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 519ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 520ea2e64f2SThomas Gleixner { 521ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 522ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 523ea2e64f2SThomas Gleixner } 524ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 525ea2e64f2SThomas Gleixner 526dc186ad7SThomas Gleixner #else 527dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 528dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 529dc186ad7SThomas Gleixner #endif 530dc186ad7SThomas Gleixner 5314e8b22bdSLi Bin /** 5324e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5334e8b22bdSLi Bin * @pool: the pool pointer of interest 5344e8b22bdSLi Bin * 5354e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5364e8b22bdSLi Bin * successfully, -errno on failure. 5374e8b22bdSLi Bin */ 5389daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5399daf9e67STejun Heo { 5409daf9e67STejun Heo int ret; 5419daf9e67STejun Heo 54268e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5435bcab335STejun Heo 5444e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5454e8b22bdSLi Bin GFP_KERNEL); 546229641a6STejun Heo if (ret >= 0) { 547e68035fbSTejun Heo pool->id = ret; 548229641a6STejun Heo return 0; 549229641a6STejun Heo } 5509daf9e67STejun Heo return ret; 5519daf9e67STejun Heo } 5529daf9e67STejun Heo 55376af4d93STejun Heo /** 554df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 555df2d5ae4STejun Heo * @wq: the target workqueue 556df2d5ae4STejun Heo * @node: the node ID 557df2d5ae4STejun Heo * 558df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 559df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 560df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 561d185af30SYacine Belkadi * 562d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 563df2d5ae4STejun Heo */ 564df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 565df2d5ae4STejun Heo int node) 566df2d5ae4STejun Heo { 567df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 568df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 569df2d5ae4STejun Heo } 570df2d5ae4STejun Heo 57173f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57273f53c4aSTejun Heo { 57373f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 57473f53c4aSTejun Heo } 57573f53c4aSTejun Heo 57673f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 57773f53c4aSTejun Heo { 57873f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 57973f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58073f53c4aSTejun Heo } 58173f53c4aSTejun Heo 58273f53c4aSTejun Heo static int work_next_color(int color) 58373f53c4aSTejun Heo { 58473f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 585a848e3b6SOleg Nesterov } 586a848e3b6SOleg Nesterov 5874594bf15SDavid Howells /* 588112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 589112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5907c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5917a22ad75STejun Heo * 592112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 593112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 594bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 595bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5967a22ad75STejun Heo * 597112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5987c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 599112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6007c3eed5cSTejun Heo * available only while the work item is queued. 601bbb68dfaSTejun Heo * 602bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 603bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 604bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 605bbb68dfaSTejun Heo * try to steal the PENDING bit. 6064594bf15SDavid Howells */ 6077a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6087a22ad75STejun Heo unsigned long flags) 6097a22ad75STejun Heo { 6106183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6117a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6127a22ad75STejun Heo } 6137a22ad75STejun Heo 614112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6154690c4abSTejun Heo unsigned long extra_flags) 616365970a1SDavid Howells { 617112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 618112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 619365970a1SDavid Howells } 620365970a1SDavid Howells 6214468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6224468a00fSLai Jiangshan int pool_id) 6234468a00fSLai Jiangshan { 6244468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6254468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6264468a00fSLai Jiangshan } 6274468a00fSLai Jiangshan 6287c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6297c3eed5cSTejun Heo int pool_id) 6304d707b9fSOleg Nesterov { 63123657bb1STejun Heo /* 63223657bb1STejun Heo * The following wmb is paired with the implied mb in 63323657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 63423657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 63523657bb1STejun Heo * owner. 63623657bb1STejun Heo */ 63723657bb1STejun Heo smp_wmb(); 6387c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6394d707b9fSOleg Nesterov } 6404d707b9fSOleg Nesterov 6417a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 642365970a1SDavid Howells { 6437c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6447c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6457a22ad75STejun Heo } 6467a22ad75STejun Heo 647112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6487a22ad75STejun Heo { 649e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6507a22ad75STejun Heo 651112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 652e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 653e120153dSTejun Heo else 654e120153dSTejun Heo return NULL; 6557a22ad75STejun Heo } 6567a22ad75STejun Heo 6577c3eed5cSTejun Heo /** 6587c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6597c3eed5cSTejun Heo * @work: the work item of interest 6607c3eed5cSTejun Heo * 66168e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 66268e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 66368e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 664fa1b54e6STejun Heo * 665fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 666fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 667fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 668fa1b54e6STejun Heo * returned pool is and stays online. 669d185af30SYacine Belkadi * 670d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6717c3eed5cSTejun Heo */ 6727c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6737a22ad75STejun Heo { 674e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6757c3eed5cSTejun Heo int pool_id; 6767a22ad75STejun Heo 67768e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 678fa1b54e6STejun Heo 679112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 680112202d9STejun Heo return ((struct pool_workqueue *) 6817c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6827a22ad75STejun Heo 6837c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6847c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6857a22ad75STejun Heo return NULL; 6867a22ad75STejun Heo 687fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6887c3eed5cSTejun Heo } 6897c3eed5cSTejun Heo 6907c3eed5cSTejun Heo /** 6917c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6927c3eed5cSTejun Heo * @work: the work item of interest 6937c3eed5cSTejun Heo * 694d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6957c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6967c3eed5cSTejun Heo */ 6977c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6987c3eed5cSTejun Heo { 69954d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7007c3eed5cSTejun Heo 701112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 702112202d9STejun Heo return ((struct pool_workqueue *) 70354d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 70454d5b7d0SLai Jiangshan 70554d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7067c3eed5cSTejun Heo } 7077c3eed5cSTejun Heo 708bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 709bbb68dfaSTejun Heo { 7107c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 711bbb68dfaSTejun Heo 7127c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7137c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 714bbb68dfaSTejun Heo } 715bbb68dfaSTejun Heo 716bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 717bbb68dfaSTejun Heo { 718bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 719bbb68dfaSTejun Heo 720112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 721bbb68dfaSTejun Heo } 722bbb68dfaSTejun Heo 723e22bee78STejun Heo /* 7243270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7253270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 726d565ed63STejun Heo * they're being called with pool->lock held. 727e22bee78STejun Heo */ 728e22bee78STejun Heo 72963d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 730649027d7STejun Heo { 731e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 732649027d7STejun Heo } 733649027d7STejun Heo 734e22bee78STejun Heo /* 735e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 736e22bee78STejun Heo * running workers. 737974271c4STejun Heo * 738974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 739706026c2STejun Heo * function will always return %true for unbound pools as long as the 740974271c4STejun Heo * worklist isn't empty. 741e22bee78STejun Heo */ 74263d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 743e22bee78STejun Heo { 74463d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 745e22bee78STejun Heo } 746e22bee78STejun Heo 747e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 74863d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 749e22bee78STejun Heo { 75063d95a91STejun Heo return pool->nr_idle; 751e22bee78STejun Heo } 752e22bee78STejun Heo 753e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 75463d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 755e22bee78STejun Heo { 756e19e397aSTejun Heo return !list_empty(&pool->worklist) && 757e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 758e22bee78STejun Heo } 759e22bee78STejun Heo 760e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 76163d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 762e22bee78STejun Heo { 76363d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 764e22bee78STejun Heo } 765e22bee78STejun Heo 766e22bee78STejun Heo /* Do I need to be the manager? */ 76763d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 768e22bee78STejun Heo { 76963d95a91STejun Heo return need_to_create_worker(pool) || 77011ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 771e22bee78STejun Heo } 772e22bee78STejun Heo 773e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 77463d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 775e22bee78STejun Heo { 77634a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 77763d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 77863d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 779e22bee78STejun Heo 780ea1abd61SLai Jiangshan /* 781ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 782ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 783ea1abd61SLai Jiangshan */ 784ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 785ea1abd61SLai Jiangshan return false; 786ea1abd61SLai Jiangshan 787e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 788e22bee78STejun Heo } 789e22bee78STejun Heo 790e22bee78STejun Heo /* 791e22bee78STejun Heo * Wake up functions. 792e22bee78STejun Heo */ 793e22bee78STejun Heo 7947e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 79563d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7967e11629dSTejun Heo { 79763d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7987e11629dSTejun Heo return NULL; 7997e11629dSTejun Heo 80063d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8017e11629dSTejun Heo } 8027e11629dSTejun Heo 8037e11629dSTejun Heo /** 8047e11629dSTejun Heo * wake_up_worker - wake up an idle worker 80563d95a91STejun Heo * @pool: worker pool to wake worker from 8067e11629dSTejun Heo * 80763d95a91STejun Heo * Wake up the first idle worker of @pool. 8087e11629dSTejun Heo * 8097e11629dSTejun Heo * CONTEXT: 810d565ed63STejun Heo * spin_lock_irq(pool->lock). 8117e11629dSTejun Heo */ 81263d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8137e11629dSTejun Heo { 81463d95a91STejun Heo struct worker *worker = first_worker(pool); 8157e11629dSTejun Heo 8167e11629dSTejun Heo if (likely(worker)) 8177e11629dSTejun Heo wake_up_process(worker->task); 8187e11629dSTejun Heo } 8197e11629dSTejun Heo 8204690c4abSTejun Heo /** 821e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 822e22bee78STejun Heo * @task: task waking up 823e22bee78STejun Heo * @cpu: CPU @task is waking up to 824e22bee78STejun Heo * 825e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 826e22bee78STejun Heo * being awoken. 827e22bee78STejun Heo * 828e22bee78STejun Heo * CONTEXT: 829e22bee78STejun Heo * spin_lock_irq(rq->lock) 830e22bee78STejun Heo */ 831d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 832e22bee78STejun Heo { 833e22bee78STejun Heo struct worker *worker = kthread_data(task); 834e22bee78STejun Heo 83536576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 836ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 837e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 838e22bee78STejun Heo } 83936576000SJoonsoo Kim } 840e22bee78STejun Heo 841e22bee78STejun Heo /** 842e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 843e22bee78STejun Heo * @task: task going to sleep 844e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 845e22bee78STejun Heo * 846e22bee78STejun Heo * This function is called during schedule() when a busy worker is 847e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 848e22bee78STejun Heo * returning pointer to its task. 849e22bee78STejun Heo * 850e22bee78STejun Heo * CONTEXT: 851e22bee78STejun Heo * spin_lock_irq(rq->lock) 852e22bee78STejun Heo * 853d185af30SYacine Belkadi * Return: 854e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 855e22bee78STejun Heo */ 856d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 857e22bee78STejun Heo { 858e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 859111c225aSTejun Heo struct worker_pool *pool; 860e22bee78STejun Heo 861111c225aSTejun Heo /* 862111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 863111c225aSTejun Heo * workers, also reach here, let's not access anything before 864111c225aSTejun Heo * checking NOT_RUNNING. 865111c225aSTejun Heo */ 8662d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 867e22bee78STejun Heo return NULL; 868e22bee78STejun Heo 869111c225aSTejun Heo pool = worker->pool; 870111c225aSTejun Heo 871e22bee78STejun Heo /* this can only happen on the local cpu */ 8726183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8736183c009STejun Heo return NULL; 874e22bee78STejun Heo 875e22bee78STejun Heo /* 876e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 877e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 878e22bee78STejun Heo * Please read comment there. 879e22bee78STejun Heo * 880628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 881628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 882628c78e7STejun Heo * disabled, which in turn means that none else could be 883d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 884628c78e7STejun Heo * lock is safe. 885e22bee78STejun Heo */ 886e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 887e19e397aSTejun Heo !list_empty(&pool->worklist)) 88863d95a91STejun Heo to_wakeup = first_worker(pool); 889e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 890e22bee78STejun Heo } 891e22bee78STejun Heo 892e22bee78STejun Heo /** 893e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 894cb444766STejun Heo * @worker: self 895d302f017STejun Heo * @flags: flags to set 896d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 897d302f017STejun Heo * 898e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 899e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 900e22bee78STejun Heo * woken up. 901d302f017STejun Heo * 902cb444766STejun Heo * CONTEXT: 903d565ed63STejun Heo * spin_lock_irq(pool->lock) 904d302f017STejun Heo */ 905d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 906d302f017STejun Heo bool wakeup) 907d302f017STejun Heo { 908bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 909e22bee78STejun Heo 910cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 911cb444766STejun Heo 912e22bee78STejun Heo /* 913e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 914e22bee78STejun Heo * wake up an idle worker as necessary if requested by 915e22bee78STejun Heo * @wakeup. 916e22bee78STejun Heo */ 917e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 918e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 919e22bee78STejun Heo if (wakeup) { 920e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 921bd7bdd43STejun Heo !list_empty(&pool->worklist)) 92263d95a91STejun Heo wake_up_worker(pool); 923e22bee78STejun Heo } else 924e19e397aSTejun Heo atomic_dec(&pool->nr_running); 925e22bee78STejun Heo } 926e22bee78STejun Heo 927d302f017STejun Heo worker->flags |= flags; 928d302f017STejun Heo } 929d302f017STejun Heo 930d302f017STejun Heo /** 931e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 932cb444766STejun Heo * @worker: self 933d302f017STejun Heo * @flags: flags to clear 934d302f017STejun Heo * 935e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 936d302f017STejun Heo * 937cb444766STejun Heo * CONTEXT: 938d565ed63STejun Heo * spin_lock_irq(pool->lock) 939d302f017STejun Heo */ 940d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 941d302f017STejun Heo { 94263d95a91STejun Heo struct worker_pool *pool = worker->pool; 943e22bee78STejun Heo unsigned int oflags = worker->flags; 944e22bee78STejun Heo 945cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 946cb444766STejun Heo 947d302f017STejun Heo worker->flags &= ~flags; 948e22bee78STejun Heo 94942c025f3STejun Heo /* 95042c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 95142c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 95242c025f3STejun Heo * of multiple flags, not a single flag. 95342c025f3STejun Heo */ 954e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 955e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 956e19e397aSTejun Heo atomic_inc(&pool->nr_running); 957d302f017STejun Heo } 958d302f017STejun Heo 959d302f017STejun Heo /** 9608cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 961c9e7cf27STejun Heo * @pool: pool of interest 9628cca0eeaSTejun Heo * @work: work to find worker for 9638cca0eeaSTejun Heo * 964c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 965c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 966a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 967a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 968a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 969a2c1c57bSTejun Heo * being executed. 970a2c1c57bSTejun Heo * 971a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 972a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 973a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 974a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 975a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 976a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 977a2c1c57bSTejun Heo * 978c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 979c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 980c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 981c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 982c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 983c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9848cca0eeaSTejun Heo * 9858cca0eeaSTejun Heo * CONTEXT: 986d565ed63STejun Heo * spin_lock_irq(pool->lock). 9878cca0eeaSTejun Heo * 988d185af30SYacine Belkadi * Return: 989d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9908cca0eeaSTejun Heo * otherwise. 9918cca0eeaSTejun Heo */ 992c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9938cca0eeaSTejun Heo struct work_struct *work) 9948cca0eeaSTejun Heo { 99542f8570fSSasha Levin struct worker *worker; 99642f8570fSSasha Levin 997b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 998a2c1c57bSTejun Heo (unsigned long)work) 999a2c1c57bSTejun Heo if (worker->current_work == work && 1000a2c1c57bSTejun Heo worker->current_func == work->func) 100142f8570fSSasha Levin return worker; 100242f8570fSSasha Levin 100342f8570fSSasha Levin return NULL; 10048cca0eeaSTejun Heo } 10058cca0eeaSTejun Heo 10068cca0eeaSTejun Heo /** 1007bf4ede01STejun Heo * move_linked_works - move linked works to a list 1008bf4ede01STejun Heo * @work: start of series of works to be scheduled 1009bf4ede01STejun Heo * @head: target list to append @work to 1010bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 1011bf4ede01STejun Heo * 1012bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1013bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1014bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1015bf4ede01STejun Heo * 1016bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1017bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1018bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1019bf4ede01STejun Heo * 1020bf4ede01STejun Heo * CONTEXT: 1021d565ed63STejun Heo * spin_lock_irq(pool->lock). 1022bf4ede01STejun Heo */ 1023bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1024bf4ede01STejun Heo struct work_struct **nextp) 1025bf4ede01STejun Heo { 1026bf4ede01STejun Heo struct work_struct *n; 1027bf4ede01STejun Heo 1028bf4ede01STejun Heo /* 1029bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1030bf4ede01STejun Heo * use NULL for list head. 1031bf4ede01STejun Heo */ 1032bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1033bf4ede01STejun Heo list_move_tail(&work->entry, head); 1034bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1035bf4ede01STejun Heo break; 1036bf4ede01STejun Heo } 1037bf4ede01STejun Heo 1038bf4ede01STejun Heo /* 1039bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1040bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1041bf4ede01STejun Heo * needs to be updated. 1042bf4ede01STejun Heo */ 1043bf4ede01STejun Heo if (nextp) 1044bf4ede01STejun Heo *nextp = n; 1045bf4ede01STejun Heo } 1046bf4ede01STejun Heo 10478864b4e5STejun Heo /** 10488864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10498864b4e5STejun Heo * @pwq: pool_workqueue to get 10508864b4e5STejun Heo * 10518864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10528864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10538864b4e5STejun Heo */ 10548864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10558864b4e5STejun Heo { 10568864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10578864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10588864b4e5STejun Heo pwq->refcnt++; 10598864b4e5STejun Heo } 10608864b4e5STejun Heo 10618864b4e5STejun Heo /** 10628864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10638864b4e5STejun Heo * @pwq: pool_workqueue to put 10648864b4e5STejun Heo * 10658864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10668864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10678864b4e5STejun Heo */ 10688864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10698864b4e5STejun Heo { 10708864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10718864b4e5STejun Heo if (likely(--pwq->refcnt)) 10728864b4e5STejun Heo return; 10738864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10748864b4e5STejun Heo return; 10758864b4e5STejun Heo /* 10768864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10778864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10788864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10798864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10808864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10818864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10828864b4e5STejun Heo */ 10838864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10848864b4e5STejun Heo } 10858864b4e5STejun Heo 1086dce90d47STejun Heo /** 1087dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1088dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1089dce90d47STejun Heo * 1090dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1091dce90d47STejun Heo */ 1092dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1093dce90d47STejun Heo { 1094dce90d47STejun Heo if (pwq) { 1095dce90d47STejun Heo /* 1096dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1097dce90d47STejun Heo * following lock operations are safe. 1098dce90d47STejun Heo */ 1099dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1100dce90d47STejun Heo put_pwq(pwq); 1101dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1102dce90d47STejun Heo } 1103dce90d47STejun Heo } 1104dce90d47STejun Heo 1105112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1106bf4ede01STejun Heo { 1107112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1108bf4ede01STejun Heo 1109bf4ede01STejun Heo trace_workqueue_activate_work(work); 1110112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1111bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1112112202d9STejun Heo pwq->nr_active++; 1113bf4ede01STejun Heo } 1114bf4ede01STejun Heo 1115112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11163aa62497SLai Jiangshan { 1117112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11183aa62497SLai Jiangshan struct work_struct, entry); 11193aa62497SLai Jiangshan 1120112202d9STejun Heo pwq_activate_delayed_work(work); 11213aa62497SLai Jiangshan } 11223aa62497SLai Jiangshan 1123bf4ede01STejun Heo /** 1124112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1125112202d9STejun Heo * @pwq: pwq of interest 1126bf4ede01STejun Heo * @color: color of work which left the queue 1127bf4ede01STejun Heo * 1128bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1129112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1130bf4ede01STejun Heo * 1131bf4ede01STejun Heo * CONTEXT: 1132d565ed63STejun Heo * spin_lock_irq(pool->lock). 