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 68bc3a1afcSTejun Heo * manager_mutex to avoid changing binding state while 6924647570STejun Heo * create_worker() is in progress. 70bc2ae0f5STejun Heo */ 7111ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 7224647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 7335b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 74db7bccf4STejun Heo 75c8e55f36STejun Heo /* worker flags */ 76c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 77c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 78c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 79e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 80fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 81f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 82a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 83e22bee78STejun Heo 84a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 85a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 86db7bccf4STejun Heo 87e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 884ce62e9eSTejun Heo 8929c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 90c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 91db7bccf4STejun Heo 92e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 93e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 94e22bee78STejun Heo 953233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 963233cdbdSTejun Heo /* call for help after 10ms 973233cdbdSTejun Heo (min two ticks) */ 98e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 99e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1001da177e4SLinus Torvalds 1011da177e4SLinus Torvalds /* 102e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 103e22bee78STejun Heo * all cpus. Give -20. 104e22bee78STejun Heo */ 105e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1063270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 107ecf6881fSTejun Heo 108ecf6881fSTejun Heo WQ_NAME_LEN = 24, 109c8e55f36STejun Heo }; 110c8e55f36STejun Heo 1111da177e4SLinus Torvalds /* 1124690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1134690c4abSTejun Heo * 114e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 115e41e704bSTejun Heo * everyone else. 1164690c4abSTejun Heo * 117e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 118e22bee78STejun Heo * only be modified and accessed from the local cpu. 119e22bee78STejun Heo * 120d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1214690c4abSTejun Heo * 122d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 123d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 124d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 125d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 126e22bee78STejun Heo * 127822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 128822d8405STejun Heo * locks. Reads can happen under either lock. 129822d8405STejun Heo * 13068e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13176af4d93STejun Heo * 13268e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1335bcab335STejun Heo * 1343c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1353c25a55dSLai Jiangshan * 136b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1372e109a28STejun Heo * 1382e109a28STejun Heo * MD: wq_mayday_lock protected. 1394690c4abSTejun Heo */ 1404690c4abSTejun Heo 1412eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 142c34056a3STejun Heo 143bd7bdd43STejun Heo struct worker_pool { 144d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 145d84ff051STejun Heo int cpu; /* I: the associated cpu */ 146f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1479daf9e67STejun Heo int id; /* I: pool ID */ 14811ebea50STejun Heo unsigned int flags; /* X: flags */ 149bd7bdd43STejun Heo 150bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 151bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 152ea1abd61SLai Jiangshan 153ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 154bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 155bd7bdd43STejun Heo 156bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 157bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 158bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 159bd7bdd43STejun Heo 160c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 161c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 162c9e7cf27STejun Heo /* L: hash of busy workers */ 163c9e7cf27STejun Heo 164bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16534a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 166bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 167822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 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 */ 23468e13a67SLai Jiangshan struct list_head list; /* PL: 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 2612728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2622728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2632728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 264df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2651da177e4SLinus Torvalds }; 2661da177e4SLinus Torvalds 267e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 268e904e6c2STejun Heo 269bce90380STejun Heo static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 270bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 271bce90380STejun Heo /* possible CPUs of each node */ 272bce90380STejun Heo 273d55262c4STejun Heo static bool wq_disable_numa; 274d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 275d55262c4STejun Heo 276cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 277cee22a15SViresh Kumar #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT 278cee22a15SViresh Kumar static bool wq_power_efficient = true; 279cee22a15SViresh Kumar #else 280cee22a15SViresh Kumar static bool wq_power_efficient; 281cee22a15SViresh Kumar #endif 282cee22a15SViresh Kumar 283cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 284cee22a15SViresh Kumar 285bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 286bce90380STejun Heo 2874c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2884c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2894c16bd32STejun Heo 29068e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2912e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2925bcab335STejun Heo 29368e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 29468e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2957d19c5ceSTejun Heo 2967d19c5ceSTejun Heo /* the per-cpu worker pools */ 2977d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2987d19c5ceSTejun Heo cpu_worker_pools); 2997d19c5ceSTejun Heo 30068e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3017d19c5ceSTejun Heo 30268e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 30329c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 30429c91e99STejun Heo 305c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 30629c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 30729c91e99STejun Heo 3088a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3098a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3108a2b7538STejun Heo 311d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 312ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 313044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3141aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 315044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 316d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 317044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 318f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 319044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 32024d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3210668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3220668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3230668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3240668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 325d320c038STejun Heo 3267d19c5ceSTejun Heo static int worker_thread(void *__worker); 3277d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3287d19c5ceSTejun Heo const struct workqueue_attrs *from); 3297d19c5ceSTejun Heo 33097bd2347STejun Heo #define CREATE_TRACE_POINTS 33197bd2347STejun Heo #include <trace/events/workqueue.h> 33297bd2347STejun Heo 33368e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3345bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 33568e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 33668e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3375bcab335STejun Heo 338b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 33976af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 340b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 341b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 34276af4d93STejun Heo 343822d8405STejun Heo #ifdef CONFIG_LOCKDEP 344822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 345519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 346519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 347822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 348822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 349822d8405STejun Heo #else 350822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 351822d8405STejun Heo #endif 352822d8405STejun Heo 353f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 354f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 355f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3567a62c2c8STejun Heo (pool)++) 3574ce62e9eSTejun Heo 35849e3cf44STejun Heo /** 35917116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 36017116969STejun Heo * @pool: iteration cursor 361611c92a0STejun Heo * @pi: integer used for iteration 362fa1b54e6STejun Heo * 36368e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 36468e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 36568e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 366fa1b54e6STejun Heo * 367fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 368fa1b54e6STejun Heo * ignored. 36917116969STejun Heo */ 370611c92a0STejun Heo #define for_each_pool(pool, pi) \ 371611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 37268e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 373fa1b54e6STejun Heo else 37417116969STejun Heo 37517116969STejun Heo /** 376822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 377822d8405STejun Heo * @worker: iteration cursor 378822d8405STejun Heo * @wi: integer used for iteration 379822d8405STejun Heo * @pool: worker_pool to iterate workers of 380822d8405STejun Heo * 381822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 382822d8405STejun Heo * 383822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 384822d8405STejun Heo * ignored. 385822d8405STejun Heo */ 386822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 387822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 388822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 389822d8405STejun Heo else 390822d8405STejun Heo 391822d8405STejun Heo /** 39249e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 39349e3cf44STejun Heo * @pwq: iteration cursor 39449e3cf44STejun Heo * @wq: the target workqueue 39576af4d93STejun Heo * 396b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 397794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 398794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 39976af4d93STejun Heo * 40076af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 40176af4d93STejun Heo * ignored. 40249e3cf44STejun Heo */ 40349e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 40476af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 405b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 40676af4d93STejun Heo else 407f3421797STejun Heo 408dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 409dc186ad7SThomas Gleixner 410dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 411dc186ad7SThomas Gleixner 41299777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 41399777288SStanislaw Gruszka { 41499777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 41599777288SStanislaw Gruszka } 41699777288SStanislaw Gruszka 417dc186ad7SThomas Gleixner /* 418dc186ad7SThomas Gleixner * fixup_init is called when: 419dc186ad7SThomas Gleixner * - an active object is initialized 420dc186ad7SThomas Gleixner */ 421dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 422dc186ad7SThomas Gleixner { 423dc186ad7SThomas Gleixner struct work_struct *work = addr; 424dc186ad7SThomas Gleixner 425dc186ad7SThomas Gleixner switch (state) { 426dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 427dc186ad7SThomas Gleixner cancel_work_sync(work); 428dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 429dc186ad7SThomas Gleixner return 1; 430dc186ad7SThomas Gleixner default: 431dc186ad7SThomas Gleixner return 0; 432dc186ad7SThomas Gleixner } 433dc186ad7SThomas Gleixner } 434dc186ad7SThomas Gleixner 435dc186ad7SThomas Gleixner /* 436dc186ad7SThomas Gleixner * fixup_activate is called when: 437dc186ad7SThomas Gleixner * - an active object is activated 438dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 439dc186ad7SThomas Gleixner */ 440dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 441dc186ad7SThomas Gleixner { 442dc186ad7SThomas Gleixner struct work_struct *work = addr; 443dc186ad7SThomas Gleixner 444dc186ad7SThomas Gleixner switch (state) { 445dc186ad7SThomas Gleixner 446dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 447dc186ad7SThomas Gleixner /* 448dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 449dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 450dc186ad7SThomas Gleixner * is tracked in the object tracker. 451dc186ad7SThomas Gleixner */ 45222df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 453dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 454dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 455dc186ad7SThomas Gleixner return 0; 456dc186ad7SThomas Gleixner } 457dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 458dc186ad7SThomas Gleixner return 0; 459dc186ad7SThomas Gleixner 460dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 461dc186ad7SThomas Gleixner WARN_ON(1); 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner default: 464dc186ad7SThomas Gleixner return 0; 465dc186ad7SThomas Gleixner } 466dc186ad7SThomas Gleixner } 467dc186ad7SThomas Gleixner 468dc186ad7SThomas Gleixner /* 469dc186ad7SThomas Gleixner * fixup_free is called when: 470dc186ad7SThomas Gleixner * - an active object is freed 471dc186ad7SThomas Gleixner */ 472dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 473dc186ad7SThomas Gleixner { 474dc186ad7SThomas Gleixner struct work_struct *work = addr; 475dc186ad7SThomas Gleixner 476dc186ad7SThomas Gleixner switch (state) { 477dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 478dc186ad7SThomas Gleixner cancel_work_sync(work); 479dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 480dc186ad7SThomas Gleixner return 1; 481dc186ad7SThomas Gleixner default: 482dc186ad7SThomas Gleixner return 0; 483dc186ad7SThomas Gleixner } 484dc186ad7SThomas Gleixner } 485dc186ad7SThomas Gleixner 486dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 487dc186ad7SThomas Gleixner .name = "work_struct", 48899777288SStanislaw Gruszka .debug_hint = work_debug_hint, 489dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 490dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 491dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 492dc186ad7SThomas Gleixner }; 493dc186ad7SThomas Gleixner 494dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 495dc186ad7SThomas Gleixner { 496dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner 499dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 500dc186ad7SThomas Gleixner { 501dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 502dc186ad7SThomas Gleixner } 503dc186ad7SThomas Gleixner 504dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 505dc186ad7SThomas Gleixner { 506dc186ad7SThomas Gleixner if (onstack) 507dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 508dc186ad7SThomas Gleixner else 509dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 510dc186ad7SThomas Gleixner } 511dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 512dc186ad7SThomas Gleixner 513dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 514dc186ad7SThomas Gleixner { 515dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 516dc186ad7SThomas Gleixner } 517dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 518dc186ad7SThomas Gleixner 519dc186ad7SThomas Gleixner #else 520dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 521dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 522dc186ad7SThomas Gleixner #endif 523dc186ad7SThomas Gleixner 5244e8b22bdSLi Bin /** 5254e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5264e8b22bdSLi Bin * @pool: the pool pointer of interest 5274e8b22bdSLi Bin * 5284e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5294e8b22bdSLi Bin * successfully, -errno on failure. 5304e8b22bdSLi Bin */ 5319daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5329daf9e67STejun Heo { 5339daf9e67STejun Heo int ret; 5349daf9e67STejun Heo 53568e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5365bcab335STejun Heo 5374e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5384e8b22bdSLi Bin GFP_KERNEL); 539229641a6STejun Heo if (ret >= 0) { 540e68035fbSTejun Heo pool->id = ret; 541229641a6STejun Heo return 0; 542229641a6STejun Heo } 5439daf9e67STejun Heo return ret; 5449daf9e67STejun Heo } 5459daf9e67STejun Heo 54676af4d93STejun Heo /** 547df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 548df2d5ae4STejun Heo * @wq: the target workqueue 549df2d5ae4STejun Heo * @node: the node ID 550df2d5ae4STejun Heo * 551df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 552df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 553df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 554d185af30SYacine Belkadi * 555d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 556df2d5ae4STejun Heo */ 557df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 558df2d5ae4STejun Heo int node) 559df2d5ae4STejun Heo { 560df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 561df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 562df2d5ae4STejun Heo } 563df2d5ae4STejun Heo 56473f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 56573f53c4aSTejun Heo { 56673f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 56773f53c4aSTejun Heo } 56873f53c4aSTejun Heo 56973f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 57073f53c4aSTejun Heo { 57173f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 57273f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 57373f53c4aSTejun Heo } 57473f53c4aSTejun Heo 57573f53c4aSTejun Heo static int work_next_color(int color) 57673f53c4aSTejun Heo { 57773f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 578a848e3b6SOleg Nesterov } 579a848e3b6SOleg Nesterov 5804594bf15SDavid Howells /* 581112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 582112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5837c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5847a22ad75STejun Heo * 585112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 586112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 587bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 588bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5897a22ad75STejun Heo * 590112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5917c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 592112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5937c3eed5cSTejun Heo * available only while the work item is queued. 594bbb68dfaSTejun Heo * 595bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 596bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 597bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 598bbb68dfaSTejun Heo * try to steal the PENDING bit. 5994594bf15SDavid Howells */ 6007a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6017a22ad75STejun Heo unsigned long flags) 6027a22ad75STejun Heo { 6036183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6047a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6057a22ad75STejun Heo } 6067a22ad75STejun Heo 607112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6084690c4abSTejun Heo unsigned long extra_flags) 609365970a1SDavid Howells { 610112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 611112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 612365970a1SDavid Howells } 613365970a1SDavid Howells 6144468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6154468a00fSLai Jiangshan int pool_id) 6164468a00fSLai Jiangshan { 6174468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6184468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6194468a00fSLai Jiangshan } 6204468a00fSLai Jiangshan 6217c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6227c3eed5cSTejun Heo int pool_id) 6234d707b9fSOleg Nesterov { 62423657bb1STejun Heo /* 62523657bb1STejun Heo * The following wmb is paired with the implied mb in 62623657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 62723657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 62823657bb1STejun Heo * owner. 62923657bb1STejun Heo */ 63023657bb1STejun Heo smp_wmb(); 6317c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6324d707b9fSOleg Nesterov } 6334d707b9fSOleg Nesterov 6347a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 635365970a1SDavid Howells { 6367c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6377c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6387a22ad75STejun Heo } 6397a22ad75STejun Heo 640112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6417a22ad75STejun Heo { 642e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6437a22ad75STejun Heo 644112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 645e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 646e120153dSTejun Heo else 647e120153dSTejun Heo return NULL; 6487a22ad75STejun Heo } 6497a22ad75STejun Heo 6507c3eed5cSTejun Heo /** 6517c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6527c3eed5cSTejun Heo * @work: the work item of interest 6537c3eed5cSTejun Heo * 65468e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 65568e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 65668e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 657fa1b54e6STejun Heo * 658fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 659fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 660fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 661fa1b54e6STejun Heo * returned pool is and stays online. 662d185af30SYacine Belkadi * 663d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6647c3eed5cSTejun Heo */ 6657c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6667a22ad75STejun Heo { 667e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6687c3eed5cSTejun Heo int pool_id; 6697a22ad75STejun Heo 67068e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 671fa1b54e6STejun Heo 672112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 673112202d9STejun Heo return ((struct pool_workqueue *) 6747c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6757a22ad75STejun Heo 6767c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6777c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6787a22ad75STejun Heo return NULL; 6797a22ad75STejun Heo 680fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6817c3eed5cSTejun Heo } 6827c3eed5cSTejun Heo 6837c3eed5cSTejun Heo /** 6847c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6857c3eed5cSTejun Heo * @work: the work item of interest 6867c3eed5cSTejun Heo * 687d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6887c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6897c3eed5cSTejun Heo */ 6907c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6917c3eed5cSTejun Heo { 69254d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6937c3eed5cSTejun Heo 694112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 695112202d9STejun Heo return ((struct pool_workqueue *) 69654d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 69754d5b7d0SLai Jiangshan 69854d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6997c3eed5cSTejun Heo } 7007c3eed5cSTejun Heo 701bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 702bbb68dfaSTejun Heo { 7037c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 704bbb68dfaSTejun Heo 7057c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7067c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 707bbb68dfaSTejun Heo } 708bbb68dfaSTejun Heo 709bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 710bbb68dfaSTejun Heo { 711bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 712bbb68dfaSTejun Heo 713112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 714bbb68dfaSTejun Heo } 715bbb68dfaSTejun Heo 716e22bee78STejun Heo /* 7173270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7183270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 719d565ed63STejun Heo * they're being called with pool->lock held. 720e22bee78STejun Heo */ 721e22bee78STejun Heo 72263d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 723649027d7STejun Heo { 724e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 725649027d7STejun Heo } 726649027d7STejun Heo 727e22bee78STejun Heo /* 728e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 729e22bee78STejun Heo * running workers. 730974271c4STejun Heo * 731974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 732706026c2STejun Heo * function will always return %true for unbound pools as long as the 733974271c4STejun Heo * worklist isn't empty. 734e22bee78STejun Heo */ 73563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 736e22bee78STejun Heo { 73763d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 738e22bee78STejun Heo } 739e22bee78STejun Heo 740e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 74163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 742e22bee78STejun Heo { 74363d95a91STejun Heo return pool->nr_idle; 744e22bee78STejun Heo } 745e22bee78STejun Heo 746e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 74763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 748e22bee78STejun Heo { 749e19e397aSTejun Heo return !list_empty(&pool->worklist) && 750e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 751e22bee78STejun Heo } 752e22bee78STejun Heo 753e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 75463d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 755e22bee78STejun Heo { 75663d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 757e22bee78STejun Heo } 758e22bee78STejun Heo 759e22bee78STejun Heo /* Do I need to be the manager? */ 76063d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 761e22bee78STejun Heo { 76263d95a91STejun Heo return need_to_create_worker(pool) || 76311ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 764e22bee78STejun Heo } 765e22bee78STejun Heo 766e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 76763d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 768e22bee78STejun Heo { 76934a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 77063d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 77163d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 772e22bee78STejun Heo 773ea1abd61SLai Jiangshan /* 774ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 775ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 776ea1abd61SLai Jiangshan */ 777ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 778ea1abd61SLai Jiangshan return false; 779ea1abd61SLai Jiangshan 780e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 781e22bee78STejun Heo } 782e22bee78STejun Heo 783e22bee78STejun Heo /* 784e22bee78STejun Heo * Wake up functions. 