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> 51e22bee78STejun Heo 52ea138446STejun Heo #include "workqueue_internal.h" 531da177e4SLinus Torvalds 54c8e55f36STejun Heo enum { 55bc2ae0f5STejun Heo /* 5624647570STejun Heo * worker_pool flags 57bc2ae0f5STejun Heo * 5824647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 59bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 60bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 61bc2ae0f5STejun Heo * is in effect. 62bc2ae0f5STejun Heo * 63bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 64bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6524647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 66bc2ae0f5STejun Heo * 67bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 6892f9c5c4SLai Jiangshan * attach_mutex to avoid changing binding state while 694736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 70bc2ae0f5STejun Heo */ 71692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7224647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 73db7bccf4STejun Heo 74c8e55f36STejun Heo /* worker flags */ 75c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 76c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 77e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 78fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 79f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 80a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 81e22bee78STejun Heo 82a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 83a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 84db7bccf4STejun Heo 85e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 864ce62e9eSTejun Heo 8729c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 88c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 89db7bccf4STejun Heo 90e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 91e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 92e22bee78STejun Heo 933233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 943233cdbdSTejun Heo /* call for help after 10ms 953233cdbdSTejun Heo (min two ticks) */ 96e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 97e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 981da177e4SLinus Torvalds 991da177e4SLinus Torvalds /* 100e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1018698a745SDongsheng Yang * all cpus. Give MIN_NICE. 102e22bee78STejun Heo */ 1038698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1048698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 105ecf6881fSTejun Heo 106ecf6881fSTejun Heo WQ_NAME_LEN = 24, 107c8e55f36STejun Heo }; 108c8e55f36STejun Heo 1091da177e4SLinus Torvalds /* 1104690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1114690c4abSTejun Heo * 112e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 113e41e704bSTejun Heo * everyone else. 1144690c4abSTejun Heo * 115e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 116e22bee78STejun Heo * only be modified and accessed from the local cpu. 117e22bee78STejun Heo * 118d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1194690c4abSTejun Heo * 120d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 121d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 122d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 123d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 124e22bee78STejun Heo * 12592f9c5c4SLai Jiangshan * A: pool->attach_mutex protected. 126822d8405STejun Heo * 12768e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12876af4d93STejun Heo * 12968e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1305bcab335STejun Heo * 1315b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1325b95e1afSLai Jiangshan * 1335b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 1345b95e1afSLai Jiangshan * sched-RCU for reads. 1355b95e1afSLai Jiangshan * 1363c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1373c25a55dSLai Jiangshan * 138b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1392e109a28STejun Heo * 1402e109a28STejun Heo * MD: wq_mayday_lock protected. 1414690c4abSTejun Heo */ 1424690c4abSTejun Heo 1432eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 144c34056a3STejun Heo 145bd7bdd43STejun Heo struct worker_pool { 146d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 147d84ff051STejun Heo int cpu; /* I: the associated cpu */ 148f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1499daf9e67STejun Heo int id; /* I: pool ID */ 15011ebea50STejun Heo unsigned int flags; /* X: flags */ 151bd7bdd43STejun Heo 15282607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15382607adcSTejun Heo 154bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 155bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 156ea1abd61SLai Jiangshan 157ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 158bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 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 168bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 1692607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17092f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 17192f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17260f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 173e19e397aSTejun Heo 1747cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 175e19e397aSTejun Heo 1767a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17768e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17868e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1797a4e344cSTejun Heo 180e19e397aSTejun Heo /* 181e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 182e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 183e19e397aSTejun Heo * cacheline. 184e19e397aSTejun Heo */ 185e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18629c91e99STejun Heo 18729c91e99STejun Heo /* 18829c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18929c91e99STejun Heo * from get_work_pool(). 19029c91e99STejun Heo */ 19129c91e99STejun Heo struct rcu_head rcu; 1928b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1938b03ae3cSTejun Heo 1948b03ae3cSTejun Heo /* 195112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 196112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 197112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 198112202d9STejun Heo * number of flag bits. 1991da177e4SLinus Torvalds */ 200112202d9STejun Heo struct pool_workqueue { 201bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2024690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20373f53c4aSTejun Heo int work_color; /* L: current color */ 20473f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2058864b4e5STejun Heo int refcnt; /* L: reference count */ 20673f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20773f53c4aSTejun Heo /* L: nr of in_flight works */ 2081e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 209a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2101e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2113c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2122e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2138864b4e5STejun Heo 2148864b4e5STejun Heo /* 2158864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2168864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2178864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 218b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2198864b4e5STejun Heo */ 2208864b4e5STejun Heo struct work_struct unbound_release_work; 2218864b4e5STejun Heo struct rcu_head rcu; 222e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* 22573f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22673f53c4aSTejun Heo */ 22773f53c4aSTejun Heo struct wq_flusher { 2283c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2293c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 23073f53c4aSTejun Heo struct completion done; /* flush completion */ 23173f53c4aSTejun Heo }; 2321da177e4SLinus Torvalds 233226223abSTejun Heo struct wq_device; 234226223abSTejun Heo 23573f53c4aSTejun Heo /* 236c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 237c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2381da177e4SLinus Torvalds */ 2391da177e4SLinus Torvalds struct workqueue_struct { 2403c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 241e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 24273f53c4aSTejun Heo 2433c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2443c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2453c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 246112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2473c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2483c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2493c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 25073f53c4aSTejun Heo 2512e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 252e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 253e22bee78STejun Heo 25487fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 255a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 256226223abSTejun Heo 2575b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2585b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2596029a918STejun Heo 260226223abSTejun Heo #ifdef CONFIG_SYSFS 261226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 262226223abSTejun Heo #endif 2634e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 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 */ 3012e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 302692b4825STejun Heo static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ 3035bcab335STejun Heo 304e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30568e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3067d19c5ceSTejun Heo 307ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 308ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 309ef557180SMike Galbraith 310ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 311ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 312b05a7928SFrederic Weisbecker 313f303fccbSTejun Heo /* 314f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 315f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 316f303fccbSTejun Heo * to uncover usages which depend on it. 317f303fccbSTejun Heo */ 318f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 319f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 320f303fccbSTejun Heo #else 321f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 322f303fccbSTejun Heo #endif 323f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 324f303fccbSTejun Heo 3257d19c5ceSTejun Heo /* the per-cpu worker pools */ 32625528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3277d19c5ceSTejun Heo 32868e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3297d19c5ceSTejun Heo 33068e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 33129c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 33229c91e99STejun Heo 333c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33429c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33529c91e99STejun Heo 3368a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3378a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3388a2b7538STejun Heo 339d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 340ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 341044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3421aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 343044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 344d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 345044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 346f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 347044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 34824d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3490668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3500668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3510668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3520668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 353d320c038STejun Heo 3547d19c5ceSTejun Heo static int worker_thread(void *__worker); 3556ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3567d19c5ceSTejun Heo 35797bd2347STejun Heo #define CREATE_TRACE_POINTS 35897bd2347STejun Heo #include <trace/events/workqueue.h> 35997bd2347STejun Heo 36068e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 361f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 362f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36368e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3645bcab335STejun Heo 365b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 366f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 367f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 368b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 36976af4d93STejun Heo 3705b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 371f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 372f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 373f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3745b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3755b95e1afSLai Jiangshan 376f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 377f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 378f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3797a62c2c8STejun Heo (pool)++) 3804ce62e9eSTejun Heo 38149e3cf44STejun Heo /** 38217116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38317116969STejun Heo * @pool: iteration cursor 384611c92a0STejun Heo * @pi: integer used for iteration 385fa1b54e6STejun Heo * 38668e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 38768e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38868e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 389fa1b54e6STejun Heo * 390fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 391fa1b54e6STejun Heo * ignored. 39217116969STejun Heo */ 393611c92a0STejun Heo #define for_each_pool(pool, pi) \ 394611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39568e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 396fa1b54e6STejun Heo else 39717116969STejun Heo 39817116969STejun Heo /** 399822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 400822d8405STejun Heo * @worker: iteration cursor 401822d8405STejun Heo * @pool: worker_pool to iterate workers of 402822d8405STejun Heo * 40392f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 404822d8405STejun Heo * 405822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 406822d8405STejun Heo * ignored. 407822d8405STejun Heo */ 408da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 409da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 41092f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 411822d8405STejun Heo else 412822d8405STejun Heo 413822d8405STejun Heo /** 41449e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41549e3cf44STejun Heo * @pwq: iteration cursor 41649e3cf44STejun Heo * @wq: the target workqueue 41776af4d93STejun Heo * 418b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 419794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 420794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 42176af4d93STejun Heo * 42276af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42376af4d93STejun Heo * ignored. 42449e3cf44STejun Heo */ 42549e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42676af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 427b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 42876af4d93STejun Heo else 429f3421797STejun Heo 430dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 431dc186ad7SThomas Gleixner 432dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 433dc186ad7SThomas Gleixner 43499777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43599777288SStanislaw Gruszka { 43699777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43799777288SStanislaw Gruszka } 43899777288SStanislaw Gruszka 439b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 440b9fdac7fSDu, Changbin { 441b9fdac7fSDu, Changbin struct work_struct *work = addr; 442b9fdac7fSDu, Changbin 443b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 444b9fdac7fSDu, Changbin } 445b9fdac7fSDu, Changbin 446dc186ad7SThomas Gleixner /* 447dc186ad7SThomas Gleixner * fixup_init is called when: 448dc186ad7SThomas Gleixner * - an active object is initialized 449dc186ad7SThomas Gleixner */ 45002a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 451dc186ad7SThomas Gleixner { 452dc186ad7SThomas Gleixner struct work_struct *work = addr; 453dc186ad7SThomas Gleixner 454dc186ad7SThomas Gleixner switch (state) { 455dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 456dc186ad7SThomas Gleixner cancel_work_sync(work); 457dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45802a982a6SDu, Changbin return true; 459dc186ad7SThomas Gleixner default: 46002a982a6SDu, Changbin return false; 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner } 463dc186ad7SThomas Gleixner 464dc186ad7SThomas Gleixner /* 465dc186ad7SThomas Gleixner * fixup_free is called when: 466dc186ad7SThomas Gleixner * - an active object is freed 467dc186ad7SThomas Gleixner */ 46802a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 469dc186ad7SThomas Gleixner { 470dc186ad7SThomas Gleixner struct work_struct *work = addr; 471dc186ad7SThomas Gleixner 472dc186ad7SThomas Gleixner switch (state) { 473dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 474dc186ad7SThomas Gleixner cancel_work_sync(work); 475dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47602a982a6SDu, Changbin return true; 477dc186ad7SThomas Gleixner default: 47802a982a6SDu, Changbin return false; 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner } 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 483dc186ad7SThomas Gleixner .name = "work_struct", 48499777288SStanislaw Gruszka .debug_hint = work_debug_hint, 485b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 486dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 487dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 488dc186ad7SThomas Gleixner }; 489dc186ad7SThomas Gleixner 490dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 491dc186ad7SThomas Gleixner { 492dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 493dc186ad7SThomas Gleixner } 494dc186ad7SThomas Gleixner 495dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 496dc186ad7SThomas Gleixner { 497dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 498dc186ad7SThomas Gleixner } 499dc186ad7SThomas Gleixner 500dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 501dc186ad7SThomas Gleixner { 502dc186ad7SThomas Gleixner if (onstack) 503dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 504dc186ad7SThomas Gleixner else 505dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 506dc186ad7SThomas Gleixner } 507dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 508dc186ad7SThomas Gleixner 509dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 510dc186ad7SThomas Gleixner { 511dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 512dc186ad7SThomas Gleixner } 513dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 514dc186ad7SThomas Gleixner 515ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 516ea2e64f2SThomas Gleixner { 517ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 518ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 519ea2e64f2SThomas Gleixner } 520ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 521ea2e64f2SThomas Gleixner 522dc186ad7SThomas Gleixner #else 523dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 524dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 525dc186ad7SThomas Gleixner #endif 526dc186ad7SThomas Gleixner 5274e8b22bdSLi Bin /** 5284e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5294e8b22bdSLi Bin * @pool: the pool pointer of interest 5304e8b22bdSLi Bin * 5314e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5324e8b22bdSLi Bin * successfully, -errno on failure. 5334e8b22bdSLi Bin */ 5349daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5359daf9e67STejun Heo { 5369daf9e67STejun Heo int ret; 5379daf9e67STejun Heo 53868e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5395bcab335STejun Heo 5404e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5414e8b22bdSLi Bin GFP_KERNEL); 542229641a6STejun Heo if (ret >= 0) { 543e68035fbSTejun Heo pool->id = ret; 544229641a6STejun Heo return 0; 545229641a6STejun Heo } 5469daf9e67STejun Heo return ret; 5479daf9e67STejun Heo } 5489daf9e67STejun Heo 54976af4d93STejun Heo /** 550df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 551df2d5ae4STejun Heo * @wq: the target workqueue 552df2d5ae4STejun Heo * @node: the node ID 553df2d5ae4STejun Heo * 5545b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5555b95e1afSLai Jiangshan * read locked. 556df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 557df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 558d185af30SYacine Belkadi * 559d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 560df2d5ae4STejun Heo */ 561df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 562df2d5ae4STejun Heo int node) 563df2d5ae4STejun Heo { 5645b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 565d6e022f1STejun Heo 566d6e022f1STejun Heo /* 567d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 568d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 569d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 570d6e022f1STejun Heo * happens, this workaround can be removed. 571d6e022f1STejun Heo */ 572d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 573d6e022f1STejun Heo return wq->dfl_pwq; 574d6e022f1STejun Heo 575df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 576df2d5ae4STejun Heo } 577df2d5ae4STejun Heo 57873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57973f53c4aSTejun Heo { 58073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 58173f53c4aSTejun Heo } 58273f53c4aSTejun Heo 58373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58473f53c4aSTejun Heo { 58573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58773f53c4aSTejun Heo } 58873f53c4aSTejun Heo 58973f53c4aSTejun Heo static int work_next_color(int color) 59073f53c4aSTejun Heo { 59173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 592a848e3b6SOleg Nesterov } 593a848e3b6SOleg Nesterov 5944594bf15SDavid Howells /* 595112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 596112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5977c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5987a22ad75STejun Heo * 599112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 600112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 601bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 602bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6037a22ad75STejun Heo * 604112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6057c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 606112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6077c3eed5cSTejun Heo * available only while the work item is queued. 608bbb68dfaSTejun Heo * 609bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 610bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 611bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 612bbb68dfaSTejun Heo * try to steal the PENDING bit. 6134594bf15SDavid Howells */ 6147a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6157a22ad75STejun Heo unsigned long flags) 6167a22ad75STejun Heo { 6176183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6187a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6197a22ad75STejun Heo } 6207a22ad75STejun Heo 621112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6224690c4abSTejun Heo unsigned long extra_flags) 623365970a1SDavid Howells { 624112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 625112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 626365970a1SDavid Howells } 627365970a1SDavid Howells 6284468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6294468a00fSLai Jiangshan int pool_id) 6304468a00fSLai Jiangshan { 6314468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6324468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6334468a00fSLai Jiangshan } 6344468a00fSLai Jiangshan 6357c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6367c3eed5cSTejun Heo int pool_id) 6374d707b9fSOleg Nesterov { 63823657bb1STejun Heo /* 63923657bb1STejun Heo * The following wmb is paired with the implied mb in 64023657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 64123657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64223657bb1STejun Heo * owner. 64323657bb1STejun Heo */ 64423657bb1STejun Heo smp_wmb(); 6457c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 646346c09f8SRoman Pen /* 647346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 648346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 649346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 650346c09f8SRoman Pen * reordering can lead to a missed execution on attempt to qeueue 651346c09f8SRoman Pen * the same @work. E.g. consider this case: 652346c09f8SRoman Pen * 653346c09f8SRoman Pen * CPU#0 CPU#1 654346c09f8SRoman Pen * ---------------------------- -------------------------------- 655346c09f8SRoman Pen * 656346c09f8SRoman Pen * 1 STORE event_indicated 657346c09f8SRoman Pen * 2 queue_work_on() { 658346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 659346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 660346c09f8SRoman Pen * 5 set_work_data() # clear bit 661346c09f8SRoman Pen * 6 smp_mb() 662346c09f8SRoman Pen * 7 work->current_func() { 663346c09f8SRoman Pen * 8 LOAD event_indicated 664346c09f8SRoman Pen * } 665346c09f8SRoman Pen * 666346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 667346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 668346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 669346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 670346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 671346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 672346c09f8SRoman Pen * before actual STORE. 673346c09f8SRoman Pen */ 674346c09f8SRoman Pen smp_mb(); 6754d707b9fSOleg Nesterov } 6764d707b9fSOleg Nesterov 6777a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 678365970a1SDavid Howells { 6797c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6807c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6817a22ad75STejun Heo } 6827a22ad75STejun Heo 683112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6847a22ad75STejun Heo { 685e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6867a22ad75STejun Heo 687112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 688e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 689e120153dSTejun Heo else 690e120153dSTejun Heo return NULL; 6917a22ad75STejun Heo } 6927a22ad75STejun Heo 6937c3eed5cSTejun Heo /** 6947c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6957c3eed5cSTejun Heo * @work: the work item of interest 6967c3eed5cSTejun Heo * 69768e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69868e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 69968e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 700fa1b54e6STejun Heo * 701fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 702fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 703fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 704fa1b54e6STejun Heo * returned pool is and stays online. 705d185af30SYacine Belkadi * 706d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7077c3eed5cSTejun Heo */ 7087c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7097a22ad75STejun Heo { 710e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7117c3eed5cSTejun Heo int pool_id; 7127a22ad75STejun Heo 71368e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 714fa1b54e6STejun Heo 715112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 716112202d9STejun Heo return ((struct pool_workqueue *) 7177c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7187a22ad75STejun Heo 7197c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7207c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7217a22ad75STejun Heo return NULL; 7227a22ad75STejun Heo 723fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7247c3eed5cSTejun Heo } 7257c3eed5cSTejun Heo 7267c3eed5cSTejun Heo /** 7277c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7287c3eed5cSTejun Heo * @work: the work item of interest 7297c3eed5cSTejun Heo * 730d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7317c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7327c3eed5cSTejun Heo */ 7337c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7347c3eed5cSTejun Heo { 73554d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7367c3eed5cSTejun Heo 737112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 738112202d9STejun Heo return ((struct pool_workqueue *) 73954d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 74054d5b7d0SLai Jiangshan 74154d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7427c3eed5cSTejun Heo } 7437c3eed5cSTejun Heo 744bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 745bbb68dfaSTejun Heo { 7467c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 747bbb68dfaSTejun Heo 7487c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7497c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 750bbb68dfaSTejun Heo } 751bbb68dfaSTejun Heo 752bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 753bbb68dfaSTejun Heo { 754bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 755bbb68dfaSTejun Heo 756112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 757bbb68dfaSTejun Heo } 758bbb68dfaSTejun Heo 759e22bee78STejun Heo /* 7603270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7613270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 762d565ed63STejun Heo * they're being called with pool->lock held. 763e22bee78STejun Heo */ 764e22bee78STejun Heo 76563d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 766649027d7STejun Heo { 767e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 768649027d7STejun Heo } 769649027d7STejun Heo 770e22bee78STejun Heo /* 771e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 772e22bee78STejun Heo * running workers. 773974271c4STejun Heo * 774974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 775706026c2STejun Heo * function will always return %true for unbound pools as long as the 776974271c4STejun Heo * worklist isn't empty. 