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 * 249a261491SBenjamin Peterson * Please read Documentation/core-api/workqueue.rst 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/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4429c91e99STejun Heo #include <linux/jhash.h> 4542f8570fSSasha Levin #include <linux/hashtable.h> 4676af4d93STejun Heo #include <linux/rculist.h> 47bce90380STejun Heo #include <linux/nodemask.h> 484c16bd32STejun Heo #include <linux/moduleparam.h> 493d1cb205STejun Heo #include <linux/uaccess.h> 50c98a9805STal Shorer #include <linux/sched/isolation.h> 5162635ea8SSergey Senozhatsky #include <linux/nmi.h> 52e22bee78STejun Heo 53ea138446STejun Heo #include "workqueue_internal.h" 541da177e4SLinus Torvalds 55c8e55f36STejun Heo enum { 56bc2ae0f5STejun Heo /* 5724647570STejun Heo * worker_pool flags 58bc2ae0f5STejun Heo * 5924647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 60bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 61bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 62bc2ae0f5STejun Heo * is in effect. 63bc2ae0f5STejun Heo * 64bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 65bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6624647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 67bc2ae0f5STejun Heo * 68bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 691258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 704736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 71bc2ae0f5STejun Heo */ 72692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7324647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 74db7bccf4STejun Heo 75c8e55f36STejun Heo /* worker flags */ 76c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 77c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 78e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 79fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 80f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 81a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 82e22bee78STejun Heo 83a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 84a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 85db7bccf4STejun Heo 86e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 874ce62e9eSTejun Heo 8829c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 89c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 90db7bccf4STejun Heo 91e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 92e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 93e22bee78STejun Heo 943233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 953233cdbdSTejun Heo /* call for help after 10ms 963233cdbdSTejun Heo (min two ticks) */ 97e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 98e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 991da177e4SLinus Torvalds 1001da177e4SLinus Torvalds /* 101e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1028698a745SDongsheng Yang * all cpus. Give MIN_NICE. 103e22bee78STejun Heo */ 1048698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1058698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 106ecf6881fSTejun Heo 107ecf6881fSTejun Heo WQ_NAME_LEN = 24, 108c8e55f36STejun Heo }; 109c8e55f36STejun Heo 1101da177e4SLinus Torvalds /* 1114690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1124690c4abSTejun Heo * 113e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 114e41e704bSTejun Heo * everyone else. 1154690c4abSTejun Heo * 116e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 117e22bee78STejun Heo * only be modified and accessed from the local cpu. 118e22bee78STejun Heo * 119d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1204690c4abSTejun Heo * 121d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 122d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 123d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 124d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 125e22bee78STejun Heo * 1261258fae7STejun Heo * A: wq_pool_attach_mutex protected. 127822d8405STejun Heo * 12868e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12976af4d93STejun Heo * 13068e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1315bcab335STejun Heo * 1325b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1335b95e1afSLai Jiangshan * 1345b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 1355b95e1afSLai Jiangshan * sched-RCU for reads. 1365b95e1afSLai Jiangshan * 1373c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1383c25a55dSLai Jiangshan * 139b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1402e109a28STejun Heo * 1412e109a28STejun Heo * MD: wq_mayday_lock protected. 1424690c4abSTejun Heo */ 1434690c4abSTejun Heo 1442eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 145c34056a3STejun Heo 146bd7bdd43STejun Heo struct worker_pool { 147d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 148d84ff051STejun Heo int cpu; /* I: the associated cpu */ 149f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1509daf9e67STejun Heo int id; /* I: pool ID */ 15111ebea50STejun Heo unsigned int flags; /* X: flags */ 152bd7bdd43STejun Heo 15382607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15482607adcSTejun Heo 155bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 156ea1abd61SLai Jiangshan 1575826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1585826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 159bd7bdd43STejun Heo 160bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 161bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 162bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 163bd7bdd43STejun Heo 164c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 165c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 166c9e7cf27STejun Heo /* L: hash of busy workers */ 167c9e7cf27STejun Heo 1682607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 16992f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17060f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 171e19e397aSTejun Heo 1727cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 173e19e397aSTejun Heo 1747a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17568e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17668e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1777a4e344cSTejun Heo 178e19e397aSTejun Heo /* 179e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 180e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 181e19e397aSTejun Heo * cacheline. 182e19e397aSTejun Heo */ 183e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18429c91e99STejun Heo 18529c91e99STejun Heo /* 18629c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18729c91e99STejun Heo * from get_work_pool(). 18829c91e99STejun Heo */ 18929c91e99STejun Heo struct rcu_head rcu; 1908b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1918b03ae3cSTejun Heo 1928b03ae3cSTejun Heo /* 193112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 194112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 195112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 196112202d9STejun Heo * number of flag bits. 1971da177e4SLinus Torvalds */ 198112202d9STejun Heo struct pool_workqueue { 199bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2004690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20173f53c4aSTejun Heo int work_color; /* L: current color */ 20273f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2038864b4e5STejun Heo int refcnt; /* L: reference count */ 20473f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20573f53c4aSTejun Heo /* L: nr of in_flight works */ 2061e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 207a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2081e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2093c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2102e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2118864b4e5STejun Heo 2128864b4e5STejun Heo /* 2138864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2148864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2158864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 216b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2178864b4e5STejun Heo */ 2188864b4e5STejun Heo struct work_struct unbound_release_work; 2198864b4e5STejun Heo struct rcu_head rcu; 220e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2211da177e4SLinus Torvalds 2221da177e4SLinus Torvalds /* 22373f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22473f53c4aSTejun Heo */ 22573f53c4aSTejun Heo struct wq_flusher { 2263c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2273c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22873f53c4aSTejun Heo struct completion done; /* flush completion */ 22973f53c4aSTejun Heo }; 2301da177e4SLinus Torvalds 231226223abSTejun Heo struct wq_device; 232226223abSTejun Heo 23373f53c4aSTejun Heo /* 234c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 235c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2361da177e4SLinus Torvalds */ 2371da177e4SLinus Torvalds struct workqueue_struct { 2383c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 239e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 24073f53c4aSTejun Heo 2413c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2423c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2433c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 244112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2453c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2463c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2473c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24873f53c4aSTejun Heo 2492e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 250e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 251e22bee78STejun Heo 25287fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 253a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 254226223abSTejun Heo 2555b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2565b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2576029a918STejun Heo 258226223abSTejun Heo #ifdef CONFIG_SYSFS 259226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 260226223abSTejun Heo #endif 2614e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2624e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2634e6045f1SJohannes Berg #endif 264ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2652728fd2fSTejun Heo 266e2dca7adSTejun Heo /* 267e2dca7adSTejun Heo * Destruction of workqueue_struct is sched-RCU protected to allow 268e2dca7adSTejun Heo * walking the workqueues list without grabbing wq_pool_mutex. 269e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 270e2dca7adSTejun Heo */ 271e2dca7adSTejun Heo struct rcu_head rcu; 272e2dca7adSTejun Heo 2732728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2742728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2752728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2765b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2771da177e4SLinus Torvalds }; 2781da177e4SLinus Torvalds 279e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 280e904e6c2STejun Heo 281bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 282bce90380STejun Heo /* possible CPUs of each node */ 283bce90380STejun Heo 284d55262c4STejun Heo static bool wq_disable_numa; 285d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 286d55262c4STejun Heo 287cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 288552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 289cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 290cee22a15SViresh Kumar 291863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 2923347fa09STejun Heo 293bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 294bce90380STejun Heo 2954c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2964c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2974c16bd32STejun Heo 29868e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2991258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 3002e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 301692b4825STejun Heo static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ 3025bcab335STejun Heo 303e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30468e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3057d19c5ceSTejun Heo 306ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 307ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 308ef557180SMike Galbraith 309ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 310ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 311b05a7928SFrederic Weisbecker 312f303fccbSTejun Heo /* 313f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 314f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 315f303fccbSTejun Heo * to uncover usages which depend on it. 316f303fccbSTejun Heo */ 317f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 318f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 319f303fccbSTejun Heo #else 320f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 321f303fccbSTejun Heo #endif 322f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 323f303fccbSTejun Heo 3247d19c5ceSTejun Heo /* the per-cpu worker pools */ 32525528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3267d19c5ceSTejun Heo 32768e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3287d19c5ceSTejun Heo 32968e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 33029c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 33129c91e99STejun Heo 332c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33329c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33429c91e99STejun Heo 3358a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3368a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3378a2b7538STejun Heo 338d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 339ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 340044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3411aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 342044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 343d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 344044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 345f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 346044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 34724d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3480668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3490668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3500668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3510668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 352d320c038STejun Heo 3537d19c5ceSTejun Heo static int worker_thread(void *__worker); 3546ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3557d19c5ceSTejun Heo 35697bd2347STejun Heo #define CREATE_TRACE_POINTS 35797bd2347STejun Heo #include <trace/events/workqueue.h> 35897bd2347STejun Heo 35968e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 360f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 361f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36268e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3635bcab335STejun Heo 364b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 365f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 366f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 367b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 36876af4d93STejun Heo 3695b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 370f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 371f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 372f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3735b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3745b95e1afSLai Jiangshan 375f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 376f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 377f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3787a62c2c8STejun Heo (pool)++) 3794ce62e9eSTejun Heo 38049e3cf44STejun Heo /** 38117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38217116969STejun Heo * @pool: iteration cursor 383611c92a0STejun Heo * @pi: integer used for iteration 384fa1b54e6STejun Heo * 38568e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 38668e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38768e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 388fa1b54e6STejun Heo * 389fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 390fa1b54e6STejun Heo * ignored. 39117116969STejun Heo */ 392611c92a0STejun Heo #define for_each_pool(pool, pi) \ 393611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39468e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 395fa1b54e6STejun Heo else 39617116969STejun Heo 39717116969STejun Heo /** 398822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 399822d8405STejun Heo * @worker: iteration cursor 400822d8405STejun Heo * @pool: worker_pool to iterate workers of 401822d8405STejun Heo * 4021258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 403822d8405STejun Heo * 404822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 405822d8405STejun Heo * ignored. 406822d8405STejun Heo */ 407da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 408da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4091258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 410822d8405STejun Heo else 411822d8405STejun Heo 412822d8405STejun Heo /** 41349e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41449e3cf44STejun Heo * @pwq: iteration cursor 41549e3cf44STejun Heo * @wq: the target workqueue 41676af4d93STejun Heo * 417b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 418794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 419794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 42076af4d93STejun Heo * 42176af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42276af4d93STejun Heo * ignored. 42349e3cf44STejun Heo */ 42449e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42576af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 426b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 42776af4d93STejun Heo else 428f3421797STejun Heo 429dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 430dc186ad7SThomas Gleixner 431dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 432dc186ad7SThomas Gleixner 43399777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43499777288SStanislaw Gruszka { 43599777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43699777288SStanislaw Gruszka } 43799777288SStanislaw Gruszka 438b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 439b9fdac7fSDu, Changbin { 440b9fdac7fSDu, Changbin struct work_struct *work = addr; 441b9fdac7fSDu, Changbin 442b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 443b9fdac7fSDu, Changbin } 444b9fdac7fSDu, Changbin 445dc186ad7SThomas Gleixner /* 446dc186ad7SThomas Gleixner * fixup_init is called when: 447dc186ad7SThomas Gleixner * - an active object is initialized 448dc186ad7SThomas Gleixner */ 44902a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 450dc186ad7SThomas Gleixner { 451dc186ad7SThomas Gleixner struct work_struct *work = addr; 452dc186ad7SThomas Gleixner 453dc186ad7SThomas Gleixner switch (state) { 454dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 455dc186ad7SThomas Gleixner cancel_work_sync(work); 456dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45702a982a6SDu, Changbin return true; 458dc186ad7SThomas Gleixner default: 45902a982a6SDu, Changbin return false; 460dc186ad7SThomas Gleixner } 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner /* 464dc186ad7SThomas Gleixner * fixup_free is called when: 465dc186ad7SThomas Gleixner * - an active object is freed 466dc186ad7SThomas Gleixner */ 46702a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 468dc186ad7SThomas Gleixner { 469dc186ad7SThomas Gleixner struct work_struct *work = addr; 470dc186ad7SThomas Gleixner 471dc186ad7SThomas Gleixner switch (state) { 472dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 473dc186ad7SThomas Gleixner cancel_work_sync(work); 474dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47502a982a6SDu, Changbin return true; 476dc186ad7SThomas Gleixner default: 47702a982a6SDu, Changbin return false; 478dc186ad7SThomas Gleixner } 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner 481dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 482dc186ad7SThomas Gleixner .name = "work_struct", 48399777288SStanislaw Gruszka .debug_hint = work_debug_hint, 484b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 485dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 486dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 487dc186ad7SThomas Gleixner }; 488dc186ad7SThomas Gleixner 489dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 490dc186ad7SThomas Gleixner { 491dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 492dc186ad7SThomas Gleixner } 493dc186ad7SThomas Gleixner 494dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 495dc186ad7SThomas Gleixner { 496dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner 499dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 500dc186ad7SThomas Gleixner { 501dc186ad7SThomas Gleixner if (onstack) 502dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 503dc186ad7SThomas Gleixner else 504dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 505dc186ad7SThomas Gleixner } 506dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 507dc186ad7SThomas Gleixner 508dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 511dc186ad7SThomas Gleixner } 512dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 513dc186ad7SThomas Gleixner 514ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 515ea2e64f2SThomas Gleixner { 516ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 517ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 518ea2e64f2SThomas Gleixner } 519ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 520ea2e64f2SThomas Gleixner 521dc186ad7SThomas Gleixner #else 522dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 523dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 524dc186ad7SThomas Gleixner #endif 525dc186ad7SThomas Gleixner 5264e8b22bdSLi Bin /** 5274e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5284e8b22bdSLi Bin * @pool: the pool pointer of interest 5294e8b22bdSLi Bin * 5304e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5314e8b22bdSLi Bin * successfully, -errno on failure. 5324e8b22bdSLi Bin */ 5339daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5349daf9e67STejun Heo { 5359daf9e67STejun Heo int ret; 5369daf9e67STejun Heo 53768e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5385bcab335STejun Heo 5394e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5404e8b22bdSLi Bin GFP_KERNEL); 541229641a6STejun Heo if (ret >= 0) { 542e68035fbSTejun Heo pool->id = ret; 543229641a6STejun Heo return 0; 544229641a6STejun Heo } 5459daf9e67STejun Heo return ret; 5469daf9e67STejun Heo } 5479daf9e67STejun Heo 54876af4d93STejun Heo /** 549df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 550df2d5ae4STejun Heo * @wq: the target workqueue 551df2d5ae4STejun Heo * @node: the node ID 552df2d5ae4STejun Heo * 5535b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5545b95e1afSLai Jiangshan * read locked. 555df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 556df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 557d185af30SYacine Belkadi * 558d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 559df2d5ae4STejun Heo */ 560df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 561df2d5ae4STejun Heo int node) 562df2d5ae4STejun Heo { 5635b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 564d6e022f1STejun Heo 565d6e022f1STejun Heo /* 566d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 567d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 568d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 569d6e022f1STejun Heo * happens, this workaround can be removed. 570d6e022f1STejun Heo */ 571d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 572d6e022f1STejun Heo return wq->dfl_pwq; 573d6e022f1STejun Heo 574df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 575df2d5ae4STejun Heo } 576df2d5ae4STejun Heo 57773f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57873f53c4aSTejun Heo { 57973f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 58073f53c4aSTejun Heo } 58173f53c4aSTejun Heo 58273f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58373f53c4aSTejun Heo { 58473f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58573f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58673f53c4aSTejun Heo } 58773f53c4aSTejun Heo 58873f53c4aSTejun Heo static int work_next_color(int color) 58973f53c4aSTejun Heo { 59073f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 591a848e3b6SOleg Nesterov } 592a848e3b6SOleg Nesterov 5934594bf15SDavid Howells /* 594112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 595112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5967c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5977a22ad75STejun Heo * 598112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 599112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 600bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 601bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6027a22ad75STejun Heo * 603112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6047c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 605112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6067c3eed5cSTejun Heo * available only while the work item is queued. 607bbb68dfaSTejun Heo * 608bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 609bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 610bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 611bbb68dfaSTejun Heo * try to steal the PENDING bit. 6124594bf15SDavid Howells */ 6137a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6147a22ad75STejun Heo unsigned long flags) 6157a22ad75STejun Heo { 6166183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6177a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6187a22ad75STejun Heo } 6197a22ad75STejun Heo 620112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6214690c4abSTejun Heo unsigned long extra_flags) 622365970a1SDavid Howells { 623112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 624112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 625365970a1SDavid Howells } 626365970a1SDavid Howells 6274468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6284468a00fSLai Jiangshan int pool_id) 6294468a00fSLai Jiangshan { 6304468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6314468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6324468a00fSLai Jiangshan } 6334468a00fSLai Jiangshan 6347c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6357c3eed5cSTejun Heo int pool_id) 6364d707b9fSOleg Nesterov { 63723657bb1STejun Heo /* 63823657bb1STejun Heo * The following wmb is paired with the implied mb in 63923657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 64023657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64123657bb1STejun Heo * owner. 64223657bb1STejun Heo */ 64323657bb1STejun Heo smp_wmb(); 6447c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 645346c09f8SRoman Pen /* 646346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 647346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 648346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 649346c09f8SRoman Pen * reordering can lead to a missed execution on attempt to qeueue 650346c09f8SRoman Pen * the same @work. E.g. consider this case: 651346c09f8SRoman Pen * 652346c09f8SRoman Pen * CPU#0 CPU#1 653346c09f8SRoman Pen * ---------------------------- -------------------------------- 654346c09f8SRoman Pen * 655346c09f8SRoman Pen * 1 STORE event_indicated 656346c09f8SRoman Pen * 2 queue_work_on() { 657346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 658346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 659346c09f8SRoman Pen * 5 set_work_data() # clear bit 660346c09f8SRoman Pen * 6 smp_mb() 661346c09f8SRoman Pen * 7 work->current_func() { 662346c09f8SRoman Pen * 8 LOAD event_indicated 663346c09f8SRoman Pen * } 664346c09f8SRoman Pen * 665346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 666346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 667346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 668346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 669346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 670346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 671346c09f8SRoman Pen * before actual STORE. 672346c09f8SRoman Pen */ 673346c09f8SRoman Pen smp_mb(); 6744d707b9fSOleg Nesterov } 6754d707b9fSOleg Nesterov 6767a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 677365970a1SDavid Howells { 6787c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6797c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6807a22ad75STejun Heo } 6817a22ad75STejun Heo 682112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6837a22ad75STejun Heo { 684e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6857a22ad75STejun Heo 686112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 687e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 688e120153dSTejun Heo else 689e120153dSTejun Heo return NULL; 6907a22ad75STejun Heo } 6917a22ad75STejun Heo 6927c3eed5cSTejun Heo /** 6937c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6947c3eed5cSTejun Heo * @work: the work item of interest 6957c3eed5cSTejun Heo * 69668e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69768e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 69868e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 699fa1b54e6STejun Heo * 700fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 701fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 702fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 703fa1b54e6STejun Heo * returned pool is and stays online. 704d185af30SYacine Belkadi * 705d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7067c3eed5cSTejun Heo */ 7077c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7087a22ad75STejun Heo { 709e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7107c3eed5cSTejun Heo int pool_id; 7117a22ad75STejun Heo 71268e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 713fa1b54e6STejun Heo 714112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 715112202d9STejun Heo return ((struct pool_workqueue *) 7167c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7177a22ad75STejun Heo 7187c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7197c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7207a22ad75STejun Heo return NULL; 7217a22ad75STejun Heo 722fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7237c3eed5cSTejun Heo } 7247c3eed5cSTejun Heo 7257c3eed5cSTejun Heo /** 7267c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7277c3eed5cSTejun Heo * @work: the work item of interest 7287c3eed5cSTejun Heo * 729d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7307c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7317c3eed5cSTejun Heo */ 7327c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7337c3eed5cSTejun Heo { 73454d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7357c3eed5cSTejun Heo 736112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 737112202d9STejun Heo return ((struct pool_workqueue *) 73854d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 73954d5b7d0SLai Jiangshan 74054d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7417c3eed5cSTejun Heo } 7427c3eed5cSTejun Heo 743bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 744bbb68dfaSTejun Heo { 7457c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 746bbb68dfaSTejun Heo 7477c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7487c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 749bbb68dfaSTejun Heo } 750bbb68dfaSTejun Heo 751bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 752bbb68dfaSTejun Heo { 753bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 754bbb68dfaSTejun Heo 755112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 756bbb68dfaSTejun Heo } 757bbb68dfaSTejun Heo 758e22bee78STejun Heo /* 7593270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7603270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 761d565ed63STejun Heo * they're being called with pool->lock held. 762e22bee78STejun Heo */ 763e22bee78STejun Heo 76463d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 765649027d7STejun Heo { 766e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 767649027d7STejun Heo } 768649027d7STejun Heo 769e22bee78STejun Heo /* 770e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 771e22bee78STejun Heo * running workers. 