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 262669de8bdSBart Van Assche char *lock_name; 263669de8bdSBart Van Assche struct lock_class_key key; 2644e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2654e6045f1SJohannes Berg #endif 266ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2672728fd2fSTejun Heo 268e2dca7adSTejun Heo /* 269e2dca7adSTejun Heo * Destruction of workqueue_struct is sched-RCU protected to allow 270e2dca7adSTejun Heo * walking the workqueues list without grabbing wq_pool_mutex. 271e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 272e2dca7adSTejun Heo */ 273e2dca7adSTejun Heo struct rcu_head rcu; 274e2dca7adSTejun Heo 2752728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2762728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2772728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2785b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2791da177e4SLinus Torvalds }; 2801da177e4SLinus Torvalds 281e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 282e904e6c2STejun Heo 283bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 284bce90380STejun Heo /* possible CPUs of each node */ 285bce90380STejun Heo 286d55262c4STejun Heo static bool wq_disable_numa; 287d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 288d55262c4STejun Heo 289cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 290552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 291cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 292cee22a15SViresh Kumar 293863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 2943347fa09STejun Heo 295bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 296bce90380STejun Heo 2974c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2984c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2994c16bd32STejun Heo 30068e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3011258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 3022e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 303692b4825STejun Heo static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ 3045bcab335STejun Heo 305e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30668e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3077d19c5ceSTejun Heo 308ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 309ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 310ef557180SMike Galbraith 311ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 312ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 313b05a7928SFrederic Weisbecker 314f303fccbSTejun Heo /* 315f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 316f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 317f303fccbSTejun Heo * to uncover usages which depend on it. 318f303fccbSTejun Heo */ 319f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 320f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 321f303fccbSTejun Heo #else 322f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 323f303fccbSTejun Heo #endif 324f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 325f303fccbSTejun Heo 3267d19c5ceSTejun Heo /* the per-cpu worker pools */ 32725528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3287d19c5ceSTejun Heo 32968e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3307d19c5ceSTejun Heo 33168e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 33229c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 33329c91e99STejun Heo 334c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33529c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33629c91e99STejun Heo 3378a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3388a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3398a2b7538STejun Heo 340d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 341ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 342044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3431aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 344044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 345d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 346044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 347f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 348044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 34924d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3500668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3510668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3520668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3530668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 354d320c038STejun Heo 3557d19c5ceSTejun Heo static int worker_thread(void *__worker); 3566ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3577d19c5ceSTejun Heo 35897bd2347STejun Heo #define CREATE_TRACE_POINTS 35997bd2347STejun Heo #include <trace/events/workqueue.h> 36097bd2347STejun Heo 36168e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 362f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 363f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36468e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3655bcab335STejun Heo 366b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 367f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 368f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 369b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 37076af4d93STejun Heo 3715b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 372f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 373f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 374f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3755b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3765b95e1afSLai Jiangshan 377f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 378f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 379f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3807a62c2c8STejun Heo (pool)++) 3814ce62e9eSTejun Heo 38249e3cf44STejun Heo /** 38317116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38417116969STejun Heo * @pool: iteration cursor 385611c92a0STejun Heo * @pi: integer used for iteration 386fa1b54e6STejun Heo * 38768e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 38868e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38968e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 390fa1b54e6STejun Heo * 391fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 392fa1b54e6STejun Heo * ignored. 39317116969STejun Heo */ 394611c92a0STejun Heo #define for_each_pool(pool, pi) \ 395611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39668e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 397fa1b54e6STejun Heo else 39817116969STejun Heo 39917116969STejun Heo /** 400822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 401822d8405STejun Heo * @worker: iteration cursor 402822d8405STejun Heo * @pool: worker_pool to iterate workers of 403822d8405STejun Heo * 4041258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 405822d8405STejun Heo * 406822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 407822d8405STejun Heo * ignored. 408822d8405STejun Heo */ 409da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 410da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4111258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 412822d8405STejun Heo else 413822d8405STejun Heo 414822d8405STejun Heo /** 41549e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41649e3cf44STejun Heo * @pwq: iteration cursor 41749e3cf44STejun Heo * @wq: the target workqueue 41876af4d93STejun Heo * 419b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 420794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 421794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 42276af4d93STejun Heo * 42376af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42476af4d93STejun Heo * ignored. 42549e3cf44STejun Heo */ 42649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42776af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 428b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 42976af4d93STejun Heo else 430f3421797STejun Heo 431dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 432dc186ad7SThomas Gleixner 433dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 434dc186ad7SThomas Gleixner 43599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43699777288SStanislaw Gruszka { 43799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43899777288SStanislaw Gruszka } 43999777288SStanislaw Gruszka 440b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 441b9fdac7fSDu, Changbin { 442b9fdac7fSDu, Changbin struct work_struct *work = addr; 443b9fdac7fSDu, Changbin 444b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 445b9fdac7fSDu, Changbin } 446b9fdac7fSDu, Changbin 447dc186ad7SThomas Gleixner /* 448dc186ad7SThomas Gleixner * fixup_init is called when: 449dc186ad7SThomas Gleixner * - an active object is initialized 450dc186ad7SThomas Gleixner */ 45102a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 452dc186ad7SThomas Gleixner { 453dc186ad7SThomas Gleixner struct work_struct *work = addr; 454dc186ad7SThomas Gleixner 455dc186ad7SThomas Gleixner switch (state) { 456dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 457dc186ad7SThomas Gleixner cancel_work_sync(work); 458dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45902a982a6SDu, Changbin return true; 460dc186ad7SThomas Gleixner default: 46102a982a6SDu, Changbin return false; 462dc186ad7SThomas Gleixner } 463dc186ad7SThomas Gleixner } 464dc186ad7SThomas Gleixner 465dc186ad7SThomas Gleixner /* 466dc186ad7SThomas Gleixner * fixup_free is called when: 467dc186ad7SThomas Gleixner * - an active object is freed 468dc186ad7SThomas Gleixner */ 46902a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 470dc186ad7SThomas Gleixner { 471dc186ad7SThomas Gleixner struct work_struct *work = addr; 472dc186ad7SThomas Gleixner 473dc186ad7SThomas Gleixner switch (state) { 474dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 475dc186ad7SThomas Gleixner cancel_work_sync(work); 476dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47702a982a6SDu, Changbin return true; 478dc186ad7SThomas Gleixner default: 47902a982a6SDu, Changbin return false; 480dc186ad7SThomas Gleixner } 481dc186ad7SThomas Gleixner } 482dc186ad7SThomas Gleixner 483dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 484dc186ad7SThomas Gleixner .name = "work_struct", 48599777288SStanislaw Gruszka .debug_hint = work_debug_hint, 486b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 487dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 488dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 489dc186ad7SThomas Gleixner }; 490dc186ad7SThomas Gleixner 491dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 492dc186ad7SThomas Gleixner { 493dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 494dc186ad7SThomas Gleixner } 495dc186ad7SThomas Gleixner 496dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 497dc186ad7SThomas Gleixner { 498dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 499dc186ad7SThomas Gleixner } 500dc186ad7SThomas Gleixner 501dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 502dc186ad7SThomas Gleixner { 503dc186ad7SThomas Gleixner if (onstack) 504dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 505dc186ad7SThomas Gleixner else 506dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 507dc186ad7SThomas Gleixner } 508dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 509dc186ad7SThomas Gleixner 510dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 511dc186ad7SThomas Gleixner { 512dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 513dc186ad7SThomas Gleixner } 514dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 515dc186ad7SThomas Gleixner 516ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 517ea2e64f2SThomas Gleixner { 518ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 519ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 520ea2e64f2SThomas Gleixner } 521ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 522ea2e64f2SThomas Gleixner 523dc186ad7SThomas Gleixner #else 524dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 525dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 526dc186ad7SThomas Gleixner #endif 527dc186ad7SThomas Gleixner 5284e8b22bdSLi Bin /** 5294e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5304e8b22bdSLi Bin * @pool: the pool pointer of interest 5314e8b22bdSLi Bin * 5324e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5334e8b22bdSLi Bin * successfully, -errno on failure. 5344e8b22bdSLi Bin */ 5359daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5369daf9e67STejun Heo { 5379daf9e67STejun Heo int ret; 5389daf9e67STejun Heo 53968e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5405bcab335STejun Heo 5414e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5424e8b22bdSLi Bin GFP_KERNEL); 543229641a6STejun Heo if (ret >= 0) { 544e68035fbSTejun Heo pool->id = ret; 545229641a6STejun Heo return 0; 546229641a6STejun Heo } 5479daf9e67STejun Heo return ret; 5489daf9e67STejun Heo } 5499daf9e67STejun Heo 55076af4d93STejun Heo /** 551df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 552df2d5ae4STejun Heo * @wq: the target workqueue 553df2d5ae4STejun Heo * @node: the node ID 554df2d5ae4STejun Heo * 5555b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5565b95e1afSLai Jiangshan * read locked. 557df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 558df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 559d185af30SYacine Belkadi * 560d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 561df2d5ae4STejun Heo */ 562df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 563df2d5ae4STejun Heo int node) 564df2d5ae4STejun Heo { 5655b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 566d6e022f1STejun Heo 567d6e022f1STejun Heo /* 568d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 569d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 570d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 571d6e022f1STejun Heo * happens, this workaround can be removed. 572d6e022f1STejun Heo */ 573d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 574d6e022f1STejun Heo return wq->dfl_pwq; 575d6e022f1STejun Heo 576df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 577df2d5ae4STejun Heo } 578df2d5ae4STejun Heo 57973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 58073f53c4aSTejun Heo { 58173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 58273f53c4aSTejun Heo } 58373f53c4aSTejun Heo 58473f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58573f53c4aSTejun Heo { 58673f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58873f53c4aSTejun Heo } 58973f53c4aSTejun Heo 59073f53c4aSTejun Heo static int work_next_color(int color) 59173f53c4aSTejun Heo { 59273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 593a848e3b6SOleg Nesterov } 594a848e3b6SOleg Nesterov 5954594bf15SDavid Howells /* 596112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 597112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5987c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5997a22ad75STejun Heo * 600112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 601112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 602bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 603bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6047a22ad75STejun Heo * 605112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6067c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 607112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6087c3eed5cSTejun Heo * available only while the work item is queued. 609bbb68dfaSTejun Heo * 610bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 611bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 612bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 613bbb68dfaSTejun Heo * try to steal the PENDING bit. 6144594bf15SDavid Howells */ 6157a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6167a22ad75STejun Heo unsigned long flags) 6177a22ad75STejun Heo { 6186183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6197a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6207a22ad75STejun Heo } 6217a22ad75STejun Heo 622112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6234690c4abSTejun Heo unsigned long extra_flags) 624365970a1SDavid Howells { 625112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 626112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 627365970a1SDavid Howells } 628365970a1SDavid Howells 6294468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6304468a00fSLai Jiangshan int pool_id) 6314468a00fSLai Jiangshan { 6324468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6334468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6344468a00fSLai Jiangshan } 6354468a00fSLai Jiangshan 6367c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6377c3eed5cSTejun Heo int pool_id) 6384d707b9fSOleg Nesterov { 63923657bb1STejun Heo /* 64023657bb1STejun Heo * The following wmb is paired with the implied mb in 64123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 64223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64323657bb1STejun Heo * owner. 64423657bb1STejun Heo */ 64523657bb1STejun Heo smp_wmb(); 6467c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 647346c09f8SRoman Pen /* 648346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 649346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 650346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6518bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 652346c09f8SRoman Pen * the same @work. E.g. consider this case: 653346c09f8SRoman Pen * 654346c09f8SRoman Pen * CPU#0 CPU#1 655346c09f8SRoman Pen * ---------------------------- -------------------------------- 656346c09f8SRoman Pen * 657346c09f8SRoman Pen * 1 STORE event_indicated 658346c09f8SRoman Pen * 2 queue_work_on() { 659346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 660346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 661346c09f8SRoman Pen * 5 set_work_data() # clear bit 662346c09f8SRoman Pen * 6 smp_mb() 663346c09f8SRoman Pen * 7 work->current_func() { 664346c09f8SRoman Pen * 8 LOAD event_indicated 665346c09f8SRoman Pen * } 666346c09f8SRoman Pen * 667346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 668346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 669346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 670346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 671346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 672346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 673346c09f8SRoman Pen * before actual STORE. 674346c09f8SRoman Pen */ 675346c09f8SRoman Pen smp_mb(); 6764d707b9fSOleg Nesterov } 6774d707b9fSOleg Nesterov 6787a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 679365970a1SDavid Howells { 6807c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6817c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6827a22ad75STejun Heo } 6837a22ad75STejun Heo 684112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6857a22ad75STejun Heo { 686e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6877a22ad75STejun Heo 688112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 689e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 690e120153dSTejun Heo else 691e120153dSTejun Heo return NULL; 6927a22ad75STejun Heo } 6937a22ad75STejun Heo 6947c3eed5cSTejun Heo /** 6957c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6967c3eed5cSTejun Heo * @work: the work item of interest 6977c3eed5cSTejun Heo * 69868e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69968e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 70068e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 701fa1b54e6STejun Heo * 702fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 703fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 704fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 705fa1b54e6STejun Heo * returned pool is and stays online. 706d185af30SYacine Belkadi * 707d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7087c3eed5cSTejun Heo */ 7097c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7107a22ad75STejun Heo { 711e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7127c3eed5cSTejun Heo int pool_id; 7137a22ad75STejun Heo 71468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 715fa1b54e6STejun Heo 716112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 717112202d9STejun Heo return ((struct pool_workqueue *) 7187c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7197a22ad75STejun Heo 7207c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7217c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7227a22ad75STejun Heo return NULL; 7237a22ad75STejun Heo 724fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7257c3eed5cSTejun Heo } 7267c3eed5cSTejun Heo 7277c3eed5cSTejun Heo /** 7287c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7297c3eed5cSTejun Heo * @work: the work item of interest 7307c3eed5cSTejun Heo * 731d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7327c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7337c3eed5cSTejun Heo */ 7347c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7357c3eed5cSTejun Heo { 73654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7377c3eed5cSTejun Heo 738112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 739112202d9STejun Heo return ((struct pool_workqueue *) 74054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 74154d5b7d0SLai Jiangshan 74254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7437c3eed5cSTejun Heo } 7447c3eed5cSTejun Heo 745bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 746bbb68dfaSTejun Heo { 7477c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 748bbb68dfaSTejun Heo 7497c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7507c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 751bbb68dfaSTejun Heo } 752bbb68dfaSTejun Heo 753bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 754bbb68dfaSTejun Heo { 755bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 756bbb68dfaSTejun Heo 757112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 758bbb68dfaSTejun Heo } 759bbb68dfaSTejun Heo 760e22bee78STejun Heo /* 7613270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7623270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 763d565ed63STejun Heo * they're being called with pool->lock held. 764e22bee78STejun Heo */ 765e22bee78STejun Heo 76663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 767649027d7STejun Heo { 768e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 769649027d7STejun Heo } 770649027d7STejun Heo 771e22bee78STejun Heo /* 772e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 773e22bee78STejun Heo * running workers. 774974271c4STejun Heo * 775974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 776706026c2STejun Heo * function will always return %true for unbound pools as long as the 777974271c4STejun Heo * worklist isn't empty. 778e22bee78STejun Heo */ 77963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 780e22bee78STejun Heo { 78163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 782e22bee78STejun Heo } 783e22bee78STejun Heo 784e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 786e22bee78STejun Heo { 78763d95a91STejun Heo return pool->nr_idle; 788e22bee78STejun Heo } 789e22bee78STejun Heo 790e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 79163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 792e22bee78STejun Heo { 793e19e397aSTejun Heo return !list_empty(&pool->worklist) && 794e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 795e22bee78STejun Heo } 796e22bee78STejun Heo 797e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 799e22bee78STejun Heo { 80063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 801e22bee78STejun Heo } 802e22bee78STejun Heo 803e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80463d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 805e22bee78STejun Heo { 806692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 80763d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80863d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 809e22bee78STejun Heo 810e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 811e22bee78STejun Heo } 812e22bee78STejun Heo 813e22bee78STejun Heo /* 814e22bee78STejun Heo * Wake up functions. 815e22bee78STejun Heo */ 816e22bee78STejun Heo 8171037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8181037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8197e11629dSTejun Heo { 82063d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8217e11629dSTejun Heo return NULL; 8227e11629dSTejun Heo 82363d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8247e11629dSTejun Heo } 8257e11629dSTejun Heo 8267e11629dSTejun Heo /** 8277e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82863d95a91STejun Heo * @pool: worker pool to wake worker from 8297e11629dSTejun Heo * 83063d95a91STejun Heo * Wake up the first idle worker of @pool. 8317e11629dSTejun Heo * 8327e11629dSTejun Heo * CONTEXT: 833d565ed63STejun Heo * spin_lock_irq(pool->lock). 8347e11629dSTejun Heo */ 83563d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8367e11629dSTejun Heo { 8371037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8387e11629dSTejun Heo 8397e11629dSTejun Heo if (likely(worker)) 8407e11629dSTejun Heo wake_up_process(worker->task); 8417e11629dSTejun Heo } 8427e11629dSTejun Heo 8434690c4abSTejun Heo /** 844e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 845e22bee78STejun Heo * @task: task waking up 846e22bee78STejun Heo * @cpu: CPU @task is waking up to 847e22bee78STejun Heo * 848e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 849e22bee78STejun Heo * being awoken. 850e22bee78STejun Heo * 851e22bee78STejun Heo * CONTEXT: 852e22bee78STejun Heo * spin_lock_irq(rq->lock) 853e22bee78STejun Heo */ 854d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 855e22bee78STejun Heo { 856e22bee78STejun Heo struct worker *worker = kthread_data(task); 857e22bee78STejun Heo 85836576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 859ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 860e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 861e22bee78STejun Heo } 86236576000SJoonsoo Kim } 863e22bee78STejun Heo 864e22bee78STejun Heo /** 865e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 866e22bee78STejun Heo * @task: task going to sleep 867e22bee78STejun Heo * 868e22bee78STejun Heo * This function is called during schedule() when a busy worker is 869e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 870e22bee78STejun Heo * returning pointer to its task. 871e22bee78STejun Heo * 872e22bee78STejun Heo * CONTEXT: 873e22bee78STejun Heo * spin_lock_irq(rq->lock) 874e22bee78STejun Heo * 875d185af30SYacine Belkadi * Return: 876e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 877e22bee78STejun Heo */ 8789b7f6597SAlexander Gordeev struct task_struct *wq_worker_sleeping(struct task_struct *task) 879e22bee78STejun Heo { 880e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 881111c225aSTejun Heo struct worker_pool *pool; 882e22bee78STejun Heo 883111c225aSTejun Heo /* 884111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 885111c225aSTejun Heo * workers, also reach here, let's not access anything before 886111c225aSTejun Heo * checking NOT_RUNNING. 887111c225aSTejun Heo */ 8882d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 889e22bee78STejun Heo return NULL; 890e22bee78STejun Heo 891111c225aSTejun Heo pool = worker->pool; 892111c225aSTejun Heo 893e22bee78STejun Heo /* this can only happen on the local cpu */ 8949b7f6597SAlexander Gordeev if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) 8956183c009STejun Heo return NULL; 896e22bee78STejun Heo 897e22bee78STejun Heo /* 898e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 899e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 900e22bee78STejun Heo * Please read comment there. 901e22bee78STejun Heo * 902628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 903628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 904628c78e7STejun Heo * disabled, which in turn means that none else could be 905d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 906628c78e7STejun Heo * lock is safe. 907e22bee78STejun Heo */ 908e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 909e19e397aSTejun Heo !list_empty(&pool->worklist)) 9101037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 911e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 912e22bee78STejun Heo } 913e22bee78STejun Heo 914e22bee78STejun Heo /** 9151b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 9161b69ac6bSJohannes Weiner * 9171b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 9181b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 9191b69ac6bSJohannes Weiner * 9201b69ac6bSJohannes Weiner * CONTEXT: 9211b69ac6bSJohannes Weiner * spin_lock_irq(rq->lock) 9221b69ac6bSJohannes Weiner * 9234b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 9244b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 9254b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 9264b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 9274b047002SJohannes Weiner * 9284b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 9294b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 9304b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 9314b047002SJohannes Weiner * is guaranteed to not be processing any works. 9324b047002SJohannes Weiner * 9331b69ac6bSJohannes Weiner * Return: 9341b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 9351b69ac6bSJohannes Weiner * hasn't executed any work yet. 9361b69ac6bSJohannes Weiner */ 9371b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 9381b69ac6bSJohannes Weiner { 9391b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 9401b69ac6bSJohannes Weiner 9411b69ac6bSJohannes Weiner return worker->last_func; 9421b69ac6bSJohannes Weiner } 9431b69ac6bSJohannes Weiner 9441b69ac6bSJohannes Weiner /** 945e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 946cb444766STejun Heo * @worker: self 947d302f017STejun Heo * @flags: flags to set 948d302f017STejun Heo * 949228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 950d302f017STejun Heo * 951cb444766STejun Heo * CONTEXT: 952d565ed63STejun Heo * spin_lock_irq(pool->lock) 953d302f017STejun Heo */ 954228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 955d302f017STejun Heo { 956bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 957e22bee78STejun Heo 958cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 959cb444766STejun Heo 960228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 961e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 962e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 963e19e397aSTejun Heo atomic_dec(&pool->nr_running); 964e22bee78STejun Heo } 965e22bee78STejun Heo 966d302f017STejun Heo worker->flags |= flags; 967d302f017STejun Heo } 968d302f017STejun Heo 969d302f017STejun Heo /** 970e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 971cb444766STejun Heo * @worker: self 972d302f017STejun Heo * @flags: flags to clear 973d302f017STejun Heo * 974e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 975d302f017STejun Heo * 976cb444766STejun Heo * CONTEXT: 977d565ed63STejun Heo * spin_lock_irq(pool->lock) 978d302f017STejun Heo */ 979d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 980d302f017STejun Heo { 98163d95a91STejun Heo struct worker_pool *pool = worker->pool; 982e22bee78STejun Heo unsigned int oflags = worker->flags; 983e22bee78STejun Heo 984cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 985cb444766STejun Heo 986d302f017STejun Heo worker->flags &= ~flags; 987e22bee78STejun Heo 98842c025f3STejun Heo /* 98942c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 99042c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 99142c025f3STejun Heo * of multiple flags, not a single flag. 99242c025f3STejun Heo */ 993e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 994e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 995e19e397aSTejun Heo atomic_inc(&pool->nr_running); 996d302f017STejun Heo } 997d302f017STejun Heo 998d302f017STejun Heo /** 9998cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1000c9e7cf27STejun Heo * @pool: pool of interest 10018cca0eeaSTejun Heo * @work: work to find worker for 10028cca0eeaSTejun Heo * 1003c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1004c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1005a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1006a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1007a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1008a2c1c57bSTejun Heo * being executed. 