ba94429SFrederic Weisbecker */ 49866ba94429SFrederic Weisbecker if (WARN_ON(wq->flags & __WQ_ORDERED)) 49876ba94429SFrederic Weisbecker return -EINVAL; 49886ba94429SFrederic Weisbecker 49896ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 49906ba94429SFrederic Weisbecker if (!wq_dev) 49916ba94429SFrederic Weisbecker return -ENOMEM; 49926ba94429SFrederic Weisbecker 49936ba94429SFrederic Weisbecker wq_dev->wq = wq; 49946ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 49956ba94429SFrederic Weisbecker wq_dev->dev.init_name = wq->name; 49966ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 49976ba94429SFrederic Weisbecker 49986ba94429SFrederic Weisbecker /* 49996ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 50006ba94429SFrederic Weisbecker * everything is ready. 50016ba94429SFrederic Weisbecker */ 50026ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 50036ba94429SFrederic Weisbecker 50046ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 50056ba94429SFrederic Weisbecker if (ret) { 50066ba94429SFrederic Weisbecker kfree(wq_dev); 50076ba94429SFrederic Weisbecker wq->wq_dev = NULL; 50086ba94429SFrederic Weisbecker return ret; 50096ba94429SFrederic Weisbecker } 50106ba94429SFrederic Weisbecker 50116ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 50126ba94429SFrederic Weisbecker struct device_attribute *attr; 50136ba94429SFrederic Weisbecker 50146ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 50156ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 50166ba94429SFrederic Weisbecker if (ret) { 50176ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 50186ba94429SFrederic Weisbecker wq->wq_dev = NULL; 50196ba94429SFrederic Weisbecker return ret; 50206ba94429SFrederic Weisbecker } 50216ba94429SFrederic Weisbecker } 50226ba94429SFrederic Weisbecker } 50236ba94429SFrederic Weisbecker 50246ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 50256ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 50266ba94429SFrederic Weisbecker return 0; 50276ba94429SFrederic Weisbecker } 50286ba94429SFrederic Weisbecker 50296ba94429SFrederic Weisbecker /** 50306ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 50316ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 50326ba94429SFrederic Weisbecker * 50336ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 50346ba94429SFrederic Weisbecker */ 50356ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 50366ba94429SFrederic Weisbecker { 50376ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 50386ba94429SFrederic Weisbecker 50396ba94429SFrederic Weisbecker if (!wq->wq_dev) 50406ba94429SFrederic Weisbecker return; 50416ba94429SFrederic Weisbecker 50426ba94429SFrederic Weisbecker wq->wq_dev = NULL; 50436ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 50446ba94429SFrederic Weisbecker } 50456ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 50466ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 50476ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 50486ba94429SFrederic Weisbecker 5049bce90380STejun Heo static void __init wq_numa_init(void) 5050bce90380STejun Heo { 5051bce90380STejun Heo cpumask_var_t *tbl; 5052bce90380STejun Heo int node, cpu; 5053bce90380STejun Heo 5054bce90380STejun Heo if (num_possible_nodes() <= 1) 5055bce90380STejun Heo return; 5056bce90380STejun Heo 5057d55262c4STejun Heo if (wq_disable_numa) { 5058d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5059d55262c4STejun Heo return; 5060d55262c4STejun Heo } 5061d55262c4STejun Heo 50624c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 50634c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 50644c16bd32STejun Heo 5065bce90380STejun Heo /* 5066bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5067bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5068bce90380STejun Heo * fully initialized by now. 5069bce90380STejun Heo */ 5070ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5071bce90380STejun Heo BUG_ON(!tbl); 5072bce90380STejun Heo 5073bce90380STejun Heo for_each_node(node) 50745a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 50751be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5076bce90380STejun Heo 5077bce90380STejun Heo for_each_possible_cpu(cpu) { 5078bce90380STejun Heo node = cpu_to_node(cpu); 5079bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5080bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5081bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5082bce90380STejun Heo return; 5083bce90380STejun Heo } 5084bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5085bce90380STejun Heo } 5086bce90380STejun Heo 5087bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5088bce90380STejun Heo wq_numa_enabled = true; 5089bce90380STejun Heo } 5090bce90380STejun Heo 50916ee0578bSSuresh Siddha static int __init init_workqueues(void) 50921da177e4SLinus Torvalds { 50937a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 50947a4e344cSTejun Heo int i, cpu; 5095c34056a3STejun Heo 5096e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5097e904e6c2STejun Heo 5098e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5099e904e6c2STejun Heo 510065758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 5101a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 51028b03ae3cSTejun Heo 5103bce90380STejun Heo wq_numa_init(); 5104bce90380STejun Heo 5105706026c2STejun Heo /* initialize CPU pools */ 510629c91e99STejun Heo for_each_possible_cpu(cpu) { 51074ce62e9eSTejun Heo struct worker_pool *pool; 51088b03ae3cSTejun Heo 51097a4e344cSTejun Heo i = 0; 5110f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 51117a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5112ec22ca5eSTejun Heo pool->cpu = cpu; 51137a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 51147a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5115f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 51167a4e344cSTejun Heo 51179daf9e67STejun Heo /* alloc pool ID */ 511868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 51199daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 512068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 51214ce62e9eSTejun Heo } 51228b03ae3cSTejun Heo } 51238b03ae3cSTejun Heo 5124e22bee78STejun Heo /* create the initial worker */ 512529c91e99STejun Heo for_each_online_cpu(cpu) { 51264ce62e9eSTejun Heo struct worker_pool *pool; 5127e22bee78STejun Heo 5128f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 512924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5130051e1850SLai Jiangshan BUG_ON(!create_worker(pool)); 5131e22bee78STejun Heo } 51324ce62e9eSTejun Heo } 5133e22bee78STejun Heo 51348a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 513529c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 513629c91e99STejun Heo struct workqueue_attrs *attrs; 513729c91e99STejun Heo 513829c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 513929c91e99STejun Heo attrs->nice = std_nice[i]; 514029c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 51418a2b7538STejun Heo 51428a2b7538STejun Heo /* 51438a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 51448a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 51458a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 51468a2b7538STejun Heo */ 51478a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 51488a2b7538STejun Heo attrs->nice = std_nice[i]; 51498a2b7538STejun Heo attrs->no_numa = true; 51508a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 515129c91e99STejun Heo } 515229c91e99STejun Heo 5153d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 51541aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5155d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5156f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5157f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 515824d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 515924d51addSTejun Heo WQ_FREEZABLE, 0); 51600668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 51610668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 51620668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 51630668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 51640668106cSViresh Kumar 0); 51651aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 51660668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 51670668106cSViresh Kumar !system_power_efficient_wq || 51680668106cSViresh Kumar !system_freezable_power_efficient_wq); 51696ee0578bSSuresh Siddha return 0; 51701da177e4SLinus Torvalds } 51716ee0578bSSuresh Siddha early_initcall(init_workqueues); 5172