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 */ 16292f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 16392f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 16460f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 165e19e397aSTejun Heo 1667cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 167e19e397aSTejun Heo 1687a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16968e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17068e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1717a4e344cSTejun Heo 172e19e397aSTejun Heo /* 173e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 174e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 175e19e397aSTejun Heo * cacheline. 176e19e397aSTejun Heo */ 177e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17829c91e99STejun Heo 17929c91e99STejun Heo /* 18029c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18129c91e99STejun Heo * from get_work_pool(). 18229c91e99STejun Heo */ 18329c91e99STejun Heo struct rcu_head rcu; 1848b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1858b03ae3cSTejun Heo 1868b03ae3cSTejun Heo /* 187112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 188112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 189112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 190112202d9STejun Heo * number of flag bits. 1911da177e4SLinus Torvalds */ 192112202d9STejun Heo struct pool_workqueue { 193bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1944690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19573f53c4aSTejun Heo int work_color; /* L: current color */ 19673f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1978864b4e5STejun Heo int refcnt; /* L: reference count */ 19873f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19973f53c4aSTejun Heo /* L: nr of in_flight works */ 2001e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 201a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2021e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2033c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2042e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2058864b4e5STejun Heo 2068864b4e5STejun Heo /* 2078864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2088864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2098864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 210b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2118864b4e5STejun Heo */ 2128864b4e5STejun Heo struct work_struct unbound_release_work; 2138864b4e5STejun Heo struct rcu_head rcu; 214e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2151da177e4SLinus Torvalds 2161da177e4SLinus Torvalds /* 21773f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21873f53c4aSTejun Heo */ 21973f53c4aSTejun Heo struct wq_flusher { 2203c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2213c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22273f53c4aSTejun Heo struct completion done; /* flush completion */ 22373f53c4aSTejun Heo }; 2241da177e4SLinus Torvalds 225226223abSTejun Heo struct wq_device; 226226223abSTejun Heo 22773f53c4aSTejun Heo /* 228c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 229c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2301da177e4SLinus Torvalds */ 2311da177e4SLinus Torvalds struct workqueue_struct { 2323c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 23368e13a67SLai Jiangshan struct list_head list; /* PL: list of all workqueues */ 23473f53c4aSTejun Heo 2353c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2363c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2373c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 238112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2393c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2403c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2413c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24273f53c4aSTejun Heo 2432e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 244e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 245e22bee78STejun Heo 24687fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 247a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 248226223abSTejun Heo 2496029a918STejun Heo struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ 2504c16bd32STejun Heo struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ 2516029a918STejun Heo 252226223abSTejun Heo #ifdef CONFIG_SYSFS 253226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 254226223abSTejun Heo #endif 2554e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2564e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2574e6045f1SJohannes Berg #endif 258ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2592728fd2fSTejun Heo 2602728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2612728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2622728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 263df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2641da177e4SLinus Torvalds }; 2651da177e4SLinus Torvalds 266e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 267e904e6c2STejun Heo 268bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 269bce90380STejun Heo /* possible CPUs of each node */ 270bce90380STejun Heo 271d55262c4STejun Heo static bool wq_disable_numa; 272d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 273d55262c4STejun Heo 274cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 275cee22a15SViresh Kumar #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT 276cee22a15SViresh Kumar static bool wq_power_efficient = true; 277cee22a15SViresh Kumar #else 278cee22a15SViresh Kumar static bool wq_power_efficient; 279cee22a15SViresh Kumar #endif 280cee22a15SViresh Kumar 281cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 282cee22a15SViresh Kumar 283bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 284bce90380STejun Heo 2854c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2864c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2874c16bd32STejun Heo 28868e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2892e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2905bcab335STejun Heo 29168e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 29268e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2937d19c5ceSTejun Heo 2947d19c5ceSTejun Heo /* the per-cpu worker pools */ 2957d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2967d19c5ceSTejun Heo cpu_worker_pools); 2977d19c5ceSTejun Heo 29868e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2997d19c5ceSTejun Heo 30068e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 30129c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 30229c91e99STejun Heo 303c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 30429c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 30529c91e99STejun Heo 3068a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3078a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3088a2b7538STejun Heo 309d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 310ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 311044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3121aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 313044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 314d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 315044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 316f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 317044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 31824d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3190668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3200668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3210668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3220668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 323d320c038STejun Heo 3247d19c5ceSTejun Heo static int worker_thread(void *__worker); 3257d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3267d19c5ceSTejun Heo const struct workqueue_attrs *from); 3277d19c5ceSTejun Heo 32897bd2347STejun Heo #define CREATE_TRACE_POINTS 32997bd2347STejun Heo #include <trace/events/workqueue.h> 33097bd2347STejun Heo 33168e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3325bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 33368e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 33468e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3355bcab335STejun Heo 336b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 33776af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 338b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 339b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 34076af4d93STejun Heo 341f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 342f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 343f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3447a62c2c8STejun Heo (pool)++) 3454ce62e9eSTejun Heo 34649e3cf44STejun Heo /** 34717116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 34817116969STejun Heo * @pool: iteration cursor 349611c92a0STejun Heo * @pi: integer used for iteration 350fa1b54e6STejun Heo * 35168e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 35268e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 35368e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 354fa1b54e6STejun Heo * 355fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 356fa1b54e6STejun Heo * ignored. 35717116969STejun Heo */ 358611c92a0STejun Heo #define for_each_pool(pool, pi) \ 359611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 36068e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 361fa1b54e6STejun Heo else 36217116969STejun Heo 36317116969STejun Heo /** 364822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 365822d8405STejun Heo * @worker: iteration cursor 366822d8405STejun Heo * @pool: worker_pool to iterate workers of 367822d8405STejun Heo * 36892f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 369822d8405STejun Heo * 370822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 371822d8405STejun Heo * ignored. 372822d8405STejun Heo */ 373da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 374da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 37592f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 376822d8405STejun Heo else 377822d8405STejun Heo 378822d8405STejun Heo /** 37949e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 38049e3cf44STejun Heo * @pwq: iteration cursor 38149e3cf44STejun Heo * @wq: the target workqueue 38276af4d93STejun Heo * 383b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 384794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 385794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 38676af4d93STejun Heo * 38776af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 38876af4d93STejun Heo * ignored. 38949e3cf44STejun Heo */ 39049e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 39176af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 392b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 39376af4d93STejun Heo else 394f3421797STejun Heo 395dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 396dc186ad7SThomas Gleixner 397dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 398dc186ad7SThomas Gleixner 39999777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 40099777288SStanislaw Gruszka { 40199777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 40299777288SStanislaw Gruszka } 40399777288SStanislaw Gruszka 404dc186ad7SThomas Gleixner /* 405dc186ad7SThomas Gleixner * fixup_init is called when: 406dc186ad7SThomas Gleixner * - an active object is initialized 407dc186ad7SThomas Gleixner */ 408dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 409dc186ad7SThomas Gleixner { 410dc186ad7SThomas Gleixner struct work_struct *work = addr; 411dc186ad7SThomas Gleixner 412dc186ad7SThomas Gleixner switch (state) { 413dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 414dc186ad7SThomas Gleixner cancel_work_sync(work); 415dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 416dc186ad7SThomas Gleixner return 1; 417dc186ad7SThomas Gleixner default: 418dc186ad7SThomas Gleixner return 0; 419dc186ad7SThomas Gleixner } 420dc186ad7SThomas Gleixner } 421dc186ad7SThomas Gleixner 422dc186ad7SThomas Gleixner /* 423dc186ad7SThomas Gleixner * fixup_activate is called when: 424dc186ad7SThomas Gleixner * - an active object is activated 425dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 426dc186ad7SThomas Gleixner */ 427dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 428dc186ad7SThomas Gleixner { 429dc186ad7SThomas Gleixner struct work_struct *work = addr; 430dc186ad7SThomas Gleixner 431dc186ad7SThomas Gleixner switch (state) { 432dc186ad7SThomas Gleixner 433dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 434dc186ad7SThomas Gleixner /* 435dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 436dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 437dc186ad7SThomas Gleixner * is tracked in the object tracker. 438dc186ad7SThomas Gleixner */ 43922df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 440dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 441dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 442dc186ad7SThomas Gleixner return 0; 443dc186ad7SThomas Gleixner } 444dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 445dc186ad7SThomas Gleixner return 0; 446dc186ad7SThomas Gleixner 447dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 448dc186ad7SThomas Gleixner WARN_ON(1); 449dc186ad7SThomas Gleixner 450dc186ad7SThomas Gleixner default: 451dc186ad7SThomas Gleixner return 0; 452dc186ad7SThomas Gleixner } 453dc186ad7SThomas Gleixner } 454dc186ad7SThomas Gleixner 455dc186ad7SThomas Gleixner /* 456dc186ad7SThomas Gleixner * fixup_free is called when: 457dc186ad7SThomas Gleixner * - an active object is freed 458dc186ad7SThomas Gleixner */ 459dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 460dc186ad7SThomas Gleixner { 461dc186ad7SThomas Gleixner struct work_struct *work = addr; 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner switch (state) { 464dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 465dc186ad7SThomas Gleixner cancel_work_sync(work); 466dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 467dc186ad7SThomas Gleixner return 1; 468dc186ad7SThomas Gleixner default: 469dc186ad7SThomas Gleixner return 0; 470dc186ad7SThomas Gleixner } 471dc186ad7SThomas Gleixner } 472dc186ad7SThomas Gleixner 473dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 474dc186ad7SThomas Gleixner .name = "work_struct", 47599777288SStanislaw Gruszka .debug_hint = work_debug_hint, 476dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 477dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 478dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 479dc186ad7SThomas Gleixner }; 480dc186ad7SThomas Gleixner 481dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 482dc186ad7SThomas Gleixner { 483dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 484dc186ad7SThomas Gleixner } 485dc186ad7SThomas Gleixner 486dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 487dc186ad7SThomas Gleixner { 488dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 489dc186ad7SThomas Gleixner } 490dc186ad7SThomas Gleixner 491dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 492dc186ad7SThomas Gleixner { 493dc186ad7SThomas Gleixner if (onstack) 494dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 495dc186ad7SThomas Gleixner else 496dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 499dc186ad7SThomas Gleixner 500dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 501dc186ad7SThomas Gleixner { 502dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 503dc186ad7SThomas Gleixner } 504dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 505dc186ad7SThomas Gleixner 506ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 507ea2e64f2SThomas Gleixner { 508ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 509ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 510ea2e64f2SThomas Gleixner } 511ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 512ea2e64f2SThomas Gleixner 513dc186ad7SThomas Gleixner #else 514dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 515dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 516dc186ad7SThomas Gleixner #endif 517dc186ad7SThomas Gleixner 5184e8b22bdSLi Bin /** 5194e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5204e8b22bdSLi Bin * @pool: the pool pointer of interest 5214e8b22bdSLi Bin * 5224e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5234e8b22bdSLi Bin * successfully, -errno on failure. 5244e8b22bdSLi Bin */ 5259daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5269daf9e67STejun Heo { 5279daf9e67STejun Heo int ret; 5289daf9e67STejun Heo 52968e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5305bcab335STejun Heo 5314e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5324e8b22bdSLi Bin GFP_KERNEL); 533229641a6STejun Heo if (ret >= 0) { 534e68035fbSTejun Heo pool->id = ret; 535229641a6STejun Heo return 0; 536229641a6STejun Heo } 5379daf9e67STejun Heo return ret; 5389daf9e67STejun Heo } 5399daf9e67STejun Heo 54076af4d93STejun Heo /** 541df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 542df2d5ae4STejun Heo * @wq: the target workqueue 543df2d5ae4STejun Heo * @node: the node ID 544df2d5ae4STejun Heo * 545df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 546df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 547df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 548d185af30SYacine Belkadi * 549d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 550df2d5ae4STejun Heo */ 551df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 552df2d5ae4STejun Heo int node) 553df2d5ae4STejun Heo { 554df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 555df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 556df2d5ae4STejun Heo } 557df2d5ae4STejun Heo 55873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 55973f53c4aSTejun Heo { 56073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 56173f53c4aSTejun Heo } 56273f53c4aSTejun Heo 56373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 56473f53c4aSTejun Heo { 56573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 56673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 56773f53c4aSTejun Heo } 56873f53c4aSTejun Heo 56973f53c4aSTejun Heo static int work_next_color(int color) 57073f53c4aSTejun Heo { 57173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 572a848e3b6SOleg Nesterov } 573a848e3b6SOleg Nesterov 5744594bf15SDavid Howells /* 575112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 576112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5777c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5787a22ad75STejun Heo * 579112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 580112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 581bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 582bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5837a22ad75STejun Heo * 584112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5857c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 586112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5877c3eed5cSTejun Heo * available only while the work item is queued. 588bbb68dfaSTejun Heo * 589bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 590bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 591bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 592bbb68dfaSTejun Heo * try to steal the PENDING bit. 5934594bf15SDavid Howells */ 5947a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5957a22ad75STejun Heo unsigned long flags) 5967a22ad75STejun Heo { 5976183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5987a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5997a22ad75STejun Heo } 6007a22ad75STejun Heo 601112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6024690c4abSTejun Heo unsigned long extra_flags) 603365970a1SDavid Howells { 604112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 605112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 606365970a1SDavid Howells } 607365970a1SDavid Howells 6084468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6094468a00fSLai Jiangshan int pool_id) 6104468a00fSLai Jiangshan { 6114468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6124468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6134468a00fSLai Jiangshan } 6144468a00fSLai Jiangshan 6157c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6167c3eed5cSTejun Heo int pool_id) 6174d707b9fSOleg Nesterov { 61823657bb1STejun Heo /* 61923657bb1STejun Heo * The following wmb is paired with the implied mb in 62023657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 62123657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 62223657bb1STejun Heo * owner. 62323657bb1STejun Heo */ 62423657bb1STejun Heo smp_wmb(); 6257c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6264d707b9fSOleg Nesterov } 6274d707b9fSOleg Nesterov 6287a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 629365970a1SDavid Howells { 6307c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6317c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6327a22ad75STejun Heo } 6337a22ad75STejun Heo 634112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6357a22ad75STejun Heo { 636e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6377a22ad75STejun Heo 638112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 639e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 640e120153dSTejun Heo else 641e120153dSTejun Heo return NULL; 6427a22ad75STejun Heo } 6437a22ad75STejun Heo 6447c3eed5cSTejun Heo /** 6457c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6467c3eed5cSTejun Heo * @work: the work item of interest 6477c3eed5cSTejun Heo * 64868e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 64968e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 65068e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 651fa1b54e6STejun Heo * 652fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 653fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 654fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 655fa1b54e6STejun Heo * returned pool is and stays online. 656d185af30SYacine Belkadi * 657d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6587c3eed5cSTejun Heo */ 6597c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6607a22ad75STejun Heo { 661e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6627c3eed5cSTejun Heo int pool_id; 6637a22ad75STejun Heo 66468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 665fa1b54e6STejun Heo 666112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 667112202d9STejun Heo return ((struct pool_workqueue *) 6687c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6697a22ad75STejun Heo 6707c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6717c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6727a22ad75STejun Heo return NULL; 6737a22ad75STejun Heo 674fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6757c3eed5cSTejun Heo } 6767c3eed5cSTejun Heo 6777c3eed5cSTejun Heo /** 6787c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6797c3eed5cSTejun Heo * @work: the work item of interest 6807c3eed5cSTejun Heo * 681d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6827c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6837c3eed5cSTejun Heo */ 6847c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6857c3eed5cSTejun Heo { 68654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6877c3eed5cSTejun Heo 688112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 689112202d9STejun Heo return ((struct pool_workqueue *) 69054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 69154d5b7d0SLai Jiangshan 69254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6937c3eed5cSTejun Heo } 6947c3eed5cSTejun Heo 695bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 696bbb68dfaSTejun Heo { 6977c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 698bbb68dfaSTejun Heo 6997c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7007c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 701bbb68dfaSTejun Heo } 702bbb68dfaSTejun Heo 703bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 704bbb68dfaSTejun Heo { 705bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 706bbb68dfaSTejun Heo 707112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 708bbb68dfaSTejun Heo } 709bbb68dfaSTejun Heo 710e22bee78STejun Heo /* 7113270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7123270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 713d565ed63STejun Heo * they're being called with pool->lock held. 714e22bee78STejun Heo */ 715e22bee78STejun Heo 71663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 717649027d7STejun Heo { 718e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 719649027d7STejun Heo } 720649027d7STejun Heo 721e22bee78STejun Heo /* 722e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 723e22bee78STejun Heo * running workers. 724974271c4STejun Heo * 725974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 726706026c2STejun Heo * function will always return %true for unbound pools as long as the 727974271c4STejun Heo * worklist isn't empty. 728e22bee78STejun Heo */ 72963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 730e22bee78STejun Heo { 73163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 732e22bee78STejun Heo } 733e22bee78STejun Heo 734e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 73563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 736e22bee78STejun Heo { 73763d95a91STejun Heo return pool->nr_idle; 738e22bee78STejun Heo } 739e22bee78STejun Heo 740e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 74163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 742e22bee78STejun Heo { 743e19e397aSTejun Heo return !list_empty(&pool->worklist) && 744e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 745e22bee78STejun Heo } 746e22bee78STejun Heo 747e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 74863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 749e22bee78STejun Heo { 75063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 751e22bee78STejun Heo } 752e22bee78STejun Heo 753e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 75463d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 755e22bee78STejun Heo { 75634a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 75763d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 75863d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 759e22bee78STejun Heo 760e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 761e22bee78STejun Heo } 762e22bee78STejun Heo 763e22bee78STejun Heo /* 764e22bee78STejun Heo * Wake up functions. 