11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 20b11895c4SLibin * normal work items and the other for high priority ones) and some extra 21b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 22b11895c4SLibin * number of these backing pools is dynamic. 23c54fce6eSTejun Heo * 24c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 251da177e4SLinus Torvalds */ 261da177e4SLinus Torvalds 279984de1aSPaul Gortmaker #include <linux/export.h> 281da177e4SLinus Torvalds #include <linux/kernel.h> 291da177e4SLinus Torvalds #include <linux/sched.h> 301da177e4SLinus Torvalds #include <linux/init.h> 311da177e4SLinus Torvalds #include <linux/signal.h> 321da177e4SLinus Torvalds #include <linux/completion.h> 331da177e4SLinus Torvalds #include <linux/workqueue.h> 341da177e4SLinus Torvalds #include <linux/slab.h> 351da177e4SLinus Torvalds #include <linux/cpu.h> 361da177e4SLinus Torvalds #include <linux/notifier.h> 371da177e4SLinus Torvalds #include <linux/kthread.h> 381fa44ecaSJames Bottomley #include <linux/hardirq.h> 3946934023SChristoph Lameter #include <linux/mempolicy.h> 40341a5958SRafael J. Wysocki #include <linux/freezer.h> 41d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51e22bee78STejun Heo 52ea138446STejun Heo #include "workqueue_internal.h" 531da177e4SLinus Torvalds 54c8e55f36STejun Heo enum { 55bc2ae0f5STejun Heo /* 5624647570STejun Heo * worker_pool flags 57bc2ae0f5STejun Heo * 5824647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 59bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 60bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 61bc2ae0f5STejun Heo * is in effect. 62bc2ae0f5STejun Heo * 63bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 64bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6524647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 66bc2ae0f5STejun Heo * 67bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 6892f9c5c4SLai Jiangshan * attach_mutex to avoid changing binding state while 694736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 70bc2ae0f5STejun Heo */ 7124647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 72db7bccf4STejun Heo 73c8e55f36STejun Heo /* worker flags */ 74c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 75c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 76e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 77fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 78f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 79a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 80e22bee78STejun Heo 81a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 82a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 83db7bccf4STejun Heo 84e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 854ce62e9eSTejun Heo 8629c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 87c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 88db7bccf4STejun Heo 89e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 90e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 91e22bee78STejun Heo 923233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 933233cdbdSTejun Heo /* call for help after 10ms 943233cdbdSTejun Heo (min two ticks) */ 95e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 96e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds /* 99e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1008698a745SDongsheng Yang * all cpus. Give MIN_NICE. 101e22bee78STejun Heo */ 1028698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1038698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 104ecf6881fSTejun Heo 105ecf6881fSTejun Heo WQ_NAME_LEN = 24, 106c8e55f36STejun Heo }; 107c8e55f36STejun Heo 1081da177e4SLinus Torvalds /* 1094690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1104690c4abSTejun Heo * 111e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 112e41e704bSTejun Heo * everyone else. 1134690c4abSTejun Heo * 114e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 115e22bee78STejun Heo * only be modified and accessed from the local cpu. 116e22bee78STejun Heo * 117d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1184690c4abSTejun Heo * 119d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 120d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 121d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 122d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 123e22bee78STejun Heo * 12492f9c5c4SLai Jiangshan * A: pool->attach_mutex protected. 125822d8405STejun Heo * 12668e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12776af4d93STejun Heo * 12868e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1295bcab335STejun Heo * 1303c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1313c25a55dSLai Jiangshan * 132b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1332e109a28STejun Heo * 1342e109a28STejun Heo * MD: wq_mayday_lock protected. 1354690c4abSTejun Heo */ 1364690c4abSTejun Heo 1372eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 138c34056a3STejun Heo 139bd7bdd43STejun Heo struct worker_pool { 140d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 141d84ff051STejun Heo int cpu; /* I: the associated cpu */ 142f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1439daf9e67STejun Heo int id; /* I: pool ID */ 14411ebea50STejun Heo unsigned int flags; /* X: flags */ 145bd7bdd43STejun Heo 146bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 147bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 148ea1abd61SLai Jiangshan 149ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 150bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 151bd7bdd43STejun Heo 152bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 153bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 154bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 155bd7bdd43STejun Heo 156c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 157c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 158c9e7cf27STejun Heo /* L: hash of busy workers */ 159c9e7cf27STejun Heo 160bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16134a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 1622607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 16392f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 16492f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 16560f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 166e19e397aSTejun Heo 1677cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 168e19e397aSTejun Heo 1697a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17068e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17168e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1727a4e344cSTejun Heo 173e19e397aSTejun Heo /* 174e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 175e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 176e19e397aSTejun Heo * cacheline. 177e19e397aSTejun Heo */ 178e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17929c91e99STejun Heo 18029c91e99STejun Heo /* 18129c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18229c91e99STejun Heo * from get_work_pool(). 18329c91e99STejun Heo */ 18429c91e99STejun Heo struct rcu_head rcu; 1858b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1868b03ae3cSTejun Heo 1878b03ae3cSTejun Heo /* 188112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 189112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 190112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 191112202d9STejun Heo * number of flag bits. 1921da177e4SLinus Torvalds */ 193112202d9STejun Heo struct pool_workqueue { 194bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1954690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19673f53c4aSTejun Heo int work_color; /* L: current color */ 19773f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1988864b4e5STejun Heo int refcnt; /* L: reference count */ 19973f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20073f53c4aSTejun Heo /* L: nr of in_flight works */ 2011e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 202a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2031e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2043c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2052e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2068864b4e5STejun Heo 2078864b4e5STejun Heo /* 2088864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2098864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2108864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 211b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2128864b4e5STejun Heo */ 2138864b4e5STejun Heo struct work_struct unbound_release_work; 2148864b4e5STejun Heo struct rcu_head rcu; 215e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* 21873f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21973f53c4aSTejun Heo */ 22073f53c4aSTejun Heo struct wq_flusher { 2213c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2223c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22373f53c4aSTejun Heo struct completion done; /* flush completion */ 22473f53c4aSTejun Heo }; 2251da177e4SLinus Torvalds 226226223abSTejun Heo struct wq_device; 227226223abSTejun Heo 22873f53c4aSTejun Heo /* 229c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 230c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2311da177e4SLinus Torvalds */ 2321da177e4SLinus Torvalds struct workqueue_struct { 2333c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 234e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 23573f53c4aSTejun Heo 2363c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2373c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2383c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 239112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2403c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2413c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2423c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24373f53c4aSTejun Heo 2442e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 245e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 246e22bee78STejun Heo 24787fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 248a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 249226223abSTejun Heo 2506029a918STejun Heo struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ 2514c16bd32STejun Heo struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ 2526029a918STejun Heo 253226223abSTejun Heo #ifdef CONFIG_SYSFS 254226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 255226223abSTejun Heo #endif 2564e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2574e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2584e6045f1SJohannes Berg #endif 259ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2602728fd2fSTejun Heo 261e2dca7adSTejun Heo /* 262e2dca7adSTejun Heo * Destruction of workqueue_struct is sched-RCU protected to allow 263e2dca7adSTejun Heo * walking the workqueues list without grabbing wq_pool_mutex. 264e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 265e2dca7adSTejun Heo */ 266e2dca7adSTejun Heo struct rcu_head rcu; 267e2dca7adSTejun Heo 2682728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2692728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2702728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 271df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2721da177e4SLinus Torvalds }; 2731da177e4SLinus Torvalds 274e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 275e904e6c2STejun Heo 276bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 277bce90380STejun Heo /* possible CPUs of each node */ 278bce90380STejun Heo 279d55262c4STejun Heo static bool wq_disable_numa; 280d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 281d55262c4STejun Heo 282cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 283cee22a15SViresh Kumar #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT 284cee22a15SViresh Kumar static bool wq_power_efficient = true; 285cee22a15SViresh Kumar #else 286cee22a15SViresh Kumar static bool wq_power_efficient; 287cee22a15SViresh Kumar #endif 288cee22a15SViresh Kumar 289cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 290cee22a15SViresh Kumar 291bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 292bce90380STejun Heo 2934c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2944c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2954c16bd32STejun Heo 29668e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2972e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2985bcab335STejun Heo 299e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30068e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3017d19c5ceSTejun Heo 3027d19c5ceSTejun Heo /* the per-cpu worker pools */ 3037d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 3047d19c5ceSTejun Heo cpu_worker_pools); 3057d19c5ceSTejun Heo 30668e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3077d19c5ceSTejun Heo 30868e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 30929c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 31029c91e99STejun Heo 311c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 31229c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 31329c91e99STejun Heo 3148a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3158a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3168a2b7538STejun Heo 317d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 318ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 319044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3201aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 321044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 322d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 323044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 324f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 325044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 32624d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3270668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3280668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3290668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3300668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 331d320c038STejun Heo 3327d19c5ceSTejun Heo static int worker_thread(void *__worker); 3337d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3347d19c5ceSTejun Heo const struct workqueue_attrs *from); 3357d19c5ceSTejun Heo 33697bd2347STejun Heo #define CREATE_TRACE_POINTS 33797bd2347STejun Heo #include <trace/events/workqueue.h> 33897bd2347STejun Heo 33968e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3405bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 34168e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 34268e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3435bcab335STejun Heo 344b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 34576af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 346b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 347b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 34876af4d93STejun Heo 349f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 350f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 351f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3527a62c2c8STejun Heo (pool)++) 3534ce62e9eSTejun Heo 35449e3cf44STejun Heo /** 35517116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 35617116969STejun Heo * @pool: iteration cursor 357611c92a0STejun Heo * @pi: integer used for iteration 358fa1b54e6STejun Heo * 35968e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 36068e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 36168e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 362fa1b54e6STejun Heo * 363fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 364fa1b54e6STejun Heo * ignored. 36517116969STejun Heo */ 366611c92a0STejun Heo #define for_each_pool(pool, pi) \ 367611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 36868e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 369fa1b54e6STejun Heo else 37017116969STejun Heo 37117116969STejun Heo /** 372822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 373822d8405STejun Heo * @worker: iteration cursor 374822d8405STejun Heo * @pool: worker_pool to iterate workers of 375822d8405STejun Heo * 37692f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 377822d8405STejun Heo * 378822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 379822d8405STejun Heo * ignored. 380822d8405STejun Heo */ 381da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 382da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 38392f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 384822d8405STejun Heo else 385822d8405STejun Heo 386822d8405STejun Heo /** 38749e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 38849e3cf44STejun Heo * @pwq: iteration cursor 38949e3cf44STejun Heo * @wq: the target workqueue 39076af4d93STejun Heo * 391b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 392794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 393794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 39476af4d93STejun Heo * 39576af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 39676af4d93STejun Heo * ignored. 39749e3cf44STejun Heo */ 39849e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 39976af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 400b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 40176af4d93STejun Heo else 402f3421797STejun Heo 403dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 404dc186ad7SThomas Gleixner 405dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 406dc186ad7SThomas Gleixner 40799777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 40899777288SStanislaw Gruszka { 40999777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 41099777288SStanislaw Gruszka } 41199777288SStanislaw Gruszka 412dc186ad7SThomas Gleixner /* 413dc186ad7SThomas Gleixner * fixup_init is called when: 414dc186ad7SThomas Gleixner * - an active object is initialized 415dc186ad7SThomas Gleixner */ 416dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 417dc186ad7SThomas Gleixner { 418dc186ad7SThomas Gleixner struct work_struct *work = addr; 419dc186ad7SThomas Gleixner 420dc186ad7SThomas Gleixner switch (state) { 421dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 422dc186ad7SThomas Gleixner cancel_work_sync(work); 423dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 424dc186ad7SThomas Gleixner return 1; 425dc186ad7SThomas Gleixner default: 426dc186ad7SThomas Gleixner return 0; 427dc186ad7SThomas Gleixner } 428dc186ad7SThomas Gleixner } 429dc186ad7SThomas Gleixner 430dc186ad7SThomas Gleixner /* 431dc186ad7SThomas Gleixner * fixup_activate is called when: 432dc186ad7SThomas Gleixner * - an active object is activated 433dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 434dc186ad7SThomas Gleixner */ 435dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 436dc186ad7SThomas Gleixner { 437dc186ad7SThomas Gleixner struct work_struct *work = addr; 438dc186ad7SThomas Gleixner 439dc186ad7SThomas Gleixner switch (state) { 440dc186ad7SThomas Gleixner 441dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 442dc186ad7SThomas Gleixner /* 443dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 444dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 445dc186ad7SThomas Gleixner * is tracked in the object tracker. 446dc186ad7SThomas Gleixner */ 44722df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 448dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 449dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 450dc186ad7SThomas Gleixner return 0; 451dc186ad7SThomas Gleixner } 452dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 453dc186ad7SThomas Gleixner return 0; 454dc186ad7SThomas Gleixner 455dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 456dc186ad7SThomas Gleixner WARN_ON(1); 457dc186ad7SThomas Gleixner 458dc186ad7SThomas Gleixner default: 459dc186ad7SThomas Gleixner return 0; 460dc186ad7SThomas Gleixner } 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner /* 464dc186ad7SThomas Gleixner * fixup_free is called when: 465dc186ad7SThomas Gleixner * - an active object is freed 466dc186ad7SThomas Gleixner */ 467dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 468dc186ad7SThomas Gleixner { 469dc186ad7SThomas Gleixner struct work_struct *work = addr; 470dc186ad7SThomas Gleixner 471dc186ad7SThomas Gleixner switch (state) { 472dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 473dc186ad7SThomas Gleixner cancel_work_sync(work); 474dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 475dc186ad7SThomas Gleixner return 1; 476dc186ad7SThomas Gleixner default: 477dc186ad7SThomas Gleixner return 0; 478dc186ad7SThomas Gleixner } 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner 481dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 482dc186ad7SThomas Gleixner .name = "work_struct", 48399777288SStanislaw Gruszka .debug_hint = work_debug_hint, 484dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 485dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 486dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 487dc186ad7SThomas Gleixner }; 488dc186ad7SThomas Gleixner 489dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 490dc186ad7SThomas Gleixner { 491dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 492dc186ad7SThomas Gleixner } 493dc186ad7SThomas Gleixner 494dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 495dc186ad7SThomas Gleixner { 496dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner 499dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 500dc186ad7SThomas Gleixner { 501dc186ad7SThomas Gleixner if (onstack) 502dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 503dc186ad7SThomas Gleixner else 504dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 505dc186ad7SThomas Gleixner } 506dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 507dc186ad7SThomas Gleixner 508dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 511dc186ad7SThomas Gleixner } 512dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 513dc186ad7SThomas Gleixner 514ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 515ea2e64f2SThomas Gleixner { 516ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 517ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 518ea2e64f2SThomas Gleixner } 519ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 520ea2e64f2SThomas Gleixner 521dc186ad7SThomas Gleixner #else 522dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 523dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 524dc186ad7SThomas Gleixner #endif 525dc186ad7SThomas Gleixner 5264e8b22bdSLi Bin /** 5274e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5284e8b22bdSLi Bin * @pool: the pool pointer of interest 5294e8b22bdSLi Bin * 5304e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5314e8b22bdSLi Bin * successfully, -errno on failure. 5324e8b22bdSLi Bin */ 5339daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5349daf9e67STejun Heo { 5359daf9e67STejun Heo int ret; 5369daf9e67STejun Heo 53768e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5385bcab335STejun Heo 5394e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5404e8b22bdSLi Bin GFP_KERNEL); 541229641a6STejun Heo if (ret >= 0) { 542e68035fbSTejun Heo pool->id = ret; 543229641a6STejun Heo return 0; 544229641a6STejun Heo } 5459daf9e67STejun Heo return ret; 5469daf9e67STejun Heo } 5479daf9e67STejun Heo 54876af4d93STejun Heo /** 549df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 550df2d5ae4STejun Heo * @wq: the target workqueue 551df2d5ae4STejun Heo * @node: the node ID 552df2d5ae4STejun Heo * 553df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 554df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 555df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 556d185af30SYacine Belkadi * 557d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 558df2d5ae4STejun Heo */ 559df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 560df2d5ae4STejun Heo int node) 561df2d5ae4STejun Heo { 562df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 563df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 564df2d5ae4STejun Heo } 565df2d5ae4STejun Heo 56673f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 56773f53c4aSTejun Heo { 56873f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 56973f53c4aSTejun Heo } 57073f53c4aSTejun Heo 57173f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 57273f53c4aSTejun Heo { 57373f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 57473f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 57573f53c4aSTejun Heo } 57673f53c4aSTejun Heo 57773f53c4aSTejun Heo static int work_next_color(int color) 57873f53c4aSTejun Heo { 57973f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 580a848e3b6SOleg Nesterov } 581a848e3b6SOleg Nesterov 5824594bf15SDavid Howells /* 583112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 584112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5857c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5867a22ad75STejun Heo * 587112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 588112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 589bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 590bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5917a22ad75STejun Heo * 592112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5937c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 594112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5957c3eed5cSTejun Heo * available only while the work item is queued. 596bbb68dfaSTejun Heo * 597bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 598bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 599bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 600bbb68dfaSTejun Heo * try to steal the PENDING bit. 6014594bf15SDavid Howells */ 6027a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6037a22ad75STejun Heo unsigned long flags) 6047a22ad75STejun Heo { 6056183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6067a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6077a22ad75STejun Heo } 6087a22ad75STejun Heo 609112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6104690c4abSTejun Heo unsigned long extra_flags) 611365970a1SDavid Howells { 612112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 613112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 614365970a1SDavid Howells } 615365970a1SDavid Howells 6164468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6174468a00fSLai Jiangshan int pool_id) 6184468a00fSLai Jiangshan { 6194468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6204468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6214468a00fSLai Jiangshan } 6224468a00fSLai Jiangshan 6237c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6247c3eed5cSTejun Heo int pool_id) 6254d707b9fSOleg Nesterov { 62623657bb1STejun Heo /* 62723657bb1STejun Heo * The following wmb is paired with the implied mb in 62823657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 62923657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 63023657bb1STejun Heo * owner. 63123657bb1STejun Heo */ 63223657bb1STejun Heo smp_wmb(); 6337c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6344d707b9fSOleg Nesterov } 6354d707b9fSOleg Nesterov 6367a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 637365970a1SDavid Howells { 6387c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6397c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6407a22ad75STejun Heo } 6417a22ad75STejun Heo 642112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6437a22ad75STejun Heo { 644e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6457a22ad75STejun Heo 646112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 647e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 648e120153dSTejun Heo else 649e120153dSTejun Heo return NULL; 6507a22ad75STejun Heo } 6517a22ad75STejun Heo 6527c3eed5cSTejun Heo /** 6537c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6547c3eed5cSTejun Heo * @work: the work item of interest 6557c3eed5cSTejun Heo * 65668e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 65768e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 65868e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 659fa1b54e6STejun Heo * 660fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 661fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 662fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 663fa1b54e6STejun Heo * returned pool is and stays online. 