1009a2c1c57bSTejun Heo * 1010a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1011a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1012a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1013a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1014a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1015a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1016a2c1c57bSTejun Heo * 1017c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1018c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1019c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1020c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1021c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1022c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10238cca0eeaSTejun Heo * 10248cca0eeaSTejun Heo * CONTEXT: 1025d565ed63STejun Heo * spin_lock_irq(pool->lock). 10268cca0eeaSTejun Heo * 1027d185af30SYacine Belkadi * Return: 1028d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10298cca0eeaSTejun Heo * otherwise. 10308cca0eeaSTejun Heo */ 1031c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10328cca0eeaSTejun Heo struct work_struct *work) 10338cca0eeaSTejun Heo { 103442f8570fSSasha Levin struct worker *worker; 103542f8570fSSasha Levin 1036b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1037a2c1c57bSTejun Heo (unsigned long)work) 1038a2c1c57bSTejun Heo if (worker->current_work == work && 1039a2c1c57bSTejun Heo worker->current_func == work->func) 104042f8570fSSasha Levin return worker; 104142f8570fSSasha Levin 104242f8570fSSasha Levin return NULL; 10438cca0eeaSTejun Heo } 10448cca0eeaSTejun Heo 10458cca0eeaSTejun Heo /** 1046bf4ede01STejun Heo * move_linked_works - move linked works to a list 1047bf4ede01STejun Heo * @work: start of series of works to be scheduled 1048bf4ede01STejun Heo * @head: target list to append @work to 1049402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1050bf4ede01STejun Heo * 1051bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1052bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1053bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1054bf4ede01STejun Heo * 1055bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1056bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1057bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1058bf4ede01STejun Heo * 1059bf4ede01STejun Heo * CONTEXT: 1060d565ed63STejun Heo * spin_lock_irq(pool->lock). 1061bf4ede01STejun Heo */ 1062bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1063bf4ede01STejun Heo struct work_struct **nextp) 1064bf4ede01STejun Heo { 1065bf4ede01STejun Heo struct work_struct *n; 1066bf4ede01STejun Heo 1067bf4ede01STejun Heo /* 1068bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1069bf4ede01STejun Heo * use NULL for list head. 1070bf4ede01STejun Heo */ 1071bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1072bf4ede01STejun Heo list_move_tail(&work->entry, head); 1073bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1074bf4ede01STejun Heo break; 1075bf4ede01STejun Heo } 1076bf4ede01STejun Heo 1077bf4ede01STejun Heo /* 1078bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1079bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1080bf4ede01STejun Heo * needs to be updated. 1081bf4ede01STejun Heo */ 1082bf4ede01STejun Heo if (nextp) 1083bf4ede01STejun Heo *nextp = n; 1084bf4ede01STejun Heo } 1085bf4ede01STejun Heo 10868864b4e5STejun Heo /** 10878864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10888864b4e5STejun Heo * @pwq: pool_workqueue to get 10898864b4e5STejun Heo * 10908864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10918864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10928864b4e5STejun Heo */ 10938864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10948864b4e5STejun Heo { 10958864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10968864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10978864b4e5STejun Heo pwq->refcnt++; 10988864b4e5STejun Heo } 10998864b4e5STejun Heo 11008864b4e5STejun Heo /** 11018864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 11028864b4e5STejun Heo * @pwq: pool_workqueue to put 11038864b4e5STejun Heo * 11048864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 11058864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11068864b4e5STejun Heo */ 11078864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11088864b4e5STejun Heo { 11098864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11108864b4e5STejun Heo if (likely(--pwq->refcnt)) 11118864b4e5STejun Heo return; 11128864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11138864b4e5STejun Heo return; 11148864b4e5STejun Heo /* 11158864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11168864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11178864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11188864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11198864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11208864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11218864b4e5STejun Heo */ 11228864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11238864b4e5STejun Heo } 11248864b4e5STejun Heo 1125dce90d47STejun Heo /** 1126dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1127dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1128dce90d47STejun Heo * 1129dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1130dce90d47STejun Heo */ 1131dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1132dce90d47STejun Heo { 1133dce90d47STejun Heo if (pwq) { 1134dce90d47STejun Heo /* 1135dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1136dce90d47STejun Heo * following lock operations are safe. 1137dce90d47STejun Heo */ 1138dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1139dce90d47STejun Heo put_pwq(pwq); 1140dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1141dce90d47STejun Heo } 1142dce90d47STejun Heo } 1143dce90d47STejun Heo 1144112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1145bf4ede01STejun Heo { 1146112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1147bf4ede01STejun Heo 1148bf4ede01STejun Heo trace_workqueue_activate_work(work); 114982607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 115082607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1151112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1152bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1153112202d9STejun Heo pwq->nr_active++; 1154bf4ede01STejun Heo } 1155bf4ede01STejun Heo 1156112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11573aa62497SLai Jiangshan { 1158112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11593aa62497SLai Jiangshan struct work_struct, entry); 11603aa62497SLai Jiangshan 1161112202d9STejun Heo pwq_activate_delayed_work(work); 11623aa62497SLai Jiangshan } 11633aa62497SLai Jiangshan 1164bf4ede01STejun Heo /** 1165112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1166112202d9STejun Heo * @pwq: pwq of interest 1167bf4ede01STejun Heo * @color: color of work which left the queue 1168bf4ede01STejun Heo * 1169bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1170112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1171bf4ede01STejun Heo * 1172bf4ede01STejun Heo * CONTEXT: 1173d565ed63STejun Heo * spin_lock_irq(pool->lock). 1174bf4ede01STejun Heo */ 1175112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1176bf4ede01STejun Heo { 11778864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1178bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11798864b4e5STejun Heo goto out_put; 1180bf4ede01STejun Heo 1181112202d9STejun Heo pwq->nr_in_flight[color]--; 1182bf4ede01STejun Heo 1183112202d9STejun Heo pwq->nr_active--; 1184112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1185bf4ede01STejun Heo /* one down, submit a delayed one */ 1186112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1187112202d9STejun Heo pwq_activate_first_delayed(pwq); 1188bf4ede01STejun Heo } 1189bf4ede01STejun Heo 1190bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1191112202d9STejun Heo if (likely(pwq->flush_color != color)) 11928864b4e5STejun Heo goto out_put; 1193bf4ede01STejun Heo 1194bf4ede01STejun Heo /* are there still in-flight works? */ 1195112202d9STejun Heo if (pwq->nr_in_flight[color]) 11968864b4e5STejun Heo goto out_put; 1197bf4ede01STejun Heo 1198112202d9STejun Heo /* this pwq is done, clear flush_color */ 1199112202d9STejun Heo pwq->flush_color = -1; 1200bf4ede01STejun Heo 1201bf4ede01STejun Heo /* 1202112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1203bf4ede01STejun Heo * will handle the rest. 1204bf4ede01STejun Heo */ 1205112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1206112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12078864b4e5STejun Heo out_put: 12088864b4e5STejun Heo put_pwq(pwq); 1209bf4ede01STejun Heo } 1210bf4ede01STejun Heo 121136e227d2STejun Heo /** 1212bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 121336e227d2STejun Heo * @work: work item to steal 121436e227d2STejun Heo * @is_dwork: @work is a delayed_work 1215bbb68dfaSTejun Heo * @flags: place to store irq state 121636e227d2STejun Heo * 121736e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1218d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 121936e227d2STejun Heo * 1220d185af30SYacine Belkadi * Return: 122136e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 122236e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 122336e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1224bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1225bbb68dfaSTejun Heo * for arbitrarily long 122636e227d2STejun Heo * 1227d185af30SYacine Belkadi * Note: 1228bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1229e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1230e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1231e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1232bbb68dfaSTejun Heo * 1233bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1234bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1235bbb68dfaSTejun Heo * 1236e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1237bf4ede01STejun Heo */ 1238bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1239bbb68dfaSTejun Heo unsigned long *flags) 1240bf4ede01STejun Heo { 1241d565ed63STejun Heo struct worker_pool *pool; 1242112202d9STejun Heo struct pool_workqueue *pwq; 1243bf4ede01STejun Heo 1244bbb68dfaSTejun Heo local_irq_save(*flags); 1245bbb68dfaSTejun Heo 124636e227d2STejun Heo /* try to steal the timer if it exists */ 124736e227d2STejun Heo if (is_dwork) { 124836e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 124936e227d2STejun Heo 1250e0aecdd8STejun Heo /* 1251e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1252e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1253e0aecdd8STejun Heo * running on the local CPU. 1254e0aecdd8STejun Heo */ 125536e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 125636e227d2STejun Heo return 1; 125736e227d2STejun Heo } 125836e227d2STejun Heo 125936e227d2STejun Heo /* try to claim PENDING the normal way */ 1260bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1261bf4ede01STejun Heo return 0; 1262bf4ede01STejun Heo 1263bf4ede01STejun Heo /* 1264bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1265bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1266bf4ede01STejun Heo */ 1267d565ed63STejun Heo pool = get_work_pool(work); 1268d565ed63STejun Heo if (!pool) 1269bbb68dfaSTejun Heo goto fail; 1270bf4ede01STejun Heo 1271d565ed63STejun Heo spin_lock(&pool->lock); 1272bf4ede01STejun Heo /* 1273112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1274112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1275112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1276112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1277112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12780b3dae68SLai Jiangshan * item is currently queued on that pool. 1279bf4ede01STejun Heo */ 1280112202d9STejun Heo pwq = get_work_pwq(work); 1281112202d9STejun Heo if (pwq && pwq->pool == pool) { 1282bf4ede01STejun Heo debug_work_deactivate(work); 12833aa62497SLai Jiangshan 12843aa62497SLai Jiangshan /* 128516062836STejun Heo * A delayed work item cannot be grabbed directly because 128616062836STejun Heo * it might have linked NO_COLOR work items which, if left 1287112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 128816062836STejun Heo * management later on and cause stall. Make sure the work 128916062836STejun Heo * item is activated before grabbing. 12903aa62497SLai Jiangshan */ 12913aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1292112202d9STejun Heo pwq_activate_delayed_work(work); 12933aa62497SLai Jiangshan 1294bf4ede01STejun Heo list_del_init(&work->entry); 12959c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 129636e227d2STejun Heo 1297112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12984468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12994468a00fSLai Jiangshan 1300d565ed63STejun Heo spin_unlock(&pool->lock); 130136e227d2STejun Heo return 1; 1302bf4ede01STejun Heo } 1303d565ed63STejun Heo spin_unlock(&pool->lock); 1304bbb68dfaSTejun Heo fail: 1305bbb68dfaSTejun Heo local_irq_restore(*flags); 1306bbb68dfaSTejun Heo if (work_is_canceling(work)) 1307bbb68dfaSTejun Heo return -ENOENT; 1308bbb68dfaSTejun Heo cpu_relax(); 130936e227d2STejun Heo return -EAGAIN; 1310bf4ede01STejun Heo } 1311bf4ede01STejun Heo 1312bf4ede01STejun Heo /** 1313706026c2STejun Heo * insert_work - insert a work into a pool 1314112202d9STejun Heo * @pwq: pwq @work belongs to 13154690c4abSTejun Heo * @work: work to insert 13164690c4abSTejun Heo * @head: insertion point 13174690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13184690c4abSTejun Heo * 1319112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1320706026c2STejun Heo * work_struct flags. 13214690c4abSTejun Heo * 13224690c4abSTejun Heo * CONTEXT: 1323d565ed63STejun Heo * spin_lock_irq(pool->lock). 1324365970a1SDavid Howells */ 1325112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1326112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1327b89deed3SOleg Nesterov { 1328112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1329e1d8aa9fSFrederic Weisbecker 13304690c4abSTejun Heo /* we own @work, set data and link */ 1331112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13321a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13338864b4e5STejun Heo get_pwq(pwq); 1334e22bee78STejun Heo 1335e22bee78STejun Heo /* 1336c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1337c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1338c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1339e22bee78STejun Heo */ 1340e22bee78STejun Heo smp_mb(); 1341e22bee78STejun Heo 134263d95a91STejun Heo if (__need_more_worker(pool)) 134363d95a91STejun Heo wake_up_worker(pool); 1344b89deed3SOleg Nesterov } 1345b89deed3SOleg Nesterov 1346c8efcc25STejun Heo /* 1347c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13488d03ecfeSTejun Heo * same workqueue. 1349c8efcc25STejun Heo */ 1350c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1351c8efcc25STejun Heo { 1352c8efcc25STejun Heo struct worker *worker; 1353c8efcc25STejun Heo 13548d03ecfeSTejun Heo worker = current_wq_worker(); 1355c8efcc25STejun Heo /* 1356bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 13578d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1358c8efcc25STejun Heo */ 1359112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1360c8efcc25STejun Heo } 1361c8efcc25STejun Heo 1362ef557180SMike Galbraith /* 1363ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1364ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1365ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1366ef557180SMike Galbraith */ 1367ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1368ef557180SMike Galbraith { 1369f303fccbSTejun Heo static bool printed_dbg_warning; 1370ef557180SMike Galbraith int new_cpu; 1371ef557180SMike Galbraith 1372f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1373ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1374ef557180SMike Galbraith return cpu; 1375f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1376f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1377f303fccbSTejun Heo printed_dbg_warning = true; 1378f303fccbSTejun Heo } 1379f303fccbSTejun Heo 1380ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1381ef557180SMike Galbraith return cpu; 1382ef557180SMike Galbraith 1383ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1384ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1385ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1386ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1387ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1388ef557180SMike Galbraith return cpu; 1389ef557180SMike Galbraith } 1390ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1391ef557180SMike Galbraith 1392ef557180SMike Galbraith return new_cpu; 1393ef557180SMike Galbraith } 1394ef557180SMike Galbraith 1395d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13961da177e4SLinus Torvalds struct work_struct *work) 13971da177e4SLinus Torvalds { 1398112202d9STejun Heo struct pool_workqueue *pwq; 1399c9178087STejun Heo struct worker_pool *last_pool; 14001e19ffc6STejun Heo struct list_head *worklist; 14018a2e8e5dSTejun Heo unsigned int work_flags; 1402b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 14038930cabaSTejun Heo 14048930cabaSTejun Heo /* 14058930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 14068930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 14078930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 14088930cabaSTejun Heo * happen with IRQ disabled. 14098930cabaSTejun Heo */ 14108e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 14111da177e4SLinus Torvalds 1412dc186ad7SThomas Gleixner debug_work_activate(work); 14131e19ffc6STejun Heo 14149ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1415618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1416c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1417e41e704bSTejun Heo return; 14189e8cd2f5STejun Heo retry: 1419df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1420ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1421df2d5ae4STejun Heo 1422df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1423df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1424c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1425df2d5ae4STejun Heo else 1426df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1427f3421797STejun Heo 142818aa9effSTejun Heo /* 1429c9178087STejun Heo * If @work was previously on a different pool, it might still be 1430c9178087STejun Heo * running there, in which case the work needs to be queued on that 1431c9178087STejun Heo * pool to guarantee non-reentrancy. 143218aa9effSTejun Heo */ 1433c9e7cf27STejun Heo last_pool = get_work_pool(work); 1434112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 143518aa9effSTejun Heo struct worker *worker; 143618aa9effSTejun Heo 1437d565ed63STejun Heo spin_lock(&last_pool->lock); 143818aa9effSTejun Heo 1439c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 144018aa9effSTejun Heo 1441112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1442c9178087STejun Heo pwq = worker->current_pwq; 14438594fadeSLai Jiangshan } else { 144418aa9effSTejun Heo /* meh... not running there, queue here */ 1445d565ed63STejun Heo spin_unlock(&last_pool->lock); 1446112202d9STejun Heo spin_lock(&pwq->pool->lock); 144718aa9effSTejun Heo } 14488930cabaSTejun Heo } else { 1449112202d9STejun Heo spin_lock(&pwq->pool->lock); 14508930cabaSTejun Heo } 1451502ca9d8STejun Heo 14529e8cd2f5STejun Heo /* 14539e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14549e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14559e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1456df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1457df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14589e8cd2f5STejun Heo * make forward-progress. 14599e8cd2f5STejun Heo */ 14609e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14619e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 14629e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 14639e8cd2f5STejun Heo cpu_relax(); 14649e8cd2f5STejun Heo goto retry; 14659e8cd2f5STejun Heo } 14669e8cd2f5STejun Heo /* oops */ 14679e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14689e8cd2f5STejun Heo wq->name, cpu); 14699e8cd2f5STejun Heo } 14709e8cd2f5STejun Heo 1471112202d9STejun Heo /* pwq determined, queue */ 1472112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1473502ca9d8STejun Heo 1474f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1475112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1476f5b2552bSDan Carpenter return; 1477f5b2552bSDan Carpenter } 14781e19ffc6STejun Heo 1479112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1480112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14811e19ffc6STejun Heo 1482112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1483cdadf009STejun Heo trace_workqueue_activate_work(work); 1484112202d9STejun Heo pwq->nr_active++; 1485112202d9STejun Heo worklist = &pwq->pool->worklist; 148682607adcSTejun Heo if (list_empty(worklist)) 148782607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14888a2e8e5dSTejun Heo } else { 14898a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1490112202d9STejun Heo worklist = &pwq->delayed_works; 14918a2e8e5dSTejun Heo } 14921e19ffc6STejun Heo 1493112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14941e19ffc6STejun Heo 1495112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14961da177e4SLinus Torvalds } 14971da177e4SLinus Torvalds 14980fcb78c2SRolf Eike Beer /** 1499c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1500c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1501c1a220e7SZhang Rui * @wq: workqueue to use 1502c1a220e7SZhang Rui * @work: work to queue 1503c1a220e7SZhang Rui * 1504c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1505c1a220e7SZhang Rui * can't go away. 1506d185af30SYacine Belkadi * 1507d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1508c1a220e7SZhang Rui */ 1509d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1510d4283e93STejun Heo struct work_struct *work) 1511c1a220e7SZhang Rui { 1512d4283e93STejun Heo bool ret = false; 15138930cabaSTejun Heo unsigned long flags; 15148930cabaSTejun Heo 15158930cabaSTejun Heo local_irq_save(flags); 1516c1a220e7SZhang Rui 151722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15184690c4abSTejun Heo __queue_work(cpu, wq, work); 1519d4283e93STejun Heo ret = true; 1520c1a220e7SZhang Rui } 15218930cabaSTejun Heo 15228930cabaSTejun Heo local_irq_restore(flags); 1523c1a220e7SZhang Rui return ret; 1524c1a220e7SZhang Rui } 1525ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1526c1a220e7SZhang Rui 15278204e0c1SAlexander Duyck /** 15288204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15298204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15308204e0c1SAlexander Duyck * 15318204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15328204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15338204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15348204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15358204e0c1SAlexander Duyck */ 15368204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15378204e0c1SAlexander Duyck { 15388204e0c1SAlexander Duyck int cpu; 15398204e0c1SAlexander Duyck 15408204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15418204e0c1SAlexander Duyck if (!wq_numa_enabled) 15428204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15438204e0c1SAlexander Duyck 15448204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15458204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15468204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15478204e0c1SAlexander Duyck 15488204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15498204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15508204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15518204e0c1SAlexander Duyck return cpu; 15528204e0c1SAlexander Duyck 15538204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15548204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15558204e0c1SAlexander Duyck 15568204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15578204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15588204e0c1SAlexander Duyck } 15598204e0c1SAlexander Duyck 15608204e0c1SAlexander Duyck /** 15618204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15628204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15638204e0c1SAlexander Duyck * @wq: workqueue to use 15648204e0c1SAlexander Duyck * @work: work to queue 15658204e0c1SAlexander Duyck * 15668204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 15678204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 15688204e0c1SAlexander Duyck * NUMA node. 15698204e0c1SAlexander Duyck * 15708204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 15718204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 15728204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 15738204e0c1SAlexander Duyck * 15748204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 15758204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 15768204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 15778204e0c1SAlexander Duyck * 15788204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 15798204e0c1SAlexander Duyck */ 15808204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 15818204e0c1SAlexander Duyck struct work_struct *work) 15828204e0c1SAlexander Duyck { 15838204e0c1SAlexander Duyck unsigned long flags; 15848204e0c1SAlexander Duyck bool ret = false; 15858204e0c1SAlexander Duyck 15868204e0c1SAlexander Duyck /* 15878204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 15888204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 15898204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 15908204e0c1SAlexander Duyck * 15918204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 15928204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 15938204e0c1SAlexander Duyck * some round robin type logic. 15948204e0c1SAlexander Duyck */ 15958204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 15968204e0c1SAlexander Duyck 15978204e0c1SAlexander Duyck local_irq_save(flags); 15988204e0c1SAlexander Duyck 15998204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16008204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16018204e0c1SAlexander Duyck 16028204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16038204e0c1SAlexander Duyck ret = true; 16048204e0c1SAlexander Duyck } 16058204e0c1SAlexander Duyck 16068204e0c1SAlexander Duyck local_irq_restore(flags); 16078204e0c1SAlexander Duyck return ret; 16088204e0c1SAlexander Duyck } 16098204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16108204e0c1SAlexander Duyck 16118c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16121da177e4SLinus Torvalds { 16138c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16141da177e4SLinus Torvalds 1615e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 161660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16171da177e4SLinus Torvalds } 16181438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16191da177e4SLinus Torvalds 16207beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 162152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16221da177e4SLinus Torvalds { 16237beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16247beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16251da177e4SLinus Torvalds 1626637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 1627841b86f3SKees Cook WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1628fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1629fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16307beb2edfSTejun Heo 16318852aac2STejun Heo /* 16328852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16338852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16348852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16358852aac2STejun Heo * on that there's no such delay when @delay is 0. 16368852aac2STejun Heo */ 16378852aac2STejun Heo if (!delay) { 16388852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16398852aac2STejun Heo return; 16408852aac2STejun Heo } 16418852aac2STejun Heo 164260c057bcSLai Jiangshan dwork->wq = wq; 16431265057fSTejun Heo dwork->cpu = cpu; 16447beb2edfSTejun Heo timer->expires = jiffies + delay; 16457beb2edfSTejun Heo 1646041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16477beb2edfSTejun Heo add_timer_on(timer, cpu); 1648041bd12eSTejun Heo else 1649041bd12eSTejun Heo add_timer(timer); 16507beb2edfSTejun Heo } 16511da177e4SLinus Torvalds 16520fcb78c2SRolf Eike Beer /** 16530fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16540fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16550fcb78c2SRolf Eike Beer * @wq: workqueue to use 1656af9997e4SRandy Dunlap * @dwork: work to queue 16570fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16580fcb78c2SRolf Eike Beer * 1659d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1660715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1661715f1300STejun Heo * execution. 16620fcb78c2SRolf Eike Beer */ 1663d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 166452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16657a6bc1cdSVenkatesh Pallipadi { 166652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1667d4283e93STejun Heo bool ret = false; 16688930cabaSTejun Heo unsigned long flags; 16698930cabaSTejun Heo 16708930cabaSTejun Heo /* read the comment in __queue_work() */ 16718930cabaSTejun Heo local_irq_save(flags); 16727a6bc1cdSVenkatesh Pallipadi 167322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16747beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1675d4283e93STejun Heo ret = true; 16767a6bc1cdSVenkatesh Pallipadi } 16778930cabaSTejun Heo 16788930cabaSTejun Heo local_irq_restore(flags); 16797a6bc1cdSVenkatesh Pallipadi return ret; 16807a6bc1cdSVenkatesh Pallipadi } 1681ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 16821da177e4SLinus Torvalds 1683c8e55f36STejun Heo /** 16848376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 16858376fe22STejun Heo * @cpu: CPU number to execute work on 16868376fe22STejun Heo * @wq: workqueue to use 16878376fe22STejun Heo * @dwork: work to queue 16888376fe22STejun Heo * @delay: number of jiffies to wait before queueing 16898376fe22STejun Heo * 16908376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 16918376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 16928376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 16938376fe22STejun Heo * current state. 16948376fe22STejun Heo * 1695d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 16968376fe22STejun Heo * pending and its timer was modified. 16978376fe22STejun Heo * 1698e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 16998376fe22STejun Heo * See try_to_grab_pending() for details. 