777e22bee78STejun Heo */ 77863d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 779e22bee78STejun Heo { 78063d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 781e22bee78STejun Heo } 782e22bee78STejun Heo 783e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78463d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 785e22bee78STejun Heo { 78663d95a91STejun Heo return pool->nr_idle; 787e22bee78STejun Heo } 788e22bee78STejun Heo 789e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 79063d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 791e22bee78STejun Heo { 792e19e397aSTejun Heo return !list_empty(&pool->worklist) && 793e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 794e22bee78STejun Heo } 795e22bee78STejun Heo 796e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79763d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 798e22bee78STejun Heo { 79963d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 800e22bee78STejun Heo } 801e22bee78STejun Heo 802e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 804e22bee78STejun Heo { 805692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 80663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 808e22bee78STejun Heo 809e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 810e22bee78STejun Heo } 811e22bee78STejun Heo 812e22bee78STejun Heo /* 813e22bee78STejun Heo * Wake up functions. 814e22bee78STejun Heo */ 815e22bee78STejun Heo 8161037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8171037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8187e11629dSTejun Heo { 81963d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8207e11629dSTejun Heo return NULL; 8217e11629dSTejun Heo 82263d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8237e11629dSTejun Heo } 8247e11629dSTejun Heo 8257e11629dSTejun Heo /** 8267e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82763d95a91STejun Heo * @pool: worker pool to wake worker from 8287e11629dSTejun Heo * 82963d95a91STejun Heo * Wake up the first idle worker of @pool. 8307e11629dSTejun Heo * 8317e11629dSTejun Heo * CONTEXT: 832d565ed63STejun Heo * spin_lock_irq(pool->lock). 8337e11629dSTejun Heo */ 83463d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8357e11629dSTejun Heo { 8361037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8377e11629dSTejun Heo 8387e11629dSTejun Heo if (likely(worker)) 8397e11629dSTejun Heo wake_up_process(worker->task); 8407e11629dSTejun Heo } 8417e11629dSTejun Heo 8424690c4abSTejun Heo /** 843e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 844e22bee78STejun Heo * @task: task waking up 845e22bee78STejun Heo * @cpu: CPU @task is waking up to 846e22bee78STejun Heo * 847e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 848e22bee78STejun Heo * being awoken. 849e22bee78STejun Heo * 850e22bee78STejun Heo * CONTEXT: 851e22bee78STejun Heo * spin_lock_irq(rq->lock) 852e22bee78STejun Heo */ 853d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 854e22bee78STejun Heo { 855e22bee78STejun Heo struct worker *worker = kthread_data(task); 856e22bee78STejun Heo 85736576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 858ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 859e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 860e22bee78STejun Heo } 86136576000SJoonsoo Kim } 862e22bee78STejun Heo 863e22bee78STejun Heo /** 864e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 865e22bee78STejun Heo * @task: task going to sleep 866e22bee78STejun Heo * 867e22bee78STejun Heo * This function is called during schedule() when a busy worker is 868e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 869e22bee78STejun Heo * returning pointer to its task. 870e22bee78STejun Heo * 871e22bee78STejun Heo * CONTEXT: 872e22bee78STejun Heo * spin_lock_irq(rq->lock) 873e22bee78STejun Heo * 874d185af30SYacine Belkadi * Return: 875e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 876e22bee78STejun Heo */ 8779b7f6597SAlexander Gordeev struct task_struct *wq_worker_sleeping(struct task_struct *task) 878e22bee78STejun Heo { 879e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 880111c225aSTejun Heo struct worker_pool *pool; 881e22bee78STejun Heo 882111c225aSTejun Heo /* 883111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 884111c225aSTejun Heo * workers, also reach here, let's not access anything before 885111c225aSTejun Heo * checking NOT_RUNNING. 886111c225aSTejun Heo */ 8872d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 888e22bee78STejun Heo return NULL; 889e22bee78STejun Heo 890111c225aSTejun Heo pool = worker->pool; 891111c225aSTejun Heo 892e22bee78STejun Heo /* this can only happen on the local cpu */ 8939b7f6597SAlexander Gordeev if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) 8946183c009STejun Heo return NULL; 895e22bee78STejun Heo 896e22bee78STejun Heo /* 897e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 898e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 899e22bee78STejun Heo * Please read comment there. 900e22bee78STejun Heo * 901628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 902628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 903628c78e7STejun Heo * disabled, which in turn means that none else could be 904d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 905628c78e7STejun Heo * lock is safe. 906e22bee78STejun Heo */ 907e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 908e19e397aSTejun Heo !list_empty(&pool->worklist)) 9091037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 910e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 911e22bee78STejun Heo } 912e22bee78STejun Heo 913e22bee78STejun Heo /** 914e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 915cb444766STejun Heo * @worker: self 916d302f017STejun Heo * @flags: flags to set 917d302f017STejun Heo * 918228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 919d302f017STejun Heo * 920cb444766STejun Heo * CONTEXT: 921d565ed63STejun Heo * spin_lock_irq(pool->lock) 922d302f017STejun Heo */ 923228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 924d302f017STejun Heo { 925bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 926e22bee78STejun Heo 927cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 928cb444766STejun Heo 929228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 930e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 931e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 932e19e397aSTejun Heo atomic_dec(&pool->nr_running); 933e22bee78STejun Heo } 934e22bee78STejun Heo 935d302f017STejun Heo worker->flags |= flags; 936d302f017STejun Heo } 937d302f017STejun Heo 938d302f017STejun Heo /** 939e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 940cb444766STejun Heo * @worker: self 941d302f017STejun Heo * @flags: flags to clear 942d302f017STejun Heo * 943e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 944d302f017STejun Heo * 945cb444766STejun Heo * CONTEXT: 946d565ed63STejun Heo * spin_lock_irq(pool->lock) 947d302f017STejun Heo */ 948d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 949d302f017STejun Heo { 95063d95a91STejun Heo struct worker_pool *pool = worker->pool; 951e22bee78STejun Heo unsigned int oflags = worker->flags; 952e22bee78STejun Heo 953cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 954cb444766STejun Heo 955d302f017STejun Heo worker->flags &= ~flags; 956e22bee78STejun Heo 95742c025f3STejun Heo /* 95842c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 95942c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 96042c025f3STejun Heo * of multiple flags, not a single flag. 96142c025f3STejun Heo */ 962e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 963e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 964e19e397aSTejun Heo atomic_inc(&pool->nr_running); 965d302f017STejun Heo } 966d302f017STejun Heo 967d302f017STejun Heo /** 9688cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 969c9e7cf27STejun Heo * @pool: pool of interest 9708cca0eeaSTejun Heo * @work: work to find worker for 9718cca0eeaSTejun Heo * 972c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 973c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 974a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 975a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 976a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 977a2c1c57bSTejun Heo * being executed. 978a2c1c57bSTejun Heo * 979a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 980a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 981a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 982a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 983a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 984a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 985a2c1c57bSTejun Heo * 986c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 987c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 988c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 989c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 990c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 991c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9928cca0eeaSTejun Heo * 9938cca0eeaSTejun Heo * CONTEXT: 994d565ed63STejun Heo * spin_lock_irq(pool->lock). 9958cca0eeaSTejun Heo * 996d185af30SYacine Belkadi * Return: 997d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9988cca0eeaSTejun Heo * otherwise. 9998cca0eeaSTejun Heo */ 1000c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10018cca0eeaSTejun Heo struct work_struct *work) 10028cca0eeaSTejun Heo { 100342f8570fSSasha Levin struct worker *worker; 100442f8570fSSasha Levin 1005b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1006a2c1c57bSTejun Heo (unsigned long)work) 1007a2c1c57bSTejun Heo if (worker->current_work == work && 1008a2c1c57bSTejun Heo worker->current_func == work->func) 100942f8570fSSasha Levin return worker; 101042f8570fSSasha Levin 101142f8570fSSasha Levin return NULL; 10128cca0eeaSTejun Heo } 10138cca0eeaSTejun Heo 10148cca0eeaSTejun Heo /** 1015bf4ede01STejun Heo * move_linked_works - move linked works to a list 1016bf4ede01STejun Heo * @work: start of series of works to be scheduled 1017bf4ede01STejun Heo * @head: target list to append @work to 1018402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1019bf4ede01STejun Heo * 1020bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1021bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1022bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1023bf4ede01STejun Heo * 1024bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1025bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1026bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1027bf4ede01STejun Heo * 1028bf4ede01STejun Heo * CONTEXT: 1029d565ed63STejun Heo * spin_lock_irq(pool->lock). 1030bf4ede01STejun Heo */ 1031bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1032bf4ede01STejun Heo struct work_struct **nextp) 1033bf4ede01STejun Heo { 1034bf4ede01STejun Heo struct work_struct *n; 1035bf4ede01STejun Heo 1036bf4ede01STejun Heo /* 1037bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1038bf4ede01STejun Heo * use NULL for list head. 1039bf4ede01STejun Heo */ 1040bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1041bf4ede01STejun Heo list_move_tail(&work->entry, head); 1042bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1043bf4ede01STejun Heo break; 1044bf4ede01STejun Heo } 1045bf4ede01STejun Heo 1046bf4ede01STejun Heo /* 1047bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1048bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1049bf4ede01STejun Heo * needs to be updated. 1050bf4ede01STejun Heo */ 1051bf4ede01STejun Heo if (nextp) 1052bf4ede01STejun Heo *nextp = n; 1053bf4ede01STejun Heo } 1054bf4ede01STejun Heo 10558864b4e5STejun Heo /** 10568864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10578864b4e5STejun Heo * @pwq: pool_workqueue to get 10588864b4e5STejun Heo * 10598864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10608864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10618864b4e5STejun Heo */ 10628864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10638864b4e5STejun Heo { 10648864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10658864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10668864b4e5STejun Heo pwq->refcnt++; 10678864b4e5STejun Heo } 10688864b4e5STejun Heo 10698864b4e5STejun Heo /** 10708864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10718864b4e5STejun Heo * @pwq: pool_workqueue to put 10728864b4e5STejun Heo * 10738864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10748864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10758864b4e5STejun Heo */ 10768864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10778864b4e5STejun Heo { 10788864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10798864b4e5STejun Heo if (likely(--pwq->refcnt)) 10808864b4e5STejun Heo return; 10818864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10828864b4e5STejun Heo return; 10838864b4e5STejun Heo /* 10848864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10858864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10868864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10878864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10888864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10898864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10908864b4e5STejun Heo */ 10918864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10928864b4e5STejun Heo } 10938864b4e5STejun Heo 1094dce90d47STejun Heo /** 1095dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1096dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1097dce90d47STejun Heo * 1098dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1099dce90d47STejun Heo */ 1100dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1101dce90d47STejun Heo { 1102dce90d47STejun Heo if (pwq) { 1103dce90d47STejun Heo /* 1104dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1105dce90d47STejun Heo * following lock operations are safe. 1106dce90d47STejun Heo */ 1107dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1108dce90d47STejun Heo put_pwq(pwq); 1109dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1110dce90d47STejun Heo } 1111dce90d47STejun Heo } 1112dce90d47STejun Heo 1113112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1114bf4ede01STejun Heo { 1115112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1116bf4ede01STejun Heo 1117bf4ede01STejun Heo trace_workqueue_activate_work(work); 111882607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 111982607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1120112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1121bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1122112202d9STejun Heo pwq->nr_active++; 1123bf4ede01STejun Heo } 1124bf4ede01STejun Heo 1125112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11263aa62497SLai Jiangshan { 1127112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11283aa62497SLai Jiangshan struct work_struct, entry); 11293aa62497SLai Jiangshan 1130112202d9STejun Heo pwq_activate_delayed_work(work); 11313aa62497SLai Jiangshan } 11323aa62497SLai Jiangshan 1133bf4ede01STejun Heo /** 1134112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1135112202d9STejun Heo * @pwq: pwq of interest 1136bf4ede01STejun Heo * @color: color of work which left the queue 1137bf4ede01STejun Heo * 1138bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1139112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1140bf4ede01STejun Heo * 1141bf4ede01STejun Heo * CONTEXT: 1142d565ed63STejun Heo * spin_lock_irq(pool->lock). 1143bf4ede01STejun Heo */ 1144112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1145bf4ede01STejun Heo { 11468864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1147bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11488864b4e5STejun Heo goto out_put; 1149bf4ede01STejun Heo 1150112202d9STejun Heo pwq->nr_in_flight[color]--; 1151bf4ede01STejun Heo 1152112202d9STejun Heo pwq->nr_active--; 1153112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1154bf4ede01STejun Heo /* one down, submit a delayed one */ 1155112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1156112202d9STejun Heo pwq_activate_first_delayed(pwq); 1157bf4ede01STejun Heo } 1158bf4ede01STejun Heo 1159bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1160112202d9STejun Heo if (likely(pwq->flush_color != color)) 11618864b4e5STejun Heo goto out_put; 1162bf4ede01STejun Heo 1163bf4ede01STejun Heo /* are there still in-flight works? */ 1164112202d9STejun Heo if (pwq->nr_in_flight[color]) 11658864b4e5STejun Heo goto out_put; 1166bf4ede01STejun Heo 1167112202d9STejun Heo /* this pwq is done, clear flush_color */ 1168112202d9STejun Heo pwq->flush_color = -1; 1169bf4ede01STejun Heo 1170bf4ede01STejun Heo /* 1171112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1172bf4ede01STejun Heo * will handle the rest. 1173bf4ede01STejun Heo */ 1174112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1175112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11768864b4e5STejun Heo out_put: 11778864b4e5STejun Heo put_pwq(pwq); 1178bf4ede01STejun Heo } 1179bf4ede01STejun Heo 118036e227d2STejun Heo /** 1181bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 118236e227d2STejun Heo * @work: work item to steal 118336e227d2STejun Heo * @is_dwork: @work is a delayed_work 1184bbb68dfaSTejun Heo * @flags: place to store irq state 118536e227d2STejun Heo * 118636e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1187d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 118836e227d2STejun Heo * 1189d185af30SYacine Belkadi * Return: 119036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 119136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 119236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1193bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1194bbb68dfaSTejun Heo * for arbitrarily long 119536e227d2STejun Heo * 1196d185af30SYacine Belkadi * Note: 1197bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1198e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1199e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1200e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1201bbb68dfaSTejun Heo * 1202bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1203bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1204bbb68dfaSTejun Heo * 1205e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1206bf4ede01STejun Heo */ 1207bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1208bbb68dfaSTejun Heo unsigned long *flags) 1209bf4ede01STejun Heo { 1210d565ed63STejun Heo struct worker_pool *pool; 1211112202d9STejun Heo struct pool_workqueue *pwq; 1212bf4ede01STejun Heo 1213bbb68dfaSTejun Heo local_irq_save(*flags); 1214bbb68dfaSTejun Heo 121536e227d2STejun Heo /* try to steal the timer if it exists */ 121636e227d2STejun Heo if (is_dwork) { 121736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 121836e227d2STejun Heo 1219e0aecdd8STejun Heo /* 1220e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1221e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1222e0aecdd8STejun Heo * running on the local CPU. 1223e0aecdd8STejun Heo */ 122436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 122536e227d2STejun Heo return 1; 122636e227d2STejun Heo } 122736e227d2STejun Heo 122836e227d2STejun Heo /* try to claim PENDING the normal way */ 1229bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1230bf4ede01STejun Heo return 0; 1231bf4ede01STejun Heo 1232bf4ede01STejun Heo /* 1233bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1234bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1235bf4ede01STejun Heo */ 1236d565ed63STejun Heo pool = get_work_pool(work); 1237d565ed63STejun Heo if (!pool) 1238bbb68dfaSTejun Heo goto fail; 1239bf4ede01STejun Heo 1240d565ed63STejun Heo spin_lock(&pool->lock); 1241bf4ede01STejun Heo /* 1242112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1243112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1244112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1245112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1246112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12470b3dae68SLai Jiangshan * item is currently queued on that pool. 1248bf4ede01STejun Heo */ 1249112202d9STejun Heo pwq = get_work_pwq(work); 1250112202d9STejun Heo if (pwq && pwq->pool == pool) { 1251bf4ede01STejun Heo debug_work_deactivate(work); 12523aa62497SLai Jiangshan 12533aa62497SLai Jiangshan /* 125416062836STejun Heo * A delayed work item cannot be grabbed directly because 125516062836STejun Heo * it might have linked NO_COLOR work items which, if left 1256112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 125716062836STejun Heo * management later on and cause stall. Make sure the work 125816062836STejun Heo * item is activated before grabbing. 12593aa62497SLai Jiangshan */ 12603aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1261112202d9STejun Heo pwq_activate_delayed_work(work); 12623aa62497SLai Jiangshan 1263bf4ede01STejun Heo list_del_init(&work->entry); 12649c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 126536e227d2STejun Heo 1266112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12674468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12684468a00fSLai Jiangshan 1269d565ed63STejun Heo spin_unlock(&pool->lock); 127036e227d2STejun Heo return 1; 1271bf4ede01STejun Heo } 1272d565ed63STejun Heo spin_unlock(&pool->lock); 1273bbb68dfaSTejun Heo fail: 1274bbb68dfaSTejun Heo local_irq_restore(*flags); 1275bbb68dfaSTejun Heo if (work_is_canceling(work)) 1276bbb68dfaSTejun Heo return -ENOENT; 1277bbb68dfaSTejun Heo cpu_relax(); 127836e227d2STejun Heo return -EAGAIN; 1279bf4ede01STejun Heo } 1280bf4ede01STejun Heo 1281bf4ede01STejun Heo /** 1282706026c2STejun Heo * insert_work - insert a work into a pool 1283112202d9STejun Heo * @pwq: pwq @work belongs to 12844690c4abSTejun Heo * @work: work to insert 12854690c4abSTejun Heo * @head: insertion point 12864690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12874690c4abSTejun Heo * 1288112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1289706026c2STejun Heo * work_struct flags. 12904690c4abSTejun Heo * 12914690c4abSTejun Heo * CONTEXT: 1292d565ed63STejun Heo * spin_lock_irq(pool->lock). 1293365970a1SDavid Howells */ 1294112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1295112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1296b89deed3SOleg Nesterov { 1297112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1298e1d8aa9fSFrederic Weisbecker 12994690c4abSTejun Heo /* we own @work, set data and link */ 1300112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13011a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13028864b4e5STejun Heo get_pwq(pwq); 1303e22bee78STejun Heo 1304e22bee78STejun Heo /* 1305c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1306c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1307c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1308e22bee78STejun Heo */ 1309e22bee78STejun Heo smp_mb(); 1310e22bee78STejun Heo 131163d95a91STejun Heo if (__need_more_worker(pool)) 131263d95a91STejun Heo wake_up_worker(pool); 1313b89deed3SOleg Nesterov } 1314b89deed3SOleg Nesterov 1315c8efcc25STejun Heo /* 1316c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13178d03ecfeSTejun Heo * same workqueue. 1318c8efcc25STejun Heo */ 1319c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1320c8efcc25STejun Heo { 1321c8efcc25STejun Heo struct worker *worker; 1322c8efcc25STejun Heo 13238d03ecfeSTejun Heo worker = current_wq_worker(); 1324c8efcc25STejun Heo /* 13258d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13268d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1327c8efcc25STejun Heo */ 1328112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1329c8efcc25STejun Heo } 1330c8efcc25STejun Heo 1331ef557180SMike Galbraith /* 1332ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1333ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1334ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1335ef557180SMike Galbraith */ 1336ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1337ef557180SMike Galbraith { 1338f303fccbSTejun Heo static bool printed_dbg_warning; 1339ef557180SMike Galbraith int new_cpu; 1340ef557180SMike Galbraith 1341f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1342ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1343ef557180SMike Galbraith return cpu; 1344f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1345f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1346f303fccbSTejun Heo printed_dbg_warning = true; 1347f303fccbSTejun Heo } 1348f303fccbSTejun Heo 1349ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1350ef557180SMike Galbraith return cpu; 1351ef557180SMike Galbraith 1352ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1353ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1354ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1355ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1356ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1357ef557180SMike Galbraith return cpu; 1358ef557180SMike Galbraith } 1359ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1360ef557180SMike Galbraith 1361ef557180SMike Galbraith return new_cpu; 1362ef557180SMike Galbraith } 1363ef557180SMike Galbraith 1364d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13651da177e4SLinus Torvalds struct work_struct *work) 13661da177e4SLinus Torvalds { 1367112202d9STejun Heo struct pool_workqueue *pwq; 1368c9178087STejun Heo struct worker_pool *last_pool; 13691e19ffc6STejun Heo struct list_head *worklist; 13708a2e8e5dSTejun Heo unsigned int work_flags; 1371b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13728930cabaSTejun Heo 13738930cabaSTejun Heo /* 13748930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13758930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13768930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13778930cabaSTejun Heo * happen with IRQ disabled. 13788930cabaSTejun Heo */ 13798e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 13801da177e4SLinus Torvalds 1381dc186ad7SThomas Gleixner debug_work_activate(work); 13821e19ffc6STejun Heo 13839ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1384618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1385c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1386e41e704bSTejun Heo return; 13879e8cd2f5STejun Heo retry: 1388df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1389ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1390df2d5ae4STejun Heo 1391df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1392df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1393c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1394df2d5ae4STejun Heo else 1395df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1396f3421797STejun Heo 139718aa9effSTejun Heo /* 1398c9178087STejun Heo * If @work was previously on a different pool, it might still be 1399c9178087STejun Heo * running there, in which case the work needs to be queued on that 1400c9178087STejun Heo * pool to guarantee non-reentrancy. 140118aa9effSTejun Heo */ 1402c9e7cf27STejun Heo last_pool = get_work_pool(work); 1403112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 140418aa9effSTejun Heo struct worker *worker; 140518aa9effSTejun Heo 1406d565ed63STejun Heo spin_lock(&last_pool->lock); 140718aa9effSTejun Heo 1408c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 140918aa9effSTejun Heo 1410112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1411c9178087STejun Heo pwq = worker->current_pwq; 14128594fadeSLai Jiangshan } else { 141318aa9effSTejun Heo /* meh... not running there, queue here */ 1414d565ed63STejun Heo spin_unlock(&last_pool->lock); 1415112202d9STejun Heo spin_lock(&pwq->pool->lock); 141618aa9effSTejun Heo } 14178930cabaSTejun Heo } else { 1418112202d9STejun Heo spin_lock(&pwq->pool->lock); 14198930cabaSTejun Heo } 1420502ca9d8STejun Heo 14219e8cd2f5STejun Heo /* 14229e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14239e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14249e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1425df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1426df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14279e8cd2f5STejun Heo * make forward-progress. 14289e8cd2f5STejun Heo */ 14299e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14309e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 14319e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 14329e8cd2f5STejun Heo cpu_relax(); 14339e8cd2f5STejun Heo goto retry; 14349e8cd2f5STejun Heo } 14359e8cd2f5STejun Heo /* oops */ 14369e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14379e8cd2f5STejun Heo wq->name, cpu); 14389e8cd2f5STejun Heo } 14399e8cd2f5STejun Heo 1440112202d9STejun Heo /* pwq determined, queue */ 1441112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1442502ca9d8STejun Heo 1443f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1444112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1445f5b2552bSDan Carpenter return; 1446f5b2552bSDan Carpenter } 14471e19ffc6STejun Heo 1448112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1449112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14501e19ffc6STejun Heo 1451112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1452cdadf009STejun Heo trace_workqueue_activate_work(work); 1453112202d9STejun Heo pwq->nr_active++; 1454112202d9STejun Heo worklist = &pwq->pool->worklist; 145582607adcSTejun Heo if (list_empty(worklist)) 145682607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14578a2e8e5dSTejun Heo } else { 14588a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1459112202d9STejun Heo worklist = &pwq->delayed_works; 14608a2e8e5dSTejun Heo } 14611e19ffc6STejun Heo 1462112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14631e19ffc6STejun Heo 1464112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14651da177e4SLinus Torvalds } 14661da177e4SLinus Torvalds 14670fcb78c2SRolf Eike Beer /** 1468c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1469c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1470c1a220e7SZhang Rui * @wq: workqueue to use 1471c1a220e7SZhang Rui * @work: work to queue 1472c1a220e7SZhang Rui * 1473c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1474c1a220e7SZhang Rui * can't go away. 1475d185af30SYacine Belkadi * 1476d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1477c1a220e7SZhang Rui */ 1478d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1479d4283e93STejun Heo struct work_struct *work) 1480c1a220e7SZhang Rui { 1481d4283e93STejun Heo bool ret = false; 14828930cabaSTejun Heo unsigned long flags; 14838930cabaSTejun Heo 14848930cabaSTejun Heo local_irq_save(flags); 1485c1a220e7SZhang Rui 148622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14874690c4abSTejun Heo __queue_work(cpu, wq, work); 1488d4283e93STejun Heo ret = true; 1489c1a220e7SZhang Rui } 14908930cabaSTejun Heo 14918930cabaSTejun Heo local_irq_restore(flags); 1492c1a220e7SZhang Rui return ret; 1493c1a220e7SZhang Rui } 1494ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1495c1a220e7SZhang Rui 14968c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 14971da177e4SLinus Torvalds { 14988c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 14991da177e4SLinus Torvalds 1500e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 150160c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 15021da177e4SLinus Torvalds } 15031438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 15041da177e4SLinus Torvalds 15057beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 150652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15071da177e4SLinus Torvalds { 15087beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 15097beb2edfSTejun Heo struct work_struct *work = &dwork->work; 15101da177e4SLinus Torvalds 1511637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 1512841b86f3SKees Cook WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1513fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1514fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 15157beb2edfSTejun Heo 15168852aac2STejun Heo /* 15178852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 15188852aac2STejun Heo * both optimization and correctness. The earliest @timer can 15198852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 15208852aac2STejun Heo * on that there's no such delay when @delay is 0. 