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 int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 269bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 270bce90380STejun Heo /* possible CPUs of each node */ 271bce90380STejun Heo 272d55262c4STejun Heo static bool wq_disable_numa; 273d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 274d55262c4STejun Heo 275cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 276cee22a15SViresh Kumar #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT 277cee22a15SViresh Kumar static bool wq_power_efficient = true; 278cee22a15SViresh Kumar #else 279cee22a15SViresh Kumar static bool wq_power_efficient; 280cee22a15SViresh Kumar #endif 281cee22a15SViresh Kumar 282cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 283cee22a15SViresh Kumar 284bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 285bce90380STejun Heo 2864c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2874c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2884c16bd32STejun Heo 28968e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2902e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2915bcab335STejun Heo 29268e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 29368e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2947d19c5ceSTejun Heo 2957d19c5ceSTejun Heo /* the per-cpu worker pools */ 2967d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2977d19c5ceSTejun Heo cpu_worker_pools); 2987d19c5ceSTejun Heo 29968e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3007d19c5ceSTejun Heo 30168e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 30229c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 30329c91e99STejun Heo 304c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 30529c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 30629c91e99STejun Heo 3078a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3088a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3098a2b7538STejun Heo 310d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 311ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 312044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3131aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 314044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 315d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 316044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 317f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 318044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 31924d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3200668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3210668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3220668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3230668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 324d320c038STejun Heo 3257d19c5ceSTejun Heo static int worker_thread(void *__worker); 3267d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3277d19c5ceSTejun Heo const struct workqueue_attrs *from); 3287d19c5ceSTejun Heo 32997bd2347STejun Heo #define CREATE_TRACE_POINTS 33097bd2347STejun Heo #include <trace/events/workqueue.h> 33197bd2347STejun Heo 33268e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3335bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 33468e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 33568e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3365bcab335STejun Heo 337b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 33876af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 339b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 340b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 34176af4d93STejun Heo 342f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 343f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 344f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3457a62c2c8STejun Heo (pool)++) 3464ce62e9eSTejun Heo 34749e3cf44STejun Heo /** 34817116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 34917116969STejun Heo * @pool: iteration cursor 350611c92a0STejun Heo * @pi: integer used for iteration 351fa1b54e6STejun Heo * 35268e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 35368e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 35468e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 355fa1b54e6STejun Heo * 356fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 357fa1b54e6STejun Heo * ignored. 35817116969STejun Heo */ 359611c92a0STejun Heo #define for_each_pool(pool, pi) \ 360611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 36168e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 362fa1b54e6STejun Heo else 36317116969STejun Heo 36417116969STejun Heo /** 365822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 366822d8405STejun Heo * @worker: iteration cursor 367822d8405STejun Heo * @pool: worker_pool to iterate workers of 368822d8405STejun Heo * 36992f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 370822d8405STejun Heo * 371822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 372822d8405STejun Heo * ignored. 373822d8405STejun Heo */ 374da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 375da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 37692f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 377822d8405STejun Heo else 378822d8405STejun Heo 379822d8405STejun Heo /** 38049e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 38149e3cf44STejun Heo * @pwq: iteration cursor 38249e3cf44STejun Heo * @wq: the target workqueue 38376af4d93STejun Heo * 384b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 385794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 386794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 38776af4d93STejun Heo * 38876af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 38976af4d93STejun Heo * ignored. 39049e3cf44STejun Heo */ 39149e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 39276af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 393b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 39476af4d93STejun Heo else 395f3421797STejun Heo 396dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 397dc186ad7SThomas Gleixner 398dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 399dc186ad7SThomas Gleixner 40099777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 40199777288SStanislaw Gruszka { 40299777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 40399777288SStanislaw Gruszka } 40499777288SStanislaw Gruszka 405dc186ad7SThomas Gleixner /* 406dc186ad7SThomas Gleixner * fixup_init is called when: 407dc186ad7SThomas Gleixner * - an active object is initialized 408dc186ad7SThomas Gleixner */ 409dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 410dc186ad7SThomas Gleixner { 411dc186ad7SThomas Gleixner struct work_struct *work = addr; 412dc186ad7SThomas Gleixner 413dc186ad7SThomas Gleixner switch (state) { 414dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 415dc186ad7SThomas Gleixner cancel_work_sync(work); 416dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 417dc186ad7SThomas Gleixner return 1; 418dc186ad7SThomas Gleixner default: 419dc186ad7SThomas Gleixner return 0; 420dc186ad7SThomas Gleixner } 421dc186ad7SThomas Gleixner } 422dc186ad7SThomas Gleixner 423dc186ad7SThomas Gleixner /* 424dc186ad7SThomas Gleixner * fixup_activate is called when: 425dc186ad7SThomas Gleixner * - an active object is activated 426dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 427dc186ad7SThomas Gleixner */ 428dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 429dc186ad7SThomas Gleixner { 430dc186ad7SThomas Gleixner struct work_struct *work = addr; 431dc186ad7SThomas Gleixner 432dc186ad7SThomas Gleixner switch (state) { 433dc186ad7SThomas Gleixner 434dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 435dc186ad7SThomas Gleixner /* 436dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 437dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 438dc186ad7SThomas Gleixner * is tracked in the object tracker. 439dc186ad7SThomas Gleixner */ 44022df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 441dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 442dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 443dc186ad7SThomas Gleixner return 0; 444dc186ad7SThomas Gleixner } 445dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 446dc186ad7SThomas Gleixner return 0; 447dc186ad7SThomas Gleixner 448dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 449dc186ad7SThomas Gleixner WARN_ON(1); 450dc186ad7SThomas Gleixner 451dc186ad7SThomas Gleixner default: 452dc186ad7SThomas Gleixner return 0; 453dc186ad7SThomas Gleixner } 454dc186ad7SThomas Gleixner } 455dc186ad7SThomas Gleixner 456dc186ad7SThomas Gleixner /* 457dc186ad7SThomas Gleixner * fixup_free is called when: 458dc186ad7SThomas Gleixner * - an active object is freed 459dc186ad7SThomas Gleixner */ 460dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 461dc186ad7SThomas Gleixner { 462dc186ad7SThomas Gleixner struct work_struct *work = addr; 463dc186ad7SThomas Gleixner 464dc186ad7SThomas Gleixner switch (state) { 465dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 466dc186ad7SThomas Gleixner cancel_work_sync(work); 467dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 468dc186ad7SThomas Gleixner return 1; 469dc186ad7SThomas Gleixner default: 470dc186ad7SThomas Gleixner return 0; 471dc186ad7SThomas Gleixner } 472dc186ad7SThomas Gleixner } 473dc186ad7SThomas Gleixner 474dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 475dc186ad7SThomas Gleixner .name = "work_struct", 47699777288SStanislaw Gruszka .debug_hint = work_debug_hint, 477dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 478dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 479dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 480dc186ad7SThomas Gleixner }; 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 483dc186ad7SThomas Gleixner { 484dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 485dc186ad7SThomas Gleixner } 486dc186ad7SThomas Gleixner 487dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 488dc186ad7SThomas Gleixner { 489dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 490dc186ad7SThomas Gleixner } 491dc186ad7SThomas Gleixner 492dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 493dc186ad7SThomas Gleixner { 494dc186ad7SThomas Gleixner if (onstack) 495dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 496dc186ad7SThomas Gleixner else 497dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 498dc186ad7SThomas Gleixner } 499dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 500dc186ad7SThomas Gleixner 501dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 502dc186ad7SThomas Gleixner { 503dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 504dc186ad7SThomas Gleixner } 505dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 506dc186ad7SThomas Gleixner 507ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 508ea2e64f2SThomas Gleixner { 509ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 510ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 511ea2e64f2SThomas Gleixner } 512ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 513ea2e64f2SThomas Gleixner 514dc186ad7SThomas Gleixner #else 515dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 516dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 517dc186ad7SThomas Gleixner #endif 518dc186ad7SThomas Gleixner 5194e8b22bdSLi Bin /** 5204e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5214e8b22bdSLi Bin * @pool: the pool pointer of interest 5224e8b22bdSLi Bin * 5234e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5244e8b22bdSLi Bin * successfully, -errno on failure. 5254e8b22bdSLi Bin */ 5269daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5279daf9e67STejun Heo { 5289daf9e67STejun Heo int ret; 5299daf9e67STejun Heo 53068e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5315bcab335STejun Heo 5324e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5334e8b22bdSLi Bin GFP_KERNEL); 534229641a6STejun Heo if (ret >= 0) { 535e68035fbSTejun Heo pool->id = ret; 536229641a6STejun Heo return 0; 537229641a6STejun Heo } 5389daf9e67STejun Heo return ret; 5399daf9e67STejun Heo } 5409daf9e67STejun Heo 54176af4d93STejun Heo /** 542df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 543df2d5ae4STejun Heo * @wq: the target workqueue 544df2d5ae4STejun Heo * @node: the node ID 545df2d5ae4STejun Heo * 546df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 547df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 548df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 549d185af30SYacine Belkadi * 550d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 551df2d5ae4STejun Heo */ 552df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 553df2d5ae4STejun Heo int node) 554df2d5ae4STejun Heo { 555df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 556df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 557df2d5ae4STejun Heo } 558df2d5ae4STejun Heo 55973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 56073f53c4aSTejun Heo { 56173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 56273f53c4aSTejun Heo } 56373f53c4aSTejun Heo 56473f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 56573f53c4aSTejun Heo { 56673f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 56773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 56873f53c4aSTejun Heo } 56973f53c4aSTejun Heo 57073f53c4aSTejun Heo static int work_next_color(int color) 57173f53c4aSTejun Heo { 57273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 573a848e3b6SOleg Nesterov } 574a848e3b6SOleg Nesterov 5754594bf15SDavid Howells /* 576112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 577112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5787c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5797a22ad75STejun Heo * 580112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 581112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 582bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 583bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5847a22ad75STejun Heo * 585112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5867c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 587112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5887c3eed5cSTejun Heo * available only while the work item is queued. 589bbb68dfaSTejun Heo * 590bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 591bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 592bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 593bbb68dfaSTejun Heo * try to steal the PENDING bit. 5944594bf15SDavid Howells */ 5957a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5967a22ad75STejun Heo unsigned long flags) 5977a22ad75STejun Heo { 5986183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5997a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6007a22ad75STejun Heo } 6017a22ad75STejun Heo 602112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6034690c4abSTejun Heo unsigned long extra_flags) 604365970a1SDavid Howells { 605112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 606112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 607365970a1SDavid Howells } 608365970a1SDavid Howells 6094468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6104468a00fSLai Jiangshan int pool_id) 6114468a00fSLai Jiangshan { 6124468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6134468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6144468a00fSLai Jiangshan } 6154468a00fSLai Jiangshan 6167c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6177c3eed5cSTejun Heo int pool_id) 6184d707b9fSOleg Nesterov { 61923657bb1STejun Heo /* 62023657bb1STejun Heo * The following wmb is paired with the implied mb in 62123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 62223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 62323657bb1STejun Heo * owner. 62423657bb1STejun Heo */ 62523657bb1STejun Heo smp_wmb(); 6267c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6274d707b9fSOleg Nesterov } 6284d707b9fSOleg Nesterov 6297a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 630365970a1SDavid Howells { 6317c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6327c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6337a22ad75STejun Heo } 6347a22ad75STejun Heo 635112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6367a22ad75STejun Heo { 637e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6387a22ad75STejun Heo 639112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 640e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 641e120153dSTejun Heo else 642e120153dSTejun Heo return NULL; 6437a22ad75STejun Heo } 6447a22ad75STejun Heo 6457c3eed5cSTejun Heo /** 6467c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6477c3eed5cSTejun Heo * @work: the work item of interest 6487c3eed5cSTejun Heo * 64968e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 65068e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 65168e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 652fa1b54e6STejun Heo * 653fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 654fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 655fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 656fa1b54e6STejun Heo * returned pool is and stays online. 657d185af30SYacine Belkadi * 658d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6597c3eed5cSTejun Heo */ 6607c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6617a22ad75STejun Heo { 662e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6637c3eed5cSTejun Heo int pool_id; 6647a22ad75STejun Heo 66568e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 666fa1b54e6STejun Heo 667112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 668112202d9STejun Heo return ((struct pool_workqueue *) 6697c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6707a22ad75STejun Heo 6717c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6727c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6737a22ad75STejun Heo return NULL; 6747a22ad75STejun Heo 675fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6767c3eed5cSTejun Heo } 6777c3eed5cSTejun Heo 6787c3eed5cSTejun Heo /** 6797c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6807c3eed5cSTejun Heo * @work: the work item of interest 6817c3eed5cSTejun Heo * 682d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6837c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6847c3eed5cSTejun Heo */ 6857c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6867c3eed5cSTejun Heo { 68754d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6887c3eed5cSTejun Heo 689112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 690112202d9STejun Heo return ((struct pool_workqueue *) 69154d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 69254d5b7d0SLai Jiangshan 69354d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6947c3eed5cSTejun Heo } 6957c3eed5cSTejun Heo 696bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 697bbb68dfaSTejun Heo { 6987c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 699bbb68dfaSTejun Heo 7007c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7017c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 702bbb68dfaSTejun Heo } 703bbb68dfaSTejun Heo 704bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 705bbb68dfaSTejun Heo { 706bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 707bbb68dfaSTejun Heo 708112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 709bbb68dfaSTejun Heo } 710bbb68dfaSTejun Heo 711e22bee78STejun Heo /* 7123270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7133270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 714d565ed63STejun Heo * they're being called with pool->lock held. 715e22bee78STejun Heo */ 716e22bee78STejun Heo 71763d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 718649027d7STejun Heo { 719e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 720649027d7STejun Heo } 721649027d7STejun Heo 722e22bee78STejun Heo /* 723e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 724e22bee78STejun Heo * running workers. 725974271c4STejun Heo * 726974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 727706026c2STejun Heo * function will always return %true for unbound pools as long as the 728974271c4STejun Heo * worklist isn't empty. 729e22bee78STejun Heo */ 73063d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 731e22bee78STejun Heo { 73263d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 733e22bee78STejun Heo } 734e22bee78STejun Heo 735e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 73663d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 737e22bee78STejun Heo { 73863d95a91STejun Heo return pool->nr_idle; 739e22bee78STejun Heo } 740e22bee78STejun Heo 741e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 74263d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 743e22bee78STejun Heo { 744e19e397aSTejun Heo return !list_empty(&pool->worklist) && 745e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 746e22bee78STejun Heo } 747e22bee78STejun Heo 748e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 74963d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 750e22bee78STejun Heo { 75163d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 752e22bee78STejun Heo } 753e22bee78STejun Heo 754e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 75563d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 756e22bee78STejun Heo { 75734a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 75863d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 75963d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 760e22bee78STejun Heo 761e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 762e22bee78STejun Heo } 763e22bee78STejun Heo 764e22bee78STejun Heo /* 765e22bee78STejun Heo * Wake up functions. 766e22bee78STejun Heo */ 767e22bee78STejun Heo 7681037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 7691037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 7707e11629dSTejun Heo { 77163d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7727e11629dSTejun Heo return NULL; 7737e11629dSTejun Heo 77463d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7757e11629dSTejun Heo } 7767e11629dSTejun Heo 7777e11629dSTejun Heo /** 7787e11629dSTejun Heo * wake_up_worker - wake up an idle worker 77963d95a91STejun Heo * @pool: worker pool to wake worker from 7807e11629dSTejun Heo * 78163d95a91STejun Heo * Wake up the first idle worker of @pool. 7827e11629dSTejun Heo * 7837e11629dSTejun Heo * CONTEXT: 784d565ed63STejun Heo * spin_lock_irq(pool->lock). 7857e11629dSTejun Heo */ 78663d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7877e11629dSTejun Heo { 7881037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 7897e11629dSTejun Heo 7907e11629dSTejun Heo if (likely(worker)) 7917e11629dSTejun Heo wake_up_process(worker->task); 7927e11629dSTejun Heo } 7937e11629dSTejun Heo 7944690c4abSTejun Heo /** 795e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 796e22bee78STejun Heo * @task: task waking up 797e22bee78STejun Heo * @cpu: CPU @task is waking up to 798e22bee78STejun Heo * 799e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 800e22bee78STejun Heo * being awoken. 801e22bee78STejun Heo * 802e22bee78STejun Heo * CONTEXT: 803e22bee78STejun Heo * spin_lock_irq(rq->lock) 804e22bee78STejun Heo */ 805d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 806e22bee78STejun Heo { 807e22bee78STejun Heo struct worker *worker = kthread_data(task); 808e22bee78STejun Heo 80936576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 810ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 811e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 812e22bee78STejun Heo } 81336576000SJoonsoo Kim } 814e22bee78STejun Heo 815e22bee78STejun Heo /** 816e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 817e22bee78STejun Heo * @task: task going to sleep 818e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 819e22bee78STejun Heo * 820e22bee78STejun Heo * This function is called during schedule() when a busy worker is 821e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 822e22bee78STejun Heo * returning pointer to its task. 823e22bee78STejun Heo * 824e22bee78STejun Heo * CONTEXT: 825e22bee78STejun Heo * spin_lock_irq(rq->lock) 826e22bee78STejun Heo * 827d185af30SYacine Belkadi * Return: 828e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 829e22bee78STejun Heo */ 830d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 831e22bee78STejun Heo { 832e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 833111c225aSTejun Heo struct worker_pool *pool; 834e22bee78STejun Heo 835111c225aSTejun Heo /* 836111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 837111c225aSTejun Heo * workers, also reach here, let's not access anything before 838111c225aSTejun Heo * checking NOT_RUNNING. 839111c225aSTejun Heo */ 8402d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 841e22bee78STejun Heo return NULL; 842e22bee78STejun Heo 843111c225aSTejun Heo pool = worker->pool; 844111c225aSTejun Heo 845e22bee78STejun Heo /* this can only happen on the local cpu */ 84692b69f50SLai Jiangshan if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) 8476183c009STejun Heo return NULL; 848e22bee78STejun Heo 849e22bee78STejun Heo /* 850e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 851e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 852e22bee78STejun Heo * Please read comment there. 