11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 20b11895c4SLibin * normal work items and the other for high priority ones) and some extra 21b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 22b11895c4SLibin * number of these backing pools is dynamic. 23c54fce6eSTejun Heo * 24c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 251da177e4SLinus Torvalds */ 261da177e4SLinus Torvalds 279984de1aSPaul Gortmaker #include <linux/export.h> 281da177e4SLinus Torvalds #include <linux/kernel.h> 291da177e4SLinus Torvalds #include <linux/sched.h> 301da177e4SLinus Torvalds #include <linux/init.h> 311da177e4SLinus Torvalds #include <linux/signal.h> 321da177e4SLinus Torvalds #include <linux/completion.h> 331da177e4SLinus Torvalds #include <linux/workqueue.h> 341da177e4SLinus Torvalds #include <linux/slab.h> 351da177e4SLinus Torvalds #include <linux/cpu.h> 361da177e4SLinus Torvalds #include <linux/notifier.h> 371da177e4SLinus Torvalds #include <linux/kthread.h> 381fa44ecaSJames Bottomley #include <linux/hardirq.h> 3946934023SChristoph Lameter #include <linux/mempolicy.h> 40341a5958SRafael J. Wysocki #include <linux/freezer.h> 41d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51e22bee78STejun Heo 52ea138446STejun Heo #include "workqueue_internal.h" 531da177e4SLinus Torvalds 54c8e55f36STejun Heo enum { 55bc2ae0f5STejun Heo /* 5624647570STejun Heo * worker_pool flags 57bc2ae0f5STejun Heo * 5824647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 59bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 60bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 61bc2ae0f5STejun Heo * is in effect. 62bc2ae0f5STejun Heo * 63bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 64bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6524647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 66bc2ae0f5STejun Heo * 67bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 6892f9c5c4SLai Jiangshan * attach_mutex to avoid changing binding state while 694736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 70bc2ae0f5STejun Heo */ 7124647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 72db7bccf4STejun Heo 73c8e55f36STejun Heo /* worker flags */ 74c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 75c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 76e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 77fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 78f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 79a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 80e22bee78STejun Heo 81a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 82a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 83db7bccf4STejun Heo 84e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 854ce62e9eSTejun Heo 8629c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 87c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 88db7bccf4STejun Heo 89e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 90e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 91e22bee78STejun Heo 923233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 933233cdbdSTejun Heo /* call for help after 10ms 943233cdbdSTejun Heo (min two ticks) */ 95e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 96e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds /* 99e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1008698a745SDongsheng Yang * all cpus. Give MIN_NICE. 101e22bee78STejun Heo */ 1028698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1038698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 104ecf6881fSTejun Heo 105ecf6881fSTejun Heo WQ_NAME_LEN = 24, 106c8e55f36STejun Heo }; 107c8e55f36STejun Heo 1081da177e4SLinus Torvalds /* 1094690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1104690c4abSTejun Heo * 111e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 112e41e704bSTejun Heo * everyone else. 1134690c4abSTejun Heo * 114e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 115e22bee78STejun Heo * only be modified and accessed from the local cpu. 116e22bee78STejun Heo * 117d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1184690c4abSTejun Heo * 119d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 120d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 121d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 122d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 123e22bee78STejun Heo * 12492f9c5c4SLai Jiangshan * A: pool->attach_mutex protected. 125822d8405STejun Heo * 12668e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12776af4d93STejun Heo * 12868e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1295bcab335STejun Heo * 1305b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1315b95e1afSLai Jiangshan * 1325b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 1335b95e1afSLai Jiangshan * sched-RCU for reads. 1345b95e1afSLai Jiangshan * 1353c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1363c25a55dSLai Jiangshan * 137b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1382e109a28STejun Heo * 1392e109a28STejun Heo * MD: wq_mayday_lock protected. 1404690c4abSTejun Heo */ 1414690c4abSTejun Heo 1422eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 143c34056a3STejun Heo 144bd7bdd43STejun Heo struct worker_pool { 145d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 146d84ff051STejun Heo int cpu; /* I: the associated cpu */ 147f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1489daf9e67STejun Heo int id; /* I: pool ID */ 14911ebea50STejun Heo unsigned int flags; /* X: flags */ 150bd7bdd43STejun Heo 15182607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15282607adcSTejun Heo 153bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 154bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 155ea1abd61SLai Jiangshan 156ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 157bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 158bd7bdd43STejun Heo 159bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 160bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 161bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 162bd7bdd43STejun Heo 163c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 164c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 165c9e7cf27STejun Heo /* L: hash of busy workers */ 166c9e7cf27STejun Heo 167bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16834a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 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 293bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 294bce90380STejun Heo 2954c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2964c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2974c16bd32STejun Heo 29868e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2992e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 3005bcab335STejun Heo 301e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30268e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3037d19c5ceSTejun Heo 304ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 305ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 306ef557180SMike Galbraith 307ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 308ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 309b05a7928SFrederic Weisbecker 310f303fccbSTejun Heo /* 311f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 312f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 313f303fccbSTejun Heo * to uncover usages which depend on it. 314f303fccbSTejun Heo */ 315f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 316f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 317f303fccbSTejun Heo #else 318f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 319f303fccbSTejun Heo #endif 320f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 321f303fccbSTejun Heo 3227d19c5ceSTejun Heo /* the per-cpu worker pools */ 32325528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3247d19c5ceSTejun Heo 32568e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3267d19c5ceSTejun Heo 32768e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 32829c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 32929c91e99STejun Heo 330c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33129c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33229c91e99STejun Heo 3338a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3348a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3358a2b7538STejun Heo 336d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 337ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 338044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3391aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 340044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 341d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 342044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 343f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 344044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 34524d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3460668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3470668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3480668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3490668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 350d320c038STejun Heo 3517d19c5ceSTejun Heo static int worker_thread(void *__worker); 3526ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3537d19c5ceSTejun Heo 35497bd2347STejun Heo #define CREATE_TRACE_POINTS 35597bd2347STejun Heo #include <trace/events/workqueue.h> 35697bd2347STejun Heo 35768e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 358f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 359f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36068e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3615bcab335STejun Heo 362b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 363f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 364f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 365b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 36676af4d93STejun Heo 3675b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 368f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 369f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 370f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3715b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3725b95e1afSLai Jiangshan 373f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 374f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 375f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3767a62c2c8STejun Heo (pool)++) 3774ce62e9eSTejun Heo 37849e3cf44STejun Heo /** 37917116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38017116969STejun Heo * @pool: iteration cursor 381611c92a0STejun Heo * @pi: integer used for iteration 382fa1b54e6STejun Heo * 38368e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 38468e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38568e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 386fa1b54e6STejun Heo * 387fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 388fa1b54e6STejun Heo * ignored. 38917116969STejun Heo */ 390611c92a0STejun Heo #define for_each_pool(pool, pi) \ 391611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39268e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 393fa1b54e6STejun Heo else 39417116969STejun Heo 39517116969STejun Heo /** 396822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 397822d8405STejun Heo * @worker: iteration cursor 398822d8405STejun Heo * @pool: worker_pool to iterate workers of 399822d8405STejun Heo * 40092f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 401822d8405STejun Heo * 402822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 403822d8405STejun Heo * ignored. 404822d8405STejun Heo */ 405da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 406da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 40792f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 408822d8405STejun Heo else 409822d8405STejun Heo 410822d8405STejun Heo /** 41149e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41249e3cf44STejun Heo * @pwq: iteration cursor 41349e3cf44STejun Heo * @wq: the target workqueue 41476af4d93STejun Heo * 415b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 416794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 417794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 41876af4d93STejun Heo * 41976af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42076af4d93STejun Heo * ignored. 42149e3cf44STejun Heo */ 42249e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42376af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 424b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 42576af4d93STejun Heo else 426f3421797STejun Heo 427dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 428dc186ad7SThomas Gleixner 429dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 430dc186ad7SThomas Gleixner 43199777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43299777288SStanislaw Gruszka { 43399777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43499777288SStanislaw Gruszka } 43599777288SStanislaw Gruszka 436b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 437b9fdac7fSDu, Changbin { 438b9fdac7fSDu, Changbin struct work_struct *work = addr; 439b9fdac7fSDu, Changbin 440b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 441b9fdac7fSDu, Changbin } 442b9fdac7fSDu, Changbin 443dc186ad7SThomas Gleixner /* 444dc186ad7SThomas Gleixner * fixup_init is called when: 445dc186ad7SThomas Gleixner * - an active object is initialized 446dc186ad7SThomas Gleixner */ 44702a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 448dc186ad7SThomas Gleixner { 449dc186ad7SThomas Gleixner struct work_struct *work = addr; 450dc186ad7SThomas Gleixner 451dc186ad7SThomas Gleixner switch (state) { 452dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 453dc186ad7SThomas Gleixner cancel_work_sync(work); 454dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45502a982a6SDu, Changbin return true; 456dc186ad7SThomas Gleixner default: 45702a982a6SDu, Changbin return false; 458dc186ad7SThomas Gleixner } 459dc186ad7SThomas Gleixner } 460dc186ad7SThomas Gleixner 461dc186ad7SThomas Gleixner /* 462dc186ad7SThomas Gleixner * fixup_free is called when: 463dc186ad7SThomas Gleixner * - an active object is freed 464dc186ad7SThomas Gleixner */ 46502a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 466dc186ad7SThomas Gleixner { 467dc186ad7SThomas Gleixner struct work_struct *work = addr; 468dc186ad7SThomas Gleixner 469dc186ad7SThomas Gleixner switch (state) { 470dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 471dc186ad7SThomas Gleixner cancel_work_sync(work); 472dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47302a982a6SDu, Changbin return true; 474dc186ad7SThomas Gleixner default: 47502a982a6SDu, Changbin return false; 476dc186ad7SThomas Gleixner } 477dc186ad7SThomas Gleixner } 478dc186ad7SThomas Gleixner 479dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 480dc186ad7SThomas Gleixner .name = "work_struct", 48199777288SStanislaw Gruszka .debug_hint = work_debug_hint, 482b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 483dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 484dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 485dc186ad7SThomas Gleixner }; 486dc186ad7SThomas Gleixner 487dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 488dc186ad7SThomas Gleixner { 489dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 490dc186ad7SThomas Gleixner } 491dc186ad7SThomas Gleixner 492dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 493dc186ad7SThomas Gleixner { 494dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 495dc186ad7SThomas Gleixner } 496dc186ad7SThomas Gleixner 497dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 498dc186ad7SThomas Gleixner { 499dc186ad7SThomas Gleixner if (onstack) 500dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 501dc186ad7SThomas Gleixner else 502dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 503dc186ad7SThomas Gleixner } 504dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 505dc186ad7SThomas Gleixner 506dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 507dc186ad7SThomas Gleixner { 508dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 509dc186ad7SThomas Gleixner } 510dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 511dc186ad7SThomas Gleixner 512ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 513ea2e64f2SThomas Gleixner { 514ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 515ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 516ea2e64f2SThomas Gleixner } 517ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 518ea2e64f2SThomas Gleixner 519dc186ad7SThomas Gleixner #else 520dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 521dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 522dc186ad7SThomas Gleixner #endif 523dc186ad7SThomas Gleixner 5244e8b22bdSLi Bin /** 5254e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5264e8b22bdSLi Bin * @pool: the pool pointer of interest 5274e8b22bdSLi Bin * 5284e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5294e8b22bdSLi Bin * successfully, -errno on failure. 5304e8b22bdSLi Bin */ 5319daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5329daf9e67STejun Heo { 5339daf9e67STejun Heo int ret; 5349daf9e67STejun Heo 53568e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5365bcab335STejun Heo 5374e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5384e8b22bdSLi Bin GFP_KERNEL); 539229641a6STejun Heo if (ret >= 0) { 540e68035fbSTejun Heo pool->id = ret; 541229641a6STejun Heo return 0; 542229641a6STejun Heo } 5439daf9e67STejun Heo return ret; 5449daf9e67STejun Heo } 5459daf9e67STejun Heo 54676af4d93STejun Heo /** 547df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 548df2d5ae4STejun Heo * @wq: the target workqueue 549df2d5ae4STejun Heo * @node: the node ID 550df2d5ae4STejun Heo * 5515b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5525b95e1afSLai Jiangshan * read locked. 553df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 554df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 555d185af30SYacine Belkadi * 556d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 557df2d5ae4STejun Heo */ 558df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 559df2d5ae4STejun Heo int node) 560df2d5ae4STejun Heo { 5615b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 562d6e022f1STejun Heo 563d6e022f1STejun Heo /* 564d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 565d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 566d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 567d6e022f1STejun Heo * happens, this workaround can be removed. 568d6e022f1STejun Heo */ 569d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 570d6e022f1STejun Heo return wq->dfl_pwq; 571d6e022f1STejun Heo 572df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 573df2d5ae4STejun Heo } 574df2d5ae4STejun Heo 57573f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57673f53c4aSTejun Heo { 57773f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 57873f53c4aSTejun Heo } 57973f53c4aSTejun Heo 58073f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58173f53c4aSTejun Heo { 58273f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58373f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58473f53c4aSTejun Heo } 58573f53c4aSTejun Heo 58673f53c4aSTejun Heo static int work_next_color(int color) 58773f53c4aSTejun Heo { 58873f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 589a848e3b6SOleg Nesterov } 590a848e3b6SOleg Nesterov 5914594bf15SDavid Howells /* 592112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 593112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5947c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5957a22ad75STejun Heo * 596112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 597112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 598bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 599bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6007a22ad75STejun Heo * 601112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6027c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 603112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6047c3eed5cSTejun Heo * available only while the work item is queued. 605bbb68dfaSTejun Heo * 606bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 607bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 608bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 609bbb68dfaSTejun Heo * try to steal the PENDING bit. 6104594bf15SDavid Howells */ 6117a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6127a22ad75STejun Heo unsigned long flags) 6137a22ad75STejun Heo { 6146183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6157a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6167a22ad75STejun Heo } 6177a22ad75STejun Heo 618112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6194690c4abSTejun Heo unsigned long extra_flags) 620365970a1SDavid Howells { 621112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 622112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 623365970a1SDavid Howells } 624365970a1SDavid Howells 6254468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6264468a00fSLai Jiangshan int pool_id) 6274468a00fSLai Jiangshan { 6284468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6294468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6304468a00fSLai Jiangshan } 6314468a00fSLai Jiangshan 6327c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6337c3eed5cSTejun Heo int pool_id) 6344d707b9fSOleg Nesterov { 63523657bb1STejun Heo /* 63623657bb1STejun Heo * The following wmb is paired with the implied mb in 63723657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 63823657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 63923657bb1STejun Heo * owner. 64023657bb1STejun Heo */ 64123657bb1STejun Heo smp_wmb(); 6427c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 643346c09f8SRoman Pen /* 644346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 645346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 646346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 647346c09f8SRoman Pen * reordering can lead to a missed execution on attempt to qeueue 648346c09f8SRoman Pen * the same @work. E.g. consider this case: 649346c09f8SRoman Pen * 650346c09f8SRoman Pen * CPU#0 CPU#1 651346c09f8SRoman Pen * ---------------------------- -------------------------------- 652346c09f8SRoman Pen * 653346c09f8SRoman Pen * 1 STORE event_indicated 654346c09f8SRoman Pen * 2 queue_work_on() { 655346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 656346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 657346c09f8SRoman Pen * 5 set_work_data() # clear bit 658346c09f8SRoman Pen * 6 smp_mb() 659346c09f8SRoman Pen * 7 work->current_func() { 660346c09f8SRoman Pen * 8 LOAD event_indicated 661346c09f8SRoman Pen * } 662346c09f8SRoman Pen * 663346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 664346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 665346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 666346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 667346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 668346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 669346c09f8SRoman Pen * before actual STORE. 670346c09f8SRoman Pen */ 671346c09f8SRoman Pen smp_mb(); 6724d707b9fSOleg Nesterov } 6734d707b9fSOleg Nesterov 6747a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 675365970a1SDavid Howells { 6767c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6777c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6787a22ad75STejun Heo } 6797a22ad75STejun Heo 680112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6817a22ad75STejun Heo { 682e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6837a22ad75STejun Heo 684112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 685e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 686e120153dSTejun Heo else 687e120153dSTejun Heo return NULL; 6887a22ad75STejun Heo } 6897a22ad75STejun Heo 6907c3eed5cSTejun Heo /** 6917c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6927c3eed5cSTejun Heo * @work: the work item of interest 6937c3eed5cSTejun Heo * 69468e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69568e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 69668e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 697fa1b54e6STejun Heo * 698fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 699fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 700fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 701fa1b54e6STejun Heo * returned pool is and stays online. 702d185af30SYacine Belkadi * 703d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7047c3eed5cSTejun Heo */ 7057c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7067a22ad75STejun Heo { 707e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7087c3eed5cSTejun Heo int pool_id; 7097a22ad75STejun Heo 71068e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 711fa1b54e6STejun Heo 712112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 713112202d9STejun Heo return ((struct pool_workqueue *) 7147c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7157a22ad75STejun Heo 7167c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7177c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7187a22ad75STejun Heo return NULL; 7197a22ad75STejun Heo 720fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7217c3eed5cSTejun Heo } 7227c3eed5cSTejun Heo 7237c3eed5cSTejun Heo /** 7247c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7257c3eed5cSTejun Heo * @work: the work item of interest 7267c3eed5cSTejun Heo * 727d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7287c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7297c3eed5cSTejun Heo */ 7307c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7317c3eed5cSTejun Heo { 73254d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7337c3eed5cSTejun Heo 734112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 735112202d9STejun Heo return ((struct pool_workqueue *) 73654d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 73754d5b7d0SLai Jiangshan 73854d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7397c3eed5cSTejun Heo } 7407c3eed5cSTejun Heo 741bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 742bbb68dfaSTejun Heo { 7437c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 744bbb68dfaSTejun Heo 7457c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7467c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 747bbb68dfaSTejun Heo } 748bbb68dfaSTejun Heo 749bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 750bbb68dfaSTejun Heo { 751bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 752bbb68dfaSTejun Heo 753112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 754bbb68dfaSTejun Heo } 755bbb68dfaSTejun Heo 756e22bee78STejun Heo /* 7573270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7583270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 759d565ed63STejun Heo * they're being called with pool->lock held. 760e22bee78STejun Heo */ 761e22bee78STejun Heo 76263d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 763649027d7STejun Heo { 764e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 765649027d7STejun Heo } 766649027d7STejun Heo 767e22bee78STejun Heo /* 768e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 769e22bee78STejun Heo * running workers. 770974271c4STejun Heo * 771974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 772706026c2STejun Heo * function will always return %true for unbound pools as long as the 773974271c4STejun Heo * worklist isn't empty. 774e22bee78STejun Heo */ 77563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 776e22bee78STejun Heo { 77763d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 778e22bee78STejun Heo } 779e22bee78STejun Heo 780e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 782e22bee78STejun Heo { 78363d95a91STejun Heo return pool->nr_idle; 784e22bee78STejun Heo } 785e22bee78STejun Heo 786e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 78763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 788e22bee78STejun Heo { 789e19e397aSTejun Heo return !list_empty(&pool->worklist) && 790e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 791e22bee78STejun Heo } 792e22bee78STejun Heo 793e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79463d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 795e22bee78STejun Heo { 79663d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 797e22bee78STejun Heo } 798e22bee78STejun Heo 799e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80063d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 801e22bee78STejun Heo { 80234a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 80363d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80463d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 805e22bee78STejun Heo 806e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 807e22bee78STejun Heo } 808e22bee78STejun Heo 809e22bee78STejun Heo /* 810e22bee78STejun Heo * Wake up functions. 811e22bee78STejun Heo */ 812e22bee78STejun Heo 8131037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8141037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8157e11629dSTejun Heo { 81663d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8177e11629dSTejun Heo return NULL; 8187e11629dSTejun Heo 81963d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8207e11629dSTejun Heo } 8217e11629dSTejun Heo 8227e11629dSTejun Heo /** 8237e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82463d95a91STejun Heo * @pool: worker pool to wake worker from 8257e11629dSTejun Heo * 82663d95a91STejun Heo * Wake up the first idle worker of @pool. 8277e11629dSTejun Heo * 8287e11629dSTejun Heo * CONTEXT: 829d565ed63STejun Heo * spin_lock_irq(pool->lock). 