772974271c4STejun Heo * 773974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 774706026c2STejun Heo * function will always return %true for unbound pools as long as the 775974271c4STejun Heo * worklist isn't empty. 776e22bee78STejun Heo */ 77763d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 778e22bee78STejun Heo { 77963d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 780e22bee78STejun Heo } 781e22bee78STejun Heo 782e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78363d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 784e22bee78STejun Heo { 78563d95a91STejun Heo return pool->nr_idle; 786e22bee78STejun Heo } 787e22bee78STejun Heo 788e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 78963d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 790e22bee78STejun Heo { 791e19e397aSTejun Heo return !list_empty(&pool->worklist) && 792e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 793e22bee78STejun Heo } 794e22bee78STejun Heo 795e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79663d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 797e22bee78STejun Heo { 79863d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 799e22bee78STejun Heo } 800e22bee78STejun Heo 801e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80263d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 803e22bee78STejun Heo { 804692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 80563d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80663d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 807e22bee78STejun Heo 808e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 809e22bee78STejun Heo } 810e22bee78STejun Heo 811e22bee78STejun Heo /* 812e22bee78STejun Heo * Wake up functions. 813e22bee78STejun Heo */ 814e22bee78STejun Heo 8151037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8161037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8177e11629dSTejun Heo { 81863d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8197e11629dSTejun Heo return NULL; 8207e11629dSTejun Heo 82163d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8227e11629dSTejun Heo } 8237e11629dSTejun Heo 8247e11629dSTejun Heo /** 8257e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82663d95a91STejun Heo * @pool: worker pool to wake worker from 8277e11629dSTejun Heo * 82863d95a91STejun Heo * Wake up the first idle worker of @pool. 8297e11629dSTejun Heo * 8307e11629dSTejun Heo * CONTEXT: 831d565ed63STejun Heo * spin_lock_irq(pool->lock). 8327e11629dSTejun Heo */ 83363d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8347e11629dSTejun Heo { 8351037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8367e11629dSTejun Heo 8377e11629dSTejun Heo if (likely(worker)) 8387e11629dSTejun Heo wake_up_process(worker->task); 8397e11629dSTejun Heo } 8407e11629dSTejun Heo 8414690c4abSTejun Heo /** 842e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 843e22bee78STejun Heo * @task: task waking up 844e22bee78STejun Heo * @cpu: CPU @task is waking up to 845e22bee78STejun Heo * 846e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 847e22bee78STejun Heo * being awoken. 848e22bee78STejun Heo * 849e22bee78STejun Heo * CONTEXT: 850e22bee78STejun Heo * spin_lock_irq(rq->lock) 851e22bee78STejun Heo */ 852d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 853e22bee78STejun Heo { 854e22bee78STejun Heo struct worker *worker = kthread_data(task); 855e22bee78STejun Heo 85636576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 857ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 858e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 859e22bee78STejun Heo } 86036576000SJoonsoo Kim } 861e22bee78STejun Heo 862e22bee78STejun Heo /** 863e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 864e22bee78STejun Heo * @task: task going to sleep 865e22bee78STejun Heo * 866e22bee78STejun Heo * This function is called during schedule() when a busy worker is 867e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 868e22bee78STejun Heo * returning pointer to its task. 869e22bee78STejun Heo * 870e22bee78STejun Heo * CONTEXT: 871e22bee78STejun Heo * spin_lock_irq(rq->lock) 872e22bee78STejun Heo * 873d185af30SYacine Belkadi * Return: 874e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 875e22bee78STejun Heo */ 8769b7f6597SAlexander Gordeev struct task_struct *wq_worker_sleeping(struct task_struct *task) 877e22bee78STejun Heo { 878e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 879111c225aSTejun Heo struct worker_pool *pool; 880e22bee78STejun Heo 881111c225aSTejun Heo /* 882111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 883111c225aSTejun Heo * workers, also reach here, let's not access anything before 884111c225aSTejun Heo * checking NOT_RUNNING. 885111c225aSTejun Heo */ 8862d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 887e22bee78STejun Heo return NULL; 888e22bee78STejun Heo 889111c225aSTejun Heo pool = worker->pool; 890111c225aSTejun Heo 891e22bee78STejun Heo /* this can only happen on the local cpu */ 8929b7f6597SAlexander Gordeev if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) 8936183c009STejun Heo return NULL; 894e22bee78STejun Heo 895e22bee78STejun Heo /* 896e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 897e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 898e22bee78STejun Heo * Please read comment there. 899e22bee78STejun Heo * 900628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 901628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 902628c78e7STejun Heo * disabled, which in turn means that none else could be 903d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 904628c78e7STejun Heo * lock is safe. 905e22bee78STejun Heo */ 906e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 907e19e397aSTejun Heo !list_empty(&pool->worklist)) 9081037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 909e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 910e22bee78STejun Heo } 911e22bee78STejun Heo 912e22bee78STejun Heo /** 913e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 914cb444766STejun Heo * @worker: self 915d302f017STejun Heo * @flags: flags to set 916d302f017STejun Heo * 917228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 918d302f017STejun Heo * 919cb444766STejun Heo * CONTEXT: 920d565ed63STejun Heo * spin_lock_irq(pool->lock) 921d302f017STejun Heo */ 922228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 923d302f017STejun Heo { 924bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 925e22bee78STejun Heo 926cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 927cb444766STejun Heo 928228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 929e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 930e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 931e19e397aSTejun Heo atomic_dec(&pool->nr_running); 932e22bee78STejun Heo } 933e22bee78STejun Heo 934d302f017STejun Heo worker->flags |= flags; 935d302f017STejun Heo } 936d302f017STejun Heo 937d302f017STejun Heo /** 938e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 939cb444766STejun Heo * @worker: self 940d302f017STejun Heo * @flags: flags to clear 941d302f017STejun Heo * 942e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 943d302f017STejun Heo * 944cb444766STejun Heo * CONTEXT: 945d565ed63STejun Heo * spin_lock_irq(pool->lock) 946d302f017STejun Heo */ 947d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 948d302f017STejun Heo { 94963d95a91STejun Heo struct worker_pool *pool = worker->pool; 950e22bee78STejun Heo unsigned int oflags = worker->flags; 951e22bee78STejun Heo 952cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 953cb444766STejun Heo 954d302f017STejun Heo worker->flags &= ~flags; 955e22bee78STejun Heo 95642c025f3STejun Heo /* 95742c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 95842c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 95942c025f3STejun Heo * of multiple flags, not a single flag. 96042c025f3STejun Heo */ 961e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 962e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 963e19e397aSTejun Heo atomic_inc(&pool->nr_running); 964d302f017STejun Heo } 965d302f017STejun Heo 966d302f017STejun Heo /** 9678cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 968c9e7cf27STejun Heo * @pool: pool of interest 9698cca0eeaSTejun Heo * @work: work to find worker for 9708cca0eeaSTejun Heo * 971c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 972c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 973a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 974a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 975a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 976a2c1c57bSTejun Heo * being executed. 977a2c1c57bSTejun Heo * 978a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 979a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 980a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 981a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 982a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 983a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 984a2c1c57bSTejun Heo * 985c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 986c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 987c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 988c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 989c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 990c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9918cca0eeaSTejun Heo * 9928cca0eeaSTejun Heo * CONTEXT: 993d565ed63STejun Heo * spin_lock_irq(pool->lock). 9948cca0eeaSTejun Heo * 995d185af30SYacine Belkadi * Return: 996d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9978cca0eeaSTejun Heo * otherwise. 9988cca0eeaSTejun Heo */ 999c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10008cca0eeaSTejun Heo struct work_struct *work) 10018cca0eeaSTejun Heo { 100242f8570fSSasha Levin struct worker *worker; 100342f8570fSSasha Levin 1004b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1005a2c1c57bSTejun Heo (unsigned long)work) 1006a2c1c57bSTejun Heo if (worker->current_work == work && 1007a2c1c57bSTejun Heo worker->current_func == work->func) 100842f8570fSSasha Levin return worker; 100942f8570fSSasha Levin 101042f8570fSSasha Levin return NULL; 10118cca0eeaSTejun Heo } 10128cca0eeaSTejun Heo 10138cca0eeaSTejun Heo /** 1014bf4ede01STejun Heo * move_linked_works - move linked works to a list 1015bf4ede01STejun Heo * @work: start of series of works to be scheduled 1016bf4ede01STejun Heo * @head: target list to append @work to 1017402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1018bf4ede01STejun Heo * 1019bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1020bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1021bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1022bf4ede01STejun Heo * 1023bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1024bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1025bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1026bf4ede01STejun Heo * 1027bf4ede01STejun Heo * CONTEXT: 1028d565ed63STejun Heo * spin_lock_irq(pool->lock). 1029bf4ede01STejun Heo */ 1030bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1031bf4ede01STejun Heo struct work_struct **nextp) 1032bf4ede01STejun Heo { 1033bf4ede01STejun Heo struct work_struct *n; 1034bf4ede01STejun Heo 1035bf4ede01STejun Heo /* 1036bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1037bf4ede01STejun Heo * use NULL for list head. 1038bf4ede01STejun Heo */ 1039bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1040bf4ede01STejun Heo list_move_tail(&work->entry, head); 1041bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1042bf4ede01STejun Heo break; 1043bf4ede01STejun Heo } 1044bf4ede01STejun Heo 1045bf4ede01STejun Heo /* 1046bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1047bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1048bf4ede01STejun Heo * needs to be updated. 1049bf4ede01STejun Heo */ 1050bf4ede01STejun Heo if (nextp) 1051bf4ede01STejun Heo *nextp = n; 1052bf4ede01STejun Heo } 1053bf4ede01STejun Heo 10548864b4e5STejun Heo /** 10558864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10568864b4e5STejun Heo * @pwq: pool_workqueue to get 10578864b4e5STejun Heo * 10588864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10598864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10608864b4e5STejun Heo */ 10618864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10628864b4e5STejun Heo { 10638864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10648864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10658864b4e5STejun Heo pwq->refcnt++; 10668864b4e5STejun Heo } 10678864b4e5STejun Heo 10688864b4e5STejun Heo /** 10698864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10708864b4e5STejun Heo * @pwq: pool_workqueue to put 10718864b4e5STejun Heo * 10728864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10738864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10748864b4e5STejun Heo */ 10758864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10768864b4e5STejun Heo { 10778864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10788864b4e5STejun Heo if (likely(--pwq->refcnt)) 10798864b4e5STejun Heo return; 10808864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10818864b4e5STejun Heo return; 10828864b4e5STejun Heo /* 10838864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10848864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10858864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10868864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10878864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10888864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10898864b4e5STejun Heo */ 10908864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10918864b4e5STejun Heo } 10928864b4e5STejun Heo 1093dce90d47STejun Heo /** 1094dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1095dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1096dce90d47STejun Heo * 1097dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1098dce90d47STejun Heo */ 1099dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1100dce90d47STejun Heo { 1101dce90d47STejun Heo if (pwq) { 1102dce90d47STejun Heo /* 1103dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1104dce90d47STejun Heo * following lock operations are safe. 1105dce90d47STejun Heo */ 1106dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1107dce90d47STejun Heo put_pwq(pwq); 1108dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1109dce90d47STejun Heo } 1110dce90d47STejun Heo } 1111dce90d47STejun Heo 1112112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1113bf4ede01STejun Heo { 1114112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1115bf4ede01STejun Heo 1116bf4ede01STejun Heo trace_workqueue_activate_work(work); 111782607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 111882607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1119112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1120bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1121112202d9STejun Heo pwq->nr_active++; 1122bf4ede01STejun Heo } 1123bf4ede01STejun Heo 1124112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11253aa62497SLai Jiangshan { 1126112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11273aa62497SLai Jiangshan struct work_struct, entry); 11283aa62497SLai Jiangshan 1129112202d9STejun Heo pwq_activate_delayed_work(work); 11303aa62497SLai Jiangshan } 11313aa62497SLai Jiangshan 1132bf4ede01STejun Heo /** 1133112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1134112202d9STejun Heo * @pwq: pwq of interest 1135bf4ede01STejun Heo * @color: color of work which left the queue 1136bf4ede01STejun Heo * 1137bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1138112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1139bf4ede01STejun Heo * 1140bf4ede01STejun Heo * CONTEXT: 1141d565ed63STejun Heo * spin_lock_irq(pool->lock). 1142bf4ede01STejun Heo */ 1143112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1144bf4ede01STejun Heo { 11458864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1146bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11478864b4e5STejun Heo goto out_put; 1148bf4ede01STejun Heo 1149112202d9STejun Heo pwq->nr_in_flight[color]--; 1150bf4ede01STejun Heo 1151112202d9STejun Heo pwq->nr_active--; 1152112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1153bf4ede01STejun Heo /* one down, submit a delayed one */ 1154112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1155112202d9STejun Heo pwq_activate_first_delayed(pwq); 1156bf4ede01STejun Heo } 1157bf4ede01STejun Heo 1158bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1159112202d9STejun Heo if (likely(pwq->flush_color != color)) 11608864b4e5STejun Heo goto out_put; 1161bf4ede01STejun Heo 1162bf4ede01STejun Heo /* are there still in-flight works? */ 1163112202d9STejun Heo if (pwq->nr_in_flight[color]) 11648864b4e5STejun Heo goto out_put; 1165bf4ede01STejun Heo 1166112202d9STejun Heo /* this pwq is done, clear flush_color */ 1167112202d9STejun Heo pwq->flush_color = -1; 1168bf4ede01STejun Heo 1169bf4ede01STejun Heo /* 1170112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1171bf4ede01STejun Heo * will handle the rest. 1172bf4ede01STejun Heo */ 1173112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1174112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11758864b4e5STejun Heo out_put: 11768864b4e5STejun Heo put_pwq(pwq); 1177bf4ede01STejun Heo } 1178bf4ede01STejun Heo 117936e227d2STejun Heo /** 1180bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 118136e227d2STejun Heo * @work: work item to steal 118236e227d2STejun Heo * @is_dwork: @work is a delayed_work 1183bbb68dfaSTejun Heo * @flags: place to store irq state 118436e227d2STejun Heo * 118536e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1186d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 118736e227d2STejun Heo * 1188d185af30SYacine Belkadi * Return: 118936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 119036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 119136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1192bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1193bbb68dfaSTejun Heo * for arbitrarily long 119436e227d2STejun Heo * 1195d185af30SYacine Belkadi * Note: 1196bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1197e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1198e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1199e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1200bbb68dfaSTejun Heo * 1201bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1202bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1203bbb68dfaSTejun Heo * 1204e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1205bf4ede01STejun Heo */ 1206bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1207bbb68dfaSTejun Heo unsigned long *flags) 1208bf4ede01STejun Heo { 1209d565ed63STejun Heo struct worker_pool *pool; 1210112202d9STejun Heo struct pool_workqueue *pwq; 1211bf4ede01STejun Heo 1212bbb68dfaSTejun Heo local_irq_save(*flags); 1213bbb68dfaSTejun Heo 121436e227d2STejun Heo /* try to steal the timer if it exists */ 121536e227d2STejun Heo if (is_dwork) { 121636e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 121736e227d2STejun Heo 1218e0aecdd8STejun Heo /* 1219e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1220e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1221e0aecdd8STejun Heo * running on the local CPU. 1222e0aecdd8STejun Heo */ 122336e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 122436e227d2STejun Heo return 1; 122536e227d2STejun Heo } 122636e227d2STejun Heo 122736e227d2STejun Heo /* try to claim PENDING the normal way */ 1228bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1229bf4ede01STejun Heo return 0; 1230bf4ede01STejun Heo 1231bf4ede01STejun Heo /* 1232bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1233bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1234bf4ede01STejun Heo */ 1235d565ed63STejun Heo pool = get_work_pool(work); 1236d565ed63STejun Heo if (!pool) 1237bbb68dfaSTejun Heo goto fail; 1238bf4ede01STejun Heo 1239d565ed63STejun Heo spin_lock(&pool->lock); 1240bf4ede01STejun Heo /* 1241112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1242112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1243112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1244112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1245112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12460b3dae68SLai Jiangshan * item is currently queued on that pool. 1247bf4ede01STejun Heo */ 1248112202d9STejun Heo pwq = get_work_pwq(work); 1249112202d9STejun Heo if (pwq && pwq->pool == pool) { 1250bf4ede01STejun Heo debug_work_deactivate(work); 12513aa62497SLai Jiangshan 12523aa62497SLai Jiangshan /* 125316062836STejun Heo * A delayed work item cannot be grabbed directly because 125416062836STejun Heo * it might have linked NO_COLOR work items which, if left 1255112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 125616062836STejun Heo * management later on and cause stall. Make sure the work 125716062836STejun Heo * item is activated before grabbing. 12583aa62497SLai Jiangshan */ 12593aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1260112202d9STejun Heo pwq_activate_delayed_work(work); 12613aa62497SLai Jiangshan 1262bf4ede01STejun Heo list_del_init(&work->entry); 12639c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 126436e227d2STejun Heo 1265112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12664468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12674468a00fSLai Jiangshan 1268d565ed63STejun Heo spin_unlock(&pool->lock); 126936e227d2STejun Heo return 1; 1270bf4ede01STejun Heo } 1271d565ed63STejun Heo spin_unlock(&pool->lock); 1272bbb68dfaSTejun Heo fail: 1273bbb68dfaSTejun Heo local_irq_restore(*flags); 1274bbb68dfaSTejun Heo if (work_is_canceling(work)) 1275bbb68dfaSTejun Heo return -ENOENT; 1276bbb68dfaSTejun Heo cpu_relax(); 127736e227d2STejun Heo return -EAGAIN; 1278bf4ede01STejun Heo } 1279bf4ede01STejun Heo 1280bf4ede01STejun Heo /** 1281706026c2STejun Heo * insert_work - insert a work into a pool 1282112202d9STejun Heo * @pwq: pwq @work belongs to 12834690c4abSTejun Heo * @work: work to insert 12844690c4abSTejun Heo * @head: insertion point 12854690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12864690c4abSTejun Heo * 1287112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1288706026c2STejun Heo * work_struct flags. 12894690c4abSTejun Heo * 12904690c4abSTejun Heo * CONTEXT: 1291d565ed63STejun Heo * spin_lock_irq(pool->lock). 1292365970a1SDavid Howells */ 1293112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1294112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1295b89deed3SOleg Nesterov { 1296112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1297e1d8aa9fSFrederic Weisbecker 12984690c4abSTejun Heo /* we own @work, set data and link */ 1299112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13001a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13018864b4e5STejun Heo get_pwq(pwq); 1302e22bee78STejun Heo 1303e22bee78STejun Heo /* 1304c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1305c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1306c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1307e22bee78STejun Heo */ 1308e22bee78STejun Heo smp_mb(); 1309e22bee78STejun Heo 131063d95a91STejun Heo if (__need_more_worker(pool)) 131163d95a91STejun Heo wake_up_worker(pool); 1312b89deed3SOleg Nesterov } 1313b89deed3SOleg Nesterov 1314c8efcc25STejun Heo /* 1315c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13168d03ecfeSTejun Heo * same workqueue. 1317c8efcc25STejun Heo */ 1318c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1319c8efcc25STejun Heo { 1320c8efcc25STejun Heo struct worker *worker; 1321c8efcc25STejun Heo 13228d03ecfeSTejun Heo worker = current_wq_worker(); 1323c8efcc25STejun Heo /* 13248d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13258d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1326c8efcc25STejun Heo */ 1327112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1328c8efcc25STejun Heo } 1329c8efcc25STejun Heo 1330ef557180SMike Galbraith /* 1331ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1332ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1333ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1334ef557180SMike Galbraith */ 1335ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1336ef557180SMike Galbraith { 1337f303fccbSTejun Heo static bool printed_dbg_warning; 1338ef557180SMike Galbraith int new_cpu; 1339ef557180SMike Galbraith 1340f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1341ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1342ef557180SMike Galbraith return cpu; 1343f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1344f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1345f303fccbSTejun Heo printed_dbg_warning = true; 1346f303fccbSTejun Heo } 1347f303fccbSTejun Heo 1348ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1349ef557180SMike Galbraith return cpu; 1350ef557180SMike Galbraith 1351ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1352ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1353ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1354ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1355ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1356ef557180SMike Galbraith return cpu; 1357ef557180SMike Galbraith } 1358ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1359ef557180SMike Galbraith 1360ef557180SMike Galbraith return new_cpu; 1361ef557180SMike Galbraith } 1362ef557180SMike Galbraith 1363d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13641da177e4SLinus Torvalds struct work_struct *work) 13651da177e4SLinus Torvalds { 1366112202d9STejun Heo struct pool_workqueue *pwq; 1367c9178087STejun Heo struct worker_pool *last_pool; 13681e19ffc6STejun Heo struct list_head *worklist; 13698a2e8e5dSTejun Heo unsigned int work_flags; 1370b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13718930cabaSTejun Heo 13728930cabaSTejun Heo /* 13738930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13748930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13758930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13768930cabaSTejun Heo * happen with IRQ disabled. 13778930cabaSTejun Heo */ 13788e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 13791da177e4SLinus Torvalds 1380dc186ad7SThomas Gleixner debug_work_activate(work); 13811e19ffc6STejun Heo 13829ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1383618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1384c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1385e41e704bSTejun Heo return; 13869e8cd2f5STejun Heo retry: 1387df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1388ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1389df2d5ae4STejun Heo 1390df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1391df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1392c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1393df2d5ae4STejun Heo else 1394df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1395f3421797STejun Heo 139618aa9effSTejun Heo /* 1397c9178087STejun Heo * If @work was previously on a different pool, it might still be 1398c9178087STejun Heo * running there, in which case the work needs to be queued on that 1399c9178087STejun Heo * pool to guarantee non-reentrancy. 140018aa9effSTejun Heo */ 1401c9e7cf27STejun Heo last_pool = get_work_pool(work); 1402112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 140318aa9effSTejun Heo struct worker *worker; 140418aa9effSTejun Heo 1405d565ed63STejun Heo spin_lock(&last_pool->lock); 140618aa9effSTejun Heo 1407c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 140818aa9effSTejun Heo 1409112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1410c9178087STejun Heo pwq = worker->current_pwq; 14118594fadeSLai Jiangshan } else { 141218aa9effSTejun Heo /* meh... not running there, queue here */ 1413d565ed63STejun Heo spin_unlock(&last_pool->lock); 1414112202d9STejun Heo spin_lock(&pwq->pool->lock); 141518aa9effSTejun Heo } 14168930cabaSTejun Heo } else { 1417112202d9STejun Heo spin_lock(&pwq->pool->lock); 14188930cabaSTejun Heo } 1419502ca9d8STejun Heo 14209e8cd2f5STejun Heo /* 14219e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14229e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14239e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1424df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1425df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14269e8cd2f5STejun Heo * make forward-progress. 14279e8cd2f5STejun Heo */ 14289e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14299e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 14309e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 14319e8cd2f5STejun Heo cpu_relax(); 14329e8cd2f5STejun Heo goto retry; 14339e8cd2f5STejun Heo } 14349e8cd2f5STejun Heo /* oops */ 14359e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14369e8cd2f5STejun Heo wq->name, cpu); 14379e8cd2f5STejun Heo } 14389e8cd2f5STejun Heo 1439112202d9STejun Heo /* pwq determined, queue */ 1440112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1441502ca9d8STejun Heo 1442f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1443112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1444f5b2552bSDan Carpenter return; 1445f5b2552bSDan Carpenter } 14461e19ffc6STejun Heo 1447112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1448112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14491e19ffc6STejun Heo 1450112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1451cdadf009STejun Heo trace_workqueue_activate_work(work); 1452112202d9STejun Heo pwq->nr_active++; 1453112202d9STejun Heo worklist = &pwq->pool->worklist; 145482607adcSTejun Heo if (list_empty(worklist)) 145582607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14568a2e8e5dSTejun Heo } else { 14578a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1458112202d9STejun Heo worklist = &pwq->delayed_works; 14598a2e8e5dSTejun Heo } 14601e19ffc6STejun Heo 1461112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14621e19ffc6STejun Heo 1463112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14641da177e4SLinus Torvalds } 14651da177e4SLinus Torvalds 14660fcb78c2SRolf Eike Beer /** 1467c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1468c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1469c1a220e7SZhang Rui * @wq: workqueue to use 1470c1a220e7SZhang Rui * @work: work to queue 1471c1a220e7SZhang Rui * 1472c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1473c1a220e7SZhang Rui * can't go away. 1474d185af30SYacine Belkadi * 1475d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1476c1a220e7SZhang Rui */ 1477d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1478d4283e93STejun Heo struct work_struct *work) 1479c1a220e7SZhang Rui { 1480d4283e93STejun Heo bool ret = false; 14818930cabaSTejun Heo unsigned long flags; 14828930cabaSTejun Heo 14838930cabaSTejun Heo local_irq_save(flags); 1484c1a220e7SZhang Rui 148522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14864690c4abSTejun Heo __queue_work(cpu, wq, work); 1487d4283e93STejun Heo ret = true; 1488c1a220e7SZhang Rui } 14898930cabaSTejun Heo 14908930cabaSTejun Heo local_irq_restore(flags); 1491c1a220e7SZhang Rui return ret; 1492c1a220e7SZhang Rui } 1493ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1494c1a220e7SZhang Rui 14958c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 14961da177e4SLinus Torvalds { 14978c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 14981da177e4SLinus Torvalds 1499e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 150060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 15011da177e4SLinus Torvalds } 15021438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 15031da177e4SLinus Torvalds 15047beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 150552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15061da177e4SLinus Torvalds { 15077beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 15087beb2edfSTejun Heo struct work_struct *work = &dwork->work; 15091da177e4SLinus Torvalds 1510637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 1511841b86f3SKees Cook WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1512fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1513fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 15147beb2edfSTejun Heo 15158852aac2STejun Heo /* 15168852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 15178852aac2STejun Heo * both optimization and correctness. The earliest @timer can 15188852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 15198852aac2STejun Heo * on that there's no such delay when @delay is 0. 