765e22bee78STejun Heo */ 766e22bee78STejun Heo 7671037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 7681037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 7697e11629dSTejun Heo { 77063d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7717e11629dSTejun Heo return NULL; 7727e11629dSTejun Heo 77363d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7747e11629dSTejun Heo } 7757e11629dSTejun Heo 7767e11629dSTejun Heo /** 7777e11629dSTejun Heo * wake_up_worker - wake up an idle worker 77863d95a91STejun Heo * @pool: worker pool to wake worker from 7797e11629dSTejun Heo * 78063d95a91STejun Heo * Wake up the first idle worker of @pool. 7817e11629dSTejun Heo * 7827e11629dSTejun Heo * CONTEXT: 783d565ed63STejun Heo * spin_lock_irq(pool->lock). 7847e11629dSTejun Heo */ 78563d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7867e11629dSTejun Heo { 7871037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 7887e11629dSTejun Heo 7897e11629dSTejun Heo if (likely(worker)) 7907e11629dSTejun Heo wake_up_process(worker->task); 7917e11629dSTejun Heo } 7927e11629dSTejun Heo 7934690c4abSTejun Heo /** 794e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 795e22bee78STejun Heo * @task: task waking up 796e22bee78STejun Heo * @cpu: CPU @task is waking up to 797e22bee78STejun Heo * 798e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 799e22bee78STejun Heo * being awoken. 800e22bee78STejun Heo * 801e22bee78STejun Heo * CONTEXT: 802e22bee78STejun Heo * spin_lock_irq(rq->lock) 803e22bee78STejun Heo */ 804d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 805e22bee78STejun Heo { 806e22bee78STejun Heo struct worker *worker = kthread_data(task); 807e22bee78STejun Heo 80836576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 809ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 810e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 811e22bee78STejun Heo } 81236576000SJoonsoo Kim } 813e22bee78STejun Heo 814e22bee78STejun Heo /** 815e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 816e22bee78STejun Heo * @task: task going to sleep 817e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 818e22bee78STejun Heo * 819e22bee78STejun Heo * This function is called during schedule() when a busy worker is 820e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 821e22bee78STejun Heo * returning pointer to its task. 822e22bee78STejun Heo * 823e22bee78STejun Heo * CONTEXT: 824e22bee78STejun Heo * spin_lock_irq(rq->lock) 825e22bee78STejun Heo * 826d185af30SYacine Belkadi * Return: 827e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 828e22bee78STejun Heo */ 829d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 830e22bee78STejun Heo { 831e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 832111c225aSTejun Heo struct worker_pool *pool; 833e22bee78STejun Heo 834111c225aSTejun Heo /* 835111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 836111c225aSTejun Heo * workers, also reach here, let's not access anything before 837111c225aSTejun Heo * checking NOT_RUNNING. 838111c225aSTejun Heo */ 8392d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 840e22bee78STejun Heo return NULL; 841e22bee78STejun Heo 842111c225aSTejun Heo pool = worker->pool; 843111c225aSTejun Heo 844e22bee78STejun Heo /* this can only happen on the local cpu */ 84592b69f50SLai Jiangshan if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) 8466183c009STejun Heo return NULL; 847e22bee78STejun Heo 848e22bee78STejun Heo /* 849e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 850e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 851e22bee78STejun Heo * Please read comment there. 852e22bee78STejun Heo * 853628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 854628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 855628c78e7STejun Heo * disabled, which in turn means that none else could be 856d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 857628c78e7STejun Heo * lock is safe. 858e22bee78STejun Heo */ 859e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 860e19e397aSTejun Heo !list_empty(&pool->worklist)) 8611037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 862e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 863e22bee78STejun Heo } 864e22bee78STejun Heo 865e22bee78STejun Heo /** 866e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 867cb444766STejun Heo * @worker: self 868d302f017STejun Heo * @flags: flags to set 869d302f017STejun Heo * 870228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 871d302f017STejun Heo * 872cb444766STejun Heo * CONTEXT: 873d565ed63STejun Heo * spin_lock_irq(pool->lock) 874d302f017STejun Heo */ 875228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 876d302f017STejun Heo { 877bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 878e22bee78STejun Heo 879cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 880cb444766STejun Heo 881228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 882e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 883e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 884e19e397aSTejun Heo atomic_dec(&pool->nr_running); 885e22bee78STejun Heo } 886e22bee78STejun Heo 887d302f017STejun Heo worker->flags |= flags; 888d302f017STejun Heo } 889d302f017STejun Heo 890d302f017STejun Heo /** 891e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 892cb444766STejun Heo * @worker: self 893d302f017STejun Heo * @flags: flags to clear 894d302f017STejun Heo * 895e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 896d302f017STejun Heo * 897cb444766STejun Heo * CONTEXT: 898d565ed63STejun Heo * spin_lock_irq(pool->lock) 899d302f017STejun Heo */ 900d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 901d302f017STejun Heo { 90263d95a91STejun Heo struct worker_pool *pool = worker->pool; 903e22bee78STejun Heo unsigned int oflags = worker->flags; 904e22bee78STejun Heo 905cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 906cb444766STejun Heo 907d302f017STejun Heo worker->flags &= ~flags; 908e22bee78STejun Heo 90942c025f3STejun Heo /* 91042c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 91142c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 91242c025f3STejun Heo * of multiple flags, not a single flag. 91342c025f3STejun Heo */ 914e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 915e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 916e19e397aSTejun Heo atomic_inc(&pool->nr_running); 917d302f017STejun Heo } 918d302f017STejun Heo 919d302f017STejun Heo /** 9208cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 921c9e7cf27STejun Heo * @pool: pool of interest 9228cca0eeaSTejun Heo * @work: work to find worker for 9238cca0eeaSTejun Heo * 924c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 925c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 926a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 927a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 928a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 929a2c1c57bSTejun Heo * being executed. 930a2c1c57bSTejun Heo * 931a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 932a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 933a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 934a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 935a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 936a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 937a2c1c57bSTejun Heo * 938c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 939c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 940c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 941c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 942c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 943c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9448cca0eeaSTejun Heo * 9458cca0eeaSTejun Heo * CONTEXT: 946d565ed63STejun Heo * spin_lock_irq(pool->lock). 9478cca0eeaSTejun Heo * 948d185af30SYacine Belkadi * Return: 949d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9508cca0eeaSTejun Heo * otherwise. 9518cca0eeaSTejun Heo */ 952c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9538cca0eeaSTejun Heo struct work_struct *work) 9548cca0eeaSTejun Heo { 95542f8570fSSasha Levin struct worker *worker; 95642f8570fSSasha Levin 957b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 958a2c1c57bSTejun Heo (unsigned long)work) 959a2c1c57bSTejun Heo if (worker->current_work == work && 960a2c1c57bSTejun Heo worker->current_func == work->func) 96142f8570fSSasha Levin return worker; 96242f8570fSSasha Levin 96342f8570fSSasha Levin return NULL; 9648cca0eeaSTejun Heo } 9658cca0eeaSTejun Heo 9668cca0eeaSTejun Heo /** 967bf4ede01STejun Heo * move_linked_works - move linked works to a list 968bf4ede01STejun Heo * @work: start of series of works to be scheduled 969bf4ede01STejun Heo * @head: target list to append @work to 970bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 971bf4ede01STejun Heo * 972bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 973bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 974bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 975bf4ede01STejun Heo * 976bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 977bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 978bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 979bf4ede01STejun Heo * 980bf4ede01STejun Heo * CONTEXT: 981d565ed63STejun Heo * spin_lock_irq(pool->lock). 982bf4ede01STejun Heo */ 983bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 984bf4ede01STejun Heo struct work_struct **nextp) 985bf4ede01STejun Heo { 986bf4ede01STejun Heo struct work_struct *n; 987bf4ede01STejun Heo 988bf4ede01STejun Heo /* 989bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 990bf4ede01STejun Heo * use NULL for list head. 991bf4ede01STejun Heo */ 992bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 993bf4ede01STejun Heo list_move_tail(&work->entry, head); 994bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 995bf4ede01STejun Heo break; 996bf4ede01STejun Heo } 997bf4ede01STejun Heo 998bf4ede01STejun Heo /* 999bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1000bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1001bf4ede01STejun Heo * needs to be updated. 1002bf4ede01STejun Heo */ 1003bf4ede01STejun Heo if (nextp) 1004bf4ede01STejun Heo *nextp = n; 1005bf4ede01STejun Heo } 1006bf4ede01STejun Heo 10078864b4e5STejun Heo /** 10088864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10098864b4e5STejun Heo * @pwq: pool_workqueue to get 10108864b4e5STejun Heo * 10118864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10128864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10138864b4e5STejun Heo */ 10148864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10158864b4e5STejun Heo { 10168864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10178864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10188864b4e5STejun Heo pwq->refcnt++; 10198864b4e5STejun Heo } 10208864b4e5STejun Heo 10218864b4e5STejun Heo /** 10228864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10238864b4e5STejun Heo * @pwq: pool_workqueue to put 10248864b4e5STejun Heo * 10258864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10268864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10278864b4e5STejun Heo */ 10288864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10298864b4e5STejun Heo { 10308864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10318864b4e5STejun Heo if (likely(--pwq->refcnt)) 10328864b4e5STejun Heo return; 10338864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10348864b4e5STejun Heo return; 10358864b4e5STejun Heo /* 10368864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10378864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10388864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10398864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10408864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10418864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10428864b4e5STejun Heo */ 10438864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10448864b4e5STejun Heo } 10458864b4e5STejun Heo 1046dce90d47STejun Heo /** 1047dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1048dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1049dce90d47STejun Heo * 1050dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1051dce90d47STejun Heo */ 1052dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1053dce90d47STejun Heo { 1054dce90d47STejun Heo if (pwq) { 1055dce90d47STejun Heo /* 1056dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1057dce90d47STejun Heo * following lock operations are safe. 1058dce90d47STejun Heo */ 1059dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1060dce90d47STejun Heo put_pwq(pwq); 1061dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1062dce90d47STejun Heo } 1063dce90d47STejun Heo } 1064dce90d47STejun Heo 1065112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1066bf4ede01STejun Heo { 1067112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1068bf4ede01STejun Heo 1069bf4ede01STejun Heo trace_workqueue_activate_work(work); 1070112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1071bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1072112202d9STejun Heo pwq->nr_active++; 1073bf4ede01STejun Heo } 1074bf4ede01STejun Heo 1075112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10763aa62497SLai Jiangshan { 1077112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10783aa62497SLai Jiangshan struct work_struct, entry); 10793aa62497SLai Jiangshan 1080112202d9STejun Heo pwq_activate_delayed_work(work); 10813aa62497SLai Jiangshan } 10823aa62497SLai Jiangshan 1083bf4ede01STejun Heo /** 1084112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1085112202d9STejun Heo * @pwq: pwq of interest 1086bf4ede01STejun Heo * @color: color of work which left the queue 1087bf4ede01STejun Heo * 1088bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1089112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1090bf4ede01STejun Heo * 1091bf4ede01STejun Heo * CONTEXT: 1092d565ed63STejun Heo * spin_lock_irq(pool->lock). 1093bf4ede01STejun Heo */ 1094112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1095bf4ede01STejun Heo { 10968864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1097bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10988864b4e5STejun Heo goto out_put; 1099bf4ede01STejun Heo 1100112202d9STejun Heo pwq->nr_in_flight[color]--; 1101bf4ede01STejun Heo 1102112202d9STejun Heo pwq->nr_active--; 1103112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1104bf4ede01STejun Heo /* one down, submit a delayed one */ 1105112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1106112202d9STejun Heo pwq_activate_first_delayed(pwq); 1107bf4ede01STejun Heo } 1108bf4ede01STejun Heo 1109bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1110112202d9STejun Heo if (likely(pwq->flush_color != color)) 11118864b4e5STejun Heo goto out_put; 1112bf4ede01STejun Heo 1113bf4ede01STejun Heo /* are there still in-flight works? */ 1114112202d9STejun Heo if (pwq->nr_in_flight[color]) 11158864b4e5STejun Heo goto out_put; 1116bf4ede01STejun Heo 1117112202d9STejun Heo /* this pwq is done, clear flush_color */ 1118112202d9STejun Heo pwq->flush_color = -1; 1119bf4ede01STejun Heo 1120bf4ede01STejun Heo /* 1121112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1122bf4ede01STejun Heo * will handle the rest. 1123bf4ede01STejun Heo */ 1124112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1125112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11268864b4e5STejun Heo out_put: 11278864b4e5STejun Heo put_pwq(pwq); 1128bf4ede01STejun Heo } 1129bf4ede01STejun Heo 113036e227d2STejun Heo /** 1131bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 113236e227d2STejun Heo * @work: work item to steal 113336e227d2STejun Heo * @is_dwork: @work is a delayed_work 1134bbb68dfaSTejun Heo * @flags: place to store irq state 113536e227d2STejun Heo * 113636e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1137d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 113836e227d2STejun Heo * 1139d185af30SYacine Belkadi * Return: 114036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 114136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 114236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1143bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1144bbb68dfaSTejun Heo * for arbitrarily long 114536e227d2STejun Heo * 1146d185af30SYacine Belkadi * Note: 1147bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1148e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1149e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1150e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1151bbb68dfaSTejun Heo * 1152bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1153bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1154bbb68dfaSTejun Heo * 1155e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1156bf4ede01STejun Heo */ 1157bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1158bbb68dfaSTejun Heo unsigned long *flags) 1159bf4ede01STejun Heo { 1160d565ed63STejun Heo struct worker_pool *pool; 1161112202d9STejun Heo struct pool_workqueue *pwq; 1162bf4ede01STejun Heo 1163bbb68dfaSTejun Heo local_irq_save(*flags); 1164bbb68dfaSTejun Heo 116536e227d2STejun Heo /* try to steal the timer if it exists */ 116636e227d2STejun Heo if (is_dwork) { 116736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 116836e227d2STejun Heo 1169e0aecdd8STejun Heo /* 1170e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1171e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1172e0aecdd8STejun Heo * running on the local CPU. 1173e0aecdd8STejun Heo */ 117436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 117536e227d2STejun Heo return 1; 117636e227d2STejun Heo } 117736e227d2STejun Heo 117836e227d2STejun Heo /* try to claim PENDING the normal way */ 1179bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1180bf4ede01STejun Heo return 0; 1181bf4ede01STejun Heo 1182bf4ede01STejun Heo /* 1183bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1184bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1185bf4ede01STejun Heo */ 1186d565ed63STejun Heo pool = get_work_pool(work); 1187d565ed63STejun Heo if (!pool) 1188bbb68dfaSTejun Heo goto fail; 1189bf4ede01STejun Heo 1190d565ed63STejun Heo spin_lock(&pool->lock); 1191bf4ede01STejun Heo /* 1192112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1193112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1194112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1195112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1196112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11970b3dae68SLai Jiangshan * item is currently queued on that pool. 1198bf4ede01STejun Heo */ 1199112202d9STejun Heo pwq = get_work_pwq(work); 1200112202d9STejun Heo if (pwq && pwq->pool == pool) { 1201bf4ede01STejun Heo debug_work_deactivate(work); 12023aa62497SLai Jiangshan 12033aa62497SLai Jiangshan /* 120416062836STejun Heo * A delayed work item cannot be grabbed directly because 120516062836STejun Heo * it might have linked NO_COLOR work items which, if left 1206112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 120716062836STejun Heo * management later on and cause stall. Make sure the work 120816062836STejun Heo * item is activated before grabbing. 12093aa62497SLai Jiangshan */ 12103aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1211112202d9STejun Heo pwq_activate_delayed_work(work); 12123aa62497SLai Jiangshan 1213bf4ede01STejun Heo list_del_init(&work->entry); 12149c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 121536e227d2STejun Heo 1216112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12174468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12184468a00fSLai Jiangshan 1219d565ed63STejun Heo spin_unlock(&pool->lock); 122036e227d2STejun Heo return 1; 1221bf4ede01STejun Heo } 1222d565ed63STejun Heo spin_unlock(&pool->lock); 1223bbb68dfaSTejun Heo fail: 1224bbb68dfaSTejun Heo local_irq_restore(*flags); 1225bbb68dfaSTejun Heo if (work_is_canceling(work)) 1226bbb68dfaSTejun Heo return -ENOENT; 1227bbb68dfaSTejun Heo cpu_relax(); 122836e227d2STejun Heo return -EAGAIN; 1229bf4ede01STejun Heo } 1230bf4ede01STejun Heo 1231bf4ede01STejun Heo /** 1232706026c2STejun Heo * insert_work - insert a work into a pool 1233112202d9STejun Heo * @pwq: pwq @work belongs to 12344690c4abSTejun Heo * @work: work to insert 12354690c4abSTejun Heo * @head: insertion point 12364690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12374690c4abSTejun Heo * 1238112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1239706026c2STejun Heo * work_struct flags. 12404690c4abSTejun Heo * 12414690c4abSTejun Heo * CONTEXT: 1242d565ed63STejun Heo * spin_lock_irq(pool->lock). 1243365970a1SDavid Howells */ 1244112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1245112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1246b89deed3SOleg Nesterov { 1247112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1248e1d8aa9fSFrederic Weisbecker 12494690c4abSTejun Heo /* we own @work, set data and link */ 1250112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12511a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12528864b4e5STejun Heo get_pwq(pwq); 1253e22bee78STejun Heo 1254e22bee78STejun Heo /* 1255c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1256c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1257c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1258e22bee78STejun Heo */ 1259e22bee78STejun Heo smp_mb(); 1260e22bee78STejun Heo 126163d95a91STejun Heo if (__need_more_worker(pool)) 126263d95a91STejun Heo wake_up_worker(pool); 1263b89deed3SOleg Nesterov } 1264b89deed3SOleg Nesterov 1265c8efcc25STejun Heo /* 1266c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12678d03ecfeSTejun Heo * same workqueue. 1268c8efcc25STejun Heo */ 1269c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1270c8efcc25STejun Heo { 1271c8efcc25STejun Heo struct worker *worker; 1272c8efcc25STejun Heo 12738d03ecfeSTejun Heo worker = current_wq_worker(); 1274c8efcc25STejun Heo /* 12758d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12768d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1277c8efcc25STejun Heo */ 1278112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1279c8efcc25STejun Heo } 1280c8efcc25STejun Heo 1281d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12821da177e4SLinus Torvalds struct work_struct *work) 12831da177e4SLinus Torvalds { 1284112202d9STejun Heo struct pool_workqueue *pwq; 1285c9178087STejun Heo struct worker_pool *last_pool; 12861e19ffc6STejun Heo struct list_head *worklist; 12878a2e8e5dSTejun Heo unsigned int work_flags; 1288b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12898930cabaSTejun Heo 12908930cabaSTejun Heo /* 12918930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12928930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12938930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12948930cabaSTejun Heo * happen with IRQ disabled. 12958930cabaSTejun Heo */ 12968930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12971da177e4SLinus Torvalds 1298dc186ad7SThomas Gleixner debug_work_activate(work); 12991e19ffc6STejun Heo 13009ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1301618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1302c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1303e41e704bSTejun Heo return; 13049e8cd2f5STejun Heo retry: 1305df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1306f3421797STejun Heo cpu = raw_smp_processor_id(); 1307df2d5ae4STejun Heo 1308df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1309df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1310c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1311df2d5ae4STejun Heo else 1312df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1313f3421797STejun Heo 131418aa9effSTejun Heo /* 1315c9178087STejun Heo * If @work was previously on a different pool, it might still be 1316c9178087STejun Heo * running there, in which case the work needs to be queued on that 1317c9178087STejun Heo * pool to guarantee non-reentrancy. 131818aa9effSTejun Heo */ 1319c9e7cf27STejun Heo last_pool = get_work_pool(work); 1320112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 132118aa9effSTejun Heo struct worker *worker; 132218aa9effSTejun Heo 1323d565ed63STejun Heo spin_lock(&last_pool->lock); 132418aa9effSTejun Heo 1325c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 132618aa9effSTejun Heo 1327112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1328c9178087STejun Heo pwq = worker->current_pwq; 13298594fadeSLai Jiangshan } else { 133018aa9effSTejun Heo /* meh... not running there, queue here */ 1331d565ed63STejun Heo spin_unlock(&last_pool->lock); 1332112202d9STejun Heo spin_lock(&pwq->pool->lock); 133318aa9effSTejun Heo } 13348930cabaSTejun Heo } else { 1335112202d9STejun Heo spin_lock(&pwq->pool->lock); 13368930cabaSTejun Heo } 1337502ca9d8STejun Heo 13389e8cd2f5STejun Heo /* 13399e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13409e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13419e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1342df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1343df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13449e8cd2f5STejun Heo * make forward-progress. 