17008376fe22STejun Heo */ 17018376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17028376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17038376fe22STejun Heo { 17048376fe22STejun Heo unsigned long flags; 17058376fe22STejun Heo int ret; 17068376fe22STejun Heo 17078376fe22STejun Heo do { 17088376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17098376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17108376fe22STejun Heo 17118376fe22STejun Heo if (likely(ret >= 0)) { 17128376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17138376fe22STejun Heo local_irq_restore(flags); 17148376fe22STejun Heo } 17158376fe22STejun Heo 17168376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17178376fe22STejun Heo return ret; 17188376fe22STejun Heo } 17198376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17208376fe22STejun Heo 172105f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 172205f0fe6bSTejun Heo { 172305f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 172405f0fe6bSTejun Heo 172505f0fe6bSTejun Heo /* read the comment in __queue_work() */ 172605f0fe6bSTejun Heo local_irq_disable(); 172705f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 172805f0fe6bSTejun Heo local_irq_enable(); 172905f0fe6bSTejun Heo } 173005f0fe6bSTejun Heo 173105f0fe6bSTejun Heo /** 173205f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 173305f0fe6bSTejun Heo * @wq: workqueue to use 173405f0fe6bSTejun Heo * @rwork: work to queue 173505f0fe6bSTejun Heo * 173605f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 173705f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1738bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 173905f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 174005f0fe6bSTejun Heo */ 174105f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 174205f0fe6bSTejun Heo { 174305f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 174405f0fe6bSTejun Heo 174505f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 174605f0fe6bSTejun Heo rwork->wq = wq; 174705f0fe6bSTejun Heo call_rcu(&rwork->rcu, rcu_work_rcufn); 174805f0fe6bSTejun Heo return true; 174905f0fe6bSTejun Heo } 175005f0fe6bSTejun Heo 175105f0fe6bSTejun Heo return false; 175205f0fe6bSTejun Heo } 175305f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 175405f0fe6bSTejun Heo 17558376fe22STejun Heo /** 1756c8e55f36STejun Heo * worker_enter_idle - enter idle state 1757c8e55f36STejun Heo * @worker: worker which is entering idle state 1758c8e55f36STejun Heo * 1759c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1760c8e55f36STejun Heo * necessary. 1761c8e55f36STejun Heo * 1762c8e55f36STejun Heo * LOCKING: 1763d565ed63STejun Heo * spin_lock_irq(pool->lock). 1764c8e55f36STejun Heo */ 1765c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 17661da177e4SLinus Torvalds { 1767bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1768c8e55f36STejun Heo 17696183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 17706183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 17716183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 17726183c009STejun Heo return; 1773c8e55f36STejun Heo 1774051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1775cb444766STejun Heo worker->flags |= WORKER_IDLE; 1776bd7bdd43STejun Heo pool->nr_idle++; 1777e22bee78STejun Heo worker->last_active = jiffies; 1778c8e55f36STejun Heo 1779c8e55f36STejun Heo /* idle_list is LIFO */ 1780bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1781db7bccf4STejun Heo 178263d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1783628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1784cb444766STejun Heo 1785544ecf31STejun Heo /* 1786e8b3f8dbSLai Jiangshan * Sanity check nr_running. Because unbind_workers() releases 1787d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1788628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1789628c78e7STejun Heo * unbind is not in progress. 1790544ecf31STejun Heo */ 179124647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1792bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1793e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1794c8e55f36STejun Heo } 1795c8e55f36STejun Heo 1796c8e55f36STejun Heo /** 1797c8e55f36STejun Heo * worker_leave_idle - leave idle state 1798c8e55f36STejun Heo * @worker: worker which is leaving idle state 1799c8e55f36STejun Heo * 1800c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1801c8e55f36STejun Heo * 1802c8e55f36STejun Heo * LOCKING: 1803d565ed63STejun Heo * spin_lock_irq(pool->lock). 1804c8e55f36STejun Heo */ 1805c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1806c8e55f36STejun Heo { 1807bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1808c8e55f36STejun Heo 18096183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18106183c009STejun Heo return; 1811d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1812bd7bdd43STejun Heo pool->nr_idle--; 1813c8e55f36STejun Heo list_del_init(&worker->entry); 1814c8e55f36STejun Heo } 1815c8e55f36STejun Heo 1816f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1817c34056a3STejun Heo { 1818c34056a3STejun Heo struct worker *worker; 1819c34056a3STejun Heo 1820f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1821c8e55f36STejun Heo if (worker) { 1822c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1823affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1824da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1825e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1826e22bee78STejun Heo worker->flags = WORKER_PREP; 1827c8e55f36STejun Heo } 1828c34056a3STejun Heo return worker; 1829c34056a3STejun Heo } 1830c34056a3STejun Heo 1831c34056a3STejun Heo /** 18324736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18334736cbf7SLai Jiangshan * @worker: worker to be attached 18344736cbf7SLai Jiangshan * @pool: the target pool 18354736cbf7SLai Jiangshan * 18364736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18374736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18384736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18394736cbf7SLai Jiangshan */ 18404736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18414736cbf7SLai Jiangshan struct worker_pool *pool) 18424736cbf7SLai Jiangshan { 18431258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18444736cbf7SLai Jiangshan 18454736cbf7SLai Jiangshan /* 18464736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 18474736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 18484736cbf7SLai Jiangshan */ 18494736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 18504736cbf7SLai Jiangshan 18514736cbf7SLai Jiangshan /* 18521258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18531258fae7STejun Heo * stable across this function. See the comments above the flag 18541258fae7STejun Heo * definition for details. 18554736cbf7SLai Jiangshan */ 18564736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 18574736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 18584736cbf7SLai Jiangshan 18594736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1860a2d812a2STejun Heo worker->pool = pool; 18614736cbf7SLai Jiangshan 18621258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 18634736cbf7SLai Jiangshan } 18644736cbf7SLai Jiangshan 18654736cbf7SLai Jiangshan /** 186660f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 186760f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 186860f5a4bcSLai Jiangshan * 18694736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 18704736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 18714736cbf7SLai Jiangshan * other reference to the pool. 187260f5a4bcSLai Jiangshan */ 1873a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 187460f5a4bcSLai Jiangshan { 1875a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 187660f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 187760f5a4bcSLai Jiangshan 18781258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1879a2d812a2STejun Heo 1880da028469SLai Jiangshan list_del(&worker->node); 1881a2d812a2STejun Heo worker->pool = NULL; 1882a2d812a2STejun Heo 1883da028469SLai Jiangshan if (list_empty(&pool->workers)) 188460f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 18851258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 188660f5a4bcSLai Jiangshan 1887b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1888b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1889b62c0751SLai Jiangshan 189060f5a4bcSLai Jiangshan if (detach_completion) 189160f5a4bcSLai Jiangshan complete(detach_completion); 189260f5a4bcSLai Jiangshan } 189360f5a4bcSLai Jiangshan 189460f5a4bcSLai Jiangshan /** 1895c34056a3STejun Heo * create_worker - create a new workqueue worker 189663d95a91STejun Heo * @pool: pool the new worker will belong to 1897c34056a3STejun Heo * 1898051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1899c34056a3STejun Heo * 1900c34056a3STejun Heo * CONTEXT: 1901c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1902c34056a3STejun Heo * 1903d185af30SYacine Belkadi * Return: 1904c34056a3STejun Heo * Pointer to the newly created worker. 1905c34056a3STejun Heo */ 1906bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1907c34056a3STejun Heo { 1908c34056a3STejun Heo struct worker *worker = NULL; 1909f3421797STejun Heo int id = -1; 1910e3c916a4STejun Heo char id_buf[16]; 1911c34056a3STejun Heo 19127cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 19137cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1914822d8405STejun Heo if (id < 0) 1915c34056a3STejun Heo goto fail; 1916c34056a3STejun Heo 1917f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1918c34056a3STejun Heo if (!worker) 1919c34056a3STejun Heo goto fail; 1920c34056a3STejun Heo 1921c34056a3STejun Heo worker->id = id; 1922c34056a3STejun Heo 192329c91e99STejun Heo if (pool->cpu >= 0) 1924e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1925e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1926f3421797STejun Heo else 1927e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1928e3c916a4STejun Heo 1929f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1930e3c916a4STejun Heo "kworker/%s", id_buf); 1931c34056a3STejun Heo if (IS_ERR(worker->task)) 1932c34056a3STejun Heo goto fail; 1933c34056a3STejun Heo 193491151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 193525834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 193691151228SOleg Nesterov 1937da028469SLai Jiangshan /* successful, attach the worker to the pool */ 19384736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1939822d8405STejun Heo 1940051e1850SLai Jiangshan /* start the newly created worker */ 1941051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1942051e1850SLai Jiangshan worker->pool->nr_workers++; 1943051e1850SLai Jiangshan worker_enter_idle(worker); 1944051e1850SLai Jiangshan wake_up_process(worker->task); 1945051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1946051e1850SLai Jiangshan 1947c34056a3STejun Heo return worker; 1948822d8405STejun Heo 1949c34056a3STejun Heo fail: 19509625ab17SLai Jiangshan if (id >= 0) 19517cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1952c34056a3STejun Heo kfree(worker); 1953c34056a3STejun Heo return NULL; 1954c34056a3STejun Heo } 1955c34056a3STejun Heo 1956c34056a3STejun Heo /** 1957c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1958c34056a3STejun Heo * @worker: worker to be destroyed 1959c34056a3STejun Heo * 196073eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 196173eb7fe7SLai Jiangshan * be idle. 1962c8e55f36STejun Heo * 1963c8e55f36STejun Heo * CONTEXT: 196460f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1965c34056a3STejun Heo */ 1966c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1967c34056a3STejun Heo { 1968bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1969c34056a3STejun Heo 1970cd549687STejun Heo lockdep_assert_held(&pool->lock); 1971cd549687STejun Heo 1972c34056a3STejun Heo /* sanity check frenzy */ 19736183c009STejun Heo if (WARN_ON(worker->current_work) || 197473eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 197573eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 19766183c009STejun Heo return; 1977c34056a3STejun Heo 1978bd7bdd43STejun Heo pool->nr_workers--; 1979bd7bdd43STejun Heo pool->nr_idle--; 1980c8e55f36STejun Heo 1981c8e55f36STejun Heo list_del_init(&worker->entry); 1982cb444766STejun Heo worker->flags |= WORKER_DIE; 198360f5a4bcSLai Jiangshan wake_up_process(worker->task); 1984c34056a3STejun Heo } 1985c34056a3STejun Heo 198632a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 1987e22bee78STejun Heo { 198832a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 1989e22bee78STejun Heo 1990d565ed63STejun Heo spin_lock_irq(&pool->lock); 1991e22bee78STejun Heo 19923347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1993e22bee78STejun Heo struct worker *worker; 1994e22bee78STejun Heo unsigned long expires; 1995e22bee78STejun Heo 1996e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 199763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1998e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1999e22bee78STejun Heo 20003347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 200163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 20023347fc9fSLai Jiangshan break; 2003e22bee78STejun Heo } 20043347fc9fSLai Jiangshan 20053347fc9fSLai Jiangshan destroy_worker(worker); 2006e22bee78STejun Heo } 2007e22bee78STejun Heo 2008d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2009e22bee78STejun Heo } 2010e22bee78STejun Heo 2011493a1724STejun Heo static void send_mayday(struct work_struct *work) 2012e22bee78STejun Heo { 2013112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2014112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2015493a1724STejun Heo 20162e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2017e22bee78STejun Heo 2018493008a8STejun Heo if (!wq->rescuer) 2019493a1724STejun Heo return; 2020e22bee78STejun Heo 2021e22bee78STejun Heo /* mayday mayday mayday */ 2022493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 202377668c8bSLai Jiangshan /* 202477668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 202577668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 202677668c8bSLai Jiangshan * rescuer is done with it. 202777668c8bSLai Jiangshan */ 202877668c8bSLai Jiangshan get_pwq(pwq); 2029493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2030e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2031493a1724STejun Heo } 2032e22bee78STejun Heo } 2033e22bee78STejun Heo 203432a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2035e22bee78STejun Heo { 203632a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2037e22bee78STejun Heo struct work_struct *work; 2038e22bee78STejun Heo 2039b2d82909STejun Heo spin_lock_irq(&pool->lock); 2040b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2041e22bee78STejun Heo 204263d95a91STejun Heo if (need_to_create_worker(pool)) { 2043e22bee78STejun Heo /* 2044e22bee78STejun Heo * We've been trying to create a new worker but 2045e22bee78STejun Heo * haven't been successful. We might be hitting an 2046e22bee78STejun Heo * allocation deadlock. Send distress signals to 2047e22bee78STejun Heo * rescuers. 2048e22bee78STejun Heo */ 204963d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2050e22bee78STejun Heo send_mayday(work); 2051e22bee78STejun Heo } 2052e22bee78STejun Heo 2053b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 2054b2d82909STejun Heo spin_unlock_irq(&pool->lock); 2055e22bee78STejun Heo 205663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2057e22bee78STejun Heo } 2058e22bee78STejun Heo 2059e22bee78STejun Heo /** 2060e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 206163d95a91STejun Heo * @pool: pool to create a new worker for 2062e22bee78STejun Heo * 206363d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2064e22bee78STejun Heo * have at least one idle worker on return from this function. If 2065e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 206663d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2067e22bee78STejun Heo * possible allocation deadlock. 2068e22bee78STejun Heo * 2069c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2070c5aa87bbSTejun Heo * may_start_working() %true. 2071e22bee78STejun Heo * 2072e22bee78STejun Heo * LOCKING: 2073d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2074e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2075e22bee78STejun Heo * manager. 2076e22bee78STejun Heo */ 207729187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2078d565ed63STejun Heo __releases(&pool->lock) 2079d565ed63STejun Heo __acquires(&pool->lock) 2080e22bee78STejun Heo { 2081e22bee78STejun Heo restart: 2082d565ed63STejun Heo spin_unlock_irq(&pool->lock); 20839f9c2364STejun Heo 2084e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 208563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2086e22bee78STejun Heo 2087e22bee78STejun Heo while (true) { 2088051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2089e22bee78STejun Heo break; 2090e22bee78STejun Heo 2091e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 20929f9c2364STejun Heo 209363d95a91STejun Heo if (!need_to_create_worker(pool)) 2094e22bee78STejun Heo break; 2095e22bee78STejun Heo } 2096e22bee78STejun Heo 209763d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2098d565ed63STejun Heo spin_lock_irq(&pool->lock); 2099051e1850SLai Jiangshan /* 2100051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2101051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2102051e1850SLai Jiangshan * already become busy. 2103051e1850SLai Jiangshan */ 210463d95a91STejun Heo if (need_to_create_worker(pool)) 2105e22bee78STejun Heo goto restart; 2106e22bee78STejun Heo } 2107e22bee78STejun Heo 2108e22bee78STejun Heo /** 2109e22bee78STejun Heo * manage_workers - manage worker pool 2110e22bee78STejun Heo * @worker: self 2111e22bee78STejun Heo * 2112706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2113e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2114706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2115e22bee78STejun Heo * 2116e22bee78STejun Heo * The caller can safely start processing works on false return. On 2117e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2118e22bee78STejun Heo * and may_start_working() is true. 2119e22bee78STejun Heo * 2120e22bee78STejun Heo * CONTEXT: 2121d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2122e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2123e22bee78STejun Heo * 2124d185af30SYacine Belkadi * Return: 212529187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 212629187a9eSTejun Heo * start processing works, %true if management function was performed and 212729187a9eSTejun Heo * the conditions that the caller verified before calling the function may 212829187a9eSTejun Heo * no longer be true. 2129e22bee78STejun Heo */ 2130e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2131e22bee78STejun Heo { 213263d95a91STejun Heo struct worker_pool *pool = worker->pool; 2133e22bee78STejun Heo 2134692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 213529187a9eSTejun Heo return false; 2136692b4825STejun Heo 2137692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 21382607d7a6STejun Heo pool->manager = worker; 2139e22bee78STejun Heo 214029187a9eSTejun Heo maybe_create_worker(pool); 2141e22bee78STejun Heo 21422607d7a6STejun Heo pool->manager = NULL; 2143692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2144692b4825STejun Heo wake_up(&wq_manager_wait); 214529187a9eSTejun Heo return true; 2146e22bee78STejun Heo } 2147e22bee78STejun Heo 2148a62428c0STejun Heo /** 2149a62428c0STejun Heo * process_one_work - process single work 2150c34056a3STejun Heo * @worker: self 2151a62428c0STejun Heo * @work: work to process 2152a62428c0STejun Heo * 2153a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2154a62428c0STejun Heo * process a single work including synchronization against and 2155a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2156a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2157a62428c0STejun Heo * call this function to process a work. 2158a62428c0STejun Heo * 2159a62428c0STejun Heo * CONTEXT: 2160d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2161a62428c0STejun Heo */ 2162c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2163d565ed63STejun Heo __releases(&pool->lock) 2164d565ed63STejun Heo __acquires(&pool->lock) 21651da177e4SLinus Torvalds { 2166112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2167bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2168112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 216973f53c4aSTejun Heo int work_color; 21707e11629dSTejun Heo struct worker *collision; 21714e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21724e6045f1SJohannes Berg /* 2173a62428c0STejun Heo * It is permissible to free the struct work_struct from 2174a62428c0STejun Heo * inside the function that is called from it, this we need to 2175a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2176a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2177a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21784e6045f1SJohannes Berg */ 21794d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21804d82a1deSPeter Zijlstra 21814d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21824e6045f1SJohannes Berg #endif 2183807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 218485327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2185ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 218625511a47STejun Heo 21877e11629dSTejun Heo /* 21887e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21897e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21907e11629dSTejun Heo * already processing the work. If so, defer the work to the 21917e11629dSTejun Heo * currently executing one. 21927e11629dSTejun Heo */ 2193c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21947e11629dSTejun Heo if (unlikely(collision)) { 21957e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21967e11629dSTejun Heo return; 21977e11629dSTejun Heo } 21981da177e4SLinus Torvalds 21998930cabaSTejun Heo /* claim and dequeue */ 22001da177e4SLinus Torvalds debug_work_deactivate(work); 2201c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2202c34056a3STejun Heo worker->current_work = work; 2203a2c1c57bSTejun Heo worker->current_func = work->func; 2204112202d9STejun Heo worker->current_pwq = pwq; 220573f53c4aSTejun Heo work_color = get_work_color(work); 22067a22ad75STejun Heo 22078bf89593STejun Heo /* 22088bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 22098bf89593STejun Heo * overridden through set_worker_desc(). 22108bf89593STejun Heo */ 22118bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 22128bf89593STejun Heo 2213a62428c0STejun Heo list_del_init(&work->entry); 2214a62428c0STejun Heo 2215649027d7STejun Heo /* 2216228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2217228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2218228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2219228f1d00SLai Jiangshan * execution of the pending work items. 2220fb0e7bebSTejun Heo */ 2221fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2222228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2223fb0e7bebSTejun Heo 2224974271c4STejun Heo /* 2225a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2226a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2227a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2228a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2229228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2230974271c4STejun Heo */ 2231a489a03eSLai Jiangshan if (need_more_worker(pool)) 223263d95a91STejun Heo wake_up_worker(pool); 2233974271c4STejun Heo 22348930cabaSTejun Heo /* 22357c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2236d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 223723657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 223823657bb1STejun Heo * disabled. 22398930cabaSTejun Heo */ 22407c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22411da177e4SLinus Torvalds 2242d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2243365970a1SDavid Howells 2244a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 22453295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2246e6f3faa7SPeter Zijlstra /* 2247f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2248f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2249e6f3faa7SPeter Zijlstra * 2250e6f3faa7SPeter Zijlstra * However, that would result in: 2251e6f3faa7SPeter Zijlstra * 2252e6f3faa7SPeter Zijlstra * A(W1) 2253e6f3faa7SPeter Zijlstra * WFC(C) 2254e6f3faa7SPeter Zijlstra * A(W1) 2255e6f3faa7SPeter Zijlstra * C(C) 2256e6f3faa7SPeter Zijlstra * 2257e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2258e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2259e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2260f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2261e6f3faa7SPeter Zijlstra * these locks. 2262e6f3faa7SPeter Zijlstra * 2263e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2264e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2265e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2266e6f3faa7SPeter Zijlstra */ 2267f52be570SPeter Zijlstra lockdep_invariant_state(true); 2268e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2269a2c1c57bSTejun Heo worker->current_func(work); 2270e36c886aSArjan van de Ven /* 2271e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2272e36c886aSArjan van de Ven * point will only record its address. 2273e36c886aSArjan van de Ven */ 2274e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22753295f0efSIngo Molnar lock_map_release(&lockdep_map); 2276112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22771da177e4SLinus Torvalds 2278d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2279044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2280044c782cSValentin Ilie " last function: %pf\n", 2281a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2282a2c1c57bSTejun Heo worker->current_func); 2283d5abe669SPeter Zijlstra debug_show_held_locks(current); 2284d5abe669SPeter Zijlstra dump_stack(); 2285d5abe669SPeter Zijlstra } 2286d5abe669SPeter Zijlstra 2287b22ce278STejun Heo /* 2288b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2289b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2290b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2291b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2292789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2293789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2294b22ce278STejun Heo */ 2295a7e6425eSPaul E. McKenney cond_resched(); 2296b22ce278STejun Heo 2297d565ed63STejun Heo spin_lock_irq(&pool->lock); 2298a62428c0STejun Heo 2299fb0e7bebSTejun Heo /* clear cpu intensive status */ 2300fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2301fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2302fb0e7bebSTejun Heo 23031b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 23041b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 23051b69ac6bSJohannes Weiner 2306a62428c0STejun Heo /* we're done with it, release */ 230742f8570fSSasha Levin hash_del(&worker->hentry); 2308c34056a3STejun Heo worker->current_work = NULL; 2309a2c1c57bSTejun Heo worker->current_func = NULL; 2310112202d9STejun Heo worker->current_pwq = NULL; 2311112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 23121da177e4SLinus Torvalds } 23131da177e4SLinus Torvalds 2314affee4b2STejun Heo /** 2315affee4b2STejun Heo * process_scheduled_works - process scheduled works 2316affee4b2STejun Heo * @worker: self 2317affee4b2STejun Heo * 2318affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2319affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2320affee4b2STejun Heo * fetches a work from the top and executes it. 2321affee4b2STejun Heo * 2322affee4b2STejun Heo * CONTEXT: 2323d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2324affee4b2STejun Heo * multiple times. 2325affee4b2STejun Heo */ 2326affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 23271da177e4SLinus Torvalds { 2328affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2329affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2330a62428c0STejun Heo struct work_struct, entry); 2331c34056a3STejun Heo process_one_work(worker, work); 2332a62428c0STejun Heo } 23331da177e4SLinus Torvalds } 23341da177e4SLinus Torvalds 2335197f6accSTejun Heo static void set_pf_worker(bool val) 2336197f6accSTejun Heo { 2337197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2338197f6accSTejun Heo if (val) 2339197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2340197f6accSTejun Heo else 2341197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2342197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2343197f6accSTejun Heo } 2344197f6accSTejun Heo 23454690c4abSTejun Heo /** 23464690c4abSTejun Heo * worker_thread - the worker thread function 2347c34056a3STejun Heo * @__worker: self 23484690c4abSTejun Heo * 2349c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2350c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2351c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2352c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2353c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2354d185af30SYacine Belkadi * 2355d185af30SYacine Belkadi * Return: 0 23564690c4abSTejun Heo */ 2357c34056a3STejun Heo static int worker_thread(void *__worker) 23581da177e4SLinus Torvalds { 2359c34056a3STejun Heo struct worker *worker = __worker; 2360bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 23611da177e4SLinus Torvalds 2362e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2363197f6accSTejun Heo set_pf_worker(true); 2364c8e55f36STejun Heo woke_up: 2365d565ed63STejun Heo spin_lock_irq(&pool->lock); 2366affee4b2STejun Heo 2367a9ab775bSTejun Heo /* am I supposed to die? */ 2368a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2369d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2370a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2371197f6accSTejun Heo set_pf_worker(false); 237260f5a4bcSLai Jiangshan 237360f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 23747cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 2375a2d812a2STejun Heo worker_detach_from_pool(worker); 237660f5a4bcSLai Jiangshan kfree(worker); 2377c8e55f36STejun Heo return 0; 2378c8e55f36STejun Heo } 2379c8e55f36STejun Heo 2380c8e55f36STejun Heo worker_leave_idle(worker); 2381db7bccf4STejun Heo recheck: 2382e22bee78STejun Heo /* no more worker necessary? */ 238363d95a91STejun Heo if (!need_more_worker(pool)) 2384e22bee78STejun Heo goto sleep; 2385e22bee78STejun Heo 2386e22bee78STejun Heo /* do we need to manage? */ 238763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2388e22bee78STejun Heo goto recheck; 2389e22bee78STejun Heo 2390c8e55f36STejun Heo /* 2391c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2392c8e55f36STejun Heo * preparing to process a work or actually processing it. 2393c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2394c8e55f36STejun Heo */ 23956183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2396c8e55f36STejun Heo 2397e22bee78STejun Heo /* 2398a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2399a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2400a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2401a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2402a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2403e22bee78STejun Heo */ 2404a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2405e22bee78STejun Heo 2406e22bee78STejun Heo do { 2407affee4b2STejun Heo struct work_struct *work = 2408bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2409affee4b2STejun Heo struct work_struct, entry); 2410affee4b2STejun Heo 241182607adcSTejun Heo pool->watchdog_ts = jiffies; 241282607adcSTejun Heo 2413c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2414affee4b2STejun Heo /* optimization path, not strictly necessary */ 2415affee4b2STejun Heo process_one_work(worker, work); 2416affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2417affee4b2STejun Heo process_scheduled_works(worker); 2418affee4b2STejun Heo } else { 2419c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2420affee4b2STejun Heo process_scheduled_works(worker); 2421affee4b2STejun Heo } 242263d95a91STejun Heo } while (keep_working(pool)); 2423affee4b2STejun Heo 2424228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2425d313dd85STejun Heo sleep: 2426c8e55f36STejun Heo /* 2427d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2428d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2429d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2430d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2431d565ed63STejun Heo * event. 