853e22bee78STejun Heo * 854628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 855628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 856628c78e7STejun Heo * disabled, which in turn means that none else could be 857d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 858628c78e7STejun Heo * lock is safe. 859e22bee78STejun Heo */ 860e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 861e19e397aSTejun Heo !list_empty(&pool->worklist)) 8621037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 863e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 864e22bee78STejun Heo } 865e22bee78STejun Heo 866e22bee78STejun Heo /** 867e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 868cb444766STejun Heo * @worker: self 869d302f017STejun Heo * @flags: flags to set 870d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 871d302f017STejun Heo * 872e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 873e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 874e22bee78STejun Heo * woken up. 875d302f017STejun Heo * 876cb444766STejun Heo * CONTEXT: 877d565ed63STejun Heo * spin_lock_irq(pool->lock) 878d302f017STejun Heo */ 879d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 880d302f017STejun Heo bool wakeup) 881d302f017STejun Heo { 882bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 883e22bee78STejun Heo 884cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 885cb444766STejun Heo 886e22bee78STejun Heo /* 887e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 888e22bee78STejun Heo * wake up an idle worker as necessary if requested by 889e22bee78STejun Heo * @wakeup. 890e22bee78STejun Heo */ 891e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 892e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 893e22bee78STejun Heo if (wakeup) { 894e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 895bd7bdd43STejun Heo !list_empty(&pool->worklist)) 89663d95a91STejun Heo wake_up_worker(pool); 897e22bee78STejun Heo } else 898e19e397aSTejun Heo atomic_dec(&pool->nr_running); 899e22bee78STejun Heo } 900e22bee78STejun Heo 901d302f017STejun Heo worker->flags |= flags; 902d302f017STejun Heo } 903d302f017STejun Heo 904d302f017STejun Heo /** 905e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 906cb444766STejun Heo * @worker: self 907d302f017STejun Heo * @flags: flags to clear 908d302f017STejun Heo * 909e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 910d302f017STejun Heo * 911cb444766STejun Heo * CONTEXT: 912d565ed63STejun Heo * spin_lock_irq(pool->lock) 913d302f017STejun Heo */ 914d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 915d302f017STejun Heo { 91663d95a91STejun Heo struct worker_pool *pool = worker->pool; 917e22bee78STejun Heo unsigned int oflags = worker->flags; 918e22bee78STejun Heo 919cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 920cb444766STejun Heo 921d302f017STejun Heo worker->flags &= ~flags; 922e22bee78STejun Heo 92342c025f3STejun Heo /* 92442c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 92542c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 92642c025f3STejun Heo * of multiple flags, not a single flag. 92742c025f3STejun Heo */ 928e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 929e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 930e19e397aSTejun Heo atomic_inc(&pool->nr_running); 931d302f017STejun Heo } 932d302f017STejun Heo 933d302f017STejun Heo /** 9348cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 935c9e7cf27STejun Heo * @pool: pool of interest 9368cca0eeaSTejun Heo * @work: work to find worker for 9378cca0eeaSTejun Heo * 938c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 939c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 940a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 941a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 942a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 943a2c1c57bSTejun Heo * being executed. 944a2c1c57bSTejun Heo * 945a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 946a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 947a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 948a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 949a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 950a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 951a2c1c57bSTejun Heo * 952c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 953c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 954c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 955c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 956c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 957c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9588cca0eeaSTejun Heo * 9598cca0eeaSTejun Heo * CONTEXT: 960d565ed63STejun Heo * spin_lock_irq(pool->lock). 9618cca0eeaSTejun Heo * 962d185af30SYacine Belkadi * Return: 963d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9648cca0eeaSTejun Heo * otherwise. 9658cca0eeaSTejun Heo */ 966c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9678cca0eeaSTejun Heo struct work_struct *work) 9688cca0eeaSTejun Heo { 96942f8570fSSasha Levin struct worker *worker; 97042f8570fSSasha Levin 971b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 972a2c1c57bSTejun Heo (unsigned long)work) 973a2c1c57bSTejun Heo if (worker->current_work == work && 974a2c1c57bSTejun Heo worker->current_func == work->func) 97542f8570fSSasha Levin return worker; 97642f8570fSSasha Levin 97742f8570fSSasha Levin return NULL; 9788cca0eeaSTejun Heo } 9798cca0eeaSTejun Heo 9808cca0eeaSTejun Heo /** 981bf4ede01STejun Heo * move_linked_works - move linked works to a list 982bf4ede01STejun Heo * @work: start of series of works to be scheduled 983bf4ede01STejun Heo * @head: target list to append @work to 984bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 985bf4ede01STejun Heo * 986bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 987bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 988bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 989bf4ede01STejun Heo * 990bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 991bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 992bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 993bf4ede01STejun Heo * 994bf4ede01STejun Heo * CONTEXT: 995d565ed63STejun Heo * spin_lock_irq(pool->lock). 996bf4ede01STejun Heo */ 997bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 998bf4ede01STejun Heo struct work_struct **nextp) 999bf4ede01STejun Heo { 1000bf4ede01STejun Heo struct work_struct *n; 1001bf4ede01STejun Heo 1002bf4ede01STejun Heo /* 1003bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1004bf4ede01STejun Heo * use NULL for list head. 1005bf4ede01STejun Heo */ 1006bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1007bf4ede01STejun Heo list_move_tail(&work->entry, head); 1008bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1009bf4ede01STejun Heo break; 1010bf4ede01STejun Heo } 1011bf4ede01STejun Heo 1012bf4ede01STejun Heo /* 1013bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1014bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1015bf4ede01STejun Heo * needs to be updated. 1016bf4ede01STejun Heo */ 1017bf4ede01STejun Heo if (nextp) 1018bf4ede01STejun Heo *nextp = n; 1019bf4ede01STejun Heo } 1020bf4ede01STejun Heo 10218864b4e5STejun Heo /** 10228864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10238864b4e5STejun Heo * @pwq: pool_workqueue to get 10248864b4e5STejun Heo * 10258864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10268864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10278864b4e5STejun Heo */ 10288864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10298864b4e5STejun Heo { 10308864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10318864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10328864b4e5STejun Heo pwq->refcnt++; 10338864b4e5STejun Heo } 10348864b4e5STejun Heo 10358864b4e5STejun Heo /** 10368864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10378864b4e5STejun Heo * @pwq: pool_workqueue to put 10388864b4e5STejun Heo * 10398864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10408864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10418864b4e5STejun Heo */ 10428864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10438864b4e5STejun Heo { 10448864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10458864b4e5STejun Heo if (likely(--pwq->refcnt)) 10468864b4e5STejun Heo return; 10478864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10488864b4e5STejun Heo return; 10498864b4e5STejun Heo /* 10508864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10518864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10528864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10538864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10548864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10558864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10568864b4e5STejun Heo */ 10578864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10588864b4e5STejun Heo } 10598864b4e5STejun Heo 1060dce90d47STejun Heo /** 1061dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1062dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1063dce90d47STejun Heo * 1064dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1065dce90d47STejun Heo */ 1066dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1067dce90d47STejun Heo { 1068dce90d47STejun Heo if (pwq) { 1069dce90d47STejun Heo /* 1070dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1071dce90d47STejun Heo * following lock operations are safe. 1072dce90d47STejun Heo */ 1073dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1074dce90d47STejun Heo put_pwq(pwq); 1075dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1076dce90d47STejun Heo } 1077dce90d47STejun Heo } 1078dce90d47STejun Heo 1079112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1080bf4ede01STejun Heo { 1081112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1082bf4ede01STejun Heo 1083bf4ede01STejun Heo trace_workqueue_activate_work(work); 1084112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1085bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1086112202d9STejun Heo pwq->nr_active++; 1087bf4ede01STejun Heo } 1088bf4ede01STejun Heo 1089112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10903aa62497SLai Jiangshan { 1091112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10923aa62497SLai Jiangshan struct work_struct, entry); 10933aa62497SLai Jiangshan 1094112202d9STejun Heo pwq_activate_delayed_work(work); 10953aa62497SLai Jiangshan } 10963aa62497SLai Jiangshan 1097bf4ede01STejun Heo /** 1098112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1099112202d9STejun Heo * @pwq: pwq of interest 1100bf4ede01STejun Heo * @color: color of work which left the queue 1101bf4ede01STejun Heo * 1102bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1103112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1104bf4ede01STejun Heo * 1105bf4ede01STejun Heo * CONTEXT: 1106d565ed63STejun Heo * spin_lock_irq(pool->lock). 1107bf4ede01STejun Heo */ 1108112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1109bf4ede01STejun Heo { 11108864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1111bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11128864b4e5STejun Heo goto out_put; 1113bf4ede01STejun Heo 1114112202d9STejun Heo pwq->nr_in_flight[color]--; 1115bf4ede01STejun Heo 1116112202d9STejun Heo pwq->nr_active--; 1117112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1118bf4ede01STejun Heo /* one down, submit a delayed one */ 1119112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1120112202d9STejun Heo pwq_activate_first_delayed(pwq); 1121bf4ede01STejun Heo } 1122bf4ede01STejun Heo 1123bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1124112202d9STejun Heo if (likely(pwq->flush_color != color)) 11258864b4e5STejun Heo goto out_put; 1126bf4ede01STejun Heo 1127bf4ede01STejun Heo /* are there still in-flight works? */ 1128112202d9STejun Heo if (pwq->nr_in_flight[color]) 11298864b4e5STejun Heo goto out_put; 1130bf4ede01STejun Heo 1131112202d9STejun Heo /* this pwq is done, clear flush_color */ 1132112202d9STejun Heo pwq->flush_color = -1; 1133bf4ede01STejun Heo 1134bf4ede01STejun Heo /* 1135112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1136bf4ede01STejun Heo * will handle the rest. 1137bf4ede01STejun Heo */ 1138112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1139112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11408864b4e5STejun Heo out_put: 11418864b4e5STejun Heo put_pwq(pwq); 1142bf4ede01STejun Heo } 1143bf4ede01STejun Heo 114436e227d2STejun Heo /** 1145bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 114636e227d2STejun Heo * @work: work item to steal 114736e227d2STejun Heo * @is_dwork: @work is a delayed_work 1148bbb68dfaSTejun Heo * @flags: place to store irq state 114936e227d2STejun Heo * 115036e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1151d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 115236e227d2STejun Heo * 1153d185af30SYacine Belkadi * Return: 115436e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 115536e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 115636e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1157bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1158bbb68dfaSTejun Heo * for arbitrarily long 115936e227d2STejun Heo * 1160d185af30SYacine Belkadi * Note: 1161bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1162e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1163e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1164e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1165bbb68dfaSTejun Heo * 1166bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1167bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1168bbb68dfaSTejun Heo * 1169e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1170bf4ede01STejun Heo */ 1171bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1172bbb68dfaSTejun Heo unsigned long *flags) 1173bf4ede01STejun Heo { 1174d565ed63STejun Heo struct worker_pool *pool; 1175112202d9STejun Heo struct pool_workqueue *pwq; 1176bf4ede01STejun Heo 1177bbb68dfaSTejun Heo local_irq_save(*flags); 1178bbb68dfaSTejun Heo 117936e227d2STejun Heo /* try to steal the timer if it exists */ 118036e227d2STejun Heo if (is_dwork) { 118136e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 118236e227d2STejun Heo 1183e0aecdd8STejun Heo /* 1184e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1185e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1186e0aecdd8STejun Heo * running on the local CPU. 1187e0aecdd8STejun Heo */ 118836e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 118936e227d2STejun Heo return 1; 119036e227d2STejun Heo } 119136e227d2STejun Heo 119236e227d2STejun Heo /* try to claim PENDING the normal way */ 1193bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1194bf4ede01STejun Heo return 0; 1195bf4ede01STejun Heo 1196bf4ede01STejun Heo /* 1197bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1198bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1199bf4ede01STejun Heo */ 1200d565ed63STejun Heo pool = get_work_pool(work); 1201d565ed63STejun Heo if (!pool) 1202bbb68dfaSTejun Heo goto fail; 1203bf4ede01STejun Heo 1204d565ed63STejun Heo spin_lock(&pool->lock); 1205bf4ede01STejun Heo /* 1206112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1207112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1208112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1209112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1210112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12110b3dae68SLai Jiangshan * item is currently queued on that pool. 1212bf4ede01STejun Heo */ 1213112202d9STejun Heo pwq = get_work_pwq(work); 1214112202d9STejun Heo if (pwq && pwq->pool == pool) { 1215bf4ede01STejun Heo debug_work_deactivate(work); 12163aa62497SLai Jiangshan 12173aa62497SLai Jiangshan /* 121816062836STejun Heo * A delayed work item cannot be grabbed directly because 121916062836STejun Heo * it might have linked NO_COLOR work items which, if left 1220112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 122116062836STejun Heo * management later on and cause stall. Make sure the work 122216062836STejun Heo * item is activated before grabbing. 12233aa62497SLai Jiangshan */ 12243aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1225112202d9STejun Heo pwq_activate_delayed_work(work); 12263aa62497SLai Jiangshan 1227bf4ede01STejun Heo list_del_init(&work->entry); 12289c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 122936e227d2STejun Heo 1230112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12314468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12324468a00fSLai Jiangshan 1233d565ed63STejun Heo spin_unlock(&pool->lock); 123436e227d2STejun Heo return 1; 1235bf4ede01STejun Heo } 1236d565ed63STejun Heo spin_unlock(&pool->lock); 1237bbb68dfaSTejun Heo fail: 1238bbb68dfaSTejun Heo local_irq_restore(*flags); 1239bbb68dfaSTejun Heo if (work_is_canceling(work)) 1240bbb68dfaSTejun Heo return -ENOENT; 1241bbb68dfaSTejun Heo cpu_relax(); 124236e227d2STejun Heo return -EAGAIN; 1243bf4ede01STejun Heo } 1244bf4ede01STejun Heo 1245bf4ede01STejun Heo /** 1246706026c2STejun Heo * insert_work - insert a work into a pool 1247112202d9STejun Heo * @pwq: pwq @work belongs to 12484690c4abSTejun Heo * @work: work to insert 12494690c4abSTejun Heo * @head: insertion point 12504690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12514690c4abSTejun Heo * 1252112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1253706026c2STejun Heo * work_struct flags. 12544690c4abSTejun Heo * 12554690c4abSTejun Heo * CONTEXT: 1256d565ed63STejun Heo * spin_lock_irq(pool->lock). 1257365970a1SDavid Howells */ 1258112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1259112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1260b89deed3SOleg Nesterov { 1261112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1262e1d8aa9fSFrederic Weisbecker 12634690c4abSTejun Heo /* we own @work, set data and link */ 1264112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12651a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12668864b4e5STejun Heo get_pwq(pwq); 1267e22bee78STejun Heo 1268e22bee78STejun Heo /* 1269c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1270c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1271c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1272e22bee78STejun Heo */ 1273e22bee78STejun Heo smp_mb(); 1274e22bee78STejun Heo 127563d95a91STejun Heo if (__need_more_worker(pool)) 127663d95a91STejun Heo wake_up_worker(pool); 1277b89deed3SOleg Nesterov } 1278b89deed3SOleg Nesterov 1279c8efcc25STejun Heo /* 1280c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12818d03ecfeSTejun Heo * same workqueue. 1282c8efcc25STejun Heo */ 1283c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1284c8efcc25STejun Heo { 1285c8efcc25STejun Heo struct worker *worker; 1286c8efcc25STejun Heo 12878d03ecfeSTejun Heo worker = current_wq_worker(); 1288c8efcc25STejun Heo /* 12898d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12908d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1291c8efcc25STejun Heo */ 1292112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1293c8efcc25STejun Heo } 1294c8efcc25STejun Heo 1295d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12961da177e4SLinus Torvalds struct work_struct *work) 12971da177e4SLinus Torvalds { 1298112202d9STejun Heo struct pool_workqueue *pwq; 1299c9178087STejun Heo struct worker_pool *last_pool; 13001e19ffc6STejun Heo struct list_head *worklist; 13018a2e8e5dSTejun Heo unsigned int work_flags; 1302b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13038930cabaSTejun Heo 13048930cabaSTejun Heo /* 13058930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13068930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13078930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13088930cabaSTejun Heo * happen with IRQ disabled. 13098930cabaSTejun Heo */ 13108930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13111da177e4SLinus Torvalds 1312dc186ad7SThomas Gleixner debug_work_activate(work); 13131e19ffc6STejun Heo 13149ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1315618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1316c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1317e41e704bSTejun Heo return; 13189e8cd2f5STejun Heo retry: 1319df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1320f3421797STejun Heo cpu = raw_smp_processor_id(); 1321df2d5ae4STejun Heo 1322df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1323df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1324c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1325df2d5ae4STejun Heo else 1326df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1327f3421797STejun Heo 132818aa9effSTejun Heo /* 1329c9178087STejun Heo * If @work was previously on a different pool, it might still be 1330c9178087STejun Heo * running there, in which case the work needs to be queued on that 1331c9178087STejun Heo * pool to guarantee non-reentrancy. 133218aa9effSTejun Heo */ 1333c9e7cf27STejun Heo last_pool = get_work_pool(work); 1334112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 133518aa9effSTejun Heo struct worker *worker; 133618aa9effSTejun Heo 1337d565ed63STejun Heo spin_lock(&last_pool->lock); 133818aa9effSTejun Heo 1339c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 134018aa9effSTejun Heo 1341112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1342c9178087STejun Heo pwq = worker->current_pwq; 13438594fadeSLai Jiangshan } else { 134418aa9effSTejun Heo /* meh... not running there, queue here */ 1345d565ed63STejun Heo spin_unlock(&last_pool->lock); 1346112202d9STejun Heo spin_lock(&pwq->pool->lock); 134718aa9effSTejun Heo } 13488930cabaSTejun Heo } else { 1349112202d9STejun Heo spin_lock(&pwq->pool->lock); 13508930cabaSTejun Heo } 1351502ca9d8STejun Heo 13529e8cd2f5STejun Heo /* 13539e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13549e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13559e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1356df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1357df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13589e8cd2f5STejun Heo * make forward-progress. 13599e8cd2f5STejun Heo */ 13609e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13619e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13629e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13639e8cd2f5STejun Heo cpu_relax(); 13649e8cd2f5STejun Heo goto retry; 13659e8cd2f5STejun Heo } 13669e8cd2f5STejun Heo /* oops */ 13679e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13689e8cd2f5STejun Heo wq->name, cpu); 13699e8cd2f5STejun Heo } 13709e8cd2f5STejun Heo 1371112202d9STejun Heo /* pwq determined, queue */ 1372112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1373502ca9d8STejun Heo 1374f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1375112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1376f5b2552bSDan Carpenter return; 1377f5b2552bSDan Carpenter } 13781e19ffc6STejun Heo 1379112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1380112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13811e19ffc6STejun Heo 1382112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1383cdadf009STejun Heo trace_workqueue_activate_work(work); 1384112202d9STejun Heo pwq->nr_active++; 1385112202d9STejun Heo worklist = &pwq->pool->worklist; 13868a2e8e5dSTejun Heo } else { 13878a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1388112202d9STejun Heo worklist = &pwq->delayed_works; 13898a2e8e5dSTejun Heo } 13901e19ffc6STejun Heo 1391112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13921e19ffc6STejun Heo 1393112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13941da177e4SLinus Torvalds } 13951da177e4SLinus Torvalds 13960fcb78c2SRolf Eike Beer /** 1397c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1398c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1399c1a220e7SZhang Rui * @wq: workqueue to use 1400c1a220e7SZhang Rui * @work: work to queue 1401c1a220e7SZhang Rui * 1402c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1403c1a220e7SZhang Rui * can't go away. 1404d185af30SYacine Belkadi * 1405d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1406c1a220e7SZhang Rui */ 1407d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1408d4283e93STejun Heo struct work_struct *work) 1409c1a220e7SZhang Rui { 1410d4283e93STejun Heo bool ret = false; 14118930cabaSTejun Heo unsigned long flags; 14128930cabaSTejun Heo 14138930cabaSTejun Heo local_irq_save(flags); 1414c1a220e7SZhang Rui 141522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14164690c4abSTejun Heo __queue_work(cpu, wq, work); 1417d4283e93STejun Heo ret = true; 1418c1a220e7SZhang Rui } 14198930cabaSTejun Heo 14208930cabaSTejun Heo local_irq_restore(flags); 1421c1a220e7SZhang Rui return ret; 1422c1a220e7SZhang Rui } 1423ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1424c1a220e7SZhang Rui 1425d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14261da177e4SLinus Torvalds { 142752bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14281da177e4SLinus Torvalds 1429e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 143060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14311da177e4SLinus Torvalds } 14321438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14331da177e4SLinus Torvalds 14347beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 143552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14361da177e4SLinus Torvalds { 14377beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14387beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14391da177e4SLinus Torvalds 14407beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14417beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1442fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1443fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14447beb2edfSTejun Heo 14458852aac2STejun Heo /* 14468852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14478852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14488852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14498852aac2STejun Heo * on that there's no such delay when @delay is 0. 