664d185af30SYacine Belkadi * 665d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6667c3eed5cSTejun Heo */ 6677c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6687a22ad75STejun Heo { 669e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6707c3eed5cSTejun Heo int pool_id; 6717a22ad75STejun Heo 67268e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 673fa1b54e6STejun Heo 674112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 675112202d9STejun Heo return ((struct pool_workqueue *) 6767c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6777a22ad75STejun Heo 6787c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6797c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6807a22ad75STejun Heo return NULL; 6817a22ad75STejun Heo 682fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6837c3eed5cSTejun Heo } 6847c3eed5cSTejun Heo 6857c3eed5cSTejun Heo /** 6867c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6877c3eed5cSTejun Heo * @work: the work item of interest 6887c3eed5cSTejun Heo * 689d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6907c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6917c3eed5cSTejun Heo */ 6927c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6937c3eed5cSTejun Heo { 69454d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6957c3eed5cSTejun Heo 696112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 697112202d9STejun Heo return ((struct pool_workqueue *) 69854d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 69954d5b7d0SLai Jiangshan 70054d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7017c3eed5cSTejun Heo } 7027c3eed5cSTejun Heo 703bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 704bbb68dfaSTejun Heo { 7057c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 706bbb68dfaSTejun Heo 7077c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7087c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 709bbb68dfaSTejun Heo } 710bbb68dfaSTejun Heo 711bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 712bbb68dfaSTejun Heo { 713bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 714bbb68dfaSTejun Heo 715112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 716bbb68dfaSTejun Heo } 717bbb68dfaSTejun Heo 718e22bee78STejun Heo /* 7193270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7203270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 721d565ed63STejun Heo * they're being called with pool->lock held. 722e22bee78STejun Heo */ 723e22bee78STejun Heo 72463d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 725649027d7STejun Heo { 726e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 727649027d7STejun Heo } 728649027d7STejun Heo 729e22bee78STejun Heo /* 730e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 731e22bee78STejun Heo * running workers. 732974271c4STejun Heo * 733974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 734706026c2STejun Heo * function will always return %true for unbound pools as long as the 735974271c4STejun Heo * worklist isn't empty. 736e22bee78STejun Heo */ 73763d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 738e22bee78STejun Heo { 73963d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 740e22bee78STejun Heo } 741e22bee78STejun Heo 742e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 74363d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 744e22bee78STejun Heo { 74563d95a91STejun Heo return pool->nr_idle; 746e22bee78STejun Heo } 747e22bee78STejun Heo 748e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 74963d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 750e22bee78STejun Heo { 751e19e397aSTejun Heo return !list_empty(&pool->worklist) && 752e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 753e22bee78STejun Heo } 754e22bee78STejun Heo 755e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 75663d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 757e22bee78STejun Heo { 75863d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 759e22bee78STejun Heo } 760e22bee78STejun Heo 761e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 76263d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 763e22bee78STejun Heo { 76434a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 76563d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 76663d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 767e22bee78STejun Heo 768e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 769e22bee78STejun Heo } 770e22bee78STejun Heo 771e22bee78STejun Heo /* 772e22bee78STejun Heo * Wake up functions. 773e22bee78STejun Heo */ 774e22bee78STejun Heo 7751037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 7761037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 7777e11629dSTejun Heo { 77863d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7797e11629dSTejun Heo return NULL; 7807e11629dSTejun Heo 78163d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7827e11629dSTejun Heo } 7837e11629dSTejun Heo 7847e11629dSTejun Heo /** 7857e11629dSTejun Heo * wake_up_worker - wake up an idle worker 78663d95a91STejun Heo * @pool: worker pool to wake worker from 7877e11629dSTejun Heo * 78863d95a91STejun Heo * Wake up the first idle worker of @pool. 7897e11629dSTejun Heo * 7907e11629dSTejun Heo * CONTEXT: 791d565ed63STejun Heo * spin_lock_irq(pool->lock). 7927e11629dSTejun Heo */ 79363d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7947e11629dSTejun Heo { 7951037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 7967e11629dSTejun Heo 7977e11629dSTejun Heo if (likely(worker)) 7987e11629dSTejun Heo wake_up_process(worker->task); 7997e11629dSTejun Heo } 8007e11629dSTejun Heo 8014690c4abSTejun Heo /** 802e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 803e22bee78STejun Heo * @task: task waking up 804e22bee78STejun Heo * @cpu: CPU @task is waking up to 805e22bee78STejun Heo * 806e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 807e22bee78STejun Heo * being awoken. 808e22bee78STejun Heo * 809e22bee78STejun Heo * CONTEXT: 810e22bee78STejun Heo * spin_lock_irq(rq->lock) 811e22bee78STejun Heo */ 812d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 813e22bee78STejun Heo { 814e22bee78STejun Heo struct worker *worker = kthread_data(task); 815e22bee78STejun Heo 81636576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 817ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 818e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 819e22bee78STejun Heo } 82036576000SJoonsoo Kim } 821e22bee78STejun Heo 822e22bee78STejun Heo /** 823e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 824e22bee78STejun Heo * @task: task going to sleep 825e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 826e22bee78STejun Heo * 827e22bee78STejun Heo * This function is called during schedule() when a busy worker is 828e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 829e22bee78STejun Heo * returning pointer to its task. 830e22bee78STejun Heo * 831e22bee78STejun Heo * CONTEXT: 832e22bee78STejun Heo * spin_lock_irq(rq->lock) 833e22bee78STejun Heo * 834d185af30SYacine Belkadi * Return: 835e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 836e22bee78STejun Heo */ 837d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 838e22bee78STejun Heo { 839e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 840111c225aSTejun Heo struct worker_pool *pool; 841e22bee78STejun Heo 842111c225aSTejun Heo /* 843111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 844111c225aSTejun Heo * workers, also reach here, let's not access anything before 845111c225aSTejun Heo * checking NOT_RUNNING. 846111c225aSTejun Heo */ 8472d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 848e22bee78STejun Heo return NULL; 849e22bee78STejun Heo 850111c225aSTejun Heo pool = worker->pool; 851111c225aSTejun Heo 852e22bee78STejun Heo /* this can only happen on the local cpu */ 85392b69f50SLai Jiangshan if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) 8546183c009STejun Heo return NULL; 855e22bee78STejun Heo 856e22bee78STejun Heo /* 857e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 858e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 859e22bee78STejun Heo * Please read comment there. 860e22bee78STejun Heo * 861628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 862628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 863628c78e7STejun Heo * disabled, which in turn means that none else could be 864d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 865628c78e7STejun Heo * lock is safe. 866e22bee78STejun Heo */ 867e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 868e19e397aSTejun Heo !list_empty(&pool->worklist)) 8691037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 870e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 871e22bee78STejun Heo } 872e22bee78STejun Heo 873e22bee78STejun Heo /** 874e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 875cb444766STejun Heo * @worker: self 876d302f017STejun Heo * @flags: flags to set 877d302f017STejun Heo * 878228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 879d302f017STejun Heo * 880cb444766STejun Heo * CONTEXT: 881d565ed63STejun Heo * spin_lock_irq(pool->lock) 882d302f017STejun Heo */ 883228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 884d302f017STejun Heo { 885bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 886e22bee78STejun Heo 887cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 888cb444766STejun Heo 889228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 890e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 891e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 892e19e397aSTejun Heo atomic_dec(&pool->nr_running); 893e22bee78STejun Heo } 894e22bee78STejun Heo 895d302f017STejun Heo worker->flags |= flags; 896d302f017STejun Heo } 897d302f017STejun Heo 898d302f017STejun Heo /** 899e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 900cb444766STejun Heo * @worker: self 901d302f017STejun Heo * @flags: flags to clear 902d302f017STejun Heo * 903e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 904d302f017STejun Heo * 905cb444766STejun Heo * CONTEXT: 906d565ed63STejun Heo * spin_lock_irq(pool->lock) 907d302f017STejun Heo */ 908d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 909d302f017STejun Heo { 91063d95a91STejun Heo struct worker_pool *pool = worker->pool; 911e22bee78STejun Heo unsigned int oflags = worker->flags; 912e22bee78STejun Heo 913cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 914cb444766STejun Heo 915d302f017STejun Heo worker->flags &= ~flags; 916e22bee78STejun Heo 91742c025f3STejun Heo /* 91842c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 91942c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 92042c025f3STejun Heo * of multiple flags, not a single flag. 92142c025f3STejun Heo */ 922e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 923e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 924e19e397aSTejun Heo atomic_inc(&pool->nr_running); 925d302f017STejun Heo } 926d302f017STejun Heo 927d302f017STejun Heo /** 9288cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 929c9e7cf27STejun Heo * @pool: pool of interest 9308cca0eeaSTejun Heo * @work: work to find worker for 9318cca0eeaSTejun Heo * 932c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 933c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 934a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 935a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 936a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 937a2c1c57bSTejun Heo * being executed. 938a2c1c57bSTejun Heo * 939a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 940a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 941a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 942a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 943a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 944a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 945a2c1c57bSTejun Heo * 946c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 947c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 948c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 949c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 950c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 951c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9528cca0eeaSTejun Heo * 9538cca0eeaSTejun Heo * CONTEXT: 954d565ed63STejun Heo * spin_lock_irq(pool->lock). 9558cca0eeaSTejun Heo * 956d185af30SYacine Belkadi * Return: 957d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9588cca0eeaSTejun Heo * otherwise. 9598cca0eeaSTejun Heo */ 960c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9618cca0eeaSTejun Heo struct work_struct *work) 9628cca0eeaSTejun Heo { 96342f8570fSSasha Levin struct worker *worker; 96442f8570fSSasha Levin 965b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 966a2c1c57bSTejun Heo (unsigned long)work) 967a2c1c57bSTejun Heo if (worker->current_work == work && 968a2c1c57bSTejun Heo worker->current_func == work->func) 96942f8570fSSasha Levin return worker; 97042f8570fSSasha Levin 97142f8570fSSasha Levin return NULL; 9728cca0eeaSTejun Heo } 9738cca0eeaSTejun Heo 9748cca0eeaSTejun Heo /** 975bf4ede01STejun Heo * move_linked_works - move linked works to a list 976bf4ede01STejun Heo * @work: start of series of works to be scheduled 977bf4ede01STejun Heo * @head: target list to append @work to 978bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 979bf4ede01STejun Heo * 980bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 981bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 982bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 983bf4ede01STejun Heo * 984bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 985bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 986bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 987bf4ede01STejun Heo * 988bf4ede01STejun Heo * CONTEXT: 989d565ed63STejun Heo * spin_lock_irq(pool->lock). 990bf4ede01STejun Heo */ 991bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 992bf4ede01STejun Heo struct work_struct **nextp) 993bf4ede01STejun Heo { 994bf4ede01STejun Heo struct work_struct *n; 995bf4ede01STejun Heo 996bf4ede01STejun Heo /* 997bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 998bf4ede01STejun Heo * use NULL for list head. 999bf4ede01STejun Heo */ 1000bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1001bf4ede01STejun Heo list_move_tail(&work->entry, head); 1002bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1003bf4ede01STejun Heo break; 1004bf4ede01STejun Heo } 1005bf4ede01STejun Heo 1006bf4ede01STejun Heo /* 1007bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1008bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1009bf4ede01STejun Heo * needs to be updated. 1010bf4ede01STejun Heo */ 1011bf4ede01STejun Heo if (nextp) 1012bf4ede01STejun Heo *nextp = n; 1013bf4ede01STejun Heo } 1014bf4ede01STejun Heo 10158864b4e5STejun Heo /** 10168864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10178864b4e5STejun Heo * @pwq: pool_workqueue to get 10188864b4e5STejun Heo * 10198864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10208864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10218864b4e5STejun Heo */ 10228864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10238864b4e5STejun Heo { 10248864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10258864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10268864b4e5STejun Heo pwq->refcnt++; 10278864b4e5STejun Heo } 10288864b4e5STejun Heo 10298864b4e5STejun Heo /** 10308864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10318864b4e5STejun Heo * @pwq: pool_workqueue to put 10328864b4e5STejun Heo * 10338864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10348864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10358864b4e5STejun Heo */ 10368864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10378864b4e5STejun Heo { 10388864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10398864b4e5STejun Heo if (likely(--pwq->refcnt)) 10408864b4e5STejun Heo return; 10418864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10428864b4e5STejun Heo return; 10438864b4e5STejun Heo /* 10448864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10458864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10468864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10478864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10488864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10498864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10508864b4e5STejun Heo */ 10518864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10528864b4e5STejun Heo } 10538864b4e5STejun Heo 1054dce90d47STejun Heo /** 1055dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1056dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1057dce90d47STejun Heo * 1058dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1059dce90d47STejun Heo */ 1060dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1061dce90d47STejun Heo { 1062dce90d47STejun Heo if (pwq) { 1063dce90d47STejun Heo /* 1064dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1065dce90d47STejun Heo * following lock operations are safe. 1066dce90d47STejun Heo */ 1067dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1068dce90d47STejun Heo put_pwq(pwq); 1069dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1070dce90d47STejun Heo } 1071dce90d47STejun Heo } 1072dce90d47STejun Heo 1073112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1074bf4ede01STejun Heo { 1075112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1076bf4ede01STejun Heo 1077bf4ede01STejun Heo trace_workqueue_activate_work(work); 1078112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1079bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1080112202d9STejun Heo pwq->nr_active++; 1081bf4ede01STejun Heo } 1082bf4ede01STejun Heo 1083112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10843aa62497SLai Jiangshan { 1085112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10863aa62497SLai Jiangshan struct work_struct, entry); 10873aa62497SLai Jiangshan 1088112202d9STejun Heo pwq_activate_delayed_work(work); 10893aa62497SLai Jiangshan } 10903aa62497SLai Jiangshan 1091bf4ede01STejun Heo /** 1092112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1093112202d9STejun Heo * @pwq: pwq of interest 1094bf4ede01STejun Heo * @color: color of work which left the queue 1095bf4ede01STejun Heo * 1096bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1097112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1098bf4ede01STejun Heo * 1099bf4ede01STejun Heo * CONTEXT: 1100d565ed63STejun Heo * spin_lock_irq(pool->lock). 1101bf4ede01STejun Heo */ 1102112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1103bf4ede01STejun Heo { 11048864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1105bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11068864b4e5STejun Heo goto out_put; 1107bf4ede01STejun Heo 1108112202d9STejun Heo pwq->nr_in_flight[color]--; 1109bf4ede01STejun Heo 1110112202d9STejun Heo pwq->nr_active--; 1111112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1112bf4ede01STejun Heo /* one down, submit a delayed one */ 1113112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1114112202d9STejun Heo pwq_activate_first_delayed(pwq); 1115bf4ede01STejun Heo } 1116bf4ede01STejun Heo 1117bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1118112202d9STejun Heo if (likely(pwq->flush_color != color)) 11198864b4e5STejun Heo goto out_put; 1120bf4ede01STejun Heo 1121bf4ede01STejun Heo /* are there still in-flight works? */ 1122112202d9STejun Heo if (pwq->nr_in_flight[color]) 11238864b4e5STejun Heo goto out_put; 1124bf4ede01STejun Heo 1125112202d9STejun Heo /* this pwq is done, clear flush_color */ 1126112202d9STejun Heo pwq->flush_color = -1; 1127bf4ede01STejun Heo 1128bf4ede01STejun Heo /* 1129112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1130bf4ede01STejun Heo * will handle the rest. 1131bf4ede01STejun Heo */ 1132112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1133112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11348864b4e5STejun Heo out_put: 11358864b4e5STejun Heo put_pwq(pwq); 1136bf4ede01STejun Heo } 1137bf4ede01STejun Heo 113836e227d2STejun Heo /** 1139bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 114036e227d2STejun Heo * @work: work item to steal 114136e227d2STejun Heo * @is_dwork: @work is a delayed_work 1142bbb68dfaSTejun Heo * @flags: place to store irq state 114336e227d2STejun Heo * 114436e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1145d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 114636e227d2STejun Heo * 1147d185af30SYacine Belkadi * Return: 114836e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 114936e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 115036e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1151bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1152bbb68dfaSTejun Heo * for arbitrarily long 115336e227d2STejun Heo * 1154d185af30SYacine Belkadi * Note: 1155bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1156e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1157e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1158e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1159bbb68dfaSTejun Heo * 1160bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1161bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1162bbb68dfaSTejun Heo * 1163e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1164bf4ede01STejun Heo */ 1165bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1166bbb68dfaSTejun Heo unsigned long *flags) 1167bf4ede01STejun Heo { 1168d565ed63STejun Heo struct worker_pool *pool; 1169112202d9STejun Heo struct pool_workqueue *pwq; 1170bf4ede01STejun Heo 1171bbb68dfaSTejun Heo local_irq_save(*flags); 1172bbb68dfaSTejun Heo 117336e227d2STejun Heo /* try to steal the timer if it exists */ 117436e227d2STejun Heo if (is_dwork) { 117536e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 117636e227d2STejun Heo 1177e0aecdd8STejun Heo /* 1178e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1179e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1180e0aecdd8STejun Heo * running on the local CPU. 1181e0aecdd8STejun Heo */ 118236e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 118336e227d2STejun Heo return 1; 118436e227d2STejun Heo } 118536e227d2STejun Heo 118636e227d2STejun Heo /* try to claim PENDING the normal way */ 1187bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1188bf4ede01STejun Heo return 0; 1189bf4ede01STejun Heo 1190bf4ede01STejun Heo /* 1191bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1192bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1193bf4ede01STejun Heo */ 1194d565ed63STejun Heo pool = get_work_pool(work); 1195d565ed63STejun Heo if (!pool) 1196bbb68dfaSTejun Heo goto fail; 1197bf4ede01STejun Heo 1198d565ed63STejun Heo spin_lock(&pool->lock); 1199bf4ede01STejun Heo /* 1200112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1201112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1202112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1203112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1204112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12050b3dae68SLai Jiangshan * item is currently queued on that pool. 1206bf4ede01STejun Heo */ 1207112202d9STejun Heo pwq = get_work_pwq(work); 1208112202d9STejun Heo if (pwq && pwq->pool == pool) { 1209bf4ede01STejun Heo debug_work_deactivate(work); 12103aa62497SLai Jiangshan 12113aa62497SLai Jiangshan /* 121216062836STejun Heo * A delayed work item cannot be grabbed directly because 121316062836STejun Heo * it might have linked NO_COLOR work items which, if left 1214112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 121516062836STejun Heo * management later on and cause stall. Make sure the work 121616062836STejun Heo * item is activated before grabbing. 12173aa62497SLai Jiangshan */ 12183aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1219112202d9STejun Heo pwq_activate_delayed_work(work); 12203aa62497SLai Jiangshan 1221bf4ede01STejun Heo list_del_init(&work->entry); 12229c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 122336e227d2STejun Heo 1224112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12254468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12264468a00fSLai Jiangshan 1227d565ed63STejun Heo spin_unlock(&pool->lock); 122836e227d2STejun Heo return 1; 1229bf4ede01STejun Heo } 1230d565ed63STejun Heo spin_unlock(&pool->lock); 1231bbb68dfaSTejun Heo fail: 1232bbb68dfaSTejun Heo local_irq_restore(*flags); 1233bbb68dfaSTejun Heo if (work_is_canceling(work)) 1234bbb68dfaSTejun Heo return -ENOENT; 1235bbb68dfaSTejun Heo cpu_relax(); 123636e227d2STejun Heo return -EAGAIN; 1237bf4ede01STejun Heo } 1238bf4ede01STejun Heo 1239bf4ede01STejun Heo /** 1240706026c2STejun Heo * insert_work - insert a work into a pool 1241112202d9STejun Heo * @pwq: pwq @work belongs to 12424690c4abSTejun Heo * @work: work to insert 12434690c4abSTejun Heo * @head: insertion point 12444690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12454690c4abSTejun Heo * 1246112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1247706026c2STejun Heo * work_struct flags. 12484690c4abSTejun Heo * 12494690c4abSTejun Heo * CONTEXT: 1250d565ed63STejun Heo * spin_lock_irq(pool->lock). 1251365970a1SDavid Howells */ 1252112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1253112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1254b89deed3SOleg Nesterov { 1255112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1256e1d8aa9fSFrederic Weisbecker 12574690c4abSTejun Heo /* we own @work, set data and link */ 1258112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12591a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12608864b4e5STejun Heo get_pwq(pwq); 1261e22bee78STejun Heo 1262e22bee78STejun Heo /* 1263c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1264c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1265c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1266e22bee78STejun Heo */ 1267e22bee78STejun Heo smp_mb(); 1268e22bee78STejun Heo 126963d95a91STejun Heo if (__need_more_worker(pool)) 127063d95a91STejun Heo wake_up_worker(pool); 1271b89deed3SOleg Nesterov } 1272b89deed3SOleg Nesterov 1273c8efcc25STejun Heo /* 1274c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12758d03ecfeSTejun Heo * same workqueue. 1276c8efcc25STejun Heo */ 1277c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1278c8efcc25STejun Heo { 1279c8efcc25STejun Heo struct worker *worker; 1280c8efcc25STejun Heo 12818d03ecfeSTejun Heo worker = current_wq_worker(); 1282c8efcc25STejun Heo /* 12838d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12848d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1285c8efcc25STejun Heo */ 1286112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1287c8efcc25STejun Heo } 1288c8efcc25STejun Heo 1289d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12901da177e4SLinus Torvalds struct work_struct *work) 12911da177e4SLinus Torvalds { 1292112202d9STejun Heo struct pool_workqueue *pwq; 1293c9178087STejun Heo struct worker_pool *last_pool; 12941e19ffc6STejun Heo struct list_head *worklist; 12958a2e8e5dSTejun Heo unsigned int work_flags; 1296b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12978930cabaSTejun Heo 12988930cabaSTejun Heo /* 12998930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13008930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13018930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13028930cabaSTejun Heo * happen with IRQ disabled. 