15208852aac2STejun Heo */ 15218852aac2STejun Heo if (!delay) { 15228852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 15238852aac2STejun Heo return; 15248852aac2STejun Heo } 15258852aac2STejun Heo 152660c057bcSLai Jiangshan dwork->wq = wq; 15271265057fSTejun Heo dwork->cpu = cpu; 15287beb2edfSTejun Heo timer->expires = jiffies + delay; 15297beb2edfSTejun Heo 1530041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 15317beb2edfSTejun Heo add_timer_on(timer, cpu); 1532041bd12eSTejun Heo else 1533041bd12eSTejun Heo add_timer(timer); 15347beb2edfSTejun Heo } 15351da177e4SLinus Torvalds 15360fcb78c2SRolf Eike Beer /** 15370fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 15380fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 15390fcb78c2SRolf Eike Beer * @wq: workqueue to use 1540af9997e4SRandy Dunlap * @dwork: work to queue 15410fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 15420fcb78c2SRolf Eike Beer * 1543d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1544715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1545715f1300STejun Heo * execution. 15460fcb78c2SRolf Eike Beer */ 1547d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 154852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15497a6bc1cdSVenkatesh Pallipadi { 155052bad64dSDavid Howells struct work_struct *work = &dwork->work; 1551d4283e93STejun Heo bool ret = false; 15528930cabaSTejun Heo unsigned long flags; 15538930cabaSTejun Heo 15548930cabaSTejun Heo /* read the comment in __queue_work() */ 15558930cabaSTejun Heo local_irq_save(flags); 15567a6bc1cdSVenkatesh Pallipadi 155722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15587beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1559d4283e93STejun Heo ret = true; 15607a6bc1cdSVenkatesh Pallipadi } 15618930cabaSTejun Heo 15628930cabaSTejun Heo local_irq_restore(flags); 15637a6bc1cdSVenkatesh Pallipadi return ret; 15647a6bc1cdSVenkatesh Pallipadi } 1565ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15661da177e4SLinus Torvalds 1567c8e55f36STejun Heo /** 15688376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15698376fe22STejun Heo * @cpu: CPU number to execute work on 15708376fe22STejun Heo * @wq: workqueue to use 15718376fe22STejun Heo * @dwork: work to queue 15728376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15738376fe22STejun Heo * 15748376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15758376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15768376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15778376fe22STejun Heo * current state. 15788376fe22STejun Heo * 1579d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15808376fe22STejun Heo * pending and its timer was modified. 15818376fe22STejun Heo * 1582e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15838376fe22STejun Heo * See try_to_grab_pending() for details. 15848376fe22STejun Heo */ 15858376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15868376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15878376fe22STejun Heo { 15888376fe22STejun Heo unsigned long flags; 15898376fe22STejun Heo int ret; 15908376fe22STejun Heo 15918376fe22STejun Heo do { 15928376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15938376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15948376fe22STejun Heo 15958376fe22STejun Heo if (likely(ret >= 0)) { 15968376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15978376fe22STejun Heo local_irq_restore(flags); 15988376fe22STejun Heo } 15998376fe22STejun Heo 16008376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 16018376fe22STejun Heo return ret; 16028376fe22STejun Heo } 16038376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 16048376fe22STejun Heo 160505f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 160605f0fe6bSTejun Heo { 160705f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 160805f0fe6bSTejun Heo 160905f0fe6bSTejun Heo /* read the comment in __queue_work() */ 161005f0fe6bSTejun Heo local_irq_disable(); 161105f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 161205f0fe6bSTejun Heo local_irq_enable(); 161305f0fe6bSTejun Heo } 161405f0fe6bSTejun Heo 161505f0fe6bSTejun Heo /** 161605f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 161705f0fe6bSTejun Heo * @wq: workqueue to use 161805f0fe6bSTejun Heo * @rwork: work to queue 161905f0fe6bSTejun Heo * 162005f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 162105f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 162205f0fe6bSTejun Heo * While @rwork is guarnateed to be executed after a %false return, the 162305f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 162405f0fe6bSTejun Heo */ 162505f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 162605f0fe6bSTejun Heo { 162705f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 162805f0fe6bSTejun Heo 162905f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 163005f0fe6bSTejun Heo rwork->wq = wq; 163105f0fe6bSTejun Heo call_rcu(&rwork->rcu, rcu_work_rcufn); 163205f0fe6bSTejun Heo return true; 163305f0fe6bSTejun Heo } 163405f0fe6bSTejun Heo 163505f0fe6bSTejun Heo return false; 163605f0fe6bSTejun Heo } 163705f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 163805f0fe6bSTejun Heo 16398376fe22STejun Heo /** 1640c8e55f36STejun Heo * worker_enter_idle - enter idle state 1641c8e55f36STejun Heo * @worker: worker which is entering idle state 1642c8e55f36STejun Heo * 1643c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1644c8e55f36STejun Heo * necessary. 1645c8e55f36STejun Heo * 1646c8e55f36STejun Heo * LOCKING: 1647d565ed63STejun Heo * spin_lock_irq(pool->lock). 1648c8e55f36STejun Heo */ 1649c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 16501da177e4SLinus Torvalds { 1651bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1652c8e55f36STejun Heo 16536183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 16546183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 16556183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 16566183c009STejun Heo return; 1657c8e55f36STejun Heo 1658051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1659cb444766STejun Heo worker->flags |= WORKER_IDLE; 1660bd7bdd43STejun Heo pool->nr_idle++; 1661e22bee78STejun Heo worker->last_active = jiffies; 1662c8e55f36STejun Heo 1663c8e55f36STejun Heo /* idle_list is LIFO */ 1664bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1665db7bccf4STejun Heo 166663d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1667628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1668cb444766STejun Heo 1669544ecf31STejun Heo /* 1670e8b3f8dbSLai Jiangshan * Sanity check nr_running. Because unbind_workers() releases 1671d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1672628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1673628c78e7STejun Heo * unbind is not in progress. 1674544ecf31STejun Heo */ 167524647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1676bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1677e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1678c8e55f36STejun Heo } 1679c8e55f36STejun Heo 1680c8e55f36STejun Heo /** 1681c8e55f36STejun Heo * worker_leave_idle - leave idle state 1682c8e55f36STejun Heo * @worker: worker which is leaving idle state 1683c8e55f36STejun Heo * 1684c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1685c8e55f36STejun Heo * 1686c8e55f36STejun Heo * LOCKING: 1687d565ed63STejun Heo * spin_lock_irq(pool->lock). 1688c8e55f36STejun Heo */ 1689c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1690c8e55f36STejun Heo { 1691bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1692c8e55f36STejun Heo 16936183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16946183c009STejun Heo return; 1695d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1696bd7bdd43STejun Heo pool->nr_idle--; 1697c8e55f36STejun Heo list_del_init(&worker->entry); 1698c8e55f36STejun Heo } 1699c8e55f36STejun Heo 1700f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1701c34056a3STejun Heo { 1702c34056a3STejun Heo struct worker *worker; 1703c34056a3STejun Heo 1704f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1705c8e55f36STejun Heo if (worker) { 1706c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1707affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1708da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1709e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1710e22bee78STejun Heo worker->flags = WORKER_PREP; 1711c8e55f36STejun Heo } 1712c34056a3STejun Heo return worker; 1713c34056a3STejun Heo } 1714c34056a3STejun Heo 1715c34056a3STejun Heo /** 17164736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 17174736cbf7SLai Jiangshan * @worker: worker to be attached 17184736cbf7SLai Jiangshan * @pool: the target pool 17194736cbf7SLai Jiangshan * 17204736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 17214736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 17224736cbf7SLai Jiangshan * cpu-[un]hotplugs. 17234736cbf7SLai Jiangshan */ 17244736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 17254736cbf7SLai Jiangshan struct worker_pool *pool) 17264736cbf7SLai Jiangshan { 17271258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 17284736cbf7SLai Jiangshan 17294736cbf7SLai Jiangshan /* 17304736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 17314736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 17324736cbf7SLai Jiangshan */ 17334736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17344736cbf7SLai Jiangshan 17354736cbf7SLai Jiangshan /* 17361258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 17371258fae7STejun Heo * stable across this function. See the comments above the flag 17381258fae7STejun Heo * definition for details. 17394736cbf7SLai Jiangshan */ 17404736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 17414736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 17424736cbf7SLai Jiangshan 17434736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 17444736cbf7SLai Jiangshan 17451258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 17464736cbf7SLai Jiangshan } 17474736cbf7SLai Jiangshan 17484736cbf7SLai Jiangshan /** 174960f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 175060f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 175160f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 175260f5a4bcSLai Jiangshan * 17534736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 17544736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 17554736cbf7SLai Jiangshan * other reference to the pool. 175660f5a4bcSLai Jiangshan */ 175760f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 175860f5a4bcSLai Jiangshan struct worker_pool *pool) 175960f5a4bcSLai Jiangshan { 176060f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 176160f5a4bcSLai Jiangshan 17621258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1763da028469SLai Jiangshan list_del(&worker->node); 1764da028469SLai Jiangshan if (list_empty(&pool->workers)) 176560f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 17661258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 176760f5a4bcSLai Jiangshan 1768b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1769b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1770b62c0751SLai Jiangshan 177160f5a4bcSLai Jiangshan if (detach_completion) 177260f5a4bcSLai Jiangshan complete(detach_completion); 177360f5a4bcSLai Jiangshan } 177460f5a4bcSLai Jiangshan 177560f5a4bcSLai Jiangshan /** 1776c34056a3STejun Heo * create_worker - create a new workqueue worker 177763d95a91STejun Heo * @pool: pool the new worker will belong to 1778c34056a3STejun Heo * 1779051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1780c34056a3STejun Heo * 1781c34056a3STejun Heo * CONTEXT: 1782c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1783c34056a3STejun Heo * 1784d185af30SYacine Belkadi * Return: 1785c34056a3STejun Heo * Pointer to the newly created worker. 1786c34056a3STejun Heo */ 1787bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1788c34056a3STejun Heo { 1789c34056a3STejun Heo struct worker *worker = NULL; 1790f3421797STejun Heo int id = -1; 1791e3c916a4STejun Heo char id_buf[16]; 1792c34056a3STejun Heo 17937cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17947cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1795822d8405STejun Heo if (id < 0) 1796c34056a3STejun Heo goto fail; 1797c34056a3STejun Heo 1798f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1799c34056a3STejun Heo if (!worker) 1800c34056a3STejun Heo goto fail; 1801c34056a3STejun Heo 1802bd7bdd43STejun Heo worker->pool = pool; 1803c34056a3STejun Heo worker->id = id; 1804c34056a3STejun Heo 180529c91e99STejun Heo if (pool->cpu >= 0) 1806e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1807e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1808f3421797STejun Heo else 1809e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1810e3c916a4STejun Heo 1811f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1812e3c916a4STejun Heo "kworker/%s", id_buf); 1813c34056a3STejun Heo if (IS_ERR(worker->task)) 1814c34056a3STejun Heo goto fail; 1815c34056a3STejun Heo 181691151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 181725834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 181891151228SOleg Nesterov 1819da028469SLai Jiangshan /* successful, attach the worker to the pool */ 18204736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1821822d8405STejun Heo 1822051e1850SLai Jiangshan /* start the newly created worker */ 1823051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1824051e1850SLai Jiangshan worker->pool->nr_workers++; 1825051e1850SLai Jiangshan worker_enter_idle(worker); 1826051e1850SLai Jiangshan wake_up_process(worker->task); 1827051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1828051e1850SLai Jiangshan 1829c34056a3STejun Heo return worker; 1830822d8405STejun Heo 1831c34056a3STejun Heo fail: 18329625ab17SLai Jiangshan if (id >= 0) 18337cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1834c34056a3STejun Heo kfree(worker); 1835c34056a3STejun Heo return NULL; 1836c34056a3STejun Heo } 1837c34056a3STejun Heo 1838c34056a3STejun Heo /** 1839c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1840c34056a3STejun Heo * @worker: worker to be destroyed 1841c34056a3STejun Heo * 184273eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 184373eb7fe7SLai Jiangshan * be idle. 1844c8e55f36STejun Heo * 1845c8e55f36STejun Heo * CONTEXT: 184660f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1847c34056a3STejun Heo */ 1848c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1849c34056a3STejun Heo { 1850bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1851c34056a3STejun Heo 1852cd549687STejun Heo lockdep_assert_held(&pool->lock); 1853cd549687STejun Heo 1854c34056a3STejun Heo /* sanity check frenzy */ 18556183c009STejun Heo if (WARN_ON(worker->current_work) || 185673eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 185773eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 18586183c009STejun Heo return; 1859c34056a3STejun Heo 1860bd7bdd43STejun Heo pool->nr_workers--; 1861bd7bdd43STejun Heo pool->nr_idle--; 1862c8e55f36STejun Heo 1863c8e55f36STejun Heo list_del_init(&worker->entry); 1864cb444766STejun Heo worker->flags |= WORKER_DIE; 186560f5a4bcSLai Jiangshan wake_up_process(worker->task); 1866c34056a3STejun Heo } 1867c34056a3STejun Heo 186832a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 1869e22bee78STejun Heo { 187032a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 1871e22bee78STejun Heo 1872d565ed63STejun Heo spin_lock_irq(&pool->lock); 1873e22bee78STejun Heo 18743347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1875e22bee78STejun Heo struct worker *worker; 1876e22bee78STejun Heo unsigned long expires; 1877e22bee78STejun Heo 1878e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 187963d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1880e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1881e22bee78STejun Heo 18823347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 188363d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18843347fc9fSLai Jiangshan break; 1885e22bee78STejun Heo } 18863347fc9fSLai Jiangshan 18873347fc9fSLai Jiangshan destroy_worker(worker); 1888e22bee78STejun Heo } 1889e22bee78STejun Heo 1890d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1891e22bee78STejun Heo } 1892e22bee78STejun Heo 1893493a1724STejun Heo static void send_mayday(struct work_struct *work) 1894e22bee78STejun Heo { 1895112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1896112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1897493a1724STejun Heo 18982e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1899e22bee78STejun Heo 1900493008a8STejun Heo if (!wq->rescuer) 1901493a1724STejun Heo return; 1902e22bee78STejun Heo 1903e22bee78STejun Heo /* mayday mayday mayday */ 1904493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 190577668c8bSLai Jiangshan /* 190677668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 190777668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 190877668c8bSLai Jiangshan * rescuer is done with it. 190977668c8bSLai Jiangshan */ 191077668c8bSLai Jiangshan get_pwq(pwq); 1911493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1912e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1913493a1724STejun Heo } 1914e22bee78STejun Heo } 1915e22bee78STejun Heo 191632a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 1917e22bee78STejun Heo { 191832a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 1919e22bee78STejun Heo struct work_struct *work; 1920e22bee78STejun Heo 1921b2d82909STejun Heo spin_lock_irq(&pool->lock); 1922b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1923e22bee78STejun Heo 192463d95a91STejun Heo if (need_to_create_worker(pool)) { 1925e22bee78STejun Heo /* 1926e22bee78STejun Heo * We've been trying to create a new worker but 1927e22bee78STejun Heo * haven't been successful. We might be hitting an 1928e22bee78STejun Heo * allocation deadlock. Send distress signals to 1929e22bee78STejun Heo * rescuers. 1930e22bee78STejun Heo */ 193163d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1932e22bee78STejun Heo send_mayday(work); 1933e22bee78STejun Heo } 1934e22bee78STejun Heo 1935b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1936b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1937e22bee78STejun Heo 193863d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1939e22bee78STejun Heo } 1940e22bee78STejun Heo 1941e22bee78STejun Heo /** 1942e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 194363d95a91STejun Heo * @pool: pool to create a new worker for 1944e22bee78STejun Heo * 194563d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1946e22bee78STejun Heo * have at least one idle worker on return from this function. If 1947e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 194863d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1949e22bee78STejun Heo * possible allocation deadlock. 1950e22bee78STejun Heo * 1951c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1952c5aa87bbSTejun Heo * may_start_working() %true. 1953e22bee78STejun Heo * 1954e22bee78STejun Heo * LOCKING: 1955d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1956e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1957e22bee78STejun Heo * manager. 1958e22bee78STejun Heo */ 195929187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1960d565ed63STejun Heo __releases(&pool->lock) 1961d565ed63STejun Heo __acquires(&pool->lock) 1962e22bee78STejun Heo { 1963e22bee78STejun Heo restart: 1964d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19659f9c2364STejun Heo 1966e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 196763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1968e22bee78STejun Heo 1969e22bee78STejun Heo while (true) { 1970051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1971e22bee78STejun Heo break; 1972e22bee78STejun Heo 1973e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19749f9c2364STejun Heo 197563d95a91STejun Heo if (!need_to_create_worker(pool)) 1976e22bee78STejun Heo break; 1977e22bee78STejun Heo } 1978e22bee78STejun Heo 197963d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1980d565ed63STejun Heo spin_lock_irq(&pool->lock); 1981051e1850SLai Jiangshan /* 1982051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1983051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1984051e1850SLai Jiangshan * already become busy. 1985051e1850SLai Jiangshan */ 198663d95a91STejun Heo if (need_to_create_worker(pool)) 1987e22bee78STejun Heo goto restart; 1988e22bee78STejun Heo } 1989e22bee78STejun Heo 1990e22bee78STejun Heo /** 1991e22bee78STejun Heo * manage_workers - manage worker pool 1992e22bee78STejun Heo * @worker: self 1993e22bee78STejun Heo * 1994706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1995e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1996706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1997e22bee78STejun Heo * 1998e22bee78STejun Heo * The caller can safely start processing works on false return. On 1999e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2000e22bee78STejun Heo * and may_start_working() is true. 2001e22bee78STejun Heo * 2002e22bee78STejun Heo * CONTEXT: 2003d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2004e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2005e22bee78STejun Heo * 2006d185af30SYacine Belkadi * Return: 200729187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 200829187a9eSTejun Heo * start processing works, %true if management function was performed and 200929187a9eSTejun Heo * the conditions that the caller verified before calling the function may 201029187a9eSTejun Heo * no longer be true. 2011e22bee78STejun Heo */ 2012e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2013e22bee78STejun Heo { 201463d95a91STejun Heo struct worker_pool *pool = worker->pool; 2015e22bee78STejun Heo 2016692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 201729187a9eSTejun Heo return false; 2018692b4825STejun Heo 2019692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 20202607d7a6STejun Heo pool->manager = worker; 2021e22bee78STejun Heo 202229187a9eSTejun Heo maybe_create_worker(pool); 2023e22bee78STejun Heo 20242607d7a6STejun Heo pool->manager = NULL; 2025692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2026692b4825STejun Heo wake_up(&wq_manager_wait); 202729187a9eSTejun Heo return true; 2028e22bee78STejun Heo } 2029e22bee78STejun Heo 2030a62428c0STejun Heo /** 2031a62428c0STejun Heo * process_one_work - process single work 2032c34056a3STejun Heo * @worker: self 2033a62428c0STejun Heo * @work: work to process 2034a62428c0STejun Heo * 2035a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2036a62428c0STejun Heo * process a single work including synchronization against and 2037a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2038a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2039a62428c0STejun Heo * call this function to process a work. 2040a62428c0STejun Heo * 2041a62428c0STejun Heo * CONTEXT: 2042d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2043a62428c0STejun Heo */ 2044c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2045d565ed63STejun Heo __releases(&pool->lock) 2046d565ed63STejun Heo __acquires(&pool->lock) 20471da177e4SLinus Torvalds { 2048112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2049bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2050112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 205173f53c4aSTejun Heo int work_color; 20527e11629dSTejun Heo struct worker *collision; 20534e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20544e6045f1SJohannes Berg /* 2055a62428c0STejun Heo * It is permissible to free the struct work_struct from 2056a62428c0STejun Heo * inside the function that is called from it, this we need to 2057a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2058a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2059a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20604e6045f1SJohannes Berg */ 20614d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20624d82a1deSPeter Zijlstra 20634d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20644e6045f1SJohannes Berg #endif 2065807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 206685327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2067ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 206825511a47STejun Heo 20697e11629dSTejun Heo /* 20707e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20717e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20727e11629dSTejun Heo * already processing the work. If so, defer the work to the 20737e11629dSTejun Heo * currently executing one. 20747e11629dSTejun Heo */ 2075c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20767e11629dSTejun Heo if (unlikely(collision)) { 20777e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20787e11629dSTejun Heo return; 20797e11629dSTejun Heo } 20801da177e4SLinus Torvalds 20818930cabaSTejun Heo /* claim and dequeue */ 20821da177e4SLinus Torvalds debug_work_deactivate(work); 2083c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2084c34056a3STejun Heo worker->current_work = work; 2085a2c1c57bSTejun Heo worker->current_func = work->func; 2086112202d9STejun Heo worker->current_pwq = pwq; 208773f53c4aSTejun Heo work_color = get_work_color(work); 20887a22ad75STejun Heo 2089a62428c0STejun Heo list_del_init(&work->entry); 2090a62428c0STejun Heo 2091649027d7STejun Heo /* 2092228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2093228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2094228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2095228f1d00SLai Jiangshan * execution of the pending work items. 2096fb0e7bebSTejun Heo */ 2097fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2098228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2099fb0e7bebSTejun Heo 2100974271c4STejun Heo /* 2101a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2102a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2103a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2104a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2105228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2106974271c4STejun Heo */ 2107a489a03eSLai Jiangshan if (need_more_worker(pool)) 210863d95a91STejun Heo wake_up_worker(pool); 2109974271c4STejun Heo 21108930cabaSTejun Heo /* 21117c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2112d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 211323657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 211423657bb1STejun Heo * disabled. 21158930cabaSTejun Heo */ 21167c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21171da177e4SLinus Torvalds 2118d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2119365970a1SDavid Howells 2120a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 21213295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2122e6f3faa7SPeter Zijlstra /* 2123f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2124f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2125e6f3faa7SPeter Zijlstra * 2126e6f3faa7SPeter Zijlstra * However, that would result in: 2127e6f3faa7SPeter Zijlstra * 2128e6f3faa7SPeter Zijlstra * A(W1) 2129e6f3faa7SPeter Zijlstra * WFC(C) 2130e6f3faa7SPeter Zijlstra * A(W1) 2131e6f3faa7SPeter Zijlstra * C(C) 2132e6f3faa7SPeter Zijlstra * 2133e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2134e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2135e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2136f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2137e6f3faa7SPeter Zijlstra * these locks. 2138e6f3faa7SPeter Zijlstra * 2139e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2140e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2141e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2142e6f3faa7SPeter Zijlstra */ 2143f52be570SPeter Zijlstra lockdep_invariant_state(true); 2144e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2145a2c1c57bSTejun Heo worker->current_func(work); 2146e36c886aSArjan van de Ven /* 2147e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2148e36c886aSArjan van de Ven * point will only record its address. 2149e36c886aSArjan van de Ven */ 2150e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21513295f0efSIngo Molnar lock_map_release(&lockdep_map); 2152112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21531da177e4SLinus Torvalds 2154d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2155044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2156044c782cSValentin Ilie " last function: %pf\n", 2157a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2158a2c1c57bSTejun Heo worker->current_func); 2159d5abe669SPeter Zijlstra debug_show_held_locks(current); 2160d5abe669SPeter Zijlstra dump_stack(); 2161d5abe669SPeter Zijlstra } 2162d5abe669SPeter Zijlstra 2163b22ce278STejun Heo /* 2164b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2165b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2166b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2167b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2168789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2169789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2170b22ce278STejun Heo */ 2171a7e6425eSPaul E. McKenney cond_resched(); 2172b22ce278STejun Heo 2173d565ed63STejun Heo spin_lock_irq(&pool->lock); 2174a62428c0STejun Heo 2175fb0e7bebSTejun Heo /* clear cpu intensive status */ 2176fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2177fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2178fb0e7bebSTejun Heo 2179a62428c0STejun Heo /* we're done with it, release */ 218042f8570fSSasha Levin hash_del(&worker->hentry); 2181c34056a3STejun Heo worker->current_work = NULL; 2182a2c1c57bSTejun Heo worker->current_func = NULL; 2183112202d9STejun Heo worker->current_pwq = NULL; 21843d1cb205STejun Heo worker->desc_valid = false; 2185112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21861da177e4SLinus Torvalds } 21871da177e4SLinus Torvalds 2188affee4b2STejun Heo /** 2189affee4b2STejun Heo * process_scheduled_works - process scheduled works 2190affee4b2STejun Heo * @worker: self 2191affee4b2STejun Heo * 2192affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2193affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2194affee4b2STejun Heo * fetches a work from the top and executes it. 2195affee4b2STejun Heo * 2196affee4b2STejun Heo * CONTEXT: 2197d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2198affee4b2STejun Heo * multiple times. 