13459e8cd2f5STejun Heo */ 13469e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13479e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13489e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13499e8cd2f5STejun Heo cpu_relax(); 13509e8cd2f5STejun Heo goto retry; 13519e8cd2f5STejun Heo } 13529e8cd2f5STejun Heo /* oops */ 13539e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13549e8cd2f5STejun Heo wq->name, cpu); 13559e8cd2f5STejun Heo } 13569e8cd2f5STejun Heo 1357112202d9STejun Heo /* pwq determined, queue */ 1358112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1359502ca9d8STejun Heo 1360f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1361112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1362f5b2552bSDan Carpenter return; 1363f5b2552bSDan Carpenter } 13641e19ffc6STejun Heo 1365112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1366112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13671e19ffc6STejun Heo 1368112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1369cdadf009STejun Heo trace_workqueue_activate_work(work); 1370112202d9STejun Heo pwq->nr_active++; 1371112202d9STejun Heo worklist = &pwq->pool->worklist; 13728a2e8e5dSTejun Heo } else { 13738a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1374112202d9STejun Heo worklist = &pwq->delayed_works; 13758a2e8e5dSTejun Heo } 13761e19ffc6STejun Heo 1377112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13781e19ffc6STejun Heo 1379112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13801da177e4SLinus Torvalds } 13811da177e4SLinus Torvalds 13820fcb78c2SRolf Eike Beer /** 1383c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1384c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1385c1a220e7SZhang Rui * @wq: workqueue to use 1386c1a220e7SZhang Rui * @work: work to queue 1387c1a220e7SZhang Rui * 1388c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1389c1a220e7SZhang Rui * can't go away. 1390d185af30SYacine Belkadi * 1391d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1392c1a220e7SZhang Rui */ 1393d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1394d4283e93STejun Heo struct work_struct *work) 1395c1a220e7SZhang Rui { 1396d4283e93STejun Heo bool ret = false; 13978930cabaSTejun Heo unsigned long flags; 13988930cabaSTejun Heo 13998930cabaSTejun Heo local_irq_save(flags); 1400c1a220e7SZhang Rui 140122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14024690c4abSTejun Heo __queue_work(cpu, wq, work); 1403d4283e93STejun Heo ret = true; 1404c1a220e7SZhang Rui } 14058930cabaSTejun Heo 14068930cabaSTejun Heo local_irq_restore(flags); 1407c1a220e7SZhang Rui return ret; 1408c1a220e7SZhang Rui } 1409ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1410c1a220e7SZhang Rui 1411d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14121da177e4SLinus Torvalds { 141352bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14141da177e4SLinus Torvalds 1415e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 141660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14171da177e4SLinus Torvalds } 14181438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14191da177e4SLinus Torvalds 14207beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 142152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14221da177e4SLinus Torvalds { 14237beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14247beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14251da177e4SLinus Torvalds 14267beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14277beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1428fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1429fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14307beb2edfSTejun Heo 14318852aac2STejun Heo /* 14328852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14338852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14348852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14358852aac2STejun Heo * on that there's no such delay when @delay is 0. 14368852aac2STejun Heo */ 14378852aac2STejun Heo if (!delay) { 14388852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14398852aac2STejun Heo return; 14408852aac2STejun Heo } 14418852aac2STejun Heo 14427beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14437beb2edfSTejun Heo 144460c057bcSLai Jiangshan dwork->wq = wq; 14451265057fSTejun Heo dwork->cpu = cpu; 14467beb2edfSTejun Heo timer->expires = jiffies + delay; 14477beb2edfSTejun Heo 14487beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14497beb2edfSTejun Heo add_timer_on(timer, cpu); 14507beb2edfSTejun Heo else 14517beb2edfSTejun Heo add_timer(timer); 14527beb2edfSTejun Heo } 14531da177e4SLinus Torvalds 14540fcb78c2SRolf Eike Beer /** 14550fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14560fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14570fcb78c2SRolf Eike Beer * @wq: workqueue to use 1458af9997e4SRandy Dunlap * @dwork: work to queue 14590fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14600fcb78c2SRolf Eike Beer * 1461d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1462715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1463715f1300STejun Heo * execution. 14640fcb78c2SRolf Eike Beer */ 1465d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 146652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14677a6bc1cdSVenkatesh Pallipadi { 146852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1469d4283e93STejun Heo bool ret = false; 14708930cabaSTejun Heo unsigned long flags; 14718930cabaSTejun Heo 14728930cabaSTejun Heo /* read the comment in __queue_work() */ 14738930cabaSTejun Heo local_irq_save(flags); 14747a6bc1cdSVenkatesh Pallipadi 147522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14767beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1477d4283e93STejun Heo ret = true; 14787a6bc1cdSVenkatesh Pallipadi } 14798930cabaSTejun Heo 14808930cabaSTejun Heo local_irq_restore(flags); 14817a6bc1cdSVenkatesh Pallipadi return ret; 14827a6bc1cdSVenkatesh Pallipadi } 1483ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 14841da177e4SLinus Torvalds 1485c8e55f36STejun Heo /** 14868376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14878376fe22STejun Heo * @cpu: CPU number to execute work on 14888376fe22STejun Heo * @wq: workqueue to use 14898376fe22STejun Heo * @dwork: work to queue 14908376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14918376fe22STejun Heo * 14928376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14938376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14948376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14958376fe22STejun Heo * current state. 14968376fe22STejun Heo * 1497d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 14988376fe22STejun Heo * pending and its timer was modified. 14998376fe22STejun Heo * 1500e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15018376fe22STejun Heo * See try_to_grab_pending() for details. 15028376fe22STejun Heo */ 15038376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15048376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15058376fe22STejun Heo { 15068376fe22STejun Heo unsigned long flags; 15078376fe22STejun Heo int ret; 15088376fe22STejun Heo 15098376fe22STejun Heo do { 15108376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15118376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15128376fe22STejun Heo 15138376fe22STejun Heo if (likely(ret >= 0)) { 15148376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15158376fe22STejun Heo local_irq_restore(flags); 15168376fe22STejun Heo } 15178376fe22STejun Heo 15188376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15198376fe22STejun Heo return ret; 15208376fe22STejun Heo } 15218376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15228376fe22STejun Heo 15238376fe22STejun Heo /** 1524c8e55f36STejun Heo * worker_enter_idle - enter idle state 1525c8e55f36STejun Heo * @worker: worker which is entering idle state 1526c8e55f36STejun Heo * 1527c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1528c8e55f36STejun Heo * necessary. 1529c8e55f36STejun Heo * 1530c8e55f36STejun Heo * LOCKING: 1531d565ed63STejun Heo * spin_lock_irq(pool->lock). 1532c8e55f36STejun Heo */ 1533c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15341da177e4SLinus Torvalds { 1535bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1536c8e55f36STejun Heo 15376183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15386183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15396183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15406183c009STejun Heo return; 1541c8e55f36STejun Heo 1542051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1543cb444766STejun Heo worker->flags |= WORKER_IDLE; 1544bd7bdd43STejun Heo pool->nr_idle++; 1545e22bee78STejun Heo worker->last_active = jiffies; 1546c8e55f36STejun Heo 1547c8e55f36STejun Heo /* idle_list is LIFO */ 1548bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1549db7bccf4STejun Heo 155063d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1551628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1552cb444766STejun Heo 1553544ecf31STejun Heo /* 1554706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1555d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1556628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1557628c78e7STejun Heo * unbind is not in progress. 1558544ecf31STejun Heo */ 155924647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1560bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1561e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1562c8e55f36STejun Heo } 1563c8e55f36STejun Heo 1564c8e55f36STejun Heo /** 1565c8e55f36STejun Heo * worker_leave_idle - leave idle state 1566c8e55f36STejun Heo * @worker: worker which is leaving idle state 1567c8e55f36STejun Heo * 1568c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1569c8e55f36STejun Heo * 1570c8e55f36STejun Heo * LOCKING: 1571d565ed63STejun Heo * spin_lock_irq(pool->lock). 1572c8e55f36STejun Heo */ 1573c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1574c8e55f36STejun Heo { 1575bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1576c8e55f36STejun Heo 15776183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15786183c009STejun Heo return; 1579d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1580bd7bdd43STejun Heo pool->nr_idle--; 1581c8e55f36STejun Heo list_del_init(&worker->entry); 1582c8e55f36STejun Heo } 1583c8e55f36STejun Heo 1584f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1585c34056a3STejun Heo { 1586c34056a3STejun Heo struct worker *worker; 1587c34056a3STejun Heo 1588f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1589c8e55f36STejun Heo if (worker) { 1590c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1591affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1592da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1593e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1594e22bee78STejun Heo worker->flags = WORKER_PREP; 1595c8e55f36STejun Heo } 1596c34056a3STejun Heo return worker; 1597c34056a3STejun Heo } 1598c34056a3STejun Heo 1599c34056a3STejun Heo /** 16004736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16014736cbf7SLai Jiangshan * @worker: worker to be attached 16024736cbf7SLai Jiangshan * @pool: the target pool 16034736cbf7SLai Jiangshan * 16044736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16054736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16064736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16074736cbf7SLai Jiangshan */ 16084736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16094736cbf7SLai Jiangshan struct worker_pool *pool) 16104736cbf7SLai Jiangshan { 16114736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16124736cbf7SLai Jiangshan 16134736cbf7SLai Jiangshan /* 16144736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16154736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16164736cbf7SLai Jiangshan */ 16174736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16184736cbf7SLai Jiangshan 16194736cbf7SLai Jiangshan /* 16204736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 16214736cbf7SLai Jiangshan * stable across this function. See the comments above the 16224736cbf7SLai Jiangshan * flag definition for details. 16234736cbf7SLai Jiangshan */ 16244736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 16254736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 16264736cbf7SLai Jiangshan 16274736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 16284736cbf7SLai Jiangshan 16294736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 16304736cbf7SLai Jiangshan } 16314736cbf7SLai Jiangshan 16324736cbf7SLai Jiangshan /** 163360f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 163460f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 163560f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 163660f5a4bcSLai Jiangshan * 16374736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 16384736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 16394736cbf7SLai Jiangshan * other reference to the pool. 164060f5a4bcSLai Jiangshan */ 164160f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 164260f5a4bcSLai Jiangshan struct worker_pool *pool) 164360f5a4bcSLai Jiangshan { 164460f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 164560f5a4bcSLai Jiangshan 164692f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1647da028469SLai Jiangshan list_del(&worker->node); 1648da028469SLai Jiangshan if (list_empty(&pool->workers)) 164960f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 165092f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 165160f5a4bcSLai Jiangshan 1652b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1653b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1654b62c0751SLai Jiangshan 165560f5a4bcSLai Jiangshan if (detach_completion) 165660f5a4bcSLai Jiangshan complete(detach_completion); 165760f5a4bcSLai Jiangshan } 165860f5a4bcSLai Jiangshan 165960f5a4bcSLai Jiangshan /** 1660c34056a3STejun Heo * create_worker - create a new workqueue worker 166163d95a91STejun Heo * @pool: pool the new worker will belong to 1662c34056a3STejun Heo * 1663051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1664c34056a3STejun Heo * 1665c34056a3STejun Heo * CONTEXT: 1666c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1667c34056a3STejun Heo * 1668d185af30SYacine Belkadi * Return: 1669c34056a3STejun Heo * Pointer to the newly created worker. 1670c34056a3STejun Heo */ 1671bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1672c34056a3STejun Heo { 1673c34056a3STejun Heo struct worker *worker = NULL; 1674f3421797STejun Heo int id = -1; 1675e3c916a4STejun Heo char id_buf[16]; 1676c34056a3STejun Heo 16777cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 16787cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1679822d8405STejun Heo if (id < 0) 1680c34056a3STejun Heo goto fail; 1681c34056a3STejun Heo 1682f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1683c34056a3STejun Heo if (!worker) 1684c34056a3STejun Heo goto fail; 1685c34056a3STejun Heo 1686bd7bdd43STejun Heo worker->pool = pool; 1687c34056a3STejun Heo worker->id = id; 1688c34056a3STejun Heo 168929c91e99STejun Heo if (pool->cpu >= 0) 1690e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1691e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1692f3421797STejun Heo else 1693e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1694e3c916a4STejun Heo 1695f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1696e3c916a4STejun Heo "kworker/%s", id_buf); 1697c34056a3STejun Heo if (IS_ERR(worker->task)) 1698c34056a3STejun Heo goto fail; 1699c34056a3STejun Heo 170091151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 170191151228SOleg Nesterov 170291151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 170391151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 170491151228SOleg Nesterov 1705da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17064736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1707822d8405STejun Heo 1708051e1850SLai Jiangshan /* start the newly created worker */ 1709051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1710051e1850SLai Jiangshan worker->pool->nr_workers++; 1711051e1850SLai Jiangshan worker_enter_idle(worker); 1712051e1850SLai Jiangshan wake_up_process(worker->task); 1713051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1714051e1850SLai Jiangshan 1715c34056a3STejun Heo return worker; 1716822d8405STejun Heo 1717c34056a3STejun Heo fail: 17189625ab17SLai Jiangshan if (id >= 0) 17197cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1720c34056a3STejun Heo kfree(worker); 1721c34056a3STejun Heo return NULL; 1722c34056a3STejun Heo } 1723c34056a3STejun Heo 1724c34056a3STejun Heo /** 1725c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1726c34056a3STejun Heo * @worker: worker to be destroyed 1727c34056a3STejun Heo * 172873eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 172973eb7fe7SLai Jiangshan * be idle. 1730c8e55f36STejun Heo * 1731c8e55f36STejun Heo * CONTEXT: 173260f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1733c34056a3STejun Heo */ 1734c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1735c34056a3STejun Heo { 1736bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1737c34056a3STejun Heo 1738cd549687STejun Heo lockdep_assert_held(&pool->lock); 1739cd549687STejun Heo 1740c34056a3STejun Heo /* sanity check frenzy */ 17416183c009STejun Heo if (WARN_ON(worker->current_work) || 174273eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 174373eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 17446183c009STejun Heo return; 1745c34056a3STejun Heo 1746bd7bdd43STejun Heo pool->nr_workers--; 1747bd7bdd43STejun Heo pool->nr_idle--; 1748c8e55f36STejun Heo 1749c8e55f36STejun Heo list_del_init(&worker->entry); 1750cb444766STejun Heo worker->flags |= WORKER_DIE; 175160f5a4bcSLai Jiangshan wake_up_process(worker->task); 1752c34056a3STejun Heo } 1753c34056a3STejun Heo 175463d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1755e22bee78STejun Heo { 175663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1757e22bee78STejun Heo 1758d565ed63STejun Heo spin_lock_irq(&pool->lock); 1759e22bee78STejun Heo 17603347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1761e22bee78STejun Heo struct worker *worker; 1762e22bee78STejun Heo unsigned long expires; 1763e22bee78STejun Heo 1764e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 176563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1766e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1767e22bee78STejun Heo 17683347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 176963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 17703347fc9fSLai Jiangshan break; 1771e22bee78STejun Heo } 17723347fc9fSLai Jiangshan 17733347fc9fSLai Jiangshan destroy_worker(worker); 1774e22bee78STejun Heo } 1775e22bee78STejun Heo 1776d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1777e22bee78STejun Heo } 1778e22bee78STejun Heo 1779493a1724STejun Heo static void send_mayday(struct work_struct *work) 1780e22bee78STejun Heo { 1781112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1782112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1783493a1724STejun Heo 17842e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1785e22bee78STejun Heo 1786493008a8STejun Heo if (!wq->rescuer) 1787493a1724STejun Heo return; 1788e22bee78STejun Heo 1789e22bee78STejun Heo /* mayday mayday mayday */ 1790493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 179177668c8bSLai Jiangshan /* 179277668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 179377668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 179477668c8bSLai Jiangshan * rescuer is done with it. 179577668c8bSLai Jiangshan */ 179677668c8bSLai Jiangshan get_pwq(pwq); 1797493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1798e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1799493a1724STejun Heo } 1800e22bee78STejun Heo } 1801e22bee78STejun Heo 1802706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1803e22bee78STejun Heo { 180463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1805e22bee78STejun Heo struct work_struct *work; 1806e22bee78STejun Heo 1807b2d82909STejun Heo spin_lock_irq(&pool->lock); 1808b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1809e22bee78STejun Heo 181063d95a91STejun Heo if (need_to_create_worker(pool)) { 1811e22bee78STejun Heo /* 1812e22bee78STejun Heo * We've been trying to create a new worker but 1813e22bee78STejun Heo * haven't been successful. We might be hitting an 1814e22bee78STejun Heo * allocation deadlock. Send distress signals to 1815e22bee78STejun Heo * rescuers. 1816e22bee78STejun Heo */ 181763d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1818e22bee78STejun Heo send_mayday(work); 1819e22bee78STejun Heo } 1820e22bee78STejun Heo 1821b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1822b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1823e22bee78STejun Heo 182463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1825e22bee78STejun Heo } 1826e22bee78STejun Heo 1827e22bee78STejun Heo /** 1828e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 182963d95a91STejun Heo * @pool: pool to create a new worker for 1830e22bee78STejun Heo * 183163d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1832e22bee78STejun Heo * have at least one idle worker on return from this function. If 1833e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 183463d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1835e22bee78STejun Heo * possible allocation deadlock. 1836e22bee78STejun Heo * 1837c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1838c5aa87bbSTejun Heo * may_start_working() %true. 1839e22bee78STejun Heo * 1840e22bee78STejun Heo * LOCKING: 1841d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1842e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1843e22bee78STejun Heo * manager. 1844e22bee78STejun Heo */ 184529187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1846d565ed63STejun Heo __releases(&pool->lock) 1847d565ed63STejun Heo __acquires(&pool->lock) 1848e22bee78STejun Heo { 1849e22bee78STejun Heo restart: 1850d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18519f9c2364STejun Heo 1852e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 185363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1854e22bee78STejun Heo 1855e22bee78STejun Heo while (true) { 1856051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1857e22bee78STejun Heo break; 1858e22bee78STejun Heo 1859e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 18609f9c2364STejun Heo 186163d95a91STejun Heo if (!need_to_create_worker(pool)) 1862e22bee78STejun Heo break; 1863e22bee78STejun Heo } 1864e22bee78STejun Heo 186563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1866d565ed63STejun Heo spin_lock_irq(&pool->lock); 1867051e1850SLai Jiangshan /* 1868051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1869051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1870051e1850SLai Jiangshan * already become busy. 1871051e1850SLai Jiangshan */ 187263d95a91STejun Heo if (need_to_create_worker(pool)) 1873e22bee78STejun Heo goto restart; 1874e22bee78STejun Heo } 1875e22bee78STejun Heo 1876e22bee78STejun Heo /** 1877e22bee78STejun Heo * manage_workers - manage worker pool 1878e22bee78STejun Heo * @worker: self 1879e22bee78STejun Heo * 1880706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1881e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1882706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1883e22bee78STejun Heo * 1884e22bee78STejun Heo * The caller can safely start processing works on false return. On 1885e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1886e22bee78STejun Heo * and may_start_working() is true. 1887e22bee78STejun Heo * 1888e22bee78STejun Heo * CONTEXT: 1889d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1890e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1891e22bee78STejun Heo * 1892d185af30SYacine Belkadi * Return: 189329187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 189429187a9eSTejun Heo * start processing works, %true if management function was performed and 189529187a9eSTejun Heo * the conditions that the caller verified before calling the function may 189629187a9eSTejun Heo * no longer be true. 1897e22bee78STejun Heo */ 1898e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1899e22bee78STejun Heo { 190063d95a91STejun Heo struct worker_pool *pool = worker->pool; 1901e22bee78STejun Heo 1902bc3a1afcSTejun Heo /* 1903bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1904bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1905bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1906bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1907bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1908bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1909bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1910bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1911bc3a1afcSTejun Heo */ 191234a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 191329187a9eSTejun Heo return false; 1914e22bee78STejun Heo 191529187a9eSTejun Heo maybe_create_worker(pool); 1916e22bee78STejun Heo 191734a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 191829187a9eSTejun Heo return true; 1919e22bee78STejun Heo } 1920e22bee78STejun Heo 1921a62428c0STejun Heo /** 1922a62428c0STejun Heo * process_one_work - process single work 1923c34056a3STejun Heo * @worker: self 1924a62428c0STejun Heo * @work: work to process 1925a62428c0STejun Heo * 1926a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1927a62428c0STejun Heo * process a single work including synchronization against and 1928a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1929a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1930a62428c0STejun Heo * call this function to process a work. 1931a62428c0STejun Heo * 1932a62428c0STejun Heo * CONTEXT: 1933d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1934a62428c0STejun Heo */ 1935c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 1936d565ed63STejun Heo __releases(&pool->lock) 1937d565ed63STejun Heo __acquires(&pool->lock) 19381da177e4SLinus Torvalds { 1939112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1940bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1941112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 194273f53c4aSTejun Heo int work_color; 19437e11629dSTejun Heo struct worker *collision; 19444e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 19454e6045f1SJohannes Berg /* 1946a62428c0STejun Heo * It is permissible to free the struct work_struct from 1947a62428c0STejun Heo * inside the function that is called from it, this we need to 1948a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1949a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1950a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 19514e6045f1SJohannes Berg */ 19524d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 19534d82a1deSPeter Zijlstra 19544d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 19554e6045f1SJohannes Berg #endif 1956807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 195785327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1958ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 195925511a47STejun Heo 19607e11629dSTejun Heo /* 19617e11629dSTejun Heo * A single work shouldn't be executed concurrently by 19627e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 19637e11629dSTejun Heo * already processing the work. If so, defer the work to the 19647e11629dSTejun Heo * currently executing one. 19657e11629dSTejun Heo */ 1966c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 19677e11629dSTejun Heo if (unlikely(collision)) { 19687e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 19697e11629dSTejun Heo return; 19707e11629dSTejun Heo } 19711da177e4SLinus Torvalds 19728930cabaSTejun Heo /* claim and dequeue */ 19731da177e4SLinus Torvalds debug_work_deactivate(work); 1974c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 1975c34056a3STejun Heo worker->current_work = work; 1976a2c1c57bSTejun Heo worker->current_func = work->func; 1977112202d9STejun Heo worker->current_pwq = pwq; 197873f53c4aSTejun Heo work_color = get_work_color(work); 19797a22ad75STejun Heo 1980a62428c0STejun Heo list_del_init(&work->entry); 1981a62428c0STejun Heo 1982649027d7STejun Heo /* 1983228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 1984228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 1985228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 1986228f1d00SLai Jiangshan * execution of the pending work items. 