8307e11629dSTejun Heo */ 83163d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8327e11629dSTejun Heo { 8331037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8347e11629dSTejun Heo 8357e11629dSTejun Heo if (likely(worker)) 8367e11629dSTejun Heo wake_up_process(worker->task); 8377e11629dSTejun Heo } 8387e11629dSTejun Heo 8394690c4abSTejun Heo /** 840e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 841e22bee78STejun Heo * @task: task waking up 842e22bee78STejun Heo * @cpu: CPU @task is waking up to 843e22bee78STejun Heo * 844e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 845e22bee78STejun Heo * being awoken. 846e22bee78STejun Heo * 847e22bee78STejun Heo * CONTEXT: 848e22bee78STejun Heo * spin_lock_irq(rq->lock) 849e22bee78STejun Heo */ 850d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 851e22bee78STejun Heo { 852e22bee78STejun Heo struct worker *worker = kthread_data(task); 853e22bee78STejun Heo 85436576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 855ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 856e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 857e22bee78STejun Heo } 85836576000SJoonsoo Kim } 859e22bee78STejun Heo 860e22bee78STejun Heo /** 861e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 862e22bee78STejun Heo * @task: task going to sleep 863e22bee78STejun Heo * 864e22bee78STejun Heo * This function is called during schedule() when a busy worker is 865e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 866e22bee78STejun Heo * returning pointer to its task. 867e22bee78STejun Heo * 868e22bee78STejun Heo * CONTEXT: 869e22bee78STejun Heo * spin_lock_irq(rq->lock) 870e22bee78STejun Heo * 871d185af30SYacine Belkadi * Return: 872e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 873e22bee78STejun Heo */ 8749b7f6597SAlexander Gordeev struct task_struct *wq_worker_sleeping(struct task_struct *task) 875e22bee78STejun Heo { 876e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 877111c225aSTejun Heo struct worker_pool *pool; 878e22bee78STejun Heo 879111c225aSTejun Heo /* 880111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 881111c225aSTejun Heo * workers, also reach here, let's not access anything before 882111c225aSTejun Heo * checking NOT_RUNNING. 883111c225aSTejun Heo */ 8842d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 885e22bee78STejun Heo return NULL; 886e22bee78STejun Heo 887111c225aSTejun Heo pool = worker->pool; 888111c225aSTejun Heo 889e22bee78STejun Heo /* this can only happen on the local cpu */ 8909b7f6597SAlexander Gordeev if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) 8916183c009STejun Heo return NULL; 892e22bee78STejun Heo 893e22bee78STejun Heo /* 894e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 895e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 896e22bee78STejun Heo * Please read comment there. 897e22bee78STejun Heo * 898628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 899628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 900628c78e7STejun Heo * disabled, which in turn means that none else could be 901d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 902628c78e7STejun Heo * lock is safe. 903e22bee78STejun Heo */ 904e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 905e19e397aSTejun Heo !list_empty(&pool->worklist)) 9061037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 907e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 908e22bee78STejun Heo } 909e22bee78STejun Heo 910e22bee78STejun Heo /** 911e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 912cb444766STejun Heo * @worker: self 913d302f017STejun Heo * @flags: flags to set 914d302f017STejun Heo * 915228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 916d302f017STejun Heo * 917cb444766STejun Heo * CONTEXT: 918d565ed63STejun Heo * spin_lock_irq(pool->lock) 919d302f017STejun Heo */ 920228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 921d302f017STejun Heo { 922bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 923e22bee78STejun Heo 924cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 925cb444766STejun Heo 926228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 927e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 928e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 929e19e397aSTejun Heo atomic_dec(&pool->nr_running); 930e22bee78STejun Heo } 931e22bee78STejun Heo 932d302f017STejun Heo worker->flags |= flags; 933d302f017STejun Heo } 934d302f017STejun Heo 935d302f017STejun Heo /** 936e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 937cb444766STejun Heo * @worker: self 938d302f017STejun Heo * @flags: flags to clear 939d302f017STejun Heo * 940e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 941d302f017STejun Heo * 942cb444766STejun Heo * CONTEXT: 943d565ed63STejun Heo * spin_lock_irq(pool->lock) 944d302f017STejun Heo */ 945d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 946d302f017STejun Heo { 94763d95a91STejun Heo struct worker_pool *pool = worker->pool; 948e22bee78STejun Heo unsigned int oflags = worker->flags; 949e22bee78STejun Heo 950cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 951cb444766STejun Heo 952d302f017STejun Heo worker->flags &= ~flags; 953e22bee78STejun Heo 95442c025f3STejun Heo /* 95542c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 95642c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 95742c025f3STejun Heo * of multiple flags, not a single flag. 95842c025f3STejun Heo */ 959e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 960e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 961e19e397aSTejun Heo atomic_inc(&pool->nr_running); 962d302f017STejun Heo } 963d302f017STejun Heo 964d302f017STejun Heo /** 9658cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 966c9e7cf27STejun Heo * @pool: pool of interest 9678cca0eeaSTejun Heo * @work: work to find worker for 9688cca0eeaSTejun Heo * 969c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 970c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 971a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 972a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 973a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 974a2c1c57bSTejun Heo * being executed. 975a2c1c57bSTejun Heo * 976a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 977a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 978a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 979a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 980a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 981a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 982a2c1c57bSTejun Heo * 983c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 984c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 985c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 986c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 987c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 988c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9898cca0eeaSTejun Heo * 9908cca0eeaSTejun Heo * CONTEXT: 991d565ed63STejun Heo * spin_lock_irq(pool->lock). 9928cca0eeaSTejun Heo * 993d185af30SYacine Belkadi * Return: 994d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9958cca0eeaSTejun Heo * otherwise. 9968cca0eeaSTejun Heo */ 997c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9988cca0eeaSTejun Heo struct work_struct *work) 9998cca0eeaSTejun Heo { 100042f8570fSSasha Levin struct worker *worker; 100142f8570fSSasha Levin 1002b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1003a2c1c57bSTejun Heo (unsigned long)work) 1004a2c1c57bSTejun Heo if (worker->current_work == work && 1005a2c1c57bSTejun Heo worker->current_func == work->func) 100642f8570fSSasha Levin return worker; 100742f8570fSSasha Levin 100842f8570fSSasha Levin return NULL; 10098cca0eeaSTejun Heo } 10108cca0eeaSTejun Heo 10118cca0eeaSTejun Heo /** 1012bf4ede01STejun Heo * move_linked_works - move linked works to a list 1013bf4ede01STejun Heo * @work: start of series of works to be scheduled 1014bf4ede01STejun Heo * @head: target list to append @work to 1015402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1016bf4ede01STejun Heo * 1017bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1018bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1019bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1020bf4ede01STejun Heo * 1021bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1022bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1023bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1024bf4ede01STejun Heo * 1025bf4ede01STejun Heo * CONTEXT: 1026d565ed63STejun Heo * spin_lock_irq(pool->lock). 1027bf4ede01STejun Heo */ 1028bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1029bf4ede01STejun Heo struct work_struct **nextp) 1030bf4ede01STejun Heo { 1031bf4ede01STejun Heo struct work_struct *n; 1032bf4ede01STejun Heo 1033bf4ede01STejun Heo /* 1034bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1035bf4ede01STejun Heo * use NULL for list head. 1036bf4ede01STejun Heo */ 1037bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1038bf4ede01STejun Heo list_move_tail(&work->entry, head); 1039bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1040bf4ede01STejun Heo break; 1041bf4ede01STejun Heo } 1042bf4ede01STejun Heo 1043bf4ede01STejun Heo /* 1044bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1045bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1046bf4ede01STejun Heo * needs to be updated. 1047bf4ede01STejun Heo */ 1048bf4ede01STejun Heo if (nextp) 1049bf4ede01STejun Heo *nextp = n; 1050bf4ede01STejun Heo } 1051bf4ede01STejun Heo 10528864b4e5STejun Heo /** 10538864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10548864b4e5STejun Heo * @pwq: pool_workqueue to get 10558864b4e5STejun Heo * 10568864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10578864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10588864b4e5STejun Heo */ 10598864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10608864b4e5STejun Heo { 10618864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10628864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10638864b4e5STejun Heo pwq->refcnt++; 10648864b4e5STejun Heo } 10658864b4e5STejun Heo 10668864b4e5STejun Heo /** 10678864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10688864b4e5STejun Heo * @pwq: pool_workqueue to put 10698864b4e5STejun Heo * 10708864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10718864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10728864b4e5STejun Heo */ 10738864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10748864b4e5STejun Heo { 10758864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10768864b4e5STejun Heo if (likely(--pwq->refcnt)) 10778864b4e5STejun Heo return; 10788864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10798864b4e5STejun Heo return; 10808864b4e5STejun Heo /* 10818864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10828864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10838864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10848864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10858864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10868864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10878864b4e5STejun Heo */ 10888864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10898864b4e5STejun Heo } 10908864b4e5STejun Heo 1091dce90d47STejun Heo /** 1092dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1093dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1094dce90d47STejun Heo * 1095dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1096dce90d47STejun Heo */ 1097dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1098dce90d47STejun Heo { 1099dce90d47STejun Heo if (pwq) { 1100dce90d47STejun Heo /* 1101dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1102dce90d47STejun Heo * following lock operations are safe. 1103dce90d47STejun Heo */ 1104dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1105dce90d47STejun Heo put_pwq(pwq); 1106dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1107dce90d47STejun Heo } 1108dce90d47STejun Heo } 1109dce90d47STejun Heo 1110112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1111bf4ede01STejun Heo { 1112112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1113bf4ede01STejun Heo 1114bf4ede01STejun Heo trace_workqueue_activate_work(work); 111582607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 111682607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1117112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1118bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1119112202d9STejun Heo pwq->nr_active++; 1120bf4ede01STejun Heo } 1121bf4ede01STejun Heo 1122112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11233aa62497SLai Jiangshan { 1124112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11253aa62497SLai Jiangshan struct work_struct, entry); 11263aa62497SLai Jiangshan 1127112202d9STejun Heo pwq_activate_delayed_work(work); 11283aa62497SLai Jiangshan } 11293aa62497SLai Jiangshan 1130bf4ede01STejun Heo /** 1131112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1132112202d9STejun Heo * @pwq: pwq of interest 1133bf4ede01STejun Heo * @color: color of work which left the queue 1134bf4ede01STejun Heo * 1135bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1136112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1137bf4ede01STejun Heo * 1138bf4ede01STejun Heo * CONTEXT: 1139d565ed63STejun Heo * spin_lock_irq(pool->lock). 1140bf4ede01STejun Heo */ 1141112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1142bf4ede01STejun Heo { 11438864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1144bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11458864b4e5STejun Heo goto out_put; 1146bf4ede01STejun Heo 1147112202d9STejun Heo pwq->nr_in_flight[color]--; 1148bf4ede01STejun Heo 1149112202d9STejun Heo pwq->nr_active--; 1150112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1151bf4ede01STejun Heo /* one down, submit a delayed one */ 1152112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1153112202d9STejun Heo pwq_activate_first_delayed(pwq); 1154bf4ede01STejun Heo } 1155bf4ede01STejun Heo 1156bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1157112202d9STejun Heo if (likely(pwq->flush_color != color)) 11588864b4e5STejun Heo goto out_put; 1159bf4ede01STejun Heo 1160bf4ede01STejun Heo /* are there still in-flight works? */ 1161112202d9STejun Heo if (pwq->nr_in_flight[color]) 11628864b4e5STejun Heo goto out_put; 1163bf4ede01STejun Heo 1164112202d9STejun Heo /* this pwq is done, clear flush_color */ 1165112202d9STejun Heo pwq->flush_color = -1; 1166bf4ede01STejun Heo 1167bf4ede01STejun Heo /* 1168112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1169bf4ede01STejun Heo * will handle the rest. 1170bf4ede01STejun Heo */ 1171112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1172112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11738864b4e5STejun Heo out_put: 11748864b4e5STejun Heo put_pwq(pwq); 1175bf4ede01STejun Heo } 1176bf4ede01STejun Heo 117736e227d2STejun Heo /** 1178bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 117936e227d2STejun Heo * @work: work item to steal 118036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1181bbb68dfaSTejun Heo * @flags: place to store irq state 118236e227d2STejun Heo * 118336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1184d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 118536e227d2STejun Heo * 1186d185af30SYacine Belkadi * Return: 118736e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 118836e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 118936e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1190bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1191bbb68dfaSTejun Heo * for arbitrarily long 119236e227d2STejun Heo * 1193d185af30SYacine Belkadi * Note: 1194bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1195e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1196e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1197e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1198bbb68dfaSTejun Heo * 1199bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1200bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1201bbb68dfaSTejun Heo * 1202e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1203bf4ede01STejun Heo */ 1204bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1205bbb68dfaSTejun Heo unsigned long *flags) 1206bf4ede01STejun Heo { 1207d565ed63STejun Heo struct worker_pool *pool; 1208112202d9STejun Heo struct pool_workqueue *pwq; 1209bf4ede01STejun Heo 1210bbb68dfaSTejun Heo local_irq_save(*flags); 1211bbb68dfaSTejun Heo 121236e227d2STejun Heo /* try to steal the timer if it exists */ 121336e227d2STejun Heo if (is_dwork) { 121436e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 121536e227d2STejun Heo 1216e0aecdd8STejun Heo /* 1217e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1218e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1219e0aecdd8STejun Heo * running on the local CPU. 1220e0aecdd8STejun Heo */ 122136e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 122236e227d2STejun Heo return 1; 122336e227d2STejun Heo } 122436e227d2STejun Heo 122536e227d2STejun Heo /* try to claim PENDING the normal way */ 1226bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1227bf4ede01STejun Heo return 0; 1228bf4ede01STejun Heo 1229bf4ede01STejun Heo /* 1230bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1231bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1232bf4ede01STejun Heo */ 1233d565ed63STejun Heo pool = get_work_pool(work); 1234d565ed63STejun Heo if (!pool) 1235bbb68dfaSTejun Heo goto fail; 1236bf4ede01STejun Heo 1237d565ed63STejun Heo spin_lock(&pool->lock); 1238bf4ede01STejun Heo /* 1239112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1240112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1241112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1242112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1243112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12440b3dae68SLai Jiangshan * item is currently queued on that pool. 1245bf4ede01STejun Heo */ 1246112202d9STejun Heo pwq = get_work_pwq(work); 1247112202d9STejun Heo if (pwq && pwq->pool == pool) { 1248bf4ede01STejun Heo debug_work_deactivate(work); 12493aa62497SLai Jiangshan 12503aa62497SLai Jiangshan /* 125116062836STejun Heo * A delayed work item cannot be grabbed directly because 125216062836STejun Heo * it might have linked NO_COLOR work items which, if left 1253112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 125416062836STejun Heo * management later on and cause stall. Make sure the work 125516062836STejun Heo * item is activated before grabbing. 12563aa62497SLai Jiangshan */ 12573aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1258112202d9STejun Heo pwq_activate_delayed_work(work); 12593aa62497SLai Jiangshan 1260bf4ede01STejun Heo list_del_init(&work->entry); 12619c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 126236e227d2STejun Heo 1263112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12644468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12654468a00fSLai Jiangshan 1266d565ed63STejun Heo spin_unlock(&pool->lock); 126736e227d2STejun Heo return 1; 1268bf4ede01STejun Heo } 1269d565ed63STejun Heo spin_unlock(&pool->lock); 1270bbb68dfaSTejun Heo fail: 1271bbb68dfaSTejun Heo local_irq_restore(*flags); 1272bbb68dfaSTejun Heo if (work_is_canceling(work)) 1273bbb68dfaSTejun Heo return -ENOENT; 1274bbb68dfaSTejun Heo cpu_relax(); 127536e227d2STejun Heo return -EAGAIN; 1276bf4ede01STejun Heo } 1277bf4ede01STejun Heo 1278bf4ede01STejun Heo /** 1279706026c2STejun Heo * insert_work - insert a work into a pool 1280112202d9STejun Heo * @pwq: pwq @work belongs to 12814690c4abSTejun Heo * @work: work to insert 12824690c4abSTejun Heo * @head: insertion point 12834690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12844690c4abSTejun Heo * 1285112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1286706026c2STejun Heo * work_struct flags. 12874690c4abSTejun Heo * 12884690c4abSTejun Heo * CONTEXT: 1289d565ed63STejun Heo * spin_lock_irq(pool->lock). 1290365970a1SDavid Howells */ 1291112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1292112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1293b89deed3SOleg Nesterov { 1294112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1295e1d8aa9fSFrederic Weisbecker 12964690c4abSTejun Heo /* we own @work, set data and link */ 1297112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12981a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12998864b4e5STejun Heo get_pwq(pwq); 1300e22bee78STejun Heo 1301e22bee78STejun Heo /* 1302c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1303c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1304c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1305e22bee78STejun Heo */ 1306e22bee78STejun Heo smp_mb(); 1307e22bee78STejun Heo 130863d95a91STejun Heo if (__need_more_worker(pool)) 130963d95a91STejun Heo wake_up_worker(pool); 1310b89deed3SOleg Nesterov } 1311b89deed3SOleg Nesterov 1312c8efcc25STejun Heo /* 1313c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13148d03ecfeSTejun Heo * same workqueue. 1315c8efcc25STejun Heo */ 1316c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1317c8efcc25STejun Heo { 1318c8efcc25STejun Heo struct worker *worker; 1319c8efcc25STejun Heo 13208d03ecfeSTejun Heo worker = current_wq_worker(); 1321c8efcc25STejun Heo /* 13228d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13238d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1324c8efcc25STejun Heo */ 1325112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1326c8efcc25STejun Heo } 1327c8efcc25STejun Heo 1328ef557180SMike Galbraith /* 1329ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1330ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1331ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1332ef557180SMike Galbraith */ 1333ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1334ef557180SMike Galbraith { 1335f303fccbSTejun Heo static bool printed_dbg_warning; 1336ef557180SMike Galbraith int new_cpu; 1337ef557180SMike Galbraith 1338f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1339ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1340ef557180SMike Galbraith return cpu; 1341f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1342f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1343f303fccbSTejun Heo printed_dbg_warning = true; 1344f303fccbSTejun Heo } 1345f303fccbSTejun Heo 1346ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1347ef557180SMike Galbraith return cpu; 1348ef557180SMike Galbraith 1349ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1350ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1351ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1352ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1353ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1354ef557180SMike Galbraith return cpu; 1355ef557180SMike Galbraith } 1356ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1357ef557180SMike Galbraith 1358ef557180SMike Galbraith return new_cpu; 1359ef557180SMike Galbraith } 1360ef557180SMike Galbraith 1361d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13621da177e4SLinus Torvalds struct work_struct *work) 13631da177e4SLinus Torvalds { 1364112202d9STejun Heo struct pool_workqueue *pwq; 1365c9178087STejun Heo struct worker_pool *last_pool; 13661e19ffc6STejun Heo struct list_head *worklist; 13678a2e8e5dSTejun Heo unsigned int work_flags; 1368b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13698930cabaSTejun Heo 13708930cabaSTejun Heo /* 13718930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13728930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13738930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13748930cabaSTejun Heo * happen with IRQ disabled. 13758930cabaSTejun Heo */ 13768930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13771da177e4SLinus Torvalds 1378dc186ad7SThomas Gleixner debug_work_activate(work); 13791e19ffc6STejun Heo 13809ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1381618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1382c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1383e41e704bSTejun Heo return; 13849e8cd2f5STejun Heo retry: 1385df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1386ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1387df2d5ae4STejun Heo 1388df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1389df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1390c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1391df2d5ae4STejun Heo else 1392df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1393f3421797STejun Heo 139418aa9effSTejun Heo /* 1395c9178087STejun Heo * If @work was previously on a different pool, it might still be 1396c9178087STejun Heo * running there, in which case the work needs to be queued on that 1397c9178087STejun Heo * pool to guarantee non-reentrancy. 139818aa9effSTejun Heo */ 1399c9e7cf27STejun Heo last_pool = get_work_pool(work); 1400112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 140118aa9effSTejun Heo struct worker *worker; 140218aa9effSTejun Heo 1403d565ed63STejun Heo spin_lock(&last_pool->lock); 140418aa9effSTejun Heo 1405c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 140618aa9effSTejun Heo 1407112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1408c9178087STejun Heo pwq = worker->current_pwq; 14098594fadeSLai Jiangshan } else { 141018aa9effSTejun Heo /* meh... not running there, queue here */ 1411d565ed63STejun Heo spin_unlock(&last_pool->lock); 1412112202d9STejun Heo spin_lock(&pwq->pool->lock); 141318aa9effSTejun Heo } 14148930cabaSTejun Heo } else { 1415112202d9STejun Heo spin_lock(&pwq->pool->lock); 14168930cabaSTejun Heo } 1417502ca9d8STejun Heo 14189e8cd2f5STejun Heo /* 14199e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14209e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14219e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1422df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1423df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14249e8cd2f5STejun Heo * make forward-progress. 14259e8cd2f5STejun Heo */ 14269e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14279e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 14289e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 14299e8cd2f5STejun Heo cpu_relax(); 14309e8cd2f5STejun Heo goto retry; 14319e8cd2f5STejun Heo } 14329e8cd2f5STejun Heo /* oops */ 14339e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14349e8cd2f5STejun Heo wq->name, cpu); 14359e8cd2f5STejun Heo } 14369e8cd2f5STejun Heo 1437112202d9STejun Heo /* pwq determined, queue */ 1438112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1439502ca9d8STejun Heo 1440f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1441112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1442f5b2552bSDan Carpenter return; 1443f5b2552bSDan Carpenter } 14441e19ffc6STejun Heo 1445112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1446112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14471e19ffc6STejun Heo 1448112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1449cdadf009STejun Heo trace_workqueue_activate_work(work); 1450112202d9STejun Heo pwq->nr_active++; 1451112202d9STejun Heo worklist = &pwq->pool->worklist; 145282607adcSTejun Heo if (list_empty(worklist)) 145382607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14548a2e8e5dSTejun Heo } else { 14558a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1456112202d9STejun Heo worklist = &pwq->delayed_works; 14578a2e8e5dSTejun Heo } 14581e19ffc6STejun Heo 1459112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14601e19ffc6STejun Heo 1461112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14621da177e4SLinus Torvalds } 14631da177e4SLinus Torvalds 14640fcb78c2SRolf Eike Beer /** 1465c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1466c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1467c1a220e7SZhang Rui * @wq: workqueue to use 1468c1a220e7SZhang Rui * @work: work to queue 1469c1a220e7SZhang Rui * 1470c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1471c1a220e7SZhang Rui * can't go away. 1472d185af30SYacine Belkadi * 1473d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1474c1a220e7SZhang Rui */ 1475d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1476d4283e93STejun Heo struct work_struct *work) 1477c1a220e7SZhang Rui { 1478d4283e93STejun Heo bool ret = false; 14798930cabaSTejun Heo unsigned long flags; 14808930cabaSTejun Heo 14818930cabaSTejun Heo local_irq_save(flags); 1482c1a220e7SZhang Rui 148322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14844690c4abSTejun Heo __queue_work(cpu, wq, work); 1485d4283e93STejun Heo ret = true; 1486c1a220e7SZhang Rui } 14878930cabaSTejun Heo 14888930cabaSTejun Heo local_irq_restore(flags); 1489c1a220e7SZhang Rui return ret; 1490c1a220e7SZhang Rui } 1491ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1492c1a220e7SZhang Rui 1493d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14941da177e4SLinus Torvalds { 149552bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14961da177e4SLinus Torvalds 1497e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 149860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14991da177e4SLinus Torvalds } 15001438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 15011da177e4SLinus Torvalds 15027beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 150352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15041da177e4SLinus Torvalds { 15057beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 15067beb2edfSTejun Heo struct work_struct *work = &dwork->work; 15071da177e4SLinus Torvalds 15087beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 15097beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1510fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1511fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 15127beb2edfSTejun Heo 15138852aac2STejun Heo /* 15148852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 15158852aac2STejun Heo * both optimization and correctness. The earliest @timer can 15168852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 15178852aac2STejun Heo * on that there's no such delay when @delay is 0. 