2199affee4b2STejun Heo */ 2200affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22011da177e4SLinus Torvalds { 2202affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2203affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2204a62428c0STejun Heo struct work_struct, entry); 2205c34056a3STejun Heo process_one_work(worker, work); 2206a62428c0STejun Heo } 22071da177e4SLinus Torvalds } 22081da177e4SLinus Torvalds 22094690c4abSTejun Heo /** 22104690c4abSTejun Heo * worker_thread - the worker thread function 2211c34056a3STejun Heo * @__worker: self 22124690c4abSTejun Heo * 2213c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2214c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2215c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2216c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2217c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2218d185af30SYacine Belkadi * 2219d185af30SYacine Belkadi * Return: 0 22204690c4abSTejun Heo */ 2221c34056a3STejun Heo static int worker_thread(void *__worker) 22221da177e4SLinus Torvalds { 2223c34056a3STejun Heo struct worker *worker = __worker; 2224bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22251da177e4SLinus Torvalds 2226e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2227e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2228c8e55f36STejun Heo woke_up: 2229d565ed63STejun Heo spin_lock_irq(&pool->lock); 2230affee4b2STejun Heo 2231a9ab775bSTejun Heo /* am I supposed to die? */ 2232a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2233d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2234a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2235e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 223660f5a4bcSLai Jiangshan 223760f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 22387cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 223960f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 224060f5a4bcSLai Jiangshan kfree(worker); 2241c8e55f36STejun Heo return 0; 2242c8e55f36STejun Heo } 2243c8e55f36STejun Heo 2244c8e55f36STejun Heo worker_leave_idle(worker); 2245db7bccf4STejun Heo recheck: 2246e22bee78STejun Heo /* no more worker necessary? */ 224763d95a91STejun Heo if (!need_more_worker(pool)) 2248e22bee78STejun Heo goto sleep; 2249e22bee78STejun Heo 2250e22bee78STejun Heo /* do we need to manage? */ 225163d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2252e22bee78STejun Heo goto recheck; 2253e22bee78STejun Heo 2254c8e55f36STejun Heo /* 2255c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2256c8e55f36STejun Heo * preparing to process a work or actually processing it. 2257c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2258c8e55f36STejun Heo */ 22596183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2260c8e55f36STejun Heo 2261e22bee78STejun Heo /* 2262a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2263a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2264a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2265a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2266a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2267e22bee78STejun Heo */ 2268a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2269e22bee78STejun Heo 2270e22bee78STejun Heo do { 2271affee4b2STejun Heo struct work_struct *work = 2272bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2273affee4b2STejun Heo struct work_struct, entry); 2274affee4b2STejun Heo 227582607adcSTejun Heo pool->watchdog_ts = jiffies; 227682607adcSTejun Heo 2277c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2278affee4b2STejun Heo /* optimization path, not strictly necessary */ 2279affee4b2STejun Heo process_one_work(worker, work); 2280affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2281affee4b2STejun Heo process_scheduled_works(worker); 2282affee4b2STejun Heo } else { 2283c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2284affee4b2STejun Heo process_scheduled_works(worker); 2285affee4b2STejun Heo } 228663d95a91STejun Heo } while (keep_working(pool)); 2287affee4b2STejun Heo 2288228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2289d313dd85STejun Heo sleep: 2290c8e55f36STejun Heo /* 2291d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2292d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2293d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2294d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2295d565ed63STejun Heo * event. 2296c8e55f36STejun Heo */ 2297c8e55f36STejun Heo worker_enter_idle(worker); 2298c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2299d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23001da177e4SLinus Torvalds schedule(); 2301c8e55f36STejun Heo goto woke_up; 23021da177e4SLinus Torvalds } 23031da177e4SLinus Torvalds 2304e22bee78STejun Heo /** 2305e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2306111c225aSTejun Heo * @__rescuer: self 2307e22bee78STejun Heo * 2308e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2309493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2310e22bee78STejun Heo * 2311706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2312e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2313e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2314e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2315e22bee78STejun Heo * the problem rescuer solves. 2316e22bee78STejun Heo * 2317706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2318706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2319e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2320e22bee78STejun Heo * 2321e22bee78STejun Heo * This should happen rarely. 2322d185af30SYacine Belkadi * 2323d185af30SYacine Belkadi * Return: 0 2324e22bee78STejun Heo */ 2325111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2326e22bee78STejun Heo { 2327111c225aSTejun Heo struct worker *rescuer = __rescuer; 2328111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2329e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 23304d595b86SLai Jiangshan bool should_stop; 2331e22bee78STejun Heo 2332e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2333111c225aSTejun Heo 2334111c225aSTejun Heo /* 2335111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2336111c225aSTejun Heo * doesn't participate in concurrency management. 2337111c225aSTejun Heo */ 2338111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2339e22bee78STejun Heo repeat: 2340c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 23411da177e4SLinus Torvalds 23424d595b86SLai Jiangshan /* 23434d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 23444d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 23454d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 23464d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 23474d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 23484d595b86SLai Jiangshan * list is always empty on exit. 23494d595b86SLai Jiangshan */ 23504d595b86SLai Jiangshan should_stop = kthread_should_stop(); 23511da177e4SLinus Torvalds 2352493a1724STejun Heo /* see whether any pwq is asking for help */ 23532e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2354493a1724STejun Heo 2355493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2356493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2357493a1724STejun Heo struct pool_workqueue, mayday_node); 2358112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2359e22bee78STejun Heo struct work_struct *work, *n; 236082607adcSTejun Heo bool first = true; 2361e22bee78STejun Heo 2362e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2363493a1724STejun Heo list_del_init(&pwq->mayday_node); 2364493a1724STejun Heo 23652e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2366e22bee78STejun Heo 236751697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 236851697d39SLai Jiangshan 236951697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2370b3104104SLai Jiangshan rescuer->pool = pool; 2371e22bee78STejun Heo 2372e22bee78STejun Heo /* 2373e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2374e22bee78STejun Heo * process'em. 2375e22bee78STejun Heo */ 23760479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 237782607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 237882607adcSTejun Heo if (get_work_pwq(work) == pwq) { 237982607adcSTejun Heo if (first) 238082607adcSTejun Heo pool->watchdog_ts = jiffies; 2381e22bee78STejun Heo move_linked_works(work, scheduled, &n); 238282607adcSTejun Heo } 238382607adcSTejun Heo first = false; 238482607adcSTejun Heo } 2385e22bee78STejun Heo 2386008847f6SNeilBrown if (!list_empty(scheduled)) { 2387e22bee78STejun Heo process_scheduled_works(rescuer); 23887576958aSTejun Heo 23897576958aSTejun Heo /* 2390008847f6SNeilBrown * The above execution of rescued work items could 2391008847f6SNeilBrown * have created more to rescue through 2392008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2393008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2394008847f6SNeilBrown * that such back-to-back work items, which may be 2395008847f6SNeilBrown * being used to relieve memory pressure, don't 2396008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2397008847f6SNeilBrown */ 2398008847f6SNeilBrown if (need_to_create_worker(pool)) { 2399008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2400008847f6SNeilBrown get_pwq(pwq); 2401008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2402008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2403008847f6SNeilBrown } 2404008847f6SNeilBrown } 2405008847f6SNeilBrown 2406008847f6SNeilBrown /* 240777668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 240813b1d625SLai Jiangshan * go away while we're still attached to it. 240977668c8bSLai Jiangshan */ 241077668c8bSLai Jiangshan put_pwq(pwq); 241177668c8bSLai Jiangshan 241277668c8bSLai Jiangshan /* 2413d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 24147576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24157576958aSTejun Heo * and stalling the execution. 24167576958aSTejun Heo */ 2417d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 241863d95a91STejun Heo wake_up_worker(pool); 24197576958aSTejun Heo 2420b3104104SLai Jiangshan rescuer->pool = NULL; 242113b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 242213b1d625SLai Jiangshan 242313b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 242413b1d625SLai Jiangshan 242513b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 24261da177e4SLinus Torvalds } 24271da177e4SLinus Torvalds 24282e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2429493a1724STejun Heo 24304d595b86SLai Jiangshan if (should_stop) { 24314d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 24324d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 24334d595b86SLai Jiangshan return 0; 24344d595b86SLai Jiangshan } 24354d595b86SLai Jiangshan 2436111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2437111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2438e22bee78STejun Heo schedule(); 2439e22bee78STejun Heo goto repeat; 24401da177e4SLinus Torvalds } 24411da177e4SLinus Torvalds 2442fca839c0STejun Heo /** 2443fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2444fca839c0STejun Heo * @target_wq: workqueue being flushed 2445fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2446fca839c0STejun Heo * 2447fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2448fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2449fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2450fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2451fca839c0STejun Heo * a deadlock. 2452fca839c0STejun Heo */ 2453fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2454fca839c0STejun Heo struct work_struct *target_work) 2455fca839c0STejun Heo { 2456fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2457fca839c0STejun Heo struct worker *worker; 2458fca839c0STejun Heo 2459fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2460fca839c0STejun Heo return; 2461fca839c0STejun Heo 2462fca839c0STejun Heo worker = current_wq_worker(); 2463fca839c0STejun Heo 2464fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2465fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2466fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 246723d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 246823d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2469fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2470fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2471fca839c0STejun Heo target_wq->name, target_func); 2472fca839c0STejun Heo } 2473fca839c0STejun Heo 2474fc2e4d70SOleg Nesterov struct wq_barrier { 2475fc2e4d70SOleg Nesterov struct work_struct work; 2476fc2e4d70SOleg Nesterov struct completion done; 24772607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2478fc2e4d70SOleg Nesterov }; 2479fc2e4d70SOleg Nesterov 2480fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2481fc2e4d70SOleg Nesterov { 2482fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2483fc2e4d70SOleg Nesterov complete(&barr->done); 2484fc2e4d70SOleg Nesterov } 2485fc2e4d70SOleg Nesterov 24864690c4abSTejun Heo /** 24874690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2488112202d9STejun Heo * @pwq: pwq to insert barrier into 24894690c4abSTejun Heo * @barr: wq_barrier to insert 2490affee4b2STejun Heo * @target: target work to attach @barr to 2491affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24924690c4abSTejun Heo * 2493affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2494affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2495affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2496affee4b2STejun Heo * cpu. 2497affee4b2STejun Heo * 2498affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2499affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2500affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2501affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2502affee4b2STejun Heo * after a work with LINKED flag set. 2503affee4b2STejun Heo * 2504affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2505112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 25064690c4abSTejun Heo * 25074690c4abSTejun Heo * CONTEXT: 2508d565ed63STejun Heo * spin_lock_irq(pool->lock). 25094690c4abSTejun Heo */ 2510112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2511affee4b2STejun Heo struct wq_barrier *barr, 2512affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2513fc2e4d70SOleg Nesterov { 2514affee4b2STejun Heo struct list_head *head; 2515affee4b2STejun Heo unsigned int linked = 0; 2516affee4b2STejun Heo 2517dc186ad7SThomas Gleixner /* 2518d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2519dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2520dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2521dc186ad7SThomas Gleixner * might deadlock. 2522dc186ad7SThomas Gleixner */ 2523ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 252422df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 252552fa5bc5SBoqun Feng 2526fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2527fd1a5b04SByungchul Park 25282607d7a6STejun Heo barr->task = current; 252983c22520SOleg Nesterov 2530affee4b2STejun Heo /* 2531affee4b2STejun Heo * If @target is currently being executed, schedule the 2532affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2533affee4b2STejun Heo */ 2534affee4b2STejun Heo if (worker) 2535affee4b2STejun Heo head = worker->scheduled.next; 2536affee4b2STejun Heo else { 2537affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2538affee4b2STejun Heo 2539affee4b2STejun Heo head = target->entry.next; 2540affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2541affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2542affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2543affee4b2STejun Heo } 2544affee4b2STejun Heo 2545dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2546112202d9STejun Heo insert_work(pwq, &barr->work, head, 2547affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2548fc2e4d70SOleg Nesterov } 2549fc2e4d70SOleg Nesterov 255073f53c4aSTejun Heo /** 2551112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 255273f53c4aSTejun Heo * @wq: workqueue being flushed 255373f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 255473f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 255573f53c4aSTejun Heo * 2556112202d9STejun Heo * Prepare pwqs for workqueue flushing. 255773f53c4aSTejun Heo * 2558112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2559112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2560112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2561112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2562112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 256373f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 256473f53c4aSTejun Heo * 256573f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 256673f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 256773f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 256873f53c4aSTejun Heo * is returned. 256973f53c4aSTejun Heo * 2570112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 257173f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 257273f53c4aSTejun Heo * advanced to @work_color. 257373f53c4aSTejun Heo * 257473f53c4aSTejun Heo * CONTEXT: 25753c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 257673f53c4aSTejun Heo * 2577d185af30SYacine Belkadi * Return: 257873f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 257973f53c4aSTejun Heo * otherwise. 258073f53c4aSTejun Heo */ 2581112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 258273f53c4aSTejun Heo int flush_color, int work_color) 25831da177e4SLinus Torvalds { 258473f53c4aSTejun Heo bool wait = false; 258549e3cf44STejun Heo struct pool_workqueue *pwq; 25861da177e4SLinus Torvalds 258773f53c4aSTejun Heo if (flush_color >= 0) { 25886183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2589112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2590dc186ad7SThomas Gleixner } 259114441960SOleg Nesterov 259249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2593112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25941da177e4SLinus Torvalds 2595b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 259673f53c4aSTejun Heo 259773f53c4aSTejun Heo if (flush_color >= 0) { 25986183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 259973f53c4aSTejun Heo 2600112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2601112202d9STejun Heo pwq->flush_color = flush_color; 2602112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 260373f53c4aSTejun Heo wait = true; 26041da177e4SLinus Torvalds } 260573f53c4aSTejun Heo } 260673f53c4aSTejun Heo 260773f53c4aSTejun Heo if (work_color >= 0) { 26086183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2609112202d9STejun Heo pwq->work_color = work_color; 261073f53c4aSTejun Heo } 261173f53c4aSTejun Heo 2612b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 26131da177e4SLinus Torvalds } 26141da177e4SLinus Torvalds 2615112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 261673f53c4aSTejun Heo complete(&wq->first_flusher->done); 261773f53c4aSTejun Heo 261873f53c4aSTejun Heo return wait; 261983c22520SOleg Nesterov } 26201da177e4SLinus Torvalds 26210fcb78c2SRolf Eike Beer /** 26221da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26230fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26241da177e4SLinus Torvalds * 2625c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2626c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 26271da177e4SLinus Torvalds */ 26287ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26291da177e4SLinus Torvalds { 263073f53c4aSTejun Heo struct wq_flusher this_flusher = { 263173f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 263273f53c4aSTejun Heo .flush_color = -1, 2633fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 263473f53c4aSTejun Heo }; 263573f53c4aSTejun Heo int next_color; 2636b1f4ec17SOleg Nesterov 26373347fa09STejun Heo if (WARN_ON(!wq_online)) 26383347fa09STejun Heo return; 26393347fa09STejun Heo 26403c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 264173f53c4aSTejun Heo 264273f53c4aSTejun Heo /* 264373f53c4aSTejun Heo * Start-to-wait phase 264473f53c4aSTejun Heo */ 264573f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 264673f53c4aSTejun Heo 264773f53c4aSTejun Heo if (next_color != wq->flush_color) { 264873f53c4aSTejun Heo /* 264973f53c4aSTejun Heo * Color space is not full. The current work_color 265073f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 265173f53c4aSTejun Heo * by one. 265273f53c4aSTejun Heo */ 26536183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 265473f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 265573f53c4aSTejun Heo wq->work_color = next_color; 265673f53c4aSTejun Heo 265773f53c4aSTejun Heo if (!wq->first_flusher) { 265873f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26596183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 266073f53c4aSTejun Heo 266173f53c4aSTejun Heo wq->first_flusher = &this_flusher; 266273f53c4aSTejun Heo 2663112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 266473f53c4aSTejun Heo wq->work_color)) { 266573f53c4aSTejun Heo /* nothing to flush, done */ 266673f53c4aSTejun Heo wq->flush_color = next_color; 266773f53c4aSTejun Heo wq->first_flusher = NULL; 266873f53c4aSTejun Heo goto out_unlock; 266973f53c4aSTejun Heo } 267073f53c4aSTejun Heo } else { 267173f53c4aSTejun Heo /* wait in queue */ 26726183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 267373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2674112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 267573f53c4aSTejun Heo } 267673f53c4aSTejun Heo } else { 267773f53c4aSTejun Heo /* 267873f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 267973f53c4aSTejun Heo * The next flush completion will assign us 268073f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 268173f53c4aSTejun Heo */ 268273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 268373f53c4aSTejun Heo } 268473f53c4aSTejun Heo 2685fca839c0STejun Heo check_flush_dependency(wq, NULL); 2686fca839c0STejun Heo 26873c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 268873f53c4aSTejun Heo 268973f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 269073f53c4aSTejun Heo 269173f53c4aSTejun Heo /* 269273f53c4aSTejun Heo * Wake-up-and-cascade phase 269373f53c4aSTejun Heo * 269473f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 269573f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 269673f53c4aSTejun Heo */ 269773f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 269873f53c4aSTejun Heo return; 269973f53c4aSTejun Heo 27003c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 270173f53c4aSTejun Heo 27024ce48b37STejun Heo /* we might have raced, check again with mutex held */ 27034ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 27044ce48b37STejun Heo goto out_unlock; 27054ce48b37STejun Heo 270673f53c4aSTejun Heo wq->first_flusher = NULL; 270773f53c4aSTejun Heo 27086183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 27096183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 271073f53c4aSTejun Heo 271173f53c4aSTejun Heo while (true) { 271273f53c4aSTejun Heo struct wq_flusher *next, *tmp; 271373f53c4aSTejun Heo 271473f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 271573f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 271673f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 271773f53c4aSTejun Heo break; 271873f53c4aSTejun Heo list_del_init(&next->list); 271973f53c4aSTejun Heo complete(&next->done); 272073f53c4aSTejun Heo } 272173f53c4aSTejun Heo 27226183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 272373f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 272473f53c4aSTejun Heo 272573f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 272673f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 272773f53c4aSTejun Heo 272873f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 272973f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 273073f53c4aSTejun Heo /* 273173f53c4aSTejun Heo * Assign the same color to all overflowed 273273f53c4aSTejun Heo * flushers, advance work_color and append to 273373f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 273473f53c4aSTejun Heo * phase for these overflowed flushers. 273573f53c4aSTejun Heo */ 273673f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 273773f53c4aSTejun Heo tmp->flush_color = wq->work_color; 273873f53c4aSTejun Heo 273973f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 274073f53c4aSTejun Heo 274173f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 274273f53c4aSTejun Heo &wq->flusher_queue); 2743112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 274473f53c4aSTejun Heo } 274573f53c4aSTejun Heo 274673f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27476183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 274873f53c4aSTejun Heo break; 274973f53c4aSTejun Heo } 275073f53c4aSTejun Heo 275173f53c4aSTejun Heo /* 275273f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2753112202d9STejun Heo * the new first flusher and arm pwqs. 275473f53c4aSTejun Heo */ 27556183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27566183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 275773f53c4aSTejun Heo 275873f53c4aSTejun Heo list_del_init(&next->list); 275973f53c4aSTejun Heo wq->first_flusher = next; 276073f53c4aSTejun Heo 2761112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 276273f53c4aSTejun Heo break; 276373f53c4aSTejun Heo 276473f53c4aSTejun Heo /* 276573f53c4aSTejun Heo * Meh... this color is already done, clear first 276673f53c4aSTejun Heo * flusher and repeat cascading. 276773f53c4aSTejun Heo */ 276873f53c4aSTejun Heo wq->first_flusher = NULL; 276973f53c4aSTejun Heo } 277073f53c4aSTejun Heo 277173f53c4aSTejun Heo out_unlock: 27723c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27731da177e4SLinus Torvalds } 27741dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 27751da177e4SLinus Torvalds 27769c5a2ba7STejun Heo /** 27779c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27789c5a2ba7STejun Heo * @wq: workqueue to drain 27799c5a2ba7STejun Heo * 27809c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27819c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27829c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2783b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 27849c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27859c5a2ba7STejun Heo * takes too long. 27869c5a2ba7STejun Heo */ 27879c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27889c5a2ba7STejun Heo { 27899c5a2ba7STejun Heo unsigned int flush_cnt = 0; 279049e3cf44STejun Heo struct pool_workqueue *pwq; 27919c5a2ba7STejun Heo 27929c5a2ba7STejun Heo /* 27939c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27949c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2795618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27969c5a2ba7STejun Heo */ 279787fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27989c5a2ba7STejun Heo if (!wq->nr_drainers++) 2799618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 280087fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28019c5a2ba7STejun Heo reflush: 28029c5a2ba7STejun Heo flush_workqueue(wq); 28039c5a2ba7STejun Heo 2804b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 280576af4d93STejun Heo 280649e3cf44STejun Heo for_each_pwq(pwq, wq) { 2807fa2563e4SThomas Tuttle bool drained; 28089c5a2ba7STejun Heo 2809b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2810112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2811b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2812fa2563e4SThomas Tuttle 2813fa2563e4SThomas Tuttle if (drained) 28149c5a2ba7STejun Heo continue; 28159c5a2ba7STejun Heo 28169c5a2ba7STejun Heo if (++flush_cnt == 10 || 28179c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2818c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 28199c5a2ba7STejun Heo wq->name, flush_cnt); 282076af4d93STejun Heo 2821b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 28229c5a2ba7STejun Heo goto reflush; 28239c5a2ba7STejun Heo } 28249c5a2ba7STejun Heo 28259c5a2ba7STejun Heo if (!--wq->nr_drainers) 2826618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 282787fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28289c5a2ba7STejun Heo } 28299c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28309c5a2ba7STejun Heo 2831606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2832baf59022STejun Heo { 2833baf59022STejun Heo struct worker *worker = NULL; 2834c9e7cf27STejun Heo struct worker_pool *pool; 2835112202d9STejun Heo struct pool_workqueue *pwq; 2836baf59022STejun Heo 2837baf59022STejun Heo might_sleep(); 2838baf59022STejun Heo 2839fa1b54e6STejun Heo local_irq_disable(); 2840fa1b54e6STejun Heo pool = get_work_pool(work); 2841fa1b54e6STejun Heo if (!pool) { 2842fa1b54e6STejun Heo local_irq_enable(); 2843fa1b54e6STejun Heo return false; 2844fa1b54e6STejun Heo } 2845fa1b54e6STejun Heo 2846fa1b54e6STejun Heo spin_lock(&pool->lock); 28470b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2848112202d9STejun Heo pwq = get_work_pwq(work); 2849112202d9STejun Heo if (pwq) { 2850112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2851baf59022STejun Heo goto already_gone; 2852606a5020STejun Heo } else { 2853c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2854baf59022STejun Heo if (!worker) 2855baf59022STejun Heo goto already_gone; 2856112202d9STejun Heo pwq = worker->current_pwq; 2857606a5020STejun Heo } 2858baf59022STejun Heo 2859fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2860fca839c0STejun Heo 2861112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2862d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2863baf59022STejun Heo 2864e159489bSTejun Heo /* 2865a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 2866a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 2867a1d14934SPeter Zijlstra * 2868a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 2869a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 2870a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 2871a1d14934SPeter Zijlstra * forward progress. 