15218852aac2STejun Heo */ 15228852aac2STejun Heo if (!delay) { 15238852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 15248852aac2STejun Heo return; 15258852aac2STejun Heo } 15268852aac2STejun Heo 152760c057bcSLai Jiangshan dwork->wq = wq; 15281265057fSTejun Heo dwork->cpu = cpu; 15297beb2edfSTejun Heo timer->expires = jiffies + delay; 15307beb2edfSTejun Heo 1531041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 15327beb2edfSTejun Heo add_timer_on(timer, cpu); 1533041bd12eSTejun Heo else 1534041bd12eSTejun Heo add_timer(timer); 15357beb2edfSTejun Heo } 15361da177e4SLinus Torvalds 15370fcb78c2SRolf Eike Beer /** 15380fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 15390fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 15400fcb78c2SRolf Eike Beer * @wq: workqueue to use 1541af9997e4SRandy Dunlap * @dwork: work to queue 15420fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 15430fcb78c2SRolf Eike Beer * 1544d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1545715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1546715f1300STejun Heo * execution. 15470fcb78c2SRolf Eike Beer */ 1548d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 154952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15507a6bc1cdSVenkatesh Pallipadi { 155152bad64dSDavid Howells struct work_struct *work = &dwork->work; 1552d4283e93STejun Heo bool ret = false; 15538930cabaSTejun Heo unsigned long flags; 15548930cabaSTejun Heo 15558930cabaSTejun Heo /* read the comment in __queue_work() */ 15568930cabaSTejun Heo local_irq_save(flags); 15577a6bc1cdSVenkatesh Pallipadi 155822df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15597beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1560d4283e93STejun Heo ret = true; 15617a6bc1cdSVenkatesh Pallipadi } 15628930cabaSTejun Heo 15638930cabaSTejun Heo local_irq_restore(flags); 15647a6bc1cdSVenkatesh Pallipadi return ret; 15657a6bc1cdSVenkatesh Pallipadi } 1566ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15671da177e4SLinus Torvalds 1568c8e55f36STejun Heo /** 15698376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15708376fe22STejun Heo * @cpu: CPU number to execute work on 15718376fe22STejun Heo * @wq: workqueue to use 15728376fe22STejun Heo * @dwork: work to queue 15738376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15748376fe22STejun Heo * 15758376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15768376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15778376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15788376fe22STejun Heo * current state. 15798376fe22STejun Heo * 1580d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15818376fe22STejun Heo * pending and its timer was modified. 15828376fe22STejun Heo * 1583e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15848376fe22STejun Heo * See try_to_grab_pending() for details. 15858376fe22STejun Heo */ 15868376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15878376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15888376fe22STejun Heo { 15898376fe22STejun Heo unsigned long flags; 15908376fe22STejun Heo int ret; 15918376fe22STejun Heo 15928376fe22STejun Heo do { 15938376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15948376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15958376fe22STejun Heo 15968376fe22STejun Heo if (likely(ret >= 0)) { 15978376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15988376fe22STejun Heo local_irq_restore(flags); 15998376fe22STejun Heo } 16008376fe22STejun Heo 16018376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 16028376fe22STejun Heo return ret; 16038376fe22STejun Heo } 16048376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 16058376fe22STejun Heo 16068376fe22STejun Heo /** 1607c8e55f36STejun Heo * worker_enter_idle - enter idle state 1608c8e55f36STejun Heo * @worker: worker which is entering idle state 1609c8e55f36STejun Heo * 1610c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1611c8e55f36STejun Heo * necessary. 1612c8e55f36STejun Heo * 1613c8e55f36STejun Heo * LOCKING: 1614d565ed63STejun Heo * spin_lock_irq(pool->lock). 1615c8e55f36STejun Heo */ 1616c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 16171da177e4SLinus Torvalds { 1618bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1619c8e55f36STejun Heo 16206183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 16216183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 16226183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 16236183c009STejun Heo return; 1624c8e55f36STejun Heo 1625051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1626cb444766STejun Heo worker->flags |= WORKER_IDLE; 1627bd7bdd43STejun Heo pool->nr_idle++; 1628e22bee78STejun Heo worker->last_active = jiffies; 1629c8e55f36STejun Heo 1630c8e55f36STejun Heo /* idle_list is LIFO */ 1631bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1632db7bccf4STejun Heo 163363d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1634628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1635cb444766STejun Heo 1636544ecf31STejun Heo /* 1637e8b3f8dbSLai Jiangshan * Sanity check nr_running. Because unbind_workers() releases 1638d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1639628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1640628c78e7STejun Heo * unbind is not in progress. 1641544ecf31STejun Heo */ 164224647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1643bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1644e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1645c8e55f36STejun Heo } 1646c8e55f36STejun Heo 1647c8e55f36STejun Heo /** 1648c8e55f36STejun Heo * worker_leave_idle - leave idle state 1649c8e55f36STejun Heo * @worker: worker which is leaving idle state 1650c8e55f36STejun Heo * 1651c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1652c8e55f36STejun Heo * 1653c8e55f36STejun Heo * LOCKING: 1654d565ed63STejun Heo * spin_lock_irq(pool->lock). 1655c8e55f36STejun Heo */ 1656c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1657c8e55f36STejun Heo { 1658bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1659c8e55f36STejun Heo 16606183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16616183c009STejun Heo return; 1662d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1663bd7bdd43STejun Heo pool->nr_idle--; 1664c8e55f36STejun Heo list_del_init(&worker->entry); 1665c8e55f36STejun Heo } 1666c8e55f36STejun Heo 1667f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1668c34056a3STejun Heo { 1669c34056a3STejun Heo struct worker *worker; 1670c34056a3STejun Heo 1671f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1672c8e55f36STejun Heo if (worker) { 1673c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1674affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1675da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1676e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1677e22bee78STejun Heo worker->flags = WORKER_PREP; 1678c8e55f36STejun Heo } 1679c34056a3STejun Heo return worker; 1680c34056a3STejun Heo } 1681c34056a3STejun Heo 1682c34056a3STejun Heo /** 16834736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16844736cbf7SLai Jiangshan * @worker: worker to be attached 16854736cbf7SLai Jiangshan * @pool: the target pool 16864736cbf7SLai Jiangshan * 16874736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16884736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16894736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16904736cbf7SLai Jiangshan */ 16914736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16924736cbf7SLai Jiangshan struct worker_pool *pool) 16934736cbf7SLai Jiangshan { 16944736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16954736cbf7SLai Jiangshan 16964736cbf7SLai Jiangshan /* 16974736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16984736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16994736cbf7SLai Jiangshan */ 17004736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17014736cbf7SLai Jiangshan 17024736cbf7SLai Jiangshan /* 17034736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 17044736cbf7SLai Jiangshan * stable across this function. See the comments above the 17054736cbf7SLai Jiangshan * flag definition for details. 17064736cbf7SLai Jiangshan */ 17074736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 17084736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 17094736cbf7SLai Jiangshan 17104736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 17114736cbf7SLai Jiangshan 17124736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 17134736cbf7SLai Jiangshan } 17144736cbf7SLai Jiangshan 17154736cbf7SLai Jiangshan /** 171660f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 171760f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 171860f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 171960f5a4bcSLai Jiangshan * 17204736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 17214736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 17224736cbf7SLai Jiangshan * other reference to the pool. 172360f5a4bcSLai Jiangshan */ 172460f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 172560f5a4bcSLai Jiangshan struct worker_pool *pool) 172660f5a4bcSLai Jiangshan { 172760f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 172860f5a4bcSLai Jiangshan 172992f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1730da028469SLai Jiangshan list_del(&worker->node); 1731da028469SLai Jiangshan if (list_empty(&pool->workers)) 173260f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 173392f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 173460f5a4bcSLai Jiangshan 1735b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1736b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1737b62c0751SLai Jiangshan 173860f5a4bcSLai Jiangshan if (detach_completion) 173960f5a4bcSLai Jiangshan complete(detach_completion); 174060f5a4bcSLai Jiangshan } 174160f5a4bcSLai Jiangshan 174260f5a4bcSLai Jiangshan /** 1743c34056a3STejun Heo * create_worker - create a new workqueue worker 174463d95a91STejun Heo * @pool: pool the new worker will belong to 1745c34056a3STejun Heo * 1746051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1747c34056a3STejun Heo * 1748c34056a3STejun Heo * CONTEXT: 1749c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1750c34056a3STejun Heo * 1751d185af30SYacine Belkadi * Return: 1752c34056a3STejun Heo * Pointer to the newly created worker. 1753c34056a3STejun Heo */ 1754bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1755c34056a3STejun Heo { 1756c34056a3STejun Heo struct worker *worker = NULL; 1757f3421797STejun Heo int id = -1; 1758e3c916a4STejun Heo char id_buf[16]; 1759c34056a3STejun Heo 17607cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17617cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1762822d8405STejun Heo if (id < 0) 1763c34056a3STejun Heo goto fail; 1764c34056a3STejun Heo 1765f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1766c34056a3STejun Heo if (!worker) 1767c34056a3STejun Heo goto fail; 1768c34056a3STejun Heo 1769bd7bdd43STejun Heo worker->pool = pool; 1770c34056a3STejun Heo worker->id = id; 1771c34056a3STejun Heo 177229c91e99STejun Heo if (pool->cpu >= 0) 1773e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1774e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1775f3421797STejun Heo else 1776e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1777e3c916a4STejun Heo 1778f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1779e3c916a4STejun Heo "kworker/%s", id_buf); 1780c34056a3STejun Heo if (IS_ERR(worker->task)) 1781c34056a3STejun Heo goto fail; 1782c34056a3STejun Heo 178391151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 178425834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 178591151228SOleg Nesterov 1786da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17874736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1788822d8405STejun Heo 1789051e1850SLai Jiangshan /* start the newly created worker */ 1790051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1791051e1850SLai Jiangshan worker->pool->nr_workers++; 1792051e1850SLai Jiangshan worker_enter_idle(worker); 1793051e1850SLai Jiangshan wake_up_process(worker->task); 1794051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1795051e1850SLai Jiangshan 1796c34056a3STejun Heo return worker; 1797822d8405STejun Heo 1798c34056a3STejun Heo fail: 17999625ab17SLai Jiangshan if (id >= 0) 18007cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1801c34056a3STejun Heo kfree(worker); 1802c34056a3STejun Heo return NULL; 1803c34056a3STejun Heo } 1804c34056a3STejun Heo 1805c34056a3STejun Heo /** 1806c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1807c34056a3STejun Heo * @worker: worker to be destroyed 1808c34056a3STejun Heo * 180973eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 181073eb7fe7SLai Jiangshan * be idle. 1811c8e55f36STejun Heo * 1812c8e55f36STejun Heo * CONTEXT: 181360f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1814c34056a3STejun Heo */ 1815c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1816c34056a3STejun Heo { 1817bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1818c34056a3STejun Heo 1819cd549687STejun Heo lockdep_assert_held(&pool->lock); 1820cd549687STejun Heo 1821c34056a3STejun Heo /* sanity check frenzy */ 18226183c009STejun Heo if (WARN_ON(worker->current_work) || 182373eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 182473eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 18256183c009STejun Heo return; 1826c34056a3STejun Heo 1827bd7bdd43STejun Heo pool->nr_workers--; 1828bd7bdd43STejun Heo pool->nr_idle--; 1829c8e55f36STejun Heo 1830c8e55f36STejun Heo list_del_init(&worker->entry); 1831cb444766STejun Heo worker->flags |= WORKER_DIE; 183260f5a4bcSLai Jiangshan wake_up_process(worker->task); 1833c34056a3STejun Heo } 1834c34056a3STejun Heo 183532a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 1836e22bee78STejun Heo { 183732a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 1838e22bee78STejun Heo 1839d565ed63STejun Heo spin_lock_irq(&pool->lock); 1840e22bee78STejun Heo 18413347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1842e22bee78STejun Heo struct worker *worker; 1843e22bee78STejun Heo unsigned long expires; 1844e22bee78STejun Heo 1845e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 184663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1847e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1848e22bee78STejun Heo 18493347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 185063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18513347fc9fSLai Jiangshan break; 1852e22bee78STejun Heo } 18533347fc9fSLai Jiangshan 18543347fc9fSLai Jiangshan destroy_worker(worker); 1855e22bee78STejun Heo } 1856e22bee78STejun Heo 1857d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1858e22bee78STejun Heo } 1859e22bee78STejun Heo 1860493a1724STejun Heo static void send_mayday(struct work_struct *work) 1861e22bee78STejun Heo { 1862112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1863112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1864493a1724STejun Heo 18652e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1866e22bee78STejun Heo 1867493008a8STejun Heo if (!wq->rescuer) 1868493a1724STejun Heo return; 1869e22bee78STejun Heo 1870e22bee78STejun Heo /* mayday mayday mayday */ 1871493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 187277668c8bSLai Jiangshan /* 187377668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 187477668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 187577668c8bSLai Jiangshan * rescuer is done with it. 187677668c8bSLai Jiangshan */ 187777668c8bSLai Jiangshan get_pwq(pwq); 1878493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1879e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1880493a1724STejun Heo } 1881e22bee78STejun Heo } 1882e22bee78STejun Heo 188332a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 1884e22bee78STejun Heo { 188532a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 1886e22bee78STejun Heo struct work_struct *work; 1887e22bee78STejun Heo 1888b2d82909STejun Heo spin_lock_irq(&pool->lock); 1889b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1890e22bee78STejun Heo 189163d95a91STejun Heo if (need_to_create_worker(pool)) { 1892e22bee78STejun Heo /* 1893e22bee78STejun Heo * We've been trying to create a new worker but 1894e22bee78STejun Heo * haven't been successful. We might be hitting an 1895e22bee78STejun Heo * allocation deadlock. Send distress signals to 1896e22bee78STejun Heo * rescuers. 1897e22bee78STejun Heo */ 189863d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1899e22bee78STejun Heo send_mayday(work); 1900e22bee78STejun Heo } 1901e22bee78STejun Heo 1902b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1903b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1904e22bee78STejun Heo 190563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1906e22bee78STejun Heo } 1907e22bee78STejun Heo 1908e22bee78STejun Heo /** 1909e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 191063d95a91STejun Heo * @pool: pool to create a new worker for 1911e22bee78STejun Heo * 191263d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1913e22bee78STejun Heo * have at least one idle worker on return from this function. If 1914e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 191563d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1916e22bee78STejun Heo * possible allocation deadlock. 1917e22bee78STejun Heo * 1918c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1919c5aa87bbSTejun Heo * may_start_working() %true. 1920e22bee78STejun Heo * 1921e22bee78STejun Heo * LOCKING: 1922d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1923e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1924e22bee78STejun Heo * manager. 1925e22bee78STejun Heo */ 192629187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1927d565ed63STejun Heo __releases(&pool->lock) 1928d565ed63STejun Heo __acquires(&pool->lock) 1929e22bee78STejun Heo { 1930e22bee78STejun Heo restart: 1931d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19329f9c2364STejun Heo 1933e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 193463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1935e22bee78STejun Heo 1936e22bee78STejun Heo while (true) { 1937051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1938e22bee78STejun Heo break; 1939e22bee78STejun Heo 1940e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19419f9c2364STejun Heo 194263d95a91STejun Heo if (!need_to_create_worker(pool)) 1943e22bee78STejun Heo break; 1944e22bee78STejun Heo } 1945e22bee78STejun Heo 194663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1947d565ed63STejun Heo spin_lock_irq(&pool->lock); 1948051e1850SLai Jiangshan /* 1949051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1950051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1951051e1850SLai Jiangshan * already become busy. 1952051e1850SLai Jiangshan */ 195363d95a91STejun Heo if (need_to_create_worker(pool)) 1954e22bee78STejun Heo goto restart; 1955e22bee78STejun Heo } 1956e22bee78STejun Heo 1957e22bee78STejun Heo /** 1958e22bee78STejun Heo * manage_workers - manage worker pool 1959e22bee78STejun Heo * @worker: self 1960e22bee78STejun Heo * 1961706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1962e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1963706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1964e22bee78STejun Heo * 1965e22bee78STejun Heo * The caller can safely start processing works on false return. On 1966e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1967e22bee78STejun Heo * and may_start_working() is true. 1968e22bee78STejun Heo * 1969e22bee78STejun Heo * CONTEXT: 1970d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1971e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1972e22bee78STejun Heo * 1973d185af30SYacine Belkadi * Return: 197429187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 197529187a9eSTejun Heo * start processing works, %true if management function was performed and 197629187a9eSTejun Heo * the conditions that the caller verified before calling the function may 197729187a9eSTejun Heo * no longer be true. 1978e22bee78STejun Heo */ 1979e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1980e22bee78STejun Heo { 198163d95a91STejun Heo struct worker_pool *pool = worker->pool; 1982e22bee78STejun Heo 1983692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 198429187a9eSTejun Heo return false; 1985692b4825STejun Heo 1986692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 19872607d7a6STejun Heo pool->manager = worker; 1988e22bee78STejun Heo 198929187a9eSTejun Heo maybe_create_worker(pool); 1990e22bee78STejun Heo 19912607d7a6STejun Heo pool->manager = NULL; 1992692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 1993692b4825STejun Heo wake_up(&wq_manager_wait); 199429187a9eSTejun Heo return true; 1995e22bee78STejun Heo } 1996e22bee78STejun Heo 1997a62428c0STejun Heo /** 1998a62428c0STejun Heo * process_one_work - process single work 1999c34056a3STejun Heo * @worker: self 2000a62428c0STejun Heo * @work: work to process 2001a62428c0STejun Heo * 2002a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2003a62428c0STejun Heo * process a single work including synchronization against and 2004a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2005a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2006a62428c0STejun Heo * call this function to process a work. 2007a62428c0STejun Heo * 2008a62428c0STejun Heo * CONTEXT: 2009d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2010a62428c0STejun Heo */ 2011c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2012d565ed63STejun Heo __releases(&pool->lock) 2013d565ed63STejun Heo __acquires(&pool->lock) 20141da177e4SLinus Torvalds { 2015112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2016bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2017112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 201873f53c4aSTejun Heo int work_color; 20197e11629dSTejun Heo struct worker *collision; 20204e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20214e6045f1SJohannes Berg /* 2022a62428c0STejun Heo * It is permissible to free the struct work_struct from 2023a62428c0STejun Heo * inside the function that is called from it, this we need to 2024a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2025a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2026a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20274e6045f1SJohannes Berg */ 20284d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20294d82a1deSPeter Zijlstra 20304d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20314e6045f1SJohannes Berg #endif 2032807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 203385327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2034ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 203525511a47STejun Heo 20367e11629dSTejun Heo /* 20377e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20387e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20397e11629dSTejun Heo * already processing the work. If so, defer the work to the 20407e11629dSTejun Heo * currently executing one. 20417e11629dSTejun Heo */ 2042c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20437e11629dSTejun Heo if (unlikely(collision)) { 20447e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20457e11629dSTejun Heo return; 20467e11629dSTejun Heo } 20471da177e4SLinus Torvalds 20488930cabaSTejun Heo /* claim and dequeue */ 20491da177e4SLinus Torvalds debug_work_deactivate(work); 2050c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2051c34056a3STejun Heo worker->current_work = work; 2052a2c1c57bSTejun Heo worker->current_func = work->func; 2053112202d9STejun Heo worker->current_pwq = pwq; 205473f53c4aSTejun Heo work_color = get_work_color(work); 20557a22ad75STejun Heo 2056a62428c0STejun Heo list_del_init(&work->entry); 2057a62428c0STejun Heo 2058649027d7STejun Heo /* 2059228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2060228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2061228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2062228f1d00SLai Jiangshan * execution of the pending work items. 2063fb0e7bebSTejun Heo */ 2064fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2065228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2066fb0e7bebSTejun Heo 2067974271c4STejun Heo /* 2068a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2069a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2070a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2071a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2072228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2073974271c4STejun Heo */ 2074a489a03eSLai Jiangshan if (need_more_worker(pool)) 207563d95a91STejun Heo wake_up_worker(pool); 2076974271c4STejun Heo 20778930cabaSTejun Heo /* 20787c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2079d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 208023657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 208123657bb1STejun Heo * disabled. 20828930cabaSTejun Heo */ 20837c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20841da177e4SLinus Torvalds 2085d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2086365970a1SDavid Howells 2087a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 20883295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2089e6f3faa7SPeter Zijlstra /* 2090f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2091f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2092e6f3faa7SPeter Zijlstra * 2093e6f3faa7SPeter Zijlstra * However, that would result in: 2094e6f3faa7SPeter Zijlstra * 2095e6f3faa7SPeter Zijlstra * A(W1) 2096e6f3faa7SPeter Zijlstra * WFC(C) 2097e6f3faa7SPeter Zijlstra * A(W1) 2098e6f3faa7SPeter Zijlstra * C(C) 2099e6f3faa7SPeter Zijlstra * 2100e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2101e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2102e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2103f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2104e6f3faa7SPeter Zijlstra * these locks. 2105e6f3faa7SPeter Zijlstra * 2106e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2107e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2108e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2109e6f3faa7SPeter Zijlstra */ 2110f52be570SPeter Zijlstra lockdep_invariant_state(true); 2111e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2112a2c1c57bSTejun Heo worker->current_func(work); 2113e36c886aSArjan van de Ven /* 2114e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2115e36c886aSArjan van de Ven * point will only record its address. 2116e36c886aSArjan van de Ven */ 2117e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21183295f0efSIngo Molnar lock_map_release(&lockdep_map); 2119112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21201da177e4SLinus Torvalds 2121d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2122044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2123044c782cSValentin Ilie " last function: %pf\n", 2124a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2125a2c1c57bSTejun Heo worker->current_func); 2126d5abe669SPeter Zijlstra debug_show_held_locks(current); 2127d5abe669SPeter Zijlstra dump_stack(); 2128d5abe669SPeter Zijlstra } 2129d5abe669SPeter Zijlstra 2130b22ce278STejun Heo /* 2131b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2132b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2133b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2134b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2135789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2136789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2137b22ce278STejun Heo */ 21383e28e377SJoe Lawrence cond_resched_rcu_qs(); 2139b22ce278STejun Heo 2140d565ed63STejun Heo spin_lock_irq(&pool->lock); 2141a62428c0STejun Heo 2142fb0e7bebSTejun Heo /* clear cpu intensive status */ 2143fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2144fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2145fb0e7bebSTejun Heo 2146a62428c0STejun Heo /* we're done with it, release */ 214742f8570fSSasha Levin hash_del(&worker->hentry); 2148c34056a3STejun Heo worker->current_work = NULL; 2149a2c1c57bSTejun Heo worker->current_func = NULL; 2150112202d9STejun Heo worker->current_pwq = NULL; 21513d1cb205STejun Heo worker->desc_valid = false; 2152112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21531da177e4SLinus Torvalds } 21541da177e4SLinus Torvalds 2155affee4b2STejun Heo /** 2156affee4b2STejun Heo * process_scheduled_works - process scheduled works 2157affee4b2STejun Heo * @worker: self 2158affee4b2STejun Heo * 2159affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2160affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2161affee4b2STejun Heo * fetches a work from the top and executes it. 2162affee4b2STejun Heo * 2163affee4b2STejun Heo * CONTEXT: 2164d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2165affee4b2STejun Heo * multiple times. 