14508852aac2STejun Heo */ 14518852aac2STejun Heo if (!delay) { 14528852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14538852aac2STejun Heo return; 14548852aac2STejun Heo } 14558852aac2STejun Heo 14567beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14577beb2edfSTejun Heo 145860c057bcSLai Jiangshan dwork->wq = wq; 14591265057fSTejun Heo dwork->cpu = cpu; 14607beb2edfSTejun Heo timer->expires = jiffies + delay; 14617beb2edfSTejun Heo 14627beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14637beb2edfSTejun Heo add_timer_on(timer, cpu); 14647beb2edfSTejun Heo else 14657beb2edfSTejun Heo add_timer(timer); 14667beb2edfSTejun Heo } 14671da177e4SLinus Torvalds 14680fcb78c2SRolf Eike Beer /** 14690fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14700fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14710fcb78c2SRolf Eike Beer * @wq: workqueue to use 1472af9997e4SRandy Dunlap * @dwork: work to queue 14730fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14740fcb78c2SRolf Eike Beer * 1475d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1476715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1477715f1300STejun Heo * execution. 14780fcb78c2SRolf Eike Beer */ 1479d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 148052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14817a6bc1cdSVenkatesh Pallipadi { 148252bad64dSDavid Howells struct work_struct *work = &dwork->work; 1483d4283e93STejun Heo bool ret = false; 14848930cabaSTejun Heo unsigned long flags; 14858930cabaSTejun Heo 14868930cabaSTejun Heo /* read the comment in __queue_work() */ 14878930cabaSTejun Heo local_irq_save(flags); 14887a6bc1cdSVenkatesh Pallipadi 148922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14907beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1491d4283e93STejun Heo ret = true; 14927a6bc1cdSVenkatesh Pallipadi } 14938930cabaSTejun Heo 14948930cabaSTejun Heo local_irq_restore(flags); 14957a6bc1cdSVenkatesh Pallipadi return ret; 14967a6bc1cdSVenkatesh Pallipadi } 1497ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 14981da177e4SLinus Torvalds 1499c8e55f36STejun Heo /** 15008376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15018376fe22STejun Heo * @cpu: CPU number to execute work on 15028376fe22STejun Heo * @wq: workqueue to use 15038376fe22STejun Heo * @dwork: work to queue 15048376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15058376fe22STejun Heo * 15068376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15078376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15088376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15098376fe22STejun Heo * current state. 15108376fe22STejun Heo * 1511d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15128376fe22STejun Heo * pending and its timer was modified. 15138376fe22STejun Heo * 1514e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15158376fe22STejun Heo * See try_to_grab_pending() for details. 15168376fe22STejun Heo */ 15178376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15188376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15198376fe22STejun Heo { 15208376fe22STejun Heo unsigned long flags; 15218376fe22STejun Heo int ret; 15228376fe22STejun Heo 15238376fe22STejun Heo do { 15248376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15258376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15268376fe22STejun Heo 15278376fe22STejun Heo if (likely(ret >= 0)) { 15288376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15298376fe22STejun Heo local_irq_restore(flags); 15308376fe22STejun Heo } 15318376fe22STejun Heo 15328376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15338376fe22STejun Heo return ret; 15348376fe22STejun Heo } 15358376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15368376fe22STejun Heo 15378376fe22STejun Heo /** 1538c8e55f36STejun Heo * worker_enter_idle - enter idle state 1539c8e55f36STejun Heo * @worker: worker which is entering idle state 1540c8e55f36STejun Heo * 1541c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1542c8e55f36STejun Heo * necessary. 1543c8e55f36STejun Heo * 1544c8e55f36STejun Heo * LOCKING: 1545d565ed63STejun Heo * spin_lock_irq(pool->lock). 1546c8e55f36STejun Heo */ 1547c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15481da177e4SLinus Torvalds { 1549bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1550c8e55f36STejun Heo 15516183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15526183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15536183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15546183c009STejun Heo return; 1555c8e55f36STejun Heo 1556cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1557cb444766STejun Heo worker->flags |= WORKER_IDLE; 1558bd7bdd43STejun Heo pool->nr_idle++; 1559e22bee78STejun Heo worker->last_active = jiffies; 1560c8e55f36STejun Heo 1561c8e55f36STejun Heo /* idle_list is LIFO */ 1562bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1563db7bccf4STejun Heo 156463d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1565628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1566cb444766STejun Heo 1567544ecf31STejun Heo /* 1568706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1569d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1570628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1571628c78e7STejun Heo * unbind is not in progress. 1572544ecf31STejun Heo */ 157324647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1574bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1575e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1576c8e55f36STejun Heo } 1577c8e55f36STejun Heo 1578c8e55f36STejun Heo /** 1579c8e55f36STejun Heo * worker_leave_idle - leave idle state 1580c8e55f36STejun Heo * @worker: worker which is leaving idle state 1581c8e55f36STejun Heo * 1582c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1583c8e55f36STejun Heo * 1584c8e55f36STejun Heo * LOCKING: 1585d565ed63STejun Heo * spin_lock_irq(pool->lock). 1586c8e55f36STejun Heo */ 1587c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1588c8e55f36STejun Heo { 1589bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1590c8e55f36STejun Heo 15916183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15926183c009STejun Heo return; 1593d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1594bd7bdd43STejun Heo pool->nr_idle--; 1595c8e55f36STejun Heo list_del_init(&worker->entry); 1596c8e55f36STejun Heo } 1597c8e55f36STejun Heo 1598f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1599c34056a3STejun Heo { 1600c34056a3STejun Heo struct worker *worker; 1601c34056a3STejun Heo 1602f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1603c8e55f36STejun Heo if (worker) { 1604c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1605affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1606da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1607e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1608e22bee78STejun Heo worker->flags = WORKER_PREP; 1609c8e55f36STejun Heo } 1610c34056a3STejun Heo return worker; 1611c34056a3STejun Heo } 1612c34056a3STejun Heo 1613c34056a3STejun Heo /** 16144736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16154736cbf7SLai Jiangshan * @worker: worker to be attached 16164736cbf7SLai Jiangshan * @pool: the target pool 16174736cbf7SLai Jiangshan * 16184736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16194736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16204736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16214736cbf7SLai Jiangshan */ 16224736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16234736cbf7SLai Jiangshan struct worker_pool *pool) 16244736cbf7SLai Jiangshan { 16254736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16264736cbf7SLai Jiangshan 16274736cbf7SLai Jiangshan /* 16284736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16294736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16304736cbf7SLai Jiangshan */ 16314736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16324736cbf7SLai Jiangshan 16334736cbf7SLai Jiangshan /* 16344736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 16354736cbf7SLai Jiangshan * stable across this function. See the comments above the 16364736cbf7SLai Jiangshan * flag definition for details. 16374736cbf7SLai Jiangshan */ 16384736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 16394736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 16404736cbf7SLai Jiangshan 16414736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 16424736cbf7SLai Jiangshan 16434736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 16444736cbf7SLai Jiangshan } 16454736cbf7SLai Jiangshan 16464736cbf7SLai Jiangshan /** 164760f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 164860f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 164960f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 165060f5a4bcSLai Jiangshan * 16514736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 16524736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 16534736cbf7SLai Jiangshan * other reference to the pool. 165460f5a4bcSLai Jiangshan */ 165560f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 165660f5a4bcSLai Jiangshan struct worker_pool *pool) 165760f5a4bcSLai Jiangshan { 165860f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 165960f5a4bcSLai Jiangshan 166092f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1661da028469SLai Jiangshan list_del(&worker->node); 1662da028469SLai Jiangshan if (list_empty(&pool->workers)) 166360f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 166492f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 166560f5a4bcSLai Jiangshan 1666b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1667b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1668b62c0751SLai Jiangshan 166960f5a4bcSLai Jiangshan if (detach_completion) 167060f5a4bcSLai Jiangshan complete(detach_completion); 167160f5a4bcSLai Jiangshan } 167260f5a4bcSLai Jiangshan 167360f5a4bcSLai Jiangshan /** 1674c34056a3STejun Heo * create_worker - create a new workqueue worker 167563d95a91STejun Heo * @pool: pool the new worker will belong to 1676c34056a3STejun Heo * 167773eb7fe7SLai Jiangshan * Create a new worker which is attached to @pool. The new worker must be 167873eb7fe7SLai Jiangshan * started by start_worker(). 1679c34056a3STejun Heo * 1680c34056a3STejun Heo * CONTEXT: 1681c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1682c34056a3STejun Heo * 1683d185af30SYacine Belkadi * Return: 1684c34056a3STejun Heo * Pointer to the newly created worker. 1685c34056a3STejun Heo */ 1686bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1687c34056a3STejun Heo { 1688c34056a3STejun Heo struct worker *worker = NULL; 1689f3421797STejun Heo int id = -1; 1690e3c916a4STejun Heo char id_buf[16]; 1691c34056a3STejun Heo 16927cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 16937cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1694822d8405STejun Heo if (id < 0) 1695c34056a3STejun Heo goto fail; 1696c34056a3STejun Heo 1697f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1698c34056a3STejun Heo if (!worker) 1699c34056a3STejun Heo goto fail; 1700c34056a3STejun Heo 1701bd7bdd43STejun Heo worker->pool = pool; 1702c34056a3STejun Heo worker->id = id; 1703c34056a3STejun Heo 170429c91e99STejun Heo if (pool->cpu >= 0) 1705e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1706e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1707f3421797STejun Heo else 1708e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1709e3c916a4STejun Heo 1710f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1711e3c916a4STejun Heo "kworker/%s", id_buf); 1712c34056a3STejun Heo if (IS_ERR(worker->task)) 1713c34056a3STejun Heo goto fail; 1714c34056a3STejun Heo 171591151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 171691151228SOleg Nesterov 171791151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 171891151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 171991151228SOleg Nesterov 1720da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17214736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1722822d8405STejun Heo 1723c34056a3STejun Heo return worker; 1724822d8405STejun Heo 1725c34056a3STejun Heo fail: 17269625ab17SLai Jiangshan if (id >= 0) 17277cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1728c34056a3STejun Heo kfree(worker); 1729c34056a3STejun Heo return NULL; 1730c34056a3STejun Heo } 1731c34056a3STejun Heo 1732c34056a3STejun Heo /** 1733c34056a3STejun Heo * start_worker - start a newly created worker 1734c34056a3STejun Heo * @worker: worker to start 1735c34056a3STejun Heo * 1736706026c2STejun Heo * Make the pool aware of @worker and start it. 1737c34056a3STejun Heo * 1738c34056a3STejun Heo * CONTEXT: 1739d565ed63STejun Heo * spin_lock_irq(pool->lock). 1740c34056a3STejun Heo */ 1741c34056a3STejun Heo static void start_worker(struct worker *worker) 1742c34056a3STejun Heo { 1743bd7bdd43STejun Heo worker->pool->nr_workers++; 1744c8e55f36STejun Heo worker_enter_idle(worker); 1745c34056a3STejun Heo wake_up_process(worker->task); 1746c34056a3STejun Heo } 1747c34056a3STejun Heo 1748c34056a3STejun Heo /** 1749ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1750ebf44d16STejun Heo * @pool: the target pool 1751ebf44d16STejun Heo * 1752cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1753d185af30SYacine Belkadi * 1754d185af30SYacine Belkadi * Return: 0 on success. A negative error code otherwise. 1755ebf44d16STejun Heo */ 1756ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1757ebf44d16STejun Heo { 1758ebf44d16STejun Heo struct worker *worker; 1759ebf44d16STejun Heo 1760ebf44d16STejun Heo worker = create_worker(pool); 1761ebf44d16STejun Heo if (worker) { 1762ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1763ebf44d16STejun Heo start_worker(worker); 1764ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1765ebf44d16STejun Heo } 1766ebf44d16STejun Heo 1767ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1768ebf44d16STejun Heo } 1769ebf44d16STejun Heo 1770ebf44d16STejun Heo /** 1771c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1772c34056a3STejun Heo * @worker: worker to be destroyed 1773c34056a3STejun Heo * 177473eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 177573eb7fe7SLai Jiangshan * be idle. 1776c8e55f36STejun Heo * 1777c8e55f36STejun Heo * CONTEXT: 177860f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1779c34056a3STejun Heo */ 1780c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1781c34056a3STejun Heo { 1782bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1783c34056a3STejun Heo 1784cd549687STejun Heo lockdep_assert_held(&pool->lock); 1785cd549687STejun Heo 1786c34056a3STejun Heo /* sanity check frenzy */ 17876183c009STejun Heo if (WARN_ON(worker->current_work) || 178873eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 178973eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 17906183c009STejun Heo return; 1791c34056a3STejun Heo 1792bd7bdd43STejun Heo pool->nr_workers--; 1793bd7bdd43STejun Heo pool->nr_idle--; 1794c8e55f36STejun Heo 1795c8e55f36STejun Heo list_del_init(&worker->entry); 1796cb444766STejun Heo worker->flags |= WORKER_DIE; 179760f5a4bcSLai Jiangshan wake_up_process(worker->task); 1798c34056a3STejun Heo } 1799c34056a3STejun Heo 180063d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1801e22bee78STejun Heo { 180263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1803e22bee78STejun Heo 1804d565ed63STejun Heo spin_lock_irq(&pool->lock); 1805e22bee78STejun Heo 18063347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1807e22bee78STejun Heo struct worker *worker; 1808e22bee78STejun Heo unsigned long expires; 1809e22bee78STejun Heo 1810e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 181163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1812e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1813e22bee78STejun Heo 18143347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 181563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18163347fc9fSLai Jiangshan break; 1817e22bee78STejun Heo } 18183347fc9fSLai Jiangshan 18193347fc9fSLai Jiangshan destroy_worker(worker); 1820e22bee78STejun Heo } 1821e22bee78STejun Heo 1822d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1823e22bee78STejun Heo } 1824e22bee78STejun Heo 1825493a1724STejun Heo static void send_mayday(struct work_struct *work) 1826e22bee78STejun Heo { 1827112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1828112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1829493a1724STejun Heo 18302e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1831e22bee78STejun Heo 1832493008a8STejun Heo if (!wq->rescuer) 1833493a1724STejun Heo return; 1834e22bee78STejun Heo 1835e22bee78STejun Heo /* mayday mayday mayday */ 1836493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 183777668c8bSLai Jiangshan /* 183877668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 183977668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 184077668c8bSLai Jiangshan * rescuer is done with it. 184177668c8bSLai Jiangshan */ 184277668c8bSLai Jiangshan get_pwq(pwq); 1843493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1844e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1845493a1724STejun Heo } 1846e22bee78STejun Heo } 1847e22bee78STejun Heo 1848706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1849e22bee78STejun Heo { 185063d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1851e22bee78STejun Heo struct work_struct *work; 1852e22bee78STejun Heo 18532e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1854493a1724STejun Heo spin_lock(&pool->lock); 1855e22bee78STejun Heo 185663d95a91STejun Heo if (need_to_create_worker(pool)) { 1857e22bee78STejun Heo /* 1858e22bee78STejun Heo * We've been trying to create a new worker but 1859e22bee78STejun Heo * haven't been successful. We might be hitting an 1860e22bee78STejun Heo * allocation deadlock. Send distress signals to 1861e22bee78STejun Heo * rescuers. 1862e22bee78STejun Heo */ 186363d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1864e22bee78STejun Heo send_mayday(work); 1865e22bee78STejun Heo } 1866e22bee78STejun Heo 1867493a1724STejun Heo spin_unlock(&pool->lock); 18682e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1869e22bee78STejun Heo 187063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1871e22bee78STejun Heo } 1872e22bee78STejun Heo 1873e22bee78STejun Heo /** 1874e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 187563d95a91STejun Heo * @pool: pool to create a new worker for 1876e22bee78STejun Heo * 187763d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1878e22bee78STejun Heo * have at least one idle worker on return from this function. If 1879e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 188063d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1881e22bee78STejun Heo * possible allocation deadlock. 1882e22bee78STejun Heo * 1883c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1884c5aa87bbSTejun Heo * may_start_working() %true. 1885e22bee78STejun Heo * 1886e22bee78STejun Heo * LOCKING: 1887d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1888e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1889e22bee78STejun Heo * manager. 1890e22bee78STejun Heo * 1891d185af30SYacine Belkadi * Return: 1892c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1893e22bee78STejun Heo * otherwise. 1894e22bee78STejun Heo */ 189563d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1896d565ed63STejun Heo __releases(&pool->lock) 1897d565ed63STejun Heo __acquires(&pool->lock) 1898e22bee78STejun Heo { 189963d95a91STejun Heo if (!need_to_create_worker(pool)) 1900e22bee78STejun Heo return false; 1901e22bee78STejun Heo restart: 1902d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19039f9c2364STejun Heo 1904e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 190563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1906e22bee78STejun Heo 1907e22bee78STejun Heo while (true) { 1908e22bee78STejun Heo struct worker *worker; 1909e22bee78STejun Heo 1910bc2ae0f5STejun Heo worker = create_worker(pool); 1911e22bee78STejun Heo if (worker) { 191263d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1913d565ed63STejun Heo spin_lock_irq(&pool->lock); 1914e22bee78STejun Heo start_worker(worker); 19156183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19166183c009STejun Heo goto restart; 1917e22bee78STejun Heo return true; 1918e22bee78STejun Heo } 1919e22bee78STejun Heo 192063d95a91STejun Heo if (!need_to_create_worker(pool)) 1921e22bee78STejun Heo break; 1922e22bee78STejun Heo 1923e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19249f9c2364STejun Heo 192563d95a91STejun Heo if (!need_to_create_worker(pool)) 1926e22bee78STejun Heo break; 1927e22bee78STejun Heo } 1928e22bee78STejun Heo 192963d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1930d565ed63STejun Heo spin_lock_irq(&pool->lock); 193163d95a91STejun Heo if (need_to_create_worker(pool)) 1932e22bee78STejun Heo goto restart; 1933e22bee78STejun Heo return true; 1934e22bee78STejun Heo } 1935e22bee78STejun Heo 1936e22bee78STejun Heo /** 1937e22bee78STejun Heo * manage_workers - manage worker pool 1938e22bee78STejun Heo * @worker: self 1939e22bee78STejun Heo * 1940706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1941e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1942706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1943e22bee78STejun Heo * 1944e22bee78STejun Heo * The caller can safely start processing works on false return. On 1945e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1946e22bee78STejun Heo * and may_start_working() is true. 1947e22bee78STejun Heo * 1948e22bee78STejun Heo * CONTEXT: 1949d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1950e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1951e22bee78STejun Heo * 1952d185af30SYacine Belkadi * Return: 19532d498db9SLibin * %false if the pool don't need management and the caller can safely start 19542d498db9SLibin * processing works, %true indicates that the function released pool->lock 19552d498db9SLibin * and reacquired it to perform some management function and that the 19562d498db9SLibin * conditions that the caller verified while holding the lock before 19572d498db9SLibin * calling the function might no longer be true. 1958e22bee78STejun Heo */ 1959e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1960e22bee78STejun Heo { 196163d95a91STejun Heo struct worker_pool *pool = worker->pool; 1962e22bee78STejun Heo bool ret = false; 1963e22bee78STejun Heo 1964bc3a1afcSTejun Heo /* 1965bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1966bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1967bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1968bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1969bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1970bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1971bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1972bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1973bc3a1afcSTejun Heo */ 197434a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 1975e22bee78STejun Heo return ret; 1976e22bee78STejun Heo 197763d95a91STejun Heo ret |= maybe_create_worker(pool); 1978e22bee78STejun Heo 197934a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 1980e22bee78STejun Heo return ret; 1981e22bee78STejun Heo } 1982e22bee78STejun Heo 1983a62428c0STejun Heo /** 1984a62428c0STejun Heo * process_one_work - process single work 1985c34056a3STejun Heo * @worker: self 1986a62428c0STejun Heo * @work: work to process 1987a62428c0STejun Heo * 1988a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1989a62428c0STejun Heo * process a single work including synchronization against and 1990a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1991a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1992a62428c0STejun Heo * call this function to process a work. 1993a62428c0STejun Heo * 1994a62428c0STejun Heo * CONTEXT: 1995d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1996a62428c0STejun Heo */ 1997c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 1998d565ed63STejun Heo __releases(&pool->lock) 1999d565ed63STejun Heo __acquires(&pool->lock) 20001da177e4SLinus Torvalds { 2001112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2002bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2003112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 200473f53c4aSTejun Heo int work_color; 20057e11629dSTejun Heo struct worker *collision; 20064e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20074e6045f1SJohannes Berg /* 2008a62428c0STejun Heo * It is permissible to free the struct work_struct from 2009a62428c0STejun Heo * inside the function that is called from it, this we need to 2010a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2011a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2012a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20134e6045f1SJohannes Berg */ 20144d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20154d82a1deSPeter Zijlstra 20164d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20174e6045f1SJohannes Berg #endif 201885327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2019ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 202025511a47STejun Heo 20217e11629dSTejun Heo /* 20227e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20237e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20247e11629dSTejun Heo * already processing the work. If so, defer the work to the 20257e11629dSTejun Heo * currently executing one. 20267e11629dSTejun Heo */ 2027c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20287e11629dSTejun Heo if (unlikely(collision)) { 20297e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20307e11629dSTejun Heo return; 20317e11629dSTejun Heo } 20321da177e4SLinus Torvalds 20338930cabaSTejun Heo /* claim and dequeue */ 20341da177e4SLinus Torvalds debug_work_deactivate(work); 2035c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2036c34056a3STejun Heo worker->current_work = work; 2037a2c1c57bSTejun Heo worker->current_func = work->func; 2038112202d9STejun Heo worker->current_pwq = pwq; 203973f53c4aSTejun Heo work_color = get_work_color(work); 20407a22ad75STejun Heo 2041a62428c0STejun Heo list_del_init(&work->entry); 2042a62428c0STejun Heo 2043649027d7STejun Heo /* 2044fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2045fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2046fb0e7bebSTejun Heo */ 2047fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2048fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2049fb0e7bebSTejun Heo 2050974271c4STejun Heo /* 2051d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2052974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2053974271c4STejun Heo */ 205463d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 205563d95a91STejun Heo wake_up_worker(pool); 2056974271c4STejun Heo 20578930cabaSTejun Heo /* 20587c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2059d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 206023657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 206123657bb1STejun Heo * disabled. 