15188852aac2STejun Heo */ 15198852aac2STejun Heo if (!delay) { 15208852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 15218852aac2STejun Heo return; 15228852aac2STejun Heo } 15238852aac2STejun Heo 15247beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 15257beb2edfSTejun Heo 152660c057bcSLai Jiangshan dwork->wq = wq; 15271265057fSTejun Heo dwork->cpu = cpu; 15287beb2edfSTejun Heo timer->expires = jiffies + delay; 15297beb2edfSTejun Heo 1530041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 15317beb2edfSTejun Heo add_timer_on(timer, cpu); 1532041bd12eSTejun Heo else 1533041bd12eSTejun Heo add_timer(timer); 15347beb2edfSTejun Heo } 15351da177e4SLinus Torvalds 15360fcb78c2SRolf Eike Beer /** 15370fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 15380fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 15390fcb78c2SRolf Eike Beer * @wq: workqueue to use 1540af9997e4SRandy Dunlap * @dwork: work to queue 15410fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 15420fcb78c2SRolf Eike Beer * 1543d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1544715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1545715f1300STejun Heo * execution. 15460fcb78c2SRolf Eike Beer */ 1547d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 154852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15497a6bc1cdSVenkatesh Pallipadi { 155052bad64dSDavid Howells struct work_struct *work = &dwork->work; 1551d4283e93STejun Heo bool ret = false; 15528930cabaSTejun Heo unsigned long flags; 15538930cabaSTejun Heo 15548930cabaSTejun Heo /* read the comment in __queue_work() */ 15558930cabaSTejun Heo local_irq_save(flags); 15567a6bc1cdSVenkatesh Pallipadi 155722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15587beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1559d4283e93STejun Heo ret = true; 15607a6bc1cdSVenkatesh Pallipadi } 15618930cabaSTejun Heo 15628930cabaSTejun Heo local_irq_restore(flags); 15637a6bc1cdSVenkatesh Pallipadi return ret; 15647a6bc1cdSVenkatesh Pallipadi } 1565ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15661da177e4SLinus Torvalds 1567c8e55f36STejun Heo /** 15688376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15698376fe22STejun Heo * @cpu: CPU number to execute work on 15708376fe22STejun Heo * @wq: workqueue to use 15718376fe22STejun Heo * @dwork: work to queue 15728376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15738376fe22STejun Heo * 15748376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15758376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15768376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15778376fe22STejun Heo * current state. 15788376fe22STejun Heo * 1579d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15808376fe22STejun Heo * pending and its timer was modified. 15818376fe22STejun Heo * 1582e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15838376fe22STejun Heo * See try_to_grab_pending() for details. 15848376fe22STejun Heo */ 15858376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15868376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15878376fe22STejun Heo { 15888376fe22STejun Heo unsigned long flags; 15898376fe22STejun Heo int ret; 15908376fe22STejun Heo 15918376fe22STejun Heo do { 15928376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15938376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15948376fe22STejun Heo 15958376fe22STejun Heo if (likely(ret >= 0)) { 15968376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15978376fe22STejun Heo local_irq_restore(flags); 15988376fe22STejun Heo } 15998376fe22STejun Heo 16008376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 16018376fe22STejun Heo return ret; 16028376fe22STejun Heo } 16038376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 16048376fe22STejun Heo 16058376fe22STejun Heo /** 1606c8e55f36STejun Heo * worker_enter_idle - enter idle state 1607c8e55f36STejun Heo * @worker: worker which is entering idle state 1608c8e55f36STejun Heo * 1609c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1610c8e55f36STejun Heo * necessary. 1611c8e55f36STejun Heo * 1612c8e55f36STejun Heo * LOCKING: 1613d565ed63STejun Heo * spin_lock_irq(pool->lock). 1614c8e55f36STejun Heo */ 1615c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 16161da177e4SLinus Torvalds { 1617bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1618c8e55f36STejun Heo 16196183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 16206183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 16216183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 16226183c009STejun Heo return; 1623c8e55f36STejun Heo 1624051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1625cb444766STejun Heo worker->flags |= WORKER_IDLE; 1626bd7bdd43STejun Heo pool->nr_idle++; 1627e22bee78STejun Heo worker->last_active = jiffies; 1628c8e55f36STejun Heo 1629c8e55f36STejun Heo /* idle_list is LIFO */ 1630bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1631db7bccf4STejun Heo 163263d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1633628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1634cb444766STejun Heo 1635544ecf31STejun Heo /* 1636706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1637d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1638628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1639628c78e7STejun Heo * unbind is not in progress. 1640544ecf31STejun Heo */ 164124647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1642bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1643e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1644c8e55f36STejun Heo } 1645c8e55f36STejun Heo 1646c8e55f36STejun Heo /** 1647c8e55f36STejun Heo * worker_leave_idle - leave idle state 1648c8e55f36STejun Heo * @worker: worker which is leaving idle state 1649c8e55f36STejun Heo * 1650c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1651c8e55f36STejun Heo * 1652c8e55f36STejun Heo * LOCKING: 1653d565ed63STejun Heo * spin_lock_irq(pool->lock). 1654c8e55f36STejun Heo */ 1655c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1656c8e55f36STejun Heo { 1657bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1658c8e55f36STejun Heo 16596183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16606183c009STejun Heo return; 1661d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1662bd7bdd43STejun Heo pool->nr_idle--; 1663c8e55f36STejun Heo list_del_init(&worker->entry); 1664c8e55f36STejun Heo } 1665c8e55f36STejun Heo 1666f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1667c34056a3STejun Heo { 1668c34056a3STejun Heo struct worker *worker; 1669c34056a3STejun Heo 1670f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1671c8e55f36STejun Heo if (worker) { 1672c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1673affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1674da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1675e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1676e22bee78STejun Heo worker->flags = WORKER_PREP; 1677c8e55f36STejun Heo } 1678c34056a3STejun Heo return worker; 1679c34056a3STejun Heo } 1680c34056a3STejun Heo 1681c34056a3STejun Heo /** 16824736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16834736cbf7SLai Jiangshan * @worker: worker to be attached 16844736cbf7SLai Jiangshan * @pool: the target pool 16854736cbf7SLai Jiangshan * 16864736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16874736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16884736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16894736cbf7SLai Jiangshan */ 16904736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16914736cbf7SLai Jiangshan struct worker_pool *pool) 16924736cbf7SLai Jiangshan { 16934736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16944736cbf7SLai Jiangshan 16954736cbf7SLai Jiangshan /* 16964736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16974736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16984736cbf7SLai Jiangshan */ 16994736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17004736cbf7SLai Jiangshan 17014736cbf7SLai Jiangshan /* 17024736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 17034736cbf7SLai Jiangshan * stable across this function. See the comments above the 17044736cbf7SLai Jiangshan * flag definition for details. 17054736cbf7SLai Jiangshan */ 17064736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 17074736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 17084736cbf7SLai Jiangshan 17094736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 17104736cbf7SLai Jiangshan 17114736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 17124736cbf7SLai Jiangshan } 17134736cbf7SLai Jiangshan 17144736cbf7SLai Jiangshan /** 171560f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 171660f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 171760f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 171860f5a4bcSLai Jiangshan * 17194736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 17204736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 17214736cbf7SLai Jiangshan * other reference to the pool. 172260f5a4bcSLai Jiangshan */ 172360f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 172460f5a4bcSLai Jiangshan struct worker_pool *pool) 172560f5a4bcSLai Jiangshan { 172660f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 172760f5a4bcSLai Jiangshan 172892f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1729da028469SLai Jiangshan list_del(&worker->node); 1730da028469SLai Jiangshan if (list_empty(&pool->workers)) 173160f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 173292f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 173360f5a4bcSLai Jiangshan 1734b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1735b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1736b62c0751SLai Jiangshan 173760f5a4bcSLai Jiangshan if (detach_completion) 173860f5a4bcSLai Jiangshan complete(detach_completion); 173960f5a4bcSLai Jiangshan } 174060f5a4bcSLai Jiangshan 174160f5a4bcSLai Jiangshan /** 1742c34056a3STejun Heo * create_worker - create a new workqueue worker 174363d95a91STejun Heo * @pool: pool the new worker will belong to 1744c34056a3STejun Heo * 1745051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1746c34056a3STejun Heo * 1747c34056a3STejun Heo * CONTEXT: 1748c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1749c34056a3STejun Heo * 1750d185af30SYacine Belkadi * Return: 1751c34056a3STejun Heo * Pointer to the newly created worker. 1752c34056a3STejun Heo */ 1753bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1754c34056a3STejun Heo { 1755c34056a3STejun Heo struct worker *worker = NULL; 1756f3421797STejun Heo int id = -1; 1757e3c916a4STejun Heo char id_buf[16]; 1758c34056a3STejun Heo 17597cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17607cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1761822d8405STejun Heo if (id < 0) 1762c34056a3STejun Heo goto fail; 1763c34056a3STejun Heo 1764f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1765c34056a3STejun Heo if (!worker) 1766c34056a3STejun Heo goto fail; 1767c34056a3STejun Heo 1768bd7bdd43STejun Heo worker->pool = pool; 1769c34056a3STejun Heo worker->id = id; 1770c34056a3STejun Heo 177129c91e99STejun Heo if (pool->cpu >= 0) 1772e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1773e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1774f3421797STejun Heo else 1775e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1776e3c916a4STejun Heo 1777f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1778e3c916a4STejun Heo "kworker/%s", id_buf); 1779c34056a3STejun Heo if (IS_ERR(worker->task)) 1780c34056a3STejun Heo goto fail; 1781c34056a3STejun Heo 178291151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 178325834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 178491151228SOleg Nesterov 1785da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17864736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1787822d8405STejun Heo 1788051e1850SLai Jiangshan /* start the newly created worker */ 1789051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1790051e1850SLai Jiangshan worker->pool->nr_workers++; 1791051e1850SLai Jiangshan worker_enter_idle(worker); 1792051e1850SLai Jiangshan wake_up_process(worker->task); 1793051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1794051e1850SLai Jiangshan 1795c34056a3STejun Heo return worker; 1796822d8405STejun Heo 1797c34056a3STejun Heo fail: 17989625ab17SLai Jiangshan if (id >= 0) 17997cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1800c34056a3STejun Heo kfree(worker); 1801c34056a3STejun Heo return NULL; 1802c34056a3STejun Heo } 1803c34056a3STejun Heo 1804c34056a3STejun Heo /** 1805c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1806c34056a3STejun Heo * @worker: worker to be destroyed 1807c34056a3STejun Heo * 180873eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 180973eb7fe7SLai Jiangshan * be idle. 1810c8e55f36STejun Heo * 1811c8e55f36STejun Heo * CONTEXT: 181260f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1813c34056a3STejun Heo */ 1814c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1815c34056a3STejun Heo { 1816bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1817c34056a3STejun Heo 1818cd549687STejun Heo lockdep_assert_held(&pool->lock); 1819cd549687STejun Heo 1820c34056a3STejun Heo /* sanity check frenzy */ 18216183c009STejun Heo if (WARN_ON(worker->current_work) || 182273eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 182373eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 18246183c009STejun Heo return; 1825c34056a3STejun Heo 1826bd7bdd43STejun Heo pool->nr_workers--; 1827bd7bdd43STejun Heo pool->nr_idle--; 1828c8e55f36STejun Heo 1829c8e55f36STejun Heo list_del_init(&worker->entry); 1830cb444766STejun Heo worker->flags |= WORKER_DIE; 183160f5a4bcSLai Jiangshan wake_up_process(worker->task); 1832c34056a3STejun Heo } 1833c34056a3STejun Heo 183463d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1835e22bee78STejun Heo { 183663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1837e22bee78STejun Heo 1838d565ed63STejun Heo spin_lock_irq(&pool->lock); 1839e22bee78STejun Heo 18403347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1841e22bee78STejun Heo struct worker *worker; 1842e22bee78STejun Heo unsigned long expires; 1843e22bee78STejun Heo 1844e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 184563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1846e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1847e22bee78STejun Heo 18483347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 184963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18503347fc9fSLai Jiangshan break; 1851e22bee78STejun Heo } 18523347fc9fSLai Jiangshan 18533347fc9fSLai Jiangshan destroy_worker(worker); 1854e22bee78STejun Heo } 1855e22bee78STejun Heo 1856d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1857e22bee78STejun Heo } 1858e22bee78STejun Heo 1859493a1724STejun Heo static void send_mayday(struct work_struct *work) 1860e22bee78STejun Heo { 1861112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1862112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1863493a1724STejun Heo 18642e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1865e22bee78STejun Heo 1866493008a8STejun Heo if (!wq->rescuer) 1867493a1724STejun Heo return; 1868e22bee78STejun Heo 1869e22bee78STejun Heo /* mayday mayday mayday */ 1870493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 187177668c8bSLai Jiangshan /* 187277668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 187377668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 187477668c8bSLai Jiangshan * rescuer is done with it. 187577668c8bSLai Jiangshan */ 187677668c8bSLai Jiangshan get_pwq(pwq); 1877493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1878e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1879493a1724STejun Heo } 1880e22bee78STejun Heo } 1881e22bee78STejun Heo 1882706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1883e22bee78STejun Heo { 188463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1885e22bee78STejun Heo struct work_struct *work; 1886e22bee78STejun Heo 1887b2d82909STejun Heo spin_lock_irq(&pool->lock); 1888b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1889e22bee78STejun Heo 189063d95a91STejun Heo if (need_to_create_worker(pool)) { 1891e22bee78STejun Heo /* 1892e22bee78STejun Heo * We've been trying to create a new worker but 1893e22bee78STejun Heo * haven't been successful. We might be hitting an 1894e22bee78STejun Heo * allocation deadlock. Send distress signals to 1895e22bee78STejun Heo * rescuers. 1896e22bee78STejun Heo */ 189763d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1898e22bee78STejun Heo send_mayday(work); 1899e22bee78STejun Heo } 1900e22bee78STejun Heo 1901b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1902b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1903e22bee78STejun Heo 190463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1905e22bee78STejun Heo } 1906e22bee78STejun Heo 1907e22bee78STejun Heo /** 1908e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 190963d95a91STejun Heo * @pool: pool to create a new worker for 1910e22bee78STejun Heo * 191163d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1912e22bee78STejun Heo * have at least one idle worker on return from this function. If 1913e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 191463d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1915e22bee78STejun Heo * possible allocation deadlock. 1916e22bee78STejun Heo * 1917c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1918c5aa87bbSTejun Heo * may_start_working() %true. 1919e22bee78STejun Heo * 1920e22bee78STejun Heo * LOCKING: 1921d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1922e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1923e22bee78STejun Heo * manager. 1924e22bee78STejun Heo */ 192529187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1926d565ed63STejun Heo __releases(&pool->lock) 1927d565ed63STejun Heo __acquires(&pool->lock) 1928e22bee78STejun Heo { 1929e22bee78STejun Heo restart: 1930d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19319f9c2364STejun Heo 1932e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 193363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1934e22bee78STejun Heo 1935e22bee78STejun Heo while (true) { 1936051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1937e22bee78STejun Heo break; 1938e22bee78STejun Heo 1939e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19409f9c2364STejun Heo 194163d95a91STejun Heo if (!need_to_create_worker(pool)) 1942e22bee78STejun Heo break; 1943e22bee78STejun Heo } 1944e22bee78STejun Heo 194563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1946d565ed63STejun Heo spin_lock_irq(&pool->lock); 1947051e1850SLai Jiangshan /* 1948051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1949051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1950051e1850SLai Jiangshan * already become busy. 1951051e1850SLai Jiangshan */ 195263d95a91STejun Heo if (need_to_create_worker(pool)) 1953e22bee78STejun Heo goto restart; 1954e22bee78STejun Heo } 1955e22bee78STejun Heo 1956e22bee78STejun Heo /** 1957e22bee78STejun Heo * manage_workers - manage worker pool 1958e22bee78STejun Heo * @worker: self 1959e22bee78STejun Heo * 1960706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1961e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1962706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1963e22bee78STejun Heo * 1964e22bee78STejun Heo * The caller can safely start processing works on false return. On 1965e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1966e22bee78STejun Heo * and may_start_working() is true. 1967e22bee78STejun Heo * 1968e22bee78STejun Heo * CONTEXT: 1969d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1970e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1971e22bee78STejun Heo * 1972d185af30SYacine Belkadi * Return: 197329187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 197429187a9eSTejun Heo * start processing works, %true if management function was performed and 197529187a9eSTejun Heo * the conditions that the caller verified before calling the function may 197629187a9eSTejun Heo * no longer be true. 1977e22bee78STejun Heo */ 1978e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1979e22bee78STejun Heo { 198063d95a91STejun Heo struct worker_pool *pool = worker->pool; 1981e22bee78STejun Heo 1982bc3a1afcSTejun Heo /* 1983bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1984bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1985bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1986bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1987bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1988bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1989bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1990bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1991bc3a1afcSTejun Heo */ 199234a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 199329187a9eSTejun Heo return false; 19942607d7a6STejun Heo pool->manager = worker; 1995e22bee78STejun Heo 199629187a9eSTejun Heo maybe_create_worker(pool); 1997e22bee78STejun Heo 19982607d7a6STejun Heo pool->manager = NULL; 199934a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 200029187a9eSTejun Heo return true; 2001e22bee78STejun Heo } 2002e22bee78STejun Heo 2003a62428c0STejun Heo /** 2004a62428c0STejun Heo * process_one_work - process single work 2005c34056a3STejun Heo * @worker: self 2006a62428c0STejun Heo * @work: work to process 2007a62428c0STejun Heo * 2008a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2009a62428c0STejun Heo * process a single work including synchronization against and 2010a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2011a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2012a62428c0STejun Heo * call this function to process a work. 2013a62428c0STejun Heo * 2014a62428c0STejun Heo * CONTEXT: 2015d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2016a62428c0STejun Heo */ 2017c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2018d565ed63STejun Heo __releases(&pool->lock) 2019d565ed63STejun Heo __acquires(&pool->lock) 20201da177e4SLinus Torvalds { 2021112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2022bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2023112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 202473f53c4aSTejun Heo int work_color; 20257e11629dSTejun Heo struct worker *collision; 20264e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20274e6045f1SJohannes Berg /* 2028a62428c0STejun Heo * It is permissible to free the struct work_struct from 2029a62428c0STejun Heo * inside the function that is called from it, this we need to 2030a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2031a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2032a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20334e6045f1SJohannes Berg */ 20344d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20354d82a1deSPeter Zijlstra 20364d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20374e6045f1SJohannes Berg #endif 2038807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 203985327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2040ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 204125511a47STejun Heo 20427e11629dSTejun Heo /* 20437e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20447e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20457e11629dSTejun Heo * already processing the work. If so, defer the work to the 20467e11629dSTejun Heo * currently executing one. 20477e11629dSTejun Heo */ 2048c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20497e11629dSTejun Heo if (unlikely(collision)) { 20507e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20517e11629dSTejun Heo return; 20527e11629dSTejun Heo } 20531da177e4SLinus Torvalds 20548930cabaSTejun Heo /* claim and dequeue */ 20551da177e4SLinus Torvalds debug_work_deactivate(work); 2056c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2057c34056a3STejun Heo worker->current_work = work; 2058a2c1c57bSTejun Heo worker->current_func = work->func; 2059112202d9STejun Heo worker->current_pwq = pwq; 206073f53c4aSTejun Heo work_color = get_work_color(work); 20617a22ad75STejun Heo 2062a62428c0STejun Heo list_del_init(&work->entry); 2063a62428c0STejun Heo 2064649027d7STejun Heo /* 2065228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2066228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2067228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2068228f1d00SLai Jiangshan * execution of the pending work items. 2069fb0e7bebSTejun Heo */ 2070fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2071228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2072fb0e7bebSTejun Heo 2073974271c4STejun Heo /* 2074a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2075a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2076a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2077a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2078228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2079974271c4STejun Heo */ 2080a489a03eSLai Jiangshan if (need_more_worker(pool)) 208163d95a91STejun Heo wake_up_worker(pool); 2082974271c4STejun Heo 20838930cabaSTejun Heo /* 20847c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2085d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 208623657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 208723657bb1STejun Heo * disabled. 20888930cabaSTejun Heo */ 20897c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20901da177e4SLinus Torvalds 2091d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2092365970a1SDavid Howells 2093112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20943295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2095e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2096a2c1c57bSTejun Heo worker->current_func(work); 2097e36c886aSArjan van de Ven /* 2098e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2099e36c886aSArjan van de Ven * point will only record its address. 2100e36c886aSArjan van de Ven */ 2101e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21023295f0efSIngo Molnar lock_map_release(&lockdep_map); 2103112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21041da177e4SLinus Torvalds 2105d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2106044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2107044c782cSValentin Ilie " last function: %pf\n", 2108a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2109a2c1c57bSTejun Heo worker->current_func); 2110d5abe669SPeter Zijlstra debug_show_held_locks(current); 2111d5abe669SPeter Zijlstra dump_stack(); 2112d5abe669SPeter Zijlstra } 2113d5abe669SPeter Zijlstra 2114b22ce278STejun Heo /* 2115b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2116b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2117b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2118b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2119789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2120789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2121b22ce278STejun Heo */ 21223e28e377SJoe Lawrence cond_resched_rcu_qs(); 2123b22ce278STejun Heo 2124d565ed63STejun Heo spin_lock_irq(&pool->lock); 2125a62428c0STejun Heo 2126fb0e7bebSTejun Heo /* clear cpu intensive status */ 2127fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2128fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2129fb0e7bebSTejun Heo 2130a62428c0STejun Heo /* we're done with it, release */ 213142f8570fSSasha Levin hash_del(&worker->hentry); 2132c34056a3STejun Heo worker->current_work = NULL; 2133a2c1c57bSTejun Heo worker->current_func = NULL; 2134112202d9STejun Heo worker->current_pwq = NULL; 21353d1cb205STejun Heo worker->desc_valid = false; 2136112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21371da177e4SLinus Torvalds } 21381da177e4SLinus Torvalds 2139affee4b2STejun Heo /** 2140affee4b2STejun Heo * process_scheduled_works - process scheduled works 2141affee4b2STejun Heo * @worker: self 2142affee4b2STejun Heo * 2143affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2144affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2145affee4b2STejun Heo * fetches a work from the top and executes it. 2146affee4b2STejun Heo * 2147affee4b2STejun Heo * CONTEXT: 2148d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2149affee4b2STejun Heo * multiple times. 