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 436dc186ad7SThomas Gleixner /* 437dc186ad7SThomas Gleixner * fixup_init is called when: 438dc186ad7SThomas Gleixner * - an active object is initialized 439dc186ad7SThomas Gleixner */ 440dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 441dc186ad7SThomas Gleixner { 442dc186ad7SThomas Gleixner struct work_struct *work = addr; 443dc186ad7SThomas Gleixner 444dc186ad7SThomas Gleixner switch (state) { 445dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 446dc186ad7SThomas Gleixner cancel_work_sync(work); 447dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 448dc186ad7SThomas Gleixner return 1; 449dc186ad7SThomas Gleixner default: 450dc186ad7SThomas Gleixner return 0; 451dc186ad7SThomas Gleixner } 452dc186ad7SThomas Gleixner } 453dc186ad7SThomas Gleixner 454dc186ad7SThomas Gleixner /* 455dc186ad7SThomas Gleixner * fixup_activate is called when: 456dc186ad7SThomas Gleixner * - an active object is activated 457dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 458dc186ad7SThomas Gleixner */ 459dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 460dc186ad7SThomas Gleixner { 461dc186ad7SThomas Gleixner struct work_struct *work = addr; 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner switch (state) { 464dc186ad7SThomas Gleixner 465dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 466dc186ad7SThomas Gleixner /* 467dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 468dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 469dc186ad7SThomas Gleixner * is tracked in the object tracker. 470dc186ad7SThomas Gleixner */ 47122df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 472dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 473dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 474dc186ad7SThomas Gleixner return 0; 475dc186ad7SThomas Gleixner } 476dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 477dc186ad7SThomas Gleixner return 0; 478dc186ad7SThomas Gleixner 479dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 480dc186ad7SThomas Gleixner WARN_ON(1); 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner default: 483dc186ad7SThomas Gleixner return 0; 484dc186ad7SThomas Gleixner } 485dc186ad7SThomas Gleixner } 486dc186ad7SThomas Gleixner 487dc186ad7SThomas Gleixner /* 488dc186ad7SThomas Gleixner * fixup_free is called when: 489dc186ad7SThomas Gleixner * - an active object is freed 490dc186ad7SThomas Gleixner */ 491dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 492dc186ad7SThomas Gleixner { 493dc186ad7SThomas Gleixner struct work_struct *work = addr; 494dc186ad7SThomas Gleixner 495dc186ad7SThomas Gleixner switch (state) { 496dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 497dc186ad7SThomas Gleixner cancel_work_sync(work); 498dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 499dc186ad7SThomas Gleixner return 1; 500dc186ad7SThomas Gleixner default: 501dc186ad7SThomas Gleixner return 0; 502dc186ad7SThomas Gleixner } 503dc186ad7SThomas Gleixner } 504dc186ad7SThomas Gleixner 505dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 506dc186ad7SThomas Gleixner .name = "work_struct", 50799777288SStanislaw Gruszka .debug_hint = work_debug_hint, 508dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 509dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 510dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 511dc186ad7SThomas Gleixner }; 512dc186ad7SThomas Gleixner 513dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 514dc186ad7SThomas Gleixner { 515dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 516dc186ad7SThomas Gleixner } 517dc186ad7SThomas Gleixner 518dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 519dc186ad7SThomas Gleixner { 520dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 521dc186ad7SThomas Gleixner } 522dc186ad7SThomas Gleixner 523dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 524dc186ad7SThomas Gleixner { 525dc186ad7SThomas Gleixner if (onstack) 526dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 527dc186ad7SThomas Gleixner else 528dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 529dc186ad7SThomas Gleixner } 530dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 531dc186ad7SThomas Gleixner 532dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 533dc186ad7SThomas Gleixner { 534dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 535dc186ad7SThomas Gleixner } 536dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 537dc186ad7SThomas Gleixner 538ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 539ea2e64f2SThomas Gleixner { 540ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 541ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 542ea2e64f2SThomas Gleixner } 543ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 544ea2e64f2SThomas Gleixner 545dc186ad7SThomas Gleixner #else 546dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 547dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 548dc186ad7SThomas Gleixner #endif 549dc186ad7SThomas Gleixner 5504e8b22bdSLi Bin /** 5514e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5524e8b22bdSLi Bin * @pool: the pool pointer of interest 5534e8b22bdSLi Bin * 5544e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5554e8b22bdSLi Bin * successfully, -errno on failure. 5564e8b22bdSLi Bin */ 5579daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5589daf9e67STejun Heo { 5599daf9e67STejun Heo int ret; 5609daf9e67STejun Heo 56168e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5625bcab335STejun Heo 5634e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5644e8b22bdSLi Bin GFP_KERNEL); 565229641a6STejun Heo if (ret >= 0) { 566e68035fbSTejun Heo pool->id = ret; 567229641a6STejun Heo return 0; 568229641a6STejun Heo } 5699daf9e67STejun Heo return ret; 5709daf9e67STejun Heo } 5719daf9e67STejun Heo 57276af4d93STejun Heo /** 573df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 574df2d5ae4STejun Heo * @wq: the target workqueue 575df2d5ae4STejun Heo * @node: the node ID 576df2d5ae4STejun Heo * 5775b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5785b95e1afSLai Jiangshan * read locked. 579df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 580df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 581d185af30SYacine Belkadi * 582d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 583df2d5ae4STejun Heo */ 584df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 585df2d5ae4STejun Heo int node) 586df2d5ae4STejun Heo { 5875b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 588d6e022f1STejun Heo 589d6e022f1STejun Heo /* 590d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 591d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 592d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 593d6e022f1STejun Heo * happens, this workaround can be removed. 594d6e022f1STejun Heo */ 595d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 596d6e022f1STejun Heo return wq->dfl_pwq; 597d6e022f1STejun Heo 598df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 599df2d5ae4STejun Heo } 600df2d5ae4STejun Heo 60173f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 60273f53c4aSTejun Heo { 60373f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 60473f53c4aSTejun Heo } 60573f53c4aSTejun Heo 60673f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 60773f53c4aSTejun Heo { 60873f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 60973f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 61073f53c4aSTejun Heo } 61173f53c4aSTejun Heo 61273f53c4aSTejun Heo static int work_next_color(int color) 61373f53c4aSTejun Heo { 61473f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 615a848e3b6SOleg Nesterov } 616a848e3b6SOleg Nesterov 6174594bf15SDavid Howells /* 618112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 619112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6207c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6217a22ad75STejun Heo * 622112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 623112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 624bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 625bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6267a22ad75STejun Heo * 627112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6287c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 629112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6307c3eed5cSTejun Heo * available only while the work item is queued. 631bbb68dfaSTejun Heo * 632bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 633bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 634bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 635bbb68dfaSTejun Heo * try to steal the PENDING bit. 6364594bf15SDavid Howells */ 6377a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6387a22ad75STejun Heo unsigned long flags) 6397a22ad75STejun Heo { 6406183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6417a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6427a22ad75STejun Heo } 6437a22ad75STejun Heo 644112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6454690c4abSTejun Heo unsigned long extra_flags) 646365970a1SDavid Howells { 647112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 648112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 649365970a1SDavid Howells } 650365970a1SDavid Howells 6514468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6524468a00fSLai Jiangshan int pool_id) 6534468a00fSLai Jiangshan { 6544468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6554468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6564468a00fSLai Jiangshan } 6574468a00fSLai Jiangshan 6587c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6597c3eed5cSTejun Heo int pool_id) 6604d707b9fSOleg Nesterov { 66123657bb1STejun Heo /* 66223657bb1STejun Heo * The following wmb is paired with the implied mb in 66323657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 66423657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 66523657bb1STejun Heo * owner. 66623657bb1STejun Heo */ 66723657bb1STejun Heo smp_wmb(); 6687c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6694d707b9fSOleg Nesterov } 6704d707b9fSOleg Nesterov 6717a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 672365970a1SDavid Howells { 6737c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6747c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6757a22ad75STejun Heo } 6767a22ad75STejun Heo 677112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6787a22ad75STejun Heo { 679e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6807a22ad75STejun Heo 681112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 682e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 683e120153dSTejun Heo else 684e120153dSTejun Heo return NULL; 6857a22ad75STejun Heo } 6867a22ad75STejun Heo 6877c3eed5cSTejun Heo /** 6887c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6897c3eed5cSTejun Heo * @work: the work item of interest 6907c3eed5cSTejun Heo * 69168e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69268e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 69368e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 694fa1b54e6STejun Heo * 695fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 696fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 697fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 698fa1b54e6STejun Heo * returned pool is and stays online. 699d185af30SYacine Belkadi * 700d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7017c3eed5cSTejun Heo */ 7027c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7037a22ad75STejun Heo { 704e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7057c3eed5cSTejun Heo int pool_id; 7067a22ad75STejun Heo 70768e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 708fa1b54e6STejun Heo 709112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 710112202d9STejun Heo return ((struct pool_workqueue *) 7117c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7127a22ad75STejun Heo 7137c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7147c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7157a22ad75STejun Heo return NULL; 7167a22ad75STejun Heo 717fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7187c3eed5cSTejun Heo } 7197c3eed5cSTejun Heo 7207c3eed5cSTejun Heo /** 7217c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7227c3eed5cSTejun Heo * @work: the work item of interest 7237c3eed5cSTejun Heo * 724d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7257c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7267c3eed5cSTejun Heo */ 7277c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7287c3eed5cSTejun Heo { 72954d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7307c3eed5cSTejun Heo 731112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 732112202d9STejun Heo return ((struct pool_workqueue *) 73354d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 73454d5b7d0SLai Jiangshan 73554d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7367c3eed5cSTejun Heo } 7377c3eed5cSTejun Heo 738bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 739bbb68dfaSTejun Heo { 7407c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 741bbb68dfaSTejun Heo 7427c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7437c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 744bbb68dfaSTejun Heo } 745bbb68dfaSTejun Heo 746bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 747bbb68dfaSTejun Heo { 748bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 749bbb68dfaSTejun Heo 750112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 751bbb68dfaSTejun Heo } 752bbb68dfaSTejun Heo 753e22bee78STejun Heo /* 7543270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7553270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 756d565ed63STejun Heo * they're being called with pool->lock held. 757e22bee78STejun Heo */ 758e22bee78STejun Heo 75963d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 760649027d7STejun Heo { 761e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 762649027d7STejun Heo } 763649027d7STejun Heo 764e22bee78STejun Heo /* 765e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 766e22bee78STejun Heo * running workers. 767974271c4STejun Heo * 768974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 769706026c2STejun Heo * function will always return %true for unbound pools as long as the 770974271c4STejun Heo * worklist isn't empty. 771e22bee78STejun Heo */ 77263d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 773e22bee78STejun Heo { 77463d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 775e22bee78STejun Heo } 776e22bee78STejun Heo 777e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 77863d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 779e22bee78STejun Heo { 78063d95a91STejun Heo return pool->nr_idle; 781e22bee78STejun Heo } 782e22bee78STejun Heo 783e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 78463d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 785e22bee78STejun Heo { 786e19e397aSTejun Heo return !list_empty(&pool->worklist) && 787e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 788e22bee78STejun Heo } 789e22bee78STejun Heo 790e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79163d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 792e22bee78STejun Heo { 79363d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 794e22bee78STejun Heo } 795e22bee78STejun Heo 796e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 79763d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 798e22bee78STejun Heo { 79934a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 80063d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80163d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 802e22bee78STejun Heo 803e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 804e22bee78STejun Heo } 805e22bee78STejun Heo 806e22bee78STejun Heo /* 807e22bee78STejun Heo * Wake up functions. 808e22bee78STejun Heo */ 809e22bee78STejun Heo 8101037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8111037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8127e11629dSTejun Heo { 81363d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8147e11629dSTejun Heo return NULL; 8157e11629dSTejun Heo 81663d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8177e11629dSTejun Heo } 8187e11629dSTejun Heo 8197e11629dSTejun Heo /** 8207e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82163d95a91STejun Heo * @pool: worker pool to wake worker from 8227e11629dSTejun Heo * 82363d95a91STejun Heo * Wake up the first idle worker of @pool. 8247e11629dSTejun Heo * 8257e11629dSTejun Heo * CONTEXT: 826d565ed63STejun Heo * spin_lock_irq(pool->lock). 8277e11629dSTejun Heo */ 82863d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8297e11629dSTejun Heo { 8301037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8317e11629dSTejun Heo 8327e11629dSTejun Heo if (likely(worker)) 8337e11629dSTejun Heo wake_up_process(worker->task); 8347e11629dSTejun Heo } 8357e11629dSTejun Heo 8364690c4abSTejun Heo /** 837e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 838e22bee78STejun Heo * @task: task waking up 839e22bee78STejun Heo * @cpu: CPU @task is waking up to 840e22bee78STejun Heo * 841e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 842e22bee78STejun Heo * being awoken. 843e22bee78STejun Heo * 844e22bee78STejun Heo * CONTEXT: 845e22bee78STejun Heo * spin_lock_irq(rq->lock) 846e22bee78STejun Heo */ 847d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 848e22bee78STejun Heo { 849e22bee78STejun Heo struct worker *worker = kthread_data(task); 850e22bee78STejun Heo 85136576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 852ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 853e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 854e22bee78STejun Heo } 85536576000SJoonsoo Kim } 856e22bee78STejun Heo 857e22bee78STejun Heo /** 858e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 859e22bee78STejun Heo * @task: task going to sleep 860e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 861e22bee78STejun Heo * 862e22bee78STejun Heo * This function is called during schedule() when a busy worker is 863e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 864e22bee78STejun Heo * returning pointer to its task. 865e22bee78STejun Heo * 866e22bee78STejun Heo * CONTEXT: 867e22bee78STejun Heo * spin_lock_irq(rq->lock) 868e22bee78STejun Heo * 869d185af30SYacine Belkadi * Return: 870e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 871e22bee78STejun Heo */ 872d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 873e22bee78STejun Heo { 874e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 875111c225aSTejun Heo struct worker_pool *pool; 876e22bee78STejun Heo 877111c225aSTejun Heo /* 878111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 879111c225aSTejun Heo * workers, also reach here, let's not access anything before 880111c225aSTejun Heo * checking NOT_RUNNING. 