20628930cabaSTejun Heo */ 20637c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20641da177e4SLinus Torvalds 2065d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2066365970a1SDavid Howells 2067112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20683295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2069e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2070a2c1c57bSTejun Heo worker->current_func(work); 2071e36c886aSArjan van de Ven /* 2072e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2073e36c886aSArjan van de Ven * point will only record its address. 2074e36c886aSArjan van de Ven */ 2075e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20763295f0efSIngo Molnar lock_map_release(&lockdep_map); 2077112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 20781da177e4SLinus Torvalds 2079d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2080044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2081044c782cSValentin Ilie " last function: %pf\n", 2082a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2083a2c1c57bSTejun Heo worker->current_func); 2084d5abe669SPeter Zijlstra debug_show_held_locks(current); 2085d5abe669SPeter Zijlstra dump_stack(); 2086d5abe669SPeter Zijlstra } 2087d5abe669SPeter Zijlstra 2088b22ce278STejun Heo /* 2089b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2090b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2091b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2092b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2093b22ce278STejun Heo * stop_machine. 2094b22ce278STejun Heo */ 2095b22ce278STejun Heo cond_resched(); 2096b22ce278STejun Heo 2097d565ed63STejun Heo spin_lock_irq(&pool->lock); 2098a62428c0STejun Heo 2099fb0e7bebSTejun Heo /* clear cpu intensive status */ 2100fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2101fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2102fb0e7bebSTejun Heo 2103a62428c0STejun Heo /* we're done with it, release */ 210442f8570fSSasha Levin hash_del(&worker->hentry); 2105c34056a3STejun Heo worker->current_work = NULL; 2106a2c1c57bSTejun Heo worker->current_func = NULL; 2107112202d9STejun Heo worker->current_pwq = NULL; 21083d1cb205STejun Heo worker->desc_valid = false; 2109112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21101da177e4SLinus Torvalds } 21111da177e4SLinus Torvalds 2112affee4b2STejun Heo /** 2113affee4b2STejun Heo * process_scheduled_works - process scheduled works 2114affee4b2STejun Heo * @worker: self 2115affee4b2STejun Heo * 2116affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2117affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2118affee4b2STejun Heo * fetches a work from the top and executes it. 2119affee4b2STejun Heo * 2120affee4b2STejun Heo * CONTEXT: 2121d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2122affee4b2STejun Heo * multiple times. 2123affee4b2STejun Heo */ 2124affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21251da177e4SLinus Torvalds { 2126affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2127affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2128a62428c0STejun Heo struct work_struct, entry); 2129c34056a3STejun Heo process_one_work(worker, work); 2130a62428c0STejun Heo } 21311da177e4SLinus Torvalds } 21321da177e4SLinus Torvalds 21334690c4abSTejun Heo /** 21344690c4abSTejun Heo * worker_thread - the worker thread function 2135c34056a3STejun Heo * @__worker: self 21364690c4abSTejun Heo * 2137c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2138c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2139c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2140c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2141c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2142d185af30SYacine Belkadi * 2143d185af30SYacine Belkadi * Return: 0 21444690c4abSTejun Heo */ 2145c34056a3STejun Heo static int worker_thread(void *__worker) 21461da177e4SLinus Torvalds { 2147c34056a3STejun Heo struct worker *worker = __worker; 2148bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21491da177e4SLinus Torvalds 2150e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2151e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2152c8e55f36STejun Heo woke_up: 2153d565ed63STejun Heo spin_lock_irq(&pool->lock); 2154affee4b2STejun Heo 2155a9ab775bSTejun Heo /* am I supposed to die? */ 2156a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2157d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2158a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2159e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 216060f5a4bcSLai Jiangshan 216160f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21627cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 216360f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 216460f5a4bcSLai Jiangshan kfree(worker); 2165c8e55f36STejun Heo return 0; 2166c8e55f36STejun Heo } 2167c8e55f36STejun Heo 2168c8e55f36STejun Heo worker_leave_idle(worker); 2169db7bccf4STejun Heo recheck: 2170e22bee78STejun Heo /* no more worker necessary? */ 217163d95a91STejun Heo if (!need_more_worker(pool)) 2172e22bee78STejun Heo goto sleep; 2173e22bee78STejun Heo 2174e22bee78STejun Heo /* do we need to manage? */ 217563d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2176e22bee78STejun Heo goto recheck; 2177e22bee78STejun Heo 2178c8e55f36STejun Heo /* 2179c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2180c8e55f36STejun Heo * preparing to process a work or actually processing it. 2181c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2182c8e55f36STejun Heo */ 21836183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2184c8e55f36STejun Heo 2185e22bee78STejun Heo /* 2186a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2187a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2188a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2189a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2190a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2191e22bee78STejun Heo */ 2192a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2193e22bee78STejun Heo 2194e22bee78STejun Heo do { 2195affee4b2STejun Heo struct work_struct *work = 2196bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2197affee4b2STejun Heo struct work_struct, entry); 2198affee4b2STejun Heo 2199c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2200affee4b2STejun Heo /* optimization path, not strictly necessary */ 2201affee4b2STejun Heo process_one_work(worker, work); 2202affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2203affee4b2STejun Heo process_scheduled_works(worker); 2204affee4b2STejun Heo } else { 2205c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2206affee4b2STejun Heo process_scheduled_works(worker); 2207affee4b2STejun Heo } 220863d95a91STejun Heo } while (keep_working(pool)); 2209affee4b2STejun Heo 2210e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2211d313dd85STejun Heo sleep: 2212c8e55f36STejun Heo /* 2213d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2214d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2215d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2216d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2217d565ed63STejun Heo * event. 2218c8e55f36STejun Heo */ 2219c8e55f36STejun Heo worker_enter_idle(worker); 2220c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2221d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22221da177e4SLinus Torvalds schedule(); 2223c8e55f36STejun Heo goto woke_up; 22241da177e4SLinus Torvalds } 22251da177e4SLinus Torvalds 2226e22bee78STejun Heo /** 2227e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2228111c225aSTejun Heo * @__rescuer: self 2229e22bee78STejun Heo * 2230e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2231493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2232e22bee78STejun Heo * 2233706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2234e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2235e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2236e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2237e22bee78STejun Heo * the problem rescuer solves. 2238e22bee78STejun Heo * 2239706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2240706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2241e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2242e22bee78STejun Heo * 2243e22bee78STejun Heo * This should happen rarely. 2244d185af30SYacine Belkadi * 2245d185af30SYacine Belkadi * Return: 0 2246e22bee78STejun Heo */ 2247111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2248e22bee78STejun Heo { 2249111c225aSTejun Heo struct worker *rescuer = __rescuer; 2250111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2251e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22524d595b86SLai Jiangshan bool should_stop; 2253e22bee78STejun Heo 2254e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2255111c225aSTejun Heo 2256111c225aSTejun Heo /* 2257111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2258111c225aSTejun Heo * doesn't participate in concurrency management. 2259111c225aSTejun Heo */ 2260111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2261e22bee78STejun Heo repeat: 2262e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22631da177e4SLinus Torvalds 22644d595b86SLai Jiangshan /* 22654d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22664d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22674d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22684d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22694d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22704d595b86SLai Jiangshan * list is always empty on exit. 22714d595b86SLai Jiangshan */ 22724d595b86SLai Jiangshan should_stop = kthread_should_stop(); 22731da177e4SLinus Torvalds 2274493a1724STejun Heo /* see whether any pwq is asking for help */ 22752e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2276493a1724STejun Heo 2277493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2278493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2279493a1724STejun Heo struct pool_workqueue, mayday_node); 2280112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2281e22bee78STejun Heo struct work_struct *work, *n; 2282e22bee78STejun Heo 2283e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2284493a1724STejun Heo list_del_init(&pwq->mayday_node); 2285493a1724STejun Heo 22862e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2287e22bee78STejun Heo 228851697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 228951697d39SLai Jiangshan 229051697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2291b3104104SLai Jiangshan rescuer->pool = pool; 2292e22bee78STejun Heo 2293e22bee78STejun Heo /* 2294e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2295e22bee78STejun Heo * process'em. 2296e22bee78STejun Heo */ 22976183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2298bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2299112202d9STejun Heo if (get_work_pwq(work) == pwq) 2300e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2301e22bee78STejun Heo 2302e22bee78STejun Heo process_scheduled_works(rescuer); 230351697d39SLai Jiangshan spin_unlock_irq(&pool->lock); 230451697d39SLai Jiangshan 230551697d39SLai Jiangshan worker_detach_from_pool(rescuer, pool); 230651697d39SLai Jiangshan 230751697d39SLai Jiangshan spin_lock_irq(&pool->lock); 23087576958aSTejun Heo 23097576958aSTejun Heo /* 231077668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 231177668c8bSLai Jiangshan * go away while we're holding its lock. 231277668c8bSLai Jiangshan */ 231377668c8bSLai Jiangshan put_pwq(pwq); 231477668c8bSLai Jiangshan 231577668c8bSLai Jiangshan /* 2316d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23177576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23187576958aSTejun Heo * and stalling the execution. 23197576958aSTejun Heo */ 232063d95a91STejun Heo if (keep_working(pool)) 232163d95a91STejun Heo wake_up_worker(pool); 23227576958aSTejun Heo 2323b3104104SLai Jiangshan rescuer->pool = NULL; 2324493a1724STejun Heo spin_unlock(&pool->lock); 23252e109a28STejun Heo spin_lock(&wq_mayday_lock); 23261da177e4SLinus Torvalds } 23271da177e4SLinus Torvalds 23282e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2329493a1724STejun Heo 23304d595b86SLai Jiangshan if (should_stop) { 23314d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23324d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23334d595b86SLai Jiangshan return 0; 23344d595b86SLai Jiangshan } 23354d595b86SLai Jiangshan 2336111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2337111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2338e22bee78STejun Heo schedule(); 2339e22bee78STejun Heo goto repeat; 23401da177e4SLinus Torvalds } 23411da177e4SLinus Torvalds 2342fc2e4d70SOleg Nesterov struct wq_barrier { 2343fc2e4d70SOleg Nesterov struct work_struct work; 2344fc2e4d70SOleg Nesterov struct completion done; 2345fc2e4d70SOleg Nesterov }; 2346fc2e4d70SOleg Nesterov 2347fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2348fc2e4d70SOleg Nesterov { 2349fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2350fc2e4d70SOleg Nesterov complete(&barr->done); 2351fc2e4d70SOleg Nesterov } 2352fc2e4d70SOleg Nesterov 23534690c4abSTejun Heo /** 23544690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2355112202d9STejun Heo * @pwq: pwq to insert barrier into 23564690c4abSTejun Heo * @barr: wq_barrier to insert 2357affee4b2STejun Heo * @target: target work to attach @barr to 2358affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23594690c4abSTejun Heo * 2360affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2361affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2362affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2363affee4b2STejun Heo * cpu. 2364affee4b2STejun Heo * 2365affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2366affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2367affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2368affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2369affee4b2STejun Heo * after a work with LINKED flag set. 2370affee4b2STejun Heo * 2371affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2372112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23734690c4abSTejun Heo * 23744690c4abSTejun Heo * CONTEXT: 2375d565ed63STejun Heo * spin_lock_irq(pool->lock). 23764690c4abSTejun Heo */ 2377112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2378affee4b2STejun Heo struct wq_barrier *barr, 2379affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2380fc2e4d70SOleg Nesterov { 2381affee4b2STejun Heo struct list_head *head; 2382affee4b2STejun Heo unsigned int linked = 0; 2383affee4b2STejun Heo 2384dc186ad7SThomas Gleixner /* 2385d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2386dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2387dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2388dc186ad7SThomas Gleixner * might deadlock. 2389dc186ad7SThomas Gleixner */ 2390ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 239122df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2392fc2e4d70SOleg Nesterov init_completion(&barr->done); 239383c22520SOleg Nesterov 2394affee4b2STejun Heo /* 2395affee4b2STejun Heo * If @target is currently being executed, schedule the 2396affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2397affee4b2STejun Heo */ 2398affee4b2STejun Heo if (worker) 2399affee4b2STejun Heo head = worker->scheduled.next; 2400affee4b2STejun Heo else { 2401affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2402affee4b2STejun Heo 2403affee4b2STejun Heo head = target->entry.next; 2404affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2405affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2406affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2407affee4b2STejun Heo } 2408affee4b2STejun Heo 2409dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2410112202d9STejun Heo insert_work(pwq, &barr->work, head, 2411affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2412fc2e4d70SOleg Nesterov } 2413fc2e4d70SOleg Nesterov 241473f53c4aSTejun Heo /** 2415112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 241673f53c4aSTejun Heo * @wq: workqueue being flushed 241773f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 241873f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 241973f53c4aSTejun Heo * 2420112202d9STejun Heo * Prepare pwqs for workqueue flushing. 242173f53c4aSTejun Heo * 2422112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2423112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2424112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2425112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2426112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 242773f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 242873f53c4aSTejun Heo * 242973f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 243073f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 243173f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 243273f53c4aSTejun Heo * is returned. 243373f53c4aSTejun Heo * 2434112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 243573f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 243673f53c4aSTejun Heo * advanced to @work_color. 243773f53c4aSTejun Heo * 243873f53c4aSTejun Heo * CONTEXT: 24393c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 244073f53c4aSTejun Heo * 2441d185af30SYacine Belkadi * Return: 244273f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 244373f53c4aSTejun Heo * otherwise. 244473f53c4aSTejun Heo */ 2445112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 244673f53c4aSTejun Heo int flush_color, int work_color) 24471da177e4SLinus Torvalds { 244873f53c4aSTejun Heo bool wait = false; 244949e3cf44STejun Heo struct pool_workqueue *pwq; 24501da177e4SLinus Torvalds 245173f53c4aSTejun Heo if (flush_color >= 0) { 24526183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2453112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2454dc186ad7SThomas Gleixner } 245514441960SOleg Nesterov 245649e3cf44STejun Heo for_each_pwq(pwq, wq) { 2457112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24581da177e4SLinus Torvalds 2459b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 246073f53c4aSTejun Heo 246173f53c4aSTejun Heo if (flush_color >= 0) { 24626183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 246373f53c4aSTejun Heo 2464112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2465112202d9STejun Heo pwq->flush_color = flush_color; 2466112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 246773f53c4aSTejun Heo wait = true; 24681da177e4SLinus Torvalds } 246973f53c4aSTejun Heo } 247073f53c4aSTejun Heo 247173f53c4aSTejun Heo if (work_color >= 0) { 24726183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2473112202d9STejun Heo pwq->work_color = work_color; 247473f53c4aSTejun Heo } 247573f53c4aSTejun Heo 2476b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 24771da177e4SLinus Torvalds } 24781da177e4SLinus Torvalds 2479112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 248073f53c4aSTejun Heo complete(&wq->first_flusher->done); 248173f53c4aSTejun Heo 248273f53c4aSTejun Heo return wait; 248383c22520SOleg Nesterov } 24841da177e4SLinus Torvalds 24850fcb78c2SRolf Eike Beer /** 24861da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 24870fcb78c2SRolf Eike Beer * @wq: workqueue to flush 24881da177e4SLinus Torvalds * 2489c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2490c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 24911da177e4SLinus Torvalds */ 24927ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 24931da177e4SLinus Torvalds { 249473f53c4aSTejun Heo struct wq_flusher this_flusher = { 249573f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 249673f53c4aSTejun Heo .flush_color = -1, 249773f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 249873f53c4aSTejun Heo }; 249973f53c4aSTejun Heo int next_color; 2500b1f4ec17SOleg Nesterov 25013295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25023295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 250373f53c4aSTejun Heo 25043c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 250573f53c4aSTejun Heo 250673f53c4aSTejun Heo /* 250773f53c4aSTejun Heo * Start-to-wait phase 250873f53c4aSTejun Heo */ 250973f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 251073f53c4aSTejun Heo 251173f53c4aSTejun Heo if (next_color != wq->flush_color) { 251273f53c4aSTejun Heo /* 251373f53c4aSTejun Heo * Color space is not full. The current work_color 251473f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 251573f53c4aSTejun Heo * by one. 251673f53c4aSTejun Heo */ 25176183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 251873f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 251973f53c4aSTejun Heo wq->work_color = next_color; 252073f53c4aSTejun Heo 252173f53c4aSTejun Heo if (!wq->first_flusher) { 252273f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25236183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 252473f53c4aSTejun Heo 252573f53c4aSTejun Heo wq->first_flusher = &this_flusher; 252673f53c4aSTejun Heo 2527112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 252873f53c4aSTejun Heo wq->work_color)) { 252973f53c4aSTejun Heo /* nothing to flush, done */ 253073f53c4aSTejun Heo wq->flush_color = next_color; 253173f53c4aSTejun Heo wq->first_flusher = NULL; 253273f53c4aSTejun Heo goto out_unlock; 253373f53c4aSTejun Heo } 253473f53c4aSTejun Heo } else { 253573f53c4aSTejun Heo /* wait in queue */ 25366183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 253773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2538112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 253973f53c4aSTejun Heo } 254073f53c4aSTejun Heo } else { 254173f53c4aSTejun Heo /* 254273f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 254373f53c4aSTejun Heo * The next flush completion will assign us 254473f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 254573f53c4aSTejun Heo */ 254673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 254773f53c4aSTejun Heo } 254873f53c4aSTejun Heo 25493c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 255073f53c4aSTejun Heo 255173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 255273f53c4aSTejun Heo 255373f53c4aSTejun Heo /* 255473f53c4aSTejun Heo * Wake-up-and-cascade phase 255573f53c4aSTejun Heo * 255673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 255773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 255873f53c4aSTejun Heo */ 255973f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 256073f53c4aSTejun Heo return; 256173f53c4aSTejun Heo 25623c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 256373f53c4aSTejun Heo 25644ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25654ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25664ce48b37STejun Heo goto out_unlock; 25674ce48b37STejun Heo 256873f53c4aSTejun Heo wq->first_flusher = NULL; 256973f53c4aSTejun Heo 25706183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25716183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 257273f53c4aSTejun Heo 257373f53c4aSTejun Heo while (true) { 257473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 257573f53c4aSTejun Heo 257673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 257773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 257873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 257973f53c4aSTejun Heo break; 258073f53c4aSTejun Heo list_del_init(&next->list); 258173f53c4aSTejun Heo complete(&next->done); 258273f53c4aSTejun Heo } 258373f53c4aSTejun Heo 25846183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 258573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 258873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 258973f53c4aSTejun Heo 259073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 259173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 259273f53c4aSTejun Heo /* 259373f53c4aSTejun Heo * Assign the same color to all overflowed 259473f53c4aSTejun Heo * flushers, advance work_color and append to 259573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 259673f53c4aSTejun Heo * phase for these overflowed flushers. 