2166affee4b2STejun Heo */ 2167affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21681da177e4SLinus Torvalds { 2169affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2170affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2171a62428c0STejun Heo struct work_struct, entry); 2172c34056a3STejun Heo process_one_work(worker, work); 2173a62428c0STejun Heo } 21741da177e4SLinus Torvalds } 21751da177e4SLinus Torvalds 21764690c4abSTejun Heo /** 21774690c4abSTejun Heo * worker_thread - the worker thread function 2178c34056a3STejun Heo * @__worker: self 21794690c4abSTejun Heo * 2180c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2181c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2182c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2183c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2184c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2185d185af30SYacine Belkadi * 2186d185af30SYacine Belkadi * Return: 0 21874690c4abSTejun Heo */ 2188c34056a3STejun Heo static int worker_thread(void *__worker) 21891da177e4SLinus Torvalds { 2190c34056a3STejun Heo struct worker *worker = __worker; 2191bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21921da177e4SLinus Torvalds 2193e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2194e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2195c8e55f36STejun Heo woke_up: 2196d565ed63STejun Heo spin_lock_irq(&pool->lock); 2197affee4b2STejun Heo 2198a9ab775bSTejun Heo /* am I supposed to die? */ 2199a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2200d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2201a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2202e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 220360f5a4bcSLai Jiangshan 220460f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 22057cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 220660f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 220760f5a4bcSLai Jiangshan kfree(worker); 2208c8e55f36STejun Heo return 0; 2209c8e55f36STejun Heo } 2210c8e55f36STejun Heo 2211c8e55f36STejun Heo worker_leave_idle(worker); 2212db7bccf4STejun Heo recheck: 2213e22bee78STejun Heo /* no more worker necessary? */ 221463d95a91STejun Heo if (!need_more_worker(pool)) 2215e22bee78STejun Heo goto sleep; 2216e22bee78STejun Heo 2217e22bee78STejun Heo /* do we need to manage? */ 221863d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2219e22bee78STejun Heo goto recheck; 2220e22bee78STejun Heo 2221c8e55f36STejun Heo /* 2222c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2223c8e55f36STejun Heo * preparing to process a work or actually processing it. 2224c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2225c8e55f36STejun Heo */ 22266183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2227c8e55f36STejun Heo 2228e22bee78STejun Heo /* 2229a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2230a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2231a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2232a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2233a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2234e22bee78STejun Heo */ 2235a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2236e22bee78STejun Heo 2237e22bee78STejun Heo do { 2238affee4b2STejun Heo struct work_struct *work = 2239bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2240affee4b2STejun Heo struct work_struct, entry); 2241affee4b2STejun Heo 224282607adcSTejun Heo pool->watchdog_ts = jiffies; 224382607adcSTejun Heo 2244c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2245affee4b2STejun Heo /* optimization path, not strictly necessary */ 2246affee4b2STejun Heo process_one_work(worker, work); 2247affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2248affee4b2STejun Heo process_scheduled_works(worker); 2249affee4b2STejun Heo } else { 2250c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2251affee4b2STejun Heo process_scheduled_works(worker); 2252affee4b2STejun Heo } 225363d95a91STejun Heo } while (keep_working(pool)); 2254affee4b2STejun Heo 2255228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2256d313dd85STejun Heo sleep: 2257c8e55f36STejun Heo /* 2258d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2259d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2260d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2261d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2262d565ed63STejun Heo * event. 2263c8e55f36STejun Heo */ 2264c8e55f36STejun Heo worker_enter_idle(worker); 2265c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2266d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22671da177e4SLinus Torvalds schedule(); 2268c8e55f36STejun Heo goto woke_up; 22691da177e4SLinus Torvalds } 22701da177e4SLinus Torvalds 2271e22bee78STejun Heo /** 2272e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2273111c225aSTejun Heo * @__rescuer: self 2274e22bee78STejun Heo * 2275e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2276493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2277e22bee78STejun Heo * 2278706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2279e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2280e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2281e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2282e22bee78STejun Heo * the problem rescuer solves. 2283e22bee78STejun Heo * 2284706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2285706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2286e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2287e22bee78STejun Heo * 2288e22bee78STejun Heo * This should happen rarely. 2289d185af30SYacine Belkadi * 2290d185af30SYacine Belkadi * Return: 0 2291e22bee78STejun Heo */ 2292111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2293e22bee78STejun Heo { 2294111c225aSTejun Heo struct worker *rescuer = __rescuer; 2295111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2296e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22974d595b86SLai Jiangshan bool should_stop; 2298e22bee78STejun Heo 2299e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2300111c225aSTejun Heo 2301111c225aSTejun Heo /* 2302111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2303111c225aSTejun Heo * doesn't participate in concurrency management. 2304111c225aSTejun Heo */ 2305111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2306e22bee78STejun Heo repeat: 2307c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 23081da177e4SLinus Torvalds 23094d595b86SLai Jiangshan /* 23104d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 23114d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 23124d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 23134d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 23144d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 23154d595b86SLai Jiangshan * list is always empty on exit. 23164d595b86SLai Jiangshan */ 23174d595b86SLai Jiangshan should_stop = kthread_should_stop(); 23181da177e4SLinus Torvalds 2319493a1724STejun Heo /* see whether any pwq is asking for help */ 23202e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2321493a1724STejun Heo 2322493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2323493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2324493a1724STejun Heo struct pool_workqueue, mayday_node); 2325112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2326e22bee78STejun Heo struct work_struct *work, *n; 232782607adcSTejun Heo bool first = true; 2328e22bee78STejun Heo 2329e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2330493a1724STejun Heo list_del_init(&pwq->mayday_node); 2331493a1724STejun Heo 23322e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2333e22bee78STejun Heo 233451697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 233551697d39SLai Jiangshan 233651697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2337b3104104SLai Jiangshan rescuer->pool = pool; 2338e22bee78STejun Heo 2339e22bee78STejun Heo /* 2340e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2341e22bee78STejun Heo * process'em. 2342e22bee78STejun Heo */ 23430479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 234482607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 234582607adcSTejun Heo if (get_work_pwq(work) == pwq) { 234682607adcSTejun Heo if (first) 234782607adcSTejun Heo pool->watchdog_ts = jiffies; 2348e22bee78STejun Heo move_linked_works(work, scheduled, &n); 234982607adcSTejun Heo } 235082607adcSTejun Heo first = false; 235182607adcSTejun Heo } 2352e22bee78STejun Heo 2353008847f6SNeilBrown if (!list_empty(scheduled)) { 2354e22bee78STejun Heo process_scheduled_works(rescuer); 23557576958aSTejun Heo 23567576958aSTejun Heo /* 2357008847f6SNeilBrown * The above execution of rescued work items could 2358008847f6SNeilBrown * have created more to rescue through 2359008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2360008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2361008847f6SNeilBrown * that such back-to-back work items, which may be 2362008847f6SNeilBrown * being used to relieve memory pressure, don't 2363008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2364008847f6SNeilBrown */ 2365008847f6SNeilBrown if (need_to_create_worker(pool)) { 2366008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2367008847f6SNeilBrown get_pwq(pwq); 2368008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2369008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2370008847f6SNeilBrown } 2371008847f6SNeilBrown } 2372008847f6SNeilBrown 2373008847f6SNeilBrown /* 237477668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 237513b1d625SLai Jiangshan * go away while we're still attached to it. 237677668c8bSLai Jiangshan */ 237777668c8bSLai Jiangshan put_pwq(pwq); 237877668c8bSLai Jiangshan 237977668c8bSLai Jiangshan /* 2380d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 23817576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23827576958aSTejun Heo * and stalling the execution. 23837576958aSTejun Heo */ 2384d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 238563d95a91STejun Heo wake_up_worker(pool); 23867576958aSTejun Heo 2387b3104104SLai Jiangshan rescuer->pool = NULL; 238813b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 238913b1d625SLai Jiangshan 239013b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 239113b1d625SLai Jiangshan 239213b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 23931da177e4SLinus Torvalds } 23941da177e4SLinus Torvalds 23952e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2396493a1724STejun Heo 23974d595b86SLai Jiangshan if (should_stop) { 23984d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23994d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 24004d595b86SLai Jiangshan return 0; 24014d595b86SLai Jiangshan } 24024d595b86SLai Jiangshan 2403111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2404111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2405e22bee78STejun Heo schedule(); 2406e22bee78STejun Heo goto repeat; 24071da177e4SLinus Torvalds } 24081da177e4SLinus Torvalds 2409fca839c0STejun Heo /** 2410fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2411fca839c0STejun Heo * @target_wq: workqueue being flushed 2412fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2413fca839c0STejun Heo * 2414fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2415fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2416fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2417fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2418fca839c0STejun Heo * a deadlock. 2419fca839c0STejun Heo */ 2420fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2421fca839c0STejun Heo struct work_struct *target_work) 2422fca839c0STejun Heo { 2423fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2424fca839c0STejun Heo struct worker *worker; 2425fca839c0STejun Heo 2426fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2427fca839c0STejun Heo return; 2428fca839c0STejun Heo 2429fca839c0STejun Heo worker = current_wq_worker(); 2430fca839c0STejun Heo 2431fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2432fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2433fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 243423d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 243523d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2436fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2437fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2438fca839c0STejun Heo target_wq->name, target_func); 2439fca839c0STejun Heo } 2440fca839c0STejun Heo 2441fc2e4d70SOleg Nesterov struct wq_barrier { 2442fc2e4d70SOleg Nesterov struct work_struct work; 2443fc2e4d70SOleg Nesterov struct completion done; 24442607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2445fc2e4d70SOleg Nesterov }; 2446fc2e4d70SOleg Nesterov 2447fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2448fc2e4d70SOleg Nesterov { 2449fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2450fc2e4d70SOleg Nesterov complete(&barr->done); 2451fc2e4d70SOleg Nesterov } 2452fc2e4d70SOleg Nesterov 24534690c4abSTejun Heo /** 24544690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2455112202d9STejun Heo * @pwq: pwq to insert barrier into 24564690c4abSTejun Heo * @barr: wq_barrier to insert 2457affee4b2STejun Heo * @target: target work to attach @barr to 2458affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24594690c4abSTejun Heo * 2460affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2461affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2462affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2463affee4b2STejun Heo * cpu. 2464affee4b2STejun Heo * 2465affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2466affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2467affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2468affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2469affee4b2STejun Heo * after a work with LINKED flag set. 2470affee4b2STejun Heo * 2471affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2472112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24734690c4abSTejun Heo * 24744690c4abSTejun Heo * CONTEXT: 2475d565ed63STejun Heo * spin_lock_irq(pool->lock). 24764690c4abSTejun Heo */ 2477112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2478affee4b2STejun Heo struct wq_barrier *barr, 2479affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2480fc2e4d70SOleg Nesterov { 2481affee4b2STejun Heo struct list_head *head; 2482affee4b2STejun Heo unsigned int linked = 0; 2483affee4b2STejun Heo 2484dc186ad7SThomas Gleixner /* 2485d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2486dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2487dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2488dc186ad7SThomas Gleixner * might deadlock. 2489dc186ad7SThomas Gleixner */ 2490ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 249122df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 249252fa5bc5SBoqun Feng 2493fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2494fd1a5b04SByungchul Park 24952607d7a6STejun Heo barr->task = current; 249683c22520SOleg Nesterov 2497affee4b2STejun Heo /* 2498affee4b2STejun Heo * If @target is currently being executed, schedule the 2499affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2500affee4b2STejun Heo */ 2501affee4b2STejun Heo if (worker) 2502affee4b2STejun Heo head = worker->scheduled.next; 2503affee4b2STejun Heo else { 2504affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2505affee4b2STejun Heo 2506affee4b2STejun Heo head = target->entry.next; 2507affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2508affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2509affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2510affee4b2STejun Heo } 2511affee4b2STejun Heo 2512dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2513112202d9STejun Heo insert_work(pwq, &barr->work, head, 2514affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2515fc2e4d70SOleg Nesterov } 2516fc2e4d70SOleg Nesterov 251773f53c4aSTejun Heo /** 2518112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 251973f53c4aSTejun Heo * @wq: workqueue being flushed 252073f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 252173f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 252273f53c4aSTejun Heo * 2523112202d9STejun Heo * Prepare pwqs for workqueue flushing. 252473f53c4aSTejun Heo * 2525112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2526112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2527112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2528112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2529112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 253073f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 253173f53c4aSTejun Heo * 253273f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 253373f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 253473f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 253573f53c4aSTejun Heo * is returned. 253673f53c4aSTejun Heo * 2537112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 253873f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 253973f53c4aSTejun Heo * advanced to @work_color. 254073f53c4aSTejun Heo * 254173f53c4aSTejun Heo * CONTEXT: 25423c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 254373f53c4aSTejun Heo * 2544d185af30SYacine Belkadi * Return: 254573f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 254673f53c4aSTejun Heo * otherwise. 254773f53c4aSTejun Heo */ 2548112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 254973f53c4aSTejun Heo int flush_color, int work_color) 25501da177e4SLinus Torvalds { 255173f53c4aSTejun Heo bool wait = false; 255249e3cf44STejun Heo struct pool_workqueue *pwq; 25531da177e4SLinus Torvalds 255473f53c4aSTejun Heo if (flush_color >= 0) { 25556183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2556112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2557dc186ad7SThomas Gleixner } 255814441960SOleg Nesterov 255949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2560112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25611da177e4SLinus Torvalds 2562b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 256373f53c4aSTejun Heo 256473f53c4aSTejun Heo if (flush_color >= 0) { 25656183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 256673f53c4aSTejun Heo 2567112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2568112202d9STejun Heo pwq->flush_color = flush_color; 2569112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 257073f53c4aSTejun Heo wait = true; 25711da177e4SLinus Torvalds } 257273f53c4aSTejun Heo } 257373f53c4aSTejun Heo 257473f53c4aSTejun Heo if (work_color >= 0) { 25756183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2576112202d9STejun Heo pwq->work_color = work_color; 257773f53c4aSTejun Heo } 257873f53c4aSTejun Heo 2579b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25801da177e4SLinus Torvalds } 25811da177e4SLinus Torvalds 2582112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 258373f53c4aSTejun Heo complete(&wq->first_flusher->done); 258473f53c4aSTejun Heo 258573f53c4aSTejun Heo return wait; 258683c22520SOleg Nesterov } 25871da177e4SLinus Torvalds 25880fcb78c2SRolf Eike Beer /** 25891da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25900fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25911da177e4SLinus Torvalds * 2592c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2593c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25941da177e4SLinus Torvalds */ 25957ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25961da177e4SLinus Torvalds { 259773f53c4aSTejun Heo struct wq_flusher this_flusher = { 259873f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 259973f53c4aSTejun Heo .flush_color = -1, 2600fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 260173f53c4aSTejun Heo }; 260273f53c4aSTejun Heo int next_color; 2603b1f4ec17SOleg Nesterov 26043347fa09STejun Heo if (WARN_ON(!wq_online)) 26053347fa09STejun Heo return; 26063347fa09STejun Heo 26073c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 260873f53c4aSTejun Heo 260973f53c4aSTejun Heo /* 261073f53c4aSTejun Heo * Start-to-wait phase 261173f53c4aSTejun Heo */ 261273f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 261373f53c4aSTejun Heo 261473f53c4aSTejun Heo if (next_color != wq->flush_color) { 261573f53c4aSTejun Heo /* 261673f53c4aSTejun Heo * Color space is not full. The current work_color 261773f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 261873f53c4aSTejun Heo * by one. 261973f53c4aSTejun Heo */ 26206183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 262173f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 262273f53c4aSTejun Heo wq->work_color = next_color; 262373f53c4aSTejun Heo 262473f53c4aSTejun Heo if (!wq->first_flusher) { 262573f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26266183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 262773f53c4aSTejun Heo 262873f53c4aSTejun Heo wq->first_flusher = &this_flusher; 262973f53c4aSTejun Heo 2630112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 263173f53c4aSTejun Heo wq->work_color)) { 263273f53c4aSTejun Heo /* nothing to flush, done */ 263373f53c4aSTejun Heo wq->flush_color = next_color; 263473f53c4aSTejun Heo wq->first_flusher = NULL; 263573f53c4aSTejun Heo goto out_unlock; 263673f53c4aSTejun Heo } 263773f53c4aSTejun Heo } else { 263873f53c4aSTejun Heo /* wait in queue */ 26396183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 264073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2641112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 264273f53c4aSTejun Heo } 264373f53c4aSTejun Heo } else { 264473f53c4aSTejun Heo /* 264573f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 264673f53c4aSTejun Heo * The next flush completion will assign us 264773f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 264873f53c4aSTejun Heo */ 264973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 265073f53c4aSTejun Heo } 265173f53c4aSTejun Heo 2652fca839c0STejun Heo check_flush_dependency(wq, NULL); 2653fca839c0STejun Heo 26543c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 265573f53c4aSTejun Heo 265673f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 265773f53c4aSTejun Heo 265873f53c4aSTejun Heo /* 265973f53c4aSTejun Heo * Wake-up-and-cascade phase 266073f53c4aSTejun Heo * 266173f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 266273f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 266373f53c4aSTejun Heo */ 266473f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 266573f53c4aSTejun Heo return; 266673f53c4aSTejun Heo 26673c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 266873f53c4aSTejun Heo 26694ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26704ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26714ce48b37STejun Heo goto out_unlock; 26724ce48b37STejun Heo 267373f53c4aSTejun Heo wq->first_flusher = NULL; 267473f53c4aSTejun Heo 26756183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26766183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 267773f53c4aSTejun Heo 267873f53c4aSTejun Heo while (true) { 267973f53c4aSTejun Heo struct wq_flusher *next, *tmp; 268073f53c4aSTejun Heo 268173f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 268273f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 268373f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 268473f53c4aSTejun Heo break; 268573f53c4aSTejun Heo list_del_init(&next->list); 268673f53c4aSTejun Heo complete(&next->done); 268773f53c4aSTejun Heo } 268873f53c4aSTejun Heo 26896183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 269073f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 269173f53c4aSTejun Heo 269273f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 269373f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 269473f53c4aSTejun Heo 269573f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 269673f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 269773f53c4aSTejun Heo /* 269873f53c4aSTejun Heo * Assign the same color to all overflowed 269973f53c4aSTejun Heo * flushers, advance work_color and append to 270073f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 270173f53c4aSTejun Heo * phase for these overflowed flushers. 270273f53c4aSTejun Heo */ 270373f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 270473f53c4aSTejun Heo tmp->flush_color = wq->work_color; 270573f53c4aSTejun Heo 270673f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 270773f53c4aSTejun Heo 270873f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 270973f53c4aSTejun Heo &wq->flusher_queue); 2710112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 271173f53c4aSTejun Heo } 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27146183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 271573f53c4aSTejun Heo break; 271673f53c4aSTejun Heo } 271773f53c4aSTejun Heo 271873f53c4aSTejun Heo /* 271973f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2720112202d9STejun Heo * the new first flusher and arm pwqs. 272173f53c4aSTejun Heo */ 27226183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27236183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 272473f53c4aSTejun Heo 272573f53c4aSTejun Heo list_del_init(&next->list); 272673f53c4aSTejun Heo wq->first_flusher = next; 272773f53c4aSTejun Heo 2728112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 272973f53c4aSTejun Heo break; 273073f53c4aSTejun Heo 273173f53c4aSTejun Heo /* 273273f53c4aSTejun Heo * Meh... this color is already done, clear first 273373f53c4aSTejun Heo * flusher and repeat cascading. 273473f53c4aSTejun Heo */ 273573f53c4aSTejun Heo wq->first_flusher = NULL; 273673f53c4aSTejun Heo } 273773f53c4aSTejun Heo 273873f53c4aSTejun Heo out_unlock: 27393c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27401da177e4SLinus Torvalds } 27411dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 27421da177e4SLinus Torvalds 27439c5a2ba7STejun Heo /** 27449c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27459c5a2ba7STejun Heo * @wq: workqueue to drain 27469c5a2ba7STejun Heo * 27479c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27489c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27499c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2750b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 27519c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27529c5a2ba7STejun Heo * takes too long. 27539c5a2ba7STejun Heo */ 27549c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27559c5a2ba7STejun Heo { 27569c5a2ba7STejun Heo unsigned int flush_cnt = 0; 275749e3cf44STejun Heo struct pool_workqueue *pwq; 27589c5a2ba7STejun Heo 27599c5a2ba7STejun Heo /* 27609c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27619c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2762618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27639c5a2ba7STejun Heo */ 276487fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27659c5a2ba7STejun Heo if (!wq->nr_drainers++) 2766618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 276787fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27689c5a2ba7STejun Heo reflush: 27699c5a2ba7STejun Heo flush_workqueue(wq); 27709c5a2ba7STejun Heo 2771b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 277276af4d93STejun Heo 277349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2774fa2563e4SThomas Tuttle bool drained; 27759c5a2ba7STejun Heo 2776b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2777112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2778b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2779fa2563e4SThomas Tuttle 2780fa2563e4SThomas Tuttle if (drained) 27819c5a2ba7STejun Heo continue; 27829c5a2ba7STejun Heo 27839c5a2ba7STejun Heo if (++flush_cnt == 10 || 27849c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2785c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27869c5a2ba7STejun Heo wq->name, flush_cnt); 278776af4d93STejun Heo 2788b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27899c5a2ba7STejun Heo goto reflush; 27909c5a2ba7STejun Heo } 27919c5a2ba7STejun Heo 27929c5a2ba7STejun Heo if (!--wq->nr_drainers) 2793618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 279487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27959c5a2ba7STejun Heo } 27969c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27979c5a2ba7STejun Heo 2798606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2799baf59022STejun Heo { 2800baf59022STejun Heo struct worker *worker = NULL; 2801c9e7cf27STejun Heo struct worker_pool *pool; 2802112202d9STejun Heo struct pool_workqueue *pwq; 2803baf59022STejun Heo 2804baf59022STejun Heo might_sleep(); 2805baf59022STejun Heo 2806fa1b54e6STejun Heo local_irq_disable(); 2807fa1b54e6STejun Heo pool = get_work_pool(work); 2808fa1b54e6STejun Heo if (!pool) { 2809fa1b54e6STejun Heo local_irq_enable(); 2810fa1b54e6STejun Heo return false; 2811fa1b54e6STejun Heo } 2812fa1b54e6STejun Heo 2813fa1b54e6STejun Heo spin_lock(&pool->lock); 28140b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2815112202d9STejun Heo pwq = get_work_pwq(work); 2816112202d9STejun Heo if (pwq) { 2817112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2818baf59022STejun Heo goto already_gone; 2819606a5020STejun Heo } else { 2820c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2821baf59022STejun Heo if (!worker) 2822baf59022STejun Heo goto already_gone; 2823112202d9STejun Heo pwq = worker->current_pwq; 2824606a5020STejun Heo } 2825baf59022STejun Heo 2826fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2827fca839c0STejun Heo 2828112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2829d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2830baf59022STejun Heo 2831e159489bSTejun Heo /* 2832a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 2833a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 2834a1d14934SPeter Zijlstra * 2835a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 2836a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 2837a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 2838a1d14934SPeter Zijlstra * forward progress. 2839e159489bSTejun Heo */ 2840a1d14934SPeter Zijlstra if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) { 2841112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2842112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2843a1d14934SPeter Zijlstra } 2844e159489bSTejun Heo 2845baf59022STejun Heo return true; 2846baf59022STejun Heo already_gone: 2847d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2848baf59022STejun Heo return false; 2849baf59022STejun Heo } 2850baf59022STejun Heo 2851db700897SOleg Nesterov /** 2852401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2853401a8d04STejun Heo * @work: the work to flush 2854db700897SOleg Nesterov * 2855606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2856606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2857401a8d04STejun Heo * 2858d185af30SYacine Belkadi * Return: 2859401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2860401a8d04STejun Heo * %false if it was already idle. 2861db700897SOleg Nesterov */ 2862401a8d04STejun Heo bool flush_work(struct work_struct *work) 2863db700897SOleg Nesterov { 286412997d1aSBjorn Helgaas struct wq_barrier barr; 286512997d1aSBjorn Helgaas 28663347fa09STejun Heo if (WARN_ON(!wq_online)) 28673347fa09STejun Heo return false; 28683347fa09STejun Heo 286912997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 287012997d1aSBjorn Helgaas wait_for_completion(&barr.done); 287112997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 287212997d1aSBjorn Helgaas return true; 287312997d1aSBjorn Helgaas } else { 287412997d1aSBjorn Helgaas return false; 287512997d1aSBjorn Helgaas } 2876606a5020STejun Heo } 2877db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2878db700897SOleg Nesterov 28798603e1b3STejun Heo struct cwt_wait { 2880ac6424b9SIngo Molnar wait_queue_entry_t wait; 28818603e1b3STejun Heo struct work_struct *work; 28828603e1b3STejun Heo }; 28838603e1b3STejun Heo 2884ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 28858603e1b3STejun Heo { 28868603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 28878603e1b3STejun Heo 28888603e1b3STejun Heo if (cwait->work != key) 28898603e1b3STejun Heo return 0; 28908603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 28918603e1b3STejun Heo } 28928603e1b3STejun Heo 289336e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2894401a8d04STejun Heo { 28958603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2896bbb68dfaSTejun Heo unsigned long flags; 28971f1f642eSOleg Nesterov int ret; 28981f1f642eSOleg Nesterov 28991f1f642eSOleg Nesterov do { 2900bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2901bbb68dfaSTejun Heo /* 29028603e1b3STejun Heo * If someone else is already canceling, wait for it to 29038603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 29048603e1b3STejun Heo * because we may get scheduled between @work's completion 29058603e1b3STejun Heo * and the other canceling task resuming and clearing 29068603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 29078603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 29088603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 29098603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 29108603e1b3STejun Heo * we're hogging the CPU. 29118603e1b3STejun Heo * 29128603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 29138603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 29148603e1b3STejun Heo * wake function which matches @work along with exclusive 29158603e1b3STejun Heo * wait and wakeup. 