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); 1744a2d812a2STejun Heo worker->pool = pool; 17454736cbf7SLai Jiangshan 17461258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 17474736cbf7SLai Jiangshan } 17484736cbf7SLai Jiangshan 17494736cbf7SLai Jiangshan /** 175060f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 175160f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 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 */ 1757a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 175860f5a4bcSLai Jiangshan { 1759a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 176060f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 176160f5a4bcSLai Jiangshan 17621258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1763a2d812a2STejun Heo 1764da028469SLai Jiangshan list_del(&worker->node); 1765a2d812a2STejun Heo worker->pool = NULL; 1766a2d812a2STejun Heo 1767da028469SLai Jiangshan if (list_empty(&pool->workers)) 176860f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 17691258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 177060f5a4bcSLai Jiangshan 1771b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1772b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1773b62c0751SLai Jiangshan 177460f5a4bcSLai Jiangshan if (detach_completion) 177560f5a4bcSLai Jiangshan complete(detach_completion); 177660f5a4bcSLai Jiangshan } 177760f5a4bcSLai Jiangshan 177860f5a4bcSLai Jiangshan /** 1779c34056a3STejun Heo * create_worker - create a new workqueue worker 178063d95a91STejun Heo * @pool: pool the new worker will belong to 1781c34056a3STejun Heo * 1782051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1783c34056a3STejun Heo * 1784c34056a3STejun Heo * CONTEXT: 1785c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1786c34056a3STejun Heo * 1787d185af30SYacine Belkadi * Return: 1788c34056a3STejun Heo * Pointer to the newly created worker. 1789c34056a3STejun Heo */ 1790bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1791c34056a3STejun Heo { 1792c34056a3STejun Heo struct worker *worker = NULL; 1793f3421797STejun Heo int id = -1; 1794e3c916a4STejun Heo char id_buf[16]; 1795c34056a3STejun Heo 17967cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17977cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1798822d8405STejun Heo if (id < 0) 1799c34056a3STejun Heo goto fail; 1800c34056a3STejun Heo 1801f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1802c34056a3STejun Heo if (!worker) 1803c34056a3STejun Heo goto fail; 1804c34056a3STejun Heo 1805c34056a3STejun Heo worker->id = id; 1806c34056a3STejun Heo 180729c91e99STejun Heo if (pool->cpu >= 0) 1808e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1809e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1810f3421797STejun Heo else 1811e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1812e3c916a4STejun Heo 1813f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1814e3c916a4STejun Heo "kworker/%s", id_buf); 1815c34056a3STejun Heo if (IS_ERR(worker->task)) 1816c34056a3STejun Heo goto fail; 1817c34056a3STejun Heo 181891151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 181925834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 182091151228SOleg Nesterov 1821da028469SLai Jiangshan /* successful, attach the worker to the pool */ 18224736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1823822d8405STejun Heo 1824051e1850SLai Jiangshan /* start the newly created worker */ 1825051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1826051e1850SLai Jiangshan worker->pool->nr_workers++; 1827051e1850SLai Jiangshan worker_enter_idle(worker); 1828051e1850SLai Jiangshan wake_up_process(worker->task); 1829051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1830051e1850SLai Jiangshan 1831c34056a3STejun Heo return worker; 1832822d8405STejun Heo 1833c34056a3STejun Heo fail: 18349625ab17SLai Jiangshan if (id >= 0) 18357cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1836c34056a3STejun Heo kfree(worker); 1837c34056a3STejun Heo return NULL; 1838c34056a3STejun Heo } 1839c34056a3STejun Heo 1840c34056a3STejun Heo /** 1841c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1842c34056a3STejun Heo * @worker: worker to be destroyed 1843c34056a3STejun Heo * 184473eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 184573eb7fe7SLai Jiangshan * be idle. 1846c8e55f36STejun Heo * 1847c8e55f36STejun Heo * CONTEXT: 184860f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1849c34056a3STejun Heo */ 1850c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1851c34056a3STejun Heo { 1852bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1853c34056a3STejun Heo 1854cd549687STejun Heo lockdep_assert_held(&pool->lock); 1855cd549687STejun Heo 1856c34056a3STejun Heo /* sanity check frenzy */ 18576183c009STejun Heo if (WARN_ON(worker->current_work) || 185873eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 185973eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 18606183c009STejun Heo return; 1861c34056a3STejun Heo 1862bd7bdd43STejun Heo pool->nr_workers--; 1863bd7bdd43STejun Heo pool->nr_idle--; 1864c8e55f36STejun Heo 1865c8e55f36STejun Heo list_del_init(&worker->entry); 1866cb444766STejun Heo worker->flags |= WORKER_DIE; 186760f5a4bcSLai Jiangshan wake_up_process(worker->task); 1868c34056a3STejun Heo } 1869c34056a3STejun Heo 187032a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 1871e22bee78STejun Heo { 187232a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 1873e22bee78STejun Heo 1874d565ed63STejun Heo spin_lock_irq(&pool->lock); 1875e22bee78STejun Heo 18763347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1877e22bee78STejun Heo struct worker *worker; 1878e22bee78STejun Heo unsigned long expires; 1879e22bee78STejun Heo 1880e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 188163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1882e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1883e22bee78STejun Heo 18843347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 188563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18863347fc9fSLai Jiangshan break; 1887e22bee78STejun Heo } 18883347fc9fSLai Jiangshan 18893347fc9fSLai Jiangshan destroy_worker(worker); 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 1892d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1893e22bee78STejun Heo } 1894e22bee78STejun Heo 1895493a1724STejun Heo static void send_mayday(struct work_struct *work) 1896e22bee78STejun Heo { 1897112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1898112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1899493a1724STejun Heo 19002e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1901e22bee78STejun Heo 1902493008a8STejun Heo if (!wq->rescuer) 1903493a1724STejun Heo return; 1904e22bee78STejun Heo 1905e22bee78STejun Heo /* mayday mayday mayday */ 1906493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 190777668c8bSLai Jiangshan /* 190877668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 190977668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 191077668c8bSLai Jiangshan * rescuer is done with it. 191177668c8bSLai Jiangshan */ 191277668c8bSLai Jiangshan get_pwq(pwq); 1913493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1914e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1915493a1724STejun Heo } 1916e22bee78STejun Heo } 1917e22bee78STejun Heo 191832a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 1919e22bee78STejun Heo { 192032a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 1921e22bee78STejun Heo struct work_struct *work; 1922e22bee78STejun Heo 1923b2d82909STejun Heo spin_lock_irq(&pool->lock); 1924b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1925e22bee78STejun Heo 192663d95a91STejun Heo if (need_to_create_worker(pool)) { 1927e22bee78STejun Heo /* 1928e22bee78STejun Heo * We've been trying to create a new worker but 1929e22bee78STejun Heo * haven't been successful. We might be hitting an 1930e22bee78STejun Heo * allocation deadlock. Send distress signals to 1931e22bee78STejun Heo * rescuers. 1932e22bee78STejun Heo */ 193363d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1934e22bee78STejun Heo send_mayday(work); 1935e22bee78STejun Heo } 1936e22bee78STejun Heo 1937b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1938b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1939e22bee78STejun Heo 194063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1941e22bee78STejun Heo } 1942e22bee78STejun Heo 1943e22bee78STejun Heo /** 1944e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 194563d95a91STejun Heo * @pool: pool to create a new worker for 1946e22bee78STejun Heo * 194763d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1948e22bee78STejun Heo * have at least one idle worker on return from this function. If 1949e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 195063d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1951e22bee78STejun Heo * possible allocation deadlock. 1952e22bee78STejun Heo * 1953c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1954c5aa87bbSTejun Heo * may_start_working() %true. 1955e22bee78STejun Heo * 1956e22bee78STejun Heo * LOCKING: 1957d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1958e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1959e22bee78STejun Heo * manager. 1960e22bee78STejun Heo */ 196129187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1962d565ed63STejun Heo __releases(&pool->lock) 1963d565ed63STejun Heo __acquires(&pool->lock) 1964e22bee78STejun Heo { 1965e22bee78STejun Heo restart: 1966d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19679f9c2364STejun Heo 1968e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 196963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1970e22bee78STejun Heo 1971e22bee78STejun Heo while (true) { 1972051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1973e22bee78STejun Heo break; 1974e22bee78STejun Heo 1975e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19769f9c2364STejun Heo 197763d95a91STejun Heo if (!need_to_create_worker(pool)) 1978e22bee78STejun Heo break; 1979e22bee78STejun Heo } 1980e22bee78STejun Heo 198163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1982d565ed63STejun Heo spin_lock_irq(&pool->lock); 1983051e1850SLai Jiangshan /* 1984051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1985051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1986051e1850SLai Jiangshan * already become busy. 1987051e1850SLai Jiangshan */ 198863d95a91STejun Heo if (need_to_create_worker(pool)) 1989e22bee78STejun Heo goto restart; 1990e22bee78STejun Heo } 1991e22bee78STejun Heo 1992e22bee78STejun Heo /** 1993e22bee78STejun Heo * manage_workers - manage worker pool 1994e22bee78STejun Heo * @worker: self 1995e22bee78STejun Heo * 1996706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1997e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1998706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1999e22bee78STejun Heo * 2000e22bee78STejun Heo * The caller can safely start processing works on false return. On 2001e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2002e22bee78STejun Heo * and may_start_working() is true. 2003e22bee78STejun Heo * 2004e22bee78STejun Heo * CONTEXT: 2005d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2006e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2007e22bee78STejun Heo * 2008d185af30SYacine Belkadi * Return: 200929187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 201029187a9eSTejun Heo * start processing works, %true if management function was performed and 201129187a9eSTejun Heo * the conditions that the caller verified before calling the function may 201229187a9eSTejun Heo * no longer be true. 2013e22bee78STejun Heo */ 2014e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2015e22bee78STejun Heo { 201663d95a91STejun Heo struct worker_pool *pool = worker->pool; 2017e22bee78STejun Heo 2018692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 201929187a9eSTejun Heo return false; 2020692b4825STejun Heo 2021692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 20222607d7a6STejun Heo pool->manager = worker; 2023e22bee78STejun Heo 202429187a9eSTejun Heo maybe_create_worker(pool); 2025e22bee78STejun Heo 20262607d7a6STejun Heo pool->manager = NULL; 2027692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2028692b4825STejun Heo wake_up(&wq_manager_wait); 202929187a9eSTejun Heo return true; 2030e22bee78STejun Heo } 2031e22bee78STejun Heo 2032a62428c0STejun Heo /** 2033a62428c0STejun Heo * process_one_work - process single work 2034c34056a3STejun Heo * @worker: self 2035a62428c0STejun Heo * @work: work to process 2036a62428c0STejun Heo * 2037a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2038a62428c0STejun Heo * process a single work including synchronization against and 2039a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2040a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2041a62428c0STejun Heo * call this function to process a work. 2042a62428c0STejun Heo * 2043a62428c0STejun Heo * CONTEXT: 2044d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2045a62428c0STejun Heo */ 2046c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2047d565ed63STejun Heo __releases(&pool->lock) 2048d565ed63STejun Heo __acquires(&pool->lock) 20491da177e4SLinus Torvalds { 2050112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2051bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2052112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 205373f53c4aSTejun Heo int work_color; 20547e11629dSTejun Heo struct worker *collision; 20554e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20564e6045f1SJohannes Berg /* 2057a62428c0STejun Heo * It is permissible to free the struct work_struct from 2058a62428c0STejun Heo * inside the function that is called from it, this we need to 2059a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2060a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2061a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20624e6045f1SJohannes Berg */ 20634d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20644d82a1deSPeter Zijlstra 20654d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20664e6045f1SJohannes Berg #endif 2067807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 206885327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2069ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 207025511a47STejun Heo 20717e11629dSTejun Heo /* 20727e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20737e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20747e11629dSTejun Heo * already processing the work. If so, defer the work to the 20757e11629dSTejun Heo * currently executing one. 20767e11629dSTejun Heo */ 2077c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20787e11629dSTejun Heo if (unlikely(collision)) { 20797e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20807e11629dSTejun Heo return; 20817e11629dSTejun Heo } 20821da177e4SLinus Torvalds 20838930cabaSTejun Heo /* claim and dequeue */ 20841da177e4SLinus Torvalds debug_work_deactivate(work); 2085c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2086c34056a3STejun Heo worker->current_work = work; 2087a2c1c57bSTejun Heo worker->current_func = work->func; 2088112202d9STejun Heo worker->current_pwq = pwq; 208973f53c4aSTejun Heo work_color = get_work_color(work); 20907a22ad75STejun Heo 20918bf89593STejun Heo /* 20928bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 20938bf89593STejun Heo * overridden through set_worker_desc(). 20948bf89593STejun Heo */ 20958bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 20968bf89593STejun Heo 2097a62428c0STejun Heo list_del_init(&work->entry); 2098a62428c0STejun Heo 2099649027d7STejun Heo /* 2100228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2101228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2102228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2103228f1d00SLai Jiangshan * execution of the pending work items. 2104fb0e7bebSTejun Heo */ 2105fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2106228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2107fb0e7bebSTejun Heo 2108974271c4STejun Heo /* 2109a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2110a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2111a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2112a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2113228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2114974271c4STejun Heo */ 2115a489a03eSLai Jiangshan if (need_more_worker(pool)) 211663d95a91STejun Heo wake_up_worker(pool); 2117974271c4STejun Heo 21188930cabaSTejun Heo /* 21197c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2120d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 212123657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 212223657bb1STejun Heo * disabled. 21238930cabaSTejun Heo */ 21247c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21251da177e4SLinus Torvalds 2126d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2127365970a1SDavid Howells 2128a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 21293295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2130e6f3faa7SPeter Zijlstra /* 2131f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2132f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2133e6f3faa7SPeter Zijlstra * 2134e6f3faa7SPeter Zijlstra * However, that would result in: 2135e6f3faa7SPeter Zijlstra * 2136e6f3faa7SPeter Zijlstra * A(W1) 2137e6f3faa7SPeter Zijlstra * WFC(C) 2138e6f3faa7SPeter Zijlstra * A(W1) 2139e6f3faa7SPeter Zijlstra * C(C) 2140e6f3faa7SPeter Zijlstra * 2141e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2142e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2143e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2144f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2145e6f3faa7SPeter Zijlstra * these locks. 2146e6f3faa7SPeter Zijlstra * 2147e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2148e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2149e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2150e6f3faa7SPeter Zijlstra */ 2151f52be570SPeter Zijlstra lockdep_invariant_state(true); 2152e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2153a2c1c57bSTejun Heo worker->current_func(work); 2154e36c886aSArjan van de Ven /* 2155e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2156e36c886aSArjan van de Ven * point will only record its address. 2157e36c886aSArjan van de Ven */ 2158e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21593295f0efSIngo Molnar lock_map_release(&lockdep_map); 2160112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21611da177e4SLinus Torvalds 2162d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2163044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2164044c782cSValentin Ilie " last function: %pf\n", 2165a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2166a2c1c57bSTejun Heo worker->current_func); 2167d5abe669SPeter Zijlstra debug_show_held_locks(current); 2168d5abe669SPeter Zijlstra dump_stack(); 2169d5abe669SPeter Zijlstra } 2170d5abe669SPeter Zijlstra 2171b22ce278STejun Heo /* 2172b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2173b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2174b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2175b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2176789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2177789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2178b22ce278STejun Heo */ 2179a7e6425eSPaul E. McKenney cond_resched(); 2180b22ce278STejun Heo 2181d565ed63STejun Heo spin_lock_irq(&pool->lock); 2182a62428c0STejun Heo 2183fb0e7bebSTejun Heo /* clear cpu intensive status */ 2184fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2185fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2186fb0e7bebSTejun Heo 2187a62428c0STejun Heo /* we're done with it, release */ 218842f8570fSSasha Levin hash_del(&worker->hentry); 2189c34056a3STejun Heo worker->current_work = NULL; 2190a2c1c57bSTejun Heo worker->current_func = NULL; 2191112202d9STejun Heo worker->current_pwq = NULL; 2192112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21931da177e4SLinus Torvalds } 21941da177e4SLinus Torvalds 2195affee4b2STejun Heo /** 2196affee4b2STejun Heo * process_scheduled_works - process scheduled works 2197affee4b2STejun Heo * @worker: self 2198affee4b2STejun Heo * 2199affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2200affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2201affee4b2STejun Heo * fetches a work from the top and executes it. 2202affee4b2STejun Heo * 2203affee4b2STejun Heo * CONTEXT: 2204d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2205affee4b2STejun Heo * multiple times. 