259773f53c4aSTejun Heo */ 259873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 259973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 260073f53c4aSTejun Heo 260173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 260273f53c4aSTejun Heo 260373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 260473f53c4aSTejun Heo &wq->flusher_queue); 2605112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 260673f53c4aSTejun Heo } 260773f53c4aSTejun Heo 260873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26096183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 261073f53c4aSTejun Heo break; 261173f53c4aSTejun Heo } 261273f53c4aSTejun Heo 261373f53c4aSTejun Heo /* 261473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2615112202d9STejun Heo * the new first flusher and arm pwqs. 261673f53c4aSTejun Heo */ 26176183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26186183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 261973f53c4aSTejun Heo 262073f53c4aSTejun Heo list_del_init(&next->list); 262173f53c4aSTejun Heo wq->first_flusher = next; 262273f53c4aSTejun Heo 2623112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 262473f53c4aSTejun Heo break; 262573f53c4aSTejun Heo 262673f53c4aSTejun Heo /* 262773f53c4aSTejun Heo * Meh... this color is already done, clear first 262873f53c4aSTejun Heo * flusher and repeat cascading. 262973f53c4aSTejun Heo */ 263073f53c4aSTejun Heo wq->first_flusher = NULL; 263173f53c4aSTejun Heo } 263273f53c4aSTejun Heo 263373f53c4aSTejun Heo out_unlock: 26343c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26351da177e4SLinus Torvalds } 2636ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26371da177e4SLinus Torvalds 26389c5a2ba7STejun Heo /** 26399c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26409c5a2ba7STejun Heo * @wq: workqueue to drain 26419c5a2ba7STejun Heo * 26429c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26439c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26449c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26459c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26469c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26479c5a2ba7STejun Heo * takes too long. 26489c5a2ba7STejun Heo */ 26499c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26509c5a2ba7STejun Heo { 26519c5a2ba7STejun Heo unsigned int flush_cnt = 0; 265249e3cf44STejun Heo struct pool_workqueue *pwq; 26539c5a2ba7STejun Heo 26549c5a2ba7STejun Heo /* 26559c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26569c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2657618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26589c5a2ba7STejun Heo */ 265987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26609c5a2ba7STejun Heo if (!wq->nr_drainers++) 2661618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 266287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26639c5a2ba7STejun Heo reflush: 26649c5a2ba7STejun Heo flush_workqueue(wq); 26659c5a2ba7STejun Heo 2666b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 266776af4d93STejun Heo 266849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2669fa2563e4SThomas Tuttle bool drained; 26709c5a2ba7STejun Heo 2671b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2672112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2673b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2674fa2563e4SThomas Tuttle 2675fa2563e4SThomas Tuttle if (drained) 26769c5a2ba7STejun Heo continue; 26779c5a2ba7STejun Heo 26789c5a2ba7STejun Heo if (++flush_cnt == 10 || 26799c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2680c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 26819c5a2ba7STejun Heo wq->name, flush_cnt); 268276af4d93STejun Heo 2683b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 26849c5a2ba7STejun Heo goto reflush; 26859c5a2ba7STejun Heo } 26869c5a2ba7STejun Heo 26879c5a2ba7STejun Heo if (!--wq->nr_drainers) 2688618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 268987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26909c5a2ba7STejun Heo } 26919c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 26929c5a2ba7STejun Heo 2693606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2694baf59022STejun Heo { 2695baf59022STejun Heo struct worker *worker = NULL; 2696c9e7cf27STejun Heo struct worker_pool *pool; 2697112202d9STejun Heo struct pool_workqueue *pwq; 2698baf59022STejun Heo 2699baf59022STejun Heo might_sleep(); 2700baf59022STejun Heo 2701fa1b54e6STejun Heo local_irq_disable(); 2702fa1b54e6STejun Heo pool = get_work_pool(work); 2703fa1b54e6STejun Heo if (!pool) { 2704fa1b54e6STejun Heo local_irq_enable(); 2705fa1b54e6STejun Heo return false; 2706fa1b54e6STejun Heo } 2707fa1b54e6STejun Heo 2708fa1b54e6STejun Heo spin_lock(&pool->lock); 27090b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2710112202d9STejun Heo pwq = get_work_pwq(work); 2711112202d9STejun Heo if (pwq) { 2712112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2713baf59022STejun Heo goto already_gone; 2714606a5020STejun Heo } else { 2715c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2716baf59022STejun Heo if (!worker) 2717baf59022STejun Heo goto already_gone; 2718112202d9STejun Heo pwq = worker->current_pwq; 2719606a5020STejun Heo } 2720baf59022STejun Heo 2721112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2722d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2723baf59022STejun Heo 2724e159489bSTejun Heo /* 2725e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2726e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2727e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2728e159489bSTejun Heo * access. 2729e159489bSTejun Heo */ 2730493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2731112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2732e159489bSTejun Heo else 2733112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2734112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2735e159489bSTejun Heo 2736baf59022STejun Heo return true; 2737baf59022STejun Heo already_gone: 2738d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2739baf59022STejun Heo return false; 2740baf59022STejun Heo } 2741baf59022STejun Heo 2742db700897SOleg Nesterov /** 2743401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2744401a8d04STejun Heo * @work: the work to flush 2745db700897SOleg Nesterov * 2746606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2747606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2748401a8d04STejun Heo * 2749d185af30SYacine Belkadi * Return: 2750401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2751401a8d04STejun Heo * %false if it was already idle. 2752db700897SOleg Nesterov */ 2753401a8d04STejun Heo bool flush_work(struct work_struct *work) 2754db700897SOleg Nesterov { 275512997d1aSBjorn Helgaas struct wq_barrier barr; 275612997d1aSBjorn Helgaas 27570976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27580976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27590976dfc1SStephen Boyd 276012997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 276112997d1aSBjorn Helgaas wait_for_completion(&barr.done); 276212997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 276312997d1aSBjorn Helgaas return true; 276412997d1aSBjorn Helgaas } else { 276512997d1aSBjorn Helgaas return false; 276612997d1aSBjorn Helgaas } 2767606a5020STejun Heo } 2768db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2769db700897SOleg Nesterov 277036e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2771401a8d04STejun Heo { 2772bbb68dfaSTejun Heo unsigned long flags; 27731f1f642eSOleg Nesterov int ret; 27741f1f642eSOleg Nesterov 27751f1f642eSOleg Nesterov do { 2776bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2777bbb68dfaSTejun Heo /* 2778bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2779bbb68dfaSTejun Heo * would be waiting for before retrying. 2780bbb68dfaSTejun Heo */ 2781bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2782606a5020STejun Heo flush_work(work); 27831f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 27841f1f642eSOleg Nesterov 2785bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2786bbb68dfaSTejun Heo mark_work_canceling(work); 2787bbb68dfaSTejun Heo local_irq_restore(flags); 2788bbb68dfaSTejun Heo 2789606a5020STejun Heo flush_work(work); 27907a22ad75STejun Heo clear_work_data(work); 27911f1f642eSOleg Nesterov return ret; 27921f1f642eSOleg Nesterov } 27931f1f642eSOleg Nesterov 27946e84d644SOleg Nesterov /** 2795401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2796401a8d04STejun Heo * @work: the work to cancel 27976e84d644SOleg Nesterov * 2798401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2799401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2800401a8d04STejun Heo * another workqueue. On return from this function, @work is 2801401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28021f1f642eSOleg Nesterov * 2803401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2804401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28056e84d644SOleg Nesterov * 2806401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28076e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2808401a8d04STejun Heo * 2809d185af30SYacine Belkadi * Return: 2810401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28116e84d644SOleg Nesterov */ 2812401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28136e84d644SOleg Nesterov { 281436e227d2STejun Heo return __cancel_work_timer(work, false); 2815b89deed3SOleg Nesterov } 281628e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2817b89deed3SOleg Nesterov 28186e84d644SOleg Nesterov /** 2819401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2820401a8d04STejun Heo * @dwork: the delayed work to flush 28216e84d644SOleg Nesterov * 2822401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2823401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2824401a8d04STejun Heo * considers the last queueing instance of @dwork. 28251f1f642eSOleg Nesterov * 2826d185af30SYacine Belkadi * Return: 2827401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2828401a8d04STejun Heo * %false if it was already idle. 28296e84d644SOleg Nesterov */ 2830401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2831401a8d04STejun Heo { 28328930cabaSTejun Heo local_irq_disable(); 2833401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 283460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28358930cabaSTejun Heo local_irq_enable(); 2836401a8d04STejun Heo return flush_work(&dwork->work); 2837401a8d04STejun Heo } 2838401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2839401a8d04STejun Heo 2840401a8d04STejun Heo /** 284157b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 284257b30ae7STejun Heo * @dwork: delayed_work to cancel 284309383498STejun Heo * 2844d185af30SYacine Belkadi * Kill off a pending delayed_work. 2845d185af30SYacine Belkadi * 2846d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2847d185af30SYacine Belkadi * pending. 2848d185af30SYacine Belkadi * 2849d185af30SYacine Belkadi * Note: 2850d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2851d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2852d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 285309383498STejun Heo * 285457b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 285509383498STejun Heo */ 285657b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 285709383498STejun Heo { 285857b30ae7STejun Heo unsigned long flags; 285957b30ae7STejun Heo int ret; 286057b30ae7STejun Heo 286157b30ae7STejun Heo do { 286257b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 286357b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 286457b30ae7STejun Heo 286557b30ae7STejun Heo if (unlikely(ret < 0)) 286657b30ae7STejun Heo return false; 286757b30ae7STejun Heo 28687c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28697c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 287057b30ae7STejun Heo local_irq_restore(flags); 2871c0158ca6SDan Magenheimer return ret; 287209383498STejun Heo } 287357b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 287409383498STejun Heo 287509383498STejun Heo /** 2876401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2877401a8d04STejun Heo * @dwork: the delayed work cancel 2878401a8d04STejun Heo * 2879401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2880401a8d04STejun Heo * 2881d185af30SYacine Belkadi * Return: 2882401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2883401a8d04STejun Heo */ 2884401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 28856e84d644SOleg Nesterov { 288636e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 28876e84d644SOleg Nesterov } 2888f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 28891da177e4SLinus Torvalds 28900fcb78c2SRolf Eike Beer /** 289131ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2892b6136773SAndrew Morton * @func: the function to call 2893b6136773SAndrew Morton * 289431ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 289531ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2896b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 289731ddd871STejun Heo * 2898d185af30SYacine Belkadi * Return: 289931ddd871STejun Heo * 0 on success, -errno on failure. 2900b6136773SAndrew Morton */ 290165f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 290215316ba8SChristoph Lameter { 290315316ba8SChristoph Lameter int cpu; 290438f51568SNamhyung Kim struct work_struct __percpu *works; 290515316ba8SChristoph Lameter 2906b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2907b6136773SAndrew Morton if (!works) 290815316ba8SChristoph Lameter return -ENOMEM; 2909b6136773SAndrew Morton 291095402b38SGautham R Shenoy get_online_cpus(); 291193981800STejun Heo 291215316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29139bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29149bfb1839SIngo Molnar 29159bfb1839SIngo Molnar INIT_WORK(work, func); 29168de6d308SOleg Nesterov schedule_work_on(cpu, work); 291715316ba8SChristoph Lameter } 291893981800STejun Heo 291993981800STejun Heo for_each_online_cpu(cpu) 29208616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 292193981800STejun Heo 292295402b38SGautham R Shenoy put_online_cpus(); 2923b6136773SAndrew Morton free_percpu(works); 292415316ba8SChristoph Lameter return 0; 292515316ba8SChristoph Lameter } 292615316ba8SChristoph Lameter 2927eef6a7d5SAlan Stern /** 2928eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2929eef6a7d5SAlan Stern * 2930eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2931eef6a7d5SAlan Stern * completion. 2932eef6a7d5SAlan Stern * 2933eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2934eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2935eef6a7d5SAlan Stern * will lead to deadlock: 2936eef6a7d5SAlan Stern * 2937eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2938eef6a7d5SAlan Stern * a lock held by your code or its caller. 2939eef6a7d5SAlan Stern * 2940eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2941eef6a7d5SAlan Stern * 2942eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2943eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2944eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2945eef6a7d5SAlan Stern * 2946eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2947eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2948eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2949eef6a7d5SAlan Stern * cancel_work_sync() instead. 2950eef6a7d5SAlan Stern */ 29511da177e4SLinus Torvalds void flush_scheduled_work(void) 29521da177e4SLinus Torvalds { 2953d320c038STejun Heo flush_workqueue(system_wq); 29541da177e4SLinus Torvalds } 2955ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29561da177e4SLinus Torvalds 29571da177e4SLinus Torvalds /** 29581fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29591fa44ecaSJames Bottomley * @fn: the function to execute 29601fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29611fa44ecaSJames Bottomley * be available when the work executes) 29621fa44ecaSJames Bottomley * 29631fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29641fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29651fa44ecaSJames Bottomley * 2966d185af30SYacine Belkadi * Return: 0 - function was executed 29671fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29681fa44ecaSJames Bottomley */ 296965f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29701fa44ecaSJames Bottomley { 29711fa44ecaSJames Bottomley if (!in_interrupt()) { 297265f27f38SDavid Howells fn(&ew->work); 29731fa44ecaSJames Bottomley return 0; 29741fa44ecaSJames Bottomley } 29751fa44ecaSJames Bottomley 297665f27f38SDavid Howells INIT_WORK(&ew->work, fn); 29771fa44ecaSJames Bottomley schedule_work(&ew->work); 29781fa44ecaSJames Bottomley 29791fa44ecaSJames Bottomley return 1; 29801fa44ecaSJames Bottomley } 29811fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 29821fa44ecaSJames Bottomley 2983226223abSTejun Heo #ifdef CONFIG_SYSFS 2984226223abSTejun Heo /* 2985226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 2986226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 2987226223abSTejun Heo * following attributes. 2988226223abSTejun Heo * 2989226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 2990226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 2991226223abSTejun Heo * 2992226223abSTejun Heo * Unbound workqueues have the following extra attributes. 2993226223abSTejun Heo * 2994226223abSTejun Heo * id RO int : the associated pool ID 2995226223abSTejun Heo * nice RW int : nice value of the workers 2996226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 2997226223abSTejun Heo */ 2998226223abSTejun Heo struct wq_device { 2999226223abSTejun Heo struct workqueue_struct *wq; 3000226223abSTejun Heo struct device dev; 3001226223abSTejun Heo }; 3002226223abSTejun Heo 3003226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3004226223abSTejun Heo { 3005226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3006226223abSTejun Heo 3007226223abSTejun Heo return wq_dev->wq; 3008226223abSTejun Heo } 3009226223abSTejun Heo 30101a6661daSGreg Kroah-Hartman static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 30111a6661daSGreg Kroah-Hartman char *buf) 3012226223abSTejun Heo { 3013226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3014226223abSTejun Heo 3015226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3016226223abSTejun Heo } 30171a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RO(per_cpu); 3018226223abSTejun Heo 30191a6661daSGreg Kroah-Hartman static ssize_t max_active_show(struct device *dev, 3020226223abSTejun Heo struct device_attribute *attr, char *buf) 3021226223abSTejun Heo { 3022226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3023226223abSTejun Heo 3024226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3025226223abSTejun Heo } 3026226223abSTejun Heo 30271a6661daSGreg Kroah-Hartman static ssize_t max_active_store(struct device *dev, 30281a6661daSGreg Kroah-Hartman struct device_attribute *attr, const char *buf, 30291a6661daSGreg Kroah-Hartman size_t count) 3030226223abSTejun Heo { 3031226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3032226223abSTejun Heo int val; 3033226223abSTejun Heo 3034226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3035226223abSTejun Heo return -EINVAL; 3036226223abSTejun Heo 3037226223abSTejun Heo workqueue_set_max_active(wq, val); 3038226223abSTejun Heo return count; 3039226223abSTejun Heo } 30401a6661daSGreg Kroah-Hartman static DEVICE_ATTR_RW(max_active); 3041226223abSTejun Heo 30421a6661daSGreg Kroah-Hartman static struct attribute *wq_sysfs_attrs[] = { 30431a6661daSGreg Kroah-Hartman &dev_attr_per_cpu.attr, 30441a6661daSGreg Kroah-Hartman &dev_attr_max_active.attr, 30451a6661daSGreg Kroah-Hartman NULL, 3046226223abSTejun Heo }; 30471a6661daSGreg Kroah-Hartman ATTRIBUTE_GROUPS(wq_sysfs); 3048226223abSTejun Heo 3049d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3050226223abSTejun Heo struct device_attribute *attr, char *buf) 3051226223abSTejun Heo { 3052226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3053d55262c4STejun Heo const char *delim = ""; 3054d55262c4STejun Heo int node, written = 0; 3055226223abSTejun Heo 3056226223abSTejun Heo rcu_read_lock_sched(); 3057d55262c4STejun Heo for_each_node(node) { 3058d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3059d55262c4STejun Heo "%s%d:%d", delim, node, 3060d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3061d55262c4STejun Heo delim = " "; 3062d55262c4STejun Heo } 3063d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3064226223abSTejun Heo rcu_read_unlock_sched(); 3065226223abSTejun Heo 3066226223abSTejun Heo return written; 3067226223abSTejun Heo } 3068226223abSTejun Heo 3069226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3070226223abSTejun Heo char *buf) 3071226223abSTejun Heo { 3072226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3073226223abSTejun Heo int written; 3074226223abSTejun Heo 30756029a918STejun Heo mutex_lock(&wq->mutex); 30766029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 30776029a918STejun Heo mutex_unlock(&wq->mutex); 3078226223abSTejun Heo 3079226223abSTejun Heo return written; 3080226223abSTejun Heo } 3081226223abSTejun Heo 3082226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3083226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3084226223abSTejun Heo { 3085226223abSTejun Heo struct workqueue_attrs *attrs; 3086226223abSTejun Heo 3087226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3088226223abSTejun Heo if (!attrs) 3089226223abSTejun Heo return NULL; 3090226223abSTejun Heo 30916029a918STejun Heo mutex_lock(&wq->mutex); 30926029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 30936029a918STejun Heo mutex_unlock(&wq->mutex); 3094226223abSTejun Heo return attrs; 3095226223abSTejun Heo } 3096226223abSTejun Heo 3097226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3098226223abSTejun Heo const char *buf, size_t count) 3099226223abSTejun Heo { 3100226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3101226223abSTejun Heo struct workqueue_attrs *attrs; 3102226223abSTejun Heo int ret; 3103226223abSTejun Heo 3104226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3105226223abSTejun Heo if (!attrs) 3106226223abSTejun Heo return -ENOMEM; 3107226223abSTejun Heo 3108226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 310914481842SDongsheng Yang attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 3110226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3111226223abSTejun Heo else 3112226223abSTejun Heo ret = -EINVAL; 3113226223abSTejun Heo 3114226223abSTejun Heo free_workqueue_attrs(attrs); 3115226223abSTejun Heo return ret ?: count; 3116226223abSTejun Heo } 3117226223abSTejun Heo 3118226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3119226223abSTejun Heo struct device_attribute *attr, char *buf) 3120226223abSTejun Heo { 3121226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3122226223abSTejun Heo int written; 3123226223abSTejun Heo 31246029a918STejun Heo mutex_lock(&wq->mutex); 31256029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 31266029a918STejun Heo mutex_unlock(&wq->mutex); 3127226223abSTejun Heo 3128226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3129226223abSTejun Heo return written; 3130226223abSTejun Heo } 3131226223abSTejun Heo 3132226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3133226223abSTejun Heo struct device_attribute *attr, 3134226223abSTejun Heo const char *buf, size_t count) 3135226223abSTejun Heo { 3136226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3137226223abSTejun Heo struct workqueue_attrs *attrs; 3138226223abSTejun Heo int ret; 3139226223abSTejun Heo 3140226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3141226223abSTejun Heo if (!attrs) 3142226223abSTejun Heo return -ENOMEM; 3143226223abSTejun Heo 3144226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3145226223abSTejun Heo if (!ret) 3146226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3147226223abSTejun Heo 3148226223abSTejun Heo free_workqueue_attrs(attrs); 3149226223abSTejun Heo return ret ?: count; 3150226223abSTejun Heo } 3151226223abSTejun Heo 3152d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3153d55262c4STejun Heo char *buf) 3154d55262c4STejun Heo { 3155d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3156d55262c4STejun Heo int written; 3157d55262c4STejun Heo 3158d55262c4STejun Heo mutex_lock(&wq->mutex); 3159d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3160d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3161d55262c4STejun Heo mutex_unlock(&wq->mutex); 3162d55262c4STejun Heo 3163d55262c4STejun Heo return written; 3164d55262c4STejun Heo } 3165d55262c4STejun Heo 3166d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3167d55262c4STejun Heo const char *buf, size_t count) 3168d55262c4STejun Heo { 3169d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3170d55262c4STejun Heo struct workqueue_attrs *attrs; 3171d55262c4STejun Heo int v, ret; 3172d55262c4STejun Heo 3173d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3174d55262c4STejun Heo if (!attrs) 3175d55262c4STejun Heo return -ENOMEM; 3176d55262c4STejun Heo 3177d55262c4STejun Heo ret = -EINVAL; 3178d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3179d55262c4STejun Heo attrs->no_numa = !v; 3180d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3181d55262c4STejun Heo } 3182d55262c4STejun Heo 3183d55262c4STejun Heo free_workqueue_attrs(attrs); 3184d55262c4STejun Heo return ret ?: count; 3185d55262c4STejun Heo } 3186d55262c4STejun Heo 3187226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3188d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3189226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3190226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3191d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3192226223abSTejun Heo __ATTR_NULL, 3193226223abSTejun Heo }; 3194226223abSTejun Heo 3195226223abSTejun Heo static struct bus_type wq_subsys = { 3196226223abSTejun Heo .name = "workqueue", 31971a6661daSGreg Kroah-Hartman .dev_groups = wq_sysfs_groups, 3198226223abSTejun Heo }; 3199226223abSTejun Heo 3200226223abSTejun Heo static int __init wq_sysfs_init(void) 3201226223abSTejun Heo { 3202226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3203226223abSTejun Heo } 3204226223abSTejun Heo core_initcall(wq_sysfs_init); 3205226223abSTejun Heo 3206226223abSTejun Heo static void wq_device_release(struct device *dev) 3207226223abSTejun Heo { 3208226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3209226223abSTejun Heo 3210226223abSTejun Heo kfree(wq_dev); 3211226223abSTejun Heo } 3212226223abSTejun Heo 3213226223abSTejun Heo /** 3214226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3215226223abSTejun Heo * @wq: the workqueue to register 3216226223abSTejun Heo * 3217226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3218226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3219226223abSTejun Heo * which is the preferred method. 