1987fb0e7bebSTejun Heo */ 1988fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 1989228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 1990fb0e7bebSTejun Heo 1991974271c4STejun Heo /* 1992a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 1993a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 1994a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 1995a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 1996228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 1997974271c4STejun Heo */ 1998a489a03eSLai Jiangshan if (need_more_worker(pool)) 199963d95a91STejun Heo wake_up_worker(pool); 2000974271c4STejun Heo 20018930cabaSTejun Heo /* 20027c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2003d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 200423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 200523657bb1STejun Heo * disabled. 20068930cabaSTejun Heo */ 20077c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20081da177e4SLinus Torvalds 2009d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2010365970a1SDavid Howells 2011112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20123295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2013e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2014a2c1c57bSTejun Heo worker->current_func(work); 2015e36c886aSArjan van de Ven /* 2016e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2017e36c886aSArjan van de Ven * point will only record its address. 2018e36c886aSArjan van de Ven */ 2019e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20203295f0efSIngo Molnar lock_map_release(&lockdep_map); 2021112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 20221da177e4SLinus Torvalds 2023d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2024044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2025044c782cSValentin Ilie " last function: %pf\n", 2026a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2027a2c1c57bSTejun Heo worker->current_func); 2028d5abe669SPeter Zijlstra debug_show_held_locks(current); 2029d5abe669SPeter Zijlstra dump_stack(); 2030d5abe669SPeter Zijlstra } 2031d5abe669SPeter Zijlstra 2032b22ce278STejun Heo /* 2033b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2034b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2035b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2036b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2037789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2038789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2039b22ce278STejun Heo */ 20403e28e377SJoe Lawrence cond_resched_rcu_qs(); 2041b22ce278STejun Heo 2042d565ed63STejun Heo spin_lock_irq(&pool->lock); 2043a62428c0STejun Heo 2044fb0e7bebSTejun Heo /* clear cpu intensive status */ 2045fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2046fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2047fb0e7bebSTejun Heo 2048a62428c0STejun Heo /* we're done with it, release */ 204942f8570fSSasha Levin hash_del(&worker->hentry); 2050c34056a3STejun Heo worker->current_work = NULL; 2051a2c1c57bSTejun Heo worker->current_func = NULL; 2052112202d9STejun Heo worker->current_pwq = NULL; 20533d1cb205STejun Heo worker->desc_valid = false; 2054112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 20551da177e4SLinus Torvalds } 20561da177e4SLinus Torvalds 2057affee4b2STejun Heo /** 2058affee4b2STejun Heo * process_scheduled_works - process scheduled works 2059affee4b2STejun Heo * @worker: self 2060affee4b2STejun Heo * 2061affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2062affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2063affee4b2STejun Heo * fetches a work from the top and executes it. 2064affee4b2STejun Heo * 2065affee4b2STejun Heo * CONTEXT: 2066d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2067affee4b2STejun Heo * multiple times. 2068affee4b2STejun Heo */ 2069affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 20701da177e4SLinus Torvalds { 2071affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2072affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2073a62428c0STejun Heo struct work_struct, entry); 2074c34056a3STejun Heo process_one_work(worker, work); 2075a62428c0STejun Heo } 20761da177e4SLinus Torvalds } 20771da177e4SLinus Torvalds 20784690c4abSTejun Heo /** 20794690c4abSTejun Heo * worker_thread - the worker thread function 2080c34056a3STejun Heo * @__worker: self 20814690c4abSTejun Heo * 2082c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2083c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2084c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2085c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2086c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2087d185af30SYacine Belkadi * 2088d185af30SYacine Belkadi * Return: 0 20894690c4abSTejun Heo */ 2090c34056a3STejun Heo static int worker_thread(void *__worker) 20911da177e4SLinus Torvalds { 2092c34056a3STejun Heo struct worker *worker = __worker; 2093bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 20941da177e4SLinus Torvalds 2095e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2096e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2097c8e55f36STejun Heo woke_up: 2098d565ed63STejun Heo spin_lock_irq(&pool->lock); 2099affee4b2STejun Heo 2100a9ab775bSTejun Heo /* am I supposed to die? */ 2101a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2102d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2103a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2104e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 210560f5a4bcSLai Jiangshan 210660f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21077cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 210860f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 210960f5a4bcSLai Jiangshan kfree(worker); 2110c8e55f36STejun Heo return 0; 2111c8e55f36STejun Heo } 2112c8e55f36STejun Heo 2113c8e55f36STejun Heo worker_leave_idle(worker); 2114db7bccf4STejun Heo recheck: 2115e22bee78STejun Heo /* no more worker necessary? */ 211663d95a91STejun Heo if (!need_more_worker(pool)) 2117e22bee78STejun Heo goto sleep; 2118e22bee78STejun Heo 2119e22bee78STejun Heo /* do we need to manage? */ 212063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2121e22bee78STejun Heo goto recheck; 2122e22bee78STejun Heo 2123c8e55f36STejun Heo /* 2124c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2125c8e55f36STejun Heo * preparing to process a work or actually processing it. 2126c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2127c8e55f36STejun Heo */ 21286183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2129c8e55f36STejun Heo 2130e22bee78STejun Heo /* 2131a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2132a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2133a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2134a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2135a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2136e22bee78STejun Heo */ 2137a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2138e22bee78STejun Heo 2139e22bee78STejun Heo do { 2140affee4b2STejun Heo struct work_struct *work = 2141bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2142affee4b2STejun Heo struct work_struct, entry); 2143affee4b2STejun Heo 2144c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2145affee4b2STejun Heo /* optimization path, not strictly necessary */ 2146affee4b2STejun Heo process_one_work(worker, work); 2147affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2148affee4b2STejun Heo process_scheduled_works(worker); 2149affee4b2STejun Heo } else { 2150c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2151affee4b2STejun Heo process_scheduled_works(worker); 2152affee4b2STejun Heo } 215363d95a91STejun Heo } while (keep_working(pool)); 2154affee4b2STejun Heo 2155228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2156d313dd85STejun Heo sleep: 2157c8e55f36STejun Heo /* 2158d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2159d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2160d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2161d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2162d565ed63STejun Heo * event. 2163c8e55f36STejun Heo */ 2164c8e55f36STejun Heo worker_enter_idle(worker); 2165c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2166d565ed63STejun Heo spin_unlock_irq(&pool->lock); 21671da177e4SLinus Torvalds schedule(); 2168c8e55f36STejun Heo goto woke_up; 21691da177e4SLinus Torvalds } 21701da177e4SLinus Torvalds 2171e22bee78STejun Heo /** 2172e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2173111c225aSTejun Heo * @__rescuer: self 2174e22bee78STejun Heo * 2175e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2176493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2177e22bee78STejun Heo * 2178706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2179e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2180e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2181e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2182e22bee78STejun Heo * the problem rescuer solves. 2183e22bee78STejun Heo * 2184706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2185706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2186e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2187e22bee78STejun Heo * 2188e22bee78STejun Heo * This should happen rarely. 2189d185af30SYacine Belkadi * 2190d185af30SYacine Belkadi * Return: 0 2191e22bee78STejun Heo */ 2192111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2193e22bee78STejun Heo { 2194111c225aSTejun Heo struct worker *rescuer = __rescuer; 2195111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2196e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 21974d595b86SLai Jiangshan bool should_stop; 2198e22bee78STejun Heo 2199e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2200111c225aSTejun Heo 2201111c225aSTejun Heo /* 2202111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2203111c225aSTejun Heo * doesn't participate in concurrency management. 2204111c225aSTejun Heo */ 2205111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2206e22bee78STejun Heo repeat: 2207e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22081da177e4SLinus Torvalds 22094d595b86SLai Jiangshan /* 22104d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22114d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22124d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22134d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22144d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22154d595b86SLai Jiangshan * list is always empty on exit. 22164d595b86SLai Jiangshan */ 22174d595b86SLai Jiangshan should_stop = kthread_should_stop(); 22181da177e4SLinus Torvalds 2219493a1724STejun Heo /* see whether any pwq is asking for help */ 22202e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2221493a1724STejun Heo 2222493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2223493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2224493a1724STejun Heo struct pool_workqueue, mayday_node); 2225112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2226e22bee78STejun Heo struct work_struct *work, *n; 2227e22bee78STejun Heo 2228e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2229493a1724STejun Heo list_del_init(&pwq->mayday_node); 2230493a1724STejun Heo 22312e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2232e22bee78STejun Heo 223351697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 223451697d39SLai Jiangshan 223551697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2236b3104104SLai Jiangshan rescuer->pool = pool; 2237e22bee78STejun Heo 2238e22bee78STejun Heo /* 2239e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2240e22bee78STejun Heo * process'em. 2241e22bee78STejun Heo */ 22420479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 2243bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2244112202d9STejun Heo if (get_work_pwq(work) == pwq) 2245e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2246e22bee78STejun Heo 2247008847f6SNeilBrown if (!list_empty(scheduled)) { 2248e22bee78STejun Heo process_scheduled_works(rescuer); 22497576958aSTejun Heo 22507576958aSTejun Heo /* 2251008847f6SNeilBrown * The above execution of rescued work items could 2252008847f6SNeilBrown * have created more to rescue through 2253008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2254008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2255008847f6SNeilBrown * that such back-to-back work items, which may be 2256008847f6SNeilBrown * being used to relieve memory pressure, don't 2257008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2258008847f6SNeilBrown */ 2259008847f6SNeilBrown if (need_to_create_worker(pool)) { 2260008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2261008847f6SNeilBrown get_pwq(pwq); 2262008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2263008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2264008847f6SNeilBrown } 2265008847f6SNeilBrown } 2266008847f6SNeilBrown 2267008847f6SNeilBrown /* 226877668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 226913b1d625SLai Jiangshan * go away while we're still attached to it. 227077668c8bSLai Jiangshan */ 227177668c8bSLai Jiangshan put_pwq(pwq); 227277668c8bSLai Jiangshan 227377668c8bSLai Jiangshan /* 2274d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 22757576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 22767576958aSTejun Heo * and stalling the execution. 22777576958aSTejun Heo */ 2278d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 227963d95a91STejun Heo wake_up_worker(pool); 22807576958aSTejun Heo 2281b3104104SLai Jiangshan rescuer->pool = NULL; 228213b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 228313b1d625SLai Jiangshan 228413b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 228513b1d625SLai Jiangshan 228613b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 22871da177e4SLinus Torvalds } 22881da177e4SLinus Torvalds 22892e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2290493a1724STejun Heo 22914d595b86SLai Jiangshan if (should_stop) { 22924d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 22934d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 22944d595b86SLai Jiangshan return 0; 22954d595b86SLai Jiangshan } 22964d595b86SLai Jiangshan 2297111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2298111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2299e22bee78STejun Heo schedule(); 2300e22bee78STejun Heo goto repeat; 23011da177e4SLinus Torvalds } 23021da177e4SLinus Torvalds 2303fc2e4d70SOleg Nesterov struct wq_barrier { 2304fc2e4d70SOleg Nesterov struct work_struct work; 2305fc2e4d70SOleg Nesterov struct completion done; 2306fc2e4d70SOleg Nesterov }; 2307fc2e4d70SOleg Nesterov 2308fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2309fc2e4d70SOleg Nesterov { 2310fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2311fc2e4d70SOleg Nesterov complete(&barr->done); 2312fc2e4d70SOleg Nesterov } 2313fc2e4d70SOleg Nesterov 23144690c4abSTejun Heo /** 23154690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2316112202d9STejun Heo * @pwq: pwq to insert barrier into 23174690c4abSTejun Heo * @barr: wq_barrier to insert 2318affee4b2STejun Heo * @target: target work to attach @barr to 2319affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23204690c4abSTejun Heo * 2321affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2322affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2323affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2324affee4b2STejun Heo * cpu. 2325affee4b2STejun Heo * 2326affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2327affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2328affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2329affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2330affee4b2STejun Heo * after a work with LINKED flag set. 2331affee4b2STejun Heo * 2332affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2333112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23344690c4abSTejun Heo * 23354690c4abSTejun Heo * CONTEXT: 2336d565ed63STejun Heo * spin_lock_irq(pool->lock). 23374690c4abSTejun Heo */ 2338112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2339affee4b2STejun Heo struct wq_barrier *barr, 2340affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2341fc2e4d70SOleg Nesterov { 2342affee4b2STejun Heo struct list_head *head; 2343affee4b2STejun Heo unsigned int linked = 0; 2344affee4b2STejun Heo 2345dc186ad7SThomas Gleixner /* 2346d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2347dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2348dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2349dc186ad7SThomas Gleixner * might deadlock. 2350dc186ad7SThomas Gleixner */ 2351ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 235222df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2353fc2e4d70SOleg Nesterov init_completion(&barr->done); 235483c22520SOleg Nesterov 2355affee4b2STejun Heo /* 2356affee4b2STejun Heo * If @target is currently being executed, schedule the 2357affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2358affee4b2STejun Heo */ 2359affee4b2STejun Heo if (worker) 2360affee4b2STejun Heo head = worker->scheduled.next; 2361affee4b2STejun Heo else { 2362affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2363affee4b2STejun Heo 2364affee4b2STejun Heo head = target->entry.next; 2365affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2366affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2367affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2368affee4b2STejun Heo } 2369affee4b2STejun Heo 2370dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2371112202d9STejun Heo insert_work(pwq, &barr->work, head, 2372affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2373fc2e4d70SOleg Nesterov } 2374fc2e4d70SOleg Nesterov 237573f53c4aSTejun Heo /** 2376112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 237773f53c4aSTejun Heo * @wq: workqueue being flushed 237873f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 237973f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 238073f53c4aSTejun Heo * 2381112202d9STejun Heo * Prepare pwqs for workqueue flushing. 238273f53c4aSTejun Heo * 2383112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2384112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2385112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2386112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2387112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 238873f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 238973f53c4aSTejun Heo * 239073f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 239173f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 239273f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 239373f53c4aSTejun Heo * is returned. 239473f53c4aSTejun Heo * 2395112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 239673f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 239773f53c4aSTejun Heo * advanced to @work_color. 239873f53c4aSTejun Heo * 239973f53c4aSTejun Heo * CONTEXT: 24003c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 240173f53c4aSTejun Heo * 2402d185af30SYacine Belkadi * Return: 240373f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 240473f53c4aSTejun Heo * otherwise. 240573f53c4aSTejun Heo */ 2406112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 240773f53c4aSTejun Heo int flush_color, int work_color) 24081da177e4SLinus Torvalds { 240973f53c4aSTejun Heo bool wait = false; 241049e3cf44STejun Heo struct pool_workqueue *pwq; 24111da177e4SLinus Torvalds 241273f53c4aSTejun Heo if (flush_color >= 0) { 24136183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2414112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2415dc186ad7SThomas Gleixner } 241614441960SOleg Nesterov 241749e3cf44STejun Heo for_each_pwq(pwq, wq) { 2418112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24191da177e4SLinus Torvalds 2420b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 242173f53c4aSTejun Heo 242273f53c4aSTejun Heo if (flush_color >= 0) { 24236183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 242473f53c4aSTejun Heo 2425112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2426112202d9STejun Heo pwq->flush_color = flush_color; 2427112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 242873f53c4aSTejun Heo wait = true; 24291da177e4SLinus Torvalds } 243073f53c4aSTejun Heo } 243173f53c4aSTejun Heo 243273f53c4aSTejun Heo if (work_color >= 0) { 24336183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2434112202d9STejun Heo pwq->work_color = work_color; 243573f53c4aSTejun Heo } 243673f53c4aSTejun Heo 2437b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 24381da177e4SLinus Torvalds } 24391da177e4SLinus Torvalds 2440112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 244173f53c4aSTejun Heo complete(&wq->first_flusher->done); 244273f53c4aSTejun Heo 244373f53c4aSTejun Heo return wait; 244483c22520SOleg Nesterov } 24451da177e4SLinus Torvalds 24460fcb78c2SRolf Eike Beer /** 24471da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 24480fcb78c2SRolf Eike Beer * @wq: workqueue to flush 24491da177e4SLinus Torvalds * 2450c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2451c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 24521da177e4SLinus Torvalds */ 24537ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 24541da177e4SLinus Torvalds { 245573f53c4aSTejun Heo struct wq_flusher this_flusher = { 245673f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 245773f53c4aSTejun Heo .flush_color = -1, 245873f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 245973f53c4aSTejun Heo }; 246073f53c4aSTejun Heo int next_color; 2461b1f4ec17SOleg Nesterov 24623295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 24633295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 246473f53c4aSTejun Heo 24653c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 246673f53c4aSTejun Heo 246773f53c4aSTejun Heo /* 246873f53c4aSTejun Heo * Start-to-wait phase 246973f53c4aSTejun Heo */ 247073f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 247173f53c4aSTejun Heo 247273f53c4aSTejun Heo if (next_color != wq->flush_color) { 247373f53c4aSTejun Heo /* 247473f53c4aSTejun Heo * Color space is not full. The current work_color 247573f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 247673f53c4aSTejun Heo * by one. 247773f53c4aSTejun Heo */ 24786183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 247973f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 248073f53c4aSTejun Heo wq->work_color = next_color; 248173f53c4aSTejun Heo 248273f53c4aSTejun Heo if (!wq->first_flusher) { 248373f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 24846183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 248573f53c4aSTejun Heo 248673f53c4aSTejun Heo wq->first_flusher = &this_flusher; 248773f53c4aSTejun Heo 2488112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 248973f53c4aSTejun Heo wq->work_color)) { 249073f53c4aSTejun Heo /* nothing to flush, done */ 249173f53c4aSTejun Heo wq->flush_color = next_color; 249273f53c4aSTejun Heo wq->first_flusher = NULL; 249373f53c4aSTejun Heo goto out_unlock; 249473f53c4aSTejun Heo } 249573f53c4aSTejun Heo } else { 249673f53c4aSTejun Heo /* wait in queue */ 24976183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 249873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2499112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 250073f53c4aSTejun Heo } 250173f53c4aSTejun Heo } else { 250273f53c4aSTejun Heo /* 250373f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 250473f53c4aSTejun Heo * The next flush completion will assign us 250573f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 250673f53c4aSTejun Heo */ 250773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 250873f53c4aSTejun Heo } 250973f53c4aSTejun Heo 25103c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 251173f53c4aSTejun Heo 251273f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 251373f53c4aSTejun Heo 251473f53c4aSTejun Heo /* 251573f53c4aSTejun Heo * Wake-up-and-cascade phase 251673f53c4aSTejun Heo * 251773f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 251873f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 251973f53c4aSTejun Heo */ 252073f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 252173f53c4aSTejun Heo return; 252273f53c4aSTejun Heo 25233c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 252473f53c4aSTejun Heo 25254ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25264ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25274ce48b37STejun Heo goto out_unlock; 25284ce48b37STejun Heo 252973f53c4aSTejun Heo wq->first_flusher = NULL; 253073f53c4aSTejun Heo 25316183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25326183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 253373f53c4aSTejun Heo 253473f53c4aSTejun Heo while (true) { 253573f53c4aSTejun Heo struct wq_flusher *next, *tmp; 253673f53c4aSTejun Heo 253773f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 253873f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 253973f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 254073f53c4aSTejun Heo break; 254173f53c4aSTejun Heo list_del_init(&next->list); 254273f53c4aSTejun Heo complete(&next->done); 254373f53c4aSTejun Heo } 254473f53c4aSTejun Heo 25456183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 254673f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 254773f53c4aSTejun Heo 254873f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 254973f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 255073f53c4aSTejun Heo 255173f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 255273f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 255373f53c4aSTejun Heo /* 255473f53c4aSTejun Heo * Assign the same color to all overflowed 255573f53c4aSTejun Heo * flushers, advance work_color and append to 255673f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 255773f53c4aSTejun Heo * phase for these overflowed flushers. 