881111c225aSTejun Heo */ 8822d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 883e22bee78STejun Heo return NULL; 884e22bee78STejun Heo 885111c225aSTejun Heo pool = worker->pool; 886111c225aSTejun Heo 887e22bee78STejun Heo /* this can only happen on the local cpu */ 88892b69f50SLai Jiangshan if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) 8896183c009STejun Heo return NULL; 890e22bee78STejun Heo 891e22bee78STejun Heo /* 892e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 893e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 894e22bee78STejun Heo * Please read comment there. 895e22bee78STejun Heo * 896628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 897628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 898628c78e7STejun Heo * disabled, which in turn means that none else could be 899d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 900628c78e7STejun Heo * lock is safe. 901e22bee78STejun Heo */ 902e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 903e19e397aSTejun Heo !list_empty(&pool->worklist)) 9041037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 905e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 906e22bee78STejun Heo } 907e22bee78STejun Heo 908e22bee78STejun Heo /** 909e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 910cb444766STejun Heo * @worker: self 911d302f017STejun Heo * @flags: flags to set 912d302f017STejun Heo * 913228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 914d302f017STejun Heo * 915cb444766STejun Heo * CONTEXT: 916d565ed63STejun Heo * spin_lock_irq(pool->lock) 917d302f017STejun Heo */ 918228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 919d302f017STejun Heo { 920bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 921e22bee78STejun Heo 922cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 923cb444766STejun Heo 924228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 925e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 926e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 927e19e397aSTejun Heo atomic_dec(&pool->nr_running); 928e22bee78STejun Heo } 929e22bee78STejun Heo 930d302f017STejun Heo worker->flags |= flags; 931d302f017STejun Heo } 932d302f017STejun Heo 933d302f017STejun Heo /** 934e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 935cb444766STejun Heo * @worker: self 936d302f017STejun Heo * @flags: flags to clear 937d302f017STejun Heo * 938e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 939d302f017STejun Heo * 940cb444766STejun Heo * CONTEXT: 941d565ed63STejun Heo * spin_lock_irq(pool->lock) 942d302f017STejun Heo */ 943d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 944d302f017STejun Heo { 94563d95a91STejun Heo struct worker_pool *pool = worker->pool; 946e22bee78STejun Heo unsigned int oflags = worker->flags; 947e22bee78STejun Heo 948cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 949cb444766STejun Heo 950d302f017STejun Heo worker->flags &= ~flags; 951e22bee78STejun Heo 95242c025f3STejun Heo /* 95342c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 95442c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 95542c025f3STejun Heo * of multiple flags, not a single flag. 95642c025f3STejun Heo */ 957e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 958e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 959e19e397aSTejun Heo atomic_inc(&pool->nr_running); 960d302f017STejun Heo } 961d302f017STejun Heo 962d302f017STejun Heo /** 9638cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 964c9e7cf27STejun Heo * @pool: pool of interest 9658cca0eeaSTejun Heo * @work: work to find worker for 9668cca0eeaSTejun Heo * 967c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 968c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 969a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 970a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 971a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 972a2c1c57bSTejun Heo * being executed. 973a2c1c57bSTejun Heo * 974a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 975a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 976a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 977a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 978a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 979a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 980a2c1c57bSTejun Heo * 981c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 982c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 983c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 984c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 985c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 986c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9878cca0eeaSTejun Heo * 9888cca0eeaSTejun Heo * CONTEXT: 989d565ed63STejun Heo * spin_lock_irq(pool->lock). 9908cca0eeaSTejun Heo * 991d185af30SYacine Belkadi * Return: 992d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9938cca0eeaSTejun Heo * otherwise. 9948cca0eeaSTejun Heo */ 995c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9968cca0eeaSTejun Heo struct work_struct *work) 9978cca0eeaSTejun Heo { 99842f8570fSSasha Levin struct worker *worker; 99942f8570fSSasha Levin 1000b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1001a2c1c57bSTejun Heo (unsigned long)work) 1002a2c1c57bSTejun Heo if (worker->current_work == work && 1003a2c1c57bSTejun Heo worker->current_func == work->func) 100442f8570fSSasha Levin return worker; 100542f8570fSSasha Levin 100642f8570fSSasha Levin return NULL; 10078cca0eeaSTejun Heo } 10088cca0eeaSTejun Heo 10098cca0eeaSTejun Heo /** 1010bf4ede01STejun Heo * move_linked_works - move linked works to a list 1011bf4ede01STejun Heo * @work: start of series of works to be scheduled 1012bf4ede01STejun Heo * @head: target list to append @work to 1013402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1014bf4ede01STejun Heo * 1015bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1016bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1017bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1018bf4ede01STejun Heo * 1019bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1020bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1021bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1022bf4ede01STejun Heo * 1023bf4ede01STejun Heo * CONTEXT: 1024d565ed63STejun Heo * spin_lock_irq(pool->lock). 1025bf4ede01STejun Heo */ 1026bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1027bf4ede01STejun Heo struct work_struct **nextp) 1028bf4ede01STejun Heo { 1029bf4ede01STejun Heo struct work_struct *n; 1030bf4ede01STejun Heo 1031bf4ede01STejun Heo /* 1032bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1033bf4ede01STejun Heo * use NULL for list head. 1034bf4ede01STejun Heo */ 1035bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1036bf4ede01STejun Heo list_move_tail(&work->entry, head); 1037bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1038bf4ede01STejun Heo break; 1039bf4ede01STejun Heo } 1040bf4ede01STejun Heo 1041bf4ede01STejun Heo /* 1042bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1043bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1044bf4ede01STejun Heo * needs to be updated. 1045bf4ede01STejun Heo */ 1046bf4ede01STejun Heo if (nextp) 1047bf4ede01STejun Heo *nextp = n; 1048bf4ede01STejun Heo } 1049bf4ede01STejun Heo 10508864b4e5STejun Heo /** 10518864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10528864b4e5STejun Heo * @pwq: pool_workqueue to get 10538864b4e5STejun Heo * 10548864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10558864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10568864b4e5STejun Heo */ 10578864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10588864b4e5STejun Heo { 10598864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10608864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10618864b4e5STejun Heo pwq->refcnt++; 10628864b4e5STejun Heo } 10638864b4e5STejun Heo 10648864b4e5STejun Heo /** 10658864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10668864b4e5STejun Heo * @pwq: pool_workqueue to put 10678864b4e5STejun Heo * 10688864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10698864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10708864b4e5STejun Heo */ 10718864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10728864b4e5STejun Heo { 10738864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10748864b4e5STejun Heo if (likely(--pwq->refcnt)) 10758864b4e5STejun Heo return; 10768864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10778864b4e5STejun Heo return; 10788864b4e5STejun Heo /* 10798864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10808864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10818864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10828864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10838864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10848864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10858864b4e5STejun Heo */ 10868864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10878864b4e5STejun Heo } 10888864b4e5STejun Heo 1089dce90d47STejun Heo /** 1090dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1091dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1092dce90d47STejun Heo * 1093dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1094dce90d47STejun Heo */ 1095dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1096dce90d47STejun Heo { 1097dce90d47STejun Heo if (pwq) { 1098dce90d47STejun Heo /* 1099dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1100dce90d47STejun Heo * following lock operations are safe. 1101dce90d47STejun Heo */ 1102dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1103dce90d47STejun Heo put_pwq(pwq); 1104dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1105dce90d47STejun Heo } 1106dce90d47STejun Heo } 1107dce90d47STejun Heo 1108112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1109bf4ede01STejun Heo { 1110112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1111bf4ede01STejun Heo 1112bf4ede01STejun Heo trace_workqueue_activate_work(work); 111382607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 111482607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1115112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1116bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1117112202d9STejun Heo pwq->nr_active++; 1118bf4ede01STejun Heo } 1119bf4ede01STejun Heo 1120112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11213aa62497SLai Jiangshan { 1122112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11233aa62497SLai Jiangshan struct work_struct, entry); 11243aa62497SLai Jiangshan 1125112202d9STejun Heo pwq_activate_delayed_work(work); 11263aa62497SLai Jiangshan } 11273aa62497SLai Jiangshan 1128bf4ede01STejun Heo /** 1129112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1130112202d9STejun Heo * @pwq: pwq of interest 1131bf4ede01STejun Heo * @color: color of work which left the queue 1132bf4ede01STejun Heo * 1133bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1134112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1135bf4ede01STejun Heo * 1136bf4ede01STejun Heo * CONTEXT: 1137d565ed63STejun Heo * spin_lock_irq(pool->lock). 1138bf4ede01STejun Heo */ 1139112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1140bf4ede01STejun Heo { 11418864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1142bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11438864b4e5STejun Heo goto out_put; 1144bf4ede01STejun Heo 1145112202d9STejun Heo pwq->nr_in_flight[color]--; 1146bf4ede01STejun Heo 1147112202d9STejun Heo pwq->nr_active--; 1148112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1149bf4ede01STejun Heo /* one down, submit a delayed one */ 1150112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1151112202d9STejun Heo pwq_activate_first_delayed(pwq); 1152bf4ede01STejun Heo } 1153bf4ede01STejun Heo 1154bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1155112202d9STejun Heo if (likely(pwq->flush_color != color)) 11568864b4e5STejun Heo goto out_put; 1157bf4ede01STejun Heo 1158bf4ede01STejun Heo /* are there still in-flight works? */ 1159112202d9STejun Heo if (pwq->nr_in_flight[color]) 11608864b4e5STejun Heo goto out_put; 1161bf4ede01STejun Heo 1162112202d9STejun Heo /* this pwq is done, clear flush_color */ 1163112202d9STejun Heo pwq->flush_color = -1; 1164bf4ede01STejun Heo 1165bf4ede01STejun Heo /* 1166112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1167bf4ede01STejun Heo * will handle the rest. 1168bf4ede01STejun Heo */ 1169112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1170112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11718864b4e5STejun Heo out_put: 11728864b4e5STejun Heo put_pwq(pwq); 1173bf4ede01STejun Heo } 1174bf4ede01STejun Heo 117536e227d2STejun Heo /** 1176bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 117736e227d2STejun Heo * @work: work item to steal 117836e227d2STejun Heo * @is_dwork: @work is a delayed_work 1179bbb68dfaSTejun Heo * @flags: place to store irq state 118036e227d2STejun Heo * 118136e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1182d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 118336e227d2STejun Heo * 1184d185af30SYacine Belkadi * Return: 118536e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 118636e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 118736e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1188bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1189bbb68dfaSTejun Heo * for arbitrarily long 119036e227d2STejun Heo * 1191d185af30SYacine Belkadi * Note: 1192bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1193e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1194e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1195e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1196bbb68dfaSTejun Heo * 1197bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1198bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1199bbb68dfaSTejun Heo * 1200e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1201bf4ede01STejun Heo */ 1202bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1203bbb68dfaSTejun Heo unsigned long *flags) 1204bf4ede01STejun Heo { 1205d565ed63STejun Heo struct worker_pool *pool; 1206112202d9STejun Heo struct pool_workqueue *pwq; 1207bf4ede01STejun Heo 1208bbb68dfaSTejun Heo local_irq_save(*flags); 1209bbb68dfaSTejun Heo 121036e227d2STejun Heo /* try to steal the timer if it exists */ 121136e227d2STejun Heo if (is_dwork) { 121236e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 121336e227d2STejun Heo 1214e0aecdd8STejun Heo /* 1215e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1216e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1217e0aecdd8STejun Heo * running on the local CPU. 1218e0aecdd8STejun Heo */ 121936e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 122036e227d2STejun Heo return 1; 122136e227d2STejun Heo } 122236e227d2STejun Heo 122336e227d2STejun Heo /* try to claim PENDING the normal way */ 1224bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1225bf4ede01STejun Heo return 0; 1226bf4ede01STejun Heo 1227bf4ede01STejun Heo /* 1228bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1229bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1230bf4ede01STejun Heo */ 1231d565ed63STejun Heo pool = get_work_pool(work); 1232d565ed63STejun Heo if (!pool) 1233bbb68dfaSTejun Heo goto fail; 1234bf4ede01STejun Heo 1235d565ed63STejun Heo spin_lock(&pool->lock); 1236bf4ede01STejun Heo /* 1237112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1238112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1239112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1240112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1241112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12420b3dae68SLai Jiangshan * item is currently queued on that pool. 1243bf4ede01STejun Heo */ 1244112202d9STejun Heo pwq = get_work_pwq(work); 1245112202d9STejun Heo if (pwq && pwq->pool == pool) { 1246bf4ede01STejun Heo debug_work_deactivate(work); 12473aa62497SLai Jiangshan 12483aa62497SLai Jiangshan /* 124916062836STejun Heo * A delayed work item cannot be grabbed directly because 125016062836STejun Heo * it might have linked NO_COLOR work items which, if left 1251112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 125216062836STejun Heo * management later on and cause stall. Make sure the work 125316062836STejun Heo * item is activated before grabbing. 12543aa62497SLai Jiangshan */ 12553aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1256112202d9STejun Heo pwq_activate_delayed_work(work); 12573aa62497SLai Jiangshan 1258bf4ede01STejun Heo list_del_init(&work->entry); 12599c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 126036e227d2STejun Heo 1261112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12624468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12634468a00fSLai Jiangshan 1264d565ed63STejun Heo spin_unlock(&pool->lock); 126536e227d2STejun Heo return 1; 1266bf4ede01STejun Heo } 1267d565ed63STejun Heo spin_unlock(&pool->lock); 1268bbb68dfaSTejun Heo fail: 1269bbb68dfaSTejun Heo local_irq_restore(*flags); 1270bbb68dfaSTejun Heo if (work_is_canceling(work)) 1271bbb68dfaSTejun Heo return -ENOENT; 1272bbb68dfaSTejun Heo cpu_relax(); 127336e227d2STejun Heo return -EAGAIN; 1274bf4ede01STejun Heo } 1275bf4ede01STejun Heo 1276bf4ede01STejun Heo /** 1277706026c2STejun Heo * insert_work - insert a work into a pool 1278112202d9STejun Heo * @pwq: pwq @work belongs to 12794690c4abSTejun Heo * @work: work to insert 12804690c4abSTejun Heo * @head: insertion point 12814690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12824690c4abSTejun Heo * 1283112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1284706026c2STejun Heo * work_struct flags. 12854690c4abSTejun Heo * 12864690c4abSTejun Heo * CONTEXT: 1287d565ed63STejun Heo * spin_lock_irq(pool->lock). 1288365970a1SDavid Howells */ 1289112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1290112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1291b89deed3SOleg Nesterov { 1292112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1293e1d8aa9fSFrederic Weisbecker 12944690c4abSTejun Heo /* we own @work, set data and link */ 1295112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12961a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12978864b4e5STejun Heo get_pwq(pwq); 1298e22bee78STejun Heo 1299e22bee78STejun Heo /* 1300c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1301c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1302c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1303e22bee78STejun Heo */ 1304e22bee78STejun Heo smp_mb(); 1305e22bee78STejun Heo 130663d95a91STejun Heo if (__need_more_worker(pool)) 130763d95a91STejun Heo wake_up_worker(pool); 1308b89deed3SOleg Nesterov } 1309b89deed3SOleg Nesterov 1310c8efcc25STejun Heo /* 1311c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13128d03ecfeSTejun Heo * same workqueue. 