2432c8e55f36STejun Heo */ 2433c8e55f36STejun Heo worker_enter_idle(worker); 2434c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2435d565ed63STejun Heo spin_unlock_irq(&pool->lock); 24361da177e4SLinus Torvalds schedule(); 2437c8e55f36STejun Heo goto woke_up; 24381da177e4SLinus Torvalds } 24391da177e4SLinus Torvalds 2440e22bee78STejun Heo /** 2441e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2442111c225aSTejun Heo * @__rescuer: self 2443e22bee78STejun Heo * 2444e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2445493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2446e22bee78STejun Heo * 2447706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2448e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2449e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2450e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2451e22bee78STejun Heo * the problem rescuer solves. 2452e22bee78STejun Heo * 2453706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2454706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2455e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2456e22bee78STejun Heo * 2457e22bee78STejun Heo * This should happen rarely. 2458d185af30SYacine Belkadi * 2459d185af30SYacine Belkadi * Return: 0 2460e22bee78STejun Heo */ 2461111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2462e22bee78STejun Heo { 2463111c225aSTejun Heo struct worker *rescuer = __rescuer; 2464111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2465e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 24664d595b86SLai Jiangshan bool should_stop; 2467e22bee78STejun Heo 2468e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2469111c225aSTejun Heo 2470111c225aSTejun Heo /* 2471111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2472111c225aSTejun Heo * doesn't participate in concurrency management. 2473111c225aSTejun Heo */ 2474197f6accSTejun Heo set_pf_worker(true); 2475e22bee78STejun Heo repeat: 2476c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 24771da177e4SLinus Torvalds 24784d595b86SLai Jiangshan /* 24794d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 24804d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 24814d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 24824d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 24834d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 24844d595b86SLai Jiangshan * list is always empty on exit. 24854d595b86SLai Jiangshan */ 24864d595b86SLai Jiangshan should_stop = kthread_should_stop(); 24871da177e4SLinus Torvalds 2488493a1724STejun Heo /* see whether any pwq is asking for help */ 24892e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2490493a1724STejun Heo 2491493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2492493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2493493a1724STejun Heo struct pool_workqueue, mayday_node); 2494112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2495e22bee78STejun Heo struct work_struct *work, *n; 249682607adcSTejun Heo bool first = true; 2497e22bee78STejun Heo 2498e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2499493a1724STejun Heo list_del_init(&pwq->mayday_node); 2500493a1724STejun Heo 25012e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2502e22bee78STejun Heo 250351697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 250451697d39SLai Jiangshan 250551697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2506e22bee78STejun Heo 2507e22bee78STejun Heo /* 2508e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2509e22bee78STejun Heo * process'em. 2510e22bee78STejun Heo */ 25110479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 251282607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 251382607adcSTejun Heo if (get_work_pwq(work) == pwq) { 251482607adcSTejun Heo if (first) 251582607adcSTejun Heo pool->watchdog_ts = jiffies; 2516e22bee78STejun Heo move_linked_works(work, scheduled, &n); 251782607adcSTejun Heo } 251882607adcSTejun Heo first = false; 251982607adcSTejun Heo } 2520e22bee78STejun Heo 2521008847f6SNeilBrown if (!list_empty(scheduled)) { 2522e22bee78STejun Heo process_scheduled_works(rescuer); 25237576958aSTejun Heo 25247576958aSTejun Heo /* 2525008847f6SNeilBrown * The above execution of rescued work items could 2526008847f6SNeilBrown * have created more to rescue through 2527008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2528008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2529008847f6SNeilBrown * that such back-to-back work items, which may be 2530008847f6SNeilBrown * being used to relieve memory pressure, don't 2531008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2532008847f6SNeilBrown */ 2533008847f6SNeilBrown if (need_to_create_worker(pool)) { 2534008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2535008847f6SNeilBrown get_pwq(pwq); 2536008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2537008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2538008847f6SNeilBrown } 2539008847f6SNeilBrown } 2540008847f6SNeilBrown 2541008847f6SNeilBrown /* 254277668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 254313b1d625SLai Jiangshan * go away while we're still attached to it. 254477668c8bSLai Jiangshan */ 254577668c8bSLai Jiangshan put_pwq(pwq); 254677668c8bSLai Jiangshan 254777668c8bSLai Jiangshan /* 2548d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 25497576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 25507576958aSTejun Heo * and stalling the execution. 25517576958aSTejun Heo */ 2552d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 255363d95a91STejun Heo wake_up_worker(pool); 25547576958aSTejun Heo 255513b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 255613b1d625SLai Jiangshan 2557a2d812a2STejun Heo worker_detach_from_pool(rescuer); 255813b1d625SLai Jiangshan 255913b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 25601da177e4SLinus Torvalds } 25611da177e4SLinus Torvalds 25622e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2563493a1724STejun Heo 25644d595b86SLai Jiangshan if (should_stop) { 25654d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2566197f6accSTejun Heo set_pf_worker(false); 25674d595b86SLai Jiangshan return 0; 25684d595b86SLai Jiangshan } 25694d595b86SLai Jiangshan 2570111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2571111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2572e22bee78STejun Heo schedule(); 2573e22bee78STejun Heo goto repeat; 25741da177e4SLinus Torvalds } 25751da177e4SLinus Torvalds 2576fca839c0STejun Heo /** 2577fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2578fca839c0STejun Heo * @target_wq: workqueue being flushed 2579fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2580fca839c0STejun Heo * 2581fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2582fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2583fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2584fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2585fca839c0STejun Heo * a deadlock. 2586fca839c0STejun Heo */ 2587fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2588fca839c0STejun Heo struct work_struct *target_work) 2589fca839c0STejun Heo { 2590fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2591fca839c0STejun Heo struct worker *worker; 2592fca839c0STejun Heo 2593fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2594fca839c0STejun Heo return; 2595fca839c0STejun Heo 2596fca839c0STejun Heo worker = current_wq_worker(); 2597fca839c0STejun Heo 2598fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2599fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2600fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 260123d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 260223d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2603fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2604fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2605fca839c0STejun Heo target_wq->name, target_func); 2606fca839c0STejun Heo } 2607fca839c0STejun Heo 2608fc2e4d70SOleg Nesterov struct wq_barrier { 2609fc2e4d70SOleg Nesterov struct work_struct work; 2610fc2e4d70SOleg Nesterov struct completion done; 26112607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2612fc2e4d70SOleg Nesterov }; 2613fc2e4d70SOleg Nesterov 2614fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2615fc2e4d70SOleg Nesterov { 2616fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2617fc2e4d70SOleg Nesterov complete(&barr->done); 2618fc2e4d70SOleg Nesterov } 2619fc2e4d70SOleg Nesterov 26204690c4abSTejun Heo /** 26214690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2622112202d9STejun Heo * @pwq: pwq to insert barrier into 26234690c4abSTejun Heo * @barr: wq_barrier to insert 2624affee4b2STejun Heo * @target: target work to attach @barr to 2625affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 26264690c4abSTejun Heo * 2627affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2628affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2629affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2630affee4b2STejun Heo * cpu. 2631affee4b2STejun Heo * 2632affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2633affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2634affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2635affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2636affee4b2STejun Heo * after a work with LINKED flag set. 2637affee4b2STejun Heo * 2638affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2639112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 26404690c4abSTejun Heo * 26414690c4abSTejun Heo * CONTEXT: 2642d565ed63STejun Heo * spin_lock_irq(pool->lock). 26434690c4abSTejun Heo */ 2644112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2645affee4b2STejun Heo struct wq_barrier *barr, 2646affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2647fc2e4d70SOleg Nesterov { 2648affee4b2STejun Heo struct list_head *head; 2649affee4b2STejun Heo unsigned int linked = 0; 2650affee4b2STejun Heo 2651dc186ad7SThomas Gleixner /* 2652d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2653dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2654dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2655dc186ad7SThomas Gleixner * might deadlock. 2656dc186ad7SThomas Gleixner */ 2657ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 265822df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 265952fa5bc5SBoqun Feng 2660fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2661fd1a5b04SByungchul Park 26622607d7a6STejun Heo barr->task = current; 266383c22520SOleg Nesterov 2664affee4b2STejun Heo /* 2665affee4b2STejun Heo * If @target is currently being executed, schedule the 2666affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2667affee4b2STejun Heo */ 2668affee4b2STejun Heo if (worker) 2669affee4b2STejun Heo head = worker->scheduled.next; 2670affee4b2STejun Heo else { 2671affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2672affee4b2STejun Heo 2673affee4b2STejun Heo head = target->entry.next; 2674affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2675affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2676affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2677affee4b2STejun Heo } 2678affee4b2STejun Heo 2679dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2680112202d9STejun Heo insert_work(pwq, &barr->work, head, 2681affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2682fc2e4d70SOleg Nesterov } 2683fc2e4d70SOleg Nesterov 268473f53c4aSTejun Heo /** 2685112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 268673f53c4aSTejun Heo * @wq: workqueue being flushed 268773f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 268873f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 268973f53c4aSTejun Heo * 2690112202d9STejun Heo * Prepare pwqs for workqueue flushing. 269173f53c4aSTejun Heo * 2692112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2693112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2694112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2695112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2696112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 269773f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 269873f53c4aSTejun Heo * 269973f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 270073f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 270173f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 270273f53c4aSTejun Heo * is returned. 270373f53c4aSTejun Heo * 2704112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 270573f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 270673f53c4aSTejun Heo * advanced to @work_color. 270773f53c4aSTejun Heo * 270873f53c4aSTejun Heo * CONTEXT: 27093c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 271073f53c4aSTejun Heo * 2711d185af30SYacine Belkadi * Return: 271273f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 271373f53c4aSTejun Heo * otherwise. 271473f53c4aSTejun Heo */ 2715112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 271673f53c4aSTejun Heo int flush_color, int work_color) 27171da177e4SLinus Torvalds { 271873f53c4aSTejun Heo bool wait = false; 271949e3cf44STejun Heo struct pool_workqueue *pwq; 27201da177e4SLinus Torvalds 272173f53c4aSTejun Heo if (flush_color >= 0) { 27226183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2723112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2724dc186ad7SThomas Gleixner } 272514441960SOleg Nesterov 272649e3cf44STejun Heo for_each_pwq(pwq, wq) { 2727112202d9STejun Heo struct worker_pool *pool = pwq->pool; 27281da177e4SLinus Torvalds 2729b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 273073f53c4aSTejun Heo 273173f53c4aSTejun Heo if (flush_color >= 0) { 27326183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 273373f53c4aSTejun Heo 2734112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2735112202d9STejun Heo pwq->flush_color = flush_color; 2736112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 273773f53c4aSTejun Heo wait = true; 27381da177e4SLinus Torvalds } 273973f53c4aSTejun Heo } 274073f53c4aSTejun Heo 274173f53c4aSTejun Heo if (work_color >= 0) { 27426183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2743112202d9STejun Heo pwq->work_color = work_color; 274473f53c4aSTejun Heo } 274573f53c4aSTejun Heo 2746b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 27471da177e4SLinus Torvalds } 27481da177e4SLinus Torvalds 2749112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 275073f53c4aSTejun Heo complete(&wq->first_flusher->done); 275173f53c4aSTejun Heo 275273f53c4aSTejun Heo return wait; 275383c22520SOleg Nesterov } 27541da177e4SLinus Torvalds 27550fcb78c2SRolf Eike Beer /** 27561da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 27570fcb78c2SRolf Eike Beer * @wq: workqueue to flush 27581da177e4SLinus Torvalds * 2759c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2760c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 27611da177e4SLinus Torvalds */ 27627ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 27631da177e4SLinus Torvalds { 276473f53c4aSTejun Heo struct wq_flusher this_flusher = { 276573f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 276673f53c4aSTejun Heo .flush_color = -1, 2767fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 276873f53c4aSTejun Heo }; 276973f53c4aSTejun Heo int next_color; 2770b1f4ec17SOleg Nesterov 27713347fa09STejun Heo if (WARN_ON(!wq_online)) 27723347fa09STejun Heo return; 27733347fa09STejun Heo 277487915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 277587915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 277687915adcSJohannes Berg 27773c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 277873f53c4aSTejun Heo 277973f53c4aSTejun Heo /* 278073f53c4aSTejun Heo * Start-to-wait phase 278173f53c4aSTejun Heo */ 278273f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 278373f53c4aSTejun Heo 278473f53c4aSTejun Heo if (next_color != wq->flush_color) { 278573f53c4aSTejun Heo /* 278673f53c4aSTejun Heo * Color space is not full. The current work_color 278773f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 278873f53c4aSTejun Heo * by one. 278973f53c4aSTejun Heo */ 27906183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 279173f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 279273f53c4aSTejun Heo wq->work_color = next_color; 279373f53c4aSTejun Heo 279473f53c4aSTejun Heo if (!wq->first_flusher) { 279573f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 27966183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 279773f53c4aSTejun Heo 279873f53c4aSTejun Heo wq->first_flusher = &this_flusher; 279973f53c4aSTejun Heo 2800112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 280173f53c4aSTejun Heo wq->work_color)) { 280273f53c4aSTejun Heo /* nothing to flush, done */ 280373f53c4aSTejun Heo wq->flush_color = next_color; 280473f53c4aSTejun Heo wq->first_flusher = NULL; 280573f53c4aSTejun Heo goto out_unlock; 280673f53c4aSTejun Heo } 280773f53c4aSTejun Heo } else { 280873f53c4aSTejun Heo /* wait in queue */ 28096183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 281073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2811112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 281273f53c4aSTejun Heo } 281373f53c4aSTejun Heo } else { 281473f53c4aSTejun Heo /* 281573f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 281673f53c4aSTejun Heo * The next flush completion will assign us 281773f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 281873f53c4aSTejun Heo */ 281973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 282073f53c4aSTejun Heo } 282173f53c4aSTejun Heo 2822fca839c0STejun Heo check_flush_dependency(wq, NULL); 2823fca839c0STejun Heo 28243c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 282573f53c4aSTejun Heo 282673f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 282773f53c4aSTejun Heo 282873f53c4aSTejun Heo /* 282973f53c4aSTejun Heo * Wake-up-and-cascade phase 283073f53c4aSTejun Heo * 283173f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 283273f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 283373f53c4aSTejun Heo */ 283473f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 283573f53c4aSTejun Heo return; 283673f53c4aSTejun Heo 28373c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 283873f53c4aSTejun Heo 28394ce48b37STejun Heo /* we might have raced, check again with mutex held */ 28404ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 28414ce48b37STejun Heo goto out_unlock; 28424ce48b37STejun Heo 284373f53c4aSTejun Heo wq->first_flusher = NULL; 284473f53c4aSTejun Heo 28456183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 28466183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 284773f53c4aSTejun Heo 284873f53c4aSTejun Heo while (true) { 284973f53c4aSTejun Heo struct wq_flusher *next, *tmp; 285073f53c4aSTejun Heo 285173f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 285273f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 285373f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 285473f53c4aSTejun Heo break; 285573f53c4aSTejun Heo list_del_init(&next->list); 285673f53c4aSTejun Heo complete(&next->done); 285773f53c4aSTejun Heo } 285873f53c4aSTejun Heo 28596183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 286073f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 286173f53c4aSTejun Heo 286273f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 286373f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 286473f53c4aSTejun Heo 286573f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 286673f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 286773f53c4aSTejun Heo /* 286873f53c4aSTejun Heo * Assign the same color to all overflowed 286973f53c4aSTejun Heo * flushers, advance work_color and append to 287073f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 287173f53c4aSTejun Heo * phase for these overflowed flushers. 287273f53c4aSTejun Heo */ 287373f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 287473f53c4aSTejun Heo tmp->flush_color = wq->work_color; 287573f53c4aSTejun Heo 287673f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 287773f53c4aSTejun Heo 287873f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 287973f53c4aSTejun Heo &wq->flusher_queue); 2880112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 288173f53c4aSTejun Heo } 288273f53c4aSTejun Heo 288373f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 28846183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 288573f53c4aSTejun Heo break; 288673f53c4aSTejun Heo } 288773f53c4aSTejun Heo 288873f53c4aSTejun Heo /* 288973f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2890112202d9STejun Heo * the new first flusher and arm pwqs. 289173f53c4aSTejun Heo */ 28926183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 28936183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 289473f53c4aSTejun Heo 289573f53c4aSTejun Heo list_del_init(&next->list); 289673f53c4aSTejun Heo wq->first_flusher = next; 289773f53c4aSTejun Heo 2898112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 289973f53c4aSTejun Heo break; 290073f53c4aSTejun Heo 290173f53c4aSTejun Heo /* 290273f53c4aSTejun Heo * Meh... this color is already done, clear first 290373f53c4aSTejun Heo * flusher and repeat cascading. 290473f53c4aSTejun Heo */ 290573f53c4aSTejun Heo wq->first_flusher = NULL; 290673f53c4aSTejun Heo } 290773f53c4aSTejun Heo 290873f53c4aSTejun Heo out_unlock: 29093c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 29101da177e4SLinus Torvalds } 29111dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 29121da177e4SLinus Torvalds 29139c5a2ba7STejun Heo /** 29149c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 29159c5a2ba7STejun Heo * @wq: workqueue to drain 29169c5a2ba7STejun Heo * 29179c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 29189c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 29199c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2920b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 29219c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 29229c5a2ba7STejun Heo * takes too long. 29239c5a2ba7STejun Heo */ 29249c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 29259c5a2ba7STejun Heo { 29269c5a2ba7STejun Heo unsigned int flush_cnt = 0; 292749e3cf44STejun Heo struct pool_workqueue *pwq; 29289c5a2ba7STejun Heo 29299c5a2ba7STejun Heo /* 29309c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 29319c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2932618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 29339c5a2ba7STejun Heo */ 293487fc741eSLai Jiangshan mutex_lock(&wq->mutex); 29359c5a2ba7STejun Heo if (!wq->nr_drainers++) 2936618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 293787fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29389c5a2ba7STejun Heo reflush: 29399c5a2ba7STejun Heo flush_workqueue(wq); 29409c5a2ba7STejun Heo 2941b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 294276af4d93STejun Heo 294349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2944fa2563e4SThomas Tuttle bool drained; 29459c5a2ba7STejun Heo 2946b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2947112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2948b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2949fa2563e4SThomas Tuttle 2950fa2563e4SThomas Tuttle if (drained) 29519c5a2ba7STejun Heo continue; 29529c5a2ba7STejun Heo 29539c5a2ba7STejun Heo if (++flush_cnt == 10 || 29549c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2955c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 29569c5a2ba7STejun Heo wq->name, flush_cnt); 295776af4d93STejun Heo 2958b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 29599c5a2ba7STejun Heo goto reflush; 29609c5a2ba7STejun Heo } 29619c5a2ba7STejun Heo 29629c5a2ba7STejun Heo if (!--wq->nr_drainers) 2963618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 296487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29659c5a2ba7STejun Heo } 29669c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 29679c5a2ba7STejun Heo 2968d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2969d6e89786SJohannes Berg bool from_cancel) 2970baf59022STejun Heo { 2971baf59022STejun Heo struct worker *worker = NULL; 2972c9e7cf27STejun Heo struct worker_pool *pool; 2973112202d9STejun Heo struct pool_workqueue *pwq; 2974baf59022STejun Heo 2975baf59022STejun Heo might_sleep(); 2976baf59022STejun Heo 2977fa1b54e6STejun Heo local_irq_disable(); 2978fa1b54e6STejun Heo pool = get_work_pool(work); 2979fa1b54e6STejun Heo if (!pool) { 2980fa1b54e6STejun Heo local_irq_enable(); 2981fa1b54e6STejun Heo return false; 2982fa1b54e6STejun Heo } 2983fa1b54e6STejun Heo 2984fa1b54e6STejun Heo spin_lock(&pool->lock); 29850b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2986112202d9STejun Heo pwq = get_work_pwq(work); 2987112202d9STejun Heo if (pwq) { 2988112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2989baf59022STejun Heo goto already_gone; 2990606a5020STejun Heo } else { 2991c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2992baf59022STejun Heo if (!worker) 2993baf59022STejun Heo goto already_gone; 2994112202d9STejun Heo pwq = worker->current_pwq; 2995606a5020STejun Heo } 2996baf59022STejun Heo 2997fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2998fca839c0STejun Heo 2999112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3000d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3001baf59022STejun Heo 3002e159489bSTejun Heo /* 3003a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3004a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3005a1d14934SPeter Zijlstra * 3006a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3007a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3008a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3009a1d14934SPeter Zijlstra * forward progress. 3010e159489bSTejun Heo */ 3011d6e89786SJohannes Berg if (!from_cancel && 3012d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3013112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3014112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3015a1d14934SPeter Zijlstra } 3016e159489bSTejun Heo 3017baf59022STejun Heo return true; 3018baf59022STejun Heo already_gone: 3019d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3020baf59022STejun Heo return false; 3021baf59022STejun Heo } 3022baf59022STejun Heo 3023d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3024d6e89786SJohannes Berg { 3025d6e89786SJohannes Berg struct wq_barrier barr; 3026d6e89786SJohannes Berg 3027d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3028d6e89786SJohannes Berg return false; 3029d6e89786SJohannes Berg 30304d43d395STetsuo Handa if (WARN_ON(!work->func)) 30314d43d395STetsuo Handa return false; 30324d43d395STetsuo Handa 303387915adcSJohannes Berg if (!from_cancel) { 303487915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 303587915adcSJohannes Berg lock_map_release(&work->lockdep_map); 303687915adcSJohannes Berg } 303787915adcSJohannes Berg 3038d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3039d6e89786SJohannes Berg wait_for_completion(&barr.done); 3040d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3041d6e89786SJohannes Berg return true; 3042d6e89786SJohannes Berg } else { 3043d6e89786SJohannes Berg return false; 3044d6e89786SJohannes Berg } 3045d6e89786SJohannes Berg } 3046d6e89786SJohannes Berg 3047db700897SOleg Nesterov /** 3048401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3049401a8d04STejun Heo * @work: the work to flush 3050db700897SOleg Nesterov * 3051606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3052606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3053401a8d04STejun Heo * 3054d185af30SYacine Belkadi * Return: 3055401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3056401a8d04STejun Heo * %false if it was already idle. 3057db700897SOleg Nesterov */ 3058401a8d04STejun Heo bool flush_work(struct work_struct *work) 3059db700897SOleg Nesterov { 3060d6e89786SJohannes Berg return __flush_work(work, false); 3061606a5020STejun Heo } 3062db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3063db700897SOleg Nesterov 30648603e1b3STejun Heo struct cwt_wait { 3065ac6424b9SIngo Molnar wait_queue_entry_t wait; 30668603e1b3STejun Heo struct work_struct *work; 30678603e1b3STejun Heo }; 30688603e1b3STejun Heo 3069ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 30708603e1b3STejun Heo { 30718603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 30728603e1b3STejun Heo 30738603e1b3STejun Heo if (cwait->work != key) 30748603e1b3STejun Heo return 0; 30758603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 30768603e1b3STejun Heo } 30778603e1b3STejun Heo 307836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3079401a8d04STejun Heo { 30808603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3081bbb68dfaSTejun Heo unsigned long flags; 30821f1f642eSOleg Nesterov int ret; 30831f1f642eSOleg Nesterov 30841f1f642eSOleg Nesterov do { 3085bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3086bbb68dfaSTejun Heo /* 30878603e1b3STejun Heo * If someone else is already canceling, wait for it to 30888603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 30898603e1b3STejun Heo * because we may get scheduled between @work's completion 30908603e1b3STejun Heo * and the other canceling task resuming and clearing 30918603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 30928603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 30938603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 30948603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 30958603e1b3STejun Heo * we're hogging the CPU. 30968603e1b3STejun Heo * 30978603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 30988603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 30998603e1b3STejun Heo * wake function which matches @work along with exclusive 31008603e1b3STejun Heo * wait and wakeup. 3101bbb68dfaSTejun Heo */ 31028603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 31038603e1b3STejun Heo struct cwt_wait cwait; 31048603e1b3STejun Heo 31058603e1b3STejun Heo init_wait(&cwait.wait); 31068603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 31078603e1b3STejun Heo cwait.work = work; 31088603e1b3STejun Heo 31098603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 31108603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 31118603e1b3STejun Heo if (work_is_canceling(work)) 31128603e1b3STejun Heo schedule(); 31138603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 31148603e1b3STejun Heo } 31151f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 31161f1f642eSOleg Nesterov 3117bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3118bbb68dfaSTejun Heo mark_work_canceling(work); 3119bbb68dfaSTejun Heo local_irq_restore(flags); 3120bbb68dfaSTejun Heo 31213347fa09STejun Heo /* 31223347fa09STejun Heo * This allows canceling during early boot. We know that @work 31233347fa09STejun Heo * isn't executing. 