11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 20b11895c4SLibin * normal work items and the other for high priority ones) and some extra 21b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 22b11895c4SLibin * number of these backing pools is dynamic. 23c54fce6eSTejun Heo * 24c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 251da177e4SLinus Torvalds */ 261da177e4SLinus Torvalds 279984de1aSPaul Gortmaker #include <linux/export.h> 281da177e4SLinus Torvalds #include <linux/kernel.h> 291da177e4SLinus Torvalds #include <linux/sched.h> 301da177e4SLinus Torvalds #include <linux/init.h> 311da177e4SLinus Torvalds #include <linux/signal.h> 321da177e4SLinus Torvalds #include <linux/completion.h> 331da177e4SLinus Torvalds #include <linux/workqueue.h> 341da177e4SLinus Torvalds #include <linux/slab.h> 351da177e4SLinus Torvalds #include <linux/cpu.h> 361da177e4SLinus Torvalds #include <linux/notifier.h> 371da177e4SLinus Torvalds #include <linux/kthread.h> 381fa44ecaSJames Bottomley #include <linux/hardirq.h> 3946934023SChristoph Lameter #include <linux/mempolicy.h> 40341a5958SRafael J. Wysocki #include <linux/freezer.h> 41d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51e22bee78STejun Heo 52ea138446STejun Heo #include "workqueue_internal.h" 531da177e4SLinus Torvalds 54c8e55f36STejun Heo enum { 55bc2ae0f5STejun Heo /* 5624647570STejun Heo * worker_pool flags 57bc2ae0f5STejun Heo * 5824647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 59bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 60bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 61bc2ae0f5STejun Heo * is in effect. 62bc2ae0f5STejun Heo * 63bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 64bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6524647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 66bc2ae0f5STejun Heo * 67bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 6892f9c5c4SLai Jiangshan * attach_mutex to avoid changing binding state while 694736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 70bc2ae0f5STejun Heo */ 7124647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 7235b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 73db7bccf4STejun Heo 74c8e55f36STejun Heo /* worker flags */ 75c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 76c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 77e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 78fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 79f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 80a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 81e22bee78STejun Heo 82a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 83a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 84db7bccf4STejun Heo 85e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 864ce62e9eSTejun Heo 8729c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 88c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 89db7bccf4STejun Heo 90e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 91e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 92e22bee78STejun Heo 933233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 943233cdbdSTejun Heo /* call for help after 10ms 953233cdbdSTejun Heo (min two ticks) */ 96e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 97e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 981da177e4SLinus Torvalds 991da177e4SLinus Torvalds /* 100e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 101e22bee78STejun Heo * all cpus. Give -20. 102e22bee78STejun Heo */ 103e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1043270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 105ecf6881fSTejun Heo 106ecf6881fSTejun Heo WQ_NAME_LEN = 24, 107c8e55f36STejun Heo }; 108c8e55f36STejun Heo 1091da177e4SLinus Torvalds /* 1104690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1114690c4abSTejun Heo * 112e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 113e41e704bSTejun Heo * everyone else. 1144690c4abSTejun Heo * 115e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 116e22bee78STejun Heo * only be modified and accessed from the local cpu. 117e22bee78STejun Heo * 118d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1194690c4abSTejun Heo * 120d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 121d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 122d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 123d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 124e22bee78STejun Heo * 12592f9c5c4SLai Jiangshan * A: pool->attach_mutex protected. 126822d8405STejun Heo * 12768e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12876af4d93STejun Heo * 12968e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1305bcab335STejun Heo * 1313c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1323c25a55dSLai Jiangshan * 133b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1342e109a28STejun Heo * 1352e109a28STejun Heo * MD: wq_mayday_lock protected. 1364690c4abSTejun Heo */ 1374690c4abSTejun Heo 1382eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 139c34056a3STejun Heo 140bd7bdd43STejun Heo struct worker_pool { 141d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 142d84ff051STejun Heo int cpu; /* I: the associated cpu */ 143f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1449daf9e67STejun Heo int id; /* I: pool ID */ 14511ebea50STejun Heo unsigned int flags; /* X: flags */ 146bd7bdd43STejun Heo 147bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 148bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 149ea1abd61SLai Jiangshan 150ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 151bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 152bd7bdd43STejun Heo 153bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 154bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 155bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 156bd7bdd43STejun Heo 157c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 158c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 159c9e7cf27STejun Heo /* L: hash of busy workers */ 160c9e7cf27STejun Heo 161bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16234a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 16392f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 16492f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 16560f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 166e19e397aSTejun Heo 1677cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 1687cda9aaeSLai Jiangshan 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 343f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 344f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 345f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3467a62c2c8STejun Heo (pool)++) 3474ce62e9eSTejun Heo 34849e3cf44STejun Heo /** 34917116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 35017116969STejun Heo * @pool: iteration cursor 351611c92a0STejun Heo * @pi: integer used for iteration 352fa1b54e6STejun Heo * 35368e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 35468e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 35568e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 356fa1b54e6STejun Heo * 357fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 358fa1b54e6STejun Heo * ignored. 35917116969STejun Heo */ 360611c92a0STejun Heo #define for_each_pool(pool, pi) \ 361611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 36268e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 363fa1b54e6STejun Heo else 36417116969STejun Heo 36517116969STejun Heo /** 366822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 367822d8405STejun Heo * @worker: iteration cursor 368822d8405STejun Heo * @pool: worker_pool to iterate workers of 369822d8405STejun Heo * 37092f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 371822d8405STejun Heo * 372822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 373822d8405STejun Heo * ignored. 374822d8405STejun Heo */ 375da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 376da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 37792f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 378822d8405STejun Heo else 379822d8405STejun Heo 380822d8405STejun Heo /** 38149e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 38249e3cf44STejun Heo * @pwq: iteration cursor 38349e3cf44STejun Heo * @wq: the target workqueue 38476af4d93STejun Heo * 385b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 386794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 387794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 38876af4d93STejun Heo * 38976af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 39076af4d93STejun Heo * ignored. 39149e3cf44STejun Heo */ 39249e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 39376af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 394b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 39576af4d93STejun Heo else 396f3421797STejun Heo 397dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 398dc186ad7SThomas Gleixner 399dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 400dc186ad7SThomas Gleixner 40199777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 40299777288SStanislaw Gruszka { 40399777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 40499777288SStanislaw Gruszka } 40599777288SStanislaw Gruszka 406dc186ad7SThomas Gleixner /* 407dc186ad7SThomas Gleixner * fixup_init is called when: 408dc186ad7SThomas Gleixner * - an active object is initialized 409dc186ad7SThomas Gleixner */ 410dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 411dc186ad7SThomas Gleixner { 412dc186ad7SThomas Gleixner struct work_struct *work = addr; 413dc186ad7SThomas Gleixner 414dc186ad7SThomas Gleixner switch (state) { 415dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 416dc186ad7SThomas Gleixner cancel_work_sync(work); 417dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 418dc186ad7SThomas Gleixner return 1; 419dc186ad7SThomas Gleixner default: 420dc186ad7SThomas Gleixner return 0; 421dc186ad7SThomas Gleixner } 422dc186ad7SThomas Gleixner } 423dc186ad7SThomas Gleixner 424dc186ad7SThomas Gleixner /* 425dc186ad7SThomas Gleixner * fixup_activate is called when: 426dc186ad7SThomas Gleixner * - an active object is activated 427dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 428dc186ad7SThomas Gleixner */ 429dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 430dc186ad7SThomas Gleixner { 431dc186ad7SThomas Gleixner struct work_struct *work = addr; 432dc186ad7SThomas Gleixner 433dc186ad7SThomas Gleixner switch (state) { 434dc186ad7SThomas Gleixner 435dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 436dc186ad7SThomas Gleixner /* 437dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 438dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 439dc186ad7SThomas Gleixner * is tracked in the object tracker. 440dc186ad7SThomas Gleixner */ 44122df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 442dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 443dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 444dc186ad7SThomas Gleixner return 0; 445dc186ad7SThomas Gleixner } 446dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 447dc186ad7SThomas Gleixner return 0; 448dc186ad7SThomas Gleixner 449dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 450dc186ad7SThomas Gleixner WARN_ON(1); 451dc186ad7SThomas Gleixner 452dc186ad7SThomas Gleixner default: 453dc186ad7SThomas Gleixner return 0; 454dc186ad7SThomas Gleixner } 455dc186ad7SThomas Gleixner } 456dc186ad7SThomas Gleixner 457dc186ad7SThomas Gleixner /* 458dc186ad7SThomas Gleixner * fixup_free is called when: 459dc186ad7SThomas Gleixner * - an active object is freed 460dc186ad7SThomas Gleixner */ 461dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 462dc186ad7SThomas Gleixner { 463dc186ad7SThomas Gleixner struct work_struct *work = addr; 464dc186ad7SThomas Gleixner 465dc186ad7SThomas Gleixner switch (state) { 466dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 467dc186ad7SThomas Gleixner cancel_work_sync(work); 468dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 469dc186ad7SThomas Gleixner return 1; 470dc186ad7SThomas Gleixner default: 471dc186ad7SThomas Gleixner return 0; 472dc186ad7SThomas Gleixner } 473dc186ad7SThomas Gleixner } 474dc186ad7SThomas Gleixner 475dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 476dc186ad7SThomas Gleixner .name = "work_struct", 47799777288SStanislaw Gruszka .debug_hint = work_debug_hint, 478dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 479dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 480dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 481dc186ad7SThomas Gleixner }; 482dc186ad7SThomas Gleixner 483dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 484dc186ad7SThomas Gleixner { 485dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 486dc186ad7SThomas Gleixner } 487dc186ad7SThomas Gleixner 488dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 489dc186ad7SThomas Gleixner { 490dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 491dc186ad7SThomas Gleixner } 492dc186ad7SThomas Gleixner 493dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 494dc186ad7SThomas Gleixner { 495dc186ad7SThomas Gleixner if (onstack) 496dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 497dc186ad7SThomas Gleixner else 498dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 499dc186ad7SThomas Gleixner } 500dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 501dc186ad7SThomas Gleixner 502dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 503dc186ad7SThomas Gleixner { 504dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 505dc186ad7SThomas Gleixner } 506dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 507dc186ad7SThomas Gleixner 508ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 509ea2e64f2SThomas Gleixner { 510ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 511ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 512ea2e64f2SThomas Gleixner } 513ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 514ea2e64f2SThomas Gleixner 515dc186ad7SThomas Gleixner #else 516dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 517dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 518dc186ad7SThomas Gleixner #endif 519dc186ad7SThomas Gleixner 5204e8b22bdSLi Bin /** 5214e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5224e8b22bdSLi Bin * @pool: the pool pointer of interest 5234e8b22bdSLi Bin * 5244e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5254e8b22bdSLi Bin * successfully, -errno on failure. 5264e8b22bdSLi Bin */ 5279daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5289daf9e67STejun Heo { 5299daf9e67STejun Heo int ret; 5309daf9e67STejun Heo 53168e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5325bcab335STejun Heo 5334e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5344e8b22bdSLi Bin GFP_KERNEL); 535229641a6STejun Heo if (ret >= 0) { 536e68035fbSTejun Heo pool->id = ret; 537229641a6STejun Heo return 0; 538229641a6STejun Heo } 5399daf9e67STejun Heo return ret; 5409daf9e67STejun Heo } 5419daf9e67STejun Heo 54276af4d93STejun Heo /** 543df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 544df2d5ae4STejun Heo * @wq: the target workqueue 545df2d5ae4STejun Heo * @node: the node ID 546df2d5ae4STejun Heo * 547df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 548df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 549df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 550d185af30SYacine Belkadi * 551d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 552df2d5ae4STejun Heo */ 553df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 554df2d5ae4STejun Heo int node) 555df2d5ae4STejun Heo { 556df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 557df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 558df2d5ae4STejun Heo } 559df2d5ae4STejun Heo 56073f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 56173f53c4aSTejun Heo { 56273f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 56373f53c4aSTejun Heo } 56473f53c4aSTejun Heo 56573f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 56673f53c4aSTejun Heo { 56773f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 56873f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 56973f53c4aSTejun Heo } 57073f53c4aSTejun Heo 57173f53c4aSTejun Heo static int work_next_color(int color) 57273f53c4aSTejun Heo { 57373f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 574a848e3b6SOleg Nesterov } 575a848e3b6SOleg Nesterov 5764594bf15SDavid Howells /* 577112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 578112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5797c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5807a22ad75STejun Heo * 581112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 582112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 583bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 584bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5857a22ad75STejun Heo * 586112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5877c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 588112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5897c3eed5cSTejun Heo * available only while the work item is queued. 590bbb68dfaSTejun Heo * 591bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 592bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 593bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 594bbb68dfaSTejun Heo * try to steal the PENDING bit. 5954594bf15SDavid Howells */ 5967a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5977a22ad75STejun Heo unsigned long flags) 5987a22ad75STejun Heo { 5996183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6007a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6017a22ad75STejun Heo } 6027a22ad75STejun Heo 603112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6044690c4abSTejun Heo unsigned long extra_flags) 605365970a1SDavid Howells { 606112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 607112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 608365970a1SDavid Howells } 609365970a1SDavid Howells 6104468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6114468a00fSLai Jiangshan int pool_id) 6124468a00fSLai Jiangshan { 6134468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6144468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6154468a00fSLai Jiangshan } 6164468a00fSLai Jiangshan 6177c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6187c3eed5cSTejun Heo int pool_id) 6194d707b9fSOleg Nesterov { 62023657bb1STejun Heo /* 62123657bb1STejun Heo * The following wmb is paired with the implied mb in 62223657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 62323657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 62423657bb1STejun Heo * owner. 62523657bb1STejun Heo */ 62623657bb1STejun Heo smp_wmb(); 6277c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6284d707b9fSOleg Nesterov } 6294d707b9fSOleg Nesterov 6307a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 631365970a1SDavid Howells { 6327c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6337c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6347a22ad75STejun Heo } 6357a22ad75STejun Heo 636112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6377a22ad75STejun Heo { 638e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6397a22ad75STejun Heo 640112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 641e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 642e120153dSTejun Heo else 643e120153dSTejun Heo return NULL; 6447a22ad75STejun Heo } 6457a22ad75STejun Heo 6467c3eed5cSTejun Heo /** 6477c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6487c3eed5cSTejun Heo * @work: the work item of interest 6497c3eed5cSTejun Heo * 65068e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 65168e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 65268e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 653fa1b54e6STejun Heo * 654fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 655fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 656fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 657fa1b54e6STejun Heo * returned pool is and stays online. 658d185af30SYacine Belkadi * 659d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6607c3eed5cSTejun Heo */ 6617c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6627a22ad75STejun Heo { 663e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6647c3eed5cSTejun Heo int pool_id; 6657a22ad75STejun Heo 66668e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 667fa1b54e6STejun Heo 668112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 669112202d9STejun Heo return ((struct pool_workqueue *) 6707c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6717a22ad75STejun Heo 6727c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6737c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6747a22ad75STejun Heo return NULL; 6757a22ad75STejun Heo 676fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6777c3eed5cSTejun Heo } 6787c3eed5cSTejun Heo 6797c3eed5cSTejun Heo /** 6807c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6817c3eed5cSTejun Heo * @work: the work item of interest 6827c3eed5cSTejun Heo * 683d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6847c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6857c3eed5cSTejun Heo */ 6867c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6877c3eed5cSTejun Heo { 68854d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6897c3eed5cSTejun Heo 690112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 691112202d9STejun Heo return ((struct pool_workqueue *) 69254d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 69354d5b7d0SLai Jiangshan 69454d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6957c3eed5cSTejun Heo } 6967c3eed5cSTejun Heo 697bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 698bbb68dfaSTejun Heo { 6997c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 700bbb68dfaSTejun Heo 7017c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7027c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 703bbb68dfaSTejun Heo } 704bbb68dfaSTejun Heo 705bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 706bbb68dfaSTejun Heo { 707bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 708bbb68dfaSTejun Heo 709112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 710bbb68dfaSTejun Heo } 711bbb68dfaSTejun Heo 712e22bee78STejun Heo /* 7133270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7143270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 715d565ed63STejun Heo * they're being called with pool->lock held. 716e22bee78STejun Heo */ 717e22bee78STejun Heo 71863d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 719649027d7STejun Heo { 720e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 721649027d7STejun Heo } 722649027d7STejun Heo 723e22bee78STejun Heo /* 724e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 725e22bee78STejun Heo * running workers. 726974271c4STejun Heo * 727974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 728706026c2STejun Heo * function will always return %true for unbound pools as long as the 729974271c4STejun Heo * worklist isn't empty. 730e22bee78STejun Heo */ 73163d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 732e22bee78STejun Heo { 73363d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 734e22bee78STejun Heo } 735e22bee78STejun Heo 736e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 73763d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 738e22bee78STejun Heo { 73963d95a91STejun Heo return pool->nr_idle; 740e22bee78STejun Heo } 741e22bee78STejun Heo 742e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 74363d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 744e22bee78STejun Heo { 745e19e397aSTejun Heo return !list_empty(&pool->worklist) && 746e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 747e22bee78STejun Heo } 748e22bee78STejun Heo 749e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 75063d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 751e22bee78STejun Heo { 75263d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 753e22bee78STejun Heo } 754e22bee78STejun Heo 755e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 75663d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 757e22bee78STejun Heo { 75834a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 75963d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 76063d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 761e22bee78STejun Heo 762ea1abd61SLai Jiangshan /* 763ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 764ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 765ea1abd61SLai Jiangshan */ 766ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 767ea1abd61SLai Jiangshan return false; 768ea1abd61SLai Jiangshan 769e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 770e22bee78STejun Heo } 771e22bee78STejun Heo 772e22bee78STejun Heo /* 773e22bee78STejun Heo * Wake up functions. 774e22bee78STejun Heo */ 775e22bee78STejun Heo 7767e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 77763d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7787e11629dSTejun Heo { 77963d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7807e11629dSTejun Heo return NULL; 7817e11629dSTejun Heo 78263d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7837e11629dSTejun Heo } 7847e11629dSTejun Heo 7857e11629dSTejun Heo /** 7867e11629dSTejun Heo * wake_up_worker - wake up an idle worker 78763d95a91STejun Heo * @pool: worker pool to wake worker from 7887e11629dSTejun Heo * 78963d95a91STejun Heo * Wake up the first idle worker of @pool. 7907e11629dSTejun Heo * 7917e11629dSTejun Heo * CONTEXT: 792d565ed63STejun Heo * spin_lock_irq(pool->lock). 7937e11629dSTejun Heo */ 79463d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7957e11629dSTejun Heo { 79663d95a91STejun Heo struct worker *worker = first_worker(pool); 7977e11629dSTejun Heo 7987e11629dSTejun Heo if (likely(worker)) 7997e11629dSTejun Heo wake_up_process(worker->task); 8007e11629dSTejun Heo } 8017e11629dSTejun Heo 8024690c4abSTejun Heo /** 803e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 804e22bee78STejun Heo * @task: task waking up 805e22bee78STejun Heo * @cpu: CPU @task is waking up to 806e22bee78STejun Heo * 807e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 808e22bee78STejun Heo * being awoken. 