2872e159489bSTejun Heo */ 2873a1d14934SPeter Zijlstra if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) { 2874112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2875112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2876a1d14934SPeter Zijlstra } 2877e159489bSTejun Heo 2878baf59022STejun Heo return true; 2879baf59022STejun Heo already_gone: 2880d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2881baf59022STejun Heo return false; 2882baf59022STejun Heo } 2883baf59022STejun Heo 2884db700897SOleg Nesterov /** 2885401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2886401a8d04STejun Heo * @work: the work to flush 2887db700897SOleg Nesterov * 2888606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2889606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2890401a8d04STejun Heo * 2891d185af30SYacine Belkadi * Return: 2892401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2893401a8d04STejun Heo * %false if it was already idle. 2894db700897SOleg Nesterov */ 2895401a8d04STejun Heo bool flush_work(struct work_struct *work) 2896db700897SOleg Nesterov { 289712997d1aSBjorn Helgaas struct wq_barrier barr; 289812997d1aSBjorn Helgaas 28993347fa09STejun Heo if (WARN_ON(!wq_online)) 29003347fa09STejun Heo return false; 29013347fa09STejun Heo 290212997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 290312997d1aSBjorn Helgaas wait_for_completion(&barr.done); 290412997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 290512997d1aSBjorn Helgaas return true; 290612997d1aSBjorn Helgaas } else { 290712997d1aSBjorn Helgaas return false; 290812997d1aSBjorn Helgaas } 2909606a5020STejun Heo } 2910db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2911db700897SOleg Nesterov 29128603e1b3STejun Heo struct cwt_wait { 2913ac6424b9SIngo Molnar wait_queue_entry_t wait; 29148603e1b3STejun Heo struct work_struct *work; 29158603e1b3STejun Heo }; 29168603e1b3STejun Heo 2917ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 29188603e1b3STejun Heo { 29198603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 29208603e1b3STejun Heo 29218603e1b3STejun Heo if (cwait->work != key) 29228603e1b3STejun Heo return 0; 29238603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 29248603e1b3STejun Heo } 29258603e1b3STejun Heo 292636e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2927401a8d04STejun Heo { 29288603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2929bbb68dfaSTejun Heo unsigned long flags; 29301f1f642eSOleg Nesterov int ret; 29311f1f642eSOleg Nesterov 29321f1f642eSOleg Nesterov do { 2933bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2934bbb68dfaSTejun Heo /* 29358603e1b3STejun Heo * If someone else is already canceling, wait for it to 29368603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 29378603e1b3STejun Heo * because we may get scheduled between @work's completion 29388603e1b3STejun Heo * and the other canceling task resuming and clearing 29398603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 29408603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 29418603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 29428603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 29438603e1b3STejun Heo * we're hogging the CPU. 29448603e1b3STejun Heo * 29458603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 29468603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 29478603e1b3STejun Heo * wake function which matches @work along with exclusive 29488603e1b3STejun Heo * wait and wakeup. 2949bbb68dfaSTejun Heo */ 29508603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 29518603e1b3STejun Heo struct cwt_wait cwait; 29528603e1b3STejun Heo 29538603e1b3STejun Heo init_wait(&cwait.wait); 29548603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 29558603e1b3STejun Heo cwait.work = work; 29568603e1b3STejun Heo 29578603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 29588603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 29598603e1b3STejun Heo if (work_is_canceling(work)) 29608603e1b3STejun Heo schedule(); 29618603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 29628603e1b3STejun Heo } 29631f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29641f1f642eSOleg Nesterov 2965bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2966bbb68dfaSTejun Heo mark_work_canceling(work); 2967bbb68dfaSTejun Heo local_irq_restore(flags); 2968bbb68dfaSTejun Heo 29693347fa09STejun Heo /* 29703347fa09STejun Heo * This allows canceling during early boot. We know that @work 29713347fa09STejun Heo * isn't executing. 29723347fa09STejun Heo */ 29733347fa09STejun Heo if (wq_online) 2974606a5020STejun Heo flush_work(work); 29753347fa09STejun Heo 29767a22ad75STejun Heo clear_work_data(work); 29778603e1b3STejun Heo 29788603e1b3STejun Heo /* 29798603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 29808603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 29818603e1b3STejun Heo * visible there. 29828603e1b3STejun Heo */ 29838603e1b3STejun Heo smp_mb(); 29848603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 29858603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 29868603e1b3STejun Heo 29871f1f642eSOleg Nesterov return ret; 29881f1f642eSOleg Nesterov } 29891f1f642eSOleg Nesterov 29906e84d644SOleg Nesterov /** 2991401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2992401a8d04STejun Heo * @work: the work to cancel 29936e84d644SOleg Nesterov * 2994401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2995401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2996401a8d04STejun Heo * another workqueue. On return from this function, @work is 2997401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29981f1f642eSOleg Nesterov * 2999401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3000401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 30016e84d644SOleg Nesterov * 3002401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 30036e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3004401a8d04STejun Heo * 3005d185af30SYacine Belkadi * Return: 3006401a8d04STejun Heo * %true if @work was pending, %false otherwise. 30076e84d644SOleg Nesterov */ 3008401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 30096e84d644SOleg Nesterov { 301036e227d2STejun Heo return __cancel_work_timer(work, false); 3011b89deed3SOleg Nesterov } 301228e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3013b89deed3SOleg Nesterov 30146e84d644SOleg Nesterov /** 3015401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3016401a8d04STejun Heo * @dwork: the delayed work to flush 30176e84d644SOleg Nesterov * 3018401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3019401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3020401a8d04STejun Heo * considers the last queueing instance of @dwork. 30211f1f642eSOleg Nesterov * 3022d185af30SYacine Belkadi * Return: 3023401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3024401a8d04STejun Heo * %false if it was already idle. 30256e84d644SOleg Nesterov */ 3026401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3027401a8d04STejun Heo { 30288930cabaSTejun Heo local_irq_disable(); 3029401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 303060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 30318930cabaSTejun Heo local_irq_enable(); 3032401a8d04STejun Heo return flush_work(&dwork->work); 3033401a8d04STejun Heo } 3034401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3035401a8d04STejun Heo 303605f0fe6bSTejun Heo /** 303705f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 303805f0fe6bSTejun Heo * @rwork: the rcu work to flush 303905f0fe6bSTejun Heo * 304005f0fe6bSTejun Heo * Return: 304105f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 304205f0fe6bSTejun Heo * %false if it was already idle. 304305f0fe6bSTejun Heo */ 304405f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 304505f0fe6bSTejun Heo { 304605f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 304705f0fe6bSTejun Heo rcu_barrier(); 304805f0fe6bSTejun Heo flush_work(&rwork->work); 304905f0fe6bSTejun Heo return true; 305005f0fe6bSTejun Heo } else { 305105f0fe6bSTejun Heo return flush_work(&rwork->work); 305205f0fe6bSTejun Heo } 305305f0fe6bSTejun Heo } 305405f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 305505f0fe6bSTejun Heo 3056f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3057f72b8792SJens Axboe { 3058f72b8792SJens Axboe unsigned long flags; 3059f72b8792SJens Axboe int ret; 3060f72b8792SJens Axboe 3061f72b8792SJens Axboe do { 3062f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3063f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3064f72b8792SJens Axboe 3065f72b8792SJens Axboe if (unlikely(ret < 0)) 3066f72b8792SJens Axboe return false; 3067f72b8792SJens Axboe 3068f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3069f72b8792SJens Axboe local_irq_restore(flags); 3070f72b8792SJens Axboe return ret; 3071f72b8792SJens Axboe } 3072f72b8792SJens Axboe 3073401a8d04STejun Heo /** 307457b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 307557b30ae7STejun Heo * @dwork: delayed_work to cancel 307609383498STejun Heo * 3077d185af30SYacine Belkadi * Kill off a pending delayed_work. 3078d185af30SYacine Belkadi * 3079d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3080d185af30SYacine Belkadi * pending. 3081d185af30SYacine Belkadi * 3082d185af30SYacine Belkadi * Note: 3083d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3084d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3085d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 308609383498STejun Heo * 308757b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 308809383498STejun Heo */ 308957b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 309009383498STejun Heo { 3091f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 309209383498STejun Heo } 309357b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 309409383498STejun Heo 309509383498STejun Heo /** 3096401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3097401a8d04STejun Heo * @dwork: the delayed work cancel 3098401a8d04STejun Heo * 3099401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3100401a8d04STejun Heo * 3101d185af30SYacine Belkadi * Return: 3102401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3103401a8d04STejun Heo */ 3104401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 31056e84d644SOleg Nesterov { 310636e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 31076e84d644SOleg Nesterov } 3108f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 31091da177e4SLinus Torvalds 31100fcb78c2SRolf Eike Beer /** 311131ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3112b6136773SAndrew Morton * @func: the function to call 3113b6136773SAndrew Morton * 311431ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 311531ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3116b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 311731ddd871STejun Heo * 3118d185af30SYacine Belkadi * Return: 311931ddd871STejun Heo * 0 on success, -errno on failure. 3120b6136773SAndrew Morton */ 312165f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 312215316ba8SChristoph Lameter { 312315316ba8SChristoph Lameter int cpu; 312438f51568SNamhyung Kim struct work_struct __percpu *works; 312515316ba8SChristoph Lameter 3126b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3127b6136773SAndrew Morton if (!works) 312815316ba8SChristoph Lameter return -ENOMEM; 3129b6136773SAndrew Morton 313095402b38SGautham R Shenoy get_online_cpus(); 313193981800STejun Heo 313215316ba8SChristoph Lameter for_each_online_cpu(cpu) { 31339bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 31349bfb1839SIngo Molnar 31359bfb1839SIngo Molnar INIT_WORK(work, func); 31368de6d308SOleg Nesterov schedule_work_on(cpu, work); 313715316ba8SChristoph Lameter } 313893981800STejun Heo 313993981800STejun Heo for_each_online_cpu(cpu) 31408616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 314193981800STejun Heo 314295402b38SGautham R Shenoy put_online_cpus(); 3143b6136773SAndrew Morton free_percpu(works); 314415316ba8SChristoph Lameter return 0; 314515316ba8SChristoph Lameter } 314615316ba8SChristoph Lameter 3147eef6a7d5SAlan Stern /** 31481fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31491fa44ecaSJames Bottomley * @fn: the function to execute 31501fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31511fa44ecaSJames Bottomley * be available when the work executes) 31521fa44ecaSJames Bottomley * 31531fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31541fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31551fa44ecaSJames Bottomley * 3156d185af30SYacine Belkadi * Return: 0 - function was executed 31571fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31581fa44ecaSJames Bottomley */ 315965f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31601fa44ecaSJames Bottomley { 31611fa44ecaSJames Bottomley if (!in_interrupt()) { 316265f27f38SDavid Howells fn(&ew->work); 31631fa44ecaSJames Bottomley return 0; 31641fa44ecaSJames Bottomley } 31651fa44ecaSJames Bottomley 316665f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31671fa44ecaSJames Bottomley schedule_work(&ew->work); 31681fa44ecaSJames Bottomley 31691fa44ecaSJames Bottomley return 1; 31701fa44ecaSJames Bottomley } 31711fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31721fa44ecaSJames Bottomley 31737a4e344cSTejun Heo /** 31747a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 31757a4e344cSTejun Heo * @attrs: workqueue_attrs to free 31767a4e344cSTejun Heo * 31777a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 31787a4e344cSTejun Heo */ 31797a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 31807a4e344cSTejun Heo { 31817a4e344cSTejun Heo if (attrs) { 31827a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 31837a4e344cSTejun Heo kfree(attrs); 31847a4e344cSTejun Heo } 31857a4e344cSTejun Heo } 31867a4e344cSTejun Heo 31877a4e344cSTejun Heo /** 31887a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 31897a4e344cSTejun Heo * @gfp_mask: allocation mask to use 31907a4e344cSTejun Heo * 31917a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3192d185af30SYacine Belkadi * return it. 3193d185af30SYacine Belkadi * 3194d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 31957a4e344cSTejun Heo */ 31967a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 31977a4e344cSTejun Heo { 31987a4e344cSTejun Heo struct workqueue_attrs *attrs; 31997a4e344cSTejun Heo 32007a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 32017a4e344cSTejun Heo if (!attrs) 32027a4e344cSTejun Heo goto fail; 32037a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 32047a4e344cSTejun Heo goto fail; 32057a4e344cSTejun Heo 320613e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 32077a4e344cSTejun Heo return attrs; 32087a4e344cSTejun Heo fail: 32097a4e344cSTejun Heo free_workqueue_attrs(attrs); 32107a4e344cSTejun Heo return NULL; 32117a4e344cSTejun Heo } 32127a4e344cSTejun Heo 321329c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 321429c91e99STejun Heo const struct workqueue_attrs *from) 321529c91e99STejun Heo { 321629c91e99STejun Heo to->nice = from->nice; 321729c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 32182865a8fbSShaohua Li /* 32192865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 32202865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 32212865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 32222865a8fbSShaohua Li */ 32232865a8fbSShaohua Li to->no_numa = from->no_numa; 322429c91e99STejun Heo } 322529c91e99STejun Heo 322629c91e99STejun Heo /* hash value of the content of @attr */ 322729c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 322829c91e99STejun Heo { 322929c91e99STejun Heo u32 hash = 0; 323029c91e99STejun Heo 323129c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 323213e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 323313e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 323429c91e99STejun Heo return hash; 323529c91e99STejun Heo } 323629c91e99STejun Heo 323729c91e99STejun Heo /* content equality test */ 323829c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 323929c91e99STejun Heo const struct workqueue_attrs *b) 324029c91e99STejun Heo { 324129c91e99STejun Heo if (a->nice != b->nice) 324229c91e99STejun Heo return false; 324329c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 324429c91e99STejun Heo return false; 324529c91e99STejun Heo return true; 324629c91e99STejun Heo } 324729c91e99STejun Heo 32487a4e344cSTejun Heo /** 32497a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 32507a4e344cSTejun Heo * @pool: worker_pool to initialize 32517a4e344cSTejun Heo * 3252402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3253d185af30SYacine Belkadi * 3254d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 325529c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 325629c91e99STejun Heo * on @pool safely to release it. 32577a4e344cSTejun Heo */ 32587a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 32594e1a1f9aSTejun Heo { 32604e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 326129c91e99STejun Heo pool->id = -1; 326229c91e99STejun Heo pool->cpu = -1; 3263f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 32644e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 326582607adcSTejun Heo pool->watchdog_ts = jiffies; 32664e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 32674e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 32684e1a1f9aSTejun Heo hash_init(pool->busy_hash); 32694e1a1f9aSTejun Heo 327032a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 32714e1a1f9aSTejun Heo 327232a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 32734e1a1f9aSTejun Heo 3274da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 32757a4e344cSTejun Heo 32767cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 327729c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 327829c91e99STejun Heo pool->refcnt = 1; 327929c91e99STejun Heo 328029c91e99STejun Heo /* shouldn't fail above this point */ 32817a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 32827a4e344cSTejun Heo if (!pool->attrs) 32837a4e344cSTejun Heo return -ENOMEM; 32847a4e344cSTejun Heo return 0; 32854e1a1f9aSTejun Heo } 32864e1a1f9aSTejun Heo 3287e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3288e2dca7adSTejun Heo { 3289e2dca7adSTejun Heo struct workqueue_struct *wq = 3290e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3291e2dca7adSTejun Heo 3292e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3293e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3294e2dca7adSTejun Heo else 3295e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3296e2dca7adSTejun Heo 3297e2dca7adSTejun Heo kfree(wq->rescuer); 3298e2dca7adSTejun Heo kfree(wq); 3299e2dca7adSTejun Heo } 3300e2dca7adSTejun Heo 330129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 330229c91e99STejun Heo { 330329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 330429c91e99STejun Heo 33057cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 330629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 330729c91e99STejun Heo kfree(pool); 330829c91e99STejun Heo } 330929c91e99STejun Heo 331029c91e99STejun Heo /** 331129c91e99STejun Heo * put_unbound_pool - put a worker_pool 331229c91e99STejun Heo * @pool: worker_pool to put 331329c91e99STejun Heo * 331429c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3315c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3316c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3317c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3318a892caccSTejun Heo * 3319a892caccSTejun Heo * Should be called with wq_pool_mutex held. 332029c91e99STejun Heo */ 332129c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 332229c91e99STejun Heo { 332360f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 332429c91e99STejun Heo struct worker *worker; 332529c91e99STejun Heo 3326a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3327a892caccSTejun Heo 3328a892caccSTejun Heo if (--pool->refcnt) 332929c91e99STejun Heo return; 333029c91e99STejun Heo 333129c91e99STejun Heo /* sanity checks */ 333261d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3333a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 333429c91e99STejun Heo return; 333529c91e99STejun Heo 333629c91e99STejun Heo /* release id and unhash */ 333729c91e99STejun Heo if (pool->id >= 0) 333829c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 333929c91e99STejun Heo hash_del(&pool->hash_node); 334029c91e99STejun Heo 3341c5aa87bbSTejun Heo /* 3342692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3343692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3344692b4825STejun Heo * manager and @pool gets freed with the flag set. 3345c5aa87bbSTejun Heo */ 334660f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 3347692b4825STejun Heo wait_event_lock_irq(wq_manager_wait, 3348692b4825STejun Heo !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); 3349692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3350692b4825STejun Heo 33511037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 335229c91e99STejun Heo destroy_worker(worker); 335329c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 335429c91e99STejun Heo spin_unlock_irq(&pool->lock); 335560f5a4bcSLai Jiangshan 33561258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3357da028469SLai Jiangshan if (!list_empty(&pool->workers)) 335860f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 33591258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 336060f5a4bcSLai Jiangshan 336160f5a4bcSLai Jiangshan if (pool->detach_completion) 336260f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 336360f5a4bcSLai Jiangshan 336429c91e99STejun Heo /* shut down the timers */ 336529c91e99STejun Heo del_timer_sync(&pool->idle_timer); 336629c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 336729c91e99STejun Heo 336829c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 336929c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 337029c91e99STejun Heo } 337129c91e99STejun Heo 337229c91e99STejun Heo /** 337329c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 337429c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 337529c91e99STejun Heo * 337629c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 337729c91e99STejun Heo * reference count and return it. If there already is a matching 337829c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3379d185af30SYacine Belkadi * create a new one. 3380a892caccSTejun Heo * 3381a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3382d185af30SYacine Belkadi * 3383d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3384d185af30SYacine Belkadi * On failure, %NULL. 338529c91e99STejun Heo */ 338629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 338729c91e99STejun Heo { 338829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 338929c91e99STejun Heo struct worker_pool *pool; 3390f3f90ad4STejun Heo int node; 3391e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 339229c91e99STejun Heo 3393a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 339429c91e99STejun Heo 339529c91e99STejun Heo /* do we already have a matching pool? */ 339629c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 339729c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 339829c91e99STejun Heo pool->refcnt++; 33993fb1823cSLai Jiangshan return pool; 340029c91e99STejun Heo } 340129c91e99STejun Heo } 340229c91e99STejun Heo 3403e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3404e2273584SXunlei Pang if (wq_numa_enabled) { 3405e2273584SXunlei Pang for_each_node(node) { 3406e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3407e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3408e2273584SXunlei Pang target_node = node; 3409e2273584SXunlei Pang break; 3410e2273584SXunlei Pang } 3411e2273584SXunlei Pang } 3412e2273584SXunlei Pang } 3413e2273584SXunlei Pang 341429c91e99STejun Heo /* nope, create a new one */ 3415e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 341629c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 341729c91e99STejun Heo goto fail; 341829c91e99STejun Heo 34198864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 342029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3421e2273584SXunlei Pang pool->node = target_node; 342229c91e99STejun Heo 34232865a8fbSShaohua Li /* 34242865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 34252865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 34262865a8fbSShaohua Li */ 34272865a8fbSShaohua Li pool->attrs->no_numa = false; 34282865a8fbSShaohua Li 342929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 343029c91e99STejun Heo goto fail; 343129c91e99STejun Heo 343229c91e99STejun Heo /* create and start the initial worker */ 34333347fa09STejun Heo if (wq_online && !create_worker(pool)) 343429c91e99STejun Heo goto fail; 343529c91e99STejun Heo 343629c91e99STejun Heo /* install */ 343729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 34383fb1823cSLai Jiangshan 343929c91e99STejun Heo return pool; 344029c91e99STejun Heo fail: 344129c91e99STejun Heo if (pool) 344229c91e99STejun Heo put_unbound_pool(pool); 344329c91e99STejun Heo return NULL; 344429c91e99STejun Heo } 344529c91e99STejun Heo 34468864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 34478864b4e5STejun Heo { 34488864b4e5STejun Heo kmem_cache_free(pwq_cache, 34498864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 34508864b4e5STejun Heo } 34518864b4e5STejun Heo 34528864b4e5STejun Heo /* 34538864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 34548864b4e5STejun Heo * and needs to be destroyed. 34558864b4e5STejun Heo */ 34568864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 34578864b4e5STejun Heo { 34588864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 34598864b4e5STejun Heo unbound_release_work); 34608864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 34618864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3462bc0caf09STejun Heo bool is_last; 34638864b4e5STejun Heo 34648864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 34658864b4e5STejun Heo return; 34668864b4e5STejun Heo 34673c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 34688864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3469bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 34703c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 34718864b4e5STejun Heo 3472a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 34738864b4e5STejun Heo put_unbound_pool(pool); 3474a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3475a892caccSTejun Heo 34768864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 34778864b4e5STejun Heo 34788864b4e5STejun Heo /* 34798864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3480e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 34818864b4e5STejun Heo */ 3482e2dca7adSTejun Heo if (is_last) 3483e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 34846029a918STejun Heo } 34858864b4e5STejun Heo 34860fbd95aaSTejun Heo /** 3487699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 34880fbd95aaSTejun Heo * @pwq: target pool_workqueue 34890fbd95aaSTejun Heo * 3490699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3491699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3492699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 34930fbd95aaSTejun Heo */ 3494699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 34950fbd95aaSTejun Heo { 3496699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3497699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 34983347fa09STejun Heo unsigned long flags; 3499699ce097STejun Heo 3500699ce097STejun Heo /* for @wq->saved_max_active */ 3501a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3502699ce097STejun Heo 3503699ce097STejun Heo /* fast exit for non-freezable wqs */ 3504699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3505699ce097STejun Heo return; 3506699ce097STejun Heo 35073347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 35083347fa09STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 3509699ce097STejun Heo 351074b414eaSLai Jiangshan /* 351174b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 351274b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 351374b414eaSLai Jiangshan * is updated and visible. 351474b414eaSLai Jiangshan */ 351574b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3516699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35170fbd95aaSTejun Heo 35180fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35190fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 35200fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3521951a078aSLai Jiangshan 3522951a078aSLai Jiangshan /* 3523951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3524951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3525951a078aSLai Jiangshan */ 3526951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3527699ce097STejun Heo } else { 3528699ce097STejun Heo pwq->max_active = 0; 3529699ce097STejun Heo } 3530699ce097STejun Heo 35313347fa09STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 35320fbd95aaSTejun Heo } 35330fbd95aaSTejun Heo 3534e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3535f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3536f147f29eSTejun Heo struct worker_pool *pool) 3537d2c1d404STejun Heo { 3538d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3539d2c1d404STejun Heo 3540e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3541e50aba9aSTejun Heo 3542d2c1d404STejun Heo pwq->pool = pool; 3543d2c1d404STejun Heo pwq->wq = wq; 3544d2c1d404STejun Heo pwq->flush_color = -1; 35458864b4e5STejun Heo pwq->refcnt = 1; 3546d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 35471befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3548d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 35498864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3550f147f29eSTejun Heo } 3551d2c1d404STejun Heo 3552f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 35531befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3554f147f29eSTejun Heo { 3555f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3556f147f29eSTejun Heo 3557f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 355875ccf595STejun Heo 35591befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 35601befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 35611befcf30STejun Heo return; 35621befcf30STejun Heo 356329b1cb41SLai Jiangshan /* set the matching work_color */ 356475ccf595STejun Heo pwq->work_color = wq->work_color; 3565983ca25eSTejun Heo 3566983ca25eSTejun Heo /* sync max_active to the current setting */ 3567983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3568983ca25eSTejun Heo 3569983ca25eSTejun Heo /* link in @pwq */ 35709e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3571df2d5ae4STejun Heo } 35726029a918STejun Heo 3573f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3574f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3575f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3576f147f29eSTejun Heo { 3577f147f29eSTejun Heo struct worker_pool *pool; 3578f147f29eSTejun Heo struct pool_workqueue *pwq; 3579f147f29eSTejun Heo 3580f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3581f147f29eSTejun Heo 3582f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3583f147f29eSTejun Heo if (!pool) 3584f147f29eSTejun Heo return NULL; 3585f147f29eSTejun Heo 3586e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3587f147f29eSTejun Heo if (!pwq) { 3588f147f29eSTejun Heo put_unbound_pool(pool); 3589f147f29eSTejun Heo return NULL; 3590f147f29eSTejun Heo } 3591f147f29eSTejun Heo 3592f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3593f147f29eSTejun Heo return pwq; 3594d2c1d404STejun Heo } 3595d2c1d404STejun Heo 35964c16bd32STejun Heo /** 359730186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3598042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 35994c16bd32STejun Heo * @node: the target NUMA node 36004c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 36014c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 36024c16bd32STejun Heo * 36034c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 36044c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3605d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 36064c16bd32STejun Heo * 36074c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 36084c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 36094c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 36104c16bd32STejun Heo * @attrs->cpumask. 36114c16bd32STejun Heo * 36124c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 36134c16bd32STejun Heo * stable. 3614d185af30SYacine Belkadi * 3615d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3616d185af30SYacine Belkadi * %false if equal. 