31243347fa09STejun Heo */ 31253347fa09STejun Heo if (wq_online) 3126d6e89786SJohannes Berg __flush_work(work, true); 31273347fa09STejun Heo 31287a22ad75STejun Heo clear_work_data(work); 31298603e1b3STejun Heo 31308603e1b3STejun Heo /* 31318603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 31328603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 31338603e1b3STejun Heo * visible there. 31348603e1b3STejun Heo */ 31358603e1b3STejun Heo smp_mb(); 31368603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 31378603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 31388603e1b3STejun Heo 31391f1f642eSOleg Nesterov return ret; 31401f1f642eSOleg Nesterov } 31411f1f642eSOleg Nesterov 31426e84d644SOleg Nesterov /** 3143401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3144401a8d04STejun Heo * @work: the work to cancel 31456e84d644SOleg Nesterov * 3146401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3147401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3148401a8d04STejun Heo * another workqueue. On return from this function, @work is 3149401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 31501f1f642eSOleg Nesterov * 3151401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3152401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 31536e84d644SOleg Nesterov * 3154401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 31556e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3156401a8d04STejun Heo * 3157d185af30SYacine Belkadi * Return: 3158401a8d04STejun Heo * %true if @work was pending, %false otherwise. 31596e84d644SOleg Nesterov */ 3160401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 31616e84d644SOleg Nesterov { 316236e227d2STejun Heo return __cancel_work_timer(work, false); 3163b89deed3SOleg Nesterov } 316428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3165b89deed3SOleg Nesterov 31666e84d644SOleg Nesterov /** 3167401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3168401a8d04STejun Heo * @dwork: the delayed work to flush 31696e84d644SOleg Nesterov * 3170401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3171401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3172401a8d04STejun Heo * considers the last queueing instance of @dwork. 31731f1f642eSOleg Nesterov * 3174d185af30SYacine Belkadi * Return: 3175401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3176401a8d04STejun Heo * %false if it was already idle. 31776e84d644SOleg Nesterov */ 3178401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3179401a8d04STejun Heo { 31808930cabaSTejun Heo local_irq_disable(); 3181401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 318260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 31838930cabaSTejun Heo local_irq_enable(); 3184401a8d04STejun Heo return flush_work(&dwork->work); 3185401a8d04STejun Heo } 3186401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3187401a8d04STejun Heo 318805f0fe6bSTejun Heo /** 318905f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 319005f0fe6bSTejun Heo * @rwork: the rcu work to flush 319105f0fe6bSTejun Heo * 319205f0fe6bSTejun Heo * Return: 319305f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 319405f0fe6bSTejun Heo * %false if it was already idle. 319505f0fe6bSTejun Heo */ 319605f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 319705f0fe6bSTejun Heo { 319805f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 319905f0fe6bSTejun Heo rcu_barrier(); 320005f0fe6bSTejun Heo flush_work(&rwork->work); 320105f0fe6bSTejun Heo return true; 320205f0fe6bSTejun Heo } else { 320305f0fe6bSTejun Heo return flush_work(&rwork->work); 320405f0fe6bSTejun Heo } 320505f0fe6bSTejun Heo } 320605f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 320705f0fe6bSTejun Heo 3208f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3209f72b8792SJens Axboe { 3210f72b8792SJens Axboe unsigned long flags; 3211f72b8792SJens Axboe int ret; 3212f72b8792SJens Axboe 3213f72b8792SJens Axboe do { 3214f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3215f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3216f72b8792SJens Axboe 3217f72b8792SJens Axboe if (unlikely(ret < 0)) 3218f72b8792SJens Axboe return false; 3219f72b8792SJens Axboe 3220f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3221f72b8792SJens Axboe local_irq_restore(flags); 3222f72b8792SJens Axboe return ret; 3223f72b8792SJens Axboe } 3224f72b8792SJens Axboe 3225401a8d04STejun Heo /** 322657b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 322757b30ae7STejun Heo * @dwork: delayed_work to cancel 322809383498STejun Heo * 3229d185af30SYacine Belkadi * Kill off a pending delayed_work. 3230d185af30SYacine Belkadi * 3231d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3232d185af30SYacine Belkadi * pending. 3233d185af30SYacine Belkadi * 3234d185af30SYacine Belkadi * Note: 3235d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3236d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3237d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 323809383498STejun Heo * 323957b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 324009383498STejun Heo */ 324157b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 324209383498STejun Heo { 3243f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 324409383498STejun Heo } 324557b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 324609383498STejun Heo 324709383498STejun Heo /** 3248401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3249401a8d04STejun Heo * @dwork: the delayed work cancel 3250401a8d04STejun Heo * 3251401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3252401a8d04STejun Heo * 3253d185af30SYacine Belkadi * Return: 3254401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3255401a8d04STejun Heo */ 3256401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 32576e84d644SOleg Nesterov { 325836e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 32596e84d644SOleg Nesterov } 3260f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 32611da177e4SLinus Torvalds 32620fcb78c2SRolf Eike Beer /** 326331ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3264b6136773SAndrew Morton * @func: the function to call 3265b6136773SAndrew Morton * 326631ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 326731ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3268b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 326931ddd871STejun Heo * 3270d185af30SYacine Belkadi * Return: 327131ddd871STejun Heo * 0 on success, -errno on failure. 3272b6136773SAndrew Morton */ 327365f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 327415316ba8SChristoph Lameter { 327515316ba8SChristoph Lameter int cpu; 327638f51568SNamhyung Kim struct work_struct __percpu *works; 327715316ba8SChristoph Lameter 3278b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3279b6136773SAndrew Morton if (!works) 328015316ba8SChristoph Lameter return -ENOMEM; 3281b6136773SAndrew Morton 328295402b38SGautham R Shenoy get_online_cpus(); 328393981800STejun Heo 328415316ba8SChristoph Lameter for_each_online_cpu(cpu) { 32859bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 32869bfb1839SIngo Molnar 32879bfb1839SIngo Molnar INIT_WORK(work, func); 32888de6d308SOleg Nesterov schedule_work_on(cpu, work); 328915316ba8SChristoph Lameter } 329093981800STejun Heo 329193981800STejun Heo for_each_online_cpu(cpu) 32928616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 329393981800STejun Heo 329495402b38SGautham R Shenoy put_online_cpus(); 3295b6136773SAndrew Morton free_percpu(works); 329615316ba8SChristoph Lameter return 0; 329715316ba8SChristoph Lameter } 329815316ba8SChristoph Lameter 3299eef6a7d5SAlan Stern /** 33001fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 33011fa44ecaSJames Bottomley * @fn: the function to execute 33021fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 33031fa44ecaSJames Bottomley * be available when the work executes) 33041fa44ecaSJames Bottomley * 33051fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 33061fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 33071fa44ecaSJames Bottomley * 3308d185af30SYacine Belkadi * Return: 0 - function was executed 33091fa44ecaSJames Bottomley * 1 - function was scheduled for execution 33101fa44ecaSJames Bottomley */ 331165f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 33121fa44ecaSJames Bottomley { 33131fa44ecaSJames Bottomley if (!in_interrupt()) { 331465f27f38SDavid Howells fn(&ew->work); 33151fa44ecaSJames Bottomley return 0; 33161fa44ecaSJames Bottomley } 33171fa44ecaSJames Bottomley 331865f27f38SDavid Howells INIT_WORK(&ew->work, fn); 33191fa44ecaSJames Bottomley schedule_work(&ew->work); 33201fa44ecaSJames Bottomley 33211fa44ecaSJames Bottomley return 1; 33221fa44ecaSJames Bottomley } 33231fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 33241fa44ecaSJames Bottomley 33257a4e344cSTejun Heo /** 33267a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33277a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33287a4e344cSTejun Heo * 33297a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33307a4e344cSTejun Heo */ 33317a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33327a4e344cSTejun Heo { 33337a4e344cSTejun Heo if (attrs) { 33347a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33357a4e344cSTejun Heo kfree(attrs); 33367a4e344cSTejun Heo } 33377a4e344cSTejun Heo } 33387a4e344cSTejun Heo 33397a4e344cSTejun Heo /** 33407a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33417a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33427a4e344cSTejun Heo * 33437a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3344d185af30SYacine Belkadi * return it. 3345d185af30SYacine Belkadi * 3346d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33477a4e344cSTejun Heo */ 33487a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33497a4e344cSTejun Heo { 33507a4e344cSTejun Heo struct workqueue_attrs *attrs; 33517a4e344cSTejun Heo 33527a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33537a4e344cSTejun Heo if (!attrs) 33547a4e344cSTejun Heo goto fail; 33557a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33567a4e344cSTejun Heo goto fail; 33577a4e344cSTejun Heo 335813e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33597a4e344cSTejun Heo return attrs; 33607a4e344cSTejun Heo fail: 33617a4e344cSTejun Heo free_workqueue_attrs(attrs); 33627a4e344cSTejun Heo return NULL; 33637a4e344cSTejun Heo } 33647a4e344cSTejun Heo 336529c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 336629c91e99STejun Heo const struct workqueue_attrs *from) 336729c91e99STejun Heo { 336829c91e99STejun Heo to->nice = from->nice; 336929c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 33702865a8fbSShaohua Li /* 33712865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 33722865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 33732865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 33742865a8fbSShaohua Li */ 33752865a8fbSShaohua Li to->no_numa = from->no_numa; 337629c91e99STejun Heo } 337729c91e99STejun Heo 337829c91e99STejun Heo /* hash value of the content of @attr */ 337929c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 338029c91e99STejun Heo { 338129c91e99STejun Heo u32 hash = 0; 338229c91e99STejun Heo 338329c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 338413e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 338513e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 338629c91e99STejun Heo return hash; 338729c91e99STejun Heo } 338829c91e99STejun Heo 338929c91e99STejun Heo /* content equality test */ 339029c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 339129c91e99STejun Heo const struct workqueue_attrs *b) 339229c91e99STejun Heo { 339329c91e99STejun Heo if (a->nice != b->nice) 339429c91e99STejun Heo return false; 339529c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 339629c91e99STejun Heo return false; 339729c91e99STejun Heo return true; 339829c91e99STejun Heo } 339929c91e99STejun Heo 34007a4e344cSTejun Heo /** 34017a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34027a4e344cSTejun Heo * @pool: worker_pool to initialize 34037a4e344cSTejun Heo * 3404402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3405d185af30SYacine Belkadi * 3406d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 340729c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 340829c91e99STejun Heo * on @pool safely to release it. 34097a4e344cSTejun Heo */ 34107a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34114e1a1f9aSTejun Heo { 34124e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 341329c91e99STejun Heo pool->id = -1; 341429c91e99STejun Heo pool->cpu = -1; 3415f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34164e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 341782607adcSTejun Heo pool->watchdog_ts = jiffies; 34184e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34194e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34204e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34214e1a1f9aSTejun Heo 342232a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 34234e1a1f9aSTejun Heo 342432a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 34254e1a1f9aSTejun Heo 3426da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34277a4e344cSTejun Heo 34287cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 342929c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 343029c91e99STejun Heo pool->refcnt = 1; 343129c91e99STejun Heo 343229c91e99STejun Heo /* shouldn't fail above this point */ 34337a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34347a4e344cSTejun Heo if (!pool->attrs) 34357a4e344cSTejun Heo return -ENOMEM; 34367a4e344cSTejun Heo return 0; 34374e1a1f9aSTejun Heo } 34384e1a1f9aSTejun Heo 3439669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3440669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3441669de8bdSBart Van Assche { 3442669de8bdSBart Van Assche char *lock_name; 3443669de8bdSBart Van Assche 3444669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3445669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3446669de8bdSBart Van Assche if (!lock_name) 3447669de8bdSBart Van Assche lock_name = wq->name; 344869a106c0SQian Cai 344969a106c0SQian Cai wq->lock_name = lock_name; 3450669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3451669de8bdSBart Van Assche } 3452669de8bdSBart Van Assche 3453669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3454669de8bdSBart Van Assche { 3455669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3456669de8bdSBart Van Assche } 3457669de8bdSBart Van Assche 3458669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3459669de8bdSBart Van Assche { 3460669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3461669de8bdSBart Van Assche kfree(wq->lock_name); 3462669de8bdSBart Van Assche } 3463669de8bdSBart Van Assche #else 3464669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3465669de8bdSBart Van Assche { 3466669de8bdSBart Van Assche } 3467669de8bdSBart Van Assche 3468669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3469669de8bdSBart Van Assche { 3470669de8bdSBart Van Assche } 3471669de8bdSBart Van Assche 3472669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3473669de8bdSBart Van Assche { 3474669de8bdSBart Van Assche } 3475669de8bdSBart Van Assche #endif 3476669de8bdSBart Van Assche 3477e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3478e2dca7adSTejun Heo { 3479e2dca7adSTejun Heo struct workqueue_struct *wq = 3480e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3481e2dca7adSTejun Heo 3482669de8bdSBart Van Assche wq_free_lockdep(wq); 3483669de8bdSBart Van Assche 3484e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3485e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3486e2dca7adSTejun Heo else 3487e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3488e2dca7adSTejun Heo 3489e2dca7adSTejun Heo kfree(wq->rescuer); 3490e2dca7adSTejun Heo kfree(wq); 3491e2dca7adSTejun Heo } 3492e2dca7adSTejun Heo 349329c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 349429c91e99STejun Heo { 349529c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 349629c91e99STejun Heo 34977cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 349829c91e99STejun Heo free_workqueue_attrs(pool->attrs); 349929c91e99STejun Heo kfree(pool); 350029c91e99STejun Heo } 350129c91e99STejun Heo 350229c91e99STejun Heo /** 350329c91e99STejun Heo * put_unbound_pool - put a worker_pool 350429c91e99STejun Heo * @pool: worker_pool to put 350529c91e99STejun Heo * 350629c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3507c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3508c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3509c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3510a892caccSTejun Heo * 3511a892caccSTejun Heo * Should be called with wq_pool_mutex held. 351229c91e99STejun Heo */ 351329c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 351429c91e99STejun Heo { 351560f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 351629c91e99STejun Heo struct worker *worker; 351729c91e99STejun Heo 3518a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3519a892caccSTejun Heo 3520a892caccSTejun Heo if (--pool->refcnt) 352129c91e99STejun Heo return; 352229c91e99STejun Heo 352329c91e99STejun Heo /* sanity checks */ 352461d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3525a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 352629c91e99STejun Heo return; 352729c91e99STejun Heo 352829c91e99STejun Heo /* release id and unhash */ 352929c91e99STejun Heo if (pool->id >= 0) 353029c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 353129c91e99STejun Heo hash_del(&pool->hash_node); 353229c91e99STejun Heo 3533c5aa87bbSTejun Heo /* 3534692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3535692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3536692b4825STejun Heo * manager and @pool gets freed with the flag set. 3537c5aa87bbSTejun Heo */ 353860f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 3539692b4825STejun Heo wait_event_lock_irq(wq_manager_wait, 3540692b4825STejun Heo !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); 3541692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3542692b4825STejun Heo 35431037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 354429c91e99STejun Heo destroy_worker(worker); 354529c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 354629c91e99STejun Heo spin_unlock_irq(&pool->lock); 354760f5a4bcSLai Jiangshan 35481258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3549da028469SLai Jiangshan if (!list_empty(&pool->workers)) 355060f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 35511258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 355260f5a4bcSLai Jiangshan 355360f5a4bcSLai Jiangshan if (pool->detach_completion) 355460f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 355560f5a4bcSLai Jiangshan 355629c91e99STejun Heo /* shut down the timers */ 355729c91e99STejun Heo del_timer_sync(&pool->idle_timer); 355829c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 355929c91e99STejun Heo 356029c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 356125b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 356229c91e99STejun Heo } 356329c91e99STejun Heo 356429c91e99STejun Heo /** 356529c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 356629c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 356729c91e99STejun Heo * 356829c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 356929c91e99STejun Heo * reference count and return it. If there already is a matching 357029c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3571d185af30SYacine Belkadi * create a new one. 3572a892caccSTejun Heo * 3573a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3574d185af30SYacine Belkadi * 3575d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3576d185af30SYacine Belkadi * On failure, %NULL. 357729c91e99STejun Heo */ 357829c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 357929c91e99STejun Heo { 358029c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 358129c91e99STejun Heo struct worker_pool *pool; 3582f3f90ad4STejun Heo int node; 3583e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 358429c91e99STejun Heo 3585a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 358629c91e99STejun Heo 358729c91e99STejun Heo /* do we already have a matching pool? */ 358829c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 358929c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 359029c91e99STejun Heo pool->refcnt++; 35913fb1823cSLai Jiangshan return pool; 359229c91e99STejun Heo } 359329c91e99STejun Heo } 359429c91e99STejun Heo 3595e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3596e2273584SXunlei Pang if (wq_numa_enabled) { 3597e2273584SXunlei Pang for_each_node(node) { 3598e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3599e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3600e2273584SXunlei Pang target_node = node; 3601e2273584SXunlei Pang break; 3602e2273584SXunlei Pang } 3603e2273584SXunlei Pang } 3604e2273584SXunlei Pang } 3605e2273584SXunlei Pang 360629c91e99STejun Heo /* nope, create a new one */ 3607e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 360829c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 360929c91e99STejun Heo goto fail; 361029c91e99STejun Heo 36118864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 361229c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3613e2273584SXunlei Pang pool->node = target_node; 361429c91e99STejun Heo 36152865a8fbSShaohua Li /* 36162865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36172865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36182865a8fbSShaohua Li */ 36192865a8fbSShaohua Li pool->attrs->no_numa = false; 36202865a8fbSShaohua Li 362129c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 362229c91e99STejun Heo goto fail; 362329c91e99STejun Heo 362429c91e99STejun Heo /* create and start the initial worker */ 36253347fa09STejun Heo if (wq_online && !create_worker(pool)) 362629c91e99STejun Heo goto fail; 362729c91e99STejun Heo 362829c91e99STejun Heo /* install */ 362929c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 36303fb1823cSLai Jiangshan 363129c91e99STejun Heo return pool; 363229c91e99STejun Heo fail: 363329c91e99STejun Heo if (pool) 363429c91e99STejun Heo put_unbound_pool(pool); 363529c91e99STejun Heo return NULL; 363629c91e99STejun Heo } 363729c91e99STejun Heo 36388864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36398864b4e5STejun Heo { 36408864b4e5STejun Heo kmem_cache_free(pwq_cache, 36418864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36428864b4e5STejun Heo } 36438864b4e5STejun Heo 36448864b4e5STejun Heo /* 36458864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36468864b4e5STejun Heo * and needs to be destroyed. 36478864b4e5STejun Heo */ 36488864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36498864b4e5STejun Heo { 36508864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36518864b4e5STejun Heo unbound_release_work); 36528864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36538864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3654bc0caf09STejun Heo bool is_last; 36558864b4e5STejun Heo 36568864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36578864b4e5STejun Heo return; 36588864b4e5STejun Heo 36593c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 36608864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3661bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 36623c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 36638864b4e5STejun Heo 3664a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 36658864b4e5STejun Heo put_unbound_pool(pool); 3666a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3667a892caccSTejun Heo 366825b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 36698864b4e5STejun Heo 36708864b4e5STejun Heo /* 36718864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3672e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 36738864b4e5STejun Heo */ 3674669de8bdSBart Van Assche if (is_last) { 3675669de8bdSBart Van Assche wq_unregister_lockdep(wq); 367625b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 36776029a918STejun Heo } 3678669de8bdSBart Van Assche } 36798864b4e5STejun Heo 36800fbd95aaSTejun Heo /** 3681699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36820fbd95aaSTejun Heo * @pwq: target pool_workqueue 36830fbd95aaSTejun Heo * 3684699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3685699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3686699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36870fbd95aaSTejun Heo */ 3688699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36890fbd95aaSTejun Heo { 3690699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3691699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 36923347fa09STejun Heo unsigned long flags; 3693699ce097STejun Heo 3694699ce097STejun Heo /* for @wq->saved_max_active */ 3695a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3696699ce097STejun Heo 3697699ce097STejun Heo /* fast exit for non-freezable wqs */ 3698699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3699699ce097STejun Heo return; 3700699ce097STejun Heo 37013347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 37023347fa09STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 3703699ce097STejun Heo 370474b414eaSLai Jiangshan /* 370574b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 370674b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 370774b414eaSLai Jiangshan * is updated and visible. 370874b414eaSLai Jiangshan */ 370974b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3710699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37110fbd95aaSTejun Heo 37120fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 37130fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 37140fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3715951a078aSLai Jiangshan 3716951a078aSLai Jiangshan /* 3717951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3718951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3719951a078aSLai Jiangshan */ 3720951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3721699ce097STejun Heo } else { 3722699ce097STejun Heo pwq->max_active = 0; 3723699ce097STejun Heo } 3724699ce097STejun Heo 37253347fa09STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 37260fbd95aaSTejun Heo } 37270fbd95aaSTejun Heo 3728e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3729f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3730f147f29eSTejun Heo struct worker_pool *pool) 3731d2c1d404STejun Heo { 3732d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3733d2c1d404STejun Heo 3734e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3735e50aba9aSTejun Heo 3736d2c1d404STejun Heo pwq->pool = pool; 3737d2c1d404STejun Heo pwq->wq = wq; 3738d2c1d404STejun Heo pwq->flush_color = -1; 37398864b4e5STejun Heo pwq->refcnt = 1; 3740d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 37411befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3742d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37438864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3744f147f29eSTejun Heo } 3745d2c1d404STejun Heo 3746f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37471befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3748f147f29eSTejun Heo { 3749f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3750f147f29eSTejun Heo 3751f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 375275ccf595STejun Heo 37531befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37541befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37551befcf30STejun Heo return; 37561befcf30STejun Heo 375729b1cb41SLai Jiangshan /* set the matching work_color */ 375875ccf595STejun Heo pwq->work_color = wq->work_color; 3759983ca25eSTejun Heo 3760983ca25eSTejun Heo /* sync max_active to the current setting */ 3761983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3762983ca25eSTejun Heo 3763983ca25eSTejun Heo /* link in @pwq */ 37649e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3765df2d5ae4STejun Heo } 37666029a918STejun Heo 3767f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3768f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3769f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3770f147f29eSTejun Heo { 3771f147f29eSTejun Heo struct worker_pool *pool; 3772f147f29eSTejun Heo struct pool_workqueue *pwq; 3773f147f29eSTejun Heo 3774f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3775f147f29eSTejun Heo 3776f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3777f147f29eSTejun Heo if (!pool) 3778f147f29eSTejun Heo return NULL; 3779f147f29eSTejun Heo 3780e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3781f147f29eSTejun Heo if (!pwq) { 3782f147f29eSTejun Heo put_unbound_pool(pool); 3783f147f29eSTejun Heo return NULL; 3784f147f29eSTejun Heo } 3785f147f29eSTejun Heo 3786f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3787f147f29eSTejun Heo return pwq; 3788d2c1d404STejun Heo } 3789d2c1d404STejun Heo 37904c16bd32STejun Heo /** 379130186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3792042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 37934c16bd32STejun Heo * @node: the target NUMA node 37944c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 37954c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 37964c16bd32STejun Heo * 37974c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 37984c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3799d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38004c16bd32STejun Heo * 38014c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38024c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38034c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38044c16bd32STejun Heo * @attrs->cpumask. 38054c16bd32STejun Heo * 38064c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38074c16bd32STejun Heo * stable. 3808d185af30SYacine Belkadi * 3809d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3810d185af30SYacine Belkadi * %false if equal. 