2916bbb68dfaSTejun Heo */ 29178603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 29188603e1b3STejun Heo struct cwt_wait cwait; 29198603e1b3STejun Heo 29208603e1b3STejun Heo init_wait(&cwait.wait); 29218603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 29228603e1b3STejun Heo cwait.work = work; 29238603e1b3STejun Heo 29248603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 29258603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 29268603e1b3STejun Heo if (work_is_canceling(work)) 29278603e1b3STejun Heo schedule(); 29288603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 29298603e1b3STejun Heo } 29301f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29311f1f642eSOleg Nesterov 2932bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2933bbb68dfaSTejun Heo mark_work_canceling(work); 2934bbb68dfaSTejun Heo local_irq_restore(flags); 2935bbb68dfaSTejun Heo 29363347fa09STejun Heo /* 29373347fa09STejun Heo * This allows canceling during early boot. We know that @work 29383347fa09STejun Heo * isn't executing. 29393347fa09STejun Heo */ 29403347fa09STejun Heo if (wq_online) 2941606a5020STejun Heo flush_work(work); 29423347fa09STejun Heo 29437a22ad75STejun Heo clear_work_data(work); 29448603e1b3STejun Heo 29458603e1b3STejun Heo /* 29468603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 29478603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 29488603e1b3STejun Heo * visible there. 29498603e1b3STejun Heo */ 29508603e1b3STejun Heo smp_mb(); 29518603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 29528603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 29538603e1b3STejun Heo 29541f1f642eSOleg Nesterov return ret; 29551f1f642eSOleg Nesterov } 29561f1f642eSOleg Nesterov 29576e84d644SOleg Nesterov /** 2958401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2959401a8d04STejun Heo * @work: the work to cancel 29606e84d644SOleg Nesterov * 2961401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2962401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2963401a8d04STejun Heo * another workqueue. On return from this function, @work is 2964401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29651f1f642eSOleg Nesterov * 2966401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2967401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29686e84d644SOleg Nesterov * 2969401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29706e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2971401a8d04STejun Heo * 2972d185af30SYacine Belkadi * Return: 2973401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29746e84d644SOleg Nesterov */ 2975401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29766e84d644SOleg Nesterov { 297736e227d2STejun Heo return __cancel_work_timer(work, false); 2978b89deed3SOleg Nesterov } 297928e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2980b89deed3SOleg Nesterov 29816e84d644SOleg Nesterov /** 2982401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2983401a8d04STejun Heo * @dwork: the delayed work to flush 29846e84d644SOleg Nesterov * 2985401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2986401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2987401a8d04STejun Heo * considers the last queueing instance of @dwork. 29881f1f642eSOleg Nesterov * 2989d185af30SYacine Belkadi * Return: 2990401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2991401a8d04STejun Heo * %false if it was already idle. 29926e84d644SOleg Nesterov */ 2993401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2994401a8d04STejun Heo { 29958930cabaSTejun Heo local_irq_disable(); 2996401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 299760c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29988930cabaSTejun Heo local_irq_enable(); 2999401a8d04STejun Heo return flush_work(&dwork->work); 3000401a8d04STejun Heo } 3001401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3002401a8d04STejun Heo 3003f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3004f72b8792SJens Axboe { 3005f72b8792SJens Axboe unsigned long flags; 3006f72b8792SJens Axboe int ret; 3007f72b8792SJens Axboe 3008f72b8792SJens Axboe do { 3009f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3010f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3011f72b8792SJens Axboe 3012f72b8792SJens Axboe if (unlikely(ret < 0)) 3013f72b8792SJens Axboe return false; 3014f72b8792SJens Axboe 3015f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3016f72b8792SJens Axboe local_irq_restore(flags); 3017f72b8792SJens Axboe return ret; 3018f72b8792SJens Axboe } 3019f72b8792SJens Axboe 3020f72b8792SJens Axboe /* 3021f72b8792SJens Axboe * See cancel_delayed_work() 3022f72b8792SJens Axboe */ 3023f72b8792SJens Axboe bool cancel_work(struct work_struct *work) 3024f72b8792SJens Axboe { 3025f72b8792SJens Axboe return __cancel_work(work, false); 3026f72b8792SJens Axboe } 3027f72b8792SJens Axboe 3028401a8d04STejun Heo /** 302957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 303057b30ae7STejun Heo * @dwork: delayed_work to cancel 303109383498STejun Heo * 3032d185af30SYacine Belkadi * Kill off a pending delayed_work. 3033d185af30SYacine Belkadi * 3034d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3035d185af30SYacine Belkadi * pending. 3036d185af30SYacine Belkadi * 3037d185af30SYacine Belkadi * Note: 3038d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3039d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3040d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 304109383498STejun Heo * 304257b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 304309383498STejun Heo */ 304457b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 304509383498STejun Heo { 3046f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 304709383498STejun Heo } 304857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 304909383498STejun Heo 305009383498STejun Heo /** 3051401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3052401a8d04STejun Heo * @dwork: the delayed work cancel 3053401a8d04STejun Heo * 3054401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3055401a8d04STejun Heo * 3056d185af30SYacine Belkadi * Return: 3057401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3058401a8d04STejun Heo */ 3059401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30606e84d644SOleg Nesterov { 306136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30626e84d644SOleg Nesterov } 3063f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30641da177e4SLinus Torvalds 30650fcb78c2SRolf Eike Beer /** 306631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3067b6136773SAndrew Morton * @func: the function to call 3068b6136773SAndrew Morton * 306931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 307031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3071b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 307231ddd871STejun Heo * 3073d185af30SYacine Belkadi * Return: 307431ddd871STejun Heo * 0 on success, -errno on failure. 3075b6136773SAndrew Morton */ 307665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 307715316ba8SChristoph Lameter { 307815316ba8SChristoph Lameter int cpu; 307938f51568SNamhyung Kim struct work_struct __percpu *works; 308015316ba8SChristoph Lameter 3081b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3082b6136773SAndrew Morton if (!works) 308315316ba8SChristoph Lameter return -ENOMEM; 3084b6136773SAndrew Morton 308595402b38SGautham R Shenoy get_online_cpus(); 308693981800STejun Heo 308715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30889bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30899bfb1839SIngo Molnar 30909bfb1839SIngo Molnar INIT_WORK(work, func); 30918de6d308SOleg Nesterov schedule_work_on(cpu, work); 309215316ba8SChristoph Lameter } 309393981800STejun Heo 309493981800STejun Heo for_each_online_cpu(cpu) 30958616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 309693981800STejun Heo 309795402b38SGautham R Shenoy put_online_cpus(); 3098b6136773SAndrew Morton free_percpu(works); 309915316ba8SChristoph Lameter return 0; 310015316ba8SChristoph Lameter } 310115316ba8SChristoph Lameter 3102eef6a7d5SAlan Stern /** 31031fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31041fa44ecaSJames Bottomley * @fn: the function to execute 31051fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31061fa44ecaSJames Bottomley * be available when the work executes) 31071fa44ecaSJames Bottomley * 31081fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31091fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31101fa44ecaSJames Bottomley * 3111d185af30SYacine Belkadi * Return: 0 - function was executed 31121fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31131fa44ecaSJames Bottomley */ 311465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31151fa44ecaSJames Bottomley { 31161fa44ecaSJames Bottomley if (!in_interrupt()) { 311765f27f38SDavid Howells fn(&ew->work); 31181fa44ecaSJames Bottomley return 0; 31191fa44ecaSJames Bottomley } 31201fa44ecaSJames Bottomley 312165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31221fa44ecaSJames Bottomley schedule_work(&ew->work); 31231fa44ecaSJames Bottomley 31241fa44ecaSJames Bottomley return 1; 31251fa44ecaSJames Bottomley } 31261fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31271fa44ecaSJames Bottomley 31287a4e344cSTejun Heo /** 31297a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 31307a4e344cSTejun Heo * @attrs: workqueue_attrs to free 31317a4e344cSTejun Heo * 31327a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 31337a4e344cSTejun Heo */ 31347a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 31357a4e344cSTejun Heo { 31367a4e344cSTejun Heo if (attrs) { 31377a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 31387a4e344cSTejun Heo kfree(attrs); 31397a4e344cSTejun Heo } 31407a4e344cSTejun Heo } 31417a4e344cSTejun Heo 31427a4e344cSTejun Heo /** 31437a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 31447a4e344cSTejun Heo * @gfp_mask: allocation mask to use 31457a4e344cSTejun Heo * 31467a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3147d185af30SYacine Belkadi * return it. 3148d185af30SYacine Belkadi * 3149d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 31507a4e344cSTejun Heo */ 31517a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 31527a4e344cSTejun Heo { 31537a4e344cSTejun Heo struct workqueue_attrs *attrs; 31547a4e344cSTejun Heo 31557a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 31567a4e344cSTejun Heo if (!attrs) 31577a4e344cSTejun Heo goto fail; 31587a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 31597a4e344cSTejun Heo goto fail; 31607a4e344cSTejun Heo 316113e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 31627a4e344cSTejun Heo return attrs; 31637a4e344cSTejun Heo fail: 31647a4e344cSTejun Heo free_workqueue_attrs(attrs); 31657a4e344cSTejun Heo return NULL; 31667a4e344cSTejun Heo } 31677a4e344cSTejun Heo 316829c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 316929c91e99STejun Heo const struct workqueue_attrs *from) 317029c91e99STejun Heo { 317129c91e99STejun Heo to->nice = from->nice; 317229c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 31732865a8fbSShaohua Li /* 31742865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 31752865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 31762865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 31772865a8fbSShaohua Li */ 31782865a8fbSShaohua Li to->no_numa = from->no_numa; 317929c91e99STejun Heo } 318029c91e99STejun Heo 318129c91e99STejun Heo /* hash value of the content of @attr */ 318229c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 318329c91e99STejun Heo { 318429c91e99STejun Heo u32 hash = 0; 318529c91e99STejun Heo 318629c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 318713e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 318813e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 318929c91e99STejun Heo return hash; 319029c91e99STejun Heo } 319129c91e99STejun Heo 319229c91e99STejun Heo /* content equality test */ 319329c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 319429c91e99STejun Heo const struct workqueue_attrs *b) 319529c91e99STejun Heo { 319629c91e99STejun Heo if (a->nice != b->nice) 319729c91e99STejun Heo return false; 319829c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 319929c91e99STejun Heo return false; 320029c91e99STejun Heo return true; 320129c91e99STejun Heo } 320229c91e99STejun Heo 32037a4e344cSTejun Heo /** 32047a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 32057a4e344cSTejun Heo * @pool: worker_pool to initialize 32067a4e344cSTejun Heo * 3207402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3208d185af30SYacine Belkadi * 3209d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 321029c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 321129c91e99STejun Heo * on @pool safely to release it. 32127a4e344cSTejun Heo */ 32137a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 32144e1a1f9aSTejun Heo { 32154e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 321629c91e99STejun Heo pool->id = -1; 321729c91e99STejun Heo pool->cpu = -1; 3218f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 32194e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 322082607adcSTejun Heo pool->watchdog_ts = jiffies; 32214e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 32224e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 32234e1a1f9aSTejun Heo hash_init(pool->busy_hash); 32244e1a1f9aSTejun Heo 322532a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 32264e1a1f9aSTejun Heo 322732a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 32284e1a1f9aSTejun Heo 322992f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3230da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 32317a4e344cSTejun Heo 32327cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 323329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 323429c91e99STejun Heo pool->refcnt = 1; 323529c91e99STejun Heo 323629c91e99STejun Heo /* shouldn't fail above this point */ 32377a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 32387a4e344cSTejun Heo if (!pool->attrs) 32397a4e344cSTejun Heo return -ENOMEM; 32407a4e344cSTejun Heo return 0; 32414e1a1f9aSTejun Heo } 32424e1a1f9aSTejun Heo 3243e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3244e2dca7adSTejun Heo { 3245e2dca7adSTejun Heo struct workqueue_struct *wq = 3246e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3247e2dca7adSTejun Heo 3248e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3249e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3250e2dca7adSTejun Heo else 3251e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3252e2dca7adSTejun Heo 3253e2dca7adSTejun Heo kfree(wq->rescuer); 3254e2dca7adSTejun Heo kfree(wq); 3255e2dca7adSTejun Heo } 3256e2dca7adSTejun Heo 325729c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 325829c91e99STejun Heo { 325929c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 326029c91e99STejun Heo 32617cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 326229c91e99STejun Heo free_workqueue_attrs(pool->attrs); 326329c91e99STejun Heo kfree(pool); 326429c91e99STejun Heo } 326529c91e99STejun Heo 326629c91e99STejun Heo /** 326729c91e99STejun Heo * put_unbound_pool - put a worker_pool 326829c91e99STejun Heo * @pool: worker_pool to put 326929c91e99STejun Heo * 327029c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3271c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3272c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3273c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3274a892caccSTejun Heo * 3275a892caccSTejun Heo * Should be called with wq_pool_mutex held. 327629c91e99STejun Heo */ 327729c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 327829c91e99STejun Heo { 327960f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 328029c91e99STejun Heo struct worker *worker; 328129c91e99STejun Heo 3282a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3283a892caccSTejun Heo 3284a892caccSTejun Heo if (--pool->refcnt) 328529c91e99STejun Heo return; 328629c91e99STejun Heo 328729c91e99STejun Heo /* sanity checks */ 328861d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3289a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 329029c91e99STejun Heo return; 329129c91e99STejun Heo 329229c91e99STejun Heo /* release id and unhash */ 329329c91e99STejun Heo if (pool->id >= 0) 329429c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 329529c91e99STejun Heo hash_del(&pool->hash_node); 329629c91e99STejun Heo 3297c5aa87bbSTejun Heo /* 3298692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3299692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3300692b4825STejun Heo * manager and @pool gets freed with the flag set. 3301c5aa87bbSTejun Heo */ 330260f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 3303692b4825STejun Heo wait_event_lock_irq(wq_manager_wait, 3304692b4825STejun Heo !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); 3305692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3306692b4825STejun Heo 33071037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 330829c91e99STejun Heo destroy_worker(worker); 330929c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 331029c91e99STejun Heo spin_unlock_irq(&pool->lock); 331160f5a4bcSLai Jiangshan 331292f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3313da028469SLai Jiangshan if (!list_empty(&pool->workers)) 331460f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 331592f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 331660f5a4bcSLai Jiangshan 331760f5a4bcSLai Jiangshan if (pool->detach_completion) 331860f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 331960f5a4bcSLai Jiangshan 332029c91e99STejun Heo /* shut down the timers */ 332129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 332229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 332329c91e99STejun Heo 332429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 332529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 332629c91e99STejun Heo } 332729c91e99STejun Heo 332829c91e99STejun Heo /** 332929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 333029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 333129c91e99STejun Heo * 333229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 333329c91e99STejun Heo * reference count and return it. If there already is a matching 333429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3335d185af30SYacine Belkadi * create a new one. 3336a892caccSTejun Heo * 3337a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3338d185af30SYacine Belkadi * 3339d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3340d185af30SYacine Belkadi * On failure, %NULL. 334129c91e99STejun Heo */ 334229c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 334329c91e99STejun Heo { 334429c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 334529c91e99STejun Heo struct worker_pool *pool; 3346f3f90ad4STejun Heo int node; 3347e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 334829c91e99STejun Heo 3349a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 335029c91e99STejun Heo 335129c91e99STejun Heo /* do we already have a matching pool? */ 335229c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 335329c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 335429c91e99STejun Heo pool->refcnt++; 33553fb1823cSLai Jiangshan return pool; 335629c91e99STejun Heo } 335729c91e99STejun Heo } 335829c91e99STejun Heo 3359e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3360e2273584SXunlei Pang if (wq_numa_enabled) { 3361e2273584SXunlei Pang for_each_node(node) { 3362e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3363e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3364e2273584SXunlei Pang target_node = node; 3365e2273584SXunlei Pang break; 3366e2273584SXunlei Pang } 3367e2273584SXunlei Pang } 3368e2273584SXunlei Pang } 3369e2273584SXunlei Pang 337029c91e99STejun Heo /* nope, create a new one */ 3371e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 337229c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 337329c91e99STejun Heo goto fail; 337429c91e99STejun Heo 33758864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 337629c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3377e2273584SXunlei Pang pool->node = target_node; 337829c91e99STejun Heo 33792865a8fbSShaohua Li /* 33802865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 33812865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 33822865a8fbSShaohua Li */ 33832865a8fbSShaohua Li pool->attrs->no_numa = false; 33842865a8fbSShaohua Li 338529c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 338629c91e99STejun Heo goto fail; 338729c91e99STejun Heo 338829c91e99STejun Heo /* create and start the initial worker */ 33893347fa09STejun Heo if (wq_online && !create_worker(pool)) 339029c91e99STejun Heo goto fail; 339129c91e99STejun Heo 339229c91e99STejun Heo /* install */ 339329c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 33943fb1823cSLai Jiangshan 339529c91e99STejun Heo return pool; 339629c91e99STejun Heo fail: 339729c91e99STejun Heo if (pool) 339829c91e99STejun Heo put_unbound_pool(pool); 339929c91e99STejun Heo return NULL; 340029c91e99STejun Heo } 340129c91e99STejun Heo 34028864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 34038864b4e5STejun Heo { 34048864b4e5STejun Heo kmem_cache_free(pwq_cache, 34058864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 34068864b4e5STejun Heo } 34078864b4e5STejun Heo 34088864b4e5STejun Heo /* 34098864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 34108864b4e5STejun Heo * and needs to be destroyed. 34118864b4e5STejun Heo */ 34128864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 34138864b4e5STejun Heo { 34148864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 34158864b4e5STejun Heo unbound_release_work); 34168864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 34178864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3418bc0caf09STejun Heo bool is_last; 34198864b4e5STejun Heo 34208864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 34218864b4e5STejun Heo return; 34228864b4e5STejun Heo 34233c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 34248864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3425bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 34263c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 34278864b4e5STejun Heo 3428a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 34298864b4e5STejun Heo put_unbound_pool(pool); 3430a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3431a892caccSTejun Heo 34328864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 34338864b4e5STejun Heo 34348864b4e5STejun Heo /* 34358864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3436e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 34378864b4e5STejun Heo */ 3438e2dca7adSTejun Heo if (is_last) 3439e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 34406029a918STejun Heo } 34418864b4e5STejun Heo 34420fbd95aaSTejun Heo /** 3443699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 34440fbd95aaSTejun Heo * @pwq: target pool_workqueue 34450fbd95aaSTejun Heo * 3446699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3447699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3448699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 34490fbd95aaSTejun Heo */ 3450699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 34510fbd95aaSTejun Heo { 3452699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3453699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 34543347fa09STejun Heo unsigned long flags; 3455699ce097STejun Heo 3456699ce097STejun Heo /* for @wq->saved_max_active */ 3457a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3458699ce097STejun Heo 3459699ce097STejun Heo /* fast exit for non-freezable wqs */ 3460699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3461699ce097STejun Heo return; 3462699ce097STejun Heo 34633347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 34643347fa09STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 3465699ce097STejun Heo 346674b414eaSLai Jiangshan /* 346774b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 346874b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 346974b414eaSLai Jiangshan * is updated and visible. 347074b414eaSLai Jiangshan */ 347174b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3472699ce097STejun Heo pwq->max_active = wq->saved_max_active; 34730fbd95aaSTejun Heo 34740fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 34750fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 34760fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3477951a078aSLai Jiangshan 3478951a078aSLai Jiangshan /* 3479951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3480951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3481951a078aSLai Jiangshan */ 3482951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3483699ce097STejun Heo } else { 3484699ce097STejun Heo pwq->max_active = 0; 3485699ce097STejun Heo } 3486699ce097STejun Heo 34873347fa09STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 34880fbd95aaSTejun Heo } 34890fbd95aaSTejun Heo 3490e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3491f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3492f147f29eSTejun Heo struct worker_pool *pool) 3493d2c1d404STejun Heo { 3494d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3495d2c1d404STejun Heo 3496e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3497e50aba9aSTejun Heo 3498d2c1d404STejun Heo pwq->pool = pool; 3499d2c1d404STejun Heo pwq->wq = wq; 3500d2c1d404STejun Heo pwq->flush_color = -1; 35018864b4e5STejun Heo pwq->refcnt = 1; 3502d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 35031befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3504d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 35058864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3506f147f29eSTejun Heo } 3507d2c1d404STejun Heo 3508f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 35091befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3510f147f29eSTejun Heo { 3511f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3512f147f29eSTejun Heo 3513f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 351475ccf595STejun Heo 35151befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 35161befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 35171befcf30STejun Heo return; 35181befcf30STejun Heo 351929b1cb41SLai Jiangshan /* set the matching work_color */ 352075ccf595STejun Heo pwq->work_color = wq->work_color; 3521983ca25eSTejun Heo 3522983ca25eSTejun Heo /* sync max_active to the current setting */ 3523983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3524983ca25eSTejun Heo 3525983ca25eSTejun Heo /* link in @pwq */ 35269e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3527df2d5ae4STejun Heo } 35286029a918STejun Heo 3529f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3530f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3531f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3532f147f29eSTejun Heo { 3533f147f29eSTejun Heo struct worker_pool *pool; 3534f147f29eSTejun Heo struct pool_workqueue *pwq; 3535f147f29eSTejun Heo 3536f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3537f147f29eSTejun Heo 3538f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3539f147f29eSTejun Heo if (!pool) 3540f147f29eSTejun Heo return NULL; 3541f147f29eSTejun Heo 3542e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3543f147f29eSTejun Heo if (!pwq) { 3544f147f29eSTejun Heo put_unbound_pool(pool); 3545f147f29eSTejun Heo return NULL; 3546f147f29eSTejun Heo } 3547f147f29eSTejun Heo 3548f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3549f147f29eSTejun Heo return pwq; 3550d2c1d404STejun Heo } 3551d2c1d404STejun Heo 35524c16bd32STejun Heo /** 355330186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3554042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 35554c16bd32STejun Heo * @node: the target NUMA node 35564c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 35574c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 35584c16bd32STejun Heo * 35594c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 35604c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3561d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 35624c16bd32STejun Heo * 35634c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 35644c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 35654c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 35664c16bd32STejun Heo * @attrs->cpumask. 35674c16bd32STejun Heo * 35684c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 35694c16bd32STejun Heo * stable. 3570d185af30SYacine Belkadi * 3571d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3572d185af30SYacine Belkadi * %false if equal. 