1133bf4ede01STejun Heo */ 1134112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1135bf4ede01STejun Heo { 11368864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1137bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11388864b4e5STejun Heo goto out_put; 1139bf4ede01STejun Heo 1140112202d9STejun Heo pwq->nr_in_flight[color]--; 1141bf4ede01STejun Heo 1142112202d9STejun Heo pwq->nr_active--; 1143112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1144bf4ede01STejun Heo /* one down, submit a delayed one */ 1145112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1146112202d9STejun Heo pwq_activate_first_delayed(pwq); 1147bf4ede01STejun Heo } 1148bf4ede01STejun Heo 1149bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1150112202d9STejun Heo if (likely(pwq->flush_color != color)) 11518864b4e5STejun Heo goto out_put; 1152bf4ede01STejun Heo 1153bf4ede01STejun Heo /* are there still in-flight works? */ 1154112202d9STejun Heo if (pwq->nr_in_flight[color]) 11558864b4e5STejun Heo goto out_put; 1156bf4ede01STejun Heo 1157112202d9STejun Heo /* this pwq is done, clear flush_color */ 1158112202d9STejun Heo pwq->flush_color = -1; 1159bf4ede01STejun Heo 1160bf4ede01STejun Heo /* 1161112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1162bf4ede01STejun Heo * will handle the rest. 1163bf4ede01STejun Heo */ 1164112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1165112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11668864b4e5STejun Heo out_put: 11678864b4e5STejun Heo put_pwq(pwq); 1168bf4ede01STejun Heo } 1169bf4ede01STejun Heo 117036e227d2STejun Heo /** 1171bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 117236e227d2STejun Heo * @work: work item to steal 117336e227d2STejun Heo * @is_dwork: @work is a delayed_work 1174bbb68dfaSTejun Heo * @flags: place to store irq state 117536e227d2STejun Heo * 117636e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1177d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 117836e227d2STejun Heo * 1179d185af30SYacine Belkadi * Return: 118036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 118136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 118236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1183bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1184bbb68dfaSTejun Heo * for arbitrarily long 118536e227d2STejun Heo * 1186d185af30SYacine Belkadi * Note: 1187bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1188e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1189e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1190e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1191bbb68dfaSTejun Heo * 1192bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1193bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1194bbb68dfaSTejun Heo * 1195e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1196bf4ede01STejun Heo */ 1197bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1198bbb68dfaSTejun Heo unsigned long *flags) 1199bf4ede01STejun Heo { 1200d565ed63STejun Heo struct worker_pool *pool; 1201112202d9STejun Heo struct pool_workqueue *pwq; 1202bf4ede01STejun Heo 1203bbb68dfaSTejun Heo local_irq_save(*flags); 1204bbb68dfaSTejun Heo 120536e227d2STejun Heo /* try to steal the timer if it exists */ 120636e227d2STejun Heo if (is_dwork) { 120736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 120836e227d2STejun Heo 1209e0aecdd8STejun Heo /* 1210e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1211e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1212e0aecdd8STejun Heo * running on the local CPU. 1213e0aecdd8STejun Heo */ 121436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 121536e227d2STejun Heo return 1; 121636e227d2STejun Heo } 121736e227d2STejun Heo 121836e227d2STejun Heo /* try to claim PENDING the normal way */ 1219bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1220bf4ede01STejun Heo return 0; 1221bf4ede01STejun Heo 1222bf4ede01STejun Heo /* 1223bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1224bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1225bf4ede01STejun Heo */ 1226d565ed63STejun Heo pool = get_work_pool(work); 1227d565ed63STejun Heo if (!pool) 1228bbb68dfaSTejun Heo goto fail; 1229bf4ede01STejun Heo 1230d565ed63STejun Heo spin_lock(&pool->lock); 1231bf4ede01STejun Heo /* 1232112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1233112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1234112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1235112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1236112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12370b3dae68SLai Jiangshan * item is currently queued on that pool. 1238bf4ede01STejun Heo */ 1239112202d9STejun Heo pwq = get_work_pwq(work); 1240112202d9STejun Heo if (pwq && pwq->pool == pool) { 1241bf4ede01STejun Heo debug_work_deactivate(work); 12423aa62497SLai Jiangshan 12433aa62497SLai Jiangshan /* 124416062836STejun Heo * A delayed work item cannot be grabbed directly because 124516062836STejun Heo * it might have linked NO_COLOR work items which, if left 1246112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 124716062836STejun Heo * management later on and cause stall. Make sure the work 124816062836STejun Heo * item is activated before grabbing. 12493aa62497SLai Jiangshan */ 12503aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1251112202d9STejun Heo pwq_activate_delayed_work(work); 12523aa62497SLai Jiangshan 1253bf4ede01STejun Heo list_del_init(&work->entry); 1254112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 125536e227d2STejun Heo 1256112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12574468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12584468a00fSLai Jiangshan 1259d565ed63STejun Heo spin_unlock(&pool->lock); 126036e227d2STejun Heo return 1; 1261bf4ede01STejun Heo } 1262d565ed63STejun Heo spin_unlock(&pool->lock); 1263bbb68dfaSTejun Heo fail: 1264bbb68dfaSTejun Heo local_irq_restore(*flags); 1265bbb68dfaSTejun Heo if (work_is_canceling(work)) 1266bbb68dfaSTejun Heo return -ENOENT; 1267bbb68dfaSTejun Heo cpu_relax(); 126836e227d2STejun Heo return -EAGAIN; 1269bf4ede01STejun Heo } 1270bf4ede01STejun Heo 1271bf4ede01STejun Heo /** 1272706026c2STejun Heo * insert_work - insert a work into a pool 1273112202d9STejun Heo * @pwq: pwq @work belongs to 12744690c4abSTejun Heo * @work: work to insert 12754690c4abSTejun Heo * @head: insertion point 12764690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12774690c4abSTejun Heo * 1278112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1279706026c2STejun Heo * work_struct flags. 12804690c4abSTejun Heo * 12814690c4abSTejun Heo * CONTEXT: 1282d565ed63STejun Heo * spin_lock_irq(pool->lock). 1283365970a1SDavid Howells */ 1284112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1285112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1286b89deed3SOleg Nesterov { 1287112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1288e1d8aa9fSFrederic Weisbecker 12894690c4abSTejun Heo /* we own @work, set data and link */ 1290112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12911a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12928864b4e5STejun Heo get_pwq(pwq); 1293e22bee78STejun Heo 1294e22bee78STejun Heo /* 1295c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1296c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1297c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1298e22bee78STejun Heo */ 1299e22bee78STejun Heo smp_mb(); 1300e22bee78STejun Heo 130163d95a91STejun Heo if (__need_more_worker(pool)) 130263d95a91STejun Heo wake_up_worker(pool); 1303b89deed3SOleg Nesterov } 1304b89deed3SOleg Nesterov 1305c8efcc25STejun Heo /* 1306c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13078d03ecfeSTejun Heo * same workqueue. 1308c8efcc25STejun Heo */ 1309c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1310c8efcc25STejun Heo { 1311c8efcc25STejun Heo struct worker *worker; 1312c8efcc25STejun Heo 13138d03ecfeSTejun Heo worker = current_wq_worker(); 1314c8efcc25STejun Heo /* 13158d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13168d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1317c8efcc25STejun Heo */ 1318112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1319c8efcc25STejun Heo } 1320c8efcc25STejun Heo 1321d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13221da177e4SLinus Torvalds struct work_struct *work) 13231da177e4SLinus Torvalds { 1324112202d9STejun Heo struct pool_workqueue *pwq; 1325c9178087STejun Heo struct worker_pool *last_pool; 13261e19ffc6STejun Heo struct list_head *worklist; 13278a2e8e5dSTejun Heo unsigned int work_flags; 1328b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13298930cabaSTejun Heo 13308930cabaSTejun Heo /* 13318930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13328930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13338930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13348930cabaSTejun Heo * happen with IRQ disabled. 13358930cabaSTejun Heo */ 13368930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13371da177e4SLinus Torvalds 1338dc186ad7SThomas Gleixner debug_work_activate(work); 13391e19ffc6STejun Heo 13409ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1341618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1342c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1343e41e704bSTejun Heo return; 13449e8cd2f5STejun Heo retry: 1345df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1346f3421797STejun Heo cpu = raw_smp_processor_id(); 1347df2d5ae4STejun Heo 1348df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1349df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1350c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1351df2d5ae4STejun Heo else 1352df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1353f3421797STejun Heo 135418aa9effSTejun Heo /* 1355c9178087STejun Heo * If @work was previously on a different pool, it might still be 1356c9178087STejun Heo * running there, in which case the work needs to be queued on that 1357c9178087STejun Heo * pool to guarantee non-reentrancy. 135818aa9effSTejun Heo */ 1359c9e7cf27STejun Heo last_pool = get_work_pool(work); 1360112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 136118aa9effSTejun Heo struct worker *worker; 136218aa9effSTejun Heo 1363d565ed63STejun Heo spin_lock(&last_pool->lock); 136418aa9effSTejun Heo 1365c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 136618aa9effSTejun Heo 1367112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1368c9178087STejun Heo pwq = worker->current_pwq; 13698594fadeSLai Jiangshan } else { 137018aa9effSTejun Heo /* meh... not running there, queue here */ 1371d565ed63STejun Heo spin_unlock(&last_pool->lock); 1372112202d9STejun Heo spin_lock(&pwq->pool->lock); 137318aa9effSTejun Heo } 13748930cabaSTejun Heo } else { 1375112202d9STejun Heo spin_lock(&pwq->pool->lock); 13768930cabaSTejun Heo } 1377502ca9d8STejun Heo 13789e8cd2f5STejun Heo /* 13799e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13809e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13819e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1382df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1383df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13849e8cd2f5STejun Heo * make forward-progress. 13859e8cd2f5STejun Heo */ 13869e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13879e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13889e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13899e8cd2f5STejun Heo cpu_relax(); 13909e8cd2f5STejun Heo goto retry; 13919e8cd2f5STejun Heo } 13929e8cd2f5STejun Heo /* oops */ 13939e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13949e8cd2f5STejun Heo wq->name, cpu); 13959e8cd2f5STejun Heo } 13969e8cd2f5STejun Heo 1397112202d9STejun Heo /* pwq determined, queue */ 1398112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1399502ca9d8STejun Heo 1400f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1401112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1402f5b2552bSDan Carpenter return; 1403f5b2552bSDan Carpenter } 14041e19ffc6STejun Heo 1405112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1406112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14071e19ffc6STejun Heo 1408112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1409cdadf009STejun Heo trace_workqueue_activate_work(work); 1410112202d9STejun Heo pwq->nr_active++; 1411112202d9STejun Heo worklist = &pwq->pool->worklist; 14128a2e8e5dSTejun Heo } else { 14138a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1414112202d9STejun Heo worklist = &pwq->delayed_works; 14158a2e8e5dSTejun Heo } 14161e19ffc6STejun Heo 1417112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14181e19ffc6STejun Heo 1419112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14201da177e4SLinus Torvalds } 14211da177e4SLinus Torvalds 14220fcb78c2SRolf Eike Beer /** 1423c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1424c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1425c1a220e7SZhang Rui * @wq: workqueue to use 1426c1a220e7SZhang Rui * @work: work to queue 1427c1a220e7SZhang Rui * 1428c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1429c1a220e7SZhang Rui * can't go away. 1430d185af30SYacine Belkadi * 1431d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1432c1a220e7SZhang Rui */ 1433d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1434d4283e93STejun Heo struct work_struct *work) 1435c1a220e7SZhang Rui { 1436d4283e93STejun Heo bool ret = false; 14378930cabaSTejun Heo unsigned long flags; 14388930cabaSTejun Heo 14398930cabaSTejun Heo local_irq_save(flags); 1440c1a220e7SZhang Rui 144122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14424690c4abSTejun Heo __queue_work(cpu, wq, work); 1443d4283e93STejun Heo ret = true; 1444c1a220e7SZhang Rui } 14458930cabaSTejun Heo 14468930cabaSTejun Heo local_irq_restore(flags); 1447c1a220e7SZhang Rui return ret; 1448c1a220e7SZhang Rui } 1449ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1450c1a220e7SZhang Rui 1451d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14521da177e4SLinus Torvalds { 145352bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14541da177e4SLinus Torvalds 1455e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 145660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14571da177e4SLinus Torvalds } 14581438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14591da177e4SLinus Torvalds 14607beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 146152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14621da177e4SLinus Torvalds { 14637beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14647beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14651da177e4SLinus Torvalds 14667beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14677beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1468fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1469fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14707beb2edfSTejun Heo 14718852aac2STejun Heo /* 14728852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14738852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14748852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14758852aac2STejun Heo * on that there's no such delay when @delay is 0. 14768852aac2STejun Heo */ 14778852aac2STejun Heo if (!delay) { 14788852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14798852aac2STejun Heo return; 14808852aac2STejun Heo } 14818852aac2STejun Heo 14827beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14837beb2edfSTejun Heo 148460c057bcSLai Jiangshan dwork->wq = wq; 14851265057fSTejun Heo dwork->cpu = cpu; 14867beb2edfSTejun Heo timer->expires = jiffies + delay; 14877beb2edfSTejun Heo 14887beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14897beb2edfSTejun Heo add_timer_on(timer, cpu); 14907beb2edfSTejun Heo else 14917beb2edfSTejun Heo add_timer(timer); 14927beb2edfSTejun Heo } 14931da177e4SLinus Torvalds 14940fcb78c2SRolf Eike Beer /** 14950fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14960fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14970fcb78c2SRolf Eike Beer * @wq: workqueue to use 1498af9997e4SRandy Dunlap * @dwork: work to queue 14990fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 15000fcb78c2SRolf Eike Beer * 1501d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1502715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1503715f1300STejun Heo * execution. 15040fcb78c2SRolf Eike Beer */ 1505d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 150652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15077a6bc1cdSVenkatesh Pallipadi { 150852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1509d4283e93STejun Heo bool ret = false; 15108930cabaSTejun Heo unsigned long flags; 15118930cabaSTejun Heo 15128930cabaSTejun Heo /* read the comment in __queue_work() */ 15138930cabaSTejun Heo local_irq_save(flags); 15147a6bc1cdSVenkatesh Pallipadi 151522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15167beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1517d4283e93STejun Heo ret = true; 15187a6bc1cdSVenkatesh Pallipadi } 15198930cabaSTejun Heo 15208930cabaSTejun Heo local_irq_restore(flags); 15217a6bc1cdSVenkatesh Pallipadi return ret; 15227a6bc1cdSVenkatesh Pallipadi } 1523ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15241da177e4SLinus Torvalds 1525c8e55f36STejun Heo /** 15268376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15278376fe22STejun Heo * @cpu: CPU number to execute work on 15288376fe22STejun Heo * @wq: workqueue to use 15298376fe22STejun Heo * @dwork: work to queue 15308376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15318376fe22STejun Heo * 15328376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15338376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15348376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15358376fe22STejun Heo * current state. 15368376fe22STejun Heo * 1537d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15388376fe22STejun Heo * pending and its timer was modified. 15398376fe22STejun Heo * 1540e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15418376fe22STejun Heo * See try_to_grab_pending() for details. 15428376fe22STejun Heo */ 15438376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15448376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15458376fe22STejun Heo { 15468376fe22STejun Heo unsigned long flags; 15478376fe22STejun Heo int ret; 15488376fe22STejun Heo 15498376fe22STejun Heo do { 15508376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15518376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15528376fe22STejun Heo 15538376fe22STejun Heo if (likely(ret >= 0)) { 15548376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15558376fe22STejun Heo local_irq_restore(flags); 15568376fe22STejun Heo } 15578376fe22STejun Heo 15588376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15598376fe22STejun Heo return ret; 15608376fe22STejun Heo } 15618376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15628376fe22STejun Heo 15638376fe22STejun Heo /** 1564c8e55f36STejun Heo * worker_enter_idle - enter idle state 1565c8e55f36STejun Heo * @worker: worker which is entering idle state 1566c8e55f36STejun Heo * 1567c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1568c8e55f36STejun Heo * necessary. 1569c8e55f36STejun Heo * 1570c8e55f36STejun Heo * LOCKING: 1571d565ed63STejun Heo * spin_lock_irq(pool->lock). 1572c8e55f36STejun Heo */ 1573c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15741da177e4SLinus Torvalds { 1575bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1576c8e55f36STejun Heo 15776183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15786183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15796183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15806183c009STejun Heo return; 1581c8e55f36STejun Heo 1582cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1583cb444766STejun Heo worker->flags |= WORKER_IDLE; 1584bd7bdd43STejun Heo pool->nr_idle++; 1585e22bee78STejun Heo worker->last_active = jiffies; 1586c8e55f36STejun Heo 1587c8e55f36STejun Heo /* idle_list is LIFO */ 1588bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1589db7bccf4STejun Heo 159063d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1591628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1592cb444766STejun Heo 1593544ecf31STejun Heo /* 1594706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1595d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1596628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1597628c78e7STejun Heo * unbind is not in progress. 1598544ecf31STejun Heo */ 159924647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1600bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1601e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1602c8e55f36STejun Heo } 1603c8e55f36STejun Heo 1604c8e55f36STejun Heo /** 1605c8e55f36STejun Heo * worker_leave_idle - leave idle state 1606c8e55f36STejun Heo * @worker: worker which is leaving idle state 1607c8e55f36STejun Heo * 1608c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1609c8e55f36STejun Heo * 1610c8e55f36STejun Heo * LOCKING: 1611d565ed63STejun Heo * spin_lock_irq(pool->lock). 1612c8e55f36STejun Heo */ 1613c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1614c8e55f36STejun Heo { 1615bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1616c8e55f36STejun Heo 16176183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16186183c009STejun Heo return; 1619d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1620bd7bdd43STejun Heo pool->nr_idle--; 1621c8e55f36STejun Heo list_del_init(&worker->entry); 1622c8e55f36STejun Heo } 1623c8e55f36STejun Heo 1624e22bee78STejun Heo /** 1625f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1626f36dc67bSLai Jiangshan * @pool: target worker_pool 1627f36dc67bSLai Jiangshan * 1628f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1629e22bee78STejun Heo * 1630e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1631e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1632e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1633e22bee78STejun Heo * guaranteed to execute on the cpu. 1634e22bee78STejun Heo * 1635f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1636e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1637e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1638e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1639706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1640e22bee78STejun Heo * [dis]associated in the meantime. 1641e22bee78STejun Heo * 1642706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 164324647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1644f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1645f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1646f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1647e22bee78STejun Heo * 1648e22bee78STejun Heo * CONTEXT: 1649d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1650e22bee78STejun Heo * held. 1651e22bee78STejun Heo * 1652d185af30SYacine Belkadi * Return: 1653706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1654e22bee78STejun Heo * bound), %false if offline. 1655e22bee78STejun Heo */ 1656f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1657d565ed63STejun Heo __acquires(&pool->lock) 1658e22bee78STejun Heo { 1659e22bee78STejun Heo while (true) { 1660e22bee78STejun Heo /* 1661e22bee78STejun Heo * The following call may fail, succeed or succeed 1662e22bee78STejun Heo * without actually migrating the task to the cpu if 1663e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 166424647570STejun Heo * against POOL_DISASSOCIATED. 