38114c16bd32STejun Heo */ 38124c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38134c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38144c16bd32STejun Heo { 3815d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38164c16bd32STejun Heo goto use_dfl; 38174c16bd32STejun Heo 38184c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38194c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38204c16bd32STejun Heo if (cpu_going_down >= 0) 38214c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38224c16bd32STejun Heo 38234c16bd32STejun Heo if (cpumask_empty(cpumask)) 38244c16bd32STejun Heo goto use_dfl; 38254c16bd32STejun Heo 38264c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38274c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38281ad0f0a7SMichael Bringmann 38291ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 38301ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 38311ad0f0a7SMichael Bringmann "possible intersect\n"); 38321ad0f0a7SMichael Bringmann return false; 38331ad0f0a7SMichael Bringmann } 38341ad0f0a7SMichael Bringmann 38354c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 38364c16bd32STejun Heo 38374c16bd32STejun Heo use_dfl: 38384c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38394c16bd32STejun Heo return false; 38404c16bd32STejun Heo } 38414c16bd32STejun Heo 38421befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38431befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38441befcf30STejun Heo int node, 38451befcf30STejun Heo struct pool_workqueue *pwq) 38461befcf30STejun Heo { 38471befcf30STejun Heo struct pool_workqueue *old_pwq; 38481befcf30STejun Heo 38495b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 38501befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38511befcf30STejun Heo 38521befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 38531befcf30STejun Heo link_pwq(pwq); 38541befcf30STejun Heo 38551befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 38561befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 38571befcf30STejun Heo return old_pwq; 38581befcf30STejun Heo } 38591befcf30STejun Heo 38602d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 38612d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 38622d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 38632d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3864042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 38652d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 38662d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 38672d5f0764SLai Jiangshan }; 38682d5f0764SLai Jiangshan 38692d5f0764SLai Jiangshan /* free the resources after success or abort */ 38702d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 38712d5f0764SLai Jiangshan { 38722d5f0764SLai Jiangshan if (ctx) { 38732d5f0764SLai Jiangshan int node; 38742d5f0764SLai Jiangshan 38752d5f0764SLai Jiangshan for_each_node(node) 38762d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 38772d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 38782d5f0764SLai Jiangshan 38792d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 38802d5f0764SLai Jiangshan 38812d5f0764SLai Jiangshan kfree(ctx); 38822d5f0764SLai Jiangshan } 38832d5f0764SLai Jiangshan } 38842d5f0764SLai Jiangshan 38852d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 38862d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 38872d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 38882d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 38892d5f0764SLai Jiangshan { 38902d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 38912d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 38922d5f0764SLai Jiangshan int node; 38932d5f0764SLai Jiangshan 38942d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 38952d5f0764SLai Jiangshan 3896acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 38972d5f0764SLai Jiangshan 38982d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38992d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39002d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 39012d5f0764SLai Jiangshan goto out_free; 39022d5f0764SLai Jiangshan 3903042f7df1SLai Jiangshan /* 3904042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3905042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3906042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3907042f7df1SLai Jiangshan */ 39082d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3909b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3910042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3911042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 39122d5f0764SLai Jiangshan 39132d5f0764SLai Jiangshan /* 39142d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 39152d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 39162d5f0764SLai Jiangshan * pools. 39172d5f0764SLai Jiangshan */ 39182d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 39192d5f0764SLai Jiangshan 39202d5f0764SLai Jiangshan /* 39212d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 39222d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 39232d5f0764SLai Jiangshan * it even if we don't use it immediately. 39242d5f0764SLai Jiangshan */ 39252d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39262d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 39272d5f0764SLai Jiangshan goto out_free; 39282d5f0764SLai Jiangshan 39292d5f0764SLai Jiangshan for_each_node(node) { 3930042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 39312d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39322d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 39332d5f0764SLai Jiangshan goto out_free; 39342d5f0764SLai Jiangshan } else { 39352d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 39362d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 39372d5f0764SLai Jiangshan } 39382d5f0764SLai Jiangshan } 39392d5f0764SLai Jiangshan 3940042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3941042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3942042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 39432d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3944042f7df1SLai Jiangshan 39452d5f0764SLai Jiangshan ctx->wq = wq; 39462d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 39472d5f0764SLai Jiangshan return ctx; 39482d5f0764SLai Jiangshan 39492d5f0764SLai Jiangshan out_free: 39502d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 39512d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 39522d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 39532d5f0764SLai Jiangshan return NULL; 39542d5f0764SLai Jiangshan } 39552d5f0764SLai Jiangshan 39562d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 39572d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 39582d5f0764SLai Jiangshan { 39592d5f0764SLai Jiangshan int node; 39602d5f0764SLai Jiangshan 39612d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 39622d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 39632d5f0764SLai Jiangshan 39642d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 39652d5f0764SLai Jiangshan 39662d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 39672d5f0764SLai Jiangshan for_each_node(node) 39682d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 39692d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 39702d5f0764SLai Jiangshan 39712d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 39722d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 39732d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 39742d5f0764SLai Jiangshan 39752d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 39762d5f0764SLai Jiangshan } 39772d5f0764SLai Jiangshan 3978a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3979a0111cf6SLai Jiangshan { 3980a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3981a0111cf6SLai Jiangshan get_online_cpus(); 3982a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3983a0111cf6SLai Jiangshan } 3984a0111cf6SLai Jiangshan 3985a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3986a0111cf6SLai Jiangshan { 3987a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3988a0111cf6SLai Jiangshan put_online_cpus(); 3989a0111cf6SLai Jiangshan } 3990a0111cf6SLai Jiangshan 3991a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3992a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3993a0111cf6SLai Jiangshan { 3994a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3995a0111cf6SLai Jiangshan 3996a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3997a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3998a0111cf6SLai Jiangshan return -EINVAL; 3999a0111cf6SLai Jiangshan 4000a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 40010a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 40020a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4003a0111cf6SLai Jiangshan return -EINVAL; 4004a0111cf6SLai Jiangshan 40050a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 40060a94efb5STejun Heo } 40070a94efb5STejun Heo 4008a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 40096201171eSwanghaibin if (!ctx) 40106201171eSwanghaibin return -ENOMEM; 4011a0111cf6SLai Jiangshan 4012a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4013a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4014a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4015a0111cf6SLai Jiangshan 40166201171eSwanghaibin return 0; 4017a0111cf6SLai Jiangshan } 4018a0111cf6SLai Jiangshan 40199e8cd2f5STejun Heo /** 40209e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 40219e8cd2f5STejun Heo * @wq: the target workqueue 40229e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 40239e8cd2f5STejun Heo * 40244c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 40254c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 40264c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 40274c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 40284c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 40294c16bd32STejun Heo * back-to-back will stay on its current pwq. 40309e8cd2f5STejun Heo * 4031d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4032d185af30SYacine Belkadi * 4033d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 40349e8cd2f5STejun Heo */ 40359e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 40369e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 40379e8cd2f5STejun Heo { 4038a0111cf6SLai Jiangshan int ret; 40399e8cd2f5STejun Heo 4040a0111cf6SLai Jiangshan apply_wqattrs_lock(); 4041a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4042a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 40432d5f0764SLai Jiangshan 40442d5f0764SLai Jiangshan return ret; 40459e8cd2f5STejun Heo } 40466106c0f8SNeilBrown EXPORT_SYMBOL_GPL(apply_workqueue_attrs); 40479e8cd2f5STejun Heo 40484c16bd32STejun Heo /** 40494c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 40504c16bd32STejun Heo * @wq: the target workqueue 40514c16bd32STejun Heo * @cpu: the CPU coming up or going down 40524c16bd32STejun Heo * @online: whether @cpu is coming up or going down 40534c16bd32STejun Heo * 40544c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 40554c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 40564c16bd32STejun Heo * @wq accordingly. 40574c16bd32STejun Heo * 40584c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 40594c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 40604c16bd32STejun Heo * correct. 40614c16bd32STejun Heo * 40624c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 40634c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 40644c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 40654c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 40664c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 40674c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 40684c16bd32STejun Heo * CPU_DOWN_PREPARE. 40694c16bd32STejun Heo */ 40704c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 40714c16bd32STejun Heo bool online) 40724c16bd32STejun Heo { 40734c16bd32STejun Heo int node = cpu_to_node(cpu); 40744c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 40754c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 40764c16bd32STejun Heo struct workqueue_attrs *target_attrs; 40774c16bd32STejun Heo cpumask_t *cpumask; 40784c16bd32STejun Heo 40794c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 40804c16bd32STejun Heo 4081f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4082f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 40834c16bd32STejun Heo return; 40844c16bd32STejun Heo 40854c16bd32STejun Heo /* 40864c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 40874c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 40884c16bd32STejun Heo * CPU hotplug exclusion. 40894c16bd32STejun Heo */ 40904c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 40914c16bd32STejun Heo cpumask = target_attrs->cpumask; 40924c16bd32STejun Heo 40934c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 40944c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 40954c16bd32STejun Heo 40964c16bd32STejun Heo /* 40974c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4098042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4099042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4100042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 41014c16bd32STejun Heo */ 4102042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 41034c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4104f7142ed4SLai Jiangshan return; 41054c16bd32STejun Heo } else { 41064c16bd32STejun Heo goto use_dfl_pwq; 41074c16bd32STejun Heo } 41084c16bd32STejun Heo 41094c16bd32STejun Heo /* create a new pwq */ 41104c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41114c16bd32STejun Heo if (!pwq) { 41122d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41134c16bd32STejun Heo wq->name); 411477f300b1SDaeseok Youn goto use_dfl_pwq; 41154c16bd32STejun Heo } 41164c16bd32STejun Heo 4117f7142ed4SLai Jiangshan /* Install the new pwq. */ 41184c16bd32STejun Heo mutex_lock(&wq->mutex); 41194c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41204c16bd32STejun Heo goto out_unlock; 41214c16bd32STejun Heo 41224c16bd32STejun Heo use_dfl_pwq: 4123f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 41244c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 41254c16bd32STejun Heo get_pwq(wq->dfl_pwq); 41264c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 41274c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 41284c16bd32STejun Heo out_unlock: 41294c16bd32STejun Heo mutex_unlock(&wq->mutex); 41304c16bd32STejun Heo put_pwq_unlocked(old_pwq); 41314c16bd32STejun Heo } 41324c16bd32STejun Heo 413330cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 41341da177e4SLinus Torvalds { 413549e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 41368a2b7538STejun Heo int cpu, ret; 4137e1d8aa9fSFrederic Weisbecker 413830cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4139420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4140420c0ddbSTejun Heo if (!wq->cpu_pwqs) 414130cdf249STejun Heo return -ENOMEM; 414230cdf249STejun Heo 414330cdf249STejun Heo for_each_possible_cpu(cpu) { 41447fb98ea7STejun Heo struct pool_workqueue *pwq = 41457fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 41467a62c2c8STejun Heo struct worker_pool *cpu_pools = 4147f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 414830cdf249STejun Heo 4149f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4150f147f29eSTejun Heo 4151f147f29eSTejun Heo mutex_lock(&wq->mutex); 41521befcf30STejun Heo link_pwq(pwq); 4153f147f29eSTejun Heo mutex_unlock(&wq->mutex); 415430cdf249STejun Heo } 415530cdf249STejun Heo return 0; 41568a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 41578a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 41588a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 41598a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 41608a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 41618a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 41628a2b7538STejun Heo return ret; 41639e8cd2f5STejun Heo } else { 41649e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 41659e8cd2f5STejun Heo } 41660f900049STejun Heo } 41670f900049STejun Heo 4168f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4169f3421797STejun Heo const char *name) 4170b71ab8c2STejun Heo { 4171f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4172f3421797STejun Heo 4173f3421797STejun Heo if (max_active < 1 || max_active > lim) 4174044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4175f3421797STejun Heo max_active, name, 1, lim); 4176b71ab8c2STejun Heo 4177f3421797STejun Heo return clamp_val(max_active, 1, lim); 4178b71ab8c2STejun Heo } 4179b71ab8c2STejun Heo 4180983c7515STejun Heo /* 4181983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4182983c7515STejun Heo * to guarantee forward progress. 4183983c7515STejun Heo */ 4184983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4185983c7515STejun Heo { 4186983c7515STejun Heo struct worker *rescuer; 4187983c7515STejun Heo int ret; 4188983c7515STejun Heo 4189983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4190983c7515STejun Heo return 0; 4191983c7515STejun Heo 4192983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4193983c7515STejun Heo if (!rescuer) 4194983c7515STejun Heo return -ENOMEM; 4195983c7515STejun Heo 4196983c7515STejun Heo rescuer->rescue_wq = wq; 4197983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4198983c7515STejun Heo ret = PTR_ERR_OR_ZERO(rescuer->task); 4199983c7515STejun Heo if (ret) { 4200983c7515STejun Heo kfree(rescuer); 4201983c7515STejun Heo return ret; 4202983c7515STejun Heo } 4203983c7515STejun Heo 4204983c7515STejun Heo wq->rescuer = rescuer; 4205983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4206983c7515STejun Heo wake_up_process(rescuer->task); 4207983c7515STejun Heo 4208983c7515STejun Heo return 0; 4209983c7515STejun Heo } 4210983c7515STejun Heo 4211669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 421297e37d7bSTejun Heo unsigned int flags, 4213669de8bdSBart Van Assche int max_active, ...) 42143af24433SOleg Nesterov { 4215df2d5ae4STejun Heo size_t tbl_size = 0; 4216ecf6881fSTejun Heo va_list args; 42173af24433SOleg Nesterov struct workqueue_struct *wq; 421849e3cf44STejun Heo struct pool_workqueue *pwq; 4219b196be89STejun Heo 42205c0338c6STejun Heo /* 42215c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 42225c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 42235c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 42245c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 42255c0338c6STejun Heo * on NUMA. 42265c0338c6STejun Heo */ 42275c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 42285c0338c6STejun Heo flags |= __WQ_ORDERED; 42295c0338c6STejun Heo 4230cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4231cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4232cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4233cee22a15SViresh Kumar 4234ecf6881fSTejun Heo /* allocate wq and format name */ 4235df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4236ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4237df2d5ae4STejun Heo 4238df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4239b196be89STejun Heo if (!wq) 4240d2c1d404STejun Heo return NULL; 4241b196be89STejun Heo 42426029a918STejun Heo if (flags & WQ_UNBOUND) { 42436029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 42446029a918STejun Heo if (!wq->unbound_attrs) 42456029a918STejun Heo goto err_free_wq; 42466029a918STejun Heo } 42476029a918STejun Heo 4248669de8bdSBart Van Assche va_start(args, max_active); 4249ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4250b196be89STejun Heo va_end(args); 42513af24433SOleg Nesterov 4252d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4253b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 42543af24433SOleg Nesterov 4255b196be89STejun Heo /* init wq */ 425697e37d7bSTejun Heo wq->flags = flags; 4257a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 42583c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4259112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 426030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 426173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 426273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4263493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 42643af24433SOleg Nesterov 4265669de8bdSBart Van Assche wq_init_lockdep(wq); 4266cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 42673af24433SOleg Nesterov 426830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 426982efcab3SBart Van Assche goto err_unreg_lockdep; 42701537663fSTejun Heo 427140c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4272d2c1d404STejun Heo goto err_destroy; 4273e22bee78STejun Heo 4274226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4275226223abSTejun Heo goto err_destroy; 4276226223abSTejun Heo 42776af8bf3dSOleg Nesterov /* 427868e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 427968e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 428068e13a67SLai Jiangshan * list. 42816af8bf3dSOleg Nesterov */ 428268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4283a0a1a5fdSTejun Heo 4284a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 428549e3cf44STejun Heo for_each_pwq(pwq, wq) 4286699ce097STejun Heo pwq_adjust_max_active(pwq); 4287a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4288a0a1a5fdSTejun Heo 4289e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4290a0a1a5fdSTejun Heo 429168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42923af24433SOleg Nesterov 42933af24433SOleg Nesterov return wq; 4294d2c1d404STejun Heo 429582efcab3SBart Van Assche err_unreg_lockdep: 4296009bb421SBart Van Assche wq_unregister_lockdep(wq); 4297009bb421SBart Van Assche wq_free_lockdep(wq); 429882efcab3SBart Van Assche err_free_wq: 42996029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 43004690c4abSTejun Heo kfree(wq); 4301d2c1d404STejun Heo return NULL; 4302d2c1d404STejun Heo err_destroy: 4303d2c1d404STejun Heo destroy_workqueue(wq); 43044690c4abSTejun Heo return NULL; 43051da177e4SLinus Torvalds } 4306669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 43071da177e4SLinus Torvalds 43083af24433SOleg Nesterov /** 43093af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 43103af24433SOleg Nesterov * @wq: target workqueue 43113af24433SOleg Nesterov * 43123af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 43133af24433SOleg Nesterov */ 43143af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 43153af24433SOleg Nesterov { 431649e3cf44STejun Heo struct pool_workqueue *pwq; 43174c16bd32STejun Heo int node; 43183af24433SOleg Nesterov 43199c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 43209c5a2ba7STejun Heo drain_workqueue(wq); 4321c8efcc25STejun Heo 43226183c009STejun Heo /* sanity checks */ 4323b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 432449e3cf44STejun Heo for_each_pwq(pwq, wq) { 43256183c009STejun Heo int i; 43266183c009STejun Heo 432776af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 432876af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4329b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4330fa07fb6aSTejun Heo show_workqueue_state(); 43316183c009STejun Heo return; 433276af4d93STejun Heo } 433376af4d93STejun Heo } 433476af4d93STejun Heo 43355c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 43368864b4e5STejun Heo WARN_ON(pwq->nr_active) || 433776af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4338b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4339fa07fb6aSTejun Heo show_workqueue_state(); 43406183c009STejun Heo return; 43416183c009STejun Heo } 434276af4d93STejun Heo } 4343b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 43446183c009STejun Heo 4345a0a1a5fdSTejun Heo /* 4346a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4347a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4348a0a1a5fdSTejun Heo */ 434968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4350e2dca7adSTejun Heo list_del_rcu(&wq->list); 435168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43523af24433SOleg Nesterov 4353226223abSTejun Heo workqueue_sysfs_unregister(wq); 4354226223abSTejun Heo 4355e2dca7adSTejun Heo if (wq->rescuer) 4356e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4357e22bee78STejun Heo 43588864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4359669de8bdSBart Van Assche wq_unregister_lockdep(wq); 436029c91e99STejun Heo /* 43618864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4362e2dca7adSTejun Heo * schedule RCU free. 436329c91e99STejun Heo */ 436425b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 43658864b4e5STejun Heo } else { 43668864b4e5STejun Heo /* 43678864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 43684c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 43694c16bd32STejun Heo * @wq will be freed when the last pwq is released. 43708864b4e5STejun Heo */ 43714c16bd32STejun Heo for_each_node(node) { 43724c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 43734c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 43744c16bd32STejun Heo put_pwq_unlocked(pwq); 43754c16bd32STejun Heo } 43764c16bd32STejun Heo 43774c16bd32STejun Heo /* 43784c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 43794c16bd32STejun Heo * put. Don't access it afterwards. 43804c16bd32STejun Heo */ 43814c16bd32STejun Heo pwq = wq->dfl_pwq; 43824c16bd32STejun Heo wq->dfl_pwq = NULL; 4383dce90d47STejun Heo put_pwq_unlocked(pwq); 438429c91e99STejun Heo } 43853af24433SOleg Nesterov } 43863af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 43873af24433SOleg Nesterov 4388dcd989cbSTejun Heo /** 4389dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4390dcd989cbSTejun Heo * @wq: target workqueue 4391dcd989cbSTejun Heo * @max_active: new max_active value. 4392dcd989cbSTejun Heo * 4393dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4394dcd989cbSTejun Heo * 4395dcd989cbSTejun Heo * CONTEXT: 4396dcd989cbSTejun Heo * Don't call from IRQ context. 4397dcd989cbSTejun Heo */ 4398dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4399dcd989cbSTejun Heo { 440049e3cf44STejun Heo struct pool_workqueue *pwq; 4401dcd989cbSTejun Heo 44028719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 44030a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 44048719dceaSTejun Heo return; 44058719dceaSTejun Heo 4406f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4407dcd989cbSTejun Heo 4408a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4409dcd989cbSTejun Heo 44100a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4411dcd989cbSTejun Heo wq->saved_max_active = max_active; 4412dcd989cbSTejun Heo 4413699ce097STejun Heo for_each_pwq(pwq, wq) 4414699ce097STejun Heo pwq_adjust_max_active(pwq); 4415dcd989cbSTejun Heo 4416a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4417dcd989cbSTejun Heo } 4418dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4419dcd989cbSTejun Heo 4420dcd989cbSTejun Heo /** 442127d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 442227d4ee03SLukas Wunner * 442327d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 442427d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 442527d4ee03SLukas Wunner * 442627d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 442727d4ee03SLukas Wunner */ 442827d4ee03SLukas Wunner struct work_struct *current_work(void) 442927d4ee03SLukas Wunner { 443027d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 443127d4ee03SLukas Wunner 443227d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 443327d4ee03SLukas Wunner } 443427d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 443527d4ee03SLukas Wunner 443627d4ee03SLukas Wunner /** 4437e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4438e6267616STejun Heo * 4439e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4440e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4441d185af30SYacine Belkadi * 4442d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4443e6267616STejun Heo */ 4444e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4445e6267616STejun Heo { 4446e6267616STejun Heo struct worker *worker = current_wq_worker(); 4447e6267616STejun Heo 44486a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4449e6267616STejun Heo } 4450e6267616STejun Heo 4451e6267616STejun Heo /** 4452dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4453dcd989cbSTejun Heo * @cpu: CPU in question 4454dcd989cbSTejun Heo * @wq: target workqueue 4455dcd989cbSTejun Heo * 4456dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4457dcd989cbSTejun Heo * no synchronization around this function and the test result is 4458dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4459dcd989cbSTejun Heo * 4460d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4461d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4462d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4463d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4464d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4465d3251859STejun Heo * 4466d185af30SYacine Belkadi * Return: 4467dcd989cbSTejun Heo * %true if congested, %false otherwise. 4468dcd989cbSTejun Heo */ 4469d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4470dcd989cbSTejun Heo { 44717fb98ea7STejun Heo struct pool_workqueue *pwq; 447276af4d93STejun Heo bool ret; 447376af4d93STejun Heo 447488109453SLai Jiangshan rcu_read_lock_sched(); 44757fb98ea7STejun Heo 4476d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4477d3251859STejun Heo cpu = smp_processor_id(); 4478d3251859STejun Heo 44797fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 44807fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 44817fb98ea7STejun Heo else 4482df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4483dcd989cbSTejun Heo 448476af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 448588109453SLai Jiangshan rcu_read_unlock_sched(); 448676af4d93STejun Heo 448776af4d93STejun Heo return ret; 4488dcd989cbSTejun Heo } 4489dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4490dcd989cbSTejun Heo 4491dcd989cbSTejun Heo /** 4492dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4493dcd989cbSTejun Heo * @work: the work to be tested 4494dcd989cbSTejun Heo * 4495dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4496dcd989cbSTejun Heo * synchronization around this function and the test result is 4497dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4498dcd989cbSTejun Heo * 4499d185af30SYacine Belkadi * Return: 4500dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4501dcd989cbSTejun Heo */ 4502dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4503dcd989cbSTejun Heo { 4504fa1b54e6STejun Heo struct worker_pool *pool; 4505dcd989cbSTejun Heo unsigned long flags; 4506dcd989cbSTejun Heo unsigned int ret = 0; 4507dcd989cbSTejun Heo 4508dcd989cbSTejun Heo if (work_pending(work)) 4509dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4510038366c5SLai Jiangshan 4511fa1b54e6STejun Heo local_irq_save(flags); 4512fa1b54e6STejun Heo pool = get_work_pool(work); 4513038366c5SLai Jiangshan if (pool) { 4514fa1b54e6STejun Heo spin_lock(&pool->lock); 4515c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4516dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4517fa1b54e6STejun Heo spin_unlock(&pool->lock); 4518038366c5SLai Jiangshan } 4519fa1b54e6STejun Heo local_irq_restore(flags); 4520dcd989cbSTejun Heo 4521dcd989cbSTejun Heo return ret; 4522dcd989cbSTejun Heo } 4523dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4524dcd989cbSTejun Heo 45253d1cb205STejun Heo /** 45263d1cb205STejun Heo * set_worker_desc - set description for the current work item 45273d1cb205STejun Heo * @fmt: printf-style format string 45283d1cb205STejun Heo * @...: arguments for the format string 45293d1cb205STejun Heo * 45303d1cb205STejun Heo * This function can be called by a running work function to describe what 45313d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 45323d1cb205STejun Heo * information will be printed out together to help debugging. The 45333d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 45343d1cb205STejun Heo */ 45353d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 45363d1cb205STejun Heo { 45373d1cb205STejun Heo struct worker *worker = current_wq_worker(); 45383d1cb205STejun Heo va_list args; 45393d1cb205STejun Heo 45403d1cb205STejun Heo if (worker) { 45413d1cb205STejun Heo va_start(args, fmt); 45423d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 45433d1cb205STejun Heo va_end(args); 45443d1cb205STejun Heo } 45453d1cb205STejun Heo } 45465c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 45473d1cb205STejun Heo 45483d1cb205STejun Heo /** 45493d1cb205STejun Heo * print_worker_info - print out worker information and description 45503d1cb205STejun Heo * @log_lvl: the log level to use when printing 45513d1cb205STejun Heo * @task: target task 45523d1cb205STejun Heo * 45533d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 45543d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 45553d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 45563d1cb205STejun Heo * 45573d1cb205STejun Heo * This function can be safely called on any task as long as the 45583d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 45593d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 45603d1cb205STejun Heo */ 45613d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 45623d1cb205STejun Heo { 45633d1cb205STejun Heo work_func_t *fn = NULL; 45643d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 45653d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 45663d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 45673d1cb205STejun Heo struct workqueue_struct *wq = NULL; 45683d1cb205STejun Heo struct worker *worker; 45693d1cb205STejun Heo 45703d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 45713d1cb205STejun Heo return; 45723d1cb205STejun Heo 45733d1cb205STejun Heo /* 45743d1cb205STejun Heo * This function is called without any synchronization and @task 45753d1cb205STejun Heo * could be in any state. Be careful with dereferences. 