255873f53c4aSTejun Heo */ 255973f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 256073f53c4aSTejun Heo tmp->flush_color = wq->work_color; 256173f53c4aSTejun Heo 256273f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 256373f53c4aSTejun Heo 256473f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 256573f53c4aSTejun Heo &wq->flusher_queue); 2566112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 256773f53c4aSTejun Heo } 256873f53c4aSTejun Heo 256973f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 25706183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 257173f53c4aSTejun Heo break; 257273f53c4aSTejun Heo } 257373f53c4aSTejun Heo 257473f53c4aSTejun Heo /* 257573f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2576112202d9STejun Heo * the new first flusher and arm pwqs. 257773f53c4aSTejun Heo */ 25786183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 25796183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 258073f53c4aSTejun Heo 258173f53c4aSTejun Heo list_del_init(&next->list); 258273f53c4aSTejun Heo wq->first_flusher = next; 258373f53c4aSTejun Heo 2584112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 258573f53c4aSTejun Heo break; 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo /* 258873f53c4aSTejun Heo * Meh... this color is already done, clear first 258973f53c4aSTejun Heo * flusher and repeat cascading. 259073f53c4aSTejun Heo */ 259173f53c4aSTejun Heo wq->first_flusher = NULL; 259273f53c4aSTejun Heo } 259373f53c4aSTejun Heo 259473f53c4aSTejun Heo out_unlock: 25953c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 25961da177e4SLinus Torvalds } 2597ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 25981da177e4SLinus Torvalds 25999c5a2ba7STejun Heo /** 26009c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26019c5a2ba7STejun Heo * @wq: workqueue to drain 26029c5a2ba7STejun Heo * 26039c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26049c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26059c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26069c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26079c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26089c5a2ba7STejun Heo * takes too long. 26099c5a2ba7STejun Heo */ 26109c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26119c5a2ba7STejun Heo { 26129c5a2ba7STejun Heo unsigned int flush_cnt = 0; 261349e3cf44STejun Heo struct pool_workqueue *pwq; 26149c5a2ba7STejun Heo 26159c5a2ba7STejun Heo /* 26169c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26179c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2618618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26199c5a2ba7STejun Heo */ 262087fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26219c5a2ba7STejun Heo if (!wq->nr_drainers++) 2622618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 262387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26249c5a2ba7STejun Heo reflush: 26259c5a2ba7STejun Heo flush_workqueue(wq); 26269c5a2ba7STejun Heo 2627b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 262876af4d93STejun Heo 262949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2630fa2563e4SThomas Tuttle bool drained; 26319c5a2ba7STejun Heo 2632b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2633112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2634b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2635fa2563e4SThomas Tuttle 2636fa2563e4SThomas Tuttle if (drained) 26379c5a2ba7STejun Heo continue; 26389c5a2ba7STejun Heo 26399c5a2ba7STejun Heo if (++flush_cnt == 10 || 26409c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2641c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 26429c5a2ba7STejun Heo wq->name, flush_cnt); 264376af4d93STejun Heo 2644b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 26459c5a2ba7STejun Heo goto reflush; 26469c5a2ba7STejun Heo } 26479c5a2ba7STejun Heo 26489c5a2ba7STejun Heo if (!--wq->nr_drainers) 2649618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 265087fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26519c5a2ba7STejun Heo } 26529c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 26539c5a2ba7STejun Heo 2654606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2655baf59022STejun Heo { 2656baf59022STejun Heo struct worker *worker = NULL; 2657c9e7cf27STejun Heo struct worker_pool *pool; 2658112202d9STejun Heo struct pool_workqueue *pwq; 2659baf59022STejun Heo 2660baf59022STejun Heo might_sleep(); 2661baf59022STejun Heo 2662fa1b54e6STejun Heo local_irq_disable(); 2663fa1b54e6STejun Heo pool = get_work_pool(work); 2664fa1b54e6STejun Heo if (!pool) { 2665fa1b54e6STejun Heo local_irq_enable(); 2666fa1b54e6STejun Heo return false; 2667fa1b54e6STejun Heo } 2668fa1b54e6STejun Heo 2669fa1b54e6STejun Heo spin_lock(&pool->lock); 26700b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2671112202d9STejun Heo pwq = get_work_pwq(work); 2672112202d9STejun Heo if (pwq) { 2673112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2674baf59022STejun Heo goto already_gone; 2675606a5020STejun Heo } else { 2676c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2677baf59022STejun Heo if (!worker) 2678baf59022STejun Heo goto already_gone; 2679112202d9STejun Heo pwq = worker->current_pwq; 2680606a5020STejun Heo } 2681baf59022STejun Heo 2682112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2683d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2684baf59022STejun Heo 2685e159489bSTejun Heo /* 2686e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2687e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2688e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2689e159489bSTejun Heo * access. 2690e159489bSTejun Heo */ 2691493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2692112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2693e159489bSTejun Heo else 2694112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2695112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2696e159489bSTejun Heo 2697baf59022STejun Heo return true; 2698baf59022STejun Heo already_gone: 2699d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2700baf59022STejun Heo return false; 2701baf59022STejun Heo } 2702baf59022STejun Heo 2703db700897SOleg Nesterov /** 2704401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2705401a8d04STejun Heo * @work: the work to flush 2706db700897SOleg Nesterov * 2707606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2708606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2709401a8d04STejun Heo * 2710d185af30SYacine Belkadi * Return: 2711401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2712401a8d04STejun Heo * %false if it was already idle. 2713db700897SOleg Nesterov */ 2714401a8d04STejun Heo bool flush_work(struct work_struct *work) 2715db700897SOleg Nesterov { 271612997d1aSBjorn Helgaas struct wq_barrier barr; 271712997d1aSBjorn Helgaas 27180976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27190976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27200976dfc1SStephen Boyd 272112997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 272212997d1aSBjorn Helgaas wait_for_completion(&barr.done); 272312997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 272412997d1aSBjorn Helgaas return true; 272512997d1aSBjorn Helgaas } else { 272612997d1aSBjorn Helgaas return false; 272712997d1aSBjorn Helgaas } 2728606a5020STejun Heo } 2729db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2730db700897SOleg Nesterov 273136e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2732401a8d04STejun Heo { 2733bbb68dfaSTejun Heo unsigned long flags; 27341f1f642eSOleg Nesterov int ret; 27351f1f642eSOleg Nesterov 27361f1f642eSOleg Nesterov do { 2737bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2738bbb68dfaSTejun Heo /* 2739bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2740bbb68dfaSTejun Heo * would be waiting for before retrying. 2741bbb68dfaSTejun Heo */ 2742bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2743606a5020STejun Heo flush_work(work); 27441f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 27451f1f642eSOleg Nesterov 2746bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2747bbb68dfaSTejun Heo mark_work_canceling(work); 2748bbb68dfaSTejun Heo local_irq_restore(flags); 2749bbb68dfaSTejun Heo 2750606a5020STejun Heo flush_work(work); 27517a22ad75STejun Heo clear_work_data(work); 27521f1f642eSOleg Nesterov return ret; 27531f1f642eSOleg Nesterov } 27541f1f642eSOleg Nesterov 27556e84d644SOleg Nesterov /** 2756401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2757401a8d04STejun Heo * @work: the work to cancel 27586e84d644SOleg Nesterov * 2759401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2760401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2761401a8d04STejun Heo * another workqueue. On return from this function, @work is 2762401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 27631f1f642eSOleg Nesterov * 2764401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2765401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 27666e84d644SOleg Nesterov * 2767401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 27686e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2769401a8d04STejun Heo * 2770d185af30SYacine Belkadi * Return: 2771401a8d04STejun Heo * %true if @work was pending, %false otherwise. 27726e84d644SOleg Nesterov */ 2773401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 27746e84d644SOleg Nesterov { 277536e227d2STejun Heo return __cancel_work_timer(work, false); 2776b89deed3SOleg Nesterov } 277728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2778b89deed3SOleg Nesterov 27796e84d644SOleg Nesterov /** 2780401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2781401a8d04STejun Heo * @dwork: the delayed work to flush 27826e84d644SOleg Nesterov * 2783401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2784401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2785401a8d04STejun Heo * considers the last queueing instance of @dwork. 27861f1f642eSOleg Nesterov * 2787d185af30SYacine Belkadi * Return: 2788401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2789401a8d04STejun Heo * %false if it was already idle. 27906e84d644SOleg Nesterov */ 2791401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2792401a8d04STejun Heo { 27938930cabaSTejun Heo local_irq_disable(); 2794401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 279560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 27968930cabaSTejun Heo local_irq_enable(); 2797401a8d04STejun Heo return flush_work(&dwork->work); 2798401a8d04STejun Heo } 2799401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2800401a8d04STejun Heo 2801401a8d04STejun Heo /** 280257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 280357b30ae7STejun Heo * @dwork: delayed_work to cancel 280409383498STejun Heo * 2805d185af30SYacine Belkadi * Kill off a pending delayed_work. 2806d185af30SYacine Belkadi * 2807d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2808d185af30SYacine Belkadi * pending. 2809d185af30SYacine Belkadi * 2810d185af30SYacine Belkadi * Note: 2811d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2812d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2813d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 281409383498STejun Heo * 281557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 281609383498STejun Heo */ 281757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 281809383498STejun Heo { 281957b30ae7STejun Heo unsigned long flags; 282057b30ae7STejun Heo int ret; 282157b30ae7STejun Heo 282257b30ae7STejun Heo do { 282357b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 282457b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 282557b30ae7STejun Heo 282657b30ae7STejun Heo if (unlikely(ret < 0)) 282757b30ae7STejun Heo return false; 282857b30ae7STejun Heo 28297c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28307c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 283157b30ae7STejun Heo local_irq_restore(flags); 2832c0158ca6SDan Magenheimer return ret; 283309383498STejun Heo } 283457b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 283509383498STejun Heo 283609383498STejun Heo /** 2837401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2838401a8d04STejun Heo * @dwork: the delayed work cancel 2839401a8d04STejun Heo * 2840401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2841401a8d04STejun Heo * 2842d185af30SYacine Belkadi * Return: 2843401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2844401a8d04STejun Heo */ 2845401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 28466e84d644SOleg Nesterov { 284736e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 28486e84d644SOleg Nesterov } 2849f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 28501da177e4SLinus Torvalds 28510fcb78c2SRolf Eike Beer /** 285231ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2853b6136773SAndrew Morton * @func: the function to call 2854b6136773SAndrew Morton * 285531ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 285631ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2857b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 285831ddd871STejun Heo * 2859d185af30SYacine Belkadi * Return: 286031ddd871STejun Heo * 0 on success, -errno on failure. 2861b6136773SAndrew Morton */ 286265f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 286315316ba8SChristoph Lameter { 286415316ba8SChristoph Lameter int cpu; 286538f51568SNamhyung Kim struct work_struct __percpu *works; 286615316ba8SChristoph Lameter 2867b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2868b6136773SAndrew Morton if (!works) 286915316ba8SChristoph Lameter return -ENOMEM; 2870b6136773SAndrew Morton 287195402b38SGautham R Shenoy get_online_cpus(); 287293981800STejun Heo 287315316ba8SChristoph Lameter for_each_online_cpu(cpu) { 28749bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 28759bfb1839SIngo Molnar 28769bfb1839SIngo Molnar INIT_WORK(work, func); 28778de6d308SOleg Nesterov schedule_work_on(cpu, work); 287815316ba8SChristoph Lameter } 287993981800STejun Heo 288093981800STejun Heo for_each_online_cpu(cpu) 28818616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 288293981800STejun Heo 288395402b38SGautham R Shenoy put_online_cpus(); 2884b6136773SAndrew Morton free_percpu(works); 288515316ba8SChristoph Lameter return 0; 288615316ba8SChristoph Lameter } 288715316ba8SChristoph Lameter 2888eef6a7d5SAlan Stern /** 2889eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2890eef6a7d5SAlan Stern * 2891eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2892eef6a7d5SAlan Stern * completion. 2893eef6a7d5SAlan Stern * 2894eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2895eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2896eef6a7d5SAlan Stern * will lead to deadlock: 2897eef6a7d5SAlan Stern * 2898eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2899eef6a7d5SAlan Stern * a lock held by your code or its caller. 2900eef6a7d5SAlan Stern * 2901eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2902eef6a7d5SAlan Stern * 2903eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2904eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2905eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2906eef6a7d5SAlan Stern * 2907eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2908eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2909eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2910eef6a7d5SAlan Stern * cancel_work_sync() instead. 2911eef6a7d5SAlan Stern */ 29121da177e4SLinus Torvalds void flush_scheduled_work(void) 29131da177e4SLinus Torvalds { 2914d320c038STejun Heo flush_workqueue(system_wq); 29151da177e4SLinus Torvalds } 2916ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29171da177e4SLinus Torvalds 29181da177e4SLinus Torvalds /** 29191fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29201fa44ecaSJames Bottomley * @fn: the function to execute 29211fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29221fa44ecaSJames Bottomley * be available when the work executes) 29231fa44ecaSJames Bottomley * 29241fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29251fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29261fa44ecaSJames Bottomley * 2927d185af30SYacine Belkadi * Return: 0 - function was executed 29281fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29291fa44ecaSJames Bottomley */ 293065f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29311fa44ecaSJames Bottomley { 29321fa44ecaSJames Bottomley if (!in_interrupt()) { 293365f27f38SDavid Howells fn(&ew->work); 29341fa44ecaSJames Bottomley return 0; 29351fa44ecaSJames Bottomley } 29361fa44ecaSJames Bottomley 293765f27f38SDavid Howells INIT_WORK(&ew->work, fn); 29381fa44ecaSJames Bottomley schedule_work(&ew->work); 29391fa44ecaSJames Bottomley 29401fa44ecaSJames Bottomley return 1; 29411fa44ecaSJames Bottomley } 29421fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 29431fa44ecaSJames Bottomley 2944226223abSTejun Heo #ifdef CONFIG_SYSFS 2945226223abSTejun Heo /* 2946226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 2947226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 2948226223abSTejun Heo * following attributes. 2949226223abSTejun Heo * 2950226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 2951226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 2952226223abSTejun Heo * 2953226223abSTejun Heo * Unbound workqueues have the following extra attributes. 2954226223abSTejun Heo * 2955226223abSTejun Heo * id RO int : the associated pool ID 2956226223abSTejun Heo * nice RW int : nice value of the workers 2957226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 2958226223abSTejun Heo */ 2959226223abSTejun Heo struct wq_device { 2960226223abSTejun Heo struct workqueue_struct *wq; 2961226223abSTejun Heo struct device dev; 2962226223abSTejun Heo }; 2963226223abSTejun Heo 2964226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 2965226223abSTejun Heo { 2966226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 2967226223abSTejun Heo 2968226223abSTejun Heo return wq_dev->wq; 2969226223abSTejun Heo } 2970226223abSTejun Heo 29711a6661daSGreg Kroah-Hartman static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 29721a6661daSGreg Kroah-Hartman char *buf) 2973226223abSTejun Heo { 2974226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 2975226223abSTejun Heo 2976226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 2977226223abSTejun Heo } 29781a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RO(per_cpu); 2979226223abSTejun Heo 29801a6661daSGreg Kroah-Hartman static ssize_t max_active_show(struct device *dev, 2981226223abSTejun Heo struct device_attribute *attr, char *buf) 2982226223abSTejun Heo { 2983226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 2984226223abSTejun Heo 2985226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 2986226223abSTejun Heo } 2987226223abSTejun Heo 29881a6661daSGreg Kroah-Hartman static ssize_t max_active_store(struct device *dev, 29891a6661daSGreg Kroah-Hartman struct device_attribute *attr, const char *buf, 29901a6661daSGreg Kroah-Hartman size_t count) 2991226223abSTejun Heo { 2992226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 2993226223abSTejun Heo int val; 2994226223abSTejun Heo 2995226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 2996226223abSTejun Heo return -EINVAL; 2997226223abSTejun Heo 2998226223abSTejun Heo workqueue_set_max_active(wq, val); 2999226223abSTejun Heo return count; 3000226223abSTejun Heo } 30011a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RW(max_active); 3002226223abSTejun Heo 30031a6661daSGreg Kroah-Hartman static struct attribute *wq_sysfs_attrs[] = { 30041a6661daSGreg Kroah-Hartman &dev_attr_per_cpu.attr, 30051a6661daSGreg Kroah-Hartman &dev_attr_max_active.attr, 30061a6661daSGreg Kroah-Hartman NULL, 3007226223abSTejun Heo }; 30081a6661daSGreg Kroah-Hartman ATTRIBUTE_GROUPS(wq_sysfs); 3009226223abSTejun Heo 3010d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3011226223abSTejun Heo struct device_attribute *attr, char *buf) 3012226223abSTejun Heo { 3013226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3014d55262c4STejun Heo const char *delim = ""; 3015d55262c4STejun Heo int node, written = 0; 3016226223abSTejun Heo 3017226223abSTejun Heo rcu_read_lock_sched(); 3018d55262c4STejun Heo for_each_node(node) { 3019d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3020d55262c4STejun Heo "%s%d:%d", delim, node, 3021d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3022d55262c4STejun Heo delim = " "; 3023d55262c4STejun Heo } 3024d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3025226223abSTejun Heo rcu_read_unlock_sched(); 3026226223abSTejun Heo 3027226223abSTejun Heo return written; 3028226223abSTejun Heo } 3029226223abSTejun Heo 3030226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3031226223abSTejun Heo char *buf) 3032226223abSTejun Heo { 3033226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3034226223abSTejun Heo int written; 3035226223abSTejun Heo 30366029a918STejun Heo mutex_lock(&wq->mutex); 30376029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 30386029a918STejun Heo mutex_unlock(&wq->mutex); 3039226223abSTejun Heo 3040226223abSTejun Heo return written; 3041226223abSTejun Heo } 3042226223abSTejun Heo 3043226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3044226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3045226223abSTejun Heo { 3046226223abSTejun Heo struct workqueue_attrs *attrs; 3047226223abSTejun Heo 3048226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3049226223abSTejun Heo if (!attrs) 3050226223abSTejun Heo return NULL; 3051226223abSTejun Heo 30526029a918STejun Heo mutex_lock(&wq->mutex); 30536029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 30546029a918STejun Heo mutex_unlock(&wq->mutex); 3055226223abSTejun Heo return attrs; 3056226223abSTejun Heo } 3057226223abSTejun Heo 3058226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3059226223abSTejun Heo const char *buf, size_t count) 3060226223abSTejun Heo { 3061226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3062226223abSTejun Heo struct workqueue_attrs *attrs; 3063226223abSTejun Heo int ret; 3064226223abSTejun Heo 3065226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3066226223abSTejun Heo if (!attrs) 3067226223abSTejun Heo return -ENOMEM; 3068226223abSTejun Heo 3069226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 307014481842SDongsheng Yang attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 3071226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3072226223abSTejun Heo else 3073226223abSTejun Heo ret = -EINVAL; 3074226223abSTejun Heo 3075226223abSTejun Heo free_workqueue_attrs(attrs); 3076226223abSTejun Heo return ret ?: count; 3077226223abSTejun Heo } 3078226223abSTejun Heo 3079226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3080226223abSTejun Heo struct device_attribute *attr, char *buf) 3081226223abSTejun Heo { 3082226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3083226223abSTejun Heo int written; 3084226223abSTejun Heo 30856029a918STejun Heo mutex_lock(&wq->mutex); 30866029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 30876029a918STejun Heo mutex_unlock(&wq->mutex); 3088226223abSTejun Heo 3089226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3090226223abSTejun Heo return written; 3091226223abSTejun Heo } 3092226223abSTejun Heo 3093226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3094226223abSTejun Heo struct device_attribute *attr, 3095226223abSTejun Heo const char *buf, size_t count) 3096226223abSTejun Heo { 3097226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3098226223abSTejun Heo struct workqueue_attrs *attrs; 3099226223abSTejun Heo int ret; 3100226223abSTejun Heo 3101226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3102226223abSTejun Heo if (!attrs) 3103226223abSTejun Heo return -ENOMEM; 3104226223abSTejun Heo 3105226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3106226223abSTejun Heo if (!ret) 3107226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3108226223abSTejun Heo 3109226223abSTejun Heo free_workqueue_attrs(attrs); 3110226223abSTejun Heo return ret ?: count; 3111226223abSTejun Heo } 3112226223abSTejun Heo 3113d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3114d55262c4STejun Heo char *buf) 3115d55262c4STejun Heo { 3116d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3117d55262c4STejun Heo int written; 3118d55262c4STejun Heo 3119d55262c4STejun Heo mutex_lock(&wq->mutex); 3120d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3121d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3122d55262c4STejun Heo mutex_unlock(&wq->mutex); 3123d55262c4STejun Heo 3124d55262c4STejun Heo return written; 3125d55262c4STejun Heo } 3126d55262c4STejun Heo 3127d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3128d55262c4STejun Heo const char *buf, size_t count) 3129d55262c4STejun Heo { 3130d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3131d55262c4STejun Heo struct workqueue_attrs *attrs; 3132d55262c4STejun Heo int v, ret; 3133d55262c4STejun Heo 3134d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3135d55262c4STejun Heo if (!attrs) 3136d55262c4STejun Heo return -ENOMEM; 3137d55262c4STejun Heo 3138d55262c4STejun Heo ret = -EINVAL; 3139d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3140d55262c4STejun Heo attrs->no_numa = !v; 3141d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3142d55262c4STejun Heo } 3143d55262c4STejun Heo 3144d55262c4STejun Heo free_workqueue_attrs(attrs); 3145d55262c4STejun Heo return ret ?: count; 3146d55262c4STejun Heo } 3147d55262c4STejun Heo 3148226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3149d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3150226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3151226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3152d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3153226223abSTejun Heo __ATTR_NULL, 3154226223abSTejun Heo }; 3155226223abSTejun Heo 3156226223abSTejun Heo static struct bus_type wq_subsys = { 3157226223abSTejun Heo .name = "workqueue", 31581a6661daSGreg Kroah-Hartman .dev_groups = wq_sysfs_groups, 3159226223abSTejun Heo }; 3160226223abSTejun Heo 3161226223abSTejun Heo static int __init wq_sysfs_init(void) 3162226223abSTejun Heo { 3163226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3164226223abSTejun Heo } 3165226223abSTejun Heo core_initcall(wq_sysfs_init); 3166226223abSTejun Heo 3167226223abSTejun Heo static void wq_device_release(struct device *dev) 3168226223abSTejun Heo { 3169226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3170226223abSTejun Heo 3171226223abSTejun Heo kfree(wq_dev); 3172226223abSTejun Heo } 3173226223abSTejun Heo 3174226223abSTejun Heo /** 3175226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3176226223abSTejun Heo * @wq: the workqueue to register 3177226223abSTejun Heo * 3178226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3179226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3180226223abSTejun Heo * which is the preferred method. 