1665e22bee78STejun Heo */ 166624647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16677a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1668e22bee78STejun Heo 1669d565ed63STejun Heo spin_lock_irq(&pool->lock); 167024647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1671e22bee78STejun Heo return false; 1672f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16737a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1674e22bee78STejun Heo return true; 1675d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1676e22bee78STejun Heo 16775035b20fSTejun Heo /* 16785035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16795035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16805035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16815035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16825035b20fSTejun Heo */ 1683e22bee78STejun Heo cpu_relax(); 16845035b20fSTejun Heo cond_resched(); 1685e22bee78STejun Heo } 1686e22bee78STejun Heo } 1687e22bee78STejun Heo 1688c34056a3STejun Heo static struct worker *alloc_worker(void) 1689c34056a3STejun Heo { 1690c34056a3STejun Heo struct worker *worker; 1691c34056a3STejun Heo 1692c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1693c8e55f36STejun Heo if (worker) { 1694c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1695affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1696e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1697e22bee78STejun Heo worker->flags = WORKER_PREP; 1698c8e55f36STejun Heo } 1699c34056a3STejun Heo return worker; 1700c34056a3STejun Heo } 1701c34056a3STejun Heo 1702c34056a3STejun Heo /** 1703c34056a3STejun Heo * create_worker - create a new workqueue worker 170463d95a91STejun Heo * @pool: pool the new worker will belong to 1705c34056a3STejun Heo * 170663d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1707c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1708c34056a3STejun Heo * destroy_worker(). 1709c34056a3STejun Heo * 1710c34056a3STejun Heo * CONTEXT: 1711c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1712c34056a3STejun Heo * 1713d185af30SYacine Belkadi * Return: 1714c34056a3STejun Heo * Pointer to the newly created worker. 1715c34056a3STejun Heo */ 1716bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1717c34056a3STejun Heo { 1718c34056a3STejun Heo struct worker *worker = NULL; 1719f3421797STejun Heo int id = -1; 1720e3c916a4STejun Heo char id_buf[16]; 1721c34056a3STejun Heo 1722cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1723cd549687STejun Heo 1724822d8405STejun Heo /* 1725822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1726822d8405STejun Heo * without installing the pointer. 1727822d8405STejun Heo */ 1728822d8405STejun Heo idr_preload(GFP_KERNEL); 1729d565ed63STejun Heo spin_lock_irq(&pool->lock); 1730822d8405STejun Heo 1731822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1732822d8405STejun Heo 1733d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1734822d8405STejun Heo idr_preload_end(); 1735822d8405STejun Heo if (id < 0) 1736c34056a3STejun Heo goto fail; 1737c34056a3STejun Heo 1738c34056a3STejun Heo worker = alloc_worker(); 1739c34056a3STejun Heo if (!worker) 1740c34056a3STejun Heo goto fail; 1741c34056a3STejun Heo 1742bd7bdd43STejun Heo worker->pool = pool; 1743c34056a3STejun Heo worker->id = id; 1744c34056a3STejun Heo 174529c91e99STejun Heo if (pool->cpu >= 0) 1746e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1747e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1748f3421797STejun Heo else 1749e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1750e3c916a4STejun Heo 1751f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1752e3c916a4STejun Heo "kworker/%s", id_buf); 1753c34056a3STejun Heo if (IS_ERR(worker->task)) 1754c34056a3STejun Heo goto fail; 1755c34056a3STejun Heo 175691151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 175791151228SOleg Nesterov 175891151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 175991151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 176091151228SOleg Nesterov 1761c5aa87bbSTejun Heo /* 1762c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1763c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1764c5aa87bbSTejun Heo */ 17657a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17663270476aSTejun Heo 17677a4e344cSTejun Heo /* 17687a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17697a4e344cSTejun Heo * remains stable across this function. See the comments above the 17707a4e344cSTejun Heo * flag definition for details. 17717a4e344cSTejun Heo */ 17727a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1773f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1774c34056a3STejun Heo 1775822d8405STejun Heo /* successful, commit the pointer to idr */ 1776822d8405STejun Heo spin_lock_irq(&pool->lock); 1777822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1778822d8405STejun Heo spin_unlock_irq(&pool->lock); 1779822d8405STejun Heo 1780c34056a3STejun Heo return worker; 1781822d8405STejun Heo 1782c34056a3STejun Heo fail: 1783c34056a3STejun Heo if (id >= 0) { 1784d565ed63STejun Heo spin_lock_irq(&pool->lock); 1785822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1786d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1787c34056a3STejun Heo } 1788c34056a3STejun Heo kfree(worker); 1789c34056a3STejun Heo return NULL; 1790c34056a3STejun Heo } 1791c34056a3STejun Heo 1792c34056a3STejun Heo /** 1793c34056a3STejun Heo * start_worker - start a newly created worker 1794c34056a3STejun Heo * @worker: worker to start 1795c34056a3STejun Heo * 1796706026c2STejun Heo * Make the pool aware of @worker and start it. 1797c34056a3STejun Heo * 1798c34056a3STejun Heo * CONTEXT: 1799d565ed63STejun Heo * spin_lock_irq(pool->lock). 1800c34056a3STejun Heo */ 1801c34056a3STejun Heo static void start_worker(struct worker *worker) 1802c34056a3STejun Heo { 1803cb444766STejun Heo worker->flags |= WORKER_STARTED; 1804bd7bdd43STejun Heo worker->pool->nr_workers++; 1805c8e55f36STejun Heo worker_enter_idle(worker); 1806c34056a3STejun Heo wake_up_process(worker->task); 1807c34056a3STejun Heo } 1808c34056a3STejun Heo 1809c34056a3STejun Heo /** 1810ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1811ebf44d16STejun Heo * @pool: the target pool 1812ebf44d16STejun Heo * 1813cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1814d185af30SYacine Belkadi * 1815d185af30SYacine Belkadi * Return: 0 on success. A negative error code otherwise. 1816ebf44d16STejun Heo */ 1817ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1818ebf44d16STejun Heo { 1819ebf44d16STejun Heo struct worker *worker; 1820ebf44d16STejun Heo 1821cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1822cd549687STejun Heo 1823ebf44d16STejun Heo worker = create_worker(pool); 1824ebf44d16STejun Heo if (worker) { 1825ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1826ebf44d16STejun Heo start_worker(worker); 1827ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1828ebf44d16STejun Heo } 1829ebf44d16STejun Heo 1830cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1831cd549687STejun Heo 1832ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1833ebf44d16STejun Heo } 1834ebf44d16STejun Heo 1835ebf44d16STejun Heo /** 1836c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1837c34056a3STejun Heo * @worker: worker to be destroyed 1838c34056a3STejun Heo * 1839706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1840c8e55f36STejun Heo * 1841c8e55f36STejun Heo * CONTEXT: 1842d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1843c34056a3STejun Heo */ 1844c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1845c34056a3STejun Heo { 1846bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1847c34056a3STejun Heo 1848cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1849cd549687STejun Heo lockdep_assert_held(&pool->lock); 1850cd549687STejun Heo 1851c34056a3STejun Heo /* sanity check frenzy */ 18526183c009STejun Heo if (WARN_ON(worker->current_work) || 18536183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18546183c009STejun Heo return; 1855c34056a3STejun Heo 1856c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1857bd7bdd43STejun Heo pool->nr_workers--; 1858c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1859bd7bdd43STejun Heo pool->nr_idle--; 1860c8e55f36STejun Heo 18615bdfff96SLai Jiangshan /* 18625bdfff96SLai Jiangshan * Once WORKER_DIE is set, the kworker may destroy itself at any 18635bdfff96SLai Jiangshan * point. Pin to ensure the task stays until we're done with it. 18645bdfff96SLai Jiangshan */ 18655bdfff96SLai Jiangshan get_task_struct(worker->task); 18665bdfff96SLai Jiangshan 1867c8e55f36STejun Heo list_del_init(&worker->entry); 1868cb444766STejun Heo worker->flags |= WORKER_DIE; 1869c8e55f36STejun Heo 1870822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1871822d8405STejun Heo 1872d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1873c8e55f36STejun Heo 1874c34056a3STejun Heo kthread_stop(worker->task); 18755bdfff96SLai Jiangshan put_task_struct(worker->task); 1876c34056a3STejun Heo kfree(worker); 1877c34056a3STejun Heo 1878d565ed63STejun Heo spin_lock_irq(&pool->lock); 1879c34056a3STejun Heo } 1880c34056a3STejun Heo 188163d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1882e22bee78STejun Heo { 188363d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1884e22bee78STejun Heo 1885d565ed63STejun Heo spin_lock_irq(&pool->lock); 1886e22bee78STejun Heo 188763d95a91STejun Heo if (too_many_workers(pool)) { 1888e22bee78STejun Heo struct worker *worker; 1889e22bee78STejun Heo unsigned long expires; 1890e22bee78STejun Heo 1891e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 189263d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1893e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1894e22bee78STejun Heo 1895e22bee78STejun Heo if (time_before(jiffies, expires)) 189663d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1897e22bee78STejun Heo else { 1898e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 189911ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 190063d95a91STejun Heo wake_up_worker(pool); 1901e22bee78STejun Heo } 1902e22bee78STejun Heo } 1903e22bee78STejun Heo 1904d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1905e22bee78STejun Heo } 1906e22bee78STejun Heo 1907493a1724STejun Heo static void send_mayday(struct work_struct *work) 1908e22bee78STejun Heo { 1909112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1910112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1911493a1724STejun Heo 19122e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1913e22bee78STejun Heo 1914493008a8STejun Heo if (!wq->rescuer) 1915493a1724STejun Heo return; 1916e22bee78STejun Heo 1917e22bee78STejun Heo /* mayday mayday mayday */ 1918493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1919493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1920e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1921493a1724STejun Heo } 1922e22bee78STejun Heo } 1923e22bee78STejun Heo 1924706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1925e22bee78STejun Heo { 192663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1927e22bee78STejun Heo struct work_struct *work; 1928e22bee78STejun Heo 19292e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1930493a1724STejun Heo spin_lock(&pool->lock); 1931e22bee78STejun Heo 193263d95a91STejun Heo if (need_to_create_worker(pool)) { 1933e22bee78STejun Heo /* 1934e22bee78STejun Heo * We've been trying to create a new worker but 1935e22bee78STejun Heo * haven't been successful. We might be hitting an 1936e22bee78STejun Heo * allocation deadlock. Send distress signals to 1937e22bee78STejun Heo * rescuers. 1938e22bee78STejun Heo */ 193963d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1940e22bee78STejun Heo send_mayday(work); 1941e22bee78STejun Heo } 1942e22bee78STejun Heo 1943493a1724STejun Heo spin_unlock(&pool->lock); 19442e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1945e22bee78STejun Heo 194663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1947e22bee78STejun Heo } 1948e22bee78STejun Heo 1949e22bee78STejun Heo /** 1950e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 195163d95a91STejun Heo * @pool: pool to create a new worker for 1952e22bee78STejun Heo * 195363d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1954e22bee78STejun Heo * have at least one idle worker on return from this function. If 1955e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 195663d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1957e22bee78STejun Heo * possible allocation deadlock. 1958e22bee78STejun Heo * 1959c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1960c5aa87bbSTejun Heo * may_start_working() %true. 1961e22bee78STejun Heo * 1962e22bee78STejun Heo * LOCKING: 1963d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1964e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1965e22bee78STejun Heo * manager. 1966e22bee78STejun Heo * 1967d185af30SYacine Belkadi * Return: 1968c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1969e22bee78STejun Heo * otherwise. 1970e22bee78STejun Heo */ 197163d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1972d565ed63STejun Heo __releases(&pool->lock) 1973d565ed63STejun Heo __acquires(&pool->lock) 1974e22bee78STejun Heo { 197563d95a91STejun Heo if (!need_to_create_worker(pool)) 1976e22bee78STejun Heo return false; 1977e22bee78STejun Heo restart: 1978d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19799f9c2364STejun Heo 1980e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 198163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1982e22bee78STejun Heo 1983e22bee78STejun Heo while (true) { 1984e22bee78STejun Heo struct worker *worker; 1985e22bee78STejun Heo 1986bc2ae0f5STejun Heo worker = create_worker(pool); 1987e22bee78STejun Heo if (worker) { 198863d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1989d565ed63STejun Heo spin_lock_irq(&pool->lock); 1990e22bee78STejun Heo start_worker(worker); 19916183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19926183c009STejun Heo goto restart; 1993e22bee78STejun Heo return true; 1994e22bee78STejun Heo } 1995e22bee78STejun Heo 199663d95a91STejun Heo if (!need_to_create_worker(pool)) 1997e22bee78STejun Heo break; 1998e22bee78STejun Heo 1999e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2000e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 20019f9c2364STejun Heo 200263d95a91STejun Heo if (!need_to_create_worker(pool)) 2003e22bee78STejun Heo break; 2004e22bee78STejun Heo } 2005e22bee78STejun Heo 200663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2007d565ed63STejun Heo spin_lock_irq(&pool->lock); 200863d95a91STejun Heo if (need_to_create_worker(pool)) 2009e22bee78STejun Heo goto restart; 2010e22bee78STejun Heo return true; 2011e22bee78STejun Heo } 2012e22bee78STejun Heo 2013e22bee78STejun Heo /** 2014e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 201563d95a91STejun Heo * @pool: pool to destroy workers for 2016e22bee78STejun Heo * 201763d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 2018e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 2019e22bee78STejun Heo * 2020e22bee78STejun Heo * LOCKING: 2021d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2022e22bee78STejun Heo * multiple times. Called only from manager. 2023e22bee78STejun Heo * 2024d185af30SYacine Belkadi * Return: 2025c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 2026e22bee78STejun Heo * otherwise. 2027e22bee78STejun Heo */ 202863d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 2029e22bee78STejun Heo { 2030e22bee78STejun Heo bool ret = false; 2031e22bee78STejun Heo 203263d95a91STejun Heo while (too_many_workers(pool)) { 2033e22bee78STejun Heo struct worker *worker; 2034e22bee78STejun Heo unsigned long expires; 2035e22bee78STejun Heo 203663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2037e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2038e22bee78STejun Heo 2039e22bee78STejun Heo if (time_before(jiffies, expires)) { 204063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2041e22bee78STejun Heo break; 2042e22bee78STejun Heo } 2043e22bee78STejun Heo 2044e22bee78STejun Heo destroy_worker(worker); 2045e22bee78STejun Heo ret = true; 2046e22bee78STejun Heo } 2047e22bee78STejun Heo 2048e22bee78STejun Heo return ret; 2049e22bee78STejun Heo } 2050e22bee78STejun Heo 2051e22bee78STejun Heo /** 2052e22bee78STejun Heo * manage_workers - manage worker pool 2053e22bee78STejun Heo * @worker: self 2054e22bee78STejun Heo * 2055706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2056e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2057706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2058e22bee78STejun Heo * 2059e22bee78STejun Heo * The caller can safely start processing works on false return. On 2060e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2061e22bee78STejun Heo * and may_start_working() is true. 2062e22bee78STejun Heo * 2063e22bee78STejun Heo * CONTEXT: 2064d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2065e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2066e22bee78STejun Heo * 2067d185af30SYacine Belkadi * Return: 20682d498db9SLibin * %false if the pool don't need management and the caller can safely start 20692d498db9SLibin * processing works, %true indicates that the function released pool->lock 20702d498db9SLibin * and reacquired it to perform some management function and that the 20712d498db9SLibin * conditions that the caller verified while holding the lock before 20722d498db9SLibin * calling the function might no longer be true. 2073e22bee78STejun Heo */ 2074e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2075e22bee78STejun Heo { 207663d95a91STejun Heo struct worker_pool *pool = worker->pool; 2077e22bee78STejun Heo bool ret = false; 2078e22bee78STejun Heo 2079bc3a1afcSTejun Heo /* 2080bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2081bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2082bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2083bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2084bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2085bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2086bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2087bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2088bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2089bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2090bc3a1afcSTejun Heo * 2091bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2092bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2093bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2094bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2095bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2096bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2097bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2098bc3a1afcSTejun Heo * manager_mutex. 2099bc3a1afcSTejun Heo */ 210034a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2101e22bee78STejun Heo return ret; 2102e22bee78STejun Heo 2103ee378aa4SLai Jiangshan /* 2104bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2105bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2106ee378aa4SLai Jiangshan */ 2107bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2108d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2109bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 21108f174b11SJoonsoo Kim spin_lock_irq(&pool->lock); 2111ee378aa4SLai Jiangshan ret = true; 2112ee378aa4SLai Jiangshan } 2113ee378aa4SLai Jiangshan 211411ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2115e22bee78STejun Heo 2116e22bee78STejun Heo /* 2117e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2118e22bee78STejun Heo * on return. 2119e22bee78STejun Heo */ 212063d95a91STejun Heo ret |= maybe_destroy_workers(pool); 212163d95a91STejun Heo ret |= maybe_create_worker(pool); 2122e22bee78STejun Heo 2123bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 212434a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2125e22bee78STejun Heo return ret; 2126e22bee78STejun Heo } 2127e22bee78STejun Heo 2128a62428c0STejun Heo /** 2129a62428c0STejun Heo * process_one_work - process single work 2130c34056a3STejun Heo * @worker: self 2131a62428c0STejun Heo * @work: work to process 2132a62428c0STejun Heo * 2133a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2134a62428c0STejun Heo * process a single work including synchronization against and 2135a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2136a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2137a62428c0STejun Heo * call this function to process a work. 2138a62428c0STejun Heo * 2139a62428c0STejun Heo * CONTEXT: 2140d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2141a62428c0STejun Heo */ 2142c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2143d565ed63STejun Heo __releases(&pool->lock) 2144d565ed63STejun Heo __acquires(&pool->lock) 21451da177e4SLinus Torvalds { 2146112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2147bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2148112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 214973f53c4aSTejun Heo int work_color; 21507e11629dSTejun Heo struct worker *collision; 21514e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21524e6045f1SJohannes Berg /* 2153a62428c0STejun Heo * It is permissible to free the struct work_struct from 2154a62428c0STejun Heo * inside the function that is called from it, this we need to 2155a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2156a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2157a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21584e6045f1SJohannes Berg */ 21594d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21604d82a1deSPeter Zijlstra 21614d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21624e6045f1SJohannes Berg #endif 21636fec10a1STejun Heo /* 21646fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21656fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 216624647570STejun Heo * unbound or a disassociated pool. 21676fec10a1STejun Heo */ 21685f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 216924647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2170ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 217125511a47STejun Heo 21727e11629dSTejun Heo /* 21737e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21747e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21757e11629dSTejun Heo * already processing the work. If so, defer the work to the 21767e11629dSTejun Heo * currently executing one. 21777e11629dSTejun Heo */ 2178c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21797e11629dSTejun Heo if (unlikely(collision)) { 21807e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21817e11629dSTejun Heo return; 21827e11629dSTejun Heo } 21831da177e4SLinus Torvalds 21848930cabaSTejun Heo /* claim and dequeue */ 21851da177e4SLinus Torvalds debug_work_deactivate(work); 2186c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2187c34056a3STejun Heo worker->current_work = work; 2188a2c1c57bSTejun Heo worker->current_func = work->func; 2189112202d9STejun Heo worker->current_pwq = pwq; 219073f53c4aSTejun Heo work_color = get_work_color(work); 21917a22ad75STejun Heo 2192a62428c0STejun Heo list_del_init(&work->entry); 2193a62428c0STejun Heo 2194649027d7STejun Heo /* 2195fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2196fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2197fb0e7bebSTejun Heo */ 2198fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2199fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2200fb0e7bebSTejun Heo 2201974271c4STejun Heo /* 2202d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2203974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2204974271c4STejun Heo */ 220563d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 220663d95a91STejun Heo wake_up_worker(pool); 2207974271c4STejun Heo 22088930cabaSTejun Heo /* 22097c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2210d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 221123657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 221223657bb1STejun Heo * disabled. 22138930cabaSTejun Heo */ 22147c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22151da177e4SLinus Torvalds 2216d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2217365970a1SDavid Howells 2218112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 22193295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2220e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2221a2c1c57bSTejun Heo worker->current_func(work); 2222e36c886aSArjan van de Ven /* 2223e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2224e36c886aSArjan van de Ven * point will only record its address. 2225e36c886aSArjan van de Ven */ 2226e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22273295f0efSIngo Molnar lock_map_release(&lockdep_map); 2228112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22291da177e4SLinus Torvalds 2230d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2231044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2232044c782cSValentin Ilie " last function: %pf\n", 2233a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2234a2c1c57bSTejun Heo worker->current_func); 2235d5abe669SPeter Zijlstra debug_show_held_locks(current); 2236d5abe669SPeter Zijlstra dump_stack(); 2237d5abe669SPeter Zijlstra } 2238d5abe669SPeter Zijlstra 2239b22ce278STejun Heo /* 2240b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2241b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2242b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2243b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2244b22ce278STejun Heo * stop_machine. 