785e22bee78STejun Heo */ 786e22bee78STejun Heo 7877e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 78863d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7897e11629dSTejun Heo { 79063d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7917e11629dSTejun Heo return NULL; 7927e11629dSTejun Heo 79363d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7947e11629dSTejun Heo } 7957e11629dSTejun Heo 7967e11629dSTejun Heo /** 7977e11629dSTejun Heo * wake_up_worker - wake up an idle worker 79863d95a91STejun Heo * @pool: worker pool to wake worker from 7997e11629dSTejun Heo * 80063d95a91STejun Heo * Wake up the first idle worker of @pool. 8017e11629dSTejun Heo * 8027e11629dSTejun Heo * CONTEXT: 803d565ed63STejun Heo * spin_lock_irq(pool->lock). 8047e11629dSTejun Heo */ 80563d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8067e11629dSTejun Heo { 80763d95a91STejun Heo struct worker *worker = first_worker(pool); 8087e11629dSTejun Heo 8097e11629dSTejun Heo if (likely(worker)) 8107e11629dSTejun Heo wake_up_process(worker->task); 8117e11629dSTejun Heo } 8127e11629dSTejun Heo 8134690c4abSTejun Heo /** 814e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 815e22bee78STejun Heo * @task: task waking up 816e22bee78STejun Heo * @cpu: CPU @task is waking up to 817e22bee78STejun Heo * 818e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 819e22bee78STejun Heo * being awoken. 820e22bee78STejun Heo * 821e22bee78STejun Heo * CONTEXT: 822e22bee78STejun Heo * spin_lock_irq(rq->lock) 823e22bee78STejun Heo */ 824d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 825e22bee78STejun Heo { 826e22bee78STejun Heo struct worker *worker = kthread_data(task); 827e22bee78STejun Heo 82836576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 829ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 830e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 831e22bee78STejun Heo } 83236576000SJoonsoo Kim } 833e22bee78STejun Heo 834e22bee78STejun Heo /** 835e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 836e22bee78STejun Heo * @task: task going to sleep 837e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 838e22bee78STejun Heo * 839e22bee78STejun Heo * This function is called during schedule() when a busy worker is 840e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 841e22bee78STejun Heo * returning pointer to its task. 842e22bee78STejun Heo * 843e22bee78STejun Heo * CONTEXT: 844e22bee78STejun Heo * spin_lock_irq(rq->lock) 845e22bee78STejun Heo * 846d185af30SYacine Belkadi * Return: 847e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 848e22bee78STejun Heo */ 849d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 850e22bee78STejun Heo { 851e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 852111c225aSTejun Heo struct worker_pool *pool; 853e22bee78STejun Heo 854111c225aSTejun Heo /* 855111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 856111c225aSTejun Heo * workers, also reach here, let's not access anything before 857111c225aSTejun Heo * checking NOT_RUNNING. 858111c225aSTejun Heo */ 8592d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 860e22bee78STejun Heo return NULL; 861e22bee78STejun Heo 862111c225aSTejun Heo pool = worker->pool; 863111c225aSTejun Heo 864e22bee78STejun Heo /* this can only happen on the local cpu */ 8656183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8666183c009STejun Heo return NULL; 867e22bee78STejun Heo 868e22bee78STejun Heo /* 869e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 870e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 871e22bee78STejun Heo * Please read comment there. 872e22bee78STejun Heo * 873628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 874628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 875628c78e7STejun Heo * disabled, which in turn means that none else could be 876d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 877628c78e7STejun Heo * lock is safe. 878e22bee78STejun Heo */ 879e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 880e19e397aSTejun Heo !list_empty(&pool->worklist)) 88163d95a91STejun Heo to_wakeup = first_worker(pool); 882e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 883e22bee78STejun Heo } 884e22bee78STejun Heo 885e22bee78STejun Heo /** 886e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 887cb444766STejun Heo * @worker: self 888d302f017STejun Heo * @flags: flags to set 889d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 890d302f017STejun Heo * 891e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 892e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 893e22bee78STejun Heo * woken up. 894d302f017STejun Heo * 895cb444766STejun Heo * CONTEXT: 896d565ed63STejun Heo * spin_lock_irq(pool->lock) 897d302f017STejun Heo */ 898d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 899d302f017STejun Heo bool wakeup) 900d302f017STejun Heo { 901bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 902e22bee78STejun Heo 903cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 904cb444766STejun Heo 905e22bee78STejun Heo /* 906e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 907e22bee78STejun Heo * wake up an idle worker as necessary if requested by 908e22bee78STejun Heo * @wakeup. 909e22bee78STejun Heo */ 910e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 911e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 912e22bee78STejun Heo if (wakeup) { 913e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 914bd7bdd43STejun Heo !list_empty(&pool->worklist)) 91563d95a91STejun Heo wake_up_worker(pool); 916e22bee78STejun Heo } else 917e19e397aSTejun Heo atomic_dec(&pool->nr_running); 918e22bee78STejun Heo } 919e22bee78STejun Heo 920d302f017STejun Heo worker->flags |= flags; 921d302f017STejun Heo } 922d302f017STejun Heo 923d302f017STejun Heo /** 924e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 925cb444766STejun Heo * @worker: self 926d302f017STejun Heo * @flags: flags to clear 927d302f017STejun Heo * 928e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 929d302f017STejun Heo * 930cb444766STejun Heo * CONTEXT: 931d565ed63STejun Heo * spin_lock_irq(pool->lock) 932d302f017STejun Heo */ 933d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 934d302f017STejun Heo { 93563d95a91STejun Heo struct worker_pool *pool = worker->pool; 936e22bee78STejun Heo unsigned int oflags = worker->flags; 937e22bee78STejun Heo 938cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 939cb444766STejun Heo 940d302f017STejun Heo worker->flags &= ~flags; 941e22bee78STejun Heo 94242c025f3STejun Heo /* 94342c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 94442c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 94542c025f3STejun Heo * of multiple flags, not a single flag. 94642c025f3STejun Heo */ 947e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 948e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 949e19e397aSTejun Heo atomic_inc(&pool->nr_running); 950d302f017STejun Heo } 951d302f017STejun Heo 952d302f017STejun Heo /** 9538cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 954c9e7cf27STejun Heo * @pool: pool of interest 9558cca0eeaSTejun Heo * @work: work to find worker for 9568cca0eeaSTejun Heo * 957c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 958c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 959a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 960a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 961a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 962a2c1c57bSTejun Heo * being executed. 963a2c1c57bSTejun Heo * 964a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 965a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 966a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 967a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 968a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 969a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 970a2c1c57bSTejun Heo * 971c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 972c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 973c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 974c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 975c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 976c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9778cca0eeaSTejun Heo * 9788cca0eeaSTejun Heo * CONTEXT: 979d565ed63STejun Heo * spin_lock_irq(pool->lock). 9808cca0eeaSTejun Heo * 981d185af30SYacine Belkadi * Return: 982d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9838cca0eeaSTejun Heo * otherwise. 9848cca0eeaSTejun Heo */ 985c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9868cca0eeaSTejun Heo struct work_struct *work) 9878cca0eeaSTejun Heo { 98842f8570fSSasha Levin struct worker *worker; 98942f8570fSSasha Levin 990b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 991a2c1c57bSTejun Heo (unsigned long)work) 992a2c1c57bSTejun Heo if (worker->current_work == work && 993a2c1c57bSTejun Heo worker->current_func == work->func) 99442f8570fSSasha Levin return worker; 99542f8570fSSasha Levin 99642f8570fSSasha Levin return NULL; 9978cca0eeaSTejun Heo } 9988cca0eeaSTejun Heo 9998cca0eeaSTejun Heo /** 1000bf4ede01STejun Heo * move_linked_works - move linked works to a list 1001bf4ede01STejun Heo * @work: start of series of works to be scheduled 1002bf4ede01STejun Heo * @head: target list to append @work to 1003bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 1004bf4ede01STejun Heo * 1005bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1006bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1007bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1008bf4ede01STejun Heo * 1009bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1010bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1011bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1012bf4ede01STejun Heo * 1013bf4ede01STejun Heo * CONTEXT: 1014d565ed63STejun Heo * spin_lock_irq(pool->lock). 1015bf4ede01STejun Heo */ 1016bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1017bf4ede01STejun Heo struct work_struct **nextp) 1018bf4ede01STejun Heo { 1019bf4ede01STejun Heo struct work_struct *n; 1020bf4ede01STejun Heo 1021bf4ede01STejun Heo /* 1022bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1023bf4ede01STejun Heo * use NULL for list head. 1024bf4ede01STejun Heo */ 1025bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1026bf4ede01STejun Heo list_move_tail(&work->entry, head); 1027bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1028bf4ede01STejun Heo break; 1029bf4ede01STejun Heo } 1030bf4ede01STejun Heo 1031bf4ede01STejun Heo /* 1032bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1033bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1034bf4ede01STejun Heo * needs to be updated. 1035bf4ede01STejun Heo */ 1036bf4ede01STejun Heo if (nextp) 1037bf4ede01STejun Heo *nextp = n; 1038bf4ede01STejun Heo } 1039bf4ede01STejun Heo 10408864b4e5STejun Heo /** 10418864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10428864b4e5STejun Heo * @pwq: pool_workqueue to get 10438864b4e5STejun Heo * 10448864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10458864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10468864b4e5STejun Heo */ 10478864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10488864b4e5STejun Heo { 10498864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10508864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10518864b4e5STejun Heo pwq->refcnt++; 10528864b4e5STejun Heo } 10538864b4e5STejun Heo 10548864b4e5STejun Heo /** 10558864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10568864b4e5STejun Heo * @pwq: pool_workqueue to put 10578864b4e5STejun Heo * 10588864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10598864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10608864b4e5STejun Heo */ 10618864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10628864b4e5STejun Heo { 10638864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10648864b4e5STejun Heo if (likely(--pwq->refcnt)) 10658864b4e5STejun Heo return; 10668864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10678864b4e5STejun Heo return; 10688864b4e5STejun Heo /* 10698864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10708864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10718864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10728864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10738864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10748864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10758864b4e5STejun Heo */ 10768864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10778864b4e5STejun Heo } 10788864b4e5STejun Heo 1079dce90d47STejun Heo /** 1080dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1081dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1082dce90d47STejun Heo * 1083dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1084dce90d47STejun Heo */ 1085dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1086dce90d47STejun Heo { 1087dce90d47STejun Heo if (pwq) { 1088dce90d47STejun Heo /* 1089dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1090dce90d47STejun Heo * following lock operations are safe. 1091dce90d47STejun Heo */ 1092dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1093dce90d47STejun Heo put_pwq(pwq); 1094dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1095dce90d47STejun Heo } 1096dce90d47STejun Heo } 1097dce90d47STejun Heo 1098112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1099bf4ede01STejun Heo { 1100112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1101bf4ede01STejun Heo 1102bf4ede01STejun Heo trace_workqueue_activate_work(work); 1103112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1104bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1105112202d9STejun Heo pwq->nr_active++; 1106bf4ede01STejun Heo } 1107bf4ede01STejun Heo 1108112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11093aa62497SLai Jiangshan { 1110112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11113aa62497SLai Jiangshan struct work_struct, entry); 11123aa62497SLai Jiangshan 1113112202d9STejun Heo pwq_activate_delayed_work(work); 11143aa62497SLai Jiangshan } 11153aa62497SLai Jiangshan 1116bf4ede01STejun Heo /** 1117112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1118112202d9STejun Heo * @pwq: pwq of interest 1119bf4ede01STejun Heo * @color: color of work which left the queue 1120bf4ede01STejun Heo * 1121bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1122112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1123bf4ede01STejun Heo * 1124bf4ede01STejun Heo * CONTEXT: 1125d565ed63STejun Heo * spin_lock_irq(pool->lock). 1126bf4ede01STejun Heo */ 1127112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1128bf4ede01STejun Heo { 11298864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1130bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11318864b4e5STejun Heo goto out_put; 1132bf4ede01STejun Heo 1133112202d9STejun Heo pwq->nr_in_flight[color]--; 1134bf4ede01STejun Heo 1135112202d9STejun Heo pwq->nr_active--; 1136112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1137bf4ede01STejun Heo /* one down, submit a delayed one */ 1138112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1139112202d9STejun Heo pwq_activate_first_delayed(pwq); 1140bf4ede01STejun Heo } 1141bf4ede01STejun Heo 1142bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1143112202d9STejun Heo if (likely(pwq->flush_color != color)) 11448864b4e5STejun Heo goto out_put; 1145bf4ede01STejun Heo 1146bf4ede01STejun Heo /* are there still in-flight works? */ 1147112202d9STejun Heo if (pwq->nr_in_flight[color]) 11488864b4e5STejun Heo goto out_put; 1149bf4ede01STejun Heo 1150112202d9STejun Heo /* this pwq is done, clear flush_color */ 1151112202d9STejun Heo pwq->flush_color = -1; 1152bf4ede01STejun Heo 1153bf4ede01STejun Heo /* 1154112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1155bf4ede01STejun Heo * will handle the rest. 1156bf4ede01STejun Heo */ 1157112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1158112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11598864b4e5STejun Heo out_put: 11608864b4e5STejun Heo put_pwq(pwq); 1161bf4ede01STejun Heo } 1162bf4ede01STejun Heo 116336e227d2STejun Heo /** 1164bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 116536e227d2STejun Heo * @work: work item to steal 116636e227d2STejun Heo * @is_dwork: @work is a delayed_work 1167bbb68dfaSTejun Heo * @flags: place to store irq state 116836e227d2STejun Heo * 116936e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1170d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 117136e227d2STejun Heo * 1172d185af30SYacine Belkadi * Return: 117336e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 117436e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 117536e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1176bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1177bbb68dfaSTejun Heo * for arbitrarily long 117836e227d2STejun Heo * 1179d185af30SYacine Belkadi * Note: 1180bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1181e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1182e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1183e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1184bbb68dfaSTejun Heo * 1185bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1186bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1187bbb68dfaSTejun Heo * 1188e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1189bf4ede01STejun Heo */ 1190bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1191bbb68dfaSTejun Heo unsigned long *flags) 1192bf4ede01STejun Heo { 1193d565ed63STejun Heo struct worker_pool *pool; 1194112202d9STejun Heo struct pool_workqueue *pwq; 1195bf4ede01STejun Heo 1196bbb68dfaSTejun Heo local_irq_save(*flags); 1197bbb68dfaSTejun Heo 119836e227d2STejun Heo /* try to steal the timer if it exists */ 119936e227d2STejun Heo if (is_dwork) { 120036e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 120136e227d2STejun Heo 1202e0aecdd8STejun Heo /* 1203e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1204e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1205e0aecdd8STejun Heo * running on the local CPU. 1206e0aecdd8STejun Heo */ 120736e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 120836e227d2STejun Heo return 1; 120936e227d2STejun Heo } 121036e227d2STejun Heo 121136e227d2STejun Heo /* try to claim PENDING the normal way */ 1212bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1213bf4ede01STejun Heo return 0; 1214bf4ede01STejun Heo 1215bf4ede01STejun Heo /* 1216bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1217bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1218bf4ede01STejun Heo */ 1219d565ed63STejun Heo pool = get_work_pool(work); 1220d565ed63STejun Heo if (!pool) 1221bbb68dfaSTejun Heo goto fail; 1222bf4ede01STejun Heo 1223d565ed63STejun Heo spin_lock(&pool->lock); 1224bf4ede01STejun Heo /* 1225112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1226112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1227112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1228112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1229112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12300b3dae68SLai Jiangshan * item is currently queued on that pool. 1231bf4ede01STejun Heo */ 1232112202d9STejun Heo pwq = get_work_pwq(work); 1233112202d9STejun Heo if (pwq && pwq->pool == pool) { 1234bf4ede01STejun Heo debug_work_deactivate(work); 12353aa62497SLai Jiangshan 12363aa62497SLai Jiangshan /* 123716062836STejun Heo * A delayed work item cannot be grabbed directly because 123816062836STejun Heo * it might have linked NO_COLOR work items which, if left 1239112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 124016062836STejun Heo * management later on and cause stall. Make sure the work 124116062836STejun Heo * item is activated before grabbing. 12423aa62497SLai Jiangshan */ 12433aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1244112202d9STejun Heo pwq_activate_delayed_work(work); 12453aa62497SLai Jiangshan 1246bf4ede01STejun Heo list_del_init(&work->entry); 1247112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 124836e227d2STejun Heo 1249112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12504468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12514468a00fSLai Jiangshan 1252d565ed63STejun Heo spin_unlock(&pool->lock); 125336e227d2STejun Heo return 1; 1254bf4ede01STejun Heo } 1255d565ed63STejun Heo spin_unlock(&pool->lock); 1256bbb68dfaSTejun Heo fail: 1257bbb68dfaSTejun Heo local_irq_restore(*flags); 1258bbb68dfaSTejun Heo if (work_is_canceling(work)) 1259bbb68dfaSTejun Heo return -ENOENT; 1260bbb68dfaSTejun Heo cpu_relax(); 126136e227d2STejun Heo return -EAGAIN; 1262bf4ede01STejun Heo } 1263bf4ede01STejun Heo 1264bf4ede01STejun Heo /** 1265706026c2STejun Heo * insert_work - insert a work into a pool 1266112202d9STejun Heo * @pwq: pwq @work belongs to 12674690c4abSTejun Heo * @work: work to insert 12684690c4abSTejun Heo * @head: insertion point 12694690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12704690c4abSTejun Heo * 1271112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1272706026c2STejun Heo * work_struct flags. 12734690c4abSTejun Heo * 12744690c4abSTejun Heo * CONTEXT: 1275d565ed63STejun Heo * spin_lock_irq(pool->lock). 1276365970a1SDavid Howells */ 1277112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1278112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1279b89deed3SOleg Nesterov { 1280112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1281e1d8aa9fSFrederic Weisbecker 12824690c4abSTejun Heo /* we own @work, set data and link */ 1283112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12841a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12858864b4e5STejun Heo get_pwq(pwq); 1286e22bee78STejun Heo 1287e22bee78STejun Heo /* 1288c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1289c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1290c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1291e22bee78STejun Heo */ 1292e22bee78STejun Heo smp_mb(); 1293e22bee78STejun Heo 129463d95a91STejun Heo if (__need_more_worker(pool)) 129563d95a91STejun Heo wake_up_worker(pool); 1296b89deed3SOleg Nesterov } 1297b89deed3SOleg Nesterov 1298c8efcc25STejun Heo /* 1299c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13008d03ecfeSTejun Heo * same workqueue. 1301c8efcc25STejun Heo */ 1302c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1303c8efcc25STejun Heo { 1304c8efcc25STejun Heo struct worker *worker; 1305c8efcc25STejun Heo 13068d03ecfeSTejun Heo worker = current_wq_worker(); 1307c8efcc25STejun Heo /* 13088d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13098d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1310c8efcc25STejun Heo */ 1311112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1312c8efcc25STejun Heo } 1313c8efcc25STejun Heo 1314d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13151da177e4SLinus Torvalds struct work_struct *work) 13161da177e4SLinus Torvalds { 1317112202d9STejun Heo struct pool_workqueue *pwq; 1318c9178087STejun Heo struct worker_pool *last_pool; 13191e19ffc6STejun Heo struct list_head *worklist; 13208a2e8e5dSTejun Heo unsigned int work_flags; 1321b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13228930cabaSTejun Heo 13238930cabaSTejun Heo /* 13248930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13258930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13268930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13278930cabaSTejun Heo * happen with IRQ disabled. 13288930cabaSTejun Heo */ 13298930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13301da177e4SLinus Torvalds 1331dc186ad7SThomas Gleixner debug_work_activate(work); 13321e19ffc6STejun Heo 13339ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1334618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1335c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1336e41e704bSTejun Heo return; 13379e8cd2f5STejun Heo retry: 1338df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1339f3421797STejun Heo cpu = raw_smp_processor_id(); 1340df2d5ae4STejun Heo 1341df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1342df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1343c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1344df2d5ae4STejun Heo else 1345df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1346f3421797STejun Heo 134718aa9effSTejun Heo /* 1348c9178087STejun Heo * If @work was previously on a different pool, it might still be 1349c9178087STejun Heo * running there, in which case the work needs to be queued on that 1350c9178087STejun Heo * pool to guarantee non-reentrancy. 135118aa9effSTejun Heo */ 1352c9e7cf27STejun Heo last_pool = get_work_pool(work); 1353112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 135418aa9effSTejun Heo struct worker *worker; 135518aa9effSTejun Heo 1356d565ed63STejun Heo spin_lock(&last_pool->lock); 135718aa9effSTejun Heo 1358c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 135918aa9effSTejun Heo 1360112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1361c9178087STejun Heo pwq = worker->current_pwq; 13628594fadeSLai Jiangshan } else { 136318aa9effSTejun Heo /* meh... not running there, queue here */ 1364d565ed63STejun Heo spin_unlock(&last_pool->lock); 1365112202d9STejun Heo spin_lock(&pwq->pool->lock); 136618aa9effSTejun Heo } 13678930cabaSTejun Heo } else { 1368112202d9STejun Heo spin_lock(&pwq->pool->lock); 13698930cabaSTejun Heo } 1370502ca9d8STejun Heo 13719e8cd2f5STejun Heo /* 13729e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13739e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13749e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1375df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1376df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13779e8cd2f5STejun Heo * make forward-progress. 