13038930cabaSTejun Heo */ 13048930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13051da177e4SLinus Torvalds 1306dc186ad7SThomas Gleixner debug_work_activate(work); 13071e19ffc6STejun Heo 13089ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1309618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1310c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1311e41e704bSTejun Heo return; 13129e8cd2f5STejun Heo retry: 1313df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1314f3421797STejun Heo cpu = raw_smp_processor_id(); 1315df2d5ae4STejun Heo 1316df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1317df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1318c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1319df2d5ae4STejun Heo else 1320df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1321f3421797STejun Heo 132218aa9effSTejun Heo /* 1323c9178087STejun Heo * If @work was previously on a different pool, it might still be 1324c9178087STejun Heo * running there, in which case the work needs to be queued on that 1325c9178087STejun Heo * pool to guarantee non-reentrancy. 132618aa9effSTejun Heo */ 1327c9e7cf27STejun Heo last_pool = get_work_pool(work); 1328112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 132918aa9effSTejun Heo struct worker *worker; 133018aa9effSTejun Heo 1331d565ed63STejun Heo spin_lock(&last_pool->lock); 133218aa9effSTejun Heo 1333c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 133418aa9effSTejun Heo 1335112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1336c9178087STejun Heo pwq = worker->current_pwq; 13378594fadeSLai Jiangshan } else { 133818aa9effSTejun Heo /* meh... not running there, queue here */ 1339d565ed63STejun Heo spin_unlock(&last_pool->lock); 1340112202d9STejun Heo spin_lock(&pwq->pool->lock); 134118aa9effSTejun Heo } 13428930cabaSTejun Heo } else { 1343112202d9STejun Heo spin_lock(&pwq->pool->lock); 13448930cabaSTejun Heo } 1345502ca9d8STejun Heo 13469e8cd2f5STejun Heo /* 13479e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13489e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13499e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1350df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1351df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13529e8cd2f5STejun Heo * make forward-progress. 13539e8cd2f5STejun Heo */ 13549e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13559e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13569e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13579e8cd2f5STejun Heo cpu_relax(); 13589e8cd2f5STejun Heo goto retry; 13599e8cd2f5STejun Heo } 13609e8cd2f5STejun Heo /* oops */ 13619e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13629e8cd2f5STejun Heo wq->name, cpu); 13639e8cd2f5STejun Heo } 13649e8cd2f5STejun Heo 1365112202d9STejun Heo /* pwq determined, queue */ 1366112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1367502ca9d8STejun Heo 1368f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1369112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1370f5b2552bSDan Carpenter return; 1371f5b2552bSDan Carpenter } 13721e19ffc6STejun Heo 1373112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1374112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13751e19ffc6STejun Heo 1376112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1377cdadf009STejun Heo trace_workqueue_activate_work(work); 1378112202d9STejun Heo pwq->nr_active++; 1379112202d9STejun Heo worklist = &pwq->pool->worklist; 13808a2e8e5dSTejun Heo } else { 13818a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1382112202d9STejun Heo worklist = &pwq->delayed_works; 13838a2e8e5dSTejun Heo } 13841e19ffc6STejun Heo 1385112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13861e19ffc6STejun Heo 1387112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13881da177e4SLinus Torvalds } 13891da177e4SLinus Torvalds 13900fcb78c2SRolf Eike Beer /** 1391c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1392c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1393c1a220e7SZhang Rui * @wq: workqueue to use 1394c1a220e7SZhang Rui * @work: work to queue 1395c1a220e7SZhang Rui * 1396c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1397c1a220e7SZhang Rui * can't go away. 1398d185af30SYacine Belkadi * 1399d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1400c1a220e7SZhang Rui */ 1401d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1402d4283e93STejun Heo struct work_struct *work) 1403c1a220e7SZhang Rui { 1404d4283e93STejun Heo bool ret = false; 14058930cabaSTejun Heo unsigned long flags; 14068930cabaSTejun Heo 14078930cabaSTejun Heo local_irq_save(flags); 1408c1a220e7SZhang Rui 140922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14104690c4abSTejun Heo __queue_work(cpu, wq, work); 1411d4283e93STejun Heo ret = true; 1412c1a220e7SZhang Rui } 14138930cabaSTejun Heo 14148930cabaSTejun Heo local_irq_restore(flags); 1415c1a220e7SZhang Rui return ret; 1416c1a220e7SZhang Rui } 1417ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1418c1a220e7SZhang Rui 1419d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14201da177e4SLinus Torvalds { 142152bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14221da177e4SLinus Torvalds 1423e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 142460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14251da177e4SLinus Torvalds } 14261438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14271da177e4SLinus Torvalds 14287beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 142952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14301da177e4SLinus Torvalds { 14317beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14327beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14331da177e4SLinus Torvalds 14347beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14357beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1436fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1437fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14387beb2edfSTejun Heo 14398852aac2STejun Heo /* 14408852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14418852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14428852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14438852aac2STejun Heo * on that there's no such delay when @delay is 0. 14448852aac2STejun Heo */ 14458852aac2STejun Heo if (!delay) { 14468852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14478852aac2STejun Heo return; 14488852aac2STejun Heo } 14498852aac2STejun Heo 14507beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14517beb2edfSTejun Heo 145260c057bcSLai Jiangshan dwork->wq = wq; 14531265057fSTejun Heo dwork->cpu = cpu; 14547beb2edfSTejun Heo timer->expires = jiffies + delay; 14557beb2edfSTejun Heo 14567beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14577beb2edfSTejun Heo add_timer_on(timer, cpu); 14587beb2edfSTejun Heo else 14597beb2edfSTejun Heo add_timer(timer); 14607beb2edfSTejun Heo } 14611da177e4SLinus Torvalds 14620fcb78c2SRolf Eike Beer /** 14630fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14640fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14650fcb78c2SRolf Eike Beer * @wq: workqueue to use 1466af9997e4SRandy Dunlap * @dwork: work to queue 14670fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14680fcb78c2SRolf Eike Beer * 1469d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1470715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1471715f1300STejun Heo * execution. 14720fcb78c2SRolf Eike Beer */ 1473d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 147452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14757a6bc1cdSVenkatesh Pallipadi { 147652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1477d4283e93STejun Heo bool ret = false; 14788930cabaSTejun Heo unsigned long flags; 14798930cabaSTejun Heo 14808930cabaSTejun Heo /* read the comment in __queue_work() */ 14818930cabaSTejun Heo local_irq_save(flags); 14827a6bc1cdSVenkatesh Pallipadi 148322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14847beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1485d4283e93STejun Heo ret = true; 14867a6bc1cdSVenkatesh Pallipadi } 14878930cabaSTejun Heo 14888930cabaSTejun Heo local_irq_restore(flags); 14897a6bc1cdSVenkatesh Pallipadi return ret; 14907a6bc1cdSVenkatesh Pallipadi } 1491ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 14921da177e4SLinus Torvalds 1493c8e55f36STejun Heo /** 14948376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14958376fe22STejun Heo * @cpu: CPU number to execute work on 14968376fe22STejun Heo * @wq: workqueue to use 14978376fe22STejun Heo * @dwork: work to queue 14988376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14998376fe22STejun Heo * 15008376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15018376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15028376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15038376fe22STejun Heo * current state. 15048376fe22STejun Heo * 1505d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15068376fe22STejun Heo * pending and its timer was modified. 15078376fe22STejun Heo * 1508e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15098376fe22STejun Heo * See try_to_grab_pending() for details. 15108376fe22STejun Heo */ 15118376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15128376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15138376fe22STejun Heo { 15148376fe22STejun Heo unsigned long flags; 15158376fe22STejun Heo int ret; 15168376fe22STejun Heo 15178376fe22STejun Heo do { 15188376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15198376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15208376fe22STejun Heo 15218376fe22STejun Heo if (likely(ret >= 0)) { 15228376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15238376fe22STejun Heo local_irq_restore(flags); 15248376fe22STejun Heo } 15258376fe22STejun Heo 15268376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15278376fe22STejun Heo return ret; 15288376fe22STejun Heo } 15298376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15308376fe22STejun Heo 15318376fe22STejun Heo /** 1532c8e55f36STejun Heo * worker_enter_idle - enter idle state 1533c8e55f36STejun Heo * @worker: worker which is entering idle state 1534c8e55f36STejun Heo * 1535c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1536c8e55f36STejun Heo * necessary. 1537c8e55f36STejun Heo * 1538c8e55f36STejun Heo * LOCKING: 1539d565ed63STejun Heo * spin_lock_irq(pool->lock). 1540c8e55f36STejun Heo */ 1541c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15421da177e4SLinus Torvalds { 1543bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1544c8e55f36STejun Heo 15456183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15466183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15476183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15486183c009STejun Heo return; 1549c8e55f36STejun Heo 1550051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1551cb444766STejun Heo worker->flags |= WORKER_IDLE; 1552bd7bdd43STejun Heo pool->nr_idle++; 1553e22bee78STejun Heo worker->last_active = jiffies; 1554c8e55f36STejun Heo 1555c8e55f36STejun Heo /* idle_list is LIFO */ 1556bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1557db7bccf4STejun Heo 155863d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1559628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1560cb444766STejun Heo 1561544ecf31STejun Heo /* 1562706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1563d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1564628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1565628c78e7STejun Heo * unbind is not in progress. 1566544ecf31STejun Heo */ 156724647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1568bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1569e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1570c8e55f36STejun Heo } 1571c8e55f36STejun Heo 1572c8e55f36STejun Heo /** 1573c8e55f36STejun Heo * worker_leave_idle - leave idle state 1574c8e55f36STejun Heo * @worker: worker which is leaving idle state 1575c8e55f36STejun Heo * 1576c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1577c8e55f36STejun Heo * 1578c8e55f36STejun Heo * LOCKING: 1579d565ed63STejun Heo * spin_lock_irq(pool->lock). 1580c8e55f36STejun Heo */ 1581c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1582c8e55f36STejun Heo { 1583bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1584c8e55f36STejun Heo 15856183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15866183c009STejun Heo return; 1587d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1588bd7bdd43STejun Heo pool->nr_idle--; 1589c8e55f36STejun Heo list_del_init(&worker->entry); 1590c8e55f36STejun Heo } 1591c8e55f36STejun Heo 1592f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1593c34056a3STejun Heo { 1594c34056a3STejun Heo struct worker *worker; 1595c34056a3STejun Heo 1596f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1597c8e55f36STejun Heo if (worker) { 1598c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1599affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1600da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1601e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1602e22bee78STejun Heo worker->flags = WORKER_PREP; 1603c8e55f36STejun Heo } 1604c34056a3STejun Heo return worker; 1605c34056a3STejun Heo } 1606c34056a3STejun Heo 1607c34056a3STejun Heo /** 16084736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16094736cbf7SLai Jiangshan * @worker: worker to be attached 16104736cbf7SLai Jiangshan * @pool: the target pool 16114736cbf7SLai Jiangshan * 16124736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16134736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16144736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16154736cbf7SLai Jiangshan */ 16164736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16174736cbf7SLai Jiangshan struct worker_pool *pool) 16184736cbf7SLai Jiangshan { 16194736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16204736cbf7SLai Jiangshan 16214736cbf7SLai Jiangshan /* 16224736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16234736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16244736cbf7SLai Jiangshan */ 16254736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16264736cbf7SLai Jiangshan 16274736cbf7SLai Jiangshan /* 16284736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 16294736cbf7SLai Jiangshan * stable across this function. See the comments above the 16304736cbf7SLai Jiangshan * flag definition for details. 16314736cbf7SLai Jiangshan */ 16324736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 16334736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 16344736cbf7SLai Jiangshan 16354736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 16364736cbf7SLai Jiangshan 16374736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 16384736cbf7SLai Jiangshan } 16394736cbf7SLai Jiangshan 16404736cbf7SLai Jiangshan /** 164160f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 164260f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 164360f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 164460f5a4bcSLai Jiangshan * 16454736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 16464736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 16474736cbf7SLai Jiangshan * other reference to the pool. 164860f5a4bcSLai Jiangshan */ 164960f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 165060f5a4bcSLai Jiangshan struct worker_pool *pool) 165160f5a4bcSLai Jiangshan { 165260f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 165360f5a4bcSLai Jiangshan 165492f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1655da028469SLai Jiangshan list_del(&worker->node); 1656da028469SLai Jiangshan if (list_empty(&pool->workers)) 165760f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 165892f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 165960f5a4bcSLai Jiangshan 1660b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1661b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1662b62c0751SLai Jiangshan 166360f5a4bcSLai Jiangshan if (detach_completion) 166460f5a4bcSLai Jiangshan complete(detach_completion); 166560f5a4bcSLai Jiangshan } 166660f5a4bcSLai Jiangshan 166760f5a4bcSLai Jiangshan /** 1668c34056a3STejun Heo * create_worker - create a new workqueue worker 166963d95a91STejun Heo * @pool: pool the new worker will belong to 1670c34056a3STejun Heo * 1671051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1672c34056a3STejun Heo * 1673c34056a3STejun Heo * CONTEXT: 1674c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1675c34056a3STejun Heo * 1676d185af30SYacine Belkadi * Return: 1677c34056a3STejun Heo * Pointer to the newly created worker. 1678c34056a3STejun Heo */ 1679bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1680c34056a3STejun Heo { 1681c34056a3STejun Heo struct worker *worker = NULL; 1682f3421797STejun Heo int id = -1; 1683e3c916a4STejun Heo char id_buf[16]; 1684c34056a3STejun Heo 16857cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 16867cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1687822d8405STejun Heo if (id < 0) 1688c34056a3STejun Heo goto fail; 1689c34056a3STejun Heo 1690f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1691c34056a3STejun Heo if (!worker) 1692c34056a3STejun Heo goto fail; 1693c34056a3STejun Heo 1694bd7bdd43STejun Heo worker->pool = pool; 1695c34056a3STejun Heo worker->id = id; 1696c34056a3STejun Heo 169729c91e99STejun Heo if (pool->cpu >= 0) 1698e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1699e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1700f3421797STejun Heo else 1701e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1702e3c916a4STejun Heo 1703f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1704e3c916a4STejun Heo "kworker/%s", id_buf); 1705c34056a3STejun Heo if (IS_ERR(worker->task)) 1706c34056a3STejun Heo goto fail; 1707c34056a3STejun Heo 170891151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 170991151228SOleg Nesterov 171091151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 171191151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 171291151228SOleg Nesterov 1713da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17144736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1715822d8405STejun Heo 1716051e1850SLai Jiangshan /* start the newly created worker */ 1717051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1718051e1850SLai Jiangshan worker->pool->nr_workers++; 1719051e1850SLai Jiangshan worker_enter_idle(worker); 1720051e1850SLai Jiangshan wake_up_process(worker->task); 1721051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1722051e1850SLai Jiangshan 1723c34056a3STejun Heo return worker; 1724822d8405STejun Heo 1725c34056a3STejun Heo fail: 17269625ab17SLai Jiangshan if (id >= 0) 17277cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1728c34056a3STejun Heo kfree(worker); 1729c34056a3STejun Heo return NULL; 1730c34056a3STejun Heo } 1731c34056a3STejun Heo 1732c34056a3STejun Heo /** 1733c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1734c34056a3STejun Heo * @worker: worker to be destroyed 1735c34056a3STejun Heo * 173673eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 173773eb7fe7SLai Jiangshan * be idle. 1738c8e55f36STejun Heo * 1739c8e55f36STejun Heo * CONTEXT: 174060f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1741c34056a3STejun Heo */ 1742c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1743c34056a3STejun Heo { 1744bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1745c34056a3STejun Heo 1746cd549687STejun Heo lockdep_assert_held(&pool->lock); 1747cd549687STejun Heo 1748c34056a3STejun Heo /* sanity check frenzy */ 17496183c009STejun Heo if (WARN_ON(worker->current_work) || 175073eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 175173eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 17526183c009STejun Heo return; 1753c34056a3STejun Heo 1754bd7bdd43STejun Heo pool->nr_workers--; 1755bd7bdd43STejun Heo pool->nr_idle--; 1756c8e55f36STejun Heo 1757c8e55f36STejun Heo list_del_init(&worker->entry); 1758cb444766STejun Heo worker->flags |= WORKER_DIE; 175960f5a4bcSLai Jiangshan wake_up_process(worker->task); 1760c34056a3STejun Heo } 1761c34056a3STejun Heo 176263d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1763e22bee78STejun Heo { 176463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1765e22bee78STejun Heo 1766d565ed63STejun Heo spin_lock_irq(&pool->lock); 1767e22bee78STejun Heo 17683347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1769e22bee78STejun Heo struct worker *worker; 1770e22bee78STejun Heo unsigned long expires; 1771e22bee78STejun Heo 1772e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 177363d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1774e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1775e22bee78STejun Heo 17763347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 177763d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 17783347fc9fSLai Jiangshan break; 1779e22bee78STejun Heo } 17803347fc9fSLai Jiangshan 17813347fc9fSLai Jiangshan destroy_worker(worker); 1782e22bee78STejun Heo } 1783e22bee78STejun Heo 1784d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1785e22bee78STejun Heo } 1786e22bee78STejun Heo 1787493a1724STejun Heo static void send_mayday(struct work_struct *work) 1788e22bee78STejun Heo { 1789112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1790112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1791493a1724STejun Heo 17922e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1793e22bee78STejun Heo 1794493008a8STejun Heo if (!wq->rescuer) 1795493a1724STejun Heo return; 1796e22bee78STejun Heo 1797e22bee78STejun Heo /* mayday mayday mayday */ 1798493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 179977668c8bSLai Jiangshan /* 180077668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 180177668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 180277668c8bSLai Jiangshan * rescuer is done with it. 180377668c8bSLai Jiangshan */ 180477668c8bSLai Jiangshan get_pwq(pwq); 1805493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1806e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1807493a1724STejun Heo } 1808e22bee78STejun Heo } 1809e22bee78STejun Heo 1810706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1811e22bee78STejun Heo { 181263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1813e22bee78STejun Heo struct work_struct *work; 1814e22bee78STejun Heo 1815b2d82909STejun Heo spin_lock_irq(&pool->lock); 1816b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1817e22bee78STejun Heo 181863d95a91STejun Heo if (need_to_create_worker(pool)) { 1819e22bee78STejun Heo /* 1820e22bee78STejun Heo * We've been trying to create a new worker but 1821e22bee78STejun Heo * haven't been successful. We might be hitting an 1822e22bee78STejun Heo * allocation deadlock. Send distress signals to 1823e22bee78STejun Heo * rescuers. 1824e22bee78STejun Heo */ 182563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1826e22bee78STejun Heo send_mayday(work); 1827e22bee78STejun Heo } 1828e22bee78STejun Heo 1829b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1830b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1831e22bee78STejun Heo 183263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1833e22bee78STejun Heo } 1834e22bee78STejun Heo 1835e22bee78STejun Heo /** 1836e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 183763d95a91STejun Heo * @pool: pool to create a new worker for 1838e22bee78STejun Heo * 183963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1840e22bee78STejun Heo * have at least one idle worker on return from this function. If 1841e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 184263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1843e22bee78STejun Heo * possible allocation deadlock. 1844e22bee78STejun Heo * 1845c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1846c5aa87bbSTejun Heo * may_start_working() %true. 1847e22bee78STejun Heo * 1848e22bee78STejun Heo * LOCKING: 1849d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1850e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1851e22bee78STejun Heo * manager. 1852e22bee78STejun Heo */ 185329187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1854d565ed63STejun Heo __releases(&pool->lock) 1855d565ed63STejun Heo __acquires(&pool->lock) 1856e22bee78STejun Heo { 1857e22bee78STejun Heo restart: 1858d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18599f9c2364STejun Heo 1860e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 186163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1862e22bee78STejun Heo 1863e22bee78STejun Heo while (true) { 1864051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1865e22bee78STejun Heo break; 1866e22bee78STejun Heo 1867e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 18689f9c2364STejun Heo 186963d95a91STejun Heo if (!need_to_create_worker(pool)) 1870e22bee78STejun Heo break; 1871e22bee78STejun Heo } 1872e22bee78STejun Heo 187363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1874d565ed63STejun Heo spin_lock_irq(&pool->lock); 1875051e1850SLai Jiangshan /* 1876051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1877051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1878051e1850SLai Jiangshan * already become busy. 1879051e1850SLai Jiangshan */ 188063d95a91STejun Heo if (need_to_create_worker(pool)) 1881e22bee78STejun Heo goto restart; 1882e22bee78STejun Heo } 1883e22bee78STejun Heo 1884e22bee78STejun Heo /** 1885e22bee78STejun Heo * manage_workers - manage worker pool 1886e22bee78STejun Heo * @worker: self 1887e22bee78STejun Heo * 1888706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1889e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1890706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1891e22bee78STejun Heo * 1892e22bee78STejun Heo * The caller can safely start processing works on false return. On 1893e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1894e22bee78STejun Heo * and may_start_working() is true. 1895e22bee78STejun Heo * 1896e22bee78STejun Heo * CONTEXT: 1897d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1898e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1899e22bee78STejun Heo * 1900d185af30SYacine Belkadi * Return: 190129187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 190229187a9eSTejun Heo * start processing works, %true if management function was performed and 190329187a9eSTejun Heo * the conditions that the caller verified before calling the function may 190429187a9eSTejun Heo * no longer be true. 1905e22bee78STejun Heo */ 1906e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1907e22bee78STejun Heo { 190863d95a91STejun Heo struct worker_pool *pool = worker->pool; 1909e22bee78STejun Heo 1910bc3a1afcSTejun Heo /* 1911bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1912bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1913bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1914bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1915bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1916bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1917bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1918bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1919bc3a1afcSTejun Heo */ 192034a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 192129187a9eSTejun Heo return false; 19222607d7a6STejun Heo pool->manager = worker; 1923e22bee78STejun Heo 192429187a9eSTejun Heo maybe_create_worker(pool); 1925e22bee78STejun Heo 19262607d7a6STejun Heo pool->manager = NULL; 192734a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 192829187a9eSTejun Heo return true; 1929e22bee78STejun Heo } 1930e22bee78STejun Heo 1931a62428c0STejun Heo /** 1932a62428c0STejun Heo * process_one_work - process single work 1933c34056a3STejun Heo * @worker: self 1934a62428c0STejun Heo * @work: work to process 1935a62428c0STejun Heo * 1936a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1937a62428c0STejun Heo * process a single work including synchronization against and 1938a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1939a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1940a62428c0STejun Heo * call this function to process a work. 1941a62428c0STejun Heo * 1942a62428c0STejun Heo * CONTEXT: 1943d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1944a62428c0STejun Heo */ 1945c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 1946d565ed63STejun Heo __releases(&pool->lock) 1947d565ed63STejun Heo __acquires(&pool->lock) 19481da177e4SLinus Torvalds { 1949112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1950bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1951112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 195273f53c4aSTejun Heo int work_color; 19537e11629dSTejun Heo struct worker *collision; 19544e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 19554e6045f1SJohannes Berg /* 1956a62428c0STejun Heo * It is permissible to free the struct work_struct from 1957a62428c0STejun Heo * inside the function that is called from it, this we need to 1958a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1959a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1960a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 19614e6045f1SJohannes Berg */ 19624d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 19634d82a1deSPeter Zijlstra 19644d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 19654e6045f1SJohannes Berg #endif 1966807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 196785327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1968ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 196925511a47STejun Heo 19707e11629dSTejun Heo /* 19717e11629dSTejun Heo * A single work shouldn't be executed concurrently by 19727e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 19737e11629dSTejun Heo * already processing the work. If so, defer the work to the 19747e11629dSTejun Heo * currently executing one. 19757e11629dSTejun Heo */ 1976c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 19777e11629dSTejun Heo if (unlikely(collision)) { 19787e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 19797e11629dSTejun Heo return; 19807e11629dSTejun Heo } 19811da177e4SLinus Torvalds 19828930cabaSTejun Heo /* claim and dequeue */ 19831da177e4SLinus Torvalds debug_work_deactivate(work); 1984c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 1985c34056a3STejun Heo worker->current_work = work; 1986a2c1c57bSTejun Heo worker->current_func = work->func; 1987112202d9STejun Heo worker->current_pwq = pwq; 198873f53c4aSTejun Heo work_color = get_work_color(work); 19897a22ad75STejun Heo 1990a62428c0STejun Heo list_del_init(&work->entry); 1991a62428c0STejun Heo 1992649027d7STejun Heo /* 1993228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 1994228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 1995228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 1996228f1d00SLai Jiangshan * execution of the pending work items. 