45763d1cb205STejun Heo */ 4577e700591aSPetr Mladek worker = kthread_probe_data(task); 45783d1cb205STejun Heo 45793d1cb205STejun Heo /* 45808bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 45818bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 45823d1cb205STejun Heo */ 45833d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 45843d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 45853d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 45863d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 45873d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 45883d1cb205STejun Heo 45893d1cb205STejun Heo if (fn || name[0] || desc[0]) { 45903d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 45918bf89593STejun Heo if (strcmp(name, desc)) 45923d1cb205STejun Heo pr_cont(" (%s)", desc); 45933d1cb205STejun Heo pr_cont("\n"); 45943d1cb205STejun Heo } 45953d1cb205STejun Heo } 45963d1cb205STejun Heo 45973494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 45983494fc30STejun Heo { 45993494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 46003494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 46013494fc30STejun Heo pr_cont(" node=%d", pool->node); 46023494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 46033494fc30STejun Heo } 46043494fc30STejun Heo 46053494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 46063494fc30STejun Heo { 46073494fc30STejun Heo if (work->func == wq_barrier_func) { 46083494fc30STejun Heo struct wq_barrier *barr; 46093494fc30STejun Heo 46103494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 46113494fc30STejun Heo 46123494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 46133494fc30STejun Heo task_pid_nr(barr->task)); 46143494fc30STejun Heo } else { 46153494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 46163494fc30STejun Heo } 46173494fc30STejun Heo } 46183494fc30STejun Heo 46193494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 46203494fc30STejun Heo { 46213494fc30STejun Heo struct worker_pool *pool = pwq->pool; 46223494fc30STejun Heo struct work_struct *work; 46233494fc30STejun Heo struct worker *worker; 46243494fc30STejun Heo bool has_in_flight = false, has_pending = false; 46253494fc30STejun Heo int bkt; 46263494fc30STejun Heo 46273494fc30STejun Heo pr_info(" pwq %d:", pool->id); 46283494fc30STejun Heo pr_cont_pool_info(pool); 46293494fc30STejun Heo 46303494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 46313494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 46323494fc30STejun Heo 46333494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 46343494fc30STejun Heo if (worker->current_pwq == pwq) { 46353494fc30STejun Heo has_in_flight = true; 46363494fc30STejun Heo break; 46373494fc30STejun Heo } 46383494fc30STejun Heo } 46393494fc30STejun Heo if (has_in_flight) { 46403494fc30STejun Heo bool comma = false; 46413494fc30STejun Heo 46423494fc30STejun Heo pr_info(" in-flight:"); 46433494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 46443494fc30STejun Heo if (worker->current_pwq != pwq) 46453494fc30STejun Heo continue; 46463494fc30STejun Heo 46473494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 46483494fc30STejun Heo task_pid_nr(worker->task), 46493494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 46503494fc30STejun Heo worker->current_func); 46513494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 46523494fc30STejun Heo pr_cont_work(false, work); 46533494fc30STejun Heo comma = true; 46543494fc30STejun Heo } 46553494fc30STejun Heo pr_cont("\n"); 46563494fc30STejun Heo } 46573494fc30STejun Heo 46583494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 46593494fc30STejun Heo if (get_work_pwq(work) == pwq) { 46603494fc30STejun Heo has_pending = true; 46613494fc30STejun Heo break; 46623494fc30STejun Heo } 46633494fc30STejun Heo } 46643494fc30STejun Heo if (has_pending) { 46653494fc30STejun Heo bool comma = false; 46663494fc30STejun Heo 46673494fc30STejun Heo pr_info(" pending:"); 46683494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 46693494fc30STejun Heo if (get_work_pwq(work) != pwq) 46703494fc30STejun Heo continue; 46713494fc30STejun Heo 46723494fc30STejun Heo pr_cont_work(comma, work); 46733494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 46743494fc30STejun Heo } 46753494fc30STejun Heo pr_cont("\n"); 46763494fc30STejun Heo } 46773494fc30STejun Heo 46783494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 46793494fc30STejun Heo bool comma = false; 46803494fc30STejun Heo 46813494fc30STejun Heo pr_info(" delayed:"); 46823494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 46833494fc30STejun Heo pr_cont_work(comma, work); 46843494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 46853494fc30STejun Heo } 46863494fc30STejun Heo pr_cont("\n"); 46873494fc30STejun Heo } 46883494fc30STejun Heo } 46893494fc30STejun Heo 46903494fc30STejun Heo /** 46913494fc30STejun Heo * show_workqueue_state - dump workqueue state 46923494fc30STejun Heo * 46937b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 46947b776af6SRoger Lu * all busy workqueues and pools. 46953494fc30STejun Heo */ 46963494fc30STejun Heo void show_workqueue_state(void) 46973494fc30STejun Heo { 46983494fc30STejun Heo struct workqueue_struct *wq; 46993494fc30STejun Heo struct worker_pool *pool; 47003494fc30STejun Heo unsigned long flags; 47013494fc30STejun Heo int pi; 47023494fc30STejun Heo 47033494fc30STejun Heo rcu_read_lock_sched(); 47043494fc30STejun Heo 47053494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 47063494fc30STejun Heo 47073494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 47083494fc30STejun Heo struct pool_workqueue *pwq; 47093494fc30STejun Heo bool idle = true; 47103494fc30STejun Heo 47113494fc30STejun Heo for_each_pwq(pwq, wq) { 47123494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 47133494fc30STejun Heo idle = false; 47143494fc30STejun Heo break; 47153494fc30STejun Heo } 47163494fc30STejun Heo } 47173494fc30STejun Heo if (idle) 47183494fc30STejun Heo continue; 47193494fc30STejun Heo 47203494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 47213494fc30STejun Heo 47223494fc30STejun Heo for_each_pwq(pwq, wq) { 47233494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 47243494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 47253494fc30STejun Heo show_pwq(pwq); 47263494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 472762635ea8SSergey Senozhatsky /* 472862635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 472962635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 473062635ea8SSergey Senozhatsky * hard lockup. 473162635ea8SSergey Senozhatsky */ 473262635ea8SSergey Senozhatsky touch_nmi_watchdog(); 47333494fc30STejun Heo } 47343494fc30STejun Heo } 47353494fc30STejun Heo 47363494fc30STejun Heo for_each_pool(pool, pi) { 47373494fc30STejun Heo struct worker *worker; 47383494fc30STejun Heo bool first = true; 47393494fc30STejun Heo 47403494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 47413494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 47423494fc30STejun Heo goto next_pool; 47433494fc30STejun Heo 47443494fc30STejun Heo pr_info("pool %d:", pool->id); 47453494fc30STejun Heo pr_cont_pool_info(pool); 474682607adcSTejun Heo pr_cont(" hung=%us workers=%d", 474782607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 474882607adcSTejun Heo pool->nr_workers); 47493494fc30STejun Heo if (pool->manager) 47503494fc30STejun Heo pr_cont(" manager: %d", 47513494fc30STejun Heo task_pid_nr(pool->manager->task)); 47523494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 47533494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 47543494fc30STejun Heo task_pid_nr(worker->task)); 47553494fc30STejun Heo first = false; 47563494fc30STejun Heo } 47573494fc30STejun Heo pr_cont("\n"); 47583494fc30STejun Heo next_pool: 47593494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 476062635ea8SSergey Senozhatsky /* 476162635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 476262635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 476362635ea8SSergey Senozhatsky * hard lockup. 476462635ea8SSergey Senozhatsky */ 476562635ea8SSergey Senozhatsky touch_nmi_watchdog(); 47663494fc30STejun Heo } 47673494fc30STejun Heo 47683494fc30STejun Heo rcu_read_unlock_sched(); 47693494fc30STejun Heo } 47703494fc30STejun Heo 47716b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 47726b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 47736b59808bSTejun Heo { 47746b59808bSTejun Heo int off; 47756b59808bSTejun Heo 47766b59808bSTejun Heo /* always show the actual comm */ 47776b59808bSTejun Heo off = strscpy(buf, task->comm, size); 47786b59808bSTejun Heo if (off < 0) 47796b59808bSTejun Heo return; 47806b59808bSTejun Heo 4781197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 47826b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 47836b59808bSTejun Heo 4784197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4785197f6accSTejun Heo struct worker *worker = kthread_data(task); 4786197f6accSTejun Heo struct worker_pool *pool = worker->pool; 47876b59808bSTejun Heo 47886b59808bSTejun Heo if (pool) { 47896b59808bSTejun Heo spin_lock_irq(&pool->lock); 47906b59808bSTejun Heo /* 4791197f6accSTejun Heo * ->desc tracks information (wq name or 4792197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4793197f6accSTejun Heo * current, prepend '+', otherwise '-'. 47946b59808bSTejun Heo */ 47956b59808bSTejun Heo if (worker->desc[0] != '\0') { 47966b59808bSTejun Heo if (worker->current_work) 47976b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 47986b59808bSTejun Heo worker->desc); 47996b59808bSTejun Heo else 48006b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 48016b59808bSTejun Heo worker->desc); 48026b59808bSTejun Heo } 48036b59808bSTejun Heo spin_unlock_irq(&pool->lock); 48046b59808bSTejun Heo } 4805197f6accSTejun Heo } 48066b59808bSTejun Heo 48076b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 48086b59808bSTejun Heo } 48096b59808bSTejun Heo 481066448bc2SMathieu Malaterre #ifdef CONFIG_SMP 481166448bc2SMathieu Malaterre 4812db7bccf4STejun Heo /* 4813db7bccf4STejun Heo * CPU hotplug. 4814db7bccf4STejun Heo * 4815e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4816112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4817706026c2STejun Heo * pool which make migrating pending and scheduled works very 4818e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 481994cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4820e22bee78STejun Heo * blocked draining impractical. 4821e22bee78STejun Heo * 482224647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4823628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4824628c78e7STejun Heo * cpu comes back online. 4825db7bccf4STejun Heo */ 4826db7bccf4STejun Heo 4827e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4828db7bccf4STejun Heo { 48294ce62e9eSTejun Heo struct worker_pool *pool; 4830db7bccf4STejun Heo struct worker *worker; 4831db7bccf4STejun Heo 4832f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 48331258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 483494cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4835e22bee78STejun Heo 4836f2d5a0eeSTejun Heo /* 483792f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 483894cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 483994cf58bbSTejun Heo * except for the ones which are still executing works from 484094cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 484194cf58bbSTejun Heo * this, they may become diasporas. 4842f2d5a0eeSTejun Heo */ 4843da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4844403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4845db7bccf4STejun Heo 484624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4847f2d5a0eeSTejun Heo 484894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 48491258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 4850e22bee78STejun Heo 4851e22bee78STejun Heo /* 4852eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4853eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4854eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4855eb283428SLai Jiangshan * from other cpus. 4856628c78e7STejun Heo */ 4857628c78e7STejun Heo schedule(); 4858628c78e7STejun Heo 4859628c78e7STejun Heo /* 4860eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4861eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4862eb283428SLai Jiangshan * and keep_working() are always true as long as the 4863eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4864eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4865eb283428SLai Jiangshan * are served by workers tied to the pool. 4866e22bee78STejun Heo */ 4867e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4868eb283428SLai Jiangshan 4869eb283428SLai Jiangshan /* 4870eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4871eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4872eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4873eb283428SLai Jiangshan */ 4874eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4875eb283428SLai Jiangshan wake_up_worker(pool); 4876eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4877eb283428SLai Jiangshan } 4878db7bccf4STejun Heo } 4879db7bccf4STejun Heo 4880bd7c089eSTejun Heo /** 4881bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4882bd7c089eSTejun Heo * @pool: pool of interest 4883bd7c089eSTejun Heo * 4884a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4885bd7c089eSTejun Heo */ 4886bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4887bd7c089eSTejun Heo { 4888a9ab775bSTejun Heo struct worker *worker; 4889bd7c089eSTejun Heo 48901258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 4891bd7c089eSTejun Heo 4892bd7c089eSTejun Heo /* 4893a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4894a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4895402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4896a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4897a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4898bd7c089eSTejun Heo */ 4899da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4900a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4901a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4902a9ab775bSTejun Heo 4903a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4904f7c17d26SWanpeng Li 49053de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4906a9ab775bSTejun Heo 4907da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4908a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4909a9ab775bSTejun Heo 4910a9ab775bSTejun Heo /* 4911a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4912a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4913a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4914a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4915a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4916a9ab775bSTejun Heo * be bound before @pool->lock is released. 4917a9ab775bSTejun Heo */ 4918a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4919bd7c089eSTejun Heo wake_up_process(worker->task); 4920bd7c089eSTejun Heo 4921bd7c089eSTejun Heo /* 4922a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4923a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4924a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4925a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4926a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4927a9ab775bSTejun Heo * concurrency management. Note that when or whether 4928a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4929a9ab775bSTejun Heo * 4930c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 4931a9ab775bSTejun Heo * tested without holding any lock in 4932a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4933a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4934a9ab775bSTejun Heo * management operations. 4935bd7c089eSTejun Heo */ 4936a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4937a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4938a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4939c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 4940bd7c089eSTejun Heo } 4941a9ab775bSTejun Heo 4942a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4943bd7c089eSTejun Heo } 4944bd7c089eSTejun Heo 49457dbc725eSTejun Heo /** 49467dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 49477dbc725eSTejun Heo * @pool: unbound pool of interest 49487dbc725eSTejun Heo * @cpu: the CPU which is coming up 49497dbc725eSTejun Heo * 49507dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 49517dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 49527dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 49537dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 49547dbc725eSTejun Heo */ 49557dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 49567dbc725eSTejun Heo { 49577dbc725eSTejun Heo static cpumask_t cpumask; 49587dbc725eSTejun Heo struct worker *worker; 49597dbc725eSTejun Heo 49601258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 49617dbc725eSTejun Heo 49627dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 49637dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 49647dbc725eSTejun Heo return; 49657dbc725eSTejun Heo 49667dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 49677dbc725eSTejun Heo 49687dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4969da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4970d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 49717dbc725eSTejun Heo } 49727dbc725eSTejun Heo 49737ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 49741da177e4SLinus Torvalds { 49754ce62e9eSTejun Heo struct worker_pool *pool; 49761da177e4SLinus Torvalds 4977f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49783ce63377STejun Heo if (pool->nr_workers) 49793ce63377STejun Heo continue; 4980051e1850SLai Jiangshan if (!create_worker(pool)) 49817ee681b2SThomas Gleixner return -ENOMEM; 49823af24433SOleg Nesterov } 49837ee681b2SThomas Gleixner return 0; 49847ee681b2SThomas Gleixner } 49851da177e4SLinus Torvalds 49867ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 49877ee681b2SThomas Gleixner { 49887ee681b2SThomas Gleixner struct worker_pool *pool; 49897ee681b2SThomas Gleixner struct workqueue_struct *wq; 49907ee681b2SThomas Gleixner int pi; 49917ee681b2SThomas Gleixner 499268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 49937dbc725eSTejun Heo 49947dbc725eSTejun Heo for_each_pool(pool, pi) { 49951258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 499694cf58bbSTejun Heo 4997f05b558dSLai Jiangshan if (pool->cpu == cpu) 499894cf58bbSTejun Heo rebind_workers(pool); 4999f05b558dSLai Jiangshan else if (pool->cpu < 0) 50007dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 500194cf58bbSTejun Heo 50021258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 500394cf58bbSTejun Heo } 50047dbc725eSTejun Heo 50054c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 50064c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 50074c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 50084c16bd32STejun Heo 500968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 50107ee681b2SThomas Gleixner return 0; 501165758202STejun Heo } 501265758202STejun Heo 50137ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 501465758202STejun Heo { 50154c16bd32STejun Heo struct workqueue_struct *wq; 50168db25e78STejun Heo 50174c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5018e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5019e8b3f8dbSLai Jiangshan return -1; 5020e8b3f8dbSLai Jiangshan 5021e8b3f8dbSLai Jiangshan unbind_workers(cpu); 50224c16bd32STejun Heo 50234c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 50244c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 50254c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 50264c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 50274c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 50284c16bd32STejun Heo 50297ee681b2SThomas Gleixner return 0; 503065758202STejun Heo } 503165758202STejun Heo 50322d3854a3SRusty Russell struct work_for_cpu { 5033ed48ece2STejun Heo struct work_struct work; 50342d3854a3SRusty Russell long (*fn)(void *); 50352d3854a3SRusty Russell void *arg; 50362d3854a3SRusty Russell long ret; 50372d3854a3SRusty Russell }; 50382d3854a3SRusty Russell 5039ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 50402d3854a3SRusty Russell { 5041ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5042ed48ece2STejun Heo 50432d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 50442d3854a3SRusty Russell } 50452d3854a3SRusty Russell 50462d3854a3SRusty Russell /** 504722aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 50482d3854a3SRusty Russell * @cpu: the cpu to run on 50492d3854a3SRusty Russell * @fn: the function to run 50502d3854a3SRusty Russell * @arg: the function arg 50512d3854a3SRusty Russell * 505231ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 50536b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5054d185af30SYacine Belkadi * 5055d185af30SYacine Belkadi * Return: The value @fn returns. 50562d3854a3SRusty Russell */ 5057d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 50582d3854a3SRusty Russell { 5059ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 50602d3854a3SRusty Russell 5061ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5062ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 506312997d1aSBjorn Helgaas flush_work(&wfc.work); 5064440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 50652d3854a3SRusty Russell return wfc.ret; 50662d3854a3SRusty Russell } 50672d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 50680e8d6a93SThomas Gleixner 50690e8d6a93SThomas Gleixner /** 50700e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 50710e8d6a93SThomas Gleixner * @cpu: the cpu to run on 50720e8d6a93SThomas Gleixner * @fn: the function to run 50730e8d6a93SThomas Gleixner * @arg: the function argument 50740e8d6a93SThomas Gleixner * 50750e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 50760e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 50770e8d6a93SThomas Gleixner * 50780e8d6a93SThomas Gleixner * Return: The value @fn returns. 50790e8d6a93SThomas Gleixner */ 50800e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 50810e8d6a93SThomas Gleixner { 50820e8d6a93SThomas Gleixner long ret = -ENODEV; 50830e8d6a93SThomas Gleixner 50840e8d6a93SThomas Gleixner get_online_cpus(); 50850e8d6a93SThomas Gleixner if (cpu_online(cpu)) 50860e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 50870e8d6a93SThomas Gleixner put_online_cpus(); 50880e8d6a93SThomas Gleixner return ret; 50890e8d6a93SThomas Gleixner } 50900e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 50912d3854a3SRusty Russell #endif /* CONFIG_SMP */ 50922d3854a3SRusty Russell 5093a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5094e7577c50SRusty Russell 5095a0a1a5fdSTejun Heo /** 5096a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5097a0a1a5fdSTejun Heo * 509858a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5099c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 5100706026c2STejun Heo * pool->worklist. 5101a0a1a5fdSTejun Heo * 5102a0a1a5fdSTejun Heo * CONTEXT: 5103a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5104a0a1a5fdSTejun Heo */ 5105a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5106a0a1a5fdSTejun Heo { 510724b8a847STejun Heo struct workqueue_struct *wq; 510824b8a847STejun Heo struct pool_workqueue *pwq; 5109a0a1a5fdSTejun Heo 511068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5111a0a1a5fdSTejun Heo 51126183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5113a0a1a5fdSTejun Heo workqueue_freezing = true; 5114a0a1a5fdSTejun Heo 511524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5116a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5117699ce097STejun Heo for_each_pwq(pwq, wq) 5118699ce097STejun Heo pwq_adjust_max_active(pwq); 5119a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5120a1056305STejun Heo } 51215bcab335STejun Heo 512268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5123a0a1a5fdSTejun Heo } 5124a0a1a5fdSTejun Heo 5125a0a1a5fdSTejun Heo /** 512658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5127a0a1a5fdSTejun Heo * 5128a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5129a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5130a0a1a5fdSTejun Heo * 5131a0a1a5fdSTejun Heo * CONTEXT: 513268e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5133a0a1a5fdSTejun Heo * 5134d185af30SYacine Belkadi * Return: 513558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 513658a69cb4STejun Heo * is complete. 5137a0a1a5fdSTejun Heo */ 5138a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5139a0a1a5fdSTejun Heo { 5140a0a1a5fdSTejun Heo bool busy = false; 514124b8a847STejun Heo struct workqueue_struct *wq; 514224b8a847STejun Heo struct pool_workqueue *pwq; 5143a0a1a5fdSTejun Heo 514468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5145a0a1a5fdSTejun Heo 51466183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5147a0a1a5fdSTejun Heo 514824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 514924b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 515024b8a847STejun Heo continue; 5151a0a1a5fdSTejun Heo /* 5152a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5153a0a1a5fdSTejun Heo * to peek without lock. 5154a0a1a5fdSTejun Heo */ 515588109453SLai Jiangshan rcu_read_lock_sched(); 515624b8a847STejun Heo for_each_pwq(pwq, wq) { 51576183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5158112202d9STejun Heo if (pwq->nr_active) { 5159a0a1a5fdSTejun Heo busy = true; 516088109453SLai Jiangshan rcu_read_unlock_sched(); 5161a0a1a5fdSTejun Heo goto out_unlock; 5162a0a1a5fdSTejun Heo } 5163a0a1a5fdSTejun Heo } 516488109453SLai Jiangshan rcu_read_unlock_sched(); 5165a0a1a5fdSTejun Heo } 5166a0a1a5fdSTejun Heo out_unlock: 516768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5168a0a1a5fdSTejun Heo return busy; 5169a0a1a5fdSTejun Heo } 5170a0a1a5fdSTejun Heo 5171a0a1a5fdSTejun Heo /** 5172a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5173a0a1a5fdSTejun Heo * 5174a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5175706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5176a0a1a5fdSTejun Heo * 5177a0a1a5fdSTejun Heo * CONTEXT: 5178a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5179a0a1a5fdSTejun Heo */ 5180a0a1a5fdSTejun Heo void thaw_workqueues(void) 5181a0a1a5fdSTejun Heo { 518224b8a847STejun Heo struct workqueue_struct *wq; 518324b8a847STejun Heo struct pool_workqueue *pwq; 5184a0a1a5fdSTejun Heo 518568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5186a0a1a5fdSTejun Heo 5187a0a1a5fdSTejun Heo if (!workqueue_freezing) 5188a0a1a5fdSTejun Heo goto out_unlock; 5189a0a1a5fdSTejun Heo 519074b414eaSLai Jiangshan workqueue_freezing = false; 519124b8a847STejun Heo 519224b8a847STejun Heo /* restore max_active and repopulate worklist */ 519324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5194a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5195699ce097STejun Heo for_each_pwq(pwq, wq) 5196699ce097STejun Heo pwq_adjust_max_active(pwq); 5197a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 519824b8a847STejun Heo } 519924b8a847STejun Heo 5200a0a1a5fdSTejun Heo out_unlock: 520168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5202a0a1a5fdSTejun Heo } 5203a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5204a0a1a5fdSTejun Heo 5205042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5206042f7df1SLai Jiangshan { 5207042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5208042f7df1SLai Jiangshan int ret = 0; 5209042f7df1SLai Jiangshan struct workqueue_struct *wq; 5210042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5211042f7df1SLai Jiangshan 5212042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5213042f7df1SLai Jiangshan 5214042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5215042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5216042f7df1SLai Jiangshan continue; 5217042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5218042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5219042f7df1SLai Jiangshan continue; 5220042f7df1SLai Jiangshan 5221042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5222042f7df1SLai Jiangshan if (!ctx) { 5223042f7df1SLai Jiangshan ret = -ENOMEM; 5224042f7df1SLai Jiangshan break; 5225042f7df1SLai Jiangshan } 5226042f7df1SLai Jiangshan 5227042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5228042f7df1SLai Jiangshan } 5229042f7df1SLai Jiangshan 5230042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5231042f7df1SLai Jiangshan if (!ret) 5232042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5233042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5234042f7df1SLai Jiangshan } 5235042f7df1SLai Jiangshan 5236042f7df1SLai Jiangshan return ret; 5237042f7df1SLai Jiangshan } 5238042f7df1SLai Jiangshan 5239042f7df1SLai Jiangshan /** 5240042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5241042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5242042f7df1SLai Jiangshan * 5243042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5244042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5245042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5246042f7df1SLai Jiangshan * 5247042f7df1SLai Jiangshan * Retun: 0 - Success 5248042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5249042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5250042f7df1SLai Jiangshan */ 5251042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5252042f7df1SLai Jiangshan { 5253042f7df1SLai Jiangshan int ret = -EINVAL; 5254042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5255042f7df1SLai Jiangshan 5256042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 5257042f7df1SLai Jiangshan return -ENOMEM; 5258042f7df1SLai Jiangshan 5259c98a9805STal Shorer /* 5260c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5261c98a9805STal Shorer * If the user wishes to include them, we allow that. 5262c98a9805STal Shorer */ 5263042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5264042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5265a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5266042f7df1SLai Jiangshan 5267042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5268042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5269042f7df1SLai Jiangshan 5270042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5271042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5272042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5273042f7df1SLai Jiangshan 5274042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5275042f7df1SLai Jiangshan if (ret < 0) 5276042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5277042f7df1SLai Jiangshan 5278a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5279042f7df1SLai Jiangshan } 5280042f7df1SLai Jiangshan 5281042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 5282042f7df1SLai Jiangshan return ret; 5283042f7df1SLai Jiangshan } 5284042f7df1SLai Jiangshan 52856ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 52866ba94429SFrederic Weisbecker /* 52876ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 52886ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 52896ba94429SFrederic Weisbecker * following attributes. 