2150affee4b2STejun Heo */ 2151affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21521da177e4SLinus Torvalds { 2153affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2154affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2155a62428c0STejun Heo struct work_struct, entry); 2156c34056a3STejun Heo process_one_work(worker, work); 2157a62428c0STejun Heo } 21581da177e4SLinus Torvalds } 21591da177e4SLinus Torvalds 21604690c4abSTejun Heo /** 21614690c4abSTejun Heo * worker_thread - the worker thread function 2162c34056a3STejun Heo * @__worker: self 21634690c4abSTejun Heo * 2164c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2165c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2166c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2167c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2168c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2169d185af30SYacine Belkadi * 2170d185af30SYacine Belkadi * Return: 0 21714690c4abSTejun Heo */ 2172c34056a3STejun Heo static int worker_thread(void *__worker) 21731da177e4SLinus Torvalds { 2174c34056a3STejun Heo struct worker *worker = __worker; 2175bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21761da177e4SLinus Torvalds 2177e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2178e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2179c8e55f36STejun Heo woke_up: 2180d565ed63STejun Heo spin_lock_irq(&pool->lock); 2181affee4b2STejun Heo 2182a9ab775bSTejun Heo /* am I supposed to die? */ 2183a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2184d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2185a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2186e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 218760f5a4bcSLai Jiangshan 218860f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21897cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 219060f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 219160f5a4bcSLai Jiangshan kfree(worker); 2192c8e55f36STejun Heo return 0; 2193c8e55f36STejun Heo } 2194c8e55f36STejun Heo 2195c8e55f36STejun Heo worker_leave_idle(worker); 2196db7bccf4STejun Heo recheck: 2197e22bee78STejun Heo /* no more worker necessary? */ 219863d95a91STejun Heo if (!need_more_worker(pool)) 2199e22bee78STejun Heo goto sleep; 2200e22bee78STejun Heo 2201e22bee78STejun Heo /* do we need to manage? */ 220263d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2203e22bee78STejun Heo goto recheck; 2204e22bee78STejun Heo 2205c8e55f36STejun Heo /* 2206c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2207c8e55f36STejun Heo * preparing to process a work or actually processing it. 2208c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2209c8e55f36STejun Heo */ 22106183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2211c8e55f36STejun Heo 2212e22bee78STejun Heo /* 2213a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2214a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2215a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2216a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2217a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2218e22bee78STejun Heo */ 2219a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2220e22bee78STejun Heo 2221e22bee78STejun Heo do { 2222affee4b2STejun Heo struct work_struct *work = 2223bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2224affee4b2STejun Heo struct work_struct, entry); 2225affee4b2STejun Heo 222682607adcSTejun Heo pool->watchdog_ts = jiffies; 222782607adcSTejun Heo 2228c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2229affee4b2STejun Heo /* optimization path, not strictly necessary */ 2230affee4b2STejun Heo process_one_work(worker, work); 2231affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2232affee4b2STejun Heo process_scheduled_works(worker); 2233affee4b2STejun Heo } else { 2234c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2235affee4b2STejun Heo process_scheduled_works(worker); 2236affee4b2STejun Heo } 223763d95a91STejun Heo } while (keep_working(pool)); 2238affee4b2STejun Heo 2239228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2240d313dd85STejun Heo sleep: 2241c8e55f36STejun Heo /* 2242d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2243d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2244d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2245d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2246d565ed63STejun Heo * event. 2247c8e55f36STejun Heo */ 2248c8e55f36STejun Heo worker_enter_idle(worker); 2249c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2250d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22511da177e4SLinus Torvalds schedule(); 2252c8e55f36STejun Heo goto woke_up; 22531da177e4SLinus Torvalds } 22541da177e4SLinus Torvalds 2255e22bee78STejun Heo /** 2256e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2257111c225aSTejun Heo * @__rescuer: self 2258e22bee78STejun Heo * 2259e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2260493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2261e22bee78STejun Heo * 2262706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2263e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2264e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2265e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2266e22bee78STejun Heo * the problem rescuer solves. 2267e22bee78STejun Heo * 2268706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2269706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2270e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2271e22bee78STejun Heo * 2272e22bee78STejun Heo * This should happen rarely. 2273d185af30SYacine Belkadi * 2274d185af30SYacine Belkadi * Return: 0 2275e22bee78STejun Heo */ 2276111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2277e22bee78STejun Heo { 2278111c225aSTejun Heo struct worker *rescuer = __rescuer; 2279111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2280e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22814d595b86SLai Jiangshan bool should_stop; 2282e22bee78STejun Heo 2283e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2284111c225aSTejun Heo 2285111c225aSTejun Heo /* 2286111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2287111c225aSTejun Heo * doesn't participate in concurrency management. 2288111c225aSTejun Heo */ 2289111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2290e22bee78STejun Heo repeat: 2291e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22921da177e4SLinus Torvalds 22934d595b86SLai Jiangshan /* 22944d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22954d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22964d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22974d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22984d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22994d595b86SLai Jiangshan * list is always empty on exit. 23004d595b86SLai Jiangshan */ 23014d595b86SLai Jiangshan should_stop = kthread_should_stop(); 23021da177e4SLinus Torvalds 2303493a1724STejun Heo /* see whether any pwq is asking for help */ 23042e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2305493a1724STejun Heo 2306493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2307493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2308493a1724STejun Heo struct pool_workqueue, mayday_node); 2309112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2310e22bee78STejun Heo struct work_struct *work, *n; 231182607adcSTejun Heo bool first = true; 2312e22bee78STejun Heo 2313e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2314493a1724STejun Heo list_del_init(&pwq->mayday_node); 2315493a1724STejun Heo 23162e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2317e22bee78STejun Heo 231851697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 231951697d39SLai Jiangshan 232051697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2321b3104104SLai Jiangshan rescuer->pool = pool; 2322e22bee78STejun Heo 2323e22bee78STejun Heo /* 2324e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2325e22bee78STejun Heo * process'em. 2326e22bee78STejun Heo */ 23270479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 232882607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 232982607adcSTejun Heo if (get_work_pwq(work) == pwq) { 233082607adcSTejun Heo if (first) 233182607adcSTejun Heo pool->watchdog_ts = jiffies; 2332e22bee78STejun Heo move_linked_works(work, scheduled, &n); 233382607adcSTejun Heo } 233482607adcSTejun Heo first = false; 233582607adcSTejun Heo } 2336e22bee78STejun Heo 2337008847f6SNeilBrown if (!list_empty(scheduled)) { 2338e22bee78STejun Heo process_scheduled_works(rescuer); 23397576958aSTejun Heo 23407576958aSTejun Heo /* 2341008847f6SNeilBrown * The above execution of rescued work items could 2342008847f6SNeilBrown * have created more to rescue through 2343008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2344008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2345008847f6SNeilBrown * that such back-to-back work items, which may be 2346008847f6SNeilBrown * being used to relieve memory pressure, don't 2347008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2348008847f6SNeilBrown */ 2349008847f6SNeilBrown if (need_to_create_worker(pool)) { 2350008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2351008847f6SNeilBrown get_pwq(pwq); 2352008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2353008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2354008847f6SNeilBrown } 2355008847f6SNeilBrown } 2356008847f6SNeilBrown 2357008847f6SNeilBrown /* 235877668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 235913b1d625SLai Jiangshan * go away while we're still attached to it. 236077668c8bSLai Jiangshan */ 236177668c8bSLai Jiangshan put_pwq(pwq); 236277668c8bSLai Jiangshan 236377668c8bSLai Jiangshan /* 2364d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 23657576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23667576958aSTejun Heo * and stalling the execution. 23677576958aSTejun Heo */ 2368d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 236963d95a91STejun Heo wake_up_worker(pool); 23707576958aSTejun Heo 2371b3104104SLai Jiangshan rescuer->pool = NULL; 237213b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 237313b1d625SLai Jiangshan 237413b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 237513b1d625SLai Jiangshan 237613b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 23771da177e4SLinus Torvalds } 23781da177e4SLinus Torvalds 23792e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2380493a1724STejun Heo 23814d595b86SLai Jiangshan if (should_stop) { 23824d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23834d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23844d595b86SLai Jiangshan return 0; 23854d595b86SLai Jiangshan } 23864d595b86SLai Jiangshan 2387111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2388111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2389e22bee78STejun Heo schedule(); 2390e22bee78STejun Heo goto repeat; 23911da177e4SLinus Torvalds } 23921da177e4SLinus Torvalds 2393fca839c0STejun Heo /** 2394fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2395fca839c0STejun Heo * @target_wq: workqueue being flushed 2396fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2397fca839c0STejun Heo * 2398fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2399fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2400fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2401fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2402fca839c0STejun Heo * a deadlock. 2403fca839c0STejun Heo */ 2404fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2405fca839c0STejun Heo struct work_struct *target_work) 2406fca839c0STejun Heo { 2407fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2408fca839c0STejun Heo struct worker *worker; 2409fca839c0STejun Heo 2410fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2411fca839c0STejun Heo return; 2412fca839c0STejun Heo 2413fca839c0STejun Heo worker = current_wq_worker(); 2414fca839c0STejun Heo 2415fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2416fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2417fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 241823d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 241923d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2420fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2421fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2422fca839c0STejun Heo target_wq->name, target_func); 2423fca839c0STejun Heo } 2424fca839c0STejun Heo 2425fc2e4d70SOleg Nesterov struct wq_barrier { 2426fc2e4d70SOleg Nesterov struct work_struct work; 2427fc2e4d70SOleg Nesterov struct completion done; 24282607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2429fc2e4d70SOleg Nesterov }; 2430fc2e4d70SOleg Nesterov 2431fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2432fc2e4d70SOleg Nesterov { 2433fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2434fc2e4d70SOleg Nesterov complete(&barr->done); 2435fc2e4d70SOleg Nesterov } 2436fc2e4d70SOleg Nesterov 24374690c4abSTejun Heo /** 24384690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2439112202d9STejun Heo * @pwq: pwq to insert barrier into 24404690c4abSTejun Heo * @barr: wq_barrier to insert 2441affee4b2STejun Heo * @target: target work to attach @barr to 2442affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24434690c4abSTejun Heo * 2444affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2445affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2446affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2447affee4b2STejun Heo * cpu. 2448affee4b2STejun Heo * 2449affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2450affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2451affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2452affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2453affee4b2STejun Heo * after a work with LINKED flag set. 2454affee4b2STejun Heo * 2455affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2456112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24574690c4abSTejun Heo * 24584690c4abSTejun Heo * CONTEXT: 2459d565ed63STejun Heo * spin_lock_irq(pool->lock). 24604690c4abSTejun Heo */ 2461112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2462affee4b2STejun Heo struct wq_barrier *barr, 2463affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2464fc2e4d70SOleg Nesterov { 2465affee4b2STejun Heo struct list_head *head; 2466affee4b2STejun Heo unsigned int linked = 0; 2467affee4b2STejun Heo 2468dc186ad7SThomas Gleixner /* 2469d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2470dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2471dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2472dc186ad7SThomas Gleixner * might deadlock. 2473dc186ad7SThomas Gleixner */ 2474ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 247522df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2476fc2e4d70SOleg Nesterov init_completion(&barr->done); 24772607d7a6STejun Heo barr->task = current; 247883c22520SOleg Nesterov 2479affee4b2STejun Heo /* 2480affee4b2STejun Heo * If @target is currently being executed, schedule the 2481affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2482affee4b2STejun Heo */ 2483affee4b2STejun Heo if (worker) 2484affee4b2STejun Heo head = worker->scheduled.next; 2485affee4b2STejun Heo else { 2486affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2487affee4b2STejun Heo 2488affee4b2STejun Heo head = target->entry.next; 2489affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2490affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2491affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2492affee4b2STejun Heo } 2493affee4b2STejun Heo 2494dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2495112202d9STejun Heo insert_work(pwq, &barr->work, head, 2496affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2497fc2e4d70SOleg Nesterov } 2498fc2e4d70SOleg Nesterov 249973f53c4aSTejun Heo /** 2500112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 250173f53c4aSTejun Heo * @wq: workqueue being flushed 250273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 250373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 250473f53c4aSTejun Heo * 2505112202d9STejun Heo * Prepare pwqs for workqueue flushing. 250673f53c4aSTejun Heo * 2507112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2508112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2509112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2510112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2511112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 251273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 251373f53c4aSTejun Heo * 251473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 251573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 251673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 251773f53c4aSTejun Heo * is returned. 251873f53c4aSTejun Heo * 2519112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 252073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 252173f53c4aSTejun Heo * advanced to @work_color. 252273f53c4aSTejun Heo * 252373f53c4aSTejun Heo * CONTEXT: 25243c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 252573f53c4aSTejun Heo * 2526d185af30SYacine Belkadi * Return: 252773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 252873f53c4aSTejun Heo * otherwise. 252973f53c4aSTejun Heo */ 2530112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 253173f53c4aSTejun Heo int flush_color, int work_color) 25321da177e4SLinus Torvalds { 253373f53c4aSTejun Heo bool wait = false; 253449e3cf44STejun Heo struct pool_workqueue *pwq; 25351da177e4SLinus Torvalds 253673f53c4aSTejun Heo if (flush_color >= 0) { 25376183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2538112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2539dc186ad7SThomas Gleixner } 254014441960SOleg Nesterov 254149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2542112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25431da177e4SLinus Torvalds 2544b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 254573f53c4aSTejun Heo 254673f53c4aSTejun Heo if (flush_color >= 0) { 25476183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 254873f53c4aSTejun Heo 2549112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2550112202d9STejun Heo pwq->flush_color = flush_color; 2551112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 255273f53c4aSTejun Heo wait = true; 25531da177e4SLinus Torvalds } 255473f53c4aSTejun Heo } 255573f53c4aSTejun Heo 255673f53c4aSTejun Heo if (work_color >= 0) { 25576183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2558112202d9STejun Heo pwq->work_color = work_color; 255973f53c4aSTejun Heo } 256073f53c4aSTejun Heo 2561b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25621da177e4SLinus Torvalds } 25631da177e4SLinus Torvalds 2564112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 256573f53c4aSTejun Heo complete(&wq->first_flusher->done); 256673f53c4aSTejun Heo 256773f53c4aSTejun Heo return wait; 256883c22520SOleg Nesterov } 25691da177e4SLinus Torvalds 25700fcb78c2SRolf Eike Beer /** 25711da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25720fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25731da177e4SLinus Torvalds * 2574c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2575c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25761da177e4SLinus Torvalds */ 25777ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25781da177e4SLinus Torvalds { 257973f53c4aSTejun Heo struct wq_flusher this_flusher = { 258073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 258173f53c4aSTejun Heo .flush_color = -1, 258273f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 258373f53c4aSTejun Heo }; 258473f53c4aSTejun Heo int next_color; 2585b1f4ec17SOleg Nesterov 25863295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25873295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 258873f53c4aSTejun Heo 25893c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 259073f53c4aSTejun Heo 259173f53c4aSTejun Heo /* 259273f53c4aSTejun Heo * Start-to-wait phase 259373f53c4aSTejun Heo */ 259473f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 259573f53c4aSTejun Heo 259673f53c4aSTejun Heo if (next_color != wq->flush_color) { 259773f53c4aSTejun Heo /* 259873f53c4aSTejun Heo * Color space is not full. The current work_color 259973f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 260073f53c4aSTejun Heo * by one. 260173f53c4aSTejun Heo */ 26026183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 260373f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 260473f53c4aSTejun Heo wq->work_color = next_color; 260573f53c4aSTejun Heo 260673f53c4aSTejun Heo if (!wq->first_flusher) { 260773f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26086183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 260973f53c4aSTejun Heo 261073f53c4aSTejun Heo wq->first_flusher = &this_flusher; 261173f53c4aSTejun Heo 2612112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 261373f53c4aSTejun Heo wq->work_color)) { 261473f53c4aSTejun Heo /* nothing to flush, done */ 261573f53c4aSTejun Heo wq->flush_color = next_color; 261673f53c4aSTejun Heo wq->first_flusher = NULL; 261773f53c4aSTejun Heo goto out_unlock; 261873f53c4aSTejun Heo } 261973f53c4aSTejun Heo } else { 262073f53c4aSTejun Heo /* wait in queue */ 26216183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 262273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2623112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 262473f53c4aSTejun Heo } 262573f53c4aSTejun Heo } else { 262673f53c4aSTejun Heo /* 262773f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 262873f53c4aSTejun Heo * The next flush completion will assign us 262973f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 263073f53c4aSTejun Heo */ 263173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 263273f53c4aSTejun Heo } 263373f53c4aSTejun Heo 2634fca839c0STejun Heo check_flush_dependency(wq, NULL); 2635fca839c0STejun Heo 26363c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 263973f53c4aSTejun Heo 264073f53c4aSTejun Heo /* 264173f53c4aSTejun Heo * Wake-up-and-cascade phase 264273f53c4aSTejun Heo * 264373f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 264473f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 264573f53c4aSTejun Heo */ 264673f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 264773f53c4aSTejun Heo return; 264873f53c4aSTejun Heo 26493c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 265073f53c4aSTejun Heo 26514ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26524ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26534ce48b37STejun Heo goto out_unlock; 26544ce48b37STejun Heo 265573f53c4aSTejun Heo wq->first_flusher = NULL; 265673f53c4aSTejun Heo 26576183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26586183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 265973f53c4aSTejun Heo 266073f53c4aSTejun Heo while (true) { 266173f53c4aSTejun Heo struct wq_flusher *next, *tmp; 266273f53c4aSTejun Heo 266373f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 266473f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 266573f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 266673f53c4aSTejun Heo break; 266773f53c4aSTejun Heo list_del_init(&next->list); 266873f53c4aSTejun Heo complete(&next->done); 266973f53c4aSTejun Heo } 267073f53c4aSTejun Heo 26716183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 267273f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 267373f53c4aSTejun Heo 267473f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 267573f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 267673f53c4aSTejun Heo 267773f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 267873f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 267973f53c4aSTejun Heo /* 268073f53c4aSTejun Heo * Assign the same color to all overflowed 268173f53c4aSTejun Heo * flushers, advance work_color and append to 268273f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 268373f53c4aSTejun Heo * phase for these overflowed flushers. 268473f53c4aSTejun Heo */ 268573f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 268673f53c4aSTejun Heo tmp->flush_color = wq->work_color; 268773f53c4aSTejun Heo 268873f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 268973f53c4aSTejun Heo 269073f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 269173f53c4aSTejun Heo &wq->flusher_queue); 2692112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 269373f53c4aSTejun Heo } 269473f53c4aSTejun Heo 269573f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26966183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 269773f53c4aSTejun Heo break; 269873f53c4aSTejun Heo } 269973f53c4aSTejun Heo 270073f53c4aSTejun Heo /* 270173f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2702112202d9STejun Heo * the new first flusher and arm pwqs. 270373f53c4aSTejun Heo */ 27046183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27056183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 270673f53c4aSTejun Heo 270773f53c4aSTejun Heo list_del_init(&next->list); 270873f53c4aSTejun Heo wq->first_flusher = next; 270973f53c4aSTejun Heo 2710112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 271173f53c4aSTejun Heo break; 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo /* 271473f53c4aSTejun Heo * Meh... this color is already done, clear first 271573f53c4aSTejun Heo * flusher and repeat cascading. 271673f53c4aSTejun Heo */ 271773f53c4aSTejun Heo wq->first_flusher = NULL; 271873f53c4aSTejun Heo } 271973f53c4aSTejun Heo 272073f53c4aSTejun Heo out_unlock: 27213c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27221da177e4SLinus Torvalds } 27231dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 27241da177e4SLinus Torvalds 27259c5a2ba7STejun Heo /** 27269c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27279c5a2ba7STejun Heo * @wq: workqueue to drain 27289c5a2ba7STejun Heo * 27299c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27309c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27319c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2732b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 27339c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27349c5a2ba7STejun Heo * takes too long. 27359c5a2ba7STejun Heo */ 27369c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27379c5a2ba7STejun Heo { 27389c5a2ba7STejun Heo unsigned int flush_cnt = 0; 273949e3cf44STejun Heo struct pool_workqueue *pwq; 27409c5a2ba7STejun Heo 27419c5a2ba7STejun Heo /* 27429c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27439c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2744618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27459c5a2ba7STejun Heo */ 274687fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27479c5a2ba7STejun Heo if (!wq->nr_drainers++) 2748618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 274987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27509c5a2ba7STejun Heo reflush: 27519c5a2ba7STejun Heo flush_workqueue(wq); 27529c5a2ba7STejun Heo 2753b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 275476af4d93STejun Heo 275549e3cf44STejun Heo for_each_pwq(pwq, wq) { 2756fa2563e4SThomas Tuttle bool drained; 27579c5a2ba7STejun Heo 2758b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2759112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2760b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2761fa2563e4SThomas Tuttle 2762fa2563e4SThomas Tuttle if (drained) 27639c5a2ba7STejun Heo continue; 27649c5a2ba7STejun Heo 27659c5a2ba7STejun Heo if (++flush_cnt == 10 || 27669c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2767c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27689c5a2ba7STejun Heo wq->name, flush_cnt); 276976af4d93STejun Heo 2770b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27719c5a2ba7STejun Heo goto reflush; 27729c5a2ba7STejun Heo } 27739c5a2ba7STejun Heo 27749c5a2ba7STejun Heo if (!--wq->nr_drainers) 2775618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 277687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27779c5a2ba7STejun Heo } 27789c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27799c5a2ba7STejun Heo 2780606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2781baf59022STejun Heo { 2782baf59022STejun Heo struct worker *worker = NULL; 2783c9e7cf27STejun Heo struct worker_pool *pool; 2784112202d9STejun Heo struct pool_workqueue *pwq; 2785baf59022STejun Heo 2786baf59022STejun Heo might_sleep(); 2787baf59022STejun Heo 2788fa1b54e6STejun Heo local_irq_disable(); 2789fa1b54e6STejun Heo pool = get_work_pool(work); 2790fa1b54e6STejun Heo if (!pool) { 2791fa1b54e6STejun Heo local_irq_enable(); 2792fa1b54e6STejun Heo return false; 2793fa1b54e6STejun Heo } 2794fa1b54e6STejun Heo 2795fa1b54e6STejun Heo spin_lock(&pool->lock); 27960b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2797112202d9STejun Heo pwq = get_work_pwq(work); 2798112202d9STejun Heo if (pwq) { 2799112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2800baf59022STejun Heo goto already_gone; 2801606a5020STejun Heo } else { 2802c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2803baf59022STejun Heo if (!worker) 2804baf59022STejun Heo goto already_gone; 2805112202d9STejun Heo pwq = worker->current_pwq; 2806606a5020STejun Heo } 2807baf59022STejun Heo 2808fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2809fca839c0STejun Heo 2810112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2811d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2812baf59022STejun Heo 2813e159489bSTejun Heo /* 2814e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2815e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2816e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2817e159489bSTejun Heo * access. 