13789e8cd2f5STejun Heo */ 13799e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13809e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13819e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13829e8cd2f5STejun Heo cpu_relax(); 13839e8cd2f5STejun Heo goto retry; 13849e8cd2f5STejun Heo } 13859e8cd2f5STejun Heo /* oops */ 13869e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13879e8cd2f5STejun Heo wq->name, cpu); 13889e8cd2f5STejun Heo } 13899e8cd2f5STejun Heo 1390112202d9STejun Heo /* pwq determined, queue */ 1391112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1392502ca9d8STejun Heo 1393f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1394112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1395f5b2552bSDan Carpenter return; 1396f5b2552bSDan Carpenter } 13971e19ffc6STejun Heo 1398112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1399112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14001e19ffc6STejun Heo 1401112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1402cdadf009STejun Heo trace_workqueue_activate_work(work); 1403112202d9STejun Heo pwq->nr_active++; 1404112202d9STejun Heo worklist = &pwq->pool->worklist; 14058a2e8e5dSTejun Heo } else { 14068a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1407112202d9STejun Heo worklist = &pwq->delayed_works; 14088a2e8e5dSTejun Heo } 14091e19ffc6STejun Heo 1410112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14111e19ffc6STejun Heo 1412112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14131da177e4SLinus Torvalds } 14141da177e4SLinus Torvalds 14150fcb78c2SRolf Eike Beer /** 1416c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1417c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1418c1a220e7SZhang Rui * @wq: workqueue to use 1419c1a220e7SZhang Rui * @work: work to queue 1420c1a220e7SZhang Rui * 1421c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1422c1a220e7SZhang Rui * can't go away. 1423d185af30SYacine Belkadi * 1424d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1425c1a220e7SZhang Rui */ 1426d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1427d4283e93STejun Heo struct work_struct *work) 1428c1a220e7SZhang Rui { 1429d4283e93STejun Heo bool ret = false; 14308930cabaSTejun Heo unsigned long flags; 14318930cabaSTejun Heo 14328930cabaSTejun Heo local_irq_save(flags); 1433c1a220e7SZhang Rui 143422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14354690c4abSTejun Heo __queue_work(cpu, wq, work); 1436d4283e93STejun Heo ret = true; 1437c1a220e7SZhang Rui } 14388930cabaSTejun Heo 14398930cabaSTejun Heo local_irq_restore(flags); 1440c1a220e7SZhang Rui return ret; 1441c1a220e7SZhang Rui } 1442ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1443c1a220e7SZhang Rui 1444d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14451da177e4SLinus Torvalds { 144652bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14471da177e4SLinus Torvalds 1448e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 144960c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14501da177e4SLinus Torvalds } 14511438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14521da177e4SLinus Torvalds 14537beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 145452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14551da177e4SLinus Torvalds { 14567beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14577beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14581da177e4SLinus Torvalds 14597beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14607beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1461fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1462fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14637beb2edfSTejun Heo 14648852aac2STejun Heo /* 14658852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14668852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14678852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14688852aac2STejun Heo * on that there's no such delay when @delay is 0. 14698852aac2STejun Heo */ 14708852aac2STejun Heo if (!delay) { 14718852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14728852aac2STejun Heo return; 14738852aac2STejun Heo } 14748852aac2STejun Heo 14757beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14767beb2edfSTejun Heo 147760c057bcSLai Jiangshan dwork->wq = wq; 14781265057fSTejun Heo dwork->cpu = cpu; 14797beb2edfSTejun Heo timer->expires = jiffies + delay; 14807beb2edfSTejun Heo 14817beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14827beb2edfSTejun Heo add_timer_on(timer, cpu); 14837beb2edfSTejun Heo else 14847beb2edfSTejun Heo add_timer(timer); 14857beb2edfSTejun Heo } 14861da177e4SLinus Torvalds 14870fcb78c2SRolf Eike Beer /** 14880fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14890fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14900fcb78c2SRolf Eike Beer * @wq: workqueue to use 1491af9997e4SRandy Dunlap * @dwork: work to queue 14920fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14930fcb78c2SRolf Eike Beer * 1494d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1495715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1496715f1300STejun Heo * execution. 14970fcb78c2SRolf Eike Beer */ 1498d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 149952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15007a6bc1cdSVenkatesh Pallipadi { 150152bad64dSDavid Howells struct work_struct *work = &dwork->work; 1502d4283e93STejun Heo bool ret = false; 15038930cabaSTejun Heo unsigned long flags; 15048930cabaSTejun Heo 15058930cabaSTejun Heo /* read the comment in __queue_work() */ 15068930cabaSTejun Heo local_irq_save(flags); 15077a6bc1cdSVenkatesh Pallipadi 150822df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15097beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1510d4283e93STejun Heo ret = true; 15117a6bc1cdSVenkatesh Pallipadi } 15128930cabaSTejun Heo 15138930cabaSTejun Heo local_irq_restore(flags); 15147a6bc1cdSVenkatesh Pallipadi return ret; 15157a6bc1cdSVenkatesh Pallipadi } 1516ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15171da177e4SLinus Torvalds 1518c8e55f36STejun Heo /** 15198376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15208376fe22STejun Heo * @cpu: CPU number to execute work on 15218376fe22STejun Heo * @wq: workqueue to use 15228376fe22STejun Heo * @dwork: work to queue 15238376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15248376fe22STejun Heo * 15258376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15268376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15278376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15288376fe22STejun Heo * current state. 15298376fe22STejun Heo * 1530d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15318376fe22STejun Heo * pending and its timer was modified. 15328376fe22STejun Heo * 1533e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15348376fe22STejun Heo * See try_to_grab_pending() for details. 15358376fe22STejun Heo */ 15368376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15378376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15388376fe22STejun Heo { 15398376fe22STejun Heo unsigned long flags; 15408376fe22STejun Heo int ret; 15418376fe22STejun Heo 15428376fe22STejun Heo do { 15438376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15448376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15458376fe22STejun Heo 15468376fe22STejun Heo if (likely(ret >= 0)) { 15478376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15488376fe22STejun Heo local_irq_restore(flags); 15498376fe22STejun Heo } 15508376fe22STejun Heo 15518376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15528376fe22STejun Heo return ret; 15538376fe22STejun Heo } 15548376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15558376fe22STejun Heo 15568376fe22STejun Heo /** 1557c8e55f36STejun Heo * worker_enter_idle - enter idle state 1558c8e55f36STejun Heo * @worker: worker which is entering idle state 1559c8e55f36STejun Heo * 1560c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1561c8e55f36STejun Heo * necessary. 1562c8e55f36STejun Heo * 1563c8e55f36STejun Heo * LOCKING: 1564d565ed63STejun Heo * spin_lock_irq(pool->lock). 1565c8e55f36STejun Heo */ 1566c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15671da177e4SLinus Torvalds { 1568bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1569c8e55f36STejun Heo 15706183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15716183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15726183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15736183c009STejun Heo return; 1574c8e55f36STejun Heo 1575cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1576cb444766STejun Heo worker->flags |= WORKER_IDLE; 1577bd7bdd43STejun Heo pool->nr_idle++; 1578e22bee78STejun Heo worker->last_active = jiffies; 1579c8e55f36STejun Heo 1580c8e55f36STejun Heo /* idle_list is LIFO */ 1581bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1582db7bccf4STejun Heo 158363d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1584628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1585cb444766STejun Heo 1586544ecf31STejun Heo /* 1587706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1588d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1589628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1590628c78e7STejun Heo * unbind is not in progress. 1591544ecf31STejun Heo */ 159224647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1593bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1594e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1595c8e55f36STejun Heo } 1596c8e55f36STejun Heo 1597c8e55f36STejun Heo /** 1598c8e55f36STejun Heo * worker_leave_idle - leave idle state 1599c8e55f36STejun Heo * @worker: worker which is leaving idle state 1600c8e55f36STejun Heo * 1601c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1602c8e55f36STejun Heo * 1603c8e55f36STejun Heo * LOCKING: 1604d565ed63STejun Heo * spin_lock_irq(pool->lock). 1605c8e55f36STejun Heo */ 1606c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1607c8e55f36STejun Heo { 1608bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1609c8e55f36STejun Heo 16106183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16116183c009STejun Heo return; 1612d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1613bd7bdd43STejun Heo pool->nr_idle--; 1614c8e55f36STejun Heo list_del_init(&worker->entry); 1615c8e55f36STejun Heo } 1616c8e55f36STejun Heo 1617e22bee78STejun Heo /** 1618f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1619f36dc67bSLai Jiangshan * @pool: target worker_pool 1620f36dc67bSLai Jiangshan * 1621f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1622e22bee78STejun Heo * 1623e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1624e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1625e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1626e22bee78STejun Heo * guaranteed to execute on the cpu. 1627e22bee78STejun Heo * 1628f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1629e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1630e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1631e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1632706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1633e22bee78STejun Heo * [dis]associated in the meantime. 1634e22bee78STejun Heo * 1635706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 163624647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1637f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1638f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1639f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1640e22bee78STejun Heo * 1641e22bee78STejun Heo * CONTEXT: 1642d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1643e22bee78STejun Heo * held. 1644e22bee78STejun Heo * 1645d185af30SYacine Belkadi * Return: 1646706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1647e22bee78STejun Heo * bound), %false if offline. 1648e22bee78STejun Heo */ 1649f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1650d565ed63STejun Heo __acquires(&pool->lock) 1651e22bee78STejun Heo { 1652e22bee78STejun Heo while (true) { 1653e22bee78STejun Heo /* 1654e22bee78STejun Heo * The following call may fail, succeed or succeed 1655e22bee78STejun Heo * without actually migrating the task to the cpu if 1656e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 165724647570STejun Heo * against POOL_DISASSOCIATED. 1658e22bee78STejun Heo */ 165924647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16607a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1661e22bee78STejun Heo 1662d565ed63STejun Heo spin_lock_irq(&pool->lock); 166324647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1664e22bee78STejun Heo return false; 1665f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16667a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1667e22bee78STejun Heo return true; 1668d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1669e22bee78STejun Heo 16705035b20fSTejun Heo /* 16715035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16725035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16735035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16745035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16755035b20fSTejun Heo */ 1676e22bee78STejun Heo cpu_relax(); 16775035b20fSTejun Heo cond_resched(); 1678e22bee78STejun Heo } 1679e22bee78STejun Heo } 1680e22bee78STejun Heo 1681c34056a3STejun Heo static struct worker *alloc_worker(void) 1682c34056a3STejun Heo { 1683c34056a3STejun Heo struct worker *worker; 1684c34056a3STejun Heo 1685c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1686c8e55f36STejun Heo if (worker) { 1687c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1688affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1689e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1690e22bee78STejun Heo worker->flags = WORKER_PREP; 1691c8e55f36STejun Heo } 1692c34056a3STejun Heo return worker; 1693c34056a3STejun Heo } 1694c34056a3STejun Heo 1695c34056a3STejun Heo /** 1696c34056a3STejun Heo * create_worker - create a new workqueue worker 169763d95a91STejun Heo * @pool: pool the new worker will belong to 1698c34056a3STejun Heo * 169963d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1700c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1701c34056a3STejun Heo * destroy_worker(). 1702c34056a3STejun Heo * 1703c34056a3STejun Heo * CONTEXT: 1704c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1705c34056a3STejun Heo * 1706d185af30SYacine Belkadi * Return: 1707c34056a3STejun Heo * Pointer to the newly created worker. 1708c34056a3STejun Heo */ 1709bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1710c34056a3STejun Heo { 1711c34056a3STejun Heo struct worker *worker = NULL; 1712f3421797STejun Heo int id = -1; 1713e3c916a4STejun Heo char id_buf[16]; 1714c34056a3STejun Heo 1715cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1716cd549687STejun Heo 1717822d8405STejun Heo /* 1718822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1719822d8405STejun Heo * without installing the pointer. 1720822d8405STejun Heo */ 1721822d8405STejun Heo idr_preload(GFP_KERNEL); 1722d565ed63STejun Heo spin_lock_irq(&pool->lock); 1723822d8405STejun Heo 1724822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1725822d8405STejun Heo 1726d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1727822d8405STejun Heo idr_preload_end(); 1728822d8405STejun Heo if (id < 0) 1729c34056a3STejun Heo goto fail; 1730c34056a3STejun Heo 1731c34056a3STejun Heo worker = alloc_worker(); 1732c34056a3STejun Heo if (!worker) 1733c34056a3STejun Heo goto fail; 1734c34056a3STejun Heo 1735bd7bdd43STejun Heo worker->pool = pool; 1736c34056a3STejun Heo worker->id = id; 1737c34056a3STejun Heo 173829c91e99STejun Heo if (pool->cpu >= 0) 1739e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1740e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1741f3421797STejun Heo else 1742e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1743e3c916a4STejun Heo 1744f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1745e3c916a4STejun Heo "kworker/%s", id_buf); 1746c34056a3STejun Heo if (IS_ERR(worker->task)) 1747c34056a3STejun Heo goto fail; 1748c34056a3STejun Heo 174991151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 175091151228SOleg Nesterov 175191151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 175291151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 175391151228SOleg Nesterov 1754c5aa87bbSTejun Heo /* 1755c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1756c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1757c5aa87bbSTejun Heo */ 17587a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17593270476aSTejun Heo 17607a4e344cSTejun Heo /* 17617a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17627a4e344cSTejun Heo * remains stable across this function. See the comments above the 17637a4e344cSTejun Heo * flag definition for details. 17647a4e344cSTejun Heo */ 17657a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1766f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1767c34056a3STejun Heo 1768822d8405STejun Heo /* successful, commit the pointer to idr */ 1769822d8405STejun Heo spin_lock_irq(&pool->lock); 1770822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1771822d8405STejun Heo spin_unlock_irq(&pool->lock); 1772822d8405STejun Heo 1773c34056a3STejun Heo return worker; 1774822d8405STejun Heo 1775c34056a3STejun Heo fail: 1776c34056a3STejun Heo if (id >= 0) { 1777d565ed63STejun Heo spin_lock_irq(&pool->lock); 1778822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1779d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1780c34056a3STejun Heo } 1781c34056a3STejun Heo kfree(worker); 1782c34056a3STejun Heo return NULL; 1783c34056a3STejun Heo } 1784c34056a3STejun Heo 1785c34056a3STejun Heo /** 1786c34056a3STejun Heo * start_worker - start a newly created worker 1787c34056a3STejun Heo * @worker: worker to start 1788c34056a3STejun Heo * 1789706026c2STejun Heo * Make the pool aware of @worker and start it. 1790c34056a3STejun Heo * 1791c34056a3STejun Heo * CONTEXT: 1792d565ed63STejun Heo * spin_lock_irq(pool->lock). 1793c34056a3STejun Heo */ 1794c34056a3STejun Heo static void start_worker(struct worker *worker) 1795c34056a3STejun Heo { 1796cb444766STejun Heo worker->flags |= WORKER_STARTED; 1797bd7bdd43STejun Heo worker->pool->nr_workers++; 1798c8e55f36STejun Heo worker_enter_idle(worker); 1799c34056a3STejun Heo wake_up_process(worker->task); 1800c34056a3STejun Heo } 1801c34056a3STejun Heo 1802c34056a3STejun Heo /** 1803ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1804ebf44d16STejun Heo * @pool: the target pool 1805ebf44d16STejun Heo * 1806cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1807d185af30SYacine Belkadi * 1808d185af30SYacine Belkadi * Return: 0 on success. A negative error code otherwise. 1809ebf44d16STejun Heo */ 1810ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1811ebf44d16STejun Heo { 1812ebf44d16STejun Heo struct worker *worker; 1813ebf44d16STejun Heo 1814cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1815cd549687STejun Heo 1816ebf44d16STejun Heo worker = create_worker(pool); 1817ebf44d16STejun Heo if (worker) { 1818ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1819ebf44d16STejun Heo start_worker(worker); 1820ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1821ebf44d16STejun Heo } 1822ebf44d16STejun Heo 1823cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1824cd549687STejun Heo 1825ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1826ebf44d16STejun Heo } 1827ebf44d16STejun Heo 1828ebf44d16STejun Heo /** 1829c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1830c34056a3STejun Heo * @worker: worker to be destroyed 1831c34056a3STejun Heo * 1832706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1833c8e55f36STejun Heo * 1834c8e55f36STejun Heo * CONTEXT: 1835d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1836c34056a3STejun Heo */ 1837c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1838c34056a3STejun Heo { 1839bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1840c34056a3STejun Heo 1841cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1842cd549687STejun Heo lockdep_assert_held(&pool->lock); 1843cd549687STejun Heo 1844c34056a3STejun Heo /* sanity check frenzy */ 18456183c009STejun Heo if (WARN_ON(worker->current_work) || 18466183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18476183c009STejun Heo return; 1848c34056a3STejun Heo 1849c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1850bd7bdd43STejun Heo pool->nr_workers--; 1851c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1852bd7bdd43STejun Heo pool->nr_idle--; 1853c8e55f36STejun Heo 1854c8e55f36STejun Heo list_del_init(&worker->entry); 1855cb444766STejun Heo worker->flags |= WORKER_DIE; 1856c8e55f36STejun Heo 1857822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1858822d8405STejun Heo 1859d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1860c8e55f36STejun Heo 1861c34056a3STejun Heo kthread_stop(worker->task); 1862c34056a3STejun Heo kfree(worker); 1863c34056a3STejun Heo 1864d565ed63STejun Heo spin_lock_irq(&pool->lock); 1865c34056a3STejun Heo } 1866c34056a3STejun Heo 186763d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1868e22bee78STejun Heo { 186963d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1870e22bee78STejun Heo 1871d565ed63STejun Heo spin_lock_irq(&pool->lock); 1872e22bee78STejun Heo 187363d95a91STejun Heo if (too_many_workers(pool)) { 1874e22bee78STejun Heo struct worker *worker; 1875e22bee78STejun Heo unsigned long expires; 1876e22bee78STejun Heo 1877e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 187863d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1879e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1880e22bee78STejun Heo 1881e22bee78STejun Heo if (time_before(jiffies, expires)) 188263d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1883e22bee78STejun Heo else { 1884e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 188511ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 188663d95a91STejun Heo wake_up_worker(pool); 1887e22bee78STejun Heo } 1888e22bee78STejun Heo } 1889e22bee78STejun Heo 1890d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1891e22bee78STejun Heo } 1892e22bee78STejun Heo 1893493a1724STejun Heo static void send_mayday(struct work_struct *work) 1894e22bee78STejun Heo { 1895112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1896112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1897493a1724STejun Heo 18982e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1899e22bee78STejun Heo 1900493008a8STejun Heo if (!wq->rescuer) 1901493a1724STejun Heo return; 1902e22bee78STejun Heo 1903e22bee78STejun Heo /* mayday mayday mayday */ 1904493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1905493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1906e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1907493a1724STejun Heo } 1908e22bee78STejun Heo } 1909e22bee78STejun Heo 1910706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1911e22bee78STejun Heo { 191263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1913e22bee78STejun Heo struct work_struct *work; 1914e22bee78STejun Heo 19152e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1916493a1724STejun Heo spin_lock(&pool->lock); 1917e22bee78STejun Heo 191863d95a91STejun Heo if (need_to_create_worker(pool)) { 1919e22bee78STejun Heo /* 1920e22bee78STejun Heo * We've been trying to create a new worker but 1921e22bee78STejun Heo * haven't been successful. We might be hitting an 1922e22bee78STejun Heo * allocation deadlock. Send distress signals to 1923e22bee78STejun Heo * rescuers. 1924e22bee78STejun Heo */ 192563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1926e22bee78STejun Heo send_mayday(work); 1927e22bee78STejun Heo } 1928e22bee78STejun Heo 1929493a1724STejun Heo spin_unlock(&pool->lock); 19302e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1931e22bee78STejun Heo 193263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1933e22bee78STejun Heo } 1934e22bee78STejun Heo 1935e22bee78STejun Heo /** 1936e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 193763d95a91STejun Heo * @pool: pool to create a new worker for 1938e22bee78STejun Heo * 193963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1940e22bee78STejun Heo * have at least one idle worker on return from this function. If 1941e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 194263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1943e22bee78STejun Heo * possible allocation deadlock. 1944e22bee78STejun Heo * 1945c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1946c5aa87bbSTejun Heo * may_start_working() %true. 1947e22bee78STejun Heo * 1948e22bee78STejun Heo * LOCKING: 1949d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1950e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1951e22bee78STejun Heo * manager. 1952e22bee78STejun Heo * 1953d185af30SYacine Belkadi * Return: 1954c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1955e22bee78STejun Heo * otherwise. 1956e22bee78STejun Heo */ 195763d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1958d565ed63STejun Heo __releases(&pool->lock) 1959d565ed63STejun Heo __acquires(&pool->lock) 1960e22bee78STejun Heo { 196163d95a91STejun Heo if (!need_to_create_worker(pool)) 1962e22bee78STejun Heo return false; 1963e22bee78STejun Heo restart: 1964d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19659f9c2364STejun Heo 1966e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 196763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1968e22bee78STejun Heo 1969e22bee78STejun Heo while (true) { 1970e22bee78STejun Heo struct worker *worker; 1971e22bee78STejun Heo 1972bc2ae0f5STejun Heo worker = create_worker(pool); 1973e22bee78STejun Heo if (worker) { 197463d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1975d565ed63STejun Heo spin_lock_irq(&pool->lock); 1976e22bee78STejun Heo start_worker(worker); 19776183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19786183c009STejun Heo goto restart; 1979e22bee78STejun Heo return true; 1980e22bee78STejun Heo } 1981e22bee78STejun Heo 198263d95a91STejun Heo if (!need_to_create_worker(pool)) 1983e22bee78STejun Heo break; 1984e22bee78STejun Heo 1985e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1986e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19879f9c2364STejun Heo 198863d95a91STejun Heo if (!need_to_create_worker(pool)) 1989e22bee78STejun Heo break; 1990e22bee78STejun Heo } 1991e22bee78STejun Heo 199263d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1993d565ed63STejun Heo spin_lock_irq(&pool->lock); 199463d95a91STejun Heo if (need_to_create_worker(pool)) 1995e22bee78STejun Heo goto restart; 1996e22bee78STejun Heo return true; 1997e22bee78STejun Heo } 1998e22bee78STejun Heo 1999e22bee78STejun Heo /** 2000e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 200163d95a91STejun Heo * @pool: pool to destroy workers for 2002e22bee78STejun Heo * 200363d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 2004e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 2005e22bee78STejun Heo * 2006e22bee78STejun Heo * LOCKING: 2007d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2008e22bee78STejun Heo * multiple times. Called only from manager. 2009e22bee78STejun Heo * 2010d185af30SYacine Belkadi * Return: 2011c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 2012e22bee78STejun Heo * otherwise. 