1313c8efcc25STejun Heo */ 1314c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1315c8efcc25STejun Heo { 1316c8efcc25STejun Heo struct worker *worker; 1317c8efcc25STejun Heo 13188d03ecfeSTejun Heo worker = current_wq_worker(); 1319c8efcc25STejun Heo /* 13208d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13218d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1322c8efcc25STejun Heo */ 1323112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1324c8efcc25STejun Heo } 1325c8efcc25STejun Heo 1326ef557180SMike Galbraith /* 1327ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1328ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1329ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1330ef557180SMike Galbraith */ 1331ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1332ef557180SMike Galbraith { 1333f303fccbSTejun Heo static bool printed_dbg_warning; 1334ef557180SMike Galbraith int new_cpu; 1335ef557180SMike Galbraith 1336f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1337ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1338ef557180SMike Galbraith return cpu; 1339f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1340f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1341f303fccbSTejun Heo printed_dbg_warning = true; 1342f303fccbSTejun Heo } 1343f303fccbSTejun Heo 1344ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1345ef557180SMike Galbraith return cpu; 1346ef557180SMike Galbraith 1347ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1348ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1349ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1350ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1351ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1352ef557180SMike Galbraith return cpu; 1353ef557180SMike Galbraith } 1354ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1355ef557180SMike Galbraith 1356ef557180SMike Galbraith return new_cpu; 1357ef557180SMike Galbraith } 1358ef557180SMike Galbraith 1359d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13601da177e4SLinus Torvalds struct work_struct *work) 13611da177e4SLinus Torvalds { 1362112202d9STejun Heo struct pool_workqueue *pwq; 1363c9178087STejun Heo struct worker_pool *last_pool; 13641e19ffc6STejun Heo struct list_head *worklist; 13658a2e8e5dSTejun Heo unsigned int work_flags; 1366b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13678930cabaSTejun Heo 13688930cabaSTejun Heo /* 13698930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13708930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13718930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13728930cabaSTejun Heo * happen with IRQ disabled. 13738930cabaSTejun Heo */ 13748930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13751da177e4SLinus Torvalds 1376dc186ad7SThomas Gleixner debug_work_activate(work); 13771e19ffc6STejun Heo 13789ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1379618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1380c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1381e41e704bSTejun Heo return; 13829e8cd2f5STejun Heo retry: 1383df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1384ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1385df2d5ae4STejun Heo 1386df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1387df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1388c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1389df2d5ae4STejun Heo else 1390df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1391f3421797STejun Heo 139218aa9effSTejun Heo /* 1393c9178087STejun Heo * If @work was previously on a different pool, it might still be 1394c9178087STejun Heo * running there, in which case the work needs to be queued on that 1395c9178087STejun Heo * pool to guarantee non-reentrancy. 139618aa9effSTejun Heo */ 1397c9e7cf27STejun Heo last_pool = get_work_pool(work); 1398112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 139918aa9effSTejun Heo struct worker *worker; 140018aa9effSTejun Heo 1401d565ed63STejun Heo spin_lock(&last_pool->lock); 140218aa9effSTejun Heo 1403c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 140418aa9effSTejun Heo 1405112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1406c9178087STejun Heo pwq = worker->current_pwq; 14078594fadeSLai Jiangshan } else { 140818aa9effSTejun Heo /* meh... not running there, queue here */ 1409d565ed63STejun Heo spin_unlock(&last_pool->lock); 1410112202d9STejun Heo spin_lock(&pwq->pool->lock); 141118aa9effSTejun Heo } 14128930cabaSTejun Heo } else { 1413112202d9STejun Heo spin_lock(&pwq->pool->lock); 14148930cabaSTejun Heo } 1415502ca9d8STejun Heo 14169e8cd2f5STejun Heo /* 14179e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14189e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14199e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1420df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1421df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14229e8cd2f5STejun Heo * make forward-progress. 14239e8cd2f5STejun Heo */ 14249e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14259e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 14269e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 14279e8cd2f5STejun Heo cpu_relax(); 14289e8cd2f5STejun Heo goto retry; 14299e8cd2f5STejun Heo } 14309e8cd2f5STejun Heo /* oops */ 14319e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14329e8cd2f5STejun Heo wq->name, cpu); 14339e8cd2f5STejun Heo } 14349e8cd2f5STejun Heo 1435112202d9STejun Heo /* pwq determined, queue */ 1436112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1437502ca9d8STejun Heo 1438f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1439112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1440f5b2552bSDan Carpenter return; 1441f5b2552bSDan Carpenter } 14421e19ffc6STejun Heo 1443112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1444112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14451e19ffc6STejun Heo 1446112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1447cdadf009STejun Heo trace_workqueue_activate_work(work); 1448112202d9STejun Heo pwq->nr_active++; 1449112202d9STejun Heo worklist = &pwq->pool->worklist; 145082607adcSTejun Heo if (list_empty(worklist)) 145182607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14528a2e8e5dSTejun Heo } else { 14538a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1454112202d9STejun Heo worklist = &pwq->delayed_works; 14558a2e8e5dSTejun Heo } 14561e19ffc6STejun Heo 1457112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14581e19ffc6STejun Heo 1459112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14601da177e4SLinus Torvalds } 14611da177e4SLinus Torvalds 14620fcb78c2SRolf Eike Beer /** 1463c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1464c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1465c1a220e7SZhang Rui * @wq: workqueue to use 1466c1a220e7SZhang Rui * @work: work to queue 1467c1a220e7SZhang Rui * 1468c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1469c1a220e7SZhang Rui * can't go away. 1470d185af30SYacine Belkadi * 1471d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1472c1a220e7SZhang Rui */ 1473d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1474d4283e93STejun Heo struct work_struct *work) 1475c1a220e7SZhang Rui { 1476d4283e93STejun Heo bool ret = false; 14778930cabaSTejun Heo unsigned long flags; 14788930cabaSTejun Heo 14798930cabaSTejun Heo local_irq_save(flags); 1480c1a220e7SZhang Rui 148122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14824690c4abSTejun Heo __queue_work(cpu, wq, work); 1483d4283e93STejun Heo ret = true; 1484c1a220e7SZhang Rui } 14858930cabaSTejun Heo 14868930cabaSTejun Heo local_irq_restore(flags); 1487c1a220e7SZhang Rui return ret; 1488c1a220e7SZhang Rui } 1489ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1490c1a220e7SZhang Rui 1491d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14921da177e4SLinus Torvalds { 149352bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14941da177e4SLinus Torvalds 1495e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 149660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14971da177e4SLinus Torvalds } 14981438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14991da177e4SLinus Torvalds 15007beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 150152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15021da177e4SLinus Torvalds { 15037beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 15047beb2edfSTejun Heo struct work_struct *work = &dwork->work; 15051da177e4SLinus Torvalds 15067beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 15077beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1508fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1509fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 15107beb2edfSTejun Heo 15118852aac2STejun Heo /* 15128852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 15138852aac2STejun Heo * both optimization and correctness. The earliest @timer can 15148852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 15158852aac2STejun Heo * on that there's no such delay when @delay is 0. 15168852aac2STejun Heo */ 15178852aac2STejun Heo if (!delay) { 15188852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 15198852aac2STejun Heo return; 15208852aac2STejun Heo } 15218852aac2STejun Heo 15227beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 15237beb2edfSTejun Heo 152460c057bcSLai Jiangshan dwork->wq = wq; 15251265057fSTejun Heo dwork->cpu = cpu; 15267beb2edfSTejun Heo timer->expires = jiffies + delay; 15277beb2edfSTejun Heo 1528041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 15297beb2edfSTejun Heo add_timer_on(timer, cpu); 1530041bd12eSTejun Heo else 1531041bd12eSTejun Heo add_timer(timer); 15327beb2edfSTejun Heo } 15331da177e4SLinus Torvalds 15340fcb78c2SRolf Eike Beer /** 15350fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 15360fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 15370fcb78c2SRolf Eike Beer * @wq: workqueue to use 1538af9997e4SRandy Dunlap * @dwork: work to queue 15390fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 15400fcb78c2SRolf Eike Beer * 1541d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1542715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1543715f1300STejun Heo * execution. 15440fcb78c2SRolf Eike Beer */ 1545d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 154652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15477a6bc1cdSVenkatesh Pallipadi { 154852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1549d4283e93STejun Heo bool ret = false; 15508930cabaSTejun Heo unsigned long flags; 15518930cabaSTejun Heo 15528930cabaSTejun Heo /* read the comment in __queue_work() */ 15538930cabaSTejun Heo local_irq_save(flags); 15547a6bc1cdSVenkatesh Pallipadi 155522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15567beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1557d4283e93STejun Heo ret = true; 15587a6bc1cdSVenkatesh Pallipadi } 15598930cabaSTejun Heo 15608930cabaSTejun Heo local_irq_restore(flags); 15617a6bc1cdSVenkatesh Pallipadi return ret; 15627a6bc1cdSVenkatesh Pallipadi } 1563ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15641da177e4SLinus Torvalds 1565c8e55f36STejun Heo /** 15668376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15678376fe22STejun Heo * @cpu: CPU number to execute work on 15688376fe22STejun Heo * @wq: workqueue to use 15698376fe22STejun Heo * @dwork: work to queue 15708376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15718376fe22STejun Heo * 15728376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15738376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15748376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15758376fe22STejun Heo * current state. 15768376fe22STejun Heo * 1577d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15788376fe22STejun Heo * pending and its timer was modified. 15798376fe22STejun Heo * 1580e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15818376fe22STejun Heo * See try_to_grab_pending() for details. 15828376fe22STejun Heo */ 15838376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15848376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15858376fe22STejun Heo { 15868376fe22STejun Heo unsigned long flags; 15878376fe22STejun Heo int ret; 15888376fe22STejun Heo 15898376fe22STejun Heo do { 15908376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15918376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15928376fe22STejun Heo 15938376fe22STejun Heo if (likely(ret >= 0)) { 15948376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15958376fe22STejun Heo local_irq_restore(flags); 15968376fe22STejun Heo } 15978376fe22STejun Heo 15988376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15998376fe22STejun Heo return ret; 16008376fe22STejun Heo } 16018376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 16028376fe22STejun Heo 16038376fe22STejun Heo /** 1604c8e55f36STejun Heo * worker_enter_idle - enter idle state 1605c8e55f36STejun Heo * @worker: worker which is entering idle state 1606c8e55f36STejun Heo * 1607c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1608c8e55f36STejun Heo * necessary. 1609c8e55f36STejun Heo * 1610c8e55f36STejun Heo * LOCKING: 1611d565ed63STejun Heo * spin_lock_irq(pool->lock). 1612c8e55f36STejun Heo */ 1613c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 16141da177e4SLinus Torvalds { 1615bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1616c8e55f36STejun Heo 16176183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 16186183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 16196183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 16206183c009STejun Heo return; 1621c8e55f36STejun Heo 1622051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1623cb444766STejun Heo worker->flags |= WORKER_IDLE; 1624bd7bdd43STejun Heo pool->nr_idle++; 1625e22bee78STejun Heo worker->last_active = jiffies; 1626c8e55f36STejun Heo 1627c8e55f36STejun Heo /* idle_list is LIFO */ 1628bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1629db7bccf4STejun Heo 163063d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1631628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1632cb444766STejun Heo 1633544ecf31STejun Heo /* 1634706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1635d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1636628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1637628c78e7STejun Heo * unbind is not in progress. 1638544ecf31STejun Heo */ 163924647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1640bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1641e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1642c8e55f36STejun Heo } 1643c8e55f36STejun Heo 1644c8e55f36STejun Heo /** 1645c8e55f36STejun Heo * worker_leave_idle - leave idle state 1646c8e55f36STejun Heo * @worker: worker which is leaving idle state 1647c8e55f36STejun Heo * 1648c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1649c8e55f36STejun Heo * 1650c8e55f36STejun Heo * LOCKING: 1651d565ed63STejun Heo * spin_lock_irq(pool->lock). 1652c8e55f36STejun Heo */ 1653c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1654c8e55f36STejun Heo { 1655bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1656c8e55f36STejun Heo 16576183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16586183c009STejun Heo return; 1659d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1660bd7bdd43STejun Heo pool->nr_idle--; 1661c8e55f36STejun Heo list_del_init(&worker->entry); 1662c8e55f36STejun Heo } 1663c8e55f36STejun Heo 1664f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1665c34056a3STejun Heo { 1666c34056a3STejun Heo struct worker *worker; 1667c34056a3STejun Heo 1668f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1669c8e55f36STejun Heo if (worker) { 1670c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1671affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1672da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1673e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1674e22bee78STejun Heo worker->flags = WORKER_PREP; 1675c8e55f36STejun Heo } 1676c34056a3STejun Heo return worker; 1677c34056a3STejun Heo } 1678c34056a3STejun Heo 1679c34056a3STejun Heo /** 16804736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16814736cbf7SLai Jiangshan * @worker: worker to be attached 16824736cbf7SLai Jiangshan * @pool: the target pool 16834736cbf7SLai Jiangshan * 16844736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16854736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16864736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16874736cbf7SLai Jiangshan */ 16884736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16894736cbf7SLai Jiangshan struct worker_pool *pool) 16904736cbf7SLai Jiangshan { 16914736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16924736cbf7SLai Jiangshan 16934736cbf7SLai Jiangshan /* 16944736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16954736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16964736cbf7SLai Jiangshan */ 16974736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16984736cbf7SLai Jiangshan 16994736cbf7SLai Jiangshan /* 17004736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 17014736cbf7SLai Jiangshan * stable across this function. See the comments above the 17024736cbf7SLai Jiangshan * flag definition for details. 17034736cbf7SLai Jiangshan */ 17044736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 17054736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 17064736cbf7SLai Jiangshan 17074736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 17084736cbf7SLai Jiangshan 17094736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 17104736cbf7SLai Jiangshan } 17114736cbf7SLai Jiangshan 17124736cbf7SLai Jiangshan /** 171360f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 171460f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 171560f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 171660f5a4bcSLai Jiangshan * 17174736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 17184736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 17194736cbf7SLai Jiangshan * other reference to the pool. 172060f5a4bcSLai Jiangshan */ 172160f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 172260f5a4bcSLai Jiangshan struct worker_pool *pool) 172360f5a4bcSLai Jiangshan { 172460f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 172560f5a4bcSLai Jiangshan 172692f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1727da028469SLai Jiangshan list_del(&worker->node); 1728da028469SLai Jiangshan if (list_empty(&pool->workers)) 172960f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 173092f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 173160f5a4bcSLai Jiangshan 1732b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1733b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1734b62c0751SLai Jiangshan 173560f5a4bcSLai Jiangshan if (detach_completion) 173660f5a4bcSLai Jiangshan complete(detach_completion); 173760f5a4bcSLai Jiangshan } 173860f5a4bcSLai Jiangshan 173960f5a4bcSLai Jiangshan /** 1740c34056a3STejun Heo * create_worker - create a new workqueue worker 174163d95a91STejun Heo * @pool: pool the new worker will belong to 1742c34056a3STejun Heo * 1743051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1744c34056a3STejun Heo * 1745c34056a3STejun Heo * CONTEXT: 1746c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1747c34056a3STejun Heo * 1748d185af30SYacine Belkadi * Return: 1749c34056a3STejun Heo * Pointer to the newly created worker. 