36174c16bd32STejun Heo */ 36184c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 36194c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 36204c16bd32STejun Heo { 3621d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 36224c16bd32STejun Heo goto use_dfl; 36234c16bd32STejun Heo 36244c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 36254c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 36264c16bd32STejun Heo if (cpu_going_down >= 0) 36274c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 36284c16bd32STejun Heo 36294c16bd32STejun Heo if (cpumask_empty(cpumask)) 36304c16bd32STejun Heo goto use_dfl; 36314c16bd32STejun Heo 36324c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 36334c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 36341ad0f0a7SMichael Bringmann 36351ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 36361ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 36371ad0f0a7SMichael Bringmann "possible intersect\n"); 36381ad0f0a7SMichael Bringmann return false; 36391ad0f0a7SMichael Bringmann } 36401ad0f0a7SMichael Bringmann 36414c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 36424c16bd32STejun Heo 36434c16bd32STejun Heo use_dfl: 36444c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 36454c16bd32STejun Heo return false; 36464c16bd32STejun Heo } 36474c16bd32STejun Heo 36481befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 36491befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 36501befcf30STejun Heo int node, 36511befcf30STejun Heo struct pool_workqueue *pwq) 36521befcf30STejun Heo { 36531befcf30STejun Heo struct pool_workqueue *old_pwq; 36541befcf30STejun Heo 36555b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36561befcf30STejun Heo lockdep_assert_held(&wq->mutex); 36571befcf30STejun Heo 36581befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 36591befcf30STejun Heo link_pwq(pwq); 36601befcf30STejun Heo 36611befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 36621befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 36631befcf30STejun Heo return old_pwq; 36641befcf30STejun Heo } 36651befcf30STejun Heo 36662d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 36672d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 36682d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 36692d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3670042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 36712d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 36722d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 36732d5f0764SLai Jiangshan }; 36742d5f0764SLai Jiangshan 36752d5f0764SLai Jiangshan /* free the resources after success or abort */ 36762d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 36772d5f0764SLai Jiangshan { 36782d5f0764SLai Jiangshan if (ctx) { 36792d5f0764SLai Jiangshan int node; 36802d5f0764SLai Jiangshan 36812d5f0764SLai Jiangshan for_each_node(node) 36822d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 36832d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 36842d5f0764SLai Jiangshan 36852d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 36862d5f0764SLai Jiangshan 36872d5f0764SLai Jiangshan kfree(ctx); 36882d5f0764SLai Jiangshan } 36892d5f0764SLai Jiangshan } 36902d5f0764SLai Jiangshan 36912d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 36922d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 36932d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 36942d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 36952d5f0764SLai Jiangshan { 36962d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 36972d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 36982d5f0764SLai Jiangshan int node; 36992d5f0764SLai Jiangshan 37002d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 37012d5f0764SLai Jiangshan 37022d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 37032d5f0764SLai Jiangshan GFP_KERNEL); 37042d5f0764SLai Jiangshan 37052d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37062d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37072d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 37082d5f0764SLai Jiangshan goto out_free; 37092d5f0764SLai Jiangshan 3710042f7df1SLai Jiangshan /* 3711042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3712042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3713042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3714042f7df1SLai Jiangshan */ 37152d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3716b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3717042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3718042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 37192d5f0764SLai Jiangshan 37202d5f0764SLai Jiangshan /* 37212d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 37222d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 37232d5f0764SLai Jiangshan * pools. 37242d5f0764SLai Jiangshan */ 37252d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 37262d5f0764SLai Jiangshan 37272d5f0764SLai Jiangshan /* 37282d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 37292d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 37302d5f0764SLai Jiangshan * it even if we don't use it immediately. 37312d5f0764SLai Jiangshan */ 37322d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 37332d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 37342d5f0764SLai Jiangshan goto out_free; 37352d5f0764SLai Jiangshan 37362d5f0764SLai Jiangshan for_each_node(node) { 3737042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 37382d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 37392d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 37402d5f0764SLai Jiangshan goto out_free; 37412d5f0764SLai Jiangshan } else { 37422d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 37432d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 37442d5f0764SLai Jiangshan } 37452d5f0764SLai Jiangshan } 37462d5f0764SLai Jiangshan 3747042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3748042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3749042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 37502d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3751042f7df1SLai Jiangshan 37522d5f0764SLai Jiangshan ctx->wq = wq; 37532d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37542d5f0764SLai Jiangshan return ctx; 37552d5f0764SLai Jiangshan 37562d5f0764SLai Jiangshan out_free: 37572d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37582d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 37592d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 37602d5f0764SLai Jiangshan return NULL; 37612d5f0764SLai Jiangshan } 37622d5f0764SLai Jiangshan 37632d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 37642d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 37652d5f0764SLai Jiangshan { 37662d5f0764SLai Jiangshan int node; 37672d5f0764SLai Jiangshan 37682d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 37692d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 37702d5f0764SLai Jiangshan 37712d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 37722d5f0764SLai Jiangshan 37732d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 37742d5f0764SLai Jiangshan for_each_node(node) 37752d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 37762d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 37772d5f0764SLai Jiangshan 37782d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 37792d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 37802d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 37812d5f0764SLai Jiangshan 37822d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 37832d5f0764SLai Jiangshan } 37842d5f0764SLai Jiangshan 3785a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3786a0111cf6SLai Jiangshan { 3787a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3788a0111cf6SLai Jiangshan get_online_cpus(); 3789a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3790a0111cf6SLai Jiangshan } 3791a0111cf6SLai Jiangshan 3792a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3793a0111cf6SLai Jiangshan { 3794a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3795a0111cf6SLai Jiangshan put_online_cpus(); 3796a0111cf6SLai Jiangshan } 3797a0111cf6SLai Jiangshan 3798a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3799a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3800a0111cf6SLai Jiangshan { 3801a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3802a0111cf6SLai Jiangshan 3803a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3804a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3805a0111cf6SLai Jiangshan return -EINVAL; 3806a0111cf6SLai Jiangshan 3807a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 38080a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 38090a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 3810a0111cf6SLai Jiangshan return -EINVAL; 3811a0111cf6SLai Jiangshan 38120a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 38130a94efb5STejun Heo } 38140a94efb5STejun Heo 3815a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 38166201171eSwanghaibin if (!ctx) 38176201171eSwanghaibin return -ENOMEM; 3818a0111cf6SLai Jiangshan 3819a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3820a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3821a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3822a0111cf6SLai Jiangshan 38236201171eSwanghaibin return 0; 3824a0111cf6SLai Jiangshan } 3825a0111cf6SLai Jiangshan 38269e8cd2f5STejun Heo /** 38279e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38289e8cd2f5STejun Heo * @wq: the target workqueue 38299e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38309e8cd2f5STejun Heo * 38314c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 38324c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 38334c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38344c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38354c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38364c16bd32STejun Heo * back-to-back will stay on its current pwq. 38379e8cd2f5STejun Heo * 3838d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3839d185af30SYacine Belkadi * 3840d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 38419e8cd2f5STejun Heo */ 38429e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38439e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38449e8cd2f5STejun Heo { 3845a0111cf6SLai Jiangshan int ret; 38469e8cd2f5STejun Heo 3847a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3848a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3849a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 38502d5f0764SLai Jiangshan 38512d5f0764SLai Jiangshan return ret; 38529e8cd2f5STejun Heo } 38536106c0f8SNeilBrown EXPORT_SYMBOL_GPL(apply_workqueue_attrs); 38549e8cd2f5STejun Heo 38554c16bd32STejun Heo /** 38564c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 38574c16bd32STejun Heo * @wq: the target workqueue 38584c16bd32STejun Heo * @cpu: the CPU coming up or going down 38594c16bd32STejun Heo * @online: whether @cpu is coming up or going down 38604c16bd32STejun Heo * 38614c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 38624c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 38634c16bd32STejun Heo * @wq accordingly. 38644c16bd32STejun Heo * 38654c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 38664c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 38674c16bd32STejun Heo * correct. 38684c16bd32STejun Heo * 38694c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 38704c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 38714c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 38724c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 38734c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 38744c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 38754c16bd32STejun Heo * CPU_DOWN_PREPARE. 38764c16bd32STejun Heo */ 38774c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 38784c16bd32STejun Heo bool online) 38794c16bd32STejun Heo { 38804c16bd32STejun Heo int node = cpu_to_node(cpu); 38814c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 38824c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 38834c16bd32STejun Heo struct workqueue_attrs *target_attrs; 38844c16bd32STejun Heo cpumask_t *cpumask; 38854c16bd32STejun Heo 38864c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38874c16bd32STejun Heo 3888f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3889f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 38904c16bd32STejun Heo return; 38914c16bd32STejun Heo 38924c16bd32STejun Heo /* 38934c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 38944c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 38954c16bd32STejun Heo * CPU hotplug exclusion. 38964c16bd32STejun Heo */ 38974c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 38984c16bd32STejun Heo cpumask = target_attrs->cpumask; 38994c16bd32STejun Heo 39004c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39014c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39024c16bd32STejun Heo 39034c16bd32STejun Heo /* 39044c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3905042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3906042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3907042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 39084c16bd32STejun Heo */ 3909042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 39104c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3911f7142ed4SLai Jiangshan return; 39124c16bd32STejun Heo } else { 39134c16bd32STejun Heo goto use_dfl_pwq; 39144c16bd32STejun Heo } 39154c16bd32STejun Heo 39164c16bd32STejun Heo /* create a new pwq */ 39174c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 39184c16bd32STejun Heo if (!pwq) { 39192d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 39204c16bd32STejun Heo wq->name); 392177f300b1SDaeseok Youn goto use_dfl_pwq; 39224c16bd32STejun Heo } 39234c16bd32STejun Heo 3924f7142ed4SLai Jiangshan /* Install the new pwq. */ 39254c16bd32STejun Heo mutex_lock(&wq->mutex); 39264c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 39274c16bd32STejun Heo goto out_unlock; 39284c16bd32STejun Heo 39294c16bd32STejun Heo use_dfl_pwq: 3930f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 39314c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 39324c16bd32STejun Heo get_pwq(wq->dfl_pwq); 39334c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 39344c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 39354c16bd32STejun Heo out_unlock: 39364c16bd32STejun Heo mutex_unlock(&wq->mutex); 39374c16bd32STejun Heo put_pwq_unlocked(old_pwq); 39384c16bd32STejun Heo } 39394c16bd32STejun Heo 394030cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 39411da177e4SLinus Torvalds { 394249e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 39438a2b7538STejun Heo int cpu, ret; 3944e1d8aa9fSFrederic Weisbecker 394530cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3946420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3947420c0ddbSTejun Heo if (!wq->cpu_pwqs) 394830cdf249STejun Heo return -ENOMEM; 394930cdf249STejun Heo 395030cdf249STejun Heo for_each_possible_cpu(cpu) { 39517fb98ea7STejun Heo struct pool_workqueue *pwq = 39527fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 39537a62c2c8STejun Heo struct worker_pool *cpu_pools = 3954f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 395530cdf249STejun Heo 3956f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3957f147f29eSTejun Heo 3958f147f29eSTejun Heo mutex_lock(&wq->mutex); 39591befcf30STejun Heo link_pwq(pwq); 3960f147f29eSTejun Heo mutex_unlock(&wq->mutex); 396130cdf249STejun Heo } 396230cdf249STejun Heo return 0; 39638a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 39648a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 39658a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 39668a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 39678a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 39688a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 39698a2b7538STejun Heo return ret; 39709e8cd2f5STejun Heo } else { 39719e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 39729e8cd2f5STejun Heo } 39730f900049STejun Heo } 39740f900049STejun Heo 3975f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3976f3421797STejun Heo const char *name) 3977b71ab8c2STejun Heo { 3978f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3979f3421797STejun Heo 3980f3421797STejun Heo if (max_active < 1 || max_active > lim) 3981044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3982f3421797STejun Heo max_active, name, 1, lim); 3983b71ab8c2STejun Heo 3984f3421797STejun Heo return clamp_val(max_active, 1, lim); 3985b71ab8c2STejun Heo } 3986b71ab8c2STejun Heo 3987983c7515STejun Heo /* 3988983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 3989983c7515STejun Heo * to guarantee forward progress. 3990983c7515STejun Heo */ 3991983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 3992983c7515STejun Heo { 3993983c7515STejun Heo struct worker *rescuer; 3994983c7515STejun Heo int ret; 3995983c7515STejun Heo 3996983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 3997983c7515STejun Heo return 0; 3998983c7515STejun Heo 3999983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4000983c7515STejun Heo if (!rescuer) 4001983c7515STejun Heo return -ENOMEM; 4002983c7515STejun Heo 4003983c7515STejun Heo rescuer->rescue_wq = wq; 4004983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4005983c7515STejun Heo ret = PTR_ERR_OR_ZERO(rescuer->task); 4006983c7515STejun Heo if (ret) { 4007983c7515STejun Heo kfree(rescuer); 4008983c7515STejun Heo return ret; 4009983c7515STejun Heo } 4010983c7515STejun Heo 4011983c7515STejun Heo wq->rescuer = rescuer; 4012983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4013983c7515STejun Heo wake_up_process(rescuer->task); 4014983c7515STejun Heo 4015983c7515STejun Heo return 0; 4016983c7515STejun Heo } 4017983c7515STejun Heo 4018b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 401997e37d7bSTejun Heo unsigned int flags, 40201e19ffc6STejun Heo int max_active, 4021eb13ba87SJohannes Berg struct lock_class_key *key, 4022b196be89STejun Heo const char *lock_name, ...) 40233af24433SOleg Nesterov { 4024df2d5ae4STejun Heo size_t tbl_size = 0; 4025ecf6881fSTejun Heo va_list args; 40263af24433SOleg Nesterov struct workqueue_struct *wq; 402749e3cf44STejun Heo struct pool_workqueue *pwq; 4028b196be89STejun Heo 40295c0338c6STejun Heo /* 40305c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 40315c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 40325c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 40335c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 40345c0338c6STejun Heo * on NUMA. 40355c0338c6STejun Heo */ 40365c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 40375c0338c6STejun Heo flags |= __WQ_ORDERED; 40385c0338c6STejun Heo 4039cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4040cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4041cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4042cee22a15SViresh Kumar 4043ecf6881fSTejun Heo /* allocate wq and format name */ 4044df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4045ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4046df2d5ae4STejun Heo 4047df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4048b196be89STejun Heo if (!wq) 4049d2c1d404STejun Heo return NULL; 4050b196be89STejun Heo 40516029a918STejun Heo if (flags & WQ_UNBOUND) { 40526029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40536029a918STejun Heo if (!wq->unbound_attrs) 40546029a918STejun Heo goto err_free_wq; 40556029a918STejun Heo } 40566029a918STejun Heo 4057ecf6881fSTejun Heo va_start(args, lock_name); 4058ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4059b196be89STejun Heo va_end(args); 40603af24433SOleg Nesterov 4061d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4062b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40633af24433SOleg Nesterov 4064b196be89STejun Heo /* init wq */ 406597e37d7bSTejun Heo wq->flags = flags; 4066a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 40673c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4068112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 406930cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 407073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 407173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4072493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 40733af24433SOleg Nesterov 4074eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4075cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 40763af24433SOleg Nesterov 407730cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4078d2c1d404STejun Heo goto err_free_wq; 40791537663fSTejun Heo 408040c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4081d2c1d404STejun Heo goto err_destroy; 4082e22bee78STejun Heo 4083226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4084226223abSTejun Heo goto err_destroy; 4085226223abSTejun Heo 40866af8bf3dSOleg Nesterov /* 408768e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 408868e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 408968e13a67SLai Jiangshan * list. 40906af8bf3dSOleg Nesterov */ 409168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4092a0a1a5fdSTejun Heo 4093a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 409449e3cf44STejun Heo for_each_pwq(pwq, wq) 4095699ce097STejun Heo pwq_adjust_max_active(pwq); 4096a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4097a0a1a5fdSTejun Heo 4098e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4099a0a1a5fdSTejun Heo 410068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41013af24433SOleg Nesterov 41023af24433SOleg Nesterov return wq; 4103d2c1d404STejun Heo 4104d2c1d404STejun Heo err_free_wq: 41056029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41064690c4abSTejun Heo kfree(wq); 4107d2c1d404STejun Heo return NULL; 4108d2c1d404STejun Heo err_destroy: 4109d2c1d404STejun Heo destroy_workqueue(wq); 41104690c4abSTejun Heo return NULL; 41111da177e4SLinus Torvalds } 4112d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41131da177e4SLinus Torvalds 41143af24433SOleg Nesterov /** 41153af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41163af24433SOleg Nesterov * @wq: target workqueue 41173af24433SOleg Nesterov * 41183af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41193af24433SOleg Nesterov */ 41203af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41213af24433SOleg Nesterov { 412249e3cf44STejun Heo struct pool_workqueue *pwq; 41234c16bd32STejun Heo int node; 41243af24433SOleg Nesterov 41259c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41269c5a2ba7STejun Heo drain_workqueue(wq); 4127c8efcc25STejun Heo 41286183c009STejun Heo /* sanity checks */ 4129b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 413049e3cf44STejun Heo for_each_pwq(pwq, wq) { 41316183c009STejun Heo int i; 41326183c009STejun Heo 413376af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 413476af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4135b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4136fa07fb6aSTejun Heo show_workqueue_state(); 41376183c009STejun Heo return; 413876af4d93STejun Heo } 413976af4d93STejun Heo } 414076af4d93STejun Heo 41415c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 41428864b4e5STejun Heo WARN_ON(pwq->nr_active) || 414376af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4144b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4145fa07fb6aSTejun Heo show_workqueue_state(); 41466183c009STejun Heo return; 41476183c009STejun Heo } 414876af4d93STejun Heo } 4149b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41506183c009STejun Heo 4151a0a1a5fdSTejun Heo /* 4152a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4153a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4154a0a1a5fdSTejun Heo */ 415568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4156e2dca7adSTejun Heo list_del_rcu(&wq->list); 415768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41583af24433SOleg Nesterov 4159226223abSTejun Heo workqueue_sysfs_unregister(wq); 4160226223abSTejun Heo 4161e2dca7adSTejun Heo if (wq->rescuer) 4162e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4163e22bee78STejun Heo 41648864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 416529c91e99STejun Heo /* 41668864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4167e2dca7adSTejun Heo * schedule RCU free. 416829c91e99STejun Heo */ 4169e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 41708864b4e5STejun Heo } else { 41718864b4e5STejun Heo /* 41728864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 41734c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 41744c16bd32STejun Heo * @wq will be freed when the last pwq is released. 41758864b4e5STejun Heo */ 41764c16bd32STejun Heo for_each_node(node) { 41774c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 41784c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 41794c16bd32STejun Heo put_pwq_unlocked(pwq); 41804c16bd32STejun Heo } 41814c16bd32STejun Heo 41824c16bd32STejun Heo /* 41834c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 41844c16bd32STejun Heo * put. Don't access it afterwards. 41854c16bd32STejun Heo */ 41864c16bd32STejun Heo pwq = wq->dfl_pwq; 41874c16bd32STejun Heo wq->dfl_pwq = NULL; 4188dce90d47STejun Heo put_pwq_unlocked(pwq); 418929c91e99STejun Heo } 41903af24433SOleg Nesterov } 41913af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 41923af24433SOleg Nesterov 4193dcd989cbSTejun Heo /** 4194dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4195dcd989cbSTejun Heo * @wq: target workqueue 4196dcd989cbSTejun Heo * @max_active: new max_active value. 4197dcd989cbSTejun Heo * 4198dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4199dcd989cbSTejun Heo * 4200dcd989cbSTejun Heo * CONTEXT: 4201dcd989cbSTejun Heo * Don't call from IRQ context. 4202dcd989cbSTejun Heo */ 4203dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4204dcd989cbSTejun Heo { 420549e3cf44STejun Heo struct pool_workqueue *pwq; 4206dcd989cbSTejun Heo 42078719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42080a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 42098719dceaSTejun Heo return; 42108719dceaSTejun Heo 4211f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4212dcd989cbSTejun Heo 4213a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4214dcd989cbSTejun Heo 42150a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4216dcd989cbSTejun Heo wq->saved_max_active = max_active; 4217dcd989cbSTejun Heo 4218699ce097STejun Heo for_each_pwq(pwq, wq) 4219699ce097STejun Heo pwq_adjust_max_active(pwq); 4220dcd989cbSTejun Heo 4221a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4222dcd989cbSTejun Heo } 4223dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4224dcd989cbSTejun Heo 4225dcd989cbSTejun Heo /** 422627d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 422727d4ee03SLukas Wunner * 422827d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 422927d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 423027d4ee03SLukas Wunner * 423127d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 423227d4ee03SLukas Wunner */ 423327d4ee03SLukas Wunner struct work_struct *current_work(void) 423427d4ee03SLukas Wunner { 423527d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 423627d4ee03SLukas Wunner 423727d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 423827d4ee03SLukas Wunner } 423927d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 424027d4ee03SLukas Wunner 424127d4ee03SLukas Wunner /** 4242e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4243e6267616STejun Heo * 4244e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4245e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4246d185af30SYacine Belkadi * 4247d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4248e6267616STejun Heo */ 4249e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4250e6267616STejun Heo { 4251e6267616STejun Heo struct worker *worker = current_wq_worker(); 4252e6267616STejun Heo 42536a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4254e6267616STejun Heo } 4255e6267616STejun Heo 4256e6267616STejun Heo /** 4257dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4258dcd989cbSTejun Heo * @cpu: CPU in question 4259dcd989cbSTejun Heo * @wq: target workqueue 4260dcd989cbSTejun Heo * 4261dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4262dcd989cbSTejun Heo * no synchronization around this function and the test result is 4263dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4264dcd989cbSTejun Heo * 4265d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4266d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4267d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4268d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4269d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4270d3251859STejun Heo * 4271d185af30SYacine Belkadi * Return: 4272dcd989cbSTejun Heo * %true if congested, %false otherwise. 4273dcd989cbSTejun Heo */ 4274d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4275dcd989cbSTejun Heo { 42767fb98ea7STejun Heo struct pool_workqueue *pwq; 427776af4d93STejun Heo bool ret; 427876af4d93STejun Heo 427988109453SLai Jiangshan rcu_read_lock_sched(); 42807fb98ea7STejun Heo 4281d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4282d3251859STejun Heo cpu = smp_processor_id(); 4283d3251859STejun Heo 42847fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 42857fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 42867fb98ea7STejun Heo else 4287df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4288dcd989cbSTejun Heo 428976af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 429088109453SLai Jiangshan rcu_read_unlock_sched(); 429176af4d93STejun Heo 429276af4d93STejun Heo return ret; 4293dcd989cbSTejun Heo } 4294dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4295dcd989cbSTejun Heo 4296dcd989cbSTejun Heo /** 4297dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4298dcd989cbSTejun Heo * @work: the work to be tested 4299dcd989cbSTejun Heo * 4300dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4301dcd989cbSTejun Heo * synchronization around this function and the test result is 4302dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4303dcd989cbSTejun Heo * 4304d185af30SYacine Belkadi * Return: 4305dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4306dcd989cbSTejun Heo */ 4307dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4308dcd989cbSTejun Heo { 4309fa1b54e6STejun Heo struct worker_pool *pool; 4310dcd989cbSTejun Heo unsigned long flags; 4311dcd989cbSTejun Heo unsigned int ret = 0; 4312dcd989cbSTejun Heo 4313dcd989cbSTejun Heo if (work_pending(work)) 4314dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4315038366c5SLai Jiangshan 4316fa1b54e6STejun Heo local_irq_save(flags); 4317fa1b54e6STejun Heo pool = get_work_pool(work); 4318038366c5SLai Jiangshan if (pool) { 4319fa1b54e6STejun Heo spin_lock(&pool->lock); 4320c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4321dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4322fa1b54e6STejun Heo spin_unlock(&pool->lock); 4323038366c5SLai Jiangshan } 4324fa1b54e6STejun Heo local_irq_restore(flags); 4325dcd989cbSTejun Heo 4326dcd989cbSTejun Heo return ret; 4327dcd989cbSTejun Heo } 4328dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4329dcd989cbSTejun Heo 43303d1cb205STejun Heo /** 43313d1cb205STejun Heo * set_worker_desc - set description for the current work item 43323d1cb205STejun Heo * @fmt: printf-style format string 43333d1cb205STejun Heo * @...: arguments for the format string 43343d1cb205STejun Heo * 43353d1cb205STejun Heo * This function can be called by a running work function to describe what 43363d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43373d1cb205STejun Heo * information will be printed out together to help debugging. The 43383d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43393d1cb205STejun Heo */ 43403d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43413d1cb205STejun Heo { 43423d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43433d1cb205STejun Heo va_list args; 43443d1cb205STejun Heo 43453d1cb205STejun Heo if (worker) { 43463d1cb205STejun Heo va_start(args, fmt); 43473d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 43483d1cb205STejun Heo va_end(args); 43493d1cb205STejun Heo worker->desc_valid = true; 43503d1cb205STejun Heo } 43513d1cb205STejun Heo } 43523d1cb205STejun Heo 43533d1cb205STejun Heo /** 43543d1cb205STejun Heo * print_worker_info - print out worker information and description 43553d1cb205STejun Heo * @log_lvl: the log level to use when printing 43563d1cb205STejun Heo * @task: target task 43573d1cb205STejun Heo * 43583d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 43593d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 43603d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 43613d1cb205STejun Heo * 43623d1cb205STejun Heo * This function can be safely called on any task as long as the 43633d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 43643d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 43653d1cb205STejun Heo */ 43663d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 43673d1cb205STejun Heo { 43683d1cb205STejun Heo work_func_t *fn = NULL; 43693d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 43703d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 43713d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 43723d1cb205STejun Heo struct workqueue_struct *wq = NULL; 43733d1cb205STejun Heo bool desc_valid = false; 43743d1cb205STejun Heo struct worker *worker; 43753d1cb205STejun Heo 43763d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 43773d1cb205STejun Heo return; 43783d1cb205STejun Heo 43793d1cb205STejun Heo /* 43803d1cb205STejun Heo * This function is called without any synchronization and @task 43813d1cb205STejun Heo * could be in any state. Be careful with dereferences. 