2013e22bee78STejun Heo */ 201463d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 2015e22bee78STejun Heo { 2016e22bee78STejun Heo bool ret = false; 2017e22bee78STejun Heo 201863d95a91STejun Heo while (too_many_workers(pool)) { 2019e22bee78STejun Heo struct worker *worker; 2020e22bee78STejun Heo unsigned long expires; 2021e22bee78STejun Heo 202263d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2023e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2024e22bee78STejun Heo 2025e22bee78STejun Heo if (time_before(jiffies, expires)) { 202663d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2027e22bee78STejun Heo break; 2028e22bee78STejun Heo } 2029e22bee78STejun Heo 2030e22bee78STejun Heo destroy_worker(worker); 2031e22bee78STejun Heo ret = true; 2032e22bee78STejun Heo } 2033e22bee78STejun Heo 2034e22bee78STejun Heo return ret; 2035e22bee78STejun Heo } 2036e22bee78STejun Heo 2037e22bee78STejun Heo /** 2038e22bee78STejun Heo * manage_workers - manage worker pool 2039e22bee78STejun Heo * @worker: self 2040e22bee78STejun Heo * 2041706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2042e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2043706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2044e22bee78STejun Heo * 2045e22bee78STejun Heo * The caller can safely start processing works on false return. On 2046e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2047e22bee78STejun Heo * and may_start_working() is true. 2048e22bee78STejun Heo * 2049e22bee78STejun Heo * CONTEXT: 2050d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2051e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2052e22bee78STejun Heo * 2053d185af30SYacine Belkadi * Return: 20542d498db9SLibin * %false if the pool don't need management and the caller can safely start 20552d498db9SLibin * processing works, %true indicates that the function released pool->lock 20562d498db9SLibin * and reacquired it to perform some management function and that the 20572d498db9SLibin * conditions that the caller verified while holding the lock before 20582d498db9SLibin * calling the function might no longer be true. 2059e22bee78STejun Heo */ 2060e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2061e22bee78STejun Heo { 206263d95a91STejun Heo struct worker_pool *pool = worker->pool; 2063e22bee78STejun Heo bool ret = false; 2064e22bee78STejun Heo 2065bc3a1afcSTejun Heo /* 2066bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2067bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2068bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2069bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2070bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2071bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2072bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2073bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2074bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2075bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2076bc3a1afcSTejun Heo * 2077bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2078bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2079bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2080bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2081bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2082bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2083bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2084bc3a1afcSTejun Heo * manager_mutex. 2085bc3a1afcSTejun Heo */ 208634a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2087e22bee78STejun Heo return ret; 2088e22bee78STejun Heo 2089ee378aa4SLai Jiangshan /* 2090bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2091bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2092ee378aa4SLai Jiangshan */ 2093bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2094d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2095bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 20968f174b11SJoonsoo Kim spin_lock_irq(&pool->lock); 2097ee378aa4SLai Jiangshan ret = true; 2098ee378aa4SLai Jiangshan } 2099ee378aa4SLai Jiangshan 210011ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2101e22bee78STejun Heo 2102e22bee78STejun Heo /* 2103e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2104e22bee78STejun Heo * on return. 2105e22bee78STejun Heo */ 210663d95a91STejun Heo ret |= maybe_destroy_workers(pool); 210763d95a91STejun Heo ret |= maybe_create_worker(pool); 2108e22bee78STejun Heo 2109bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 211034a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2111e22bee78STejun Heo return ret; 2112e22bee78STejun Heo } 2113e22bee78STejun Heo 2114a62428c0STejun Heo /** 2115a62428c0STejun Heo * process_one_work - process single work 2116c34056a3STejun Heo * @worker: self 2117a62428c0STejun Heo * @work: work to process 2118a62428c0STejun Heo * 2119a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2120a62428c0STejun Heo * process a single work including synchronization against and 2121a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2122a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2123a62428c0STejun Heo * call this function to process a work. 2124a62428c0STejun Heo * 2125a62428c0STejun Heo * CONTEXT: 2126d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2127a62428c0STejun Heo */ 2128c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2129d565ed63STejun Heo __releases(&pool->lock) 2130d565ed63STejun Heo __acquires(&pool->lock) 21311da177e4SLinus Torvalds { 2132112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2133bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2134112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 213573f53c4aSTejun Heo int work_color; 21367e11629dSTejun Heo struct worker *collision; 21374e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21384e6045f1SJohannes Berg /* 2139a62428c0STejun Heo * It is permissible to free the struct work_struct from 2140a62428c0STejun Heo * inside the function that is called from it, this we need to 2141a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2142a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2143a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21444e6045f1SJohannes Berg */ 21454d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21464d82a1deSPeter Zijlstra 21474d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21484e6045f1SJohannes Berg #endif 21496fec10a1STejun Heo /* 21506fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21516fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 215224647570STejun Heo * unbound or a disassociated pool. 21536fec10a1STejun Heo */ 21545f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 215524647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2156ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 215725511a47STejun Heo 21587e11629dSTejun Heo /* 21597e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21607e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21617e11629dSTejun Heo * already processing the work. If so, defer the work to the 21627e11629dSTejun Heo * currently executing one. 21637e11629dSTejun Heo */ 2164c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21657e11629dSTejun Heo if (unlikely(collision)) { 21667e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21677e11629dSTejun Heo return; 21687e11629dSTejun Heo } 21691da177e4SLinus Torvalds 21708930cabaSTejun Heo /* claim and dequeue */ 21711da177e4SLinus Torvalds debug_work_deactivate(work); 2172c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2173c34056a3STejun Heo worker->current_work = work; 2174a2c1c57bSTejun Heo worker->current_func = work->func; 2175112202d9STejun Heo worker->current_pwq = pwq; 217673f53c4aSTejun Heo work_color = get_work_color(work); 21777a22ad75STejun Heo 2178a62428c0STejun Heo list_del_init(&work->entry); 2179a62428c0STejun Heo 2180649027d7STejun Heo /* 2181fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2182fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2183fb0e7bebSTejun Heo */ 2184fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2185fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2186fb0e7bebSTejun Heo 2187974271c4STejun Heo /* 2188d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2189974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2190974271c4STejun Heo */ 219163d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 219263d95a91STejun Heo wake_up_worker(pool); 2193974271c4STejun Heo 21948930cabaSTejun Heo /* 21957c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2196d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 219723657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 219823657bb1STejun Heo * disabled. 21998930cabaSTejun Heo */ 22007c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22011da177e4SLinus Torvalds 2202d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2203365970a1SDavid Howells 2204112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 22053295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2206e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2207a2c1c57bSTejun Heo worker->current_func(work); 2208e36c886aSArjan van de Ven /* 2209e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2210e36c886aSArjan van de Ven * point will only record its address. 2211e36c886aSArjan van de Ven */ 2212e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22133295f0efSIngo Molnar lock_map_release(&lockdep_map); 2214112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22151da177e4SLinus Torvalds 2216d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2217044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2218044c782cSValentin Ilie " last function: %pf\n", 2219a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2220a2c1c57bSTejun Heo worker->current_func); 2221d5abe669SPeter Zijlstra debug_show_held_locks(current); 2222d5abe669SPeter Zijlstra dump_stack(); 2223d5abe669SPeter Zijlstra } 2224d5abe669SPeter Zijlstra 2225b22ce278STejun Heo /* 2226b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2227b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2228b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2229b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2230b22ce278STejun Heo * stop_machine. 2231b22ce278STejun Heo */ 2232b22ce278STejun Heo cond_resched(); 2233b22ce278STejun Heo 2234d565ed63STejun Heo spin_lock_irq(&pool->lock); 2235a62428c0STejun Heo 2236fb0e7bebSTejun Heo /* clear cpu intensive status */ 2237fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2238fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2239fb0e7bebSTejun Heo 2240a62428c0STejun Heo /* we're done with it, release */ 224142f8570fSSasha Levin hash_del(&worker->hentry); 2242c34056a3STejun Heo worker->current_work = NULL; 2243a2c1c57bSTejun Heo worker->current_func = NULL; 2244112202d9STejun Heo worker->current_pwq = NULL; 22453d1cb205STejun Heo worker->desc_valid = false; 2246112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 22471da177e4SLinus Torvalds } 22481da177e4SLinus Torvalds 2249affee4b2STejun Heo /** 2250affee4b2STejun Heo * process_scheduled_works - process scheduled works 2251affee4b2STejun Heo * @worker: self 2252affee4b2STejun Heo * 2253affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2254affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2255affee4b2STejun Heo * fetches a work from the top and executes it. 2256affee4b2STejun Heo * 2257affee4b2STejun Heo * CONTEXT: 2258d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2259affee4b2STejun Heo * multiple times. 2260affee4b2STejun Heo */ 2261affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22621da177e4SLinus Torvalds { 2263affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2264affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2265a62428c0STejun Heo struct work_struct, entry); 2266c34056a3STejun Heo process_one_work(worker, work); 2267a62428c0STejun Heo } 22681da177e4SLinus Torvalds } 22691da177e4SLinus Torvalds 22704690c4abSTejun Heo /** 22714690c4abSTejun Heo * worker_thread - the worker thread function 2272c34056a3STejun Heo * @__worker: self 22734690c4abSTejun Heo * 2274c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2275c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2276c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2277c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2278c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2279d185af30SYacine Belkadi * 2280d185af30SYacine Belkadi * Return: 0 22814690c4abSTejun Heo */ 2282c34056a3STejun Heo static int worker_thread(void *__worker) 22831da177e4SLinus Torvalds { 2284c34056a3STejun Heo struct worker *worker = __worker; 2285bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22861da177e4SLinus Torvalds 2287e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2288e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2289c8e55f36STejun Heo woke_up: 2290d565ed63STejun Heo spin_lock_irq(&pool->lock); 2291affee4b2STejun Heo 2292a9ab775bSTejun Heo /* am I supposed to die? */ 2293a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2294d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2295a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2296e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2297c8e55f36STejun Heo return 0; 2298c8e55f36STejun Heo } 2299c8e55f36STejun Heo 2300c8e55f36STejun Heo worker_leave_idle(worker); 2301db7bccf4STejun Heo recheck: 2302e22bee78STejun Heo /* no more worker necessary? */ 230363d95a91STejun Heo if (!need_more_worker(pool)) 2304e22bee78STejun Heo goto sleep; 2305e22bee78STejun Heo 2306e22bee78STejun Heo /* do we need to manage? */ 230763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2308e22bee78STejun Heo goto recheck; 2309e22bee78STejun Heo 2310c8e55f36STejun Heo /* 2311c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2312c8e55f36STejun Heo * preparing to process a work or actually processing it. 2313c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2314c8e55f36STejun Heo */ 23156183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2316c8e55f36STejun Heo 2317e22bee78STejun Heo /* 2318a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2319a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2320a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2321a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2322a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2323e22bee78STejun Heo */ 2324a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2325e22bee78STejun Heo 2326e22bee78STejun Heo do { 2327affee4b2STejun Heo struct work_struct *work = 2328bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2329affee4b2STejun Heo struct work_struct, entry); 2330affee4b2STejun Heo 2331c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2332affee4b2STejun Heo /* optimization path, not strictly necessary */ 2333affee4b2STejun Heo process_one_work(worker, work); 2334affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2335affee4b2STejun Heo process_scheduled_works(worker); 2336affee4b2STejun Heo } else { 2337c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2338affee4b2STejun Heo process_scheduled_works(worker); 2339affee4b2STejun Heo } 234063d95a91STejun Heo } while (keep_working(pool)); 2341affee4b2STejun Heo 2342e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2343d313dd85STejun Heo sleep: 234463d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2345e22bee78STejun Heo goto recheck; 2346d313dd85STejun Heo 2347c8e55f36STejun Heo /* 2348d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2349d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2350d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2351d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2352d565ed63STejun Heo * event. 2353c8e55f36STejun Heo */ 2354c8e55f36STejun Heo worker_enter_idle(worker); 2355c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2356d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23571da177e4SLinus Torvalds schedule(); 2358c8e55f36STejun Heo goto woke_up; 23591da177e4SLinus Torvalds } 23601da177e4SLinus Torvalds 2361e22bee78STejun Heo /** 2362e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2363111c225aSTejun Heo * @__rescuer: self 2364e22bee78STejun Heo * 2365e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2366493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2367e22bee78STejun Heo * 2368706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2369e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2370e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2371e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2372e22bee78STejun Heo * the problem rescuer solves. 2373e22bee78STejun Heo * 2374706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2375706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2376e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2377e22bee78STejun Heo * 2378e22bee78STejun Heo * This should happen rarely. 2379d185af30SYacine Belkadi * 2380d185af30SYacine Belkadi * Return: 0 2381e22bee78STejun Heo */ 2382111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2383e22bee78STejun Heo { 2384111c225aSTejun Heo struct worker *rescuer = __rescuer; 2385111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2386e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2387e22bee78STejun Heo 2388e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2389111c225aSTejun Heo 2390111c225aSTejun Heo /* 2391111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2392111c225aSTejun Heo * doesn't participate in concurrency management. 2393111c225aSTejun Heo */ 2394111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2395e22bee78STejun Heo repeat: 2396e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23971da177e4SLinus Torvalds 2398412d32e6SMike Galbraith if (kthread_should_stop()) { 2399412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2400111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2401e22bee78STejun Heo return 0; 2402412d32e6SMike Galbraith } 24031da177e4SLinus Torvalds 2404493a1724STejun Heo /* see whether any pwq is asking for help */ 24052e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2406493a1724STejun Heo 2407493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2408493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2409493a1724STejun Heo struct pool_workqueue, mayday_node); 2410112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2411e22bee78STejun Heo struct work_struct *work, *n; 2412e22bee78STejun Heo 2413e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2414493a1724STejun Heo list_del_init(&pwq->mayday_node); 2415493a1724STejun Heo 24162e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2417e22bee78STejun Heo 2418e22bee78STejun Heo /* migrate to the target cpu if possible */ 2419f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2420b3104104SLai Jiangshan rescuer->pool = pool; 2421e22bee78STejun Heo 2422e22bee78STejun Heo /* 2423e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2424e22bee78STejun Heo * process'em. 2425e22bee78STejun Heo */ 24266183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2427bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2428112202d9STejun Heo if (get_work_pwq(work) == pwq) 2429e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2430e22bee78STejun Heo 2431e22bee78STejun Heo process_scheduled_works(rescuer); 24327576958aSTejun Heo 24337576958aSTejun Heo /* 2434d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 24357576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24367576958aSTejun Heo * and stalling the execution. 24377576958aSTejun Heo */ 243863d95a91STejun Heo if (keep_working(pool)) 243963d95a91STejun Heo wake_up_worker(pool); 24407576958aSTejun Heo 2441b3104104SLai Jiangshan rescuer->pool = NULL; 2442493a1724STejun Heo spin_unlock(&pool->lock); 24432e109a28STejun Heo spin_lock(&wq_mayday_lock); 24441da177e4SLinus Torvalds } 24451da177e4SLinus Torvalds 24462e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2447493a1724STejun Heo 2448111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2449111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2450e22bee78STejun Heo schedule(); 2451e22bee78STejun Heo goto repeat; 24521da177e4SLinus Torvalds } 24531da177e4SLinus Torvalds 2454fc2e4d70SOleg Nesterov struct wq_barrier { 2455fc2e4d70SOleg Nesterov struct work_struct work; 2456fc2e4d70SOleg Nesterov struct completion done; 2457fc2e4d70SOleg Nesterov }; 2458fc2e4d70SOleg Nesterov 2459fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2460fc2e4d70SOleg Nesterov { 2461fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2462fc2e4d70SOleg Nesterov complete(&barr->done); 2463fc2e4d70SOleg Nesterov } 2464fc2e4d70SOleg Nesterov 24654690c4abSTejun Heo /** 24664690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2467112202d9STejun Heo * @pwq: pwq to insert barrier into 24684690c4abSTejun Heo * @barr: wq_barrier to insert 2469affee4b2STejun Heo * @target: target work to attach @barr to 2470affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24714690c4abSTejun Heo * 2472affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2473affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2474affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2475affee4b2STejun Heo * cpu. 2476affee4b2STejun Heo * 2477affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2478affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2479affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2480affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2481affee4b2STejun Heo * after a work with LINKED flag set. 2482affee4b2STejun Heo * 2483affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2484112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24854690c4abSTejun Heo * 24864690c4abSTejun Heo * CONTEXT: 2487d565ed63STejun Heo * spin_lock_irq(pool->lock). 24884690c4abSTejun Heo */ 2489112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2490affee4b2STejun Heo struct wq_barrier *barr, 2491affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2492fc2e4d70SOleg Nesterov { 2493affee4b2STejun Heo struct list_head *head; 2494affee4b2STejun Heo unsigned int linked = 0; 2495affee4b2STejun Heo 2496dc186ad7SThomas Gleixner /* 2497d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2498dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2499dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2500dc186ad7SThomas Gleixner * might deadlock. 2501dc186ad7SThomas Gleixner */ 2502ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 250322df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2504fc2e4d70SOleg Nesterov init_completion(&barr->done); 250583c22520SOleg Nesterov 2506affee4b2STejun Heo /* 2507affee4b2STejun Heo * If @target is currently being executed, schedule the 2508affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2509affee4b2STejun Heo */ 2510affee4b2STejun Heo if (worker) 2511affee4b2STejun Heo head = worker->scheduled.next; 2512affee4b2STejun Heo else { 2513affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2514affee4b2STejun Heo 2515affee4b2STejun Heo head = target->entry.next; 2516affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2517affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2518affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2519affee4b2STejun Heo } 2520affee4b2STejun Heo 2521dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2522112202d9STejun Heo insert_work(pwq, &barr->work, head, 2523affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2524fc2e4d70SOleg Nesterov } 2525fc2e4d70SOleg Nesterov 252673f53c4aSTejun Heo /** 2527112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 252873f53c4aSTejun Heo * @wq: workqueue being flushed 252973f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 253073f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 253173f53c4aSTejun Heo * 2532112202d9STejun Heo * Prepare pwqs for workqueue flushing. 253373f53c4aSTejun Heo * 2534112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2535112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2536112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2537112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2538112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 253973f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 254073f53c4aSTejun Heo * 254173f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 254273f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 254373f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 254473f53c4aSTejun Heo * is returned. 254573f53c4aSTejun Heo * 2546112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 254773f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 254873f53c4aSTejun Heo * advanced to @work_color. 254973f53c4aSTejun Heo * 255073f53c4aSTejun Heo * CONTEXT: 25513c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 255273f53c4aSTejun Heo * 2553d185af30SYacine Belkadi * Return: 255473f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 255573f53c4aSTejun Heo * otherwise. 