35734c16bd32STejun Heo */ 35744c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 35754c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 35764c16bd32STejun Heo { 3577d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 35784c16bd32STejun Heo goto use_dfl; 35794c16bd32STejun Heo 35804c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 35814c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 35824c16bd32STejun Heo if (cpu_going_down >= 0) 35834c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 35844c16bd32STejun Heo 35854c16bd32STejun Heo if (cpumask_empty(cpumask)) 35864c16bd32STejun Heo goto use_dfl; 35874c16bd32STejun Heo 35884c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 35894c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 35901ad0f0a7SMichael Bringmann 35911ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 35921ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 35931ad0f0a7SMichael Bringmann "possible intersect\n"); 35941ad0f0a7SMichael Bringmann return false; 35951ad0f0a7SMichael Bringmann } 35961ad0f0a7SMichael Bringmann 35974c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 35984c16bd32STejun Heo 35994c16bd32STejun Heo use_dfl: 36004c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 36014c16bd32STejun Heo return false; 36024c16bd32STejun Heo } 36034c16bd32STejun Heo 36041befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 36051befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 36061befcf30STejun Heo int node, 36071befcf30STejun Heo struct pool_workqueue *pwq) 36081befcf30STejun Heo { 36091befcf30STejun Heo struct pool_workqueue *old_pwq; 36101befcf30STejun Heo 36115b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36121befcf30STejun Heo lockdep_assert_held(&wq->mutex); 36131befcf30STejun Heo 36141befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 36151befcf30STejun Heo link_pwq(pwq); 36161befcf30STejun Heo 36171befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 36181befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 36191befcf30STejun Heo return old_pwq; 36201befcf30STejun Heo } 36211befcf30STejun Heo 36222d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 36232d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 36242d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 36252d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3626042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 36272d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 36282d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 36292d5f0764SLai Jiangshan }; 36302d5f0764SLai Jiangshan 36312d5f0764SLai Jiangshan /* free the resources after success or abort */ 36322d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 36332d5f0764SLai Jiangshan { 36342d5f0764SLai Jiangshan if (ctx) { 36352d5f0764SLai Jiangshan int node; 36362d5f0764SLai Jiangshan 36372d5f0764SLai Jiangshan for_each_node(node) 36382d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 36392d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 36402d5f0764SLai Jiangshan 36412d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 36422d5f0764SLai Jiangshan 36432d5f0764SLai Jiangshan kfree(ctx); 36442d5f0764SLai Jiangshan } 36452d5f0764SLai Jiangshan } 36462d5f0764SLai Jiangshan 36472d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 36482d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 36492d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 36502d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 36512d5f0764SLai Jiangshan { 36522d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 36532d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 36542d5f0764SLai Jiangshan int node; 36552d5f0764SLai Jiangshan 36562d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36572d5f0764SLai Jiangshan 36582d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 36592d5f0764SLai Jiangshan GFP_KERNEL); 36602d5f0764SLai Jiangshan 36612d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36622d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36632d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 36642d5f0764SLai Jiangshan goto out_free; 36652d5f0764SLai Jiangshan 3666042f7df1SLai Jiangshan /* 3667042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3668042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3669042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3670042f7df1SLai Jiangshan */ 36712d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3672b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3673042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3674042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 36752d5f0764SLai Jiangshan 36762d5f0764SLai Jiangshan /* 36772d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 36782d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 36792d5f0764SLai Jiangshan * pools. 36802d5f0764SLai Jiangshan */ 36812d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 36822d5f0764SLai Jiangshan 36832d5f0764SLai Jiangshan /* 36842d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 36852d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 36862d5f0764SLai Jiangshan * it even if we don't use it immediately. 36872d5f0764SLai Jiangshan */ 36882d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 36892d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 36902d5f0764SLai Jiangshan goto out_free; 36912d5f0764SLai Jiangshan 36922d5f0764SLai Jiangshan for_each_node(node) { 3693042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 36942d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 36952d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 36962d5f0764SLai Jiangshan goto out_free; 36972d5f0764SLai Jiangshan } else { 36982d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 36992d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 37002d5f0764SLai Jiangshan } 37012d5f0764SLai Jiangshan } 37022d5f0764SLai Jiangshan 3703042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3704042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3705042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 37062d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3707042f7df1SLai Jiangshan 37082d5f0764SLai Jiangshan ctx->wq = wq; 37092d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37102d5f0764SLai Jiangshan return ctx; 37112d5f0764SLai Jiangshan 37122d5f0764SLai Jiangshan out_free: 37132d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37142d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 37152d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 37162d5f0764SLai Jiangshan return NULL; 37172d5f0764SLai Jiangshan } 37182d5f0764SLai Jiangshan 37192d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 37202d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 37212d5f0764SLai Jiangshan { 37222d5f0764SLai Jiangshan int node; 37232d5f0764SLai Jiangshan 37242d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 37252d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 37262d5f0764SLai Jiangshan 37272d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 37282d5f0764SLai Jiangshan 37292d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 37302d5f0764SLai Jiangshan for_each_node(node) 37312d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 37322d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 37332d5f0764SLai Jiangshan 37342d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 37352d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 37362d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 37372d5f0764SLai Jiangshan 37382d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 37392d5f0764SLai Jiangshan } 37402d5f0764SLai Jiangshan 3741a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3742a0111cf6SLai Jiangshan { 3743a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3744a0111cf6SLai Jiangshan get_online_cpus(); 3745a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3746a0111cf6SLai Jiangshan } 3747a0111cf6SLai Jiangshan 3748a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3749a0111cf6SLai Jiangshan { 3750a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3751a0111cf6SLai Jiangshan put_online_cpus(); 3752a0111cf6SLai Jiangshan } 3753a0111cf6SLai Jiangshan 3754a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3755a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3756a0111cf6SLai Jiangshan { 3757a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3758a0111cf6SLai Jiangshan 3759a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3760a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3761a0111cf6SLai Jiangshan return -EINVAL; 3762a0111cf6SLai Jiangshan 3763a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 37640a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 37650a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 3766a0111cf6SLai Jiangshan return -EINVAL; 3767a0111cf6SLai Jiangshan 37680a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 37690a94efb5STejun Heo } 37700a94efb5STejun Heo 3771a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 37726201171eSwanghaibin if (!ctx) 37736201171eSwanghaibin return -ENOMEM; 3774a0111cf6SLai Jiangshan 3775a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3776a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3777a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3778a0111cf6SLai Jiangshan 37796201171eSwanghaibin return 0; 3780a0111cf6SLai Jiangshan } 3781a0111cf6SLai Jiangshan 37829e8cd2f5STejun Heo /** 37839e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37849e8cd2f5STejun Heo * @wq: the target workqueue 37859e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37869e8cd2f5STejun Heo * 37874c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 37884c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 37894c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 37904c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 37914c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 37924c16bd32STejun Heo * back-to-back will stay on its current pwq. 37939e8cd2f5STejun Heo * 3794d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3795d185af30SYacine Belkadi * 3796d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 37979e8cd2f5STejun Heo */ 37989e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 37999e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38009e8cd2f5STejun Heo { 3801a0111cf6SLai Jiangshan int ret; 38029e8cd2f5STejun Heo 3803a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3804a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3805a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 38062d5f0764SLai Jiangshan 38072d5f0764SLai Jiangshan return ret; 38089e8cd2f5STejun Heo } 38099e8cd2f5STejun Heo 38104c16bd32STejun Heo /** 38114c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 38124c16bd32STejun Heo * @wq: the target workqueue 38134c16bd32STejun Heo * @cpu: the CPU coming up or going down 38144c16bd32STejun Heo * @online: whether @cpu is coming up or going down 38154c16bd32STejun Heo * 38164c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 38174c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 38184c16bd32STejun Heo * @wq accordingly. 38194c16bd32STejun Heo * 38204c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 38214c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 38224c16bd32STejun Heo * correct. 38234c16bd32STejun Heo * 38244c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 38254c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 38264c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 38274c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 38284c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 38294c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 38304c16bd32STejun Heo * CPU_DOWN_PREPARE. 38314c16bd32STejun Heo */ 38324c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 38334c16bd32STejun Heo bool online) 38344c16bd32STejun Heo { 38354c16bd32STejun Heo int node = cpu_to_node(cpu); 38364c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 38374c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 38384c16bd32STejun Heo struct workqueue_attrs *target_attrs; 38394c16bd32STejun Heo cpumask_t *cpumask; 38404c16bd32STejun Heo 38414c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38424c16bd32STejun Heo 3843f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3844f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 38454c16bd32STejun Heo return; 38464c16bd32STejun Heo 38474c16bd32STejun Heo /* 38484c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 38494c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 38504c16bd32STejun Heo * CPU hotplug exclusion. 38514c16bd32STejun Heo */ 38524c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 38534c16bd32STejun Heo cpumask = target_attrs->cpumask; 38544c16bd32STejun Heo 38554c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 38564c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 38574c16bd32STejun Heo 38584c16bd32STejun Heo /* 38594c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3860042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3861042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3862042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 38634c16bd32STejun Heo */ 3864042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 38654c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3866f7142ed4SLai Jiangshan return; 38674c16bd32STejun Heo } else { 38684c16bd32STejun Heo goto use_dfl_pwq; 38694c16bd32STejun Heo } 38704c16bd32STejun Heo 38714c16bd32STejun Heo /* create a new pwq */ 38724c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 38734c16bd32STejun Heo if (!pwq) { 38742d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 38754c16bd32STejun Heo wq->name); 387677f300b1SDaeseok Youn goto use_dfl_pwq; 38774c16bd32STejun Heo } 38784c16bd32STejun Heo 3879f7142ed4SLai Jiangshan /* Install the new pwq. */ 38804c16bd32STejun Heo mutex_lock(&wq->mutex); 38814c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 38824c16bd32STejun Heo goto out_unlock; 38834c16bd32STejun Heo 38844c16bd32STejun Heo use_dfl_pwq: 3885f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 38864c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 38874c16bd32STejun Heo get_pwq(wq->dfl_pwq); 38884c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 38894c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 38904c16bd32STejun Heo out_unlock: 38914c16bd32STejun Heo mutex_unlock(&wq->mutex); 38924c16bd32STejun Heo put_pwq_unlocked(old_pwq); 38934c16bd32STejun Heo } 38944c16bd32STejun Heo 389530cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 38961da177e4SLinus Torvalds { 389749e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 38988a2b7538STejun Heo int cpu, ret; 3899e1d8aa9fSFrederic Weisbecker 390030cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3901420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3902420c0ddbSTejun Heo if (!wq->cpu_pwqs) 390330cdf249STejun Heo return -ENOMEM; 390430cdf249STejun Heo 390530cdf249STejun Heo for_each_possible_cpu(cpu) { 39067fb98ea7STejun Heo struct pool_workqueue *pwq = 39077fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 39087a62c2c8STejun Heo struct worker_pool *cpu_pools = 3909f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 391030cdf249STejun Heo 3911f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3912f147f29eSTejun Heo 3913f147f29eSTejun Heo mutex_lock(&wq->mutex); 39141befcf30STejun Heo link_pwq(pwq); 3915f147f29eSTejun Heo mutex_unlock(&wq->mutex); 391630cdf249STejun Heo } 391730cdf249STejun Heo return 0; 39188a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 39198a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 39208a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 39218a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 39228a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 39238a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 39248a2b7538STejun Heo return ret; 39259e8cd2f5STejun Heo } else { 39269e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 39279e8cd2f5STejun Heo } 39280f900049STejun Heo } 39290f900049STejun Heo 3930f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3931f3421797STejun Heo const char *name) 3932b71ab8c2STejun Heo { 3933f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3934f3421797STejun Heo 3935f3421797STejun Heo if (max_active < 1 || max_active > lim) 3936044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3937f3421797STejun Heo max_active, name, 1, lim); 3938b71ab8c2STejun Heo 3939f3421797STejun Heo return clamp_val(max_active, 1, lim); 3940b71ab8c2STejun Heo } 3941b71ab8c2STejun Heo 3942983c7515STejun Heo /* 3943983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 3944983c7515STejun Heo * to guarantee forward progress. 3945983c7515STejun Heo */ 3946983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 3947983c7515STejun Heo { 3948983c7515STejun Heo struct worker *rescuer; 3949983c7515STejun Heo int ret; 3950983c7515STejun Heo 3951983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 3952983c7515STejun Heo return 0; 3953983c7515STejun Heo 3954983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 3955983c7515STejun Heo if (!rescuer) 3956983c7515STejun Heo return -ENOMEM; 3957983c7515STejun Heo 3958983c7515STejun Heo rescuer->rescue_wq = wq; 3959983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 3960983c7515STejun Heo ret = PTR_ERR_OR_ZERO(rescuer->task); 3961983c7515STejun Heo if (ret) { 3962983c7515STejun Heo kfree(rescuer); 3963983c7515STejun Heo return ret; 3964983c7515STejun Heo } 3965983c7515STejun Heo 3966983c7515STejun Heo wq->rescuer = rescuer; 3967983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 3968983c7515STejun Heo wake_up_process(rescuer->task); 3969983c7515STejun Heo 3970983c7515STejun Heo return 0; 3971983c7515STejun Heo } 3972983c7515STejun Heo 3973b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 397497e37d7bSTejun Heo unsigned int flags, 39751e19ffc6STejun Heo int max_active, 3976eb13ba87SJohannes Berg struct lock_class_key *key, 3977b196be89STejun Heo const char *lock_name, ...) 39783af24433SOleg Nesterov { 3979df2d5ae4STejun Heo size_t tbl_size = 0; 3980ecf6881fSTejun Heo va_list args; 39813af24433SOleg Nesterov struct workqueue_struct *wq; 398249e3cf44STejun Heo struct pool_workqueue *pwq; 3983b196be89STejun Heo 39845c0338c6STejun Heo /* 39855c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 39865c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 39875c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 39885c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 39895c0338c6STejun Heo * on NUMA. 39905c0338c6STejun Heo */ 39915c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 39925c0338c6STejun Heo flags |= __WQ_ORDERED; 39935c0338c6STejun Heo 3994cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 3995cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 3996cee22a15SViresh Kumar flags |= WQ_UNBOUND; 3997cee22a15SViresh Kumar 3998ecf6881fSTejun Heo /* allocate wq and format name */ 3999df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4000ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4001df2d5ae4STejun Heo 4002df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4003b196be89STejun Heo if (!wq) 4004d2c1d404STejun Heo return NULL; 4005b196be89STejun Heo 40066029a918STejun Heo if (flags & WQ_UNBOUND) { 40076029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40086029a918STejun Heo if (!wq->unbound_attrs) 40096029a918STejun Heo goto err_free_wq; 40106029a918STejun Heo } 40116029a918STejun Heo 4012ecf6881fSTejun Heo va_start(args, lock_name); 4013ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4014b196be89STejun Heo va_end(args); 40153af24433SOleg Nesterov 4016d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4017b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40183af24433SOleg Nesterov 4019b196be89STejun Heo /* init wq */ 402097e37d7bSTejun Heo wq->flags = flags; 4021a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 40223c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4023112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 402430cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 402573f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 402673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4027493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 40283af24433SOleg Nesterov 4029eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4030cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 40313af24433SOleg Nesterov 403230cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4033d2c1d404STejun Heo goto err_free_wq; 40341537663fSTejun Heo 4035983c7515STejun Heo if (init_rescuer(wq) < 0) 4036d2c1d404STejun Heo goto err_destroy; 4037e22bee78STejun Heo 4038226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4039226223abSTejun Heo goto err_destroy; 4040226223abSTejun Heo 40416af8bf3dSOleg Nesterov /* 404268e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 404368e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 404468e13a67SLai Jiangshan * list. 40456af8bf3dSOleg Nesterov */ 404668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4047a0a1a5fdSTejun Heo 4048a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 404949e3cf44STejun Heo for_each_pwq(pwq, wq) 4050699ce097STejun Heo pwq_adjust_max_active(pwq); 4051a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4052a0a1a5fdSTejun Heo 4053e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4054a0a1a5fdSTejun Heo 405568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 40563af24433SOleg Nesterov 40573af24433SOleg Nesterov return wq; 4058d2c1d404STejun Heo 4059d2c1d404STejun Heo err_free_wq: 40606029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 40614690c4abSTejun Heo kfree(wq); 4062d2c1d404STejun Heo return NULL; 4063d2c1d404STejun Heo err_destroy: 4064d2c1d404STejun Heo destroy_workqueue(wq); 40654690c4abSTejun Heo return NULL; 40661da177e4SLinus Torvalds } 4067d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 40681da177e4SLinus Torvalds 40693af24433SOleg Nesterov /** 40703af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 40713af24433SOleg Nesterov * @wq: target workqueue 40723af24433SOleg Nesterov * 40733af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 40743af24433SOleg Nesterov */ 40753af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 40763af24433SOleg Nesterov { 407749e3cf44STejun Heo struct pool_workqueue *pwq; 40784c16bd32STejun Heo int node; 40793af24433SOleg Nesterov 40809c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 40819c5a2ba7STejun Heo drain_workqueue(wq); 4082c8efcc25STejun Heo 40836183c009STejun Heo /* sanity checks */ 4084b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 408549e3cf44STejun Heo for_each_pwq(pwq, wq) { 40866183c009STejun Heo int i; 40876183c009STejun Heo 408876af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 408976af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4090b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4091fa07fb6aSTejun Heo show_workqueue_state(); 40926183c009STejun Heo return; 409376af4d93STejun Heo } 409476af4d93STejun Heo } 409576af4d93STejun Heo 40965c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 40978864b4e5STejun Heo WARN_ON(pwq->nr_active) || 409876af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4099b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4100fa07fb6aSTejun Heo show_workqueue_state(); 41016183c009STejun Heo return; 41026183c009STejun Heo } 410376af4d93STejun Heo } 4104b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41056183c009STejun Heo 4106a0a1a5fdSTejun Heo /* 4107a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4108a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4109a0a1a5fdSTejun Heo */ 411068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4111e2dca7adSTejun Heo list_del_rcu(&wq->list); 411268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41133af24433SOleg Nesterov 4114226223abSTejun Heo workqueue_sysfs_unregister(wq); 4115226223abSTejun Heo 4116e2dca7adSTejun Heo if (wq->rescuer) 4117e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4118e22bee78STejun Heo 41198864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 412029c91e99STejun Heo /* 41218864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4122e2dca7adSTejun Heo * schedule RCU free. 412329c91e99STejun Heo */ 4124e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 41258864b4e5STejun Heo } else { 41268864b4e5STejun Heo /* 41278864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 41284c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 41294c16bd32STejun Heo * @wq will be freed when the last pwq is released. 41308864b4e5STejun Heo */ 41314c16bd32STejun Heo for_each_node(node) { 41324c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 41334c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 41344c16bd32STejun Heo put_pwq_unlocked(pwq); 41354c16bd32STejun Heo } 41364c16bd32STejun Heo 41374c16bd32STejun Heo /* 41384c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 41394c16bd32STejun Heo * put. Don't access it afterwards. 41404c16bd32STejun Heo */ 41414c16bd32STejun Heo pwq = wq->dfl_pwq; 41424c16bd32STejun Heo wq->dfl_pwq = NULL; 4143dce90d47STejun Heo put_pwq_unlocked(pwq); 414429c91e99STejun Heo } 41453af24433SOleg Nesterov } 41463af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 41473af24433SOleg Nesterov 4148dcd989cbSTejun Heo /** 4149dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4150dcd989cbSTejun Heo * @wq: target workqueue 4151dcd989cbSTejun Heo * @max_active: new max_active value. 4152dcd989cbSTejun Heo * 4153dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4154dcd989cbSTejun Heo * 4155dcd989cbSTejun Heo * CONTEXT: 4156dcd989cbSTejun Heo * Don't call from IRQ context. 4157dcd989cbSTejun Heo */ 4158dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4159dcd989cbSTejun Heo { 416049e3cf44STejun Heo struct pool_workqueue *pwq; 4161dcd989cbSTejun Heo 41628719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 41630a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 41648719dceaSTejun Heo return; 41658719dceaSTejun Heo 4166f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4167dcd989cbSTejun Heo 4168a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4169dcd989cbSTejun Heo 41700a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4171dcd989cbSTejun Heo wq->saved_max_active = max_active; 4172dcd989cbSTejun Heo 4173699ce097STejun Heo for_each_pwq(pwq, wq) 4174699ce097STejun Heo pwq_adjust_max_active(pwq); 4175dcd989cbSTejun Heo 4176a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4177dcd989cbSTejun Heo } 4178dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4179dcd989cbSTejun Heo 4180dcd989cbSTejun Heo /** 4181e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4182e6267616STejun Heo * 4183e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4184e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4185d185af30SYacine Belkadi * 4186d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4187e6267616STejun Heo */ 4188e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4189e6267616STejun Heo { 4190e6267616STejun Heo struct worker *worker = current_wq_worker(); 4191e6267616STejun Heo 41926a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4193e6267616STejun Heo } 4194e6267616STejun Heo 4195e6267616STejun Heo /** 4196dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4197dcd989cbSTejun Heo * @cpu: CPU in question 4198dcd989cbSTejun Heo * @wq: target workqueue 4199dcd989cbSTejun Heo * 4200dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4201dcd989cbSTejun Heo * no synchronization around this function and the test result is 4202dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4203dcd989cbSTejun Heo * 4204d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4205d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4206d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4207d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4208d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4209d3251859STejun Heo * 4210d185af30SYacine Belkadi * Return: 4211dcd989cbSTejun Heo * %true if congested, %false otherwise. 4212dcd989cbSTejun Heo */ 4213d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4214dcd989cbSTejun Heo { 42157fb98ea7STejun Heo struct pool_workqueue *pwq; 421676af4d93STejun Heo bool ret; 421776af4d93STejun Heo 421888109453SLai Jiangshan rcu_read_lock_sched(); 42197fb98ea7STejun Heo 4220d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4221d3251859STejun Heo cpu = smp_processor_id(); 4222d3251859STejun Heo 42237fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 42247fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 42257fb98ea7STejun Heo else 4226df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4227dcd989cbSTejun Heo 422876af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 422988109453SLai Jiangshan rcu_read_unlock_sched(); 423076af4d93STejun Heo 423176af4d93STejun Heo return ret; 4232dcd989cbSTejun Heo } 4233dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4234dcd989cbSTejun Heo 4235dcd989cbSTejun Heo /** 4236dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4237dcd989cbSTejun Heo * @work: the work to be tested 4238dcd989cbSTejun Heo * 4239dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4240dcd989cbSTejun Heo * synchronization around this function and the test result is 4241dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4242dcd989cbSTejun Heo * 4243d185af30SYacine Belkadi * Return: 4244dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4245dcd989cbSTejun Heo */ 4246dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4247dcd989cbSTejun Heo { 4248fa1b54e6STejun Heo struct worker_pool *pool; 4249dcd989cbSTejun Heo unsigned long flags; 4250dcd989cbSTejun Heo unsigned int ret = 0; 4251dcd989cbSTejun Heo 4252dcd989cbSTejun Heo if (work_pending(work)) 4253dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4254038366c5SLai Jiangshan 4255fa1b54e6STejun Heo local_irq_save(flags); 4256fa1b54e6STejun Heo pool = get_work_pool(work); 4257038366c5SLai Jiangshan if (pool) { 4258fa1b54e6STejun Heo spin_lock(&pool->lock); 4259c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4260dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4261fa1b54e6STejun Heo spin_unlock(&pool->lock); 4262038366c5SLai Jiangshan } 4263fa1b54e6STejun Heo local_irq_restore(flags); 4264dcd989cbSTejun Heo 4265dcd989cbSTejun Heo return ret; 4266dcd989cbSTejun Heo } 4267dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4268dcd989cbSTejun Heo 42693d1cb205STejun Heo /** 42703d1cb205STejun Heo * set_worker_desc - set description for the current work item 42713d1cb205STejun Heo * @fmt: printf-style format string 42723d1cb205STejun Heo * @...: arguments for the format string 42733d1cb205STejun Heo * 42743d1cb205STejun Heo * This function can be called by a running work function to describe what 42753d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 42763d1cb205STejun Heo * information will be printed out together to help debugging. The 42773d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 42783d1cb205STejun Heo */ 42793d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 42803d1cb205STejun Heo { 42813d1cb205STejun Heo struct worker *worker = current_wq_worker(); 42823d1cb205STejun Heo va_list args; 42833d1cb205STejun Heo 42843d1cb205STejun Heo if (worker) { 42853d1cb205STejun Heo va_start(args, fmt); 42863d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 42873d1cb205STejun Heo va_end(args); 42883d1cb205STejun Heo worker->desc_valid = true; 42893d1cb205STejun Heo } 42903d1cb205STejun Heo } 42913d1cb205STejun Heo 42923d1cb205STejun Heo /** 42933d1cb205STejun Heo * print_worker_info - print out worker information and description 42943d1cb205STejun Heo * @log_lvl: the log level to use when printing 42953d1cb205STejun Heo * @task: target task 42963d1cb205STejun Heo * 42973d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 42983d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 42993d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 43003d1cb205STejun Heo * 43013d1cb205STejun Heo * This function can be safely called on any task as long as the 43023d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 43033d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 43043d1cb205STejun Heo */ 43053d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 43063d1cb205STejun Heo { 43073d1cb205STejun Heo work_func_t *fn = NULL; 43083d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 43093d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 43103d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 43113d1cb205STejun Heo struct workqueue_struct *wq = NULL; 43123d1cb205STejun Heo bool desc_valid = false; 43133d1cb205STejun Heo struct worker *worker; 43143d1cb205STejun Heo 43153d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 43163d1cb205STejun Heo return; 43173d1cb205STejun Heo 43183d1cb205STejun Heo /* 43193d1cb205STejun Heo * This function is called without any synchronization and @task 43203d1cb205STejun Heo * could be in any state. Be careful with dereferences. 