1997fb0e7bebSTejun Heo */ 1998fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 1999228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2000fb0e7bebSTejun Heo 2001974271c4STejun Heo /* 2002a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2003a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2004a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2005a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2006228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2007974271c4STejun Heo */ 2008a489a03eSLai Jiangshan if (need_more_worker(pool)) 200963d95a91STejun Heo wake_up_worker(pool); 2010974271c4STejun Heo 20118930cabaSTejun Heo /* 20127c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2013d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 201423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 201523657bb1STejun Heo * disabled. 20168930cabaSTejun Heo */ 20177c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20181da177e4SLinus Torvalds 2019d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2020365970a1SDavid Howells 2021112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20223295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2023e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2024a2c1c57bSTejun Heo worker->current_func(work); 2025e36c886aSArjan van de Ven /* 2026e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2027e36c886aSArjan van de Ven * point will only record its address. 2028e36c886aSArjan van de Ven */ 2029e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20303295f0efSIngo Molnar lock_map_release(&lockdep_map); 2031112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 20321da177e4SLinus Torvalds 2033d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2034044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2035044c782cSValentin Ilie " last function: %pf\n", 2036a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2037a2c1c57bSTejun Heo worker->current_func); 2038d5abe669SPeter Zijlstra debug_show_held_locks(current); 2039d5abe669SPeter Zijlstra dump_stack(); 2040d5abe669SPeter Zijlstra } 2041d5abe669SPeter Zijlstra 2042b22ce278STejun Heo /* 2043b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2044b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2045b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2046b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2047789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2048789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2049b22ce278STejun Heo */ 20503e28e377SJoe Lawrence cond_resched_rcu_qs(); 2051b22ce278STejun Heo 2052d565ed63STejun Heo spin_lock_irq(&pool->lock); 2053a62428c0STejun Heo 2054fb0e7bebSTejun Heo /* clear cpu intensive status */ 2055fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2056fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2057fb0e7bebSTejun Heo 2058a62428c0STejun Heo /* we're done with it, release */ 205942f8570fSSasha Levin hash_del(&worker->hentry); 2060c34056a3STejun Heo worker->current_work = NULL; 2061a2c1c57bSTejun Heo worker->current_func = NULL; 2062112202d9STejun Heo worker->current_pwq = NULL; 20633d1cb205STejun Heo worker->desc_valid = false; 2064112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 20651da177e4SLinus Torvalds } 20661da177e4SLinus Torvalds 2067affee4b2STejun Heo /** 2068affee4b2STejun Heo * process_scheduled_works - process scheduled works 2069affee4b2STejun Heo * @worker: self 2070affee4b2STejun Heo * 2071affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2072affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2073affee4b2STejun Heo * fetches a work from the top and executes it. 2074affee4b2STejun Heo * 2075affee4b2STejun Heo * CONTEXT: 2076d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2077affee4b2STejun Heo * multiple times. 2078affee4b2STejun Heo */ 2079affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 20801da177e4SLinus Torvalds { 2081affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2082affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2083a62428c0STejun Heo struct work_struct, entry); 2084c34056a3STejun Heo process_one_work(worker, work); 2085a62428c0STejun Heo } 20861da177e4SLinus Torvalds } 20871da177e4SLinus Torvalds 20884690c4abSTejun Heo /** 20894690c4abSTejun Heo * worker_thread - the worker thread function 2090c34056a3STejun Heo * @__worker: self 20914690c4abSTejun Heo * 2092c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2093c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2094c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2095c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2096c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2097d185af30SYacine Belkadi * 2098d185af30SYacine Belkadi * Return: 0 20994690c4abSTejun Heo */ 2100c34056a3STejun Heo static int worker_thread(void *__worker) 21011da177e4SLinus Torvalds { 2102c34056a3STejun Heo struct worker *worker = __worker; 2103bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21041da177e4SLinus Torvalds 2105e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2106e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2107c8e55f36STejun Heo woke_up: 2108d565ed63STejun Heo spin_lock_irq(&pool->lock); 2109affee4b2STejun Heo 2110a9ab775bSTejun Heo /* am I supposed to die? */ 2111a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2112d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2113a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2114e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 211560f5a4bcSLai Jiangshan 211660f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21177cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 211860f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 211960f5a4bcSLai Jiangshan kfree(worker); 2120c8e55f36STejun Heo return 0; 2121c8e55f36STejun Heo } 2122c8e55f36STejun Heo 2123c8e55f36STejun Heo worker_leave_idle(worker); 2124db7bccf4STejun Heo recheck: 2125e22bee78STejun Heo /* no more worker necessary? */ 212663d95a91STejun Heo if (!need_more_worker(pool)) 2127e22bee78STejun Heo goto sleep; 2128e22bee78STejun Heo 2129e22bee78STejun Heo /* do we need to manage? */ 213063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2131e22bee78STejun Heo goto recheck; 2132e22bee78STejun Heo 2133c8e55f36STejun Heo /* 2134c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2135c8e55f36STejun Heo * preparing to process a work or actually processing it. 2136c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2137c8e55f36STejun Heo */ 21386183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2139c8e55f36STejun Heo 2140e22bee78STejun Heo /* 2141a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2142a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2143a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2144a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2145a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2146e22bee78STejun Heo */ 2147a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2148e22bee78STejun Heo 2149e22bee78STejun Heo do { 2150affee4b2STejun Heo struct work_struct *work = 2151bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2152affee4b2STejun Heo struct work_struct, entry); 2153affee4b2STejun Heo 2154c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2155affee4b2STejun Heo /* optimization path, not strictly necessary */ 2156affee4b2STejun Heo process_one_work(worker, work); 2157affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2158affee4b2STejun Heo process_scheduled_works(worker); 2159affee4b2STejun Heo } else { 2160c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2161affee4b2STejun Heo process_scheduled_works(worker); 2162affee4b2STejun Heo } 216363d95a91STejun Heo } while (keep_working(pool)); 2164affee4b2STejun Heo 2165228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2166d313dd85STejun Heo sleep: 2167c8e55f36STejun Heo /* 2168d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2169d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2170d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2171d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2172d565ed63STejun Heo * event. 2173c8e55f36STejun Heo */ 2174c8e55f36STejun Heo worker_enter_idle(worker); 2175c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2176d565ed63STejun Heo spin_unlock_irq(&pool->lock); 21771da177e4SLinus Torvalds schedule(); 2178c8e55f36STejun Heo goto woke_up; 21791da177e4SLinus Torvalds } 21801da177e4SLinus Torvalds 2181e22bee78STejun Heo /** 2182e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2183111c225aSTejun Heo * @__rescuer: self 2184e22bee78STejun Heo * 2185e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2186493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2187e22bee78STejun Heo * 2188706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2189e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2190e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2191e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2192e22bee78STejun Heo * the problem rescuer solves. 2193e22bee78STejun Heo * 2194706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2195706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2196e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2197e22bee78STejun Heo * 2198e22bee78STejun Heo * This should happen rarely. 2199d185af30SYacine Belkadi * 2200d185af30SYacine Belkadi * Return: 0 2201e22bee78STejun Heo */ 2202111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2203e22bee78STejun Heo { 2204111c225aSTejun Heo struct worker *rescuer = __rescuer; 2205111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2206e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22074d595b86SLai Jiangshan bool should_stop; 2208e22bee78STejun Heo 2209e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2210111c225aSTejun Heo 2211111c225aSTejun Heo /* 2212111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2213111c225aSTejun Heo * doesn't participate in concurrency management. 2214111c225aSTejun Heo */ 2215111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2216e22bee78STejun Heo repeat: 2217e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22181da177e4SLinus Torvalds 22194d595b86SLai Jiangshan /* 22204d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22214d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22224d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22234d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22244d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22254d595b86SLai Jiangshan * list is always empty on exit. 22264d595b86SLai Jiangshan */ 22274d595b86SLai Jiangshan should_stop = kthread_should_stop(); 22281da177e4SLinus Torvalds 2229493a1724STejun Heo /* see whether any pwq is asking for help */ 22302e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2231493a1724STejun Heo 2232493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2233493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2234493a1724STejun Heo struct pool_workqueue, mayday_node); 2235112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2236e22bee78STejun Heo struct work_struct *work, *n; 2237e22bee78STejun Heo 2238e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2239493a1724STejun Heo list_del_init(&pwq->mayday_node); 2240493a1724STejun Heo 22412e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2242e22bee78STejun Heo 224351697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 224451697d39SLai Jiangshan 224551697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2246b3104104SLai Jiangshan rescuer->pool = pool; 2247e22bee78STejun Heo 2248e22bee78STejun Heo /* 2249e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2250e22bee78STejun Heo * process'em. 2251e22bee78STejun Heo */ 22520479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 2253bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2254112202d9STejun Heo if (get_work_pwq(work) == pwq) 2255e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2256e22bee78STejun Heo 2257008847f6SNeilBrown if (!list_empty(scheduled)) { 2258e22bee78STejun Heo process_scheduled_works(rescuer); 22597576958aSTejun Heo 22607576958aSTejun Heo /* 2261008847f6SNeilBrown * The above execution of rescued work items could 2262008847f6SNeilBrown * have created more to rescue through 2263008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2264008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2265008847f6SNeilBrown * that such back-to-back work items, which may be 2266008847f6SNeilBrown * being used to relieve memory pressure, don't 2267008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2268008847f6SNeilBrown */ 2269008847f6SNeilBrown if (need_to_create_worker(pool)) { 2270008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2271008847f6SNeilBrown get_pwq(pwq); 2272008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2273008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2274008847f6SNeilBrown } 2275008847f6SNeilBrown } 2276008847f6SNeilBrown 2277008847f6SNeilBrown /* 227877668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 227913b1d625SLai Jiangshan * go away while we're still attached to it. 228077668c8bSLai Jiangshan */ 228177668c8bSLai Jiangshan put_pwq(pwq); 228277668c8bSLai Jiangshan 228377668c8bSLai Jiangshan /* 2284d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 22857576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 22867576958aSTejun Heo * and stalling the execution. 22877576958aSTejun Heo */ 2288d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 228963d95a91STejun Heo wake_up_worker(pool); 22907576958aSTejun Heo 2291b3104104SLai Jiangshan rescuer->pool = NULL; 229213b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 229313b1d625SLai Jiangshan 229413b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 229513b1d625SLai Jiangshan 229613b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 22971da177e4SLinus Torvalds } 22981da177e4SLinus Torvalds 22992e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2300493a1724STejun Heo 23014d595b86SLai Jiangshan if (should_stop) { 23024d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23034d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23044d595b86SLai Jiangshan return 0; 23054d595b86SLai Jiangshan } 23064d595b86SLai Jiangshan 2307111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2308111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2309e22bee78STejun Heo schedule(); 2310e22bee78STejun Heo goto repeat; 23111da177e4SLinus Torvalds } 23121da177e4SLinus Torvalds 2313fc2e4d70SOleg Nesterov struct wq_barrier { 2314fc2e4d70SOleg Nesterov struct work_struct work; 2315fc2e4d70SOleg Nesterov struct completion done; 23162607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2317fc2e4d70SOleg Nesterov }; 2318fc2e4d70SOleg Nesterov 2319fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2320fc2e4d70SOleg Nesterov { 2321fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2322fc2e4d70SOleg Nesterov complete(&barr->done); 2323fc2e4d70SOleg Nesterov } 2324fc2e4d70SOleg Nesterov 23254690c4abSTejun Heo /** 23264690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2327112202d9STejun Heo * @pwq: pwq to insert barrier into 23284690c4abSTejun Heo * @barr: wq_barrier to insert 2329affee4b2STejun Heo * @target: target work to attach @barr to 2330affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23314690c4abSTejun Heo * 2332affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2333affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2334affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2335affee4b2STejun Heo * cpu. 2336affee4b2STejun Heo * 2337affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2338affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2339affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2340affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2341affee4b2STejun Heo * after a work with LINKED flag set. 2342affee4b2STejun Heo * 2343affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2344112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23454690c4abSTejun Heo * 23464690c4abSTejun Heo * CONTEXT: 2347d565ed63STejun Heo * spin_lock_irq(pool->lock). 23484690c4abSTejun Heo */ 2349112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2350affee4b2STejun Heo struct wq_barrier *barr, 2351affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2352fc2e4d70SOleg Nesterov { 2353affee4b2STejun Heo struct list_head *head; 2354affee4b2STejun Heo unsigned int linked = 0; 2355affee4b2STejun Heo 2356dc186ad7SThomas Gleixner /* 2357d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2358dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2359dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2360dc186ad7SThomas Gleixner * might deadlock. 2361dc186ad7SThomas Gleixner */ 2362ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 236322df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2364fc2e4d70SOleg Nesterov init_completion(&barr->done); 23652607d7a6STejun Heo barr->task = current; 236683c22520SOleg Nesterov 2367affee4b2STejun Heo /* 2368affee4b2STejun Heo * If @target is currently being executed, schedule the 2369affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2370affee4b2STejun Heo */ 2371affee4b2STejun Heo if (worker) 2372affee4b2STejun Heo head = worker->scheduled.next; 2373affee4b2STejun Heo else { 2374affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2375affee4b2STejun Heo 2376affee4b2STejun Heo head = target->entry.next; 2377affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2378affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2379affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2380affee4b2STejun Heo } 2381affee4b2STejun Heo 2382dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2383112202d9STejun Heo insert_work(pwq, &barr->work, head, 2384affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2385fc2e4d70SOleg Nesterov } 2386fc2e4d70SOleg Nesterov 238773f53c4aSTejun Heo /** 2388112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 238973f53c4aSTejun Heo * @wq: workqueue being flushed 239073f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 239173f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 239273f53c4aSTejun Heo * 2393112202d9STejun Heo * Prepare pwqs for workqueue flushing. 239473f53c4aSTejun Heo * 2395112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2396112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2397112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2398112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2399112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 240073f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 240173f53c4aSTejun Heo * 240273f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 240373f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 240473f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 240573f53c4aSTejun Heo * is returned. 240673f53c4aSTejun Heo * 2407112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 240873f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 240973f53c4aSTejun Heo * advanced to @work_color. 241073f53c4aSTejun Heo * 241173f53c4aSTejun Heo * CONTEXT: 24123c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 241373f53c4aSTejun Heo * 2414d185af30SYacine Belkadi * Return: 241573f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 241673f53c4aSTejun Heo * otherwise. 241773f53c4aSTejun Heo */ 2418112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 241973f53c4aSTejun Heo int flush_color, int work_color) 24201da177e4SLinus Torvalds { 242173f53c4aSTejun Heo bool wait = false; 242249e3cf44STejun Heo struct pool_workqueue *pwq; 24231da177e4SLinus Torvalds 242473f53c4aSTejun Heo if (flush_color >= 0) { 24256183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2426112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2427dc186ad7SThomas Gleixner } 242814441960SOleg Nesterov 242949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2430112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24311da177e4SLinus Torvalds 2432b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 243373f53c4aSTejun Heo 243473f53c4aSTejun Heo if (flush_color >= 0) { 24356183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 243673f53c4aSTejun Heo 2437112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2438112202d9STejun Heo pwq->flush_color = flush_color; 2439112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 244073f53c4aSTejun Heo wait = true; 24411da177e4SLinus Torvalds } 244273f53c4aSTejun Heo } 244373f53c4aSTejun Heo 244473f53c4aSTejun Heo if (work_color >= 0) { 24456183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2446112202d9STejun Heo pwq->work_color = work_color; 244773f53c4aSTejun Heo } 244873f53c4aSTejun Heo 2449b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 24501da177e4SLinus Torvalds } 24511da177e4SLinus Torvalds 2452112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 245373f53c4aSTejun Heo complete(&wq->first_flusher->done); 245473f53c4aSTejun Heo 245573f53c4aSTejun Heo return wait; 245683c22520SOleg Nesterov } 24571da177e4SLinus Torvalds 24580fcb78c2SRolf Eike Beer /** 24591da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 24600fcb78c2SRolf Eike Beer * @wq: workqueue to flush 24611da177e4SLinus Torvalds * 2462c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2463c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 24641da177e4SLinus Torvalds */ 24657ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 24661da177e4SLinus Torvalds { 246773f53c4aSTejun Heo struct wq_flusher this_flusher = { 246873f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 246973f53c4aSTejun Heo .flush_color = -1, 247073f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 247173f53c4aSTejun Heo }; 247273f53c4aSTejun Heo int next_color; 2473b1f4ec17SOleg Nesterov 24743295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 24753295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 247673f53c4aSTejun Heo 24773c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 247873f53c4aSTejun Heo 247973f53c4aSTejun Heo /* 248073f53c4aSTejun Heo * Start-to-wait phase 248173f53c4aSTejun Heo */ 248273f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 248373f53c4aSTejun Heo 248473f53c4aSTejun Heo if (next_color != wq->flush_color) { 248573f53c4aSTejun Heo /* 248673f53c4aSTejun Heo * Color space is not full. The current work_color 248773f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 248873f53c4aSTejun Heo * by one. 248973f53c4aSTejun Heo */ 24906183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 249173f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 249273f53c4aSTejun Heo wq->work_color = next_color; 249373f53c4aSTejun Heo 249473f53c4aSTejun Heo if (!wq->first_flusher) { 249573f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 24966183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 249773f53c4aSTejun Heo 249873f53c4aSTejun Heo wq->first_flusher = &this_flusher; 249973f53c4aSTejun Heo 2500112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 250173f53c4aSTejun Heo wq->work_color)) { 250273f53c4aSTejun Heo /* nothing to flush, done */ 250373f53c4aSTejun Heo wq->flush_color = next_color; 250473f53c4aSTejun Heo wq->first_flusher = NULL; 250573f53c4aSTejun Heo goto out_unlock; 250673f53c4aSTejun Heo } 250773f53c4aSTejun Heo } else { 250873f53c4aSTejun Heo /* wait in queue */ 25096183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 251073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2511112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 251273f53c4aSTejun Heo } 251373f53c4aSTejun Heo } else { 251473f53c4aSTejun Heo /* 251573f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 251673f53c4aSTejun Heo * The next flush completion will assign us 251773f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 251873f53c4aSTejun Heo */ 251973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 252073f53c4aSTejun Heo } 252173f53c4aSTejun Heo 25223c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 252373f53c4aSTejun Heo 252473f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 252573f53c4aSTejun Heo 252673f53c4aSTejun Heo /* 252773f53c4aSTejun Heo * Wake-up-and-cascade phase 252873f53c4aSTejun Heo * 252973f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 253073f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 253173f53c4aSTejun Heo */ 253273f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 253373f53c4aSTejun Heo return; 253473f53c4aSTejun Heo 25353c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 253673f53c4aSTejun Heo 25374ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25384ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25394ce48b37STejun Heo goto out_unlock; 25404ce48b37STejun Heo 254173f53c4aSTejun Heo wq->first_flusher = NULL; 254273f53c4aSTejun Heo 25436183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25446183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 254573f53c4aSTejun Heo 254673f53c4aSTejun Heo while (true) { 254773f53c4aSTejun Heo struct wq_flusher *next, *tmp; 254873f53c4aSTejun Heo 254973f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 255073f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 255173f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 255273f53c4aSTejun Heo break; 255373f53c4aSTejun Heo list_del_init(&next->list); 255473f53c4aSTejun Heo complete(&next->done); 255573f53c4aSTejun Heo } 255673f53c4aSTejun Heo 25576183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 255873f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 255973f53c4aSTejun Heo 256073f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 256173f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 256273f53c4aSTejun Heo 256373f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 256473f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 256573f53c4aSTejun Heo /* 256673f53c4aSTejun Heo * Assign the same color to all overflowed 256773f53c4aSTejun Heo * flushers, advance work_color and append to 256873f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 256973f53c4aSTejun Heo * phase for these overflowed flushers. 