3181226223abSTejun Heo * 3182226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3183226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3184226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3185226223abSTejun Heo * attributes. 3186226223abSTejun Heo * 3187d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. 3188226223abSTejun Heo */ 3189226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3190226223abSTejun Heo { 3191226223abSTejun Heo struct wq_device *wq_dev; 3192226223abSTejun Heo int ret; 3193226223abSTejun Heo 3194226223abSTejun Heo /* 3195226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3196226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3197226223abSTejun Heo * workqueues. 3198226223abSTejun Heo */ 3199226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3200226223abSTejun Heo return -EINVAL; 3201226223abSTejun Heo 3202226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3203226223abSTejun Heo if (!wq_dev) 3204226223abSTejun Heo return -ENOMEM; 3205226223abSTejun Heo 3206226223abSTejun Heo wq_dev->wq = wq; 3207226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3208226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3209226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3210226223abSTejun Heo 3211226223abSTejun Heo /* 3212226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3213226223abSTejun Heo * everything is ready. 3214226223abSTejun Heo */ 3215226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3216226223abSTejun Heo 3217226223abSTejun Heo ret = device_register(&wq_dev->dev); 3218226223abSTejun Heo if (ret) { 3219226223abSTejun Heo kfree(wq_dev); 3220226223abSTejun Heo wq->wq_dev = NULL; 3221226223abSTejun Heo return ret; 3222226223abSTejun Heo } 3223226223abSTejun Heo 3224226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3225226223abSTejun Heo struct device_attribute *attr; 3226226223abSTejun Heo 3227226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3228226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3229226223abSTejun Heo if (ret) { 3230226223abSTejun Heo device_unregister(&wq_dev->dev); 3231226223abSTejun Heo wq->wq_dev = NULL; 3232226223abSTejun Heo return ret; 3233226223abSTejun Heo } 3234226223abSTejun Heo } 3235226223abSTejun Heo } 3236226223abSTejun Heo 3237bddbceb6SMaxime Bizon dev_set_uevent_suppress(&wq_dev->dev, false); 3238226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3239226223abSTejun Heo return 0; 3240226223abSTejun Heo } 3241226223abSTejun Heo 3242226223abSTejun Heo /** 3243226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3244226223abSTejun Heo * @wq: the workqueue to unregister 3245226223abSTejun Heo * 3246226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3247226223abSTejun Heo */ 3248226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3249226223abSTejun Heo { 3250226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3251226223abSTejun Heo 3252226223abSTejun Heo if (!wq->wq_dev) 3253226223abSTejun Heo return; 3254226223abSTejun Heo 3255226223abSTejun Heo wq->wq_dev = NULL; 3256226223abSTejun Heo device_unregister(&wq_dev->dev); 3257226223abSTejun Heo } 3258226223abSTejun Heo #else /* CONFIG_SYSFS */ 3259226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3260226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3261226223abSTejun Heo 32627a4e344cSTejun Heo /** 32637a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 32647a4e344cSTejun Heo * @attrs: workqueue_attrs to free 32657a4e344cSTejun Heo * 32667a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 32677a4e344cSTejun Heo */ 32687a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 32697a4e344cSTejun Heo { 32707a4e344cSTejun Heo if (attrs) { 32717a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 32727a4e344cSTejun Heo kfree(attrs); 32737a4e344cSTejun Heo } 32747a4e344cSTejun Heo } 32757a4e344cSTejun Heo 32767a4e344cSTejun Heo /** 32777a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 32787a4e344cSTejun Heo * @gfp_mask: allocation mask to use 32797a4e344cSTejun Heo * 32807a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3281d185af30SYacine Belkadi * return it. 3282d185af30SYacine Belkadi * 3283d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 32847a4e344cSTejun Heo */ 32857a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 32867a4e344cSTejun Heo { 32877a4e344cSTejun Heo struct workqueue_attrs *attrs; 32887a4e344cSTejun Heo 32897a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 32907a4e344cSTejun Heo if (!attrs) 32917a4e344cSTejun Heo goto fail; 32927a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 32937a4e344cSTejun Heo goto fail; 32947a4e344cSTejun Heo 329513e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 32967a4e344cSTejun Heo return attrs; 32977a4e344cSTejun Heo fail: 32987a4e344cSTejun Heo free_workqueue_attrs(attrs); 32997a4e344cSTejun Heo return NULL; 33007a4e344cSTejun Heo } 33017a4e344cSTejun Heo 330229c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 330329c91e99STejun Heo const struct workqueue_attrs *from) 330429c91e99STejun Heo { 330529c91e99STejun Heo to->nice = from->nice; 330629c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 33072865a8fbSShaohua Li /* 33082865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 33092865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 33102865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 33112865a8fbSShaohua Li */ 33122865a8fbSShaohua Li to->no_numa = from->no_numa; 331329c91e99STejun Heo } 331429c91e99STejun Heo 331529c91e99STejun Heo /* hash value of the content of @attr */ 331629c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 331729c91e99STejun Heo { 331829c91e99STejun Heo u32 hash = 0; 331929c91e99STejun Heo 332029c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 332113e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 332213e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 332329c91e99STejun Heo return hash; 332429c91e99STejun Heo } 332529c91e99STejun Heo 332629c91e99STejun Heo /* content equality test */ 332729c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 332829c91e99STejun Heo const struct workqueue_attrs *b) 332929c91e99STejun Heo { 333029c91e99STejun Heo if (a->nice != b->nice) 333129c91e99STejun Heo return false; 333229c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 333329c91e99STejun Heo return false; 333429c91e99STejun Heo return true; 333529c91e99STejun Heo } 333629c91e99STejun Heo 33377a4e344cSTejun Heo /** 33387a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33397a4e344cSTejun Heo * @pool: worker_pool to initialize 33407a4e344cSTejun Heo * 33417a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 3342d185af30SYacine Belkadi * 3343d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 334429c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 334529c91e99STejun Heo * on @pool safely to release it. 33467a4e344cSTejun Heo */ 33477a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33484e1a1f9aSTejun Heo { 33494e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 335029c91e99STejun Heo pool->id = -1; 335129c91e99STejun Heo pool->cpu = -1; 3352f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 33534e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33544e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33554e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33564e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33574e1a1f9aSTejun Heo 33584e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33594e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33604e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33614e1a1f9aSTejun Heo 33624e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 33634e1a1f9aSTejun Heo (unsigned long)pool); 33644e1a1f9aSTejun Heo 33654e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 336692f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3367da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 33687a4e344cSTejun Heo 33697cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 337029c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 337129c91e99STejun Heo pool->refcnt = 1; 337229c91e99STejun Heo 337329c91e99STejun Heo /* shouldn't fail above this point */ 33747a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 33757a4e344cSTejun Heo if (!pool->attrs) 33767a4e344cSTejun Heo return -ENOMEM; 33777a4e344cSTejun Heo return 0; 33784e1a1f9aSTejun Heo } 33794e1a1f9aSTejun Heo 338029c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 338129c91e99STejun Heo { 338229c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 338329c91e99STejun Heo 33847cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 338529c91e99STejun Heo free_workqueue_attrs(pool->attrs); 338629c91e99STejun Heo kfree(pool); 338729c91e99STejun Heo } 338829c91e99STejun Heo 338929c91e99STejun Heo /** 339029c91e99STejun Heo * put_unbound_pool - put a worker_pool 339129c91e99STejun Heo * @pool: worker_pool to put 339229c91e99STejun Heo * 339329c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3394c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3395c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3396c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3397a892caccSTejun Heo * 3398a892caccSTejun Heo * Should be called with wq_pool_mutex held. 339929c91e99STejun Heo */ 340029c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 340129c91e99STejun Heo { 340260f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 340329c91e99STejun Heo struct worker *worker; 340429c91e99STejun Heo 3405a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3406a892caccSTejun Heo 3407a892caccSTejun Heo if (--pool->refcnt) 340829c91e99STejun Heo return; 340929c91e99STejun Heo 341029c91e99STejun Heo /* sanity checks */ 341161d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3412a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 341329c91e99STejun Heo return; 341429c91e99STejun Heo 341529c91e99STejun Heo /* release id and unhash */ 341629c91e99STejun Heo if (pool->id >= 0) 341729c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 341829c91e99STejun Heo hash_del(&pool->hash_node); 341929c91e99STejun Heo 3420c5aa87bbSTejun Heo /* 3421c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3422c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 342392f9c5c4SLai Jiangshan * attach_mutex. 3424c5aa87bbSTejun Heo */ 342529c91e99STejun Heo mutex_lock(&pool->manager_arb); 342629c91e99STejun Heo 342760f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 34281037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 342929c91e99STejun Heo destroy_worker(worker); 343029c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 343129c91e99STejun Heo spin_unlock_irq(&pool->lock); 343260f5a4bcSLai Jiangshan 343392f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3434da028469SLai Jiangshan if (!list_empty(&pool->workers)) 343560f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 343692f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 343760f5a4bcSLai Jiangshan 343860f5a4bcSLai Jiangshan if (pool->detach_completion) 343960f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 344060f5a4bcSLai Jiangshan 344129c91e99STejun Heo mutex_unlock(&pool->manager_arb); 344229c91e99STejun Heo 344329c91e99STejun Heo /* shut down the timers */ 344429c91e99STejun Heo del_timer_sync(&pool->idle_timer); 344529c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 344629c91e99STejun Heo 344729c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 344829c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 344929c91e99STejun Heo } 345029c91e99STejun Heo 345129c91e99STejun Heo /** 345229c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 345329c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 345429c91e99STejun Heo * 345529c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 345629c91e99STejun Heo * reference count and return it. If there already is a matching 345729c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3458d185af30SYacine Belkadi * create a new one. 3459a892caccSTejun Heo * 3460a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3461d185af30SYacine Belkadi * 3462d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3463d185af30SYacine Belkadi * On failure, %NULL. 346429c91e99STejun Heo */ 346529c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 346629c91e99STejun Heo { 346729c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 346829c91e99STejun Heo struct worker_pool *pool; 3469f3f90ad4STejun Heo int node; 347029c91e99STejun Heo 3471a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 347229c91e99STejun Heo 347329c91e99STejun Heo /* do we already have a matching pool? */ 347429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 347529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 347629c91e99STejun Heo pool->refcnt++; 34773fb1823cSLai Jiangshan return pool; 347829c91e99STejun Heo } 347929c91e99STejun Heo } 348029c91e99STejun Heo 348129c91e99STejun Heo /* nope, create a new one */ 348229c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 348329c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 348429c91e99STejun Heo goto fail; 348529c91e99STejun Heo 34868864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 348729c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 348829c91e99STejun Heo 34892865a8fbSShaohua Li /* 34902865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 34912865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 34922865a8fbSShaohua Li */ 34932865a8fbSShaohua Li pool->attrs->no_numa = false; 34942865a8fbSShaohua Li 3495f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3496f3f90ad4STejun Heo if (wq_numa_enabled) { 3497f3f90ad4STejun Heo for_each_node(node) { 3498f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3499f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3500f3f90ad4STejun Heo pool->node = node; 3501f3f90ad4STejun Heo break; 3502f3f90ad4STejun Heo } 3503f3f90ad4STejun Heo } 3504f3f90ad4STejun Heo } 3505f3f90ad4STejun Heo 350629c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 350729c91e99STejun Heo goto fail; 350829c91e99STejun Heo 350929c91e99STejun Heo /* create and start the initial worker */ 3510051e1850SLai Jiangshan if (!create_worker(pool)) 351129c91e99STejun Heo goto fail; 351229c91e99STejun Heo 351329c91e99STejun Heo /* install */ 351429c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 35153fb1823cSLai Jiangshan 351629c91e99STejun Heo return pool; 351729c91e99STejun Heo fail: 351829c91e99STejun Heo if (pool) 351929c91e99STejun Heo put_unbound_pool(pool); 352029c91e99STejun Heo return NULL; 352129c91e99STejun Heo } 352229c91e99STejun Heo 35238864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35248864b4e5STejun Heo { 35258864b4e5STejun Heo kmem_cache_free(pwq_cache, 35268864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35278864b4e5STejun Heo } 35288864b4e5STejun Heo 35298864b4e5STejun Heo /* 35308864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35318864b4e5STejun Heo * and needs to be destroyed. 35328864b4e5STejun Heo */ 35338864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35348864b4e5STejun Heo { 35358864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35368864b4e5STejun Heo unbound_release_work); 35378864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35388864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3539bc0caf09STejun Heo bool is_last; 35408864b4e5STejun Heo 35418864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35428864b4e5STejun Heo return; 35438864b4e5STejun Heo 35443c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 35458864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3546bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35473c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35488864b4e5STejun Heo 3549a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35508864b4e5STejun Heo put_unbound_pool(pool); 3551a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3552a892caccSTejun Heo 35538864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35548864b4e5STejun Heo 35558864b4e5STejun Heo /* 35568864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 35578864b4e5STejun Heo * is gonna access it anymore. Free it. 35588864b4e5STejun Heo */ 35596029a918STejun Heo if (is_last) { 35606029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 35618864b4e5STejun Heo kfree(wq); 35628864b4e5STejun Heo } 35636029a918STejun Heo } 35648864b4e5STejun Heo 35650fbd95aaSTejun Heo /** 3566699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35670fbd95aaSTejun Heo * @pwq: target pool_workqueue 35680fbd95aaSTejun Heo * 3569699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3570699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3571699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35720fbd95aaSTejun Heo */ 3573699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35740fbd95aaSTejun Heo { 3575699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3576699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3577699ce097STejun Heo 3578699ce097STejun Heo /* for @wq->saved_max_active */ 3579a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3580699ce097STejun Heo 3581699ce097STejun Heo /* fast exit for non-freezable wqs */ 3582699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3583699ce097STejun Heo return; 3584699ce097STejun Heo 3585a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3586699ce097STejun Heo 358774b414eaSLai Jiangshan /* 358874b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 358974b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 359074b414eaSLai Jiangshan * is updated and visible. 359174b414eaSLai Jiangshan */ 359274b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3593699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35940fbd95aaSTejun Heo 35950fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35960fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 35970fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3598951a078aSLai Jiangshan 3599951a078aSLai Jiangshan /* 3600951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3601951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3602951a078aSLai Jiangshan */ 3603951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3604699ce097STejun Heo } else { 3605699ce097STejun Heo pwq->max_active = 0; 3606699ce097STejun Heo } 3607699ce097STejun Heo 3608a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36090fbd95aaSTejun Heo } 36100fbd95aaSTejun Heo 3611e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3612f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3613f147f29eSTejun Heo struct worker_pool *pool) 3614d2c1d404STejun Heo { 3615d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3616d2c1d404STejun Heo 3617e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3618e50aba9aSTejun Heo 3619d2c1d404STejun Heo pwq->pool = pool; 3620d2c1d404STejun Heo pwq->wq = wq; 3621d2c1d404STejun Heo pwq->flush_color = -1; 36228864b4e5STejun Heo pwq->refcnt = 1; 3623d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 36241befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3625d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36268864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3627f147f29eSTejun Heo } 3628d2c1d404STejun Heo 3629f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 36301befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3631f147f29eSTejun Heo { 3632f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3633f147f29eSTejun Heo 3634f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 363575ccf595STejun Heo 36361befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 36371befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 36381befcf30STejun Heo return; 36391befcf30STejun Heo 364029b1cb41SLai Jiangshan /* set the matching work_color */ 364175ccf595STejun Heo pwq->work_color = wq->work_color; 3642983ca25eSTejun Heo 3643983ca25eSTejun Heo /* sync max_active to the current setting */ 3644983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3645983ca25eSTejun Heo 3646983ca25eSTejun Heo /* link in @pwq */ 36479e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3648df2d5ae4STejun Heo } 36496029a918STejun Heo 3650f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3651f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3652f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3653f147f29eSTejun Heo { 3654f147f29eSTejun Heo struct worker_pool *pool; 3655f147f29eSTejun Heo struct pool_workqueue *pwq; 3656f147f29eSTejun Heo 3657f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3658f147f29eSTejun Heo 3659f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3660f147f29eSTejun Heo if (!pool) 3661f147f29eSTejun Heo return NULL; 3662f147f29eSTejun Heo 3663e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3664f147f29eSTejun Heo if (!pwq) { 3665f147f29eSTejun Heo put_unbound_pool(pool); 3666f147f29eSTejun Heo return NULL; 3667f147f29eSTejun Heo } 3668f147f29eSTejun Heo 3669f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3670f147f29eSTejun Heo return pwq; 3671d2c1d404STejun Heo } 3672d2c1d404STejun Heo 36734c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 36744c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 36754c16bd32STejun Heo { 36764c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 36774c16bd32STejun Heo 36784c16bd32STejun Heo if (pwq) { 36794c16bd32STejun Heo put_unbound_pool(pwq->pool); 3680cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 36814c16bd32STejun Heo } 36824c16bd32STejun Heo } 36834c16bd32STejun Heo 36844c16bd32STejun Heo /** 36854c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 36864c16bd32STejun Heo * @attrs: the wq_attrs of interest 36874c16bd32STejun Heo * @node: the target NUMA node 36884c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 36894c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 36904c16bd32STejun Heo * 36914c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 36924c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3693d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 36944c16bd32STejun Heo * 36954c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 36964c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 36974c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 36984c16bd32STejun Heo * @attrs->cpumask. 36994c16bd32STejun Heo * 37004c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 37014c16bd32STejun Heo * stable. 3702d185af30SYacine Belkadi * 3703d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3704d185af30SYacine Belkadi * %false if equal. 