52906ba94429SFrederic Weisbecker * 52916ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 52926ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 52936ba94429SFrederic Weisbecker * 52946ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 52956ba94429SFrederic Weisbecker * 52969a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 52976ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 52986ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 52999a19b463SWang Long * numa RW bool : whether enable NUMA affinity 53006ba94429SFrederic Weisbecker */ 53016ba94429SFrederic Weisbecker struct wq_device { 53026ba94429SFrederic Weisbecker struct workqueue_struct *wq; 53036ba94429SFrederic Weisbecker struct device dev; 53046ba94429SFrederic Weisbecker }; 53056ba94429SFrederic Weisbecker 53066ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 53076ba94429SFrederic Weisbecker { 53086ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 53096ba94429SFrederic Weisbecker 53106ba94429SFrederic Weisbecker return wq_dev->wq; 53116ba94429SFrederic Weisbecker } 53126ba94429SFrederic Weisbecker 53136ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 53146ba94429SFrederic Weisbecker char *buf) 53156ba94429SFrederic Weisbecker { 53166ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53176ba94429SFrederic Weisbecker 53186ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 53196ba94429SFrederic Weisbecker } 53206ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 53216ba94429SFrederic Weisbecker 53226ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 53236ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 53246ba94429SFrederic Weisbecker { 53256ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53266ba94429SFrederic Weisbecker 53276ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 53286ba94429SFrederic Weisbecker } 53296ba94429SFrederic Weisbecker 53306ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 53316ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 53326ba94429SFrederic Weisbecker size_t count) 53336ba94429SFrederic Weisbecker { 53346ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53356ba94429SFrederic Weisbecker int val; 53366ba94429SFrederic Weisbecker 53376ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 53386ba94429SFrederic Weisbecker return -EINVAL; 53396ba94429SFrederic Weisbecker 53406ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 53416ba94429SFrederic Weisbecker return count; 53426ba94429SFrederic Weisbecker } 53436ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 53446ba94429SFrederic Weisbecker 53456ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 53466ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 53476ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 53486ba94429SFrederic Weisbecker NULL, 53496ba94429SFrederic Weisbecker }; 53506ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 53516ba94429SFrederic Weisbecker 53526ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 53536ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 53546ba94429SFrederic Weisbecker { 53556ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53566ba94429SFrederic Weisbecker const char *delim = ""; 53576ba94429SFrederic Weisbecker int node, written = 0; 53586ba94429SFrederic Weisbecker 53596ba94429SFrederic Weisbecker rcu_read_lock_sched(); 53606ba94429SFrederic Weisbecker for_each_node(node) { 53616ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 53626ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 53636ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 53646ba94429SFrederic Weisbecker delim = " "; 53656ba94429SFrederic Weisbecker } 53666ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 53676ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 53686ba94429SFrederic Weisbecker 53696ba94429SFrederic Weisbecker return written; 53706ba94429SFrederic Weisbecker } 53716ba94429SFrederic Weisbecker 53726ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 53736ba94429SFrederic Weisbecker char *buf) 53746ba94429SFrederic Weisbecker { 53756ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 53766ba94429SFrederic Weisbecker int written; 53776ba94429SFrederic Weisbecker 53786ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 53796ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 53806ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 53816ba94429SFrederic Weisbecker 53826ba94429SFrederic Weisbecker return written; 53836ba94429SFrederic Weisbecker } 53846ba94429SFrederic Weisbecker 53856ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 53866ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 53876ba94429SFrederic Weisbecker { 53886ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 53896ba94429SFrederic Weisbecker 5390899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5391899a94feSLai Jiangshan 53926ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 53936ba94429SFrederic Weisbecker if (!attrs) 53946ba94429SFrederic Weisbecker return NULL; 53956ba94429SFrederic Weisbecker 53966ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 53976ba94429SFrederic Weisbecker return attrs; 53986ba94429SFrederic Weisbecker } 53996ba94429SFrederic Weisbecker 54006ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 54016ba94429SFrederic Weisbecker const char *buf, size_t count) 54026ba94429SFrederic Weisbecker { 54036ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54046ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5405d4d3e257SLai Jiangshan int ret = -ENOMEM; 5406d4d3e257SLai Jiangshan 5407d4d3e257SLai Jiangshan apply_wqattrs_lock(); 54086ba94429SFrederic Weisbecker 54096ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 54106ba94429SFrederic Weisbecker if (!attrs) 5411d4d3e257SLai Jiangshan goto out_unlock; 54126ba94429SFrederic Weisbecker 54136ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 54146ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5415d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 54166ba94429SFrederic Weisbecker else 54176ba94429SFrederic Weisbecker ret = -EINVAL; 54186ba94429SFrederic Weisbecker 5419d4d3e257SLai Jiangshan out_unlock: 5420d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 54216ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 54226ba94429SFrederic Weisbecker return ret ?: count; 54236ba94429SFrederic Weisbecker } 54246ba94429SFrederic Weisbecker 54256ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 54266ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54276ba94429SFrederic Weisbecker { 54286ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54296ba94429SFrederic Weisbecker int written; 54306ba94429SFrederic Weisbecker 54316ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 54326ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 54336ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 54346ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 54356ba94429SFrederic Weisbecker return written; 54366ba94429SFrederic Weisbecker } 54376ba94429SFrederic Weisbecker 54386ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 54396ba94429SFrederic Weisbecker struct device_attribute *attr, 54406ba94429SFrederic Weisbecker const char *buf, size_t count) 54416ba94429SFrederic Weisbecker { 54426ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54436ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5444d4d3e257SLai Jiangshan int ret = -ENOMEM; 5445d4d3e257SLai Jiangshan 5446d4d3e257SLai Jiangshan apply_wqattrs_lock(); 54476ba94429SFrederic Weisbecker 54486ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 54496ba94429SFrederic Weisbecker if (!attrs) 5450d4d3e257SLai Jiangshan goto out_unlock; 54516ba94429SFrederic Weisbecker 54526ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 54536ba94429SFrederic Weisbecker if (!ret) 5454d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 54556ba94429SFrederic Weisbecker 5456d4d3e257SLai Jiangshan out_unlock: 5457d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 54586ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 54596ba94429SFrederic Weisbecker return ret ?: count; 54606ba94429SFrederic Weisbecker } 54616ba94429SFrederic Weisbecker 54626ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 54636ba94429SFrederic Weisbecker char *buf) 54646ba94429SFrederic Weisbecker { 54656ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54666ba94429SFrederic Weisbecker int written; 54676ba94429SFrederic Weisbecker 54686ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 54696ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 54706ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 54716ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 54726ba94429SFrederic Weisbecker 54736ba94429SFrederic Weisbecker return written; 54746ba94429SFrederic Weisbecker } 54756ba94429SFrederic Weisbecker 54766ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 54776ba94429SFrederic Weisbecker const char *buf, size_t count) 54786ba94429SFrederic Weisbecker { 54796ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54806ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5481d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5482d4d3e257SLai Jiangshan 5483d4d3e257SLai Jiangshan apply_wqattrs_lock(); 54846ba94429SFrederic Weisbecker 54856ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 54866ba94429SFrederic Weisbecker if (!attrs) 5487d4d3e257SLai Jiangshan goto out_unlock; 54886ba94429SFrederic Weisbecker 54896ba94429SFrederic Weisbecker ret = -EINVAL; 54906ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 54916ba94429SFrederic Weisbecker attrs->no_numa = !v; 5492d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 54936ba94429SFrederic Weisbecker } 54946ba94429SFrederic Weisbecker 5495d4d3e257SLai Jiangshan out_unlock: 5496d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 54976ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 54986ba94429SFrederic Weisbecker return ret ?: count; 54996ba94429SFrederic Weisbecker } 55006ba94429SFrederic Weisbecker 55016ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 55026ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 55036ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 55046ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 55056ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 55066ba94429SFrederic Weisbecker __ATTR_NULL, 55076ba94429SFrederic Weisbecker }; 55086ba94429SFrederic Weisbecker 55096ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 55106ba94429SFrederic Weisbecker .name = "workqueue", 55116ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 55126ba94429SFrederic Weisbecker }; 55136ba94429SFrederic Weisbecker 5514b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5515b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5516b05a7928SFrederic Weisbecker { 5517b05a7928SFrederic Weisbecker int written; 5518b05a7928SFrederic Weisbecker 5519042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5520b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5521b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5522042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5523b05a7928SFrederic Weisbecker 5524b05a7928SFrederic Weisbecker return written; 5525b05a7928SFrederic Weisbecker } 5526b05a7928SFrederic Weisbecker 5527042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5528042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5529042f7df1SLai Jiangshan { 5530042f7df1SLai Jiangshan cpumask_var_t cpumask; 5531042f7df1SLai Jiangshan int ret; 5532042f7df1SLai Jiangshan 5533042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5534042f7df1SLai Jiangshan return -ENOMEM; 5535042f7df1SLai Jiangshan 5536042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5537042f7df1SLai Jiangshan if (!ret) 5538042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5539042f7df1SLai Jiangshan 5540042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5541042f7df1SLai Jiangshan return ret ? ret : count; 5542042f7df1SLai Jiangshan } 5543042f7df1SLai Jiangshan 5544b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5545042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5546042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5547b05a7928SFrederic Weisbecker 55486ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 55496ba94429SFrederic Weisbecker { 5550b05a7928SFrederic Weisbecker int err; 5551b05a7928SFrederic Weisbecker 5552b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5553b05a7928SFrederic Weisbecker if (err) 5554b05a7928SFrederic Weisbecker return err; 5555b05a7928SFrederic Weisbecker 5556b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 55576ba94429SFrederic Weisbecker } 55586ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 55596ba94429SFrederic Weisbecker 55606ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 55616ba94429SFrederic Weisbecker { 55626ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 55636ba94429SFrederic Weisbecker 55646ba94429SFrederic Weisbecker kfree(wq_dev); 55656ba94429SFrederic Weisbecker } 55666ba94429SFrederic Weisbecker 55676ba94429SFrederic Weisbecker /** 55686ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 55696ba94429SFrederic Weisbecker * @wq: the workqueue to register 55706ba94429SFrederic Weisbecker * 55716ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 55726ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 55736ba94429SFrederic Weisbecker * which is the preferred method. 55746ba94429SFrederic Weisbecker * 55756ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 55766ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 55776ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 55786ba94429SFrederic Weisbecker * attributes. 55796ba94429SFrederic Weisbecker * 55806ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 55816ba94429SFrederic Weisbecker */ 55826ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 55836ba94429SFrederic Weisbecker { 55846ba94429SFrederic Weisbecker struct wq_device *wq_dev; 55856ba94429SFrederic Weisbecker int ret; 55866ba94429SFrederic Weisbecker 55876ba94429SFrederic Weisbecker /* 5588402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 55896ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 55906ba94429SFrederic Weisbecker * workqueues. 55916ba94429SFrederic Weisbecker */ 55920a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 55936ba94429SFrederic Weisbecker return -EINVAL; 55946ba94429SFrederic Weisbecker 55956ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 55966ba94429SFrederic Weisbecker if (!wq_dev) 55976ba94429SFrederic Weisbecker return -ENOMEM; 55986ba94429SFrederic Weisbecker 55996ba94429SFrederic Weisbecker wq_dev->wq = wq; 56006ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 56016ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 560223217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 56036ba94429SFrederic Weisbecker 56046ba94429SFrederic Weisbecker /* 56056ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 56066ba94429SFrederic Weisbecker * everything is ready. 56076ba94429SFrederic Weisbecker */ 56086ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 56096ba94429SFrederic Weisbecker 56106ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 56116ba94429SFrederic Weisbecker if (ret) { 5612537f4146SArvind Yadav put_device(&wq_dev->dev); 56136ba94429SFrederic Weisbecker wq->wq_dev = NULL; 56146ba94429SFrederic Weisbecker return ret; 56156ba94429SFrederic Weisbecker } 56166ba94429SFrederic Weisbecker 56176ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 56186ba94429SFrederic Weisbecker struct device_attribute *attr; 56196ba94429SFrederic Weisbecker 56206ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 56216ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 56226ba94429SFrederic Weisbecker if (ret) { 56236ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 56246ba94429SFrederic Weisbecker wq->wq_dev = NULL; 56256ba94429SFrederic Weisbecker return ret; 56266ba94429SFrederic Weisbecker } 56276ba94429SFrederic Weisbecker } 56286ba94429SFrederic Weisbecker } 56296ba94429SFrederic Weisbecker 56306ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 56316ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 56326ba94429SFrederic Weisbecker return 0; 56336ba94429SFrederic Weisbecker } 56346ba94429SFrederic Weisbecker 56356ba94429SFrederic Weisbecker /** 56366ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 56376ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 56386ba94429SFrederic Weisbecker * 56396ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 56406ba94429SFrederic Weisbecker */ 56416ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 56426ba94429SFrederic Weisbecker { 56436ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 56446ba94429SFrederic Weisbecker 56456ba94429SFrederic Weisbecker if (!wq->wq_dev) 56466ba94429SFrederic Weisbecker return; 56476ba94429SFrederic Weisbecker 56486ba94429SFrederic Weisbecker wq->wq_dev = NULL; 56496ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 56506ba94429SFrederic Weisbecker } 56516ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 56526ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 56536ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 56546ba94429SFrederic Weisbecker 565582607adcSTejun Heo /* 565682607adcSTejun Heo * Workqueue watchdog. 565782607adcSTejun Heo * 565882607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 565982607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 566082607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 566182607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 566282607adcSTejun Heo * largely opaque. 566382607adcSTejun Heo * 566482607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 566582607adcSTejun Heo * state if some pools failed to make forward progress for a while where 566682607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 566782607adcSTejun Heo * 566882607adcSTejun Heo * This mechanism is controlled through the kernel parameter 566982607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 567082607adcSTejun Heo * corresponding sysfs parameter file. 567182607adcSTejun Heo */ 567282607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 567382607adcSTejun Heo 567482607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 56755cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 567682607adcSTejun Heo 567782607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 567882607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 567982607adcSTejun Heo 568082607adcSTejun Heo static void wq_watchdog_reset_touched(void) 568182607adcSTejun Heo { 568282607adcSTejun Heo int cpu; 568382607adcSTejun Heo 568482607adcSTejun Heo wq_watchdog_touched = jiffies; 568582607adcSTejun Heo for_each_possible_cpu(cpu) 568682607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 568782607adcSTejun Heo } 568882607adcSTejun Heo 56895cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 569082607adcSTejun Heo { 569182607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 569282607adcSTejun Heo bool lockup_detected = false; 569382607adcSTejun Heo struct worker_pool *pool; 569482607adcSTejun Heo int pi; 569582607adcSTejun Heo 569682607adcSTejun Heo if (!thresh) 569782607adcSTejun Heo return; 569882607adcSTejun Heo 569982607adcSTejun Heo rcu_read_lock(); 570082607adcSTejun Heo 570182607adcSTejun Heo for_each_pool(pool, pi) { 570282607adcSTejun Heo unsigned long pool_ts, touched, ts; 570382607adcSTejun Heo 570482607adcSTejun Heo if (list_empty(&pool->worklist)) 570582607adcSTejun Heo continue; 570682607adcSTejun Heo 570782607adcSTejun Heo /* get the latest of pool and touched timestamps */ 570882607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 570982607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 571082607adcSTejun Heo 571182607adcSTejun Heo if (time_after(pool_ts, touched)) 571282607adcSTejun Heo ts = pool_ts; 571382607adcSTejun Heo else 571482607adcSTejun Heo ts = touched; 571582607adcSTejun Heo 571682607adcSTejun Heo if (pool->cpu >= 0) { 571782607adcSTejun Heo unsigned long cpu_touched = 571882607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 571982607adcSTejun Heo pool->cpu)); 572082607adcSTejun Heo if (time_after(cpu_touched, ts)) 572182607adcSTejun Heo ts = cpu_touched; 572282607adcSTejun Heo } 572382607adcSTejun Heo 572482607adcSTejun Heo /* did we stall? */ 572582607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 572682607adcSTejun Heo lockup_detected = true; 572782607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 572882607adcSTejun Heo pr_cont_pool_info(pool); 572982607adcSTejun Heo pr_cont(" stuck for %us!\n", 573082607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 573182607adcSTejun Heo } 573282607adcSTejun Heo } 573382607adcSTejun Heo 573482607adcSTejun Heo rcu_read_unlock(); 573582607adcSTejun Heo 573682607adcSTejun Heo if (lockup_detected) 573782607adcSTejun Heo show_workqueue_state(); 573882607adcSTejun Heo 573982607adcSTejun Heo wq_watchdog_reset_touched(); 574082607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 574182607adcSTejun Heo } 574282607adcSTejun Heo 5743cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 574482607adcSTejun Heo { 574582607adcSTejun Heo if (cpu >= 0) 574682607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 574782607adcSTejun Heo else 574882607adcSTejun Heo wq_watchdog_touched = jiffies; 574982607adcSTejun Heo } 575082607adcSTejun Heo 575182607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 575282607adcSTejun Heo { 575382607adcSTejun Heo wq_watchdog_thresh = 0; 575482607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 575582607adcSTejun Heo 575682607adcSTejun Heo if (thresh) { 575782607adcSTejun Heo wq_watchdog_thresh = thresh; 575882607adcSTejun Heo wq_watchdog_reset_touched(); 575982607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 576082607adcSTejun Heo } 576182607adcSTejun Heo } 576282607adcSTejun Heo 576382607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 576482607adcSTejun Heo const struct kernel_param *kp) 576582607adcSTejun Heo { 576682607adcSTejun Heo unsigned long thresh; 576782607adcSTejun Heo int ret; 576882607adcSTejun Heo 576982607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 577082607adcSTejun Heo if (ret) 577182607adcSTejun Heo return ret; 577282607adcSTejun Heo 577382607adcSTejun Heo if (system_wq) 577482607adcSTejun Heo wq_watchdog_set_thresh(thresh); 577582607adcSTejun Heo else 577682607adcSTejun Heo wq_watchdog_thresh = thresh; 577782607adcSTejun Heo 577882607adcSTejun Heo return 0; 577982607adcSTejun Heo } 578082607adcSTejun Heo 578182607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 578282607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 578382607adcSTejun Heo .get = param_get_ulong, 578482607adcSTejun Heo }; 578582607adcSTejun Heo 578682607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 578782607adcSTejun Heo 0644); 578882607adcSTejun Heo 578982607adcSTejun Heo static void wq_watchdog_init(void) 579082607adcSTejun Heo { 57915cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 579282607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 579382607adcSTejun Heo } 579482607adcSTejun Heo 579582607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 579682607adcSTejun Heo 579782607adcSTejun Heo static inline void wq_watchdog_init(void) { } 579882607adcSTejun Heo 579982607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 580082607adcSTejun Heo 5801bce90380STejun Heo static void __init wq_numa_init(void) 5802bce90380STejun Heo { 5803bce90380STejun Heo cpumask_var_t *tbl; 5804bce90380STejun Heo int node, cpu; 5805bce90380STejun Heo 5806bce90380STejun Heo if (num_possible_nodes() <= 1) 5807bce90380STejun Heo return; 5808bce90380STejun Heo 5809d55262c4STejun Heo if (wq_disable_numa) { 5810d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5811d55262c4STejun Heo return; 5812d55262c4STejun Heo } 5813d55262c4STejun Heo 58144c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 58154c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 58164c16bd32STejun Heo 5817bce90380STejun Heo /* 5818bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5819bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5820bce90380STejun Heo * fully initialized by now. 5821bce90380STejun Heo */ 58226396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5823bce90380STejun Heo BUG_ON(!tbl); 5824bce90380STejun Heo 5825bce90380STejun Heo for_each_node(node) 58265a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 58271be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5828bce90380STejun Heo 5829bce90380STejun Heo for_each_possible_cpu(cpu) { 5830bce90380STejun Heo node = cpu_to_node(cpu); 5831bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5832bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5833bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5834bce90380STejun Heo return; 5835bce90380STejun Heo } 5836bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5837bce90380STejun Heo } 5838bce90380STejun Heo 5839bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5840bce90380STejun Heo wq_numa_enabled = true; 5841bce90380STejun Heo } 5842bce90380STejun Heo 58433347fa09STejun Heo /** 58443347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 58453347fa09STejun Heo * 58463347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 58473347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 58483347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 58493347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 58503347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 58513347fa09STejun Heo * created and scheduled right before early initcalls. 58523347fa09STejun Heo */ 58533347fa09STejun Heo int __init workqueue_init_early(void) 58541da177e4SLinus Torvalds { 58557a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 58561bda3f80SFrederic Weisbecker int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 58577a4e344cSTejun Heo int i, cpu; 5858c34056a3STejun Heo 5859e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5860e904e6c2STejun Heo 5861b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 58621bda3f80SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 5863b05a7928SFrederic Weisbecker 5864e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5865e904e6c2STejun Heo 5866706026c2STejun Heo /* initialize CPU pools */ 586729c91e99STejun Heo for_each_possible_cpu(cpu) { 58684ce62e9eSTejun Heo struct worker_pool *pool; 58698b03ae3cSTejun Heo 58707a4e344cSTejun Heo i = 0; 5871f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 58727a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5873ec22ca5eSTejun Heo pool->cpu = cpu; 58747a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 58757a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5876f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 58777a4e344cSTejun Heo 58789daf9e67STejun Heo /* alloc pool ID */ 587968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 58809daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 588168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 58824ce62e9eSTejun Heo } 58838b03ae3cSTejun Heo } 58848b03ae3cSTejun Heo 58858a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 588629c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 588729c91e99STejun Heo struct workqueue_attrs *attrs; 588829c91e99STejun Heo 588929c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 589029c91e99STejun Heo attrs->nice = std_nice[i]; 589129c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 58928a2b7538STejun Heo 58938a2b7538STejun Heo /* 58948a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 58958a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 58968a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 58978a2b7538STejun Heo */ 58988a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 58998a2b7538STejun Heo attrs->nice = std_nice[i]; 59008a2b7538STejun Heo attrs->no_numa = true; 59018a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 590229c91e99STejun Heo } 590329c91e99STejun Heo 5904d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 59051aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5906d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5907f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5908f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 590924d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 591024d51addSTejun Heo WQ_FREEZABLE, 0); 59110668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 59120668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 59130668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 59140668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 59150668106cSViresh Kumar 0); 59161aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 59170668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 59180668106cSViresh Kumar !system_power_efficient_wq || 59190668106cSViresh Kumar !system_freezable_power_efficient_wq); 592082607adcSTejun Heo 59213347fa09STejun Heo return 0; 59223347fa09STejun Heo } 59233347fa09STejun Heo 59243347fa09STejun Heo /** 59253347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 59263347fa09STejun Heo * 59273347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 59283347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 59293347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 59303347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 59313347fa09STejun Heo * with the initial workers and enable future kworker creations. 59323347fa09STejun Heo */ 59333347fa09STejun Heo int __init workqueue_init(void) 59343347fa09STejun Heo { 59352186d9f9STejun Heo struct workqueue_struct *wq; 59363347fa09STejun Heo struct worker_pool *pool; 59373347fa09STejun Heo int cpu, bkt; 59383347fa09STejun Heo 59392186d9f9STejun Heo /* 59402186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 59412186d9f9STejun Heo * CPU to node mapping may not be available that early on some 59422186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 59432186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 59442186d9f9STejun Heo * affinity, fix them up. 594540c17f75STejun Heo * 594640c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 59472186d9f9STejun Heo */ 59482186d9f9STejun Heo wq_numa_init(); 59492186d9f9STejun Heo 59502186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 59512186d9f9STejun Heo 59522186d9f9STejun Heo for_each_possible_cpu(cpu) { 59532186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 59542186d9f9STejun Heo pool->node = cpu_to_node(cpu); 59552186d9f9STejun Heo } 59562186d9f9STejun Heo } 59572186d9f9STejun Heo 595840c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 59592186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 596040c17f75STejun Heo WARN(init_rescuer(wq), 596140c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 596240c17f75STejun Heo wq->name); 596340c17f75STejun Heo } 59642186d9f9STejun Heo 59652186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 59662186d9f9STejun Heo 59673347fa09STejun Heo /* create the initial workers */ 59683347fa09STejun Heo for_each_online_cpu(cpu) { 59693347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 59703347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 59713347fa09STejun Heo BUG_ON(!create_worker(pool)); 59723347fa09STejun Heo } 59733347fa09STejun Heo } 59743347fa09STejun Heo 59753347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 59763347fa09STejun Heo BUG_ON(!create_worker(pool)); 59773347fa09STejun Heo 59783347fa09STejun Heo wq_online = true; 597982607adcSTejun Heo wq_watchdog_init(); 598082607adcSTejun Heo 59816ee0578bSSuresh Siddha return 0; 59821da177e4SLinus Torvalds } 5983