809e22bee78STejun Heo * 810e22bee78STejun Heo * CONTEXT: 811e22bee78STejun Heo * spin_lock_irq(rq->lock) 812e22bee78STejun Heo */ 813d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 814e22bee78STejun Heo { 815e22bee78STejun Heo struct worker *worker = kthread_data(task); 816e22bee78STejun Heo 81736576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 818ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 819e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 820e22bee78STejun Heo } 82136576000SJoonsoo Kim } 822e22bee78STejun Heo 823e22bee78STejun Heo /** 824e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 825e22bee78STejun Heo * @task: task going to sleep 826e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 827e22bee78STejun Heo * 828e22bee78STejun Heo * This function is called during schedule() when a busy worker is 829e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 830e22bee78STejun Heo * returning pointer to its task. 831e22bee78STejun Heo * 832e22bee78STejun Heo * CONTEXT: 833e22bee78STejun Heo * spin_lock_irq(rq->lock) 834e22bee78STejun Heo * 835d185af30SYacine Belkadi * Return: 836e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 837e22bee78STejun Heo */ 838d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 839e22bee78STejun Heo { 840e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 841111c225aSTejun Heo struct worker_pool *pool; 842e22bee78STejun Heo 843111c225aSTejun Heo /* 844111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 845111c225aSTejun Heo * workers, also reach here, let's not access anything before 846111c225aSTejun Heo * checking NOT_RUNNING. 847111c225aSTejun Heo */ 8482d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 849e22bee78STejun Heo return NULL; 850e22bee78STejun Heo 851111c225aSTejun Heo pool = worker->pool; 852111c225aSTejun Heo 853e22bee78STejun Heo /* this can only happen on the local cpu */ 8546183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8556183c009STejun Heo return NULL; 856e22bee78STejun Heo 857e22bee78STejun Heo /* 858e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 859e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 860e22bee78STejun Heo * Please read comment there. 861e22bee78STejun Heo * 862628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 863628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 864628c78e7STejun Heo * disabled, which in turn means that none else could be 865d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 866628c78e7STejun Heo * lock is safe. 867e22bee78STejun Heo */ 868e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 869e19e397aSTejun Heo !list_empty(&pool->worklist)) 87063d95a91STejun Heo to_wakeup = first_worker(pool); 871e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 872e22bee78STejun Heo } 873e22bee78STejun Heo 874e22bee78STejun Heo /** 875e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 876cb444766STejun Heo * @worker: self 877d302f017STejun Heo * @flags: flags to set 878d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 879d302f017STejun Heo * 880e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 881e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 882e22bee78STejun Heo * woken up. 883d302f017STejun Heo * 884cb444766STejun Heo * CONTEXT: 885d565ed63STejun Heo * spin_lock_irq(pool->lock) 886d302f017STejun Heo */ 887d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 888d302f017STejun Heo bool wakeup) 889d302f017STejun Heo { 890bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 891e22bee78STejun Heo 892cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 893cb444766STejun Heo 894e22bee78STejun Heo /* 895e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 896e22bee78STejun Heo * wake up an idle worker as necessary if requested by 897e22bee78STejun Heo * @wakeup. 898e22bee78STejun Heo */ 899e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 900e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 901e22bee78STejun Heo if (wakeup) { 902e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 903bd7bdd43STejun Heo !list_empty(&pool->worklist)) 90463d95a91STejun Heo wake_up_worker(pool); 905e22bee78STejun Heo } else 906e19e397aSTejun Heo atomic_dec(&pool->nr_running); 907e22bee78STejun Heo } 908e22bee78STejun Heo 909d302f017STejun Heo worker->flags |= flags; 910d302f017STejun Heo } 911d302f017STejun Heo 912d302f017STejun Heo /** 913e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 914cb444766STejun Heo * @worker: self 915d302f017STejun Heo * @flags: flags to clear 916d302f017STejun Heo * 917e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 918d302f017STejun Heo * 919cb444766STejun Heo * CONTEXT: 920d565ed63STejun Heo * spin_lock_irq(pool->lock) 921d302f017STejun Heo */ 922d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 923d302f017STejun Heo { 92463d95a91STejun Heo struct worker_pool *pool = worker->pool; 925e22bee78STejun Heo unsigned int oflags = worker->flags; 926e22bee78STejun Heo 927cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 928cb444766STejun Heo 929d302f017STejun Heo worker->flags &= ~flags; 930e22bee78STejun Heo 93142c025f3STejun Heo /* 93242c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 93342c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 93442c025f3STejun Heo * of multiple flags, not a single flag. 93542c025f3STejun Heo */ 936e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 937e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 938e19e397aSTejun Heo atomic_inc(&pool->nr_running); 939d302f017STejun Heo } 940d302f017STejun Heo 941d302f017STejun Heo /** 9428cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 943c9e7cf27STejun Heo * @pool: pool of interest 9448cca0eeaSTejun Heo * @work: work to find worker for 9458cca0eeaSTejun Heo * 946c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 947c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 948a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 949a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 950a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 951a2c1c57bSTejun Heo * being executed. 952a2c1c57bSTejun Heo * 953a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 954a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 955a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 956a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 957a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 958a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 959a2c1c57bSTejun Heo * 960c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 961c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 962c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 963c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 964c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 965c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9668cca0eeaSTejun Heo * 9678cca0eeaSTejun Heo * CONTEXT: 968d565ed63STejun Heo * spin_lock_irq(pool->lock). 9698cca0eeaSTejun Heo * 970d185af30SYacine Belkadi * Return: 971d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9728cca0eeaSTejun Heo * otherwise. 9738cca0eeaSTejun Heo */ 974c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9758cca0eeaSTejun Heo struct work_struct *work) 9768cca0eeaSTejun Heo { 97742f8570fSSasha Levin struct worker *worker; 97842f8570fSSasha Levin 979b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 980a2c1c57bSTejun Heo (unsigned long)work) 981a2c1c57bSTejun Heo if (worker->current_work == work && 982a2c1c57bSTejun Heo worker->current_func == work->func) 98342f8570fSSasha Levin return worker; 98442f8570fSSasha Levin 98542f8570fSSasha Levin return NULL; 9868cca0eeaSTejun Heo } 9878cca0eeaSTejun Heo 9888cca0eeaSTejun Heo /** 989bf4ede01STejun Heo * move_linked_works - move linked works to a list 990bf4ede01STejun Heo * @work: start of series of works to be scheduled 991bf4ede01STejun Heo * @head: target list to append @work to 992bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 993bf4ede01STejun Heo * 994bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 995bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 996bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 997bf4ede01STejun Heo * 998bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 999bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1000bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1001bf4ede01STejun Heo * 1002bf4ede01STejun Heo * CONTEXT: 1003d565ed63STejun Heo * spin_lock_irq(pool->lock). 1004bf4ede01STejun Heo */ 1005bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1006bf4ede01STejun Heo struct work_struct **nextp) 1007bf4ede01STejun Heo { 1008bf4ede01STejun Heo struct work_struct *n; 1009bf4ede01STejun Heo 1010bf4ede01STejun Heo /* 1011bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1012bf4ede01STejun Heo * use NULL for list head. 1013bf4ede01STejun Heo */ 1014bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1015bf4ede01STejun Heo list_move_tail(&work->entry, head); 1016bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1017bf4ede01STejun Heo break; 1018bf4ede01STejun Heo } 1019bf4ede01STejun Heo 1020bf4ede01STejun Heo /* 1021bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1022bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1023bf4ede01STejun Heo * needs to be updated. 1024bf4ede01STejun Heo */ 1025bf4ede01STejun Heo if (nextp) 1026bf4ede01STejun Heo *nextp = n; 1027bf4ede01STejun Heo } 1028bf4ede01STejun Heo 10298864b4e5STejun Heo /** 10308864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10318864b4e5STejun Heo * @pwq: pool_workqueue to get 10328864b4e5STejun Heo * 10338864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10348864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10358864b4e5STejun Heo */ 10368864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10378864b4e5STejun Heo { 10388864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10398864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10408864b4e5STejun Heo pwq->refcnt++; 10418864b4e5STejun Heo } 10428864b4e5STejun Heo 10438864b4e5STejun Heo /** 10448864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10458864b4e5STejun Heo * @pwq: pool_workqueue to put 10468864b4e5STejun Heo * 10478864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10488864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10498864b4e5STejun Heo */ 10508864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10518864b4e5STejun Heo { 10528864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10538864b4e5STejun Heo if (likely(--pwq->refcnt)) 10548864b4e5STejun Heo return; 10558864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10568864b4e5STejun Heo return; 10578864b4e5STejun Heo /* 10588864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10598864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10608864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10618864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10628864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10638864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10648864b4e5STejun Heo */ 10658864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10668864b4e5STejun Heo } 10678864b4e5STejun Heo 1068dce90d47STejun Heo /** 1069dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1070dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1071dce90d47STejun Heo * 1072dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1073dce90d47STejun Heo */ 1074dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1075dce90d47STejun Heo { 1076dce90d47STejun Heo if (pwq) { 1077dce90d47STejun Heo /* 1078dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1079dce90d47STejun Heo * following lock operations are safe. 1080dce90d47STejun Heo */ 1081dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1082dce90d47STejun Heo put_pwq(pwq); 1083dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1084dce90d47STejun Heo } 1085dce90d47STejun Heo } 1086dce90d47STejun Heo 1087112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1088bf4ede01STejun Heo { 1089112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1090bf4ede01STejun Heo 1091bf4ede01STejun Heo trace_workqueue_activate_work(work); 1092112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1093bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1094112202d9STejun Heo pwq->nr_active++; 1095bf4ede01STejun Heo } 1096bf4ede01STejun Heo 1097112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10983aa62497SLai Jiangshan { 1099112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11003aa62497SLai Jiangshan struct work_struct, entry); 11013aa62497SLai Jiangshan 1102112202d9STejun Heo pwq_activate_delayed_work(work); 11033aa62497SLai Jiangshan } 11043aa62497SLai Jiangshan 1105bf4ede01STejun Heo /** 1106112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1107112202d9STejun Heo * @pwq: pwq of interest 1108bf4ede01STejun Heo * @color: color of work which left the queue 1109bf4ede01STejun Heo * 1110bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1111112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1112bf4ede01STejun Heo * 1113bf4ede01STejun Heo * CONTEXT: 1114d565ed63STejun Heo * spin_lock_irq(pool->lock). 1115bf4ede01STejun Heo */ 1116112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1117bf4ede01STejun Heo { 11188864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1119bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11208864b4e5STejun Heo goto out_put; 1121bf4ede01STejun Heo 1122112202d9STejun Heo pwq->nr_in_flight[color]--; 1123bf4ede01STejun Heo 1124112202d9STejun Heo pwq->nr_active--; 1125112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1126bf4ede01STejun Heo /* one down, submit a delayed one */ 1127112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1128112202d9STejun Heo pwq_activate_first_delayed(pwq); 1129bf4ede01STejun Heo } 1130bf4ede01STejun Heo 1131bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1132112202d9STejun Heo if (likely(pwq->flush_color != color)) 11338864b4e5STejun Heo goto out_put; 1134bf4ede01STejun Heo 1135bf4ede01STejun Heo /* are there still in-flight works? */ 1136112202d9STejun Heo if (pwq->nr_in_flight[color]) 11378864b4e5STejun Heo goto out_put; 1138bf4ede01STejun Heo 1139112202d9STejun Heo /* this pwq is done, clear flush_color */ 1140112202d9STejun Heo pwq->flush_color = -1; 1141bf4ede01STejun Heo 1142bf4ede01STejun Heo /* 1143112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1144bf4ede01STejun Heo * will handle the rest. 1145bf4ede01STejun Heo */ 1146112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1147112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11488864b4e5STejun Heo out_put: 11498864b4e5STejun Heo put_pwq(pwq); 1150bf4ede01STejun Heo } 1151bf4ede01STejun Heo 115236e227d2STejun Heo /** 1153bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 115436e227d2STejun Heo * @work: work item to steal 115536e227d2STejun Heo * @is_dwork: @work is a delayed_work 1156bbb68dfaSTejun Heo * @flags: place to store irq state 115736e227d2STejun Heo * 115836e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1159d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 116036e227d2STejun Heo * 1161d185af30SYacine Belkadi * Return: 116236e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 116336e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 116436e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1165bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1166bbb68dfaSTejun Heo * for arbitrarily long 116736e227d2STejun Heo * 1168d185af30SYacine Belkadi * Note: 1169bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1170e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1171e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1172e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1173bbb68dfaSTejun Heo * 1174bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1175bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1176bbb68dfaSTejun Heo * 1177e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1178bf4ede01STejun Heo */ 1179bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1180bbb68dfaSTejun Heo unsigned long *flags) 1181bf4ede01STejun Heo { 1182d565ed63STejun Heo struct worker_pool *pool; 1183112202d9STejun Heo struct pool_workqueue *pwq; 1184bf4ede01STejun Heo 1185bbb68dfaSTejun Heo local_irq_save(*flags); 1186bbb68dfaSTejun Heo 118736e227d2STejun Heo /* try to steal the timer if it exists */ 118836e227d2STejun Heo if (is_dwork) { 118936e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 119036e227d2STejun Heo 1191e0aecdd8STejun Heo /* 1192e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1193e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1194e0aecdd8STejun Heo * running on the local CPU. 1195e0aecdd8STejun Heo */ 119636e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 119736e227d2STejun Heo return 1; 119836e227d2STejun Heo } 119936e227d2STejun Heo 120036e227d2STejun Heo /* try to claim PENDING the normal way */ 1201bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1202bf4ede01STejun Heo return 0; 1203bf4ede01STejun Heo 1204bf4ede01STejun Heo /* 1205bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1206bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1207bf4ede01STejun Heo */ 1208d565ed63STejun Heo pool = get_work_pool(work); 1209d565ed63STejun Heo if (!pool) 1210bbb68dfaSTejun Heo goto fail; 1211bf4ede01STejun Heo 1212d565ed63STejun Heo spin_lock(&pool->lock); 1213bf4ede01STejun Heo /* 1214112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1215112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1216112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1217112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1218112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12190b3dae68SLai Jiangshan * item is currently queued on that pool. 1220bf4ede01STejun Heo */ 1221112202d9STejun Heo pwq = get_work_pwq(work); 1222112202d9STejun Heo if (pwq && pwq->pool == pool) { 1223bf4ede01STejun Heo debug_work_deactivate(work); 12243aa62497SLai Jiangshan 12253aa62497SLai Jiangshan /* 122616062836STejun Heo * A delayed work item cannot be grabbed directly because 122716062836STejun Heo * it might have linked NO_COLOR work items which, if left 1228112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 122916062836STejun Heo * management later on and cause stall. Make sure the work 123016062836STejun Heo * item is activated before grabbing. 12313aa62497SLai Jiangshan */ 12323aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1233112202d9STejun Heo pwq_activate_delayed_work(work); 12343aa62497SLai Jiangshan 1235bf4ede01STejun Heo list_del_init(&work->entry); 1236112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 123736e227d2STejun Heo 1238112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12394468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12404468a00fSLai Jiangshan 1241d565ed63STejun Heo spin_unlock(&pool->lock); 124236e227d2STejun Heo return 1; 1243bf4ede01STejun Heo } 1244d565ed63STejun Heo spin_unlock(&pool->lock); 1245bbb68dfaSTejun Heo fail: 1246bbb68dfaSTejun Heo local_irq_restore(*flags); 1247bbb68dfaSTejun Heo if (work_is_canceling(work)) 1248bbb68dfaSTejun Heo return -ENOENT; 1249bbb68dfaSTejun Heo cpu_relax(); 125036e227d2STejun Heo return -EAGAIN; 1251bf4ede01STejun Heo } 1252bf4ede01STejun Heo 1253bf4ede01STejun Heo /** 1254706026c2STejun Heo * insert_work - insert a work into a pool 1255112202d9STejun Heo * @pwq: pwq @work belongs to 12564690c4abSTejun Heo * @work: work to insert 12574690c4abSTejun Heo * @head: insertion point 12584690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12594690c4abSTejun Heo * 1260112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1261706026c2STejun Heo * work_struct flags. 12624690c4abSTejun Heo * 12634690c4abSTejun Heo * CONTEXT: 1264d565ed63STejun Heo * spin_lock_irq(pool->lock). 1265365970a1SDavid Howells */ 1266112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1267112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1268b89deed3SOleg Nesterov { 1269112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1270e1d8aa9fSFrederic Weisbecker 12714690c4abSTejun Heo /* we own @work, set data and link */ 1272112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12731a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12748864b4e5STejun Heo get_pwq(pwq); 1275e22bee78STejun Heo 1276e22bee78STejun Heo /* 1277c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1278c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1279c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1280e22bee78STejun Heo */ 1281e22bee78STejun Heo smp_mb(); 1282e22bee78STejun Heo 128363d95a91STejun Heo if (__need_more_worker(pool)) 128463d95a91STejun Heo wake_up_worker(pool); 1285b89deed3SOleg Nesterov } 1286b89deed3SOleg Nesterov 1287c8efcc25STejun Heo /* 1288c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12898d03ecfeSTejun Heo * same workqueue. 1290c8efcc25STejun Heo */ 1291c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1292c8efcc25STejun Heo { 1293c8efcc25STejun Heo struct worker *worker; 1294c8efcc25STejun Heo 12958d03ecfeSTejun Heo worker = current_wq_worker(); 1296c8efcc25STejun Heo /* 12978d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12988d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1299c8efcc25STejun Heo */ 1300112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1301c8efcc25STejun Heo } 1302c8efcc25STejun Heo 1303d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13041da177e4SLinus Torvalds struct work_struct *work) 13051da177e4SLinus Torvalds { 1306112202d9STejun Heo struct pool_workqueue *pwq; 1307c9178087STejun Heo struct worker_pool *last_pool; 13081e19ffc6STejun Heo struct list_head *worklist; 13098a2e8e5dSTejun Heo unsigned int work_flags; 1310b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13118930cabaSTejun Heo 13128930cabaSTejun Heo /* 13138930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13148930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13158930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13168930cabaSTejun Heo * happen with IRQ disabled. 13178930cabaSTejun Heo */ 13188930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13191da177e4SLinus Torvalds 1320dc186ad7SThomas Gleixner debug_work_activate(work); 13211e19ffc6STejun Heo 13229ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1323618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1324c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1325e41e704bSTejun Heo return; 13269e8cd2f5STejun Heo retry: 1327df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1328f3421797STejun Heo cpu = raw_smp_processor_id(); 1329df2d5ae4STejun Heo 1330df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1331df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1332c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1333df2d5ae4STejun Heo else 1334df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1335f3421797STejun Heo 133618aa9effSTejun Heo /* 1337c9178087STejun Heo * If @work was previously on a different pool, it might still be 1338c9178087STejun Heo * running there, in which case the work needs to be queued on that 1339c9178087STejun Heo * pool to guarantee non-reentrancy. 134018aa9effSTejun Heo */ 1341c9e7cf27STejun Heo last_pool = get_work_pool(work); 1342112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 134318aa9effSTejun Heo struct worker *worker; 134418aa9effSTejun Heo 1345d565ed63STejun Heo spin_lock(&last_pool->lock); 134618aa9effSTejun Heo 1347c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 134818aa9effSTejun Heo 1349112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1350c9178087STejun Heo pwq = worker->current_pwq; 13518594fadeSLai Jiangshan } else { 135218aa9effSTejun Heo /* meh... not running there, queue here */ 1353d565ed63STejun Heo spin_unlock(&last_pool->lock); 1354112202d9STejun Heo spin_lock(&pwq->pool->lock); 135518aa9effSTejun Heo } 13568930cabaSTejun Heo } else { 1357112202d9STejun Heo spin_lock(&pwq->pool->lock); 13588930cabaSTejun Heo } 1359502ca9d8STejun Heo 13609e8cd2f5STejun Heo /* 13619e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13629e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13639e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1364df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1365df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13669e8cd2f5STejun Heo * make forward-progress. 