257073f53c4aSTejun Heo */ 257173f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 257273f53c4aSTejun Heo tmp->flush_color = wq->work_color; 257373f53c4aSTejun Heo 257473f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 257573f53c4aSTejun Heo 257673f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 257773f53c4aSTejun Heo &wq->flusher_queue); 2578112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 257973f53c4aSTejun Heo } 258073f53c4aSTejun Heo 258173f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 25826183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 258373f53c4aSTejun Heo break; 258473f53c4aSTejun Heo } 258573f53c4aSTejun Heo 258673f53c4aSTejun Heo /* 258773f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2588112202d9STejun Heo * the new first flusher and arm pwqs. 258973f53c4aSTejun Heo */ 25906183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 25916183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 259273f53c4aSTejun Heo 259373f53c4aSTejun Heo list_del_init(&next->list); 259473f53c4aSTejun Heo wq->first_flusher = next; 259573f53c4aSTejun Heo 2596112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 259773f53c4aSTejun Heo break; 259873f53c4aSTejun Heo 259973f53c4aSTejun Heo /* 260073f53c4aSTejun Heo * Meh... this color is already done, clear first 260173f53c4aSTejun Heo * flusher and repeat cascading. 260273f53c4aSTejun Heo */ 260373f53c4aSTejun Heo wq->first_flusher = NULL; 260473f53c4aSTejun Heo } 260573f53c4aSTejun Heo 260673f53c4aSTejun Heo out_unlock: 26073c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26081da177e4SLinus Torvalds } 2609ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26101da177e4SLinus Torvalds 26119c5a2ba7STejun Heo /** 26129c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26139c5a2ba7STejun Heo * @wq: workqueue to drain 26149c5a2ba7STejun Heo * 26159c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26169c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26179c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26189c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26199c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26209c5a2ba7STejun Heo * takes too long. 26219c5a2ba7STejun Heo */ 26229c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26239c5a2ba7STejun Heo { 26249c5a2ba7STejun Heo unsigned int flush_cnt = 0; 262549e3cf44STejun Heo struct pool_workqueue *pwq; 26269c5a2ba7STejun Heo 26279c5a2ba7STejun Heo /* 26289c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26299c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2630618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26319c5a2ba7STejun Heo */ 263287fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26339c5a2ba7STejun Heo if (!wq->nr_drainers++) 2634618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 263587fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26369c5a2ba7STejun Heo reflush: 26379c5a2ba7STejun Heo flush_workqueue(wq); 26389c5a2ba7STejun Heo 2639b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 264076af4d93STejun Heo 264149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2642fa2563e4SThomas Tuttle bool drained; 26439c5a2ba7STejun Heo 2644b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2645112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2646b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2647fa2563e4SThomas Tuttle 2648fa2563e4SThomas Tuttle if (drained) 26499c5a2ba7STejun Heo continue; 26509c5a2ba7STejun Heo 26519c5a2ba7STejun Heo if (++flush_cnt == 10 || 26529c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2653c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 26549c5a2ba7STejun Heo wq->name, flush_cnt); 265576af4d93STejun Heo 2656b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 26579c5a2ba7STejun Heo goto reflush; 26589c5a2ba7STejun Heo } 26599c5a2ba7STejun Heo 26609c5a2ba7STejun Heo if (!--wq->nr_drainers) 2661618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 266287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26639c5a2ba7STejun Heo } 26649c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 26659c5a2ba7STejun Heo 2666606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2667baf59022STejun Heo { 2668baf59022STejun Heo struct worker *worker = NULL; 2669c9e7cf27STejun Heo struct worker_pool *pool; 2670112202d9STejun Heo struct pool_workqueue *pwq; 2671baf59022STejun Heo 2672baf59022STejun Heo might_sleep(); 2673baf59022STejun Heo 2674fa1b54e6STejun Heo local_irq_disable(); 2675fa1b54e6STejun Heo pool = get_work_pool(work); 2676fa1b54e6STejun Heo if (!pool) { 2677fa1b54e6STejun Heo local_irq_enable(); 2678fa1b54e6STejun Heo return false; 2679fa1b54e6STejun Heo } 2680fa1b54e6STejun Heo 2681fa1b54e6STejun Heo spin_lock(&pool->lock); 26820b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2683112202d9STejun Heo pwq = get_work_pwq(work); 2684112202d9STejun Heo if (pwq) { 2685112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2686baf59022STejun Heo goto already_gone; 2687606a5020STejun Heo } else { 2688c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2689baf59022STejun Heo if (!worker) 2690baf59022STejun Heo goto already_gone; 2691112202d9STejun Heo pwq = worker->current_pwq; 2692606a5020STejun Heo } 2693baf59022STejun Heo 2694112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2695d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2696baf59022STejun Heo 2697e159489bSTejun Heo /* 2698e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2699e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2700e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2701e159489bSTejun Heo * access. 2702e159489bSTejun Heo */ 2703493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2704112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2705e159489bSTejun Heo else 2706112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2707112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2708e159489bSTejun Heo 2709baf59022STejun Heo return true; 2710baf59022STejun Heo already_gone: 2711d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2712baf59022STejun Heo return false; 2713baf59022STejun Heo } 2714baf59022STejun Heo 2715db700897SOleg Nesterov /** 2716401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2717401a8d04STejun Heo * @work: the work to flush 2718db700897SOleg Nesterov * 2719606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2720606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2721401a8d04STejun Heo * 2722d185af30SYacine Belkadi * Return: 2723401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2724401a8d04STejun Heo * %false if it was already idle. 2725db700897SOleg Nesterov */ 2726401a8d04STejun Heo bool flush_work(struct work_struct *work) 2727db700897SOleg Nesterov { 272812997d1aSBjorn Helgaas struct wq_barrier barr; 272912997d1aSBjorn Helgaas 27300976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27310976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27320976dfc1SStephen Boyd 273312997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 273412997d1aSBjorn Helgaas wait_for_completion(&barr.done); 273512997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 273612997d1aSBjorn Helgaas return true; 273712997d1aSBjorn Helgaas } else { 273812997d1aSBjorn Helgaas return false; 273912997d1aSBjorn Helgaas } 2740606a5020STejun Heo } 2741db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2742db700897SOleg Nesterov 27438603e1b3STejun Heo struct cwt_wait { 27448603e1b3STejun Heo wait_queue_t wait; 27458603e1b3STejun Heo struct work_struct *work; 27468603e1b3STejun Heo }; 27478603e1b3STejun Heo 27488603e1b3STejun Heo static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) 27498603e1b3STejun Heo { 27508603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 27518603e1b3STejun Heo 27528603e1b3STejun Heo if (cwait->work != key) 27538603e1b3STejun Heo return 0; 27548603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 27558603e1b3STejun Heo } 27568603e1b3STejun Heo 275736e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2758401a8d04STejun Heo { 27598603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2760bbb68dfaSTejun Heo unsigned long flags; 27611f1f642eSOleg Nesterov int ret; 27621f1f642eSOleg Nesterov 27631f1f642eSOleg Nesterov do { 2764bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2765bbb68dfaSTejun Heo /* 27668603e1b3STejun Heo * If someone else is already canceling, wait for it to 27678603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 27688603e1b3STejun Heo * because we may get scheduled between @work's completion 27698603e1b3STejun Heo * and the other canceling task resuming and clearing 27708603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 27718603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 27728603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 27738603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 27748603e1b3STejun Heo * we're hogging the CPU. 27758603e1b3STejun Heo * 27768603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 27778603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 27788603e1b3STejun Heo * wake function which matches @work along with exclusive 27798603e1b3STejun Heo * wait and wakeup. 2780bbb68dfaSTejun Heo */ 27818603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 27828603e1b3STejun Heo struct cwt_wait cwait; 27838603e1b3STejun Heo 27848603e1b3STejun Heo init_wait(&cwait.wait); 27858603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 27868603e1b3STejun Heo cwait.work = work; 27878603e1b3STejun Heo 27888603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 27898603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 27908603e1b3STejun Heo if (work_is_canceling(work)) 27918603e1b3STejun Heo schedule(); 27928603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 27938603e1b3STejun Heo } 27941f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 27951f1f642eSOleg Nesterov 2796bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2797bbb68dfaSTejun Heo mark_work_canceling(work); 2798bbb68dfaSTejun Heo local_irq_restore(flags); 2799bbb68dfaSTejun Heo 2800606a5020STejun Heo flush_work(work); 28017a22ad75STejun Heo clear_work_data(work); 28028603e1b3STejun Heo 28038603e1b3STejun Heo /* 28048603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 28058603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 28068603e1b3STejun Heo * visible there. 28078603e1b3STejun Heo */ 28088603e1b3STejun Heo smp_mb(); 28098603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 28108603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 28118603e1b3STejun Heo 28121f1f642eSOleg Nesterov return ret; 28131f1f642eSOleg Nesterov } 28141f1f642eSOleg Nesterov 28156e84d644SOleg Nesterov /** 2816401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2817401a8d04STejun Heo * @work: the work to cancel 28186e84d644SOleg Nesterov * 2819401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2820401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2821401a8d04STejun Heo * another workqueue. On return from this function, @work is 2822401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28231f1f642eSOleg Nesterov * 2824401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2825401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28266e84d644SOleg Nesterov * 2827401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28286e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2829401a8d04STejun Heo * 2830d185af30SYacine Belkadi * Return: 2831401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28326e84d644SOleg Nesterov */ 2833401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28346e84d644SOleg Nesterov { 283536e227d2STejun Heo return __cancel_work_timer(work, false); 2836b89deed3SOleg Nesterov } 283728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2838b89deed3SOleg Nesterov 28396e84d644SOleg Nesterov /** 2840401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2841401a8d04STejun Heo * @dwork: the delayed work to flush 28426e84d644SOleg Nesterov * 2843401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2844401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2845401a8d04STejun Heo * considers the last queueing instance of @dwork. 28461f1f642eSOleg Nesterov * 2847d185af30SYacine Belkadi * Return: 2848401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2849401a8d04STejun Heo * %false if it was already idle. 28506e84d644SOleg Nesterov */ 2851401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2852401a8d04STejun Heo { 28538930cabaSTejun Heo local_irq_disable(); 2854401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 285560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28568930cabaSTejun Heo local_irq_enable(); 2857401a8d04STejun Heo return flush_work(&dwork->work); 2858401a8d04STejun Heo } 2859401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2860401a8d04STejun Heo 2861401a8d04STejun Heo /** 286257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 286357b30ae7STejun Heo * @dwork: delayed_work to cancel 286409383498STejun Heo * 2865d185af30SYacine Belkadi * Kill off a pending delayed_work. 2866d185af30SYacine Belkadi * 2867d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2868d185af30SYacine Belkadi * pending. 2869d185af30SYacine Belkadi * 2870d185af30SYacine Belkadi * Note: 2871d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2872d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2873d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 287409383498STejun Heo * 287557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 287609383498STejun Heo */ 287757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 287809383498STejun Heo { 287957b30ae7STejun Heo unsigned long flags; 288057b30ae7STejun Heo int ret; 288157b30ae7STejun Heo 288257b30ae7STejun Heo do { 288357b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 288457b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 288557b30ae7STejun Heo 288657b30ae7STejun Heo if (unlikely(ret < 0)) 288757b30ae7STejun Heo return false; 288857b30ae7STejun Heo 28897c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28907c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 289157b30ae7STejun Heo local_irq_restore(flags); 2892c0158ca6SDan Magenheimer return ret; 289309383498STejun Heo } 289457b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 289509383498STejun Heo 289609383498STejun Heo /** 2897401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2898401a8d04STejun Heo * @dwork: the delayed work cancel 2899401a8d04STejun Heo * 2900401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2901401a8d04STejun Heo * 2902d185af30SYacine Belkadi * Return: 2903401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2904401a8d04STejun Heo */ 2905401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29066e84d644SOleg Nesterov { 290736e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29086e84d644SOleg Nesterov } 2909f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29101da177e4SLinus Torvalds 29110fcb78c2SRolf Eike Beer /** 291231ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2913b6136773SAndrew Morton * @func: the function to call 2914b6136773SAndrew Morton * 291531ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 291631ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2917b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 291831ddd871STejun Heo * 2919d185af30SYacine Belkadi * Return: 292031ddd871STejun Heo * 0 on success, -errno on failure. 2921b6136773SAndrew Morton */ 292265f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 292315316ba8SChristoph Lameter { 292415316ba8SChristoph Lameter int cpu; 292538f51568SNamhyung Kim struct work_struct __percpu *works; 292615316ba8SChristoph Lameter 2927b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2928b6136773SAndrew Morton if (!works) 292915316ba8SChristoph Lameter return -ENOMEM; 2930b6136773SAndrew Morton 293195402b38SGautham R Shenoy get_online_cpus(); 293293981800STejun Heo 293315316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29349bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29359bfb1839SIngo Molnar 29369bfb1839SIngo Molnar INIT_WORK(work, func); 29378de6d308SOleg Nesterov schedule_work_on(cpu, work); 293815316ba8SChristoph Lameter } 293993981800STejun Heo 294093981800STejun Heo for_each_online_cpu(cpu) 29418616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 294293981800STejun Heo 294395402b38SGautham R Shenoy put_online_cpus(); 2944b6136773SAndrew Morton free_percpu(works); 294515316ba8SChristoph Lameter return 0; 294615316ba8SChristoph Lameter } 294715316ba8SChristoph Lameter 2948eef6a7d5SAlan Stern /** 2949eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2950eef6a7d5SAlan Stern * 2951eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2952eef6a7d5SAlan Stern * completion. 2953eef6a7d5SAlan Stern * 2954eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2955eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2956eef6a7d5SAlan Stern * will lead to deadlock: 2957eef6a7d5SAlan Stern * 2958eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2959eef6a7d5SAlan Stern * a lock held by your code or its caller. 2960eef6a7d5SAlan Stern * 2961eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2962eef6a7d5SAlan Stern * 2963eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2964eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2965eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2966eef6a7d5SAlan Stern * 2967eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2968eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2969eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2970eef6a7d5SAlan Stern * cancel_work_sync() instead. 2971eef6a7d5SAlan Stern */ 29721da177e4SLinus Torvalds void flush_scheduled_work(void) 29731da177e4SLinus Torvalds { 2974d320c038STejun Heo flush_workqueue(system_wq); 29751da177e4SLinus Torvalds } 2976ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29771da177e4SLinus Torvalds 29781da177e4SLinus Torvalds /** 29791fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29801fa44ecaSJames Bottomley * @fn: the function to execute 29811fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29821fa44ecaSJames Bottomley * be available when the work executes) 29831fa44ecaSJames Bottomley * 29841fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29851fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29861fa44ecaSJames Bottomley * 2987d185af30SYacine Belkadi * Return: 0 - function was executed 29881fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29891fa44ecaSJames Bottomley */ 299065f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29911fa44ecaSJames Bottomley { 29921fa44ecaSJames Bottomley if (!in_interrupt()) { 299365f27f38SDavid Howells fn(&ew->work); 29941fa44ecaSJames Bottomley return 0; 29951fa44ecaSJames Bottomley } 29961fa44ecaSJames Bottomley 299765f27f38SDavid Howells INIT_WORK(&ew->work, fn); 29981fa44ecaSJames Bottomley schedule_work(&ew->work); 29991fa44ecaSJames Bottomley 30001fa44ecaSJames Bottomley return 1; 30011fa44ecaSJames Bottomley } 30021fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30031fa44ecaSJames Bottomley 3004226223abSTejun Heo #ifdef CONFIG_SYSFS 3005226223abSTejun Heo /* 3006226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3007226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3008226223abSTejun Heo * following attributes. 3009226223abSTejun Heo * 3010226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3011226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3012226223abSTejun Heo * 3013226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3014226223abSTejun Heo * 3015226223abSTejun Heo * id RO int : the associated pool ID 3016226223abSTejun Heo * nice RW int : nice value of the workers 3017226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3018226223abSTejun Heo */ 3019226223abSTejun Heo struct wq_device { 3020226223abSTejun Heo struct workqueue_struct *wq; 3021226223abSTejun Heo struct device dev; 3022226223abSTejun Heo }; 3023226223abSTejun Heo 3024226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3025226223abSTejun Heo { 3026226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3027226223abSTejun Heo 3028226223abSTejun Heo return wq_dev->wq; 3029226223abSTejun Heo } 3030226223abSTejun Heo 30311a6661daSGreg Kroah-Hartman static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 30321a6661daSGreg Kroah-Hartman 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", (bool)!(wq->flags & WQ_UNBOUND)); 3037226223abSTejun Heo } 30381a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RO(per_cpu); 3039226223abSTejun Heo 30401a6661daSGreg Kroah-Hartman static ssize_t max_active_show(struct device *dev, 3041226223abSTejun Heo struct device_attribute *attr, char *buf) 3042226223abSTejun Heo { 3043226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3044226223abSTejun Heo 3045226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3046226223abSTejun Heo } 3047226223abSTejun Heo 30481a6661daSGreg Kroah-Hartman static ssize_t max_active_store(struct device *dev, 30491a6661daSGreg Kroah-Hartman struct device_attribute *attr, const char *buf, 30501a6661daSGreg Kroah-Hartman size_t count) 3051226223abSTejun Heo { 3052226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3053226223abSTejun Heo int val; 3054226223abSTejun Heo 3055226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3056226223abSTejun Heo return -EINVAL; 3057226223abSTejun Heo 3058226223abSTejun Heo workqueue_set_max_active(wq, val); 3059226223abSTejun Heo return count; 3060226223abSTejun Heo } 30611a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RW(max_active); 3062226223abSTejun Heo 30631a6661daSGreg Kroah-Hartman static struct attribute *wq_sysfs_attrs[] = { 30641a6661daSGreg Kroah-Hartman &dev_attr_per_cpu.attr, 30651a6661daSGreg Kroah-Hartman &dev_attr_max_active.attr, 30661a6661daSGreg Kroah-Hartman NULL, 3067226223abSTejun Heo }; 30681a6661daSGreg Kroah-Hartman ATTRIBUTE_GROUPS(wq_sysfs); 3069226223abSTejun Heo 3070d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3071226223abSTejun Heo struct device_attribute *attr, char *buf) 3072226223abSTejun Heo { 3073226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3074d55262c4STejun Heo const char *delim = ""; 3075d55262c4STejun Heo int node, written = 0; 3076226223abSTejun Heo 3077226223abSTejun Heo rcu_read_lock_sched(); 3078d55262c4STejun Heo for_each_node(node) { 3079d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3080d55262c4STejun Heo "%s%d:%d", delim, node, 3081d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3082d55262c4STejun Heo delim = " "; 3083d55262c4STejun Heo } 3084d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3085226223abSTejun Heo rcu_read_unlock_sched(); 3086226223abSTejun Heo 3087226223abSTejun Heo return written; 3088226223abSTejun Heo } 3089226223abSTejun Heo 3090226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3091226223abSTejun Heo char *buf) 3092226223abSTejun Heo { 3093226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3094226223abSTejun Heo int written; 3095226223abSTejun Heo 30966029a918STejun Heo mutex_lock(&wq->mutex); 30976029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 30986029a918STejun Heo mutex_unlock(&wq->mutex); 3099226223abSTejun Heo 3100226223abSTejun Heo return written; 3101226223abSTejun Heo } 3102226223abSTejun Heo 3103226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3104226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3105226223abSTejun Heo { 3106226223abSTejun Heo struct workqueue_attrs *attrs; 3107226223abSTejun Heo 3108226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3109226223abSTejun Heo if (!attrs) 3110226223abSTejun Heo return NULL; 3111226223abSTejun Heo 31126029a918STejun Heo mutex_lock(&wq->mutex); 31136029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 31146029a918STejun Heo mutex_unlock(&wq->mutex); 3115226223abSTejun Heo return attrs; 3116226223abSTejun Heo } 3117226223abSTejun Heo 3118226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3119226223abSTejun Heo const char *buf, size_t count) 3120226223abSTejun Heo { 3121226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3122226223abSTejun Heo struct workqueue_attrs *attrs; 3123226223abSTejun Heo int ret; 3124226223abSTejun Heo 3125226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3126226223abSTejun Heo if (!attrs) 3127226223abSTejun Heo return -ENOMEM; 3128226223abSTejun Heo 3129226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 313014481842SDongsheng Yang attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 3131226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3132226223abSTejun Heo else 3133226223abSTejun Heo ret = -EINVAL; 3134226223abSTejun Heo 3135226223abSTejun Heo free_workqueue_attrs(attrs); 3136226223abSTejun Heo return ret ?: count; 3137226223abSTejun Heo } 3138226223abSTejun Heo 3139226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3140226223abSTejun Heo struct device_attribute *attr, char *buf) 3141226223abSTejun Heo { 3142226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3143226223abSTejun Heo int written; 3144226223abSTejun Heo 31456029a918STejun Heo mutex_lock(&wq->mutex); 3146dfbcbf42STejun Heo written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 3147dfbcbf42STejun Heo cpumask_pr_args(wq->unbound_attrs->cpumask)); 31486029a918STejun Heo mutex_unlock(&wq->mutex); 3149226223abSTejun Heo return written; 3150226223abSTejun Heo } 3151226223abSTejun Heo 3152226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3153226223abSTejun Heo struct device_attribute *attr, 3154226223abSTejun Heo const char *buf, size_t count) 3155226223abSTejun Heo { 3156226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3157226223abSTejun Heo struct workqueue_attrs *attrs; 3158226223abSTejun Heo int ret; 3159226223abSTejun Heo 3160226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3161226223abSTejun Heo if (!attrs) 3162226223abSTejun Heo return -ENOMEM; 3163226223abSTejun Heo 3164226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3165226223abSTejun Heo if (!ret) 3166226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3167226223abSTejun Heo 3168226223abSTejun Heo free_workqueue_attrs(attrs); 3169226223abSTejun Heo return ret ?: count; 3170226223abSTejun Heo } 3171226223abSTejun Heo 3172d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3173d55262c4STejun Heo char *buf) 3174d55262c4STejun Heo { 3175d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3176d55262c4STejun Heo int written; 3177d55262c4STejun Heo 3178d55262c4STejun Heo mutex_lock(&wq->mutex); 3179d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3180d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3181d55262c4STejun Heo mutex_unlock(&wq->mutex); 3182d55262c4STejun Heo 3183d55262c4STejun Heo return written; 3184d55262c4STejun Heo } 3185d55262c4STejun Heo 3186d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3187d55262c4STejun Heo const char *buf, size_t count) 3188d55262c4STejun Heo { 3189d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3190d55262c4STejun Heo struct workqueue_attrs *attrs; 3191d55262c4STejun Heo int v, ret; 3192d55262c4STejun Heo 3193d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3194d55262c4STejun Heo if (!attrs) 3195d55262c4STejun Heo return -ENOMEM; 3196d55262c4STejun Heo 3197d55262c4STejun Heo ret = -EINVAL; 3198d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3199d55262c4STejun Heo attrs->no_numa = !v; 3200d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3201d55262c4STejun Heo } 3202d55262c4STejun Heo 3203d55262c4STejun Heo free_workqueue_attrs(attrs); 3204d55262c4STejun Heo return ret ?: count; 3205d55262c4STejun Heo } 3206d55262c4STejun Heo 3207226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3208d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3209226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3210226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3211d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3212226223abSTejun Heo __ATTR_NULL, 3213226223abSTejun Heo }; 3214226223abSTejun Heo 3215226223abSTejun Heo static struct bus_type wq_subsys = { 3216226223abSTejun Heo .name = "workqueue", 32171a6661daSGreg Kroah-Hartman .dev_groups = wq_sysfs_groups, 3218226223abSTejun Heo }; 3219226223abSTejun Heo 3220226223abSTejun Heo static int __init wq_sysfs_init(void) 3221226223abSTejun Heo { 3222226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3223226223abSTejun Heo } 3224226223abSTejun Heo core_initcall(wq_sysfs_init); 3225226223abSTejun Heo 3226226223abSTejun Heo static void wq_device_release(struct device *dev) 3227226223abSTejun Heo { 3228226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3229226223abSTejun Heo 3230226223abSTejun Heo kfree(wq_dev); 3231226223abSTejun Heo } 3232226223abSTejun Heo 3233226223abSTejun Heo /** 3234226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3235226223abSTejun Heo * @wq: the workqueue to register 3236226223abSTejun Heo * 3237226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3238226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3239226223abSTejun Heo * which is the preferred method. 