2245b22ce278STejun Heo */ 2246b22ce278STejun Heo cond_resched(); 2247b22ce278STejun Heo 2248d565ed63STejun Heo spin_lock_irq(&pool->lock); 2249a62428c0STejun Heo 2250fb0e7bebSTejun Heo /* clear cpu intensive status */ 2251fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2252fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2253fb0e7bebSTejun Heo 2254a62428c0STejun Heo /* we're done with it, release */ 225542f8570fSSasha Levin hash_del(&worker->hentry); 2256c34056a3STejun Heo worker->current_work = NULL; 2257a2c1c57bSTejun Heo worker->current_func = NULL; 2258112202d9STejun Heo worker->current_pwq = NULL; 22593d1cb205STejun Heo worker->desc_valid = false; 2260112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 22611da177e4SLinus Torvalds } 22621da177e4SLinus Torvalds 2263affee4b2STejun Heo /** 2264affee4b2STejun Heo * process_scheduled_works - process scheduled works 2265affee4b2STejun Heo * @worker: self 2266affee4b2STejun Heo * 2267affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2268affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2269affee4b2STejun Heo * fetches a work from the top and executes it. 2270affee4b2STejun Heo * 2271affee4b2STejun Heo * CONTEXT: 2272d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2273affee4b2STejun Heo * multiple times. 2274affee4b2STejun Heo */ 2275affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22761da177e4SLinus Torvalds { 2277affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2278affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2279a62428c0STejun Heo struct work_struct, entry); 2280c34056a3STejun Heo process_one_work(worker, work); 2281a62428c0STejun Heo } 22821da177e4SLinus Torvalds } 22831da177e4SLinus Torvalds 22844690c4abSTejun Heo /** 22854690c4abSTejun Heo * worker_thread - the worker thread function 2286c34056a3STejun Heo * @__worker: self 22874690c4abSTejun Heo * 2288c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2289c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2290c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2291c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2292c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2293d185af30SYacine Belkadi * 2294d185af30SYacine Belkadi * Return: 0 22954690c4abSTejun Heo */ 2296c34056a3STejun Heo static int worker_thread(void *__worker) 22971da177e4SLinus Torvalds { 2298c34056a3STejun Heo struct worker *worker = __worker; 2299bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 23001da177e4SLinus Torvalds 2301e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2302e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2303c8e55f36STejun Heo woke_up: 2304d565ed63STejun Heo spin_lock_irq(&pool->lock); 2305affee4b2STejun Heo 2306a9ab775bSTejun Heo /* am I supposed to die? */ 2307a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2308d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2309a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2310e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2311c8e55f36STejun Heo return 0; 2312c8e55f36STejun Heo } 2313c8e55f36STejun Heo 2314c8e55f36STejun Heo worker_leave_idle(worker); 2315db7bccf4STejun Heo recheck: 2316e22bee78STejun Heo /* no more worker necessary? */ 231763d95a91STejun Heo if (!need_more_worker(pool)) 2318e22bee78STejun Heo goto sleep; 2319e22bee78STejun Heo 2320e22bee78STejun Heo /* do we need to manage? */ 232163d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2322e22bee78STejun Heo goto recheck; 2323e22bee78STejun Heo 2324c8e55f36STejun Heo /* 2325c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2326c8e55f36STejun Heo * preparing to process a work or actually processing it. 2327c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2328c8e55f36STejun Heo */ 23296183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2330c8e55f36STejun Heo 2331e22bee78STejun Heo /* 2332a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2333a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2334a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2335a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2336a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2337e22bee78STejun Heo */ 2338a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2339e22bee78STejun Heo 2340e22bee78STejun Heo do { 2341affee4b2STejun Heo struct work_struct *work = 2342bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2343affee4b2STejun Heo struct work_struct, entry); 2344affee4b2STejun Heo 2345c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2346affee4b2STejun Heo /* optimization path, not strictly necessary */ 2347affee4b2STejun Heo process_one_work(worker, work); 2348affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2349affee4b2STejun Heo process_scheduled_works(worker); 2350affee4b2STejun Heo } else { 2351c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2352affee4b2STejun Heo process_scheduled_works(worker); 2353affee4b2STejun Heo } 235463d95a91STejun Heo } while (keep_working(pool)); 2355affee4b2STejun Heo 2356e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2357d313dd85STejun Heo sleep: 235863d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2359e22bee78STejun Heo goto recheck; 2360d313dd85STejun Heo 2361c8e55f36STejun Heo /* 2362d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2363d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2364d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2365d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2366d565ed63STejun Heo * event. 2367c8e55f36STejun Heo */ 2368c8e55f36STejun Heo worker_enter_idle(worker); 2369c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2370d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23711da177e4SLinus Torvalds schedule(); 2372c8e55f36STejun Heo goto woke_up; 23731da177e4SLinus Torvalds } 23741da177e4SLinus Torvalds 2375e22bee78STejun Heo /** 2376e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2377111c225aSTejun Heo * @__rescuer: self 2378e22bee78STejun Heo * 2379e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2380493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2381e22bee78STejun Heo * 2382706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2383e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2384e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2385e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2386e22bee78STejun Heo * the problem rescuer solves. 2387e22bee78STejun Heo * 2388706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2389706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2390e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2391e22bee78STejun Heo * 2392e22bee78STejun Heo * This should happen rarely. 2393d185af30SYacine Belkadi * 2394d185af30SYacine Belkadi * Return: 0 2395e22bee78STejun Heo */ 2396111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2397e22bee78STejun Heo { 2398111c225aSTejun Heo struct worker *rescuer = __rescuer; 2399111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2400e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2401e22bee78STejun Heo 2402e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2403111c225aSTejun Heo 2404111c225aSTejun Heo /* 2405111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2406111c225aSTejun Heo * doesn't participate in concurrency management. 2407111c225aSTejun Heo */ 2408111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2409e22bee78STejun Heo repeat: 2410e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 24111da177e4SLinus Torvalds 2412412d32e6SMike Galbraith if (kthread_should_stop()) { 2413412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2414111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2415e22bee78STejun Heo return 0; 2416412d32e6SMike Galbraith } 24171da177e4SLinus Torvalds 2418493a1724STejun Heo /* see whether any pwq is asking for help */ 24192e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2420493a1724STejun Heo 2421493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2422493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2423493a1724STejun Heo struct pool_workqueue, mayday_node); 2424112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2425e22bee78STejun Heo struct work_struct *work, *n; 2426e22bee78STejun Heo 2427e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2428493a1724STejun Heo list_del_init(&pwq->mayday_node); 2429493a1724STejun Heo 24302e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2431e22bee78STejun Heo 2432e22bee78STejun Heo /* migrate to the target cpu if possible */ 2433f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2434b3104104SLai Jiangshan rescuer->pool = pool; 2435e22bee78STejun Heo 2436e22bee78STejun Heo /* 2437e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2438e22bee78STejun Heo * process'em. 2439e22bee78STejun Heo */ 24406183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2441bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2442112202d9STejun Heo if (get_work_pwq(work) == pwq) 2443e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2444e22bee78STejun Heo 2445e22bee78STejun Heo process_scheduled_works(rescuer); 24467576958aSTejun Heo 24477576958aSTejun Heo /* 2448d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 24497576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24507576958aSTejun Heo * and stalling the execution. 24517576958aSTejun Heo */ 245263d95a91STejun Heo if (keep_working(pool)) 245363d95a91STejun Heo wake_up_worker(pool); 24547576958aSTejun Heo 2455b3104104SLai Jiangshan rescuer->pool = NULL; 2456493a1724STejun Heo spin_unlock(&pool->lock); 24572e109a28STejun Heo spin_lock(&wq_mayday_lock); 24581da177e4SLinus Torvalds } 24591da177e4SLinus Torvalds 24602e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2461493a1724STejun Heo 2462111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2463111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2464e22bee78STejun Heo schedule(); 2465e22bee78STejun Heo goto repeat; 24661da177e4SLinus Torvalds } 24671da177e4SLinus Torvalds 2468fc2e4d70SOleg Nesterov struct wq_barrier { 2469fc2e4d70SOleg Nesterov struct work_struct work; 2470fc2e4d70SOleg Nesterov struct completion done; 2471fc2e4d70SOleg Nesterov }; 2472fc2e4d70SOleg Nesterov 2473fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2474fc2e4d70SOleg Nesterov { 2475fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2476fc2e4d70SOleg Nesterov complete(&barr->done); 2477fc2e4d70SOleg Nesterov } 2478fc2e4d70SOleg Nesterov 24794690c4abSTejun Heo /** 24804690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2481112202d9STejun Heo * @pwq: pwq to insert barrier into 24824690c4abSTejun Heo * @barr: wq_barrier to insert 2483affee4b2STejun Heo * @target: target work to attach @barr to 2484affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24854690c4abSTejun Heo * 2486affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2487affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2488affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2489affee4b2STejun Heo * cpu. 2490affee4b2STejun Heo * 2491affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2492affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2493affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2494affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2495affee4b2STejun Heo * after a work with LINKED flag set. 2496affee4b2STejun Heo * 2497affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2498112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24994690c4abSTejun Heo * 25004690c4abSTejun Heo * CONTEXT: 2501d565ed63STejun Heo * spin_lock_irq(pool->lock). 25024690c4abSTejun Heo */ 2503112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2504affee4b2STejun Heo struct wq_barrier *barr, 2505affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2506fc2e4d70SOleg Nesterov { 2507affee4b2STejun Heo struct list_head *head; 2508affee4b2STejun Heo unsigned int linked = 0; 2509affee4b2STejun Heo 2510dc186ad7SThomas Gleixner /* 2511d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2512dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2513dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2514dc186ad7SThomas Gleixner * might deadlock. 2515dc186ad7SThomas Gleixner */ 2516ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 251722df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2518fc2e4d70SOleg Nesterov init_completion(&barr->done); 251983c22520SOleg Nesterov 2520affee4b2STejun Heo /* 2521affee4b2STejun Heo * If @target is currently being executed, schedule the 2522affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2523affee4b2STejun Heo */ 2524affee4b2STejun Heo if (worker) 2525affee4b2STejun Heo head = worker->scheduled.next; 2526affee4b2STejun Heo else { 2527affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2528affee4b2STejun Heo 2529affee4b2STejun Heo head = target->entry.next; 2530affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2531affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2532affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2533affee4b2STejun Heo } 2534affee4b2STejun Heo 2535dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2536112202d9STejun Heo insert_work(pwq, &barr->work, head, 2537affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2538fc2e4d70SOleg Nesterov } 2539fc2e4d70SOleg Nesterov 254073f53c4aSTejun Heo /** 2541112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 254273f53c4aSTejun Heo * @wq: workqueue being flushed 254373f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 254473f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 254573f53c4aSTejun Heo * 2546112202d9STejun Heo * Prepare pwqs for workqueue flushing. 254773f53c4aSTejun Heo * 2548112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2549112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2550112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2551112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2552112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 255373f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 255473f53c4aSTejun Heo * 255573f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 255673f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 255773f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 255873f53c4aSTejun Heo * is returned. 255973f53c4aSTejun Heo * 2560112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 256173f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 256273f53c4aSTejun Heo * advanced to @work_color. 256373f53c4aSTejun Heo * 256473f53c4aSTejun Heo * CONTEXT: 25653c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 256673f53c4aSTejun Heo * 2567d185af30SYacine Belkadi * Return: 256873f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 256973f53c4aSTejun Heo * otherwise. 257073f53c4aSTejun Heo */ 2571112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 257273f53c4aSTejun Heo int flush_color, int work_color) 25731da177e4SLinus Torvalds { 257473f53c4aSTejun Heo bool wait = false; 257549e3cf44STejun Heo struct pool_workqueue *pwq; 25761da177e4SLinus Torvalds 257773f53c4aSTejun Heo if (flush_color >= 0) { 25786183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2579112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2580dc186ad7SThomas Gleixner } 258114441960SOleg Nesterov 258249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2583112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25841da177e4SLinus Torvalds 2585b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo if (flush_color >= 0) { 25886183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 258973f53c4aSTejun Heo 2590112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2591112202d9STejun Heo pwq->flush_color = flush_color; 2592112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 259373f53c4aSTejun Heo wait = true; 25941da177e4SLinus Torvalds } 259573f53c4aSTejun Heo } 259673f53c4aSTejun Heo 259773f53c4aSTejun Heo if (work_color >= 0) { 25986183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2599112202d9STejun Heo pwq->work_color = work_color; 260073f53c4aSTejun Heo } 260173f53c4aSTejun Heo 2602b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 26031da177e4SLinus Torvalds } 26041da177e4SLinus Torvalds 2605112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 260673f53c4aSTejun Heo complete(&wq->first_flusher->done); 260773f53c4aSTejun Heo 260873f53c4aSTejun Heo return wait; 260983c22520SOleg Nesterov } 26101da177e4SLinus Torvalds 26110fcb78c2SRolf Eike Beer /** 26121da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26130fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26141da177e4SLinus Torvalds * 2615c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2616c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 26171da177e4SLinus Torvalds */ 26187ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26191da177e4SLinus Torvalds { 262073f53c4aSTejun Heo struct wq_flusher this_flusher = { 262173f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 262273f53c4aSTejun Heo .flush_color = -1, 262373f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 262473f53c4aSTejun Heo }; 262573f53c4aSTejun Heo int next_color; 2626b1f4ec17SOleg Nesterov 26273295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 26283295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 262973f53c4aSTejun Heo 26303c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 263173f53c4aSTejun Heo 263273f53c4aSTejun Heo /* 263373f53c4aSTejun Heo * Start-to-wait phase 263473f53c4aSTejun Heo */ 263573f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 263673f53c4aSTejun Heo 263773f53c4aSTejun Heo if (next_color != wq->flush_color) { 263873f53c4aSTejun Heo /* 263973f53c4aSTejun Heo * Color space is not full. The current work_color 264073f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 264173f53c4aSTejun Heo * by one. 264273f53c4aSTejun Heo */ 26436183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 264473f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 264573f53c4aSTejun Heo wq->work_color = next_color; 264673f53c4aSTejun Heo 264773f53c4aSTejun Heo if (!wq->first_flusher) { 264873f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26496183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 265073f53c4aSTejun Heo 265173f53c4aSTejun Heo wq->first_flusher = &this_flusher; 265273f53c4aSTejun Heo 2653112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 265473f53c4aSTejun Heo wq->work_color)) { 265573f53c4aSTejun Heo /* nothing to flush, done */ 265673f53c4aSTejun Heo wq->flush_color = next_color; 265773f53c4aSTejun Heo wq->first_flusher = NULL; 265873f53c4aSTejun Heo goto out_unlock; 265973f53c4aSTejun Heo } 266073f53c4aSTejun Heo } else { 266173f53c4aSTejun Heo /* wait in queue */ 26626183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 266373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2664112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 266573f53c4aSTejun Heo } 266673f53c4aSTejun Heo } else { 266773f53c4aSTejun Heo /* 266873f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 266973f53c4aSTejun Heo * The next flush completion will assign us 267073f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 267173f53c4aSTejun Heo */ 267273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 267373f53c4aSTejun Heo } 267473f53c4aSTejun Heo 26753c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 267673f53c4aSTejun Heo 267773f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 267873f53c4aSTejun Heo 267973f53c4aSTejun Heo /* 268073f53c4aSTejun Heo * Wake-up-and-cascade phase 268173f53c4aSTejun Heo * 268273f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 268373f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 268473f53c4aSTejun Heo */ 268573f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 268673f53c4aSTejun Heo return; 268773f53c4aSTejun Heo 26883c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 268973f53c4aSTejun Heo 26904ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26914ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26924ce48b37STejun Heo goto out_unlock; 26934ce48b37STejun Heo 269473f53c4aSTejun Heo wq->first_flusher = NULL; 269573f53c4aSTejun Heo 26966183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26976183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 269873f53c4aSTejun Heo 269973f53c4aSTejun Heo while (true) { 270073f53c4aSTejun Heo struct wq_flusher *next, *tmp; 270173f53c4aSTejun Heo 270273f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 270373f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 270473f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 270573f53c4aSTejun Heo break; 270673f53c4aSTejun Heo list_del_init(&next->list); 270773f53c4aSTejun Heo complete(&next->done); 270873f53c4aSTejun Heo } 270973f53c4aSTejun Heo 27106183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 271173f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 271473f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 271573f53c4aSTejun Heo 271673f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 271773f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 271873f53c4aSTejun Heo /* 271973f53c4aSTejun Heo * Assign the same color to all overflowed 272073f53c4aSTejun Heo * flushers, advance work_color and append to 272173f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 272273f53c4aSTejun Heo * phase for these overflowed flushers. 272373f53c4aSTejun Heo */ 272473f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 272573f53c4aSTejun Heo tmp->flush_color = wq->work_color; 272673f53c4aSTejun Heo 272773f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 272873f53c4aSTejun Heo 272973f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 273073f53c4aSTejun Heo &wq->flusher_queue); 2731112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 273273f53c4aSTejun Heo } 273373f53c4aSTejun Heo 273473f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27356183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 273673f53c4aSTejun Heo break; 273773f53c4aSTejun Heo } 273873f53c4aSTejun Heo 273973f53c4aSTejun Heo /* 274073f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2741112202d9STejun Heo * the new first flusher and arm pwqs. 274273f53c4aSTejun Heo */ 27436183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27446183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 274573f53c4aSTejun Heo 274673f53c4aSTejun Heo list_del_init(&next->list); 274773f53c4aSTejun Heo wq->first_flusher = next; 274873f53c4aSTejun Heo 2749112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 275073f53c4aSTejun Heo break; 275173f53c4aSTejun Heo 275273f53c4aSTejun Heo /* 275373f53c4aSTejun Heo * Meh... this color is already done, clear first 275473f53c4aSTejun Heo * flusher and repeat cascading. 275573f53c4aSTejun Heo */ 275673f53c4aSTejun Heo wq->first_flusher = NULL; 275773f53c4aSTejun Heo } 275873f53c4aSTejun Heo 275973f53c4aSTejun Heo out_unlock: 27603c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27611da177e4SLinus Torvalds } 2762ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27631da177e4SLinus Torvalds 27649c5a2ba7STejun Heo /** 27659c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27669c5a2ba7STejun Heo * @wq: workqueue to drain 27679c5a2ba7STejun Heo * 27689c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27699c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27709c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27719c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27729c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27739c5a2ba7STejun Heo * takes too long. 27749c5a2ba7STejun Heo */ 27759c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27769c5a2ba7STejun Heo { 27779c5a2ba7STejun Heo unsigned int flush_cnt = 0; 277849e3cf44STejun Heo struct pool_workqueue *pwq; 27799c5a2ba7STejun Heo 27809c5a2ba7STejun Heo /* 27819c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27829c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2783618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27849c5a2ba7STejun Heo */ 278587fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27869c5a2ba7STejun Heo if (!wq->nr_drainers++) 2787618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 278887fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27899c5a2ba7STejun Heo reflush: 27909c5a2ba7STejun Heo flush_workqueue(wq); 27919c5a2ba7STejun Heo 2792b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 279376af4d93STejun Heo 279449e3cf44STejun Heo for_each_pwq(pwq, wq) { 2795fa2563e4SThomas Tuttle bool drained; 27969c5a2ba7STejun Heo 2797b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2798112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2799b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2800fa2563e4SThomas Tuttle 2801fa2563e4SThomas Tuttle if (drained) 28029c5a2ba7STejun Heo continue; 28039c5a2ba7STejun Heo 28049c5a2ba7STejun Heo if (++flush_cnt == 10 || 28059c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2806c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 28079c5a2ba7STejun Heo wq->name, flush_cnt); 280876af4d93STejun Heo 2809b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 28109c5a2ba7STejun Heo goto reflush; 28119c5a2ba7STejun Heo } 28129c5a2ba7STejun Heo 28139c5a2ba7STejun Heo if (!