13679e8cd2f5STejun Heo */ 13689e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13699e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13709e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13719e8cd2f5STejun Heo cpu_relax(); 13729e8cd2f5STejun Heo goto retry; 13739e8cd2f5STejun Heo } 13749e8cd2f5STejun Heo /* oops */ 13759e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13769e8cd2f5STejun Heo wq->name, cpu); 13779e8cd2f5STejun Heo } 13789e8cd2f5STejun Heo 1379112202d9STejun Heo /* pwq determined, queue */ 1380112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1381502ca9d8STejun Heo 1382f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1383112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1384f5b2552bSDan Carpenter return; 1385f5b2552bSDan Carpenter } 13861e19ffc6STejun Heo 1387112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1388112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13891e19ffc6STejun Heo 1390112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1391cdadf009STejun Heo trace_workqueue_activate_work(work); 1392112202d9STejun Heo pwq->nr_active++; 1393112202d9STejun Heo worklist = &pwq->pool->worklist; 13948a2e8e5dSTejun Heo } else { 13958a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1396112202d9STejun Heo worklist = &pwq->delayed_works; 13978a2e8e5dSTejun Heo } 13981e19ffc6STejun Heo 1399112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14001e19ffc6STejun Heo 1401112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14021da177e4SLinus Torvalds } 14031da177e4SLinus Torvalds 14040fcb78c2SRolf Eike Beer /** 1405c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1406c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1407c1a220e7SZhang Rui * @wq: workqueue to use 1408c1a220e7SZhang Rui * @work: work to queue 1409c1a220e7SZhang Rui * 1410c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1411c1a220e7SZhang Rui * can't go away. 1412d185af30SYacine Belkadi * 1413d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1414c1a220e7SZhang Rui */ 1415d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1416d4283e93STejun Heo struct work_struct *work) 1417c1a220e7SZhang Rui { 1418d4283e93STejun Heo bool ret = false; 14198930cabaSTejun Heo unsigned long flags; 14208930cabaSTejun Heo 14218930cabaSTejun Heo local_irq_save(flags); 1422c1a220e7SZhang Rui 142322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14244690c4abSTejun Heo __queue_work(cpu, wq, work); 1425d4283e93STejun Heo ret = true; 1426c1a220e7SZhang Rui } 14278930cabaSTejun Heo 14288930cabaSTejun Heo local_irq_restore(flags); 1429c1a220e7SZhang Rui return ret; 1430c1a220e7SZhang Rui } 1431ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1432c1a220e7SZhang Rui 1433d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14341da177e4SLinus Torvalds { 143552bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14361da177e4SLinus Torvalds 1437e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 143860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14391da177e4SLinus Torvalds } 14401438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14411da177e4SLinus Torvalds 14427beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 144352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14441da177e4SLinus Torvalds { 14457beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14467beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14471da177e4SLinus Torvalds 14487beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14497beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1450fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1451fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14527beb2edfSTejun Heo 14538852aac2STejun Heo /* 14548852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14558852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14568852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14578852aac2STejun Heo * on that there's no such delay when @delay is 0. 14588852aac2STejun Heo */ 14598852aac2STejun Heo if (!delay) { 14608852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14618852aac2STejun Heo return; 14628852aac2STejun Heo } 14638852aac2STejun Heo 14647beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14657beb2edfSTejun Heo 146660c057bcSLai Jiangshan dwork->wq = wq; 14671265057fSTejun Heo dwork->cpu = cpu; 14687beb2edfSTejun Heo timer->expires = jiffies + delay; 14697beb2edfSTejun Heo 14707beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14717beb2edfSTejun Heo add_timer_on(timer, cpu); 14727beb2edfSTejun Heo else 14737beb2edfSTejun Heo add_timer(timer); 14747beb2edfSTejun Heo } 14751da177e4SLinus Torvalds 14760fcb78c2SRolf Eike Beer /** 14770fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14780fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14790fcb78c2SRolf Eike Beer * @wq: workqueue to use 1480af9997e4SRandy Dunlap * @dwork: work to queue 14810fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14820fcb78c2SRolf Eike Beer * 1483d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1484715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1485715f1300STejun Heo * execution. 14860fcb78c2SRolf Eike Beer */ 1487d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 148852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14897a6bc1cdSVenkatesh Pallipadi { 149052bad64dSDavid Howells struct work_struct *work = &dwork->work; 1491d4283e93STejun Heo bool ret = false; 14928930cabaSTejun Heo unsigned long flags; 14938930cabaSTejun Heo 14948930cabaSTejun Heo /* read the comment in __queue_work() */ 14958930cabaSTejun Heo local_irq_save(flags); 14967a6bc1cdSVenkatesh Pallipadi 149722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14987beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1499d4283e93STejun Heo ret = true; 15007a6bc1cdSVenkatesh Pallipadi } 15018930cabaSTejun Heo 15028930cabaSTejun Heo local_irq_restore(flags); 15037a6bc1cdSVenkatesh Pallipadi return ret; 15047a6bc1cdSVenkatesh Pallipadi } 1505ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15061da177e4SLinus Torvalds 1507c8e55f36STejun Heo /** 15088376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15098376fe22STejun Heo * @cpu: CPU number to execute work on 15108376fe22STejun Heo * @wq: workqueue to use 15118376fe22STejun Heo * @dwork: work to queue 15128376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15138376fe22STejun Heo * 15148376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15158376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15168376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15178376fe22STejun Heo * current state. 15188376fe22STejun Heo * 1519d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15208376fe22STejun Heo * pending and its timer was modified. 15218376fe22STejun Heo * 1522e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15238376fe22STejun Heo * See try_to_grab_pending() for details. 15248376fe22STejun Heo */ 15258376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15268376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15278376fe22STejun Heo { 15288376fe22STejun Heo unsigned long flags; 15298376fe22STejun Heo int ret; 15308376fe22STejun Heo 15318376fe22STejun Heo do { 15328376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15338376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15348376fe22STejun Heo 15358376fe22STejun Heo if (likely(ret >= 0)) { 15368376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15378376fe22STejun Heo local_irq_restore(flags); 15388376fe22STejun Heo } 15398376fe22STejun Heo 15408376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15418376fe22STejun Heo return ret; 15428376fe22STejun Heo } 15438376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15448376fe22STejun Heo 15458376fe22STejun Heo /** 1546c8e55f36STejun Heo * worker_enter_idle - enter idle state 1547c8e55f36STejun Heo * @worker: worker which is entering idle state 1548c8e55f36STejun Heo * 1549c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1550c8e55f36STejun Heo * necessary. 1551c8e55f36STejun Heo * 1552c8e55f36STejun Heo * LOCKING: 1553d565ed63STejun Heo * spin_lock_irq(pool->lock). 1554c8e55f36STejun Heo */ 1555c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15561da177e4SLinus Torvalds { 1557bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1558c8e55f36STejun Heo 15596183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15606183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15616183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15626183c009STejun Heo return; 1563c8e55f36STejun Heo 1564cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1565cb444766STejun Heo worker->flags |= WORKER_IDLE; 1566bd7bdd43STejun Heo pool->nr_idle++; 1567e22bee78STejun Heo worker->last_active = jiffies; 1568c8e55f36STejun Heo 1569c8e55f36STejun Heo /* idle_list is LIFO */ 1570bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1571db7bccf4STejun Heo 157263d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1573628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1574cb444766STejun Heo 1575544ecf31STejun Heo /* 1576706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1577d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1578628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1579628c78e7STejun Heo * unbind is not in progress. 1580544ecf31STejun Heo */ 158124647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1582bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1583e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1584c8e55f36STejun Heo } 1585c8e55f36STejun Heo 1586c8e55f36STejun Heo /** 1587c8e55f36STejun Heo * worker_leave_idle - leave idle state 1588c8e55f36STejun Heo * @worker: worker which is leaving idle state 1589c8e55f36STejun Heo * 1590c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1591c8e55f36STejun Heo * 1592c8e55f36STejun Heo * LOCKING: 1593d565ed63STejun Heo * spin_lock_irq(pool->lock). 1594c8e55f36STejun Heo */ 1595c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1596c8e55f36STejun Heo { 1597bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1598c8e55f36STejun Heo 15996183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16006183c009STejun Heo return; 1601d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1602bd7bdd43STejun Heo pool->nr_idle--; 1603c8e55f36STejun Heo list_del_init(&worker->entry); 1604c8e55f36STejun Heo } 1605c8e55f36STejun Heo 1606c34056a3STejun Heo static struct worker *alloc_worker(void) 1607c34056a3STejun Heo { 1608c34056a3STejun Heo struct worker *worker; 1609c34056a3STejun Heo 1610c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1611c8e55f36STejun Heo if (worker) { 1612c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1613affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1614da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1615e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1616e22bee78STejun Heo worker->flags = WORKER_PREP; 1617c8e55f36STejun Heo } 1618c34056a3STejun Heo return worker; 1619c34056a3STejun Heo } 1620c34056a3STejun Heo 1621c34056a3STejun Heo /** 16224736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16234736cbf7SLai Jiangshan * @worker: worker to be attached 16244736cbf7SLai Jiangshan * @pool: the target pool 16254736cbf7SLai Jiangshan * 16264736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16274736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16284736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16294736cbf7SLai Jiangshan */ 16304736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16314736cbf7SLai Jiangshan struct worker_pool *pool) 16324736cbf7SLai Jiangshan { 16334736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16344736cbf7SLai Jiangshan 16354736cbf7SLai Jiangshan /* 16364736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16374736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16384736cbf7SLai Jiangshan */ 16394736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16404736cbf7SLai Jiangshan 16414736cbf7SLai Jiangshan /* 16424736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 16434736cbf7SLai Jiangshan * stable across this function. See the comments above the 16444736cbf7SLai Jiangshan * flag definition for details. 16454736cbf7SLai Jiangshan */ 16464736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 16474736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 16484736cbf7SLai Jiangshan 16494736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 16504736cbf7SLai Jiangshan 16514736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 16524736cbf7SLai Jiangshan } 16534736cbf7SLai Jiangshan 16544736cbf7SLai Jiangshan /** 165560f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 165660f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 165760f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 165860f5a4bcSLai Jiangshan * 16594736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 16604736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 16614736cbf7SLai Jiangshan * other reference to the pool. 166260f5a4bcSLai Jiangshan */ 166360f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 166460f5a4bcSLai Jiangshan struct worker_pool *pool) 166560f5a4bcSLai Jiangshan { 166660f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 166760f5a4bcSLai Jiangshan 166892f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1669da028469SLai Jiangshan list_del(&worker->node); 1670da028469SLai Jiangshan if (list_empty(&pool->workers)) 167160f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 167292f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 167360f5a4bcSLai Jiangshan 167460f5a4bcSLai Jiangshan if (detach_completion) 167560f5a4bcSLai Jiangshan complete(detach_completion); 167660f5a4bcSLai Jiangshan } 167760f5a4bcSLai Jiangshan 167860f5a4bcSLai Jiangshan /** 1679c34056a3STejun Heo * create_worker - create a new workqueue worker 168063d95a91STejun Heo * @pool: pool the new worker will belong to 1681c34056a3STejun Heo * 168273eb7fe7SLai Jiangshan * Create a new worker which is attached to @pool. The new worker must be 168373eb7fe7SLai Jiangshan * started by start_worker(). 1684c34056a3STejun Heo * 1685c34056a3STejun Heo * CONTEXT: 1686c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1687c34056a3STejun Heo * 1688d185af30SYacine Belkadi * Return: 1689c34056a3STejun Heo * Pointer to the newly created worker. 1690c34056a3STejun Heo */ 1691bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1692c34056a3STejun Heo { 1693c34056a3STejun Heo struct worker *worker = NULL; 1694f3421797STejun Heo int id = -1; 1695e3c916a4STejun Heo char id_buf[16]; 1696c34056a3STejun Heo 16977cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 16987cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1699822d8405STejun Heo if (id < 0) 1700c34056a3STejun Heo goto fail; 1701c34056a3STejun Heo 1702c34056a3STejun Heo worker = alloc_worker(); 1703c34056a3STejun Heo if (!worker) 1704c34056a3STejun Heo goto fail; 1705c34056a3STejun Heo 1706bd7bdd43STejun Heo worker->pool = pool; 1707c34056a3STejun Heo worker->id = id; 1708c34056a3STejun Heo 170929c91e99STejun Heo if (pool->cpu >= 0) 1710e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1711e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1712f3421797STejun Heo else 1713e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1714e3c916a4STejun Heo 1715f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1716e3c916a4STejun Heo "kworker/%s", id_buf); 1717c34056a3STejun Heo if (IS_ERR(worker->task)) 1718c34056a3STejun Heo goto fail; 1719c34056a3STejun Heo 172091151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 172191151228SOleg Nesterov 172291151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 172391151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 172491151228SOleg Nesterov 1725da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17264736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 17274d757c5cSLai Jiangshan 1728c34056a3STejun Heo return worker; 1729822d8405STejun Heo 1730c34056a3STejun Heo fail: 17319625ab17SLai Jiangshan if (id >= 0) 17327cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1733c34056a3STejun Heo kfree(worker); 1734c34056a3STejun Heo return NULL; 1735c34056a3STejun Heo } 1736c34056a3STejun Heo 1737c34056a3STejun Heo /** 1738c34056a3STejun Heo * start_worker - start a newly created worker 1739c34056a3STejun Heo * @worker: worker to start 1740c34056a3STejun Heo * 1741706026c2STejun Heo * Make the pool aware of @worker and start it. 1742c34056a3STejun Heo * 1743c34056a3STejun Heo * CONTEXT: 1744d565ed63STejun Heo * spin_lock_irq(pool->lock). 1745c34056a3STejun Heo */ 1746c34056a3STejun Heo static void start_worker(struct worker *worker) 1747c34056a3STejun Heo { 1748bd7bdd43STejun Heo worker->pool->nr_workers++; 1749c8e55f36STejun Heo worker_enter_idle(worker); 1750c34056a3STejun Heo wake_up_process(worker->task); 1751c34056a3STejun Heo } 1752c34056a3STejun Heo 1753c34056a3STejun Heo /** 1754ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1755ebf44d16STejun Heo * @pool: the target pool 1756ebf44d16STejun Heo * 1757cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1758d185af30SYacine Belkadi * 1759d185af30SYacine Belkadi * Return: 0 on success. A negative error code otherwise. 1760ebf44d16STejun Heo */ 1761ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1762ebf44d16STejun Heo { 1763ebf44d16STejun Heo struct worker *worker; 1764ebf44d16STejun Heo 1765ebf44d16STejun Heo worker = create_worker(pool); 1766ebf44d16STejun Heo if (worker) { 1767ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1768ebf44d16STejun Heo start_worker(worker); 1769ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1770ebf44d16STejun Heo } 1771ebf44d16STejun Heo 1772ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1773ebf44d16STejun Heo } 1774ebf44d16STejun Heo 1775ebf44d16STejun Heo /** 1776c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1777c34056a3STejun Heo * @worker: worker to be destroyed 1778c34056a3STejun Heo * 177973eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 178073eb7fe7SLai Jiangshan * be idle. 1781c8e55f36STejun Heo * 1782c8e55f36STejun Heo * CONTEXT: 178360f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1784c34056a3STejun Heo */ 1785c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1786c34056a3STejun Heo { 1787bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1788c34056a3STejun Heo 1789cd549687STejun Heo lockdep_assert_held(&pool->lock); 1790cd549687STejun Heo 1791c34056a3STejun Heo /* sanity check frenzy */ 17926183c009STejun Heo if (WARN_ON(worker->current_work) || 179373eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 179473eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 17956183c009STejun Heo return; 1796c34056a3STejun Heo 1797bd7bdd43STejun Heo pool->nr_workers--; 1798bd7bdd43STejun Heo pool->nr_idle--; 1799c8e55f36STejun Heo 1800c8e55f36STejun Heo list_del_init(&worker->entry); 1801cb444766STejun Heo worker->flags |= WORKER_DIE; 180260f5a4bcSLai Jiangshan wake_up_process(worker->task); 1803c34056a3STejun Heo } 1804c34056a3STejun Heo 180563d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1806e22bee78STejun Heo { 180763d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1808e22bee78STejun Heo 1809d565ed63STejun Heo spin_lock_irq(&pool->lock); 1810e22bee78STejun Heo 18113347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1812e22bee78STejun Heo struct worker *worker; 1813e22bee78STejun Heo unsigned long expires; 1814e22bee78STejun Heo 1815e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 181663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1817e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1818e22bee78STejun Heo 18193347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 182063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18213347fc9fSLai Jiangshan break; 1822e22bee78STejun Heo } 18233347fc9fSLai Jiangshan 18243347fc9fSLai Jiangshan destroy_worker(worker); 1825e22bee78STejun Heo } 1826e22bee78STejun Heo 1827d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1828e22bee78STejun Heo } 1829e22bee78STejun Heo 1830493a1724STejun Heo static void send_mayday(struct work_struct *work) 1831e22bee78STejun Heo { 1832112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1833112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1834493a1724STejun Heo 18352e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1836e22bee78STejun Heo 1837493008a8STejun Heo if (!wq->rescuer) 1838493a1724STejun Heo return; 1839e22bee78STejun Heo 1840e22bee78STejun Heo /* mayday mayday mayday */ 1841493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 184277668c8bSLai Jiangshan /* 184377668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 184477668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 184577668c8bSLai Jiangshan * rescuer is done with it. 184677668c8bSLai Jiangshan */ 184777668c8bSLai Jiangshan get_pwq(pwq); 1848493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1849e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1850493a1724STejun Heo } 1851e22bee78STejun Heo } 1852e22bee78STejun Heo 1853706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1854e22bee78STejun Heo { 185563d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1856e22bee78STejun Heo struct work_struct *work; 1857e22bee78STejun Heo 18582e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1859493a1724STejun Heo spin_lock(&pool->lock); 1860e22bee78STejun Heo 186163d95a91STejun Heo if (need_to_create_worker(pool)) { 1862e22bee78STejun Heo /* 1863e22bee78STejun Heo * We've been trying to create a new worker but 1864e22bee78STejun Heo * haven't been successful. We might be hitting an 1865e22bee78STejun Heo * allocation deadlock. Send distress signals to 1866e22bee78STejun Heo * rescuers. 1867e22bee78STejun Heo */ 186863d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1869e22bee78STejun Heo send_mayday(work); 1870e22bee78STejun Heo } 1871e22bee78STejun Heo 1872493a1724STejun Heo spin_unlock(&pool->lock); 18732e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1874e22bee78STejun Heo 187563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1876e22bee78STejun Heo } 1877e22bee78STejun Heo 1878e22bee78STejun Heo /** 1879e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 188063d95a91STejun Heo * @pool: pool to create a new worker for 1881e22bee78STejun Heo * 188263d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1883e22bee78STejun Heo * have at least one idle worker on return from this function. If 1884e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 188563d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1886e22bee78STejun Heo * possible allocation deadlock. 1887e22bee78STejun Heo * 1888c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1889c5aa87bbSTejun Heo * may_start_working() %true. 1890e22bee78STejun Heo * 1891e22bee78STejun Heo * LOCKING: 1892d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1893e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1894e22bee78STejun Heo * manager. 1895e22bee78STejun Heo * 1896d185af30SYacine Belkadi * Return: 1897c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1898e22bee78STejun Heo * otherwise. 1899e22bee78STejun Heo */ 190063d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1901d565ed63STejun Heo __releases(&pool->lock) 1902d565ed63STejun Heo __acquires(&pool->lock) 1903e22bee78STejun Heo { 190463d95a91STejun Heo if (!need_to_create_worker(pool)) 1905e22bee78STejun Heo return false; 1906e22bee78STejun Heo restart: 1907d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19089f9c2364STejun Heo 1909e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 191063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1911e22bee78STejun Heo 1912e22bee78STejun Heo while (true) { 1913e22bee78STejun Heo struct worker *worker; 1914e22bee78STejun Heo 1915bc2ae0f5STejun Heo worker = create_worker(pool); 1916e22bee78STejun Heo if (worker) { 191763d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1918d565ed63STejun Heo spin_lock_irq(&pool->lock); 1919e22bee78STejun Heo start_worker(worker); 19206183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19216183c009STejun Heo goto restart; 1922e22bee78STejun Heo return true; 1923e22bee78STejun Heo } 1924e22bee78STejun Heo 192563d95a91STejun Heo if (!need_to_create_worker(pool)) 1926e22bee78STejun Heo break; 1927e22bee78STejun Heo 1928e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1929e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19309f9c2364STejun Heo 193163d95a91STejun Heo if (!need_to_create_worker(pool)) 1932e22bee78STejun Heo break; 1933e22bee78STejun Heo } 1934e22bee78STejun Heo 193563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1936d565ed63STejun Heo spin_lock_irq(&pool->lock); 193763d95a91STejun Heo if (need_to_create_worker(pool)) 1938e22bee78STejun Heo goto restart; 1939e22bee78STejun Heo return true; 1940e22bee78STejun Heo } 1941e22bee78STejun Heo 1942e22bee78STejun Heo /** 1943e22bee78STejun Heo * manage_workers - manage worker pool 1944e22bee78STejun Heo * @worker: self 1945e22bee78STejun Heo * 1946706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1947e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1948706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1949e22bee78STejun Heo * 1950e22bee78STejun Heo * The caller can safely start processing works on false return. On 1951e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1952e22bee78STejun Heo * and may_start_working() is true. 1953e22bee78STejun Heo * 1954e22bee78STejun Heo * CONTEXT: 1955d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1956e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1957e22bee78STejun Heo * 1958d185af30SYacine Belkadi * Return: 19592d498db9SLibin * %false if the pool don't need management and the caller can safely start 19602d498db9SLibin * processing works, %true indicates that the function released pool->lock 19612d498db9SLibin * and reacquired it to perform some management function and that the 19622d498db9SLibin * conditions that the caller verified while holding the lock before 19632d498db9SLibin * calling the function might no longer be true. 1964e22bee78STejun Heo */ 1965e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1966e22bee78STejun Heo { 196763d95a91STejun Heo struct worker_pool *pool = worker->pool; 1968e22bee78STejun Heo bool ret = false; 1969e22bee78STejun Heo 1970bc3a1afcSTejun Heo /* 1971bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1972bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1973bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1974bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1975bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1976bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1977bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1978bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1979bc3a1afcSTejun Heo */ 198034a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 1981e22bee78STejun Heo return ret; 1982e22bee78STejun Heo 198363d95a91STejun Heo ret |= maybe_create_worker(pool); 1984e22bee78STejun Heo 198534a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 1986e22bee78STejun Heo return ret; 1987e22bee78STejun Heo } 1988e22bee78STejun Heo 1989a62428c0STejun Heo /** 1990a62428c0STejun Heo * process_one_work - process single work 1991c34056a3STejun Heo * @worker: self 1992a62428c0STejun Heo * @work: work to process 1993a62428c0STejun Heo * 1994a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1995a62428c0STejun Heo * process a single work including synchronization against and 1996a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1997a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1998a62428c0STejun Heo * call this function to process a work. 1999a62428c0STejun Heo * 2000a62428c0STejun Heo * CONTEXT: 2001d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2002a62428c0STejun Heo */ 2003c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2004d565ed63STejun Heo __releases(&pool->lock) 2005d565ed63STejun Heo __acquires(&pool->lock) 20061da177e4SLinus Torvalds { 2007112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2008bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2009112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 201073f53c4aSTejun Heo int work_color; 20117e11629dSTejun Heo struct worker *collision; 20124e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20134e6045f1SJohannes Berg /* 2014a62428c0STejun Heo * It is permissible to free the struct work_struct from 2015a62428c0STejun Heo * inside the function that is called from it, this we need to 2016a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2017a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2018a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20194e6045f1SJohannes Berg */ 20204d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20214d82a1deSPeter Zijlstra 20224d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20234e6045f1SJohannes Berg #endif 20246fec10a1STejun Heo /* 20256fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 20266fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 202724647570STejun Heo * unbound or a disassociated pool. 20286fec10a1STejun Heo */ 20295f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 203024647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2031ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 203225511a47STejun Heo 20337e11629dSTejun Heo /* 20347e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20357e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20367e11629dSTejun Heo * already processing the work. If so, defer the work to the 20377e11629dSTejun Heo * currently executing one. 20387e11629dSTejun Heo */ 2039c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20407e11629dSTejun Heo if (unlikely(collision)) { 20417e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20427e11629dSTejun Heo return; 20437e11629dSTejun Heo } 20441da177e4SLinus Torvalds 20458930cabaSTejun Heo /* claim and dequeue */ 20461da177e4SLinus Torvalds debug_work_deactivate(work); 2047c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2048c34056a3STejun Heo worker->current_work = work; 2049a2c1c57bSTejun Heo worker->current_func = work->func; 2050112202d9STejun Heo worker->current_pwq = pwq; 205173f53c4aSTejun Heo work_color = get_work_color(work); 20527a22ad75STejun Heo 2053a62428c0STejun Heo list_del_init(&work->entry); 2054a62428c0STejun Heo 2055649027d7STejun Heo /* 2056fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2057fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2058fb0e7bebSTejun Heo */ 2059fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2060fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2061fb0e7bebSTejun Heo 2062974271c4STejun Heo /* 2063d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2064974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2065974271c4STejun Heo */ 206663d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 206763d95a91STejun Heo wake_up_worker(pool); 2068974271c4STejun Heo 20698930cabaSTejun Heo /* 20707c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2071d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 207223657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 207323657bb1STejun Heo * disabled. 