2206affee4b2STejun Heo */ 2207affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22081da177e4SLinus Torvalds { 2209affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2210affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2211a62428c0STejun Heo struct work_struct, entry); 2212c34056a3STejun Heo process_one_work(worker, work); 2213a62428c0STejun Heo } 22141da177e4SLinus Torvalds } 22151da177e4SLinus Torvalds 2216197f6accSTejun Heo static void set_pf_worker(bool val) 2217197f6accSTejun Heo { 2218197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2219197f6accSTejun Heo if (val) 2220197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2221197f6accSTejun Heo else 2222197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2223197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2224197f6accSTejun Heo } 2225197f6accSTejun Heo 22264690c4abSTejun Heo /** 22274690c4abSTejun Heo * worker_thread - the worker thread function 2228c34056a3STejun Heo * @__worker: self 22294690c4abSTejun Heo * 2230c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2231c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2232c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2233c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2234c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2235d185af30SYacine Belkadi * 2236d185af30SYacine Belkadi * Return: 0 22374690c4abSTejun Heo */ 2238c34056a3STejun Heo static int worker_thread(void *__worker) 22391da177e4SLinus Torvalds { 2240c34056a3STejun Heo struct worker *worker = __worker; 2241bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22421da177e4SLinus Torvalds 2243e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2244197f6accSTejun Heo set_pf_worker(true); 2245c8e55f36STejun Heo woke_up: 2246d565ed63STejun Heo spin_lock_irq(&pool->lock); 2247affee4b2STejun Heo 2248a9ab775bSTejun Heo /* am I supposed to die? */ 2249a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2250d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2251a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2252197f6accSTejun Heo set_pf_worker(false); 225360f5a4bcSLai Jiangshan 225460f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 22557cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 2256a2d812a2STejun Heo worker_detach_from_pool(worker); 225760f5a4bcSLai Jiangshan kfree(worker); 2258c8e55f36STejun Heo return 0; 2259c8e55f36STejun Heo } 2260c8e55f36STejun Heo 2261c8e55f36STejun Heo worker_leave_idle(worker); 2262db7bccf4STejun Heo recheck: 2263e22bee78STejun Heo /* no more worker necessary? */ 226463d95a91STejun Heo if (!need_more_worker(pool)) 2265e22bee78STejun Heo goto sleep; 2266e22bee78STejun Heo 2267e22bee78STejun Heo /* do we need to manage? */ 226863d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2269e22bee78STejun Heo goto recheck; 2270e22bee78STejun Heo 2271c8e55f36STejun Heo /* 2272c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2273c8e55f36STejun Heo * preparing to process a work or actually processing it. 2274c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2275c8e55f36STejun Heo */ 22766183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2277c8e55f36STejun Heo 2278e22bee78STejun Heo /* 2279a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2280a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2281a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2282a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2283a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2284e22bee78STejun Heo */ 2285a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2286e22bee78STejun Heo 2287e22bee78STejun Heo do { 2288affee4b2STejun Heo struct work_struct *work = 2289bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2290affee4b2STejun Heo struct work_struct, entry); 2291affee4b2STejun Heo 229282607adcSTejun Heo pool->watchdog_ts = jiffies; 229382607adcSTejun Heo 2294c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2295affee4b2STejun Heo /* optimization path, not strictly necessary */ 2296affee4b2STejun Heo process_one_work(worker, work); 2297affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2298affee4b2STejun Heo process_scheduled_works(worker); 2299affee4b2STejun Heo } else { 2300c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2301affee4b2STejun Heo process_scheduled_works(worker); 2302affee4b2STejun Heo } 230363d95a91STejun Heo } while (keep_working(pool)); 2304affee4b2STejun Heo 2305228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2306d313dd85STejun Heo sleep: 2307c8e55f36STejun Heo /* 2308d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2309d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2310d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2311d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2312d565ed63STejun Heo * event. 2313c8e55f36STejun Heo */ 2314c8e55f36STejun Heo worker_enter_idle(worker); 2315c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2316d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23171da177e4SLinus Torvalds schedule(); 2318c8e55f36STejun Heo goto woke_up; 23191da177e4SLinus Torvalds } 23201da177e4SLinus Torvalds 2321e22bee78STejun Heo /** 2322e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2323111c225aSTejun Heo * @__rescuer: self 2324e22bee78STejun Heo * 2325e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2326493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2327e22bee78STejun Heo * 2328706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2329e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2330e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2331e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2332e22bee78STejun Heo * the problem rescuer solves. 2333e22bee78STejun Heo * 2334706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2335706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2336e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2337e22bee78STejun Heo * 2338e22bee78STejun Heo * This should happen rarely. 2339d185af30SYacine Belkadi * 2340d185af30SYacine Belkadi * Return: 0 2341e22bee78STejun Heo */ 2342111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2343e22bee78STejun Heo { 2344111c225aSTejun Heo struct worker *rescuer = __rescuer; 2345111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2346e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 23474d595b86SLai Jiangshan bool should_stop; 2348e22bee78STejun Heo 2349e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2350111c225aSTejun Heo 2351111c225aSTejun Heo /* 2352111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2353111c225aSTejun Heo * doesn't participate in concurrency management. 2354111c225aSTejun Heo */ 2355197f6accSTejun Heo set_pf_worker(true); 2356e22bee78STejun Heo repeat: 2357c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 23581da177e4SLinus Torvalds 23594d595b86SLai Jiangshan /* 23604d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 23614d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 23624d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 23634d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 23644d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 23654d595b86SLai Jiangshan * list is always empty on exit. 23664d595b86SLai Jiangshan */ 23674d595b86SLai Jiangshan should_stop = kthread_should_stop(); 23681da177e4SLinus Torvalds 2369493a1724STejun Heo /* see whether any pwq is asking for help */ 23702e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2371493a1724STejun Heo 2372493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2373493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2374493a1724STejun Heo struct pool_workqueue, mayday_node); 2375112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2376e22bee78STejun Heo struct work_struct *work, *n; 237782607adcSTejun Heo bool first = true; 2378e22bee78STejun Heo 2379e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2380493a1724STejun Heo list_del_init(&pwq->mayday_node); 2381493a1724STejun Heo 23822e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2383e22bee78STejun Heo 238451697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 238551697d39SLai Jiangshan 238651697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2387e22bee78STejun Heo 2388e22bee78STejun Heo /* 2389e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2390e22bee78STejun Heo * process'em. 2391e22bee78STejun Heo */ 23920479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 239382607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 239482607adcSTejun Heo if (get_work_pwq(work) == pwq) { 239582607adcSTejun Heo if (first) 239682607adcSTejun Heo pool->watchdog_ts = jiffies; 2397e22bee78STejun Heo move_linked_works(work, scheduled, &n); 239882607adcSTejun Heo } 239982607adcSTejun Heo first = false; 240082607adcSTejun Heo } 2401e22bee78STejun Heo 2402008847f6SNeilBrown if (!list_empty(scheduled)) { 2403e22bee78STejun Heo process_scheduled_works(rescuer); 24047576958aSTejun Heo 24057576958aSTejun Heo /* 2406008847f6SNeilBrown * The above execution of rescued work items could 2407008847f6SNeilBrown * have created more to rescue through 2408008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2409008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2410008847f6SNeilBrown * that such back-to-back work items, which may be 2411008847f6SNeilBrown * being used to relieve memory pressure, don't 2412008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2413008847f6SNeilBrown */ 2414008847f6SNeilBrown if (need_to_create_worker(pool)) { 2415008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2416008847f6SNeilBrown get_pwq(pwq); 2417008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2418008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2419008847f6SNeilBrown } 2420008847f6SNeilBrown } 2421008847f6SNeilBrown 2422008847f6SNeilBrown /* 242377668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 242413b1d625SLai Jiangshan * go away while we're still attached to it. 242577668c8bSLai Jiangshan */ 242677668c8bSLai Jiangshan put_pwq(pwq); 242777668c8bSLai Jiangshan 242877668c8bSLai Jiangshan /* 2429d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 24307576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24317576958aSTejun Heo * and stalling the execution. 24327576958aSTejun Heo */ 2433d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 243463d95a91STejun Heo wake_up_worker(pool); 24357576958aSTejun Heo 243613b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 243713b1d625SLai Jiangshan 2438a2d812a2STejun Heo worker_detach_from_pool(rescuer); 243913b1d625SLai Jiangshan 244013b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 24411da177e4SLinus Torvalds } 24421da177e4SLinus Torvalds 24432e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2444493a1724STejun Heo 24454d595b86SLai Jiangshan if (should_stop) { 24464d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2447197f6accSTejun Heo set_pf_worker(false); 24484d595b86SLai Jiangshan return 0; 24494d595b86SLai Jiangshan } 24504d595b86SLai Jiangshan 2451111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2452111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2453e22bee78STejun Heo schedule(); 2454e22bee78STejun Heo goto repeat; 24551da177e4SLinus Torvalds } 24561da177e4SLinus Torvalds 2457fca839c0STejun Heo /** 2458fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2459fca839c0STejun Heo * @target_wq: workqueue being flushed 2460fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2461fca839c0STejun Heo * 2462fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2463fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2464fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2465fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2466fca839c0STejun Heo * a deadlock. 2467fca839c0STejun Heo */ 2468fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2469fca839c0STejun Heo struct work_struct *target_work) 2470fca839c0STejun Heo { 2471fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2472fca839c0STejun Heo struct worker *worker; 2473fca839c0STejun Heo 2474fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2475fca839c0STejun Heo return; 2476fca839c0STejun Heo 2477fca839c0STejun Heo worker = current_wq_worker(); 2478fca839c0STejun Heo 2479fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2480fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2481fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 248223d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 248323d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2484fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2485fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2486fca839c0STejun Heo target_wq->name, target_func); 2487fca839c0STejun Heo } 2488fca839c0STejun Heo 2489fc2e4d70SOleg Nesterov struct wq_barrier { 2490fc2e4d70SOleg Nesterov struct work_struct work; 2491fc2e4d70SOleg Nesterov struct completion done; 24922607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2493fc2e4d70SOleg Nesterov }; 2494fc2e4d70SOleg Nesterov 2495fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2496fc2e4d70SOleg Nesterov { 2497fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2498fc2e4d70SOleg Nesterov complete(&barr->done); 2499fc2e4d70SOleg Nesterov } 2500fc2e4d70SOleg Nesterov 25014690c4abSTejun Heo /** 25024690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2503112202d9STejun Heo * @pwq: pwq to insert barrier into 25044690c4abSTejun Heo * @barr: wq_barrier to insert 2505affee4b2STejun Heo * @target: target work to attach @barr to 2506affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 25074690c4abSTejun Heo * 2508affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2509affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2510affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2511affee4b2STejun Heo * cpu. 2512affee4b2STejun Heo * 2513affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2514affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2515affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2516affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2517affee4b2STejun Heo * after a work with LINKED flag set. 2518affee4b2STejun Heo * 2519affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2520112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 25214690c4abSTejun Heo * 25224690c4abSTejun Heo * CONTEXT: 2523d565ed63STejun Heo * spin_lock_irq(pool->lock). 25244690c4abSTejun Heo */ 2525112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2526affee4b2STejun Heo struct wq_barrier *barr, 2527affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2528fc2e4d70SOleg Nesterov { 2529affee4b2STejun Heo struct list_head *head; 2530affee4b2STejun Heo unsigned int linked = 0; 2531affee4b2STejun Heo 2532dc186ad7SThomas Gleixner /* 2533d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2534dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2535dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2536dc186ad7SThomas Gleixner * might deadlock. 2537dc186ad7SThomas Gleixner */ 2538ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 253922df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 254052fa5bc5SBoqun Feng 2541fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2542fd1a5b04SByungchul Park 25432607d7a6STejun Heo barr->task = current; 254483c22520SOleg Nesterov 2545affee4b2STejun Heo /* 2546affee4b2STejun Heo * If @target is currently being executed, schedule the 2547affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2548affee4b2STejun Heo */ 2549affee4b2STejun Heo if (worker) 2550affee4b2STejun Heo head = worker->scheduled.next; 2551affee4b2STejun Heo else { 2552affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2553affee4b2STejun Heo 2554affee4b2STejun Heo head = target->entry.next; 2555affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2556affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2557affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2558affee4b2STejun Heo } 2559affee4b2STejun Heo 2560dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2561112202d9STejun Heo insert_work(pwq, &barr->work, head, 2562affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2563fc2e4d70SOleg Nesterov } 2564fc2e4d70SOleg Nesterov 256573f53c4aSTejun Heo /** 2566112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 256773f53c4aSTejun Heo * @wq: workqueue being flushed 256873f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 256973f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 257073f53c4aSTejun Heo * 2571112202d9STejun Heo * Prepare pwqs for workqueue flushing. 257273f53c4aSTejun Heo * 2573112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2574112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2575112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2576112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2577112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 257873f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 257973f53c4aSTejun Heo * 258073f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 258173f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 258273f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 258373f53c4aSTejun Heo * is returned. 258473f53c4aSTejun Heo * 2585112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 258673f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 258773f53c4aSTejun Heo * advanced to @work_color. 258873f53c4aSTejun Heo * 258973f53c4aSTejun Heo * CONTEXT: 25903c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 259173f53c4aSTejun Heo * 2592d185af30SYacine Belkadi * Return: 259373f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 259473f53c4aSTejun Heo * otherwise. 259573f53c4aSTejun Heo */ 2596112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 259773f53c4aSTejun Heo int flush_color, int work_color) 25981da177e4SLinus Torvalds { 259973f53c4aSTejun Heo bool wait = false; 260049e3cf44STejun Heo struct pool_workqueue *pwq; 26011da177e4SLinus Torvalds 260273f53c4aSTejun Heo if (flush_color >= 0) { 26036183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2604112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2605dc186ad7SThomas Gleixner } 260614441960SOleg Nesterov 260749e3cf44STejun Heo for_each_pwq(pwq, wq) { 2608112202d9STejun Heo struct worker_pool *pool = pwq->pool; 26091da177e4SLinus Torvalds 2610b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 261173f53c4aSTejun Heo 261273f53c4aSTejun Heo if (flush_color >= 0) { 26136183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 261473f53c4aSTejun Heo 2615112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2616112202d9STejun Heo pwq->flush_color = flush_color; 2617112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 261873f53c4aSTejun Heo wait = true; 26191da177e4SLinus Torvalds } 262073f53c4aSTejun Heo } 262173f53c4aSTejun Heo 262273f53c4aSTejun Heo if (work_color >= 0) { 26236183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2624112202d9STejun Heo pwq->work_color = work_color; 262573f53c4aSTejun Heo } 262673f53c4aSTejun Heo 2627b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 26281da177e4SLinus Torvalds } 26291da177e4SLinus Torvalds 2630112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 263173f53c4aSTejun Heo complete(&wq->first_flusher->done); 263273f53c4aSTejun Heo 263373f53c4aSTejun Heo return wait; 263483c22520SOleg Nesterov } 26351da177e4SLinus Torvalds 26360fcb78c2SRolf Eike Beer /** 26371da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26380fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26391da177e4SLinus Torvalds * 2640c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2641c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 26421da177e4SLinus Torvalds */ 26437ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26441da177e4SLinus Torvalds { 264573f53c4aSTejun Heo struct wq_flusher this_flusher = { 264673f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 264773f53c4aSTejun Heo .flush_color = -1, 2648fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 264973f53c4aSTejun Heo }; 265073f53c4aSTejun Heo int next_color; 2651b1f4ec17SOleg Nesterov 26523347fa09STejun Heo if (WARN_ON(!wq_online)) 26533347fa09STejun Heo return; 26543347fa09STejun Heo 265587915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 265687915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 265787915adcSJohannes Berg 26583c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 265973f53c4aSTejun Heo 266073f53c4aSTejun Heo /* 266173f53c4aSTejun Heo * Start-to-wait phase 266273f53c4aSTejun Heo */ 266373f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 266473f53c4aSTejun Heo 266573f53c4aSTejun Heo if (next_color != wq->flush_color) { 266673f53c4aSTejun Heo /* 266773f53c4aSTejun Heo * Color space is not full. The current work_color 266873f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 266973f53c4aSTejun Heo * by one. 267073f53c4aSTejun Heo */ 26716183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 267273f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 267373f53c4aSTejun Heo wq->work_color = next_color; 267473f53c4aSTejun Heo 267573f53c4aSTejun Heo if (!wq->first_flusher) { 267673f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26776183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 267873f53c4aSTejun Heo 267973f53c4aSTejun Heo wq->first_flusher = &this_flusher; 268073f53c4aSTejun Heo 2681112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 268273f53c4aSTejun Heo wq->work_color)) { 268373f53c4aSTejun Heo /* nothing to flush, done */ 268473f53c4aSTejun Heo wq->flush_color = next_color; 268573f53c4aSTejun Heo wq->first_flusher = NULL; 268673f53c4aSTejun Heo goto out_unlock; 268773f53c4aSTejun Heo } 268873f53c4aSTejun Heo } else { 268973f53c4aSTejun Heo /* wait in queue */ 26906183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 269173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2692112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 269373f53c4aSTejun Heo } 269473f53c4aSTejun Heo } else { 269573f53c4aSTejun Heo /* 269673f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 269773f53c4aSTejun Heo * The next flush completion will assign us 269873f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 269973f53c4aSTejun Heo */ 270073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 270173f53c4aSTejun Heo } 270273f53c4aSTejun Heo 2703fca839c0STejun Heo check_flush_dependency(wq, NULL); 2704fca839c0STejun Heo 27053c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 270673f53c4aSTejun Heo 270773f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 270873f53c4aSTejun Heo 270973f53c4aSTejun Heo /* 271073f53c4aSTejun Heo * Wake-up-and-cascade phase 271173f53c4aSTejun Heo * 271273f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 271373f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 271473f53c4aSTejun Heo */ 271573f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 271673f53c4aSTejun Heo return; 271773f53c4aSTejun Heo 27183c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 271973f53c4aSTejun Heo 27204ce48b37STejun Heo /* we might have raced, check again with mutex held */ 27214ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 27224ce48b37STejun Heo goto out_unlock; 27234ce48b37STejun Heo 272473f53c4aSTejun Heo wq->first_flusher = NULL; 272573f53c4aSTejun Heo 27266183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 27276183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 272873f53c4aSTejun Heo 272973f53c4aSTejun Heo while (true) { 273073f53c4aSTejun Heo struct wq_flusher *next, *tmp; 273173f53c4aSTejun Heo 273273f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 273373f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 273473f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 273573f53c4aSTejun Heo break; 273673f53c4aSTejun Heo list_del_init(&next->list); 273773f53c4aSTejun Heo complete(&next->done); 273873f53c4aSTejun Heo } 273973f53c4aSTejun Heo 27406183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 274173f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 274273f53c4aSTejun Heo 274373f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 274473f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 274573f53c4aSTejun Heo 274673f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 274773f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 274873f53c4aSTejun Heo /* 274973f53c4aSTejun Heo * Assign the same color to all overflowed 275073f53c4aSTejun Heo * flushers, advance work_color and append to 275173f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 275273f53c4aSTejun Heo * phase for these overflowed flushers. 