255673f53c4aSTejun Heo */ 2557112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 255873f53c4aSTejun Heo int flush_color, int work_color) 25591da177e4SLinus Torvalds { 256073f53c4aSTejun Heo bool wait = false; 256149e3cf44STejun Heo struct pool_workqueue *pwq; 25621da177e4SLinus Torvalds 256373f53c4aSTejun Heo if (flush_color >= 0) { 25646183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2565112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2566dc186ad7SThomas Gleixner } 256714441960SOleg Nesterov 256849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2569112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25701da177e4SLinus Torvalds 2571b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 257273f53c4aSTejun Heo 257373f53c4aSTejun Heo if (flush_color >= 0) { 25746183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 257573f53c4aSTejun Heo 2576112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2577112202d9STejun Heo pwq->flush_color = flush_color; 2578112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 257973f53c4aSTejun Heo wait = true; 25801da177e4SLinus Torvalds } 258173f53c4aSTejun Heo } 258273f53c4aSTejun Heo 258373f53c4aSTejun Heo if (work_color >= 0) { 25846183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2585112202d9STejun Heo pwq->work_color = work_color; 258673f53c4aSTejun Heo } 258773f53c4aSTejun Heo 2588b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25891da177e4SLinus Torvalds } 25901da177e4SLinus Torvalds 2591112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 259273f53c4aSTejun Heo complete(&wq->first_flusher->done); 259373f53c4aSTejun Heo 259473f53c4aSTejun Heo return wait; 259583c22520SOleg Nesterov } 25961da177e4SLinus Torvalds 25970fcb78c2SRolf Eike Beer /** 25981da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25990fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26001da177e4SLinus Torvalds * 2601c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2602c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 26031da177e4SLinus Torvalds */ 26047ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26051da177e4SLinus Torvalds { 260673f53c4aSTejun Heo struct wq_flusher this_flusher = { 260773f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 260873f53c4aSTejun Heo .flush_color = -1, 260973f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 261073f53c4aSTejun Heo }; 261173f53c4aSTejun Heo int next_color; 2612b1f4ec17SOleg Nesterov 26133295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 26143295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 261573f53c4aSTejun Heo 26163c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 261773f53c4aSTejun Heo 261873f53c4aSTejun Heo /* 261973f53c4aSTejun Heo * Start-to-wait phase 262073f53c4aSTejun Heo */ 262173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 262273f53c4aSTejun Heo 262373f53c4aSTejun Heo if (next_color != wq->flush_color) { 262473f53c4aSTejun Heo /* 262573f53c4aSTejun Heo * Color space is not full. The current work_color 262673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 262773f53c4aSTejun Heo * by one. 262873f53c4aSTejun Heo */ 26296183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 263073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 263173f53c4aSTejun Heo wq->work_color = next_color; 263273f53c4aSTejun Heo 263373f53c4aSTejun Heo if (!wq->first_flusher) { 263473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26356183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 263673f53c4aSTejun Heo 263773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 263873f53c4aSTejun Heo 2639112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 264073f53c4aSTejun Heo wq->work_color)) { 264173f53c4aSTejun Heo /* nothing to flush, done */ 264273f53c4aSTejun Heo wq->flush_color = next_color; 264373f53c4aSTejun Heo wq->first_flusher = NULL; 264473f53c4aSTejun Heo goto out_unlock; 264573f53c4aSTejun Heo } 264673f53c4aSTejun Heo } else { 264773f53c4aSTejun Heo /* wait in queue */ 26486183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 264973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2650112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 265173f53c4aSTejun Heo } 265273f53c4aSTejun Heo } else { 265373f53c4aSTejun Heo /* 265473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 265573f53c4aSTejun Heo * The next flush completion will assign us 265673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 265773f53c4aSTejun Heo */ 265873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 265973f53c4aSTejun Heo } 266073f53c4aSTejun Heo 26613c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 266273f53c4aSTejun Heo 266373f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 266473f53c4aSTejun Heo 266573f53c4aSTejun Heo /* 266673f53c4aSTejun Heo * Wake-up-and-cascade phase 266773f53c4aSTejun Heo * 266873f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 266973f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 267073f53c4aSTejun Heo */ 267173f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 267273f53c4aSTejun Heo return; 267373f53c4aSTejun Heo 26743c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 267573f53c4aSTejun Heo 26764ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26774ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26784ce48b37STejun Heo goto out_unlock; 26794ce48b37STejun Heo 268073f53c4aSTejun Heo wq->first_flusher = NULL; 268173f53c4aSTejun Heo 26826183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26836183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 268473f53c4aSTejun Heo 268573f53c4aSTejun Heo while (true) { 268673f53c4aSTejun Heo struct wq_flusher *next, *tmp; 268773f53c4aSTejun Heo 268873f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 268973f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 269073f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 269173f53c4aSTejun Heo break; 269273f53c4aSTejun Heo list_del_init(&next->list); 269373f53c4aSTejun Heo complete(&next->done); 269473f53c4aSTejun Heo } 269573f53c4aSTejun Heo 26966183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 269773f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 269873f53c4aSTejun Heo 269973f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 270073f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 270173f53c4aSTejun Heo 270273f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 270373f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 270473f53c4aSTejun Heo /* 270573f53c4aSTejun Heo * Assign the same color to all overflowed 270673f53c4aSTejun Heo * flushers, advance work_color and append to 270773f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 270873f53c4aSTejun Heo * phase for these overflowed flushers. 270973f53c4aSTejun Heo */ 271073f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 271173f53c4aSTejun Heo tmp->flush_color = wq->work_color; 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 271473f53c4aSTejun Heo 271573f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 271673f53c4aSTejun Heo &wq->flusher_queue); 2717112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 271873f53c4aSTejun Heo } 271973f53c4aSTejun Heo 272073f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27216183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 272273f53c4aSTejun Heo break; 272373f53c4aSTejun Heo } 272473f53c4aSTejun Heo 272573f53c4aSTejun Heo /* 272673f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2727112202d9STejun Heo * the new first flusher and arm pwqs. 272873f53c4aSTejun Heo */ 27296183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27306183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 273173f53c4aSTejun Heo 273273f53c4aSTejun Heo list_del_init(&next->list); 273373f53c4aSTejun Heo wq->first_flusher = next; 273473f53c4aSTejun Heo 2735112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 273673f53c4aSTejun Heo break; 273773f53c4aSTejun Heo 273873f53c4aSTejun Heo /* 273973f53c4aSTejun Heo * Meh... this color is already done, clear first 274073f53c4aSTejun Heo * flusher and repeat cascading. 274173f53c4aSTejun Heo */ 274273f53c4aSTejun Heo wq->first_flusher = NULL; 274373f53c4aSTejun Heo } 274473f53c4aSTejun Heo 274573f53c4aSTejun Heo out_unlock: 27463c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27471da177e4SLinus Torvalds } 2748ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27491da177e4SLinus Torvalds 27509c5a2ba7STejun Heo /** 27519c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27529c5a2ba7STejun Heo * @wq: workqueue to drain 27539c5a2ba7STejun Heo * 27549c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27559c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27569c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27579c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27589c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27599c5a2ba7STejun Heo * takes too long. 27609c5a2ba7STejun Heo */ 27619c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27629c5a2ba7STejun Heo { 27639c5a2ba7STejun Heo unsigned int flush_cnt = 0; 276449e3cf44STejun Heo struct pool_workqueue *pwq; 27659c5a2ba7STejun Heo 27669c5a2ba7STejun Heo /* 27679c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27689c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2769618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27709c5a2ba7STejun Heo */ 277187fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27729c5a2ba7STejun Heo if (!wq->nr_drainers++) 2773618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 277487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27759c5a2ba7STejun Heo reflush: 27769c5a2ba7STejun Heo flush_workqueue(wq); 27779c5a2ba7STejun Heo 2778b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 277976af4d93STejun Heo 278049e3cf44STejun Heo for_each_pwq(pwq, wq) { 2781fa2563e4SThomas Tuttle bool drained; 27829c5a2ba7STejun Heo 2783b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2784112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2785b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2786fa2563e4SThomas Tuttle 2787fa2563e4SThomas Tuttle if (drained) 27889c5a2ba7STejun Heo continue; 27899c5a2ba7STejun Heo 27909c5a2ba7STejun Heo if (++flush_cnt == 10 || 27919c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2792c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27939c5a2ba7STejun Heo wq->name, flush_cnt); 279476af4d93STejun Heo 2795b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27969c5a2ba7STejun Heo goto reflush; 27979c5a2ba7STejun Heo } 27989c5a2ba7STejun Heo 27999c5a2ba7STejun Heo if (!--wq->nr_drainers) 2800618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 280187fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28029c5a2ba7STejun Heo } 28039c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28049c5a2ba7STejun Heo 2805606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2806baf59022STejun Heo { 2807baf59022STejun Heo struct worker *worker = NULL; 2808c9e7cf27STejun Heo struct worker_pool *pool; 2809112202d9STejun Heo struct pool_workqueue *pwq; 2810baf59022STejun Heo 2811baf59022STejun Heo might_sleep(); 2812baf59022STejun Heo 2813fa1b54e6STejun Heo local_irq_disable(); 2814fa1b54e6STejun Heo pool = get_work_pool(work); 2815fa1b54e6STejun Heo if (!pool) { 2816fa1b54e6STejun Heo local_irq_enable(); 2817fa1b54e6STejun Heo return false; 2818fa1b54e6STejun Heo } 2819fa1b54e6STejun Heo 2820fa1b54e6STejun Heo spin_lock(&pool->lock); 28210b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2822112202d9STejun Heo pwq = get_work_pwq(work); 2823112202d9STejun Heo if (pwq) { 2824112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2825baf59022STejun Heo goto already_gone; 2826606a5020STejun Heo } else { 2827c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2828baf59022STejun Heo if (!worker) 2829baf59022STejun Heo goto already_gone; 2830112202d9STejun Heo pwq = worker->current_pwq; 2831606a5020STejun Heo } 2832baf59022STejun Heo 2833112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2834d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2835baf59022STejun Heo 2836e159489bSTejun Heo /* 2837e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2838e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2839e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2840e159489bSTejun Heo * access. 2841e159489bSTejun Heo */ 2842493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2843112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2844e159489bSTejun Heo else 2845112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2846112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2847e159489bSTejun Heo 2848baf59022STejun Heo return true; 2849baf59022STejun Heo already_gone: 2850d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2851baf59022STejun Heo return false; 2852baf59022STejun Heo } 2853baf59022STejun Heo 2854c2fda509SLai Jiangshan static bool __flush_work(struct work_struct *work) 2855c2fda509SLai Jiangshan { 2856c2fda509SLai Jiangshan struct wq_barrier barr; 2857c2fda509SLai Jiangshan 2858c2fda509SLai Jiangshan if (start_flush_work(work, &barr)) { 2859c2fda509SLai Jiangshan wait_for_completion(&barr.done); 2860c2fda509SLai Jiangshan destroy_work_on_stack(&barr.work); 2861c2fda509SLai Jiangshan return true; 2862c2fda509SLai Jiangshan } else { 2863c2fda509SLai Jiangshan return false; 2864c2fda509SLai Jiangshan } 2865c2fda509SLai Jiangshan } 2866c2fda509SLai Jiangshan 2867db700897SOleg Nesterov /** 2868401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2869401a8d04STejun Heo * @work: the work to flush 2870db700897SOleg Nesterov * 2871606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2872606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2873401a8d04STejun Heo * 2874d185af30SYacine Belkadi * Return: 2875401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2876401a8d04STejun Heo * %false if it was already idle. 2877db700897SOleg Nesterov */ 2878401a8d04STejun Heo bool flush_work(struct work_struct *work) 2879db700897SOleg Nesterov { 28800976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28810976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28820976dfc1SStephen Boyd 2883c2fda509SLai Jiangshan return __flush_work(work); 2884606a5020STejun Heo } 2885db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2886db700897SOleg Nesterov 288736e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2888401a8d04STejun Heo { 2889bbb68dfaSTejun Heo unsigned long flags; 28901f1f642eSOleg Nesterov int ret; 28911f1f642eSOleg Nesterov 28921f1f642eSOleg Nesterov do { 2893bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2894bbb68dfaSTejun Heo /* 2895bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2896bbb68dfaSTejun Heo * would be waiting for before retrying. 2897bbb68dfaSTejun Heo */ 2898bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2899606a5020STejun Heo flush_work(work); 29001f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29011f1f642eSOleg Nesterov 2902bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2903bbb68dfaSTejun Heo mark_work_canceling(work); 2904bbb68dfaSTejun Heo local_irq_restore(flags); 2905bbb68dfaSTejun Heo 2906606a5020STejun Heo flush_work(work); 29077a22ad75STejun Heo clear_work_data(work); 29081f1f642eSOleg Nesterov return ret; 29091f1f642eSOleg Nesterov } 29101f1f642eSOleg Nesterov 29116e84d644SOleg Nesterov /** 2912401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2913401a8d04STejun Heo * @work: the work to cancel 29146e84d644SOleg Nesterov * 2915401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2916401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2917401a8d04STejun Heo * another workqueue. On return from this function, @work is 2918401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29191f1f642eSOleg Nesterov * 2920401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2921401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29226e84d644SOleg Nesterov * 2923401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29246e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2925401a8d04STejun Heo * 2926d185af30SYacine Belkadi * Return: 2927401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29286e84d644SOleg Nesterov */ 2929401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29306e84d644SOleg Nesterov { 293136e227d2STejun Heo return __cancel_work_timer(work, false); 2932b89deed3SOleg Nesterov } 293328e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2934b89deed3SOleg Nesterov 29356e84d644SOleg Nesterov /** 2936401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2937401a8d04STejun Heo * @dwork: the delayed work to flush 29386e84d644SOleg Nesterov * 2939401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2940401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2941401a8d04STejun Heo * considers the last queueing instance of @dwork. 29421f1f642eSOleg Nesterov * 2943d185af30SYacine Belkadi * Return: 2944401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2945401a8d04STejun Heo * %false if it was already idle. 29466e84d644SOleg Nesterov */ 2947401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2948401a8d04STejun Heo { 29498930cabaSTejun Heo local_irq_disable(); 2950401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 295160c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29528930cabaSTejun Heo local_irq_enable(); 2953401a8d04STejun Heo return flush_work(&dwork->work); 2954401a8d04STejun Heo } 2955401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2956401a8d04STejun Heo 2957401a8d04STejun Heo /** 295857b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 295957b30ae7STejun Heo * @dwork: delayed_work to cancel 296009383498STejun Heo * 2961d185af30SYacine Belkadi * Kill off a pending delayed_work. 2962d185af30SYacine Belkadi * 2963d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2964d185af30SYacine Belkadi * pending. 2965d185af30SYacine Belkadi * 2966d185af30SYacine Belkadi * Note: 2967d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2968d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2969d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 297009383498STejun Heo * 297157b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 297209383498STejun Heo */ 297357b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 297409383498STejun Heo { 297557b30ae7STejun Heo unsigned long flags; 297657b30ae7STejun Heo int ret; 297757b30ae7STejun Heo 297857b30ae7STejun Heo do { 297957b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 298057b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 298157b30ae7STejun Heo 298257b30ae7STejun Heo if (unlikely(ret < 0)) 298357b30ae7STejun Heo return false; 298457b30ae7STejun Heo 29857c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29867c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 298757b30ae7STejun Heo local_irq_restore(flags); 2988c0158ca6SDan Magenheimer return ret; 298909383498STejun Heo } 299057b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 299109383498STejun Heo 299209383498STejun Heo /** 2993401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2994401a8d04STejun Heo * @dwork: the delayed work cancel 2995401a8d04STejun Heo * 2996401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2997401a8d04STejun Heo * 2998d185af30SYacine Belkadi * Return: 2999401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3000401a8d04STejun Heo */ 3001401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30026e84d644SOleg Nesterov { 300336e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30046e84d644SOleg Nesterov } 3005f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30061da177e4SLinus Torvalds 30070fcb78c2SRolf Eike Beer /** 300831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3009b6136773SAndrew Morton * @func: the function to call 3010b6136773SAndrew Morton * 301131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 301231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3013b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 301431ddd871STejun Heo * 3015d185af30SYacine Belkadi * Return: 301631ddd871STejun Heo * 0 on success, -errno on failure. 3017b6136773SAndrew Morton */ 301865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 301915316ba8SChristoph Lameter { 302015316ba8SChristoph Lameter int cpu; 302138f51568SNamhyung Kim struct work_struct __percpu *works; 302215316ba8SChristoph Lameter 3023b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3024b6136773SAndrew Morton if (!works) 302515316ba8SChristoph Lameter return -ENOMEM; 3026b6136773SAndrew Morton 302795402b38SGautham R Shenoy get_online_cpus(); 302893981800STejun Heo 302915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30309bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30319bfb1839SIngo Molnar 30329bfb1839SIngo Molnar INIT_WORK(work, func); 30338de6d308SOleg Nesterov schedule_work_on(cpu, work); 303415316ba8SChristoph Lameter } 303593981800STejun Heo 303693981800STejun Heo for_each_online_cpu(cpu) 30378616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 303893981800STejun Heo 303995402b38SGautham R Shenoy put_online_cpus(); 3040b6136773SAndrew Morton free_percpu(works); 304115316ba8SChristoph Lameter return 0; 304215316ba8SChristoph Lameter } 304315316ba8SChristoph Lameter 3044eef6a7d5SAlan Stern /** 3045eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3046eef6a7d5SAlan Stern * 3047eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3048eef6a7d5SAlan Stern * completion. 3049eef6a7d5SAlan Stern * 3050eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3051eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3052eef6a7d5SAlan Stern * will lead to deadlock: 3053eef6a7d5SAlan Stern * 3054eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3055eef6a7d5SAlan Stern * a lock held by your code or its caller. 3056eef6a7d5SAlan Stern * 3057eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3058eef6a7d5SAlan Stern * 3059eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3060eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3061eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3062eef6a7d5SAlan Stern * 3063eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3064eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3065eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3066eef6a7d5SAlan Stern * cancel_work_sync() instead. 3067eef6a7d5SAlan Stern */ 30681da177e4SLinus Torvalds void flush_scheduled_work(void) 30691da177e4SLinus Torvalds { 3070d320c038STejun Heo flush_workqueue(system_wq); 30711da177e4SLinus Torvalds } 3072ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30731da177e4SLinus Torvalds 30741da177e4SLinus Torvalds /** 30751fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30761fa44ecaSJames Bottomley * @fn: the function to execute 30771fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30781fa44ecaSJames Bottomley * be available when the work executes) 30791fa44ecaSJames Bottomley * 30801fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30811fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30821fa44ecaSJames Bottomley * 3083d185af30SYacine Belkadi * Return: 0 - function was executed 30841fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30851fa44ecaSJames Bottomley */ 308665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30871fa44ecaSJames Bottomley { 30881fa44ecaSJames Bottomley if (!in_interrupt()) { 308965f27f38SDavid Howells fn(&ew->work); 30901fa44ecaSJames Bottomley return 0; 30911fa44ecaSJames Bottomley } 30921fa44ecaSJames Bottomley 309365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30941fa44ecaSJames Bottomley schedule_work(&ew->work); 30951fa44ecaSJames Bottomley 30961fa44ecaSJames Bottomley return 1; 30971fa44ecaSJames Bottomley } 30981fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30991fa44ecaSJames Bottomley 3100226223abSTejun Heo #ifdef CONFIG_SYSFS 3101226223abSTejun Heo /* 3102226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3103226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3104226223abSTejun Heo * following attributes. 3105226223abSTejun Heo * 3106226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3107226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3108226223abSTejun Heo * 3109226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3110226223abSTejun Heo * 3111226223abSTejun Heo * id RO int : the associated pool ID 3112226223abSTejun Heo * nice RW int : nice value of the workers 3113226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3114226223abSTejun Heo */ 3115226223abSTejun Heo struct wq_device { 3116226223abSTejun Heo struct workqueue_struct *wq; 3117226223abSTejun Heo struct device dev; 3118226223abSTejun Heo }; 3119226223abSTejun Heo 3120226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3121226223abSTejun Heo { 3122226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3123226223abSTejun Heo 3124226223abSTejun Heo return wq_dev->wq; 3125226223abSTejun Heo } 3126226223abSTejun Heo 31271a6661daSGreg Kroah-Hartman static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 31281a6661daSGreg Kroah-Hartman char *buf) 3129226223abSTejun Heo { 3130226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3131226223abSTejun Heo 3132226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3133226223abSTejun Heo } 31341a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RO(per_cpu); 3135226223abSTejun Heo 31361a6661daSGreg Kroah-Hartman static ssize_t max_active_show(struct device *dev, 3137226223abSTejun Heo struct device_attribute *attr, char *buf) 3138226223abSTejun Heo { 3139226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3140226223abSTejun Heo 3141226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3142226223abSTejun Heo } 3143226223abSTejun Heo 31441a6661daSGreg Kroah-Hartman static ssize_t max_active_store(struct device *dev, 31451a6661daSGreg Kroah-Hartman struct device_attribute *attr, const char *buf, 31461a6661daSGreg Kroah-Hartman size_t count) 3147226223abSTejun Heo { 3148226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3149226223abSTejun Heo int val; 3150226223abSTejun Heo 3151226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3152226223abSTejun Heo return -EINVAL; 3153226223abSTejun Heo 3154226223abSTejun Heo workqueue_set_max_active(wq, val); 3155226223abSTejun Heo return count; 3156226223abSTejun Heo } 31571a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RW(max_active); 3158226223abSTejun Heo 31591a6661daSGreg Kroah-Hartman static struct attribute *wq_sysfs_attrs[] = { 31601a6661daSGreg Kroah-Hartman &dev_attr_per_cpu.attr, 31611a6661daSGreg Kroah-Hartman &dev_attr_max_active.attr, 31621a6661daSGreg Kroah-Hartman NULL, 3163226223abSTejun Heo }; 31641a6661daSGreg Kroah-Hartman ATTRIBUTE_GROUPS(wq_sysfs); 3165226223abSTejun Heo 3166d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3167226223abSTejun Heo struct device_attribute *attr, char *buf) 3168226223abSTejun Heo { 3169226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3170d55262c4STejun Heo const char *delim = ""; 3171d55262c4STejun Heo int node, written = 0; 3172226223abSTejun Heo 3173226223abSTejun Heo rcu_read_lock_sched(); 3174d55262c4STejun Heo for_each_node(node) { 3175d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3176d55262c4STejun Heo "%s%d:%d", delim, node, 3177d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3178d55262c4STejun Heo delim = " "; 3179d55262c4STejun Heo } 3180d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3181226223abSTejun Heo rcu_read_unlock_sched(); 3182226223abSTejun Heo 3183226223abSTejun Heo return written; 3184226223abSTejun Heo } 3185226223abSTejun Heo 3186226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3187226223abSTejun Heo char *buf) 3188226223abSTejun Heo { 3189226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3190226223abSTejun Heo int written; 3191226223abSTejun Heo 31926029a918STejun Heo mutex_lock(&wq->mutex); 31936029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 31946029a918STejun Heo mutex_unlock(&wq->mutex); 3195226223abSTejun Heo 3196226223abSTejun Heo return written; 3197226223abSTejun Heo } 3198226223abSTejun Heo 3199226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3200226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3201226223abSTejun Heo { 3202226223abSTejun Heo struct workqueue_attrs *attrs; 3203226223abSTejun Heo 3204226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3205226223abSTejun Heo if (!attrs) 3206226223abSTejun Heo return NULL; 3207226223abSTejun Heo 32086029a918STejun Heo mutex_lock(&wq->mutex); 32096029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 32106029a918STejun Heo mutex_unlock(&wq->mutex); 3211226223abSTejun Heo return attrs; 3212226223abSTejun Heo } 3213226223abSTejun Heo 3214226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3215226223abSTejun Heo const char *buf, size_t count) 3216226223abSTejun Heo { 3217226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3218226223abSTejun Heo struct workqueue_attrs *attrs; 3219226223abSTejun Heo int ret; 3220226223abSTejun Heo 3221226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3222226223abSTejun Heo if (!attrs) 3223226223abSTejun Heo return -ENOMEM; 3224226223abSTejun Heo 3225226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3226226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3227226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3228226223abSTejun Heo else 3229226223abSTejun Heo ret = -EINVAL; 3230226223abSTejun Heo 3231226223abSTejun Heo free_workqueue_attrs(attrs); 3232226223abSTejun Heo return ret ?: count; 3233226223abSTejun Heo } 3234226223abSTejun Heo 3235226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3236226223abSTejun Heo struct device_attribute *attr, char *buf) 3237226223abSTejun Heo { 3238226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3239226223abSTejun Heo int written; 3240226223abSTejun Heo 32416029a918STejun Heo mutex_lock(&wq->mutex); 32426029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 32436029a918STejun Heo mutex_unlock(&wq->mutex); 3244226223abSTejun Heo 3245226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3246226223abSTejun Heo return written; 3247226223abSTejun Heo } 3248226223abSTejun Heo 3249226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3250226223abSTejun Heo struct device_attribute *attr, 3251226223abSTejun Heo const char *buf, size_t count) 3252226223abSTejun Heo { 3253226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3254226223abSTejun Heo struct workqueue_attrs *attrs; 3255226223abSTejun Heo int ret; 3256226223abSTejun Heo 3257226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3258226223abSTejun Heo if (!attrs) 3259226223abSTejun Heo return -ENOMEM; 3260226223abSTejun Heo 3261226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3262226223abSTejun Heo if (!ret) 3263226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3264226223abSTejun Heo 3265226223abSTejun Heo free_workqueue_attrs(attrs); 3266226223abSTejun Heo return ret ?: count; 3267226223abSTejun Heo } 3268226223abSTejun Heo 3269d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3270d55262c4STejun Heo char *buf) 3271d55262c4STejun Heo { 3272d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3273d55262c4STejun Heo int written; 3274d55262c4STejun Heo 3275d55262c4STejun Heo mutex_lock(&wq->mutex); 3276d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3277d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3278d55262c4STejun Heo mutex_unlock(&wq->mutex); 3279d55262c4STejun Heo 3280d55262c4STejun Heo return written; 3281d55262c4STejun Heo } 3282d55262c4STejun Heo 3283d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3284d55262c4STejun Heo const char *buf, size_t count) 3285d55262c4STejun Heo { 3286d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3287d55262c4STejun Heo struct workqueue_attrs *attrs; 3288d55262c4STejun Heo int v, ret; 3289d55262c4STejun Heo 3290d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3291d55262c4STejun Heo if (!attrs) 3292d55262c4STejun Heo return -ENOMEM; 3293d55262c4STejun Heo 3294d55262c4STejun Heo ret = -EINVAL; 3295d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3296d55262c4STejun Heo attrs->no_numa = !v; 3297d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3298d55262c4STejun Heo } 3299d55262c4STejun Heo 3300d55262c4STejun Heo free_workqueue_attrs(attrs); 3301d55262c4STejun Heo return ret ?: count; 3302d55262c4STejun Heo } 3303d55262c4STejun Heo 3304226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3305d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3306226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3307226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3308d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3309226223abSTejun Heo __ATTR_NULL, 3310226223abSTejun Heo }; 3311226223abSTejun Heo 3312226223abSTejun Heo static struct bus_type wq_subsys = { 3313226223abSTejun Heo .name = "workqueue", 33141a6661daSGreg Kroah-Hartman .dev_groups = wq_sysfs_groups, 3315226223abSTejun Heo }; 3316226223abSTejun Heo 3317226223abSTejun Heo static int __init wq_sysfs_init(void) 3318226223abSTejun Heo { 3319226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3320226223abSTejun Heo } 3321226223abSTejun Heo core_initcall(wq_sysfs_init); 3322226223abSTejun Heo 3323226223abSTejun Heo static void wq_device_release(struct device *dev) 3324226223abSTejun Heo { 3325226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3326226223abSTejun Heo 3327226223abSTejun Heo kfree(wq_dev); 3328226223abSTejun Heo } 3329226223abSTejun Heo 3330226223abSTejun Heo /** 3331226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3332226223abSTejun Heo * @wq: the workqueue to register 3333226223abSTejun Heo * 3334226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3335226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3336226223abSTejun Heo * which is the preferred method. 