20748930cabaSTejun Heo */ 20757c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20761da177e4SLinus Torvalds 2077d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2078365970a1SDavid Howells 2079112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20803295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2081e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2082a2c1c57bSTejun Heo worker->current_func(work); 2083e36c886aSArjan van de Ven /* 2084e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2085e36c886aSArjan van de Ven * point will only record its address. 2086e36c886aSArjan van de Ven */ 2087e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20883295f0efSIngo Molnar lock_map_release(&lockdep_map); 2089112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 20901da177e4SLinus Torvalds 2091d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2092044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2093044c782cSValentin Ilie " last function: %pf\n", 2094a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2095a2c1c57bSTejun Heo worker->current_func); 2096d5abe669SPeter Zijlstra debug_show_held_locks(current); 2097d5abe669SPeter Zijlstra dump_stack(); 2098d5abe669SPeter Zijlstra } 2099d5abe669SPeter Zijlstra 2100b22ce278STejun Heo /* 2101b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2102b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2103b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2104b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2105b22ce278STejun Heo * stop_machine. 2106b22ce278STejun Heo */ 2107b22ce278STejun Heo cond_resched(); 2108b22ce278STejun Heo 2109d565ed63STejun Heo spin_lock_irq(&pool->lock); 2110a62428c0STejun Heo 2111fb0e7bebSTejun Heo /* clear cpu intensive status */ 2112fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2113fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2114fb0e7bebSTejun Heo 2115a62428c0STejun Heo /* we're done with it, release */ 211642f8570fSSasha Levin hash_del(&worker->hentry); 2117c34056a3STejun Heo worker->current_work = NULL; 2118a2c1c57bSTejun Heo worker->current_func = NULL; 2119112202d9STejun Heo worker->current_pwq = NULL; 21203d1cb205STejun Heo worker->desc_valid = false; 2121112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21221da177e4SLinus Torvalds } 21231da177e4SLinus Torvalds 2124affee4b2STejun Heo /** 2125affee4b2STejun Heo * process_scheduled_works - process scheduled works 2126affee4b2STejun Heo * @worker: self 2127affee4b2STejun Heo * 2128affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2129affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2130affee4b2STejun Heo * fetches a work from the top and executes it. 2131affee4b2STejun Heo * 2132affee4b2STejun Heo * CONTEXT: 2133d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2134affee4b2STejun Heo * multiple times. 2135affee4b2STejun Heo */ 2136affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21371da177e4SLinus Torvalds { 2138affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2139affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2140a62428c0STejun Heo struct work_struct, entry); 2141c34056a3STejun Heo process_one_work(worker, work); 2142a62428c0STejun Heo } 21431da177e4SLinus Torvalds } 21441da177e4SLinus Torvalds 21454690c4abSTejun Heo /** 21464690c4abSTejun Heo * worker_thread - the worker thread function 2147c34056a3STejun Heo * @__worker: self 21484690c4abSTejun Heo * 2149c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2150c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2151c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2152c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2153c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2154d185af30SYacine Belkadi * 2155d185af30SYacine Belkadi * Return: 0 21564690c4abSTejun Heo */ 2157c34056a3STejun Heo static int worker_thread(void *__worker) 21581da177e4SLinus Torvalds { 2159c34056a3STejun Heo struct worker *worker = __worker; 2160bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21611da177e4SLinus Torvalds 2162e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2163e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2164c8e55f36STejun Heo woke_up: 2165d565ed63STejun Heo spin_lock_irq(&pool->lock); 2166affee4b2STejun Heo 2167a9ab775bSTejun Heo /* am I supposed to die? */ 2168a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2169d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2170a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2171e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 217260f5a4bcSLai Jiangshan 217360f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21747cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 217560f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 217660f5a4bcSLai Jiangshan kfree(worker); 2177c8e55f36STejun Heo return 0; 2178c8e55f36STejun Heo } 2179c8e55f36STejun Heo 2180c8e55f36STejun Heo worker_leave_idle(worker); 2181db7bccf4STejun Heo recheck: 2182e22bee78STejun Heo /* no more worker necessary? */ 218363d95a91STejun Heo if (!need_more_worker(pool)) 2184e22bee78STejun Heo goto sleep; 2185e22bee78STejun Heo 2186e22bee78STejun Heo /* do we need to manage? */ 218763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2188e22bee78STejun Heo goto recheck; 2189e22bee78STejun Heo 2190c8e55f36STejun Heo /* 2191c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2192c8e55f36STejun Heo * preparing to process a work or actually processing it. 2193c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2194c8e55f36STejun Heo */ 21956183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2196c8e55f36STejun Heo 2197e22bee78STejun Heo /* 2198a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2199a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2200a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2201a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2202a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2203e22bee78STejun Heo */ 2204a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2205e22bee78STejun Heo 2206e22bee78STejun Heo do { 2207affee4b2STejun Heo struct work_struct *work = 2208bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2209affee4b2STejun Heo struct work_struct, entry); 2210affee4b2STejun Heo 2211c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2212affee4b2STejun Heo /* optimization path, not strictly necessary */ 2213affee4b2STejun Heo process_one_work(worker, work); 2214affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2215affee4b2STejun Heo process_scheduled_works(worker); 2216affee4b2STejun Heo } else { 2217c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2218affee4b2STejun Heo process_scheduled_works(worker); 2219affee4b2STejun Heo } 222063d95a91STejun Heo } while (keep_working(pool)); 2221affee4b2STejun Heo 2222e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2223d313dd85STejun Heo sleep: 2224c8e55f36STejun Heo /* 2225d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2226d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2227d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2228d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2229d565ed63STejun Heo * event. 2230c8e55f36STejun Heo */ 2231c8e55f36STejun Heo worker_enter_idle(worker); 2232c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2233d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22341da177e4SLinus Torvalds schedule(); 2235c8e55f36STejun Heo goto woke_up; 22361da177e4SLinus Torvalds } 22371da177e4SLinus Torvalds 2238e22bee78STejun Heo /** 2239e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2240111c225aSTejun Heo * @__rescuer: self 2241e22bee78STejun Heo * 2242e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2243493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2244e22bee78STejun Heo * 2245706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2246e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2247e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2248e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2249e22bee78STejun Heo * the problem rescuer solves. 2250e22bee78STejun Heo * 2251706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2252706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2253e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2254e22bee78STejun Heo * 2255e22bee78STejun Heo * This should happen rarely. 2256d185af30SYacine Belkadi * 2257d185af30SYacine Belkadi * Return: 0 2258e22bee78STejun Heo */ 2259111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2260e22bee78STejun Heo { 2261111c225aSTejun Heo struct worker *rescuer = __rescuer; 2262111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2263e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22644d595b86SLai Jiangshan bool should_stop; 2265e22bee78STejun Heo 2266e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2267111c225aSTejun Heo 2268111c225aSTejun Heo /* 2269111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2270111c225aSTejun Heo * doesn't participate in concurrency management. 2271111c225aSTejun Heo */ 2272111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2273e22bee78STejun Heo repeat: 2274e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22751da177e4SLinus Torvalds 22764d595b86SLai Jiangshan /* 22774d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22784d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22794d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22804d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22814d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22824d595b86SLai Jiangshan * list is always empty on exit. 22834d595b86SLai Jiangshan */ 22844d595b86SLai Jiangshan should_stop = kthread_should_stop(); 22851da177e4SLinus Torvalds 2286493a1724STejun Heo /* see whether any pwq is asking for help */ 22872e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2288493a1724STejun Heo 2289493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2290493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2291493a1724STejun Heo struct pool_workqueue, mayday_node); 2292112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2293e22bee78STejun Heo struct work_struct *work, *n; 2294e22bee78STejun Heo 2295e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2296493a1724STejun Heo list_del_init(&pwq->mayday_node); 2297493a1724STejun Heo 22982e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2299e22bee78STejun Heo 230051697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 230151697d39SLai Jiangshan 230251697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2303b3104104SLai Jiangshan rescuer->pool = pool; 2304e22bee78STejun Heo 2305e22bee78STejun Heo /* 2306e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2307e22bee78STejun Heo * process'em. 2308e22bee78STejun Heo */ 23096183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2310bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2311112202d9STejun Heo if (get_work_pwq(work) == pwq) 2312e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2313e22bee78STejun Heo 2314e22bee78STejun Heo process_scheduled_works(rescuer); 231551697d39SLai Jiangshan spin_unlock_irq(&pool->lock); 231651697d39SLai Jiangshan 231751697d39SLai Jiangshan worker_detach_from_pool(rescuer, pool); 231851697d39SLai Jiangshan 231951697d39SLai Jiangshan spin_lock_irq(&pool->lock); 23207576958aSTejun Heo 23217576958aSTejun Heo /* 232277668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 232377668c8bSLai Jiangshan * go away while we're holding its lock. 232477668c8bSLai Jiangshan */ 232577668c8bSLai Jiangshan put_pwq(pwq); 232677668c8bSLai Jiangshan 232777668c8bSLai Jiangshan /* 2328d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23297576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23307576958aSTejun Heo * and stalling the execution. 23317576958aSTejun Heo */ 233263d95a91STejun Heo if (keep_working(pool)) 233363d95a91STejun Heo wake_up_worker(pool); 23347576958aSTejun Heo 2335b3104104SLai Jiangshan rescuer->pool = NULL; 2336493a1724STejun Heo spin_unlock(&pool->lock); 23372e109a28STejun Heo spin_lock(&wq_mayday_lock); 23381da177e4SLinus Torvalds } 23391da177e4SLinus Torvalds 23402e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2341493a1724STejun Heo 23424d595b86SLai Jiangshan if (should_stop) { 23434d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23444d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23454d595b86SLai Jiangshan return 0; 23464d595b86SLai Jiangshan } 23474d595b86SLai Jiangshan 2348111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2349111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2350e22bee78STejun Heo schedule(); 2351e22bee78STejun Heo goto repeat; 23521da177e4SLinus Torvalds } 23531da177e4SLinus Torvalds 2354fc2e4d70SOleg Nesterov struct wq_barrier { 2355fc2e4d70SOleg Nesterov struct work_struct work; 2356fc2e4d70SOleg Nesterov struct completion done; 2357fc2e4d70SOleg Nesterov }; 2358fc2e4d70SOleg Nesterov 2359fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2360fc2e4d70SOleg Nesterov { 2361fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2362fc2e4d70SOleg Nesterov complete(&barr->done); 2363fc2e4d70SOleg Nesterov } 2364fc2e4d70SOleg Nesterov 23654690c4abSTejun Heo /** 23664690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2367112202d9STejun Heo * @pwq: pwq to insert barrier into 23684690c4abSTejun Heo * @barr: wq_barrier to insert 2369affee4b2STejun Heo * @target: target work to attach @barr to 2370affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23714690c4abSTejun Heo * 2372affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2373affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2374affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2375affee4b2STejun Heo * cpu. 2376affee4b2STejun Heo * 2377affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2378affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2379affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2380affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2381affee4b2STejun Heo * after a work with LINKED flag set. 2382affee4b2STejun Heo * 2383affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2384112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23854690c4abSTejun Heo * 23864690c4abSTejun Heo * CONTEXT: 2387d565ed63STejun Heo * spin_lock_irq(pool->lock). 23884690c4abSTejun Heo */ 2389112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2390affee4b2STejun Heo struct wq_barrier *barr, 2391affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2392fc2e4d70SOleg Nesterov { 2393affee4b2STejun Heo struct list_head *head; 2394affee4b2STejun Heo unsigned int linked = 0; 2395affee4b2STejun Heo 2396dc186ad7SThomas Gleixner /* 2397d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2398dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2399dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2400dc186ad7SThomas Gleixner * might deadlock. 2401dc186ad7SThomas Gleixner */ 2402ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 240322df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2404fc2e4d70SOleg Nesterov init_completion(&barr->done); 240583c22520SOleg Nesterov 2406affee4b2STejun Heo /* 2407affee4b2STejun Heo * If @target is currently being executed, schedule the 2408affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2409affee4b2STejun Heo */ 2410affee4b2STejun Heo if (worker) 2411affee4b2STejun Heo head = worker->scheduled.next; 2412affee4b2STejun Heo else { 2413affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2414affee4b2STejun Heo 2415affee4b2STejun Heo head = target->entry.next; 2416affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2417affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2418affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2419affee4b2STejun Heo } 2420affee4b2STejun Heo 2421dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2422112202d9STejun Heo insert_work(pwq, &barr->work, head, 2423affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2424fc2e4d70SOleg Nesterov } 2425fc2e4d70SOleg Nesterov 242673f53c4aSTejun Heo /** 2427112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 242873f53c4aSTejun Heo * @wq: workqueue being flushed 242973f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 243073f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 243173f53c4aSTejun Heo * 2432112202d9STejun Heo * Prepare pwqs for workqueue flushing. 243373f53c4aSTejun Heo * 2434112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2435112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2436112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2437112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2438112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 243973f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 244073f53c4aSTejun Heo * 244173f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 244273f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 244373f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 244473f53c4aSTejun Heo * is returned. 244573f53c4aSTejun Heo * 2446112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 244773f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 244873f53c4aSTejun Heo * advanced to @work_color. 244973f53c4aSTejun Heo * 245073f53c4aSTejun Heo * CONTEXT: 24513c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 245273f53c4aSTejun Heo * 2453d185af30SYacine Belkadi * Return: 245473f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 245573f53c4aSTejun Heo * otherwise. 245673f53c4aSTejun Heo */ 2457112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 245873f53c4aSTejun Heo int flush_color, int work_color) 24591da177e4SLinus Torvalds { 246073f53c4aSTejun Heo bool wait = false; 246149e3cf44STejun Heo struct pool_workqueue *pwq; 24621da177e4SLinus Torvalds 246373f53c4aSTejun Heo if (flush_color >= 0) { 24646183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2465112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2466dc186ad7SThomas Gleixner } 246714441960SOleg Nesterov 246849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2469112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24701da177e4SLinus Torvalds 2471b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 247273f53c4aSTejun Heo 247373f53c4aSTejun Heo if (flush_color >= 0) { 24746183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 247573f53c4aSTejun Heo 2476112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2477112202d9STejun Heo pwq->flush_color = flush_color; 2478112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 247973f53c4aSTejun Heo wait = true; 24801da177e4SLinus Torvalds } 248173f53c4aSTejun Heo } 248273f53c4aSTejun Heo 248373f53c4aSTejun Heo if (work_color >= 0) { 24846183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2485112202d9STejun Heo pwq->work_color = work_color; 248673f53c4aSTejun Heo } 248773f53c4aSTejun Heo 2488b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 24891da177e4SLinus Torvalds } 24901da177e4SLinus Torvalds 2491112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 249273f53c4aSTejun Heo complete(&wq->first_flusher->done); 249373f53c4aSTejun Heo 249473f53c4aSTejun Heo return wait; 249583c22520SOleg Nesterov } 24961da177e4SLinus Torvalds 24970fcb78c2SRolf Eike Beer /** 24981da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 24990fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25001da177e4SLinus Torvalds * 2501c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2502c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25031da177e4SLinus Torvalds */ 25047ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25051da177e4SLinus Torvalds { 250673f53c4aSTejun Heo struct wq_flusher this_flusher = { 250773f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 250873f53c4aSTejun Heo .flush_color = -1, 250973f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 251073f53c4aSTejun Heo }; 251173f53c4aSTejun Heo int next_color; 2512b1f4ec17SOleg Nesterov 25133295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25143295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 251573f53c4aSTejun Heo 25163c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 251773f53c4aSTejun Heo 251873f53c4aSTejun Heo /* 251973f53c4aSTejun Heo * Start-to-wait phase 252073f53c4aSTejun Heo */ 252173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 252273f53c4aSTejun Heo 252373f53c4aSTejun Heo if (next_color != wq->flush_color) { 252473f53c4aSTejun Heo /* 252573f53c4aSTejun Heo * Color space is not full. The current work_color 252673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 252773f53c4aSTejun Heo * by one. 252873f53c4aSTejun Heo */ 25296183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 253073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 253173f53c4aSTejun Heo wq->work_color = next_color; 253273f53c4aSTejun Heo 253373f53c4aSTejun Heo if (!wq->first_flusher) { 253473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25356183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 253673f53c4aSTejun Heo 253773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 253873f53c4aSTejun Heo 2539112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 254073f53c4aSTejun Heo wq->work_color)) { 254173f53c4aSTejun Heo /* nothing to flush, done */ 254273f53c4aSTejun Heo wq->flush_color = next_color; 254373f53c4aSTejun Heo wq->first_flusher = NULL; 254473f53c4aSTejun Heo goto out_unlock; 254573f53c4aSTejun Heo } 254673f53c4aSTejun Heo } else { 254773f53c4aSTejun Heo /* wait in queue */ 25486183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 254973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2550112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 255173f53c4aSTejun Heo } 255273f53c4aSTejun Heo } else { 255373f53c4aSTejun Heo /* 255473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 255573f53c4aSTejun Heo * The next flush completion will assign us 255673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 255773f53c4aSTejun Heo */ 255873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 255973f53c4aSTejun Heo } 256073f53c4aSTejun Heo 25613c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 256273f53c4aSTejun Heo 256373f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 256473f53c4aSTejun Heo 256573f53c4aSTejun Heo /* 256673f53c4aSTejun Heo * Wake-up-and-cascade phase 256773f53c4aSTejun Heo * 256873f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 256973f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 257073f53c4aSTejun Heo */ 257173f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 257273f53c4aSTejun Heo return; 257373f53c4aSTejun Heo 25743c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 257573f53c4aSTejun Heo 25764ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25774ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25784ce48b37STejun Heo goto out_unlock; 25794ce48b37STejun Heo 258073f53c4aSTejun Heo wq->first_flusher = NULL; 258173f53c4aSTejun Heo 25826183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25836183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 258473f53c4aSTejun Heo 258573f53c4aSTejun Heo while (true) { 258673f53c4aSTejun Heo struct wq_flusher *next, *tmp; 258773f53c4aSTejun Heo 258873f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 258973f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 259073f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 259173f53c4aSTejun Heo break; 259273f53c4aSTejun Heo list_del_init(&next->list); 259373f53c4aSTejun Heo complete(&next->done); 259473f53c4aSTejun Heo } 259573f53c4aSTejun Heo 25966183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 259773f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 259873f53c4aSTejun Heo 259973f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 260073f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 260173f53c4aSTejun Heo 260273f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 260373f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 260473f53c4aSTejun Heo /* 260573f53c4aSTejun Heo * Assign the same color to all overflowed 260673f53c4aSTejun Heo * flushers, advance work_color and append to 260773f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 260873f53c4aSTejun Heo * phase for these overflowed flushers. 