--wq->nr_drainers) 2814618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 281587fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28169c5a2ba7STejun Heo } 28179c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28189c5a2ba7STejun Heo 2819606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2820baf59022STejun Heo { 2821baf59022STejun Heo struct worker *worker = NULL; 2822c9e7cf27STejun Heo struct worker_pool *pool; 2823112202d9STejun Heo struct pool_workqueue *pwq; 2824baf59022STejun Heo 2825baf59022STejun Heo might_sleep(); 2826baf59022STejun Heo 2827fa1b54e6STejun Heo local_irq_disable(); 2828fa1b54e6STejun Heo pool = get_work_pool(work); 2829fa1b54e6STejun Heo if (!pool) { 2830fa1b54e6STejun Heo local_irq_enable(); 2831fa1b54e6STejun Heo return false; 2832fa1b54e6STejun Heo } 2833fa1b54e6STejun Heo 2834fa1b54e6STejun Heo spin_lock(&pool->lock); 28350b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2836112202d9STejun Heo pwq = get_work_pwq(work); 2837112202d9STejun Heo if (pwq) { 2838112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2839baf59022STejun Heo goto already_gone; 2840606a5020STejun Heo } else { 2841c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2842baf59022STejun Heo if (!worker) 2843baf59022STejun Heo goto already_gone; 2844112202d9STejun Heo pwq = worker->current_pwq; 2845606a5020STejun Heo } 2846baf59022STejun Heo 2847112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2848d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2849baf59022STejun Heo 2850e159489bSTejun Heo /* 2851e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2852e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2853e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2854e159489bSTejun Heo * access. 2855e159489bSTejun Heo */ 2856493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2857112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2858e159489bSTejun Heo else 2859112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2860112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2861e159489bSTejun Heo 2862baf59022STejun Heo return true; 2863baf59022STejun Heo already_gone: 2864d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2865baf59022STejun Heo return false; 2866baf59022STejun Heo } 2867baf59022STejun Heo 2868db700897SOleg Nesterov /** 2869401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2870401a8d04STejun Heo * @work: the work to flush 2871db700897SOleg Nesterov * 2872606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2873606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2874401a8d04STejun Heo * 2875d185af30SYacine Belkadi * Return: 2876401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2877401a8d04STejun Heo * %false if it was already idle. 2878db700897SOleg Nesterov */ 2879401a8d04STejun Heo bool flush_work(struct work_struct *work) 2880db700897SOleg Nesterov { 288112997d1aSBjorn Helgaas struct wq_barrier barr; 288212997d1aSBjorn Helgaas 28830976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28840976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28850976dfc1SStephen Boyd 288612997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 288712997d1aSBjorn Helgaas wait_for_completion(&barr.done); 288812997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 288912997d1aSBjorn Helgaas return true; 289012997d1aSBjorn Helgaas } else { 289112997d1aSBjorn Helgaas return false; 289212997d1aSBjorn Helgaas } 2893606a5020STejun Heo } 2894db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2895db700897SOleg Nesterov 289636e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2897401a8d04STejun Heo { 2898bbb68dfaSTejun Heo unsigned long flags; 28991f1f642eSOleg Nesterov int ret; 29001f1f642eSOleg Nesterov 29011f1f642eSOleg Nesterov do { 2902bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2903bbb68dfaSTejun Heo /* 2904bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2905bbb68dfaSTejun Heo * would be waiting for before retrying. 2906bbb68dfaSTejun Heo */ 2907bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2908606a5020STejun Heo flush_work(work); 29091f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29101f1f642eSOleg Nesterov 2911bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2912bbb68dfaSTejun Heo mark_work_canceling(work); 2913bbb68dfaSTejun Heo local_irq_restore(flags); 2914bbb68dfaSTejun Heo 2915606a5020STejun Heo flush_work(work); 29167a22ad75STejun Heo clear_work_data(work); 29171f1f642eSOleg Nesterov return ret; 29181f1f642eSOleg Nesterov } 29191f1f642eSOleg Nesterov 29206e84d644SOleg Nesterov /** 2921401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2922401a8d04STejun Heo * @work: the work to cancel 29236e84d644SOleg Nesterov * 2924401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2925401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2926401a8d04STejun Heo * another workqueue. On return from this function, @work is 2927401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29281f1f642eSOleg Nesterov * 2929401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2930401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29316e84d644SOleg Nesterov * 2932401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29336e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2934401a8d04STejun Heo * 2935d185af30SYacine Belkadi * Return: 2936401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29376e84d644SOleg Nesterov */ 2938401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29396e84d644SOleg Nesterov { 294036e227d2STejun Heo return __cancel_work_timer(work, false); 2941b89deed3SOleg Nesterov } 294228e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2943b89deed3SOleg Nesterov 29446e84d644SOleg Nesterov /** 2945401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2946401a8d04STejun Heo * @dwork: the delayed work to flush 29476e84d644SOleg Nesterov * 2948401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2949401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2950401a8d04STejun Heo * considers the last queueing instance of @dwork. 29511f1f642eSOleg Nesterov * 2952d185af30SYacine Belkadi * Return: 2953401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2954401a8d04STejun Heo * %false if it was already idle. 29556e84d644SOleg Nesterov */ 2956401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2957401a8d04STejun Heo { 29588930cabaSTejun Heo local_irq_disable(); 2959401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 296060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29618930cabaSTejun Heo local_irq_enable(); 2962401a8d04STejun Heo return flush_work(&dwork->work); 2963401a8d04STejun Heo } 2964401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2965401a8d04STejun Heo 2966401a8d04STejun Heo /** 296757b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 296857b30ae7STejun Heo * @dwork: delayed_work to cancel 296909383498STejun Heo * 2970d185af30SYacine Belkadi * Kill off a pending delayed_work. 2971d185af30SYacine Belkadi * 2972d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2973d185af30SYacine Belkadi * pending. 2974d185af30SYacine Belkadi * 2975d185af30SYacine Belkadi * Note: 2976d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2977d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2978d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 297909383498STejun Heo * 298057b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 298109383498STejun Heo */ 298257b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 298309383498STejun Heo { 298457b30ae7STejun Heo unsigned long flags; 298557b30ae7STejun Heo int ret; 298657b30ae7STejun Heo 298757b30ae7STejun Heo do { 298857b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 298957b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 299057b30ae7STejun Heo 299157b30ae7STejun Heo if (unlikely(ret < 0)) 299257b30ae7STejun Heo return false; 299357b30ae7STejun Heo 29947c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29957c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 299657b30ae7STejun Heo local_irq_restore(flags); 2997c0158ca6SDan Magenheimer return ret; 299809383498STejun Heo } 299957b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 300009383498STejun Heo 300109383498STejun Heo /** 3002401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3003401a8d04STejun Heo * @dwork: the delayed work cancel 3004401a8d04STejun Heo * 3005401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3006401a8d04STejun Heo * 3007d185af30SYacine Belkadi * Return: 3008401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3009401a8d04STejun Heo */ 3010401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30116e84d644SOleg Nesterov { 301236e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30136e84d644SOleg Nesterov } 3014f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30151da177e4SLinus Torvalds 30160fcb78c2SRolf Eike Beer /** 301731ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3018b6136773SAndrew Morton * @func: the function to call 3019b6136773SAndrew Morton * 302031ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 302131ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3022b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 302331ddd871STejun Heo * 3024d185af30SYacine Belkadi * Return: 302531ddd871STejun Heo * 0 on success, -errno on failure. 3026b6136773SAndrew Morton */ 302765f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 302815316ba8SChristoph Lameter { 302915316ba8SChristoph Lameter int cpu; 303038f51568SNamhyung Kim struct work_struct __percpu *works; 303115316ba8SChristoph Lameter 3032b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3033b6136773SAndrew Morton if (!works) 303415316ba8SChristoph Lameter return -ENOMEM; 3035b6136773SAndrew Morton 303695402b38SGautham R Shenoy get_online_cpus(); 303793981800STejun Heo 303815316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30399bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30409bfb1839SIngo Molnar 30419bfb1839SIngo Molnar INIT_WORK(work, func); 30428de6d308SOleg Nesterov schedule_work_on(cpu, work); 304315316ba8SChristoph Lameter } 304493981800STejun Heo 304593981800STejun Heo for_each_online_cpu(cpu) 30468616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 304793981800STejun Heo 304895402b38SGautham R Shenoy put_online_cpus(); 3049b6136773SAndrew Morton free_percpu(works); 305015316ba8SChristoph Lameter return 0; 305115316ba8SChristoph Lameter } 305215316ba8SChristoph Lameter 3053eef6a7d5SAlan Stern /** 3054eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3055eef6a7d5SAlan Stern * 3056eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3057eef6a7d5SAlan Stern * completion. 3058eef6a7d5SAlan Stern * 3059eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3060eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3061eef6a7d5SAlan Stern * will lead to deadlock: 3062eef6a7d5SAlan Stern * 3063eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3064eef6a7d5SAlan Stern * a lock held by your code or its caller. 3065eef6a7d5SAlan Stern * 3066eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3067eef6a7d5SAlan Stern * 3068eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3069eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3070eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3071eef6a7d5SAlan Stern * 3072eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3073eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3074eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3075eef6a7d5SAlan Stern * cancel_work_sync() instead. 3076eef6a7d5SAlan Stern */ 30771da177e4SLinus Torvalds void flush_scheduled_work(void) 30781da177e4SLinus Torvalds { 3079d320c038STejun Heo flush_workqueue(system_wq); 30801da177e4SLinus Torvalds } 3081ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30821da177e4SLinus Torvalds 30831da177e4SLinus Torvalds /** 30841fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30851fa44ecaSJames Bottomley * @fn: the function to execute 30861fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30871fa44ecaSJames Bottomley * be available when the work executes) 30881fa44ecaSJames Bottomley * 30891fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30901fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30911fa44ecaSJames Bottomley * 3092d185af30SYacine Belkadi * Return: 0 - function was executed 30931fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30941fa44ecaSJames Bottomley */ 309565f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30961fa44ecaSJames Bottomley { 30971fa44ecaSJames Bottomley if (!in_interrupt()) { 309865f27f38SDavid Howells fn(&ew->work); 30991fa44ecaSJames Bottomley return 0; 31001fa44ecaSJames Bottomley } 31011fa44ecaSJames Bottomley 310265f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31031fa44ecaSJames Bottomley schedule_work(&ew->work); 31041fa44ecaSJames Bottomley 31051fa44ecaSJames Bottomley return 1; 31061fa44ecaSJames Bottomley } 31071fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31081fa44ecaSJames Bottomley 3109226223abSTejun Heo #ifdef CONFIG_SYSFS 3110226223abSTejun Heo /* 3111226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3112226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3113226223abSTejun Heo * following attributes. 3114226223abSTejun Heo * 3115226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3116226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3117226223abSTejun Heo * 3118226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3119226223abSTejun Heo * 3120226223abSTejun Heo * id RO int : the associated pool ID 3121226223abSTejun Heo * nice RW int : nice value of the workers 3122226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3123226223abSTejun Heo */ 3124226223abSTejun Heo struct wq_device { 3125226223abSTejun Heo struct workqueue_struct *wq; 3126226223abSTejun Heo struct device dev; 3127226223abSTejun Heo }; 3128226223abSTejun Heo 3129226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3130226223abSTejun Heo { 3131226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3132226223abSTejun Heo 3133226223abSTejun Heo return wq_dev->wq; 3134226223abSTejun Heo } 3135226223abSTejun Heo 31361a6661daSGreg Kroah-Hartman static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 31371a6661daSGreg Kroah-Hartman 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", (bool)!(wq->flags & WQ_UNBOUND)); 3142226223abSTejun Heo } 31431a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RO(per_cpu); 3144226223abSTejun Heo 31451a6661daSGreg Kroah-Hartman static ssize_t max_active_show(struct device *dev, 3146226223abSTejun Heo struct device_attribute *attr, char *buf) 3147226223abSTejun Heo { 3148226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3149226223abSTejun Heo 3150226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3151226223abSTejun Heo } 3152226223abSTejun Heo 31531a6661daSGreg Kroah-Hartman static ssize_t max_active_store(struct device *dev, 31541a6661daSGreg Kroah-Hartman struct device_attribute *attr, const char *buf, 31551a6661daSGreg Kroah-Hartman size_t count) 3156226223abSTejun Heo { 3157226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3158226223abSTejun Heo int val; 3159226223abSTejun Heo 3160226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3161226223abSTejun Heo return -EINVAL; 3162226223abSTejun Heo 3163226223abSTejun Heo workqueue_set_max_active(wq, val); 3164226223abSTejun Heo return count; 3165226223abSTejun Heo } 31661a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RW(max_active); 3167226223abSTejun Heo 31681a6661daSGreg Kroah-Hartman static struct attribute *wq_sysfs_attrs[] = { 31691a6661daSGreg Kroah-Hartman &dev_attr_per_cpu.attr, 31701a6661daSGreg Kroah-Hartman &dev_attr_max_active.attr, 31711a6661daSGreg Kroah-Hartman NULL, 3172226223abSTejun Heo }; 31731a6661daSGreg Kroah-Hartman ATTRIBUTE_GROUPS(wq_sysfs); 3174226223abSTejun Heo 3175d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3176226223abSTejun Heo struct device_attribute *attr, char *buf) 3177226223abSTejun Heo { 3178226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3179d55262c4STejun Heo const char *delim = ""; 3180d55262c4STejun Heo int node, written = 0; 3181226223abSTejun Heo 3182226223abSTejun Heo rcu_read_lock_sched(); 3183d55262c4STejun Heo for_each_node(node) { 3184d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3185d55262c4STejun Heo "%s%d:%d", delim, node, 3186d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3187d55262c4STejun Heo delim = " "; 3188d55262c4STejun Heo } 3189d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3190226223abSTejun Heo rcu_read_unlock_sched(); 3191226223abSTejun Heo 3192226223abSTejun Heo return written; 3193226223abSTejun Heo } 3194226223abSTejun Heo 3195226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3196226223abSTejun Heo char *buf) 3197226223abSTejun Heo { 3198226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3199226223abSTejun Heo int written; 3200226223abSTejun Heo 32016029a918STejun Heo mutex_lock(&wq->mutex); 32026029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 32036029a918STejun Heo mutex_unlock(&wq->mutex); 3204226223abSTejun Heo 3205226223abSTejun Heo return written; 3206226223abSTejun Heo } 3207226223abSTejun Heo 3208226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3209226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3210226223abSTejun Heo { 3211226223abSTejun Heo struct workqueue_attrs *attrs; 3212226223abSTejun Heo 3213226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3214226223abSTejun Heo if (!attrs) 3215226223abSTejun Heo return NULL; 3216226223abSTejun Heo 32176029a918STejun Heo mutex_lock(&wq->mutex); 32186029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 32196029a918STejun Heo mutex_unlock(&wq->mutex); 3220226223abSTejun Heo return attrs; 3221226223abSTejun Heo } 3222226223abSTejun Heo 3223226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3224226223abSTejun Heo const char *buf, size_t count) 3225226223abSTejun Heo { 3226226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3227226223abSTejun Heo struct workqueue_attrs *attrs; 3228226223abSTejun Heo int ret; 3229226223abSTejun Heo 3230226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3231226223abSTejun Heo if (!attrs) 3232226223abSTejun Heo return -ENOMEM; 3233226223abSTejun Heo 3234226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 323514481842SDongsheng Yang attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 3236226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3237226223abSTejun Heo else 3238226223abSTejun Heo ret = -EINVAL; 3239226223abSTejun Heo 3240226223abSTejun Heo free_workqueue_attrs(attrs); 3241226223abSTejun Heo return ret ?: count; 3242226223abSTejun Heo } 3243226223abSTejun Heo 3244226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3245226223abSTejun Heo struct device_attribute *attr, char *buf) 3246226223abSTejun Heo { 3247226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3248226223abSTejun Heo int written; 3249226223abSTejun Heo 32506029a918STejun Heo mutex_lock(&wq->mutex); 32516029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 32526029a918STejun Heo mutex_unlock(&wq->mutex); 3253226223abSTejun Heo 3254226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3255226223abSTejun Heo return written; 3256226223abSTejun Heo } 3257226223abSTejun Heo 3258226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3259226223abSTejun Heo struct device_attribute *attr, 3260226223abSTejun Heo const char *buf, size_t count) 3261226223abSTejun Heo { 3262226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3263226223abSTejun Heo struct workqueue_attrs *attrs; 3264226223abSTejun Heo int ret; 3265226223abSTejun Heo 3266226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3267226223abSTejun Heo if (!attrs) 3268226223abSTejun Heo return -ENOMEM; 3269226223abSTejun Heo 3270226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3271226223abSTejun Heo if (!ret) 3272226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3273226223abSTejun Heo 3274226223abSTejun Heo free_workqueue_attrs(attrs); 3275226223abSTejun Heo return ret ?: count; 3276226223abSTejun Heo } 3277226223abSTejun Heo 3278d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3279d55262c4STejun Heo char *buf) 3280d55262c4STejun Heo { 3281d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3282d55262c4STejun Heo int written; 3283d55262c4STejun Heo 3284d55262c4STejun Heo mutex_lock(&wq->mutex); 3285d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3286d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3287d55262c4STejun Heo mutex_unlock(&wq->mutex); 3288d55262c4STejun Heo 3289d55262c4STejun Heo return written; 3290d55262c4STejun Heo } 3291d55262c4STejun Heo 3292d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3293d55262c4STejun Heo const char *buf, size_t count) 3294d55262c4STejun Heo { 3295d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3296d55262c4STejun Heo struct workqueue_attrs *attrs; 3297d55262c4STejun Heo int v, ret; 3298d55262c4STejun Heo 3299d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3300d55262c4STejun Heo if (!attrs) 3301d55262c4STejun Heo return -ENOMEM; 3302d55262c4STejun Heo 3303d55262c4STejun Heo ret = -EINVAL; 3304d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3305d55262c4STejun Heo attrs->no_numa = !v; 3306d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3307d55262c4STejun Heo } 3308d55262c4STejun Heo 3309d55262c4STejun Heo free_workqueue_attrs(attrs); 3310d55262c4STejun Heo return ret ?: count; 3311d55262c4STejun Heo } 3312d55262c4STejun Heo 3313226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3314d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3315226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3316226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3317d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3318226223abSTejun Heo __ATTR_NULL, 3319226223abSTejun Heo }; 3320226223abSTejun Heo 3321226223abSTejun Heo static struct bus_type wq_subsys = { 3322226223abSTejun Heo .name = "workqueue", 33231a6661daSGreg Kroah-Hartman .dev_groups = wq_sysfs_groups, 3324226223abSTejun Heo }; 3325226223abSTejun Heo 3326226223abSTejun Heo static int __init wq_sysfs_init(void) 3327226223abSTejun Heo { 3328226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3329226223abSTejun Heo } 3330226223abSTejun Heo core_initcall(wq_sysfs_init); 3331226223abSTejun Heo 3332226223abSTejun Heo static void wq_device_release(struct device *dev) 3333226223abSTejun Heo { 3334226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3335226223abSTejun Heo 3336226223abSTejun Heo kfree(wq_dev); 3337226223abSTejun Heo } 3338226223abSTejun Heo 3339226223abSTejun Heo /** 3340226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3341226223abSTejun Heo * @wq: the workqueue to register 3342226223abSTejun Heo * 3343226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3344226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3345226223abSTejun Heo * which is the preferred method. 3346226223abSTejun Heo * 3347226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3348226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3349226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3350226223abSTejun Heo * attributes. 3351226223abSTejun Heo * 3352d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. 3353226223abSTejun Heo */ 3354226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3355226223abSTejun Heo { 3356226223abSTejun Heo struct wq_device *wq_dev; 3357226223abSTejun Heo int ret; 3358226223abSTejun Heo 3359226223abSTejun Heo /* 3360226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3361226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3362226223abSTejun Heo * workqueues. 3363226223abSTejun Heo */ 3364226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3365226223abSTejun Heo return -EINVAL; 3366226223abSTejun Heo 3367226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3368226223abSTejun Heo if (!wq_dev) 3369226223abSTejun Heo return -ENOMEM; 3370226223abSTejun Heo 3371226223abSTejun Heo wq_dev->wq = wq; 3372226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3373226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3374226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3375226223abSTejun Heo 3376226223abSTejun Heo /* 3377226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3378226223abSTejun Heo * everything is ready. 