43823d1cb205STejun Heo */ 4383e700591aSPetr Mladek worker = kthread_probe_data(task); 43843d1cb205STejun Heo 43853d1cb205STejun Heo /* 43863d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 43873d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 43883d1cb205STejun Heo */ 43893d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 43903d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 43913d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 43923d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 43933d1cb205STejun Heo 43943d1cb205STejun Heo /* copy worker description */ 43953d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 43963d1cb205STejun Heo if (desc_valid) 43973d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 43983d1cb205STejun Heo 43993d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44003d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 44013d1cb205STejun Heo if (desc[0]) 44023d1cb205STejun Heo pr_cont(" (%s)", desc); 44033d1cb205STejun Heo pr_cont("\n"); 44043d1cb205STejun Heo } 44053d1cb205STejun Heo } 44063d1cb205STejun Heo 44073494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 44083494fc30STejun Heo { 44093494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 44103494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 44113494fc30STejun Heo pr_cont(" node=%d", pool->node); 44123494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 44133494fc30STejun Heo } 44143494fc30STejun Heo 44153494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 44163494fc30STejun Heo { 44173494fc30STejun Heo if (work->func == wq_barrier_func) { 44183494fc30STejun Heo struct wq_barrier *barr; 44193494fc30STejun Heo 44203494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 44213494fc30STejun Heo 44223494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 44233494fc30STejun Heo task_pid_nr(barr->task)); 44243494fc30STejun Heo } else { 44253494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 44263494fc30STejun Heo } 44273494fc30STejun Heo } 44283494fc30STejun Heo 44293494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 44303494fc30STejun Heo { 44313494fc30STejun Heo struct worker_pool *pool = pwq->pool; 44323494fc30STejun Heo struct work_struct *work; 44333494fc30STejun Heo struct worker *worker; 44343494fc30STejun Heo bool has_in_flight = false, has_pending = false; 44353494fc30STejun Heo int bkt; 44363494fc30STejun Heo 44373494fc30STejun Heo pr_info(" pwq %d:", pool->id); 44383494fc30STejun Heo pr_cont_pool_info(pool); 44393494fc30STejun Heo 44403494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 44413494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 44423494fc30STejun Heo 44433494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 44443494fc30STejun Heo if (worker->current_pwq == pwq) { 44453494fc30STejun Heo has_in_flight = true; 44463494fc30STejun Heo break; 44473494fc30STejun Heo } 44483494fc30STejun Heo } 44493494fc30STejun Heo if (has_in_flight) { 44503494fc30STejun Heo bool comma = false; 44513494fc30STejun Heo 44523494fc30STejun Heo pr_info(" in-flight:"); 44533494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 44543494fc30STejun Heo if (worker->current_pwq != pwq) 44553494fc30STejun Heo continue; 44563494fc30STejun Heo 44573494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 44583494fc30STejun Heo task_pid_nr(worker->task), 44593494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 44603494fc30STejun Heo worker->current_func); 44613494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 44623494fc30STejun Heo pr_cont_work(false, work); 44633494fc30STejun Heo comma = true; 44643494fc30STejun Heo } 44653494fc30STejun Heo pr_cont("\n"); 44663494fc30STejun Heo } 44673494fc30STejun Heo 44683494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 44693494fc30STejun Heo if (get_work_pwq(work) == pwq) { 44703494fc30STejun Heo has_pending = true; 44713494fc30STejun Heo break; 44723494fc30STejun Heo } 44733494fc30STejun Heo } 44743494fc30STejun Heo if (has_pending) { 44753494fc30STejun Heo bool comma = false; 44763494fc30STejun Heo 44773494fc30STejun Heo pr_info(" pending:"); 44783494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 44793494fc30STejun Heo if (get_work_pwq(work) != pwq) 44803494fc30STejun Heo continue; 44813494fc30STejun Heo 44823494fc30STejun Heo pr_cont_work(comma, work); 44833494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 44843494fc30STejun Heo } 44853494fc30STejun Heo pr_cont("\n"); 44863494fc30STejun Heo } 44873494fc30STejun Heo 44883494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 44893494fc30STejun Heo bool comma = false; 44903494fc30STejun Heo 44913494fc30STejun Heo pr_info(" delayed:"); 44923494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 44933494fc30STejun Heo pr_cont_work(comma, work); 44943494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 44953494fc30STejun Heo } 44963494fc30STejun Heo pr_cont("\n"); 44973494fc30STejun Heo } 44983494fc30STejun Heo } 44993494fc30STejun Heo 45003494fc30STejun Heo /** 45013494fc30STejun Heo * show_workqueue_state - dump workqueue state 45023494fc30STejun Heo * 45037b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 45047b776af6SRoger Lu * all busy workqueues and pools. 45053494fc30STejun Heo */ 45063494fc30STejun Heo void show_workqueue_state(void) 45073494fc30STejun Heo { 45083494fc30STejun Heo struct workqueue_struct *wq; 45093494fc30STejun Heo struct worker_pool *pool; 45103494fc30STejun Heo unsigned long flags; 45113494fc30STejun Heo int pi; 45123494fc30STejun Heo 45133494fc30STejun Heo rcu_read_lock_sched(); 45143494fc30STejun Heo 45153494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 45163494fc30STejun Heo 45173494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 45183494fc30STejun Heo struct pool_workqueue *pwq; 45193494fc30STejun Heo bool idle = true; 45203494fc30STejun Heo 45213494fc30STejun Heo for_each_pwq(pwq, wq) { 45223494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 45233494fc30STejun Heo idle = false; 45243494fc30STejun Heo break; 45253494fc30STejun Heo } 45263494fc30STejun Heo } 45273494fc30STejun Heo if (idle) 45283494fc30STejun Heo continue; 45293494fc30STejun Heo 45303494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 45313494fc30STejun Heo 45323494fc30STejun Heo for_each_pwq(pwq, wq) { 45333494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 45343494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 45353494fc30STejun Heo show_pwq(pwq); 45363494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 453762635ea8SSergey Senozhatsky /* 453862635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 453962635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 454062635ea8SSergey Senozhatsky * hard lockup. 454162635ea8SSergey Senozhatsky */ 454262635ea8SSergey Senozhatsky touch_nmi_watchdog(); 45433494fc30STejun Heo } 45443494fc30STejun Heo } 45453494fc30STejun Heo 45463494fc30STejun Heo for_each_pool(pool, pi) { 45473494fc30STejun Heo struct worker *worker; 45483494fc30STejun Heo bool first = true; 45493494fc30STejun Heo 45503494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 45513494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 45523494fc30STejun Heo goto next_pool; 45533494fc30STejun Heo 45543494fc30STejun Heo pr_info("pool %d:", pool->id); 45553494fc30STejun Heo pr_cont_pool_info(pool); 455682607adcSTejun Heo pr_cont(" hung=%us workers=%d", 455782607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 455882607adcSTejun Heo pool->nr_workers); 45593494fc30STejun Heo if (pool->manager) 45603494fc30STejun Heo pr_cont(" manager: %d", 45613494fc30STejun Heo task_pid_nr(pool->manager->task)); 45623494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 45633494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 45643494fc30STejun Heo task_pid_nr(worker->task)); 45653494fc30STejun Heo first = false; 45663494fc30STejun Heo } 45673494fc30STejun Heo pr_cont("\n"); 45683494fc30STejun Heo next_pool: 45693494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 457062635ea8SSergey Senozhatsky /* 457162635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 457262635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 457362635ea8SSergey Senozhatsky * hard lockup. 457462635ea8SSergey Senozhatsky */ 457562635ea8SSergey Senozhatsky touch_nmi_watchdog(); 45763494fc30STejun Heo } 45773494fc30STejun Heo 45783494fc30STejun Heo rcu_read_unlock_sched(); 45793494fc30STejun Heo } 45803494fc30STejun Heo 4581db7bccf4STejun Heo /* 4582db7bccf4STejun Heo * CPU hotplug. 4583db7bccf4STejun Heo * 4584e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4585112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4586706026c2STejun Heo * pool which make migrating pending and scheduled works very 4587e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 458894cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4589e22bee78STejun Heo * blocked draining impractical. 4590e22bee78STejun Heo * 459124647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4592628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4593628c78e7STejun Heo * cpu comes back online. 4594db7bccf4STejun Heo */ 4595db7bccf4STejun Heo 4596e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4597db7bccf4STejun Heo { 45984ce62e9eSTejun Heo struct worker_pool *pool; 4599db7bccf4STejun Heo struct worker *worker; 4600db7bccf4STejun Heo 4601f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46021258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 460394cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4604e22bee78STejun Heo 4605f2d5a0eeSTejun Heo /* 460692f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 460794cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 460894cf58bbSTejun Heo * except for the ones which are still executing works from 460994cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 461094cf58bbSTejun Heo * this, they may become diasporas. 4611f2d5a0eeSTejun Heo */ 4612da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4613403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4614db7bccf4STejun Heo 461524647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4616f2d5a0eeSTejun Heo 461794cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 46181258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 4619e22bee78STejun Heo 4620e22bee78STejun Heo /* 4621eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4622eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4623eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4624eb283428SLai Jiangshan * from other cpus. 4625628c78e7STejun Heo */ 4626628c78e7STejun Heo schedule(); 4627628c78e7STejun Heo 4628628c78e7STejun Heo /* 4629eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4630eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4631eb283428SLai Jiangshan * and keep_working() are always true as long as the 4632eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4633eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4634eb283428SLai Jiangshan * are served by workers tied to the pool. 4635e22bee78STejun Heo */ 4636e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4637eb283428SLai Jiangshan 4638eb283428SLai Jiangshan /* 4639eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4640eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4641eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4642eb283428SLai Jiangshan */ 4643eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4644eb283428SLai Jiangshan wake_up_worker(pool); 4645eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4646eb283428SLai Jiangshan } 4647db7bccf4STejun Heo } 4648db7bccf4STejun Heo 4649bd7c089eSTejun Heo /** 4650bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4651bd7c089eSTejun Heo * @pool: pool of interest 4652bd7c089eSTejun Heo * 4653a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4654bd7c089eSTejun Heo */ 4655bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4656bd7c089eSTejun Heo { 4657a9ab775bSTejun Heo struct worker *worker; 4658bd7c089eSTejun Heo 46591258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 4660bd7c089eSTejun Heo 4661bd7c089eSTejun Heo /* 4662a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4663a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4664402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4665a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4666a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4667bd7c089eSTejun Heo */ 4668da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4669a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4670a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4671a9ab775bSTejun Heo 4672a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4673f7c17d26SWanpeng Li 46743de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4675a9ab775bSTejun Heo 4676da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4677a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4678a9ab775bSTejun Heo 4679a9ab775bSTejun Heo /* 4680a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4681a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4682a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4683a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4684a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4685a9ab775bSTejun Heo * be bound before @pool->lock is released. 4686a9ab775bSTejun Heo */ 4687a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4688bd7c089eSTejun Heo wake_up_process(worker->task); 4689bd7c089eSTejun Heo 4690bd7c089eSTejun Heo /* 4691a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4692a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4693a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4694a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4695a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4696a9ab775bSTejun Heo * concurrency management. Note that when or whether 4697a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4698a9ab775bSTejun Heo * 4699c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 4700a9ab775bSTejun Heo * tested without holding any lock in 4701a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4702a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4703a9ab775bSTejun Heo * management operations. 4704bd7c089eSTejun Heo */ 4705a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4706a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4707a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4708c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 4709bd7c089eSTejun Heo } 4710a9ab775bSTejun Heo 4711a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4712bd7c089eSTejun Heo } 4713bd7c089eSTejun Heo 47147dbc725eSTejun Heo /** 47157dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 47167dbc725eSTejun Heo * @pool: unbound pool of interest 47177dbc725eSTejun Heo * @cpu: the CPU which is coming up 47187dbc725eSTejun Heo * 47197dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 47207dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 47217dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 47227dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 47237dbc725eSTejun Heo */ 47247dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 47257dbc725eSTejun Heo { 47267dbc725eSTejun Heo static cpumask_t cpumask; 47277dbc725eSTejun Heo struct worker *worker; 47287dbc725eSTejun Heo 47291258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 47307dbc725eSTejun Heo 47317dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 47327dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 47337dbc725eSTejun Heo return; 47347dbc725eSTejun Heo 47357dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 47367dbc725eSTejun Heo 47377dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4738da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4739d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 47407dbc725eSTejun Heo } 47417dbc725eSTejun Heo 47427ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 47431da177e4SLinus Torvalds { 47444ce62e9eSTejun Heo struct worker_pool *pool; 47451da177e4SLinus Torvalds 4746f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 47473ce63377STejun Heo if (pool->nr_workers) 47483ce63377STejun Heo continue; 4749051e1850SLai Jiangshan if (!create_worker(pool)) 47507ee681b2SThomas Gleixner return -ENOMEM; 47513af24433SOleg Nesterov } 47527ee681b2SThomas Gleixner return 0; 47537ee681b2SThomas Gleixner } 47541da177e4SLinus Torvalds 47557ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 47567ee681b2SThomas Gleixner { 47577ee681b2SThomas Gleixner struct worker_pool *pool; 47587ee681b2SThomas Gleixner struct workqueue_struct *wq; 47597ee681b2SThomas Gleixner int pi; 47607ee681b2SThomas Gleixner 476168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 47627dbc725eSTejun Heo 47637dbc725eSTejun Heo for_each_pool(pool, pi) { 47641258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 476594cf58bbSTejun Heo 4766f05b558dSLai Jiangshan if (pool->cpu == cpu) 476794cf58bbSTejun Heo rebind_workers(pool); 4768f05b558dSLai Jiangshan else if (pool->cpu < 0) 47697dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 477094cf58bbSTejun Heo 47711258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 477294cf58bbSTejun Heo } 47737dbc725eSTejun Heo 47744c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47754c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47764c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 47774c16bd32STejun Heo 477868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47797ee681b2SThomas Gleixner return 0; 478065758202STejun Heo } 478165758202STejun Heo 47827ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 478365758202STejun Heo { 47844c16bd32STejun Heo struct workqueue_struct *wq; 47858db25e78STejun Heo 47864c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4787e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 4788e8b3f8dbSLai Jiangshan return -1; 4789e8b3f8dbSLai Jiangshan 4790e8b3f8dbSLai Jiangshan unbind_workers(cpu); 47914c16bd32STejun Heo 47924c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47934c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 47944c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47954c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 47964c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 47974c16bd32STejun Heo 47987ee681b2SThomas Gleixner return 0; 479965758202STejun Heo } 480065758202STejun Heo 48012d3854a3SRusty Russell #ifdef CONFIG_SMP 48028ccad40dSRusty Russell 48032d3854a3SRusty Russell struct work_for_cpu { 4804ed48ece2STejun Heo struct work_struct work; 48052d3854a3SRusty Russell long (*fn)(void *); 48062d3854a3SRusty Russell void *arg; 48072d3854a3SRusty Russell long ret; 48082d3854a3SRusty Russell }; 48092d3854a3SRusty Russell 4810ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 48112d3854a3SRusty Russell { 4812ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4813ed48ece2STejun Heo 48142d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 48152d3854a3SRusty Russell } 48162d3854a3SRusty Russell 48172d3854a3SRusty Russell /** 481822aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 48192d3854a3SRusty Russell * @cpu: the cpu to run on 48202d3854a3SRusty Russell * @fn: the function to run 48212d3854a3SRusty Russell * @arg: the function arg 48222d3854a3SRusty Russell * 482331ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 48246b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4825d185af30SYacine Belkadi * 4826d185af30SYacine Belkadi * Return: The value @fn returns. 48272d3854a3SRusty Russell */ 4828d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 48292d3854a3SRusty Russell { 4830ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 48312d3854a3SRusty Russell 4832ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4833ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 483412997d1aSBjorn Helgaas flush_work(&wfc.work); 4835440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 48362d3854a3SRusty Russell return wfc.ret; 48372d3854a3SRusty Russell } 48382d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 48390e8d6a93SThomas Gleixner 48400e8d6a93SThomas Gleixner /** 48410e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 48420e8d6a93SThomas Gleixner * @cpu: the cpu to run on 48430e8d6a93SThomas Gleixner * @fn: the function to run 48440e8d6a93SThomas Gleixner * @arg: the function argument 48450e8d6a93SThomas Gleixner * 48460e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 48470e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 48480e8d6a93SThomas Gleixner * 48490e8d6a93SThomas Gleixner * Return: The value @fn returns. 48500e8d6a93SThomas Gleixner */ 48510e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 48520e8d6a93SThomas Gleixner { 48530e8d6a93SThomas Gleixner long ret = -ENODEV; 48540e8d6a93SThomas Gleixner 48550e8d6a93SThomas Gleixner get_online_cpus(); 48560e8d6a93SThomas Gleixner if (cpu_online(cpu)) 48570e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 48580e8d6a93SThomas Gleixner put_online_cpus(); 48590e8d6a93SThomas Gleixner return ret; 48600e8d6a93SThomas Gleixner } 48610e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 48622d3854a3SRusty Russell #endif /* CONFIG_SMP */ 48632d3854a3SRusty Russell 4864a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4865e7577c50SRusty Russell 4866a0a1a5fdSTejun Heo /** 4867a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4868a0a1a5fdSTejun Heo * 486958a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4870c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4871706026c2STejun Heo * pool->worklist. 4872a0a1a5fdSTejun Heo * 4873a0a1a5fdSTejun Heo * CONTEXT: 4874a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4875a0a1a5fdSTejun Heo */ 4876a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4877a0a1a5fdSTejun Heo { 487824b8a847STejun Heo struct workqueue_struct *wq; 487924b8a847STejun Heo struct pool_workqueue *pwq; 4880a0a1a5fdSTejun Heo 488168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4882a0a1a5fdSTejun Heo 48836183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4884a0a1a5fdSTejun Heo workqueue_freezing = true; 4885a0a1a5fdSTejun Heo 488624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4887a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4888699ce097STejun Heo for_each_pwq(pwq, wq) 4889699ce097STejun Heo pwq_adjust_max_active(pwq); 4890a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4891a1056305STejun Heo } 48925bcab335STejun Heo 489368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4894a0a1a5fdSTejun Heo } 4895a0a1a5fdSTejun Heo 4896a0a1a5fdSTejun Heo /** 489758a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4898a0a1a5fdSTejun Heo * 4899a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4900a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4901a0a1a5fdSTejun Heo * 4902a0a1a5fdSTejun Heo * CONTEXT: 490368e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4904a0a1a5fdSTejun Heo * 4905d185af30SYacine Belkadi * Return: 490658a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 490758a69cb4STejun Heo * is complete. 4908a0a1a5fdSTejun Heo */ 4909a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4910a0a1a5fdSTejun Heo { 4911a0a1a5fdSTejun Heo bool busy = false; 491224b8a847STejun Heo struct workqueue_struct *wq; 491324b8a847STejun Heo struct pool_workqueue *pwq; 4914a0a1a5fdSTejun Heo 491568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4916a0a1a5fdSTejun Heo 49176183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4918a0a1a5fdSTejun Heo 491924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 492024b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 492124b8a847STejun Heo continue; 4922a0a1a5fdSTejun Heo /* 4923a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4924a0a1a5fdSTejun Heo * to peek without lock. 4925a0a1a5fdSTejun Heo */ 492688109453SLai Jiangshan rcu_read_lock_sched(); 492724b8a847STejun Heo for_each_pwq(pwq, wq) { 49286183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4929112202d9STejun Heo if (pwq->nr_active) { 4930a0a1a5fdSTejun Heo busy = true; 493188109453SLai Jiangshan rcu_read_unlock_sched(); 4932a0a1a5fdSTejun Heo goto out_unlock; 4933a0a1a5fdSTejun Heo } 4934a0a1a5fdSTejun Heo } 493588109453SLai Jiangshan rcu_read_unlock_sched(); 4936a0a1a5fdSTejun Heo } 4937a0a1a5fdSTejun Heo out_unlock: 493868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4939a0a1a5fdSTejun Heo return busy; 4940a0a1a5fdSTejun Heo } 4941a0a1a5fdSTejun Heo 4942a0a1a5fdSTejun Heo /** 4943a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4944a0a1a5fdSTejun Heo * 4945a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4946706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4947a0a1a5fdSTejun Heo * 4948a0a1a5fdSTejun Heo * CONTEXT: 4949a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4950a0a1a5fdSTejun Heo */ 4951a0a1a5fdSTejun Heo void thaw_workqueues(void) 4952a0a1a5fdSTejun Heo { 495324b8a847STejun Heo struct workqueue_struct *wq; 495424b8a847STejun Heo struct pool_workqueue *pwq; 4955a0a1a5fdSTejun Heo 495668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4957a0a1a5fdSTejun Heo 4958a0a1a5fdSTejun Heo if (!workqueue_freezing) 4959a0a1a5fdSTejun Heo goto out_unlock; 4960a0a1a5fdSTejun Heo 496174b414eaSLai Jiangshan workqueue_freezing = false; 496224b8a847STejun Heo 496324b8a847STejun Heo /* restore max_active and repopulate worklist */ 496424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4965a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4966699ce097STejun Heo for_each_pwq(pwq, wq) 4967699ce097STejun Heo pwq_adjust_max_active(pwq); 4968a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 496924b8a847STejun Heo } 497024b8a847STejun Heo 4971a0a1a5fdSTejun Heo out_unlock: 497268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4973a0a1a5fdSTejun Heo } 4974a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4975a0a1a5fdSTejun Heo 4976042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4977042f7df1SLai Jiangshan { 4978042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4979042f7df1SLai Jiangshan int ret = 0; 4980042f7df1SLai Jiangshan struct workqueue_struct *wq; 4981042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4982042f7df1SLai Jiangshan 4983042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4984042f7df1SLai Jiangshan 4985042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4986042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4987042f7df1SLai Jiangshan continue; 4988042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 4989042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 4990042f7df1SLai Jiangshan continue; 4991042f7df1SLai Jiangshan 4992042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 4993042f7df1SLai Jiangshan if (!ctx) { 4994042f7df1SLai Jiangshan ret = -ENOMEM; 4995042f7df1SLai Jiangshan break; 4996042f7df1SLai Jiangshan } 4997042f7df1SLai Jiangshan 4998042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 4999042f7df1SLai Jiangshan } 5000042f7df1SLai Jiangshan 5001042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5002042f7df1SLai Jiangshan if (!ret) 5003042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5004042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5005042f7df1SLai Jiangshan } 5006042f7df1SLai Jiangshan 5007042f7df1SLai Jiangshan return ret; 5008042f7df1SLai Jiangshan } 5009042f7df1SLai Jiangshan 5010042f7df1SLai Jiangshan /** 5011042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5012042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5013042f7df1SLai Jiangshan * 5014042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5015042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5016042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5017042f7df1SLai Jiangshan * 5018042f7df1SLai Jiangshan * Retun: 0 - Success 5019042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5020042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5021042f7df1SLai Jiangshan */ 5022042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5023042f7df1SLai Jiangshan { 5024042f7df1SLai Jiangshan int ret = -EINVAL; 5025042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5026042f7df1SLai Jiangshan 5027042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 5028042f7df1SLai Jiangshan return -ENOMEM; 5029042f7df1SLai Jiangshan 5030c98a9805STal Shorer /* 5031c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5032c98a9805STal Shorer * If the user wishes to include them, we allow that. 5033c98a9805STal Shorer */ 5034042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5035042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5036a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5037042f7df1SLai Jiangshan 5038042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5039042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5040042f7df1SLai Jiangshan 5041042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5042042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5043042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5044042f7df1SLai Jiangshan 5045042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5046042f7df1SLai Jiangshan if (ret < 0) 5047042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5048042f7df1SLai Jiangshan 5049a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5050042f7df1SLai Jiangshan } 5051042f7df1SLai Jiangshan 5052042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 5053042f7df1SLai Jiangshan return ret; 5054042f7df1SLai Jiangshan } 5055042f7df1SLai Jiangshan 50566ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 50576ba94429SFrederic Weisbecker /* 50586ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 50596ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 50606ba94429SFrederic Weisbecker * following attributes. 