2818e159489bSTejun Heo */ 2819493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2820112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2821e159489bSTejun Heo else 2822112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2823112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2824e159489bSTejun Heo 2825baf59022STejun Heo return true; 2826baf59022STejun Heo already_gone: 2827d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2828baf59022STejun Heo return false; 2829baf59022STejun Heo } 2830baf59022STejun Heo 2831db700897SOleg Nesterov /** 2832401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2833401a8d04STejun Heo * @work: the work to flush 2834db700897SOleg Nesterov * 2835606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2836606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2837401a8d04STejun Heo * 2838d185af30SYacine Belkadi * Return: 2839401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2840401a8d04STejun Heo * %false if it was already idle. 2841db700897SOleg Nesterov */ 2842401a8d04STejun Heo bool flush_work(struct work_struct *work) 2843db700897SOleg Nesterov { 284412997d1aSBjorn Helgaas struct wq_barrier barr; 284512997d1aSBjorn Helgaas 28460976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28470976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28480976dfc1SStephen Boyd 284912997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 285012997d1aSBjorn Helgaas wait_for_completion(&barr.done); 285112997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 285212997d1aSBjorn Helgaas return true; 285312997d1aSBjorn Helgaas } else { 285412997d1aSBjorn Helgaas return false; 285512997d1aSBjorn Helgaas } 2856606a5020STejun Heo } 2857db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2858db700897SOleg Nesterov 28598603e1b3STejun Heo struct cwt_wait { 28608603e1b3STejun Heo wait_queue_t wait; 28618603e1b3STejun Heo struct work_struct *work; 28628603e1b3STejun Heo }; 28638603e1b3STejun Heo 28648603e1b3STejun Heo static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) 28658603e1b3STejun Heo { 28668603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 28678603e1b3STejun Heo 28688603e1b3STejun Heo if (cwait->work != key) 28698603e1b3STejun Heo return 0; 28708603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 28718603e1b3STejun Heo } 28728603e1b3STejun Heo 287336e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2874401a8d04STejun Heo { 28758603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2876bbb68dfaSTejun Heo unsigned long flags; 28771f1f642eSOleg Nesterov int ret; 28781f1f642eSOleg Nesterov 28791f1f642eSOleg Nesterov do { 2880bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2881bbb68dfaSTejun Heo /* 28828603e1b3STejun Heo * If someone else is already canceling, wait for it to 28838603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 28848603e1b3STejun Heo * because we may get scheduled between @work's completion 28858603e1b3STejun Heo * and the other canceling task resuming and clearing 28868603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 28878603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 28888603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 28898603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 28908603e1b3STejun Heo * we're hogging the CPU. 28918603e1b3STejun Heo * 28928603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 28938603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 28948603e1b3STejun Heo * wake function which matches @work along with exclusive 28958603e1b3STejun Heo * wait and wakeup. 2896bbb68dfaSTejun Heo */ 28978603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 28988603e1b3STejun Heo struct cwt_wait cwait; 28998603e1b3STejun Heo 29008603e1b3STejun Heo init_wait(&cwait.wait); 29018603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 29028603e1b3STejun Heo cwait.work = work; 29038603e1b3STejun Heo 29048603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 29058603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 29068603e1b3STejun Heo if (work_is_canceling(work)) 29078603e1b3STejun Heo schedule(); 29088603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 29098603e1b3STejun Heo } 29101f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29111f1f642eSOleg Nesterov 2912bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2913bbb68dfaSTejun Heo mark_work_canceling(work); 2914bbb68dfaSTejun Heo local_irq_restore(flags); 2915bbb68dfaSTejun Heo 2916606a5020STejun Heo flush_work(work); 29177a22ad75STejun Heo clear_work_data(work); 29188603e1b3STejun Heo 29198603e1b3STejun Heo /* 29208603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 29218603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 29228603e1b3STejun Heo * visible there. 29238603e1b3STejun Heo */ 29248603e1b3STejun Heo smp_mb(); 29258603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 29268603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 29278603e1b3STejun Heo 29281f1f642eSOleg Nesterov return ret; 29291f1f642eSOleg Nesterov } 29301f1f642eSOleg Nesterov 29316e84d644SOleg Nesterov /** 2932401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2933401a8d04STejun Heo * @work: the work to cancel 29346e84d644SOleg Nesterov * 2935401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2936401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2937401a8d04STejun Heo * another workqueue. On return from this function, @work is 2938401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29391f1f642eSOleg Nesterov * 2940401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2941401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29426e84d644SOleg Nesterov * 2943401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29446e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2945401a8d04STejun Heo * 2946d185af30SYacine Belkadi * Return: 2947401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29486e84d644SOleg Nesterov */ 2949401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29506e84d644SOleg Nesterov { 295136e227d2STejun Heo return __cancel_work_timer(work, false); 2952b89deed3SOleg Nesterov } 295328e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2954b89deed3SOleg Nesterov 29556e84d644SOleg Nesterov /** 2956401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2957401a8d04STejun Heo * @dwork: the delayed work to flush 29586e84d644SOleg Nesterov * 2959401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2960401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2961401a8d04STejun Heo * considers the last queueing instance of @dwork. 29621f1f642eSOleg Nesterov * 2963d185af30SYacine Belkadi * Return: 2964401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2965401a8d04STejun Heo * %false if it was already idle. 29666e84d644SOleg Nesterov */ 2967401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2968401a8d04STejun Heo { 29698930cabaSTejun Heo local_irq_disable(); 2970401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 297160c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29728930cabaSTejun Heo local_irq_enable(); 2973401a8d04STejun Heo return flush_work(&dwork->work); 2974401a8d04STejun Heo } 2975401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2976401a8d04STejun Heo 2977f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 2978f72b8792SJens Axboe { 2979f72b8792SJens Axboe unsigned long flags; 2980f72b8792SJens Axboe int ret; 2981f72b8792SJens Axboe 2982f72b8792SJens Axboe do { 2983f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 2984f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 2985f72b8792SJens Axboe 2986f72b8792SJens Axboe if (unlikely(ret < 0)) 2987f72b8792SJens Axboe return false; 2988f72b8792SJens Axboe 2989f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 2990f72b8792SJens Axboe local_irq_restore(flags); 2991f72b8792SJens Axboe return ret; 2992f72b8792SJens Axboe } 2993f72b8792SJens Axboe 2994f72b8792SJens Axboe /* 2995f72b8792SJens Axboe * See cancel_delayed_work() 2996f72b8792SJens Axboe */ 2997f72b8792SJens Axboe bool cancel_work(struct work_struct *work) 2998f72b8792SJens Axboe { 2999f72b8792SJens Axboe return __cancel_work(work, false); 3000f72b8792SJens Axboe } 3001f72b8792SJens Axboe 3002401a8d04STejun Heo /** 300357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 300457b30ae7STejun Heo * @dwork: delayed_work to cancel 300509383498STejun Heo * 3006d185af30SYacine Belkadi * Kill off a pending delayed_work. 3007d185af30SYacine Belkadi * 3008d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3009d185af30SYacine Belkadi * pending. 3010d185af30SYacine Belkadi * 3011d185af30SYacine Belkadi * Note: 3012d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3013d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3014d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 301509383498STejun Heo * 301657b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 301709383498STejun Heo */ 301857b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 301909383498STejun Heo { 3020f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 302109383498STejun Heo } 302257b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 302309383498STejun Heo 302409383498STejun Heo /** 3025401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3026401a8d04STejun Heo * @dwork: the delayed work cancel 3027401a8d04STejun Heo * 3028401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3029401a8d04STejun Heo * 3030d185af30SYacine Belkadi * Return: 3031401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3032401a8d04STejun Heo */ 3033401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30346e84d644SOleg Nesterov { 303536e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30366e84d644SOleg Nesterov } 3037f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30381da177e4SLinus Torvalds 30390fcb78c2SRolf Eike Beer /** 304031ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3041b6136773SAndrew Morton * @func: the function to call 3042b6136773SAndrew Morton * 304331ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 304431ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3045b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 304631ddd871STejun Heo * 3047d185af30SYacine Belkadi * Return: 304831ddd871STejun Heo * 0 on success, -errno on failure. 3049b6136773SAndrew Morton */ 305065f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 305115316ba8SChristoph Lameter { 305215316ba8SChristoph Lameter int cpu; 305338f51568SNamhyung Kim struct work_struct __percpu *works; 305415316ba8SChristoph Lameter 3055b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3056b6136773SAndrew Morton if (!works) 305715316ba8SChristoph Lameter return -ENOMEM; 3058b6136773SAndrew Morton 305995402b38SGautham R Shenoy get_online_cpus(); 306093981800STejun Heo 306115316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30629bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30639bfb1839SIngo Molnar 30649bfb1839SIngo Molnar INIT_WORK(work, func); 30658de6d308SOleg Nesterov schedule_work_on(cpu, work); 306615316ba8SChristoph Lameter } 306793981800STejun Heo 306893981800STejun Heo for_each_online_cpu(cpu) 30698616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 307093981800STejun Heo 307195402b38SGautham R Shenoy put_online_cpus(); 3072b6136773SAndrew Morton free_percpu(works); 307315316ba8SChristoph Lameter return 0; 307415316ba8SChristoph Lameter } 307515316ba8SChristoph Lameter 3076eef6a7d5SAlan Stern /** 30771fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30781fa44ecaSJames Bottomley * @fn: the function to execute 30791fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30801fa44ecaSJames Bottomley * be available when the work executes) 30811fa44ecaSJames Bottomley * 30821fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30831fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30841fa44ecaSJames Bottomley * 3085d185af30SYacine Belkadi * Return: 0 - function was executed 30861fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30871fa44ecaSJames Bottomley */ 308865f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30891fa44ecaSJames Bottomley { 30901fa44ecaSJames Bottomley if (!in_interrupt()) { 309165f27f38SDavid Howells fn(&ew->work); 30921fa44ecaSJames Bottomley return 0; 30931fa44ecaSJames Bottomley } 30941fa44ecaSJames Bottomley 309565f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30961fa44ecaSJames Bottomley schedule_work(&ew->work); 30971fa44ecaSJames Bottomley 30981fa44ecaSJames Bottomley return 1; 30991fa44ecaSJames Bottomley } 31001fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31011fa44ecaSJames Bottomley 31027a4e344cSTejun Heo /** 31037a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 31047a4e344cSTejun Heo * @attrs: workqueue_attrs to free 31057a4e344cSTejun Heo * 31067a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 31077a4e344cSTejun Heo */ 31087a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 31097a4e344cSTejun Heo { 31107a4e344cSTejun Heo if (attrs) { 31117a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 31127a4e344cSTejun Heo kfree(attrs); 31137a4e344cSTejun Heo } 31147a4e344cSTejun Heo } 31157a4e344cSTejun Heo 31167a4e344cSTejun Heo /** 31177a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 31187a4e344cSTejun Heo * @gfp_mask: allocation mask to use 31197a4e344cSTejun Heo * 31207a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3121d185af30SYacine Belkadi * return it. 3122d185af30SYacine Belkadi * 3123d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 31247a4e344cSTejun Heo */ 31257a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 31267a4e344cSTejun Heo { 31277a4e344cSTejun Heo struct workqueue_attrs *attrs; 31287a4e344cSTejun Heo 31297a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 31307a4e344cSTejun Heo if (!attrs) 31317a4e344cSTejun Heo goto fail; 31327a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 31337a4e344cSTejun Heo goto fail; 31347a4e344cSTejun Heo 313513e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 31367a4e344cSTejun Heo return attrs; 31377a4e344cSTejun Heo fail: 31387a4e344cSTejun Heo free_workqueue_attrs(attrs); 31397a4e344cSTejun Heo return NULL; 31407a4e344cSTejun Heo } 31417a4e344cSTejun Heo 314229c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 314329c91e99STejun Heo const struct workqueue_attrs *from) 314429c91e99STejun Heo { 314529c91e99STejun Heo to->nice = from->nice; 314629c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 31472865a8fbSShaohua Li /* 31482865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 31492865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 31502865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 31512865a8fbSShaohua Li */ 31522865a8fbSShaohua Li to->no_numa = from->no_numa; 315329c91e99STejun Heo } 315429c91e99STejun Heo 315529c91e99STejun Heo /* hash value of the content of @attr */ 315629c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 315729c91e99STejun Heo { 315829c91e99STejun Heo u32 hash = 0; 315929c91e99STejun Heo 316029c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 316113e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 316213e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 316329c91e99STejun Heo return hash; 316429c91e99STejun Heo } 316529c91e99STejun Heo 316629c91e99STejun Heo /* content equality test */ 316729c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 316829c91e99STejun Heo const struct workqueue_attrs *b) 316929c91e99STejun Heo { 317029c91e99STejun Heo if (a->nice != b->nice) 317129c91e99STejun Heo return false; 317229c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 317329c91e99STejun Heo return false; 317429c91e99STejun Heo return true; 317529c91e99STejun Heo } 317629c91e99STejun Heo 31777a4e344cSTejun Heo /** 31787a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 31797a4e344cSTejun Heo * @pool: worker_pool to initialize 31807a4e344cSTejun Heo * 3181402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3182d185af30SYacine Belkadi * 3183d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 318429c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 318529c91e99STejun Heo * on @pool safely to release it. 31867a4e344cSTejun Heo */ 31877a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 31884e1a1f9aSTejun Heo { 31894e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 319029c91e99STejun Heo pool->id = -1; 319129c91e99STejun Heo pool->cpu = -1; 3192f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 31934e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 319482607adcSTejun Heo pool->watchdog_ts = jiffies; 31954e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 31964e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 31974e1a1f9aSTejun Heo hash_init(pool->busy_hash); 31984e1a1f9aSTejun Heo 31994e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 32004e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 32014e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 32024e1a1f9aSTejun Heo 32034e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 32044e1a1f9aSTejun Heo (unsigned long)pool); 32054e1a1f9aSTejun Heo 32064e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 320792f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3208da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 32097a4e344cSTejun Heo 32107cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 321129c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 321229c91e99STejun Heo pool->refcnt = 1; 321329c91e99STejun Heo 321429c91e99STejun Heo /* shouldn't fail above this point */ 32157a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 32167a4e344cSTejun Heo if (!pool->attrs) 32177a4e344cSTejun Heo return -ENOMEM; 32187a4e344cSTejun Heo return 0; 32194e1a1f9aSTejun Heo } 32204e1a1f9aSTejun Heo 3221e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3222e2dca7adSTejun Heo { 3223e2dca7adSTejun Heo struct workqueue_struct *wq = 3224e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3225e2dca7adSTejun Heo 3226e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3227e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3228e2dca7adSTejun Heo else 3229e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3230e2dca7adSTejun Heo 3231e2dca7adSTejun Heo kfree(wq->rescuer); 3232e2dca7adSTejun Heo kfree(wq); 3233e2dca7adSTejun Heo } 3234e2dca7adSTejun Heo 323529c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 323629c91e99STejun Heo { 323729c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 323829c91e99STejun Heo 32397cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 324029c91e99STejun Heo free_workqueue_attrs(pool->attrs); 324129c91e99STejun Heo kfree(pool); 324229c91e99STejun Heo } 324329c91e99STejun Heo 324429c91e99STejun Heo /** 324529c91e99STejun Heo * put_unbound_pool - put a worker_pool 324629c91e99STejun Heo * @pool: worker_pool to put 324729c91e99STejun Heo * 324829c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3249c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3250c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3251c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3252a892caccSTejun Heo * 3253a892caccSTejun Heo * Should be called with wq_pool_mutex held. 325429c91e99STejun Heo */ 325529c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 325629c91e99STejun Heo { 325760f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 325829c91e99STejun Heo struct worker *worker; 325929c91e99STejun Heo 3260a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3261a892caccSTejun Heo 3262a892caccSTejun Heo if (--pool->refcnt) 326329c91e99STejun Heo return; 326429c91e99STejun Heo 326529c91e99STejun Heo /* sanity checks */ 326661d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3267a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 326829c91e99STejun Heo return; 326929c91e99STejun Heo 327029c91e99STejun Heo /* release id and unhash */ 327129c91e99STejun Heo if (pool->id >= 0) 327229c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 327329c91e99STejun Heo hash_del(&pool->hash_node); 327429c91e99STejun Heo 3275c5aa87bbSTejun Heo /* 3276c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3277c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 327892f9c5c4SLai Jiangshan * attach_mutex. 3279c5aa87bbSTejun Heo */ 328029c91e99STejun Heo mutex_lock(&pool->manager_arb); 328129c91e99STejun Heo 328260f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 32831037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 328429c91e99STejun Heo destroy_worker(worker); 328529c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 328629c91e99STejun Heo spin_unlock_irq(&pool->lock); 328760f5a4bcSLai Jiangshan 328892f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3289da028469SLai Jiangshan if (!list_empty(&pool->workers)) 329060f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 329192f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 329260f5a4bcSLai Jiangshan 329360f5a4bcSLai Jiangshan if (pool->detach_completion) 329460f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 329560f5a4bcSLai Jiangshan 329629c91e99STejun Heo mutex_unlock(&pool->manager_arb); 329729c91e99STejun Heo 329829c91e99STejun Heo /* shut down the timers */ 329929c91e99STejun Heo del_timer_sync(&pool->idle_timer); 330029c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 330129c91e99STejun Heo 330229c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 330329c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 330429c91e99STejun Heo } 330529c91e99STejun Heo 330629c91e99STejun Heo /** 330729c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 330829c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 330929c91e99STejun Heo * 331029c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 331129c91e99STejun Heo * reference count and return it. If there already is a matching 331229c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3313d185af30SYacine Belkadi * create a new one. 3314a892caccSTejun Heo * 3315a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3316d185af30SYacine Belkadi * 3317d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3318d185af30SYacine Belkadi * On failure, %NULL. 331929c91e99STejun Heo */ 332029c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 332129c91e99STejun Heo { 332229c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 332329c91e99STejun Heo struct worker_pool *pool; 3324f3f90ad4STejun Heo int node; 3325e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 332629c91e99STejun Heo 3327a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 332829c91e99STejun Heo 332929c91e99STejun Heo /* do we already have a matching pool? */ 333029c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 333129c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 333229c91e99STejun Heo pool->refcnt++; 33333fb1823cSLai Jiangshan return pool; 333429c91e99STejun Heo } 333529c91e99STejun Heo } 333629c91e99STejun Heo 3337e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3338e2273584SXunlei Pang if (wq_numa_enabled) { 3339e2273584SXunlei Pang for_each_node(node) { 3340e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3341e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3342e2273584SXunlei Pang target_node = node; 3343e2273584SXunlei Pang break; 3344e2273584SXunlei Pang } 3345e2273584SXunlei Pang } 3346e2273584SXunlei Pang } 3347e2273584SXunlei Pang 334829c91e99STejun Heo /* nope, create a new one */ 3349e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 335029c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 335129c91e99STejun Heo goto fail; 335229c91e99STejun Heo 33538864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 335429c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3355e2273584SXunlei Pang pool->node = target_node; 335629c91e99STejun Heo 33572865a8fbSShaohua Li /* 33582865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 33592865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 33602865a8fbSShaohua Li */ 33612865a8fbSShaohua Li pool->attrs->no_numa = false; 33622865a8fbSShaohua Li 336329c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 336429c91e99STejun Heo goto fail; 336529c91e99STejun Heo 336629c91e99STejun Heo /* create and start the initial worker */ 3367051e1850SLai Jiangshan if (!create_worker(pool)) 336829c91e99STejun Heo goto fail; 336929c91e99STejun Heo 337029c91e99STejun Heo /* install */ 337129c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 33723fb1823cSLai Jiangshan 337329c91e99STejun Heo return pool; 337429c91e99STejun Heo fail: 337529c91e99STejun Heo if (pool) 337629c91e99STejun Heo put_unbound_pool(pool); 337729c91e99STejun Heo return NULL; 337829c91e99STejun Heo } 337929c91e99STejun Heo 33808864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 33818864b4e5STejun Heo { 33828864b4e5STejun Heo kmem_cache_free(pwq_cache, 33838864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 33848864b4e5STejun Heo } 33858864b4e5STejun Heo 33868864b4e5STejun Heo /* 33878864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 33888864b4e5STejun Heo * and needs to be destroyed. 33898864b4e5STejun Heo */ 33908864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 33918864b4e5STejun Heo { 33928864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 33938864b4e5STejun Heo unbound_release_work); 33948864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 33958864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3396bc0caf09STejun Heo bool is_last; 33978864b4e5STejun Heo 33988864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 33998864b4e5STejun Heo return; 34008864b4e5STejun Heo 34013c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 34028864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3403bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 34043c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 34058864b4e5STejun Heo 3406a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 34078864b4e5STejun Heo put_unbound_pool(pool); 3408a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3409a892caccSTejun Heo 34108864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 34118864b4e5STejun Heo 34128864b4e5STejun Heo /* 34138864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3414e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 34158864b4e5STejun Heo */ 3416e2dca7adSTejun Heo if (is_last) 3417e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 34186029a918STejun Heo } 34198864b4e5STejun Heo 34200fbd95aaSTejun Heo /** 3421699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 34220fbd95aaSTejun Heo * @pwq: target pool_workqueue 34230fbd95aaSTejun Heo * 3424699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3425699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3426699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 34270fbd95aaSTejun Heo */ 3428699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 34290fbd95aaSTejun Heo { 3430699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3431699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3432699ce097STejun Heo 3433699ce097STejun Heo /* for @wq->saved_max_active */ 3434a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3435699ce097STejun Heo 3436699ce097STejun Heo /* fast exit for non-freezable wqs */ 3437699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3438699ce097STejun Heo return; 3439699ce097STejun Heo 3440a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3441699ce097STejun Heo 344274b414eaSLai Jiangshan /* 344374b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 344474b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 344574b414eaSLai Jiangshan * is updated and visible. 344674b414eaSLai Jiangshan */ 344774b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3448699ce097STejun Heo pwq->max_active = wq->saved_max_active; 34490fbd95aaSTejun Heo 34500fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 34510fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 34520fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3453951a078aSLai Jiangshan 3454951a078aSLai Jiangshan /* 3455951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3456951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3457951a078aSLai Jiangshan */ 3458951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3459699ce097STejun Heo } else { 3460699ce097STejun Heo pwq->max_active = 0; 3461699ce097STejun Heo } 3462699ce097STejun Heo 3463a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 34640fbd95aaSTejun Heo } 34650fbd95aaSTejun Heo 3466e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3467f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3468f147f29eSTejun Heo struct worker_pool *pool) 3469d2c1d404STejun Heo { 3470d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3471d2c1d404STejun Heo 3472e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3473e50aba9aSTejun Heo 3474d2c1d404STejun Heo pwq->pool = pool; 3475d2c1d404STejun Heo pwq->wq = wq; 3476d2c1d404STejun Heo pwq->flush_color = -1; 34778864b4e5STejun Heo pwq->refcnt = 1; 3478d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 34791befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3480d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 34818864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3482f147f29eSTejun Heo } 3483d2c1d404STejun Heo 3484f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 34851befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3486f147f29eSTejun Heo { 3487f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3488f147f29eSTejun Heo 3489f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 349075ccf595STejun Heo 34911befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 34921befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 34931befcf30STejun Heo return; 34941befcf30STejun Heo 349529b1cb41SLai Jiangshan /* set the matching work_color */ 349675ccf595STejun Heo pwq->work_color = wq->work_color; 3497983ca25eSTejun Heo 3498983ca25eSTejun Heo /* sync max_active to the current setting */ 3499983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3500983ca25eSTejun Heo 3501983ca25eSTejun Heo /* link in @pwq */ 35029e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3503df2d5ae4STejun Heo } 35046029a918STejun Heo 3505f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3506f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3507f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3508f147f29eSTejun Heo { 3509f147f29eSTejun Heo struct worker_pool *pool; 3510f147f29eSTejun Heo struct pool_workqueue *pwq; 3511f147f29eSTejun Heo 3512f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3513f147f29eSTejun Heo 3514f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3515f147f29eSTejun Heo if (!pool) 3516f147f29eSTejun Heo return NULL; 3517f147f29eSTejun Heo 3518e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3519f147f29eSTejun Heo if (!pwq) { 3520f147f29eSTejun Heo put_unbound_pool(pool); 3521f147f29eSTejun Heo return NULL; 3522f147f29eSTejun Heo } 3523f147f29eSTejun Heo 3524f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3525f147f29eSTejun Heo return pwq; 3526d2c1d404STejun Heo } 3527d2c1d404STejun Heo 35284c16bd32STejun Heo /** 352930186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3530042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 35314c16bd32STejun Heo * @node: the target NUMA node 35324c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 35334c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 35344c16bd32STejun Heo * 35354c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 35364c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3537d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 35384c16bd32STejun Heo * 35394c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 35404c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 35414c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 35424c16bd32STejun Heo * @attrs->cpumask. 