260973f53c4aSTejun Heo */ 261073f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 261173f53c4aSTejun Heo tmp->flush_color = wq->work_color; 261273f53c4aSTejun Heo 261373f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 261473f53c4aSTejun Heo 261573f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 261673f53c4aSTejun Heo &wq->flusher_queue); 2617112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 261873f53c4aSTejun Heo } 261973f53c4aSTejun Heo 262073f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26216183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 262273f53c4aSTejun Heo break; 262373f53c4aSTejun Heo } 262473f53c4aSTejun Heo 262573f53c4aSTejun Heo /* 262673f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2627112202d9STejun Heo * the new first flusher and arm pwqs. 262873f53c4aSTejun Heo */ 26296183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26306183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 263173f53c4aSTejun Heo 263273f53c4aSTejun Heo list_del_init(&next->list); 263373f53c4aSTejun Heo wq->first_flusher = next; 263473f53c4aSTejun Heo 2635112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 263673f53c4aSTejun Heo break; 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo /* 263973f53c4aSTejun Heo * Meh... this color is already done, clear first 264073f53c4aSTejun Heo * flusher and repeat cascading. 264173f53c4aSTejun Heo */ 264273f53c4aSTejun Heo wq->first_flusher = NULL; 264373f53c4aSTejun Heo } 264473f53c4aSTejun Heo 264573f53c4aSTejun Heo out_unlock: 26463c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26471da177e4SLinus Torvalds } 2648ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26491da177e4SLinus Torvalds 26509c5a2ba7STejun Heo /** 26519c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26529c5a2ba7STejun Heo * @wq: workqueue to drain 26539c5a2ba7STejun Heo * 26549c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26559c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26569c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26579c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26589c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26599c5a2ba7STejun Heo * takes too long. 26609c5a2ba7STejun Heo */ 26619c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26629c5a2ba7STejun Heo { 26639c5a2ba7STejun Heo unsigned int flush_cnt = 0; 266449e3cf44STejun Heo struct pool_workqueue *pwq; 26659c5a2ba7STejun Heo 26669c5a2ba7STejun Heo /* 26679c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26689c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2669618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26709c5a2ba7STejun Heo */ 267187fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26729c5a2ba7STejun Heo if (!wq->nr_drainers++) 2673618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 267487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26759c5a2ba7STejun Heo reflush: 26769c5a2ba7STejun Heo flush_workqueue(wq); 26779c5a2ba7STejun Heo 2678b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 267976af4d93STejun Heo 268049e3cf44STejun Heo for_each_pwq(pwq, wq) { 2681fa2563e4SThomas Tuttle bool drained; 26829c5a2ba7STejun Heo 2683b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2684112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2685b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2686fa2563e4SThomas Tuttle 2687fa2563e4SThomas Tuttle if (drained) 26889c5a2ba7STejun Heo continue; 26899c5a2ba7STejun Heo 26909c5a2ba7STejun Heo if (++flush_cnt == 10 || 26919c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2692c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 26939c5a2ba7STejun Heo wq->name, flush_cnt); 269476af4d93STejun Heo 2695b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 26969c5a2ba7STejun Heo goto reflush; 26979c5a2ba7STejun Heo } 26989c5a2ba7STejun Heo 26999c5a2ba7STejun Heo if (!--wq->nr_drainers) 2700618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 270187fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27029c5a2ba7STejun Heo } 27039c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27049c5a2ba7STejun Heo 2705606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2706baf59022STejun Heo { 2707baf59022STejun Heo struct worker *worker = NULL; 2708c9e7cf27STejun Heo struct worker_pool *pool; 2709112202d9STejun Heo struct pool_workqueue *pwq; 2710baf59022STejun Heo 2711baf59022STejun Heo might_sleep(); 2712baf59022STejun Heo 2713fa1b54e6STejun Heo local_irq_disable(); 2714fa1b54e6STejun Heo pool = get_work_pool(work); 2715fa1b54e6STejun Heo if (!pool) { 2716fa1b54e6STejun Heo local_irq_enable(); 2717fa1b54e6STejun Heo return false; 2718fa1b54e6STejun Heo } 2719fa1b54e6STejun Heo 2720fa1b54e6STejun Heo spin_lock(&pool->lock); 27210b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2722112202d9STejun Heo pwq = get_work_pwq(work); 2723112202d9STejun Heo if (pwq) { 2724112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2725baf59022STejun Heo goto already_gone; 2726606a5020STejun Heo } else { 2727c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2728baf59022STejun Heo if (!worker) 2729baf59022STejun Heo goto already_gone; 2730112202d9STejun Heo pwq = worker->current_pwq; 2731606a5020STejun Heo } 2732baf59022STejun Heo 2733112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2734d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2735baf59022STejun Heo 2736e159489bSTejun Heo /* 2737e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2738e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2739e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2740e159489bSTejun Heo * access. 2741e159489bSTejun Heo */ 2742493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2743112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2744e159489bSTejun Heo else 2745112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2746112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2747e159489bSTejun Heo 2748baf59022STejun Heo return true; 2749baf59022STejun Heo already_gone: 2750d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2751baf59022STejun Heo return false; 2752baf59022STejun Heo } 2753baf59022STejun Heo 2754db700897SOleg Nesterov /** 2755401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2756401a8d04STejun Heo * @work: the work to flush 2757db700897SOleg Nesterov * 2758606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2759606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2760401a8d04STejun Heo * 2761d185af30SYacine Belkadi * Return: 2762401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2763401a8d04STejun Heo * %false if it was already idle. 2764db700897SOleg Nesterov */ 2765401a8d04STejun Heo bool flush_work(struct work_struct *work) 2766db700897SOleg Nesterov { 276712997d1aSBjorn Helgaas struct wq_barrier barr; 276812997d1aSBjorn Helgaas 27690976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27700976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27710976dfc1SStephen Boyd 277212997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 277312997d1aSBjorn Helgaas wait_for_completion(&barr.done); 277412997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 277512997d1aSBjorn Helgaas return true; 277612997d1aSBjorn Helgaas } else { 277712997d1aSBjorn Helgaas return false; 277812997d1aSBjorn Helgaas } 2779606a5020STejun Heo } 2780db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2781db700897SOleg Nesterov 278236e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2783401a8d04STejun Heo { 2784bbb68dfaSTejun Heo unsigned long flags; 27851f1f642eSOleg Nesterov int ret; 27861f1f642eSOleg Nesterov 27871f1f642eSOleg Nesterov do { 2788bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2789bbb68dfaSTejun Heo /* 2790bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2791bbb68dfaSTejun Heo * would be waiting for before retrying. 2792bbb68dfaSTejun Heo */ 2793bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2794606a5020STejun Heo flush_work(work); 27951f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 27961f1f642eSOleg Nesterov 2797bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2798bbb68dfaSTejun Heo mark_work_canceling(work); 2799bbb68dfaSTejun Heo local_irq_restore(flags); 2800bbb68dfaSTejun Heo 2801606a5020STejun Heo flush_work(work); 28027a22ad75STejun Heo clear_work_data(work); 28031f1f642eSOleg Nesterov return ret; 28041f1f642eSOleg Nesterov } 28051f1f642eSOleg Nesterov 28066e84d644SOleg Nesterov /** 2807401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2808401a8d04STejun Heo * @work: the work to cancel 28096e84d644SOleg Nesterov * 2810401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2811401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2812401a8d04STejun Heo * another workqueue. On return from this function, @work is 2813401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28141f1f642eSOleg Nesterov * 2815401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2816401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28176e84d644SOleg Nesterov * 2818401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28196e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2820401a8d04STejun Heo * 2821d185af30SYacine Belkadi * Return: 2822401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28236e84d644SOleg Nesterov */ 2824401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28256e84d644SOleg Nesterov { 282636e227d2STejun Heo return __cancel_work_timer(work, false); 2827b89deed3SOleg Nesterov } 282828e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2829b89deed3SOleg Nesterov 28306e84d644SOleg Nesterov /** 2831401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2832401a8d04STejun Heo * @dwork: the delayed work to flush 28336e84d644SOleg Nesterov * 2834401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2835401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2836401a8d04STejun Heo * considers the last queueing instance of @dwork. 28371f1f642eSOleg Nesterov * 2838d185af30SYacine Belkadi * Return: 2839401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2840401a8d04STejun Heo * %false if it was already idle. 28416e84d644SOleg Nesterov */ 2842401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2843401a8d04STejun Heo { 28448930cabaSTejun Heo local_irq_disable(); 2845401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 284660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28478930cabaSTejun Heo local_irq_enable(); 2848401a8d04STejun Heo return flush_work(&dwork->work); 2849401a8d04STejun Heo } 2850401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2851401a8d04STejun Heo 2852401a8d04STejun Heo /** 285357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 285457b30ae7STejun Heo * @dwork: delayed_work to cancel 285509383498STejun Heo * 2856d185af30SYacine Belkadi * Kill off a pending delayed_work. 2857d185af30SYacine Belkadi * 2858d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2859d185af30SYacine Belkadi * pending. 2860d185af30SYacine Belkadi * 2861d185af30SYacine Belkadi * Note: 2862d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2863d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2864d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 286509383498STejun Heo * 286657b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 286709383498STejun Heo */ 286857b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 286909383498STejun Heo { 287057b30ae7STejun Heo unsigned long flags; 287157b30ae7STejun Heo int ret; 287257b30ae7STejun Heo 287357b30ae7STejun Heo do { 287457b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 287557b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 287657b30ae7STejun Heo 287757b30ae7STejun Heo if (unlikely(ret < 0)) 287857b30ae7STejun Heo return false; 287957b30ae7STejun Heo 28807c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28817c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 288257b30ae7STejun Heo local_irq_restore(flags); 2883c0158ca6SDan Magenheimer return ret; 288409383498STejun Heo } 288557b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 288609383498STejun Heo 288709383498STejun Heo /** 2888401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2889401a8d04STejun Heo * @dwork: the delayed work cancel 2890401a8d04STejun Heo * 2891401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2892401a8d04STejun Heo * 2893d185af30SYacine Belkadi * Return: 2894401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2895401a8d04STejun Heo */ 2896401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 28976e84d644SOleg Nesterov { 289836e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 28996e84d644SOleg Nesterov } 2900f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29011da177e4SLinus Torvalds 29020fcb78c2SRolf Eike Beer /** 290331ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2904b6136773SAndrew Morton * @func: the function to call 2905b6136773SAndrew Morton * 290631ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 290731ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2908b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 290931ddd871STejun Heo * 2910d185af30SYacine Belkadi * Return: 291131ddd871STejun Heo * 0 on success, -errno on failure. 2912b6136773SAndrew Morton */ 291365f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 291415316ba8SChristoph Lameter { 291515316ba8SChristoph Lameter int cpu; 291638f51568SNamhyung Kim struct work_struct __percpu *works; 291715316ba8SChristoph Lameter 2918b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2919b6136773SAndrew Morton if (!works) 292015316ba8SChristoph Lameter return -ENOMEM; 2921b6136773SAndrew Morton 292295402b38SGautham R Shenoy get_online_cpus(); 292393981800STejun Heo 292415316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29259bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29269bfb1839SIngo Molnar 29279bfb1839SIngo Molnar INIT_WORK(work, func); 29288de6d308SOleg Nesterov schedule_work_on(cpu, work); 292915316ba8SChristoph Lameter } 293093981800STejun Heo 293193981800STejun Heo for_each_online_cpu(cpu) 29328616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 293393981800STejun Heo 293495402b38SGautham R Shenoy put_online_cpus(); 2935b6136773SAndrew Morton free_percpu(works); 293615316ba8SChristoph Lameter return 0; 293715316ba8SChristoph Lameter } 293815316ba8SChristoph Lameter 2939eef6a7d5SAlan Stern /** 2940eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2941eef6a7d5SAlan Stern * 2942eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2943eef6a7d5SAlan Stern * completion. 2944eef6a7d5SAlan Stern * 2945eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2946eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2947eef6a7d5SAlan Stern * will lead to deadlock: 2948eef6a7d5SAlan Stern * 2949eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2950eef6a7d5SAlan Stern * a lock held by your code or its caller. 2951eef6a7d5SAlan Stern * 2952eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2953eef6a7d5SAlan Stern * 2954eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2955eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2956eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2957eef6a7d5SAlan Stern * 2958eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2959eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2960eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2961eef6a7d5SAlan Stern * cancel_work_sync() instead. 2962eef6a7d5SAlan Stern */ 29631da177e4SLinus Torvalds void flush_scheduled_work(void) 29641da177e4SLinus Torvalds { 2965d320c038STejun Heo flush_workqueue(system_wq); 29661da177e4SLinus Torvalds } 2967ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29681da177e4SLinus Torvalds 29691da177e4SLinus Torvalds /** 29701fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29711fa44ecaSJames Bottomley * @fn: the function to execute 29721fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29731fa44ecaSJames Bottomley * be available when the work executes) 29741fa44ecaSJames Bottomley * 29751fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29761fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29771fa44ecaSJames Bottomley * 2978d185af30SYacine Belkadi * Return: 0 - function was executed 29791fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29801fa44ecaSJames Bottomley */ 298165f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29821fa44ecaSJames Bottomley { 29831fa44ecaSJames Bottomley if (!in_interrupt()) { 298465f27f38SDavid Howells fn(&ew->work); 29851fa44ecaSJames Bottomley return 0; 29861fa44ecaSJames Bottomley } 29871fa44ecaSJames Bottomley 298865f27f38SDavid Howells INIT_WORK(&ew->work, fn); 29891fa44ecaSJames Bottomley schedule_work(&ew->work); 29901fa44ecaSJames Bottomley 29911fa44ecaSJames Bottomley return 1; 29921fa44ecaSJames Bottomley } 29931fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 29941fa44ecaSJames Bottomley 2995226223abSTejun Heo #ifdef CONFIG_SYSFS 2996226223abSTejun Heo /* 2997226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 2998226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 2999226223abSTejun Heo * following attributes. 3000226223abSTejun Heo * 3001226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3002226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3003226223abSTejun Heo * 3004226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3005226223abSTejun Heo * 3006226223abSTejun Heo * id RO int : the associated pool ID 3007226223abSTejun Heo * nice RW int : nice value of the workers 3008226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3009226223abSTejun Heo */ 3010226223abSTejun Heo struct wq_device { 3011226223abSTejun Heo struct workqueue_struct *wq; 3012226223abSTejun Heo struct device dev; 3013226223abSTejun Heo }; 3014226223abSTejun Heo 3015226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3016226223abSTejun Heo { 3017226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3018226223abSTejun Heo 3019226223abSTejun Heo return wq_dev->wq; 3020226223abSTejun Heo } 3021226223abSTejun Heo 30221a6661daSGreg Kroah-Hartman static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 30231a6661daSGreg Kroah-Hartman char *buf) 3024226223abSTejun Heo { 3025226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3026226223abSTejun Heo 3027226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3028226223abSTejun Heo } 30291a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RO(per_cpu); 3030226223abSTejun Heo 30311a6661daSGreg Kroah-Hartman static ssize_t max_active_show(struct device *dev, 3032226223abSTejun Heo struct device_attribute *attr, char *buf) 3033226223abSTejun Heo { 3034226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3035226223abSTejun Heo 3036226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3037226223abSTejun Heo } 3038226223abSTejun Heo 30391a6661daSGreg Kroah-Hartman static ssize_t max_active_store(struct device *dev, 30401a6661daSGreg Kroah-Hartman struct device_attribute *attr, const char *buf, 30411a6661daSGreg Kroah-Hartman size_t count) 3042226223abSTejun Heo { 3043226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3044226223abSTejun Heo int val; 3045226223abSTejun Heo 3046226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3047226223abSTejun Heo return -EINVAL; 3048226223abSTejun Heo 3049226223abSTejun Heo workqueue_set_max_active(wq, val); 3050226223abSTejun Heo return count; 3051226223abSTejun Heo } 30521a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RW(max_active); 3053226223abSTejun Heo 30541a6661daSGreg Kroah-Hartman static struct attribute *wq_sysfs_attrs[] = { 30551a6661daSGreg Kroah-Hartman &dev_attr_per_cpu.attr, 30561a6661daSGreg Kroah-Hartman &dev_attr_max_active.attr, 30571a6661daSGreg Kroah-Hartman NULL, 3058226223abSTejun Heo }; 30591a6661daSGreg Kroah-Hartman ATTRIBUTE_GROUPS(wq_sysfs); 3060226223abSTejun Heo 3061d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3062226223abSTejun Heo struct device_attribute *attr, char *buf) 3063226223abSTejun Heo { 3064226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3065d55262c4STejun Heo const char *delim = ""; 3066d55262c4STejun Heo int node, written = 0; 3067226223abSTejun Heo 3068226223abSTejun Heo rcu_read_lock_sched(); 3069d55262c4STejun Heo for_each_node(node) { 3070d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3071d55262c4STejun Heo "%s%d:%d", delim, node, 3072d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3073d55262c4STejun Heo delim = " "; 3074d55262c4STejun Heo } 3075d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3076226223abSTejun Heo rcu_read_unlock_sched(); 3077226223abSTejun Heo 3078226223abSTejun Heo return written; 3079226223abSTejun Heo } 3080226223abSTejun Heo 3081226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3082226223abSTejun Heo char *buf) 3083226223abSTejun Heo { 3084226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3085226223abSTejun Heo int written; 3086226223abSTejun Heo 30876029a918STejun Heo mutex_lock(&wq->mutex); 30886029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 30896029a918STejun Heo mutex_unlock(&wq->mutex); 3090226223abSTejun Heo 3091226223abSTejun Heo return written; 3092226223abSTejun Heo } 3093226223abSTejun Heo 3094226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3095226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3096226223abSTejun Heo { 3097226223abSTejun Heo struct workqueue_attrs *attrs; 3098226223abSTejun Heo 3099226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3100226223abSTejun Heo if (!attrs) 3101226223abSTejun Heo return NULL; 3102226223abSTejun Heo 31036029a918STejun Heo mutex_lock(&wq->mutex); 31046029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 31056029a918STejun Heo mutex_unlock(&wq->mutex); 3106226223abSTejun Heo return attrs; 3107226223abSTejun Heo } 3108226223abSTejun Heo 3109226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3110226223abSTejun Heo const char *buf, size_t count) 3111226223abSTejun Heo { 3112226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3113226223abSTejun Heo struct workqueue_attrs *attrs; 3114226223abSTejun Heo int ret; 3115226223abSTejun Heo 3116226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3117226223abSTejun Heo if (!attrs) 3118226223abSTejun Heo return -ENOMEM; 3119226223abSTejun Heo 3120226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 312114481842SDongsheng Yang attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 3122226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3123226223abSTejun Heo else 3124226223abSTejun Heo ret = -EINVAL; 3125226223abSTejun Heo 3126226223abSTejun Heo free_workqueue_attrs(attrs); 3127226223abSTejun Heo return ret ?: count; 3128226223abSTejun Heo } 3129226223abSTejun Heo 3130226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3131226223abSTejun Heo struct device_attribute *attr, char *buf) 3132226223abSTejun Heo { 3133226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3134226223abSTejun Heo int written; 3135226223abSTejun Heo 31366029a918STejun Heo mutex_lock(&wq->mutex); 31376029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 31386029a918STejun Heo mutex_unlock(&wq->mutex); 3139226223abSTejun Heo 3140226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3141226223abSTejun Heo return written; 3142226223abSTejun Heo } 3143226223abSTejun Heo 3144226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3145226223abSTejun Heo struct device_attribute *attr, 3146226223abSTejun Heo const char *buf, size_t count) 3147226223abSTejun Heo { 3148226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3149226223abSTejun Heo struct workqueue_attrs *attrs; 3150226223abSTejun Heo int ret; 3151226223abSTejun Heo 3152226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3153226223abSTejun Heo if (!attrs) 3154226223abSTejun Heo return -ENOMEM; 3155226223abSTejun Heo 3156226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3157226223abSTejun Heo if (!ret) 3158226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3159226223abSTejun Heo 3160226223abSTejun Heo free_workqueue_attrs(attrs); 3161226223abSTejun Heo return ret ?: count; 3162226223abSTejun Heo } 3163226223abSTejun Heo 3164d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3165d55262c4STejun Heo char *buf) 3166d55262c4STejun Heo { 3167d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3168d55262c4STejun Heo int written; 3169d55262c4STejun Heo 3170d55262c4STejun Heo mutex_lock(&wq->mutex); 3171d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3172d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3173d55262c4STejun Heo mutex_unlock(&wq->mutex); 3174d55262c4STejun Heo 3175d55262c4STejun Heo return written; 3176d55262c4STejun Heo } 3177d55262c4STejun Heo 3178d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3179d55262c4STejun Heo const char *buf, size_t count) 3180d55262c4STejun Heo { 3181d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3182d55262c4STejun Heo struct workqueue_attrs *attrs; 3183d55262c4STejun Heo int v, ret; 3184d55262c4STejun Heo 3185d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3186d55262c4STejun Heo if (!attrs) 3187d55262c4STejun Heo return -ENOMEM; 3188d55262c4STejun Heo 3189d55262c4STejun Heo ret = -EINVAL; 3190d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3191d55262c4STejun Heo attrs->no_numa = !v; 3192d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3193d55262c4STejun Heo } 3194d55262c4STejun Heo 3195d55262c4STejun Heo free_workqueue_attrs(attrs); 3196d55262c4STejun Heo return ret ?: count; 3197d55262c4STejun Heo } 3198d55262c4STejun Heo 3199226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3200d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3201226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3202226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3203d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3204226223abSTejun Heo __ATTR_NULL, 3205226223abSTejun Heo }; 3206226223abSTejun Heo 3207226223abSTejun Heo static struct bus_type wq_subsys = { 3208226223abSTejun Heo .name = "workqueue", 32091a6661daSGreg Kroah-Hartman .dev_groups = wq_sysfs_groups, 3210226223abSTejun Heo }; 3211226223abSTejun Heo 3212226223abSTejun Heo static int __init wq_sysfs_init(void) 3213226223abSTejun Heo { 3214226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3215226223abSTejun Heo } 3216226223abSTejun Heo core_initcall(wq_sysfs_init); 3217226223abSTejun Heo 3218226223abSTejun Heo static void wq_device_release(struct device *dev) 3219226223abSTejun Heo { 3220226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3221226223abSTejun Heo 3222226223abSTejun Heo kfree(wq_dev); 3223226223abSTejun Heo } 3224226223abSTejun Heo 3225226223abSTejun Heo /** 3226226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3227226223abSTejun Heo * @wq: the workqueue to register 3228226223abSTejun Heo * 3229226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3230226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3231226223abSTejun Heo * which is the preferred method. 