50616ba94429SFrederic Weisbecker * 50626ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 50636ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 50646ba94429SFrederic Weisbecker * 50656ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 50666ba94429SFrederic Weisbecker * 50679a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 50686ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 50696ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 50709a19b463SWang Long * numa RW bool : whether enable NUMA affinity 50716ba94429SFrederic Weisbecker */ 50726ba94429SFrederic Weisbecker struct wq_device { 50736ba94429SFrederic Weisbecker struct workqueue_struct *wq; 50746ba94429SFrederic Weisbecker struct device dev; 50756ba94429SFrederic Weisbecker }; 50766ba94429SFrederic Weisbecker 50776ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 50786ba94429SFrederic Weisbecker { 50796ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 50806ba94429SFrederic Weisbecker 50816ba94429SFrederic Weisbecker return wq_dev->wq; 50826ba94429SFrederic Weisbecker } 50836ba94429SFrederic Weisbecker 50846ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 50856ba94429SFrederic Weisbecker char *buf) 50866ba94429SFrederic Weisbecker { 50876ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50886ba94429SFrederic Weisbecker 50896ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 50906ba94429SFrederic Weisbecker } 50916ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 50926ba94429SFrederic Weisbecker 50936ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 50946ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 50956ba94429SFrederic Weisbecker { 50966ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50976ba94429SFrederic Weisbecker 50986ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 50996ba94429SFrederic Weisbecker } 51006ba94429SFrederic Weisbecker 51016ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 51026ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 51036ba94429SFrederic Weisbecker size_t count) 51046ba94429SFrederic Weisbecker { 51056ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51066ba94429SFrederic Weisbecker int val; 51076ba94429SFrederic Weisbecker 51086ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 51096ba94429SFrederic Weisbecker return -EINVAL; 51106ba94429SFrederic Weisbecker 51116ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 51126ba94429SFrederic Weisbecker return count; 51136ba94429SFrederic Weisbecker } 51146ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 51156ba94429SFrederic Weisbecker 51166ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 51176ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 51186ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 51196ba94429SFrederic Weisbecker NULL, 51206ba94429SFrederic Weisbecker }; 51216ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 51226ba94429SFrederic Weisbecker 51236ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 51246ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 51256ba94429SFrederic Weisbecker { 51266ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51276ba94429SFrederic Weisbecker const char *delim = ""; 51286ba94429SFrederic Weisbecker int node, written = 0; 51296ba94429SFrederic Weisbecker 51306ba94429SFrederic Weisbecker rcu_read_lock_sched(); 51316ba94429SFrederic Weisbecker for_each_node(node) { 51326ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 51336ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 51346ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 51356ba94429SFrederic Weisbecker delim = " "; 51366ba94429SFrederic Weisbecker } 51376ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 51386ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 51396ba94429SFrederic Weisbecker 51406ba94429SFrederic Weisbecker return written; 51416ba94429SFrederic Weisbecker } 51426ba94429SFrederic Weisbecker 51436ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 51446ba94429SFrederic Weisbecker char *buf) 51456ba94429SFrederic Weisbecker { 51466ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51476ba94429SFrederic Weisbecker int written; 51486ba94429SFrederic Weisbecker 51496ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 51506ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 51516ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 51526ba94429SFrederic Weisbecker 51536ba94429SFrederic Weisbecker return written; 51546ba94429SFrederic Weisbecker } 51556ba94429SFrederic Weisbecker 51566ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 51576ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 51586ba94429SFrederic Weisbecker { 51596ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 51606ba94429SFrederic Weisbecker 5161899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5162899a94feSLai Jiangshan 51636ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 51646ba94429SFrederic Weisbecker if (!attrs) 51656ba94429SFrederic Weisbecker return NULL; 51666ba94429SFrederic Weisbecker 51676ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 51686ba94429SFrederic Weisbecker return attrs; 51696ba94429SFrederic Weisbecker } 51706ba94429SFrederic Weisbecker 51716ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 51726ba94429SFrederic Weisbecker const char *buf, size_t count) 51736ba94429SFrederic Weisbecker { 51746ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51756ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5176d4d3e257SLai Jiangshan int ret = -ENOMEM; 5177d4d3e257SLai Jiangshan 5178d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51796ba94429SFrederic Weisbecker 51806ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51816ba94429SFrederic Weisbecker if (!attrs) 5182d4d3e257SLai Jiangshan goto out_unlock; 51836ba94429SFrederic Weisbecker 51846ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 51856ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5186d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51876ba94429SFrederic Weisbecker else 51886ba94429SFrederic Weisbecker ret = -EINVAL; 51896ba94429SFrederic Weisbecker 5190d4d3e257SLai Jiangshan out_unlock: 5191d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 51926ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 51936ba94429SFrederic Weisbecker return ret ?: count; 51946ba94429SFrederic Weisbecker } 51956ba94429SFrederic Weisbecker 51966ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 51976ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 51986ba94429SFrederic Weisbecker { 51996ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52006ba94429SFrederic Weisbecker int written; 52016ba94429SFrederic Weisbecker 52026ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 52036ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 52046ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 52056ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 52066ba94429SFrederic Weisbecker return written; 52076ba94429SFrederic Weisbecker } 52086ba94429SFrederic Weisbecker 52096ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 52106ba94429SFrederic Weisbecker struct device_attribute *attr, 52116ba94429SFrederic Weisbecker const char *buf, size_t count) 52126ba94429SFrederic Weisbecker { 52136ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52146ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5215d4d3e257SLai Jiangshan int ret = -ENOMEM; 5216d4d3e257SLai Jiangshan 5217d4d3e257SLai Jiangshan apply_wqattrs_lock(); 52186ba94429SFrederic Weisbecker 52196ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 52206ba94429SFrederic Weisbecker if (!attrs) 5221d4d3e257SLai Jiangshan goto out_unlock; 52226ba94429SFrederic Weisbecker 52236ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 52246ba94429SFrederic Weisbecker if (!ret) 5225d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 52266ba94429SFrederic Weisbecker 5227d4d3e257SLai Jiangshan out_unlock: 5228d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 52296ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 52306ba94429SFrederic Weisbecker return ret ?: count; 52316ba94429SFrederic Weisbecker } 52326ba94429SFrederic Weisbecker 52336ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 52346ba94429SFrederic Weisbecker char *buf) 52356ba94429SFrederic Weisbecker { 52366ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52376ba94429SFrederic Weisbecker int written; 52386ba94429SFrederic Weisbecker 52396ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 52406ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 52416ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 52426ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 52436ba94429SFrederic Weisbecker 52446ba94429SFrederic Weisbecker return written; 52456ba94429SFrederic Weisbecker } 52466ba94429SFrederic Weisbecker 52476ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 52486ba94429SFrederic Weisbecker const char *buf, size_t count) 52496ba94429SFrederic Weisbecker { 52506ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52516ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5252d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5253d4d3e257SLai Jiangshan 5254d4d3e257SLai Jiangshan apply_wqattrs_lock(); 52556ba94429SFrederic Weisbecker 52566ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 52576ba94429SFrederic Weisbecker if (!attrs) 5258d4d3e257SLai Jiangshan goto out_unlock; 52596ba94429SFrederic Weisbecker 52606ba94429SFrederic Weisbecker ret = -EINVAL; 52616ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 52626ba94429SFrederic Weisbecker attrs->no_numa = !v; 5263d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 52646ba94429SFrederic Weisbecker } 52656ba94429SFrederic Weisbecker 5266d4d3e257SLai Jiangshan out_unlock: 5267d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 52686ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 52696ba94429SFrederic Weisbecker return ret ?: count; 52706ba94429SFrederic Weisbecker } 52716ba94429SFrederic Weisbecker 52726ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 52736ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 52746ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 52756ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 52766ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 52776ba94429SFrederic Weisbecker __ATTR_NULL, 52786ba94429SFrederic Weisbecker }; 52796ba94429SFrederic Weisbecker 52806ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 52816ba94429SFrederic Weisbecker .name = "workqueue", 52826ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 52836ba94429SFrederic Weisbecker }; 52846ba94429SFrederic Weisbecker 5285b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5286b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5287b05a7928SFrederic Weisbecker { 5288b05a7928SFrederic Weisbecker int written; 5289b05a7928SFrederic Weisbecker 5290042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5291b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5292b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5293042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5294b05a7928SFrederic Weisbecker 5295b05a7928SFrederic Weisbecker return written; 5296b05a7928SFrederic Weisbecker } 5297b05a7928SFrederic Weisbecker 5298042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5299042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5300042f7df1SLai Jiangshan { 5301042f7df1SLai Jiangshan cpumask_var_t cpumask; 5302042f7df1SLai Jiangshan int ret; 5303042f7df1SLai Jiangshan 5304042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5305042f7df1SLai Jiangshan return -ENOMEM; 5306042f7df1SLai Jiangshan 5307042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5308042f7df1SLai Jiangshan if (!ret) 5309042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5310042f7df1SLai Jiangshan 5311042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5312042f7df1SLai Jiangshan return ret ? ret : count; 5313042f7df1SLai Jiangshan } 5314042f7df1SLai Jiangshan 5315b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5316042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5317042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5318b05a7928SFrederic Weisbecker 53196ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 53206ba94429SFrederic Weisbecker { 5321b05a7928SFrederic Weisbecker int err; 5322b05a7928SFrederic Weisbecker 5323b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5324b05a7928SFrederic Weisbecker if (err) 5325b05a7928SFrederic Weisbecker return err; 5326b05a7928SFrederic Weisbecker 5327b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 53286ba94429SFrederic Weisbecker } 53296ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 53306ba94429SFrederic Weisbecker 53316ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 53326ba94429SFrederic Weisbecker { 53336ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 53346ba94429SFrederic Weisbecker 53356ba94429SFrederic Weisbecker kfree(wq_dev); 53366ba94429SFrederic Weisbecker } 53376ba94429SFrederic Weisbecker 53386ba94429SFrederic Weisbecker /** 53396ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 53406ba94429SFrederic Weisbecker * @wq: the workqueue to register 53416ba94429SFrederic Weisbecker * 53426ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 53436ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 53446ba94429SFrederic Weisbecker * which is the preferred method. 53456ba94429SFrederic Weisbecker * 53466ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 53476ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 53486ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 53496ba94429SFrederic Weisbecker * attributes. 53506ba94429SFrederic Weisbecker * 53516ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 53526ba94429SFrederic Weisbecker */ 53536ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 53546ba94429SFrederic Weisbecker { 53556ba94429SFrederic Weisbecker struct wq_device *wq_dev; 53566ba94429SFrederic Weisbecker int ret; 53576ba94429SFrederic Weisbecker 53586ba94429SFrederic Weisbecker /* 5359402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 53606ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 53616ba94429SFrederic Weisbecker * workqueues. 53626ba94429SFrederic Weisbecker */ 53630a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 53646ba94429SFrederic Weisbecker return -EINVAL; 53656ba94429SFrederic Weisbecker 53666ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 53676ba94429SFrederic Weisbecker if (!wq_dev) 53686ba94429SFrederic Weisbecker return -ENOMEM; 53696ba94429SFrederic Weisbecker 53706ba94429SFrederic Weisbecker wq_dev->wq = wq; 53716ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 53726ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 537323217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 53746ba94429SFrederic Weisbecker 53756ba94429SFrederic Weisbecker /* 53766ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 53776ba94429SFrederic Weisbecker * everything is ready. 53786ba94429SFrederic Weisbecker */ 53796ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 53806ba94429SFrederic Weisbecker 53816ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 53826ba94429SFrederic Weisbecker if (ret) { 5383537f4146SArvind Yadav put_device(&wq_dev->dev); 53846ba94429SFrederic Weisbecker wq->wq_dev = NULL; 53856ba94429SFrederic Weisbecker return ret; 53866ba94429SFrederic Weisbecker } 53876ba94429SFrederic Weisbecker 53886ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 53896ba94429SFrederic Weisbecker struct device_attribute *attr; 53906ba94429SFrederic Weisbecker 53916ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 53926ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 53936ba94429SFrederic Weisbecker if (ret) { 53946ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 53956ba94429SFrederic Weisbecker wq->wq_dev = NULL; 53966ba94429SFrederic Weisbecker return ret; 53976ba94429SFrederic Weisbecker } 53986ba94429SFrederic Weisbecker } 53996ba94429SFrederic Weisbecker } 54006ba94429SFrederic Weisbecker 54016ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 54026ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 54036ba94429SFrederic Weisbecker return 0; 54046ba94429SFrederic Weisbecker } 54056ba94429SFrederic Weisbecker 54066ba94429SFrederic Weisbecker /** 54076ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 54086ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 54096ba94429SFrederic Weisbecker * 54106ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 54116ba94429SFrederic Weisbecker */ 54126ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 54136ba94429SFrederic Weisbecker { 54146ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 54156ba94429SFrederic Weisbecker 54166ba94429SFrederic Weisbecker if (!wq->wq_dev) 54176ba94429SFrederic Weisbecker return; 54186ba94429SFrederic Weisbecker 54196ba94429SFrederic Weisbecker wq->wq_dev = NULL; 54206ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 54216ba94429SFrederic Weisbecker } 54226ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 54236ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 54246ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 54256ba94429SFrederic Weisbecker 542682607adcSTejun Heo /* 542782607adcSTejun Heo * Workqueue watchdog. 542882607adcSTejun Heo * 542982607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 543082607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 543182607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 543282607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 543382607adcSTejun Heo * largely opaque. 543482607adcSTejun Heo * 543582607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 543682607adcSTejun Heo * state if some pools failed to make forward progress for a while where 543782607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 543882607adcSTejun Heo * 543982607adcSTejun Heo * This mechanism is controlled through the kernel parameter 544082607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 544182607adcSTejun Heo * corresponding sysfs parameter file. 544282607adcSTejun Heo */ 544382607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 544482607adcSTejun Heo 544582607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 54465cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 544782607adcSTejun Heo 544882607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 544982607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 545082607adcSTejun Heo 545182607adcSTejun Heo static void wq_watchdog_reset_touched(void) 545282607adcSTejun Heo { 545382607adcSTejun Heo int cpu; 545482607adcSTejun Heo 545582607adcSTejun Heo wq_watchdog_touched = jiffies; 545682607adcSTejun Heo for_each_possible_cpu(cpu) 545782607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 545882607adcSTejun Heo } 545982607adcSTejun Heo 54605cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 546182607adcSTejun Heo { 546282607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 546382607adcSTejun Heo bool lockup_detected = false; 546482607adcSTejun Heo struct worker_pool *pool; 546582607adcSTejun Heo int pi; 546682607adcSTejun Heo 546782607adcSTejun Heo if (!thresh) 546882607adcSTejun Heo return; 546982607adcSTejun Heo 547082607adcSTejun Heo rcu_read_lock(); 547182607adcSTejun Heo 547282607adcSTejun Heo for_each_pool(pool, pi) { 547382607adcSTejun Heo unsigned long pool_ts, touched, ts; 547482607adcSTejun Heo 547582607adcSTejun Heo if (list_empty(&pool->worklist)) 547682607adcSTejun Heo continue; 547782607adcSTejun Heo 547882607adcSTejun Heo /* get the latest of pool and touched timestamps */ 547982607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 548082607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 548182607adcSTejun Heo 548282607adcSTejun Heo if (time_after(pool_ts, touched)) 548382607adcSTejun Heo ts = pool_ts; 548482607adcSTejun Heo else 548582607adcSTejun Heo ts = touched; 548682607adcSTejun Heo 548782607adcSTejun Heo if (pool->cpu >= 0) { 548882607adcSTejun Heo unsigned long cpu_touched = 548982607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 549082607adcSTejun Heo pool->cpu)); 549182607adcSTejun Heo if (time_after(cpu_touched, ts)) 549282607adcSTejun Heo ts = cpu_touched; 549382607adcSTejun Heo } 549482607adcSTejun Heo 549582607adcSTejun Heo /* did we stall? */ 549682607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 549782607adcSTejun Heo lockup_detected = true; 549882607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 549982607adcSTejun Heo pr_cont_pool_info(pool); 550082607adcSTejun Heo pr_cont(" stuck for %us!\n", 550182607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 550282607adcSTejun Heo } 550382607adcSTejun Heo } 550482607adcSTejun Heo 550582607adcSTejun Heo rcu_read_unlock(); 550682607adcSTejun Heo 550782607adcSTejun Heo if (lockup_detected) 550882607adcSTejun Heo show_workqueue_state(); 550982607adcSTejun Heo 551082607adcSTejun Heo wq_watchdog_reset_touched(); 551182607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 551282607adcSTejun Heo } 551382607adcSTejun Heo 551482607adcSTejun Heo void wq_watchdog_touch(int cpu) 551582607adcSTejun Heo { 551682607adcSTejun Heo if (cpu >= 0) 551782607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 551882607adcSTejun Heo else 551982607adcSTejun Heo wq_watchdog_touched = jiffies; 552082607adcSTejun Heo } 552182607adcSTejun Heo 552282607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 552382607adcSTejun Heo { 552482607adcSTejun Heo wq_watchdog_thresh = 0; 552582607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 552682607adcSTejun Heo 552782607adcSTejun Heo if (thresh) { 552882607adcSTejun Heo wq_watchdog_thresh = thresh; 552982607adcSTejun Heo wq_watchdog_reset_touched(); 553082607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 553182607adcSTejun Heo } 553282607adcSTejun Heo } 553382607adcSTejun Heo 553482607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 553582607adcSTejun Heo const struct kernel_param *kp) 553682607adcSTejun Heo { 553782607adcSTejun Heo unsigned long thresh; 553882607adcSTejun Heo int ret; 553982607adcSTejun Heo 554082607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 554182607adcSTejun Heo if (ret) 554282607adcSTejun Heo return ret; 554382607adcSTejun Heo 554482607adcSTejun Heo if (system_wq) 554582607adcSTejun Heo wq_watchdog_set_thresh(thresh); 554682607adcSTejun Heo else 554782607adcSTejun Heo wq_watchdog_thresh = thresh; 554882607adcSTejun Heo 554982607adcSTejun Heo return 0; 555082607adcSTejun Heo } 555182607adcSTejun Heo 555282607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 555382607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 555482607adcSTejun Heo .get = param_get_ulong, 555582607adcSTejun Heo }; 555682607adcSTejun Heo 555782607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 555882607adcSTejun Heo 0644); 555982607adcSTejun Heo 556082607adcSTejun Heo static void wq_watchdog_init(void) 556182607adcSTejun Heo { 55625cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 556382607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 556482607adcSTejun Heo } 556582607adcSTejun Heo 556682607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 556782607adcSTejun Heo 556882607adcSTejun Heo static inline void wq_watchdog_init(void) { } 556982607adcSTejun Heo 557082607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 557182607adcSTejun Heo 5572bce90380STejun Heo static void __init wq_numa_init(void) 5573bce90380STejun Heo { 5574bce90380STejun Heo cpumask_var_t *tbl; 5575bce90380STejun Heo int node, cpu; 5576bce90380STejun Heo 5577bce90380STejun Heo if (num_possible_nodes() <= 1) 5578bce90380STejun Heo return; 5579bce90380STejun Heo 5580d55262c4STejun Heo if (wq_disable_numa) { 5581d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5582d55262c4STejun Heo return; 5583d55262c4STejun Heo } 5584d55262c4STejun Heo 55854c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 55864c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 55874c16bd32STejun Heo 5588bce90380STejun Heo /* 5589bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5590bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5591bce90380STejun Heo * fully initialized by now. 5592bce90380STejun Heo */ 5593ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5594bce90380STejun Heo BUG_ON(!tbl); 5595bce90380STejun Heo 5596bce90380STejun Heo for_each_node(node) 55975a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 55981be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5599bce90380STejun Heo 5600bce90380STejun Heo for_each_possible_cpu(cpu) { 5601bce90380STejun Heo node = cpu_to_node(cpu); 5602bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5603bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5604bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5605bce90380STejun Heo return; 5606bce90380STejun Heo } 5607bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5608bce90380STejun Heo } 5609bce90380STejun Heo 5610bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5611bce90380STejun Heo wq_numa_enabled = true; 5612bce90380STejun Heo } 5613bce90380STejun Heo 56143347fa09STejun Heo /** 56153347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 56163347fa09STejun Heo * 56173347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 56183347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 56193347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 56203347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 56213347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 56223347fa09STejun Heo * created and scheduled right before early initcalls. 56233347fa09STejun Heo */ 56243347fa09STejun Heo int __init workqueue_init_early(void) 56251da177e4SLinus Torvalds { 56267a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 56271bda3f80SFrederic Weisbecker int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 56287a4e344cSTejun Heo int i, cpu; 5629c34056a3STejun Heo 5630e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5631e904e6c2STejun Heo 5632b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 56331bda3f80SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 5634b05a7928SFrederic Weisbecker 5635e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5636e904e6c2STejun Heo 5637706026c2STejun Heo /* initialize CPU pools */ 563829c91e99STejun Heo for_each_possible_cpu(cpu) { 56394ce62e9eSTejun Heo struct worker_pool *pool; 56408b03ae3cSTejun Heo 56417a4e344cSTejun Heo i = 0; 5642f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 56437a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5644ec22ca5eSTejun Heo pool->cpu = cpu; 56457a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 56467a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5647f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 56487a4e344cSTejun Heo 56499daf9e67STejun Heo /* alloc pool ID */ 565068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 56519daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 565268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 56534ce62e9eSTejun Heo } 56548b03ae3cSTejun Heo } 56558b03ae3cSTejun Heo 56568a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 565729c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 565829c91e99STejun Heo struct workqueue_attrs *attrs; 565929c91e99STejun Heo 566029c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 566129c91e99STejun Heo attrs->nice = std_nice[i]; 566229c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 56638a2b7538STejun Heo 56648a2b7538STejun Heo /* 56658a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 56668a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 56678a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 56688a2b7538STejun Heo */ 56698a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 56708a2b7538STejun Heo attrs->nice = std_nice[i]; 56718a2b7538STejun Heo attrs->no_numa = true; 56728a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 567329c91e99STejun Heo } 567429c91e99STejun Heo 5675d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 56761aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5677d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5678f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5679f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 568024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 568124d51addSTejun Heo WQ_FREEZABLE, 0); 56820668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 56830668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 56840668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 56850668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 56860668106cSViresh Kumar 0); 56871aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 56880668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 56890668106cSViresh Kumar !system_power_efficient_wq || 56900668106cSViresh Kumar !system_freezable_power_efficient_wq); 569182607adcSTejun Heo 56923347fa09STejun Heo return 0; 56933347fa09STejun Heo } 56943347fa09STejun Heo 56953347fa09STejun Heo /** 56963347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 56973347fa09STejun Heo * 56983347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 56993347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 57003347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 57013347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 57023347fa09STejun Heo * with the initial workers and enable future kworker creations. 57033347fa09STejun Heo */ 57043347fa09STejun Heo int __init workqueue_init(void) 57053347fa09STejun Heo { 57062186d9f9STejun Heo struct workqueue_struct *wq; 57073347fa09STejun Heo struct worker_pool *pool; 57083347fa09STejun Heo int cpu, bkt; 57093347fa09STejun Heo 57102186d9f9STejun Heo /* 57112186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 57122186d9f9STejun Heo * CPU to node mapping may not be available that early on some 57132186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 57142186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 57152186d9f9STejun Heo * affinity, fix them up. 571640c17f75STejun Heo * 571740c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 57182186d9f9STejun Heo */ 57192186d9f9STejun Heo wq_numa_init(); 57202186d9f9STejun Heo 57212186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 57222186d9f9STejun Heo 57232186d9f9STejun Heo for_each_possible_cpu(cpu) { 57242186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 57252186d9f9STejun Heo pool->node = cpu_to_node(cpu); 57262186d9f9STejun Heo } 57272186d9f9STejun Heo } 57282186d9f9STejun Heo 572940c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 57302186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 573140c17f75STejun Heo WARN(init_rescuer(wq), 573240c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 573340c17f75STejun Heo wq->name); 573440c17f75STejun Heo } 57352186d9f9STejun Heo 57362186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 57372186d9f9STejun Heo 57383347fa09STejun Heo /* create the initial workers */ 57393347fa09STejun Heo for_each_online_cpu(cpu) { 57403347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 57413347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 57423347fa09STejun Heo BUG_ON(!create_worker(pool)); 57433347fa09STejun Heo } 57443347fa09STejun Heo } 57453347fa09STejun Heo 57463347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 57473347fa09STejun Heo BUG_ON(!create_worker(pool)); 57483347fa09STejun Heo 57493347fa09STejun Heo wq_online = true; 575082607adcSTejun Heo wq_watchdog_init(); 575182607adcSTejun Heo 57526ee0578bSSuresh Siddha return 0; 57531da177e4SLinus Torvalds } 5754