3240226223abSTejun Heo * 3241226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3242226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3243226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3244226223abSTejun Heo * attributes. 3245226223abSTejun Heo * 3246d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. 3247226223abSTejun Heo */ 3248226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3249226223abSTejun Heo { 3250226223abSTejun Heo struct wq_device *wq_dev; 3251226223abSTejun Heo int ret; 3252226223abSTejun Heo 3253226223abSTejun Heo /* 3254226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3255226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3256226223abSTejun Heo * workqueues. 3257226223abSTejun Heo */ 3258226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3259226223abSTejun Heo return -EINVAL; 3260226223abSTejun Heo 3261226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3262226223abSTejun Heo if (!wq_dev) 3263226223abSTejun Heo return -ENOMEM; 3264226223abSTejun Heo 3265226223abSTejun Heo wq_dev->wq = wq; 3266226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3267226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3268226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3269226223abSTejun Heo 3270226223abSTejun Heo /* 3271226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3272226223abSTejun Heo * everything is ready. 3273226223abSTejun Heo */ 3274226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3275226223abSTejun Heo 3276226223abSTejun Heo ret = device_register(&wq_dev->dev); 3277226223abSTejun Heo if (ret) { 3278226223abSTejun Heo kfree(wq_dev); 3279226223abSTejun Heo wq->wq_dev = NULL; 3280226223abSTejun Heo return ret; 3281226223abSTejun Heo } 3282226223abSTejun Heo 3283226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3284226223abSTejun Heo struct device_attribute *attr; 3285226223abSTejun Heo 3286226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3287226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3288226223abSTejun Heo if (ret) { 3289226223abSTejun Heo device_unregister(&wq_dev->dev); 3290226223abSTejun Heo wq->wq_dev = NULL; 3291226223abSTejun Heo return ret; 3292226223abSTejun Heo } 3293226223abSTejun Heo } 3294226223abSTejun Heo } 3295226223abSTejun Heo 3296bddbceb6SMaxime Bizon dev_set_uevent_suppress(&wq_dev->dev, false); 3297226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3298226223abSTejun Heo return 0; 3299226223abSTejun Heo } 3300226223abSTejun Heo 3301226223abSTejun Heo /** 3302226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3303226223abSTejun Heo * @wq: the workqueue to unregister 3304226223abSTejun Heo * 3305226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3306226223abSTejun Heo */ 3307226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3308226223abSTejun Heo { 3309226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3310226223abSTejun Heo 3311226223abSTejun Heo if (!wq->wq_dev) 3312226223abSTejun Heo return; 3313226223abSTejun Heo 3314226223abSTejun Heo wq->wq_dev = NULL; 3315226223abSTejun Heo device_unregister(&wq_dev->dev); 3316226223abSTejun Heo } 3317226223abSTejun Heo #else /* CONFIG_SYSFS */ 3318226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3319226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3320226223abSTejun Heo 33217a4e344cSTejun Heo /** 33227a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33237a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33247a4e344cSTejun Heo * 33257a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33267a4e344cSTejun Heo */ 33277a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33287a4e344cSTejun Heo { 33297a4e344cSTejun Heo if (attrs) { 33307a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33317a4e344cSTejun Heo kfree(attrs); 33327a4e344cSTejun Heo } 33337a4e344cSTejun Heo } 33347a4e344cSTejun Heo 33357a4e344cSTejun Heo /** 33367a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33377a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33387a4e344cSTejun Heo * 33397a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3340d185af30SYacine Belkadi * return it. 3341d185af30SYacine Belkadi * 3342d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33437a4e344cSTejun Heo */ 33447a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33457a4e344cSTejun Heo { 33467a4e344cSTejun Heo struct workqueue_attrs *attrs; 33477a4e344cSTejun Heo 33487a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33497a4e344cSTejun Heo if (!attrs) 33507a4e344cSTejun Heo goto fail; 33517a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33527a4e344cSTejun Heo goto fail; 33537a4e344cSTejun Heo 335413e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33557a4e344cSTejun Heo return attrs; 33567a4e344cSTejun Heo fail: 33577a4e344cSTejun Heo free_workqueue_attrs(attrs); 33587a4e344cSTejun Heo return NULL; 33597a4e344cSTejun Heo } 33607a4e344cSTejun Heo 336129c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 336229c91e99STejun Heo const struct workqueue_attrs *from) 336329c91e99STejun Heo { 336429c91e99STejun Heo to->nice = from->nice; 336529c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 33662865a8fbSShaohua Li /* 33672865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 33682865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 33692865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 33702865a8fbSShaohua Li */ 33712865a8fbSShaohua Li to->no_numa = from->no_numa; 337229c91e99STejun Heo } 337329c91e99STejun Heo 337429c91e99STejun Heo /* hash value of the content of @attr */ 337529c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 337629c91e99STejun Heo { 337729c91e99STejun Heo u32 hash = 0; 337829c91e99STejun Heo 337929c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 338013e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 338113e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 338229c91e99STejun Heo return hash; 338329c91e99STejun Heo } 338429c91e99STejun Heo 338529c91e99STejun Heo /* content equality test */ 338629c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 338729c91e99STejun Heo const struct workqueue_attrs *b) 338829c91e99STejun Heo { 338929c91e99STejun Heo if (a->nice != b->nice) 339029c91e99STejun Heo return false; 339129c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 339229c91e99STejun Heo return false; 339329c91e99STejun Heo return true; 339429c91e99STejun Heo } 339529c91e99STejun Heo 33967a4e344cSTejun Heo /** 33977a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33987a4e344cSTejun Heo * @pool: worker_pool to initialize 33997a4e344cSTejun Heo * 34007a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 3401d185af30SYacine Belkadi * 3402d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 340329c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 340429c91e99STejun Heo * on @pool safely to release it. 34057a4e344cSTejun Heo */ 34067a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34074e1a1f9aSTejun Heo { 34084e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 340929c91e99STejun Heo pool->id = -1; 341029c91e99STejun Heo pool->cpu = -1; 3411f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34124e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 34134e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34144e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34154e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34164e1a1f9aSTejun Heo 34174e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 34184e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 34194e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 34204e1a1f9aSTejun Heo 34214e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 34224e1a1f9aSTejun Heo (unsigned long)pool); 34234e1a1f9aSTejun Heo 34244e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 342592f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3426da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34277a4e344cSTejun Heo 34287cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 342929c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 343029c91e99STejun Heo pool->refcnt = 1; 343129c91e99STejun Heo 343229c91e99STejun Heo /* shouldn't fail above this point */ 34337a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34347a4e344cSTejun Heo if (!pool->attrs) 34357a4e344cSTejun Heo return -ENOMEM; 34367a4e344cSTejun Heo return 0; 34374e1a1f9aSTejun Heo } 34384e1a1f9aSTejun Heo 3439e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3440e2dca7adSTejun Heo { 3441e2dca7adSTejun Heo struct workqueue_struct *wq = 3442e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3443e2dca7adSTejun Heo 3444e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3445e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3446e2dca7adSTejun Heo else 3447e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3448e2dca7adSTejun Heo 3449e2dca7adSTejun Heo kfree(wq->rescuer); 3450e2dca7adSTejun Heo kfree(wq); 3451e2dca7adSTejun Heo } 3452e2dca7adSTejun Heo 345329c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 345429c91e99STejun Heo { 345529c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 345629c91e99STejun Heo 34577cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 345829c91e99STejun Heo free_workqueue_attrs(pool->attrs); 345929c91e99STejun Heo kfree(pool); 346029c91e99STejun Heo } 346129c91e99STejun Heo 346229c91e99STejun Heo /** 346329c91e99STejun Heo * put_unbound_pool - put a worker_pool 346429c91e99STejun Heo * @pool: worker_pool to put 346529c91e99STejun Heo * 346629c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3467c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3468c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3469c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3470a892caccSTejun Heo * 3471a892caccSTejun Heo * Should be called with wq_pool_mutex held. 347229c91e99STejun Heo */ 347329c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 347429c91e99STejun Heo { 347560f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 347629c91e99STejun Heo struct worker *worker; 347729c91e99STejun Heo 3478a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3479a892caccSTejun Heo 3480a892caccSTejun Heo if (--pool->refcnt) 348129c91e99STejun Heo return; 348229c91e99STejun Heo 348329c91e99STejun Heo /* sanity checks */ 348461d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3485a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 348629c91e99STejun Heo return; 348729c91e99STejun Heo 348829c91e99STejun Heo /* release id and unhash */ 348929c91e99STejun Heo if (pool->id >= 0) 349029c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 349129c91e99STejun Heo hash_del(&pool->hash_node); 349229c91e99STejun Heo 3493c5aa87bbSTejun Heo /* 3494c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3495c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 349692f9c5c4SLai Jiangshan * attach_mutex. 3497c5aa87bbSTejun Heo */ 349829c91e99STejun Heo mutex_lock(&pool->manager_arb); 349929c91e99STejun Heo 350060f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 35011037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 350229c91e99STejun Heo destroy_worker(worker); 350329c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 350429c91e99STejun Heo spin_unlock_irq(&pool->lock); 350560f5a4bcSLai Jiangshan 350692f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3507da028469SLai Jiangshan if (!list_empty(&pool->workers)) 350860f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 350992f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 351060f5a4bcSLai Jiangshan 351160f5a4bcSLai Jiangshan if (pool->detach_completion) 351260f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 351360f5a4bcSLai Jiangshan 351429c91e99STejun Heo mutex_unlock(&pool->manager_arb); 351529c91e99STejun Heo 351629c91e99STejun Heo /* shut down the timers */ 351729c91e99STejun Heo del_timer_sync(&pool->idle_timer); 351829c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 351929c91e99STejun Heo 352029c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 352129c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 352229c91e99STejun Heo } 352329c91e99STejun Heo 352429c91e99STejun Heo /** 352529c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 352629c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 352729c91e99STejun Heo * 352829c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 352929c91e99STejun Heo * reference count and return it. If there already is a matching 353029c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3531d185af30SYacine Belkadi * create a new one. 3532a892caccSTejun Heo * 3533a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3534d185af30SYacine Belkadi * 3535d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3536d185af30SYacine Belkadi * On failure, %NULL. 353729c91e99STejun Heo */ 353829c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 353929c91e99STejun Heo { 354029c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 354129c91e99STejun Heo struct worker_pool *pool; 3542f3f90ad4STejun Heo int node; 354329c91e99STejun Heo 3544a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 354529c91e99STejun Heo 354629c91e99STejun Heo /* do we already have a matching pool? */ 354729c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 354829c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 354929c91e99STejun Heo pool->refcnt++; 35503fb1823cSLai Jiangshan return pool; 355129c91e99STejun Heo } 355229c91e99STejun Heo } 355329c91e99STejun Heo 355429c91e99STejun Heo /* nope, create a new one */ 355529c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 355629c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 355729c91e99STejun Heo goto fail; 355829c91e99STejun Heo 35598864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 356029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 356129c91e99STejun Heo 35622865a8fbSShaohua Li /* 35632865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 35642865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 35652865a8fbSShaohua Li */ 35662865a8fbSShaohua Li pool->attrs->no_numa = false; 35672865a8fbSShaohua Li 3568f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3569f3f90ad4STejun Heo if (wq_numa_enabled) { 3570f3f90ad4STejun Heo for_each_node(node) { 3571f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3572f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3573f3f90ad4STejun Heo pool->node = node; 3574f3f90ad4STejun Heo break; 3575f3f90ad4STejun Heo } 3576f3f90ad4STejun Heo } 3577f3f90ad4STejun Heo } 3578f3f90ad4STejun Heo 357929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 358029c91e99STejun Heo goto fail; 358129c91e99STejun Heo 358229c91e99STejun Heo /* create and start the initial worker */ 3583051e1850SLai Jiangshan if (!create_worker(pool)) 358429c91e99STejun Heo goto fail; 358529c91e99STejun Heo 358629c91e99STejun Heo /* install */ 358729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 35883fb1823cSLai Jiangshan 358929c91e99STejun Heo return pool; 359029c91e99STejun Heo fail: 359129c91e99STejun Heo if (pool) 359229c91e99STejun Heo put_unbound_pool(pool); 359329c91e99STejun Heo return NULL; 359429c91e99STejun Heo } 359529c91e99STejun Heo 35968864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35978864b4e5STejun Heo { 35988864b4e5STejun Heo kmem_cache_free(pwq_cache, 35998864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36008864b4e5STejun Heo } 36018864b4e5STejun Heo 36028864b4e5STejun Heo /* 36038864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36048864b4e5STejun Heo * and needs to be destroyed. 36058864b4e5STejun Heo */ 36068864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36078864b4e5STejun Heo { 36088864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36098864b4e5STejun Heo unbound_release_work); 36108864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36118864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3612bc0caf09STejun Heo bool is_last; 36138864b4e5STejun Heo 36148864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36158864b4e5STejun Heo return; 36168864b4e5STejun Heo 36173c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 36188864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3619bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 36203c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 36218864b4e5STejun Heo 3622a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 36238864b4e5STejun Heo put_unbound_pool(pool); 3624a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3625a892caccSTejun Heo 36268864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 36278864b4e5STejun Heo 36288864b4e5STejun Heo /* 36298864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3630e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 36318864b4e5STejun Heo */ 3632e2dca7adSTejun Heo if (is_last) 3633e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 36346029a918STejun Heo } 36358864b4e5STejun Heo 36360fbd95aaSTejun Heo /** 3637699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36380fbd95aaSTejun Heo * @pwq: target pool_workqueue 36390fbd95aaSTejun Heo * 3640699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3641699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3642699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36430fbd95aaSTejun Heo */ 3644699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36450fbd95aaSTejun Heo { 3646699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3647699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3648699ce097STejun Heo 3649699ce097STejun Heo /* for @wq->saved_max_active */ 3650a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3651699ce097STejun Heo 3652699ce097STejun Heo /* fast exit for non-freezable wqs */ 3653699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3654699ce097STejun Heo return; 3655699ce097STejun Heo 3656a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3657699ce097STejun Heo 365874b414eaSLai Jiangshan /* 365974b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 366074b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 366174b414eaSLai Jiangshan * is updated and visible. 366274b414eaSLai Jiangshan */ 366374b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3664699ce097STejun Heo pwq->max_active = wq->saved_max_active; 36650fbd95aaSTejun Heo 36660fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36670fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36680fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3669951a078aSLai Jiangshan 3670951a078aSLai Jiangshan /* 3671951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3672951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3673951a078aSLai Jiangshan */ 3674951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3675699ce097STejun Heo } else { 3676699ce097STejun Heo pwq->max_active = 0; 3677699ce097STejun Heo } 3678699ce097STejun Heo 3679a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36800fbd95aaSTejun Heo } 36810fbd95aaSTejun Heo 3682e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3683f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3684f147f29eSTejun Heo struct worker_pool *pool) 3685d2c1d404STejun Heo { 3686d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3687d2c1d404STejun Heo 3688e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3689e50aba9aSTejun Heo 3690d2c1d404STejun Heo pwq->pool = pool; 3691d2c1d404STejun Heo pwq->wq = wq; 3692d2c1d404STejun Heo pwq->flush_color = -1; 36938864b4e5STejun Heo pwq->refcnt = 1; 3694d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 36951befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3696d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36978864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3698f147f29eSTejun Heo } 3699d2c1d404STejun Heo 3700f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37011befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3702f147f29eSTejun Heo { 3703f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3704f147f29eSTejun Heo 3705f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 370675ccf595STejun Heo 37071befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37081befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37091befcf30STejun Heo return; 37101befcf30STejun Heo 371129b1cb41SLai Jiangshan /* set the matching work_color */ 371275ccf595STejun Heo pwq->work_color = wq->work_color; 3713983ca25eSTejun Heo 3714983ca25eSTejun Heo /* sync max_active to the current setting */ 3715983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3716983ca25eSTejun Heo 3717983ca25eSTejun Heo /* link in @pwq */ 37189e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3719df2d5ae4STejun Heo } 37206029a918STejun Heo 3721f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3722f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3723f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3724f147f29eSTejun Heo { 3725f147f29eSTejun Heo struct worker_pool *pool; 3726f147f29eSTejun Heo struct pool_workqueue *pwq; 3727f147f29eSTejun Heo 3728f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3729f147f29eSTejun Heo 3730f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3731f147f29eSTejun Heo if (!pool) 3732f147f29eSTejun Heo return NULL; 3733f147f29eSTejun Heo 3734e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3735f147f29eSTejun Heo if (!pwq) { 3736f147f29eSTejun Heo put_unbound_pool(pool); 3737f147f29eSTejun Heo return NULL; 3738f147f29eSTejun Heo } 3739f147f29eSTejun Heo 3740f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3741f147f29eSTejun Heo return pwq; 3742d2c1d404STejun Heo } 3743d2c1d404STejun Heo 37444c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 37454c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 37464c16bd32STejun Heo { 37474c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37484c16bd32STejun Heo 37494c16bd32STejun Heo if (pwq) { 37504c16bd32STejun Heo put_unbound_pool(pwq->pool); 3751cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 37524c16bd32STejun Heo } 37534c16bd32STejun Heo } 37544c16bd32STejun Heo 37554c16bd32STejun Heo /** 37564c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 37574c16bd32STejun Heo * @attrs: the wq_attrs of interest 37584c16bd32STejun Heo * @node: the target NUMA node 37594c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 37604c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 37614c16bd32STejun Heo * 37624c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 37634c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3764d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 37654c16bd32STejun Heo * 37664c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 37674c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 37684c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 37694c16bd32STejun Heo * @attrs->cpumask. 37704c16bd32STejun Heo * 37714c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 37724c16bd32STejun Heo * stable. 3773d185af30SYacine Belkadi * 3774d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3775d185af30SYacine Belkadi * %false if equal. 37764c16bd32STejun Heo */ 37774c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 37784c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 37794c16bd32STejun Heo { 3780d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 37814c16bd32STejun Heo goto use_dfl; 37824c16bd32STejun Heo 37834c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 37844c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 37854c16bd32STejun Heo if (cpu_going_down >= 0) 37864c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 37874c16bd32STejun Heo 37884c16bd32STejun Heo if (cpumask_empty(cpumask)) 37894c16bd32STejun Heo goto use_dfl; 37904c16bd32STejun Heo 37914c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 37924c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 37934c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 37944c16bd32STejun Heo 37954c16bd32STejun Heo use_dfl: 37964c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 37974c16bd32STejun Heo return false; 37984c16bd32STejun Heo } 37994c16bd32STejun Heo 38001befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38011befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38021befcf30STejun Heo int node, 38031befcf30STejun Heo struct pool_workqueue *pwq) 38041befcf30STejun Heo { 38051befcf30STejun Heo struct pool_workqueue *old_pwq; 38061befcf30STejun Heo 38071befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38081befcf30STejun Heo 38091befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 38101befcf30STejun Heo link_pwq(pwq); 38111befcf30STejun Heo 38121befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 38131befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 38141befcf30STejun Heo return old_pwq; 38151befcf30STejun Heo } 38161befcf30STejun Heo 38179e8cd2f5STejun Heo /** 38189e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38199e8cd2f5STejun Heo * @wq: the target workqueue 38209e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38219e8cd2f5STejun Heo * 38224c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 38234c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 38244c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38254c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38264c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38274c16bd32STejun Heo * back-to-back will stay on its current pwq. 38289e8cd2f5STejun Heo * 3829d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3830d185af30SYacine Belkadi * 3831d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 38329e8cd2f5STejun Heo */ 38339e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38349e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38359e8cd2f5STejun Heo { 38364c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 38374c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3838f147f29eSTejun Heo int node, ret; 38399e8cd2f5STejun Heo 38408719dceaSTejun Heo /* only unbound workqueues can change attributes */ 38419e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 38429e8cd2f5STejun Heo return -EINVAL; 38439e8cd2f5STejun Heo 38448719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 38458719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 38468719dceaSTejun Heo return -EINVAL; 38478719dceaSTejun Heo 3848ddcb57e2SLai Jiangshan pwq_tbl = kzalloc(nr_node_ids * sizeof(pwq_tbl[0]), GFP_KERNEL); 384913e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38504c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38514c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 385213e2e556STejun Heo goto enomem; 385313e2e556STejun Heo 38544c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 385513e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 385613e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 385713e2e556STejun Heo 38584c16bd32STejun Heo /* 38594c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 38604c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 38614c16bd32STejun Heo * pools. 