43213d1cb205STejun Heo */ 4322e700591aSPetr Mladek worker = kthread_probe_data(task); 43233d1cb205STejun Heo 43243d1cb205STejun Heo /* 43253d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 43263d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 43273d1cb205STejun Heo */ 43283d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 43293d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 43303d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 43313d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 43323d1cb205STejun Heo 43333d1cb205STejun Heo /* copy worker description */ 43343d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 43353d1cb205STejun Heo if (desc_valid) 43363d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 43373d1cb205STejun Heo 43383d1cb205STejun Heo if (fn || name[0] || desc[0]) { 43393d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 43403d1cb205STejun Heo if (desc[0]) 43413d1cb205STejun Heo pr_cont(" (%s)", desc); 43423d1cb205STejun Heo pr_cont("\n"); 43433d1cb205STejun Heo } 43443d1cb205STejun Heo } 43453d1cb205STejun Heo 43463494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 43473494fc30STejun Heo { 43483494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 43493494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 43503494fc30STejun Heo pr_cont(" node=%d", pool->node); 43513494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 43523494fc30STejun Heo } 43533494fc30STejun Heo 43543494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 43553494fc30STejun Heo { 43563494fc30STejun Heo if (work->func == wq_barrier_func) { 43573494fc30STejun Heo struct wq_barrier *barr; 43583494fc30STejun Heo 43593494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 43603494fc30STejun Heo 43613494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 43623494fc30STejun Heo task_pid_nr(barr->task)); 43633494fc30STejun Heo } else { 43643494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 43653494fc30STejun Heo } 43663494fc30STejun Heo } 43673494fc30STejun Heo 43683494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 43693494fc30STejun Heo { 43703494fc30STejun Heo struct worker_pool *pool = pwq->pool; 43713494fc30STejun Heo struct work_struct *work; 43723494fc30STejun Heo struct worker *worker; 43733494fc30STejun Heo bool has_in_flight = false, has_pending = false; 43743494fc30STejun Heo int bkt; 43753494fc30STejun Heo 43763494fc30STejun Heo pr_info(" pwq %d:", pool->id); 43773494fc30STejun Heo pr_cont_pool_info(pool); 43783494fc30STejun Heo 43793494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 43803494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 43813494fc30STejun Heo 43823494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43833494fc30STejun Heo if (worker->current_pwq == pwq) { 43843494fc30STejun Heo has_in_flight = true; 43853494fc30STejun Heo break; 43863494fc30STejun Heo } 43873494fc30STejun Heo } 43883494fc30STejun Heo if (has_in_flight) { 43893494fc30STejun Heo bool comma = false; 43903494fc30STejun Heo 43913494fc30STejun Heo pr_info(" in-flight:"); 43923494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43933494fc30STejun Heo if (worker->current_pwq != pwq) 43943494fc30STejun Heo continue; 43953494fc30STejun Heo 43963494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 43973494fc30STejun Heo task_pid_nr(worker->task), 43983494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 43993494fc30STejun Heo worker->current_func); 44003494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 44013494fc30STejun Heo pr_cont_work(false, work); 44023494fc30STejun Heo comma = true; 44033494fc30STejun Heo } 44043494fc30STejun Heo pr_cont("\n"); 44053494fc30STejun Heo } 44063494fc30STejun Heo 44073494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 44083494fc30STejun Heo if (get_work_pwq(work) == pwq) { 44093494fc30STejun Heo has_pending = true; 44103494fc30STejun Heo break; 44113494fc30STejun Heo } 44123494fc30STejun Heo } 44133494fc30STejun Heo if (has_pending) { 44143494fc30STejun Heo bool comma = false; 44153494fc30STejun Heo 44163494fc30STejun Heo pr_info(" pending:"); 44173494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 44183494fc30STejun Heo if (get_work_pwq(work) != pwq) 44193494fc30STejun Heo continue; 44203494fc30STejun Heo 44213494fc30STejun Heo pr_cont_work(comma, work); 44223494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 44233494fc30STejun Heo } 44243494fc30STejun Heo pr_cont("\n"); 44253494fc30STejun Heo } 44263494fc30STejun Heo 44273494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 44283494fc30STejun Heo bool comma = false; 44293494fc30STejun Heo 44303494fc30STejun Heo pr_info(" delayed:"); 44313494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 44323494fc30STejun Heo pr_cont_work(comma, work); 44333494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 44343494fc30STejun Heo } 44353494fc30STejun Heo pr_cont("\n"); 44363494fc30STejun Heo } 44373494fc30STejun Heo } 44383494fc30STejun Heo 44393494fc30STejun Heo /** 44403494fc30STejun Heo * show_workqueue_state - dump workqueue state 44413494fc30STejun Heo * 44427b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 44437b776af6SRoger Lu * all busy workqueues and pools. 44443494fc30STejun Heo */ 44453494fc30STejun Heo void show_workqueue_state(void) 44463494fc30STejun Heo { 44473494fc30STejun Heo struct workqueue_struct *wq; 44483494fc30STejun Heo struct worker_pool *pool; 44493494fc30STejun Heo unsigned long flags; 44503494fc30STejun Heo int pi; 44513494fc30STejun Heo 44523494fc30STejun Heo rcu_read_lock_sched(); 44533494fc30STejun Heo 44543494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 44553494fc30STejun Heo 44563494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 44573494fc30STejun Heo struct pool_workqueue *pwq; 44583494fc30STejun Heo bool idle = true; 44593494fc30STejun Heo 44603494fc30STejun Heo for_each_pwq(pwq, wq) { 44613494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 44623494fc30STejun Heo idle = false; 44633494fc30STejun Heo break; 44643494fc30STejun Heo } 44653494fc30STejun Heo } 44663494fc30STejun Heo if (idle) 44673494fc30STejun Heo continue; 44683494fc30STejun Heo 44693494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 44703494fc30STejun Heo 44713494fc30STejun Heo for_each_pwq(pwq, wq) { 44723494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 44733494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 44743494fc30STejun Heo show_pwq(pwq); 44753494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 44763494fc30STejun Heo } 44773494fc30STejun Heo } 44783494fc30STejun Heo 44793494fc30STejun Heo for_each_pool(pool, pi) { 44803494fc30STejun Heo struct worker *worker; 44813494fc30STejun Heo bool first = true; 44823494fc30STejun Heo 44833494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 44843494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 44853494fc30STejun Heo goto next_pool; 44863494fc30STejun Heo 44873494fc30STejun Heo pr_info("pool %d:", pool->id); 44883494fc30STejun Heo pr_cont_pool_info(pool); 448982607adcSTejun Heo pr_cont(" hung=%us workers=%d", 449082607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 449182607adcSTejun Heo pool->nr_workers); 44923494fc30STejun Heo if (pool->manager) 44933494fc30STejun Heo pr_cont(" manager: %d", 44943494fc30STejun Heo task_pid_nr(pool->manager->task)); 44953494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 44963494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 44973494fc30STejun Heo task_pid_nr(worker->task)); 44983494fc30STejun Heo first = false; 44993494fc30STejun Heo } 45003494fc30STejun Heo pr_cont("\n"); 45013494fc30STejun Heo next_pool: 45023494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 45033494fc30STejun Heo } 45043494fc30STejun Heo 45053494fc30STejun Heo rcu_read_unlock_sched(); 45063494fc30STejun Heo } 45073494fc30STejun Heo 4508db7bccf4STejun Heo /* 4509db7bccf4STejun Heo * CPU hotplug. 4510db7bccf4STejun Heo * 4511e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4512112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4513706026c2STejun Heo * pool which make migrating pending and scheduled works very 4514e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 451594cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4516e22bee78STejun Heo * blocked draining impractical. 4517e22bee78STejun Heo * 451824647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4519628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4520628c78e7STejun Heo * cpu comes back online. 4521db7bccf4STejun Heo */ 4522db7bccf4STejun Heo 4523e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4524db7bccf4STejun Heo { 45254ce62e9eSTejun Heo struct worker_pool *pool; 4526db7bccf4STejun Heo struct worker *worker; 4527db7bccf4STejun Heo 4528f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 452992f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 453094cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4531e22bee78STejun Heo 4532f2d5a0eeSTejun Heo /* 453392f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 453494cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 453594cf58bbSTejun Heo * except for the ones which are still executing works from 453694cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 453794cf58bbSTejun Heo * this, they may become diasporas. 4538f2d5a0eeSTejun Heo */ 4539da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4540403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4541db7bccf4STejun Heo 454224647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4543f2d5a0eeSTejun Heo 454494cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 454592f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4546e22bee78STejun Heo 4547e22bee78STejun Heo /* 4548eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4549eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4550eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4551eb283428SLai Jiangshan * from other cpus. 4552628c78e7STejun Heo */ 4553628c78e7STejun Heo schedule(); 4554628c78e7STejun Heo 4555628c78e7STejun Heo /* 4556eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4557eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4558eb283428SLai Jiangshan * and keep_working() are always true as long as the 4559eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4560eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4561eb283428SLai Jiangshan * are served by workers tied to the pool. 4562e22bee78STejun Heo */ 4563e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4564eb283428SLai Jiangshan 4565eb283428SLai Jiangshan /* 4566eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4567eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4568eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4569eb283428SLai Jiangshan */ 4570eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4571eb283428SLai Jiangshan wake_up_worker(pool); 4572eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4573eb283428SLai Jiangshan } 4574db7bccf4STejun Heo } 4575db7bccf4STejun Heo 4576bd7c089eSTejun Heo /** 4577bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4578bd7c089eSTejun Heo * @pool: pool of interest 4579bd7c089eSTejun Heo * 4580a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4581bd7c089eSTejun Heo */ 4582bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4583bd7c089eSTejun Heo { 4584a9ab775bSTejun Heo struct worker *worker; 4585bd7c089eSTejun Heo 458692f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4587bd7c089eSTejun Heo 4588bd7c089eSTejun Heo /* 4589a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4590a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4591402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4592a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4593a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4594bd7c089eSTejun Heo */ 4595da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4596a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4597a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4598a9ab775bSTejun Heo 4599a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4600f7c17d26SWanpeng Li 46013de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4602a9ab775bSTejun Heo 4603da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4604a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4605a9ab775bSTejun Heo 4606a9ab775bSTejun Heo /* 4607a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4608a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4609a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4610a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4611a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4612a9ab775bSTejun Heo * be bound before @pool->lock is released. 4613a9ab775bSTejun Heo */ 4614a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4615bd7c089eSTejun Heo wake_up_process(worker->task); 4616bd7c089eSTejun Heo 4617bd7c089eSTejun Heo /* 4618a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4619a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4620a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4621a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4622a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4623a9ab775bSTejun Heo * concurrency management. Note that when or whether 4624a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4625a9ab775bSTejun Heo * 4626c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 4627a9ab775bSTejun Heo * tested without holding any lock in 4628a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4629a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4630a9ab775bSTejun Heo * management operations. 4631bd7c089eSTejun Heo */ 4632a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4633a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4634a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4635c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 4636bd7c089eSTejun Heo } 4637a9ab775bSTejun Heo 4638a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4639bd7c089eSTejun Heo } 4640bd7c089eSTejun Heo 46417dbc725eSTejun Heo /** 46427dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 46437dbc725eSTejun Heo * @pool: unbound pool of interest 46447dbc725eSTejun Heo * @cpu: the CPU which is coming up 46457dbc725eSTejun Heo * 46467dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 46477dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 46487dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 46497dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 46507dbc725eSTejun Heo */ 46517dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 46527dbc725eSTejun Heo { 46537dbc725eSTejun Heo static cpumask_t cpumask; 46547dbc725eSTejun Heo struct worker *worker; 46557dbc725eSTejun Heo 465692f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 46577dbc725eSTejun Heo 46587dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 46597dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 46607dbc725eSTejun Heo return; 46617dbc725eSTejun Heo 46627dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 46637dbc725eSTejun Heo 46647dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4665da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4666d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 46677dbc725eSTejun Heo } 46687dbc725eSTejun Heo 46697ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 46701da177e4SLinus Torvalds { 46714ce62e9eSTejun Heo struct worker_pool *pool; 46721da177e4SLinus Torvalds 4673f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46743ce63377STejun Heo if (pool->nr_workers) 46753ce63377STejun Heo continue; 4676051e1850SLai Jiangshan if (!create_worker(pool)) 46777ee681b2SThomas Gleixner return -ENOMEM; 46783af24433SOleg Nesterov } 46797ee681b2SThomas Gleixner return 0; 46807ee681b2SThomas Gleixner } 46811da177e4SLinus Torvalds 46827ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 46837ee681b2SThomas Gleixner { 46847ee681b2SThomas Gleixner struct worker_pool *pool; 46857ee681b2SThomas Gleixner struct workqueue_struct *wq; 46867ee681b2SThomas Gleixner int pi; 46877ee681b2SThomas Gleixner 468868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46897dbc725eSTejun Heo 46907dbc725eSTejun Heo for_each_pool(pool, pi) { 469192f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 469294cf58bbSTejun Heo 4693f05b558dSLai Jiangshan if (pool->cpu == cpu) 469494cf58bbSTejun Heo rebind_workers(pool); 4695f05b558dSLai Jiangshan else if (pool->cpu < 0) 46967dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 469794cf58bbSTejun Heo 469892f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 469994cf58bbSTejun Heo } 47007dbc725eSTejun Heo 47014c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47024c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47034c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 47044c16bd32STejun Heo 470568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47067ee681b2SThomas Gleixner return 0; 470765758202STejun Heo } 470865758202STejun Heo 47097ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 471065758202STejun Heo { 47114c16bd32STejun Heo struct workqueue_struct *wq; 47128db25e78STejun Heo 47134c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4714e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 4715e8b3f8dbSLai Jiangshan return -1; 4716e8b3f8dbSLai Jiangshan 4717e8b3f8dbSLai Jiangshan unbind_workers(cpu); 47184c16bd32STejun Heo 47194c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47204c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 47214c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47224c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 47234c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 47244c16bd32STejun Heo 47257ee681b2SThomas Gleixner return 0; 472665758202STejun Heo } 472765758202STejun Heo 47282d3854a3SRusty Russell #ifdef CONFIG_SMP 47298ccad40dSRusty Russell 47302d3854a3SRusty Russell struct work_for_cpu { 4731ed48ece2STejun Heo struct work_struct work; 47322d3854a3SRusty Russell long (*fn)(void *); 47332d3854a3SRusty Russell void *arg; 47342d3854a3SRusty Russell long ret; 47352d3854a3SRusty Russell }; 47362d3854a3SRusty Russell 4737ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 47382d3854a3SRusty Russell { 4739ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4740ed48ece2STejun Heo 47412d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 47422d3854a3SRusty Russell } 47432d3854a3SRusty Russell 47442d3854a3SRusty Russell /** 474522aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 47462d3854a3SRusty Russell * @cpu: the cpu to run on 47472d3854a3SRusty Russell * @fn: the function to run 47482d3854a3SRusty Russell * @arg: the function arg 47492d3854a3SRusty Russell * 475031ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47516b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4752d185af30SYacine Belkadi * 4753d185af30SYacine Belkadi * Return: The value @fn returns. 47542d3854a3SRusty Russell */ 4755d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47562d3854a3SRusty Russell { 4757ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47582d3854a3SRusty Russell 4759ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4760ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 476112997d1aSBjorn Helgaas flush_work(&wfc.work); 4762440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 47632d3854a3SRusty Russell return wfc.ret; 47642d3854a3SRusty Russell } 47652d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47660e8d6a93SThomas Gleixner 47670e8d6a93SThomas Gleixner /** 47680e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 47690e8d6a93SThomas Gleixner * @cpu: the cpu to run on 47700e8d6a93SThomas Gleixner * @fn: the function to run 47710e8d6a93SThomas Gleixner * @arg: the function argument 47720e8d6a93SThomas Gleixner * 47730e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 47740e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 47750e8d6a93SThomas Gleixner * 47760e8d6a93SThomas Gleixner * Return: The value @fn returns. 47770e8d6a93SThomas Gleixner */ 47780e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 47790e8d6a93SThomas Gleixner { 47800e8d6a93SThomas Gleixner long ret = -ENODEV; 47810e8d6a93SThomas Gleixner 47820e8d6a93SThomas Gleixner get_online_cpus(); 47830e8d6a93SThomas Gleixner if (cpu_online(cpu)) 47840e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 47850e8d6a93SThomas Gleixner put_online_cpus(); 47860e8d6a93SThomas Gleixner return ret; 47870e8d6a93SThomas Gleixner } 47880e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 47892d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47902d3854a3SRusty Russell 4791a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4792e7577c50SRusty Russell 4793a0a1a5fdSTejun Heo /** 4794a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4795a0a1a5fdSTejun Heo * 479658a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4797c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4798706026c2STejun Heo * pool->worklist. 4799a0a1a5fdSTejun Heo * 4800a0a1a5fdSTejun Heo * CONTEXT: 4801a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4802a0a1a5fdSTejun Heo */ 4803a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4804a0a1a5fdSTejun Heo { 480524b8a847STejun Heo struct workqueue_struct *wq; 480624b8a847STejun Heo struct pool_workqueue *pwq; 4807a0a1a5fdSTejun Heo 480868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4809a0a1a5fdSTejun Heo 48106183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4811a0a1a5fdSTejun Heo workqueue_freezing = true; 4812a0a1a5fdSTejun Heo 481324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4814a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4815699ce097STejun Heo for_each_pwq(pwq, wq) 4816699ce097STejun Heo pwq_adjust_max_active(pwq); 4817a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4818a1056305STejun Heo } 48195bcab335STejun Heo 482068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4821a0a1a5fdSTejun Heo } 4822a0a1a5fdSTejun Heo 4823a0a1a5fdSTejun Heo /** 482458a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4825a0a1a5fdSTejun Heo * 4826a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4827a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4828a0a1a5fdSTejun Heo * 4829a0a1a5fdSTejun Heo * CONTEXT: 483068e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4831a0a1a5fdSTejun Heo * 4832d185af30SYacine Belkadi * Return: 483358a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 483458a69cb4STejun Heo * is complete. 4835a0a1a5fdSTejun Heo */ 4836a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4837a0a1a5fdSTejun Heo { 4838a0a1a5fdSTejun Heo bool busy = false; 483924b8a847STejun Heo struct workqueue_struct *wq; 484024b8a847STejun Heo struct pool_workqueue *pwq; 4841a0a1a5fdSTejun Heo 484268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4843a0a1a5fdSTejun Heo 48446183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4845a0a1a5fdSTejun Heo 484624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 484724b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 484824b8a847STejun Heo continue; 4849a0a1a5fdSTejun Heo /* 4850a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4851a0a1a5fdSTejun Heo * to peek without lock. 4852a0a1a5fdSTejun Heo */ 485388109453SLai Jiangshan rcu_read_lock_sched(); 485424b8a847STejun Heo for_each_pwq(pwq, wq) { 48556183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4856112202d9STejun Heo if (pwq->nr_active) { 4857a0a1a5fdSTejun Heo busy = true; 485888109453SLai Jiangshan rcu_read_unlock_sched(); 4859a0a1a5fdSTejun Heo goto out_unlock; 4860a0a1a5fdSTejun Heo } 4861a0a1a5fdSTejun Heo } 486288109453SLai Jiangshan rcu_read_unlock_sched(); 4863a0a1a5fdSTejun Heo } 4864a0a1a5fdSTejun Heo out_unlock: 486568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4866a0a1a5fdSTejun Heo return busy; 4867a0a1a5fdSTejun Heo } 4868a0a1a5fdSTejun Heo 4869a0a1a5fdSTejun Heo /** 4870a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4871a0a1a5fdSTejun Heo * 4872a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4873706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4874a0a1a5fdSTejun Heo * 4875a0a1a5fdSTejun Heo * CONTEXT: 4876a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4877a0a1a5fdSTejun Heo */ 4878a0a1a5fdSTejun Heo void thaw_workqueues(void) 4879a0a1a5fdSTejun Heo { 488024b8a847STejun Heo struct workqueue_struct *wq; 488124b8a847STejun Heo struct pool_workqueue *pwq; 4882a0a1a5fdSTejun Heo 488368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4884a0a1a5fdSTejun Heo 4885a0a1a5fdSTejun Heo if (!workqueue_freezing) 4886a0a1a5fdSTejun Heo goto out_unlock; 4887a0a1a5fdSTejun Heo 488874b414eaSLai Jiangshan workqueue_freezing = false; 488924b8a847STejun Heo 489024b8a847STejun Heo /* restore max_active and repopulate worklist */ 489124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4892a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4893699ce097STejun Heo for_each_pwq(pwq, wq) 4894699ce097STejun Heo pwq_adjust_max_active(pwq); 4895a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 489624b8a847STejun Heo } 489724b8a847STejun Heo 4898a0a1a5fdSTejun Heo out_unlock: 489968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4900a0a1a5fdSTejun Heo } 4901a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4902a0a1a5fdSTejun Heo 4903042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4904042f7df1SLai Jiangshan { 4905042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4906042f7df1SLai Jiangshan int ret = 0; 4907042f7df1SLai Jiangshan struct workqueue_struct *wq; 4908042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4909042f7df1SLai Jiangshan 4910042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4911042f7df1SLai Jiangshan 4912042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4913042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4914042f7df1SLai Jiangshan continue; 4915042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 4916042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 4917042f7df1SLai Jiangshan continue; 4918042f7df1SLai Jiangshan 4919042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 4920042f7df1SLai Jiangshan if (!ctx) { 4921042f7df1SLai Jiangshan ret = -ENOMEM; 4922042f7df1SLai Jiangshan break; 4923042f7df1SLai Jiangshan } 4924042f7df1SLai Jiangshan 4925042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 4926042f7df1SLai Jiangshan } 4927042f7df1SLai Jiangshan 4928042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 4929042f7df1SLai Jiangshan if (!ret) 4930042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 4931042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 4932042f7df1SLai Jiangshan } 4933042f7df1SLai Jiangshan 4934042f7df1SLai Jiangshan return ret; 4935042f7df1SLai Jiangshan } 4936042f7df1SLai Jiangshan 4937042f7df1SLai Jiangshan /** 4938042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 4939042f7df1SLai Jiangshan * @cpumask: the cpumask to set 4940042f7df1SLai Jiangshan * 4941042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 4942042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 4943042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 4944042f7df1SLai Jiangshan * 4945042f7df1SLai Jiangshan * Retun: 0 - Success 4946042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 4947042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 4948042f7df1SLai Jiangshan */ 4949042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 4950042f7df1SLai Jiangshan { 4951042f7df1SLai Jiangshan int ret = -EINVAL; 4952042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 4953042f7df1SLai Jiangshan 4954042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 4955042f7df1SLai Jiangshan return -ENOMEM; 4956042f7df1SLai Jiangshan 4957c98a9805STal Shorer /* 4958c98a9805STal Shorer * Not excluding isolated cpus on purpose. 4959c98a9805STal Shorer * If the user wishes to include them, we allow that. 4960c98a9805STal Shorer */ 4961042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 4962042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 4963a0111cf6SLai Jiangshan apply_wqattrs_lock(); 4964042f7df1SLai Jiangshan 4965042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 4966042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 4967042f7df1SLai Jiangshan 4968042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 4969042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 4970042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 4971042f7df1SLai Jiangshan 4972042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 4973042f7df1SLai Jiangshan if (ret < 0) 4974042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 4975042f7df1SLai Jiangshan 4976a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 4977042f7df1SLai Jiangshan } 4978042f7df1SLai Jiangshan 4979042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 4980042f7df1SLai Jiangshan return ret; 4981042f7df1SLai Jiangshan } 4982042f7df1SLai Jiangshan 49836ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 49846ba94429SFrederic Weisbecker /* 49856ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 49866ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 49876ba94429SFrederic Weisbecker * following attributes. 