275373f53c4aSTejun Heo */ 275473f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 275573f53c4aSTejun Heo tmp->flush_color = wq->work_color; 275673f53c4aSTejun Heo 275773f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 275873f53c4aSTejun Heo 275973f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 276073f53c4aSTejun Heo &wq->flusher_queue); 2761112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 276273f53c4aSTejun Heo } 276373f53c4aSTejun Heo 276473f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27656183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 276673f53c4aSTejun Heo break; 276773f53c4aSTejun Heo } 276873f53c4aSTejun Heo 276973f53c4aSTejun Heo /* 277073f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2771112202d9STejun Heo * the new first flusher and arm pwqs. 277273f53c4aSTejun Heo */ 27736183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27746183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 277573f53c4aSTejun Heo 277673f53c4aSTejun Heo list_del_init(&next->list); 277773f53c4aSTejun Heo wq->first_flusher = next; 277873f53c4aSTejun Heo 2779112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 278073f53c4aSTejun Heo break; 278173f53c4aSTejun Heo 278273f53c4aSTejun Heo /* 278373f53c4aSTejun Heo * Meh... this color is already done, clear first 278473f53c4aSTejun Heo * flusher and repeat cascading. 278573f53c4aSTejun Heo */ 278673f53c4aSTejun Heo wq->first_flusher = NULL; 278773f53c4aSTejun Heo } 278873f53c4aSTejun Heo 278973f53c4aSTejun Heo out_unlock: 27903c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27911da177e4SLinus Torvalds } 27921dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 27931da177e4SLinus Torvalds 27949c5a2ba7STejun Heo /** 27959c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27969c5a2ba7STejun Heo * @wq: workqueue to drain 27979c5a2ba7STejun Heo * 27989c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27999c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 28009c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2801b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 28029c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 28039c5a2ba7STejun Heo * takes too long. 28049c5a2ba7STejun Heo */ 28059c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 28069c5a2ba7STejun Heo { 28079c5a2ba7STejun Heo unsigned int flush_cnt = 0; 280849e3cf44STejun Heo struct pool_workqueue *pwq; 28099c5a2ba7STejun Heo 28109c5a2ba7STejun Heo /* 28119c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 28129c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2813618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 28149c5a2ba7STejun Heo */ 281587fc741eSLai Jiangshan mutex_lock(&wq->mutex); 28169c5a2ba7STejun Heo if (!wq->nr_drainers++) 2817618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 281887fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28199c5a2ba7STejun Heo reflush: 28209c5a2ba7STejun Heo flush_workqueue(wq); 28219c5a2ba7STejun Heo 2822b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 282376af4d93STejun Heo 282449e3cf44STejun Heo for_each_pwq(pwq, wq) { 2825fa2563e4SThomas Tuttle bool drained; 28269c5a2ba7STejun Heo 2827b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2828112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2829b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2830fa2563e4SThomas Tuttle 2831fa2563e4SThomas Tuttle if (drained) 28329c5a2ba7STejun Heo continue; 28339c5a2ba7STejun Heo 28349c5a2ba7STejun Heo if (++flush_cnt == 10 || 28359c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2836c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 28379c5a2ba7STejun Heo wq->name, flush_cnt); 283876af4d93STejun Heo 2839b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 28409c5a2ba7STejun Heo goto reflush; 28419c5a2ba7STejun Heo } 28429c5a2ba7STejun Heo 28439c5a2ba7STejun Heo if (!--wq->nr_drainers) 2844618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 284587fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28469c5a2ba7STejun Heo } 28479c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28489c5a2ba7STejun Heo 2849d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2850d6e89786SJohannes Berg bool from_cancel) 2851baf59022STejun Heo { 2852baf59022STejun Heo struct worker *worker = NULL; 2853c9e7cf27STejun Heo struct worker_pool *pool; 2854112202d9STejun Heo struct pool_workqueue *pwq; 2855baf59022STejun Heo 2856baf59022STejun Heo might_sleep(); 2857baf59022STejun Heo 2858fa1b54e6STejun Heo local_irq_disable(); 2859fa1b54e6STejun Heo pool = get_work_pool(work); 2860fa1b54e6STejun Heo if (!pool) { 2861fa1b54e6STejun Heo local_irq_enable(); 2862fa1b54e6STejun Heo return false; 2863fa1b54e6STejun Heo } 2864fa1b54e6STejun Heo 2865fa1b54e6STejun Heo spin_lock(&pool->lock); 28660b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2867112202d9STejun Heo pwq = get_work_pwq(work); 2868112202d9STejun Heo if (pwq) { 2869112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2870baf59022STejun Heo goto already_gone; 2871606a5020STejun Heo } else { 2872c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2873baf59022STejun Heo if (!worker) 2874baf59022STejun Heo goto already_gone; 2875112202d9STejun Heo pwq = worker->current_pwq; 2876606a5020STejun Heo } 2877baf59022STejun Heo 2878fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2879fca839c0STejun Heo 2880112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2881d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2882baf59022STejun Heo 2883e159489bSTejun Heo /* 2884a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 2885a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 2886a1d14934SPeter Zijlstra * 2887a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 2888a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 2889a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 2890a1d14934SPeter Zijlstra * forward progress. 2891e159489bSTejun Heo */ 2892d6e89786SJohannes Berg if (!from_cancel && 2893d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 2894112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2895112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2896a1d14934SPeter Zijlstra } 2897e159489bSTejun Heo 2898baf59022STejun Heo return true; 2899baf59022STejun Heo already_gone: 2900d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2901baf59022STejun Heo return false; 2902baf59022STejun Heo } 2903baf59022STejun Heo 2904d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 2905d6e89786SJohannes Berg { 2906d6e89786SJohannes Berg struct wq_barrier barr; 2907d6e89786SJohannes Berg 2908d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 2909d6e89786SJohannes Berg return false; 2910d6e89786SJohannes Berg 291187915adcSJohannes Berg if (!from_cancel) { 291287915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 291387915adcSJohannes Berg lock_map_release(&work->lockdep_map); 291487915adcSJohannes Berg } 291587915adcSJohannes Berg 2916d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 2917d6e89786SJohannes Berg wait_for_completion(&barr.done); 2918d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 2919d6e89786SJohannes Berg return true; 2920d6e89786SJohannes Berg } else { 2921d6e89786SJohannes Berg return false; 2922d6e89786SJohannes Berg } 2923d6e89786SJohannes Berg } 2924d6e89786SJohannes Berg 2925db700897SOleg Nesterov /** 2926401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2927401a8d04STejun Heo * @work: the work to flush 2928db700897SOleg Nesterov * 2929606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2930606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2931401a8d04STejun Heo * 2932d185af30SYacine Belkadi * Return: 2933401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2934401a8d04STejun Heo * %false if it was already idle. 2935db700897SOleg Nesterov */ 2936401a8d04STejun Heo bool flush_work(struct work_struct *work) 2937db700897SOleg Nesterov { 2938d6e89786SJohannes Berg return __flush_work(work, false); 2939606a5020STejun Heo } 2940db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2941db700897SOleg Nesterov 29428603e1b3STejun Heo struct cwt_wait { 2943ac6424b9SIngo Molnar wait_queue_entry_t wait; 29448603e1b3STejun Heo struct work_struct *work; 29458603e1b3STejun Heo }; 29468603e1b3STejun Heo 2947ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 29488603e1b3STejun Heo { 29498603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 29508603e1b3STejun Heo 29518603e1b3STejun Heo if (cwait->work != key) 29528603e1b3STejun Heo return 0; 29538603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 29548603e1b3STejun Heo } 29558603e1b3STejun Heo 295636e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2957401a8d04STejun Heo { 29588603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2959bbb68dfaSTejun Heo unsigned long flags; 29601f1f642eSOleg Nesterov int ret; 29611f1f642eSOleg Nesterov 29621f1f642eSOleg Nesterov do { 2963bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2964bbb68dfaSTejun Heo /* 29658603e1b3STejun Heo * If someone else is already canceling, wait for it to 29668603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 29678603e1b3STejun Heo * because we may get scheduled between @work's completion 29688603e1b3STejun Heo * and the other canceling task resuming and clearing 29698603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 29708603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 29718603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 29728603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 29738603e1b3STejun Heo * we're hogging the CPU. 29748603e1b3STejun Heo * 29758603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 29768603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 29778603e1b3STejun Heo * wake function which matches @work along with exclusive 29788603e1b3STejun Heo * wait and wakeup. 2979bbb68dfaSTejun Heo */ 29808603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 29818603e1b3STejun Heo struct cwt_wait cwait; 29828603e1b3STejun Heo 29838603e1b3STejun Heo init_wait(&cwait.wait); 29848603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 29858603e1b3STejun Heo cwait.work = work; 29868603e1b3STejun Heo 29878603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 29888603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 29898603e1b3STejun Heo if (work_is_canceling(work)) 29908603e1b3STejun Heo schedule(); 29918603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 29928603e1b3STejun Heo } 29931f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29941f1f642eSOleg Nesterov 2995bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2996bbb68dfaSTejun Heo mark_work_canceling(work); 2997bbb68dfaSTejun Heo local_irq_restore(flags); 2998bbb68dfaSTejun Heo 29993347fa09STejun Heo /* 30003347fa09STejun Heo * This allows canceling during early boot. We know that @work 30013347fa09STejun Heo * isn't executing. 30023347fa09STejun Heo */ 30033347fa09STejun Heo if (wq_online) 3004d6e89786SJohannes Berg __flush_work(work, true); 30053347fa09STejun Heo 30067a22ad75STejun Heo clear_work_data(work); 30078603e1b3STejun Heo 30088603e1b3STejun Heo /* 30098603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 30108603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 30118603e1b3STejun Heo * visible there. 30128603e1b3STejun Heo */ 30138603e1b3STejun Heo smp_mb(); 30148603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 30158603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 30168603e1b3STejun Heo 30171f1f642eSOleg Nesterov return ret; 30181f1f642eSOleg Nesterov } 30191f1f642eSOleg Nesterov 30206e84d644SOleg Nesterov /** 3021401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3022401a8d04STejun Heo * @work: the work to cancel 30236e84d644SOleg Nesterov * 3024401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3025401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3026401a8d04STejun Heo * another workqueue. On return from this function, @work is 3027401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 30281f1f642eSOleg Nesterov * 3029401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3030401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 30316e84d644SOleg Nesterov * 3032401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 30336e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3034401a8d04STejun Heo * 3035d185af30SYacine Belkadi * Return: 3036401a8d04STejun Heo * %true if @work was pending, %false otherwise. 30376e84d644SOleg Nesterov */ 3038401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 30396e84d644SOleg Nesterov { 304036e227d2STejun Heo return __cancel_work_timer(work, false); 3041b89deed3SOleg Nesterov } 304228e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3043b89deed3SOleg Nesterov 30446e84d644SOleg Nesterov /** 3045401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3046401a8d04STejun Heo * @dwork: the delayed work to flush 30476e84d644SOleg Nesterov * 3048401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3049401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3050401a8d04STejun Heo * considers the last queueing instance of @dwork. 30511f1f642eSOleg Nesterov * 3052d185af30SYacine Belkadi * Return: 3053401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3054401a8d04STejun Heo * %false if it was already idle. 30556e84d644SOleg Nesterov */ 3056401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3057401a8d04STejun Heo { 30588930cabaSTejun Heo local_irq_disable(); 3059401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 306060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 30618930cabaSTejun Heo local_irq_enable(); 3062401a8d04STejun Heo return flush_work(&dwork->work); 3063401a8d04STejun Heo } 3064401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3065401a8d04STejun Heo 306605f0fe6bSTejun Heo /** 306705f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 306805f0fe6bSTejun Heo * @rwork: the rcu work to flush 306905f0fe6bSTejun Heo * 307005f0fe6bSTejun Heo * Return: 307105f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 307205f0fe6bSTejun Heo * %false if it was already idle. 307305f0fe6bSTejun Heo */ 307405f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 307505f0fe6bSTejun Heo { 307605f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 307705f0fe6bSTejun Heo rcu_barrier(); 307805f0fe6bSTejun Heo flush_work(&rwork->work); 307905f0fe6bSTejun Heo return true; 308005f0fe6bSTejun Heo } else { 308105f0fe6bSTejun Heo return flush_work(&rwork->work); 308205f0fe6bSTejun Heo } 308305f0fe6bSTejun Heo } 308405f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 308505f0fe6bSTejun Heo 3086f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3087f72b8792SJens Axboe { 3088f72b8792SJens Axboe unsigned long flags; 3089f72b8792SJens Axboe int ret; 3090f72b8792SJens Axboe 3091f72b8792SJens Axboe do { 3092f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3093f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3094f72b8792SJens Axboe 3095f72b8792SJens Axboe if (unlikely(ret < 0)) 3096f72b8792SJens Axboe return false; 3097f72b8792SJens Axboe 3098f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3099f72b8792SJens Axboe local_irq_restore(flags); 3100f72b8792SJens Axboe return ret; 3101f72b8792SJens Axboe } 3102f72b8792SJens Axboe 3103401a8d04STejun Heo /** 310457b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 310557b30ae7STejun Heo * @dwork: delayed_work to cancel 310609383498STejun Heo * 3107d185af30SYacine Belkadi * Kill off a pending delayed_work. 3108d185af30SYacine Belkadi * 3109d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3110d185af30SYacine Belkadi * pending. 3111d185af30SYacine Belkadi * 3112d185af30SYacine Belkadi * Note: 3113d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3114d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3115d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 311609383498STejun Heo * 311757b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 311809383498STejun Heo */ 311957b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 312009383498STejun Heo { 3121f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 312209383498STejun Heo } 312357b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 312409383498STejun Heo 312509383498STejun Heo /** 3126401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3127401a8d04STejun Heo * @dwork: the delayed work cancel 3128401a8d04STejun Heo * 3129401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3130401a8d04STejun Heo * 3131d185af30SYacine Belkadi * Return: 3132401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3133401a8d04STejun Heo */ 3134401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 31356e84d644SOleg Nesterov { 313636e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 31376e84d644SOleg Nesterov } 3138f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 31391da177e4SLinus Torvalds 31400fcb78c2SRolf Eike Beer /** 314131ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3142b6136773SAndrew Morton * @func: the function to call 3143b6136773SAndrew Morton * 314431ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 314531ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3146b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 314731ddd871STejun Heo * 3148d185af30SYacine Belkadi * Return: 314931ddd871STejun Heo * 0 on success, -errno on failure. 3150b6136773SAndrew Morton */ 315165f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 315215316ba8SChristoph Lameter { 315315316ba8SChristoph Lameter int cpu; 315438f51568SNamhyung Kim struct work_struct __percpu *works; 315515316ba8SChristoph Lameter 3156b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3157b6136773SAndrew Morton if (!works) 315815316ba8SChristoph Lameter return -ENOMEM; 3159b6136773SAndrew Morton 316095402b38SGautham R Shenoy get_online_cpus(); 316193981800STejun Heo 316215316ba8SChristoph Lameter for_each_online_cpu(cpu) { 31639bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 31649bfb1839SIngo Molnar 31659bfb1839SIngo Molnar INIT_WORK(work, func); 31668de6d308SOleg Nesterov schedule_work_on(cpu, work); 316715316ba8SChristoph Lameter } 316893981800STejun Heo 316993981800STejun Heo for_each_online_cpu(cpu) 31708616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 317193981800STejun Heo 317295402b38SGautham R Shenoy put_online_cpus(); 3173b6136773SAndrew Morton free_percpu(works); 317415316ba8SChristoph Lameter return 0; 317515316ba8SChristoph Lameter } 317615316ba8SChristoph Lameter 3177eef6a7d5SAlan Stern /** 31781fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31791fa44ecaSJames Bottomley * @fn: the function to execute 31801fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31811fa44ecaSJames Bottomley * be available when the work executes) 31821fa44ecaSJames Bottomley * 31831fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31841fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31851fa44ecaSJames Bottomley * 3186d185af30SYacine Belkadi * Return: 0 - function was executed 31871fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31881fa44ecaSJames Bottomley */ 318965f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31901fa44ecaSJames Bottomley { 31911fa44ecaSJames Bottomley if (!in_interrupt()) { 319265f27f38SDavid Howells fn(&ew->work); 31931fa44ecaSJames Bottomley return 0; 31941fa44ecaSJames Bottomley } 31951fa44ecaSJames Bottomley 319665f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31971fa44ecaSJames Bottomley schedule_work(&ew->work); 31981fa44ecaSJames Bottomley 31991fa44ecaSJames Bottomley return 1; 32001fa44ecaSJames Bottomley } 32011fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 32021fa44ecaSJames Bottomley 32037a4e344cSTejun Heo /** 32047a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 32057a4e344cSTejun Heo * @attrs: workqueue_attrs to free 32067a4e344cSTejun Heo * 32077a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 32087a4e344cSTejun Heo */ 32097a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 32107a4e344cSTejun Heo { 32117a4e344cSTejun Heo if (attrs) { 32127a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 32137a4e344cSTejun Heo kfree(attrs); 32147a4e344cSTejun Heo } 32157a4e344cSTejun Heo } 32167a4e344cSTejun Heo 32177a4e344cSTejun Heo /** 32187a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 32197a4e344cSTejun Heo * @gfp_mask: allocation mask to use 32207a4e344cSTejun Heo * 32217a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3222d185af30SYacine Belkadi * return it. 3223d185af30SYacine Belkadi * 3224d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 32257a4e344cSTejun Heo */ 32267a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 32277a4e344cSTejun Heo { 32287a4e344cSTejun Heo struct workqueue_attrs *attrs; 32297a4e344cSTejun Heo 32307a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 32317a4e344cSTejun Heo if (!attrs) 32327a4e344cSTejun Heo goto fail; 32337a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 32347a4e344cSTejun Heo goto fail; 32357a4e344cSTejun Heo 323613e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 32377a4e344cSTejun Heo return attrs; 32387a4e344cSTejun Heo fail: 32397a4e344cSTejun Heo free_workqueue_attrs(attrs); 32407a4e344cSTejun Heo return NULL; 32417a4e344cSTejun Heo } 32427a4e344cSTejun Heo 324329c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 324429c91e99STejun Heo const struct workqueue_attrs *from) 324529c91e99STejun Heo { 324629c91e99STejun Heo to->nice = from->nice; 324729c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 32482865a8fbSShaohua Li /* 32492865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 32502865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 32512865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 32522865a8fbSShaohua Li */ 32532865a8fbSShaohua Li to->no_numa = from->no_numa; 325429c91e99STejun Heo } 325529c91e99STejun Heo 325629c91e99STejun Heo /* hash value of the content of @attr */ 325729c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 325829c91e99STejun Heo { 325929c91e99STejun Heo u32 hash = 0; 326029c91e99STejun Heo 326129c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 326213e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 326313e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 326429c91e99STejun Heo return hash; 326529c91e99STejun Heo } 326629c91e99STejun Heo 326729c91e99STejun Heo /* content equality test */ 326829c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 326929c91e99STejun Heo const struct workqueue_attrs *b) 327029c91e99STejun Heo { 327129c91e99STejun Heo if (a->nice != b->nice) 327229c91e99STejun Heo return false; 327329c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 