3379226223abSTejun Heo */ 3380226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3381226223abSTejun Heo 3382226223abSTejun Heo ret = device_register(&wq_dev->dev); 3383226223abSTejun Heo if (ret) { 3384226223abSTejun Heo kfree(wq_dev); 3385226223abSTejun Heo wq->wq_dev = NULL; 3386226223abSTejun Heo return ret; 3387226223abSTejun Heo } 3388226223abSTejun Heo 3389226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3390226223abSTejun Heo struct device_attribute *attr; 3391226223abSTejun Heo 3392226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3393226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3394226223abSTejun Heo if (ret) { 3395226223abSTejun Heo device_unregister(&wq_dev->dev); 3396226223abSTejun Heo wq->wq_dev = NULL; 3397226223abSTejun Heo return ret; 3398226223abSTejun Heo } 3399226223abSTejun Heo } 3400226223abSTejun Heo } 3401226223abSTejun Heo 3402226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3403226223abSTejun Heo return 0; 3404226223abSTejun Heo } 3405226223abSTejun Heo 3406226223abSTejun Heo /** 3407226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3408226223abSTejun Heo * @wq: the workqueue to unregister 3409226223abSTejun Heo * 3410226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3411226223abSTejun Heo */ 3412226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3413226223abSTejun Heo { 3414226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3415226223abSTejun Heo 3416226223abSTejun Heo if (!wq->wq_dev) 3417226223abSTejun Heo return; 3418226223abSTejun Heo 3419226223abSTejun Heo wq->wq_dev = NULL; 3420226223abSTejun Heo device_unregister(&wq_dev->dev); 3421226223abSTejun Heo } 3422226223abSTejun Heo #else /* CONFIG_SYSFS */ 3423226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3424226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3425226223abSTejun Heo 34267a4e344cSTejun Heo /** 34277a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 34287a4e344cSTejun Heo * @attrs: workqueue_attrs to free 34297a4e344cSTejun Heo * 34307a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 34317a4e344cSTejun Heo */ 34327a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 34337a4e344cSTejun Heo { 34347a4e344cSTejun Heo if (attrs) { 34357a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 34367a4e344cSTejun Heo kfree(attrs); 34377a4e344cSTejun Heo } 34387a4e344cSTejun Heo } 34397a4e344cSTejun Heo 34407a4e344cSTejun Heo /** 34417a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 34427a4e344cSTejun Heo * @gfp_mask: allocation mask to use 34437a4e344cSTejun Heo * 34447a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3445d185af30SYacine Belkadi * return it. 3446d185af30SYacine Belkadi * 3447d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 34487a4e344cSTejun Heo */ 34497a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 34507a4e344cSTejun Heo { 34517a4e344cSTejun Heo struct workqueue_attrs *attrs; 34527a4e344cSTejun Heo 34537a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 34547a4e344cSTejun Heo if (!attrs) 34557a4e344cSTejun Heo goto fail; 34567a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 34577a4e344cSTejun Heo goto fail; 34587a4e344cSTejun Heo 345913e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 34607a4e344cSTejun Heo return attrs; 34617a4e344cSTejun Heo fail: 34627a4e344cSTejun Heo free_workqueue_attrs(attrs); 34637a4e344cSTejun Heo return NULL; 34647a4e344cSTejun Heo } 34657a4e344cSTejun Heo 346629c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 346729c91e99STejun Heo const struct workqueue_attrs *from) 346829c91e99STejun Heo { 346929c91e99STejun Heo to->nice = from->nice; 347029c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 34712865a8fbSShaohua Li /* 34722865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 34732865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 34742865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 34752865a8fbSShaohua Li */ 34762865a8fbSShaohua Li to->no_numa = from->no_numa; 347729c91e99STejun Heo } 347829c91e99STejun Heo 347929c91e99STejun Heo /* hash value of the content of @attr */ 348029c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 348129c91e99STejun Heo { 348229c91e99STejun Heo u32 hash = 0; 348329c91e99STejun Heo 348429c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 348513e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 348613e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 348729c91e99STejun Heo return hash; 348829c91e99STejun Heo } 348929c91e99STejun Heo 349029c91e99STejun Heo /* content equality test */ 349129c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 349229c91e99STejun Heo const struct workqueue_attrs *b) 349329c91e99STejun Heo { 349429c91e99STejun Heo if (a->nice != b->nice) 349529c91e99STejun Heo return false; 349629c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 349729c91e99STejun Heo return false; 349829c91e99STejun Heo return true; 349929c91e99STejun Heo } 350029c91e99STejun Heo 35017a4e344cSTejun Heo /** 35027a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 35037a4e344cSTejun Heo * @pool: worker_pool to initialize 35047a4e344cSTejun Heo * 35057a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 3506d185af30SYacine Belkadi * 3507d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 350829c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 350929c91e99STejun Heo * on @pool safely to release it. 35107a4e344cSTejun Heo */ 35117a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 35124e1a1f9aSTejun Heo { 35134e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 351429c91e99STejun Heo pool->id = -1; 351529c91e99STejun Heo pool->cpu = -1; 3516f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 35174e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 35184e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 35194e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 35204e1a1f9aSTejun Heo hash_init(pool->busy_hash); 35214e1a1f9aSTejun Heo 35224e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 35234e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 35244e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 35254e1a1f9aSTejun Heo 35264e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 35274e1a1f9aSTejun Heo (unsigned long)pool); 35284e1a1f9aSTejun Heo 35294e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3530bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3531822d8405STejun Heo idr_init(&pool->worker_idr); 35327a4e344cSTejun Heo 353329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 353429c91e99STejun Heo pool->refcnt = 1; 353529c91e99STejun Heo 353629c91e99STejun Heo /* shouldn't fail above this point */ 35377a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 35387a4e344cSTejun Heo if (!pool->attrs) 35397a4e344cSTejun Heo return -ENOMEM; 35407a4e344cSTejun Heo return 0; 35414e1a1f9aSTejun Heo } 35424e1a1f9aSTejun Heo 354329c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 354429c91e99STejun Heo { 354529c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 354629c91e99STejun Heo 3547822d8405STejun Heo idr_destroy(&pool->worker_idr); 354829c91e99STejun Heo free_workqueue_attrs(pool->attrs); 354929c91e99STejun Heo kfree(pool); 355029c91e99STejun Heo } 355129c91e99STejun Heo 355229c91e99STejun Heo /** 355329c91e99STejun Heo * put_unbound_pool - put a worker_pool 355429c91e99STejun Heo * @pool: worker_pool to put 355529c91e99STejun Heo * 355629c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3557c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3558c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3559c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3560a892caccSTejun Heo * 3561a892caccSTejun Heo * Should be called with wq_pool_mutex held. 356229c91e99STejun Heo */ 356329c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 356429c91e99STejun Heo { 356529c91e99STejun Heo struct worker *worker; 356629c91e99STejun Heo 3567a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3568a892caccSTejun Heo 3569a892caccSTejun Heo if (--pool->refcnt) 357029c91e99STejun Heo return; 357129c91e99STejun Heo 357229c91e99STejun Heo /* sanity checks */ 357329c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3574a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 357529c91e99STejun Heo return; 357629c91e99STejun Heo 357729c91e99STejun Heo /* release id and unhash */ 357829c91e99STejun Heo if (pool->id >= 0) 357929c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 358029c91e99STejun Heo hash_del(&pool->hash_node); 358129c91e99STejun Heo 3582c5aa87bbSTejun Heo /* 3583c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3584c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3585c5aa87bbSTejun Heo * manager_mutex. 3586c5aa87bbSTejun Heo */ 358729c91e99STejun Heo mutex_lock(&pool->manager_arb); 3588cd549687STejun Heo mutex_lock(&pool->manager_mutex); 358929c91e99STejun Heo spin_lock_irq(&pool->lock); 359029c91e99STejun Heo 359129c91e99STejun Heo while ((worker = first_worker(pool))) 359229c91e99STejun Heo destroy_worker(worker); 359329c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 359429c91e99STejun Heo 359529c91e99STejun Heo spin_unlock_irq(&pool->lock); 3596cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 359729c91e99STejun Heo mutex_unlock(&pool->manager_arb); 359829c91e99STejun Heo 359929c91e99STejun Heo /* shut down the timers */ 360029c91e99STejun Heo del_timer_sync(&pool->idle_timer); 360129c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 360229c91e99STejun Heo 360329c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 360429c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 360529c91e99STejun Heo } 360629c91e99STejun Heo 360729c91e99STejun Heo /** 360829c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 360929c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 361029c91e99STejun Heo * 361129c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 361229c91e99STejun Heo * reference count and return it. If there already is a matching 361329c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3614d185af30SYacine Belkadi * create a new one. 3615a892caccSTejun Heo * 3616a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3617d185af30SYacine Belkadi * 3618d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3619d185af30SYacine Belkadi * On failure, %NULL. 362029c91e99STejun Heo */ 362129c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 362229c91e99STejun Heo { 362329c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 362429c91e99STejun Heo struct worker_pool *pool; 3625f3f90ad4STejun Heo int node; 362629c91e99STejun Heo 3627a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 362829c91e99STejun Heo 362929c91e99STejun Heo /* do we already have a matching pool? */ 363029c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 363129c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 363229c91e99STejun Heo pool->refcnt++; 363329c91e99STejun Heo goto out_unlock; 363429c91e99STejun Heo } 363529c91e99STejun Heo } 363629c91e99STejun Heo 363729c91e99STejun Heo /* nope, create a new one */ 363829c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 363929c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 364029c91e99STejun Heo goto fail; 364129c91e99STejun Heo 364212ee4fc6SLai Jiangshan if (workqueue_freezing) 364312ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 364412ee4fc6SLai Jiangshan 36458864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 364629c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 364729c91e99STejun Heo 36482865a8fbSShaohua Li /* 36492865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36502865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36512865a8fbSShaohua Li */ 36522865a8fbSShaohua Li pool->attrs->no_numa = false; 36532865a8fbSShaohua Li 3654f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3655f3f90ad4STejun Heo if (wq_numa_enabled) { 3656f3f90ad4STejun Heo for_each_node(node) { 3657f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3658f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3659f3f90ad4STejun Heo pool->node = node; 3660f3f90ad4STejun Heo break; 3661f3f90ad4STejun Heo } 3662f3f90ad4STejun Heo } 3663f3f90ad4STejun Heo } 3664f3f90ad4STejun Heo 366529c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 366629c91e99STejun Heo goto fail; 366729c91e99STejun Heo 366829c91e99STejun Heo /* create and start the initial worker */ 3669ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 367029c91e99STejun Heo goto fail; 367129c91e99STejun Heo 367229c91e99STejun Heo /* install */ 367329c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 367429c91e99STejun Heo out_unlock: 367529c91e99STejun Heo return pool; 367629c91e99STejun Heo fail: 367729c91e99STejun Heo if (pool) 367829c91e99STejun Heo put_unbound_pool(pool); 367929c91e99STejun Heo return NULL; 368029c91e99STejun Heo } 368129c91e99STejun Heo 36828864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36838864b4e5STejun Heo { 36848864b4e5STejun Heo kmem_cache_free(pwq_cache, 36858864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36868864b4e5STejun Heo } 36878864b4e5STejun Heo 36888864b4e5STejun Heo /* 36898864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36908864b4e5STejun Heo * and needs to be destroyed. 36918864b4e5STejun Heo */ 36928864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36938864b4e5STejun Heo { 36948864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36958864b4e5STejun Heo unbound_release_work); 36968864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36978864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3698bc0caf09STejun Heo bool is_last; 36998864b4e5STejun Heo 37008864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 37018864b4e5STejun Heo return; 37028864b4e5STejun Heo 370375ccf595STejun Heo /* 37043c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 370575ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 370675ccf595STejun Heo * and consistent with the linking path. 370775ccf595STejun Heo */ 37083c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 37098864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3710bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 37113c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 37128864b4e5STejun Heo 3713a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 37148864b4e5STejun Heo put_unbound_pool(pool); 3715a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3716a892caccSTejun Heo 37178864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 37188864b4e5STejun Heo 37198864b4e5STejun Heo /* 37208864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 37218864b4e5STejun Heo * is gonna access it anymore. Free it. 37228864b4e5STejun Heo */ 37236029a918STejun Heo if (is_last) { 37246029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 37258864b4e5STejun Heo kfree(wq); 37268864b4e5STejun Heo } 37276029a918STejun Heo } 37288864b4e5STejun Heo 37290fbd95aaSTejun Heo /** 3730699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37310fbd95aaSTejun Heo * @pwq: target pool_workqueue 37320fbd95aaSTejun Heo * 3733699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3734699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3735699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37360fbd95aaSTejun Heo */ 3737699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37380fbd95aaSTejun Heo { 3739699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3740699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3741699ce097STejun Heo 3742699ce097STejun Heo /* for @wq->saved_max_active */ 3743a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3744699ce097STejun Heo 3745699ce097STejun Heo /* fast exit for non-freezable wqs */ 3746699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3747699ce097STejun Heo return; 3748699ce097STejun Heo 3749a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3750699ce097STejun Heo 3751699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3752699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37530fbd95aaSTejun Heo 37540fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 37550fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 37560fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3757951a078aSLai Jiangshan 3758951a078aSLai Jiangshan /* 3759951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3760951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3761951a078aSLai Jiangshan */ 3762951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3763699ce097STejun Heo } else { 3764699ce097STejun Heo pwq->max_active = 0; 3765699ce097STejun Heo } 3766699ce097STejun Heo 3767a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 37680fbd95aaSTejun Heo } 37690fbd95aaSTejun Heo 3770e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3771f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3772f147f29eSTejun Heo struct worker_pool *pool) 3773d2c1d404STejun Heo { 3774d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3775d2c1d404STejun Heo 3776e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3777e50aba9aSTejun Heo 3778d2c1d404STejun Heo pwq->pool = pool; 3779d2c1d404STejun Heo pwq->wq = wq; 3780d2c1d404STejun Heo pwq->flush_color = -1; 37818864b4e5STejun Heo pwq->refcnt = 1; 3782d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 37831befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3784d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37858864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3786f147f29eSTejun Heo } 3787d2c1d404STejun Heo 3788f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37891befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3790f147f29eSTejun Heo { 3791f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3792f147f29eSTejun Heo 3793f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 379475ccf595STejun Heo 37951befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37961befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37971befcf30STejun Heo return; 37981befcf30STejun Heo 3799983ca25eSTejun Heo /* 3800983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 38013c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3802983ca25eSTejun Heo */ 380375ccf595STejun Heo pwq->work_color = wq->work_color; 3804983ca25eSTejun Heo 3805983ca25eSTejun Heo /* sync max_active to the current setting */ 3806983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3807983ca25eSTejun Heo 3808983ca25eSTejun Heo /* link in @pwq */ 38099e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3810df2d5ae4STejun Heo } 38116029a918STejun Heo 3812f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3813f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3814f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3815f147f29eSTejun Heo { 3816f147f29eSTejun Heo struct worker_pool *pool; 3817f147f29eSTejun Heo struct pool_workqueue *pwq; 3818f147f29eSTejun Heo 3819f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3820f147f29eSTejun Heo 3821f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3822f147f29eSTejun Heo if (!pool) 3823f147f29eSTejun Heo return NULL; 3824f147f29eSTejun Heo 3825e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3826f147f29eSTejun Heo if (!pwq) { 3827f147f29eSTejun Heo put_unbound_pool(pool); 3828f147f29eSTejun Heo return NULL; 3829f147f29eSTejun Heo } 3830f147f29eSTejun Heo 3831f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3832f147f29eSTejun Heo return pwq; 3833d2c1d404STejun Heo } 3834d2c1d404STejun Heo 38354c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 38364c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 38374c16bd32STejun Heo { 38384c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38394c16bd32STejun Heo 38404c16bd32STejun Heo if (pwq) { 38414c16bd32STejun Heo put_unbound_pool(pwq->pool); 3842cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 38434c16bd32STejun Heo } 38444c16bd32STejun Heo } 38454c16bd32STejun Heo 38464c16bd32STejun Heo /** 38474c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 38484c16bd32STejun Heo * @attrs: the wq_attrs of interest 38494c16bd32STejun Heo * @node: the target NUMA node 38504c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38514c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38524c16bd32STejun Heo * 38534c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38544c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3855d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38564c16bd32STejun Heo * 38574c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38584c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38594c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38604c16bd32STejun Heo * @attrs->cpumask. 38614c16bd32STejun Heo * 38624c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38634c16bd32STejun Heo * stable. 3864d185af30SYacine Belkadi * 3865d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3866d185af30SYacine Belkadi * %false if equal. 38674c16bd32STejun Heo */ 38684c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38694c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38704c16bd32STejun Heo { 3871d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38724c16bd32STejun Heo goto use_dfl; 38734c16bd32STejun Heo 38744c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38754c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38764c16bd32STejun Heo if (cpu_going_down >= 0) 38774c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38784c16bd32STejun Heo 38794c16bd32STejun Heo if (cpumask_empty(cpumask)) 38804c16bd32STejun Heo goto use_dfl; 38814c16bd32STejun Heo 38824c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38834c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38844c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 38854c16bd32STejun Heo 38864c16bd32STejun Heo use_dfl: 38874c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38884c16bd32STejun Heo return false; 38894c16bd32STejun Heo } 38904c16bd32STejun Heo 38911befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38921befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38931befcf30STejun Heo int node, 38941befcf30STejun Heo struct pool_workqueue *pwq) 38951befcf30STejun Heo { 38961befcf30STejun Heo struct pool_workqueue *old_pwq; 38971befcf30STejun Heo 38981befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38991befcf30STejun Heo 39001befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 39011befcf30STejun Heo link_pwq(pwq); 39021befcf30STejun Heo 39031befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 39041befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 39051befcf30STejun Heo return old_pwq; 39061befcf30STejun Heo } 39071befcf30STejun Heo 39089e8cd2f5STejun Heo /** 39099e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 39109e8cd2f5STejun Heo * @wq: the target workqueue 39119e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 39129e8cd2f5STejun Heo * 39134c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 39144c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 39154c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 39164c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 39174c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 39184c16bd32STejun Heo * back-to-back will stay on its current pwq. 39199e8cd2f5STejun Heo * 3920d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3921d185af30SYacine Belkadi * 3922d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 39239e8cd2f5STejun Heo */ 39249e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 39259e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 39269e8cd2f5STejun Heo { 39274c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 39284c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3929f147f29eSTejun Heo int node, ret; 39309e8cd2f5STejun Heo 39318719dceaSTejun Heo /* only unbound workqueues can change attributes */ 39329e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 39339e8cd2f5STejun Heo return -EINVAL; 39349e8cd2f5STejun Heo 39358719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 39368719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 39378719dceaSTejun Heo return -EINVAL; 39388719dceaSTejun Heo 39394c16bd32STejun Heo pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); 394013e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39414c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39424c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 394313e2e556STejun Heo goto enomem; 394413e2e556STejun Heo 39454c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 394613e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 394713e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 394813e2e556STejun Heo 39494c16bd32STejun Heo /* 39504c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 39514c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 39524c16bd32STejun Heo * pools. 39534c16bd32STejun Heo */ 39544c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 39559e8cd2f5STejun Heo 39564c16bd32STejun Heo /* 39574c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 39584c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 39594c16bd32STejun Heo * pwqs accordingly. 39604c16bd32STejun Heo */ 39614c16bd32STejun Heo get_online_cpus(); 39624c16bd32STejun Heo 39634c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 39644c16bd32STejun Heo 39654c16bd32STejun Heo /* 39664c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 39674c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 39684c16bd32STejun Heo * it even if we don't use it immediately. 