35434c16bd32STejun Heo * 35444c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 35454c16bd32STejun Heo * stable. 3546d185af30SYacine Belkadi * 3547d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3548d185af30SYacine Belkadi * %false if equal. 35494c16bd32STejun Heo */ 35504c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 35514c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 35524c16bd32STejun Heo { 3553d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 35544c16bd32STejun Heo goto use_dfl; 35554c16bd32STejun Heo 35564c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 35574c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 35584c16bd32STejun Heo if (cpu_going_down >= 0) 35594c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 35604c16bd32STejun Heo 35614c16bd32STejun Heo if (cpumask_empty(cpumask)) 35624c16bd32STejun Heo goto use_dfl; 35634c16bd32STejun Heo 35644c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 35654c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 35664c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 35674c16bd32STejun Heo 35684c16bd32STejun Heo use_dfl: 35694c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 35704c16bd32STejun Heo return false; 35714c16bd32STejun Heo } 35724c16bd32STejun Heo 35731befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 35741befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 35751befcf30STejun Heo int node, 35761befcf30STejun Heo struct pool_workqueue *pwq) 35771befcf30STejun Heo { 35781befcf30STejun Heo struct pool_workqueue *old_pwq; 35791befcf30STejun Heo 35805b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 35811befcf30STejun Heo lockdep_assert_held(&wq->mutex); 35821befcf30STejun Heo 35831befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 35841befcf30STejun Heo link_pwq(pwq); 35851befcf30STejun Heo 35861befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 35871befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 35881befcf30STejun Heo return old_pwq; 35891befcf30STejun Heo } 35901befcf30STejun Heo 35912d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 35922d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 35932d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 35942d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3595042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 35962d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 35972d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 35982d5f0764SLai Jiangshan }; 35992d5f0764SLai Jiangshan 36002d5f0764SLai Jiangshan /* free the resources after success or abort */ 36012d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 36022d5f0764SLai Jiangshan { 36032d5f0764SLai Jiangshan if (ctx) { 36042d5f0764SLai Jiangshan int node; 36052d5f0764SLai Jiangshan 36062d5f0764SLai Jiangshan for_each_node(node) 36072d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 36082d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 36092d5f0764SLai Jiangshan 36102d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 36112d5f0764SLai Jiangshan 36122d5f0764SLai Jiangshan kfree(ctx); 36132d5f0764SLai Jiangshan } 36142d5f0764SLai Jiangshan } 36152d5f0764SLai Jiangshan 36162d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 36172d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 36182d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 36192d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 36202d5f0764SLai Jiangshan { 36212d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 36222d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 36232d5f0764SLai Jiangshan int node; 36242d5f0764SLai Jiangshan 36252d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36262d5f0764SLai Jiangshan 36272d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 36282d5f0764SLai Jiangshan GFP_KERNEL); 36292d5f0764SLai Jiangshan 36302d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36312d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36322d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 36332d5f0764SLai Jiangshan goto out_free; 36342d5f0764SLai Jiangshan 3635042f7df1SLai Jiangshan /* 3636042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3637042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3638042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3639042f7df1SLai Jiangshan */ 36402d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3641b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3642042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3643042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 36442d5f0764SLai Jiangshan 36452d5f0764SLai Jiangshan /* 36462d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 36472d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 36482d5f0764SLai Jiangshan * pools. 36492d5f0764SLai Jiangshan */ 36502d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 36512d5f0764SLai Jiangshan 36522d5f0764SLai Jiangshan /* 36532d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 36542d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 36552d5f0764SLai Jiangshan * it even if we don't use it immediately. 36562d5f0764SLai Jiangshan */ 36572d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 36582d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 36592d5f0764SLai Jiangshan goto out_free; 36602d5f0764SLai Jiangshan 36612d5f0764SLai Jiangshan for_each_node(node) { 3662042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 36632d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 36642d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 36652d5f0764SLai Jiangshan goto out_free; 36662d5f0764SLai Jiangshan } else { 36672d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 36682d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 36692d5f0764SLai Jiangshan } 36702d5f0764SLai Jiangshan } 36712d5f0764SLai Jiangshan 3672042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3673042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3674042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 36752d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3676042f7df1SLai Jiangshan 36772d5f0764SLai Jiangshan ctx->wq = wq; 36782d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36792d5f0764SLai Jiangshan return ctx; 36802d5f0764SLai Jiangshan 36812d5f0764SLai Jiangshan out_free: 36822d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36832d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 36842d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 36852d5f0764SLai Jiangshan return NULL; 36862d5f0764SLai Jiangshan } 36872d5f0764SLai Jiangshan 36882d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 36892d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 36902d5f0764SLai Jiangshan { 36912d5f0764SLai Jiangshan int node; 36922d5f0764SLai Jiangshan 36932d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 36942d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 36952d5f0764SLai Jiangshan 36962d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 36972d5f0764SLai Jiangshan 36982d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 36992d5f0764SLai Jiangshan for_each_node(node) 37002d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 37012d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 37022d5f0764SLai Jiangshan 37032d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 37042d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 37052d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 37062d5f0764SLai Jiangshan 37072d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 37082d5f0764SLai Jiangshan } 37092d5f0764SLai Jiangshan 3710a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3711a0111cf6SLai Jiangshan { 3712a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3713a0111cf6SLai Jiangshan get_online_cpus(); 3714a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3715a0111cf6SLai Jiangshan } 3716a0111cf6SLai Jiangshan 3717a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3718a0111cf6SLai Jiangshan { 3719a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3720a0111cf6SLai Jiangshan put_online_cpus(); 3721a0111cf6SLai Jiangshan } 3722a0111cf6SLai Jiangshan 3723a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3724a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3725a0111cf6SLai Jiangshan { 3726a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3727a0111cf6SLai Jiangshan 3728a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3729a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3730a0111cf6SLai Jiangshan return -EINVAL; 3731a0111cf6SLai Jiangshan 3732a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 3733a0111cf6SLai Jiangshan if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 3734a0111cf6SLai Jiangshan return -EINVAL; 3735a0111cf6SLai Jiangshan 3736a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 37376201171eSwanghaibin if (!ctx) 37386201171eSwanghaibin return -ENOMEM; 3739a0111cf6SLai Jiangshan 3740a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3741a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3742a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3743a0111cf6SLai Jiangshan 37446201171eSwanghaibin return 0; 3745a0111cf6SLai Jiangshan } 3746a0111cf6SLai Jiangshan 37479e8cd2f5STejun Heo /** 37489e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37499e8cd2f5STejun Heo * @wq: the target workqueue 37509e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37519e8cd2f5STejun Heo * 37524c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 37534c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 37544c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 37554c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 37564c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 37574c16bd32STejun Heo * back-to-back will stay on its current pwq. 37589e8cd2f5STejun Heo * 3759d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3760d185af30SYacine Belkadi * 3761d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 37629e8cd2f5STejun Heo */ 37639e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 37649e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 37659e8cd2f5STejun Heo { 3766a0111cf6SLai Jiangshan int ret; 37679e8cd2f5STejun Heo 3768a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3769a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3770a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 37712d5f0764SLai Jiangshan 37722d5f0764SLai Jiangshan return ret; 37739e8cd2f5STejun Heo } 37749e8cd2f5STejun Heo 37754c16bd32STejun Heo /** 37764c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 37774c16bd32STejun Heo * @wq: the target workqueue 37784c16bd32STejun Heo * @cpu: the CPU coming up or going down 37794c16bd32STejun Heo * @online: whether @cpu is coming up or going down 37804c16bd32STejun Heo * 37814c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 37824c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 37834c16bd32STejun Heo * @wq accordingly. 37844c16bd32STejun Heo * 37854c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 37864c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 37874c16bd32STejun Heo * correct. 37884c16bd32STejun Heo * 37894c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 37904c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 37914c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 37924c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 37934c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 37944c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 37954c16bd32STejun Heo * CPU_DOWN_PREPARE. 37964c16bd32STejun Heo */ 37974c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 37984c16bd32STejun Heo bool online) 37994c16bd32STejun Heo { 38004c16bd32STejun Heo int node = cpu_to_node(cpu); 38014c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 38024c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 38034c16bd32STejun Heo struct workqueue_attrs *target_attrs; 38044c16bd32STejun Heo cpumask_t *cpumask; 38054c16bd32STejun Heo 38064c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38074c16bd32STejun Heo 3808f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3809f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 38104c16bd32STejun Heo return; 38114c16bd32STejun Heo 38124c16bd32STejun Heo /* 38134c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 38144c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 38154c16bd32STejun Heo * CPU hotplug exclusion. 38164c16bd32STejun Heo */ 38174c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 38184c16bd32STejun Heo cpumask = target_attrs->cpumask; 38194c16bd32STejun Heo 38204c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 38214c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 38224c16bd32STejun Heo 38234c16bd32STejun Heo /* 38244c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3825042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3826042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3827042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 38284c16bd32STejun Heo */ 3829042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 38304c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3831f7142ed4SLai Jiangshan return; 38324c16bd32STejun Heo } else { 38334c16bd32STejun Heo goto use_dfl_pwq; 38344c16bd32STejun Heo } 38354c16bd32STejun Heo 38364c16bd32STejun Heo /* create a new pwq */ 38374c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 38384c16bd32STejun Heo if (!pwq) { 38392d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 38404c16bd32STejun Heo wq->name); 384177f300b1SDaeseok Youn goto use_dfl_pwq; 38424c16bd32STejun Heo } 38434c16bd32STejun Heo 3844f7142ed4SLai Jiangshan /* Install the new pwq. */ 38454c16bd32STejun Heo mutex_lock(&wq->mutex); 38464c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 38474c16bd32STejun Heo goto out_unlock; 38484c16bd32STejun Heo 38494c16bd32STejun Heo use_dfl_pwq: 3850f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 38514c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 38524c16bd32STejun Heo get_pwq(wq->dfl_pwq); 38534c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 38544c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 38554c16bd32STejun Heo out_unlock: 38564c16bd32STejun Heo mutex_unlock(&wq->mutex); 38574c16bd32STejun Heo put_pwq_unlocked(old_pwq); 38584c16bd32STejun Heo } 38594c16bd32STejun Heo 386030cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 38611da177e4SLinus Torvalds { 386249e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 38638a2b7538STejun Heo int cpu, ret; 3864e1d8aa9fSFrederic Weisbecker 386530cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3866420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3867420c0ddbSTejun Heo if (!wq->cpu_pwqs) 386830cdf249STejun Heo return -ENOMEM; 386930cdf249STejun Heo 387030cdf249STejun Heo for_each_possible_cpu(cpu) { 38717fb98ea7STejun Heo struct pool_workqueue *pwq = 38727fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 38737a62c2c8STejun Heo struct worker_pool *cpu_pools = 3874f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 387530cdf249STejun Heo 3876f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3877f147f29eSTejun Heo 3878f147f29eSTejun Heo mutex_lock(&wq->mutex); 38791befcf30STejun Heo link_pwq(pwq); 3880f147f29eSTejun Heo mutex_unlock(&wq->mutex); 388130cdf249STejun Heo } 388230cdf249STejun Heo return 0; 38838a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 38848a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 38858a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 38868a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 38878a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 38888a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 38898a2b7538STejun Heo return ret; 38909e8cd2f5STejun Heo } else { 38919e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 38929e8cd2f5STejun Heo } 38930f900049STejun Heo } 38940f900049STejun Heo 3895f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3896f3421797STejun Heo const char *name) 3897b71ab8c2STejun Heo { 3898f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3899f3421797STejun Heo 3900f3421797STejun Heo if (max_active < 1 || max_active > lim) 3901044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3902f3421797STejun Heo max_active, name, 1, lim); 3903b71ab8c2STejun Heo 3904f3421797STejun Heo return clamp_val(max_active, 1, lim); 3905b71ab8c2STejun Heo } 3906b71ab8c2STejun Heo 3907b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 390897e37d7bSTejun Heo unsigned int flags, 39091e19ffc6STejun Heo int max_active, 3910eb13ba87SJohannes Berg struct lock_class_key *key, 3911b196be89STejun Heo const char *lock_name, ...) 39123af24433SOleg Nesterov { 3913df2d5ae4STejun Heo size_t tbl_size = 0; 3914ecf6881fSTejun Heo va_list args; 39153af24433SOleg Nesterov struct workqueue_struct *wq; 391649e3cf44STejun Heo struct pool_workqueue *pwq; 3917b196be89STejun Heo 3918cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 3919cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 3920cee22a15SViresh Kumar flags |= WQ_UNBOUND; 3921cee22a15SViresh Kumar 3922ecf6881fSTejun Heo /* allocate wq and format name */ 3923df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 3924ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 3925df2d5ae4STejun Heo 3926df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 3927b196be89STejun Heo if (!wq) 3928d2c1d404STejun Heo return NULL; 3929b196be89STejun Heo 39306029a918STejun Heo if (flags & WQ_UNBOUND) { 39316029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39326029a918STejun Heo if (!wq->unbound_attrs) 39336029a918STejun Heo goto err_free_wq; 39346029a918STejun Heo } 39356029a918STejun Heo 3936ecf6881fSTejun Heo va_start(args, lock_name); 3937ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3938b196be89STejun Heo va_end(args); 39393af24433SOleg Nesterov 3940d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3941b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 39423af24433SOleg Nesterov 3943b196be89STejun Heo /* init wq */ 394497e37d7bSTejun Heo wq->flags = flags; 3945a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 39463c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3947112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 394830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 394973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 395073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3951493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 39523af24433SOleg Nesterov 3953eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3954cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 39553af24433SOleg Nesterov 395630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3957d2c1d404STejun Heo goto err_free_wq; 39581537663fSTejun Heo 3959493008a8STejun Heo /* 3960493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3961493008a8STejun Heo * have a rescuer to guarantee forward progress. 3962493008a8STejun Heo */ 3963493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3964e22bee78STejun Heo struct worker *rescuer; 3965e22bee78STejun Heo 3966f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 3967e22bee78STejun Heo if (!rescuer) 3968d2c1d404STejun Heo goto err_destroy; 3969e22bee78STejun Heo 3970111c225aSTejun Heo rescuer->rescue_wq = wq; 3971111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3972b196be89STejun Heo wq->name); 3973d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3974d2c1d404STejun Heo kfree(rescuer); 3975d2c1d404STejun Heo goto err_destroy; 3976d2c1d404STejun Heo } 3977e22bee78STejun Heo 3978d2c1d404STejun Heo wq->rescuer = rescuer; 397925834c73SPeter Zijlstra kthread_bind_mask(rescuer->task, cpu_possible_mask); 3980e22bee78STejun Heo wake_up_process(rescuer->task); 39813af24433SOleg Nesterov } 39821537663fSTejun Heo 3983226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3984226223abSTejun Heo goto err_destroy; 3985226223abSTejun Heo 39863af24433SOleg Nesterov /* 398768e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 398868e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 398968e13a67SLai Jiangshan * list. 39903af24433SOleg Nesterov */ 399168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3992a0a1a5fdSTejun Heo 3993a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 399449e3cf44STejun Heo for_each_pwq(pwq, wq) 3995699ce097STejun Heo pwq_adjust_max_active(pwq); 3996a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3997a0a1a5fdSTejun Heo 3998e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 3999a0a1a5fdSTejun Heo 400068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 40013af24433SOleg Nesterov 40023af24433SOleg Nesterov return wq; 4003d2c1d404STejun Heo 4004d2c1d404STejun Heo err_free_wq: 40056029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 40064690c4abSTejun Heo kfree(wq); 4007d2c1d404STejun Heo return NULL; 4008d2c1d404STejun Heo err_destroy: 4009d2c1d404STejun Heo destroy_workqueue(wq); 40104690c4abSTejun Heo return NULL; 40111da177e4SLinus Torvalds } 4012d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 40131da177e4SLinus Torvalds 40143af24433SOleg Nesterov /** 40153af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 40163af24433SOleg Nesterov * @wq: target workqueue 40173af24433SOleg Nesterov * 40183af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 40193af24433SOleg Nesterov */ 40203af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 40213af24433SOleg Nesterov { 402249e3cf44STejun Heo struct pool_workqueue *pwq; 40234c16bd32STejun Heo int node; 40243af24433SOleg Nesterov 40259c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 40269c5a2ba7STejun Heo drain_workqueue(wq); 4027c8efcc25STejun Heo 40286183c009STejun Heo /* sanity checks */ 4029b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 403049e3cf44STejun Heo for_each_pwq(pwq, wq) { 40316183c009STejun Heo int i; 40326183c009STejun Heo 403376af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 403476af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4035b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 40366183c009STejun Heo return; 403776af4d93STejun Heo } 403876af4d93STejun Heo } 403976af4d93STejun Heo 40405c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 40418864b4e5STejun Heo WARN_ON(pwq->nr_active) || 404276af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4043b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 40446183c009STejun Heo return; 40456183c009STejun Heo } 404676af4d93STejun Heo } 4047b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 40486183c009STejun Heo 4049a0a1a5fdSTejun Heo /* 4050a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4051a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4052a0a1a5fdSTejun Heo */ 405368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4054e2dca7adSTejun Heo list_del_rcu(&wq->list); 405568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 40563af24433SOleg Nesterov 4057226223abSTejun Heo workqueue_sysfs_unregister(wq); 4058226223abSTejun Heo 4059e2dca7adSTejun Heo if (wq->rescuer) 4060e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4061e22bee78STejun Heo 40628864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 406329c91e99STejun Heo /* 40648864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4065e2dca7adSTejun Heo * schedule RCU free. 406629c91e99STejun Heo */ 4067e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 40688864b4e5STejun Heo } else { 40698864b4e5STejun Heo /* 40708864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 40714c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 40724c16bd32STejun Heo * @wq will be freed when the last pwq is released. 40738864b4e5STejun Heo */ 40744c16bd32STejun Heo for_each_node(node) { 40754c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 40764c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 40774c16bd32STejun Heo put_pwq_unlocked(pwq); 40784c16bd32STejun Heo } 40794c16bd32STejun Heo 40804c16bd32STejun Heo /* 40814c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 40824c16bd32STejun Heo * put. Don't access it afterwards. 40834c16bd32STejun Heo */ 40844c16bd32STejun Heo pwq = wq->dfl_pwq; 40854c16bd32STejun Heo wq->dfl_pwq = NULL; 4086dce90d47STejun Heo put_pwq_unlocked(pwq); 408729c91e99STejun Heo } 40883af24433SOleg Nesterov } 40893af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 40903af24433SOleg Nesterov 4091dcd989cbSTejun Heo /** 4092dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4093dcd989cbSTejun Heo * @wq: target workqueue 4094dcd989cbSTejun Heo * @max_active: new max_active value. 4095dcd989cbSTejun Heo * 4096dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4097dcd989cbSTejun Heo * 4098dcd989cbSTejun Heo * CONTEXT: 4099dcd989cbSTejun Heo * Don't call from IRQ context. 4100dcd989cbSTejun Heo */ 4101dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4102dcd989cbSTejun Heo { 410349e3cf44STejun Heo struct pool_workqueue *pwq; 4104dcd989cbSTejun Heo 41058719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 41068719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 41078719dceaSTejun Heo return; 41088719dceaSTejun Heo 4109f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4110dcd989cbSTejun Heo 4111a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4112dcd989cbSTejun Heo 4113dcd989cbSTejun Heo wq->saved_max_active = max_active; 4114dcd989cbSTejun Heo 4115699ce097STejun Heo for_each_pwq(pwq, wq) 4116699ce097STejun Heo pwq_adjust_max_active(pwq); 4117dcd989cbSTejun Heo 4118a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4119dcd989cbSTejun Heo } 4120dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4121dcd989cbSTejun Heo 4122dcd989cbSTejun Heo /** 4123e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4124e6267616STejun Heo * 4125e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4126e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4127d185af30SYacine Belkadi * 4128d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4129e6267616STejun Heo */ 4130e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4131e6267616STejun Heo { 4132e6267616STejun Heo struct worker *worker = current_wq_worker(); 4133e6267616STejun Heo 41346a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4135e6267616STejun Heo } 4136e6267616STejun Heo 4137e6267616STejun Heo /** 4138dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4139dcd989cbSTejun Heo * @cpu: CPU in question 4140dcd989cbSTejun Heo * @wq: target workqueue 4141dcd989cbSTejun Heo * 4142dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4143dcd989cbSTejun Heo * no synchronization around this function and the test result is 4144dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4145dcd989cbSTejun Heo * 4146d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4147d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4148d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4149d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4150d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4151d3251859STejun Heo * 4152d185af30SYacine Belkadi * Return: 4153dcd989cbSTejun Heo * %true if congested, %false otherwise. 4154dcd989cbSTejun Heo */ 4155d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4156dcd989cbSTejun Heo { 41577fb98ea7STejun Heo struct pool_workqueue *pwq; 415876af4d93STejun Heo bool ret; 415976af4d93STejun Heo 416088109453SLai Jiangshan rcu_read_lock_sched(); 41617fb98ea7STejun Heo 4162d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4163d3251859STejun Heo cpu = smp_processor_id(); 4164d3251859STejun Heo 41657fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 41667fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 41677fb98ea7STejun Heo else 4168df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4169dcd989cbSTejun Heo 417076af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 417188109453SLai Jiangshan rcu_read_unlock_sched(); 417276af4d93STejun Heo 417376af4d93STejun Heo return ret; 4174dcd989cbSTejun Heo } 4175dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4176dcd989cbSTejun Heo 4177dcd989cbSTejun Heo /** 4178dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4179dcd989cbSTejun Heo * @work: the work to be tested 4180dcd989cbSTejun Heo * 4181dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4182dcd989cbSTejun Heo * synchronization around this function and the test result is 4183dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4184dcd989cbSTejun Heo * 4185d185af30SYacine Belkadi * Return: 4186dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4187dcd989cbSTejun Heo */ 4188dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4189dcd989cbSTejun Heo { 4190fa1b54e6STejun Heo struct worker_pool *pool; 4191dcd989cbSTejun Heo unsigned long flags; 4192dcd989cbSTejun Heo unsigned int ret = 0; 4193dcd989cbSTejun Heo 4194dcd989cbSTejun Heo if (work_pending(work)) 4195dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4196038366c5SLai Jiangshan 4197fa1b54e6STejun Heo local_irq_save(flags); 4198fa1b54e6STejun Heo pool = get_work_pool(work); 4199038366c5SLai Jiangshan if (pool) { 4200fa1b54e6STejun Heo spin_lock(&pool->lock); 4201c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4202dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4203fa1b54e6STejun Heo spin_unlock(&pool->lock); 4204038366c5SLai Jiangshan } 4205fa1b54e6STejun Heo local_irq_restore(flags); 4206dcd989cbSTejun Heo 4207dcd989cbSTejun Heo return ret; 4208dcd989cbSTejun Heo } 4209dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4210dcd989cbSTejun Heo 42113d1cb205STejun Heo /** 42123d1cb205STejun Heo * set_worker_desc - set description for the current work item 42133d1cb205STejun Heo * @fmt: printf-style format string 42143d1cb205STejun Heo * @...: arguments for the format string 42153d1cb205STejun Heo * 42163d1cb205STejun Heo * This function can be called by a running work function to describe what 42173d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 42183d1cb205STejun Heo * information will be printed out together to help debugging. The 42193d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 42203d1cb205STejun Heo */ 42213d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 42223d1cb205STejun Heo { 42233d1cb205STejun Heo struct worker *worker = current_wq_worker(); 42243d1cb205STejun Heo va_list args; 42253d1cb205STejun Heo 42263d1cb205STejun Heo if (worker) { 42273d1cb205STejun Heo va_start(args, fmt); 42283d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 42293d1cb205STejun Heo va_end(args); 42303d1cb205STejun Heo worker->desc_valid = true; 42313d1cb205STejun Heo } 42323d1cb205STejun Heo } 42333d1cb205STejun Heo 42343d1cb205STejun Heo /** 42353d1cb205STejun Heo * print_worker_info - print out worker information and description 42363d1cb205STejun Heo * @log_lvl: the log level to use when printing 42373d1cb205STejun Heo * @task: target task 42383d1cb205STejun Heo * 42393d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 42403d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 42413d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 42423d1cb205STejun Heo * 42433d1cb205STejun Heo * This function can be safely called on any task as long as the 42443d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 42453d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 42463d1cb205STejun Heo */ 42473d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 42483d1cb205STejun Heo { 42493d1cb205STejun Heo work_func_t *fn = NULL; 42503d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 42513d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 42523d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 42533d1cb205STejun Heo struct workqueue_struct *wq = NULL; 42543d1cb205STejun Heo bool desc_valid = false; 42553d1cb205STejun Heo struct worker *worker; 42563d1cb205STejun Heo 42573d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 42583d1cb205STejun Heo return; 42593d1cb205STejun Heo 42603d1cb205STejun Heo /* 42613d1cb205STejun Heo * This function is called without any synchronization and @task 42623d1cb205STejun Heo * could be in any state. Be careful with dereferences. 