327429c91e99STejun Heo return false; 327529c91e99STejun Heo return true; 327629c91e99STejun Heo } 327729c91e99STejun Heo 32787a4e344cSTejun Heo /** 32797a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 32807a4e344cSTejun Heo * @pool: worker_pool to initialize 32817a4e344cSTejun Heo * 3282402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3283d185af30SYacine Belkadi * 3284d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 328529c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 328629c91e99STejun Heo * on @pool safely to release it. 32877a4e344cSTejun Heo */ 32887a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 32894e1a1f9aSTejun Heo { 32904e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 329129c91e99STejun Heo pool->id = -1; 329229c91e99STejun Heo pool->cpu = -1; 3293f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 32944e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 329582607adcSTejun Heo pool->watchdog_ts = jiffies; 32964e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 32974e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 32984e1a1f9aSTejun Heo hash_init(pool->busy_hash); 32994e1a1f9aSTejun Heo 330032a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 33014e1a1f9aSTejun Heo 330232a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 33034e1a1f9aSTejun Heo 3304da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 33057a4e344cSTejun Heo 33067cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 330729c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 330829c91e99STejun Heo pool->refcnt = 1; 330929c91e99STejun Heo 331029c91e99STejun Heo /* shouldn't fail above this point */ 33117a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 33127a4e344cSTejun Heo if (!pool->attrs) 33137a4e344cSTejun Heo return -ENOMEM; 33147a4e344cSTejun Heo return 0; 33154e1a1f9aSTejun Heo } 33164e1a1f9aSTejun Heo 3317e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3318e2dca7adSTejun Heo { 3319e2dca7adSTejun Heo struct workqueue_struct *wq = 3320e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3321e2dca7adSTejun Heo 3322e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3323e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3324e2dca7adSTejun Heo else 3325e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3326e2dca7adSTejun Heo 3327e2dca7adSTejun Heo kfree(wq->rescuer); 3328e2dca7adSTejun Heo kfree(wq); 3329e2dca7adSTejun Heo } 3330e2dca7adSTejun Heo 333129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 333229c91e99STejun Heo { 333329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 333429c91e99STejun Heo 33357cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 333629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 333729c91e99STejun Heo kfree(pool); 333829c91e99STejun Heo } 333929c91e99STejun Heo 334029c91e99STejun Heo /** 334129c91e99STejun Heo * put_unbound_pool - put a worker_pool 334229c91e99STejun Heo * @pool: worker_pool to put 334329c91e99STejun Heo * 334429c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3345c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3346c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3347c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3348a892caccSTejun Heo * 3349a892caccSTejun Heo * Should be called with wq_pool_mutex held. 335029c91e99STejun Heo */ 335129c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 335229c91e99STejun Heo { 335360f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 335429c91e99STejun Heo struct worker *worker; 335529c91e99STejun Heo 3356a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3357a892caccSTejun Heo 3358a892caccSTejun Heo if (--pool->refcnt) 335929c91e99STejun Heo return; 336029c91e99STejun Heo 336129c91e99STejun Heo /* sanity checks */ 336261d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3363a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 336429c91e99STejun Heo return; 336529c91e99STejun Heo 336629c91e99STejun Heo /* release id and unhash */ 336729c91e99STejun Heo if (pool->id >= 0) 336829c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 336929c91e99STejun Heo hash_del(&pool->hash_node); 337029c91e99STejun Heo 3371c5aa87bbSTejun Heo /* 3372692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3373692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3374692b4825STejun Heo * manager and @pool gets freed with the flag set. 3375c5aa87bbSTejun Heo */ 337660f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 3377692b4825STejun Heo wait_event_lock_irq(wq_manager_wait, 3378692b4825STejun Heo !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); 3379692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3380692b4825STejun Heo 33811037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 338229c91e99STejun Heo destroy_worker(worker); 338329c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 338429c91e99STejun Heo spin_unlock_irq(&pool->lock); 338560f5a4bcSLai Jiangshan 33861258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3387da028469SLai Jiangshan if (!list_empty(&pool->workers)) 338860f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 33891258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 339060f5a4bcSLai Jiangshan 339160f5a4bcSLai Jiangshan if (pool->detach_completion) 339260f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 339360f5a4bcSLai Jiangshan 339429c91e99STejun Heo /* shut down the timers */ 339529c91e99STejun Heo del_timer_sync(&pool->idle_timer); 339629c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 339729c91e99STejun Heo 339829c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 339929c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 340029c91e99STejun Heo } 340129c91e99STejun Heo 340229c91e99STejun Heo /** 340329c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 340429c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 340529c91e99STejun Heo * 340629c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 340729c91e99STejun Heo * reference count and return it. If there already is a matching 340829c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3409d185af30SYacine Belkadi * create a new one. 3410a892caccSTejun Heo * 3411a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3412d185af30SYacine Belkadi * 3413d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3414d185af30SYacine Belkadi * On failure, %NULL. 341529c91e99STejun Heo */ 341629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 341729c91e99STejun Heo { 341829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 341929c91e99STejun Heo struct worker_pool *pool; 3420f3f90ad4STejun Heo int node; 3421e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 342229c91e99STejun Heo 3423a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 342429c91e99STejun Heo 342529c91e99STejun Heo /* do we already have a matching pool? */ 342629c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 342729c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 342829c91e99STejun Heo pool->refcnt++; 34293fb1823cSLai Jiangshan return pool; 343029c91e99STejun Heo } 343129c91e99STejun Heo } 343229c91e99STejun Heo 3433e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3434e2273584SXunlei Pang if (wq_numa_enabled) { 3435e2273584SXunlei Pang for_each_node(node) { 3436e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3437e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3438e2273584SXunlei Pang target_node = node; 3439e2273584SXunlei Pang break; 3440e2273584SXunlei Pang } 3441e2273584SXunlei Pang } 3442e2273584SXunlei Pang } 3443e2273584SXunlei Pang 344429c91e99STejun Heo /* nope, create a new one */ 3445e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 344629c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 344729c91e99STejun Heo goto fail; 344829c91e99STejun Heo 34498864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 345029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3451e2273584SXunlei Pang pool->node = target_node; 345229c91e99STejun Heo 34532865a8fbSShaohua Li /* 34542865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 34552865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 34562865a8fbSShaohua Li */ 34572865a8fbSShaohua Li pool->attrs->no_numa = false; 34582865a8fbSShaohua Li 345929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 346029c91e99STejun Heo goto fail; 346129c91e99STejun Heo 346229c91e99STejun Heo /* create and start the initial worker */ 34633347fa09STejun Heo if (wq_online && !create_worker(pool)) 346429c91e99STejun Heo goto fail; 346529c91e99STejun Heo 346629c91e99STejun Heo /* install */ 346729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 34683fb1823cSLai Jiangshan 346929c91e99STejun Heo return pool; 347029c91e99STejun Heo fail: 347129c91e99STejun Heo if (pool) 347229c91e99STejun Heo put_unbound_pool(pool); 347329c91e99STejun Heo return NULL; 347429c91e99STejun Heo } 347529c91e99STejun Heo 34768864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 34778864b4e5STejun Heo { 34788864b4e5STejun Heo kmem_cache_free(pwq_cache, 34798864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 34808864b4e5STejun Heo } 34818864b4e5STejun Heo 34828864b4e5STejun Heo /* 34838864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 34848864b4e5STejun Heo * and needs to be destroyed. 34858864b4e5STejun Heo */ 34868864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 34878864b4e5STejun Heo { 34888864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 34898864b4e5STejun Heo unbound_release_work); 34908864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 34918864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3492bc0caf09STejun Heo bool is_last; 34938864b4e5STejun Heo 34948864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 34958864b4e5STejun Heo return; 34968864b4e5STejun Heo 34973c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 34988864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3499bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35003c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35018864b4e5STejun Heo 3502a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35038864b4e5STejun Heo put_unbound_pool(pool); 3504a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3505a892caccSTejun Heo 35068864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35078864b4e5STejun Heo 35088864b4e5STejun Heo /* 35098864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3510e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 35118864b4e5STejun Heo */ 3512e2dca7adSTejun Heo if (is_last) 3513e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 35146029a918STejun Heo } 35158864b4e5STejun Heo 35160fbd95aaSTejun Heo /** 3517699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35180fbd95aaSTejun Heo * @pwq: target pool_workqueue 35190fbd95aaSTejun Heo * 3520699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3521699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3522699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35230fbd95aaSTejun Heo */ 3524699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35250fbd95aaSTejun Heo { 3526699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3527699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 35283347fa09STejun Heo unsigned long flags; 3529699ce097STejun Heo 3530699ce097STejun Heo /* for @wq->saved_max_active */ 3531a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3532699ce097STejun Heo 3533699ce097STejun Heo /* fast exit for non-freezable wqs */ 3534699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3535699ce097STejun Heo return; 3536699ce097STejun Heo 35373347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 35383347fa09STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 3539699ce097STejun Heo 354074b414eaSLai Jiangshan /* 354174b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 354274b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 354374b414eaSLai Jiangshan * is updated and visible. 354474b414eaSLai Jiangshan */ 354574b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3546699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35470fbd95aaSTejun Heo 35480fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35490fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 35500fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3551951a078aSLai Jiangshan 3552951a078aSLai Jiangshan /* 3553951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3554951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3555951a078aSLai Jiangshan */ 3556951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3557699ce097STejun Heo } else { 3558699ce097STejun Heo pwq->max_active = 0; 3559699ce097STejun Heo } 3560699ce097STejun Heo 35613347fa09STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 35620fbd95aaSTejun Heo } 35630fbd95aaSTejun Heo 3564e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3565f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3566f147f29eSTejun Heo struct worker_pool *pool) 3567d2c1d404STejun Heo { 3568d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3569d2c1d404STejun Heo 3570e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3571e50aba9aSTejun Heo 3572d2c1d404STejun Heo pwq->pool = pool; 3573d2c1d404STejun Heo pwq->wq = wq; 3574d2c1d404STejun Heo pwq->flush_color = -1; 35758864b4e5STejun Heo pwq->refcnt = 1; 3576d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 35771befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3578d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 35798864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3580f147f29eSTejun Heo } 3581d2c1d404STejun Heo 3582f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 35831befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3584f147f29eSTejun Heo { 3585f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3586f147f29eSTejun Heo 3587f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 358875ccf595STejun Heo 35891befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 35901befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 35911befcf30STejun Heo return; 35921befcf30STejun Heo 359329b1cb41SLai Jiangshan /* set the matching work_color */ 359475ccf595STejun Heo pwq->work_color = wq->work_color; 3595983ca25eSTejun Heo 3596983ca25eSTejun Heo /* sync max_active to the current setting */ 3597983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3598983ca25eSTejun Heo 3599983ca25eSTejun Heo /* link in @pwq */ 36009e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3601df2d5ae4STejun Heo } 36026029a918STejun Heo 3603f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3604f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3605f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3606f147f29eSTejun Heo { 3607f147f29eSTejun Heo struct worker_pool *pool; 3608f147f29eSTejun Heo struct pool_workqueue *pwq; 3609f147f29eSTejun Heo 3610f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3611f147f29eSTejun Heo 3612f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3613f147f29eSTejun Heo if (!pool) 3614f147f29eSTejun Heo return NULL; 3615f147f29eSTejun Heo 3616e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3617f147f29eSTejun Heo if (!pwq) { 3618f147f29eSTejun Heo put_unbound_pool(pool); 3619f147f29eSTejun Heo return NULL; 3620f147f29eSTejun Heo } 3621f147f29eSTejun Heo 3622f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3623f147f29eSTejun Heo return pwq; 3624d2c1d404STejun Heo } 3625d2c1d404STejun Heo 36264c16bd32STejun Heo /** 362730186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3628042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 36294c16bd32STejun Heo * @node: the target NUMA node 36304c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 36314c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 36324c16bd32STejun Heo * 36334c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 36344c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3635d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 36364c16bd32STejun Heo * 36374c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 36384c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 36394c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 36404c16bd32STejun Heo * @attrs->cpumask. 36414c16bd32STejun Heo * 36424c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 36434c16bd32STejun Heo * stable. 3644d185af30SYacine Belkadi * 3645d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3646d185af30SYacine Belkadi * %false if equal. 36474c16bd32STejun Heo */ 36484c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 36494c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 36504c16bd32STejun Heo { 3651d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 36524c16bd32STejun Heo goto use_dfl; 36534c16bd32STejun Heo 36544c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 36554c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 36564c16bd32STejun Heo if (cpu_going_down >= 0) 36574c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 36584c16bd32STejun Heo 36594c16bd32STejun Heo if (cpumask_empty(cpumask)) 36604c16bd32STejun Heo goto use_dfl; 36614c16bd32STejun Heo 36624c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 36634c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 36641ad0f0a7SMichael Bringmann 36651ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 36661ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 36671ad0f0a7SMichael Bringmann "possible intersect\n"); 36681ad0f0a7SMichael Bringmann return false; 36691ad0f0a7SMichael Bringmann } 36701ad0f0a7SMichael Bringmann 36714c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 36724c16bd32STejun Heo 36734c16bd32STejun Heo use_dfl: 36744c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 36754c16bd32STejun Heo return false; 36764c16bd32STejun Heo } 36774c16bd32STejun Heo 36781befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 36791befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 36801befcf30STejun Heo int node, 36811befcf30STejun Heo struct pool_workqueue *pwq) 36821befcf30STejun Heo { 36831befcf30STejun Heo struct pool_workqueue *old_pwq; 36841befcf30STejun Heo 36855b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36861befcf30STejun Heo lockdep_assert_held(&wq->mutex); 36871befcf30STejun Heo 36881befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 36891befcf30STejun Heo link_pwq(pwq); 36901befcf30STejun Heo 36911befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 36921befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 36931befcf30STejun Heo return old_pwq; 36941befcf30STejun Heo } 36951befcf30STejun Heo 36962d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 36972d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 36982d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 36992d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3700042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 37012d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 37022d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 37032d5f0764SLai Jiangshan }; 37042d5f0764SLai Jiangshan 37052d5f0764SLai Jiangshan /* free the resources after success or abort */ 37062d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 37072d5f0764SLai Jiangshan { 37082d5f0764SLai Jiangshan if (ctx) { 37092d5f0764SLai Jiangshan int node; 37102d5f0764SLai Jiangshan 37112d5f0764SLai Jiangshan for_each_node(node) 37122d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 37132d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 37142d5f0764SLai Jiangshan 37152d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 37162d5f0764SLai Jiangshan 37172d5f0764SLai Jiangshan kfree(ctx); 37182d5f0764SLai Jiangshan } 37192d5f0764SLai Jiangshan } 37202d5f0764SLai Jiangshan 37212d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 37222d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 37232d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 37242d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 37252d5f0764SLai Jiangshan { 37262d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 37272d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 37282d5f0764SLai Jiangshan int node; 37292d5f0764SLai Jiangshan 37302d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 37312d5f0764SLai Jiangshan 3732acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 37332d5f0764SLai Jiangshan 37342d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37352d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37362d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 37372d5f0764SLai Jiangshan goto out_free; 37382d5f0764SLai Jiangshan 3739042f7df1SLai Jiangshan /* 3740042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3741042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3742042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3743042f7df1SLai Jiangshan */ 37442d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3745b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3746042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3747042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 37482d5f0764SLai Jiangshan 37492d5f0764SLai Jiangshan /* 37502d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 37512d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 37522d5f0764SLai Jiangshan * pools. 37532d5f0764SLai Jiangshan */ 37542d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 37552d5f0764SLai Jiangshan 37562d5f0764SLai Jiangshan /* 37572d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 37582d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 37592d5f0764SLai Jiangshan * it even if we don't use it immediately. 