38624c16bd32STejun Heo */ 38634c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 38649e8cd2f5STejun Heo 38654c16bd32STejun Heo /* 38664c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 38674c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 38684c16bd32STejun Heo * pwqs accordingly. 38694c16bd32STejun Heo */ 38704c16bd32STejun Heo get_online_cpus(); 38714c16bd32STejun Heo 38724c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 38734c16bd32STejun Heo 38744c16bd32STejun Heo /* 38754c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 38764c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 38774c16bd32STejun Heo * it even if we don't use it immediately. 38784c16bd32STejun Heo */ 38794c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 38804c16bd32STejun Heo if (!dfl_pwq) 38814c16bd32STejun Heo goto enomem_pwq; 38824c16bd32STejun Heo 38834c16bd32STejun Heo for_each_node(node) { 38844c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 38854c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 38864c16bd32STejun Heo if (!pwq_tbl[node]) 38874c16bd32STejun Heo goto enomem_pwq; 38884c16bd32STejun Heo } else { 38894c16bd32STejun Heo dfl_pwq->refcnt++; 38904c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 38914c16bd32STejun Heo } 38924c16bd32STejun Heo } 38934c16bd32STejun Heo 38944c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 38954c16bd32STejun Heo 38964c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3897f147f29eSTejun Heo mutex_lock(&wq->mutex); 3898a892caccSTejun Heo 3899f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 39004c16bd32STejun Heo 39014c16bd32STejun Heo /* save the previous pwq and install the new one */ 3902f147f29eSTejun Heo for_each_node(node) 39034c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 39044c16bd32STejun Heo 39054c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 39064c16bd32STejun Heo link_pwq(dfl_pwq); 39074c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3908f147f29eSTejun Heo 3909f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3910f147f29eSTejun Heo 39114c16bd32STejun Heo /* put the old pwqs */ 39124c16bd32STejun Heo for_each_node(node) 39134c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 39144c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 39154c16bd32STejun Heo 39164c16bd32STejun Heo put_online_cpus(); 39174862125bSTejun Heo ret = 0; 39184862125bSTejun Heo /* fall through */ 39194862125bSTejun Heo out_free: 39204c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 39214862125bSTejun Heo free_workqueue_attrs(new_attrs); 39224c16bd32STejun Heo kfree(pwq_tbl); 39234862125bSTejun Heo return ret; 392413e2e556STejun Heo 39254c16bd32STejun Heo enomem_pwq: 39264c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 39274c16bd32STejun Heo for_each_node(node) 39284c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 39294c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 39304c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39314c16bd32STejun Heo put_online_cpus(); 393213e2e556STejun Heo enomem: 39334862125bSTejun Heo ret = -ENOMEM; 39344862125bSTejun Heo goto out_free; 39359e8cd2f5STejun Heo } 39369e8cd2f5STejun Heo 39374c16bd32STejun Heo /** 39384c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 39394c16bd32STejun Heo * @wq: the target workqueue 39404c16bd32STejun Heo * @cpu: the CPU coming up or going down 39414c16bd32STejun Heo * @online: whether @cpu is coming up or going down 39424c16bd32STejun Heo * 39434c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 39444c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 39454c16bd32STejun Heo * @wq accordingly. 39464c16bd32STejun Heo * 39474c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 39484c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 39494c16bd32STejun Heo * correct. 39504c16bd32STejun Heo * 39514c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 39524c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 39534c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 39544c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 39554c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 39564c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 39574c16bd32STejun Heo * CPU_DOWN_PREPARE. 39584c16bd32STejun Heo */ 39594c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 39604c16bd32STejun Heo bool online) 39614c16bd32STejun Heo { 39624c16bd32STejun Heo int node = cpu_to_node(cpu); 39634c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 39644c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 39654c16bd32STejun Heo struct workqueue_attrs *target_attrs; 39664c16bd32STejun Heo cpumask_t *cpumask; 39674c16bd32STejun Heo 39684c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 39694c16bd32STejun Heo 39704c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 39714c16bd32STejun Heo return; 39724c16bd32STejun Heo 39734c16bd32STejun Heo /* 39744c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39754c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39764c16bd32STejun Heo * CPU hotplug exclusion. 39774c16bd32STejun Heo */ 39784c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39794c16bd32STejun Heo cpumask = target_attrs->cpumask; 39804c16bd32STejun Heo 39814c16bd32STejun Heo mutex_lock(&wq->mutex); 3982d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 3983d55262c4STejun Heo goto out_unlock; 39844c16bd32STejun Heo 39854c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39864c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39874c16bd32STejun Heo 39884c16bd32STejun Heo /* 39894c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 39904c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 39914c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 3992534a3fbbSDaeseok Youn * wq's, the default pwq should be used. 39934c16bd32STejun Heo */ 39944c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 39954c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 39964c16bd32STejun Heo goto out_unlock; 39974c16bd32STejun Heo } else { 39984c16bd32STejun Heo goto use_dfl_pwq; 39994c16bd32STejun Heo } 40004c16bd32STejun Heo 40014c16bd32STejun Heo mutex_unlock(&wq->mutex); 40024c16bd32STejun Heo 40034c16bd32STejun Heo /* create a new pwq */ 40044c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 40054c16bd32STejun Heo if (!pwq) { 40062d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 40074c16bd32STejun Heo wq->name); 400877f300b1SDaeseok Youn mutex_lock(&wq->mutex); 400977f300b1SDaeseok Youn goto use_dfl_pwq; 40104c16bd32STejun Heo } 40114c16bd32STejun Heo 40124c16bd32STejun Heo /* 40134c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 40144c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 40154c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 40164c16bd32STejun Heo * inbetween. 40174c16bd32STejun Heo */ 40184c16bd32STejun Heo mutex_lock(&wq->mutex); 40194c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 40204c16bd32STejun Heo goto out_unlock; 40214c16bd32STejun Heo 40224c16bd32STejun Heo use_dfl_pwq: 40234c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 40244c16bd32STejun Heo get_pwq(wq->dfl_pwq); 40254c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 40264c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 40274c16bd32STejun Heo out_unlock: 40284c16bd32STejun Heo mutex_unlock(&wq->mutex); 40294c16bd32STejun Heo put_pwq_unlocked(old_pwq); 40304c16bd32STejun Heo } 40314c16bd32STejun Heo 403230cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 40331da177e4SLinus Torvalds { 403449e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 40358a2b7538STejun Heo int cpu, ret; 4036e1d8aa9fSFrederic Weisbecker 403730cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4038420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4039420c0ddbSTejun Heo if (!wq->cpu_pwqs) 404030cdf249STejun Heo return -ENOMEM; 404130cdf249STejun Heo 404230cdf249STejun Heo for_each_possible_cpu(cpu) { 40437fb98ea7STejun Heo struct pool_workqueue *pwq = 40447fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 40457a62c2c8STejun Heo struct worker_pool *cpu_pools = 4046f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 404730cdf249STejun Heo 4048f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4049f147f29eSTejun Heo 4050f147f29eSTejun Heo mutex_lock(&wq->mutex); 40511befcf30STejun Heo link_pwq(pwq); 4052f147f29eSTejun Heo mutex_unlock(&wq->mutex); 405330cdf249STejun Heo } 405430cdf249STejun Heo return 0; 40558a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 40568a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 40578a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 40588a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 40598a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 40608a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 40618a2b7538STejun Heo return ret; 40629e8cd2f5STejun Heo } else { 40639e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 40649e8cd2f5STejun Heo } 40650f900049STejun Heo } 40660f900049STejun Heo 4067f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4068f3421797STejun Heo const char *name) 4069b71ab8c2STejun Heo { 4070f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4071f3421797STejun Heo 4072f3421797STejun Heo if (max_active < 1 || max_active > lim) 4073044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4074f3421797STejun Heo max_active, name, 1, lim); 4075b71ab8c2STejun Heo 4076f3421797STejun Heo return clamp_val(max_active, 1, lim); 4077b71ab8c2STejun Heo } 4078b71ab8c2STejun Heo 4079b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 408097e37d7bSTejun Heo unsigned int flags, 40811e19ffc6STejun Heo int max_active, 4082eb13ba87SJohannes Berg struct lock_class_key *key, 4083b196be89STejun Heo const char *lock_name, ...) 40843af24433SOleg Nesterov { 4085df2d5ae4STejun Heo size_t tbl_size = 0; 4086ecf6881fSTejun Heo va_list args; 40873af24433SOleg Nesterov struct workqueue_struct *wq; 408849e3cf44STejun Heo struct pool_workqueue *pwq; 4089b196be89STejun Heo 4090cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4091cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4092cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4093cee22a15SViresh Kumar 4094ecf6881fSTejun Heo /* allocate wq and format name */ 4095df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4096ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4097df2d5ae4STejun Heo 4098df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4099b196be89STejun Heo if (!wq) 4100d2c1d404STejun Heo return NULL; 4101b196be89STejun Heo 41026029a918STejun Heo if (flags & WQ_UNBOUND) { 41036029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 41046029a918STejun Heo if (!wq->unbound_attrs) 41056029a918STejun Heo goto err_free_wq; 41066029a918STejun Heo } 41076029a918STejun Heo 4108ecf6881fSTejun Heo va_start(args, lock_name); 4109ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4110b196be89STejun Heo va_end(args); 41113af24433SOleg Nesterov 4112d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4113b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 41143af24433SOleg Nesterov 4115b196be89STejun Heo /* init wq */ 411697e37d7bSTejun Heo wq->flags = flags; 4117a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 41183c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4119112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 412030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 412173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 412273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4123493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 41243af24433SOleg Nesterov 4125eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4126cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 41273af24433SOleg Nesterov 412830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4129d2c1d404STejun Heo goto err_free_wq; 41301537663fSTejun Heo 4131493008a8STejun Heo /* 4132493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4133493008a8STejun Heo * have a rescuer to guarantee forward progress. 4134493008a8STejun Heo */ 4135493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4136e22bee78STejun Heo struct worker *rescuer; 4137e22bee78STejun Heo 4138f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 4139e22bee78STejun Heo if (!rescuer) 4140d2c1d404STejun Heo goto err_destroy; 4141e22bee78STejun Heo 4142111c225aSTejun Heo rescuer->rescue_wq = wq; 4143111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4144b196be89STejun Heo wq->name); 4145d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4146d2c1d404STejun Heo kfree(rescuer); 4147d2c1d404STejun Heo goto err_destroy; 4148d2c1d404STejun Heo } 4149e22bee78STejun Heo 4150d2c1d404STejun Heo wq->rescuer = rescuer; 415114a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4152e22bee78STejun Heo wake_up_process(rescuer->task); 41533af24433SOleg Nesterov } 41541537663fSTejun Heo 4155226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4156226223abSTejun Heo goto err_destroy; 4157226223abSTejun Heo 41583af24433SOleg Nesterov /* 415968e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 416068e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 416168e13a67SLai Jiangshan * list. 41623af24433SOleg Nesterov */ 416368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4164a0a1a5fdSTejun Heo 4165a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 416649e3cf44STejun Heo for_each_pwq(pwq, wq) 4167699ce097STejun Heo pwq_adjust_max_active(pwq); 4168a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4169a0a1a5fdSTejun Heo 4170e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4171a0a1a5fdSTejun Heo 417268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41733af24433SOleg Nesterov 41743af24433SOleg Nesterov return wq; 4175d2c1d404STejun Heo 4176d2c1d404STejun Heo err_free_wq: 41776029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41784690c4abSTejun Heo kfree(wq); 4179d2c1d404STejun Heo return NULL; 4180d2c1d404STejun Heo err_destroy: 4181d2c1d404STejun Heo destroy_workqueue(wq); 41824690c4abSTejun Heo return NULL; 41831da177e4SLinus Torvalds } 4184d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41851da177e4SLinus Torvalds 41863af24433SOleg Nesterov /** 41873af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41883af24433SOleg Nesterov * @wq: target workqueue 41893af24433SOleg Nesterov * 41903af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41913af24433SOleg Nesterov */ 41923af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41933af24433SOleg Nesterov { 419449e3cf44STejun Heo struct pool_workqueue *pwq; 41954c16bd32STejun Heo int node; 41963af24433SOleg Nesterov 41979c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41989c5a2ba7STejun Heo drain_workqueue(wq); 4199c8efcc25STejun Heo 42006183c009STejun Heo /* sanity checks */ 4201b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 420249e3cf44STejun Heo for_each_pwq(pwq, wq) { 42036183c009STejun Heo int i; 42046183c009STejun Heo 420576af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 420676af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4207b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42086183c009STejun Heo return; 420976af4d93STejun Heo } 421076af4d93STejun Heo } 421176af4d93STejun Heo 42125c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 42138864b4e5STejun Heo WARN_ON(pwq->nr_active) || 421476af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4215b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42166183c009STejun Heo return; 42176183c009STejun Heo } 421876af4d93STejun Heo } 4219b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42206183c009STejun Heo 4221a0a1a5fdSTejun Heo /* 4222a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4223a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4224a0a1a5fdSTejun Heo */ 422568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4226e2dca7adSTejun Heo list_del_rcu(&wq->list); 422768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42283af24433SOleg Nesterov 4229226223abSTejun Heo workqueue_sysfs_unregister(wq); 4230226223abSTejun Heo 4231e2dca7adSTejun Heo if (wq->rescuer) 4232e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4233e22bee78STejun Heo 42348864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 423529c91e99STejun Heo /* 42368864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4237e2dca7adSTejun Heo * schedule RCU free. 423829c91e99STejun Heo */ 4239e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 42408864b4e5STejun Heo } else { 42418864b4e5STejun Heo /* 42428864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 42434c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 42444c16bd32STejun Heo * @wq will be freed when the last pwq is released. 42458864b4e5STejun Heo */ 42464c16bd32STejun Heo for_each_node(node) { 42474c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42484c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 42494c16bd32STejun Heo put_pwq_unlocked(pwq); 42504c16bd32STejun Heo } 42514c16bd32STejun Heo 42524c16bd32STejun Heo /* 42534c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 42544c16bd32STejun Heo * put. Don't access it afterwards. 42554c16bd32STejun Heo */ 42564c16bd32STejun Heo pwq = wq->dfl_pwq; 42574c16bd32STejun Heo wq->dfl_pwq = NULL; 4258dce90d47STejun Heo put_pwq_unlocked(pwq); 425929c91e99STejun Heo } 42603af24433SOleg Nesterov } 42613af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 42623af24433SOleg Nesterov 4263dcd989cbSTejun Heo /** 4264dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4265dcd989cbSTejun Heo * @wq: target workqueue 4266dcd989cbSTejun Heo * @max_active: new max_active value. 4267dcd989cbSTejun Heo * 4268dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4269dcd989cbSTejun Heo * 4270dcd989cbSTejun Heo * CONTEXT: 4271dcd989cbSTejun Heo * Don't call from IRQ context. 4272dcd989cbSTejun Heo */ 4273dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4274dcd989cbSTejun Heo { 427549e3cf44STejun Heo struct pool_workqueue *pwq; 4276dcd989cbSTejun Heo 42778719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42788719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 42798719dceaSTejun Heo return; 42808719dceaSTejun Heo 4281f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4282dcd989cbSTejun Heo 4283a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4284dcd989cbSTejun Heo 4285dcd989cbSTejun Heo wq->saved_max_active = max_active; 4286dcd989cbSTejun Heo 4287699ce097STejun Heo for_each_pwq(pwq, wq) 4288699ce097STejun Heo pwq_adjust_max_active(pwq); 4289dcd989cbSTejun Heo 4290a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4291dcd989cbSTejun Heo } 4292dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4293dcd989cbSTejun Heo 4294dcd989cbSTejun Heo /** 4295e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4296e6267616STejun Heo * 4297e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4298e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4299d185af30SYacine Belkadi * 4300d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4301e6267616STejun Heo */ 4302e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4303e6267616STejun Heo { 4304e6267616STejun Heo struct worker *worker = current_wq_worker(); 4305e6267616STejun Heo 43066a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4307e6267616STejun Heo } 4308e6267616STejun Heo 4309e6267616STejun Heo /** 4310dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4311dcd989cbSTejun Heo * @cpu: CPU in question 4312dcd989cbSTejun Heo * @wq: target workqueue 4313dcd989cbSTejun Heo * 4314dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4315dcd989cbSTejun Heo * no synchronization around this function and the test result is 4316dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4317dcd989cbSTejun Heo * 4318d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4319d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4320d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4321d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4322d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4323d3251859STejun Heo * 4324d185af30SYacine Belkadi * Return: 4325dcd989cbSTejun Heo * %true if congested, %false otherwise. 4326dcd989cbSTejun Heo */ 4327d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4328dcd989cbSTejun Heo { 43297fb98ea7STejun Heo struct pool_workqueue *pwq; 433076af4d93STejun Heo bool ret; 433176af4d93STejun Heo 433288109453SLai Jiangshan rcu_read_lock_sched(); 43337fb98ea7STejun Heo 4334d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4335d3251859STejun Heo cpu = smp_processor_id(); 4336d3251859STejun Heo 43377fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 43387fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 43397fb98ea7STejun Heo else 4340df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4341dcd989cbSTejun Heo 434276af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 434388109453SLai Jiangshan rcu_read_unlock_sched(); 434476af4d93STejun Heo 434576af4d93STejun Heo return ret; 4346dcd989cbSTejun Heo } 4347dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4348dcd989cbSTejun Heo 4349dcd989cbSTejun Heo /** 4350dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4351dcd989cbSTejun Heo * @work: the work to be tested 4352dcd989cbSTejun Heo * 4353dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4354dcd989cbSTejun Heo * synchronization around this function and the test result is 4355dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4356dcd989cbSTejun Heo * 4357d185af30SYacine Belkadi * Return: 4358dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4359dcd989cbSTejun Heo */ 4360dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4361dcd989cbSTejun Heo { 4362fa1b54e6STejun Heo struct worker_pool *pool; 4363dcd989cbSTejun Heo unsigned long flags; 4364dcd989cbSTejun Heo unsigned int ret = 0; 4365dcd989cbSTejun Heo 4366dcd989cbSTejun Heo if (work_pending(work)) 4367dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4368038366c5SLai Jiangshan 4369fa1b54e6STejun Heo local_irq_save(flags); 4370fa1b54e6STejun Heo pool = get_work_pool(work); 4371038366c5SLai Jiangshan if (pool) { 4372fa1b54e6STejun Heo spin_lock(&pool->lock); 4373c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4374dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4375fa1b54e6STejun Heo spin_unlock(&pool->lock); 4376038366c5SLai Jiangshan } 4377fa1b54e6STejun Heo local_irq_restore(flags); 4378dcd989cbSTejun Heo 4379dcd989cbSTejun Heo return ret; 4380dcd989cbSTejun Heo } 4381dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4382dcd989cbSTejun Heo 43833d1cb205STejun Heo /** 43843d1cb205STejun Heo * set_worker_desc - set description for the current work item 43853d1cb205STejun Heo * @fmt: printf-style format string 43863d1cb205STejun Heo * @...: arguments for the format string 43873d1cb205STejun Heo * 43883d1cb205STejun Heo * This function can be called by a running work function to describe what 43893d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43903d1cb205STejun Heo * information will be printed out together to help debugging. The 43913d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43923d1cb205STejun Heo */ 43933d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43943d1cb205STejun Heo { 43953d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43963d1cb205STejun Heo va_list args; 43973d1cb205STejun Heo 43983d1cb205STejun Heo if (worker) { 43993d1cb205STejun Heo va_start(args, fmt); 44003d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 44013d1cb205STejun Heo va_end(args); 44023d1cb205STejun Heo worker->desc_valid = true; 44033d1cb205STejun Heo } 44043d1cb205STejun Heo } 44053d1cb205STejun Heo 44063d1cb205STejun Heo /** 44073d1cb205STejun Heo * print_worker_info - print out worker information and description 44083d1cb205STejun Heo * @log_lvl: the log level to use when printing 44093d1cb205STejun Heo * @task: target task 44103d1cb205STejun Heo * 44113d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 44123d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 44133d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 44143d1cb205STejun Heo * 44153d1cb205STejun Heo * This function can be safely called on any task as long as the 44163d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 44173d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 44183d1cb205STejun Heo */ 44193d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 44203d1cb205STejun Heo { 44213d1cb205STejun Heo work_func_t *fn = NULL; 44223d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 44233d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 44243d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 44253d1cb205STejun Heo struct workqueue_struct *wq = NULL; 44263d1cb205STejun Heo bool desc_valid = false; 44273d1cb205STejun Heo struct worker *worker; 44283d1cb205STejun Heo 44293d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 44303d1cb205STejun Heo return; 44313d1cb205STejun Heo 44323d1cb205STejun Heo /* 44333d1cb205STejun Heo * This function is called without any synchronization and @task 44343d1cb205STejun Heo * could be in any state. Be careful with dereferences. 44353d1cb205STejun Heo */ 44363d1cb205STejun Heo worker = probe_kthread_data(task); 44373d1cb205STejun Heo 44383d1cb205STejun Heo /* 44393d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 44403d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 44413d1cb205STejun Heo */ 44423d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 44433d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 44443d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 44453d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 44463d1cb205STejun Heo 44473d1cb205STejun Heo /* copy worker description */ 44483d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 44493d1cb205STejun Heo if (desc_valid) 44503d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 44513d1cb205STejun Heo 44523d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44533d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 44543d1cb205STejun Heo if (desc[0]) 44553d1cb205STejun Heo pr_cont(" (%s)", desc); 44563d1cb205STejun Heo pr_cont("\n"); 44573d1cb205STejun Heo } 44583d1cb205STejun Heo } 44593d1cb205STejun Heo 44603494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 44613494fc30STejun Heo { 44623494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 44633494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 44643494fc30STejun Heo pr_cont(" node=%d", pool->node); 44653494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 44663494fc30STejun Heo } 44673494fc30STejun Heo 44683494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 44693494fc30STejun Heo { 44703494fc30STejun Heo if (work->func == wq_barrier_func) { 44713494fc30STejun Heo struct wq_barrier *barr; 44723494fc30STejun Heo 44733494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 44743494fc30STejun Heo 44753494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 44763494fc30STejun Heo task_pid_nr(barr->task)); 44773494fc30STejun Heo } else { 44783494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 44793494fc30STejun Heo } 44803494fc30STejun Heo } 44813494fc30STejun Heo 44823494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 44833494fc30STejun Heo { 44843494fc30STejun Heo struct worker_pool *pool = pwq->pool; 44853494fc30STejun Heo struct work_struct *work; 44863494fc30STejun Heo struct worker *worker; 44873494fc30STejun Heo bool has_in_flight = false, has_pending = false; 44883494fc30STejun Heo int bkt; 44893494fc30STejun Heo 44903494fc30STejun Heo pr_info(" pwq %d:", pool->id); 44913494fc30STejun Heo pr_cont_pool_info(pool); 44923494fc30STejun Heo 44933494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 44943494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 44953494fc30STejun Heo 44963494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 44973494fc30STejun Heo if (worker->current_pwq == pwq) { 44983494fc30STejun Heo has_in_flight = true; 44993494fc30STejun Heo break; 45003494fc30STejun Heo } 45013494fc30STejun Heo } 45023494fc30STejun Heo if (has_in_flight) { 45033494fc30STejun Heo bool comma = false; 45043494fc30STejun Heo 45053494fc30STejun Heo pr_info(" in-flight:"); 45063494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 45073494fc30STejun Heo if (worker->current_pwq != pwq) 45083494fc30STejun Heo continue; 45093494fc30STejun Heo 45103494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 45113494fc30STejun Heo task_pid_nr(worker->task), 45123494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 45133494fc30STejun Heo worker->current_func); 45143494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 45153494fc30STejun Heo pr_cont_work(false, work); 45163494fc30STejun Heo comma = true; 45173494fc30STejun Heo } 45183494fc30STejun Heo pr_cont("\n"); 45193494fc30STejun Heo } 45203494fc30STejun Heo 45213494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 45223494fc30STejun Heo if (get_work_pwq(work) == pwq) { 45233494fc30STejun Heo has_pending = true; 45243494fc30STejun Heo break; 45253494fc30STejun Heo } 45263494fc30STejun Heo } 45273494fc30STejun Heo if (has_pending) { 45283494fc30STejun Heo bool comma = false; 45293494fc30STejun Heo 45303494fc30STejun Heo pr_info(" pending:"); 45313494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 45323494fc30STejun Heo if (get_work_pwq(work) != pwq) 45333494fc30STejun Heo continue; 45343494fc30STejun Heo 45353494fc30STejun Heo pr_cont_work(comma, work); 45363494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 45373494fc30STejun Heo } 45383494fc30STejun Heo pr_cont("\n"); 45393494fc30STejun Heo } 45403494fc30STejun Heo 45413494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 45423494fc30STejun Heo bool comma = false; 45433494fc30STejun Heo 45443494fc30STejun Heo pr_info(" delayed:"); 45453494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 45463494fc30STejun Heo pr_cont_work(comma, work); 45473494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 45483494fc30STejun Heo } 45493494fc30STejun Heo pr_cont("\n"); 45503494fc30STejun Heo } 45513494fc30STejun Heo } 45523494fc30STejun Heo 45533494fc30STejun Heo /** 45543494fc30STejun Heo * show_workqueue_state - dump workqueue state 45553494fc30STejun Heo * 45563494fc30STejun Heo * Called from a sysrq handler and prints out all busy workqueues and 45573494fc30STejun Heo * pools. 