1750c34056a3STejun Heo */ 1751bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1752c34056a3STejun Heo { 1753c34056a3STejun Heo struct worker *worker = NULL; 1754f3421797STejun Heo int id = -1; 1755e3c916a4STejun Heo char id_buf[16]; 1756c34056a3STejun Heo 17577cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17587cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1759822d8405STejun Heo if (id < 0) 1760c34056a3STejun Heo goto fail; 1761c34056a3STejun Heo 1762f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1763c34056a3STejun Heo if (!worker) 1764c34056a3STejun Heo goto fail; 1765c34056a3STejun Heo 1766bd7bdd43STejun Heo worker->pool = pool; 1767c34056a3STejun Heo worker->id = id; 1768c34056a3STejun Heo 176929c91e99STejun Heo if (pool->cpu >= 0) 1770e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1771e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1772f3421797STejun Heo else 1773e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1774e3c916a4STejun Heo 1775f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1776e3c916a4STejun Heo "kworker/%s", id_buf); 1777c34056a3STejun Heo if (IS_ERR(worker->task)) 1778c34056a3STejun Heo goto fail; 1779c34056a3STejun Heo 178091151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 178125834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 178291151228SOleg Nesterov 1783da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17844736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1785822d8405STejun Heo 1786051e1850SLai Jiangshan /* start the newly created worker */ 1787051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1788051e1850SLai Jiangshan worker->pool->nr_workers++; 1789051e1850SLai Jiangshan worker_enter_idle(worker); 1790051e1850SLai Jiangshan wake_up_process(worker->task); 1791051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1792051e1850SLai Jiangshan 1793c34056a3STejun Heo return worker; 1794822d8405STejun Heo 1795c34056a3STejun Heo fail: 17969625ab17SLai Jiangshan if (id >= 0) 17977cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1798c34056a3STejun Heo kfree(worker); 1799c34056a3STejun Heo return NULL; 1800c34056a3STejun Heo } 1801c34056a3STejun Heo 1802c34056a3STejun Heo /** 1803c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1804c34056a3STejun Heo * @worker: worker to be destroyed 1805c34056a3STejun Heo * 180673eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 180773eb7fe7SLai Jiangshan * be idle. 1808c8e55f36STejun Heo * 1809c8e55f36STejun Heo * CONTEXT: 181060f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1811c34056a3STejun Heo */ 1812c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1813c34056a3STejun Heo { 1814bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1815c34056a3STejun Heo 1816cd549687STejun Heo lockdep_assert_held(&pool->lock); 1817cd549687STejun Heo 1818c34056a3STejun Heo /* sanity check frenzy */ 18196183c009STejun Heo if (WARN_ON(worker->current_work) || 182073eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 182173eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 18226183c009STejun Heo return; 1823c34056a3STejun Heo 1824bd7bdd43STejun Heo pool->nr_workers--; 1825bd7bdd43STejun Heo pool->nr_idle--; 1826c8e55f36STejun Heo 1827c8e55f36STejun Heo list_del_init(&worker->entry); 1828cb444766STejun Heo worker->flags |= WORKER_DIE; 182960f5a4bcSLai Jiangshan wake_up_process(worker->task); 1830c34056a3STejun Heo } 1831c34056a3STejun Heo 183263d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1833e22bee78STejun Heo { 183463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1835e22bee78STejun Heo 1836d565ed63STejun Heo spin_lock_irq(&pool->lock); 1837e22bee78STejun Heo 18383347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1839e22bee78STejun Heo struct worker *worker; 1840e22bee78STejun Heo unsigned long expires; 1841e22bee78STejun Heo 1842e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 184363d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1844e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1845e22bee78STejun Heo 18463347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 184763d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18483347fc9fSLai Jiangshan break; 1849e22bee78STejun Heo } 18503347fc9fSLai Jiangshan 18513347fc9fSLai Jiangshan destroy_worker(worker); 1852e22bee78STejun Heo } 1853e22bee78STejun Heo 1854d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1855e22bee78STejun Heo } 1856e22bee78STejun Heo 1857493a1724STejun Heo static void send_mayday(struct work_struct *work) 1858e22bee78STejun Heo { 1859112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1860112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1861493a1724STejun Heo 18622e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1863e22bee78STejun Heo 1864493008a8STejun Heo if (!wq->rescuer) 1865493a1724STejun Heo return; 1866e22bee78STejun Heo 1867e22bee78STejun Heo /* mayday mayday mayday */ 1868493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 186977668c8bSLai Jiangshan /* 187077668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 187177668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 187277668c8bSLai Jiangshan * rescuer is done with it. 187377668c8bSLai Jiangshan */ 187477668c8bSLai Jiangshan get_pwq(pwq); 1875493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1876e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1877493a1724STejun Heo } 1878e22bee78STejun Heo } 1879e22bee78STejun Heo 1880706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1881e22bee78STejun Heo { 188263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1883e22bee78STejun Heo struct work_struct *work; 1884e22bee78STejun Heo 1885b2d82909STejun Heo spin_lock_irq(&pool->lock); 1886b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1887e22bee78STejun Heo 188863d95a91STejun Heo if (need_to_create_worker(pool)) { 1889e22bee78STejun Heo /* 1890e22bee78STejun Heo * We've been trying to create a new worker but 1891e22bee78STejun Heo * haven't been successful. We might be hitting an 1892e22bee78STejun Heo * allocation deadlock. Send distress signals to 1893e22bee78STejun Heo * rescuers. 1894e22bee78STejun Heo */ 189563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1896e22bee78STejun Heo send_mayday(work); 1897e22bee78STejun Heo } 1898e22bee78STejun Heo 1899b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1900b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1901e22bee78STejun Heo 190263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1903e22bee78STejun Heo } 1904e22bee78STejun Heo 1905e22bee78STejun Heo /** 1906e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 190763d95a91STejun Heo * @pool: pool to create a new worker for 1908e22bee78STejun Heo * 190963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1910e22bee78STejun Heo * have at least one idle worker on return from this function. If 1911e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 191263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1913e22bee78STejun Heo * possible allocation deadlock. 1914e22bee78STejun Heo * 1915c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1916c5aa87bbSTejun Heo * may_start_working() %true. 1917e22bee78STejun Heo * 1918e22bee78STejun Heo * LOCKING: 1919d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1920e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1921e22bee78STejun Heo * manager. 1922e22bee78STejun Heo */ 192329187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1924d565ed63STejun Heo __releases(&pool->lock) 1925d565ed63STejun Heo __acquires(&pool->lock) 1926e22bee78STejun Heo { 1927e22bee78STejun Heo restart: 1928d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19299f9c2364STejun Heo 1930e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 193163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1932e22bee78STejun Heo 1933e22bee78STejun Heo while (true) { 1934051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1935e22bee78STejun Heo break; 1936e22bee78STejun Heo 1937e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19389f9c2364STejun Heo 193963d95a91STejun Heo if (!need_to_create_worker(pool)) 1940e22bee78STejun Heo break; 1941e22bee78STejun Heo } 1942e22bee78STejun Heo 194363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1944d565ed63STejun Heo spin_lock_irq(&pool->lock); 1945051e1850SLai Jiangshan /* 1946051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1947051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1948051e1850SLai Jiangshan * already become busy. 1949051e1850SLai Jiangshan */ 195063d95a91STejun Heo if (need_to_create_worker(pool)) 1951e22bee78STejun Heo goto restart; 1952e22bee78STejun Heo } 1953e22bee78STejun Heo 1954e22bee78STejun Heo /** 1955e22bee78STejun Heo * manage_workers - manage worker pool 1956e22bee78STejun Heo * @worker: self 1957e22bee78STejun Heo * 1958706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1959e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1960706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1961e22bee78STejun Heo * 1962e22bee78STejun Heo * The caller can safely start processing works on false return. On 1963e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1964e22bee78STejun Heo * and may_start_working() is true. 1965e22bee78STejun Heo * 1966e22bee78STejun Heo * CONTEXT: 1967d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1968e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1969e22bee78STejun Heo * 1970d185af30SYacine Belkadi * Return: 197129187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 197229187a9eSTejun Heo * start processing works, %true if management function was performed and 197329187a9eSTejun Heo * the conditions that the caller verified before calling the function may 197429187a9eSTejun Heo * no longer be true. 1975e22bee78STejun Heo */ 1976e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1977e22bee78STejun Heo { 197863d95a91STejun Heo struct worker_pool *pool = worker->pool; 1979e22bee78STejun Heo 1980bc3a1afcSTejun Heo /* 1981bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1982bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1983bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1984bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1985bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1986bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1987bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1988bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1989bc3a1afcSTejun Heo */ 199034a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 199129187a9eSTejun Heo return false; 19922607d7a6STejun Heo pool->manager = worker; 1993e22bee78STejun Heo 199429187a9eSTejun Heo maybe_create_worker(pool); 1995e22bee78STejun Heo 19962607d7a6STejun Heo pool->manager = NULL; 199734a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 199829187a9eSTejun Heo return true; 1999e22bee78STejun Heo } 2000e22bee78STejun Heo 2001a62428c0STejun Heo /** 2002a62428c0STejun Heo * process_one_work - process single work 2003c34056a3STejun Heo * @worker: self 2004a62428c0STejun Heo * @work: work to process 2005a62428c0STejun Heo * 2006a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2007a62428c0STejun Heo * process a single work including synchronization against and 2008a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2009a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2010a62428c0STejun Heo * call this function to process a work. 2011a62428c0STejun Heo * 2012a62428c0STejun Heo * CONTEXT: 2013d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2014a62428c0STejun Heo */ 2015c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2016d565ed63STejun Heo __releases(&pool->lock) 2017d565ed63STejun Heo __acquires(&pool->lock) 20181da177e4SLinus Torvalds { 2019112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2020bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2021112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 202273f53c4aSTejun Heo int work_color; 20237e11629dSTejun Heo struct worker *collision; 20244e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20254e6045f1SJohannes Berg /* 2026a62428c0STejun Heo * It is permissible to free the struct work_struct from 2027a62428c0STejun Heo * inside the function that is called from it, this we need to 2028a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2029a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2030a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20314e6045f1SJohannes Berg */ 20324d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20334d82a1deSPeter Zijlstra 20344d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20354e6045f1SJohannes Berg #endif 2036807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 203785327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2038ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 203925511a47STejun Heo 20407e11629dSTejun Heo /* 20417e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20427e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20437e11629dSTejun Heo * already processing the work. If so, defer the work to the 20447e11629dSTejun Heo * currently executing one. 20457e11629dSTejun Heo */ 2046c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20477e11629dSTejun Heo if (unlikely(collision)) { 20487e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20497e11629dSTejun Heo return; 20507e11629dSTejun Heo } 20511da177e4SLinus Torvalds 20528930cabaSTejun Heo /* claim and dequeue */ 20531da177e4SLinus Torvalds debug_work_deactivate(work); 2054c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2055c34056a3STejun Heo worker->current_work = work; 2056a2c1c57bSTejun Heo worker->current_func = work->func; 2057112202d9STejun Heo worker->current_pwq = pwq; 205873f53c4aSTejun Heo work_color = get_work_color(work); 20597a22ad75STejun Heo 2060a62428c0STejun Heo list_del_init(&work->entry); 2061a62428c0STejun Heo 2062649027d7STejun Heo /* 2063228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2064228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2065228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2066228f1d00SLai Jiangshan * execution of the pending work items. 2067fb0e7bebSTejun Heo */ 2068fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2069228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2070fb0e7bebSTejun Heo 2071974271c4STejun Heo /* 2072a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2073a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2074a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2075a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2076228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2077974271c4STejun Heo */ 2078a489a03eSLai Jiangshan if (need_more_worker(pool)) 207963d95a91STejun Heo wake_up_worker(pool); 2080974271c4STejun Heo 20818930cabaSTejun Heo /* 20827c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2083d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 208423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 208523657bb1STejun Heo * disabled. 20868930cabaSTejun Heo */ 20877c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20881da177e4SLinus Torvalds 2089d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2090365970a1SDavid Howells 2091112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20923295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2093e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2094a2c1c57bSTejun Heo worker->current_func(work); 2095e36c886aSArjan van de Ven /* 2096e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2097e36c886aSArjan van de Ven * point will only record its address. 2098e36c886aSArjan van de Ven */ 2099e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21003295f0efSIngo Molnar lock_map_release(&lockdep_map); 2101112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21021da177e4SLinus Torvalds 2103d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2104044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2105044c782cSValentin Ilie " last function: %pf\n", 2106a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2107a2c1c57bSTejun Heo worker->current_func); 2108d5abe669SPeter Zijlstra debug_show_held_locks(current); 2109d5abe669SPeter Zijlstra dump_stack(); 2110d5abe669SPeter Zijlstra } 2111d5abe669SPeter Zijlstra 2112b22ce278STejun Heo /* 2113b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2114b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2115b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2116b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2117789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2118789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2119b22ce278STejun Heo */ 21203e28e377SJoe Lawrence cond_resched_rcu_qs(); 2121b22ce278STejun Heo 2122d565ed63STejun Heo spin_lock_irq(&pool->lock); 2123a62428c0STejun Heo 2124fb0e7bebSTejun Heo /* clear cpu intensive status */ 2125fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2126fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2127fb0e7bebSTejun Heo 2128a62428c0STejun Heo /* we're done with it, release */ 212942f8570fSSasha Levin hash_del(&worker->hentry); 2130c34056a3STejun Heo worker->current_work = NULL; 2131a2c1c57bSTejun Heo worker->current_func = NULL; 2132112202d9STejun Heo worker->current_pwq = NULL; 21333d1cb205STejun Heo worker->desc_valid = false; 2134112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21351da177e4SLinus Torvalds } 21361da177e4SLinus Torvalds 2137affee4b2STejun Heo /** 2138affee4b2STejun Heo * process_scheduled_works - process scheduled works 2139affee4b2STejun Heo * @worker: self 2140affee4b2STejun Heo * 2141affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2142affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2143affee4b2STejun Heo * fetches a work from the top and executes it. 2144affee4b2STejun Heo * 2145affee4b2STejun Heo * CONTEXT: 2146d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2147affee4b2STejun Heo * multiple times. 