42633d1cb205STejun Heo */ 4264e700591aSPetr Mladek worker = kthread_probe_data(task); 42653d1cb205STejun Heo 42663d1cb205STejun Heo /* 42673d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 42683d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 42693d1cb205STejun Heo */ 42703d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 42713d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 42723d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 42733d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 42743d1cb205STejun Heo 42753d1cb205STejun Heo /* copy worker description */ 42763d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 42773d1cb205STejun Heo if (desc_valid) 42783d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 42793d1cb205STejun Heo 42803d1cb205STejun Heo if (fn || name[0] || desc[0]) { 42813d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 42823d1cb205STejun Heo if (desc[0]) 42833d1cb205STejun Heo pr_cont(" (%s)", desc); 42843d1cb205STejun Heo pr_cont("\n"); 42853d1cb205STejun Heo } 42863d1cb205STejun Heo } 42873d1cb205STejun Heo 42883494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 42893494fc30STejun Heo { 42903494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 42913494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 42923494fc30STejun Heo pr_cont(" node=%d", pool->node); 42933494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 42943494fc30STejun Heo } 42953494fc30STejun Heo 42963494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 42973494fc30STejun Heo { 42983494fc30STejun Heo if (work->func == wq_barrier_func) { 42993494fc30STejun Heo struct wq_barrier *barr; 43003494fc30STejun Heo 43013494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 43023494fc30STejun Heo 43033494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 43043494fc30STejun Heo task_pid_nr(barr->task)); 43053494fc30STejun Heo } else { 43063494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 43073494fc30STejun Heo } 43083494fc30STejun Heo } 43093494fc30STejun Heo 43103494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 43113494fc30STejun Heo { 43123494fc30STejun Heo struct worker_pool *pool = pwq->pool; 43133494fc30STejun Heo struct work_struct *work; 43143494fc30STejun Heo struct worker *worker; 43153494fc30STejun Heo bool has_in_flight = false, has_pending = false; 43163494fc30STejun Heo int bkt; 43173494fc30STejun Heo 43183494fc30STejun Heo pr_info(" pwq %d:", pool->id); 43193494fc30STejun Heo pr_cont_pool_info(pool); 43203494fc30STejun Heo 43213494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 43223494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 43233494fc30STejun Heo 43243494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43253494fc30STejun Heo if (worker->current_pwq == pwq) { 43263494fc30STejun Heo has_in_flight = true; 43273494fc30STejun Heo break; 43283494fc30STejun Heo } 43293494fc30STejun Heo } 43303494fc30STejun Heo if (has_in_flight) { 43313494fc30STejun Heo bool comma = false; 43323494fc30STejun Heo 43333494fc30STejun Heo pr_info(" in-flight:"); 43343494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43353494fc30STejun Heo if (worker->current_pwq != pwq) 43363494fc30STejun Heo continue; 43373494fc30STejun Heo 43383494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 43393494fc30STejun Heo task_pid_nr(worker->task), 43403494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 43413494fc30STejun Heo worker->current_func); 43423494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 43433494fc30STejun Heo pr_cont_work(false, work); 43443494fc30STejun Heo comma = true; 43453494fc30STejun Heo } 43463494fc30STejun Heo pr_cont("\n"); 43473494fc30STejun Heo } 43483494fc30STejun Heo 43493494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 43503494fc30STejun Heo if (get_work_pwq(work) == pwq) { 43513494fc30STejun Heo has_pending = true; 43523494fc30STejun Heo break; 43533494fc30STejun Heo } 43543494fc30STejun Heo } 43553494fc30STejun Heo if (has_pending) { 43563494fc30STejun Heo bool comma = false; 43573494fc30STejun Heo 43583494fc30STejun Heo pr_info(" pending:"); 43593494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 43603494fc30STejun Heo if (get_work_pwq(work) != pwq) 43613494fc30STejun Heo continue; 43623494fc30STejun Heo 43633494fc30STejun Heo pr_cont_work(comma, work); 43643494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 43653494fc30STejun Heo } 43663494fc30STejun Heo pr_cont("\n"); 43673494fc30STejun Heo } 43683494fc30STejun Heo 43693494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 43703494fc30STejun Heo bool comma = false; 43713494fc30STejun Heo 43723494fc30STejun Heo pr_info(" delayed:"); 43733494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 43743494fc30STejun Heo pr_cont_work(comma, work); 43753494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 43763494fc30STejun Heo } 43773494fc30STejun Heo pr_cont("\n"); 43783494fc30STejun Heo } 43793494fc30STejun Heo } 43803494fc30STejun Heo 43813494fc30STejun Heo /** 43823494fc30STejun Heo * show_workqueue_state - dump workqueue state 43833494fc30STejun Heo * 43847b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 43857b776af6SRoger Lu * all busy workqueues and pools. 43863494fc30STejun Heo */ 43873494fc30STejun Heo void show_workqueue_state(void) 43883494fc30STejun Heo { 43893494fc30STejun Heo struct workqueue_struct *wq; 43903494fc30STejun Heo struct worker_pool *pool; 43913494fc30STejun Heo unsigned long flags; 43923494fc30STejun Heo int pi; 43933494fc30STejun Heo 43943494fc30STejun Heo rcu_read_lock_sched(); 43953494fc30STejun Heo 43963494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 43973494fc30STejun Heo 43983494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 43993494fc30STejun Heo struct pool_workqueue *pwq; 44003494fc30STejun Heo bool idle = true; 44013494fc30STejun Heo 44023494fc30STejun Heo for_each_pwq(pwq, wq) { 44033494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 44043494fc30STejun Heo idle = false; 44053494fc30STejun Heo break; 44063494fc30STejun Heo } 44073494fc30STejun Heo } 44083494fc30STejun Heo if (idle) 44093494fc30STejun Heo continue; 44103494fc30STejun Heo 44113494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 44123494fc30STejun Heo 44133494fc30STejun Heo for_each_pwq(pwq, wq) { 44143494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 44153494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 44163494fc30STejun Heo show_pwq(pwq); 44173494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 44183494fc30STejun Heo } 44193494fc30STejun Heo } 44203494fc30STejun Heo 44213494fc30STejun Heo for_each_pool(pool, pi) { 44223494fc30STejun Heo struct worker *worker; 44233494fc30STejun Heo bool first = true; 44243494fc30STejun Heo 44253494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 44263494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 44273494fc30STejun Heo goto next_pool; 44283494fc30STejun Heo 44293494fc30STejun Heo pr_info("pool %d:", pool->id); 44303494fc30STejun Heo pr_cont_pool_info(pool); 443182607adcSTejun Heo pr_cont(" hung=%us workers=%d", 443282607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 443382607adcSTejun Heo pool->nr_workers); 44343494fc30STejun Heo if (pool->manager) 44353494fc30STejun Heo pr_cont(" manager: %d", 44363494fc30STejun Heo task_pid_nr(pool->manager->task)); 44373494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 44383494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 44393494fc30STejun Heo task_pid_nr(worker->task)); 44403494fc30STejun Heo first = false; 44413494fc30STejun Heo } 44423494fc30STejun Heo pr_cont("\n"); 44433494fc30STejun Heo next_pool: 44443494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 44453494fc30STejun Heo } 44463494fc30STejun Heo 44473494fc30STejun Heo rcu_read_unlock_sched(); 44483494fc30STejun Heo } 44493494fc30STejun Heo 4450db7bccf4STejun Heo /* 4451db7bccf4STejun Heo * CPU hotplug. 4452db7bccf4STejun Heo * 4453e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4454112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4455706026c2STejun Heo * pool which make migrating pending and scheduled works very 4456e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 445794cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4458e22bee78STejun Heo * blocked draining impractical. 4459e22bee78STejun Heo * 446024647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4461628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4462628c78e7STejun Heo * cpu comes back online. 4463db7bccf4STejun Heo */ 4464db7bccf4STejun Heo 4465706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4466db7bccf4STejun Heo { 446738db41d9STejun Heo int cpu = smp_processor_id(); 44684ce62e9eSTejun Heo struct worker_pool *pool; 4469db7bccf4STejun Heo struct worker *worker; 4470db7bccf4STejun Heo 4471f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 447292f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 447394cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4474e22bee78STejun Heo 4475f2d5a0eeSTejun Heo /* 447692f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 447794cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 447894cf58bbSTejun Heo * except for the ones which are still executing works from 447994cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 448094cf58bbSTejun Heo * this, they may become diasporas. 4481f2d5a0eeSTejun Heo */ 4482da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4483403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4484db7bccf4STejun Heo 448524647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4486f2d5a0eeSTejun Heo 448794cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 448892f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4489e22bee78STejun Heo 4490e22bee78STejun Heo /* 4491eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4492eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4493eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4494eb283428SLai Jiangshan * from other cpus. 4495628c78e7STejun Heo */ 4496628c78e7STejun Heo schedule(); 4497628c78e7STejun Heo 4498628c78e7STejun Heo /* 4499eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4500eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4501eb283428SLai Jiangshan * and keep_working() are always true as long as the 4502eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4503eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4504eb283428SLai Jiangshan * are served by workers tied to the pool. 4505e22bee78STejun Heo */ 4506e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4507eb283428SLai Jiangshan 4508eb283428SLai Jiangshan /* 4509eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4510eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4511eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4512eb283428SLai Jiangshan */ 4513eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4514eb283428SLai Jiangshan wake_up_worker(pool); 4515eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4516eb283428SLai Jiangshan } 4517db7bccf4STejun Heo } 4518db7bccf4STejun Heo 4519bd7c089eSTejun Heo /** 4520bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4521bd7c089eSTejun Heo * @pool: pool of interest 4522bd7c089eSTejun Heo * 4523a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4524bd7c089eSTejun Heo */ 4525bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4526bd7c089eSTejun Heo { 4527a9ab775bSTejun Heo struct worker *worker; 4528bd7c089eSTejun Heo 452992f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4530bd7c089eSTejun Heo 4531bd7c089eSTejun Heo /* 4532a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4533a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4534402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4535a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4536a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4537bd7c089eSTejun Heo */ 4538da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4539a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4540a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4541a9ab775bSTejun Heo 4542a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4543f7c17d26SWanpeng Li 4544f7c17d26SWanpeng Li /* 4545f7c17d26SWanpeng Li * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED 4546f7c17d26SWanpeng Li * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is 4547f7c17d26SWanpeng Li * being reworked and this can go away in time. 4548f7c17d26SWanpeng Li */ 4549f7c17d26SWanpeng Li if (!(pool->flags & POOL_DISASSOCIATED)) { 4550f7c17d26SWanpeng Li spin_unlock_irq(&pool->lock); 4551f7c17d26SWanpeng Li return; 4552f7c17d26SWanpeng Li } 4553f7c17d26SWanpeng Li 45543de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4555a9ab775bSTejun Heo 4556da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4557a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4558a9ab775bSTejun Heo 4559a9ab775bSTejun Heo /* 4560a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4561a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4562a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4563a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4564a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4565a9ab775bSTejun Heo * be bound before @pool->lock is released. 4566a9ab775bSTejun Heo */ 4567a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4568bd7c089eSTejun Heo wake_up_process(worker->task); 4569bd7c089eSTejun Heo 4570bd7c089eSTejun Heo /* 4571a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4572a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4573a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4574a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4575a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4576a9ab775bSTejun Heo * concurrency management. Note that when or whether 4577a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4578a9ab775bSTejun Heo * 4579a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4580a9ab775bSTejun Heo * tested without holding any lock in 4581a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4582a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4583a9ab775bSTejun Heo * management operations. 4584bd7c089eSTejun Heo */ 4585a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4586a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4587a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4588a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4589bd7c089eSTejun Heo } 4590a9ab775bSTejun Heo 4591a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4592bd7c089eSTejun Heo } 4593bd7c089eSTejun Heo 45947dbc725eSTejun Heo /** 45957dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 45967dbc725eSTejun Heo * @pool: unbound pool of interest 45977dbc725eSTejun Heo * @cpu: the CPU which is coming up 45987dbc725eSTejun Heo * 45997dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 46007dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 46017dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 46027dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 46037dbc725eSTejun Heo */ 46047dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 46057dbc725eSTejun Heo { 46067dbc725eSTejun Heo static cpumask_t cpumask; 46077dbc725eSTejun Heo struct worker *worker; 46087dbc725eSTejun Heo 460992f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 46107dbc725eSTejun Heo 46117dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 46127dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 46137dbc725eSTejun Heo return; 46147dbc725eSTejun Heo 46157dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 46167dbc725eSTejun Heo 46177dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4618da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4619d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 46207dbc725eSTejun Heo } 46217dbc725eSTejun Heo 46227ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 46231da177e4SLinus Torvalds { 46244ce62e9eSTejun Heo struct worker_pool *pool; 46251da177e4SLinus Torvalds 4626f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46273ce63377STejun Heo if (pool->nr_workers) 46283ce63377STejun Heo continue; 4629051e1850SLai Jiangshan if (!create_worker(pool)) 46307ee681b2SThomas Gleixner return -ENOMEM; 46313af24433SOleg Nesterov } 46327ee681b2SThomas Gleixner return 0; 46337ee681b2SThomas Gleixner } 46341da177e4SLinus Torvalds 46357ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 46367ee681b2SThomas Gleixner { 46377ee681b2SThomas Gleixner struct worker_pool *pool; 46387ee681b2SThomas Gleixner struct workqueue_struct *wq; 46397ee681b2SThomas Gleixner int pi; 46407ee681b2SThomas Gleixner 464168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46427dbc725eSTejun Heo 46437dbc725eSTejun Heo for_each_pool(pool, pi) { 464492f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 464594cf58bbSTejun Heo 4646f05b558dSLai Jiangshan if (pool->cpu == cpu) 464794cf58bbSTejun Heo rebind_workers(pool); 4648f05b558dSLai Jiangshan else if (pool->cpu < 0) 46497dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 465094cf58bbSTejun Heo 465192f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 465294cf58bbSTejun Heo } 46537dbc725eSTejun Heo 46544c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46554c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46564c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 46574c16bd32STejun Heo 465868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46597ee681b2SThomas Gleixner return 0; 466065758202STejun Heo } 466165758202STejun Heo 46627ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 466365758202STejun Heo { 46648db25e78STejun Heo struct work_struct unbind_work; 46654c16bd32STejun Heo struct workqueue_struct *wq; 46668db25e78STejun Heo 46674c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4668706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46697635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46704c16bd32STejun Heo 46714c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46724c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46734c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46744c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 46754c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 46764c16bd32STejun Heo 46774c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 46788db25e78STejun Heo flush_work(&unbind_work); 4679440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 46807ee681b2SThomas Gleixner return 0; 468165758202STejun Heo } 468265758202STejun Heo 46832d3854a3SRusty Russell #ifdef CONFIG_SMP 46848ccad40dSRusty Russell 46852d3854a3SRusty Russell struct work_for_cpu { 4686ed48ece2STejun Heo struct work_struct work; 46872d3854a3SRusty Russell long (*fn)(void *); 46882d3854a3SRusty Russell void *arg; 46892d3854a3SRusty Russell long ret; 46902d3854a3SRusty Russell }; 46912d3854a3SRusty Russell 4692ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 46932d3854a3SRusty Russell { 4694ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4695ed48ece2STejun Heo 46962d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 46972d3854a3SRusty Russell } 46982d3854a3SRusty Russell 46992d3854a3SRusty Russell /** 470022aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 47012d3854a3SRusty Russell * @cpu: the cpu to run on 47022d3854a3SRusty Russell * @fn: the function to run 47032d3854a3SRusty Russell * @arg: the function arg 47042d3854a3SRusty Russell * 470531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47066b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4707d185af30SYacine Belkadi * 4708d185af30SYacine Belkadi * Return: The value @fn returns. 47092d3854a3SRusty Russell */ 4710d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47112d3854a3SRusty Russell { 4712ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47132d3854a3SRusty Russell 4714ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4715ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 471612997d1aSBjorn Helgaas flush_work(&wfc.work); 4717440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 47182d3854a3SRusty Russell return wfc.ret; 47192d3854a3SRusty Russell } 47202d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47212d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47222d3854a3SRusty Russell 4723a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4724e7577c50SRusty Russell 4725a0a1a5fdSTejun Heo /** 4726a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4727a0a1a5fdSTejun Heo * 472858a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4729c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4730706026c2STejun Heo * pool->worklist. 4731a0a1a5fdSTejun Heo * 4732a0a1a5fdSTejun Heo * CONTEXT: 4733a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4734a0a1a5fdSTejun Heo */ 4735a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4736a0a1a5fdSTejun Heo { 473724b8a847STejun Heo struct workqueue_struct *wq; 473824b8a847STejun Heo struct pool_workqueue *pwq; 4739a0a1a5fdSTejun Heo 474068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4741a0a1a5fdSTejun Heo 47426183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4743a0a1a5fdSTejun Heo workqueue_freezing = true; 4744a0a1a5fdSTejun Heo 474524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4746a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4747699ce097STejun Heo for_each_pwq(pwq, wq) 4748699ce097STejun Heo pwq_adjust_max_active(pwq); 4749a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4750a1056305STejun Heo } 47515bcab335STejun Heo 475268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4753a0a1a5fdSTejun Heo } 4754a0a1a5fdSTejun Heo 4755a0a1a5fdSTejun Heo /** 475658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4757a0a1a5fdSTejun Heo * 4758a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4759a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4760a0a1a5fdSTejun Heo * 4761a0a1a5fdSTejun Heo * CONTEXT: 476268e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4763a0a1a5fdSTejun Heo * 4764d185af30SYacine Belkadi * Return: 476558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 476658a69cb4STejun Heo * is complete. 4767a0a1a5fdSTejun Heo */ 4768a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4769a0a1a5fdSTejun Heo { 4770a0a1a5fdSTejun Heo bool busy = false; 477124b8a847STejun Heo struct workqueue_struct *wq; 477224b8a847STejun Heo struct pool_workqueue *pwq; 4773a0a1a5fdSTejun Heo 477468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4775a0a1a5fdSTejun Heo 47766183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4777a0a1a5fdSTejun Heo 477824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 477924b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 478024b8a847STejun Heo continue; 4781a0a1a5fdSTejun Heo /* 4782a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4783a0a1a5fdSTejun Heo * to peek without lock. 4784a0a1a5fdSTejun Heo */ 478588109453SLai Jiangshan rcu_read_lock_sched(); 478624b8a847STejun Heo for_each_pwq(pwq, wq) { 47876183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4788112202d9STejun Heo if (pwq->nr_active) { 4789a0a1a5fdSTejun Heo busy = true; 479088109453SLai Jiangshan rcu_read_unlock_sched(); 4791a0a1a5fdSTejun Heo goto out_unlock; 4792a0a1a5fdSTejun Heo } 4793a0a1a5fdSTejun Heo } 479488109453SLai Jiangshan rcu_read_unlock_sched(); 4795a0a1a5fdSTejun Heo } 4796a0a1a5fdSTejun Heo out_unlock: 479768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4798a0a1a5fdSTejun Heo return busy; 4799a0a1a5fdSTejun Heo } 4800a0a1a5fdSTejun Heo 4801a0a1a5fdSTejun Heo /** 4802a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4803a0a1a5fdSTejun Heo * 4804a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4805706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4806a0a1a5fdSTejun Heo * 4807a0a1a5fdSTejun Heo * CONTEXT: 4808a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4809a0a1a5fdSTejun Heo */ 4810a0a1a5fdSTejun Heo void thaw_workqueues(void) 4811a0a1a5fdSTejun Heo { 481224b8a847STejun Heo struct workqueue_struct *wq; 481324b8a847STejun Heo struct pool_workqueue *pwq; 4814a0a1a5fdSTejun Heo 481568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4816a0a1a5fdSTejun Heo 4817a0a1a5fdSTejun Heo if (!workqueue_freezing) 4818a0a1a5fdSTejun Heo goto out_unlock; 4819a0a1a5fdSTejun Heo 482074b414eaSLai Jiangshan workqueue_freezing = false; 482124b8a847STejun Heo 482224b8a847STejun Heo /* restore max_active and repopulate worklist */ 482324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4824a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4825699ce097STejun Heo for_each_pwq(pwq, wq) 4826699ce097STejun Heo pwq_adjust_max_active(pwq); 4827a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 482824b8a847STejun Heo } 482924b8a847STejun Heo 4830a0a1a5fdSTejun Heo out_unlock: 483168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4832a0a1a5fdSTejun Heo } 4833a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4834a0a1a5fdSTejun Heo 4835042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4836042f7df1SLai Jiangshan { 4837042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4838042f7df1SLai Jiangshan int ret = 0; 4839042f7df1SLai Jiangshan struct workqueue_struct *wq; 4840042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4841042f7df1SLai Jiangshan 4842042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4843042f7df1SLai Jiangshan 4844042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4845042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4846042f7df1SLai Jiangshan continue; 4847042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 4848042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 4849042f7df1SLai Jiangshan continue; 4850042f7df1SLai Jiangshan 4851042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 4852042f7df1SLai Jiangshan if (!ctx) { 4853042f7df1SLai Jiangshan ret = -ENOMEM; 4854042f7df1SLai Jiangshan break; 4855042f7df1SLai Jiangshan } 4856042f7df1SLai Jiangshan 4857042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 4858042f7df1SLai Jiangshan } 4859042f7df1SLai Jiangshan 4860042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 4861042f7df1SLai Jiangshan if (!ret) 4862042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 4863042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 4864042f7df1SLai Jiangshan } 4865042f7df1SLai Jiangshan 4866042f7df1SLai Jiangshan return ret; 4867042f7df1SLai Jiangshan } 4868042f7df1SLai Jiangshan 4869042f7df1SLai Jiangshan /** 4870042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 4871042f7df1SLai Jiangshan * @cpumask: the cpumask to set 4872042f7df1SLai Jiangshan * 4873042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 4874042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 4875042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 4876042f7df1SLai Jiangshan * 4877042f7df1SLai Jiangshan * Retun: 0 - Success 4878042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 4879042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 4880042f7df1SLai Jiangshan */ 4881042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 4882042f7df1SLai Jiangshan { 4883042f7df1SLai Jiangshan int ret = -EINVAL; 4884042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 4885042f7df1SLai Jiangshan 4886042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 4887042f7df1SLai Jiangshan return -ENOMEM; 4888042f7df1SLai Jiangshan 4889042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 4890042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 4891a0111cf6SLai Jiangshan apply_wqattrs_lock(); 4892042f7df1SLai Jiangshan 4893042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 4894042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 4895042f7df1SLai Jiangshan 4896042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 4897042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 4898042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 4899042f7df1SLai Jiangshan 4900042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 4901042f7df1SLai Jiangshan if (ret < 0) 4902042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 4903042f7df1SLai Jiangshan 4904a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 4905042f7df1SLai Jiangshan } 4906042f7df1SLai Jiangshan 4907042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 4908042f7df1SLai Jiangshan return ret; 4909042f7df1SLai Jiangshan } 4910042f7df1SLai Jiangshan 49116ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 49126ba94429SFrederic Weisbecker /* 49136ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 49146ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 49156ba94429SFrederic Weisbecker * following attributes. 