39694c16bd32STejun Heo */ 39704c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39714c16bd32STejun Heo if (!dfl_pwq) 39724c16bd32STejun Heo goto enomem_pwq; 39734c16bd32STejun Heo 39744c16bd32STejun Heo for_each_node(node) { 39754c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 39764c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39774c16bd32STejun Heo if (!pwq_tbl[node]) 39784c16bd32STejun Heo goto enomem_pwq; 39794c16bd32STejun Heo } else { 39804c16bd32STejun Heo dfl_pwq->refcnt++; 39814c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 39824c16bd32STejun Heo } 39834c16bd32STejun Heo } 39844c16bd32STejun Heo 39854c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39864c16bd32STejun Heo 39874c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3988f147f29eSTejun Heo mutex_lock(&wq->mutex); 3989a892caccSTejun Heo 3990f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 39914c16bd32STejun Heo 39924c16bd32STejun Heo /* save the previous pwq and install the new one */ 3993f147f29eSTejun Heo for_each_node(node) 39944c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 39954c16bd32STejun Heo 39964c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 39974c16bd32STejun Heo link_pwq(dfl_pwq); 39984c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3999f147f29eSTejun Heo 4000f147f29eSTejun Heo mutex_unlock(&wq->mutex); 4001f147f29eSTejun Heo 40024c16bd32STejun Heo /* put the old pwqs */ 40034c16bd32STejun Heo for_each_node(node) 40044c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 40054c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 40064c16bd32STejun Heo 40074c16bd32STejun Heo put_online_cpus(); 40084862125bSTejun Heo ret = 0; 40094862125bSTejun Heo /* fall through */ 40104862125bSTejun Heo out_free: 40114c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 40124862125bSTejun Heo free_workqueue_attrs(new_attrs); 40134c16bd32STejun Heo kfree(pwq_tbl); 40144862125bSTejun Heo return ret; 401513e2e556STejun Heo 40164c16bd32STejun Heo enomem_pwq: 40174c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 40184c16bd32STejun Heo for_each_node(node) 40194c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 40204c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 40214c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 40224c16bd32STejun Heo put_online_cpus(); 402313e2e556STejun Heo enomem: 40244862125bSTejun Heo ret = -ENOMEM; 40254862125bSTejun Heo goto out_free; 40269e8cd2f5STejun Heo } 40279e8cd2f5STejun Heo 40284c16bd32STejun Heo /** 40294c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 40304c16bd32STejun Heo * @wq: the target workqueue 40314c16bd32STejun Heo * @cpu: the CPU coming up or going down 40324c16bd32STejun Heo * @online: whether @cpu is coming up or going down 40334c16bd32STejun Heo * 40344c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 40354c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 40364c16bd32STejun Heo * @wq accordingly. 40374c16bd32STejun Heo * 40384c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 40394c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 40404c16bd32STejun Heo * correct. 40414c16bd32STejun Heo * 40424c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 40434c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 40444c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 40454c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 40464c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 40474c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 40484c16bd32STejun Heo * CPU_DOWN_PREPARE. 40494c16bd32STejun Heo */ 40504c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 40514c16bd32STejun Heo bool online) 40524c16bd32STejun Heo { 40534c16bd32STejun Heo int node = cpu_to_node(cpu); 40544c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 40554c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 40564c16bd32STejun Heo struct workqueue_attrs *target_attrs; 40574c16bd32STejun Heo cpumask_t *cpumask; 40584c16bd32STejun Heo 40594c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 40604c16bd32STejun Heo 40614c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 40624c16bd32STejun Heo return; 40634c16bd32STejun Heo 40644c16bd32STejun Heo /* 40654c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 40664c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 40674c16bd32STejun Heo * CPU hotplug exclusion. 40684c16bd32STejun Heo */ 40694c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 40704c16bd32STejun Heo cpumask = target_attrs->cpumask; 40714c16bd32STejun Heo 40724c16bd32STejun Heo mutex_lock(&wq->mutex); 4073d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 4074d55262c4STejun Heo goto out_unlock; 40754c16bd32STejun Heo 40764c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 40774c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 40784c16bd32STejun Heo 40794c16bd32STejun Heo /* 40804c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 40814c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 40824c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 40834c16bd32STejun Heo * wq's, the default pwq should be used. If @pwq is already the 40844c16bd32STejun Heo * default one, nothing to do; otherwise, install the default one. 40854c16bd32STejun Heo */ 40864c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 40874c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 40884c16bd32STejun Heo goto out_unlock; 40894c16bd32STejun Heo } else { 40904c16bd32STejun Heo if (pwq == wq->dfl_pwq) 40914c16bd32STejun Heo goto out_unlock; 40924c16bd32STejun Heo else 40934c16bd32STejun Heo goto use_dfl_pwq; 40944c16bd32STejun Heo } 40954c16bd32STejun Heo 40964c16bd32STejun Heo mutex_unlock(&wq->mutex); 40974c16bd32STejun Heo 40984c16bd32STejun Heo /* create a new pwq */ 40994c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41004c16bd32STejun Heo if (!pwq) { 41014c16bd32STejun Heo pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41024c16bd32STejun Heo wq->name); 410377f300b1SDaeseok Youn mutex_lock(&wq->mutex); 410477f300b1SDaeseok Youn goto use_dfl_pwq; 41054c16bd32STejun Heo } 41064c16bd32STejun Heo 41074c16bd32STejun Heo /* 41084c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 41094c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 41104c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 41114c16bd32STejun Heo * inbetween. 41124c16bd32STejun Heo */ 41134c16bd32STejun Heo mutex_lock(&wq->mutex); 41144c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41154c16bd32STejun Heo goto out_unlock; 41164c16bd32STejun Heo 41174c16bd32STejun Heo use_dfl_pwq: 41184c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 41194c16bd32STejun Heo get_pwq(wq->dfl_pwq); 41204c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 41214c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 41224c16bd32STejun Heo out_unlock: 41234c16bd32STejun Heo mutex_unlock(&wq->mutex); 41244c16bd32STejun Heo put_pwq_unlocked(old_pwq); 41254c16bd32STejun Heo } 41264c16bd32STejun Heo 412730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 41281da177e4SLinus Torvalds { 412949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 41308a2b7538STejun Heo int cpu, ret; 4131e1d8aa9fSFrederic Weisbecker 413230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4133420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4134420c0ddbSTejun Heo if (!wq->cpu_pwqs) 413530cdf249STejun Heo return -ENOMEM; 413630cdf249STejun Heo 413730cdf249STejun Heo for_each_possible_cpu(cpu) { 41387fb98ea7STejun Heo struct pool_workqueue *pwq = 41397fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 41407a62c2c8STejun Heo struct worker_pool *cpu_pools = 4141f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 414230cdf249STejun Heo 4143f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4144f147f29eSTejun Heo 4145f147f29eSTejun Heo mutex_lock(&wq->mutex); 41461befcf30STejun Heo link_pwq(pwq); 4147f147f29eSTejun Heo mutex_unlock(&wq->mutex); 414830cdf249STejun Heo } 414930cdf249STejun Heo return 0; 41508a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 41518a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 41528a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 41538a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 41548a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 41558a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 41568a2b7538STejun Heo return ret; 41579e8cd2f5STejun Heo } else { 41589e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 41599e8cd2f5STejun Heo } 41600f900049STejun Heo } 41610f900049STejun Heo 4162f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4163f3421797STejun Heo const char *name) 4164b71ab8c2STejun Heo { 4165f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4166f3421797STejun Heo 4167f3421797STejun Heo if (max_active < 1 || max_active > lim) 4168044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4169f3421797STejun Heo max_active, name, 1, lim); 4170b71ab8c2STejun Heo 4171f3421797STejun Heo return clamp_val(max_active, 1, lim); 4172b71ab8c2STejun Heo } 4173b71ab8c2STejun Heo 4174b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 417597e37d7bSTejun Heo unsigned int flags, 41761e19ffc6STejun Heo int max_active, 4177eb13ba87SJohannes Berg struct lock_class_key *key, 4178b196be89STejun Heo const char *lock_name, ...) 41793af24433SOleg Nesterov { 4180df2d5ae4STejun Heo size_t tbl_size = 0; 4181ecf6881fSTejun Heo va_list args; 41823af24433SOleg Nesterov struct workqueue_struct *wq; 418349e3cf44STejun Heo struct pool_workqueue *pwq; 4184b196be89STejun Heo 4185cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4186cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4187cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4188cee22a15SViresh Kumar 4189ecf6881fSTejun Heo /* allocate wq and format name */ 4190df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4191df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 4192df2d5ae4STejun Heo 4193df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4194b196be89STejun Heo if (!wq) 4195d2c1d404STejun Heo return NULL; 4196b196be89STejun Heo 41976029a918STejun Heo if (flags & WQ_UNBOUND) { 41986029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 41996029a918STejun Heo if (!wq->unbound_attrs) 42006029a918STejun Heo goto err_free_wq; 42016029a918STejun Heo } 42026029a918STejun Heo 4203ecf6881fSTejun Heo va_start(args, lock_name); 4204ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4205b196be89STejun Heo va_end(args); 42063af24433SOleg Nesterov 4207d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4208b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 42093af24433SOleg Nesterov 4210b196be89STejun Heo /* init wq */ 421197e37d7bSTejun Heo wq->flags = flags; 4212a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 42133c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4214112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 421530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 421673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 421773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4218493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 42193af24433SOleg Nesterov 4220eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4221cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 42223af24433SOleg Nesterov 422330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4224d2c1d404STejun Heo goto err_free_wq; 42251537663fSTejun Heo 4226493008a8STejun Heo /* 4227493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4228493008a8STejun Heo * have a rescuer to guarantee forward progress. 4229493008a8STejun Heo */ 4230493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4231e22bee78STejun Heo struct worker *rescuer; 4232e22bee78STejun Heo 4233d2c1d404STejun Heo rescuer = alloc_worker(); 4234e22bee78STejun Heo if (!rescuer) 4235d2c1d404STejun Heo goto err_destroy; 4236e22bee78STejun Heo 4237111c225aSTejun Heo rescuer->rescue_wq = wq; 4238111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4239b196be89STejun Heo wq->name); 4240d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4241d2c1d404STejun Heo kfree(rescuer); 4242d2c1d404STejun Heo goto err_destroy; 4243d2c1d404STejun Heo } 4244e22bee78STejun Heo 4245d2c1d404STejun Heo wq->rescuer = rescuer; 424614a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4247e22bee78STejun Heo wake_up_process(rescuer->task); 42483af24433SOleg Nesterov } 42491537663fSTejun Heo 4250226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4251226223abSTejun Heo goto err_destroy; 4252226223abSTejun Heo 42533af24433SOleg Nesterov /* 425468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 425568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 425668e13a67SLai Jiangshan * list. 42573af24433SOleg Nesterov */ 425868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4259a0a1a5fdSTejun Heo 4260a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 426149e3cf44STejun Heo for_each_pwq(pwq, wq) 4262699ce097STejun Heo pwq_adjust_max_active(pwq); 4263a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4264a0a1a5fdSTejun Heo 42653af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4266a0a1a5fdSTejun Heo 426768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42683af24433SOleg Nesterov 42693af24433SOleg Nesterov return wq; 4270d2c1d404STejun Heo 4271d2c1d404STejun Heo err_free_wq: 42726029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 42734690c4abSTejun Heo kfree(wq); 4274d2c1d404STejun Heo return NULL; 4275d2c1d404STejun Heo err_destroy: 4276d2c1d404STejun Heo destroy_workqueue(wq); 42774690c4abSTejun Heo return NULL; 42781da177e4SLinus Torvalds } 4279d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 42801da177e4SLinus Torvalds 42813af24433SOleg Nesterov /** 42823af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 42833af24433SOleg Nesterov * @wq: target workqueue 42843af24433SOleg Nesterov * 42853af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 42863af24433SOleg Nesterov */ 42873af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 42883af24433SOleg Nesterov { 428949e3cf44STejun Heo struct pool_workqueue *pwq; 42904c16bd32STejun Heo int node; 42913af24433SOleg Nesterov 42929c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 42939c5a2ba7STejun Heo drain_workqueue(wq); 4294c8efcc25STejun Heo 42956183c009STejun Heo /* sanity checks */ 4296b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 429749e3cf44STejun Heo for_each_pwq(pwq, wq) { 42986183c009STejun Heo int i; 42996183c009STejun Heo 430076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 430176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4302b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 43036183c009STejun Heo return; 430476af4d93STejun Heo } 430576af4d93STejun Heo } 430676af4d93STejun Heo 43075c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 43088864b4e5STejun Heo WARN_ON(pwq->nr_active) || 430976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4310b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 43116183c009STejun Heo return; 43126183c009STejun Heo } 431376af4d93STejun Heo } 4314b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 43156183c009STejun Heo 4316a0a1a5fdSTejun Heo /* 4317a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4318a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4319a0a1a5fdSTejun Heo */ 432068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4321d2c1d404STejun Heo list_del_init(&wq->list); 432268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43233af24433SOleg Nesterov 4324226223abSTejun Heo workqueue_sysfs_unregister(wq); 4325226223abSTejun Heo 4326493008a8STejun Heo if (wq->rescuer) { 4327e22bee78STejun Heo kthread_stop(wq->rescuer->task); 43288d9df9f0SXiaotian Feng kfree(wq->rescuer); 4329493008a8STejun Heo wq->rescuer = NULL; 4330e22bee78STejun Heo } 4331e22bee78STejun Heo 43328864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 433329c91e99STejun Heo /* 43348864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 43358864b4e5STejun Heo * free the pwqs and wq. 433629c91e99STejun Heo */ 43378864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 43388864b4e5STejun Heo kfree(wq); 43398864b4e5STejun Heo } else { 43408864b4e5STejun Heo /* 43418864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 43424c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 43434c16bd32STejun Heo * @wq will be freed when the last pwq is released. 43448864b4e5STejun Heo */ 43454c16bd32STejun Heo for_each_node(node) { 43464c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 43474c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 43484c16bd32STejun Heo put_pwq_unlocked(pwq); 43494c16bd32STejun Heo } 43504c16bd32STejun Heo 43514c16bd32STejun Heo /* 43524c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 43534c16bd32STejun Heo * put. Don't access it afterwards. 43544c16bd32STejun Heo */ 43554c16bd32STejun Heo pwq = wq->dfl_pwq; 43564c16bd32STejun Heo wq->dfl_pwq = NULL; 4357dce90d47STejun Heo put_pwq_unlocked(pwq); 435829c91e99STejun Heo } 43593af24433SOleg Nesterov } 43603af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 43613af24433SOleg Nesterov 4362dcd989cbSTejun Heo /** 4363dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4364dcd989cbSTejun Heo * @wq: target workqueue 4365dcd989cbSTejun Heo * @max_active: new max_active value. 4366dcd989cbSTejun Heo * 4367dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4368dcd989cbSTejun Heo * 4369dcd989cbSTejun Heo * CONTEXT: 4370dcd989cbSTejun Heo * Don't call from IRQ context. 4371dcd989cbSTejun Heo */ 4372dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4373dcd989cbSTejun Heo { 437449e3cf44STejun Heo struct pool_workqueue *pwq; 4375dcd989cbSTejun Heo 43768719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 43778719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 43788719dceaSTejun Heo return; 43798719dceaSTejun Heo 4380f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4381dcd989cbSTejun Heo 4382a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4383dcd989cbSTejun Heo 4384dcd989cbSTejun Heo wq->saved_max_active = max_active; 4385dcd989cbSTejun Heo 4386699ce097STejun Heo for_each_pwq(pwq, wq) 4387699ce097STejun Heo pwq_adjust_max_active(pwq); 4388dcd989cbSTejun Heo 4389a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4390dcd989cbSTejun Heo } 4391dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4392dcd989cbSTejun Heo 4393dcd989cbSTejun Heo /** 4394e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4395e6267616STejun Heo * 4396e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4397e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4398d185af30SYacine Belkadi * 4399d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4400e6267616STejun Heo */ 4401e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4402e6267616STejun Heo { 4403e6267616STejun Heo struct worker *worker = current_wq_worker(); 4404e6267616STejun Heo 44056a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4406e6267616STejun Heo } 4407e6267616STejun Heo 4408e6267616STejun Heo /** 4409dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4410dcd989cbSTejun Heo * @cpu: CPU in question 4411dcd989cbSTejun Heo * @wq: target workqueue 4412dcd989cbSTejun Heo * 4413dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4414dcd989cbSTejun Heo * no synchronization around this function and the test result is 4415dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4416dcd989cbSTejun Heo * 4417d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4418d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4419d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4420d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4421d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4422d3251859STejun Heo * 4423d185af30SYacine Belkadi * Return: 4424dcd989cbSTejun Heo * %true if congested, %false otherwise. 4425dcd989cbSTejun Heo */ 4426d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4427dcd989cbSTejun Heo { 44287fb98ea7STejun Heo struct pool_workqueue *pwq; 442976af4d93STejun Heo bool ret; 443076af4d93STejun Heo 443188109453SLai Jiangshan rcu_read_lock_sched(); 44327fb98ea7STejun Heo 4433d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4434d3251859STejun Heo cpu = smp_processor_id(); 4435d3251859STejun Heo 44367fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 44377fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 44387fb98ea7STejun Heo else 4439df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4440dcd989cbSTejun Heo 444176af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 444288109453SLai Jiangshan rcu_read_unlock_sched(); 444376af4d93STejun Heo 444476af4d93STejun Heo return ret; 4445dcd989cbSTejun Heo } 4446dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4447dcd989cbSTejun Heo 4448dcd989cbSTejun Heo /** 4449dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4450dcd989cbSTejun Heo * @work: the work to be tested 4451dcd989cbSTejun Heo * 4452dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4453dcd989cbSTejun Heo * synchronization around this function and the test result is 4454dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4455dcd989cbSTejun Heo * 4456d185af30SYacine Belkadi * Return: 4457dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4458dcd989cbSTejun Heo */ 4459dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4460dcd989cbSTejun Heo { 4461fa1b54e6STejun Heo struct worker_pool *pool; 4462dcd989cbSTejun Heo unsigned long flags; 4463dcd989cbSTejun Heo unsigned int ret = 0; 4464dcd989cbSTejun Heo 4465dcd989cbSTejun Heo if (work_pending(work)) 4466dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4467038366c5SLai Jiangshan 4468fa1b54e6STejun Heo local_irq_save(flags); 4469fa1b54e6STejun Heo pool = get_work_pool(work); 4470038366c5SLai Jiangshan if (pool) { 4471fa1b54e6STejun Heo spin_lock(&pool->lock); 4472c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4473dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4474fa1b54e6STejun Heo spin_unlock(&pool->lock); 4475038366c5SLai Jiangshan } 4476fa1b54e6STejun Heo local_irq_restore(flags); 4477dcd989cbSTejun Heo 4478dcd989cbSTejun Heo return ret; 4479dcd989cbSTejun Heo } 4480dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4481dcd989cbSTejun Heo 44823d1cb205STejun Heo /** 44833d1cb205STejun Heo * set_worker_desc - set description for the current work item 44843d1cb205STejun Heo * @fmt: printf-style format string 44853d1cb205STejun Heo * @...: arguments for the format string 44863d1cb205STejun Heo * 44873d1cb205STejun Heo * This function can be called by a running work function to describe what 44883d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 44893d1cb205STejun Heo * information will be printed out together to help debugging. The 44903d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 44913d1cb205STejun Heo */ 44923d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 44933d1cb205STejun Heo { 44943d1cb205STejun Heo struct worker *worker = current_wq_worker(); 44953d1cb205STejun Heo va_list args; 44963d1cb205STejun Heo 44973d1cb205STejun Heo if (worker) { 44983d1cb205STejun Heo va_start(args, fmt); 44993d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 45003d1cb205STejun Heo va_end(args); 45013d1cb205STejun Heo worker->desc_valid = true; 45023d1cb205STejun Heo } 45033d1cb205STejun Heo } 45043d1cb205STejun Heo 45053d1cb205STejun Heo /** 45063d1cb205STejun Heo * print_worker_info - print out worker information and description 45073d1cb205STejun Heo * @log_lvl: the log level to use when printing 45083d1cb205STejun Heo * @task: target task 45093d1cb205STejun Heo * 45103d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 45113d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 45123d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 45133d1cb205STejun Heo * 45143d1cb205STejun Heo * This function can be safely called on any task as long as the 45153d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 45163d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 45173d1cb205STejun Heo */ 45183d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 45193d1cb205STejun Heo { 45203d1cb205STejun Heo work_func_t *fn = NULL; 45213d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 45223d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 45233d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 45243d1cb205STejun Heo struct workqueue_struct *wq = NULL; 45253d1cb205STejun Heo bool desc_valid = false; 45263d1cb205STejun Heo struct worker *worker; 45273d1cb205STejun Heo 45283d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 45293d1cb205STejun Heo return; 45303d1cb205STejun Heo 45313d1cb205STejun Heo /* 45323d1cb205STejun Heo * This function is called without any synchronization and @task 45333d1cb205STejun Heo * could be in any state. Be careful with dereferences. 