2148affee4b2STejun Heo */ 2149affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21501da177e4SLinus Torvalds { 2151affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2152affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2153a62428c0STejun Heo struct work_struct, entry); 2154c34056a3STejun Heo process_one_work(worker, work); 2155a62428c0STejun Heo } 21561da177e4SLinus Torvalds } 21571da177e4SLinus Torvalds 21584690c4abSTejun Heo /** 21594690c4abSTejun Heo * worker_thread - the worker thread function 2160c34056a3STejun Heo * @__worker: self 21614690c4abSTejun Heo * 2162c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2163c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2164c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2165c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2166c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2167d185af30SYacine Belkadi * 2168d185af30SYacine Belkadi * Return: 0 21694690c4abSTejun Heo */ 2170c34056a3STejun Heo static int worker_thread(void *__worker) 21711da177e4SLinus Torvalds { 2172c34056a3STejun Heo struct worker *worker = __worker; 2173bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21741da177e4SLinus Torvalds 2175e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2176e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2177c8e55f36STejun Heo woke_up: 2178d565ed63STejun Heo spin_lock_irq(&pool->lock); 2179affee4b2STejun Heo 2180a9ab775bSTejun Heo /* am I supposed to die? */ 2181a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2182d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2183a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2184e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 218560f5a4bcSLai Jiangshan 218660f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21877cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 218860f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 218960f5a4bcSLai Jiangshan kfree(worker); 2190c8e55f36STejun Heo return 0; 2191c8e55f36STejun Heo } 2192c8e55f36STejun Heo 2193c8e55f36STejun Heo worker_leave_idle(worker); 2194db7bccf4STejun Heo recheck: 2195e22bee78STejun Heo /* no more worker necessary? */ 219663d95a91STejun Heo if (!need_more_worker(pool)) 2197e22bee78STejun Heo goto sleep; 2198e22bee78STejun Heo 2199e22bee78STejun Heo /* do we need to manage? */ 220063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2201e22bee78STejun Heo goto recheck; 2202e22bee78STejun Heo 2203c8e55f36STejun Heo /* 2204c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2205c8e55f36STejun Heo * preparing to process a work or actually processing it. 2206c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2207c8e55f36STejun Heo */ 22086183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2209c8e55f36STejun Heo 2210e22bee78STejun Heo /* 2211a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2212a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2213a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2214a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2215a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2216e22bee78STejun Heo */ 2217a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2218e22bee78STejun Heo 2219e22bee78STejun Heo do { 2220affee4b2STejun Heo struct work_struct *work = 2221bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2222affee4b2STejun Heo struct work_struct, entry); 2223affee4b2STejun Heo 222482607adcSTejun Heo pool->watchdog_ts = jiffies; 222582607adcSTejun Heo 2226c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2227affee4b2STejun Heo /* optimization path, not strictly necessary */ 2228affee4b2STejun Heo process_one_work(worker, work); 2229affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2230affee4b2STejun Heo process_scheduled_works(worker); 2231affee4b2STejun Heo } else { 2232c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2233affee4b2STejun Heo process_scheduled_works(worker); 2234affee4b2STejun Heo } 223563d95a91STejun Heo } while (keep_working(pool)); 2236affee4b2STejun Heo 2237228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2238d313dd85STejun Heo sleep: 2239c8e55f36STejun Heo /* 2240d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2241d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2242d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2243d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2244d565ed63STejun Heo * event. 2245c8e55f36STejun Heo */ 2246c8e55f36STejun Heo worker_enter_idle(worker); 2247c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2248d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22491da177e4SLinus Torvalds schedule(); 2250c8e55f36STejun Heo goto woke_up; 22511da177e4SLinus Torvalds } 22521da177e4SLinus Torvalds 2253e22bee78STejun Heo /** 2254e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2255111c225aSTejun Heo * @__rescuer: self 2256e22bee78STejun Heo * 2257e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2258493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2259e22bee78STejun Heo * 2260706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2261e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2262e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2263e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2264e22bee78STejun Heo * the problem rescuer solves. 2265e22bee78STejun Heo * 2266706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2267706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2268e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2269e22bee78STejun Heo * 2270e22bee78STejun Heo * This should happen rarely. 2271d185af30SYacine Belkadi * 2272d185af30SYacine Belkadi * Return: 0 2273e22bee78STejun Heo */ 2274111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2275e22bee78STejun Heo { 2276111c225aSTejun Heo struct worker *rescuer = __rescuer; 2277111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2278e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22794d595b86SLai Jiangshan bool should_stop; 2280e22bee78STejun Heo 2281e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2282111c225aSTejun Heo 2283111c225aSTejun Heo /* 2284111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2285111c225aSTejun Heo * doesn't participate in concurrency management. 2286111c225aSTejun Heo */ 2287111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2288e22bee78STejun Heo repeat: 2289e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22901da177e4SLinus Torvalds 22914d595b86SLai Jiangshan /* 22924d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22934d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22944d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22954d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22964d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22974d595b86SLai Jiangshan * list is always empty on exit. 22984d595b86SLai Jiangshan */ 22994d595b86SLai Jiangshan should_stop = kthread_should_stop(); 23001da177e4SLinus Torvalds 2301493a1724STejun Heo /* see whether any pwq is asking for help */ 23022e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2303493a1724STejun Heo 2304493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2305493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2306493a1724STejun Heo struct pool_workqueue, mayday_node); 2307112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2308e22bee78STejun Heo struct work_struct *work, *n; 230982607adcSTejun Heo bool first = true; 2310e22bee78STejun Heo 2311e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2312493a1724STejun Heo list_del_init(&pwq->mayday_node); 2313493a1724STejun Heo 23142e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2315e22bee78STejun Heo 231651697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 231751697d39SLai Jiangshan 231851697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2319b3104104SLai Jiangshan rescuer->pool = pool; 2320e22bee78STejun Heo 2321e22bee78STejun Heo /* 2322e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2323e22bee78STejun Heo * process'em. 2324e22bee78STejun Heo */ 23250479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 232682607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 232782607adcSTejun Heo if (get_work_pwq(work) == pwq) { 232882607adcSTejun Heo if (first) 232982607adcSTejun Heo pool->watchdog_ts = jiffies; 2330e22bee78STejun Heo move_linked_works(work, scheduled, &n); 233182607adcSTejun Heo } 233282607adcSTejun Heo first = false; 233382607adcSTejun Heo } 2334e22bee78STejun Heo 2335008847f6SNeilBrown if (!list_empty(scheduled)) { 2336e22bee78STejun Heo process_scheduled_works(rescuer); 23377576958aSTejun Heo 23387576958aSTejun Heo /* 2339008847f6SNeilBrown * The above execution of rescued work items could 2340008847f6SNeilBrown * have created more to rescue through 2341008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2342008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2343008847f6SNeilBrown * that such back-to-back work items, which may be 2344008847f6SNeilBrown * being used to relieve memory pressure, don't 2345008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2346008847f6SNeilBrown */ 2347008847f6SNeilBrown if (need_to_create_worker(pool)) { 2348008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2349008847f6SNeilBrown get_pwq(pwq); 2350008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2351008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2352008847f6SNeilBrown } 2353008847f6SNeilBrown } 2354008847f6SNeilBrown 2355008847f6SNeilBrown /* 235677668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 235713b1d625SLai Jiangshan * go away while we're still attached to it. 235877668c8bSLai Jiangshan */ 235977668c8bSLai Jiangshan put_pwq(pwq); 236077668c8bSLai Jiangshan 236177668c8bSLai Jiangshan /* 2362d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 23637576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23647576958aSTejun Heo * and stalling the execution. 23657576958aSTejun Heo */ 2366d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 236763d95a91STejun Heo wake_up_worker(pool); 23687576958aSTejun Heo 2369b3104104SLai Jiangshan rescuer->pool = NULL; 237013b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 237113b1d625SLai Jiangshan 237213b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 237313b1d625SLai Jiangshan 237413b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 23751da177e4SLinus Torvalds } 23761da177e4SLinus Torvalds 23772e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2378493a1724STejun Heo 23794d595b86SLai Jiangshan if (should_stop) { 23804d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23814d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23824d595b86SLai Jiangshan return 0; 23834d595b86SLai Jiangshan } 23844d595b86SLai Jiangshan 2385111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2386111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2387e22bee78STejun Heo schedule(); 2388e22bee78STejun Heo goto repeat; 23891da177e4SLinus Torvalds } 23901da177e4SLinus Torvalds 2391fca839c0STejun Heo /** 2392fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2393fca839c0STejun Heo * @target_wq: workqueue being flushed 2394fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2395fca839c0STejun Heo * 2396fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2397fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2398fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2399fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2400fca839c0STejun Heo * a deadlock. 2401fca839c0STejun Heo */ 2402fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2403fca839c0STejun Heo struct work_struct *target_work) 2404fca839c0STejun Heo { 2405fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2406fca839c0STejun Heo struct worker *worker; 2407fca839c0STejun Heo 2408fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2409fca839c0STejun Heo return; 2410fca839c0STejun Heo 2411fca839c0STejun Heo worker = current_wq_worker(); 2412fca839c0STejun Heo 2413fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2414fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2415fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 241623d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 241723d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2418fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2419fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2420fca839c0STejun Heo target_wq->name, target_func); 2421fca839c0STejun Heo } 2422fca839c0STejun Heo 2423fc2e4d70SOleg Nesterov struct wq_barrier { 2424fc2e4d70SOleg Nesterov struct work_struct work; 2425fc2e4d70SOleg Nesterov struct completion done; 24262607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2427fc2e4d70SOleg Nesterov }; 2428fc2e4d70SOleg Nesterov 2429fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2430fc2e4d70SOleg Nesterov { 2431fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2432fc2e4d70SOleg Nesterov complete(&barr->done); 2433fc2e4d70SOleg Nesterov } 2434fc2e4d70SOleg Nesterov 24354690c4abSTejun Heo /** 24364690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2437112202d9STejun Heo * @pwq: pwq to insert barrier into 24384690c4abSTejun Heo * @barr: wq_barrier to insert 2439affee4b2STejun Heo * @target: target work to attach @barr to 2440affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24414690c4abSTejun Heo * 2442affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2443affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2444affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2445affee4b2STejun Heo * cpu. 2446affee4b2STejun Heo * 2447affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2448affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2449affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2450affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2451affee4b2STejun Heo * after a work with LINKED flag set. 2452affee4b2STejun Heo * 2453affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2454112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24554690c4abSTejun Heo * 24564690c4abSTejun Heo * CONTEXT: 2457d565ed63STejun Heo * spin_lock_irq(pool->lock). 24584690c4abSTejun Heo */ 2459112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2460affee4b2STejun Heo struct wq_barrier *barr, 2461affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2462fc2e4d70SOleg Nesterov { 2463affee4b2STejun Heo struct list_head *head; 2464affee4b2STejun Heo unsigned int linked = 0; 2465affee4b2STejun Heo 2466dc186ad7SThomas Gleixner /* 2467d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2468dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2469dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2470dc186ad7SThomas Gleixner * might deadlock. 2471dc186ad7SThomas Gleixner */ 2472ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 247322df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2474fc2e4d70SOleg Nesterov init_completion(&barr->done); 24752607d7a6STejun Heo barr->task = current; 247683c22520SOleg Nesterov 2477affee4b2STejun Heo /* 2478affee4b2STejun Heo * If @target is currently being executed, schedule the 2479affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2480affee4b2STejun Heo */ 2481affee4b2STejun Heo if (worker) 2482affee4b2STejun Heo head = worker->scheduled.next; 2483affee4b2STejun Heo else { 2484affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2485affee4b2STejun Heo 2486affee4b2STejun Heo head = target->entry.next; 2487affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2488affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2489affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2490affee4b2STejun Heo } 2491affee4b2STejun Heo 2492dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2493112202d9STejun Heo insert_work(pwq, &barr->work, head, 2494affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2495fc2e4d70SOleg Nesterov } 2496fc2e4d70SOleg Nesterov 249773f53c4aSTejun Heo /** 2498112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 249973f53c4aSTejun Heo * @wq: workqueue being flushed 250073f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 250173f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 250273f53c4aSTejun Heo * 2503112202d9STejun Heo * Prepare pwqs for workqueue flushing. 250473f53c4aSTejun Heo * 2505112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2506112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2507112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2508112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2509112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 251073f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 251173f53c4aSTejun Heo * 251273f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 251373f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 251473f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 251573f53c4aSTejun Heo * is returned. 251673f53c4aSTejun Heo * 2517112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 251873f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 251973f53c4aSTejun Heo * advanced to @work_color. 252073f53c4aSTejun Heo * 252173f53c4aSTejun Heo * CONTEXT: 25223c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 252373f53c4aSTejun Heo * 2524d185af30SYacine Belkadi * Return: 252573f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 252673f53c4aSTejun Heo * otherwise. 