3232226223abSTejun Heo * 3233226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3234226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3235226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3236226223abSTejun Heo * attributes. 3237226223abSTejun Heo * 3238d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. 3239226223abSTejun Heo */ 3240226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3241226223abSTejun Heo { 3242226223abSTejun Heo struct wq_device *wq_dev; 3243226223abSTejun Heo int ret; 3244226223abSTejun Heo 3245226223abSTejun Heo /* 3246226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3247226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3248226223abSTejun Heo * workqueues. 3249226223abSTejun Heo */ 3250226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3251226223abSTejun Heo return -EINVAL; 3252226223abSTejun Heo 3253226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3254226223abSTejun Heo if (!wq_dev) 3255226223abSTejun Heo return -ENOMEM; 3256226223abSTejun Heo 3257226223abSTejun Heo wq_dev->wq = wq; 3258226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3259226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3260226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3261226223abSTejun Heo 3262226223abSTejun Heo /* 3263226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3264226223abSTejun Heo * everything is ready. 3265226223abSTejun Heo */ 3266226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3267226223abSTejun Heo 3268226223abSTejun Heo ret = device_register(&wq_dev->dev); 3269226223abSTejun Heo if (ret) { 3270226223abSTejun Heo kfree(wq_dev); 3271226223abSTejun Heo wq->wq_dev = NULL; 3272226223abSTejun Heo return ret; 3273226223abSTejun Heo } 3274226223abSTejun Heo 3275226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3276226223abSTejun Heo struct device_attribute *attr; 3277226223abSTejun Heo 3278226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3279226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3280226223abSTejun Heo if (ret) { 3281226223abSTejun Heo device_unregister(&wq_dev->dev); 3282226223abSTejun Heo wq->wq_dev = NULL; 3283226223abSTejun Heo return ret; 3284226223abSTejun Heo } 3285226223abSTejun Heo } 3286226223abSTejun Heo } 3287226223abSTejun Heo 3288226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3289226223abSTejun Heo return 0; 3290226223abSTejun Heo } 3291226223abSTejun Heo 3292226223abSTejun Heo /** 3293226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3294226223abSTejun Heo * @wq: the workqueue to unregister 3295226223abSTejun Heo * 3296226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3297226223abSTejun Heo */ 3298226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3299226223abSTejun Heo { 3300226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3301226223abSTejun Heo 3302226223abSTejun Heo if (!wq->wq_dev) 3303226223abSTejun Heo return; 3304226223abSTejun Heo 3305226223abSTejun Heo wq->wq_dev = NULL; 3306226223abSTejun Heo device_unregister(&wq_dev->dev); 3307226223abSTejun Heo } 3308226223abSTejun Heo #else /* CONFIG_SYSFS */ 3309226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3310226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3311226223abSTejun Heo 33127a4e344cSTejun Heo /** 33137a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33147a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33157a4e344cSTejun Heo * 33167a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33177a4e344cSTejun Heo */ 33187a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33197a4e344cSTejun Heo { 33207a4e344cSTejun Heo if (attrs) { 33217a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33227a4e344cSTejun Heo kfree(attrs); 33237a4e344cSTejun Heo } 33247a4e344cSTejun Heo } 33257a4e344cSTejun Heo 33267a4e344cSTejun Heo /** 33277a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33287a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33297a4e344cSTejun Heo * 33307a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3331d185af30SYacine Belkadi * return it. 3332d185af30SYacine Belkadi * 3333d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33347a4e344cSTejun Heo */ 33357a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33367a4e344cSTejun Heo { 33377a4e344cSTejun Heo struct workqueue_attrs *attrs; 33387a4e344cSTejun Heo 33397a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33407a4e344cSTejun Heo if (!attrs) 33417a4e344cSTejun Heo goto fail; 33427a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33437a4e344cSTejun Heo goto fail; 33447a4e344cSTejun Heo 334513e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33467a4e344cSTejun Heo return attrs; 33477a4e344cSTejun Heo fail: 33487a4e344cSTejun Heo free_workqueue_attrs(attrs); 33497a4e344cSTejun Heo return NULL; 33507a4e344cSTejun Heo } 33517a4e344cSTejun Heo 335229c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 335329c91e99STejun Heo const struct workqueue_attrs *from) 335429c91e99STejun Heo { 335529c91e99STejun Heo to->nice = from->nice; 335629c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 33572865a8fbSShaohua Li /* 33582865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 33592865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 33602865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 33612865a8fbSShaohua Li */ 33622865a8fbSShaohua Li to->no_numa = from->no_numa; 336329c91e99STejun Heo } 336429c91e99STejun Heo 336529c91e99STejun Heo /* hash value of the content of @attr */ 336629c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 336729c91e99STejun Heo { 336829c91e99STejun Heo u32 hash = 0; 336929c91e99STejun Heo 337029c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 337113e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 337213e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 337329c91e99STejun Heo return hash; 337429c91e99STejun Heo } 337529c91e99STejun Heo 337629c91e99STejun Heo /* content equality test */ 337729c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 337829c91e99STejun Heo const struct workqueue_attrs *b) 337929c91e99STejun Heo { 338029c91e99STejun Heo if (a->nice != b->nice) 338129c91e99STejun Heo return false; 338229c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 338329c91e99STejun Heo return false; 338429c91e99STejun Heo return true; 338529c91e99STejun Heo } 338629c91e99STejun Heo 33877a4e344cSTejun Heo /** 33887a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33897a4e344cSTejun Heo * @pool: worker_pool to initialize 33907a4e344cSTejun Heo * 33917a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 3392d185af30SYacine Belkadi * 3393d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 339429c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 339529c91e99STejun Heo * on @pool safely to release it. 33967a4e344cSTejun Heo */ 33977a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33984e1a1f9aSTejun Heo { 33994e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 340029c91e99STejun Heo pool->id = -1; 340129c91e99STejun Heo pool->cpu = -1; 3402f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34034e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 34044e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34054e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34064e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34074e1a1f9aSTejun Heo 34084e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 34094e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 34104e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 34114e1a1f9aSTejun Heo 34124e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 34134e1a1f9aSTejun Heo (unsigned long)pool); 34144e1a1f9aSTejun Heo 34154e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 341692f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3417da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34187a4e344cSTejun Heo 34197cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 342029c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 342129c91e99STejun Heo pool->refcnt = 1; 342229c91e99STejun Heo 342329c91e99STejun Heo /* shouldn't fail above this point */ 34247a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34257a4e344cSTejun Heo if (!pool->attrs) 34267a4e344cSTejun Heo return -ENOMEM; 34277a4e344cSTejun Heo return 0; 34284e1a1f9aSTejun Heo } 34294e1a1f9aSTejun Heo 343029c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 343129c91e99STejun Heo { 343229c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 343329c91e99STejun Heo 34347cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 343529c91e99STejun Heo free_workqueue_attrs(pool->attrs); 343629c91e99STejun Heo kfree(pool); 343729c91e99STejun Heo } 343829c91e99STejun Heo 343929c91e99STejun Heo /** 344029c91e99STejun Heo * put_unbound_pool - put a worker_pool 344129c91e99STejun Heo * @pool: worker_pool to put 344229c91e99STejun Heo * 344329c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3444c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3445c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3446c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3447a892caccSTejun Heo * 3448a892caccSTejun Heo * Should be called with wq_pool_mutex held. 344929c91e99STejun Heo */ 345029c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 345129c91e99STejun Heo { 345260f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 345329c91e99STejun Heo struct worker *worker; 345429c91e99STejun Heo 3455a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3456a892caccSTejun Heo 3457a892caccSTejun Heo if (--pool->refcnt) 345829c91e99STejun Heo return; 345929c91e99STejun Heo 346029c91e99STejun Heo /* sanity checks */ 346129c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3462a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 346329c91e99STejun Heo return; 346429c91e99STejun Heo 346529c91e99STejun Heo /* release id and unhash */ 346629c91e99STejun Heo if (pool->id >= 0) 346729c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 346829c91e99STejun Heo hash_del(&pool->hash_node); 346929c91e99STejun Heo 3470c5aa87bbSTejun Heo /* 3471c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3472c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 347392f9c5c4SLai Jiangshan * attach_mutex. 3474c5aa87bbSTejun Heo */ 347529c91e99STejun Heo mutex_lock(&pool->manager_arb); 347629c91e99STejun Heo 347760f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 347829c91e99STejun Heo while ((worker = first_worker(pool))) 347929c91e99STejun Heo destroy_worker(worker); 348029c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 348129c91e99STejun Heo spin_unlock_irq(&pool->lock); 348260f5a4bcSLai Jiangshan 348392f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3484da028469SLai Jiangshan if (!list_empty(&pool->workers)) 348560f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 348692f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 348760f5a4bcSLai Jiangshan 348860f5a4bcSLai Jiangshan if (pool->detach_completion) 348960f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 349060f5a4bcSLai Jiangshan 349129c91e99STejun Heo mutex_unlock(&pool->manager_arb); 349229c91e99STejun Heo 349329c91e99STejun Heo /* shut down the timers */ 349429c91e99STejun Heo del_timer_sync(&pool->idle_timer); 349529c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 349629c91e99STejun Heo 349729c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 349829c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 349929c91e99STejun Heo } 350029c91e99STejun Heo 350129c91e99STejun Heo /** 350229c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 350329c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 350429c91e99STejun Heo * 350529c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 350629c91e99STejun Heo * reference count and return it. If there already is a matching 350729c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3508d185af30SYacine Belkadi * create a new one. 3509a892caccSTejun Heo * 3510a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3511d185af30SYacine Belkadi * 3512d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3513d185af30SYacine Belkadi * On failure, %NULL. 351429c91e99STejun Heo */ 351529c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 351629c91e99STejun Heo { 351729c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 351829c91e99STejun Heo struct worker_pool *pool; 3519f3f90ad4STejun Heo int node; 352029c91e99STejun Heo 3521a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 352229c91e99STejun Heo 352329c91e99STejun Heo /* do we already have a matching pool? */ 352429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 352529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 352629c91e99STejun Heo pool->refcnt++; 352729c91e99STejun Heo goto out_unlock; 352829c91e99STejun Heo } 352929c91e99STejun Heo } 353029c91e99STejun Heo 353129c91e99STejun Heo /* nope, create a new one */ 353229c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 353329c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 353429c91e99STejun Heo goto fail; 353529c91e99STejun Heo 353612ee4fc6SLai Jiangshan if (workqueue_freezing) 353712ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 353812ee4fc6SLai Jiangshan 35398864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 354029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 354129c91e99STejun Heo 35422865a8fbSShaohua Li /* 35432865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 35442865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 35452865a8fbSShaohua Li */ 35462865a8fbSShaohua Li pool->attrs->no_numa = false; 35472865a8fbSShaohua Li 3548f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3549f3f90ad4STejun Heo if (wq_numa_enabled) { 3550f3f90ad4STejun Heo for_each_node(node) { 3551f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3552f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3553f3f90ad4STejun Heo pool->node = node; 3554f3f90ad4STejun Heo break; 3555f3f90ad4STejun Heo } 3556f3f90ad4STejun Heo } 3557f3f90ad4STejun Heo } 3558f3f90ad4STejun Heo 355929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 356029c91e99STejun Heo goto fail; 356129c91e99STejun Heo 356229c91e99STejun Heo /* create and start the initial worker */ 3563ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 356429c91e99STejun Heo goto fail; 356529c91e99STejun Heo 356629c91e99STejun Heo /* install */ 356729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 356829c91e99STejun Heo out_unlock: 356929c91e99STejun Heo return pool; 357029c91e99STejun Heo fail: 357129c91e99STejun Heo if (pool) 357229c91e99STejun Heo put_unbound_pool(pool); 357329c91e99STejun Heo return NULL; 357429c91e99STejun Heo } 357529c91e99STejun Heo 35768864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35778864b4e5STejun Heo { 35788864b4e5STejun Heo kmem_cache_free(pwq_cache, 35798864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35808864b4e5STejun Heo } 35818864b4e5STejun Heo 35828864b4e5STejun Heo /* 35838864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35848864b4e5STejun Heo * and needs to be destroyed. 35858864b4e5STejun Heo */ 35868864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35878864b4e5STejun Heo { 35888864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35898864b4e5STejun Heo unbound_release_work); 35908864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35918864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3592bc0caf09STejun Heo bool is_last; 35938864b4e5STejun Heo 35948864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35958864b4e5STejun Heo return; 35968864b4e5STejun Heo 359775ccf595STejun Heo /* 35983c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 359975ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 360075ccf595STejun Heo * and consistent with the linking path. 360175ccf595STejun Heo */ 36023c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 36038864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3604bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 36053c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 36068864b4e5STejun Heo 3607a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 36088864b4e5STejun Heo put_unbound_pool(pool); 3609a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3610a892caccSTejun Heo 36118864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 36128864b4e5STejun Heo 36138864b4e5STejun Heo /* 36148864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 36158864b4e5STejun Heo * is gonna access it anymore. Free it. 36168864b4e5STejun Heo */ 36176029a918STejun Heo if (is_last) { 36186029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 36198864b4e5STejun Heo kfree(wq); 36208864b4e5STejun Heo } 36216029a918STejun Heo } 36228864b4e5STejun Heo 36230fbd95aaSTejun Heo /** 3624699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36250fbd95aaSTejun Heo * @pwq: target pool_workqueue 36260fbd95aaSTejun Heo * 3627699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3628699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3629699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36300fbd95aaSTejun Heo */ 3631699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36320fbd95aaSTejun Heo { 3633699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3634699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3635699ce097STejun Heo 3636699ce097STejun Heo /* for @wq->saved_max_active */ 3637a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3638699ce097STejun Heo 3639699ce097STejun Heo /* fast exit for non-freezable wqs */ 3640699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3641699ce097STejun Heo return; 3642699ce097STejun Heo 3643a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3644699ce097STejun Heo 3645699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3646699ce097STejun Heo pwq->max_active = wq->saved_max_active; 36470fbd95aaSTejun Heo 36480fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36490fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36500fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3651951a078aSLai Jiangshan 3652951a078aSLai Jiangshan /* 3653951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3654951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3655951a078aSLai Jiangshan */ 3656951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3657699ce097STejun Heo } else { 3658699ce097STejun Heo pwq->max_active = 0; 3659699ce097STejun Heo } 3660699ce097STejun Heo 3661a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36620fbd95aaSTejun Heo } 36630fbd95aaSTejun Heo 3664e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3665f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3666f147f29eSTejun Heo struct worker_pool *pool) 3667d2c1d404STejun Heo { 3668d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3669d2c1d404STejun Heo 3670e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3671e50aba9aSTejun Heo 3672d2c1d404STejun Heo pwq->pool = pool; 3673d2c1d404STejun Heo pwq->wq = wq; 3674d2c1d404STejun Heo pwq->flush_color = -1; 36758864b4e5STejun Heo pwq->refcnt = 1; 3676d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 36771befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3678d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36798864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3680f147f29eSTejun Heo } 3681d2c1d404STejun Heo 3682f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 36831befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3684f147f29eSTejun Heo { 3685f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3686f147f29eSTejun Heo 3687f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 368875ccf595STejun Heo 36891befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 36901befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 36911befcf30STejun Heo return; 36921befcf30STejun Heo 3693983ca25eSTejun Heo /* 3694983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 36953c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3696983ca25eSTejun Heo */ 369775ccf595STejun Heo pwq->work_color = wq->work_color; 3698983ca25eSTejun Heo 3699983ca25eSTejun Heo /* sync max_active to the current setting */ 3700983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3701983ca25eSTejun Heo 3702983ca25eSTejun Heo /* link in @pwq */ 37039e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3704df2d5ae4STejun Heo } 37056029a918STejun Heo 3706f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3707f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3708f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3709f147f29eSTejun Heo { 3710f147f29eSTejun Heo struct worker_pool *pool; 3711f147f29eSTejun Heo struct pool_workqueue *pwq; 3712f147f29eSTejun Heo 3713f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3714f147f29eSTejun Heo 3715f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3716f147f29eSTejun Heo if (!pool) 3717f147f29eSTejun Heo return NULL; 3718f147f29eSTejun Heo 3719e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3720f147f29eSTejun Heo if (!pwq) { 3721f147f29eSTejun Heo put_unbound_pool(pool); 3722f147f29eSTejun Heo return NULL; 3723f147f29eSTejun Heo } 3724f147f29eSTejun Heo 3725f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3726f147f29eSTejun Heo return pwq; 3727d2c1d404STejun Heo } 3728d2c1d404STejun Heo 37294c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 37304c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 37314c16bd32STejun Heo { 37324c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37334c16bd32STejun Heo 37344c16bd32STejun Heo if (pwq) { 37354c16bd32STejun Heo put_unbound_pool(pwq->pool); 3736cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 37374c16bd32STejun Heo } 37384c16bd32STejun Heo } 37394c16bd32STejun Heo 37404c16bd32STejun Heo /** 37414c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 37424c16bd32STejun Heo * @attrs: the wq_attrs of interest 37434c16bd32STejun Heo * @node: the target NUMA node 37444c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 37454c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 37464c16bd32STejun Heo * 37474c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 37484c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3749d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 37504c16bd32STejun Heo * 37514c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 37524c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 37534c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 37544c16bd32STejun Heo * @attrs->cpumask. 37554c16bd32STejun Heo * 37564c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 37574c16bd32STejun Heo * stable. 3758d185af30SYacine Belkadi * 3759d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3760d185af30SYacine Belkadi * %false if equal. 37614c16bd32STejun Heo */ 37624c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 37634c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 37644c16bd32STejun Heo { 3765d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 37664c16bd32STejun Heo goto use_dfl; 37674c16bd32STejun Heo 37684c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 37694c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 37704c16bd32STejun Heo if (cpu_going_down >= 0) 37714c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 37724c16bd32STejun Heo 37734c16bd32STejun Heo if (cpumask_empty(cpumask)) 37744c16bd32STejun Heo goto use_dfl; 37754c16bd32STejun Heo 37764c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 37774c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 37784c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 37794c16bd32STejun Heo 37804c16bd32STejun Heo use_dfl: 37814c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 37824c16bd32STejun Heo return false; 37834c16bd32STejun Heo } 37844c16bd32STejun Heo 37851befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 37861befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 37871befcf30STejun Heo int node, 37881befcf30STejun Heo struct pool_workqueue *pwq) 37891befcf30STejun Heo { 37901befcf30STejun Heo struct pool_workqueue *old_pwq; 37911befcf30STejun Heo 37921befcf30STejun Heo lockdep_assert_held(&wq->mutex); 37931befcf30STejun Heo 37941befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 37951befcf30STejun Heo link_pwq(pwq); 37961befcf30STejun Heo 37971befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 37981befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 37991befcf30STejun Heo return old_pwq; 38001befcf30STejun Heo } 38011befcf30STejun Heo 38029e8cd2f5STejun Heo /** 38039e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38049e8cd2f5STejun Heo * @wq: the target workqueue 38059e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38069e8cd2f5STejun Heo * 38074c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 38084c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 38094c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38104c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38114c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38124c16bd32STejun Heo * back-to-back will stay on its current pwq. 38139e8cd2f5STejun Heo * 3814d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3815d185af30SYacine Belkadi * 3816d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 38179e8cd2f5STejun Heo */ 38189e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38199e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38209e8cd2f5STejun Heo { 38214c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 38224c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3823f147f29eSTejun Heo int node, ret; 38249e8cd2f5STejun Heo 38258719dceaSTejun Heo /* only unbound workqueues can change attributes */ 38269e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 38279e8cd2f5STejun Heo return -EINVAL; 38289e8cd2f5STejun Heo 38298719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 38308719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 38318719dceaSTejun Heo return -EINVAL; 38328719dceaSTejun Heo 38334c16bd32STejun Heo pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); 383413e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38354c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38364c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 383713e2e556STejun Heo goto enomem; 383813e2e556STejun Heo 38394c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 384013e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 384113e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 384213e2e556STejun Heo 38434c16bd32STejun Heo /* 38444c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 38454c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 38464c16bd32STejun Heo * pools. 