45343d1cb205STejun Heo */ 45353d1cb205STejun Heo worker = probe_kthread_data(task); 45363d1cb205STejun Heo 45373d1cb205STejun Heo /* 45383d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 45393d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 45403d1cb205STejun Heo */ 45413d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 45423d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 45433d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 45443d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 45453d1cb205STejun Heo 45463d1cb205STejun Heo /* copy worker description */ 45473d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 45483d1cb205STejun Heo if (desc_valid) 45493d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 45503d1cb205STejun Heo 45513d1cb205STejun Heo if (fn || name[0] || desc[0]) { 45523d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 45533d1cb205STejun Heo if (desc[0]) 45543d1cb205STejun Heo pr_cont(" (%s)", desc); 45553d1cb205STejun Heo pr_cont("\n"); 45563d1cb205STejun Heo } 45573d1cb205STejun Heo } 45583d1cb205STejun Heo 4559db7bccf4STejun Heo /* 4560db7bccf4STejun Heo * CPU hotplug. 4561db7bccf4STejun Heo * 4562e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4563112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4564706026c2STejun Heo * pool which make migrating pending and scheduled works very 4565e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 456694cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4567e22bee78STejun Heo * blocked draining impractical. 4568e22bee78STejun Heo * 456924647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4570628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4571628c78e7STejun Heo * cpu comes back online. 4572db7bccf4STejun Heo */ 4573db7bccf4STejun Heo 4574706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4575db7bccf4STejun Heo { 457638db41d9STejun Heo int cpu = smp_processor_id(); 45774ce62e9eSTejun Heo struct worker_pool *pool; 4578db7bccf4STejun Heo struct worker *worker; 4579a9ab775bSTejun Heo int wi; 4580db7bccf4STejun Heo 4581f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45826183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4583db7bccf4STejun Heo 4584bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 458594cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4586e22bee78STejun Heo 4587f2d5a0eeSTejun Heo /* 4588bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 458994cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 459094cf58bbSTejun Heo * except for the ones which are still executing works from 459194cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 459294cf58bbSTejun Heo * this, they may become diasporas. 4593f2d5a0eeSTejun Heo */ 4594a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4595403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4596db7bccf4STejun Heo 459724647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4598f2d5a0eeSTejun Heo 459994cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4600bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 4601e22bee78STejun Heo 4602e22bee78STejun Heo /* 4603eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4604eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4605eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4606eb283428SLai Jiangshan * from other cpus. 4607628c78e7STejun Heo */ 4608628c78e7STejun Heo schedule(); 4609628c78e7STejun Heo 4610628c78e7STejun Heo /* 4611eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4612eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4613eb283428SLai Jiangshan * and keep_working() are always true as long as the 4614eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4615eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4616eb283428SLai Jiangshan * are served by workers tied to the pool. 4617e22bee78STejun Heo */ 4618e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4619eb283428SLai Jiangshan 4620eb283428SLai Jiangshan /* 4621eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4622eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4623eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4624eb283428SLai Jiangshan */ 4625eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4626eb283428SLai Jiangshan wake_up_worker(pool); 4627eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4628eb283428SLai Jiangshan } 4629db7bccf4STejun Heo } 4630db7bccf4STejun Heo 4631bd7c089eSTejun Heo /** 4632bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4633bd7c089eSTejun Heo * @pool: pool of interest 4634bd7c089eSTejun Heo * 4635a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4636bd7c089eSTejun Heo */ 4637bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4638bd7c089eSTejun Heo { 4639a9ab775bSTejun Heo struct worker *worker; 4640a9ab775bSTejun Heo int wi; 4641bd7c089eSTejun Heo 4642bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4643bd7c089eSTejun Heo 4644bd7c089eSTejun Heo /* 4645a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4646a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4647a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4648a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4649a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4650bd7c089eSTejun Heo */ 4651a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4652a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4653a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4654a9ab775bSTejun Heo 4655a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4656a9ab775bSTejun Heo 4657a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4658a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4659a9ab775bSTejun Heo 4660a9ab775bSTejun Heo /* 4661a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4662a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4663a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4664a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4665a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4666a9ab775bSTejun Heo * be bound before @pool->lock is released. 4667a9ab775bSTejun Heo */ 4668a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4669bd7c089eSTejun Heo wake_up_process(worker->task); 4670bd7c089eSTejun Heo 4671bd7c089eSTejun Heo /* 4672a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4673a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4674a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4675a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4676a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4677a9ab775bSTejun Heo * concurrency management. Note that when or whether 4678a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4679a9ab775bSTejun Heo * 4680a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4681a9ab775bSTejun Heo * tested without holding any lock in 4682a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4683a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4684a9ab775bSTejun Heo * management operations. 4685bd7c089eSTejun Heo */ 4686a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4687a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4688a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4689a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4690bd7c089eSTejun Heo } 4691a9ab775bSTejun Heo 4692a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4693bd7c089eSTejun Heo } 4694bd7c089eSTejun Heo 46957dbc725eSTejun Heo /** 46967dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 46977dbc725eSTejun Heo * @pool: unbound pool of interest 46987dbc725eSTejun Heo * @cpu: the CPU which is coming up 46997dbc725eSTejun Heo * 47007dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 47017dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 47027dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 47037dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 47047dbc725eSTejun Heo */ 47057dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 47067dbc725eSTejun Heo { 47077dbc725eSTejun Heo static cpumask_t cpumask; 47087dbc725eSTejun Heo struct worker *worker; 47097dbc725eSTejun Heo int wi; 47107dbc725eSTejun Heo 47117dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 47127dbc725eSTejun Heo 47137dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 47147dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 47157dbc725eSTejun Heo return; 47167dbc725eSTejun Heo 47177dbc725eSTejun Heo /* is @cpu the only online CPU? */ 47187dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 47197dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 47207dbc725eSTejun Heo return; 47217dbc725eSTejun Heo 47227dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 47237dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 47247dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 47257dbc725eSTejun Heo pool->attrs->cpumask) < 0); 47267dbc725eSTejun Heo } 47277dbc725eSTejun Heo 47288db25e78STejun Heo /* 47298db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 47308db25e78STejun Heo * This will be registered high priority CPU notifier. 47318db25e78STejun Heo */ 47320db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 47331da177e4SLinus Torvalds unsigned long action, 47341da177e4SLinus Torvalds void *hcpu) 47351da177e4SLinus Torvalds { 4736d84ff051STejun Heo int cpu = (unsigned long)hcpu; 47374ce62e9eSTejun Heo struct worker_pool *pool; 47384c16bd32STejun Heo struct workqueue_struct *wq; 47397dbc725eSTejun Heo int pi; 47401da177e4SLinus Torvalds 47418db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 47423af24433SOleg Nesterov case CPU_UP_PREPARE: 4743f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 47443ce63377STejun Heo if (pool->nr_workers) 47453ce63377STejun Heo continue; 4746ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 47473ce63377STejun Heo return NOTIFY_BAD; 47483af24433SOleg Nesterov } 47491da177e4SLinus Torvalds break; 47501da177e4SLinus Torvalds 475165758202STejun Heo case CPU_DOWN_FAILED: 475265758202STejun Heo case CPU_ONLINE: 475368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 47547dbc725eSTejun Heo 47557dbc725eSTejun Heo for_each_pool(pool, pi) { 4756bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 475794cf58bbSTejun Heo 47587dbc725eSTejun Heo if (pool->cpu == cpu) { 4759a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 476024647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4761a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4762a9ab775bSTejun Heo 476394cf58bbSTejun Heo rebind_workers(pool); 47647dbc725eSTejun Heo } else if (pool->cpu < 0) { 47657dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 47667dbc725eSTejun Heo } 476794cf58bbSTejun Heo 4768bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 476994cf58bbSTejun Heo } 47707dbc725eSTejun Heo 47714c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47724c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47734c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 47744c16bd32STejun Heo 477568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47768db25e78STejun Heo break; 477765758202STejun Heo } 477865758202STejun Heo return NOTIFY_OK; 477965758202STejun Heo } 478065758202STejun Heo 478165758202STejun Heo /* 478265758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 478365758202STejun Heo * This will be registered as low priority CPU notifier. 478465758202STejun Heo */ 47850db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 478665758202STejun Heo unsigned long action, 478765758202STejun Heo void *hcpu) 478865758202STejun Heo { 4789d84ff051STejun Heo int cpu = (unsigned long)hcpu; 47908db25e78STejun Heo struct work_struct unbind_work; 47914c16bd32STejun Heo struct workqueue_struct *wq; 47928db25e78STejun Heo 479365758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 479465758202STejun Heo case CPU_DOWN_PREPARE: 47954c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4796706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 47977635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 47984c16bd32STejun Heo 47994c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 48004c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 48014c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 48024c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 48034c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 48044c16bd32STejun Heo 48054c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 48068db25e78STejun Heo flush_work(&unbind_work); 4807440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 48088db25e78STejun Heo break; 480965758202STejun Heo } 481065758202STejun Heo return NOTIFY_OK; 481165758202STejun Heo } 481265758202STejun Heo 48132d3854a3SRusty Russell #ifdef CONFIG_SMP 48148ccad40dSRusty Russell 48152d3854a3SRusty Russell struct work_for_cpu { 4816ed48ece2STejun Heo struct work_struct work; 48172d3854a3SRusty Russell long (*fn)(void *); 48182d3854a3SRusty Russell void *arg; 48192d3854a3SRusty Russell long ret; 48202d3854a3SRusty Russell }; 48212d3854a3SRusty Russell 4822ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 48232d3854a3SRusty Russell { 4824ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4825ed48ece2STejun Heo 48262d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 48272d3854a3SRusty Russell } 48282d3854a3SRusty Russell 48292d3854a3SRusty Russell /** 48302d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 48312d3854a3SRusty Russell * @cpu: the cpu to run on 48322d3854a3SRusty Russell * @fn: the function to run 48332d3854a3SRusty Russell * @arg: the function arg 48342d3854a3SRusty Russell * 483531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 48366b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4837d185af30SYacine Belkadi * 4838d185af30SYacine Belkadi * Return: The value @fn returns. 48392d3854a3SRusty Russell */ 4840d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 48412d3854a3SRusty Russell { 4842ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 48432d3854a3SRusty Russell 4844ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4845ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 484612997d1aSBjorn Helgaas flush_work(&wfc.work); 4847440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 48482d3854a3SRusty Russell return wfc.ret; 48492d3854a3SRusty Russell } 48502d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 48512d3854a3SRusty Russell #endif /* CONFIG_SMP */ 48522d3854a3SRusty Russell 4853a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4854e7577c50SRusty Russell 4855a0a1a5fdSTejun Heo /** 4856a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4857a0a1a5fdSTejun Heo * 485858a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4859c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4860706026c2STejun Heo * pool->worklist. 4861a0a1a5fdSTejun Heo * 4862a0a1a5fdSTejun Heo * CONTEXT: 4863a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4864a0a1a5fdSTejun Heo */ 4865a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4866a0a1a5fdSTejun Heo { 486717116969STejun Heo struct worker_pool *pool; 486824b8a847STejun Heo struct workqueue_struct *wq; 486924b8a847STejun Heo struct pool_workqueue *pwq; 4870611c92a0STejun Heo int pi; 4871a0a1a5fdSTejun Heo 487268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4873a0a1a5fdSTejun Heo 48746183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4875a0a1a5fdSTejun Heo workqueue_freezing = true; 4876a0a1a5fdSTejun Heo 487724b8a847STejun Heo /* set FREEZING */ 4878611c92a0STejun Heo for_each_pool(pool, pi) { 48795bcab335STejun Heo spin_lock_irq(&pool->lock); 488035b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 488135b6bb63STejun Heo pool->flags |= POOL_FREEZING; 48825bcab335STejun Heo spin_unlock_irq(&pool->lock); 48831da177e4SLinus Torvalds } 48848b03ae3cSTejun Heo 488524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4886a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4887699ce097STejun Heo for_each_pwq(pwq, wq) 4888699ce097STejun Heo pwq_adjust_max_active(pwq); 4889a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4890a1056305STejun Heo } 48915bcab335STejun Heo 489268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4893a0a1a5fdSTejun Heo } 4894a0a1a5fdSTejun Heo 4895a0a1a5fdSTejun Heo /** 489658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4897a0a1a5fdSTejun Heo * 4898a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4899a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4900a0a1a5fdSTejun Heo * 4901a0a1a5fdSTejun Heo * CONTEXT: 490268e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4903a0a1a5fdSTejun Heo * 4904d185af30SYacine Belkadi * Return: 490558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 490658a69cb4STejun Heo * is complete. 4907a0a1a5fdSTejun Heo */ 4908a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4909a0a1a5fdSTejun Heo { 4910a0a1a5fdSTejun Heo bool busy = false; 491124b8a847STejun Heo struct workqueue_struct *wq; 491224b8a847STejun Heo struct pool_workqueue *pwq; 4913a0a1a5fdSTejun Heo 491468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4915a0a1a5fdSTejun Heo 49166183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4917a0a1a5fdSTejun Heo 491824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 491924b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 492024b8a847STejun Heo continue; 4921a0a1a5fdSTejun Heo /* 4922a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4923a0a1a5fdSTejun Heo * to peek without lock. 4924a0a1a5fdSTejun Heo */ 492588109453SLai Jiangshan rcu_read_lock_sched(); 492624b8a847STejun Heo for_each_pwq(pwq, wq) { 49276183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4928112202d9STejun Heo if (pwq->nr_active) { 4929a0a1a5fdSTejun Heo busy = true; 493088109453SLai Jiangshan rcu_read_unlock_sched(); 4931a0a1a5fdSTejun Heo goto out_unlock; 4932a0a1a5fdSTejun Heo } 4933a0a1a5fdSTejun Heo } 493488109453SLai Jiangshan rcu_read_unlock_sched(); 4935a0a1a5fdSTejun Heo } 4936a0a1a5fdSTejun Heo out_unlock: 493768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4938a0a1a5fdSTejun Heo return busy; 4939a0a1a5fdSTejun Heo } 4940a0a1a5fdSTejun Heo 4941a0a1a5fdSTejun Heo /** 4942a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4943a0a1a5fdSTejun Heo * 4944a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4945706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4946a0a1a5fdSTejun Heo * 4947a0a1a5fdSTejun Heo * CONTEXT: 4948a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4949a0a1a5fdSTejun Heo */ 4950a0a1a5fdSTejun Heo void thaw_workqueues(void) 4951a0a1a5fdSTejun Heo { 495224b8a847STejun Heo struct workqueue_struct *wq; 495324b8a847STejun Heo struct pool_workqueue *pwq; 495424b8a847STejun Heo struct worker_pool *pool; 4955611c92a0STejun Heo int pi; 4956a0a1a5fdSTejun Heo 495768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4958a0a1a5fdSTejun Heo 4959a0a1a5fdSTejun Heo if (!workqueue_freezing) 4960a0a1a5fdSTejun Heo goto out_unlock; 4961a0a1a5fdSTejun Heo 496224b8a847STejun Heo /* clear FREEZING */ 4963611c92a0STejun Heo for_each_pool(pool, pi) { 49645bcab335STejun Heo spin_lock_irq(&pool->lock); 496535b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 496635b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 49675bcab335STejun Heo spin_unlock_irq(&pool->lock); 4968d565ed63STejun Heo } 496924b8a847STejun Heo 497024b8a847STejun Heo /* restore max_active and repopulate worklist */ 497124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4972a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4973699ce097STejun Heo for_each_pwq(pwq, wq) 4974699ce097STejun Heo pwq_adjust_max_active(pwq); 4975a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 497624b8a847STejun Heo } 497724b8a847STejun Heo 4978a0a1a5fdSTejun Heo workqueue_freezing = false; 4979a0a1a5fdSTejun Heo out_unlock: 498068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4981a0a1a5fdSTejun Heo } 4982a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4983a0a1a5fdSTejun Heo 4984bce90380STejun Heo static void __init wq_numa_init(void) 4985bce90380STejun Heo { 4986bce90380STejun Heo cpumask_var_t *tbl; 4987bce90380STejun Heo int node, cpu; 4988bce90380STejun Heo 4989bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4990bce90380STejun Heo for_each_node(node) 4991bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4992bce90380STejun Heo 4993bce90380STejun Heo if (num_possible_nodes() <= 1) 4994bce90380STejun Heo return; 4995bce90380STejun Heo 4996d55262c4STejun Heo if (wq_disable_numa) { 4997d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 4998d55262c4STejun Heo return; 4999d55262c4STejun Heo } 5000d55262c4STejun Heo 50014c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 50024c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 50034c16bd32STejun Heo 5004bce90380STejun Heo /* 5005bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5006bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5007bce90380STejun Heo * fully initialized by now. 5008bce90380STejun Heo */ 5009bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 5010bce90380STejun Heo BUG_ON(!tbl); 5011bce90380STejun Heo 5012bce90380STejun Heo for_each_node(node) 50131be0c25dSTejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 50141be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5015bce90380STejun Heo 5016bce90380STejun Heo for_each_possible_cpu(cpu) { 5017bce90380STejun Heo node = cpu_to_node(cpu); 5018bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5019bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5020bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5021bce90380STejun Heo return; 5022bce90380STejun Heo } 5023bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5024bce90380STejun Heo } 5025bce90380STejun Heo 5026bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5027bce90380STejun Heo wq_numa_enabled = true; 5028bce90380STejun Heo } 5029bce90380STejun Heo 50306ee0578bSSuresh Siddha static int __init init_workqueues(void) 50311da177e4SLinus Torvalds { 50327a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 50337a4e344cSTejun Heo int i, cpu; 5034c34056a3STejun Heo 5035e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5036e904e6c2STejun Heo 5037e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5038e904e6c2STejun Heo 503965758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 5040a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 50418b03ae3cSTejun Heo 5042bce90380STejun Heo wq_numa_init(); 5043bce90380STejun Heo 5044706026c2STejun Heo /* initialize CPU pools */ 504529c91e99STejun Heo for_each_possible_cpu(cpu) { 50464ce62e9eSTejun Heo struct worker_pool *pool; 50478b03ae3cSTejun Heo 50487a4e344cSTejun Heo i = 0; 5049f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 50507a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5051ec22ca5eSTejun Heo pool->cpu = cpu; 50527a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 50537a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5054f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 50557a4e344cSTejun Heo 50569daf9e67STejun Heo /* alloc pool ID */ 505768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 50589daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 505968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 50604ce62e9eSTejun Heo } 50618b03ae3cSTejun Heo } 50628b03ae3cSTejun Heo 5063e22bee78STejun Heo /* create the initial worker */ 506429c91e99STejun Heo for_each_online_cpu(cpu) { 50654ce62e9eSTejun Heo struct worker_pool *pool; 5066e22bee78STejun Heo 5067f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 506824647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5069ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 5070e22bee78STejun Heo } 50714ce62e9eSTejun Heo } 5072e22bee78STejun Heo 50738a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 507429c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 507529c91e99STejun Heo struct workqueue_attrs *attrs; 507629c91e99STejun Heo 507729c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 507829c91e99STejun Heo attrs->nice = std_nice[i]; 507929c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 50808a2b7538STejun Heo 50818a2b7538STejun Heo /* 50828a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 50838a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 50848a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 50858a2b7538STejun Heo */ 50868a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 50878a2b7538STejun Heo attrs->nice = std_nice[i]; 50888a2b7538STejun Heo attrs->no_numa = true; 50898a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 509029c91e99STejun Heo } 509129c91e99STejun Heo 5092d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 50931aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5094d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5095f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5096f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 509724d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 509824d51addSTejun Heo WQ_FREEZABLE, 0); 50990668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 51000668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 51010668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 51020668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 51030668106cSViresh Kumar 0); 51041aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 51050668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 51060668106cSViresh Kumar !system_power_efficient_wq || 51070668106cSViresh Kumar !system_freezable_power_efficient_wq); 51086ee0578bSSuresh Siddha return 0; 51091da177e4SLinus Torvalds } 51106ee0578bSSuresh Siddha early_initcall(init_workqueues); 5111