37602d5f0764SLai Jiangshan */ 37612d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 37622d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 37632d5f0764SLai Jiangshan goto out_free; 37642d5f0764SLai Jiangshan 37652d5f0764SLai Jiangshan for_each_node(node) { 3766042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 37672d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 37682d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 37692d5f0764SLai Jiangshan goto out_free; 37702d5f0764SLai Jiangshan } else { 37712d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 37722d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 37732d5f0764SLai Jiangshan } 37742d5f0764SLai Jiangshan } 37752d5f0764SLai Jiangshan 3776042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3777042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3778042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 37792d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3780042f7df1SLai Jiangshan 37812d5f0764SLai Jiangshan ctx->wq = wq; 37822d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37832d5f0764SLai Jiangshan return ctx; 37842d5f0764SLai Jiangshan 37852d5f0764SLai Jiangshan out_free: 37862d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37872d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 37882d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 37892d5f0764SLai Jiangshan return NULL; 37902d5f0764SLai Jiangshan } 37912d5f0764SLai Jiangshan 37922d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 37932d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 37942d5f0764SLai Jiangshan { 37952d5f0764SLai Jiangshan int node; 37962d5f0764SLai Jiangshan 37972d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 37982d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 37992d5f0764SLai Jiangshan 38002d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 38012d5f0764SLai Jiangshan 38022d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 38032d5f0764SLai Jiangshan for_each_node(node) 38042d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 38052d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 38062d5f0764SLai Jiangshan 38072d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 38082d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 38092d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 38102d5f0764SLai Jiangshan 38112d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 38122d5f0764SLai Jiangshan } 38132d5f0764SLai Jiangshan 3814a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3815a0111cf6SLai Jiangshan { 3816a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3817a0111cf6SLai Jiangshan get_online_cpus(); 3818a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3819a0111cf6SLai Jiangshan } 3820a0111cf6SLai Jiangshan 3821a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3822a0111cf6SLai Jiangshan { 3823a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3824a0111cf6SLai Jiangshan put_online_cpus(); 3825a0111cf6SLai Jiangshan } 3826a0111cf6SLai Jiangshan 3827a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3828a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3829a0111cf6SLai Jiangshan { 3830a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3831a0111cf6SLai Jiangshan 3832a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3833a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3834a0111cf6SLai Jiangshan return -EINVAL; 3835a0111cf6SLai Jiangshan 3836a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 38370a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 38380a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 3839a0111cf6SLai Jiangshan return -EINVAL; 3840a0111cf6SLai Jiangshan 38410a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 38420a94efb5STejun Heo } 38430a94efb5STejun Heo 3844a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 38456201171eSwanghaibin if (!ctx) 38466201171eSwanghaibin return -ENOMEM; 3847a0111cf6SLai Jiangshan 3848a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3849a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3850a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3851a0111cf6SLai Jiangshan 38526201171eSwanghaibin return 0; 3853a0111cf6SLai Jiangshan } 3854a0111cf6SLai Jiangshan 38559e8cd2f5STejun Heo /** 38569e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38579e8cd2f5STejun Heo * @wq: the target workqueue 38589e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38599e8cd2f5STejun Heo * 38604c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 38614c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 38624c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38634c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38644c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38654c16bd32STejun Heo * back-to-back will stay on its current pwq. 38669e8cd2f5STejun Heo * 3867d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3868d185af30SYacine Belkadi * 3869d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 38709e8cd2f5STejun Heo */ 38719e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38729e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38739e8cd2f5STejun Heo { 3874a0111cf6SLai Jiangshan int ret; 38759e8cd2f5STejun Heo 3876a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3877a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3878a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 38792d5f0764SLai Jiangshan 38802d5f0764SLai Jiangshan return ret; 38819e8cd2f5STejun Heo } 38826106c0f8SNeilBrown EXPORT_SYMBOL_GPL(apply_workqueue_attrs); 38839e8cd2f5STejun Heo 38844c16bd32STejun Heo /** 38854c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 38864c16bd32STejun Heo * @wq: the target workqueue 38874c16bd32STejun Heo * @cpu: the CPU coming up or going down 38884c16bd32STejun Heo * @online: whether @cpu is coming up or going down 38894c16bd32STejun Heo * 38904c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 38914c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 38924c16bd32STejun Heo * @wq accordingly. 38934c16bd32STejun Heo * 38944c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 38954c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 38964c16bd32STejun Heo * correct. 38974c16bd32STejun Heo * 38984c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 38994c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 39004c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 39014c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 39024c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 39034c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 39044c16bd32STejun Heo * CPU_DOWN_PREPARE. 39054c16bd32STejun Heo */ 39064c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 39074c16bd32STejun Heo bool online) 39084c16bd32STejun Heo { 39094c16bd32STejun Heo int node = cpu_to_node(cpu); 39104c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 39114c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 39124c16bd32STejun Heo struct workqueue_attrs *target_attrs; 39134c16bd32STejun Heo cpumask_t *cpumask; 39144c16bd32STejun Heo 39154c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 39164c16bd32STejun Heo 3917f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3918f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 39194c16bd32STejun Heo return; 39204c16bd32STejun Heo 39214c16bd32STejun Heo /* 39224c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39234c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39244c16bd32STejun Heo * CPU hotplug exclusion. 39254c16bd32STejun Heo */ 39264c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39274c16bd32STejun Heo cpumask = target_attrs->cpumask; 39284c16bd32STejun Heo 39294c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39304c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39314c16bd32STejun Heo 39324c16bd32STejun Heo /* 39334c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3934042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3935042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3936042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 39374c16bd32STejun Heo */ 3938042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 39394c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3940f7142ed4SLai Jiangshan return; 39414c16bd32STejun Heo } else { 39424c16bd32STejun Heo goto use_dfl_pwq; 39434c16bd32STejun Heo } 39444c16bd32STejun Heo 39454c16bd32STejun Heo /* create a new pwq */ 39464c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 39474c16bd32STejun Heo if (!pwq) { 39482d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 39494c16bd32STejun Heo wq->name); 395077f300b1SDaeseok Youn goto use_dfl_pwq; 39514c16bd32STejun Heo } 39524c16bd32STejun Heo 3953f7142ed4SLai Jiangshan /* Install the new pwq. */ 39544c16bd32STejun Heo mutex_lock(&wq->mutex); 39554c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 39564c16bd32STejun Heo goto out_unlock; 39574c16bd32STejun Heo 39584c16bd32STejun Heo use_dfl_pwq: 3959f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 39604c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 39614c16bd32STejun Heo get_pwq(wq->dfl_pwq); 39624c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 39634c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 39644c16bd32STejun Heo out_unlock: 39654c16bd32STejun Heo mutex_unlock(&wq->mutex); 39664c16bd32STejun Heo put_pwq_unlocked(old_pwq); 39674c16bd32STejun Heo } 39684c16bd32STejun Heo 396930cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 39701da177e4SLinus Torvalds { 397149e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 39728a2b7538STejun Heo int cpu, ret; 3973e1d8aa9fSFrederic Weisbecker 397430cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3975420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3976420c0ddbSTejun Heo if (!wq->cpu_pwqs) 397730cdf249STejun Heo return -ENOMEM; 397830cdf249STejun Heo 397930cdf249STejun Heo for_each_possible_cpu(cpu) { 39807fb98ea7STejun Heo struct pool_workqueue *pwq = 39817fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 39827a62c2c8STejun Heo struct worker_pool *cpu_pools = 3983f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 398430cdf249STejun Heo 3985f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3986f147f29eSTejun Heo 3987f147f29eSTejun Heo mutex_lock(&wq->mutex); 39881befcf30STejun Heo link_pwq(pwq); 3989f147f29eSTejun Heo mutex_unlock(&wq->mutex); 399030cdf249STejun Heo } 399130cdf249STejun Heo return 0; 39928a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 39938a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 39948a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 39958a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 39968a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 39978a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 39988a2b7538STejun Heo return ret; 39999e8cd2f5STejun Heo } else { 40009e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 40019e8cd2f5STejun Heo } 40020f900049STejun Heo } 40030f900049STejun Heo 4004f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4005f3421797STejun Heo const char *name) 4006b71ab8c2STejun Heo { 4007f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4008f3421797STejun Heo 4009f3421797STejun Heo if (max_active < 1 || max_active > lim) 4010044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4011f3421797STejun Heo max_active, name, 1, lim); 4012b71ab8c2STejun Heo 4013f3421797STejun Heo return clamp_val(max_active, 1, lim); 4014b71ab8c2STejun Heo } 4015b71ab8c2STejun Heo 4016983c7515STejun Heo /* 4017983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4018983c7515STejun Heo * to guarantee forward progress. 4019983c7515STejun Heo */ 4020983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4021983c7515STejun Heo { 4022983c7515STejun Heo struct worker *rescuer; 4023983c7515STejun Heo int ret; 4024983c7515STejun Heo 4025983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4026983c7515STejun Heo return 0; 4027983c7515STejun Heo 4028983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4029983c7515STejun Heo if (!rescuer) 4030983c7515STejun Heo return -ENOMEM; 4031983c7515STejun Heo 4032983c7515STejun Heo rescuer->rescue_wq = wq; 4033983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4034983c7515STejun Heo ret = PTR_ERR_OR_ZERO(rescuer->task); 4035983c7515STejun Heo if (ret) { 4036983c7515STejun Heo kfree(rescuer); 4037983c7515STejun Heo return ret; 4038983c7515STejun Heo } 4039983c7515STejun Heo 4040983c7515STejun Heo wq->rescuer = rescuer; 4041983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4042983c7515STejun Heo wake_up_process(rescuer->task); 4043983c7515STejun Heo 4044983c7515STejun Heo return 0; 4045983c7515STejun Heo } 4046983c7515STejun Heo 4047b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 404897e37d7bSTejun Heo unsigned int flags, 40491e19ffc6STejun Heo int max_active, 4050eb13ba87SJohannes Berg struct lock_class_key *key, 4051b196be89STejun Heo const char *lock_name, ...) 40523af24433SOleg Nesterov { 4053df2d5ae4STejun Heo size_t tbl_size = 0; 4054ecf6881fSTejun Heo va_list args; 40553af24433SOleg Nesterov struct workqueue_struct *wq; 405649e3cf44STejun Heo struct pool_workqueue *pwq; 4057b196be89STejun Heo 40585c0338c6STejun Heo /* 40595c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 40605c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 40615c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 40625c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 40635c0338c6STejun Heo * on NUMA. 40645c0338c6STejun Heo */ 40655c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 40665c0338c6STejun Heo flags |= __WQ_ORDERED; 40675c0338c6STejun Heo 4068cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4069cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4070cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4071cee22a15SViresh Kumar 4072ecf6881fSTejun Heo /* allocate wq and format name */ 4073df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4074ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4075df2d5ae4STejun Heo 4076df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4077b196be89STejun Heo if (!wq) 4078d2c1d404STejun Heo return NULL; 4079b196be89STejun Heo 40806029a918STejun Heo if (flags & WQ_UNBOUND) { 40816029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40826029a918STejun Heo if (!wq->unbound_attrs) 40836029a918STejun Heo goto err_free_wq; 40846029a918STejun Heo } 40856029a918STejun Heo 4086ecf6881fSTejun Heo va_start(args, lock_name); 4087ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4088b196be89STejun Heo va_end(args); 40893af24433SOleg Nesterov 4090d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4091b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40923af24433SOleg Nesterov 4093b196be89STejun Heo /* init wq */ 409497e37d7bSTejun Heo wq->flags = flags; 4095a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 40963c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4097112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 409830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 409973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 410073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4101493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 41023af24433SOleg Nesterov 4103eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4104cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 41053af24433SOleg Nesterov 410630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4107d2c1d404STejun Heo goto err_free_wq; 41081537663fSTejun Heo 410940c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4110d2c1d404STejun Heo goto err_destroy; 4111e22bee78STejun Heo 4112226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4113226223abSTejun Heo goto err_destroy; 4114226223abSTejun Heo 41156af8bf3dSOleg Nesterov /* 411668e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 411768e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 411868e13a67SLai Jiangshan * list. 41196af8bf3dSOleg Nesterov */ 412068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4121a0a1a5fdSTejun Heo 4122a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 412349e3cf44STejun Heo for_each_pwq(pwq, wq) 4124699ce097STejun Heo pwq_adjust_max_active(pwq); 4125a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4126a0a1a5fdSTejun Heo 4127e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4128a0a1a5fdSTejun Heo 412968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41303af24433SOleg Nesterov 41313af24433SOleg Nesterov return wq; 4132d2c1d404STejun Heo 4133d2c1d404STejun Heo err_free_wq: 41346029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41354690c4abSTejun Heo kfree(wq); 4136d2c1d404STejun Heo return NULL; 4137d2c1d404STejun Heo err_destroy: 4138d2c1d404STejun Heo destroy_workqueue(wq); 41394690c4abSTejun Heo return NULL; 41401da177e4SLinus Torvalds } 4141d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41421da177e4SLinus Torvalds 41433af24433SOleg Nesterov /** 41443af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41453af24433SOleg Nesterov * @wq: target workqueue 41463af24433SOleg Nesterov * 41473af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41483af24433SOleg Nesterov */ 41493af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41503af24433SOleg Nesterov { 415149e3cf44STejun Heo struct pool_workqueue *pwq; 41524c16bd32STejun Heo int node; 41533af24433SOleg Nesterov 41549c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41559c5a2ba7STejun Heo drain_workqueue(wq); 4156c8efcc25STejun Heo 41576183c009STejun Heo /* sanity checks */ 4158b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 415949e3cf44STejun Heo for_each_pwq(pwq, wq) { 41606183c009STejun Heo int i; 41616183c009STejun Heo 416276af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 416376af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4164b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4165fa07fb6aSTejun Heo show_workqueue_state(); 41666183c009STejun Heo return; 416776af4d93STejun Heo } 416876af4d93STejun Heo } 416976af4d93STejun Heo 41705c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 41718864b4e5STejun Heo WARN_ON(pwq->nr_active) || 417276af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4173b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4174fa07fb6aSTejun Heo show_workqueue_state(); 41756183c009STejun Heo return; 41766183c009STejun Heo } 417776af4d93STejun Heo } 4178b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41796183c009STejun Heo 4180a0a1a5fdSTejun Heo /* 4181a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4182a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4183a0a1a5fdSTejun Heo */ 418468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4185e2dca7adSTejun Heo list_del_rcu(&wq->list); 418668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41873af24433SOleg Nesterov 4188226223abSTejun Heo workqueue_sysfs_unregister(wq); 4189226223abSTejun Heo 4190e2dca7adSTejun Heo if (wq->rescuer) 4191e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4192e22bee78STejun Heo 41938864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 419429c91e99STejun Heo /* 41958864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4196e2dca7adSTejun Heo * schedule RCU free. 419729c91e99STejun Heo */ 4198e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 41998864b4e5STejun Heo } else { 42008864b4e5STejun Heo /* 42018864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 42024c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 42034c16bd32STejun Heo * @wq will be freed when the last pwq is released. 42048864b4e5STejun Heo */ 42054c16bd32STejun Heo for_each_node(node) { 42064c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42074c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 42084c16bd32STejun Heo put_pwq_unlocked(pwq); 42094c16bd32STejun Heo } 42104c16bd32STejun Heo 42114c16bd32STejun Heo /* 42124c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 42134c16bd32STejun Heo * put. Don't access it afterwards. 42144c16bd32STejun Heo */ 42154c16bd32STejun Heo pwq = wq->dfl_pwq; 42164c16bd32STejun Heo wq->dfl_pwq = NULL; 4217dce90d47STejun Heo put_pwq_unlocked(pwq); 421829c91e99STejun Heo } 42193af24433SOleg Nesterov } 42203af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 42213af24433SOleg Nesterov 4222dcd989cbSTejun Heo /** 4223dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4224dcd989cbSTejun Heo * @wq: target workqueue 4225dcd989cbSTejun Heo * @max_active: new max_active value. 4226dcd989cbSTejun Heo * 4227dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4228dcd989cbSTejun Heo * 4229dcd989cbSTejun Heo * CONTEXT: 4230dcd989cbSTejun Heo * Don't call from IRQ context. 4231dcd989cbSTejun Heo */ 4232dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4233dcd989cbSTejun Heo { 423449e3cf44STejun Heo struct pool_workqueue *pwq; 4235dcd989cbSTejun Heo 42368719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42370a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 42388719dceaSTejun Heo return; 42398719dceaSTejun Heo 4240f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4241dcd989cbSTejun Heo 4242a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4243dcd989cbSTejun Heo 42440a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4245dcd989cbSTejun Heo wq->saved_max_active = max_active; 4246dcd989cbSTejun Heo 4247699ce097STejun Heo for_each_pwq(pwq, wq) 4248699ce097STejun Heo pwq_adjust_max_active(pwq); 4249dcd989cbSTejun Heo 4250a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4251dcd989cbSTejun Heo } 4252dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4253dcd989cbSTejun Heo 4254dcd989cbSTejun Heo /** 425527d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 425627d4ee03SLukas Wunner * 425727d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 425827d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 425927d4ee03SLukas Wunner * 426027d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 426127d4ee03SLukas Wunner */ 426227d4ee03SLukas Wunner struct work_struct *current_work(void) 426327d4ee03SLukas Wunner { 426427d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 426527d4ee03SLukas Wunner 426627d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 426727d4ee03SLukas Wunner } 426827d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 426927d4ee03SLukas Wunner 427027d4ee03SLukas Wunner /** 4271e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4272e6267616STejun Heo * 4273e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4274e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4275d185af30SYacine Belkadi * 4276d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4277e6267616STejun Heo */ 4278e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4279e6267616STejun Heo { 4280e6267616STejun Heo struct worker *worker = current_wq_worker(); 4281e6267616STejun Heo 42826a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4283e6267616STejun Heo } 4284e6267616STejun Heo 4285e6267616STejun Heo /** 4286dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4287dcd989cbSTejun Heo * @cpu: CPU in question 4288dcd989cbSTejun Heo * @wq: target workqueue 4289dcd989cbSTejun Heo * 4290dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4291dcd989cbSTejun Heo * no synchronization around this function and the test result is 4292dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4293dcd989cbSTejun Heo * 4294d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4295d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4296d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4297d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4298d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4299d3251859STejun Heo * 4300d185af30SYacine Belkadi * Return: 4301dcd989cbSTejun Heo * %true if congested, %false otherwise. 