38474c16bd32STejun Heo */ 38484c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 38499e8cd2f5STejun Heo 38504c16bd32STejun Heo /* 38514c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 38524c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 38534c16bd32STejun Heo * pwqs accordingly. 38544c16bd32STejun Heo */ 38554c16bd32STejun Heo get_online_cpus(); 38564c16bd32STejun Heo 38574c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 38584c16bd32STejun Heo 38594c16bd32STejun Heo /* 38604c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 38614c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 38624c16bd32STejun Heo * it even if we don't use it immediately. 38634c16bd32STejun Heo */ 38644c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 38654c16bd32STejun Heo if (!dfl_pwq) 38664c16bd32STejun Heo goto enomem_pwq; 38674c16bd32STejun Heo 38684c16bd32STejun Heo for_each_node(node) { 38694c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 38704c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 38714c16bd32STejun Heo if (!pwq_tbl[node]) 38724c16bd32STejun Heo goto enomem_pwq; 38734c16bd32STejun Heo } else { 38744c16bd32STejun Heo dfl_pwq->refcnt++; 38754c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 38764c16bd32STejun Heo } 38774c16bd32STejun Heo } 38784c16bd32STejun Heo 38794c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 38804c16bd32STejun Heo 38814c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3882f147f29eSTejun Heo mutex_lock(&wq->mutex); 3883a892caccSTejun Heo 3884f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 38854c16bd32STejun Heo 38864c16bd32STejun Heo /* save the previous pwq and install the new one */ 3887f147f29eSTejun Heo for_each_node(node) 38884c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 38894c16bd32STejun Heo 38904c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 38914c16bd32STejun Heo link_pwq(dfl_pwq); 38924c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3893f147f29eSTejun Heo 3894f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3895f147f29eSTejun Heo 38964c16bd32STejun Heo /* put the old pwqs */ 38974c16bd32STejun Heo for_each_node(node) 38984c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 38994c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 39004c16bd32STejun Heo 39014c16bd32STejun Heo put_online_cpus(); 39024862125bSTejun Heo ret = 0; 39034862125bSTejun Heo /* fall through */ 39044862125bSTejun Heo out_free: 39054c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 39064862125bSTejun Heo free_workqueue_attrs(new_attrs); 39074c16bd32STejun Heo kfree(pwq_tbl); 39084862125bSTejun Heo return ret; 390913e2e556STejun Heo 39104c16bd32STejun Heo enomem_pwq: 39114c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 39124c16bd32STejun Heo for_each_node(node) 39134c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 39144c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 39154c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39164c16bd32STejun Heo put_online_cpus(); 391713e2e556STejun Heo enomem: 39184862125bSTejun Heo ret = -ENOMEM; 39194862125bSTejun Heo goto out_free; 39209e8cd2f5STejun Heo } 39219e8cd2f5STejun Heo 39224c16bd32STejun Heo /** 39234c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 39244c16bd32STejun Heo * @wq: the target workqueue 39254c16bd32STejun Heo * @cpu: the CPU coming up or going down 39264c16bd32STejun Heo * @online: whether @cpu is coming up or going down 39274c16bd32STejun Heo * 39284c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 39294c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 39304c16bd32STejun Heo * @wq accordingly. 39314c16bd32STejun Heo * 39324c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 39334c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 39344c16bd32STejun Heo * correct. 39354c16bd32STejun Heo * 39364c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 39374c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 39384c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 39394c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 39404c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 39414c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 39424c16bd32STejun Heo * CPU_DOWN_PREPARE. 39434c16bd32STejun Heo */ 39444c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 39454c16bd32STejun Heo bool online) 39464c16bd32STejun Heo { 39474c16bd32STejun Heo int node = cpu_to_node(cpu); 39484c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 39494c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 39504c16bd32STejun Heo struct workqueue_attrs *target_attrs; 39514c16bd32STejun Heo cpumask_t *cpumask; 39524c16bd32STejun Heo 39534c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 39544c16bd32STejun Heo 39554c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 39564c16bd32STejun Heo return; 39574c16bd32STejun Heo 39584c16bd32STejun Heo /* 39594c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39604c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39614c16bd32STejun Heo * CPU hotplug exclusion. 39624c16bd32STejun Heo */ 39634c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39644c16bd32STejun Heo cpumask = target_attrs->cpumask; 39654c16bd32STejun Heo 39664c16bd32STejun Heo mutex_lock(&wq->mutex); 3967d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 3968d55262c4STejun Heo goto out_unlock; 39694c16bd32STejun Heo 39704c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39714c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39724c16bd32STejun Heo 39734c16bd32STejun Heo /* 39744c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 39754c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 39764c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 3977534a3fbbSDaeseok Youn * wq's, the default pwq should be used. 39784c16bd32STejun Heo */ 39794c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 39804c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 39814c16bd32STejun Heo goto out_unlock; 39824c16bd32STejun Heo } else { 39834c16bd32STejun Heo goto use_dfl_pwq; 39844c16bd32STejun Heo } 39854c16bd32STejun Heo 39864c16bd32STejun Heo mutex_unlock(&wq->mutex); 39874c16bd32STejun Heo 39884c16bd32STejun Heo /* create a new pwq */ 39894c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 39904c16bd32STejun Heo if (!pwq) { 39912d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 39924c16bd32STejun Heo wq->name); 399377f300b1SDaeseok Youn mutex_lock(&wq->mutex); 399477f300b1SDaeseok Youn goto use_dfl_pwq; 39954c16bd32STejun Heo } 39964c16bd32STejun Heo 39974c16bd32STejun Heo /* 39984c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 39994c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 40004c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 40014c16bd32STejun Heo * inbetween. 40024c16bd32STejun Heo */ 40034c16bd32STejun Heo mutex_lock(&wq->mutex); 40044c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 40054c16bd32STejun Heo goto out_unlock; 40064c16bd32STejun Heo 40074c16bd32STejun Heo use_dfl_pwq: 40084c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 40094c16bd32STejun Heo get_pwq(wq->dfl_pwq); 40104c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 40114c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 40124c16bd32STejun Heo out_unlock: 40134c16bd32STejun Heo mutex_unlock(&wq->mutex); 40144c16bd32STejun Heo put_pwq_unlocked(old_pwq); 40154c16bd32STejun Heo } 40164c16bd32STejun Heo 401730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 40181da177e4SLinus Torvalds { 401949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 40208a2b7538STejun Heo int cpu, ret; 4021e1d8aa9fSFrederic Weisbecker 402230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4023420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4024420c0ddbSTejun Heo if (!wq->cpu_pwqs) 402530cdf249STejun Heo return -ENOMEM; 402630cdf249STejun Heo 402730cdf249STejun Heo for_each_possible_cpu(cpu) { 40287fb98ea7STejun Heo struct pool_workqueue *pwq = 40297fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 40307a62c2c8STejun Heo struct worker_pool *cpu_pools = 4031f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 403230cdf249STejun Heo 4033f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4034f147f29eSTejun Heo 4035f147f29eSTejun Heo mutex_lock(&wq->mutex); 40361befcf30STejun Heo link_pwq(pwq); 4037f147f29eSTejun Heo mutex_unlock(&wq->mutex); 403830cdf249STejun Heo } 403930cdf249STejun Heo return 0; 40408a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 40418a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 40428a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 40438a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 40448a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 40458a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 40468a2b7538STejun Heo return ret; 40479e8cd2f5STejun Heo } else { 40489e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 40499e8cd2f5STejun Heo } 40500f900049STejun Heo } 40510f900049STejun Heo 4052f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4053f3421797STejun Heo const char *name) 4054b71ab8c2STejun Heo { 4055f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4056f3421797STejun Heo 4057f3421797STejun Heo if (max_active < 1 || max_active > lim) 4058044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4059f3421797STejun Heo max_active, name, 1, lim); 4060b71ab8c2STejun Heo 4061f3421797STejun Heo return clamp_val(max_active, 1, lim); 4062b71ab8c2STejun Heo } 4063b71ab8c2STejun Heo 4064b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 406597e37d7bSTejun Heo unsigned int flags, 40661e19ffc6STejun Heo int max_active, 4067eb13ba87SJohannes Berg struct lock_class_key *key, 4068b196be89STejun Heo const char *lock_name, ...) 40693af24433SOleg Nesterov { 4070df2d5ae4STejun Heo size_t tbl_size = 0; 4071ecf6881fSTejun Heo va_list args; 40723af24433SOleg Nesterov struct workqueue_struct *wq; 407349e3cf44STejun Heo struct pool_workqueue *pwq; 4074b196be89STejun Heo 4075cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4076cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4077cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4078cee22a15SViresh Kumar 4079ecf6881fSTejun Heo /* allocate wq and format name */ 4080df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4081df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 4082df2d5ae4STejun Heo 4083df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4084b196be89STejun Heo if (!wq) 4085d2c1d404STejun Heo return NULL; 4086b196be89STejun Heo 40876029a918STejun Heo if (flags & WQ_UNBOUND) { 40886029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40896029a918STejun Heo if (!wq->unbound_attrs) 40906029a918STejun Heo goto err_free_wq; 40916029a918STejun Heo } 40926029a918STejun Heo 4093ecf6881fSTejun Heo va_start(args, lock_name); 4094ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4095b196be89STejun Heo va_end(args); 40963af24433SOleg Nesterov 4097d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4098b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40993af24433SOleg Nesterov 4100b196be89STejun Heo /* init wq */ 410197e37d7bSTejun Heo wq->flags = flags; 4102a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 41033c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4104112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 410530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 410673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 410773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4108493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 41093af24433SOleg Nesterov 4110eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4111cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 41123af24433SOleg Nesterov 411330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4114d2c1d404STejun Heo goto err_free_wq; 41151537663fSTejun Heo 4116493008a8STejun Heo /* 4117493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4118493008a8STejun Heo * have a rescuer to guarantee forward progress. 4119493008a8STejun Heo */ 4120493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4121e22bee78STejun Heo struct worker *rescuer; 4122e22bee78STejun Heo 4123d2c1d404STejun Heo rescuer = alloc_worker(); 4124e22bee78STejun Heo if (!rescuer) 4125d2c1d404STejun Heo goto err_destroy; 4126e22bee78STejun Heo 4127111c225aSTejun Heo rescuer->rescue_wq = wq; 4128111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4129b196be89STejun Heo wq->name); 4130d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4131d2c1d404STejun Heo kfree(rescuer); 4132d2c1d404STejun Heo goto err_destroy; 4133d2c1d404STejun Heo } 4134e22bee78STejun Heo 4135d2c1d404STejun Heo wq->rescuer = rescuer; 413614a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4137e22bee78STejun Heo wake_up_process(rescuer->task); 41383af24433SOleg Nesterov } 41391537663fSTejun Heo 4140226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4141226223abSTejun Heo goto err_destroy; 4142226223abSTejun Heo 41433af24433SOleg Nesterov /* 414468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 414568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 414668e13a67SLai Jiangshan * list. 41473af24433SOleg Nesterov */ 414868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4149a0a1a5fdSTejun Heo 4150a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 415149e3cf44STejun Heo for_each_pwq(pwq, wq) 4152699ce097STejun Heo pwq_adjust_max_active(pwq); 4153a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4154a0a1a5fdSTejun Heo 41553af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4156a0a1a5fdSTejun Heo 415768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41583af24433SOleg Nesterov 41593af24433SOleg Nesterov return wq; 4160d2c1d404STejun Heo 4161d2c1d404STejun Heo err_free_wq: 41626029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41634690c4abSTejun Heo kfree(wq); 4164d2c1d404STejun Heo return NULL; 4165d2c1d404STejun Heo err_destroy: 4166d2c1d404STejun Heo destroy_workqueue(wq); 41674690c4abSTejun Heo return NULL; 41681da177e4SLinus Torvalds } 4169d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41701da177e4SLinus Torvalds 41713af24433SOleg Nesterov /** 41723af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41733af24433SOleg Nesterov * @wq: target workqueue 41743af24433SOleg Nesterov * 41753af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41763af24433SOleg Nesterov */ 41773af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41783af24433SOleg Nesterov { 417949e3cf44STejun Heo struct pool_workqueue *pwq; 41804c16bd32STejun Heo int node; 41813af24433SOleg Nesterov 41829c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41839c5a2ba7STejun Heo drain_workqueue(wq); 4184c8efcc25STejun Heo 41856183c009STejun Heo /* sanity checks */ 4186b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 418749e3cf44STejun Heo for_each_pwq(pwq, wq) { 41886183c009STejun Heo int i; 41896183c009STejun Heo 419076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 419176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4192b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41936183c009STejun Heo return; 419476af4d93STejun Heo } 419576af4d93STejun Heo } 419676af4d93STejun Heo 41975c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 41988864b4e5STejun Heo WARN_ON(pwq->nr_active) || 419976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4200b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42016183c009STejun Heo return; 42026183c009STejun Heo } 420376af4d93STejun Heo } 4204b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42056183c009STejun Heo 4206a0a1a5fdSTejun Heo /* 4207a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4208a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4209a0a1a5fdSTejun Heo */ 421068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4211d2c1d404STejun Heo list_del_init(&wq->list); 421268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42133af24433SOleg Nesterov 4214226223abSTejun Heo workqueue_sysfs_unregister(wq); 4215226223abSTejun Heo 4216493008a8STejun Heo if (wq->rescuer) { 4217e22bee78STejun Heo kthread_stop(wq->rescuer->task); 42188d9df9f0SXiaotian Feng kfree(wq->rescuer); 4219493008a8STejun Heo wq->rescuer = NULL; 4220e22bee78STejun Heo } 4221e22bee78STejun Heo 42228864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 422329c91e99STejun Heo /* 42248864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 42258864b4e5STejun Heo * free the pwqs and wq. 422629c91e99STejun Heo */ 42278864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 42288864b4e5STejun Heo kfree(wq); 42298864b4e5STejun Heo } else { 42308864b4e5STejun Heo /* 42318864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 42324c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 42334c16bd32STejun Heo * @wq will be freed when the last pwq is released. 42348864b4e5STejun Heo */ 42354c16bd32STejun Heo for_each_node(node) { 42364c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42374c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 42384c16bd32STejun Heo put_pwq_unlocked(pwq); 42394c16bd32STejun Heo } 42404c16bd32STejun Heo 42414c16bd32STejun Heo /* 42424c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 42434c16bd32STejun Heo * put. Don't access it afterwards. 42444c16bd32STejun Heo */ 42454c16bd32STejun Heo pwq = wq->dfl_pwq; 42464c16bd32STejun Heo wq->dfl_pwq = NULL; 4247dce90d47STejun Heo put_pwq_unlocked(pwq); 424829c91e99STejun Heo } 42493af24433SOleg Nesterov } 42503af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 42513af24433SOleg Nesterov 4252dcd989cbSTejun Heo /** 4253dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4254dcd989cbSTejun Heo * @wq: target workqueue 4255dcd989cbSTejun Heo * @max_active: new max_active value. 4256dcd989cbSTejun Heo * 4257dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4258dcd989cbSTejun Heo * 4259dcd989cbSTejun Heo * CONTEXT: 4260dcd989cbSTejun Heo * Don't call from IRQ context. 4261dcd989cbSTejun Heo */ 4262dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4263dcd989cbSTejun Heo { 426449e3cf44STejun Heo struct pool_workqueue *pwq; 4265dcd989cbSTejun Heo 42668719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42678719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 42688719dceaSTejun Heo return; 42698719dceaSTejun Heo 4270f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4271dcd989cbSTejun Heo 4272a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4273dcd989cbSTejun Heo 4274dcd989cbSTejun Heo wq->saved_max_active = max_active; 4275dcd989cbSTejun Heo 4276699ce097STejun Heo for_each_pwq(pwq, wq) 4277699ce097STejun Heo pwq_adjust_max_active(pwq); 4278dcd989cbSTejun Heo 4279a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4280dcd989cbSTejun Heo } 4281dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4282dcd989cbSTejun Heo 4283dcd989cbSTejun Heo /** 4284e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4285e6267616STejun Heo * 4286e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4287e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4288d185af30SYacine Belkadi * 4289d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4290e6267616STejun Heo */ 4291e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4292e6267616STejun Heo { 4293e6267616STejun Heo struct worker *worker = current_wq_worker(); 4294e6267616STejun Heo 42956a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4296e6267616STejun Heo } 4297e6267616STejun Heo 4298e6267616STejun Heo /** 4299dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4300dcd989cbSTejun Heo * @cpu: CPU in question 4301dcd989cbSTejun Heo * @wq: target workqueue 4302dcd989cbSTejun Heo * 4303dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4304dcd989cbSTejun Heo * no synchronization around this function and the test result is 4305dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4306dcd989cbSTejun Heo * 4307d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4308d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4309d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4310d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4311d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4312d3251859STejun Heo * 4313d185af30SYacine Belkadi * Return: 4314dcd989cbSTejun Heo * %true if congested, %false otherwise. 4315dcd989cbSTejun Heo */ 4316d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4317dcd989cbSTejun Heo { 43187fb98ea7STejun Heo struct pool_workqueue *pwq; 431976af4d93STejun Heo bool ret; 432076af4d93STejun Heo 432188109453SLai Jiangshan rcu_read_lock_sched(); 43227fb98ea7STejun Heo 4323d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4324d3251859STejun Heo cpu = smp_processor_id(); 4325d3251859STejun Heo 43267fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 43277fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 43287fb98ea7STejun Heo else 4329df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4330dcd989cbSTejun Heo 433176af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 433288109453SLai Jiangshan rcu_read_unlock_sched(); 433376af4d93STejun Heo 433476af4d93STejun Heo return ret; 4335dcd989cbSTejun Heo } 4336dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4337dcd989cbSTejun Heo 4338dcd989cbSTejun Heo /** 4339dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4340dcd989cbSTejun Heo * @work: the work to be tested 4341dcd989cbSTejun Heo * 4342dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4343dcd989cbSTejun Heo * synchronization around this function and the test result is 4344dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4345dcd989cbSTejun Heo * 4346d185af30SYacine Belkadi * Return: 4347dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4348dcd989cbSTejun Heo */ 4349dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4350dcd989cbSTejun Heo { 4351fa1b54e6STejun Heo struct worker_pool *pool; 4352dcd989cbSTejun Heo unsigned long flags; 4353dcd989cbSTejun Heo unsigned int ret = 0; 4354dcd989cbSTejun Heo 4355dcd989cbSTejun Heo if (work_pending(work)) 4356dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4357038366c5SLai Jiangshan 4358fa1b54e6STejun Heo local_irq_save(flags); 4359fa1b54e6STejun Heo pool = get_work_pool(work); 4360038366c5SLai Jiangshan if (pool) { 4361fa1b54e6STejun Heo spin_lock(&pool->lock); 4362c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4363dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4364fa1b54e6STejun Heo spin_unlock(&pool->lock); 4365038366c5SLai Jiangshan } 4366fa1b54e6STejun Heo local_irq_restore(flags); 4367dcd989cbSTejun Heo 4368dcd989cbSTejun Heo return ret; 4369dcd989cbSTejun Heo } 4370dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4371dcd989cbSTejun Heo 43723d1cb205STejun Heo /** 43733d1cb205STejun Heo * set_worker_desc - set description for the current work item 43743d1cb205STejun Heo * @fmt: printf-style format string 43753d1cb205STejun Heo * @...: arguments for the format string 43763d1cb205STejun Heo * 43773d1cb205STejun Heo * This function can be called by a running work function to describe what 43783d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43793d1cb205STejun Heo * information will be printed out together to help debugging. The 43803d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43813d1cb205STejun Heo */ 43823d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43833d1cb205STejun Heo { 43843d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43853d1cb205STejun Heo va_list args; 43863d1cb205STejun Heo 43873d1cb205STejun Heo if (worker) { 43883d1cb205STejun Heo va_start(args, fmt); 43893d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 43903d1cb205STejun Heo va_end(args); 43913d1cb205STejun Heo worker->desc_valid = true; 43923d1cb205STejun Heo } 43933d1cb205STejun Heo } 43943d1cb205STejun Heo 43953d1cb205STejun Heo /** 43963d1cb205STejun Heo * print_worker_info - print out worker information and description 43973d1cb205STejun Heo * @log_lvl: the log level to use when printing 43983d1cb205STejun Heo * @task: target task 43993d1cb205STejun Heo * 44003d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 44013d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 44023d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 44033d1cb205STejun Heo * 44043d1cb205STejun Heo * This function can be safely called on any task as long as the 44053d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 44063d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 44073d1cb205STejun Heo */ 44083d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 44093d1cb205STejun Heo { 44103d1cb205STejun Heo work_func_t *fn = NULL; 44113d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 44123d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 44133d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 44143d1cb205STejun Heo struct workqueue_struct *wq = NULL; 44153d1cb205STejun Heo bool desc_valid = false; 44163d1cb205STejun Heo struct worker *worker; 44173d1cb205STejun Heo 44183d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 44193d1cb205STejun Heo return; 44203d1cb205STejun Heo 44213d1cb205STejun Heo /* 44223d1cb205STejun Heo * This function is called without any synchronization and @task 44233d1cb205STejun Heo * could be in any state. Be careful with dereferences. 