3337226223abSTejun Heo * 3338226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3339226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3340226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3341226223abSTejun Heo * attributes. 3342226223abSTejun Heo * 3343d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. 3344226223abSTejun Heo */ 3345226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3346226223abSTejun Heo { 3347226223abSTejun Heo struct wq_device *wq_dev; 3348226223abSTejun Heo int ret; 3349226223abSTejun Heo 3350226223abSTejun Heo /* 3351226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3352226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3353226223abSTejun Heo * workqueues. 3354226223abSTejun Heo */ 3355226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3356226223abSTejun Heo return -EINVAL; 3357226223abSTejun Heo 3358226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3359226223abSTejun Heo if (!wq_dev) 3360226223abSTejun Heo return -ENOMEM; 3361226223abSTejun Heo 3362226223abSTejun Heo wq_dev->wq = wq; 3363226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3364226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3365226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3366226223abSTejun Heo 3367226223abSTejun Heo /* 3368226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3369226223abSTejun Heo * everything is ready. 3370226223abSTejun Heo */ 3371226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3372226223abSTejun Heo 3373226223abSTejun Heo ret = device_register(&wq_dev->dev); 3374226223abSTejun Heo if (ret) { 3375226223abSTejun Heo kfree(wq_dev); 3376226223abSTejun Heo wq->wq_dev = NULL; 3377226223abSTejun Heo return ret; 3378226223abSTejun Heo } 3379226223abSTejun Heo 3380226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3381226223abSTejun Heo struct device_attribute *attr; 3382226223abSTejun Heo 3383226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3384226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3385226223abSTejun Heo if (ret) { 3386226223abSTejun Heo device_unregister(&wq_dev->dev); 3387226223abSTejun Heo wq->wq_dev = NULL; 3388226223abSTejun Heo return ret; 3389226223abSTejun Heo } 3390226223abSTejun Heo } 3391226223abSTejun Heo } 3392226223abSTejun Heo 3393226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3394226223abSTejun Heo return 0; 3395226223abSTejun Heo } 3396226223abSTejun Heo 3397226223abSTejun Heo /** 3398226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3399226223abSTejun Heo * @wq: the workqueue to unregister 3400226223abSTejun Heo * 3401226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3402226223abSTejun Heo */ 3403226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3404226223abSTejun Heo { 3405226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3406226223abSTejun Heo 3407226223abSTejun Heo if (!wq->wq_dev) 3408226223abSTejun Heo return; 3409226223abSTejun Heo 3410226223abSTejun Heo wq->wq_dev = NULL; 3411226223abSTejun Heo device_unregister(&wq_dev->dev); 3412226223abSTejun Heo } 3413226223abSTejun Heo #else /* CONFIG_SYSFS */ 3414226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3415226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3416226223abSTejun Heo 34177a4e344cSTejun Heo /** 34187a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 34197a4e344cSTejun Heo * @attrs: workqueue_attrs to free 34207a4e344cSTejun Heo * 34217a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 34227a4e344cSTejun Heo */ 34237a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 34247a4e344cSTejun Heo { 34257a4e344cSTejun Heo if (attrs) { 34267a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 34277a4e344cSTejun Heo kfree(attrs); 34287a4e344cSTejun Heo } 34297a4e344cSTejun Heo } 34307a4e344cSTejun Heo 34317a4e344cSTejun Heo /** 34327a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 34337a4e344cSTejun Heo * @gfp_mask: allocation mask to use 34347a4e344cSTejun Heo * 34357a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3436d185af30SYacine Belkadi * return it. 3437d185af30SYacine Belkadi * 3438d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 34397a4e344cSTejun Heo */ 34407a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 34417a4e344cSTejun Heo { 34427a4e344cSTejun Heo struct workqueue_attrs *attrs; 34437a4e344cSTejun Heo 34447a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 34457a4e344cSTejun Heo if (!attrs) 34467a4e344cSTejun Heo goto fail; 34477a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 34487a4e344cSTejun Heo goto fail; 34497a4e344cSTejun Heo 345013e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 34517a4e344cSTejun Heo return attrs; 34527a4e344cSTejun Heo fail: 34537a4e344cSTejun Heo free_workqueue_attrs(attrs); 34547a4e344cSTejun Heo return NULL; 34557a4e344cSTejun Heo } 34567a4e344cSTejun Heo 345729c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 345829c91e99STejun Heo const struct workqueue_attrs *from) 345929c91e99STejun Heo { 346029c91e99STejun Heo to->nice = from->nice; 346129c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 34622865a8fbSShaohua Li /* 34632865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 34642865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 34652865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 34662865a8fbSShaohua Li */ 34672865a8fbSShaohua Li to->no_numa = from->no_numa; 346829c91e99STejun Heo } 346929c91e99STejun Heo 347029c91e99STejun Heo /* hash value of the content of @attr */ 347129c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 347229c91e99STejun Heo { 347329c91e99STejun Heo u32 hash = 0; 347429c91e99STejun Heo 347529c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 347613e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 347713e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 347829c91e99STejun Heo return hash; 347929c91e99STejun Heo } 348029c91e99STejun Heo 348129c91e99STejun Heo /* content equality test */ 348229c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 348329c91e99STejun Heo const struct workqueue_attrs *b) 348429c91e99STejun Heo { 348529c91e99STejun Heo if (a->nice != b->nice) 348629c91e99STejun Heo return false; 348729c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 348829c91e99STejun Heo return false; 348929c91e99STejun Heo return true; 349029c91e99STejun Heo } 349129c91e99STejun Heo 34927a4e344cSTejun Heo /** 34937a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34947a4e344cSTejun Heo * @pool: worker_pool to initialize 34957a4e344cSTejun Heo * 34967a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 3497d185af30SYacine Belkadi * 3498d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 349929c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 350029c91e99STejun Heo * on @pool safely to release it. 35017a4e344cSTejun Heo */ 35027a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 35034e1a1f9aSTejun Heo { 35044e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 350529c91e99STejun Heo pool->id = -1; 350629c91e99STejun Heo pool->cpu = -1; 3507f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 35084e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 35094e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 35104e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 35114e1a1f9aSTejun Heo hash_init(pool->busy_hash); 35124e1a1f9aSTejun Heo 35134e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 35144e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 35154e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 35164e1a1f9aSTejun Heo 35174e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 35184e1a1f9aSTejun Heo (unsigned long)pool); 35194e1a1f9aSTejun Heo 35204e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3521bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3522822d8405STejun Heo idr_init(&pool->worker_idr); 35237a4e344cSTejun Heo 352429c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 352529c91e99STejun Heo pool->refcnt = 1; 352629c91e99STejun Heo 352729c91e99STejun Heo /* shouldn't fail above this point */ 35287a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 35297a4e344cSTejun Heo if (!pool->attrs) 35307a4e344cSTejun Heo return -ENOMEM; 35317a4e344cSTejun Heo return 0; 35324e1a1f9aSTejun Heo } 35334e1a1f9aSTejun Heo 353429c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 353529c91e99STejun Heo { 353629c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 353729c91e99STejun Heo 3538822d8405STejun Heo idr_destroy(&pool->worker_idr); 353929c91e99STejun Heo free_workqueue_attrs(pool->attrs); 354029c91e99STejun Heo kfree(pool); 354129c91e99STejun Heo } 354229c91e99STejun Heo 354329c91e99STejun Heo /** 354429c91e99STejun Heo * put_unbound_pool - put a worker_pool 354529c91e99STejun Heo * @pool: worker_pool to put 354629c91e99STejun Heo * 354729c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3548c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3549c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3550c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3551a892caccSTejun Heo * 3552a892caccSTejun Heo * Should be called with wq_pool_mutex held. 355329c91e99STejun Heo */ 355429c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 355529c91e99STejun Heo { 355629c91e99STejun Heo struct worker *worker; 355729c91e99STejun Heo 3558a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3559a892caccSTejun Heo 3560a892caccSTejun Heo if (--pool->refcnt) 356129c91e99STejun Heo return; 356229c91e99STejun Heo 356329c91e99STejun Heo /* sanity checks */ 356429c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3565a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 356629c91e99STejun Heo return; 356729c91e99STejun Heo 356829c91e99STejun Heo /* release id and unhash */ 356929c91e99STejun Heo if (pool->id >= 0) 357029c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 357129c91e99STejun Heo hash_del(&pool->hash_node); 357229c91e99STejun Heo 3573c5aa87bbSTejun Heo /* 3574c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3575c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3576c5aa87bbSTejun Heo * manager_mutex. 3577c5aa87bbSTejun Heo */ 357829c91e99STejun Heo mutex_lock(&pool->manager_arb); 3579cd549687STejun Heo mutex_lock(&pool->manager_mutex); 358029c91e99STejun Heo spin_lock_irq(&pool->lock); 358129c91e99STejun Heo 358229c91e99STejun Heo while ((worker = first_worker(pool))) 358329c91e99STejun Heo destroy_worker(worker); 358429c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 358529c91e99STejun Heo 358629c91e99STejun Heo spin_unlock_irq(&pool->lock); 3587cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 358829c91e99STejun Heo mutex_unlock(&pool->manager_arb); 358929c91e99STejun Heo 359029c91e99STejun Heo /* shut down the timers */ 359129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 359229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 359329c91e99STejun Heo 359429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 359529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 359629c91e99STejun Heo } 359729c91e99STejun Heo 359829c91e99STejun Heo /** 359929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 360029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 360129c91e99STejun Heo * 360229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 360329c91e99STejun Heo * reference count and return it. If there already is a matching 360429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3605d185af30SYacine Belkadi * create a new one. 3606a892caccSTejun Heo * 3607a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3608d185af30SYacine Belkadi * 3609d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3610d185af30SYacine Belkadi * On failure, %NULL. 361129c91e99STejun Heo */ 361229c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 361329c91e99STejun Heo { 361429c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 361529c91e99STejun Heo struct worker_pool *pool; 3616f3f90ad4STejun Heo int node; 361729c91e99STejun Heo 3618a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 361929c91e99STejun Heo 362029c91e99STejun Heo /* do we already have a matching pool? */ 362129c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 362229c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 362329c91e99STejun Heo pool->refcnt++; 362429c91e99STejun Heo goto out_unlock; 362529c91e99STejun Heo } 362629c91e99STejun Heo } 362729c91e99STejun Heo 362829c91e99STejun Heo /* nope, create a new one */ 362929c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 363029c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 363129c91e99STejun Heo goto fail; 363229c91e99STejun Heo 363312ee4fc6SLai Jiangshan if (workqueue_freezing) 363412ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 363512ee4fc6SLai Jiangshan 36368864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 363729c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 363829c91e99STejun Heo 36392865a8fbSShaohua Li /* 36402865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36412865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36422865a8fbSShaohua Li */ 36432865a8fbSShaohua Li pool->attrs->no_numa = false; 36442865a8fbSShaohua Li 3645f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3646f3f90ad4STejun Heo if (wq_numa_enabled) { 3647f3f90ad4STejun Heo for_each_node(node) { 3648f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3649f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3650f3f90ad4STejun Heo pool->node = node; 3651f3f90ad4STejun Heo break; 3652f3f90ad4STejun Heo } 3653f3f90ad4STejun Heo } 3654f3f90ad4STejun Heo } 3655f3f90ad4STejun Heo 365629c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 365729c91e99STejun Heo goto fail; 365829c91e99STejun Heo 365929c91e99STejun Heo /* create and start the initial worker */ 3660ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 366129c91e99STejun Heo goto fail; 366229c91e99STejun Heo 366329c91e99STejun Heo /* install */ 366429c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 366529c91e99STejun Heo out_unlock: 366629c91e99STejun Heo return pool; 366729c91e99STejun Heo fail: 366829c91e99STejun Heo if (pool) 366929c91e99STejun Heo put_unbound_pool(pool); 367029c91e99STejun Heo return NULL; 367129c91e99STejun Heo } 367229c91e99STejun Heo 36738864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36748864b4e5STejun Heo { 36758864b4e5STejun Heo kmem_cache_free(pwq_cache, 36768864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36778864b4e5STejun Heo } 36788864b4e5STejun Heo 36798864b4e5STejun Heo /* 36808864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36818864b4e5STejun Heo * and needs to be destroyed. 36828864b4e5STejun Heo */ 36838864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36848864b4e5STejun Heo { 36858864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36868864b4e5STejun Heo unbound_release_work); 36878864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36888864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3689bc0caf09STejun Heo bool is_last; 36908864b4e5STejun Heo 36918864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36928864b4e5STejun Heo return; 36938864b4e5STejun Heo 369475ccf595STejun Heo /* 36953c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 369675ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 369775ccf595STejun Heo * and consistent with the linking path. 369875ccf595STejun Heo */ 36993c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 37008864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3701bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 37023c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 37038864b4e5STejun Heo 3704a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 37058864b4e5STejun Heo put_unbound_pool(pool); 3706a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3707a892caccSTejun Heo 37088864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 37098864b4e5STejun Heo 37108864b4e5STejun Heo /* 37118864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 37128864b4e5STejun Heo * is gonna access it anymore. Free it. 37138864b4e5STejun Heo */ 37146029a918STejun Heo if (is_last) { 37156029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 37168864b4e5STejun Heo kfree(wq); 37178864b4e5STejun Heo } 37186029a918STejun Heo } 37198864b4e5STejun Heo 37200fbd95aaSTejun Heo /** 3721699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37220fbd95aaSTejun Heo * @pwq: target pool_workqueue 37230fbd95aaSTejun Heo * 3724699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3725699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3726699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37270fbd95aaSTejun Heo */ 3728699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37290fbd95aaSTejun Heo { 3730699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3731699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3732699ce097STejun Heo 3733699ce097STejun Heo /* for @wq->saved_max_active */ 3734a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3735699ce097STejun Heo 3736699ce097STejun Heo /* fast exit for non-freezable wqs */ 3737699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3738699ce097STejun Heo return; 3739699ce097STejun Heo 3740a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3741699ce097STejun Heo 3742699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3743699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37440fbd95aaSTejun Heo 37450fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 37460fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 37470fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3748951a078aSLai Jiangshan 3749951a078aSLai Jiangshan /* 3750951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3751951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3752951a078aSLai Jiangshan */ 3753951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3754699ce097STejun Heo } else { 3755699ce097STejun Heo pwq->max_active = 0; 3756699ce097STejun Heo } 3757699ce097STejun Heo 3758a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 37590fbd95aaSTejun Heo } 37600fbd95aaSTejun Heo 3761e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3762f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3763f147f29eSTejun Heo struct worker_pool *pool) 3764d2c1d404STejun Heo { 3765d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3766d2c1d404STejun Heo 3767e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3768e50aba9aSTejun Heo 3769d2c1d404STejun Heo pwq->pool = pool; 3770d2c1d404STejun Heo pwq->wq = wq; 3771d2c1d404STejun Heo pwq->flush_color = -1; 37728864b4e5STejun Heo pwq->refcnt = 1; 3773d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 37741befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3775d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37768864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3777f147f29eSTejun Heo } 3778d2c1d404STejun Heo 3779f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37801befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3781f147f29eSTejun Heo { 3782f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3783f147f29eSTejun Heo 3784f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 378575ccf595STejun Heo 37861befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37871befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37881befcf30STejun Heo return; 37891befcf30STejun Heo 3790983ca25eSTejun Heo /* 3791983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 37923c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3793983ca25eSTejun Heo */ 379475ccf595STejun Heo pwq->work_color = wq->work_color; 3795983ca25eSTejun Heo 3796983ca25eSTejun Heo /* sync max_active to the current setting */ 3797983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3798983ca25eSTejun Heo 3799983ca25eSTejun Heo /* link in @pwq */ 38009e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3801df2d5ae4STejun Heo } 38026029a918STejun Heo 3803f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3804f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3805f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3806f147f29eSTejun Heo { 3807f147f29eSTejun Heo struct worker_pool *pool; 3808f147f29eSTejun Heo struct pool_workqueue *pwq; 3809f147f29eSTejun Heo 3810f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3811f147f29eSTejun Heo 3812f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3813f147f29eSTejun Heo if (!pool) 3814f147f29eSTejun Heo return NULL; 3815f147f29eSTejun Heo 3816e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3817f147f29eSTejun Heo if (!pwq) { 3818f147f29eSTejun Heo put_unbound_pool(pool); 3819f147f29eSTejun Heo return NULL; 3820f147f29eSTejun Heo } 3821f147f29eSTejun Heo 3822f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3823f147f29eSTejun Heo return pwq; 3824d2c1d404STejun Heo } 3825d2c1d404STejun Heo 38264c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 38274c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 38284c16bd32STejun Heo { 38294c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38304c16bd32STejun Heo 38314c16bd32STejun Heo if (pwq) { 38324c16bd32STejun Heo put_unbound_pool(pwq->pool); 3833cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 38344c16bd32STejun Heo } 38354c16bd32STejun Heo } 38364c16bd32STejun Heo 38374c16bd32STejun Heo /** 38384c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 38394c16bd32STejun Heo * @attrs: the wq_attrs of interest 38404c16bd32STejun Heo * @node: the target NUMA node 38414c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38424c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38434c16bd32STejun Heo * 38444c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38454c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3846d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38474c16bd32STejun Heo * 38484c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38494c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38504c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38514c16bd32STejun Heo * @attrs->cpumask. 38524c16bd32STejun Heo * 38534c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38544c16bd32STejun Heo * stable. 3855d185af30SYacine Belkadi * 3856d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3857d185af30SYacine Belkadi * %false if equal. 