45583494fc30STejun Heo */ 45593494fc30STejun Heo void show_workqueue_state(void) 45603494fc30STejun Heo { 45613494fc30STejun Heo struct workqueue_struct *wq; 45623494fc30STejun Heo struct worker_pool *pool; 45633494fc30STejun Heo unsigned long flags; 45643494fc30STejun Heo int pi; 45653494fc30STejun Heo 45663494fc30STejun Heo rcu_read_lock_sched(); 45673494fc30STejun Heo 45683494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 45693494fc30STejun Heo 45703494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 45713494fc30STejun Heo struct pool_workqueue *pwq; 45723494fc30STejun Heo bool idle = true; 45733494fc30STejun Heo 45743494fc30STejun Heo for_each_pwq(pwq, wq) { 45753494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 45763494fc30STejun Heo idle = false; 45773494fc30STejun Heo break; 45783494fc30STejun Heo } 45793494fc30STejun Heo } 45803494fc30STejun Heo if (idle) 45813494fc30STejun Heo continue; 45823494fc30STejun Heo 45833494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 45843494fc30STejun Heo 45853494fc30STejun Heo for_each_pwq(pwq, wq) { 45863494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 45873494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 45883494fc30STejun Heo show_pwq(pwq); 45893494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 45903494fc30STejun Heo } 45913494fc30STejun Heo } 45923494fc30STejun Heo 45933494fc30STejun Heo for_each_pool(pool, pi) { 45943494fc30STejun Heo struct worker *worker; 45953494fc30STejun Heo bool first = true; 45963494fc30STejun Heo 45973494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 45983494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 45993494fc30STejun Heo goto next_pool; 46003494fc30STejun Heo 46013494fc30STejun Heo pr_info("pool %d:", pool->id); 46023494fc30STejun Heo pr_cont_pool_info(pool); 46033494fc30STejun Heo pr_cont(" workers=%d", pool->nr_workers); 46043494fc30STejun Heo if (pool->manager) 46053494fc30STejun Heo pr_cont(" manager: %d", 46063494fc30STejun Heo task_pid_nr(pool->manager->task)); 46073494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 46083494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 46093494fc30STejun Heo task_pid_nr(worker->task)); 46103494fc30STejun Heo first = false; 46113494fc30STejun Heo } 46123494fc30STejun Heo pr_cont("\n"); 46133494fc30STejun Heo next_pool: 46143494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 46153494fc30STejun Heo } 46163494fc30STejun Heo 46173494fc30STejun Heo rcu_read_unlock_sched(); 46183494fc30STejun Heo } 46193494fc30STejun Heo 4620db7bccf4STejun Heo /* 4621db7bccf4STejun Heo * CPU hotplug. 4622db7bccf4STejun Heo * 4623e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4624112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4625706026c2STejun Heo * pool which make migrating pending and scheduled works very 4626e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 462794cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4628e22bee78STejun Heo * blocked draining impractical. 4629e22bee78STejun Heo * 463024647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4631628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4632628c78e7STejun Heo * cpu comes back online. 4633db7bccf4STejun Heo */ 4634db7bccf4STejun Heo 4635706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4636db7bccf4STejun Heo { 463738db41d9STejun Heo int cpu = smp_processor_id(); 46384ce62e9eSTejun Heo struct worker_pool *pool; 4639db7bccf4STejun Heo struct worker *worker; 4640db7bccf4STejun Heo 4641f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 464292f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 464394cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4644e22bee78STejun Heo 4645f2d5a0eeSTejun Heo /* 464692f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 464794cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 464894cf58bbSTejun Heo * except for the ones which are still executing works from 464994cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 465094cf58bbSTejun Heo * this, they may become diasporas. 4651f2d5a0eeSTejun Heo */ 4652da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4653403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4654db7bccf4STejun Heo 465524647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4656f2d5a0eeSTejun Heo 465794cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 465892f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4659e22bee78STejun Heo 4660e22bee78STejun Heo /* 4661eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4662eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4663eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4664eb283428SLai Jiangshan * from other cpus. 4665628c78e7STejun Heo */ 4666628c78e7STejun Heo schedule(); 4667628c78e7STejun Heo 4668628c78e7STejun Heo /* 4669eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4670eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4671eb283428SLai Jiangshan * and keep_working() are always true as long as the 4672eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4673eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4674eb283428SLai Jiangshan * are served by workers tied to the pool. 4675e22bee78STejun Heo */ 4676e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4677eb283428SLai Jiangshan 4678eb283428SLai Jiangshan /* 4679eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4680eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4681eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4682eb283428SLai Jiangshan */ 4683eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4684eb283428SLai Jiangshan wake_up_worker(pool); 4685eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4686eb283428SLai Jiangshan } 4687db7bccf4STejun Heo } 4688db7bccf4STejun Heo 4689bd7c089eSTejun Heo /** 4690bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4691bd7c089eSTejun Heo * @pool: pool of interest 4692bd7c089eSTejun Heo * 4693a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4694bd7c089eSTejun Heo */ 4695bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4696bd7c089eSTejun Heo { 4697a9ab775bSTejun Heo struct worker *worker; 4698bd7c089eSTejun Heo 469992f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4700bd7c089eSTejun Heo 4701bd7c089eSTejun Heo /* 4702a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4703a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4704a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4705a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4706a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4707bd7c089eSTejun Heo */ 4708da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4709a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4710a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4711a9ab775bSTejun Heo 4712a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 47133de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4714a9ab775bSTejun Heo 4715da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4716a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4717a9ab775bSTejun Heo 4718a9ab775bSTejun Heo /* 4719a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4720a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4721a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4722a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4723a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4724a9ab775bSTejun Heo * be bound before @pool->lock is released. 4725a9ab775bSTejun Heo */ 4726a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4727bd7c089eSTejun Heo wake_up_process(worker->task); 4728bd7c089eSTejun Heo 4729bd7c089eSTejun Heo /* 4730a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4731a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4732a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4733a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4734a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4735a9ab775bSTejun Heo * concurrency management. Note that when or whether 4736a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4737a9ab775bSTejun Heo * 4738a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4739a9ab775bSTejun Heo * tested without holding any lock in 4740a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4741a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4742a9ab775bSTejun Heo * management operations. 4743bd7c089eSTejun Heo */ 4744a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4745a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4746a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4747a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4748bd7c089eSTejun Heo } 4749a9ab775bSTejun Heo 4750a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4751bd7c089eSTejun Heo } 4752bd7c089eSTejun Heo 47537dbc725eSTejun Heo /** 47547dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 47557dbc725eSTejun Heo * @pool: unbound pool of interest 47567dbc725eSTejun Heo * @cpu: the CPU which is coming up 47577dbc725eSTejun Heo * 47587dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 47597dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 47607dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 47617dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 47627dbc725eSTejun Heo */ 47637dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 47647dbc725eSTejun Heo { 47657dbc725eSTejun Heo static cpumask_t cpumask; 47667dbc725eSTejun Heo struct worker *worker; 47677dbc725eSTejun Heo 476892f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 47697dbc725eSTejun Heo 47707dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 47717dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 47727dbc725eSTejun Heo return; 47737dbc725eSTejun Heo 47747dbc725eSTejun Heo /* is @cpu the only online CPU? */ 47757dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 47767dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 47777dbc725eSTejun Heo return; 47787dbc725eSTejun Heo 47797dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4780da028469SLai Jiangshan for_each_pool_worker(worker, pool) 47817dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 47827dbc725eSTejun Heo pool->attrs->cpumask) < 0); 47837dbc725eSTejun Heo } 47847dbc725eSTejun Heo 47858db25e78STejun Heo /* 47868db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 47878db25e78STejun Heo * This will be registered high priority CPU notifier. 47888db25e78STejun Heo */ 47890db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 47901da177e4SLinus Torvalds unsigned long action, 47911da177e4SLinus Torvalds void *hcpu) 47921da177e4SLinus Torvalds { 4793d84ff051STejun Heo int cpu = (unsigned long)hcpu; 47944ce62e9eSTejun Heo struct worker_pool *pool; 47954c16bd32STejun Heo struct workqueue_struct *wq; 47967dbc725eSTejun Heo int pi; 47971da177e4SLinus Torvalds 47988db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 47993af24433SOleg Nesterov case CPU_UP_PREPARE: 4800f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 48013ce63377STejun Heo if (pool->nr_workers) 48023ce63377STejun Heo continue; 4803051e1850SLai Jiangshan if (!create_worker(pool)) 48043ce63377STejun Heo return NOTIFY_BAD; 48053af24433SOleg Nesterov } 48061da177e4SLinus Torvalds break; 48071da177e4SLinus Torvalds 480865758202STejun Heo case CPU_DOWN_FAILED: 480965758202STejun Heo case CPU_ONLINE: 481068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 48117dbc725eSTejun Heo 48127dbc725eSTejun Heo for_each_pool(pool, pi) { 481392f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 481494cf58bbSTejun Heo 4815f05b558dSLai Jiangshan if (pool->cpu == cpu) 481694cf58bbSTejun Heo rebind_workers(pool); 4817f05b558dSLai Jiangshan else if (pool->cpu < 0) 48187dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 481994cf58bbSTejun Heo 482092f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 482194cf58bbSTejun Heo } 48227dbc725eSTejun Heo 48234c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 48244c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 48254c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 48264c16bd32STejun Heo 482768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 48288db25e78STejun Heo break; 482965758202STejun Heo } 483065758202STejun Heo return NOTIFY_OK; 483165758202STejun Heo } 483265758202STejun Heo 483365758202STejun Heo /* 483465758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 483565758202STejun Heo * This will be registered as low priority CPU notifier. 483665758202STejun Heo */ 48370db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 483865758202STejun Heo unsigned long action, 483965758202STejun Heo void *hcpu) 484065758202STejun Heo { 4841d84ff051STejun Heo int cpu = (unsigned long)hcpu; 48428db25e78STejun Heo struct work_struct unbind_work; 48434c16bd32STejun Heo struct workqueue_struct *wq; 48448db25e78STejun Heo 484565758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 484665758202STejun Heo case CPU_DOWN_PREPARE: 48474c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4848706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 48497635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 48504c16bd32STejun Heo 48514c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 48524c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 48534c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 48544c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 48554c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 48564c16bd32STejun Heo 48574c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 48588db25e78STejun Heo flush_work(&unbind_work); 4859440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 48608db25e78STejun Heo break; 486165758202STejun Heo } 486265758202STejun Heo return NOTIFY_OK; 486365758202STejun Heo } 486465758202STejun Heo 48652d3854a3SRusty Russell #ifdef CONFIG_SMP 48668ccad40dSRusty Russell 48672d3854a3SRusty Russell struct work_for_cpu { 4868ed48ece2STejun Heo struct work_struct work; 48692d3854a3SRusty Russell long (*fn)(void *); 48702d3854a3SRusty Russell void *arg; 48712d3854a3SRusty Russell long ret; 48722d3854a3SRusty Russell }; 48732d3854a3SRusty Russell 4874ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 48752d3854a3SRusty Russell { 4876ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4877ed48ece2STejun Heo 48782d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 48792d3854a3SRusty Russell } 48802d3854a3SRusty Russell 48812d3854a3SRusty Russell /** 48822d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 48832d3854a3SRusty Russell * @cpu: the cpu to run on 48842d3854a3SRusty Russell * @fn: the function to run 48852d3854a3SRusty Russell * @arg: the function arg 48862d3854a3SRusty Russell * 488731ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 48886b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4889d185af30SYacine Belkadi * 4890d185af30SYacine Belkadi * Return: The value @fn returns. 48912d3854a3SRusty Russell */ 4892d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 48932d3854a3SRusty Russell { 4894ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 48952d3854a3SRusty Russell 4896ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4897ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 489812997d1aSBjorn Helgaas flush_work(&wfc.work); 4899440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 49002d3854a3SRusty Russell return wfc.ret; 49012d3854a3SRusty Russell } 49022d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 49032d3854a3SRusty Russell #endif /* CONFIG_SMP */ 49042d3854a3SRusty Russell 4905a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4906e7577c50SRusty Russell 4907a0a1a5fdSTejun Heo /** 4908a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4909a0a1a5fdSTejun Heo * 491058a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4911c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4912706026c2STejun Heo * pool->worklist. 4913a0a1a5fdSTejun Heo * 4914a0a1a5fdSTejun Heo * CONTEXT: 4915a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4916a0a1a5fdSTejun Heo */ 4917a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4918a0a1a5fdSTejun Heo { 491924b8a847STejun Heo struct workqueue_struct *wq; 492024b8a847STejun Heo struct pool_workqueue *pwq; 4921a0a1a5fdSTejun Heo 492268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4923a0a1a5fdSTejun Heo 49246183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4925a0a1a5fdSTejun Heo workqueue_freezing = true; 4926a0a1a5fdSTejun Heo 492724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4928a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4929699ce097STejun Heo for_each_pwq(pwq, wq) 4930699ce097STejun Heo pwq_adjust_max_active(pwq); 4931a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4932a1056305STejun Heo } 49335bcab335STejun Heo 493468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4935a0a1a5fdSTejun Heo } 4936a0a1a5fdSTejun Heo 4937a0a1a5fdSTejun Heo /** 493858a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4939a0a1a5fdSTejun Heo * 4940a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4941a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4942a0a1a5fdSTejun Heo * 4943a0a1a5fdSTejun Heo * CONTEXT: 494468e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4945a0a1a5fdSTejun Heo * 4946d185af30SYacine Belkadi * Return: 494758a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 494858a69cb4STejun Heo * is complete. 4949a0a1a5fdSTejun Heo */ 4950a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4951a0a1a5fdSTejun Heo { 4952a0a1a5fdSTejun Heo bool busy = false; 495324b8a847STejun Heo struct workqueue_struct *wq; 495424b8a847STejun Heo struct pool_workqueue *pwq; 4955a0a1a5fdSTejun Heo 495668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4957a0a1a5fdSTejun Heo 49586183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4959a0a1a5fdSTejun Heo 496024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 496124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 496224b8a847STejun Heo continue; 4963a0a1a5fdSTejun Heo /* 4964a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4965a0a1a5fdSTejun Heo * to peek without lock. 4966a0a1a5fdSTejun Heo */ 496788109453SLai Jiangshan rcu_read_lock_sched(); 496824b8a847STejun Heo for_each_pwq(pwq, wq) { 49696183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4970112202d9STejun Heo if (pwq->nr_active) { 4971a0a1a5fdSTejun Heo busy = true; 497288109453SLai Jiangshan rcu_read_unlock_sched(); 4973a0a1a5fdSTejun Heo goto out_unlock; 4974a0a1a5fdSTejun Heo } 4975a0a1a5fdSTejun Heo } 497688109453SLai Jiangshan rcu_read_unlock_sched(); 4977a0a1a5fdSTejun Heo } 4978a0a1a5fdSTejun Heo out_unlock: 497968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4980a0a1a5fdSTejun Heo return busy; 4981a0a1a5fdSTejun Heo } 4982a0a1a5fdSTejun Heo 4983a0a1a5fdSTejun Heo /** 4984a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4985a0a1a5fdSTejun Heo * 4986a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4987706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4988a0a1a5fdSTejun Heo * 4989a0a1a5fdSTejun Heo * CONTEXT: 4990a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4991a0a1a5fdSTejun Heo */ 4992a0a1a5fdSTejun Heo void thaw_workqueues(void) 4993a0a1a5fdSTejun Heo { 499424b8a847STejun Heo struct workqueue_struct *wq; 499524b8a847STejun Heo struct pool_workqueue *pwq; 4996a0a1a5fdSTejun Heo 499768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4998a0a1a5fdSTejun Heo 4999a0a1a5fdSTejun Heo if (!workqueue_freezing) 5000a0a1a5fdSTejun Heo goto out_unlock; 5001a0a1a5fdSTejun Heo 500274b414eaSLai Jiangshan workqueue_freezing = false; 500324b8a847STejun Heo 500424b8a847STejun Heo /* restore max_active and repopulate worklist */ 500524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5006a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5007699ce097STejun Heo for_each_pwq(pwq, wq) 5008699ce097STejun Heo pwq_adjust_max_active(pwq); 5009a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 501024b8a847STejun Heo } 501124b8a847STejun Heo 5012a0a1a5fdSTejun Heo out_unlock: 501368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5014a0a1a5fdSTejun Heo } 5015a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5016a0a1a5fdSTejun Heo 5017bce90380STejun Heo static void __init wq_numa_init(void) 5018bce90380STejun Heo { 5019bce90380STejun Heo cpumask_var_t *tbl; 5020bce90380STejun Heo int node, cpu; 5021bce90380STejun Heo 5022bce90380STejun Heo if (num_possible_nodes() <= 1) 5023bce90380STejun Heo return; 5024bce90380STejun Heo 5025d55262c4STejun Heo if (wq_disable_numa) { 5026d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5027d55262c4STejun Heo return; 5028d55262c4STejun Heo } 5029d55262c4STejun Heo 50304c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 50314c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 50324c16bd32STejun Heo 5033bce90380STejun Heo /* 5034bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5035bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5036bce90380STejun Heo * fully initialized by now. 5037bce90380STejun Heo */ 5038ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5039bce90380STejun Heo BUG_ON(!tbl); 5040bce90380STejun Heo 5041bce90380STejun Heo for_each_node(node) 50425a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 50431be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5044bce90380STejun Heo 5045bce90380STejun Heo for_each_possible_cpu(cpu) { 5046bce90380STejun Heo node = cpu_to_node(cpu); 5047bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5048bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5049bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5050bce90380STejun Heo return; 5051bce90380STejun Heo } 5052bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5053bce90380STejun Heo } 5054bce90380STejun Heo 5055bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5056bce90380STejun Heo wq_numa_enabled = true; 5057bce90380STejun Heo } 5058bce90380STejun Heo 50596ee0578bSSuresh Siddha static int __init init_workqueues(void) 50601da177e4SLinus Torvalds { 50617a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 50627a4e344cSTejun Heo int i, cpu; 5063c34056a3STejun Heo 5064e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5065e904e6c2STejun Heo 5066e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5067e904e6c2STejun Heo 506865758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 5069a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 50708b03ae3cSTejun Heo 5071bce90380STejun Heo wq_numa_init(); 5072bce90380STejun Heo 5073706026c2STejun Heo /* initialize CPU pools */ 507429c91e99STejun Heo for_each_possible_cpu(cpu) { 50754ce62e9eSTejun Heo struct worker_pool *pool; 50768b03ae3cSTejun Heo 50777a4e344cSTejun Heo i = 0; 5078f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 50797a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5080ec22ca5eSTejun Heo pool->cpu = cpu; 50817a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 50827a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5083f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 50847a4e344cSTejun Heo 50859daf9e67STejun Heo /* alloc pool ID */ 508668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 50879daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 508868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 50894ce62e9eSTejun Heo } 50908b03ae3cSTejun Heo } 50918b03ae3cSTejun Heo 5092e22bee78STejun Heo /* create the initial worker */ 509329c91e99STejun Heo for_each_online_cpu(cpu) { 50944ce62e9eSTejun Heo struct worker_pool *pool; 5095e22bee78STejun Heo 5096f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 509724647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5098051e1850SLai Jiangshan BUG_ON(!create_worker(pool)); 5099e22bee78STejun Heo } 51004ce62e9eSTejun Heo } 5101e22bee78STejun Heo 51028a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 510329c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 510429c91e99STejun Heo struct workqueue_attrs *attrs; 510529c91e99STejun Heo 510629c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 510729c91e99STejun Heo attrs->nice = std_nice[i]; 510829c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 51098a2b7538STejun Heo 51108a2b7538STejun Heo /* 51118a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 51128a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 51138a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 51148a2b7538STejun Heo */ 51158a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 51168a2b7538STejun Heo attrs->nice = std_nice[i]; 51178a2b7538STejun Heo attrs->no_numa = true; 51188a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 511929c91e99STejun Heo } 512029c91e99STejun Heo 5121d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 51221aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5123d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5124f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5125f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 512624d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 512724d51addSTejun Heo WQ_FREEZABLE, 0); 51280668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 51290668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 51300668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 51310668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 51320668106cSViresh Kumar 0); 51331aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 51340668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 51350668106cSViresh Kumar !system_power_efficient_wq || 51360668106cSViresh Kumar !system_freezable_power_efficient_wq); 51376ee0578bSSuresh Siddha return 0; 51381da177e4SLinus Torvalds } 51396ee0578bSSuresh Siddha early_initcall(init_workqueues); 5140