44243d1cb205STejun Heo */ 44253d1cb205STejun Heo worker = probe_kthread_data(task); 44263d1cb205STejun Heo 44273d1cb205STejun Heo /* 44283d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 44293d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 44303d1cb205STejun Heo */ 44313d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 44323d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 44333d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 44343d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 44353d1cb205STejun Heo 44363d1cb205STejun Heo /* copy worker description */ 44373d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 44383d1cb205STejun Heo if (desc_valid) 44393d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 44403d1cb205STejun Heo 44413d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44422d916033SFabian Frederick pr_info("%sWorkqueue: %s %pf", log_lvl, name, fn); 44433d1cb205STejun Heo if (desc[0]) 44443d1cb205STejun Heo pr_cont(" (%s)", desc); 44453d1cb205STejun Heo pr_cont("\n"); 44463d1cb205STejun Heo } 44473d1cb205STejun Heo } 44483d1cb205STejun Heo 4449db7bccf4STejun Heo /* 4450db7bccf4STejun Heo * CPU hotplug. 4451db7bccf4STejun Heo * 4452e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4453112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4454706026c2STejun Heo * pool which make migrating pending and scheduled works very 4455e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 445694cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4457e22bee78STejun Heo * blocked draining impractical. 4458e22bee78STejun Heo * 445924647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4460628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4461628c78e7STejun Heo * cpu comes back online. 4462db7bccf4STejun Heo */ 4463db7bccf4STejun Heo 4464706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4465db7bccf4STejun Heo { 446638db41d9STejun Heo int cpu = smp_processor_id(); 44674ce62e9eSTejun Heo struct worker_pool *pool; 4468db7bccf4STejun Heo struct worker *worker; 4469db7bccf4STejun Heo 4470f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44716183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4472db7bccf4STejun Heo 447392f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 447494cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4475e22bee78STejun Heo 4476f2d5a0eeSTejun Heo /* 447792f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 447894cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 447994cf58bbSTejun Heo * except for the ones which are still executing works from 448094cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 448194cf58bbSTejun Heo * this, they may become diasporas. 4482f2d5a0eeSTejun Heo */ 4483da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4484403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4485db7bccf4STejun Heo 448624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4487f2d5a0eeSTejun Heo 448894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 448992f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4490e22bee78STejun Heo 4491e22bee78STejun Heo /* 4492eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4493eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4494eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4495eb283428SLai Jiangshan * from other cpus. 4496628c78e7STejun Heo */ 4497628c78e7STejun Heo schedule(); 4498628c78e7STejun Heo 4499628c78e7STejun Heo /* 4500eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4501eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4502eb283428SLai Jiangshan * and keep_working() are always true as long as the 4503eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4504eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4505eb283428SLai Jiangshan * are served by workers tied to the pool. 4506e22bee78STejun Heo */ 4507e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4508eb283428SLai Jiangshan 4509eb283428SLai Jiangshan /* 4510eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4511eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4512eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4513eb283428SLai Jiangshan */ 4514eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4515eb283428SLai Jiangshan wake_up_worker(pool); 4516eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4517eb283428SLai Jiangshan } 4518db7bccf4STejun Heo } 4519db7bccf4STejun Heo 4520bd7c089eSTejun Heo /** 4521bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4522bd7c089eSTejun Heo * @pool: pool of interest 4523bd7c089eSTejun Heo * 4524a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4525bd7c089eSTejun Heo */ 4526bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4527bd7c089eSTejun Heo { 4528a9ab775bSTejun Heo struct worker *worker; 4529bd7c089eSTejun Heo 453092f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4531bd7c089eSTejun Heo 4532bd7c089eSTejun Heo /* 4533a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4534a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4535a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4536a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4537a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4538bd7c089eSTejun Heo */ 4539da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4540a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4541a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4542a9ab775bSTejun Heo 4543a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4544a9ab775bSTejun Heo 4545da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4546a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4547a9ab775bSTejun Heo 4548a9ab775bSTejun Heo /* 4549a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4550a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4551a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4552a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4553a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4554a9ab775bSTejun Heo * be bound before @pool->lock is released. 4555a9ab775bSTejun Heo */ 4556a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4557bd7c089eSTejun Heo wake_up_process(worker->task); 4558bd7c089eSTejun Heo 4559bd7c089eSTejun Heo /* 4560a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4561a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4562a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4563a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4564a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4565a9ab775bSTejun Heo * concurrency management. Note that when or whether 4566a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4567a9ab775bSTejun Heo * 4568a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4569a9ab775bSTejun Heo * tested without holding any lock in 4570a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4571a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4572a9ab775bSTejun Heo * management operations. 4573bd7c089eSTejun Heo */ 4574a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4575a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4576a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4577a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4578bd7c089eSTejun Heo } 4579a9ab775bSTejun Heo 4580a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4581bd7c089eSTejun Heo } 4582bd7c089eSTejun Heo 45837dbc725eSTejun Heo /** 45847dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 45857dbc725eSTejun Heo * @pool: unbound pool of interest 45867dbc725eSTejun Heo * @cpu: the CPU which is coming up 45877dbc725eSTejun Heo * 45887dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 45897dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 45907dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 45917dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 45927dbc725eSTejun Heo */ 45937dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 45947dbc725eSTejun Heo { 45957dbc725eSTejun Heo static cpumask_t cpumask; 45967dbc725eSTejun Heo struct worker *worker; 45977dbc725eSTejun Heo 459892f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 45997dbc725eSTejun Heo 46007dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 46017dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 46027dbc725eSTejun Heo return; 46037dbc725eSTejun Heo 46047dbc725eSTejun Heo /* is @cpu the only online CPU? */ 46057dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 46067dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 46077dbc725eSTejun Heo return; 46087dbc725eSTejun Heo 46097dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4610da028469SLai Jiangshan for_each_pool_worker(worker, pool) 46117dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 46127dbc725eSTejun Heo pool->attrs->cpumask) < 0); 46137dbc725eSTejun Heo } 46147dbc725eSTejun Heo 46158db25e78STejun Heo /* 46168db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 46178db25e78STejun Heo * This will be registered high priority CPU notifier. 46188db25e78STejun Heo */ 46190db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 46201da177e4SLinus Torvalds unsigned long action, 46211da177e4SLinus Torvalds void *hcpu) 46221da177e4SLinus Torvalds { 4623d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46244ce62e9eSTejun Heo struct worker_pool *pool; 46254c16bd32STejun Heo struct workqueue_struct *wq; 46267dbc725eSTejun Heo int pi; 46271da177e4SLinus Torvalds 46288db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 46293af24433SOleg Nesterov case CPU_UP_PREPARE: 4630f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46313ce63377STejun Heo if (pool->nr_workers) 46323ce63377STejun Heo continue; 4633ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 46343ce63377STejun Heo return NOTIFY_BAD; 46353af24433SOleg Nesterov } 46361da177e4SLinus Torvalds break; 46371da177e4SLinus Torvalds 463865758202STejun Heo case CPU_DOWN_FAILED: 463965758202STejun Heo case CPU_ONLINE: 464068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46417dbc725eSTejun Heo 46427dbc725eSTejun Heo for_each_pool(pool, pi) { 464392f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 464494cf58bbSTejun Heo 46457dbc725eSTejun Heo if (pool->cpu == cpu) { 4646a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 464724647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4648a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4649a9ab775bSTejun Heo 465094cf58bbSTejun Heo rebind_workers(pool); 46517dbc725eSTejun Heo } else if (pool->cpu < 0) { 46527dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 46537dbc725eSTejun Heo } 465494cf58bbSTejun Heo 465592f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 465694cf58bbSTejun Heo } 46577dbc725eSTejun Heo 46584c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46594c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46604c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 46614c16bd32STejun Heo 466268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46638db25e78STejun Heo break; 466465758202STejun Heo } 466565758202STejun Heo return NOTIFY_OK; 466665758202STejun Heo } 466765758202STejun Heo 466865758202STejun Heo /* 466965758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 467065758202STejun Heo * This will be registered as low priority CPU notifier. 467165758202STejun Heo */ 46720db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 467365758202STejun Heo unsigned long action, 467465758202STejun Heo void *hcpu) 467565758202STejun Heo { 4676d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46778db25e78STejun Heo struct work_struct unbind_work; 46784c16bd32STejun Heo struct workqueue_struct *wq; 46798db25e78STejun Heo 468065758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 468165758202STejun Heo case CPU_DOWN_PREPARE: 46824c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4683706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46847635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46854c16bd32STejun Heo 46864c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46874c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46884c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46894c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 46904c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 46914c16bd32STejun Heo 46924c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 46938db25e78STejun Heo flush_work(&unbind_work); 4694440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 46958db25e78STejun Heo break; 469665758202STejun Heo } 469765758202STejun Heo return NOTIFY_OK; 469865758202STejun Heo } 469965758202STejun Heo 47002d3854a3SRusty Russell #ifdef CONFIG_SMP 47018ccad40dSRusty Russell 47022d3854a3SRusty Russell struct work_for_cpu { 4703ed48ece2STejun Heo struct work_struct work; 47042d3854a3SRusty Russell long (*fn)(void *); 47052d3854a3SRusty Russell void *arg; 47062d3854a3SRusty Russell long ret; 47072d3854a3SRusty Russell }; 47082d3854a3SRusty Russell 4709ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 47102d3854a3SRusty Russell { 4711ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4712ed48ece2STejun Heo 47132d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 47142d3854a3SRusty Russell } 47152d3854a3SRusty Russell 47162d3854a3SRusty Russell /** 47172d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 47182d3854a3SRusty Russell * @cpu: the cpu to run on 47192d3854a3SRusty Russell * @fn: the function to run 47202d3854a3SRusty Russell * @arg: the function arg 47212d3854a3SRusty Russell * 472231ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47236b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4724d185af30SYacine Belkadi * 4725d185af30SYacine Belkadi * Return: The value @fn returns. 47262d3854a3SRusty Russell */ 4727d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47282d3854a3SRusty Russell { 4729ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47302d3854a3SRusty Russell 4731ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4732ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 473312997d1aSBjorn Helgaas flush_work(&wfc.work); 4734440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 47352d3854a3SRusty Russell return wfc.ret; 47362d3854a3SRusty Russell } 47372d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47382d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47392d3854a3SRusty Russell 4740a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4741e7577c50SRusty Russell 4742a0a1a5fdSTejun Heo /** 4743a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4744a0a1a5fdSTejun Heo * 474558a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4746c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4747706026c2STejun Heo * pool->worklist. 4748a0a1a5fdSTejun Heo * 4749a0a1a5fdSTejun Heo * CONTEXT: 4750a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4751a0a1a5fdSTejun Heo */ 4752a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4753a0a1a5fdSTejun Heo { 475417116969STejun Heo struct worker_pool *pool; 475524b8a847STejun Heo struct workqueue_struct *wq; 475624b8a847STejun Heo struct pool_workqueue *pwq; 4757611c92a0STejun Heo int pi; 4758a0a1a5fdSTejun Heo 475968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4760a0a1a5fdSTejun Heo 47616183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4762a0a1a5fdSTejun Heo workqueue_freezing = true; 4763a0a1a5fdSTejun Heo 476424b8a847STejun Heo /* set FREEZING */ 4765611c92a0STejun Heo for_each_pool(pool, pi) { 47665bcab335STejun Heo spin_lock_irq(&pool->lock); 476735b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 476835b6bb63STejun Heo pool->flags |= POOL_FREEZING; 47695bcab335STejun Heo spin_unlock_irq(&pool->lock); 47701da177e4SLinus Torvalds } 47718b03ae3cSTejun Heo 477224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4773a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4774699ce097STejun Heo for_each_pwq(pwq, wq) 4775699ce097STejun Heo pwq_adjust_max_active(pwq); 4776a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4777a1056305STejun Heo } 47785bcab335STejun Heo 477968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4780a0a1a5fdSTejun Heo } 4781a0a1a5fdSTejun Heo 4782a0a1a5fdSTejun Heo /** 478358a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4784a0a1a5fdSTejun Heo * 4785a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4786a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4787a0a1a5fdSTejun Heo * 4788a0a1a5fdSTejun Heo * CONTEXT: 478968e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4790a0a1a5fdSTejun Heo * 4791d185af30SYacine Belkadi * Return: 479258a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 479358a69cb4STejun Heo * is complete. 4794a0a1a5fdSTejun Heo */ 4795a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4796a0a1a5fdSTejun Heo { 4797a0a1a5fdSTejun Heo bool busy = false; 479824b8a847STejun Heo struct workqueue_struct *wq; 479924b8a847STejun Heo struct pool_workqueue *pwq; 4800a0a1a5fdSTejun Heo 480168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4802a0a1a5fdSTejun Heo 48036183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4804a0a1a5fdSTejun Heo 480524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 480624b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 480724b8a847STejun Heo continue; 4808a0a1a5fdSTejun Heo /* 4809a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4810a0a1a5fdSTejun Heo * to peek without lock. 4811a0a1a5fdSTejun Heo */ 481288109453SLai Jiangshan rcu_read_lock_sched(); 481324b8a847STejun Heo for_each_pwq(pwq, wq) { 48146183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4815112202d9STejun Heo if (pwq->nr_active) { 4816a0a1a5fdSTejun Heo busy = true; 481788109453SLai Jiangshan rcu_read_unlock_sched(); 4818a0a1a5fdSTejun Heo goto out_unlock; 4819a0a1a5fdSTejun Heo } 4820a0a1a5fdSTejun Heo } 482188109453SLai Jiangshan rcu_read_unlock_sched(); 4822a0a1a5fdSTejun Heo } 4823a0a1a5fdSTejun Heo out_unlock: 482468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4825a0a1a5fdSTejun Heo return busy; 4826a0a1a5fdSTejun Heo } 4827a0a1a5fdSTejun Heo 4828a0a1a5fdSTejun Heo /** 4829a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4830a0a1a5fdSTejun Heo * 4831a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4832706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4833a0a1a5fdSTejun Heo * 4834a0a1a5fdSTejun Heo * CONTEXT: 4835a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4836a0a1a5fdSTejun Heo */ 4837a0a1a5fdSTejun Heo void thaw_workqueues(void) 4838a0a1a5fdSTejun Heo { 483924b8a847STejun Heo struct workqueue_struct *wq; 484024b8a847STejun Heo struct pool_workqueue *pwq; 484124b8a847STejun Heo struct worker_pool *pool; 4842611c92a0STejun Heo int pi; 4843a0a1a5fdSTejun Heo 484468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4845a0a1a5fdSTejun Heo 4846a0a1a5fdSTejun Heo if (!workqueue_freezing) 4847a0a1a5fdSTejun Heo goto out_unlock; 4848a0a1a5fdSTejun Heo 484924b8a847STejun Heo /* clear FREEZING */ 4850611c92a0STejun Heo for_each_pool(pool, pi) { 48515bcab335STejun Heo spin_lock_irq(&pool->lock); 485235b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 485335b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 48545bcab335STejun Heo spin_unlock_irq(&pool->lock); 4855d565ed63STejun Heo } 485624b8a847STejun Heo 485724b8a847STejun Heo /* restore max_active and repopulate worklist */ 485824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4859a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4860699ce097STejun Heo for_each_pwq(pwq, wq) 4861699ce097STejun Heo pwq_adjust_max_active(pwq); 4862a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 486324b8a847STejun Heo } 486424b8a847STejun Heo 4865a0a1a5fdSTejun Heo workqueue_freezing = false; 4866a0a1a5fdSTejun Heo out_unlock: 486768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4868a0a1a5fdSTejun Heo } 4869a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4870a0a1a5fdSTejun Heo 4871bce90380STejun Heo static void __init wq_numa_init(void) 4872bce90380STejun Heo { 4873bce90380STejun Heo cpumask_var_t *tbl; 4874bce90380STejun Heo int node, cpu; 4875bce90380STejun Heo 4876bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4877bce90380STejun Heo for_each_node(node) 4878bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4879bce90380STejun Heo 4880bce90380STejun Heo if (num_possible_nodes() <= 1) 4881bce90380STejun Heo return; 4882bce90380STejun Heo 4883d55262c4STejun Heo if (wq_disable_numa) { 4884d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 4885d55262c4STejun Heo return; 4886d55262c4STejun Heo } 4887d55262c4STejun Heo 48884c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 48894c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 48904c16bd32STejun Heo 4891bce90380STejun Heo /* 4892bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4893bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4894bce90380STejun Heo * fully initialized by now. 4895bce90380STejun Heo */ 4896bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4897bce90380STejun Heo BUG_ON(!tbl); 4898bce90380STejun Heo 4899bce90380STejun Heo for_each_node(node) 49001be0c25dSTejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 49011be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 4902bce90380STejun Heo 4903bce90380STejun Heo for_each_possible_cpu(cpu) { 4904bce90380STejun Heo node = cpu_to_node(cpu); 4905bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4906bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4907bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4908bce90380STejun Heo return; 4909bce90380STejun Heo } 4910bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4911bce90380STejun Heo } 4912bce90380STejun Heo 4913bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4914bce90380STejun Heo wq_numa_enabled = true; 4915bce90380STejun Heo } 4916bce90380STejun Heo 49176ee0578bSSuresh Siddha static int __init init_workqueues(void) 49181da177e4SLinus Torvalds { 49197a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 49207a4e344cSTejun Heo int i, cpu; 4921c34056a3STejun Heo 4922e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4923e904e6c2STejun Heo 4924e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4925e904e6c2STejun Heo 492665758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4927a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 49288b03ae3cSTejun Heo 4929bce90380STejun Heo wq_numa_init(); 4930bce90380STejun Heo 4931706026c2STejun Heo /* initialize CPU pools */ 493229c91e99STejun Heo for_each_possible_cpu(cpu) { 49334ce62e9eSTejun Heo struct worker_pool *pool; 49348b03ae3cSTejun Heo 49357a4e344cSTejun Heo i = 0; 4936f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49377a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4938ec22ca5eSTejun Heo pool->cpu = cpu; 49397a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 49407a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4941f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 49427a4e344cSTejun Heo 49439daf9e67STejun Heo /* alloc pool ID */ 494468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 49459daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 494668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 49474ce62e9eSTejun Heo } 49488b03ae3cSTejun Heo } 49498b03ae3cSTejun Heo 4950e22bee78STejun Heo /* create the initial worker */ 495129c91e99STejun Heo for_each_online_cpu(cpu) { 49524ce62e9eSTejun Heo struct worker_pool *pool; 4953e22bee78STejun Heo 4954f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 495524647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4956ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4957e22bee78STejun Heo } 49584ce62e9eSTejun Heo } 4959e22bee78STejun Heo 49608a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 496129c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 496229c91e99STejun Heo struct workqueue_attrs *attrs; 496329c91e99STejun Heo 496429c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 496529c91e99STejun Heo attrs->nice = std_nice[i]; 496629c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 49678a2b7538STejun Heo 49688a2b7538STejun Heo /* 49698a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 49708a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 49718a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 49728a2b7538STejun Heo */ 49738a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 49748a2b7538STejun Heo attrs->nice = std_nice[i]; 49758a2b7538STejun Heo attrs->no_numa = true; 49768a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 497729c91e99STejun Heo } 497829c91e99STejun Heo 4979d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 49801aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4981d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4982f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4983f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 498424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 498524d51addSTejun Heo WQ_FREEZABLE, 0); 49860668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 49870668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 49880668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 49890668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 49900668106cSViresh Kumar 0); 49911aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 49920668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 49930668106cSViresh Kumar !system_power_efficient_wq || 49940668106cSViresh Kumar !system_freezable_power_efficient_wq); 49956ee0578bSSuresh Siddha return 0; 49961da177e4SLinus Torvalds } 49976ee0578bSSuresh Siddha early_initcall(init_workqueues); 4998