3220226223abSTejun Heo * 3221226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3222226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3223226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3224226223abSTejun Heo * attributes. 3225226223abSTejun Heo * 3226d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. 3227226223abSTejun Heo */ 3228226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3229226223abSTejun Heo { 3230226223abSTejun Heo struct wq_device *wq_dev; 3231226223abSTejun Heo int ret; 3232226223abSTejun Heo 3233226223abSTejun Heo /* 3234226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3235226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3236226223abSTejun Heo * workqueues. 3237226223abSTejun Heo */ 3238226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3239226223abSTejun Heo return -EINVAL; 3240226223abSTejun Heo 3241226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3242226223abSTejun Heo if (!wq_dev) 3243226223abSTejun Heo return -ENOMEM; 3244226223abSTejun Heo 3245226223abSTejun Heo wq_dev->wq = wq; 3246226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3247226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3248226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3249226223abSTejun Heo 3250226223abSTejun Heo /* 3251226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3252226223abSTejun Heo * everything is ready. 3253226223abSTejun Heo */ 3254226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3255226223abSTejun Heo 3256226223abSTejun Heo ret = device_register(&wq_dev->dev); 3257226223abSTejun Heo if (ret) { 3258226223abSTejun Heo kfree(wq_dev); 3259226223abSTejun Heo wq->wq_dev = NULL; 3260226223abSTejun Heo return ret; 3261226223abSTejun Heo } 3262226223abSTejun Heo 3263226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3264226223abSTejun Heo struct device_attribute *attr; 3265226223abSTejun Heo 3266226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3267226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3268226223abSTejun Heo if (ret) { 3269226223abSTejun Heo device_unregister(&wq_dev->dev); 3270226223abSTejun Heo wq->wq_dev = NULL; 3271226223abSTejun Heo return ret; 3272226223abSTejun Heo } 3273226223abSTejun Heo } 3274226223abSTejun Heo } 3275226223abSTejun Heo 3276226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3277226223abSTejun Heo return 0; 3278226223abSTejun Heo } 3279226223abSTejun Heo 3280226223abSTejun Heo /** 3281226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3282226223abSTejun Heo * @wq: the workqueue to unregister 3283226223abSTejun Heo * 3284226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3285226223abSTejun Heo */ 3286226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3287226223abSTejun Heo { 3288226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3289226223abSTejun Heo 3290226223abSTejun Heo if (!wq->wq_dev) 3291226223abSTejun Heo return; 3292226223abSTejun Heo 3293226223abSTejun Heo wq->wq_dev = NULL; 3294226223abSTejun Heo device_unregister(&wq_dev->dev); 3295226223abSTejun Heo } 3296226223abSTejun Heo #else /* CONFIG_SYSFS */ 3297226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3298226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3299226223abSTejun Heo 33007a4e344cSTejun Heo /** 33017a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33027a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33037a4e344cSTejun Heo * 33047a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33057a4e344cSTejun Heo */ 33067a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33077a4e344cSTejun Heo { 33087a4e344cSTejun Heo if (attrs) { 33097a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33107a4e344cSTejun Heo kfree(attrs); 33117a4e344cSTejun Heo } 33127a4e344cSTejun Heo } 33137a4e344cSTejun Heo 33147a4e344cSTejun Heo /** 33157a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33167a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33177a4e344cSTejun Heo * 33187a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3319d185af30SYacine Belkadi * return it. 3320d185af30SYacine Belkadi * 3321d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33227a4e344cSTejun Heo */ 33237a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33247a4e344cSTejun Heo { 33257a4e344cSTejun Heo struct workqueue_attrs *attrs; 33267a4e344cSTejun Heo 33277a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33287a4e344cSTejun Heo if (!attrs) 33297a4e344cSTejun Heo goto fail; 33307a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33317a4e344cSTejun Heo goto fail; 33327a4e344cSTejun Heo 333313e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33347a4e344cSTejun Heo return attrs; 33357a4e344cSTejun Heo fail: 33367a4e344cSTejun Heo free_workqueue_attrs(attrs); 33377a4e344cSTejun Heo return NULL; 33387a4e344cSTejun Heo } 33397a4e344cSTejun Heo 334029c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 334129c91e99STejun Heo const struct workqueue_attrs *from) 334229c91e99STejun Heo { 334329c91e99STejun Heo to->nice = from->nice; 334429c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 33452865a8fbSShaohua Li /* 33462865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 33472865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 33482865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 33492865a8fbSShaohua Li */ 33502865a8fbSShaohua Li to->no_numa = from->no_numa; 335129c91e99STejun Heo } 335229c91e99STejun Heo 335329c91e99STejun Heo /* hash value of the content of @attr */ 335429c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 335529c91e99STejun Heo { 335629c91e99STejun Heo u32 hash = 0; 335729c91e99STejun Heo 335829c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 335913e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 336013e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 336129c91e99STejun Heo return hash; 336229c91e99STejun Heo } 336329c91e99STejun Heo 336429c91e99STejun Heo /* content equality test */ 336529c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 336629c91e99STejun Heo const struct workqueue_attrs *b) 336729c91e99STejun Heo { 336829c91e99STejun Heo if (a->nice != b->nice) 336929c91e99STejun Heo return false; 337029c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 337129c91e99STejun Heo return false; 337229c91e99STejun Heo return true; 337329c91e99STejun Heo } 337429c91e99STejun Heo 33757a4e344cSTejun Heo /** 33767a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33777a4e344cSTejun Heo * @pool: worker_pool to initialize 33787a4e344cSTejun Heo * 33797a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 3380d185af30SYacine Belkadi * 3381d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 338229c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 338329c91e99STejun Heo * on @pool safely to release it. 33847a4e344cSTejun Heo */ 33857a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33864e1a1f9aSTejun Heo { 33874e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 338829c91e99STejun Heo pool->id = -1; 338929c91e99STejun Heo pool->cpu = -1; 3390f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 33914e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33924e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33934e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33944e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33954e1a1f9aSTejun Heo 33964e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33974e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33984e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33994e1a1f9aSTejun Heo 34004e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 34014e1a1f9aSTejun Heo (unsigned long)pool); 34024e1a1f9aSTejun Heo 34034e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 340492f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3405da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34067a4e344cSTejun Heo 34077cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 340829c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 340929c91e99STejun Heo pool->refcnt = 1; 341029c91e99STejun Heo 341129c91e99STejun Heo /* shouldn't fail above this point */ 34127a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34137a4e344cSTejun Heo if (!pool->attrs) 34147a4e344cSTejun Heo return -ENOMEM; 34157a4e344cSTejun Heo return 0; 34164e1a1f9aSTejun Heo } 34174e1a1f9aSTejun Heo 341829c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 341929c91e99STejun Heo { 342029c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 342129c91e99STejun Heo 34227cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 342329c91e99STejun Heo free_workqueue_attrs(pool->attrs); 342429c91e99STejun Heo kfree(pool); 342529c91e99STejun Heo } 342629c91e99STejun Heo 342729c91e99STejun Heo /** 342829c91e99STejun Heo * put_unbound_pool - put a worker_pool 342929c91e99STejun Heo * @pool: worker_pool to put 343029c91e99STejun Heo * 343129c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3432c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3433c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3434c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3435a892caccSTejun Heo * 3436a892caccSTejun Heo * Should be called with wq_pool_mutex held. 343729c91e99STejun Heo */ 343829c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 343929c91e99STejun Heo { 344060f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 344129c91e99STejun Heo struct worker *worker; 344229c91e99STejun Heo 3443a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3444a892caccSTejun Heo 3445a892caccSTejun Heo if (--pool->refcnt) 344629c91e99STejun Heo return; 344729c91e99STejun Heo 344829c91e99STejun Heo /* sanity checks */ 344961d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3450a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 345129c91e99STejun Heo return; 345229c91e99STejun Heo 345329c91e99STejun Heo /* release id and unhash */ 345429c91e99STejun Heo if (pool->id >= 0) 345529c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 345629c91e99STejun Heo hash_del(&pool->hash_node); 345729c91e99STejun Heo 3458c5aa87bbSTejun Heo /* 3459c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3460c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 346192f9c5c4SLai Jiangshan * attach_mutex. 3462c5aa87bbSTejun Heo */ 346329c91e99STejun Heo mutex_lock(&pool->manager_arb); 346429c91e99STejun Heo 346560f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 34661037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 346729c91e99STejun Heo destroy_worker(worker); 346829c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 346929c91e99STejun Heo spin_unlock_irq(&pool->lock); 347060f5a4bcSLai Jiangshan 347192f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3472da028469SLai Jiangshan if (!list_empty(&pool->workers)) 347360f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 347492f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 347560f5a4bcSLai Jiangshan 347660f5a4bcSLai Jiangshan if (pool->detach_completion) 347760f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 347860f5a4bcSLai Jiangshan 347929c91e99STejun Heo mutex_unlock(&pool->manager_arb); 348029c91e99STejun Heo 348129c91e99STejun Heo /* shut down the timers */ 348229c91e99STejun Heo del_timer_sync(&pool->idle_timer); 348329c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 348429c91e99STejun Heo 348529c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 348629c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 348729c91e99STejun Heo } 348829c91e99STejun Heo 348929c91e99STejun Heo /** 349029c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 349129c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 349229c91e99STejun Heo * 349329c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 349429c91e99STejun Heo * reference count and return it. If there already is a matching 349529c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3496d185af30SYacine Belkadi * create a new one. 3497a892caccSTejun Heo * 3498a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3499d185af30SYacine Belkadi * 3500d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3501d185af30SYacine Belkadi * On failure, %NULL. 350229c91e99STejun Heo */ 350329c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 350429c91e99STejun Heo { 350529c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 350629c91e99STejun Heo struct worker_pool *pool; 3507f3f90ad4STejun Heo int node; 350829c91e99STejun Heo 3509a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 351029c91e99STejun Heo 351129c91e99STejun Heo /* do we already have a matching pool? */ 351229c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 351329c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 351429c91e99STejun Heo pool->refcnt++; 351529c91e99STejun Heo goto out_unlock; 351629c91e99STejun Heo } 351729c91e99STejun Heo } 351829c91e99STejun Heo 351929c91e99STejun Heo /* nope, create a new one */ 352029c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 352129c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 352229c91e99STejun Heo goto fail; 352329c91e99STejun Heo 35248864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 352529c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 352629c91e99STejun Heo 35272865a8fbSShaohua Li /* 35282865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 35292865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 35302865a8fbSShaohua Li */ 35312865a8fbSShaohua Li pool->attrs->no_numa = false; 35322865a8fbSShaohua Li 3533f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3534f3f90ad4STejun Heo if (wq_numa_enabled) { 3535f3f90ad4STejun Heo for_each_node(node) { 3536f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3537f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3538f3f90ad4STejun Heo pool->node = node; 3539f3f90ad4STejun Heo break; 3540f3f90ad4STejun Heo } 3541f3f90ad4STejun Heo } 3542f3f90ad4STejun Heo } 3543f3f90ad4STejun Heo 354429c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 354529c91e99STejun Heo goto fail; 354629c91e99STejun Heo 354729c91e99STejun Heo /* create and start the initial worker */ 3548ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 354929c91e99STejun Heo goto fail; 355029c91e99STejun Heo 355129c91e99STejun Heo /* install */ 355229c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 355329c91e99STejun Heo out_unlock: 355429c91e99STejun Heo return pool; 355529c91e99STejun Heo fail: 355629c91e99STejun Heo if (pool) 355729c91e99STejun Heo put_unbound_pool(pool); 355829c91e99STejun Heo return NULL; 355929c91e99STejun Heo } 356029c91e99STejun Heo 35618864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35628864b4e5STejun Heo { 35638864b4e5STejun Heo kmem_cache_free(pwq_cache, 35648864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35658864b4e5STejun Heo } 35668864b4e5STejun Heo 35678864b4e5STejun Heo /* 35688864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35698864b4e5STejun Heo * and needs to be destroyed. 35708864b4e5STejun Heo */ 35718864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35728864b4e5STejun Heo { 35738864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35748864b4e5STejun Heo unbound_release_work); 35758864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35768864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3577bc0caf09STejun Heo bool is_last; 35788864b4e5STejun Heo 35798864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35808864b4e5STejun Heo return; 35818864b4e5STejun Heo 358275ccf595STejun Heo /* 35833c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 358475ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 358575ccf595STejun Heo * and consistent with the linking path. 358675ccf595STejun Heo */ 35873c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 35888864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3589bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35903c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35918864b4e5STejun Heo 3592a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35938864b4e5STejun Heo put_unbound_pool(pool); 3594a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3595a892caccSTejun Heo 35968864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35978864b4e5STejun Heo 35988864b4e5STejun Heo /* 35998864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 36008864b4e5STejun Heo * is gonna access it anymore. Free it. 36018864b4e5STejun Heo */ 36026029a918STejun Heo if (is_last) { 36036029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 36048864b4e5STejun Heo kfree(wq); 36058864b4e5STejun Heo } 36066029a918STejun Heo } 36078864b4e5STejun Heo 36080fbd95aaSTejun Heo /** 3609699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36100fbd95aaSTejun Heo * @pwq: target pool_workqueue 36110fbd95aaSTejun Heo * 3612699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3613699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3614699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36150fbd95aaSTejun Heo */ 3616699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36170fbd95aaSTejun Heo { 3618699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3619699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3620699ce097STejun Heo 3621699ce097STejun Heo /* for @wq->saved_max_active */ 3622a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3623699ce097STejun Heo 3624699ce097STejun Heo /* fast exit for non-freezable wqs */ 3625699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3626699ce097STejun Heo return; 3627699ce097STejun Heo 3628a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3629699ce097STejun Heo 363074b414eaSLai Jiangshan /* 363174b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 363274b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 363374b414eaSLai Jiangshan * is updated and visible. 363474b414eaSLai Jiangshan */ 363574b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3636699ce097STejun Heo pwq->max_active = wq->saved_max_active; 36370fbd95aaSTejun Heo 36380fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36390fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36400fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3641951a078aSLai Jiangshan 3642951a078aSLai Jiangshan /* 3643951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3644951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3645951a078aSLai Jiangshan */ 3646951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3647699ce097STejun Heo } else { 3648699ce097STejun Heo pwq->max_active = 0; 3649699ce097STejun Heo } 3650699ce097STejun Heo 3651a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36520fbd95aaSTejun Heo } 36530fbd95aaSTejun Heo 3654e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3655f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3656f147f29eSTejun Heo struct worker_pool *pool) 3657d2c1d404STejun Heo { 3658d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3659d2c1d404STejun Heo 3660e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3661e50aba9aSTejun Heo 3662d2c1d404STejun Heo pwq->pool = pool; 3663d2c1d404STejun Heo pwq->wq = wq; 3664d2c1d404STejun Heo pwq->flush_color = -1; 36658864b4e5STejun Heo pwq->refcnt = 1; 3666d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 36671befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3668d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36698864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3670f147f29eSTejun Heo } 3671d2c1d404STejun Heo 3672f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 36731befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3674f147f29eSTejun Heo { 3675f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3676f147f29eSTejun Heo 3677f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 367875ccf595STejun Heo 36791befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 36801befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 36811befcf30STejun Heo return; 36821befcf30STejun Heo 3683983ca25eSTejun Heo /* 3684983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 36853c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3686983ca25eSTejun Heo */ 368775ccf595STejun Heo pwq->work_color = wq->work_color; 3688983ca25eSTejun Heo 3689983ca25eSTejun Heo /* sync max_active to the current setting */ 3690983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3691983ca25eSTejun Heo 3692983ca25eSTejun Heo /* link in @pwq */ 36939e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3694df2d5ae4STejun Heo } 36956029a918STejun Heo 3696f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3697f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3698f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3699f147f29eSTejun Heo { 3700f147f29eSTejun Heo struct worker_pool *pool; 3701f147f29eSTejun Heo struct pool_workqueue *pwq; 3702f147f29eSTejun Heo 3703f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3704f147f29eSTejun Heo 3705f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3706f147f29eSTejun Heo if (!pool) 3707f147f29eSTejun Heo return NULL; 3708f147f29eSTejun Heo 3709e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3710f147f29eSTejun Heo if (!pwq) { 3711f147f29eSTejun Heo put_unbound_pool(pool); 3712f147f29eSTejun Heo return NULL; 3713f147f29eSTejun Heo } 3714f147f29eSTejun Heo 3715f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3716f147f29eSTejun Heo return pwq; 3717d2c1d404STejun Heo } 3718d2c1d404STejun Heo 37194c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 37204c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 37214c16bd32STejun Heo { 37224c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37234c16bd32STejun Heo 37244c16bd32STejun Heo if (pwq) { 37254c16bd32STejun Heo put_unbound_pool(pwq->pool); 3726cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 37274c16bd32STejun Heo } 37284c16bd32STejun Heo } 37294c16bd32STejun Heo 37304c16bd32STejun Heo /** 37314c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 37324c16bd32STejun Heo * @attrs: the wq_attrs of interest 37334c16bd32STejun Heo * @node: the target NUMA node 37344c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 37354c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 37364c16bd32STejun Heo * 37374c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 37384c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3739d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 37404c16bd32STejun Heo * 37414c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 37424c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 37434c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 37444c16bd32STejun Heo * @attrs->cpumask. 37454c16bd32STejun Heo * 37464c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 37474c16bd32STejun Heo * stable. 3748d185af30SYacine Belkadi * 3749d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3750d185af30SYacine Belkadi * %false if equal. 37514c16bd32STejun Heo */ 37524c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 37534c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 37544c16bd32STejun Heo { 3755d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 37564c16bd32STejun Heo goto use_dfl; 37574c16bd32STejun Heo 37584c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 37594c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 37604c16bd32STejun Heo if (cpu_going_down >= 0) 37614c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 37624c16bd32STejun Heo 37634c16bd32STejun Heo if (cpumask_empty(cpumask)) 37644c16bd32STejun Heo goto use_dfl; 37654c16bd32STejun Heo 37664c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 37674c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 37684c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 37694c16bd32STejun Heo 37704c16bd32STejun Heo use_dfl: 37714c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 37724c16bd32STejun Heo return false; 37734c16bd32STejun Heo } 37744c16bd32STejun Heo 37751befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 37761befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 37771befcf30STejun Heo int node, 37781befcf30STejun Heo struct pool_workqueue *pwq) 37791befcf30STejun Heo { 37801befcf30STejun Heo struct pool_workqueue *old_pwq; 37811befcf30STejun Heo 37821befcf30STejun Heo lockdep_assert_held(&wq->mutex); 37831befcf30STejun Heo 37841befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 37851befcf30STejun Heo link_pwq(pwq); 37861befcf30STejun Heo 37871befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 37881befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 37891befcf30STejun Heo return old_pwq; 37901befcf30STejun Heo } 37911befcf30STejun Heo 37929e8cd2f5STejun Heo /** 37939e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37949e8cd2f5STejun Heo * @wq: the target workqueue 37959e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37969e8cd2f5STejun Heo * 37974c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 37984c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 37994c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38004c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38014c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38024c16bd32STejun Heo * back-to-back will stay on its current pwq. 38039e8cd2f5STejun Heo * 3804d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3805d185af30SYacine Belkadi * 3806d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 38079e8cd2f5STejun Heo */ 38089e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38099e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38109e8cd2f5STejun Heo { 38114c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 38124c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3813f147f29eSTejun Heo int node, ret; 38149e8cd2f5STejun Heo 38158719dceaSTejun Heo /* only unbound workqueues can change attributes */ 38169e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 38179e8cd2f5STejun Heo return -EINVAL; 38189e8cd2f5STejun Heo 38198719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 38208719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 38218719dceaSTejun Heo return -EINVAL; 38228719dceaSTejun Heo 38234c16bd32STejun Heo pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); 382413e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38254c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38264c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 382713e2e556STejun Heo goto enomem; 382813e2e556STejun Heo 38294c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 383013e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 383113e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 383213e2e556STejun Heo 38334c16bd32STejun Heo /* 38344c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 38354c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 38364c16bd32STejun Heo * pools. 