37054c16bd32STejun Heo */ 37064c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 37074c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 37084c16bd32STejun Heo { 3709d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 37104c16bd32STejun Heo goto use_dfl; 37114c16bd32STejun Heo 37124c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 37134c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 37144c16bd32STejun Heo if (cpu_going_down >= 0) 37154c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 37164c16bd32STejun Heo 37174c16bd32STejun Heo if (cpumask_empty(cpumask)) 37184c16bd32STejun Heo goto use_dfl; 37194c16bd32STejun Heo 37204c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 37214c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 37224c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 37234c16bd32STejun Heo 37244c16bd32STejun Heo use_dfl: 37254c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 37264c16bd32STejun Heo return false; 37274c16bd32STejun Heo } 37284c16bd32STejun Heo 37291befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 37301befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 37311befcf30STejun Heo int node, 37321befcf30STejun Heo struct pool_workqueue *pwq) 37331befcf30STejun Heo { 37341befcf30STejun Heo struct pool_workqueue *old_pwq; 37351befcf30STejun Heo 37361befcf30STejun Heo lockdep_assert_held(&wq->mutex); 37371befcf30STejun Heo 37381befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 37391befcf30STejun Heo link_pwq(pwq); 37401befcf30STejun Heo 37411befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 37421befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 37431befcf30STejun Heo return old_pwq; 37441befcf30STejun Heo } 37451befcf30STejun Heo 37469e8cd2f5STejun Heo /** 37479e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37489e8cd2f5STejun Heo * @wq: the target workqueue 37499e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37509e8cd2f5STejun Heo * 37514c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 37524c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 37534c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 37544c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 37554c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 37564c16bd32STejun Heo * back-to-back will stay on its current pwq. 37579e8cd2f5STejun Heo * 3758d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3759d185af30SYacine Belkadi * 3760d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 37619e8cd2f5STejun Heo */ 37629e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 37639e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 37649e8cd2f5STejun Heo { 37654c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 37664c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3767f147f29eSTejun Heo int node, ret; 37689e8cd2f5STejun Heo 37698719dceaSTejun Heo /* only unbound workqueues can change attributes */ 37709e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 37719e8cd2f5STejun Heo return -EINVAL; 37729e8cd2f5STejun Heo 37738719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 37748719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 37758719dceaSTejun Heo return -EINVAL; 37768719dceaSTejun Heo 3777ddcb57e2SLai Jiangshan pwq_tbl = kzalloc(nr_node_ids * sizeof(pwq_tbl[0]), GFP_KERNEL); 377813e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37794c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37804c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 378113e2e556STejun Heo goto enomem; 378213e2e556STejun Heo 37834c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 378413e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 378513e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 378613e2e556STejun Heo 37874c16bd32STejun Heo /* 37884c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 37894c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 37904c16bd32STejun Heo * pools. 37914c16bd32STejun Heo */ 37924c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 37939e8cd2f5STejun Heo 37944c16bd32STejun Heo /* 37954c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 37964c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 37974c16bd32STejun Heo * pwqs accordingly. 37984c16bd32STejun Heo */ 37994c16bd32STejun Heo get_online_cpus(); 38004c16bd32STejun Heo 38014c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 38024c16bd32STejun Heo 38034c16bd32STejun Heo /* 38044c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 38054c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 38064c16bd32STejun Heo * it even if we don't use it immediately. 38074c16bd32STejun Heo */ 38084c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 38094c16bd32STejun Heo if (!dfl_pwq) 38104c16bd32STejun Heo goto enomem_pwq; 38114c16bd32STejun Heo 38124c16bd32STejun Heo for_each_node(node) { 38134c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 38144c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 38154c16bd32STejun Heo if (!pwq_tbl[node]) 38164c16bd32STejun Heo goto enomem_pwq; 38174c16bd32STejun Heo } else { 38184c16bd32STejun Heo dfl_pwq->refcnt++; 38194c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 38204c16bd32STejun Heo } 38214c16bd32STejun Heo } 38224c16bd32STejun Heo 38234c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 38244c16bd32STejun Heo 38254c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3826f147f29eSTejun Heo mutex_lock(&wq->mutex); 3827a892caccSTejun Heo 3828f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 38294c16bd32STejun Heo 38304c16bd32STejun Heo /* save the previous pwq and install the new one */ 3831f147f29eSTejun Heo for_each_node(node) 38324c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 38334c16bd32STejun Heo 38344c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 38354c16bd32STejun Heo link_pwq(dfl_pwq); 38364c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3837f147f29eSTejun Heo 3838f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3839f147f29eSTejun Heo 38404c16bd32STejun Heo /* put the old pwqs */ 38414c16bd32STejun Heo for_each_node(node) 38424c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 38434c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 38444c16bd32STejun Heo 38454c16bd32STejun Heo put_online_cpus(); 38464862125bSTejun Heo ret = 0; 38474862125bSTejun Heo /* fall through */ 38484862125bSTejun Heo out_free: 38494c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 38504862125bSTejun Heo free_workqueue_attrs(new_attrs); 38514c16bd32STejun Heo kfree(pwq_tbl); 38524862125bSTejun Heo return ret; 385313e2e556STejun Heo 38544c16bd32STejun Heo enomem_pwq: 38554c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 38564c16bd32STejun Heo for_each_node(node) 38574c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 38584c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 38594c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 38604c16bd32STejun Heo put_online_cpus(); 386113e2e556STejun Heo enomem: 38624862125bSTejun Heo ret = -ENOMEM; 38634862125bSTejun Heo goto out_free; 38649e8cd2f5STejun Heo } 38659e8cd2f5STejun Heo 38664c16bd32STejun Heo /** 38674c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 38684c16bd32STejun Heo * @wq: the target workqueue 38694c16bd32STejun Heo * @cpu: the CPU coming up or going down 38704c16bd32STejun Heo * @online: whether @cpu is coming up or going down 38714c16bd32STejun Heo * 38724c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 38734c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 38744c16bd32STejun Heo * @wq accordingly. 38754c16bd32STejun Heo * 38764c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 38774c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 38784c16bd32STejun Heo * correct. 38794c16bd32STejun Heo * 38804c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 38814c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 38824c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 38834c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 38844c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 38854c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 38864c16bd32STejun Heo * CPU_DOWN_PREPARE. 38874c16bd32STejun Heo */ 38884c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 38894c16bd32STejun Heo bool online) 38904c16bd32STejun Heo { 38914c16bd32STejun Heo int node = cpu_to_node(cpu); 38924c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 38934c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 38944c16bd32STejun Heo struct workqueue_attrs *target_attrs; 38954c16bd32STejun Heo cpumask_t *cpumask; 38964c16bd32STejun Heo 38974c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38984c16bd32STejun Heo 38994c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 39004c16bd32STejun Heo return; 39014c16bd32STejun Heo 39024c16bd32STejun Heo /* 39034c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39044c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39054c16bd32STejun Heo * CPU hotplug exclusion. 39064c16bd32STejun Heo */ 39074c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39084c16bd32STejun Heo cpumask = target_attrs->cpumask; 39094c16bd32STejun Heo 39104c16bd32STejun Heo mutex_lock(&wq->mutex); 3911d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 3912d55262c4STejun Heo goto out_unlock; 39134c16bd32STejun Heo 39144c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39154c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39164c16bd32STejun Heo 39174c16bd32STejun Heo /* 39184c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 39194c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 39204c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 3921534a3fbbSDaeseok Youn * wq's, the default pwq should be used. 39224c16bd32STejun Heo */ 39234c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 39244c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 39254c16bd32STejun Heo goto out_unlock; 39264c16bd32STejun Heo } else { 39274c16bd32STejun Heo goto use_dfl_pwq; 39284c16bd32STejun Heo } 39294c16bd32STejun Heo 39304c16bd32STejun Heo mutex_unlock(&wq->mutex); 39314c16bd32STejun Heo 39324c16bd32STejun Heo /* create a new pwq */ 39334c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 39344c16bd32STejun Heo if (!pwq) { 39352d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 39364c16bd32STejun Heo wq->name); 393777f300b1SDaeseok Youn mutex_lock(&wq->mutex); 393877f300b1SDaeseok Youn goto use_dfl_pwq; 39394c16bd32STejun Heo } 39404c16bd32STejun Heo 39414c16bd32STejun Heo /* 39424c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 39434c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 39444c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 39454c16bd32STejun Heo * inbetween. 39464c16bd32STejun Heo */ 39474c16bd32STejun Heo mutex_lock(&wq->mutex); 39484c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 39494c16bd32STejun Heo goto out_unlock; 39504c16bd32STejun Heo 39514c16bd32STejun Heo use_dfl_pwq: 39524c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 39534c16bd32STejun Heo get_pwq(wq->dfl_pwq); 39544c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 39554c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 39564c16bd32STejun Heo out_unlock: 39574c16bd32STejun Heo mutex_unlock(&wq->mutex); 39584c16bd32STejun Heo put_pwq_unlocked(old_pwq); 39594c16bd32STejun Heo } 39604c16bd32STejun Heo 396130cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 39621da177e4SLinus Torvalds { 396349e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 39648a2b7538STejun Heo int cpu, ret; 3965e1d8aa9fSFrederic Weisbecker 396630cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3967420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3968420c0ddbSTejun Heo if (!wq->cpu_pwqs) 396930cdf249STejun Heo return -ENOMEM; 397030cdf249STejun Heo 397130cdf249STejun Heo for_each_possible_cpu(cpu) { 39727fb98ea7STejun Heo struct pool_workqueue *pwq = 39737fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 39747a62c2c8STejun Heo struct worker_pool *cpu_pools = 3975f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 397630cdf249STejun Heo 3977f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3978f147f29eSTejun Heo 3979f147f29eSTejun Heo mutex_lock(&wq->mutex); 39801befcf30STejun Heo link_pwq(pwq); 3981f147f29eSTejun Heo mutex_unlock(&wq->mutex); 398230cdf249STejun Heo } 398330cdf249STejun Heo return 0; 39848a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 39858a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 39868a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 39878a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 39888a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 39898a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 39908a2b7538STejun Heo return ret; 39919e8cd2f5STejun Heo } else { 39929e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 39939e8cd2f5STejun Heo } 39940f900049STejun Heo } 39950f900049STejun Heo 3996f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3997f3421797STejun Heo const char *name) 3998b71ab8c2STejun Heo { 3999f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4000f3421797STejun Heo 4001f3421797STejun Heo if (max_active < 1 || max_active > lim) 4002044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4003f3421797STejun Heo max_active, name, 1, lim); 4004b71ab8c2STejun Heo 4005f3421797STejun Heo return clamp_val(max_active, 1, lim); 4006b71ab8c2STejun Heo } 4007b71ab8c2STejun Heo 4008b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 400997e37d7bSTejun Heo unsigned int flags, 40101e19ffc6STejun Heo int max_active, 4011eb13ba87SJohannes Berg struct lock_class_key *key, 4012b196be89STejun Heo const char *lock_name, ...) 40133af24433SOleg Nesterov { 4014df2d5ae4STejun Heo size_t tbl_size = 0; 4015ecf6881fSTejun Heo va_list args; 40163af24433SOleg Nesterov struct workqueue_struct *wq; 401749e3cf44STejun Heo struct pool_workqueue *pwq; 4018b196be89STejun Heo 4019cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4020cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4021cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4022cee22a15SViresh Kumar 4023ecf6881fSTejun Heo /* allocate wq and format name */ 4024df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4025ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4026df2d5ae4STejun Heo 4027df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4028b196be89STejun Heo if (!wq) 4029d2c1d404STejun Heo return NULL; 4030b196be89STejun Heo 40316029a918STejun Heo if (flags & WQ_UNBOUND) { 40326029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40336029a918STejun Heo if (!wq->unbound_attrs) 40346029a918STejun Heo goto err_free_wq; 40356029a918STejun Heo } 40366029a918STejun Heo 4037ecf6881fSTejun Heo va_start(args, lock_name); 4038ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4039b196be89STejun Heo va_end(args); 40403af24433SOleg Nesterov 4041d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4042b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40433af24433SOleg Nesterov 4044b196be89STejun Heo /* init wq */ 404597e37d7bSTejun Heo wq->flags = flags; 4046a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 40473c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4048112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 404930cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 405073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 405173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4052493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 40533af24433SOleg Nesterov 4054eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4055cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 40563af24433SOleg Nesterov 405730cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4058d2c1d404STejun Heo goto err_free_wq; 40591537663fSTejun Heo 4060493008a8STejun Heo /* 4061493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4062493008a8STejun Heo * have a rescuer to guarantee forward progress. 4063493008a8STejun Heo */ 4064493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4065e22bee78STejun Heo struct worker *rescuer; 4066e22bee78STejun Heo 4067f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 4068e22bee78STejun Heo if (!rescuer) 4069d2c1d404STejun Heo goto err_destroy; 4070e22bee78STejun Heo 4071111c225aSTejun Heo rescuer->rescue_wq = wq; 4072111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4073b196be89STejun Heo wq->name); 4074d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4075d2c1d404STejun Heo kfree(rescuer); 4076d2c1d404STejun Heo goto err_destroy; 4077d2c1d404STejun Heo } 4078e22bee78STejun Heo 4079d2c1d404STejun Heo wq->rescuer = rescuer; 408014a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4081e22bee78STejun Heo wake_up_process(rescuer->task); 40823af24433SOleg Nesterov } 40831537663fSTejun Heo 4084226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4085226223abSTejun Heo goto err_destroy; 4086226223abSTejun Heo 40873af24433SOleg Nesterov /* 408868e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 408968e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 409068e13a67SLai Jiangshan * list. 40913af24433SOleg Nesterov */ 409268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4093a0a1a5fdSTejun Heo 4094a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 409549e3cf44STejun Heo for_each_pwq(pwq, wq) 4096699ce097STejun Heo pwq_adjust_max_active(pwq); 4097a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4098a0a1a5fdSTejun Heo 40993af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4100a0a1a5fdSTejun Heo 410168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41023af24433SOleg Nesterov 41033af24433SOleg Nesterov return wq; 4104d2c1d404STejun Heo 4105d2c1d404STejun Heo err_free_wq: 41066029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41074690c4abSTejun Heo kfree(wq); 4108d2c1d404STejun Heo return NULL; 4109d2c1d404STejun Heo err_destroy: 4110d2c1d404STejun Heo destroy_workqueue(wq); 41114690c4abSTejun Heo return NULL; 41121da177e4SLinus Torvalds } 4113d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41141da177e4SLinus Torvalds 41153af24433SOleg Nesterov /** 41163af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41173af24433SOleg Nesterov * @wq: target workqueue 41183af24433SOleg Nesterov * 41193af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41203af24433SOleg Nesterov */ 41213af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41223af24433SOleg Nesterov { 412349e3cf44STejun Heo struct pool_workqueue *pwq; 41244c16bd32STejun Heo int node; 41253af24433SOleg Nesterov 41269c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41279c5a2ba7STejun Heo drain_workqueue(wq); 4128c8efcc25STejun Heo 41296183c009STejun Heo /* sanity checks */ 4130b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 413149e3cf44STejun Heo for_each_pwq(pwq, wq) { 41326183c009STejun Heo int i; 41336183c009STejun Heo 413476af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 413576af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4136b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41376183c009STejun Heo return; 413876af4d93STejun Heo } 413976af4d93STejun Heo } 414076af4d93STejun Heo 41415c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 41428864b4e5STejun Heo WARN_ON(pwq->nr_active) || 414376af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4144b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41456183c009STejun Heo return; 41466183c009STejun Heo } 414776af4d93STejun Heo } 4148b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41496183c009STejun Heo 4150a0a1a5fdSTejun Heo /* 4151a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4152a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4153a0a1a5fdSTejun Heo */ 415468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4155d2c1d404STejun Heo list_del_init(&wq->list); 415668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41573af24433SOleg Nesterov 4158226223abSTejun Heo workqueue_sysfs_unregister(wq); 4159226223abSTejun Heo 4160493008a8STejun Heo if (wq->rescuer) { 4161e22bee78STejun Heo kthread_stop(wq->rescuer->task); 41628d9df9f0SXiaotian Feng kfree(wq->rescuer); 4163493008a8STejun Heo wq->rescuer = NULL; 4164e22bee78STejun Heo } 4165e22bee78STejun Heo 41668864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 416729c91e99STejun Heo /* 41688864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 41698864b4e5STejun Heo * free the pwqs and wq. 417029c91e99STejun Heo */ 41718864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 41728864b4e5STejun Heo kfree(wq); 41738864b4e5STejun Heo } else { 41748864b4e5STejun Heo /* 41758864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 41764c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 41774c16bd32STejun Heo * @wq will be freed when the last pwq is released. 41788864b4e5STejun Heo */ 41794c16bd32STejun Heo for_each_node(node) { 41804c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 41814c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 41824c16bd32STejun Heo put_pwq_unlocked(pwq); 41834c16bd32STejun Heo } 41844c16bd32STejun Heo 41854c16bd32STejun Heo /* 41864c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 41874c16bd32STejun Heo * put. Don't access it afterwards. 41884c16bd32STejun Heo */ 41894c16bd32STejun Heo pwq = wq->dfl_pwq; 41904c16bd32STejun Heo wq->dfl_pwq = NULL; 4191dce90d47STejun Heo put_pwq_unlocked(pwq); 419229c91e99STejun Heo } 41933af24433SOleg Nesterov } 41943af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 41953af24433SOleg Nesterov 4196dcd989cbSTejun Heo /** 4197dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4198dcd989cbSTejun Heo * @wq: target workqueue 4199dcd989cbSTejun Heo * @max_active: new max_active value. 4200dcd989cbSTejun Heo * 4201dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4202dcd989cbSTejun Heo * 4203dcd989cbSTejun Heo * CONTEXT: 4204dcd989cbSTejun Heo * Don't call from IRQ context. 4205dcd989cbSTejun Heo */ 4206dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4207dcd989cbSTejun Heo { 420849e3cf44STejun Heo struct pool_workqueue *pwq; 4209dcd989cbSTejun Heo 42108719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42118719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 42128719dceaSTejun Heo return; 42138719dceaSTejun Heo 4214f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4215dcd989cbSTejun Heo 4216a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4217dcd989cbSTejun Heo 4218dcd989cbSTejun Heo wq->saved_max_active = max_active; 4219dcd989cbSTejun Heo 4220699ce097STejun Heo for_each_pwq(pwq, wq) 4221699ce097STejun Heo pwq_adjust_max_active(pwq); 4222dcd989cbSTejun Heo 4223a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4224dcd989cbSTejun Heo } 4225dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4226dcd989cbSTejun Heo 4227dcd989cbSTejun Heo /** 4228e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4229e6267616STejun Heo * 4230e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4231e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4232d185af30SYacine Belkadi * 4233d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4234e6267616STejun Heo */ 4235e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4236e6267616STejun Heo { 4237e6267616STejun Heo struct worker *worker = current_wq_worker(); 4238e6267616STejun Heo 42396a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4240e6267616STejun Heo } 4241e6267616STejun Heo 4242e6267616STejun Heo /** 4243dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4244dcd989cbSTejun Heo * @cpu: CPU in question 4245dcd989cbSTejun Heo * @wq: target workqueue 4246dcd989cbSTejun Heo * 4247dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4248dcd989cbSTejun Heo * no synchronization around this function and the test result is 4249dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4250dcd989cbSTejun Heo * 4251d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4252d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4253d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4254d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4255d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4256d3251859STejun Heo * 4257d185af30SYacine Belkadi * Return: 4258dcd989cbSTejun Heo * %true if congested, %false otherwise. 4259dcd989cbSTejun Heo */ 4260d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4261dcd989cbSTejun Heo { 42627fb98ea7STejun Heo struct pool_workqueue *pwq; 426376af4d93STejun Heo bool ret; 426476af4d93STejun Heo 426588109453SLai Jiangshan rcu_read_lock_sched(); 42667fb98ea7STejun Heo 4267d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4268d3251859STejun Heo cpu = smp_processor_id(); 4269d3251859STejun Heo 42707fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 42717fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 42727fb98ea7STejun Heo else 4273df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4274dcd989cbSTejun Heo 427576af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 427688109453SLai Jiangshan rcu_read_unlock_sched(); 427776af4d93STejun Heo 427876af4d93STejun Heo return ret; 4279dcd989cbSTejun Heo } 4280dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4281dcd989cbSTejun Heo 4282dcd989cbSTejun Heo /** 4283dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4284dcd989cbSTejun Heo * @work: the work to be tested 4285dcd989cbSTejun Heo * 4286dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4287dcd989cbSTejun Heo * synchronization around this function and the test result is 4288dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4289dcd989cbSTejun Heo * 4290d185af30SYacine Belkadi * Return: 4291dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4292dcd989cbSTejun Heo */ 4293dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4294dcd989cbSTejun Heo { 4295fa1b54e6STejun Heo struct worker_pool *pool; 4296dcd989cbSTejun Heo unsigned long flags; 4297dcd989cbSTejun Heo unsigned int ret = 0; 4298dcd989cbSTejun Heo 4299dcd989cbSTejun Heo if (work_pending(work)) 4300dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4301038366c5SLai Jiangshan 4302fa1b54e6STejun Heo local_irq_save(flags); 4303fa1b54e6STejun Heo pool = get_work_pool(work); 4304038366c5SLai Jiangshan if (pool) { 4305fa1b54e6STejun Heo spin_lock(&pool->lock); 4306c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4307dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4308fa1b54e6STejun Heo spin_unlock(&pool->lock); 4309038366c5SLai Jiangshan } 4310fa1b54e6STejun Heo local_irq_restore(flags); 4311dcd989cbSTejun Heo 4312dcd989cbSTejun Heo return ret; 4313dcd989cbSTejun Heo } 4314dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4315dcd989cbSTejun Heo 43163d1cb205STejun Heo /** 43173d1cb205STejun Heo * set_worker_desc - set description for the current work item 43183d1cb205STejun Heo * @fmt: printf-style format string 43193d1cb205STejun Heo * @...: arguments for the format string 43203d1cb205STejun Heo * 43213d1cb205STejun Heo * This function can be called by a running work function to describe what 43223d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43233d1cb205STejun Heo * information will be printed out together to help debugging. The 43243d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43253d1cb205STejun Heo */ 43263d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43273d1cb205STejun Heo { 43283d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43293d1cb205STejun Heo va_list args; 43303d1cb205STejun Heo 43313d1cb205STejun Heo if (worker) { 43323d1cb205STejun Heo va_start(args, fmt); 43333d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 43343d1cb205STejun Heo va_end(args); 43353d1cb205STejun Heo worker->desc_valid = true; 43363d1cb205STejun Heo } 43373d1cb205STejun Heo } 43383d1cb205STejun Heo 43393d1cb205STejun Heo /** 43403d1cb205STejun Heo * print_worker_info - print out worker information and description 43413d1cb205STejun Heo * @log_lvl: the log level to use when printing 43423d1cb205STejun Heo * @task: target task 43433d1cb205STejun Heo * 43443d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 43453d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 43463d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 43473d1cb205STejun Heo * 43483d1cb205STejun Heo * This function can be safely called on any task as long as the 43493d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 43503d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 43513d1cb205STejun Heo */ 43523d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 43533d1cb205STejun Heo { 43543d1cb205STejun Heo work_func_t *fn = NULL; 43553d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 43563d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 43573d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 43583d1cb205STejun Heo struct workqueue_struct *wq = NULL; 43593d1cb205STejun Heo bool desc_valid = false; 43603d1cb205STejun Heo struct worker *worker; 43613d1cb205STejun Heo 43623d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 43633d1cb205STejun Heo return; 43643d1cb205STejun Heo 43653d1cb205STejun Heo /* 43663d1cb205STejun Heo * This function is called without any synchronization and @task 43673d1cb205STejun Heo * could be in any state. Be careful with dereferences. 