252773f53c4aSTejun Heo */ 2528112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 252973f53c4aSTejun Heo int flush_color, int work_color) 25301da177e4SLinus Torvalds { 253173f53c4aSTejun Heo bool wait = false; 253249e3cf44STejun Heo struct pool_workqueue *pwq; 25331da177e4SLinus Torvalds 253473f53c4aSTejun Heo if (flush_color >= 0) { 25356183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2536112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2537dc186ad7SThomas Gleixner } 253814441960SOleg Nesterov 253949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2540112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25411da177e4SLinus Torvalds 2542b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 254373f53c4aSTejun Heo 254473f53c4aSTejun Heo if (flush_color >= 0) { 25456183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 254673f53c4aSTejun Heo 2547112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2548112202d9STejun Heo pwq->flush_color = flush_color; 2549112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 255073f53c4aSTejun Heo wait = true; 25511da177e4SLinus Torvalds } 255273f53c4aSTejun Heo } 255373f53c4aSTejun Heo 255473f53c4aSTejun Heo if (work_color >= 0) { 25556183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2556112202d9STejun Heo pwq->work_color = work_color; 255773f53c4aSTejun Heo } 255873f53c4aSTejun Heo 2559b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25601da177e4SLinus Torvalds } 25611da177e4SLinus Torvalds 2562112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 256373f53c4aSTejun Heo complete(&wq->first_flusher->done); 256473f53c4aSTejun Heo 256573f53c4aSTejun Heo return wait; 256683c22520SOleg Nesterov } 25671da177e4SLinus Torvalds 25680fcb78c2SRolf Eike Beer /** 25691da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25700fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25711da177e4SLinus Torvalds * 2572c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2573c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25741da177e4SLinus Torvalds */ 25757ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25761da177e4SLinus Torvalds { 257773f53c4aSTejun Heo struct wq_flusher this_flusher = { 257873f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 257973f53c4aSTejun Heo .flush_color = -1, 258073f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 258173f53c4aSTejun Heo }; 258273f53c4aSTejun Heo int next_color; 2583b1f4ec17SOleg Nesterov 25843295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25853295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 258673f53c4aSTejun Heo 25873c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 258873f53c4aSTejun Heo 258973f53c4aSTejun Heo /* 259073f53c4aSTejun Heo * Start-to-wait phase 259173f53c4aSTejun Heo */ 259273f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 259373f53c4aSTejun Heo 259473f53c4aSTejun Heo if (next_color != wq->flush_color) { 259573f53c4aSTejun Heo /* 259673f53c4aSTejun Heo * Color space is not full. The current work_color 259773f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 259873f53c4aSTejun Heo * by one. 259973f53c4aSTejun Heo */ 26006183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 260173f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 260273f53c4aSTejun Heo wq->work_color = next_color; 260373f53c4aSTejun Heo 260473f53c4aSTejun Heo if (!wq->first_flusher) { 260573f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26066183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 260773f53c4aSTejun Heo 260873f53c4aSTejun Heo wq->first_flusher = &this_flusher; 260973f53c4aSTejun Heo 2610112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 261173f53c4aSTejun Heo wq->work_color)) { 261273f53c4aSTejun Heo /* nothing to flush, done */ 261373f53c4aSTejun Heo wq->flush_color = next_color; 261473f53c4aSTejun Heo wq->first_flusher = NULL; 261573f53c4aSTejun Heo goto out_unlock; 261673f53c4aSTejun Heo } 261773f53c4aSTejun Heo } else { 261873f53c4aSTejun Heo /* wait in queue */ 26196183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 262073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2621112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 262273f53c4aSTejun Heo } 262373f53c4aSTejun Heo } else { 262473f53c4aSTejun Heo /* 262573f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 262673f53c4aSTejun Heo * The next flush completion will assign us 262773f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 262873f53c4aSTejun Heo */ 262973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 263073f53c4aSTejun Heo } 263173f53c4aSTejun Heo 2632fca839c0STejun Heo check_flush_dependency(wq, NULL); 2633fca839c0STejun Heo 26343c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 263573f53c4aSTejun Heo 263673f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo /* 263973f53c4aSTejun Heo * Wake-up-and-cascade phase 264073f53c4aSTejun Heo * 264173f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 264273f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 264373f53c4aSTejun Heo */ 264473f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 264573f53c4aSTejun Heo return; 264673f53c4aSTejun Heo 26473c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 264873f53c4aSTejun Heo 26494ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26504ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26514ce48b37STejun Heo goto out_unlock; 26524ce48b37STejun Heo 265373f53c4aSTejun Heo wq->first_flusher = NULL; 265473f53c4aSTejun Heo 26556183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26566183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 265773f53c4aSTejun Heo 265873f53c4aSTejun Heo while (true) { 265973f53c4aSTejun Heo struct wq_flusher *next, *tmp; 266073f53c4aSTejun Heo 266173f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 266273f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 266373f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 266473f53c4aSTejun Heo break; 266573f53c4aSTejun Heo list_del_init(&next->list); 266673f53c4aSTejun Heo complete(&next->done); 266773f53c4aSTejun Heo } 266873f53c4aSTejun Heo 26696183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 267073f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 267173f53c4aSTejun Heo 267273f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 267373f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 267473f53c4aSTejun Heo 267573f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 267673f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 267773f53c4aSTejun Heo /* 267873f53c4aSTejun Heo * Assign the same color to all overflowed 267973f53c4aSTejun Heo * flushers, advance work_color and append to 268073f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 268173f53c4aSTejun Heo * phase for these overflowed flushers. 268273f53c4aSTejun Heo */ 268373f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 268473f53c4aSTejun Heo tmp->flush_color = wq->work_color; 268573f53c4aSTejun Heo 268673f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 268773f53c4aSTejun Heo 268873f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 268973f53c4aSTejun Heo &wq->flusher_queue); 2690112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 269173f53c4aSTejun Heo } 269273f53c4aSTejun Heo 269373f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26946183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 269573f53c4aSTejun Heo break; 269673f53c4aSTejun Heo } 269773f53c4aSTejun Heo 269873f53c4aSTejun Heo /* 269973f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2700112202d9STejun Heo * the new first flusher and arm pwqs. 270173f53c4aSTejun Heo */ 27026183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27036183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 270473f53c4aSTejun Heo 270573f53c4aSTejun Heo list_del_init(&next->list); 270673f53c4aSTejun Heo wq->first_flusher = next; 270773f53c4aSTejun Heo 2708112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 270973f53c4aSTejun Heo break; 271073f53c4aSTejun Heo 271173f53c4aSTejun Heo /* 271273f53c4aSTejun Heo * Meh... this color is already done, clear first 271373f53c4aSTejun Heo * flusher and repeat cascading. 271473f53c4aSTejun Heo */ 271573f53c4aSTejun Heo wq->first_flusher = NULL; 271673f53c4aSTejun Heo } 271773f53c4aSTejun Heo 271873f53c4aSTejun Heo out_unlock: 27193c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27201da177e4SLinus Torvalds } 27211dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 27221da177e4SLinus Torvalds 27239c5a2ba7STejun Heo /** 27249c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27259c5a2ba7STejun Heo * @wq: workqueue to drain 27269c5a2ba7STejun Heo * 27279c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27289c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27299c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2730b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 27319c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27329c5a2ba7STejun Heo * takes too long. 27339c5a2ba7STejun Heo */ 27349c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27359c5a2ba7STejun Heo { 27369c5a2ba7STejun Heo unsigned int flush_cnt = 0; 273749e3cf44STejun Heo struct pool_workqueue *pwq; 27389c5a2ba7STejun Heo 27399c5a2ba7STejun Heo /* 27409c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27419c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2742618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27439c5a2ba7STejun Heo */ 274487fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27459c5a2ba7STejun Heo if (!wq->nr_drainers++) 2746618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 274787fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27489c5a2ba7STejun Heo reflush: 27499c5a2ba7STejun Heo flush_workqueue(wq); 27509c5a2ba7STejun Heo 2751b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 275276af4d93STejun Heo 275349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2754fa2563e4SThomas Tuttle bool drained; 27559c5a2ba7STejun Heo 2756b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2757112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2758b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2759fa2563e4SThomas Tuttle 2760fa2563e4SThomas Tuttle if (drained) 27619c5a2ba7STejun Heo continue; 27629c5a2ba7STejun Heo 27639c5a2ba7STejun Heo if (++flush_cnt == 10 || 27649c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2765c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27669c5a2ba7STejun Heo wq->name, flush_cnt); 276776af4d93STejun Heo 2768b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27699c5a2ba7STejun Heo goto reflush; 27709c5a2ba7STejun Heo } 27719c5a2ba7STejun Heo 27729c5a2ba7STejun Heo if (!--wq->nr_drainers) 2773618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 277487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27759c5a2ba7STejun Heo } 27769c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27779c5a2ba7STejun Heo 2778606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2779baf59022STejun Heo { 2780baf59022STejun Heo struct worker *worker = NULL; 2781c9e7cf27STejun Heo struct worker_pool *pool; 2782112202d9STejun Heo struct pool_workqueue *pwq; 2783baf59022STejun Heo 2784baf59022STejun Heo might_sleep(); 2785baf59022STejun Heo 2786fa1b54e6STejun Heo local_irq_disable(); 2787fa1b54e6STejun Heo pool = get_work_pool(work); 2788fa1b54e6STejun Heo if (!pool) { 2789fa1b54e6STejun Heo local_irq_enable(); 2790fa1b54e6STejun Heo return false; 2791fa1b54e6STejun Heo } 2792fa1b54e6STejun Heo 2793fa1b54e6STejun Heo spin_lock(&pool->lock); 27940b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2795112202d9STejun Heo pwq = get_work_pwq(work); 2796112202d9STejun Heo if (pwq) { 2797112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2798baf59022STejun Heo goto already_gone; 2799606a5020STejun Heo } else { 2800c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2801baf59022STejun Heo if (!worker) 2802baf59022STejun Heo goto already_gone; 2803112202d9STejun Heo pwq = worker->current_pwq; 2804606a5020STejun Heo } 2805baf59022STejun Heo 2806fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2807fca839c0STejun Heo 2808112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2809d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2810baf59022STejun Heo 2811e159489bSTejun Heo /* 2812e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2813e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2814e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2815e159489bSTejun Heo * access. 2816e159489bSTejun Heo */ 2817493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2818112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2819e159489bSTejun Heo else 2820112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2821112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2822e159489bSTejun Heo 2823baf59022STejun Heo return true; 2824baf59022STejun Heo already_gone: 2825d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2826baf59022STejun Heo return false; 2827baf59022STejun Heo } 2828baf59022STejun Heo 2829db700897SOleg Nesterov /** 2830401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2831401a8d04STejun Heo * @work: the work to flush 2832db700897SOleg Nesterov * 2833606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2834606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2835401a8d04STejun Heo * 2836d185af30SYacine Belkadi * Return: 2837401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2838401a8d04STejun Heo * %false if it was already idle. 2839db700897SOleg Nesterov */ 2840401a8d04STejun Heo bool flush_work(struct work_struct *work) 2841db700897SOleg Nesterov { 284212997d1aSBjorn Helgaas struct wq_barrier barr; 284312997d1aSBjorn Helgaas 28440976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28450976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28460976dfc1SStephen Boyd 284712997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 284812997d1aSBjorn Helgaas wait_for_completion(&barr.done); 284912997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 285012997d1aSBjorn Helgaas return true; 285112997d1aSBjorn Helgaas } else { 285212997d1aSBjorn Helgaas return false; 285312997d1aSBjorn Helgaas } 2854606a5020STejun Heo } 2855db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2856db700897SOleg Nesterov 28578603e1b3STejun Heo struct cwt_wait { 28588603e1b3STejun Heo wait_queue_t wait; 28598603e1b3STejun Heo struct work_struct *work; 28608603e1b3STejun Heo }; 28618603e1b3STejun Heo 28628603e1b3STejun Heo static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) 28638603e1b3STejun Heo { 28648603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 28658603e1b3STejun Heo 28668603e1b3STejun Heo if (cwait->work != key) 28678603e1b3STejun Heo return 0; 28688603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 28698603e1b3STejun Heo } 28708603e1b3STejun Heo 287136e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2872401a8d04STejun Heo { 28738603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2874bbb68dfaSTejun Heo unsigned long flags; 28751f1f642eSOleg Nesterov int ret; 28761f1f642eSOleg Nesterov 28771f1f642eSOleg Nesterov do { 2878bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2879bbb68dfaSTejun Heo /* 28808603e1b3STejun Heo * If someone else is already canceling, wait for it to 28818603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 28828603e1b3STejun Heo * because we may get scheduled between @work's completion 28838603e1b3STejun Heo * and the other canceling task resuming and clearing 28848603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 28858603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 28868603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 28878603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 28888603e1b3STejun Heo * we're hogging the CPU. 28898603e1b3STejun Heo * 28908603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 28918603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 28928603e1b3STejun Heo * wake function which matches @work along with exclusive 28938603e1b3STejun Heo * wait and wakeup. 2894bbb68dfaSTejun Heo */ 28958603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 28968603e1b3STejun Heo struct cwt_wait cwait; 28978603e1b3STejun Heo 28988603e1b3STejun Heo init_wait(&cwait.wait); 28998603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 29008603e1b3STejun Heo cwait.work = work; 29018603e1b3STejun Heo 29028603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 29038603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 29048603e1b3STejun Heo if (work_is_canceling(work)) 29058603e1b3STejun Heo schedule(); 29068603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 29078603e1b3STejun Heo } 29081f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29091f1f642eSOleg Nesterov 2910bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2911bbb68dfaSTejun Heo mark_work_canceling(work); 2912bbb68dfaSTejun Heo local_irq_restore(flags); 2913bbb68dfaSTejun Heo 2914606a5020STejun Heo flush_work(work); 29157a22ad75STejun Heo clear_work_data(work); 29168603e1b3STejun Heo 29178603e1b3STejun Heo /* 29188603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 29198603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 29208603e1b3STejun Heo * visible there. 29218603e1b3STejun Heo */ 29228603e1b3STejun Heo smp_mb(); 29238603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 29248603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 29258603e1b3STejun Heo 29261f1f642eSOleg Nesterov return ret; 29271f1f642eSOleg Nesterov } 29281f1f642eSOleg Nesterov 29296e84d644SOleg Nesterov /** 2930401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2931401a8d04STejun Heo * @work: the work to cancel 29326e84d644SOleg Nesterov * 2933401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2934401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2935401a8d04STejun Heo * another workqueue. On return from this function, @work is 2936401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29371f1f642eSOleg Nesterov * 2938401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2939401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29406e84d644SOleg Nesterov * 2941401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29426e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2943401a8d04STejun Heo * 2944d185af30SYacine Belkadi * Return: 2945401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29466e84d644SOleg Nesterov */ 2947401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29486e84d644SOleg Nesterov { 294936e227d2STejun Heo return __cancel_work_timer(work, false); 2950b89deed3SOleg Nesterov } 295128e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2952b89deed3SOleg Nesterov 29536e84d644SOleg Nesterov /** 2954401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2955401a8d04STejun Heo * @dwork: the delayed work to flush 29566e84d644SOleg Nesterov * 2957401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2958401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2959401a8d04STejun Heo * considers the last queueing instance of @dwork. 29601f1f642eSOleg Nesterov * 2961d185af30SYacine Belkadi * Return: 2962401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2963401a8d04STejun Heo * %false if it was already idle. 29646e84d644SOleg Nesterov */ 2965401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2966401a8d04STejun Heo { 29678930cabaSTejun Heo local_irq_disable(); 2968401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 296960c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29708930cabaSTejun Heo local_irq_enable(); 2971401a8d04STejun Heo return flush_work(&dwork->work); 2972401a8d04STejun Heo } 2973401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2974401a8d04STejun Heo 2975401a8d04STejun Heo /** 297657b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 297757b30ae7STejun Heo * @dwork: delayed_work to cancel 297809383498STejun Heo * 2979d185af30SYacine Belkadi * Kill off a pending delayed_work. 2980d185af30SYacine Belkadi * 2981d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2982d185af30SYacine Belkadi * pending. 2983d185af30SYacine Belkadi * 2984d185af30SYacine Belkadi * Note: 2985d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2986d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2987d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 298809383498STejun Heo * 298957b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 299009383498STejun Heo */ 299157b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 299209383498STejun Heo { 299357b30ae7STejun Heo unsigned long flags; 299457b30ae7STejun Heo int ret; 299557b30ae7STejun Heo 299657b30ae7STejun Heo do { 299757b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 299857b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 299957b30ae7STejun Heo 300057b30ae7STejun Heo if (unlikely(ret < 0)) 300157b30ae7STejun Heo return false; 300257b30ae7STejun Heo 30037c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 30047c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 300557b30ae7STejun Heo local_irq_restore(flags); 3006c0158ca6SDan Magenheimer return ret; 300709383498STejun Heo } 300857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 300909383498STejun Heo 301009383498STejun Heo /** 3011401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3012401a8d04STejun Heo * @dwork: the delayed work cancel 3013401a8d04STejun Heo * 3014401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3015401a8d04STejun Heo * 3016d185af30SYacine Belkadi * Return: 3017401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3018401a8d04STejun Heo */ 3019401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30206e84d644SOleg Nesterov { 302136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30226e84d644SOleg Nesterov } 3023f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30241da177e4SLinus Torvalds 30250fcb78c2SRolf Eike Beer /** 302631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3027b6136773SAndrew Morton * @func: the function to call 3028b6136773SAndrew Morton * 302931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 303031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3031b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 303231ddd871STejun Heo * 3033d185af30SYacine Belkadi * Return: 303431ddd871STejun Heo * 0 on success, -errno on failure. 