4302dcd989cbSTejun Heo */ 4303d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4304dcd989cbSTejun Heo { 43057fb98ea7STejun Heo struct pool_workqueue *pwq; 430676af4d93STejun Heo bool ret; 430776af4d93STejun Heo 430888109453SLai Jiangshan rcu_read_lock_sched(); 43097fb98ea7STejun Heo 4310d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4311d3251859STejun Heo cpu = smp_processor_id(); 4312d3251859STejun Heo 43137fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 43147fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 43157fb98ea7STejun Heo else 4316df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4317dcd989cbSTejun Heo 431876af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 431988109453SLai Jiangshan rcu_read_unlock_sched(); 432076af4d93STejun Heo 432176af4d93STejun Heo return ret; 4322dcd989cbSTejun Heo } 4323dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4324dcd989cbSTejun Heo 4325dcd989cbSTejun Heo /** 4326dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4327dcd989cbSTejun Heo * @work: the work to be tested 4328dcd989cbSTejun Heo * 4329dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4330dcd989cbSTejun Heo * synchronization around this function and the test result is 4331dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4332dcd989cbSTejun Heo * 4333d185af30SYacine Belkadi * Return: 4334dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4335dcd989cbSTejun Heo */ 4336dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4337dcd989cbSTejun Heo { 4338fa1b54e6STejun Heo struct worker_pool *pool; 4339dcd989cbSTejun Heo unsigned long flags; 4340dcd989cbSTejun Heo unsigned int ret = 0; 4341dcd989cbSTejun Heo 4342dcd989cbSTejun Heo if (work_pending(work)) 4343dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4344038366c5SLai Jiangshan 4345fa1b54e6STejun Heo local_irq_save(flags); 4346fa1b54e6STejun Heo pool = get_work_pool(work); 4347038366c5SLai Jiangshan if (pool) { 4348fa1b54e6STejun Heo spin_lock(&pool->lock); 4349c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4350dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4351fa1b54e6STejun Heo spin_unlock(&pool->lock); 4352038366c5SLai Jiangshan } 4353fa1b54e6STejun Heo local_irq_restore(flags); 4354dcd989cbSTejun Heo 4355dcd989cbSTejun Heo return ret; 4356dcd989cbSTejun Heo } 4357dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4358dcd989cbSTejun Heo 43593d1cb205STejun Heo /** 43603d1cb205STejun Heo * set_worker_desc - set description for the current work item 43613d1cb205STejun Heo * @fmt: printf-style format string 43623d1cb205STejun Heo * @...: arguments for the format string 43633d1cb205STejun Heo * 43643d1cb205STejun Heo * This function can be called by a running work function to describe what 43653d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43663d1cb205STejun Heo * information will be printed out together to help debugging. The 43673d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43683d1cb205STejun Heo */ 43693d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43703d1cb205STejun Heo { 43713d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43723d1cb205STejun Heo va_list args; 43733d1cb205STejun Heo 43743d1cb205STejun Heo if (worker) { 43753d1cb205STejun Heo va_start(args, fmt); 43763d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 43773d1cb205STejun Heo va_end(args); 43783d1cb205STejun Heo } 43793d1cb205STejun Heo } 43805c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 43813d1cb205STejun Heo 43823d1cb205STejun Heo /** 43833d1cb205STejun Heo * print_worker_info - print out worker information and description 43843d1cb205STejun Heo * @log_lvl: the log level to use when printing 43853d1cb205STejun Heo * @task: target task 43863d1cb205STejun Heo * 43873d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 43883d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 43893d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 43903d1cb205STejun Heo * 43913d1cb205STejun Heo * This function can be safely called on any task as long as the 43923d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 43933d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 43943d1cb205STejun Heo */ 43953d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 43963d1cb205STejun Heo { 43973d1cb205STejun Heo work_func_t *fn = NULL; 43983d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 43993d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 44003d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 44013d1cb205STejun Heo struct workqueue_struct *wq = NULL; 44023d1cb205STejun Heo struct worker *worker; 44033d1cb205STejun Heo 44043d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 44053d1cb205STejun Heo return; 44063d1cb205STejun Heo 44073d1cb205STejun Heo /* 44083d1cb205STejun Heo * This function is called without any synchronization and @task 44093d1cb205STejun Heo * could be in any state. Be careful with dereferences. 44103d1cb205STejun Heo */ 4411e700591aSPetr Mladek worker = kthread_probe_data(task); 44123d1cb205STejun Heo 44133d1cb205STejun Heo /* 44148bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 44158bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 44163d1cb205STejun Heo */ 44173d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 44183d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 44193d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 44203d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 44213d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 44223d1cb205STejun Heo 44233d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44243d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 44258bf89593STejun Heo if (strcmp(name, desc)) 44263d1cb205STejun Heo pr_cont(" (%s)", desc); 44273d1cb205STejun Heo pr_cont("\n"); 44283d1cb205STejun Heo } 44293d1cb205STejun Heo } 44303d1cb205STejun Heo 44313494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 44323494fc30STejun Heo { 44333494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 44343494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 44353494fc30STejun Heo pr_cont(" node=%d", pool->node); 44363494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 44373494fc30STejun Heo } 44383494fc30STejun Heo 44393494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 44403494fc30STejun Heo { 44413494fc30STejun Heo if (work->func == wq_barrier_func) { 44423494fc30STejun Heo struct wq_barrier *barr; 44433494fc30STejun Heo 44443494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 44453494fc30STejun Heo 44463494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 44473494fc30STejun Heo task_pid_nr(barr->task)); 44483494fc30STejun Heo } else { 44493494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 44503494fc30STejun Heo } 44513494fc30STejun Heo } 44523494fc30STejun Heo 44533494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 44543494fc30STejun Heo { 44553494fc30STejun Heo struct worker_pool *pool = pwq->pool; 44563494fc30STejun Heo struct work_struct *work; 44573494fc30STejun Heo struct worker *worker; 44583494fc30STejun Heo bool has_in_flight = false, has_pending = false; 44593494fc30STejun Heo int bkt; 44603494fc30STejun Heo 44613494fc30STejun Heo pr_info(" pwq %d:", pool->id); 44623494fc30STejun Heo pr_cont_pool_info(pool); 44633494fc30STejun Heo 44643494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 44653494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 44663494fc30STejun Heo 44673494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 44683494fc30STejun Heo if (worker->current_pwq == pwq) { 44693494fc30STejun Heo has_in_flight = true; 44703494fc30STejun Heo break; 44713494fc30STejun Heo } 44723494fc30STejun Heo } 44733494fc30STejun Heo if (has_in_flight) { 44743494fc30STejun Heo bool comma = false; 44753494fc30STejun Heo 44763494fc30STejun Heo pr_info(" in-flight:"); 44773494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 44783494fc30STejun Heo if (worker->current_pwq != pwq) 44793494fc30STejun Heo continue; 44803494fc30STejun Heo 44813494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 44823494fc30STejun Heo task_pid_nr(worker->task), 44833494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 44843494fc30STejun Heo worker->current_func); 44853494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 44863494fc30STejun Heo pr_cont_work(false, work); 44873494fc30STejun Heo comma = true; 44883494fc30STejun Heo } 44893494fc30STejun Heo pr_cont("\n"); 44903494fc30STejun Heo } 44913494fc30STejun Heo 44923494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 44933494fc30STejun Heo if (get_work_pwq(work) == pwq) { 44943494fc30STejun Heo has_pending = true; 44953494fc30STejun Heo break; 44963494fc30STejun Heo } 44973494fc30STejun Heo } 44983494fc30STejun Heo if (has_pending) { 44993494fc30STejun Heo bool comma = false; 45003494fc30STejun Heo 45013494fc30STejun Heo pr_info(" pending:"); 45023494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 45033494fc30STejun Heo if (get_work_pwq(work) != pwq) 45043494fc30STejun Heo continue; 45053494fc30STejun Heo 45063494fc30STejun Heo pr_cont_work(comma, work); 45073494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 45083494fc30STejun Heo } 45093494fc30STejun Heo pr_cont("\n"); 45103494fc30STejun Heo } 45113494fc30STejun Heo 45123494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 45133494fc30STejun Heo bool comma = false; 45143494fc30STejun Heo 45153494fc30STejun Heo pr_info(" delayed:"); 45163494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 45173494fc30STejun Heo pr_cont_work(comma, work); 45183494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 45193494fc30STejun Heo } 45203494fc30STejun Heo pr_cont("\n"); 45213494fc30STejun Heo } 45223494fc30STejun Heo } 45233494fc30STejun Heo 45243494fc30STejun Heo /** 45253494fc30STejun Heo * show_workqueue_state - dump workqueue state 45263494fc30STejun Heo * 45277b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 45287b776af6SRoger Lu * all busy workqueues and pools. 45293494fc30STejun Heo */ 45303494fc30STejun Heo void show_workqueue_state(void) 45313494fc30STejun Heo { 45323494fc30STejun Heo struct workqueue_struct *wq; 45333494fc30STejun Heo struct worker_pool *pool; 45343494fc30STejun Heo unsigned long flags; 45353494fc30STejun Heo int pi; 45363494fc30STejun Heo 45373494fc30STejun Heo rcu_read_lock_sched(); 45383494fc30STejun Heo 45393494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 45403494fc30STejun Heo 45413494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 45423494fc30STejun Heo struct pool_workqueue *pwq; 45433494fc30STejun Heo bool idle = true; 45443494fc30STejun Heo 45453494fc30STejun Heo for_each_pwq(pwq, wq) { 45463494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 45473494fc30STejun Heo idle = false; 45483494fc30STejun Heo break; 45493494fc30STejun Heo } 45503494fc30STejun Heo } 45513494fc30STejun Heo if (idle) 45523494fc30STejun Heo continue; 45533494fc30STejun Heo 45543494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 45553494fc30STejun Heo 45563494fc30STejun Heo for_each_pwq(pwq, wq) { 45573494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 45583494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 45593494fc30STejun Heo show_pwq(pwq); 45603494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 456162635ea8SSergey Senozhatsky /* 456262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 456362635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 456462635ea8SSergey Senozhatsky * hard lockup. 456562635ea8SSergey Senozhatsky */ 456662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 45673494fc30STejun Heo } 45683494fc30STejun Heo } 45693494fc30STejun Heo 45703494fc30STejun Heo for_each_pool(pool, pi) { 45713494fc30STejun Heo struct worker *worker; 45723494fc30STejun Heo bool first = true; 45733494fc30STejun Heo 45743494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 45753494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 45763494fc30STejun Heo goto next_pool; 45773494fc30STejun Heo 45783494fc30STejun Heo pr_info("pool %d:", pool->id); 45793494fc30STejun Heo pr_cont_pool_info(pool); 458082607adcSTejun Heo pr_cont(" hung=%us workers=%d", 458182607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 458282607adcSTejun Heo pool->nr_workers); 45833494fc30STejun Heo if (pool->manager) 45843494fc30STejun Heo pr_cont(" manager: %d", 45853494fc30STejun Heo task_pid_nr(pool->manager->task)); 45863494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 45873494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 45883494fc30STejun Heo task_pid_nr(worker->task)); 45893494fc30STejun Heo first = false; 45903494fc30STejun Heo } 45913494fc30STejun Heo pr_cont("\n"); 45923494fc30STejun Heo next_pool: 45933494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 459462635ea8SSergey Senozhatsky /* 459562635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 459662635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 459762635ea8SSergey Senozhatsky * hard lockup. 459862635ea8SSergey Senozhatsky */ 459962635ea8SSergey Senozhatsky touch_nmi_watchdog(); 46003494fc30STejun Heo } 46013494fc30STejun Heo 46023494fc30STejun Heo rcu_read_unlock_sched(); 46033494fc30STejun Heo } 46043494fc30STejun Heo 46056b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 46066b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 46076b59808bSTejun Heo { 46086b59808bSTejun Heo int off; 46096b59808bSTejun Heo 46106b59808bSTejun Heo /* always show the actual comm */ 46116b59808bSTejun Heo off = strscpy(buf, task->comm, size); 46126b59808bSTejun Heo if (off < 0) 46136b59808bSTejun Heo return; 46146b59808bSTejun Heo 4615197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 46166b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 46176b59808bSTejun Heo 4618197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4619197f6accSTejun Heo struct worker *worker = kthread_data(task); 4620197f6accSTejun Heo struct worker_pool *pool = worker->pool; 46216b59808bSTejun Heo 46226b59808bSTejun Heo if (pool) { 46236b59808bSTejun Heo spin_lock_irq(&pool->lock); 46246b59808bSTejun Heo /* 4625197f6accSTejun Heo * ->desc tracks information (wq name or 4626197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4627197f6accSTejun Heo * current, prepend '+', otherwise '-'. 46286b59808bSTejun Heo */ 46296b59808bSTejun Heo if (worker->desc[0] != '\0') { 46306b59808bSTejun Heo if (worker->current_work) 46316b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 46326b59808bSTejun Heo worker->desc); 46336b59808bSTejun Heo else 46346b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 46356b59808bSTejun Heo worker->desc); 46366b59808bSTejun Heo } 46376b59808bSTejun Heo spin_unlock_irq(&pool->lock); 46386b59808bSTejun Heo } 4639197f6accSTejun Heo } 46406b59808bSTejun Heo 46416b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 46426b59808bSTejun Heo } 46436b59808bSTejun Heo 464466448bc2SMathieu Malaterre #ifdef CONFIG_SMP 464566448bc2SMathieu Malaterre 4646db7bccf4STejun Heo /* 4647db7bccf4STejun Heo * CPU hotplug. 4648db7bccf4STejun Heo * 4649e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4650112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4651706026c2STejun Heo * pool which make migrating pending and scheduled works very 4652e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 465394cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4654e22bee78STejun Heo * blocked draining impractical. 4655e22bee78STejun Heo * 465624647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4657628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4658628c78e7STejun Heo * cpu comes back online. 4659db7bccf4STejun Heo */ 4660db7bccf4STejun Heo 4661e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4662db7bccf4STejun Heo { 46634ce62e9eSTejun Heo struct worker_pool *pool; 4664db7bccf4STejun Heo struct worker *worker; 4665db7bccf4STejun Heo 4666f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46671258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 466894cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4669e22bee78STejun Heo 4670f2d5a0eeSTejun Heo /* 467192f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 467294cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 467394cf58bbSTejun Heo * except for the ones which are still executing works from 467494cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 467594cf58bbSTejun Heo * this, they may become diasporas. 4676f2d5a0eeSTejun Heo */ 4677da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4678403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4679db7bccf4STejun Heo 468024647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4681f2d5a0eeSTejun Heo 468294cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 46831258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 4684e22bee78STejun Heo 4685e22bee78STejun Heo /* 4686eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4687eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4688eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4689eb283428SLai Jiangshan * from other cpus. 4690628c78e7STejun Heo */ 4691628c78e7STejun Heo schedule(); 4692628c78e7STejun Heo 4693628c78e7STejun Heo /* 4694eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4695eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4696eb283428SLai Jiangshan * and keep_working() are always true as long as the 4697eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4698eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4699eb283428SLai Jiangshan * are served by workers tied to the pool. 4700e22bee78STejun Heo */ 4701e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4702eb283428SLai Jiangshan 4703eb283428SLai Jiangshan /* 4704eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4705eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4706eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4707eb283428SLai Jiangshan */ 4708eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4709eb283428SLai Jiangshan wake_up_worker(pool); 4710eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4711eb283428SLai Jiangshan } 4712db7bccf4STejun Heo } 4713db7bccf4STejun Heo 4714bd7c089eSTejun Heo /** 4715bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4716bd7c089eSTejun Heo * @pool: pool of interest 4717bd7c089eSTejun Heo * 4718a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4719bd7c089eSTejun Heo */ 4720bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4721bd7c089eSTejun Heo { 4722a9ab775bSTejun Heo struct worker *worker; 4723bd7c089eSTejun Heo 47241258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 4725bd7c089eSTejun Heo 4726bd7c089eSTejun Heo /* 4727a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4728a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4729402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4730a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4731a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4732bd7c089eSTejun Heo */ 4733da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4734a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4735a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4736a9ab775bSTejun Heo 4737a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4738f7c17d26SWanpeng Li 47393de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4740a9ab775bSTejun Heo 4741da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4742a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4743a9ab775bSTejun Heo 4744a9ab775bSTejun Heo /* 4745a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4746a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4747a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4748a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4749a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4750a9ab775bSTejun Heo * be bound before @pool->lock is released. 4751a9ab775bSTejun Heo */ 4752a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4753bd7c089eSTejun Heo wake_up_process(worker->task); 4754bd7c089eSTejun Heo 4755bd7c089eSTejun Heo /* 4756a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4757a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4758a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4759a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4760a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4761a9ab775bSTejun Heo * concurrency management. Note that when or whether 4762a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4763a9ab775bSTejun Heo * 4764c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 4765a9ab775bSTejun Heo * tested without holding any lock in 4766a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4767a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4768a9ab775bSTejun Heo * management operations. 