38374c16bd32STejun Heo */ 38384c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 38399e8cd2f5STejun Heo 38404c16bd32STejun Heo /* 38414c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 38424c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 38434c16bd32STejun Heo * pwqs accordingly. 38444c16bd32STejun Heo */ 38454c16bd32STejun Heo get_online_cpus(); 38464c16bd32STejun Heo 38474c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 38484c16bd32STejun Heo 38494c16bd32STejun Heo /* 38504c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 38514c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 38524c16bd32STejun Heo * it even if we don't use it immediately. 38534c16bd32STejun Heo */ 38544c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 38554c16bd32STejun Heo if (!dfl_pwq) 38564c16bd32STejun Heo goto enomem_pwq; 38574c16bd32STejun Heo 38584c16bd32STejun Heo for_each_node(node) { 38594c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 38604c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 38614c16bd32STejun Heo if (!pwq_tbl[node]) 38624c16bd32STejun Heo goto enomem_pwq; 38634c16bd32STejun Heo } else { 38644c16bd32STejun Heo dfl_pwq->refcnt++; 38654c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 38664c16bd32STejun Heo } 38674c16bd32STejun Heo } 38684c16bd32STejun Heo 38694c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 38704c16bd32STejun Heo 38714c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3872f147f29eSTejun Heo mutex_lock(&wq->mutex); 3873a892caccSTejun Heo 3874f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 38754c16bd32STejun Heo 38764c16bd32STejun Heo /* save the previous pwq and install the new one */ 3877f147f29eSTejun Heo for_each_node(node) 38784c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 38794c16bd32STejun Heo 38804c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 38814c16bd32STejun Heo link_pwq(dfl_pwq); 38824c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3883f147f29eSTejun Heo 3884f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3885f147f29eSTejun Heo 38864c16bd32STejun Heo /* put the old pwqs */ 38874c16bd32STejun Heo for_each_node(node) 38884c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 38894c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 38904c16bd32STejun Heo 38914c16bd32STejun Heo put_online_cpus(); 38924862125bSTejun Heo ret = 0; 38934862125bSTejun Heo /* fall through */ 38944862125bSTejun Heo out_free: 38954c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 38964862125bSTejun Heo free_workqueue_attrs(new_attrs); 38974c16bd32STejun Heo kfree(pwq_tbl); 38984862125bSTejun Heo return ret; 389913e2e556STejun Heo 39004c16bd32STejun Heo enomem_pwq: 39014c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 39024c16bd32STejun Heo for_each_node(node) 39034c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 39044c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 39054c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39064c16bd32STejun Heo put_online_cpus(); 390713e2e556STejun Heo enomem: 39084862125bSTejun Heo ret = -ENOMEM; 39094862125bSTejun Heo goto out_free; 39109e8cd2f5STejun Heo } 39119e8cd2f5STejun Heo 39124c16bd32STejun Heo /** 39134c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 39144c16bd32STejun Heo * @wq: the target workqueue 39154c16bd32STejun Heo * @cpu: the CPU coming up or going down 39164c16bd32STejun Heo * @online: whether @cpu is coming up or going down 39174c16bd32STejun Heo * 39184c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 39194c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 39204c16bd32STejun Heo * @wq accordingly. 39214c16bd32STejun Heo * 39224c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 39234c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 39244c16bd32STejun Heo * correct. 39254c16bd32STejun Heo * 39264c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 39274c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 39284c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 39294c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 39304c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 39314c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 39324c16bd32STejun Heo * CPU_DOWN_PREPARE. 39334c16bd32STejun Heo */ 39344c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 39354c16bd32STejun Heo bool online) 39364c16bd32STejun Heo { 39374c16bd32STejun Heo int node = cpu_to_node(cpu); 39384c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 39394c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 39404c16bd32STejun Heo struct workqueue_attrs *target_attrs; 39414c16bd32STejun Heo cpumask_t *cpumask; 39424c16bd32STejun Heo 39434c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 39444c16bd32STejun Heo 39454c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 39464c16bd32STejun Heo return; 39474c16bd32STejun Heo 39484c16bd32STejun Heo /* 39494c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39504c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39514c16bd32STejun Heo * CPU hotplug exclusion. 39524c16bd32STejun Heo */ 39534c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39544c16bd32STejun Heo cpumask = target_attrs->cpumask; 39554c16bd32STejun Heo 39564c16bd32STejun Heo mutex_lock(&wq->mutex); 3957d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 3958d55262c4STejun Heo goto out_unlock; 39594c16bd32STejun Heo 39604c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39614c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39624c16bd32STejun Heo 39634c16bd32STejun Heo /* 39644c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 39654c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 39664c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 3967534a3fbbSDaeseok Youn * wq's, the default pwq should be used. 39684c16bd32STejun Heo */ 39694c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 39704c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 39714c16bd32STejun Heo goto out_unlock; 39724c16bd32STejun Heo } else { 39734c16bd32STejun Heo goto use_dfl_pwq; 39744c16bd32STejun Heo } 39754c16bd32STejun Heo 39764c16bd32STejun Heo mutex_unlock(&wq->mutex); 39774c16bd32STejun Heo 39784c16bd32STejun Heo /* create a new pwq */ 39794c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 39804c16bd32STejun Heo if (!pwq) { 39812d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 39824c16bd32STejun Heo wq->name); 398377f300b1SDaeseok Youn mutex_lock(&wq->mutex); 398477f300b1SDaeseok Youn goto use_dfl_pwq; 39854c16bd32STejun Heo } 39864c16bd32STejun Heo 39874c16bd32STejun Heo /* 39884c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 39894c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 39904c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 39914c16bd32STejun Heo * inbetween. 39924c16bd32STejun Heo */ 39934c16bd32STejun Heo mutex_lock(&wq->mutex); 39944c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 39954c16bd32STejun Heo goto out_unlock; 39964c16bd32STejun Heo 39974c16bd32STejun Heo use_dfl_pwq: 39984c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 39994c16bd32STejun Heo get_pwq(wq->dfl_pwq); 40004c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 40014c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 40024c16bd32STejun Heo out_unlock: 40034c16bd32STejun Heo mutex_unlock(&wq->mutex); 40044c16bd32STejun Heo put_pwq_unlocked(old_pwq); 40054c16bd32STejun Heo } 40064c16bd32STejun Heo 400730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 40081da177e4SLinus Torvalds { 400949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 40108a2b7538STejun Heo int cpu, ret; 4011e1d8aa9fSFrederic Weisbecker 401230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4013420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4014420c0ddbSTejun Heo if (!wq->cpu_pwqs) 401530cdf249STejun Heo return -ENOMEM; 401630cdf249STejun Heo 401730cdf249STejun Heo for_each_possible_cpu(cpu) { 40187fb98ea7STejun Heo struct pool_workqueue *pwq = 40197fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 40207a62c2c8STejun Heo struct worker_pool *cpu_pools = 4021f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 402230cdf249STejun Heo 4023f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4024f147f29eSTejun Heo 4025f147f29eSTejun Heo mutex_lock(&wq->mutex); 40261befcf30STejun Heo link_pwq(pwq); 4027f147f29eSTejun Heo mutex_unlock(&wq->mutex); 402830cdf249STejun Heo } 402930cdf249STejun Heo return 0; 40308a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 40318a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 40328a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 40338a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 40348a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 40358a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 40368a2b7538STejun Heo return ret; 40379e8cd2f5STejun Heo } else { 40389e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 40399e8cd2f5STejun Heo } 40400f900049STejun Heo } 40410f900049STejun Heo 4042f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4043f3421797STejun Heo const char *name) 4044b71ab8c2STejun Heo { 4045f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4046f3421797STejun Heo 4047f3421797STejun Heo if (max_active < 1 || max_active > lim) 4048044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4049f3421797STejun Heo max_active, name, 1, lim); 4050b71ab8c2STejun Heo 4051f3421797STejun Heo return clamp_val(max_active, 1, lim); 4052b71ab8c2STejun Heo } 4053b71ab8c2STejun Heo 4054b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 405597e37d7bSTejun Heo unsigned int flags, 40561e19ffc6STejun Heo int max_active, 4057eb13ba87SJohannes Berg struct lock_class_key *key, 4058b196be89STejun Heo const char *lock_name, ...) 40593af24433SOleg Nesterov { 4060df2d5ae4STejun Heo size_t tbl_size = 0; 4061ecf6881fSTejun Heo va_list args; 40623af24433SOleg Nesterov struct workqueue_struct *wq; 406349e3cf44STejun Heo struct pool_workqueue *pwq; 4064b196be89STejun Heo 4065cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4066cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4067cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4068cee22a15SViresh Kumar 4069ecf6881fSTejun Heo /* allocate wq and format name */ 4070df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4071df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 4072df2d5ae4STejun Heo 4073df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4074b196be89STejun Heo if (!wq) 4075d2c1d404STejun Heo return NULL; 4076b196be89STejun Heo 40776029a918STejun Heo if (flags & WQ_UNBOUND) { 40786029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40796029a918STejun Heo if (!wq->unbound_attrs) 40806029a918STejun Heo goto err_free_wq; 40816029a918STejun Heo } 40826029a918STejun Heo 4083ecf6881fSTejun Heo va_start(args, lock_name); 4084ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4085b196be89STejun Heo va_end(args); 40863af24433SOleg Nesterov 4087d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4088b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40893af24433SOleg Nesterov 4090b196be89STejun Heo /* init wq */ 409197e37d7bSTejun Heo wq->flags = flags; 4092a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 40933c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4094112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 409530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 409673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 409773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4098493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 40993af24433SOleg Nesterov 4100eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4101cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 41023af24433SOleg Nesterov 410330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4104d2c1d404STejun Heo goto err_free_wq; 41051537663fSTejun Heo 4106493008a8STejun Heo /* 4107493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4108493008a8STejun Heo * have a rescuer to guarantee forward progress. 4109493008a8STejun Heo */ 4110493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4111e22bee78STejun Heo struct worker *rescuer; 4112e22bee78STejun Heo 4113f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 4114e22bee78STejun Heo if (!rescuer) 4115d2c1d404STejun Heo goto err_destroy; 4116e22bee78STejun Heo 4117111c225aSTejun Heo rescuer->rescue_wq = wq; 4118111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4119b196be89STejun Heo wq->name); 4120d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4121d2c1d404STejun Heo kfree(rescuer); 4122d2c1d404STejun Heo goto err_destroy; 4123d2c1d404STejun Heo } 4124e22bee78STejun Heo 4125d2c1d404STejun Heo wq->rescuer = rescuer; 412614a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4127e22bee78STejun Heo wake_up_process(rescuer->task); 41283af24433SOleg Nesterov } 41291537663fSTejun Heo 4130226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4131226223abSTejun Heo goto err_destroy; 4132226223abSTejun Heo 41333af24433SOleg Nesterov /* 413468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 413568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 413668e13a67SLai Jiangshan * list. 41373af24433SOleg Nesterov */ 413868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4139a0a1a5fdSTejun Heo 4140a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 414149e3cf44STejun Heo for_each_pwq(pwq, wq) 4142699ce097STejun Heo pwq_adjust_max_active(pwq); 4143a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4144a0a1a5fdSTejun Heo 41453af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4146a0a1a5fdSTejun Heo 414768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41483af24433SOleg Nesterov 41493af24433SOleg Nesterov return wq; 4150d2c1d404STejun Heo 4151d2c1d404STejun Heo err_free_wq: 41526029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41534690c4abSTejun Heo kfree(wq); 4154d2c1d404STejun Heo return NULL; 4155d2c1d404STejun Heo err_destroy: 4156d2c1d404STejun Heo destroy_workqueue(wq); 41574690c4abSTejun Heo return NULL; 41581da177e4SLinus Torvalds } 4159d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41601da177e4SLinus Torvalds 41613af24433SOleg Nesterov /** 41623af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41633af24433SOleg Nesterov * @wq: target workqueue 41643af24433SOleg Nesterov * 41653af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41663af24433SOleg Nesterov */ 41673af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41683af24433SOleg Nesterov { 416949e3cf44STejun Heo struct pool_workqueue *pwq; 41704c16bd32STejun Heo int node; 41713af24433SOleg Nesterov 41729c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41739c5a2ba7STejun Heo drain_workqueue(wq); 4174c8efcc25STejun Heo 41756183c009STejun Heo /* sanity checks */ 4176b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 417749e3cf44STejun Heo for_each_pwq(pwq, wq) { 41786183c009STejun Heo int i; 41796183c009STejun Heo 418076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 418176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4182b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41836183c009STejun Heo return; 418476af4d93STejun Heo } 418576af4d93STejun Heo } 418676af4d93STejun Heo 41875c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 41888864b4e5STejun Heo WARN_ON(pwq->nr_active) || 418976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4190b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41916183c009STejun Heo return; 41926183c009STejun Heo } 419376af4d93STejun Heo } 4194b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41956183c009STejun Heo 4196a0a1a5fdSTejun Heo /* 4197a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4198a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4199a0a1a5fdSTejun Heo */ 420068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4201d2c1d404STejun Heo list_del_init(&wq->list); 420268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42033af24433SOleg Nesterov 4204226223abSTejun Heo workqueue_sysfs_unregister(wq); 4205226223abSTejun Heo 4206493008a8STejun Heo if (wq->rescuer) { 4207e22bee78STejun Heo kthread_stop(wq->rescuer->task); 42088d9df9f0SXiaotian Feng kfree(wq->rescuer); 4209493008a8STejun Heo wq->rescuer = NULL; 4210e22bee78STejun Heo } 4211e22bee78STejun Heo 42128864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 421329c91e99STejun Heo /* 42148864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 42158864b4e5STejun Heo * free the pwqs and wq. 421629c91e99STejun Heo */ 42178864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 42188864b4e5STejun Heo kfree(wq); 42198864b4e5STejun Heo } else { 42208864b4e5STejun Heo /* 42218864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 42224c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 42234c16bd32STejun Heo * @wq will be freed when the last pwq is released. 42248864b4e5STejun Heo */ 42254c16bd32STejun Heo for_each_node(node) { 42264c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42274c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 42284c16bd32STejun Heo put_pwq_unlocked(pwq); 42294c16bd32STejun Heo } 42304c16bd32STejun Heo 42314c16bd32STejun Heo /* 42324c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 42334c16bd32STejun Heo * put. Don't access it afterwards. 42344c16bd32STejun Heo */ 42354c16bd32STejun Heo pwq = wq->dfl_pwq; 42364c16bd32STejun Heo wq->dfl_pwq = NULL; 4237dce90d47STejun Heo put_pwq_unlocked(pwq); 423829c91e99STejun Heo } 42393af24433SOleg Nesterov } 42403af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 42413af24433SOleg Nesterov 4242dcd989cbSTejun Heo /** 4243dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4244dcd989cbSTejun Heo * @wq: target workqueue 4245dcd989cbSTejun Heo * @max_active: new max_active value. 4246dcd989cbSTejun Heo * 4247dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4248dcd989cbSTejun Heo * 4249dcd989cbSTejun Heo * CONTEXT: 4250dcd989cbSTejun Heo * Don't call from IRQ context. 4251dcd989cbSTejun Heo */ 4252dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4253dcd989cbSTejun Heo { 425449e3cf44STejun Heo struct pool_workqueue *pwq; 4255dcd989cbSTejun Heo 42568719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42578719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 42588719dceaSTejun Heo return; 42598719dceaSTejun Heo 4260f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4261dcd989cbSTejun Heo 4262a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4263dcd989cbSTejun Heo 4264dcd989cbSTejun Heo wq->saved_max_active = max_active; 4265dcd989cbSTejun Heo 4266699ce097STejun Heo for_each_pwq(pwq, wq) 4267699ce097STejun Heo pwq_adjust_max_active(pwq); 4268dcd989cbSTejun Heo 4269a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4270dcd989cbSTejun Heo } 4271dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4272dcd989cbSTejun Heo 4273dcd989cbSTejun Heo /** 4274e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4275e6267616STejun Heo * 4276e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4277e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4278d185af30SYacine Belkadi * 4279d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4280e6267616STejun Heo */ 4281e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4282e6267616STejun Heo { 4283e6267616STejun Heo struct worker *worker = current_wq_worker(); 4284e6267616STejun Heo 42856a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4286e6267616STejun Heo } 4287e6267616STejun Heo 4288e6267616STejun Heo /** 4289dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4290dcd989cbSTejun Heo * @cpu: CPU in question 4291dcd989cbSTejun Heo * @wq: target workqueue 4292dcd989cbSTejun Heo * 4293dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4294dcd989cbSTejun Heo * no synchronization around this function and the test result is 4295dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4296dcd989cbSTejun Heo * 4297d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4298d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4299d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4300d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4301d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4302d3251859STejun Heo * 4303d185af30SYacine Belkadi * Return: 4304dcd989cbSTejun Heo * %true if congested, %false otherwise. 4305dcd989cbSTejun Heo */ 4306d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4307dcd989cbSTejun Heo { 43087fb98ea7STejun Heo struct pool_workqueue *pwq; 430976af4d93STejun Heo bool ret; 431076af4d93STejun Heo 431188109453SLai Jiangshan rcu_read_lock_sched(); 43127fb98ea7STejun Heo 4313d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4314d3251859STejun Heo cpu = smp_processor_id(); 4315d3251859STejun Heo 43167fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 43177fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 43187fb98ea7STejun Heo else 4319df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4320dcd989cbSTejun Heo 432176af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 432288109453SLai Jiangshan rcu_read_unlock_sched(); 432376af4d93STejun Heo 432476af4d93STejun Heo return ret; 4325dcd989cbSTejun Heo } 4326dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4327dcd989cbSTejun Heo 4328dcd989cbSTejun Heo /** 4329dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4330dcd989cbSTejun Heo * @work: the work to be tested 4331dcd989cbSTejun Heo * 4332dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4333dcd989cbSTejun Heo * synchronization around this function and the test result is 4334dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4335dcd989cbSTejun Heo * 4336d185af30SYacine Belkadi * Return: 4337dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4338dcd989cbSTejun Heo */ 4339dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4340dcd989cbSTejun Heo { 4341fa1b54e6STejun Heo struct worker_pool *pool; 4342dcd989cbSTejun Heo unsigned long flags; 4343dcd989cbSTejun Heo unsigned int ret = 0; 4344dcd989cbSTejun Heo 4345dcd989cbSTejun Heo if (work_pending(work)) 4346dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4347038366c5SLai Jiangshan 4348fa1b54e6STejun Heo local_irq_save(flags); 4349fa1b54e6STejun Heo pool = get_work_pool(work); 4350038366c5SLai Jiangshan if (pool) { 4351fa1b54e6STejun Heo spin_lock(&pool->lock); 4352c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4353dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4354fa1b54e6STejun Heo spin_unlock(&pool->lock); 4355038366c5SLai Jiangshan } 4356fa1b54e6STejun Heo local_irq_restore(flags); 4357dcd989cbSTejun Heo 4358dcd989cbSTejun Heo return ret; 4359dcd989cbSTejun Heo } 4360dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4361dcd989cbSTejun Heo 43623d1cb205STejun Heo /** 43633d1cb205STejun Heo * set_worker_desc - set description for the current work item 43643d1cb205STejun Heo * @fmt: printf-style format string 43653d1cb205STejun Heo * @...: arguments for the format string 43663d1cb205STejun Heo * 43673d1cb205STejun Heo * This function can be called by a running work function to describe what 43683d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43693d1cb205STejun Heo * information will be printed out together to help debugging. The 43703d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43713d1cb205STejun Heo */ 43723d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43733d1cb205STejun Heo { 43743d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43753d1cb205STejun Heo va_list args; 43763d1cb205STejun Heo 43773d1cb205STejun Heo if (worker) { 43783d1cb205STejun Heo va_start(args, fmt); 43793d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 43803d1cb205STejun Heo va_end(args); 43813d1cb205STejun Heo worker->desc_valid = true; 43823d1cb205STejun Heo } 43833d1cb205STejun Heo } 43843d1cb205STejun Heo 43853d1cb205STejun Heo /** 43863d1cb205STejun Heo * print_worker_info - print out worker information and description 43873d1cb205STejun Heo * @log_lvl: the log level to use when printing 43883d1cb205STejun Heo * @task: target task 43893d1cb205STejun Heo * 43903d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 43913d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 43923d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 43933d1cb205STejun Heo * 43943d1cb205STejun Heo * This function can be safely called on any task as long as the 43953d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 43963d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 43973d1cb205STejun Heo */ 43983d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 43993d1cb205STejun Heo { 44003d1cb205STejun Heo work_func_t *fn = NULL; 44013d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 44023d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 44033d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 44043d1cb205STejun Heo struct workqueue_struct *wq = NULL; 44053d1cb205STejun Heo bool desc_valid = false; 44063d1cb205STejun Heo struct worker *worker; 44073d1cb205STejun Heo 44083d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 44093d1cb205STejun Heo return; 44103d1cb205STejun Heo 44113d1cb205STejun Heo /* 44123d1cb205STejun Heo * This function is called without any synchronization and @task 44133d1cb205STejun Heo * could be in any state. Be careful with dereferences. 