38584c16bd32STejun Heo */ 38594c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38604c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38614c16bd32STejun Heo { 3862d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38634c16bd32STejun Heo goto use_dfl; 38644c16bd32STejun Heo 38654c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38664c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38674c16bd32STejun Heo if (cpu_going_down >= 0) 38684c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38694c16bd32STejun Heo 38704c16bd32STejun Heo if (cpumask_empty(cpumask)) 38714c16bd32STejun Heo goto use_dfl; 38724c16bd32STejun Heo 38734c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38744c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38754c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 38764c16bd32STejun Heo 38774c16bd32STejun Heo use_dfl: 38784c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38794c16bd32STejun Heo return false; 38804c16bd32STejun Heo } 38814c16bd32STejun Heo 38821befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38831befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38841befcf30STejun Heo int node, 38851befcf30STejun Heo struct pool_workqueue *pwq) 38861befcf30STejun Heo { 38871befcf30STejun Heo struct pool_workqueue *old_pwq; 38881befcf30STejun Heo 38891befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38901befcf30STejun Heo 38911befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 38921befcf30STejun Heo link_pwq(pwq); 38931befcf30STejun Heo 38941befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 38951befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 38961befcf30STejun Heo return old_pwq; 38971befcf30STejun Heo } 38981befcf30STejun Heo 38999e8cd2f5STejun Heo /** 39009e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 39019e8cd2f5STejun Heo * @wq: the target workqueue 39029e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 39039e8cd2f5STejun Heo * 39044c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 39054c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 39064c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 39074c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 39084c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 39094c16bd32STejun Heo * back-to-back will stay on its current pwq. 39109e8cd2f5STejun Heo * 3911d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3912d185af30SYacine Belkadi * 3913d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 39149e8cd2f5STejun Heo */ 39159e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 39169e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 39179e8cd2f5STejun Heo { 39184c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 39194c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3920f147f29eSTejun Heo int node, ret; 39219e8cd2f5STejun Heo 39228719dceaSTejun Heo /* only unbound workqueues can change attributes */ 39239e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 39249e8cd2f5STejun Heo return -EINVAL; 39259e8cd2f5STejun Heo 39268719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 39278719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 39288719dceaSTejun Heo return -EINVAL; 39298719dceaSTejun Heo 39304c16bd32STejun Heo pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); 393113e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39324c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39334c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 393413e2e556STejun Heo goto enomem; 393513e2e556STejun Heo 39364c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 393713e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 393813e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 393913e2e556STejun Heo 39404c16bd32STejun Heo /* 39414c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 39424c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 39434c16bd32STejun Heo * pools. 39444c16bd32STejun Heo */ 39454c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 39469e8cd2f5STejun Heo 39474c16bd32STejun Heo /* 39484c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 39494c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 39504c16bd32STejun Heo * pwqs accordingly. 39514c16bd32STejun Heo */ 39524c16bd32STejun Heo get_online_cpus(); 39534c16bd32STejun Heo 39544c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 39554c16bd32STejun Heo 39564c16bd32STejun Heo /* 39574c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 39584c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 39594c16bd32STejun Heo * it even if we don't use it immediately. 39604c16bd32STejun Heo */ 39614c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39624c16bd32STejun Heo if (!dfl_pwq) 39634c16bd32STejun Heo goto enomem_pwq; 39644c16bd32STejun Heo 39654c16bd32STejun Heo for_each_node(node) { 39664c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 39674c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39684c16bd32STejun Heo if (!pwq_tbl[node]) 39694c16bd32STejun Heo goto enomem_pwq; 39704c16bd32STejun Heo } else { 39714c16bd32STejun Heo dfl_pwq->refcnt++; 39724c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 39734c16bd32STejun Heo } 39744c16bd32STejun Heo } 39754c16bd32STejun Heo 39764c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39774c16bd32STejun Heo 39784c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3979f147f29eSTejun Heo mutex_lock(&wq->mutex); 3980a892caccSTejun Heo 3981f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 39824c16bd32STejun Heo 39834c16bd32STejun Heo /* save the previous pwq and install the new one */ 3984f147f29eSTejun Heo for_each_node(node) 39854c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 39864c16bd32STejun Heo 39874c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 39884c16bd32STejun Heo link_pwq(dfl_pwq); 39894c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3990f147f29eSTejun Heo 3991f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3992f147f29eSTejun Heo 39934c16bd32STejun Heo /* put the old pwqs */ 39944c16bd32STejun Heo for_each_node(node) 39954c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 39964c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 39974c16bd32STejun Heo 39984c16bd32STejun Heo put_online_cpus(); 39994862125bSTejun Heo ret = 0; 40004862125bSTejun Heo /* fall through */ 40014862125bSTejun Heo out_free: 40024c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 40034862125bSTejun Heo free_workqueue_attrs(new_attrs); 40044c16bd32STejun Heo kfree(pwq_tbl); 40054862125bSTejun Heo return ret; 400613e2e556STejun Heo 40074c16bd32STejun Heo enomem_pwq: 40084c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 40094c16bd32STejun Heo for_each_node(node) 40104c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 40114c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 40124c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 40134c16bd32STejun Heo put_online_cpus(); 401413e2e556STejun Heo enomem: 40154862125bSTejun Heo ret = -ENOMEM; 40164862125bSTejun Heo goto out_free; 40179e8cd2f5STejun Heo } 40189e8cd2f5STejun Heo 40194c16bd32STejun Heo /** 40204c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 40214c16bd32STejun Heo * @wq: the target workqueue 40224c16bd32STejun Heo * @cpu: the CPU coming up or going down 40234c16bd32STejun Heo * @online: whether @cpu is coming up or going down 40244c16bd32STejun Heo * 40254c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 40264c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 40274c16bd32STejun Heo * @wq accordingly. 40284c16bd32STejun Heo * 40294c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 40304c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 40314c16bd32STejun Heo * correct. 40324c16bd32STejun Heo * 40334c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 40344c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 40354c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 40364c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 40374c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 40384c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 40394c16bd32STejun Heo * CPU_DOWN_PREPARE. 40404c16bd32STejun Heo */ 40414c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 40424c16bd32STejun Heo bool online) 40434c16bd32STejun Heo { 40444c16bd32STejun Heo int node = cpu_to_node(cpu); 40454c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 40464c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 40474c16bd32STejun Heo struct workqueue_attrs *target_attrs; 40484c16bd32STejun Heo cpumask_t *cpumask; 40494c16bd32STejun Heo 40504c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 40514c16bd32STejun Heo 40524c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 40534c16bd32STejun Heo return; 40544c16bd32STejun Heo 40554c16bd32STejun Heo /* 40564c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 40574c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 40584c16bd32STejun Heo * CPU hotplug exclusion. 40594c16bd32STejun Heo */ 40604c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 40614c16bd32STejun Heo cpumask = target_attrs->cpumask; 40624c16bd32STejun Heo 40634c16bd32STejun Heo mutex_lock(&wq->mutex); 4064d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 4065d55262c4STejun Heo goto out_unlock; 40664c16bd32STejun Heo 40674c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 40684c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 40694c16bd32STejun Heo 40704c16bd32STejun Heo /* 40714c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 40724c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 40734c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 40744c16bd32STejun Heo * wq's, the default pwq should be used. If @pwq is already the 40754c16bd32STejun Heo * default one, nothing to do; otherwise, install the default one. 40764c16bd32STejun Heo */ 40774c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 40784c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 40794c16bd32STejun Heo goto out_unlock; 40804c16bd32STejun Heo } else { 40814c16bd32STejun Heo if (pwq == wq->dfl_pwq) 40824c16bd32STejun Heo goto out_unlock; 40834c16bd32STejun Heo else 40844c16bd32STejun Heo goto use_dfl_pwq; 40854c16bd32STejun Heo } 40864c16bd32STejun Heo 40874c16bd32STejun Heo mutex_unlock(&wq->mutex); 40884c16bd32STejun Heo 40894c16bd32STejun Heo /* create a new pwq */ 40904c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 40914c16bd32STejun Heo if (!pwq) { 40924c16bd32STejun Heo pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 40934c16bd32STejun Heo wq->name); 40944c16bd32STejun Heo goto out_unlock; 40954c16bd32STejun Heo } 40964c16bd32STejun Heo 40974c16bd32STejun Heo /* 40984c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 40994c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 41004c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 41014c16bd32STejun Heo * inbetween. 41024c16bd32STejun Heo */ 41034c16bd32STejun Heo mutex_lock(&wq->mutex); 41044c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41054c16bd32STejun Heo goto out_unlock; 41064c16bd32STejun Heo 41074c16bd32STejun Heo use_dfl_pwq: 41084c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 41094c16bd32STejun Heo get_pwq(wq->dfl_pwq); 41104c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 41114c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 41124c16bd32STejun Heo out_unlock: 41134c16bd32STejun Heo mutex_unlock(&wq->mutex); 41144c16bd32STejun Heo put_pwq_unlocked(old_pwq); 41154c16bd32STejun Heo } 41164c16bd32STejun Heo 411730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 41181da177e4SLinus Torvalds { 411949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 41208a2b7538STejun Heo int cpu, ret; 4121e1d8aa9fSFrederic Weisbecker 412230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4123420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4124420c0ddbSTejun Heo if (!wq->cpu_pwqs) 412530cdf249STejun Heo return -ENOMEM; 412630cdf249STejun Heo 412730cdf249STejun Heo for_each_possible_cpu(cpu) { 41287fb98ea7STejun Heo struct pool_workqueue *pwq = 41297fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 41307a62c2c8STejun Heo struct worker_pool *cpu_pools = 4131f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 413230cdf249STejun Heo 4133f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4134f147f29eSTejun Heo 4135f147f29eSTejun Heo mutex_lock(&wq->mutex); 41361befcf30STejun Heo link_pwq(pwq); 4137f147f29eSTejun Heo mutex_unlock(&wq->mutex); 413830cdf249STejun Heo } 413930cdf249STejun Heo return 0; 41408a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 41418a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 41428a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 41438a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 41448a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 41458a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 41468a2b7538STejun Heo return ret; 41479e8cd2f5STejun Heo } else { 41489e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 41499e8cd2f5STejun Heo } 41500f900049STejun Heo } 41510f900049STejun Heo 4152f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4153f3421797STejun Heo const char *name) 4154b71ab8c2STejun Heo { 4155f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4156f3421797STejun Heo 4157f3421797STejun Heo if (max_active < 1 || max_active > lim) 4158044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4159f3421797STejun Heo max_active, name, 1, lim); 4160b71ab8c2STejun Heo 4161f3421797STejun Heo return clamp_val(max_active, 1, lim); 4162b71ab8c2STejun Heo } 4163b71ab8c2STejun Heo 4164b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 416597e37d7bSTejun Heo unsigned int flags, 41661e19ffc6STejun Heo int max_active, 4167eb13ba87SJohannes Berg struct lock_class_key *key, 4168b196be89STejun Heo const char *lock_name, ...) 41693af24433SOleg Nesterov { 4170df2d5ae4STejun Heo size_t tbl_size = 0; 4171ecf6881fSTejun Heo va_list args; 41723af24433SOleg Nesterov struct workqueue_struct *wq; 417349e3cf44STejun Heo struct pool_workqueue *pwq; 4174b196be89STejun Heo 4175cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4176cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4177cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4178cee22a15SViresh Kumar 4179ecf6881fSTejun Heo /* allocate wq and format name */ 4180df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4181df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 4182df2d5ae4STejun Heo 4183df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4184b196be89STejun Heo if (!wq) 4185d2c1d404STejun Heo return NULL; 4186b196be89STejun Heo 41876029a918STejun Heo if (flags & WQ_UNBOUND) { 41886029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 41896029a918STejun Heo if (!wq->unbound_attrs) 41906029a918STejun Heo goto err_free_wq; 41916029a918STejun Heo } 41926029a918STejun Heo 4193ecf6881fSTejun Heo va_start(args, lock_name); 4194ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4195b196be89STejun Heo va_end(args); 41963af24433SOleg Nesterov 4197d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4198b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 41993af24433SOleg Nesterov 4200b196be89STejun Heo /* init wq */ 420197e37d7bSTejun Heo wq->flags = flags; 4202a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 42033c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4204112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 420530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 420673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 420773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4208493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 42093af24433SOleg Nesterov 4210eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4211cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 42123af24433SOleg Nesterov 421330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4214d2c1d404STejun Heo goto err_free_wq; 42151537663fSTejun Heo 4216493008a8STejun Heo /* 4217493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4218493008a8STejun Heo * have a rescuer to guarantee forward progress. 4219493008a8STejun Heo */ 4220493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4221e22bee78STejun Heo struct worker *rescuer; 4222e22bee78STejun Heo 4223d2c1d404STejun Heo rescuer = alloc_worker(); 4224e22bee78STejun Heo if (!rescuer) 4225d2c1d404STejun Heo goto err_destroy; 4226e22bee78STejun Heo 4227111c225aSTejun Heo rescuer->rescue_wq = wq; 4228111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4229b196be89STejun Heo wq->name); 4230d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4231d2c1d404STejun Heo kfree(rescuer); 4232d2c1d404STejun Heo goto err_destroy; 4233d2c1d404STejun Heo } 4234e22bee78STejun Heo 4235d2c1d404STejun Heo wq->rescuer = rescuer; 423614a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4237e22bee78STejun Heo wake_up_process(rescuer->task); 42383af24433SOleg Nesterov } 42391537663fSTejun Heo 4240226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4241226223abSTejun Heo goto err_destroy; 4242226223abSTejun Heo 42433af24433SOleg Nesterov /* 424468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 424568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 424668e13a67SLai Jiangshan * list. 42473af24433SOleg Nesterov */ 424868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4249a0a1a5fdSTejun Heo 4250a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 425149e3cf44STejun Heo for_each_pwq(pwq, wq) 4252699ce097STejun Heo pwq_adjust_max_active(pwq); 4253a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4254a0a1a5fdSTejun Heo 42553af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4256a0a1a5fdSTejun Heo 425768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42583af24433SOleg Nesterov 42593af24433SOleg Nesterov return wq; 4260d2c1d404STejun Heo 4261d2c1d404STejun Heo err_free_wq: 42626029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 42634690c4abSTejun Heo kfree(wq); 4264d2c1d404STejun Heo return NULL; 4265d2c1d404STejun Heo err_destroy: 4266d2c1d404STejun Heo destroy_workqueue(wq); 42674690c4abSTejun Heo return NULL; 42681da177e4SLinus Torvalds } 4269d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 42701da177e4SLinus Torvalds 42713af24433SOleg Nesterov /** 42723af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 42733af24433SOleg Nesterov * @wq: target workqueue 42743af24433SOleg Nesterov * 42753af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 42763af24433SOleg Nesterov */ 42773af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 42783af24433SOleg Nesterov { 427949e3cf44STejun Heo struct pool_workqueue *pwq; 42804c16bd32STejun Heo int node; 42813af24433SOleg Nesterov 42829c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 42839c5a2ba7STejun Heo drain_workqueue(wq); 4284c8efcc25STejun Heo 42856183c009STejun Heo /* sanity checks */ 4286b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 428749e3cf44STejun Heo for_each_pwq(pwq, wq) { 42886183c009STejun Heo int i; 42896183c009STejun Heo 429076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 429176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4292b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42936183c009STejun Heo return; 429476af4d93STejun Heo } 429576af4d93STejun Heo } 429676af4d93STejun Heo 42975c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 42988864b4e5STejun Heo WARN_ON(pwq->nr_active) || 429976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4300b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 43016183c009STejun Heo return; 43026183c009STejun Heo } 430376af4d93STejun Heo } 4304b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 43056183c009STejun Heo 4306a0a1a5fdSTejun Heo /* 4307a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4308a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4309a0a1a5fdSTejun Heo */ 431068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4311d2c1d404STejun Heo list_del_init(&wq->list); 431268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43133af24433SOleg Nesterov 4314226223abSTejun Heo workqueue_sysfs_unregister(wq); 4315226223abSTejun Heo 4316493008a8STejun Heo if (wq->rescuer) { 4317e22bee78STejun Heo kthread_stop(wq->rescuer->task); 43188d9df9f0SXiaotian Feng kfree(wq->rescuer); 4319493008a8STejun Heo wq->rescuer = NULL; 4320e22bee78STejun Heo } 4321e22bee78STejun Heo 43228864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 432329c91e99STejun Heo /* 43248864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 43258864b4e5STejun Heo * free the pwqs and wq. 432629c91e99STejun Heo */ 43278864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 43288864b4e5STejun Heo kfree(wq); 43298864b4e5STejun Heo } else { 43308864b4e5STejun Heo /* 43318864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 43324c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 43334c16bd32STejun Heo * @wq will be freed when the last pwq is released. 43348864b4e5STejun Heo */ 43354c16bd32STejun Heo for_each_node(node) { 43364c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 43374c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 43384c16bd32STejun Heo put_pwq_unlocked(pwq); 43394c16bd32STejun Heo } 43404c16bd32STejun Heo 43414c16bd32STejun Heo /* 43424c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 43434c16bd32STejun Heo * put. Don't access it afterwards. 43444c16bd32STejun Heo */ 43454c16bd32STejun Heo pwq = wq->dfl_pwq; 43464c16bd32STejun Heo wq->dfl_pwq = NULL; 4347dce90d47STejun Heo put_pwq_unlocked(pwq); 434829c91e99STejun Heo } 43493af24433SOleg Nesterov } 43503af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 43513af24433SOleg Nesterov 4352dcd989cbSTejun Heo /** 4353dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4354dcd989cbSTejun Heo * @wq: target workqueue 4355dcd989cbSTejun Heo * @max_active: new max_active value. 4356dcd989cbSTejun Heo * 4357dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4358dcd989cbSTejun Heo * 4359dcd989cbSTejun Heo * CONTEXT: 4360dcd989cbSTejun Heo * Don't call from IRQ context. 4361dcd989cbSTejun Heo */ 4362dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4363dcd989cbSTejun Heo { 436449e3cf44STejun Heo struct pool_workqueue *pwq; 4365dcd989cbSTejun Heo 43668719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 43678719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 43688719dceaSTejun Heo return; 43698719dceaSTejun Heo 4370f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4371dcd989cbSTejun Heo 4372a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4373dcd989cbSTejun Heo 4374dcd989cbSTejun Heo wq->saved_max_active = max_active; 4375dcd989cbSTejun Heo 4376699ce097STejun Heo for_each_pwq(pwq, wq) 4377699ce097STejun Heo pwq_adjust_max_active(pwq); 4378dcd989cbSTejun Heo 4379a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4380dcd989cbSTejun Heo } 4381dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4382dcd989cbSTejun Heo 4383dcd989cbSTejun Heo /** 4384e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4385e6267616STejun Heo * 4386e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4387e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4388d185af30SYacine Belkadi * 4389d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4390e6267616STejun Heo */ 4391e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4392e6267616STejun Heo { 4393e6267616STejun Heo struct worker *worker = current_wq_worker(); 4394e6267616STejun Heo 43956a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4396e6267616STejun Heo } 4397e6267616STejun Heo 4398e6267616STejun Heo /** 4399dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4400dcd989cbSTejun Heo * @cpu: CPU in question 4401dcd989cbSTejun Heo * @wq: target workqueue 4402dcd989cbSTejun Heo * 4403dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4404dcd989cbSTejun Heo * no synchronization around this function and the test result is 4405dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4406dcd989cbSTejun Heo * 4407d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4408d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4409d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4410d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4411d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4412d3251859STejun Heo * 4413d185af30SYacine Belkadi * Return: 4414dcd989cbSTejun Heo * %true if congested, %false otherwise. 4415dcd989cbSTejun Heo */ 4416d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4417dcd989cbSTejun Heo { 44187fb98ea7STejun Heo struct pool_workqueue *pwq; 441976af4d93STejun Heo bool ret; 442076af4d93STejun Heo 442188109453SLai Jiangshan rcu_read_lock_sched(); 44227fb98ea7STejun Heo 4423d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4424d3251859STejun Heo cpu = smp_processor_id(); 4425d3251859STejun Heo 44267fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 44277fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 44287fb98ea7STejun Heo else 4429df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4430dcd989cbSTejun Heo 443176af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 443288109453SLai Jiangshan rcu_read_unlock_sched(); 443376af4d93STejun Heo 443476af4d93STejun Heo return ret; 4435dcd989cbSTejun Heo } 4436dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4437dcd989cbSTejun Heo 4438dcd989cbSTejun Heo /** 4439dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4440dcd989cbSTejun Heo * @work: the work to be tested 4441dcd989cbSTejun Heo * 4442dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4443dcd989cbSTejun Heo * synchronization around this function and the test result is 4444dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4445dcd989cbSTejun Heo * 4446d185af30SYacine Belkadi * Return: 4447dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4448dcd989cbSTejun Heo */ 4449dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4450dcd989cbSTejun Heo { 4451fa1b54e6STejun Heo struct worker_pool *pool; 4452dcd989cbSTejun Heo unsigned long flags; 4453dcd989cbSTejun Heo unsigned int ret = 0; 4454dcd989cbSTejun Heo 4455dcd989cbSTejun Heo if (work_pending(work)) 4456dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4457038366c5SLai Jiangshan 4458fa1b54e6STejun Heo local_irq_save(flags); 4459fa1b54e6STejun Heo pool = get_work_pool(work); 4460038366c5SLai Jiangshan if (pool) { 4461fa1b54e6STejun Heo spin_lock(&pool->lock); 4462c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4463dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4464fa1b54e6STejun Heo spin_unlock(&pool->lock); 4465038366c5SLai Jiangshan } 4466fa1b54e6STejun Heo local_irq_restore(flags); 4467dcd989cbSTejun Heo 4468dcd989cbSTejun Heo return ret; 4469dcd989cbSTejun Heo } 4470dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4471dcd989cbSTejun Heo 44723d1cb205STejun Heo /** 44733d1cb205STejun Heo * set_worker_desc - set description for the current work item 44743d1cb205STejun Heo * @fmt: printf-style format string 44753d1cb205STejun Heo * @...: arguments for the format string 44763d1cb205STejun Heo * 44773d1cb205STejun Heo * This function can be called by a running work function to describe what 44783d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 44793d1cb205STejun Heo * information will be printed out together to help debugging. The 44803d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 44813d1cb205STejun Heo */ 44823d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 44833d1cb205STejun Heo { 44843d1cb205STejun Heo struct worker *worker = current_wq_worker(); 44853d1cb205STejun Heo va_list args; 44863d1cb205STejun Heo 44873d1cb205STejun Heo if (worker) { 44883d1cb205STejun Heo va_start(args, fmt); 44893d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 44903d1cb205STejun Heo va_end(args); 44913d1cb205STejun Heo worker->desc_valid = true; 44923d1cb205STejun Heo } 44933d1cb205STejun Heo } 44943d1cb205STejun Heo 44953d1cb205STejun Heo /** 44963d1cb205STejun Heo * print_worker_info - print out worker information and description 44973d1cb205STejun Heo * @log_lvl: the log level to use when printing 44983d1cb205STejun Heo * @task: target task 44993d1cb205STejun Heo * 45003d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 45013d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 45023d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 45033d1cb205STejun Heo * 45043d1cb205STejun Heo * This function can be safely called on any task as long as the 45053d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 45063d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 45073d1cb205STejun Heo */ 45083d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 45093d1cb205STejun Heo { 45103d1cb205STejun Heo work_func_t *fn = NULL; 45113d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 45123d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 45133d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 45143d1cb205STejun Heo struct workqueue_struct *wq = NULL; 45153d1cb205STejun Heo bool desc_valid = false; 45163d1cb205STejun Heo struct worker *worker; 45173d1cb205STejun Heo 45183d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 45193d1cb205STejun Heo return; 45203d1cb205STejun Heo 45213d1cb205STejun Heo /* 45223d1cb205STejun Heo * This function is called without any synchronization and @task 45233d1cb205STejun Heo * could be in any state. Be careful with dereferences. 