4769bd7c089eSTejun Heo */ 4770a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4771a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4772a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4773c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 4774bd7c089eSTejun Heo } 4775a9ab775bSTejun Heo 4776a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4777bd7c089eSTejun Heo } 4778bd7c089eSTejun Heo 47797dbc725eSTejun Heo /** 47807dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 47817dbc725eSTejun Heo * @pool: unbound pool of interest 47827dbc725eSTejun Heo * @cpu: the CPU which is coming up 47837dbc725eSTejun Heo * 47847dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 47857dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 47867dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 47877dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 47887dbc725eSTejun Heo */ 47897dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 47907dbc725eSTejun Heo { 47917dbc725eSTejun Heo static cpumask_t cpumask; 47927dbc725eSTejun Heo struct worker *worker; 47937dbc725eSTejun Heo 47941258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 47957dbc725eSTejun Heo 47967dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 47977dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 47987dbc725eSTejun Heo return; 47997dbc725eSTejun Heo 48007dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 48017dbc725eSTejun Heo 48027dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4803da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4804d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 48057dbc725eSTejun Heo } 48067dbc725eSTejun Heo 48077ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 48081da177e4SLinus Torvalds { 48094ce62e9eSTejun Heo struct worker_pool *pool; 48101da177e4SLinus Torvalds 4811f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 48123ce63377STejun Heo if (pool->nr_workers) 48133ce63377STejun Heo continue; 4814051e1850SLai Jiangshan if (!create_worker(pool)) 48157ee681b2SThomas Gleixner return -ENOMEM; 48163af24433SOleg Nesterov } 48177ee681b2SThomas Gleixner return 0; 48187ee681b2SThomas Gleixner } 48191da177e4SLinus Torvalds 48207ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 48217ee681b2SThomas Gleixner { 48227ee681b2SThomas Gleixner struct worker_pool *pool; 48237ee681b2SThomas Gleixner struct workqueue_struct *wq; 48247ee681b2SThomas Gleixner int pi; 48257ee681b2SThomas Gleixner 482668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 48277dbc725eSTejun Heo 48287dbc725eSTejun Heo for_each_pool(pool, pi) { 48291258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 483094cf58bbSTejun Heo 4831f05b558dSLai Jiangshan if (pool->cpu == cpu) 483294cf58bbSTejun Heo rebind_workers(pool); 4833f05b558dSLai Jiangshan else if (pool->cpu < 0) 48347dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 483594cf58bbSTejun Heo 48361258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 483794cf58bbSTejun Heo } 48387dbc725eSTejun Heo 48394c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 48404c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 48414c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 48424c16bd32STejun Heo 484368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 48447ee681b2SThomas Gleixner return 0; 484565758202STejun Heo } 484665758202STejun Heo 48477ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 484865758202STejun Heo { 48494c16bd32STejun Heo struct workqueue_struct *wq; 48508db25e78STejun Heo 48514c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4852e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 4853e8b3f8dbSLai Jiangshan return -1; 4854e8b3f8dbSLai Jiangshan 4855e8b3f8dbSLai Jiangshan unbind_workers(cpu); 48564c16bd32STejun Heo 48574c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 48584c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 48594c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 48604c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 48614c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 48624c16bd32STejun Heo 48637ee681b2SThomas Gleixner return 0; 486465758202STejun Heo } 486565758202STejun Heo 48662d3854a3SRusty Russell struct work_for_cpu { 4867ed48ece2STejun Heo struct work_struct work; 48682d3854a3SRusty Russell long (*fn)(void *); 48692d3854a3SRusty Russell void *arg; 48702d3854a3SRusty Russell long ret; 48712d3854a3SRusty Russell }; 48722d3854a3SRusty Russell 4873ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 48742d3854a3SRusty Russell { 4875ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4876ed48ece2STejun Heo 48772d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 48782d3854a3SRusty Russell } 48792d3854a3SRusty Russell 48802d3854a3SRusty Russell /** 488122aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 48822d3854a3SRusty Russell * @cpu: the cpu to run on 48832d3854a3SRusty Russell * @fn: the function to run 48842d3854a3SRusty Russell * @arg: the function arg 48852d3854a3SRusty Russell * 488631ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 48876b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4888d185af30SYacine Belkadi * 4889d185af30SYacine Belkadi * Return: The value @fn returns. 48902d3854a3SRusty Russell */ 4891d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 48922d3854a3SRusty Russell { 4893ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 48942d3854a3SRusty Russell 4895ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4896ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 489712997d1aSBjorn Helgaas flush_work(&wfc.work); 4898440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 48992d3854a3SRusty Russell return wfc.ret; 49002d3854a3SRusty Russell } 49012d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 49020e8d6a93SThomas Gleixner 49030e8d6a93SThomas Gleixner /** 49040e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 49050e8d6a93SThomas Gleixner * @cpu: the cpu to run on 49060e8d6a93SThomas Gleixner * @fn: the function to run 49070e8d6a93SThomas Gleixner * @arg: the function argument 49080e8d6a93SThomas Gleixner * 49090e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 49100e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 49110e8d6a93SThomas Gleixner * 49120e8d6a93SThomas Gleixner * Return: The value @fn returns. 49130e8d6a93SThomas Gleixner */ 49140e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 49150e8d6a93SThomas Gleixner { 49160e8d6a93SThomas Gleixner long ret = -ENODEV; 49170e8d6a93SThomas Gleixner 49180e8d6a93SThomas Gleixner get_online_cpus(); 49190e8d6a93SThomas Gleixner if (cpu_online(cpu)) 49200e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 49210e8d6a93SThomas Gleixner put_online_cpus(); 49220e8d6a93SThomas Gleixner return ret; 49230e8d6a93SThomas Gleixner } 49240e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 49252d3854a3SRusty Russell #endif /* CONFIG_SMP */ 49262d3854a3SRusty Russell 4927a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4928e7577c50SRusty Russell 4929a0a1a5fdSTejun Heo /** 4930a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4931a0a1a5fdSTejun Heo * 493258a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4933c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4934706026c2STejun Heo * pool->worklist. 4935a0a1a5fdSTejun Heo * 4936a0a1a5fdSTejun Heo * CONTEXT: 4937a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4938a0a1a5fdSTejun Heo */ 4939a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4940a0a1a5fdSTejun Heo { 494124b8a847STejun Heo struct workqueue_struct *wq; 494224b8a847STejun Heo struct pool_workqueue *pwq; 4943a0a1a5fdSTejun Heo 494468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4945a0a1a5fdSTejun Heo 49466183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4947a0a1a5fdSTejun Heo workqueue_freezing = true; 4948a0a1a5fdSTejun Heo 494924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4950a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4951699ce097STejun Heo for_each_pwq(pwq, wq) 4952699ce097STejun Heo pwq_adjust_max_active(pwq); 4953a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4954a1056305STejun Heo } 49555bcab335STejun Heo 495668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4957a0a1a5fdSTejun Heo } 4958a0a1a5fdSTejun Heo 4959a0a1a5fdSTejun Heo /** 496058a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4961a0a1a5fdSTejun Heo * 4962a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4963a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4964a0a1a5fdSTejun Heo * 4965a0a1a5fdSTejun Heo * CONTEXT: 496668e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4967a0a1a5fdSTejun Heo * 4968d185af30SYacine Belkadi * Return: 496958a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 497058a69cb4STejun Heo * is complete. 4971a0a1a5fdSTejun Heo */ 4972a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4973a0a1a5fdSTejun Heo { 4974a0a1a5fdSTejun Heo bool busy = false; 497524b8a847STejun Heo struct workqueue_struct *wq; 497624b8a847STejun Heo struct pool_workqueue *pwq; 4977a0a1a5fdSTejun Heo 497868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4979a0a1a5fdSTejun Heo 49806183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4981a0a1a5fdSTejun Heo 498224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 498324b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 498424b8a847STejun Heo continue; 4985a0a1a5fdSTejun Heo /* 4986a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4987a0a1a5fdSTejun Heo * to peek without lock. 4988a0a1a5fdSTejun Heo */ 498988109453SLai Jiangshan rcu_read_lock_sched(); 499024b8a847STejun Heo for_each_pwq(pwq, wq) { 49916183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4992112202d9STejun Heo if (pwq->nr_active) { 4993a0a1a5fdSTejun Heo busy = true; 499488109453SLai Jiangshan rcu_read_unlock_sched(); 4995a0a1a5fdSTejun Heo goto out_unlock; 4996a0a1a5fdSTejun Heo } 4997a0a1a5fdSTejun Heo } 499888109453SLai Jiangshan rcu_read_unlock_sched(); 4999a0a1a5fdSTejun Heo } 5000a0a1a5fdSTejun Heo out_unlock: 500168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5002a0a1a5fdSTejun Heo return busy; 5003a0a1a5fdSTejun Heo } 5004a0a1a5fdSTejun Heo 5005a0a1a5fdSTejun Heo /** 5006a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5007a0a1a5fdSTejun Heo * 5008a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5009706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5010a0a1a5fdSTejun Heo * 5011a0a1a5fdSTejun Heo * CONTEXT: 5012a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5013a0a1a5fdSTejun Heo */ 5014a0a1a5fdSTejun Heo void thaw_workqueues(void) 5015a0a1a5fdSTejun Heo { 501624b8a847STejun Heo struct workqueue_struct *wq; 501724b8a847STejun Heo struct pool_workqueue *pwq; 5018a0a1a5fdSTejun Heo 501968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5020a0a1a5fdSTejun Heo 5021a0a1a5fdSTejun Heo if (!workqueue_freezing) 5022a0a1a5fdSTejun Heo goto out_unlock; 5023a0a1a5fdSTejun Heo 502474b414eaSLai Jiangshan workqueue_freezing = false; 502524b8a847STejun Heo 502624b8a847STejun Heo /* restore max_active and repopulate worklist */ 502724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5028a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5029699ce097STejun Heo for_each_pwq(pwq, wq) 5030699ce097STejun Heo pwq_adjust_max_active(pwq); 5031a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 503224b8a847STejun Heo } 503324b8a847STejun Heo 5034a0a1a5fdSTejun Heo out_unlock: 503568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5036a0a1a5fdSTejun Heo } 5037a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5038a0a1a5fdSTejun Heo 5039042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5040042f7df1SLai Jiangshan { 5041042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5042042f7df1SLai Jiangshan int ret = 0; 5043042f7df1SLai Jiangshan struct workqueue_struct *wq; 5044042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5045042f7df1SLai Jiangshan 5046042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5047042f7df1SLai Jiangshan 5048042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5049042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5050042f7df1SLai Jiangshan continue; 5051042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5052042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5053042f7df1SLai Jiangshan continue; 5054042f7df1SLai Jiangshan 5055042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5056042f7df1SLai Jiangshan if (!ctx) { 5057042f7df1SLai Jiangshan ret = -ENOMEM; 5058042f7df1SLai Jiangshan break; 5059042f7df1SLai Jiangshan } 5060042f7df1SLai Jiangshan 5061042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5062042f7df1SLai Jiangshan } 5063042f7df1SLai Jiangshan 5064042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5065042f7df1SLai Jiangshan if (!ret) 5066042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5067042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5068042f7df1SLai Jiangshan } 5069042f7df1SLai Jiangshan 5070042f7df1SLai Jiangshan return ret; 5071042f7df1SLai Jiangshan } 5072042f7df1SLai Jiangshan 5073042f7df1SLai Jiangshan /** 5074042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5075042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5076042f7df1SLai Jiangshan * 5077042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5078042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5079042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5080042f7df1SLai Jiangshan * 5081042f7df1SLai Jiangshan * Retun: 0 - Success 5082042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5083042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5084042f7df1SLai Jiangshan */ 5085042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5086042f7df1SLai Jiangshan { 5087042f7df1SLai Jiangshan int ret = -EINVAL; 5088042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5089042f7df1SLai Jiangshan 5090042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 5091042f7df1SLai Jiangshan return -ENOMEM; 5092042f7df1SLai Jiangshan 5093c98a9805STal Shorer /* 5094c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5095c98a9805STal Shorer * If the user wishes to include them, we allow that. 5096c98a9805STal Shorer */ 5097042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5098042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5099a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5100042f7df1SLai Jiangshan 5101042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5102042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5103042f7df1SLai Jiangshan 5104042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5105042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5106042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5107042f7df1SLai Jiangshan 5108042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5109042f7df1SLai Jiangshan if (ret < 0) 5110042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5111042f7df1SLai Jiangshan 5112a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5113042f7df1SLai Jiangshan } 5114042f7df1SLai Jiangshan 5115042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 5116042f7df1SLai Jiangshan return ret; 5117042f7df1SLai Jiangshan } 5118042f7df1SLai Jiangshan 51196ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 51206ba94429SFrederic Weisbecker /* 51216ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 51226ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 51236ba94429SFrederic Weisbecker * following attributes. 51246ba94429SFrederic Weisbecker * 51256ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 51266ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 51276ba94429SFrederic Weisbecker * 51286ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 51296ba94429SFrederic Weisbecker * 51309a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 51316ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 51326ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 51339a19b463SWang Long * numa RW bool : whether enable NUMA affinity 51346ba94429SFrederic Weisbecker */ 51356ba94429SFrederic Weisbecker struct wq_device { 51366ba94429SFrederic Weisbecker struct workqueue_struct *wq; 51376ba94429SFrederic Weisbecker struct device dev; 51386ba94429SFrederic Weisbecker }; 51396ba94429SFrederic Weisbecker 51406ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 51416ba94429SFrederic Weisbecker { 51426ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 51436ba94429SFrederic Weisbecker 51446ba94429SFrederic Weisbecker return wq_dev->wq; 51456ba94429SFrederic Weisbecker } 51466ba94429SFrederic Weisbecker 51476ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 51486ba94429SFrederic Weisbecker char *buf) 51496ba94429SFrederic Weisbecker { 51506ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51516ba94429SFrederic Weisbecker 51526ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 51536ba94429SFrederic Weisbecker } 51546ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 51556ba94429SFrederic Weisbecker 51566ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 51576ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 51586ba94429SFrederic Weisbecker { 51596ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51606ba94429SFrederic Weisbecker 51616ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 51626ba94429SFrederic Weisbecker } 51636ba94429SFrederic Weisbecker 51646ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 51656ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 51666ba94429SFrederic Weisbecker size_t count) 51676ba94429SFrederic Weisbecker { 51686ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51696ba94429SFrederic Weisbecker int val; 51706ba94429SFrederic Weisbecker 51716ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 51726ba94429SFrederic Weisbecker return -EINVAL; 51736ba94429SFrederic Weisbecker 51746ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 51756ba94429SFrederic Weisbecker return count; 51766ba94429SFrederic Weisbecker } 51776ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 51786ba94429SFrederic Weisbecker 51796ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 51806ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 51816ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 51826ba94429SFrederic Weisbecker NULL, 51836ba94429SFrederic Weisbecker }; 51846ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 51856ba94429SFrederic Weisbecker 51866ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 51876ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 51886ba94429SFrederic Weisbecker { 51896ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51906ba94429SFrederic Weisbecker const char *delim = ""; 51916ba94429SFrederic Weisbecker int node, written = 0; 51926ba94429SFrederic Weisbecker 51936ba94429SFrederic Weisbecker rcu_read_lock_sched(); 51946ba94429SFrederic Weisbecker for_each_node(node) { 51956ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 51966ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 51976ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 51986ba94429SFrederic Weisbecker delim = " "; 51996ba94429SFrederic Weisbecker } 52006ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 52016ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 52026ba94429SFrederic Weisbecker 52036ba94429SFrederic Weisbecker return written; 52046ba94429SFrederic Weisbecker } 52056ba94429SFrederic Weisbecker 52066ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 52076ba94429SFrederic Weisbecker char *buf) 52086ba94429SFrederic Weisbecker { 52096ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52106ba94429SFrederic Weisbecker int written; 52116ba94429SFrederic Weisbecker 52126ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 52136ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 52146ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 52156ba94429SFrederic Weisbecker 52166ba94429SFrederic Weisbecker return written; 52176ba94429SFrederic Weisbecker } 52186ba94429SFrederic Weisbecker 52196ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 52206ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 52216ba94429SFrederic Weisbecker { 52226ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 52236ba94429SFrederic Weisbecker 5224899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5225899a94feSLai Jiangshan 52266ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 52276ba94429SFrederic Weisbecker if (!attrs) 52286ba94429SFrederic Weisbecker return NULL; 52296ba94429SFrederic Weisbecker 52306ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 52316ba94429SFrederic Weisbecker return attrs; 52326ba94429SFrederic Weisbecker } 52336ba94429SFrederic Weisbecker 52346ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 52356ba94429SFrederic Weisbecker const char *buf, size_t count) 52366ba94429SFrederic Weisbecker { 52376ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52386ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5239d4d3e257SLai Jiangshan int ret = -ENOMEM; 5240d4d3e257SLai Jiangshan 5241d4d3e257SLai Jiangshan apply_wqattrs_lock(); 52426ba94429SFrederic Weisbecker 52436ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 52446ba94429SFrederic Weisbecker if (!attrs) 5245d4d3e257SLai Jiangshan goto out_unlock; 52466ba94429SFrederic Weisbecker 52476ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 52486ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5249d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 52506ba94429SFrederic Weisbecker else 52516ba94429SFrederic Weisbecker ret = -EINVAL; 52526ba94429SFrederic Weisbecker 5253d4d3e257SLai Jiangshan out_unlock: 5254d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 52556ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 52566ba94429SFrederic Weisbecker return ret ?: count; 52576ba94429SFrederic Weisbecker } 52586ba94429SFrederic Weisbecker 52596ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 52606ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 52616ba94429SFrederic Weisbecker { 52626ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52636ba94429SFrederic Weisbecker int written; 52646ba94429SFrederic Weisbecker 52656ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 52666ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 52676ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 52686ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 52696ba94429SFrederic Weisbecker return written; 52706ba94429SFrederic Weisbecker } 52716ba94429SFrederic Weisbecker 52726ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 52736ba94429SFrederic Weisbecker struct device_attribute *attr, 52746ba94429SFrederic Weisbecker const char *buf, size_t count) 52756ba94429SFrederic Weisbecker { 52766ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52776ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5278d4d3e257SLai Jiangshan int ret = -ENOMEM; 5279d4d3e257SLai Jiangshan 5280d4d3e257SLai Jiangshan apply_wqattrs_lock(); 52816ba94429SFrederic Weisbecker 52826ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 52836ba94429SFrederic Weisbecker if (!attrs) 5284d4d3e257SLai Jiangshan goto out_unlock; 52856ba94429SFrederic Weisbecker 52866ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 52876ba94429SFrederic Weisbecker if (!ret) 5288d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 52896ba94429SFrederic Weisbecker 5290d4d3e257SLai Jiangshan out_unlock: 5291d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 52926ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 52936ba94429SFrederic Weisbecker return ret ?: count; 52946ba94429SFrederic Weisbecker } 52956ba94429SFrederic Weisbecker 52966ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 52976ba94429SFrederic Weisbecker char *buf) 52986ba94429SFrederic Weisbecker { 52996ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53006ba94429SFrederic Weisbecker int written; 53016ba94429SFrederic Weisbecker 53026ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 53036ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 53046ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 53056ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 53066ba94429SFrederic Weisbecker 53076ba94429SFrederic Weisbecker return written; 53086ba94429SFrederic Weisbecker } 53096ba94429SFrederic Weisbecker 53106ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 53116ba94429SFrederic Weisbecker const char *buf, size_t count) 53126ba94429SFrederic Weisbecker { 