3035b6136773SAndrew Morton */ 303665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 303715316ba8SChristoph Lameter { 303815316ba8SChristoph Lameter int cpu; 303938f51568SNamhyung Kim struct work_struct __percpu *works; 304015316ba8SChristoph Lameter 3041b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3042b6136773SAndrew Morton if (!works) 304315316ba8SChristoph Lameter return -ENOMEM; 3044b6136773SAndrew Morton 304595402b38SGautham R Shenoy get_online_cpus(); 304693981800STejun Heo 304715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30489bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30499bfb1839SIngo Molnar 30509bfb1839SIngo Molnar INIT_WORK(work, func); 30518de6d308SOleg Nesterov schedule_work_on(cpu, work); 305215316ba8SChristoph Lameter } 305393981800STejun Heo 305493981800STejun Heo for_each_online_cpu(cpu) 30558616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 305693981800STejun Heo 305795402b38SGautham R Shenoy put_online_cpus(); 3058b6136773SAndrew Morton free_percpu(works); 305915316ba8SChristoph Lameter return 0; 306015316ba8SChristoph Lameter } 306115316ba8SChristoph Lameter 3062eef6a7d5SAlan Stern /** 30631fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30641fa44ecaSJames Bottomley * @fn: the function to execute 30651fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30661fa44ecaSJames Bottomley * be available when the work executes) 30671fa44ecaSJames Bottomley * 30681fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30691fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30701fa44ecaSJames Bottomley * 3071d185af30SYacine Belkadi * Return: 0 - function was executed 30721fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30731fa44ecaSJames Bottomley */ 307465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30751fa44ecaSJames Bottomley { 30761fa44ecaSJames Bottomley if (!in_interrupt()) { 307765f27f38SDavid Howells fn(&ew->work); 30781fa44ecaSJames Bottomley return 0; 30791fa44ecaSJames Bottomley } 30801fa44ecaSJames Bottomley 308165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30821fa44ecaSJames Bottomley schedule_work(&ew->work); 30831fa44ecaSJames Bottomley 30841fa44ecaSJames Bottomley return 1; 30851fa44ecaSJames Bottomley } 30861fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30871fa44ecaSJames Bottomley 30887a4e344cSTejun Heo /** 30897a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 30907a4e344cSTejun Heo * @attrs: workqueue_attrs to free 30917a4e344cSTejun Heo * 30927a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 30937a4e344cSTejun Heo */ 30947a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 30957a4e344cSTejun Heo { 30967a4e344cSTejun Heo if (attrs) { 30977a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 30987a4e344cSTejun Heo kfree(attrs); 30997a4e344cSTejun Heo } 31007a4e344cSTejun Heo } 31017a4e344cSTejun Heo 31027a4e344cSTejun Heo /** 31037a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 31047a4e344cSTejun Heo * @gfp_mask: allocation mask to use 31057a4e344cSTejun Heo * 31067a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3107d185af30SYacine Belkadi * return it. 3108d185af30SYacine Belkadi * 3109d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 31107a4e344cSTejun Heo */ 31117a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 31127a4e344cSTejun Heo { 31137a4e344cSTejun Heo struct workqueue_attrs *attrs; 31147a4e344cSTejun Heo 31157a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 31167a4e344cSTejun Heo if (!attrs) 31177a4e344cSTejun Heo goto fail; 31187a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 31197a4e344cSTejun Heo goto fail; 31207a4e344cSTejun Heo 312113e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 31227a4e344cSTejun Heo return attrs; 31237a4e344cSTejun Heo fail: 31247a4e344cSTejun Heo free_workqueue_attrs(attrs); 31257a4e344cSTejun Heo return NULL; 31267a4e344cSTejun Heo } 31277a4e344cSTejun Heo 312829c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 312929c91e99STejun Heo const struct workqueue_attrs *from) 313029c91e99STejun Heo { 313129c91e99STejun Heo to->nice = from->nice; 313229c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 31332865a8fbSShaohua Li /* 31342865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 31352865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 31362865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 31372865a8fbSShaohua Li */ 31382865a8fbSShaohua Li to->no_numa = from->no_numa; 313929c91e99STejun Heo } 314029c91e99STejun Heo 314129c91e99STejun Heo /* hash value of the content of @attr */ 314229c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 314329c91e99STejun Heo { 314429c91e99STejun Heo u32 hash = 0; 314529c91e99STejun Heo 314629c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 314713e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 314813e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 314929c91e99STejun Heo return hash; 315029c91e99STejun Heo } 315129c91e99STejun Heo 315229c91e99STejun Heo /* content equality test */ 315329c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 315429c91e99STejun Heo const struct workqueue_attrs *b) 315529c91e99STejun Heo { 315629c91e99STejun Heo if (a->nice != b->nice) 315729c91e99STejun Heo return false; 315829c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 315929c91e99STejun Heo return false; 316029c91e99STejun Heo return true; 316129c91e99STejun Heo } 316229c91e99STejun Heo 31637a4e344cSTejun Heo /** 31647a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 31657a4e344cSTejun Heo * @pool: worker_pool to initialize 31667a4e344cSTejun Heo * 3167402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3168d185af30SYacine Belkadi * 3169d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 317029c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 317129c91e99STejun Heo * on @pool safely to release it. 31727a4e344cSTejun Heo */ 31737a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 31744e1a1f9aSTejun Heo { 31754e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 317629c91e99STejun Heo pool->id = -1; 317729c91e99STejun Heo pool->cpu = -1; 3178f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 31794e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 318082607adcSTejun Heo pool->watchdog_ts = jiffies; 31814e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 31824e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 31834e1a1f9aSTejun Heo hash_init(pool->busy_hash); 31844e1a1f9aSTejun Heo 31854e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 31864e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 31874e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 31884e1a1f9aSTejun Heo 31894e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 31904e1a1f9aSTejun Heo (unsigned long)pool); 31914e1a1f9aSTejun Heo 31924e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 319392f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3194da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 31957a4e344cSTejun Heo 31967cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 319729c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 319829c91e99STejun Heo pool->refcnt = 1; 319929c91e99STejun Heo 320029c91e99STejun Heo /* shouldn't fail above this point */ 32017a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 32027a4e344cSTejun Heo if (!pool->attrs) 32037a4e344cSTejun Heo return -ENOMEM; 32047a4e344cSTejun Heo return 0; 32054e1a1f9aSTejun Heo } 32064e1a1f9aSTejun Heo 3207e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3208e2dca7adSTejun Heo { 3209e2dca7adSTejun Heo struct workqueue_struct *wq = 3210e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3211e2dca7adSTejun Heo 3212e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3213e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3214e2dca7adSTejun Heo else 3215e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3216e2dca7adSTejun Heo 3217e2dca7adSTejun Heo kfree(wq->rescuer); 3218e2dca7adSTejun Heo kfree(wq); 3219e2dca7adSTejun Heo } 3220e2dca7adSTejun Heo 322129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 322229c91e99STejun Heo { 322329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 322429c91e99STejun Heo 32257cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 322629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 322729c91e99STejun Heo kfree(pool); 322829c91e99STejun Heo } 322929c91e99STejun Heo 323029c91e99STejun Heo /** 323129c91e99STejun Heo * put_unbound_pool - put a worker_pool 323229c91e99STejun Heo * @pool: worker_pool to put 323329c91e99STejun Heo * 323429c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3235c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3236c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3237c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3238a892caccSTejun Heo * 3239a892caccSTejun Heo * Should be called with wq_pool_mutex held. 324029c91e99STejun Heo */ 324129c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 324229c91e99STejun Heo { 324360f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 324429c91e99STejun Heo struct worker *worker; 324529c91e99STejun Heo 3246a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3247a892caccSTejun Heo 3248a892caccSTejun Heo if (--pool->refcnt) 324929c91e99STejun Heo return; 325029c91e99STejun Heo 325129c91e99STejun Heo /* sanity checks */ 325261d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3253a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 325429c91e99STejun Heo return; 325529c91e99STejun Heo 325629c91e99STejun Heo /* release id and unhash */ 325729c91e99STejun Heo if (pool->id >= 0) 325829c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 325929c91e99STejun Heo hash_del(&pool->hash_node); 326029c91e99STejun Heo 3261c5aa87bbSTejun Heo /* 3262c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3263c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 326492f9c5c4SLai Jiangshan * attach_mutex. 3265c5aa87bbSTejun Heo */ 326629c91e99STejun Heo mutex_lock(&pool->manager_arb); 326729c91e99STejun Heo 326860f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 32691037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 327029c91e99STejun Heo destroy_worker(worker); 327129c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 327229c91e99STejun Heo spin_unlock_irq(&pool->lock); 327360f5a4bcSLai Jiangshan 327492f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3275da028469SLai Jiangshan if (!list_empty(&pool->workers)) 327660f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 327792f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 327860f5a4bcSLai Jiangshan 327960f5a4bcSLai Jiangshan if (pool->detach_completion) 328060f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 328160f5a4bcSLai Jiangshan 328229c91e99STejun Heo mutex_unlock(&pool->manager_arb); 328329c91e99STejun Heo 328429c91e99STejun Heo /* shut down the timers */ 328529c91e99STejun Heo del_timer_sync(&pool->idle_timer); 328629c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 328729c91e99STejun Heo 328829c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 328929c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 329029c91e99STejun Heo } 329129c91e99STejun Heo 329229c91e99STejun Heo /** 329329c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 329429c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 329529c91e99STejun Heo * 329629c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 329729c91e99STejun Heo * reference count and return it. If there already is a matching 329829c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3299d185af30SYacine Belkadi * create a new one. 3300a892caccSTejun Heo * 3301a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3302d185af30SYacine Belkadi * 3303d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3304d185af30SYacine Belkadi * On failure, %NULL. 330529c91e99STejun Heo */ 330629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 330729c91e99STejun Heo { 330829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 330929c91e99STejun Heo struct worker_pool *pool; 3310f3f90ad4STejun Heo int node; 3311e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 331229c91e99STejun Heo 3313a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 331429c91e99STejun Heo 331529c91e99STejun Heo /* do we already have a matching pool? */ 331629c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 331729c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 331829c91e99STejun Heo pool->refcnt++; 33193fb1823cSLai Jiangshan return pool; 332029c91e99STejun Heo } 332129c91e99STejun Heo } 332229c91e99STejun Heo 3323e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3324e2273584SXunlei Pang if (wq_numa_enabled) { 3325e2273584SXunlei Pang for_each_node(node) { 3326e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3327e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3328e2273584SXunlei Pang target_node = node; 3329e2273584SXunlei Pang break; 3330e2273584SXunlei Pang } 3331e2273584SXunlei Pang } 3332e2273584SXunlei Pang } 3333e2273584SXunlei Pang 333429c91e99STejun Heo /* nope, create a new one */ 3335e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 333629c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 333729c91e99STejun Heo goto fail; 333829c91e99STejun Heo 33398864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 334029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3341e2273584SXunlei Pang pool->node = target_node; 334229c91e99STejun Heo 33432865a8fbSShaohua Li /* 33442865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 33452865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 33462865a8fbSShaohua Li */ 33472865a8fbSShaohua Li pool->attrs->no_numa = false; 33482865a8fbSShaohua Li 334929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 335029c91e99STejun Heo goto fail; 335129c91e99STejun Heo 335229c91e99STejun Heo /* create and start the initial worker */ 3353051e1850SLai Jiangshan if (!create_worker(pool)) 335429c91e99STejun Heo goto fail; 335529c91e99STejun Heo 335629c91e99STejun Heo /* install */ 335729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 33583fb1823cSLai Jiangshan 335929c91e99STejun Heo return pool; 336029c91e99STejun Heo fail: 336129c91e99STejun Heo if (pool) 336229c91e99STejun Heo put_unbound_pool(pool); 336329c91e99STejun Heo return NULL; 336429c91e99STejun Heo } 336529c91e99STejun Heo 33668864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 33678864b4e5STejun Heo { 33688864b4e5STejun Heo kmem_cache_free(pwq_cache, 33698864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 33708864b4e5STejun Heo } 33718864b4e5STejun Heo 33728864b4e5STejun Heo /* 33738864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 33748864b4e5STejun Heo * and needs to be destroyed. 33758864b4e5STejun Heo */ 33768864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 33778864b4e5STejun Heo { 33788864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 33798864b4e5STejun Heo unbound_release_work); 33808864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 33818864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3382bc0caf09STejun Heo bool is_last; 33838864b4e5STejun Heo 33848864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 33858864b4e5STejun Heo return; 33868864b4e5STejun Heo 33873c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 33888864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3389bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 33903c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 33918864b4e5STejun Heo 3392a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 33938864b4e5STejun Heo put_unbound_pool(pool); 3394a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3395a892caccSTejun Heo 33968864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 33978864b4e5STejun Heo 33988864b4e5STejun Heo /* 33998864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3400e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 34018864b4e5STejun Heo */ 3402e2dca7adSTejun Heo if (is_last) 3403e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 34046029a918STejun Heo } 34058864b4e5STejun Heo 34060fbd95aaSTejun Heo /** 3407699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 34080fbd95aaSTejun Heo * @pwq: target pool_workqueue 34090fbd95aaSTejun Heo * 3410699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3411699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3412699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 34130fbd95aaSTejun Heo */ 3414699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 34150fbd95aaSTejun Heo { 3416699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3417699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3418699ce097STejun Heo 3419699ce097STejun Heo /* for @wq->saved_max_active */ 3420a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3421699ce097STejun Heo 3422699ce097STejun Heo /* fast exit for non-freezable wqs */ 3423699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3424699ce097STejun Heo return; 3425699ce097STejun Heo 3426a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3427699ce097STejun Heo 342874b414eaSLai Jiangshan /* 342974b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 343074b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 343174b414eaSLai Jiangshan * is updated and visible. 