43683d1cb205STejun Heo */ 43693d1cb205STejun Heo worker = probe_kthread_data(task); 43703d1cb205STejun Heo 43713d1cb205STejun Heo /* 43723d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 43733d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 43743d1cb205STejun Heo */ 43753d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 43763d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 43773d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 43783d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 43793d1cb205STejun Heo 43803d1cb205STejun Heo /* copy worker description */ 43813d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 43823d1cb205STejun Heo if (desc_valid) 43833d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 43843d1cb205STejun Heo 43853d1cb205STejun Heo if (fn || name[0] || desc[0]) { 43863d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 43873d1cb205STejun Heo if (desc[0]) 43883d1cb205STejun Heo pr_cont(" (%s)", desc); 43893d1cb205STejun Heo pr_cont("\n"); 43903d1cb205STejun Heo } 43913d1cb205STejun Heo } 43923d1cb205STejun Heo 4393db7bccf4STejun Heo /* 4394db7bccf4STejun Heo * CPU hotplug. 4395db7bccf4STejun Heo * 4396e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4397112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4398706026c2STejun Heo * pool which make migrating pending and scheduled works very 4399e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 440094cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4401e22bee78STejun Heo * blocked draining impractical. 4402e22bee78STejun Heo * 440324647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4404628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4405628c78e7STejun Heo * cpu comes back online. 4406db7bccf4STejun Heo */ 4407db7bccf4STejun Heo 4408706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4409db7bccf4STejun Heo { 441038db41d9STejun Heo int cpu = smp_processor_id(); 44114ce62e9eSTejun Heo struct worker_pool *pool; 4412db7bccf4STejun Heo struct worker *worker; 4413db7bccf4STejun Heo 4414f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 441592f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 441694cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4417e22bee78STejun Heo 4418f2d5a0eeSTejun Heo /* 441992f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 442094cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 442194cf58bbSTejun Heo * except for the ones which are still executing works from 442294cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 442394cf58bbSTejun Heo * this, they may become diasporas. 4424f2d5a0eeSTejun Heo */ 4425da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4426403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4427db7bccf4STejun Heo 442824647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4429f2d5a0eeSTejun Heo 443094cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 443192f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4432e22bee78STejun Heo 4433e22bee78STejun Heo /* 4434eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4435eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4436eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4437eb283428SLai Jiangshan * from other cpus. 4438628c78e7STejun Heo */ 4439628c78e7STejun Heo schedule(); 4440628c78e7STejun Heo 4441628c78e7STejun Heo /* 4442eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4443eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4444eb283428SLai Jiangshan * and keep_working() are always true as long as the 4445eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4446eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4447eb283428SLai Jiangshan * are served by workers tied to the pool. 4448e22bee78STejun Heo */ 4449e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4450eb283428SLai Jiangshan 4451eb283428SLai Jiangshan /* 4452eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4453eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4454eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4455eb283428SLai Jiangshan */ 4456eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4457eb283428SLai Jiangshan wake_up_worker(pool); 4458eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4459eb283428SLai Jiangshan } 4460db7bccf4STejun Heo } 4461db7bccf4STejun Heo 4462bd7c089eSTejun Heo /** 4463bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4464bd7c089eSTejun Heo * @pool: pool of interest 4465bd7c089eSTejun Heo * 4466a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4467bd7c089eSTejun Heo */ 4468bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4469bd7c089eSTejun Heo { 4470a9ab775bSTejun Heo struct worker *worker; 4471bd7c089eSTejun Heo 447292f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4473bd7c089eSTejun Heo 4474bd7c089eSTejun Heo /* 4475a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4476a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4477a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4478a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4479a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4480bd7c089eSTejun Heo */ 4481da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4482a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4483a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4484a9ab775bSTejun Heo 4485a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 44863de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4487a9ab775bSTejun Heo 4488da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4489a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4490a9ab775bSTejun Heo 4491a9ab775bSTejun Heo /* 4492a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4493a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4494a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4495a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4496a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4497a9ab775bSTejun Heo * be bound before @pool->lock is released. 4498a9ab775bSTejun Heo */ 4499a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4500bd7c089eSTejun Heo wake_up_process(worker->task); 4501bd7c089eSTejun Heo 4502bd7c089eSTejun Heo /* 4503a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4504a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4505a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4506a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4507a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4508a9ab775bSTejun Heo * concurrency management. Note that when or whether 4509a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4510a9ab775bSTejun Heo * 4511a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4512a9ab775bSTejun Heo * tested without holding any lock in 4513a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4514a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4515a9ab775bSTejun Heo * management operations. 4516bd7c089eSTejun Heo */ 4517a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4518a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4519a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4520a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4521bd7c089eSTejun Heo } 4522a9ab775bSTejun Heo 4523a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4524bd7c089eSTejun Heo } 4525bd7c089eSTejun Heo 45267dbc725eSTejun Heo /** 45277dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 45287dbc725eSTejun Heo * @pool: unbound pool of interest 45297dbc725eSTejun Heo * @cpu: the CPU which is coming up 45307dbc725eSTejun Heo * 45317dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 45327dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 45337dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 45347dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 45357dbc725eSTejun Heo */ 45367dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 45377dbc725eSTejun Heo { 45387dbc725eSTejun Heo static cpumask_t cpumask; 45397dbc725eSTejun Heo struct worker *worker; 45407dbc725eSTejun Heo 454192f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 45427dbc725eSTejun Heo 45437dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 45447dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 45457dbc725eSTejun Heo return; 45467dbc725eSTejun Heo 45477dbc725eSTejun Heo /* is @cpu the only online CPU? */ 45487dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 45497dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 45507dbc725eSTejun Heo return; 45517dbc725eSTejun Heo 45527dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4553da028469SLai Jiangshan for_each_pool_worker(worker, pool) 45547dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 45557dbc725eSTejun Heo pool->attrs->cpumask) < 0); 45567dbc725eSTejun Heo } 45577dbc725eSTejun Heo 45588db25e78STejun Heo /* 45598db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 45608db25e78STejun Heo * This will be registered high priority CPU notifier. 45618db25e78STejun Heo */ 45620db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 45631da177e4SLinus Torvalds unsigned long action, 45641da177e4SLinus Torvalds void *hcpu) 45651da177e4SLinus Torvalds { 4566d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45674ce62e9eSTejun Heo struct worker_pool *pool; 45684c16bd32STejun Heo struct workqueue_struct *wq; 45697dbc725eSTejun Heo int pi; 45701da177e4SLinus Torvalds 45718db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 45723af24433SOleg Nesterov case CPU_UP_PREPARE: 4573f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45743ce63377STejun Heo if (pool->nr_workers) 45753ce63377STejun Heo continue; 4576051e1850SLai Jiangshan if (!create_worker(pool)) 45773ce63377STejun Heo return NOTIFY_BAD; 45783af24433SOleg Nesterov } 45791da177e4SLinus Torvalds break; 45801da177e4SLinus Torvalds 458165758202STejun Heo case CPU_DOWN_FAILED: 458265758202STejun Heo case CPU_ONLINE: 458368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45847dbc725eSTejun Heo 45857dbc725eSTejun Heo for_each_pool(pool, pi) { 458692f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 458794cf58bbSTejun Heo 4588f05b558dSLai Jiangshan if (pool->cpu == cpu) 458994cf58bbSTejun Heo rebind_workers(pool); 4590f05b558dSLai Jiangshan else if (pool->cpu < 0) 45917dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 459294cf58bbSTejun Heo 459392f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 459494cf58bbSTejun Heo } 45957dbc725eSTejun Heo 45964c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45974c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45984c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 45994c16bd32STejun Heo 460068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46018db25e78STejun Heo break; 460265758202STejun Heo } 460365758202STejun Heo return NOTIFY_OK; 460465758202STejun Heo } 460565758202STejun Heo 460665758202STejun Heo /* 460765758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 460865758202STejun Heo * This will be registered as low priority CPU notifier. 460965758202STejun Heo */ 46100db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 461165758202STejun Heo unsigned long action, 461265758202STejun Heo void *hcpu) 461365758202STejun Heo { 4614d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46158db25e78STejun Heo struct work_struct unbind_work; 46164c16bd32STejun Heo struct workqueue_struct *wq; 46178db25e78STejun Heo 461865758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 461965758202STejun Heo case CPU_DOWN_PREPARE: 46204c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4621706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46227635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46234c16bd32STejun Heo 46244c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46254c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46264c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46274c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 46284c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 46294c16bd32STejun Heo 46304c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 46318db25e78STejun Heo flush_work(&unbind_work); 4632440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 46338db25e78STejun Heo break; 463465758202STejun Heo } 463565758202STejun Heo return NOTIFY_OK; 463665758202STejun Heo } 463765758202STejun Heo 46382d3854a3SRusty Russell #ifdef CONFIG_SMP 46398ccad40dSRusty Russell 46402d3854a3SRusty Russell struct work_for_cpu { 4641ed48ece2STejun Heo struct work_struct work; 46422d3854a3SRusty Russell long (*fn)(void *); 46432d3854a3SRusty Russell void *arg; 46442d3854a3SRusty Russell long ret; 46452d3854a3SRusty Russell }; 46462d3854a3SRusty Russell 4647ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 46482d3854a3SRusty Russell { 4649ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4650ed48ece2STejun Heo 46512d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 46522d3854a3SRusty Russell } 46532d3854a3SRusty Russell 46542d3854a3SRusty Russell /** 46552d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 46562d3854a3SRusty Russell * @cpu: the cpu to run on 46572d3854a3SRusty Russell * @fn: the function to run 46582d3854a3SRusty Russell * @arg: the function arg 46592d3854a3SRusty Russell * 466031ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 46616b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4662d185af30SYacine Belkadi * 4663d185af30SYacine Belkadi * Return: The value @fn returns. 46642d3854a3SRusty Russell */ 4665d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 46662d3854a3SRusty Russell { 4667ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 46682d3854a3SRusty Russell 4669ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4670ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 467112997d1aSBjorn Helgaas flush_work(&wfc.work); 4672440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 46732d3854a3SRusty Russell return wfc.ret; 46742d3854a3SRusty Russell } 46752d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 46762d3854a3SRusty Russell #endif /* CONFIG_SMP */ 46772d3854a3SRusty Russell 4678a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4679e7577c50SRusty Russell 4680a0a1a5fdSTejun Heo /** 4681a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4682a0a1a5fdSTejun Heo * 468358a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4684c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4685706026c2STejun Heo * pool->worklist. 4686a0a1a5fdSTejun Heo * 4687a0a1a5fdSTejun Heo * CONTEXT: 4688a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4689a0a1a5fdSTejun Heo */ 4690a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4691a0a1a5fdSTejun Heo { 469224b8a847STejun Heo struct workqueue_struct *wq; 469324b8a847STejun Heo struct pool_workqueue *pwq; 4694a0a1a5fdSTejun Heo 469568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4696a0a1a5fdSTejun Heo 46976183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4698a0a1a5fdSTejun Heo workqueue_freezing = true; 4699a0a1a5fdSTejun Heo 470024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4701a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4702699ce097STejun Heo for_each_pwq(pwq, wq) 4703699ce097STejun Heo pwq_adjust_max_active(pwq); 4704a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4705a1056305STejun Heo } 47065bcab335STejun Heo 470768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4708a0a1a5fdSTejun Heo } 4709a0a1a5fdSTejun Heo 4710a0a1a5fdSTejun Heo /** 471158a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4712a0a1a5fdSTejun Heo * 4713a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4714a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4715a0a1a5fdSTejun Heo * 4716a0a1a5fdSTejun Heo * CONTEXT: 471768e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4718a0a1a5fdSTejun Heo * 4719d185af30SYacine Belkadi * Return: 472058a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 472158a69cb4STejun Heo * is complete. 4722a0a1a5fdSTejun Heo */ 4723a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4724a0a1a5fdSTejun Heo { 4725a0a1a5fdSTejun Heo bool busy = false; 472624b8a847STejun Heo struct workqueue_struct *wq; 472724b8a847STejun Heo struct pool_workqueue *pwq; 4728a0a1a5fdSTejun Heo 472968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4730a0a1a5fdSTejun Heo 47316183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4732a0a1a5fdSTejun Heo 473324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 473424b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 473524b8a847STejun Heo continue; 4736a0a1a5fdSTejun Heo /* 4737a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4738a0a1a5fdSTejun Heo * to peek without lock. 4739a0a1a5fdSTejun Heo */ 474088109453SLai Jiangshan rcu_read_lock_sched(); 474124b8a847STejun Heo for_each_pwq(pwq, wq) { 47426183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4743112202d9STejun Heo if (pwq->nr_active) { 4744a0a1a5fdSTejun Heo busy = true; 474588109453SLai Jiangshan rcu_read_unlock_sched(); 4746a0a1a5fdSTejun Heo goto out_unlock; 4747a0a1a5fdSTejun Heo } 4748a0a1a5fdSTejun Heo } 474988109453SLai Jiangshan rcu_read_unlock_sched(); 4750a0a1a5fdSTejun Heo } 4751a0a1a5fdSTejun Heo out_unlock: 475268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4753a0a1a5fdSTejun Heo return busy; 4754a0a1a5fdSTejun Heo } 4755a0a1a5fdSTejun Heo 4756a0a1a5fdSTejun Heo /** 4757a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4758a0a1a5fdSTejun Heo * 4759a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4760706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4761a0a1a5fdSTejun Heo * 4762a0a1a5fdSTejun Heo * CONTEXT: 4763a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4764a0a1a5fdSTejun Heo */ 4765a0a1a5fdSTejun Heo void thaw_workqueues(void) 4766a0a1a5fdSTejun Heo { 476724b8a847STejun Heo struct workqueue_struct *wq; 476824b8a847STejun Heo struct pool_workqueue *pwq; 4769a0a1a5fdSTejun Heo 477068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4771a0a1a5fdSTejun Heo 4772a0a1a5fdSTejun Heo if (!workqueue_freezing) 4773a0a1a5fdSTejun Heo goto out_unlock; 4774a0a1a5fdSTejun Heo 477574b414eaSLai Jiangshan workqueue_freezing = false; 477624b8a847STejun Heo 477724b8a847STejun Heo /* restore max_active and repopulate worklist */ 477824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4779a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4780699ce097STejun Heo for_each_pwq(pwq, wq) 4781699ce097STejun Heo pwq_adjust_max_active(pwq); 4782a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 478324b8a847STejun Heo } 478424b8a847STejun Heo 4785a0a1a5fdSTejun Heo out_unlock: 478668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4787a0a1a5fdSTejun Heo } 4788a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4789a0a1a5fdSTejun Heo 4790bce90380STejun Heo static void __init wq_numa_init(void) 4791bce90380STejun Heo { 4792bce90380STejun Heo cpumask_var_t *tbl; 4793bce90380STejun Heo int node, cpu; 4794bce90380STejun Heo 4795bce90380STejun Heo if (num_possible_nodes() <= 1) 4796bce90380STejun Heo return; 4797bce90380STejun Heo 4798d55262c4STejun Heo if (wq_disable_numa) { 4799d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 4800d55262c4STejun Heo return; 4801d55262c4STejun Heo } 4802d55262c4STejun Heo 48034c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 48044c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 48054c16bd32STejun Heo 4806bce90380STejun Heo /* 4807bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4808bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4809bce90380STejun Heo * fully initialized by now. 4810bce90380STejun Heo */ 4811ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 4812bce90380STejun Heo BUG_ON(!tbl); 4813bce90380STejun Heo 4814bce90380STejun Heo for_each_node(node) 48155a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 48161be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 4817bce90380STejun Heo 4818bce90380STejun Heo for_each_possible_cpu(cpu) { 4819bce90380STejun Heo node = cpu_to_node(cpu); 4820bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4821bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4822bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4823bce90380STejun Heo return; 4824bce90380STejun Heo } 4825bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4826bce90380STejun Heo } 4827bce90380STejun Heo 4828bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4829bce90380STejun Heo wq_numa_enabled = true; 4830bce90380STejun Heo } 4831bce90380STejun Heo 48326ee0578bSSuresh Siddha static int __init init_workqueues(void) 48331da177e4SLinus Torvalds { 48347a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 48357a4e344cSTejun Heo int i, cpu; 4836c34056a3STejun Heo 4837e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4838e904e6c2STejun Heo 4839e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4840e904e6c2STejun Heo 484165758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4842a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 48438b03ae3cSTejun Heo 4844bce90380STejun Heo wq_numa_init(); 4845bce90380STejun Heo 4846706026c2STejun Heo /* initialize CPU pools */ 484729c91e99STejun Heo for_each_possible_cpu(cpu) { 48484ce62e9eSTejun Heo struct worker_pool *pool; 48498b03ae3cSTejun Heo 48507a4e344cSTejun Heo i = 0; 4851f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 48527a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4853ec22ca5eSTejun Heo pool->cpu = cpu; 48547a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 48557a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4856f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 48577a4e344cSTejun Heo 48589daf9e67STejun Heo /* alloc pool ID */ 485968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 48609daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 486168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 48624ce62e9eSTejun Heo } 48638b03ae3cSTejun Heo } 48648b03ae3cSTejun Heo 4865e22bee78STejun Heo /* create the initial worker */ 486629c91e99STejun Heo for_each_online_cpu(cpu) { 48674ce62e9eSTejun Heo struct worker_pool *pool; 4868e22bee78STejun Heo 4869f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 487024647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4871051e1850SLai Jiangshan BUG_ON(!create_worker(pool)); 4872e22bee78STejun Heo } 48734ce62e9eSTejun Heo } 4874e22bee78STejun Heo 48758a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 487629c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 487729c91e99STejun Heo struct workqueue_attrs *attrs; 487829c91e99STejun Heo 487929c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 488029c91e99STejun Heo attrs->nice = std_nice[i]; 488129c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 48828a2b7538STejun Heo 48838a2b7538STejun Heo /* 48848a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 48858a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 48868a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 48878a2b7538STejun Heo */ 48888a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 48898a2b7538STejun Heo attrs->nice = std_nice[i]; 48908a2b7538STejun Heo attrs->no_numa = true; 48918a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 489229c91e99STejun Heo } 489329c91e99STejun Heo 4894d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 48951aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4896d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4897f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4898f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 489924d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 490024d51addSTejun Heo WQ_FREEZABLE, 0); 49010668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 49020668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 49030668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 49040668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 49050668106cSViresh Kumar 0); 49061aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 49070668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 49080668106cSViresh Kumar !system_power_efficient_wq || 49090668106cSViresh Kumar !system_freezable_power_efficient_wq); 49106ee0578bSSuresh Siddha return 0; 49111da177e4SLinus Torvalds } 49126ee0578bSSuresh Siddha early_initcall(init_workqueues); 4913