53136ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53146ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5315d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5316d4d3e257SLai Jiangshan 5317d4d3e257SLai Jiangshan apply_wqattrs_lock(); 53186ba94429SFrederic Weisbecker 53196ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 53206ba94429SFrederic Weisbecker if (!attrs) 5321d4d3e257SLai Jiangshan goto out_unlock; 53226ba94429SFrederic Weisbecker 53236ba94429SFrederic Weisbecker ret = -EINVAL; 53246ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 53256ba94429SFrederic Weisbecker attrs->no_numa = !v; 5326d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 53276ba94429SFrederic Weisbecker } 53286ba94429SFrederic Weisbecker 5329d4d3e257SLai Jiangshan out_unlock: 5330d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 53316ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 53326ba94429SFrederic Weisbecker return ret ?: count; 53336ba94429SFrederic Weisbecker } 53346ba94429SFrederic Weisbecker 53356ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 53366ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 53376ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 53386ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 53396ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 53406ba94429SFrederic Weisbecker __ATTR_NULL, 53416ba94429SFrederic Weisbecker }; 53426ba94429SFrederic Weisbecker 53436ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 53446ba94429SFrederic Weisbecker .name = "workqueue", 53456ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 53466ba94429SFrederic Weisbecker }; 53476ba94429SFrederic Weisbecker 5348b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5349b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5350b05a7928SFrederic Weisbecker { 5351b05a7928SFrederic Weisbecker int written; 5352b05a7928SFrederic Weisbecker 5353042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5354b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5355b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5356042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5357b05a7928SFrederic Weisbecker 5358b05a7928SFrederic Weisbecker return written; 5359b05a7928SFrederic Weisbecker } 5360b05a7928SFrederic Weisbecker 5361042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5362042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5363042f7df1SLai Jiangshan { 5364042f7df1SLai Jiangshan cpumask_var_t cpumask; 5365042f7df1SLai Jiangshan int ret; 5366042f7df1SLai Jiangshan 5367042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5368042f7df1SLai Jiangshan return -ENOMEM; 5369042f7df1SLai Jiangshan 5370042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5371042f7df1SLai Jiangshan if (!ret) 5372042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5373042f7df1SLai Jiangshan 5374042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5375042f7df1SLai Jiangshan return ret ? ret : count; 5376042f7df1SLai Jiangshan } 5377042f7df1SLai Jiangshan 5378b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5379042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5380042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5381b05a7928SFrederic Weisbecker 53826ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 53836ba94429SFrederic Weisbecker { 5384b05a7928SFrederic Weisbecker int err; 5385b05a7928SFrederic Weisbecker 5386b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5387b05a7928SFrederic Weisbecker if (err) 5388b05a7928SFrederic Weisbecker return err; 5389b05a7928SFrederic Weisbecker 5390b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 53916ba94429SFrederic Weisbecker } 53926ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 53936ba94429SFrederic Weisbecker 53946ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 53956ba94429SFrederic Weisbecker { 53966ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 53976ba94429SFrederic Weisbecker 53986ba94429SFrederic Weisbecker kfree(wq_dev); 53996ba94429SFrederic Weisbecker } 54006ba94429SFrederic Weisbecker 54016ba94429SFrederic Weisbecker /** 54026ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 54036ba94429SFrederic Weisbecker * @wq: the workqueue to register 54046ba94429SFrederic Weisbecker * 54056ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 54066ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 54076ba94429SFrederic Weisbecker * which is the preferred method. 54086ba94429SFrederic Weisbecker * 54096ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 54106ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 54116ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 54126ba94429SFrederic Weisbecker * attributes. 54136ba94429SFrederic Weisbecker * 54146ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 54156ba94429SFrederic Weisbecker */ 54166ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 54176ba94429SFrederic Weisbecker { 54186ba94429SFrederic Weisbecker struct wq_device *wq_dev; 54196ba94429SFrederic Weisbecker int ret; 54206ba94429SFrederic Weisbecker 54216ba94429SFrederic Weisbecker /* 5422402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 54236ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 54246ba94429SFrederic Weisbecker * workqueues. 54256ba94429SFrederic Weisbecker */ 54260a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 54276ba94429SFrederic Weisbecker return -EINVAL; 54286ba94429SFrederic Weisbecker 54296ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 54306ba94429SFrederic Weisbecker if (!wq_dev) 54316ba94429SFrederic Weisbecker return -ENOMEM; 54326ba94429SFrederic Weisbecker 54336ba94429SFrederic Weisbecker wq_dev->wq = wq; 54346ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 54356ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 543623217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 54376ba94429SFrederic Weisbecker 54386ba94429SFrederic Weisbecker /* 54396ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 54406ba94429SFrederic Weisbecker * everything is ready. 54416ba94429SFrederic Weisbecker */ 54426ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 54436ba94429SFrederic Weisbecker 54446ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 54456ba94429SFrederic Weisbecker if (ret) { 5446537f4146SArvind Yadav put_device(&wq_dev->dev); 54476ba94429SFrederic Weisbecker wq->wq_dev = NULL; 54486ba94429SFrederic Weisbecker return ret; 54496ba94429SFrederic Weisbecker } 54506ba94429SFrederic Weisbecker 54516ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 54526ba94429SFrederic Weisbecker struct device_attribute *attr; 54536ba94429SFrederic Weisbecker 54546ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 54556ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 54566ba94429SFrederic Weisbecker if (ret) { 54576ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 54586ba94429SFrederic Weisbecker wq->wq_dev = NULL; 54596ba94429SFrederic Weisbecker return ret; 54606ba94429SFrederic Weisbecker } 54616ba94429SFrederic Weisbecker } 54626ba94429SFrederic Weisbecker } 54636ba94429SFrederic Weisbecker 54646ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 54656ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 54666ba94429SFrederic Weisbecker return 0; 54676ba94429SFrederic Weisbecker } 54686ba94429SFrederic Weisbecker 54696ba94429SFrederic Weisbecker /** 54706ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 54716ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 54726ba94429SFrederic Weisbecker * 54736ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 54746ba94429SFrederic Weisbecker */ 54756ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 54766ba94429SFrederic Weisbecker { 54776ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 54786ba94429SFrederic Weisbecker 54796ba94429SFrederic Weisbecker if (!wq->wq_dev) 54806ba94429SFrederic Weisbecker return; 54816ba94429SFrederic Weisbecker 54826ba94429SFrederic Weisbecker wq->wq_dev = NULL; 54836ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 54846ba94429SFrederic Weisbecker } 54856ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 54866ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 54876ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 54886ba94429SFrederic Weisbecker 548982607adcSTejun Heo /* 549082607adcSTejun Heo * Workqueue watchdog. 549182607adcSTejun Heo * 549282607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 549382607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 549482607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 549582607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 549682607adcSTejun Heo * largely opaque. 549782607adcSTejun Heo * 549882607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 549982607adcSTejun Heo * state if some pools failed to make forward progress for a while where 550082607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 550182607adcSTejun Heo * 550282607adcSTejun Heo * This mechanism is controlled through the kernel parameter 550382607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 550482607adcSTejun Heo * corresponding sysfs parameter file. 550582607adcSTejun Heo */ 550682607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 550782607adcSTejun Heo 550882607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 55095cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 551082607adcSTejun Heo 551182607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 551282607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 551382607adcSTejun Heo 551482607adcSTejun Heo static void wq_watchdog_reset_touched(void) 551582607adcSTejun Heo { 551682607adcSTejun Heo int cpu; 551782607adcSTejun Heo 551882607adcSTejun Heo wq_watchdog_touched = jiffies; 551982607adcSTejun Heo for_each_possible_cpu(cpu) 552082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 552182607adcSTejun Heo } 552282607adcSTejun Heo 55235cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 552482607adcSTejun Heo { 552582607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 552682607adcSTejun Heo bool lockup_detected = false; 552782607adcSTejun Heo struct worker_pool *pool; 552882607adcSTejun Heo int pi; 552982607adcSTejun Heo 553082607adcSTejun Heo if (!thresh) 553182607adcSTejun Heo return; 553282607adcSTejun Heo 553382607adcSTejun Heo rcu_read_lock(); 553482607adcSTejun Heo 553582607adcSTejun Heo for_each_pool(pool, pi) { 553682607adcSTejun Heo unsigned long pool_ts, touched, ts; 553782607adcSTejun Heo 553882607adcSTejun Heo if (list_empty(&pool->worklist)) 553982607adcSTejun Heo continue; 554082607adcSTejun Heo 554182607adcSTejun Heo /* get the latest of pool and touched timestamps */ 554282607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 554382607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 554482607adcSTejun Heo 554582607adcSTejun Heo if (time_after(pool_ts, touched)) 554682607adcSTejun Heo ts = pool_ts; 554782607adcSTejun Heo else 554882607adcSTejun Heo ts = touched; 554982607adcSTejun Heo 555082607adcSTejun Heo if (pool->cpu >= 0) { 555182607adcSTejun Heo unsigned long cpu_touched = 555282607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 555382607adcSTejun Heo pool->cpu)); 555482607adcSTejun Heo if (time_after(cpu_touched, ts)) 555582607adcSTejun Heo ts = cpu_touched; 555682607adcSTejun Heo } 555782607adcSTejun Heo 555882607adcSTejun Heo /* did we stall? */ 555982607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 556082607adcSTejun Heo lockup_detected = true; 556182607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 556282607adcSTejun Heo pr_cont_pool_info(pool); 556382607adcSTejun Heo pr_cont(" stuck for %us!\n", 556482607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 556582607adcSTejun Heo } 556682607adcSTejun Heo } 556782607adcSTejun Heo 556882607adcSTejun Heo rcu_read_unlock(); 556982607adcSTejun Heo 557082607adcSTejun Heo if (lockup_detected) 557182607adcSTejun Heo show_workqueue_state(); 557282607adcSTejun Heo 557382607adcSTejun Heo wq_watchdog_reset_touched(); 557482607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 557582607adcSTejun Heo } 557682607adcSTejun Heo 5577cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 557882607adcSTejun Heo { 557982607adcSTejun Heo if (cpu >= 0) 558082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 558182607adcSTejun Heo else 558282607adcSTejun Heo wq_watchdog_touched = jiffies; 558382607adcSTejun Heo } 558482607adcSTejun Heo 558582607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 558682607adcSTejun Heo { 558782607adcSTejun Heo wq_watchdog_thresh = 0; 558882607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 558982607adcSTejun Heo 559082607adcSTejun Heo if (thresh) { 559182607adcSTejun Heo wq_watchdog_thresh = thresh; 559282607adcSTejun Heo wq_watchdog_reset_touched(); 559382607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 559482607adcSTejun Heo } 559582607adcSTejun Heo } 559682607adcSTejun Heo 559782607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 559882607adcSTejun Heo const struct kernel_param *kp) 559982607adcSTejun Heo { 560082607adcSTejun Heo unsigned long thresh; 560182607adcSTejun Heo int ret; 560282607adcSTejun Heo 560382607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 560482607adcSTejun Heo if (ret) 560582607adcSTejun Heo return ret; 560682607adcSTejun Heo 560782607adcSTejun Heo if (system_wq) 560882607adcSTejun Heo wq_watchdog_set_thresh(thresh); 560982607adcSTejun Heo else 561082607adcSTejun Heo wq_watchdog_thresh = thresh; 561182607adcSTejun Heo 561282607adcSTejun Heo return 0; 561382607adcSTejun Heo } 561482607adcSTejun Heo 561582607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 561682607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 561782607adcSTejun Heo .get = param_get_ulong, 561882607adcSTejun Heo }; 561982607adcSTejun Heo 562082607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 562182607adcSTejun Heo 0644); 562282607adcSTejun Heo 562382607adcSTejun Heo static void wq_watchdog_init(void) 562482607adcSTejun Heo { 56255cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 562682607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 562782607adcSTejun Heo } 562882607adcSTejun Heo 562982607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 563082607adcSTejun Heo 563182607adcSTejun Heo static inline void wq_watchdog_init(void) { } 563282607adcSTejun Heo 563382607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 563482607adcSTejun Heo 5635bce90380STejun Heo static void __init wq_numa_init(void) 5636bce90380STejun Heo { 5637bce90380STejun Heo cpumask_var_t *tbl; 5638bce90380STejun Heo int node, cpu; 5639bce90380STejun Heo 5640bce90380STejun Heo if (num_possible_nodes() <= 1) 5641bce90380STejun Heo return; 5642bce90380STejun Heo 5643d55262c4STejun Heo if (wq_disable_numa) { 5644d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5645d55262c4STejun Heo return; 5646d55262c4STejun Heo } 5647d55262c4STejun Heo 56484c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 56494c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 56504c16bd32STejun Heo 5651bce90380STejun Heo /* 5652bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5653bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5654bce90380STejun Heo * fully initialized by now. 5655bce90380STejun Heo */ 56566396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5657bce90380STejun Heo BUG_ON(!tbl); 5658bce90380STejun Heo 5659bce90380STejun Heo for_each_node(node) 56605a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 56611be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5662bce90380STejun Heo 5663bce90380STejun Heo for_each_possible_cpu(cpu) { 5664bce90380STejun Heo node = cpu_to_node(cpu); 5665bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5666bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5667bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5668bce90380STejun Heo return; 5669bce90380STejun Heo } 5670bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5671bce90380STejun Heo } 5672bce90380STejun Heo 5673bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5674bce90380STejun Heo wq_numa_enabled = true; 5675bce90380STejun Heo } 5676bce90380STejun Heo 56773347fa09STejun Heo /** 56783347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 56793347fa09STejun Heo * 56803347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 56813347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 56823347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 56833347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 56843347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 56853347fa09STejun Heo * created and scheduled right before early initcalls. 56863347fa09STejun Heo */ 56873347fa09STejun Heo int __init workqueue_init_early(void) 56881da177e4SLinus Torvalds { 56897a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 56901bda3f80SFrederic Weisbecker int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 56917a4e344cSTejun Heo int i, cpu; 5692c34056a3STejun Heo 5693e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5694e904e6c2STejun Heo 5695b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 56961bda3f80SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 5697b05a7928SFrederic Weisbecker 5698e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5699e904e6c2STejun Heo 5700706026c2STejun Heo /* initialize CPU pools */ 570129c91e99STejun Heo for_each_possible_cpu(cpu) { 57024ce62e9eSTejun Heo struct worker_pool *pool; 57038b03ae3cSTejun Heo 57047a4e344cSTejun Heo i = 0; 5705f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 57067a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5707ec22ca5eSTejun Heo pool->cpu = cpu; 57087a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 57097a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5710f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 57117a4e344cSTejun Heo 57129daf9e67STejun Heo /* alloc pool ID */ 571368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 57149daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 571568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 57164ce62e9eSTejun Heo } 57178b03ae3cSTejun Heo } 57188b03ae3cSTejun Heo 57198a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 572029c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 572129c91e99STejun Heo struct workqueue_attrs *attrs; 572229c91e99STejun Heo 572329c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 572429c91e99STejun Heo attrs->nice = std_nice[i]; 572529c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 57268a2b7538STejun Heo 57278a2b7538STejun Heo /* 57288a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 57298a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 57308a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 57318a2b7538STejun Heo */ 57328a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 57338a2b7538STejun Heo attrs->nice = std_nice[i]; 57348a2b7538STejun Heo attrs->no_numa = true; 57358a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 573629c91e99STejun Heo } 573729c91e99STejun Heo 5738d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 57391aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5740d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5741f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5742f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 574324d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 574424d51addSTejun Heo WQ_FREEZABLE, 0); 57450668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 57460668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 57470668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 57480668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 57490668106cSViresh Kumar 0); 57501aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 57510668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 57520668106cSViresh Kumar !system_power_efficient_wq || 57530668106cSViresh Kumar !system_freezable_power_efficient_wq); 575482607adcSTejun Heo 57553347fa09STejun Heo return 0; 57563347fa09STejun Heo } 57573347fa09STejun Heo 57583347fa09STejun Heo /** 57593347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 57603347fa09STejun Heo * 57613347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 57623347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 57633347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 57643347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 57653347fa09STejun Heo * with the initial workers and enable future kworker creations. 57663347fa09STejun Heo */ 57673347fa09STejun Heo int __init workqueue_init(void) 57683347fa09STejun Heo { 57692186d9f9STejun Heo struct workqueue_struct *wq; 57703347fa09STejun Heo struct worker_pool *pool; 57713347fa09STejun Heo int cpu, bkt; 57723347fa09STejun Heo 57732186d9f9STejun Heo /* 57742186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 57752186d9f9STejun Heo * CPU to node mapping may not be available that early on some 57762186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 57772186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 57782186d9f9STejun Heo * affinity, fix them up. 577940c17f75STejun Heo * 578040c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 57812186d9f9STejun Heo */ 57822186d9f9STejun Heo wq_numa_init(); 57832186d9f9STejun Heo 57842186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 57852186d9f9STejun Heo 57862186d9f9STejun Heo for_each_possible_cpu(cpu) { 57872186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 57882186d9f9STejun Heo pool->node = cpu_to_node(cpu); 57892186d9f9STejun Heo } 57902186d9f9STejun Heo } 57912186d9f9STejun Heo 579240c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 57932186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 579440c17f75STejun Heo WARN(init_rescuer(wq), 579540c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 579640c17f75STejun Heo wq->name); 579740c17f75STejun Heo } 57982186d9f9STejun Heo 57992186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 58002186d9f9STejun Heo 58013347fa09STejun Heo /* create the initial workers */ 58023347fa09STejun Heo for_each_online_cpu(cpu) { 58033347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 58043347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 58053347fa09STejun Heo BUG_ON(!create_worker(pool)); 58063347fa09STejun Heo } 58073347fa09STejun Heo } 58083347fa09STejun Heo 58093347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 58103347fa09STejun Heo BUG_ON(!create_worker(pool)); 58113347fa09STejun Heo 58123347fa09STejun Heo wq_online = true; 581382607adcSTejun Heo wq_watchdog_init(); 581482607adcSTejun Heo 58156ee0578bSSuresh Siddha return 0; 58161da177e4SLinus Torvalds } 5817