44143d1cb205STejun Heo */ 44153d1cb205STejun Heo worker = probe_kthread_data(task); 44163d1cb205STejun Heo 44173d1cb205STejun Heo /* 44183d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 44193d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 44203d1cb205STejun Heo */ 44213d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 44223d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 44233d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 44243d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 44253d1cb205STejun Heo 44263d1cb205STejun Heo /* copy worker description */ 44273d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 44283d1cb205STejun Heo if (desc_valid) 44293d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 44303d1cb205STejun Heo 44313d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44323d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 44333d1cb205STejun Heo if (desc[0]) 44343d1cb205STejun Heo pr_cont(" (%s)", desc); 44353d1cb205STejun Heo pr_cont("\n"); 44363d1cb205STejun Heo } 44373d1cb205STejun Heo } 44383d1cb205STejun Heo 4439db7bccf4STejun Heo /* 4440db7bccf4STejun Heo * CPU hotplug. 4441db7bccf4STejun Heo * 4442e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4443112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4444706026c2STejun Heo * pool which make migrating pending and scheduled works very 4445e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 444694cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4447e22bee78STejun Heo * blocked draining impractical. 4448e22bee78STejun Heo * 444924647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4450628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4451628c78e7STejun Heo * cpu comes back online. 4452db7bccf4STejun Heo */ 4453db7bccf4STejun Heo 4454706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4455db7bccf4STejun Heo { 445638db41d9STejun Heo int cpu = smp_processor_id(); 44574ce62e9eSTejun Heo struct worker_pool *pool; 4458db7bccf4STejun Heo struct worker *worker; 4459db7bccf4STejun Heo 4460f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 446192f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 446294cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4463e22bee78STejun Heo 4464f2d5a0eeSTejun Heo /* 446592f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 446694cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 446794cf58bbSTejun Heo * except for the ones which are still executing works from 446894cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 446994cf58bbSTejun Heo * this, they may become diasporas. 4470f2d5a0eeSTejun Heo */ 4471da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4472403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4473db7bccf4STejun Heo 447424647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4475f2d5a0eeSTejun Heo 447694cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 447792f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4478e22bee78STejun Heo 4479e22bee78STejun Heo /* 4480eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4481eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4482eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4483eb283428SLai Jiangshan * from other cpus. 4484628c78e7STejun Heo */ 4485628c78e7STejun Heo schedule(); 4486628c78e7STejun Heo 4487628c78e7STejun Heo /* 4488eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4489eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4490eb283428SLai Jiangshan * and keep_working() are always true as long as the 4491eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4492eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4493eb283428SLai Jiangshan * are served by workers tied to the pool. 4494e22bee78STejun Heo */ 4495e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4496eb283428SLai Jiangshan 4497eb283428SLai Jiangshan /* 4498eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4499eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4500eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4501eb283428SLai Jiangshan */ 4502eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4503eb283428SLai Jiangshan wake_up_worker(pool); 4504eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4505eb283428SLai Jiangshan } 4506db7bccf4STejun Heo } 4507db7bccf4STejun Heo 4508bd7c089eSTejun Heo /** 4509bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4510bd7c089eSTejun Heo * @pool: pool of interest 4511bd7c089eSTejun Heo * 4512a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4513bd7c089eSTejun Heo */ 4514bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4515bd7c089eSTejun Heo { 4516a9ab775bSTejun Heo struct worker *worker; 4517bd7c089eSTejun Heo 451892f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4519bd7c089eSTejun Heo 4520bd7c089eSTejun Heo /* 4521a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4522a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4523a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4524a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4525a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4526bd7c089eSTejun Heo */ 4527da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4528a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4529a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4530a9ab775bSTejun Heo 4531a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 45323de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4533a9ab775bSTejun Heo 4534da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4535a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4536a9ab775bSTejun Heo 4537a9ab775bSTejun Heo /* 4538a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4539a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4540a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4541a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4542a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4543a9ab775bSTejun Heo * be bound before @pool->lock is released. 4544a9ab775bSTejun Heo */ 4545a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4546bd7c089eSTejun Heo wake_up_process(worker->task); 4547bd7c089eSTejun Heo 4548bd7c089eSTejun Heo /* 4549a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4550a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4551a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4552a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4553a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4554a9ab775bSTejun Heo * concurrency management. Note that when or whether 4555a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4556a9ab775bSTejun Heo * 4557a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4558a9ab775bSTejun Heo * tested without holding any lock in 4559a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4560a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4561a9ab775bSTejun Heo * management operations. 4562bd7c089eSTejun Heo */ 4563a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4564a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4565a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4566a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4567bd7c089eSTejun Heo } 4568a9ab775bSTejun Heo 4569a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4570bd7c089eSTejun Heo } 4571bd7c089eSTejun Heo 45727dbc725eSTejun Heo /** 45737dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 45747dbc725eSTejun Heo * @pool: unbound pool of interest 45757dbc725eSTejun Heo * @cpu: the CPU which is coming up 45767dbc725eSTejun Heo * 45777dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 45787dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 45797dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 45807dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 45817dbc725eSTejun Heo */ 45827dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 45837dbc725eSTejun Heo { 45847dbc725eSTejun Heo static cpumask_t cpumask; 45857dbc725eSTejun Heo struct worker *worker; 45867dbc725eSTejun Heo 458792f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 45887dbc725eSTejun Heo 45897dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 45907dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 45917dbc725eSTejun Heo return; 45927dbc725eSTejun Heo 45937dbc725eSTejun Heo /* is @cpu the only online CPU? */ 45947dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 45957dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 45967dbc725eSTejun Heo return; 45977dbc725eSTejun Heo 45987dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4599da028469SLai Jiangshan for_each_pool_worker(worker, pool) 46007dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 46017dbc725eSTejun Heo pool->attrs->cpumask) < 0); 46027dbc725eSTejun Heo } 46037dbc725eSTejun Heo 46048db25e78STejun Heo /* 46058db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 46068db25e78STejun Heo * This will be registered high priority CPU notifier. 46078db25e78STejun Heo */ 46080db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 46091da177e4SLinus Torvalds unsigned long action, 46101da177e4SLinus Torvalds void *hcpu) 46111da177e4SLinus Torvalds { 4612d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46134ce62e9eSTejun Heo struct worker_pool *pool; 46144c16bd32STejun Heo struct workqueue_struct *wq; 46157dbc725eSTejun Heo int pi; 46161da177e4SLinus Torvalds 46178db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 46183af24433SOleg Nesterov case CPU_UP_PREPARE: 4619f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46203ce63377STejun Heo if (pool->nr_workers) 46213ce63377STejun Heo continue; 4622ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 46233ce63377STejun Heo return NOTIFY_BAD; 46243af24433SOleg Nesterov } 46251da177e4SLinus Torvalds break; 46261da177e4SLinus Torvalds 462765758202STejun Heo case CPU_DOWN_FAILED: 462865758202STejun Heo case CPU_ONLINE: 462968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46307dbc725eSTejun Heo 46317dbc725eSTejun Heo for_each_pool(pool, pi) { 463292f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 463394cf58bbSTejun Heo 46347dbc725eSTejun Heo if (pool->cpu == cpu) { 463594cf58bbSTejun Heo rebind_workers(pool); 46367dbc725eSTejun Heo } else if (pool->cpu < 0) { 46377dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 46387dbc725eSTejun Heo } 463994cf58bbSTejun Heo 464092f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 464194cf58bbSTejun Heo } 46427dbc725eSTejun Heo 46434c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46444c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46454c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 46464c16bd32STejun Heo 464768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46488db25e78STejun Heo break; 464965758202STejun Heo } 465065758202STejun Heo return NOTIFY_OK; 465165758202STejun Heo } 465265758202STejun Heo 465365758202STejun Heo /* 465465758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 465565758202STejun Heo * This will be registered as low priority CPU notifier. 465665758202STejun Heo */ 46570db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 465865758202STejun Heo unsigned long action, 465965758202STejun Heo void *hcpu) 466065758202STejun Heo { 4661d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46628db25e78STejun Heo struct work_struct unbind_work; 46634c16bd32STejun Heo struct workqueue_struct *wq; 46648db25e78STejun Heo 466565758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 466665758202STejun Heo case CPU_DOWN_PREPARE: 46674c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4668706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46697635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46704c16bd32STejun Heo 46714c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46724c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46734c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46744c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 46754c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 46764c16bd32STejun Heo 46774c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 46788db25e78STejun Heo flush_work(&unbind_work); 4679440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 46808db25e78STejun Heo break; 468165758202STejun Heo } 468265758202STejun Heo return NOTIFY_OK; 468365758202STejun Heo } 468465758202STejun Heo 46852d3854a3SRusty Russell #ifdef CONFIG_SMP 46868ccad40dSRusty Russell 46872d3854a3SRusty Russell struct work_for_cpu { 4688ed48ece2STejun Heo struct work_struct work; 46892d3854a3SRusty Russell long (*fn)(void *); 46902d3854a3SRusty Russell void *arg; 46912d3854a3SRusty Russell long ret; 46922d3854a3SRusty Russell }; 46932d3854a3SRusty Russell 4694ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 46952d3854a3SRusty Russell { 4696ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4697ed48ece2STejun Heo 46982d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 46992d3854a3SRusty Russell } 47002d3854a3SRusty Russell 47012d3854a3SRusty Russell /** 47022d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 47032d3854a3SRusty Russell * @cpu: the cpu to run on 47042d3854a3SRusty Russell * @fn: the function to run 47052d3854a3SRusty Russell * @arg: the function arg 47062d3854a3SRusty Russell * 470731ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47086b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4709d185af30SYacine Belkadi * 4710d185af30SYacine Belkadi * Return: The value @fn returns. 47112d3854a3SRusty Russell */ 4712d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47132d3854a3SRusty Russell { 4714ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47152d3854a3SRusty Russell 4716ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4717ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 471812997d1aSBjorn Helgaas flush_work(&wfc.work); 4719440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 47202d3854a3SRusty Russell return wfc.ret; 47212d3854a3SRusty Russell } 47222d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47232d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47242d3854a3SRusty Russell 4725a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4726e7577c50SRusty Russell 4727a0a1a5fdSTejun Heo /** 4728a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4729a0a1a5fdSTejun Heo * 473058a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4731c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4732706026c2STejun Heo * pool->worklist. 4733a0a1a5fdSTejun Heo * 4734a0a1a5fdSTejun Heo * CONTEXT: 4735a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4736a0a1a5fdSTejun Heo */ 4737a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4738a0a1a5fdSTejun Heo { 473924b8a847STejun Heo struct workqueue_struct *wq; 474024b8a847STejun Heo struct pool_workqueue *pwq; 4741a0a1a5fdSTejun Heo 474268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4743a0a1a5fdSTejun Heo 47446183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4745a0a1a5fdSTejun Heo workqueue_freezing = true; 4746a0a1a5fdSTejun Heo 474724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4748a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4749699ce097STejun Heo for_each_pwq(pwq, wq) 4750699ce097STejun Heo pwq_adjust_max_active(pwq); 4751a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4752a1056305STejun Heo } 47535bcab335STejun Heo 475468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4755a0a1a5fdSTejun Heo } 4756a0a1a5fdSTejun Heo 4757a0a1a5fdSTejun Heo /** 475858a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4759a0a1a5fdSTejun Heo * 4760a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4761a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4762a0a1a5fdSTejun Heo * 4763a0a1a5fdSTejun Heo * CONTEXT: 476468e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4765a0a1a5fdSTejun Heo * 4766d185af30SYacine Belkadi * Return: 476758a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 476858a69cb4STejun Heo * is complete. 4769a0a1a5fdSTejun Heo */ 4770a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4771a0a1a5fdSTejun Heo { 4772a0a1a5fdSTejun Heo bool busy = false; 477324b8a847STejun Heo struct workqueue_struct *wq; 477424b8a847STejun Heo struct pool_workqueue *pwq; 4775a0a1a5fdSTejun Heo 477668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4777a0a1a5fdSTejun Heo 47786183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4779a0a1a5fdSTejun Heo 478024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 478124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 478224b8a847STejun Heo continue; 4783a0a1a5fdSTejun Heo /* 4784a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4785a0a1a5fdSTejun Heo * to peek without lock. 4786a0a1a5fdSTejun Heo */ 478788109453SLai Jiangshan rcu_read_lock_sched(); 478824b8a847STejun Heo for_each_pwq(pwq, wq) { 47896183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4790112202d9STejun Heo if (pwq->nr_active) { 4791a0a1a5fdSTejun Heo busy = true; 479288109453SLai Jiangshan rcu_read_unlock_sched(); 4793a0a1a5fdSTejun Heo goto out_unlock; 4794a0a1a5fdSTejun Heo } 4795a0a1a5fdSTejun Heo } 479688109453SLai Jiangshan rcu_read_unlock_sched(); 4797a0a1a5fdSTejun Heo } 4798a0a1a5fdSTejun Heo out_unlock: 479968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4800a0a1a5fdSTejun Heo return busy; 4801a0a1a5fdSTejun Heo } 4802a0a1a5fdSTejun Heo 4803a0a1a5fdSTejun Heo /** 4804a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4805a0a1a5fdSTejun Heo * 4806a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4807706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4808a0a1a5fdSTejun Heo * 4809a0a1a5fdSTejun Heo * CONTEXT: 4810a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4811a0a1a5fdSTejun Heo */ 4812a0a1a5fdSTejun Heo void thaw_workqueues(void) 4813a0a1a5fdSTejun Heo { 481424b8a847STejun Heo struct workqueue_struct *wq; 481524b8a847STejun Heo struct pool_workqueue *pwq; 4816a0a1a5fdSTejun Heo 481768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4818a0a1a5fdSTejun Heo 4819a0a1a5fdSTejun Heo if (!workqueue_freezing) 4820a0a1a5fdSTejun Heo goto out_unlock; 4821a0a1a5fdSTejun Heo 482274b414eaSLai Jiangshan workqueue_freezing = false; 482324b8a847STejun Heo 482424b8a847STejun Heo /* restore max_active and repopulate worklist */ 482524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4826a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4827699ce097STejun Heo for_each_pwq(pwq, wq) 4828699ce097STejun Heo pwq_adjust_max_active(pwq); 4829a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 483024b8a847STejun Heo } 483124b8a847STejun Heo 4832a0a1a5fdSTejun Heo out_unlock: 483368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4834a0a1a5fdSTejun Heo } 4835a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4836a0a1a5fdSTejun Heo 4837bce90380STejun Heo static void __init wq_numa_init(void) 4838bce90380STejun Heo { 4839bce90380STejun Heo cpumask_var_t *tbl; 4840bce90380STejun Heo int node, cpu; 4841bce90380STejun Heo 4842bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4843bce90380STejun Heo for_each_node(node) 4844bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4845bce90380STejun Heo 4846bce90380STejun Heo if (num_possible_nodes() <= 1) 4847bce90380STejun Heo return; 4848bce90380STejun Heo 4849d55262c4STejun Heo if (wq_disable_numa) { 4850d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 4851d55262c4STejun Heo return; 4852d55262c4STejun Heo } 4853d55262c4STejun Heo 48544c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 48554c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 48564c16bd32STejun Heo 4857bce90380STejun Heo /* 4858bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4859bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4860bce90380STejun Heo * fully initialized by now. 4861bce90380STejun Heo */ 4862bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4863bce90380STejun Heo BUG_ON(!tbl); 4864bce90380STejun Heo 4865bce90380STejun Heo for_each_node(node) 48661be0c25dSTejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 48671be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 4868bce90380STejun Heo 4869bce90380STejun Heo for_each_possible_cpu(cpu) { 4870bce90380STejun Heo node = cpu_to_node(cpu); 4871bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4872bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4873bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4874bce90380STejun Heo return; 4875bce90380STejun Heo } 4876bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4877bce90380STejun Heo } 4878bce90380STejun Heo 4879bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4880bce90380STejun Heo wq_numa_enabled = true; 4881bce90380STejun Heo } 4882bce90380STejun Heo 48836ee0578bSSuresh Siddha static int __init init_workqueues(void) 48841da177e4SLinus Torvalds { 48857a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 48867a4e344cSTejun Heo int i, cpu; 4887c34056a3STejun Heo 4888e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4889e904e6c2STejun Heo 4890e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4891e904e6c2STejun Heo 489265758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4893a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 48948b03ae3cSTejun Heo 4895bce90380STejun Heo wq_numa_init(); 4896bce90380STejun Heo 4897706026c2STejun Heo /* initialize CPU pools */ 489829c91e99STejun Heo for_each_possible_cpu(cpu) { 48994ce62e9eSTejun Heo struct worker_pool *pool; 49008b03ae3cSTejun Heo 49017a4e344cSTejun Heo i = 0; 4902f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49037a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4904ec22ca5eSTejun Heo pool->cpu = cpu; 49057a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 49067a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4907f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 49087a4e344cSTejun Heo 49099daf9e67STejun Heo /* alloc pool ID */ 491068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 49119daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 491268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 49134ce62e9eSTejun Heo } 49148b03ae3cSTejun Heo } 49158b03ae3cSTejun Heo 4916e22bee78STejun Heo /* create the initial worker */ 491729c91e99STejun Heo for_each_online_cpu(cpu) { 49184ce62e9eSTejun Heo struct worker_pool *pool; 4919e22bee78STejun Heo 4920f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 492124647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4922ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4923e22bee78STejun Heo } 49244ce62e9eSTejun Heo } 4925e22bee78STejun Heo 49268a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 492729c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 492829c91e99STejun Heo struct workqueue_attrs *attrs; 492929c91e99STejun Heo 493029c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 493129c91e99STejun Heo attrs->nice = std_nice[i]; 493229c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 49338a2b7538STejun Heo 49348a2b7538STejun Heo /* 49358a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 49368a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 49378a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 49388a2b7538STejun Heo */ 49398a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 49408a2b7538STejun Heo attrs->nice = std_nice[i]; 49418a2b7538STejun Heo attrs->no_numa = true; 49428a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 494329c91e99STejun Heo } 494429c91e99STejun Heo 4945d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 49461aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4947d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4948f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4949f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 495024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 495124d51addSTejun Heo WQ_FREEZABLE, 0); 49520668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 49530668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 49540668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 49550668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 49560668106cSViresh Kumar 0); 49571aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 49580668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 49590668106cSViresh Kumar !system_power_efficient_wq || 49600668106cSViresh Kumar !system_freezable_power_efficient_wq); 49616ee0578bSSuresh Siddha return 0; 49621da177e4SLinus Torvalds } 49636ee0578bSSuresh Siddha early_initcall(init_workqueues); 4964