45243d1cb205STejun Heo */ 45253d1cb205STejun Heo worker = probe_kthread_data(task); 45263d1cb205STejun Heo 45273d1cb205STejun Heo /* 45283d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 45293d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 45303d1cb205STejun Heo */ 45313d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 45323d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 45333d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 45343d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 45353d1cb205STejun Heo 45363d1cb205STejun Heo /* copy worker description */ 45373d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 45383d1cb205STejun Heo if (desc_valid) 45393d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 45403d1cb205STejun Heo 45413d1cb205STejun Heo if (fn || name[0] || desc[0]) { 45423d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 45433d1cb205STejun Heo if (desc[0]) 45443d1cb205STejun Heo pr_cont(" (%s)", desc); 45453d1cb205STejun Heo pr_cont("\n"); 45463d1cb205STejun Heo } 45473d1cb205STejun Heo } 45483d1cb205STejun Heo 4549db7bccf4STejun Heo /* 4550db7bccf4STejun Heo * CPU hotplug. 4551db7bccf4STejun Heo * 4552e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4553112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4554706026c2STejun Heo * pool which make migrating pending and scheduled works very 4555e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 455694cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4557e22bee78STejun Heo * blocked draining impractical. 4558e22bee78STejun Heo * 455924647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4560628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4561628c78e7STejun Heo * cpu comes back online. 4562db7bccf4STejun Heo */ 4563db7bccf4STejun Heo 4564706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4565db7bccf4STejun Heo { 456638db41d9STejun Heo int cpu = smp_processor_id(); 45674ce62e9eSTejun Heo struct worker_pool *pool; 4568db7bccf4STejun Heo struct worker *worker; 4569a9ab775bSTejun Heo int wi; 4570db7bccf4STejun Heo 4571f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45726183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4573db7bccf4STejun Heo 4574bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 457594cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4576e22bee78STejun Heo 4577f2d5a0eeSTejun Heo /* 4578bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 457994cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 458094cf58bbSTejun Heo * except for the ones which are still executing works from 458194cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 458294cf58bbSTejun Heo * this, they may become diasporas. 4583f2d5a0eeSTejun Heo */ 4584a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4585403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4586db7bccf4STejun Heo 458724647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4588f2d5a0eeSTejun Heo 458994cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4590bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 4591e22bee78STejun Heo 4592e22bee78STejun Heo /* 4593eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4594eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4595eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4596eb283428SLai Jiangshan * from other cpus. 4597628c78e7STejun Heo */ 4598628c78e7STejun Heo schedule(); 4599628c78e7STejun Heo 4600628c78e7STejun Heo /* 4601eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4602eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4603eb283428SLai Jiangshan * and keep_working() are always true as long as the 4604eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4605eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4606eb283428SLai Jiangshan * are served by workers tied to the pool. 4607e22bee78STejun Heo */ 4608e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4609eb283428SLai Jiangshan 4610eb283428SLai Jiangshan /* 4611eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4612eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4613eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4614eb283428SLai Jiangshan */ 4615eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4616eb283428SLai Jiangshan wake_up_worker(pool); 4617eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4618eb283428SLai Jiangshan } 4619db7bccf4STejun Heo } 4620db7bccf4STejun Heo 4621bd7c089eSTejun Heo /** 4622bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4623bd7c089eSTejun Heo * @pool: pool of interest 4624bd7c089eSTejun Heo * 4625a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4626bd7c089eSTejun Heo */ 4627bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4628bd7c089eSTejun Heo { 4629a9ab775bSTejun Heo struct worker *worker; 4630a9ab775bSTejun Heo int wi; 4631bd7c089eSTejun Heo 4632bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4633bd7c089eSTejun Heo 4634bd7c089eSTejun Heo /* 4635a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4636a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4637a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4638a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4639a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4640bd7c089eSTejun Heo */ 4641a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4642a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4643a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4644a9ab775bSTejun Heo 4645a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4646a9ab775bSTejun Heo 4647a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4648a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4649a9ab775bSTejun Heo 4650a9ab775bSTejun Heo /* 4651a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4652a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4653a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4654a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4655a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4656a9ab775bSTejun Heo * be bound before @pool->lock is released. 4657a9ab775bSTejun Heo */ 4658a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4659bd7c089eSTejun Heo wake_up_process(worker->task); 4660bd7c089eSTejun Heo 4661bd7c089eSTejun Heo /* 4662a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4663a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4664a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4665a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4666a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4667a9ab775bSTejun Heo * concurrency management. Note that when or whether 4668a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4669a9ab775bSTejun Heo * 4670a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4671a9ab775bSTejun Heo * tested without holding any lock in 4672a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4673a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4674a9ab775bSTejun Heo * management operations. 4675bd7c089eSTejun Heo */ 4676a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4677a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4678a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4679a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4680bd7c089eSTejun Heo } 4681a9ab775bSTejun Heo 4682a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4683bd7c089eSTejun Heo } 4684bd7c089eSTejun Heo 46857dbc725eSTejun Heo /** 46867dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 46877dbc725eSTejun Heo * @pool: unbound pool of interest 46887dbc725eSTejun Heo * @cpu: the CPU which is coming up 46897dbc725eSTejun Heo * 46907dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 46917dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 46927dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 46937dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 46947dbc725eSTejun Heo */ 46957dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 46967dbc725eSTejun Heo { 46977dbc725eSTejun Heo static cpumask_t cpumask; 46987dbc725eSTejun Heo struct worker *worker; 46997dbc725eSTejun Heo int wi; 47007dbc725eSTejun Heo 47017dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 47027dbc725eSTejun Heo 47037dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 47047dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 47057dbc725eSTejun Heo return; 47067dbc725eSTejun Heo 47077dbc725eSTejun Heo /* is @cpu the only online CPU? */ 47087dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 47097dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 47107dbc725eSTejun Heo return; 47117dbc725eSTejun Heo 47127dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 47137dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 47147dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 47157dbc725eSTejun Heo pool->attrs->cpumask) < 0); 47167dbc725eSTejun Heo } 47177dbc725eSTejun Heo 47188db25e78STejun Heo /* 47198db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 47208db25e78STejun Heo * This will be registered high priority CPU notifier. 47218db25e78STejun Heo */ 47220db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 47231da177e4SLinus Torvalds unsigned long action, 47241da177e4SLinus Torvalds void *hcpu) 47251da177e4SLinus Torvalds { 4726d84ff051STejun Heo int cpu = (unsigned long)hcpu; 47274ce62e9eSTejun Heo struct worker_pool *pool; 47284c16bd32STejun Heo struct workqueue_struct *wq; 47297dbc725eSTejun Heo int pi; 47301da177e4SLinus Torvalds 47318db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 47323af24433SOleg Nesterov case CPU_UP_PREPARE: 4733f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 47343ce63377STejun Heo if (pool->nr_workers) 47353ce63377STejun Heo continue; 4736ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 47373ce63377STejun Heo return NOTIFY_BAD; 47383af24433SOleg Nesterov } 47391da177e4SLinus Torvalds break; 47401da177e4SLinus Torvalds 474165758202STejun Heo case CPU_DOWN_FAILED: 474265758202STejun Heo case CPU_ONLINE: 474368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 47447dbc725eSTejun Heo 47457dbc725eSTejun Heo for_each_pool(pool, pi) { 4746bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 474794cf58bbSTejun Heo 47487dbc725eSTejun Heo if (pool->cpu == cpu) { 4749a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 475024647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4751a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4752a9ab775bSTejun Heo 475394cf58bbSTejun Heo rebind_workers(pool); 47547dbc725eSTejun Heo } else if (pool->cpu < 0) { 47557dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 47567dbc725eSTejun Heo } 475794cf58bbSTejun Heo 4758bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 475994cf58bbSTejun Heo } 47607dbc725eSTejun Heo 47614c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47624c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47634c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 47644c16bd32STejun Heo 476568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47668db25e78STejun Heo break; 476765758202STejun Heo } 476865758202STejun Heo return NOTIFY_OK; 476965758202STejun Heo } 477065758202STejun Heo 477165758202STejun Heo /* 477265758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 477365758202STejun Heo * This will be registered as low priority CPU notifier. 477465758202STejun Heo */ 47750db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 477665758202STejun Heo unsigned long action, 477765758202STejun Heo void *hcpu) 477865758202STejun Heo { 4779d84ff051STejun Heo int cpu = (unsigned long)hcpu; 47808db25e78STejun Heo struct work_struct unbind_work; 47814c16bd32STejun Heo struct workqueue_struct *wq; 47828db25e78STejun Heo 478365758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 478465758202STejun Heo case CPU_DOWN_PREPARE: 47854c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4786706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 47877635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 47884c16bd32STejun Heo 47894c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47904c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 47914c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47924c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 47934c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 47944c16bd32STejun Heo 47954c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 47968db25e78STejun Heo flush_work(&unbind_work); 47978db25e78STejun Heo break; 479865758202STejun Heo } 479965758202STejun Heo return NOTIFY_OK; 480065758202STejun Heo } 480165758202STejun Heo 48022d3854a3SRusty Russell #ifdef CONFIG_SMP 48038ccad40dSRusty Russell 48042d3854a3SRusty Russell struct work_for_cpu { 4805ed48ece2STejun Heo struct work_struct work; 48062d3854a3SRusty Russell long (*fn)(void *); 48072d3854a3SRusty Russell void *arg; 48082d3854a3SRusty Russell long ret; 48092d3854a3SRusty Russell }; 48102d3854a3SRusty Russell 4811ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 48122d3854a3SRusty Russell { 4813ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4814ed48ece2STejun Heo 48152d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 48162d3854a3SRusty Russell } 48172d3854a3SRusty Russell 48182d3854a3SRusty Russell /** 48192d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 48202d3854a3SRusty Russell * @cpu: the cpu to run on 48212d3854a3SRusty Russell * @fn: the function to run 48222d3854a3SRusty Russell * @arg: the function arg 48232d3854a3SRusty Russell * 482431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 48256b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4826d185af30SYacine Belkadi * 4827d185af30SYacine Belkadi * Return: The value @fn returns. 48282d3854a3SRusty Russell */ 4829d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 48302d3854a3SRusty Russell { 4831ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 48322d3854a3SRusty Russell 4833ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4834ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4835c2fda509SLai Jiangshan 4836c2fda509SLai Jiangshan /* 4837c2fda509SLai Jiangshan * The work item is on-stack and can't lead to deadlock through 4838c2fda509SLai Jiangshan * flushing. Use __flush_work() to avoid spurious lockdep warnings 4839c2fda509SLai Jiangshan * when work_on_cpu()s are nested. 4840c2fda509SLai Jiangshan */ 4841c2fda509SLai Jiangshan __flush_work(&wfc.work); 4842c2fda509SLai Jiangshan 48432d3854a3SRusty Russell return wfc.ret; 48442d3854a3SRusty Russell } 48452d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 48462d3854a3SRusty Russell #endif /* CONFIG_SMP */ 48472d3854a3SRusty Russell 4848a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4849e7577c50SRusty Russell 4850a0a1a5fdSTejun Heo /** 4851a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4852a0a1a5fdSTejun Heo * 485358a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4854c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4855706026c2STejun Heo * pool->worklist. 4856a0a1a5fdSTejun Heo * 4857a0a1a5fdSTejun Heo * CONTEXT: 4858a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4859a0a1a5fdSTejun Heo */ 4860a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4861a0a1a5fdSTejun Heo { 486217116969STejun Heo struct worker_pool *pool; 486324b8a847STejun Heo struct workqueue_struct *wq; 486424b8a847STejun Heo struct pool_workqueue *pwq; 4865611c92a0STejun Heo int pi; 4866a0a1a5fdSTejun Heo 486768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4868a0a1a5fdSTejun Heo 48696183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4870a0a1a5fdSTejun Heo workqueue_freezing = true; 4871a0a1a5fdSTejun Heo 487224b8a847STejun Heo /* set FREEZING */ 4873611c92a0STejun Heo for_each_pool(pool, pi) { 48745bcab335STejun Heo spin_lock_irq(&pool->lock); 487535b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 487635b6bb63STejun Heo pool->flags |= POOL_FREEZING; 48775bcab335STejun Heo spin_unlock_irq(&pool->lock); 48781da177e4SLinus Torvalds } 48798b03ae3cSTejun Heo 488024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4881a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4882699ce097STejun Heo for_each_pwq(pwq, wq) 4883699ce097STejun Heo pwq_adjust_max_active(pwq); 4884a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4885a1056305STejun Heo } 48865bcab335STejun Heo 488768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4888a0a1a5fdSTejun Heo } 4889a0a1a5fdSTejun Heo 4890a0a1a5fdSTejun Heo /** 489158a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4892a0a1a5fdSTejun Heo * 4893a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4894a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4895a0a1a5fdSTejun Heo * 4896a0a1a5fdSTejun Heo * CONTEXT: 489768e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4898a0a1a5fdSTejun Heo * 4899d185af30SYacine Belkadi * Return: 490058a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 490158a69cb4STejun Heo * is complete. 4902a0a1a5fdSTejun Heo */ 4903a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4904a0a1a5fdSTejun Heo { 4905a0a1a5fdSTejun Heo bool busy = false; 490624b8a847STejun Heo struct workqueue_struct *wq; 490724b8a847STejun Heo struct pool_workqueue *pwq; 4908a0a1a5fdSTejun Heo 490968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4910a0a1a5fdSTejun Heo 49116183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4912a0a1a5fdSTejun Heo 491324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 491424b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 491524b8a847STejun Heo continue; 4916a0a1a5fdSTejun Heo /* 4917a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4918a0a1a5fdSTejun Heo * to peek without lock. 4919a0a1a5fdSTejun Heo */ 492088109453SLai Jiangshan rcu_read_lock_sched(); 492124b8a847STejun Heo for_each_pwq(pwq, wq) { 49226183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4923112202d9STejun Heo if (pwq->nr_active) { 4924a0a1a5fdSTejun Heo busy = true; 492588109453SLai Jiangshan rcu_read_unlock_sched(); 4926a0a1a5fdSTejun Heo goto out_unlock; 4927a0a1a5fdSTejun Heo } 4928a0a1a5fdSTejun Heo } 492988109453SLai Jiangshan rcu_read_unlock_sched(); 4930a0a1a5fdSTejun Heo } 4931a0a1a5fdSTejun Heo out_unlock: 493268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4933a0a1a5fdSTejun Heo return busy; 4934a0a1a5fdSTejun Heo } 4935a0a1a5fdSTejun Heo 4936a0a1a5fdSTejun Heo /** 4937a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4938a0a1a5fdSTejun Heo * 4939a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4940706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4941a0a1a5fdSTejun Heo * 4942a0a1a5fdSTejun Heo * CONTEXT: 4943a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4944a0a1a5fdSTejun Heo */ 4945a0a1a5fdSTejun Heo void thaw_workqueues(void) 4946a0a1a5fdSTejun Heo { 494724b8a847STejun Heo struct workqueue_struct *wq; 494824b8a847STejun Heo struct pool_workqueue *pwq; 494924b8a847STejun Heo struct worker_pool *pool; 4950611c92a0STejun Heo int pi; 4951a0a1a5fdSTejun Heo 495268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4953a0a1a5fdSTejun Heo 4954a0a1a5fdSTejun Heo if (!workqueue_freezing) 4955a0a1a5fdSTejun Heo goto out_unlock; 4956a0a1a5fdSTejun Heo 495724b8a847STejun Heo /* clear FREEZING */ 4958611c92a0STejun Heo for_each_pool(pool, pi) { 49595bcab335STejun Heo spin_lock_irq(&pool->lock); 496035b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 496135b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 49625bcab335STejun Heo spin_unlock_irq(&pool->lock); 4963d565ed63STejun Heo } 496424b8a847STejun Heo 496524b8a847STejun Heo /* restore max_active and repopulate worklist */ 496624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4967a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4968699ce097STejun Heo for_each_pwq(pwq, wq) 4969699ce097STejun Heo pwq_adjust_max_active(pwq); 4970a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 497124b8a847STejun Heo } 497224b8a847STejun Heo 4973a0a1a5fdSTejun Heo workqueue_freezing = false; 4974a0a1a5fdSTejun Heo out_unlock: 497568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4976a0a1a5fdSTejun Heo } 4977a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4978a0a1a5fdSTejun Heo 4979bce90380STejun Heo static void __init wq_numa_init(void) 4980bce90380STejun Heo { 4981bce90380STejun Heo cpumask_var_t *tbl; 4982bce90380STejun Heo int node, cpu; 4983bce90380STejun Heo 4984bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4985bce90380STejun Heo for_each_node(node) 4986bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4987bce90380STejun Heo 4988bce90380STejun Heo if (num_possible_nodes() <= 1) 4989bce90380STejun Heo return; 4990bce90380STejun Heo 4991d55262c4STejun Heo if (wq_disable_numa) { 4992d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 4993d55262c4STejun Heo return; 4994d55262c4STejun Heo } 4995d55262c4STejun Heo 49964c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 49974c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 49984c16bd32STejun Heo 4999bce90380STejun Heo /* 5000bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5001bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5002bce90380STejun Heo * fully initialized by now. 5003bce90380STejun Heo */ 5004bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 5005bce90380STejun Heo BUG_ON(!tbl); 5006bce90380STejun Heo 5007bce90380STejun Heo for_each_node(node) 50081be0c25dSTejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 50091be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5010bce90380STejun Heo 5011bce90380STejun Heo for_each_possible_cpu(cpu) { 5012bce90380STejun Heo node = cpu_to_node(cpu); 5013bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5014bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5015bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5016bce90380STejun Heo return; 5017bce90380STejun Heo } 5018bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5019bce90380STejun Heo } 5020bce90380STejun Heo 5021bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5022bce90380STejun Heo wq_numa_enabled = true; 5023bce90380STejun Heo } 5024bce90380STejun Heo 50256ee0578bSSuresh Siddha static int __init init_workqueues(void) 50261da177e4SLinus Torvalds { 50277a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 50287a4e344cSTejun Heo int i, cpu; 5029c34056a3STejun Heo 5030e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5031e904e6c2STejun Heo 5032e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5033e904e6c2STejun Heo 503465758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 5035a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 50368b03ae3cSTejun Heo 5037bce90380STejun Heo wq_numa_init(); 5038bce90380STejun Heo 5039706026c2STejun Heo /* initialize CPU pools */ 504029c91e99STejun Heo for_each_possible_cpu(cpu) { 50414ce62e9eSTejun Heo struct worker_pool *pool; 50428b03ae3cSTejun Heo 50437a4e344cSTejun Heo i = 0; 5044f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 50457a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5046ec22ca5eSTejun Heo pool->cpu = cpu; 50477a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 50487a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5049f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 50507a4e344cSTejun Heo 50519daf9e67STejun Heo /* alloc pool ID */ 505268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 50539daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 505468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 50554ce62e9eSTejun Heo } 50568b03ae3cSTejun Heo } 50578b03ae3cSTejun Heo 5058e22bee78STejun Heo /* create the initial worker */ 505929c91e99STejun Heo for_each_online_cpu(cpu) { 50604ce62e9eSTejun Heo struct worker_pool *pool; 5061e22bee78STejun Heo 5062f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 506324647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5064ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 5065e22bee78STejun Heo } 50664ce62e9eSTejun Heo } 5067e22bee78STejun Heo 50688a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 506929c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 507029c91e99STejun Heo struct workqueue_attrs *attrs; 507129c91e99STejun Heo 507229c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 507329c91e99STejun Heo attrs->nice = std_nice[i]; 507429c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 50758a2b7538STejun Heo 50768a2b7538STejun Heo /* 50778a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 50788a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 50798a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 50808a2b7538STejun Heo */ 50818a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 50828a2b7538STejun Heo attrs->nice = std_nice[i]; 50838a2b7538STejun Heo attrs->no_numa = true; 50848a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 508529c91e99STejun Heo } 508629c91e99STejun Heo 5087d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 50881aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5089d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5090f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5091f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 509224d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 509324d51addSTejun Heo WQ_FREEZABLE, 0); 50940668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 50950668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 50960668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 50970668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 50980668106cSViresh Kumar 0); 50991aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 51000668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 51010668106cSViresh Kumar !system_power_efficient_wq || 51020668106cSViresh Kumar !system_freezable_power_efficient_wq); 51036ee0578bSSuresh Siddha return 0; 51041da177e4SLinus Torvalds } 51056ee0578bSSuresh Siddha early_initcall(init_workqueues); 5106