49166ba94429SFrederic Weisbecker * 49176ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 49186ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 49196ba94429SFrederic Weisbecker * 49206ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 49216ba94429SFrederic Weisbecker * 49226ba94429SFrederic Weisbecker * id RO int : the associated pool ID 49236ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 49246ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 49256ba94429SFrederic Weisbecker */ 49266ba94429SFrederic Weisbecker struct wq_device { 49276ba94429SFrederic Weisbecker struct workqueue_struct *wq; 49286ba94429SFrederic Weisbecker struct device dev; 49296ba94429SFrederic Weisbecker }; 49306ba94429SFrederic Weisbecker 49316ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 49326ba94429SFrederic Weisbecker { 49336ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 49346ba94429SFrederic Weisbecker 49356ba94429SFrederic Weisbecker return wq_dev->wq; 49366ba94429SFrederic Weisbecker } 49376ba94429SFrederic Weisbecker 49386ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 49396ba94429SFrederic Weisbecker char *buf) 49406ba94429SFrederic Weisbecker { 49416ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49426ba94429SFrederic Weisbecker 49436ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 49446ba94429SFrederic Weisbecker } 49456ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 49466ba94429SFrederic Weisbecker 49476ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 49486ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49496ba94429SFrederic Weisbecker { 49506ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49516ba94429SFrederic Weisbecker 49526ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 49536ba94429SFrederic Weisbecker } 49546ba94429SFrederic Weisbecker 49556ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 49566ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 49576ba94429SFrederic Weisbecker size_t count) 49586ba94429SFrederic Weisbecker { 49596ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49606ba94429SFrederic Weisbecker int val; 49616ba94429SFrederic Weisbecker 49626ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 49636ba94429SFrederic Weisbecker return -EINVAL; 49646ba94429SFrederic Weisbecker 49656ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 49666ba94429SFrederic Weisbecker return count; 49676ba94429SFrederic Weisbecker } 49686ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 49696ba94429SFrederic Weisbecker 49706ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 49716ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 49726ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 49736ba94429SFrederic Weisbecker NULL, 49746ba94429SFrederic Weisbecker }; 49756ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 49766ba94429SFrederic Weisbecker 49776ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 49786ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49796ba94429SFrederic Weisbecker { 49806ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49816ba94429SFrederic Weisbecker const char *delim = ""; 49826ba94429SFrederic Weisbecker int node, written = 0; 49836ba94429SFrederic Weisbecker 49846ba94429SFrederic Weisbecker rcu_read_lock_sched(); 49856ba94429SFrederic Weisbecker for_each_node(node) { 49866ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 49876ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 49886ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 49896ba94429SFrederic Weisbecker delim = " "; 49906ba94429SFrederic Weisbecker } 49916ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 49926ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 49936ba94429SFrederic Weisbecker 49946ba94429SFrederic Weisbecker return written; 49956ba94429SFrederic Weisbecker } 49966ba94429SFrederic Weisbecker 49976ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 49986ba94429SFrederic Weisbecker char *buf) 49996ba94429SFrederic Weisbecker { 50006ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50016ba94429SFrederic Weisbecker int written; 50026ba94429SFrederic Weisbecker 50036ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50046ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 50056ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50066ba94429SFrederic Weisbecker 50076ba94429SFrederic Weisbecker return written; 50086ba94429SFrederic Weisbecker } 50096ba94429SFrederic Weisbecker 50106ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 50116ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 50126ba94429SFrederic Weisbecker { 50136ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 50146ba94429SFrederic Weisbecker 5015899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5016899a94feSLai Jiangshan 50176ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 50186ba94429SFrederic Weisbecker if (!attrs) 50196ba94429SFrederic Weisbecker return NULL; 50206ba94429SFrederic Weisbecker 50216ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 50226ba94429SFrederic Weisbecker return attrs; 50236ba94429SFrederic Weisbecker } 50246ba94429SFrederic Weisbecker 50256ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 50266ba94429SFrederic Weisbecker const char *buf, size_t count) 50276ba94429SFrederic Weisbecker { 50286ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50296ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5030d4d3e257SLai Jiangshan int ret = -ENOMEM; 5031d4d3e257SLai Jiangshan 5032d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50336ba94429SFrederic Weisbecker 50346ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50356ba94429SFrederic Weisbecker if (!attrs) 5036d4d3e257SLai Jiangshan goto out_unlock; 50376ba94429SFrederic Weisbecker 50386ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 50396ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5040d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50416ba94429SFrederic Weisbecker else 50426ba94429SFrederic Weisbecker ret = -EINVAL; 50436ba94429SFrederic Weisbecker 5044d4d3e257SLai Jiangshan out_unlock: 5045d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50466ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50476ba94429SFrederic Weisbecker return ret ?: count; 50486ba94429SFrederic Weisbecker } 50496ba94429SFrederic Weisbecker 50506ba94429SFrederic Weisbecker static ssize_t wq_cpumask_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 int written; 50556ba94429SFrederic Weisbecker 50566ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50576ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 50586ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 50596ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50606ba94429SFrederic Weisbecker return written; 50616ba94429SFrederic Weisbecker } 50626ba94429SFrederic Weisbecker 50636ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 50646ba94429SFrederic Weisbecker struct device_attribute *attr, 50656ba94429SFrederic Weisbecker const char *buf, size_t count) 50666ba94429SFrederic Weisbecker { 50676ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50686ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5069d4d3e257SLai Jiangshan int ret = -ENOMEM; 5070d4d3e257SLai Jiangshan 5071d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50726ba94429SFrederic Weisbecker 50736ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50746ba94429SFrederic Weisbecker if (!attrs) 5075d4d3e257SLai Jiangshan goto out_unlock; 50766ba94429SFrederic Weisbecker 50776ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 50786ba94429SFrederic Weisbecker if (!ret) 5079d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50806ba94429SFrederic Weisbecker 5081d4d3e257SLai Jiangshan out_unlock: 5082d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50836ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50846ba94429SFrederic Weisbecker return ret ?: count; 50856ba94429SFrederic Weisbecker } 50866ba94429SFrederic Weisbecker 50876ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 50886ba94429SFrederic Weisbecker char *buf) 50896ba94429SFrederic Weisbecker { 50906ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50916ba94429SFrederic Weisbecker int written; 50926ba94429SFrederic Weisbecker 50936ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50946ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 50956ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 50966ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50976ba94429SFrederic Weisbecker 50986ba94429SFrederic Weisbecker return written; 50996ba94429SFrederic Weisbecker } 51006ba94429SFrederic Weisbecker 51016ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 51026ba94429SFrederic Weisbecker const char *buf, size_t count) 51036ba94429SFrederic Weisbecker { 51046ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51056ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5106d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5107d4d3e257SLai Jiangshan 5108d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51096ba94429SFrederic Weisbecker 51106ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51116ba94429SFrederic Weisbecker if (!attrs) 5112d4d3e257SLai Jiangshan goto out_unlock; 51136ba94429SFrederic Weisbecker 51146ba94429SFrederic Weisbecker ret = -EINVAL; 51156ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 51166ba94429SFrederic Weisbecker attrs->no_numa = !v; 5117d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51186ba94429SFrederic Weisbecker } 51196ba94429SFrederic Weisbecker 5120d4d3e257SLai Jiangshan out_unlock: 5121d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 51226ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 51236ba94429SFrederic Weisbecker return ret ?: count; 51246ba94429SFrederic Weisbecker } 51256ba94429SFrederic Weisbecker 51266ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 51276ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 51286ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 51296ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 51306ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 51316ba94429SFrederic Weisbecker __ATTR_NULL, 51326ba94429SFrederic Weisbecker }; 51336ba94429SFrederic Weisbecker 51346ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 51356ba94429SFrederic Weisbecker .name = "workqueue", 51366ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 51376ba94429SFrederic Weisbecker }; 51386ba94429SFrederic Weisbecker 5139b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5140b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5141b05a7928SFrederic Weisbecker { 5142b05a7928SFrederic Weisbecker int written; 5143b05a7928SFrederic Weisbecker 5144042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5145b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5146b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5147042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5148b05a7928SFrederic Weisbecker 5149b05a7928SFrederic Weisbecker return written; 5150b05a7928SFrederic Weisbecker } 5151b05a7928SFrederic Weisbecker 5152042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5153042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5154042f7df1SLai Jiangshan { 5155042f7df1SLai Jiangshan cpumask_var_t cpumask; 5156042f7df1SLai Jiangshan int ret; 5157042f7df1SLai Jiangshan 5158042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5159042f7df1SLai Jiangshan return -ENOMEM; 5160042f7df1SLai Jiangshan 5161042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5162042f7df1SLai Jiangshan if (!ret) 5163042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5164042f7df1SLai Jiangshan 5165042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5166042f7df1SLai Jiangshan return ret ? ret : count; 5167042f7df1SLai Jiangshan } 5168042f7df1SLai Jiangshan 5169b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5170042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5171042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5172b05a7928SFrederic Weisbecker 51736ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 51746ba94429SFrederic Weisbecker { 5175b05a7928SFrederic Weisbecker int err; 5176b05a7928SFrederic Weisbecker 5177b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5178b05a7928SFrederic Weisbecker if (err) 5179b05a7928SFrederic Weisbecker return err; 5180b05a7928SFrederic Weisbecker 5181b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 51826ba94429SFrederic Weisbecker } 51836ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 51846ba94429SFrederic Weisbecker 51856ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 51866ba94429SFrederic Weisbecker { 51876ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 51886ba94429SFrederic Weisbecker 51896ba94429SFrederic Weisbecker kfree(wq_dev); 51906ba94429SFrederic Weisbecker } 51916ba94429SFrederic Weisbecker 51926ba94429SFrederic Weisbecker /** 51936ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 51946ba94429SFrederic Weisbecker * @wq: the workqueue to register 51956ba94429SFrederic Weisbecker * 51966ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 51976ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 51986ba94429SFrederic Weisbecker * which is the preferred method. 51996ba94429SFrederic Weisbecker * 52006ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 52016ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 52026ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 52036ba94429SFrederic Weisbecker * attributes. 52046ba94429SFrederic Weisbecker * 52056ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 52066ba94429SFrederic Weisbecker */ 52076ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 52086ba94429SFrederic Weisbecker { 52096ba94429SFrederic Weisbecker struct wq_device *wq_dev; 52106ba94429SFrederic Weisbecker int ret; 52116ba94429SFrederic Weisbecker 52126ba94429SFrederic Weisbecker /* 5213402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 52146ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 52156ba94429SFrederic Weisbecker * workqueues. 52166ba94429SFrederic Weisbecker */ 52176ba94429SFrederic Weisbecker if (WARN_ON(wq->flags & __WQ_ORDERED)) 52186ba94429SFrederic Weisbecker return -EINVAL; 52196ba94429SFrederic Weisbecker 52206ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 52216ba94429SFrederic Weisbecker if (!wq_dev) 52226ba94429SFrederic Weisbecker return -ENOMEM; 52236ba94429SFrederic Weisbecker 52246ba94429SFrederic Weisbecker wq_dev->wq = wq; 52256ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 52266ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 522723217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 52286ba94429SFrederic Weisbecker 52296ba94429SFrederic Weisbecker /* 52306ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 52316ba94429SFrederic Weisbecker * everything is ready. 52326ba94429SFrederic Weisbecker */ 52336ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 52346ba94429SFrederic Weisbecker 52356ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 52366ba94429SFrederic Weisbecker if (ret) { 52376ba94429SFrederic Weisbecker kfree(wq_dev); 52386ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52396ba94429SFrederic Weisbecker return ret; 52406ba94429SFrederic Weisbecker } 52416ba94429SFrederic Weisbecker 52426ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 52436ba94429SFrederic Weisbecker struct device_attribute *attr; 52446ba94429SFrederic Weisbecker 52456ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 52466ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 52476ba94429SFrederic Weisbecker if (ret) { 52486ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 52496ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52506ba94429SFrederic Weisbecker return ret; 52516ba94429SFrederic Weisbecker } 52526ba94429SFrederic Weisbecker } 52536ba94429SFrederic Weisbecker } 52546ba94429SFrederic Weisbecker 52556ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 52566ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 52576ba94429SFrederic Weisbecker return 0; 52586ba94429SFrederic Weisbecker } 52596ba94429SFrederic Weisbecker 52606ba94429SFrederic Weisbecker /** 52616ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 52626ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 52636ba94429SFrederic Weisbecker * 52646ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 52656ba94429SFrederic Weisbecker */ 52666ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 52676ba94429SFrederic Weisbecker { 52686ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 52696ba94429SFrederic Weisbecker 52706ba94429SFrederic Weisbecker if (!wq->wq_dev) 52716ba94429SFrederic Weisbecker return; 52726ba94429SFrederic Weisbecker 52736ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52746ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 52756ba94429SFrederic Weisbecker } 52766ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 52776ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 52786ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 52796ba94429SFrederic Weisbecker 528082607adcSTejun Heo /* 528182607adcSTejun Heo * Workqueue watchdog. 528282607adcSTejun Heo * 528382607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 528482607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 528582607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 528682607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 528782607adcSTejun Heo * largely opaque. 528882607adcSTejun Heo * 528982607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 529082607adcSTejun Heo * state if some pools failed to make forward progress for a while where 529182607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 529282607adcSTejun Heo * 529382607adcSTejun Heo * This mechanism is controlled through the kernel parameter 529482607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 529582607adcSTejun Heo * corresponding sysfs parameter file. 529682607adcSTejun Heo */ 529782607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 529882607adcSTejun Heo 529982607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data); 530082607adcSTejun Heo 530182607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 530282607adcSTejun Heo static struct timer_list wq_watchdog_timer = 530382607adcSTejun Heo TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0); 530482607adcSTejun Heo 530582607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 530682607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 530782607adcSTejun Heo 530882607adcSTejun Heo static void wq_watchdog_reset_touched(void) 530982607adcSTejun Heo { 531082607adcSTejun Heo int cpu; 531182607adcSTejun Heo 531282607adcSTejun Heo wq_watchdog_touched = jiffies; 531382607adcSTejun Heo for_each_possible_cpu(cpu) 531482607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 531582607adcSTejun Heo } 531682607adcSTejun Heo 531782607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data) 531882607adcSTejun Heo { 531982607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 532082607adcSTejun Heo bool lockup_detected = false; 532182607adcSTejun Heo struct worker_pool *pool; 532282607adcSTejun Heo int pi; 532382607adcSTejun Heo 532482607adcSTejun Heo if (!thresh) 532582607adcSTejun Heo return; 532682607adcSTejun Heo 532782607adcSTejun Heo rcu_read_lock(); 532882607adcSTejun Heo 532982607adcSTejun Heo for_each_pool(pool, pi) { 533082607adcSTejun Heo unsigned long pool_ts, touched, ts; 533182607adcSTejun Heo 533282607adcSTejun Heo if (list_empty(&pool->worklist)) 533382607adcSTejun Heo continue; 533482607adcSTejun Heo 533582607adcSTejun Heo /* get the latest of pool and touched timestamps */ 533682607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 533782607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 533882607adcSTejun Heo 533982607adcSTejun Heo if (time_after(pool_ts, touched)) 534082607adcSTejun Heo ts = pool_ts; 534182607adcSTejun Heo else 534282607adcSTejun Heo ts = touched; 534382607adcSTejun Heo 534482607adcSTejun Heo if (pool->cpu >= 0) { 534582607adcSTejun Heo unsigned long cpu_touched = 534682607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 534782607adcSTejun Heo pool->cpu)); 534882607adcSTejun Heo if (time_after(cpu_touched, ts)) 534982607adcSTejun Heo ts = cpu_touched; 535082607adcSTejun Heo } 535182607adcSTejun Heo 535282607adcSTejun Heo /* did we stall? */ 535382607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 535482607adcSTejun Heo lockup_detected = true; 535582607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 535682607adcSTejun Heo pr_cont_pool_info(pool); 535782607adcSTejun Heo pr_cont(" stuck for %us!\n", 535882607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 535982607adcSTejun Heo } 536082607adcSTejun Heo } 536182607adcSTejun Heo 536282607adcSTejun Heo rcu_read_unlock(); 536382607adcSTejun Heo 536482607adcSTejun Heo if (lockup_detected) 536582607adcSTejun Heo show_workqueue_state(); 536682607adcSTejun Heo 536782607adcSTejun Heo wq_watchdog_reset_touched(); 536882607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 536982607adcSTejun Heo } 537082607adcSTejun Heo 537182607adcSTejun Heo void wq_watchdog_touch(int cpu) 537282607adcSTejun Heo { 537382607adcSTejun Heo if (cpu >= 0) 537482607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 537582607adcSTejun Heo else 537682607adcSTejun Heo wq_watchdog_touched = jiffies; 537782607adcSTejun Heo } 537882607adcSTejun Heo 537982607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 538082607adcSTejun Heo { 538182607adcSTejun Heo wq_watchdog_thresh = 0; 538282607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 538382607adcSTejun Heo 538482607adcSTejun Heo if (thresh) { 538582607adcSTejun Heo wq_watchdog_thresh = thresh; 538682607adcSTejun Heo wq_watchdog_reset_touched(); 538782607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 538882607adcSTejun Heo } 538982607adcSTejun Heo } 539082607adcSTejun Heo 539182607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 539282607adcSTejun Heo const struct kernel_param *kp) 539382607adcSTejun Heo { 539482607adcSTejun Heo unsigned long thresh; 539582607adcSTejun Heo int ret; 539682607adcSTejun Heo 539782607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 539882607adcSTejun Heo if (ret) 539982607adcSTejun Heo return ret; 540082607adcSTejun Heo 540182607adcSTejun Heo if (system_wq) 540282607adcSTejun Heo wq_watchdog_set_thresh(thresh); 540382607adcSTejun Heo else 540482607adcSTejun Heo wq_watchdog_thresh = thresh; 540582607adcSTejun Heo 540682607adcSTejun Heo return 0; 540782607adcSTejun Heo } 540882607adcSTejun Heo 540982607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 541082607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 541182607adcSTejun Heo .get = param_get_ulong, 541282607adcSTejun Heo }; 541382607adcSTejun Heo 541482607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 541582607adcSTejun Heo 0644); 541682607adcSTejun Heo 541782607adcSTejun Heo static void wq_watchdog_init(void) 541882607adcSTejun Heo { 541982607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 542082607adcSTejun Heo } 542182607adcSTejun Heo 542282607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 542382607adcSTejun Heo 542482607adcSTejun Heo static inline void wq_watchdog_init(void) { } 542582607adcSTejun Heo 542682607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 542782607adcSTejun Heo 5428bce90380STejun Heo static void __init wq_numa_init(void) 5429bce90380STejun Heo { 5430bce90380STejun Heo cpumask_var_t *tbl; 5431bce90380STejun Heo int node, cpu; 5432bce90380STejun Heo 5433bce90380STejun Heo if (num_possible_nodes() <= 1) 5434bce90380STejun Heo return; 5435bce90380STejun Heo 5436d55262c4STejun Heo if (wq_disable_numa) { 5437d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5438d55262c4STejun Heo return; 5439d55262c4STejun Heo } 5440d55262c4STejun Heo 54414c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 54424c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 54434c16bd32STejun Heo 5444bce90380STejun Heo /* 5445bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5446bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5447bce90380STejun Heo * fully initialized by now. 5448bce90380STejun Heo */ 5449ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5450bce90380STejun Heo BUG_ON(!tbl); 5451bce90380STejun Heo 5452bce90380STejun Heo for_each_node(node) 54535a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 54541be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5455bce90380STejun Heo 5456bce90380STejun Heo for_each_possible_cpu(cpu) { 5457bce90380STejun Heo node = cpu_to_node(cpu); 5458bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5459bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5460bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5461bce90380STejun Heo return; 5462bce90380STejun Heo } 5463bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5464bce90380STejun Heo } 5465bce90380STejun Heo 5466bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5467bce90380STejun Heo wq_numa_enabled = true; 5468bce90380STejun Heo } 5469bce90380STejun Heo 54706ee0578bSSuresh Siddha static int __init init_workqueues(void) 54711da177e4SLinus Torvalds { 54727a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 54737a4e344cSTejun Heo int i, cpu; 5474c34056a3STejun Heo 5475e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5476e904e6c2STejun Heo 5477b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5478b05a7928SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 5479b05a7928SFrederic Weisbecker 5480e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5481e904e6c2STejun Heo 5482bce90380STejun Heo wq_numa_init(); 5483bce90380STejun Heo 5484706026c2STejun Heo /* initialize CPU pools */ 548529c91e99STejun Heo for_each_possible_cpu(cpu) { 54864ce62e9eSTejun Heo struct worker_pool *pool; 54878b03ae3cSTejun Heo 54887a4e344cSTejun Heo i = 0; 5489f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54907a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5491ec22ca5eSTejun Heo pool->cpu = cpu; 54927a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 54937a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5494f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 54957a4e344cSTejun Heo 54969daf9e67STejun Heo /* alloc pool ID */ 549768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 54989daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 549968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55004ce62e9eSTejun Heo } 55018b03ae3cSTejun Heo } 55028b03ae3cSTejun Heo 5503e22bee78STejun Heo /* create the initial worker */ 550429c91e99STejun Heo for_each_online_cpu(cpu) { 55054ce62e9eSTejun Heo struct worker_pool *pool; 5506e22bee78STejun Heo 5507f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 550824647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5509051e1850SLai Jiangshan BUG_ON(!create_worker(pool)); 5510e22bee78STejun Heo } 55114ce62e9eSTejun Heo } 5512e22bee78STejun Heo 55138a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 551429c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 551529c91e99STejun Heo struct workqueue_attrs *attrs; 551629c91e99STejun Heo 551729c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 551829c91e99STejun Heo attrs->nice = std_nice[i]; 551929c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 55208a2b7538STejun Heo 55218a2b7538STejun Heo /* 55228a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 55238a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 55248a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 55258a2b7538STejun Heo */ 55268a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 55278a2b7538STejun Heo attrs->nice = std_nice[i]; 55288a2b7538STejun Heo attrs->no_numa = true; 55298a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 553029c91e99STejun Heo } 553129c91e99STejun Heo 5532d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 55331aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5534d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5535f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5536f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 553724d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 553824d51addSTejun Heo WQ_FREEZABLE, 0); 55390668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 55400668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 55410668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 55420668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 55430668106cSViresh Kumar 0); 55441aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 55450668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 55460668106cSViresh Kumar !system_power_efficient_wq || 55470668106cSViresh Kumar !system_freezable_power_efficient_wq); 554882607adcSTejun Heo 554982607adcSTejun Heo wq_watchdog_init(); 555082607adcSTejun Heo 55516ee0578bSSuresh Siddha return 0; 55521da177e4SLinus Torvalds } 55536ee0578bSSuresh Siddha early_initcall(init_workqueues); 5554