49886ba94429SFrederic Weisbecker * 49896ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 49906ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 49916ba94429SFrederic Weisbecker * 49926ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 49936ba94429SFrederic Weisbecker * 49949a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 49956ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 49966ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 49979a19b463SWang Long * numa RW bool : whether enable NUMA affinity 49986ba94429SFrederic Weisbecker */ 49996ba94429SFrederic Weisbecker struct wq_device { 50006ba94429SFrederic Weisbecker struct workqueue_struct *wq; 50016ba94429SFrederic Weisbecker struct device dev; 50026ba94429SFrederic Weisbecker }; 50036ba94429SFrederic Weisbecker 50046ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 50056ba94429SFrederic Weisbecker { 50066ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 50076ba94429SFrederic Weisbecker 50086ba94429SFrederic Weisbecker return wq_dev->wq; 50096ba94429SFrederic Weisbecker } 50106ba94429SFrederic Weisbecker 50116ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 50126ba94429SFrederic Weisbecker char *buf) 50136ba94429SFrederic Weisbecker { 50146ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50156ba94429SFrederic Weisbecker 50166ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 50176ba94429SFrederic Weisbecker } 50186ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 50196ba94429SFrederic Weisbecker 50206ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 50216ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 50226ba94429SFrederic Weisbecker { 50236ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50246ba94429SFrederic Weisbecker 50256ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 50266ba94429SFrederic Weisbecker } 50276ba94429SFrederic Weisbecker 50286ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 50296ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 50306ba94429SFrederic Weisbecker size_t count) 50316ba94429SFrederic Weisbecker { 50326ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50336ba94429SFrederic Weisbecker int val; 50346ba94429SFrederic Weisbecker 50356ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 50366ba94429SFrederic Weisbecker return -EINVAL; 50376ba94429SFrederic Weisbecker 50386ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 50396ba94429SFrederic Weisbecker return count; 50406ba94429SFrederic Weisbecker } 50416ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 50426ba94429SFrederic Weisbecker 50436ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 50446ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 50456ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 50466ba94429SFrederic Weisbecker NULL, 50476ba94429SFrederic Weisbecker }; 50486ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 50496ba94429SFrederic Weisbecker 50506ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 50516ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 50526ba94429SFrederic Weisbecker { 50536ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50546ba94429SFrederic Weisbecker const char *delim = ""; 50556ba94429SFrederic Weisbecker int node, written = 0; 50566ba94429SFrederic Weisbecker 50576ba94429SFrederic Weisbecker rcu_read_lock_sched(); 50586ba94429SFrederic Weisbecker for_each_node(node) { 50596ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 50606ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 50616ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 50626ba94429SFrederic Weisbecker delim = " "; 50636ba94429SFrederic Weisbecker } 50646ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 50656ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 50666ba94429SFrederic Weisbecker 50676ba94429SFrederic Weisbecker return written; 50686ba94429SFrederic Weisbecker } 50696ba94429SFrederic Weisbecker 50706ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 50716ba94429SFrederic Weisbecker char *buf) 50726ba94429SFrederic Weisbecker { 50736ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50746ba94429SFrederic Weisbecker int written; 50756ba94429SFrederic Weisbecker 50766ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50776ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 50786ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50796ba94429SFrederic Weisbecker 50806ba94429SFrederic Weisbecker return written; 50816ba94429SFrederic Weisbecker } 50826ba94429SFrederic Weisbecker 50836ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 50846ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 50856ba94429SFrederic Weisbecker { 50866ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 50876ba94429SFrederic Weisbecker 5088899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5089899a94feSLai Jiangshan 50906ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 50916ba94429SFrederic Weisbecker if (!attrs) 50926ba94429SFrederic Weisbecker return NULL; 50936ba94429SFrederic Weisbecker 50946ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 50956ba94429SFrederic Weisbecker return attrs; 50966ba94429SFrederic Weisbecker } 50976ba94429SFrederic Weisbecker 50986ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 50996ba94429SFrederic Weisbecker const char *buf, size_t count) 51006ba94429SFrederic Weisbecker { 51016ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51026ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5103d4d3e257SLai Jiangshan int ret = -ENOMEM; 5104d4d3e257SLai Jiangshan 5105d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51066ba94429SFrederic Weisbecker 51076ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51086ba94429SFrederic Weisbecker if (!attrs) 5109d4d3e257SLai Jiangshan goto out_unlock; 51106ba94429SFrederic Weisbecker 51116ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 51126ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5113d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51146ba94429SFrederic Weisbecker else 51156ba94429SFrederic Weisbecker ret = -EINVAL; 51166ba94429SFrederic Weisbecker 5117d4d3e257SLai Jiangshan out_unlock: 5118d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 51196ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 51206ba94429SFrederic Weisbecker return ret ?: count; 51216ba94429SFrederic Weisbecker } 51226ba94429SFrederic Weisbecker 51236ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 51246ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 51256ba94429SFrederic Weisbecker { 51266ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51276ba94429SFrederic Weisbecker int written; 51286ba94429SFrederic Weisbecker 51296ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 51306ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 51316ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 51326ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 51336ba94429SFrederic Weisbecker return written; 51346ba94429SFrederic Weisbecker } 51356ba94429SFrederic Weisbecker 51366ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 51376ba94429SFrederic Weisbecker struct device_attribute *attr, 51386ba94429SFrederic Weisbecker const char *buf, size_t count) 51396ba94429SFrederic Weisbecker { 51406ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51416ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5142d4d3e257SLai Jiangshan int ret = -ENOMEM; 5143d4d3e257SLai Jiangshan 5144d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51456ba94429SFrederic Weisbecker 51466ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51476ba94429SFrederic Weisbecker if (!attrs) 5148d4d3e257SLai Jiangshan goto out_unlock; 51496ba94429SFrederic Weisbecker 51506ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 51516ba94429SFrederic Weisbecker if (!ret) 5152d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51536ba94429SFrederic Weisbecker 5154d4d3e257SLai Jiangshan out_unlock: 5155d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 51566ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 51576ba94429SFrederic Weisbecker return ret ?: count; 51586ba94429SFrederic Weisbecker } 51596ba94429SFrederic Weisbecker 51606ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 51616ba94429SFrederic Weisbecker char *buf) 51626ba94429SFrederic Weisbecker { 51636ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51646ba94429SFrederic Weisbecker int written; 51656ba94429SFrederic Weisbecker 51666ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 51676ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 51686ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 51696ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 51706ba94429SFrederic Weisbecker 51716ba94429SFrederic Weisbecker return written; 51726ba94429SFrederic Weisbecker } 51736ba94429SFrederic Weisbecker 51746ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 51756ba94429SFrederic Weisbecker const char *buf, size_t count) 51766ba94429SFrederic Weisbecker { 51776ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51786ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5179d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5180d4d3e257SLai Jiangshan 5181d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51826ba94429SFrederic Weisbecker 51836ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51846ba94429SFrederic Weisbecker if (!attrs) 5185d4d3e257SLai Jiangshan goto out_unlock; 51866ba94429SFrederic Weisbecker 51876ba94429SFrederic Weisbecker ret = -EINVAL; 51886ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 51896ba94429SFrederic Weisbecker attrs->no_numa = !v; 5190d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51916ba94429SFrederic Weisbecker } 51926ba94429SFrederic Weisbecker 5193d4d3e257SLai Jiangshan out_unlock: 5194d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 51956ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 51966ba94429SFrederic Weisbecker return ret ?: count; 51976ba94429SFrederic Weisbecker } 51986ba94429SFrederic Weisbecker 51996ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 52006ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 52016ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 52026ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 52036ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 52046ba94429SFrederic Weisbecker __ATTR_NULL, 52056ba94429SFrederic Weisbecker }; 52066ba94429SFrederic Weisbecker 52076ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 52086ba94429SFrederic Weisbecker .name = "workqueue", 52096ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 52106ba94429SFrederic Weisbecker }; 52116ba94429SFrederic Weisbecker 5212b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5213b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5214b05a7928SFrederic Weisbecker { 5215b05a7928SFrederic Weisbecker int written; 5216b05a7928SFrederic Weisbecker 5217042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5218b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5219b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5220042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5221b05a7928SFrederic Weisbecker 5222b05a7928SFrederic Weisbecker return written; 5223b05a7928SFrederic Weisbecker } 5224b05a7928SFrederic Weisbecker 5225042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5226042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5227042f7df1SLai Jiangshan { 5228042f7df1SLai Jiangshan cpumask_var_t cpumask; 5229042f7df1SLai Jiangshan int ret; 5230042f7df1SLai Jiangshan 5231042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5232042f7df1SLai Jiangshan return -ENOMEM; 5233042f7df1SLai Jiangshan 5234042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5235042f7df1SLai Jiangshan if (!ret) 5236042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5237042f7df1SLai Jiangshan 5238042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5239042f7df1SLai Jiangshan return ret ? ret : count; 5240042f7df1SLai Jiangshan } 5241042f7df1SLai Jiangshan 5242b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5243042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5244042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5245b05a7928SFrederic Weisbecker 52466ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 52476ba94429SFrederic Weisbecker { 5248b05a7928SFrederic Weisbecker int err; 5249b05a7928SFrederic Weisbecker 5250b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5251b05a7928SFrederic Weisbecker if (err) 5252b05a7928SFrederic Weisbecker return err; 5253b05a7928SFrederic Weisbecker 5254b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 52556ba94429SFrederic Weisbecker } 52566ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 52576ba94429SFrederic Weisbecker 52586ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 52596ba94429SFrederic Weisbecker { 52606ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 52616ba94429SFrederic Weisbecker 52626ba94429SFrederic Weisbecker kfree(wq_dev); 52636ba94429SFrederic Weisbecker } 52646ba94429SFrederic Weisbecker 52656ba94429SFrederic Weisbecker /** 52666ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 52676ba94429SFrederic Weisbecker * @wq: the workqueue to register 52686ba94429SFrederic Weisbecker * 52696ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 52706ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 52716ba94429SFrederic Weisbecker * which is the preferred method. 52726ba94429SFrederic Weisbecker * 52736ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 52746ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 52756ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 52766ba94429SFrederic Weisbecker * attributes. 52776ba94429SFrederic Weisbecker * 52786ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 52796ba94429SFrederic Weisbecker */ 52806ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 52816ba94429SFrederic Weisbecker { 52826ba94429SFrederic Weisbecker struct wq_device *wq_dev; 52836ba94429SFrederic Weisbecker int ret; 52846ba94429SFrederic Weisbecker 52856ba94429SFrederic Weisbecker /* 5286402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 52876ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 52886ba94429SFrederic Weisbecker * workqueues. 52896ba94429SFrederic Weisbecker */ 52900a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 52916ba94429SFrederic Weisbecker return -EINVAL; 52926ba94429SFrederic Weisbecker 52936ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 52946ba94429SFrederic Weisbecker if (!wq_dev) 52956ba94429SFrederic Weisbecker return -ENOMEM; 52966ba94429SFrederic Weisbecker 52976ba94429SFrederic Weisbecker wq_dev->wq = wq; 52986ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 52996ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 530023217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 53016ba94429SFrederic Weisbecker 53026ba94429SFrederic Weisbecker /* 53036ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 53046ba94429SFrederic Weisbecker * everything is ready. 53056ba94429SFrederic Weisbecker */ 53066ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 53076ba94429SFrederic Weisbecker 53086ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 53096ba94429SFrederic Weisbecker if (ret) { 53106ba94429SFrederic Weisbecker kfree(wq_dev); 53116ba94429SFrederic Weisbecker wq->wq_dev = NULL; 53126ba94429SFrederic Weisbecker return ret; 53136ba94429SFrederic Weisbecker } 53146ba94429SFrederic Weisbecker 53156ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 53166ba94429SFrederic Weisbecker struct device_attribute *attr; 53176ba94429SFrederic Weisbecker 53186ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 53196ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 53206ba94429SFrederic Weisbecker if (ret) { 53216ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 53226ba94429SFrederic Weisbecker wq->wq_dev = NULL; 53236ba94429SFrederic Weisbecker return ret; 53246ba94429SFrederic Weisbecker } 53256ba94429SFrederic Weisbecker } 53266ba94429SFrederic Weisbecker } 53276ba94429SFrederic Weisbecker 53286ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 53296ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 53306ba94429SFrederic Weisbecker return 0; 53316ba94429SFrederic Weisbecker } 53326ba94429SFrederic Weisbecker 53336ba94429SFrederic Weisbecker /** 53346ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 53356ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 53366ba94429SFrederic Weisbecker * 53376ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 53386ba94429SFrederic Weisbecker */ 53396ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 53406ba94429SFrederic Weisbecker { 53416ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 53426ba94429SFrederic Weisbecker 53436ba94429SFrederic Weisbecker if (!wq->wq_dev) 53446ba94429SFrederic Weisbecker return; 53456ba94429SFrederic Weisbecker 53466ba94429SFrederic Weisbecker wq->wq_dev = NULL; 53476ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 53486ba94429SFrederic Weisbecker } 53496ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 53506ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 53516ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 53526ba94429SFrederic Weisbecker 535382607adcSTejun Heo /* 535482607adcSTejun Heo * Workqueue watchdog. 535582607adcSTejun Heo * 535682607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 535782607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 535882607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 535982607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 536082607adcSTejun Heo * largely opaque. 536182607adcSTejun Heo * 536282607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 536382607adcSTejun Heo * state if some pools failed to make forward progress for a while where 536482607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 536582607adcSTejun Heo * 536682607adcSTejun Heo * This mechanism is controlled through the kernel parameter 536782607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 536882607adcSTejun Heo * corresponding sysfs parameter file. 536982607adcSTejun Heo */ 537082607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 537182607adcSTejun Heo 537282607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 53735cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 537482607adcSTejun Heo 537582607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 537682607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 537782607adcSTejun Heo 537882607adcSTejun Heo static void wq_watchdog_reset_touched(void) 537982607adcSTejun Heo { 538082607adcSTejun Heo int cpu; 538182607adcSTejun Heo 538282607adcSTejun Heo wq_watchdog_touched = jiffies; 538382607adcSTejun Heo for_each_possible_cpu(cpu) 538482607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 538582607adcSTejun Heo } 538682607adcSTejun Heo 53875cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 538882607adcSTejun Heo { 538982607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 539082607adcSTejun Heo bool lockup_detected = false; 539182607adcSTejun Heo struct worker_pool *pool; 539282607adcSTejun Heo int pi; 539382607adcSTejun Heo 539482607adcSTejun Heo if (!thresh) 539582607adcSTejun Heo return; 539682607adcSTejun Heo 539782607adcSTejun Heo rcu_read_lock(); 539882607adcSTejun Heo 539982607adcSTejun Heo for_each_pool(pool, pi) { 540082607adcSTejun Heo unsigned long pool_ts, touched, ts; 540182607adcSTejun Heo 540282607adcSTejun Heo if (list_empty(&pool->worklist)) 540382607adcSTejun Heo continue; 540482607adcSTejun Heo 540582607adcSTejun Heo /* get the latest of pool and touched timestamps */ 540682607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 540782607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 540882607adcSTejun Heo 540982607adcSTejun Heo if (time_after(pool_ts, touched)) 541082607adcSTejun Heo ts = pool_ts; 541182607adcSTejun Heo else 541282607adcSTejun Heo ts = touched; 541382607adcSTejun Heo 541482607adcSTejun Heo if (pool->cpu >= 0) { 541582607adcSTejun Heo unsigned long cpu_touched = 541682607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 541782607adcSTejun Heo pool->cpu)); 541882607adcSTejun Heo if (time_after(cpu_touched, ts)) 541982607adcSTejun Heo ts = cpu_touched; 542082607adcSTejun Heo } 542182607adcSTejun Heo 542282607adcSTejun Heo /* did we stall? */ 542382607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 542482607adcSTejun Heo lockup_detected = true; 542582607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 542682607adcSTejun Heo pr_cont_pool_info(pool); 542782607adcSTejun Heo pr_cont(" stuck for %us!\n", 542882607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 542982607adcSTejun Heo } 543082607adcSTejun Heo } 543182607adcSTejun Heo 543282607adcSTejun Heo rcu_read_unlock(); 543382607adcSTejun Heo 543482607adcSTejun Heo if (lockup_detected) 543582607adcSTejun Heo show_workqueue_state(); 543682607adcSTejun Heo 543782607adcSTejun Heo wq_watchdog_reset_touched(); 543882607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 543982607adcSTejun Heo } 544082607adcSTejun Heo 544182607adcSTejun Heo void wq_watchdog_touch(int cpu) 544282607adcSTejun Heo { 544382607adcSTejun Heo if (cpu >= 0) 544482607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 544582607adcSTejun Heo else 544682607adcSTejun Heo wq_watchdog_touched = jiffies; 544782607adcSTejun Heo } 544882607adcSTejun Heo 544982607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 545082607adcSTejun Heo { 545182607adcSTejun Heo wq_watchdog_thresh = 0; 545282607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 545382607adcSTejun Heo 545482607adcSTejun Heo if (thresh) { 545582607adcSTejun Heo wq_watchdog_thresh = thresh; 545682607adcSTejun Heo wq_watchdog_reset_touched(); 545782607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 545882607adcSTejun Heo } 545982607adcSTejun Heo } 546082607adcSTejun Heo 546182607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 546282607adcSTejun Heo const struct kernel_param *kp) 546382607adcSTejun Heo { 546482607adcSTejun Heo unsigned long thresh; 546582607adcSTejun Heo int ret; 546682607adcSTejun Heo 546782607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 546882607adcSTejun Heo if (ret) 546982607adcSTejun Heo return ret; 547082607adcSTejun Heo 547182607adcSTejun Heo if (system_wq) 547282607adcSTejun Heo wq_watchdog_set_thresh(thresh); 547382607adcSTejun Heo else 547482607adcSTejun Heo wq_watchdog_thresh = thresh; 547582607adcSTejun Heo 547682607adcSTejun Heo return 0; 547782607adcSTejun Heo } 547882607adcSTejun Heo 547982607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 548082607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 548182607adcSTejun Heo .get = param_get_ulong, 548282607adcSTejun Heo }; 548382607adcSTejun Heo 548482607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 548582607adcSTejun Heo 0644); 548682607adcSTejun Heo 548782607adcSTejun Heo static void wq_watchdog_init(void) 548882607adcSTejun Heo { 54895cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 549082607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 549182607adcSTejun Heo } 549282607adcSTejun Heo 549382607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 549482607adcSTejun Heo 549582607adcSTejun Heo static inline void wq_watchdog_init(void) { } 549682607adcSTejun Heo 549782607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 549882607adcSTejun Heo 5499bce90380STejun Heo static void __init wq_numa_init(void) 5500bce90380STejun Heo { 5501bce90380STejun Heo cpumask_var_t *tbl; 5502bce90380STejun Heo int node, cpu; 5503bce90380STejun Heo 5504bce90380STejun Heo if (num_possible_nodes() <= 1) 5505bce90380STejun Heo return; 5506bce90380STejun Heo 5507d55262c4STejun Heo if (wq_disable_numa) { 5508d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5509d55262c4STejun Heo return; 5510d55262c4STejun Heo } 5511d55262c4STejun Heo 55124c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 55134c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 55144c16bd32STejun Heo 5515bce90380STejun Heo /* 5516bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5517bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5518bce90380STejun Heo * fully initialized by now. 5519bce90380STejun Heo */ 5520ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5521bce90380STejun Heo BUG_ON(!tbl); 5522bce90380STejun Heo 5523bce90380STejun Heo for_each_node(node) 55245a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 55251be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5526bce90380STejun Heo 5527bce90380STejun Heo for_each_possible_cpu(cpu) { 5528bce90380STejun Heo node = cpu_to_node(cpu); 5529bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5530bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5531bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5532bce90380STejun Heo return; 5533bce90380STejun Heo } 5534bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5535bce90380STejun Heo } 5536bce90380STejun Heo 5537bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5538bce90380STejun Heo wq_numa_enabled = true; 5539bce90380STejun Heo } 5540bce90380STejun Heo 55413347fa09STejun Heo /** 55423347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 55433347fa09STejun Heo * 55443347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 55453347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 55463347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 55473347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 55483347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 55493347fa09STejun Heo * created and scheduled right before early initcalls. 55503347fa09STejun Heo */ 55513347fa09STejun Heo int __init workqueue_init_early(void) 55521da177e4SLinus Torvalds { 55537a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 55547a4e344cSTejun Heo int i, cpu; 5555c34056a3STejun Heo 5556e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5557e904e6c2STejun Heo 5558b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5559c98a9805STal Shorer cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_FLAG_DOMAIN)); 5560b05a7928SFrederic Weisbecker 5561e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5562e904e6c2STejun Heo 5563706026c2STejun Heo /* initialize CPU pools */ 556429c91e99STejun Heo for_each_possible_cpu(cpu) { 55654ce62e9eSTejun Heo struct worker_pool *pool; 55668b03ae3cSTejun Heo 55677a4e344cSTejun Heo i = 0; 5568f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 55697a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5570ec22ca5eSTejun Heo pool->cpu = cpu; 55717a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 55727a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5573f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 55747a4e344cSTejun Heo 55759daf9e67STejun Heo /* alloc pool ID */ 557668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 55779daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 557868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55794ce62e9eSTejun Heo } 55808b03ae3cSTejun Heo } 55818b03ae3cSTejun Heo 55828a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 558329c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 558429c91e99STejun Heo struct workqueue_attrs *attrs; 558529c91e99STejun Heo 558629c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 558729c91e99STejun Heo attrs->nice = std_nice[i]; 558829c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 55898a2b7538STejun Heo 55908a2b7538STejun Heo /* 55918a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 55928a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 55938a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 55948a2b7538STejun Heo */ 55958a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 55968a2b7538STejun Heo attrs->nice = std_nice[i]; 55978a2b7538STejun Heo attrs->no_numa = true; 55988a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 559929c91e99STejun Heo } 560029c91e99STejun Heo 5601d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 56021aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5603d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5604f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5605f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 560624d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 560724d51addSTejun Heo WQ_FREEZABLE, 0); 56080668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 56090668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 56100668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 56110668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 56120668106cSViresh Kumar 0); 56131aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 56140668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 56150668106cSViresh Kumar !system_power_efficient_wq || 56160668106cSViresh Kumar !system_freezable_power_efficient_wq); 561782607adcSTejun Heo 56183347fa09STejun Heo return 0; 56193347fa09STejun Heo } 56203347fa09STejun Heo 56213347fa09STejun Heo /** 56223347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 56233347fa09STejun Heo * 56243347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 56253347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 56263347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 56273347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 56283347fa09STejun Heo * with the initial workers and enable future kworker creations. 56293347fa09STejun Heo */ 56303347fa09STejun Heo int __init workqueue_init(void) 56313347fa09STejun Heo { 56322186d9f9STejun Heo struct workqueue_struct *wq; 56333347fa09STejun Heo struct worker_pool *pool; 56343347fa09STejun Heo int cpu, bkt; 56353347fa09STejun Heo 56362186d9f9STejun Heo /* 56372186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 56382186d9f9STejun Heo * CPU to node mapping may not be available that early on some 56392186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 56402186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 56412186d9f9STejun Heo * affinity, fix them up. 56422186d9f9STejun Heo */ 56432186d9f9STejun Heo wq_numa_init(); 56442186d9f9STejun Heo 56452186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 56462186d9f9STejun Heo 56472186d9f9STejun Heo for_each_possible_cpu(cpu) { 56482186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 56492186d9f9STejun Heo pool->node = cpu_to_node(cpu); 56502186d9f9STejun Heo } 56512186d9f9STejun Heo } 56522186d9f9STejun Heo 56532186d9f9STejun Heo list_for_each_entry(wq, &workqueues, list) 56542186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 56552186d9f9STejun Heo 56562186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 56572186d9f9STejun Heo 56583347fa09STejun Heo /* create the initial workers */ 56593347fa09STejun Heo for_each_online_cpu(cpu) { 56603347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 56613347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 56623347fa09STejun Heo BUG_ON(!create_worker(pool)); 56633347fa09STejun Heo } 56643347fa09STejun Heo } 56653347fa09STejun Heo 56663347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 56673347fa09STejun Heo BUG_ON(!create_worker(pool)); 56683347fa09STejun Heo 56693347fa09STejun Heo wq_online = true; 567082607adcSTejun Heo wq_watchdog_init(); 567182607adcSTejun Heo 56726ee0578bSSuresh Siddha return 0; 56731da177e4SLinus Torvalds } 5674