343274b414eaSLai Jiangshan */ 343374b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3434699ce097STejun Heo pwq->max_active = wq->saved_max_active; 34350fbd95aaSTejun Heo 34360fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 34370fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 34380fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3439951a078aSLai Jiangshan 3440951a078aSLai Jiangshan /* 3441951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3442951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3443951a078aSLai Jiangshan */ 3444951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3445699ce097STejun Heo } else { 3446699ce097STejun Heo pwq->max_active = 0; 3447699ce097STejun Heo } 3448699ce097STejun Heo 3449a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 34500fbd95aaSTejun Heo } 34510fbd95aaSTejun Heo 3452e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3453f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3454f147f29eSTejun Heo struct worker_pool *pool) 3455d2c1d404STejun Heo { 3456d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3457d2c1d404STejun Heo 3458e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3459e50aba9aSTejun Heo 3460d2c1d404STejun Heo pwq->pool = pool; 3461d2c1d404STejun Heo pwq->wq = wq; 3462d2c1d404STejun Heo pwq->flush_color = -1; 34638864b4e5STejun Heo pwq->refcnt = 1; 3464d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 34651befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3466d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 34678864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3468f147f29eSTejun Heo } 3469d2c1d404STejun Heo 3470f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 34711befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3472f147f29eSTejun Heo { 3473f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3474f147f29eSTejun Heo 3475f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 347675ccf595STejun Heo 34771befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 34781befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 34791befcf30STejun Heo return; 34801befcf30STejun Heo 348129b1cb41SLai Jiangshan /* set the matching work_color */ 348275ccf595STejun Heo pwq->work_color = wq->work_color; 3483983ca25eSTejun Heo 3484983ca25eSTejun Heo /* sync max_active to the current setting */ 3485983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3486983ca25eSTejun Heo 3487983ca25eSTejun Heo /* link in @pwq */ 34889e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3489df2d5ae4STejun Heo } 34906029a918STejun Heo 3491f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3492f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3493f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3494f147f29eSTejun Heo { 3495f147f29eSTejun Heo struct worker_pool *pool; 3496f147f29eSTejun Heo struct pool_workqueue *pwq; 3497f147f29eSTejun Heo 3498f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3499f147f29eSTejun Heo 3500f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3501f147f29eSTejun Heo if (!pool) 3502f147f29eSTejun Heo return NULL; 3503f147f29eSTejun Heo 3504e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3505f147f29eSTejun Heo if (!pwq) { 3506f147f29eSTejun Heo put_unbound_pool(pool); 3507f147f29eSTejun Heo return NULL; 3508f147f29eSTejun Heo } 3509f147f29eSTejun Heo 3510f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3511f147f29eSTejun Heo return pwq; 3512d2c1d404STejun Heo } 3513d2c1d404STejun Heo 35144c16bd32STejun Heo /** 351530186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3516042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 35174c16bd32STejun Heo * @node: the target NUMA node 35184c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 35194c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 35204c16bd32STejun Heo * 35214c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 35224c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3523d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 35244c16bd32STejun Heo * 35254c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 35264c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 35274c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 35284c16bd32STejun Heo * @attrs->cpumask. 35294c16bd32STejun Heo * 35304c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 35314c16bd32STejun Heo * stable. 3532d185af30SYacine Belkadi * 3533d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3534d185af30SYacine Belkadi * %false if equal. 35354c16bd32STejun Heo */ 35364c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 35374c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 35384c16bd32STejun Heo { 3539d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 35404c16bd32STejun Heo goto use_dfl; 35414c16bd32STejun Heo 35424c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 35434c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 35444c16bd32STejun Heo if (cpu_going_down >= 0) 35454c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 35464c16bd32STejun Heo 35474c16bd32STejun Heo if (cpumask_empty(cpumask)) 35484c16bd32STejun Heo goto use_dfl; 35494c16bd32STejun Heo 35504c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 35514c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 35524c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 35534c16bd32STejun Heo 35544c16bd32STejun Heo use_dfl: 35554c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 35564c16bd32STejun Heo return false; 35574c16bd32STejun Heo } 35584c16bd32STejun Heo 35591befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 35601befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 35611befcf30STejun Heo int node, 35621befcf30STejun Heo struct pool_workqueue *pwq) 35631befcf30STejun Heo { 35641befcf30STejun Heo struct pool_workqueue *old_pwq; 35651befcf30STejun Heo 35665b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 35671befcf30STejun Heo lockdep_assert_held(&wq->mutex); 35681befcf30STejun Heo 35691befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 35701befcf30STejun Heo link_pwq(pwq); 35711befcf30STejun Heo 35721befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 35731befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 35741befcf30STejun Heo return old_pwq; 35751befcf30STejun Heo } 35761befcf30STejun Heo 35772d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 35782d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 35792d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 35802d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3581042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 35822d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 35832d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 35842d5f0764SLai Jiangshan }; 35852d5f0764SLai Jiangshan 35862d5f0764SLai Jiangshan /* free the resources after success or abort */ 35872d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 35882d5f0764SLai Jiangshan { 35892d5f0764SLai Jiangshan if (ctx) { 35902d5f0764SLai Jiangshan int node; 35912d5f0764SLai Jiangshan 35922d5f0764SLai Jiangshan for_each_node(node) 35932d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 35942d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 35952d5f0764SLai Jiangshan 35962d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 35972d5f0764SLai Jiangshan 35982d5f0764SLai Jiangshan kfree(ctx); 35992d5f0764SLai Jiangshan } 36002d5f0764SLai Jiangshan } 36012d5f0764SLai Jiangshan 36022d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 36032d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 36042d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 36052d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 36062d5f0764SLai Jiangshan { 36072d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 36082d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 36092d5f0764SLai Jiangshan int node; 36102d5f0764SLai Jiangshan 36112d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36122d5f0764SLai Jiangshan 36132d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 36142d5f0764SLai Jiangshan GFP_KERNEL); 36152d5f0764SLai Jiangshan 36162d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36172d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36182d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 36192d5f0764SLai Jiangshan goto out_free; 36202d5f0764SLai Jiangshan 3621042f7df1SLai Jiangshan /* 3622042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3623042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3624042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3625042f7df1SLai Jiangshan */ 36262d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3627b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3628042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3629042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 36302d5f0764SLai Jiangshan 36312d5f0764SLai Jiangshan /* 36322d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 36332d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 36342d5f0764SLai Jiangshan * pools. 36352d5f0764SLai Jiangshan */ 36362d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 36372d5f0764SLai Jiangshan 36382d5f0764SLai Jiangshan /* 36392d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 36402d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 36412d5f0764SLai Jiangshan * it even if we don't use it immediately. 36422d5f0764SLai Jiangshan */ 36432d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 36442d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 36452d5f0764SLai Jiangshan goto out_free; 36462d5f0764SLai Jiangshan 36472d5f0764SLai Jiangshan for_each_node(node) { 3648042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 36492d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 36502d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 36512d5f0764SLai Jiangshan goto out_free; 36522d5f0764SLai Jiangshan } else { 36532d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 36542d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 36552d5f0764SLai Jiangshan } 36562d5f0764SLai Jiangshan } 36572d5f0764SLai Jiangshan 3658042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3659042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3660042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 36612d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3662042f7df1SLai Jiangshan 36632d5f0764SLai Jiangshan ctx->wq = wq; 36642d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36652d5f0764SLai Jiangshan return ctx; 36662d5f0764SLai Jiangshan 36672d5f0764SLai Jiangshan out_free: 36682d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36692d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 36702d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 36712d5f0764SLai Jiangshan return NULL; 36722d5f0764SLai Jiangshan } 36732d5f0764SLai Jiangshan 36742d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 36752d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 36762d5f0764SLai Jiangshan { 36772d5f0764SLai Jiangshan int node; 36782d5f0764SLai Jiangshan 36792d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 36802d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 36812d5f0764SLai Jiangshan 36822d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 36832d5f0764SLai Jiangshan 36842d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 36852d5f0764SLai Jiangshan for_each_node(node) 36862d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 36872d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 36882d5f0764SLai Jiangshan 36892d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 36902d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 36912d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 36922d5f0764SLai Jiangshan 36932d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 36942d5f0764SLai Jiangshan } 36952d5f0764SLai Jiangshan 3696a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3697a0111cf6SLai Jiangshan { 3698a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3699a0111cf6SLai Jiangshan get_online_cpus(); 3700a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3701a0111cf6SLai Jiangshan } 3702a0111cf6SLai Jiangshan 3703a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3704a0111cf6SLai Jiangshan { 3705a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3706a0111cf6SLai Jiangshan put_online_cpus(); 3707a0111cf6SLai Jiangshan } 3708a0111cf6SLai Jiangshan 3709a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3710a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3711a0111cf6SLai Jiangshan { 3712a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3713a0111cf6SLai Jiangshan 3714a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3715a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3716a0111cf6SLai Jiangshan return -EINVAL; 3717a0111cf6SLai Jiangshan 3718a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 3719a0111cf6SLai Jiangshan if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 3720a0111cf6SLai Jiangshan return -EINVAL; 3721a0111cf6SLai Jiangshan 3722a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 37236201171eSwanghaibin if (!ctx) 37246201171eSwanghaibin return -ENOMEM; 3725a0111cf6SLai Jiangshan 3726a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3727a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3728a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3729a0111cf6SLai Jiangshan 37306201171eSwanghaibin return 0; 3731a0111cf6SLai Jiangshan } 3732a0111cf6SLai Jiangshan 37339e8cd2f5STejun Heo /** 37349e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37359e8cd2f5STejun Heo * @wq: the target workqueue 37369e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37379e8cd2f5STejun Heo * 37384c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 37394c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 37404c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 37414c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 37424c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 37434c16bd32STejun Heo * back-to-back will stay on its current pwq. 37449e8cd2f5STejun Heo * 3745d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3746d185af30SYacine Belkadi * 3747d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 37489e8cd2f5STejun Heo */ 37499e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 37509e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 37519e8cd2f5STejun Heo { 3752a0111cf6SLai Jiangshan int ret; 37539e8cd2f5STejun Heo 3754a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3755a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3756a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 37572d5f0764SLai Jiangshan 37582d5f0764SLai Jiangshan return ret; 37599e8cd2f5STejun Heo } 37609e8cd2f5STejun Heo 37614c16bd32STejun Heo /** 37624c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 37634c16bd32STejun Heo * @wq: the target workqueue 37644c16bd32STejun Heo * @cpu: the CPU coming up or going down 37654c16bd32STejun Heo * @online: whether @cpu is coming up or going down 37664c16bd32STejun Heo * 37674c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 37684c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 37694c16bd32STejun Heo * @wq accordingly. 37704c16bd32STejun Heo * 37714c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 37724c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 37734c16bd32STejun Heo * correct. 37744c16bd32STejun Heo * 37754c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 37764c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 37774c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 37784c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 37794c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 37804c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 37814c16bd32STejun Heo * CPU_DOWN_PREPARE. 37824c16bd32STejun Heo */ 37834c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 37844c16bd32STejun Heo bool online) 37854c16bd32STejun Heo { 37864c16bd32STejun Heo int node = cpu_to_node(cpu); 37874c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 37884c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 37894c16bd32STejun Heo struct workqueue_attrs *target_attrs; 37904c16bd32STejun Heo cpumask_t *cpumask; 37914c16bd32STejun Heo 37924c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37934c16bd32STejun Heo 3794f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3795f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 37964c16bd32STejun Heo return; 37974c16bd32STejun Heo 37984c16bd32STejun Heo /* 37994c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 38004c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 38014c16bd32STejun Heo * CPU hotplug exclusion. 38024c16bd32STejun Heo */ 38034c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 38044c16bd32STejun Heo cpumask = target_attrs->cpumask; 38054c16bd32STejun Heo 38064c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 38074c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 38084c16bd32STejun Heo 38094c16bd32STejun Heo /* 38104c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3811042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3812042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3813042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 38144c16bd32STejun Heo */ 3815042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 38164c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3817f7142ed4SLai Jiangshan return; 38184c16bd32STejun Heo } else { 38194c16bd32STejun Heo goto use_dfl_pwq; 38204c16bd32STejun Heo } 38214c16bd32STejun Heo 38224c16bd32STejun Heo /* create a new pwq */ 38234c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 38244c16bd32STejun Heo if (!pwq) { 38252d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 38264c16bd32STejun Heo wq->name); 382777f300b1SDaeseok Youn goto use_dfl_pwq; 38284c16bd32STejun Heo } 38294c16bd32STejun Heo 3830f7142ed4SLai Jiangshan /* Install the new pwq. */ 38314c16bd32STejun Heo mutex_lock(&wq->mutex); 38324c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 38334c16bd32STejun Heo goto out_unlock; 38344c16bd32STejun Heo 38354c16bd32STejun Heo use_dfl_pwq: 3836f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 38374c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 38384c16bd32STejun Heo get_pwq(wq->dfl_pwq); 38394c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 38404c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 38414c16bd32STejun Heo out_unlock: 38424c16bd32STejun Heo mutex_unlock(&wq->mutex); 38434c16bd32STejun Heo put_pwq_unlocked(old_pwq); 38444c16bd32STejun Heo } 38454c16bd32STejun Heo 384630cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 38471da177e4SLinus Torvalds { 384849e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 38498a2b7538STejun Heo int cpu, ret; 3850e1d8aa9fSFrederic Weisbecker 385130cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3852420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3853420c0ddbSTejun Heo if (!wq->cpu_pwqs) 385430cdf249STejun Heo return -ENOMEM; 385530cdf249STejun Heo 385630cdf249STejun Heo for_each_possible_cpu(cpu) { 38577fb98ea7STejun Heo struct pool_workqueue *pwq = 38587fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 38597a62c2c8STejun Heo struct worker_pool *cpu_pools = 3860f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 386130cdf249STejun Heo 3862f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3863f147f29eSTejun Heo 3864f147f29eSTejun Heo mutex_lock(&wq->mutex); 38651befcf30STejun Heo link_pwq(pwq); 3866f147f29eSTejun Heo mutex_unlock(&wq->mutex); 386730cdf249STejun Heo } 386830cdf249STejun Heo return 0; 38698a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 38708a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 38718a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 38728a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 38738a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 38748a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 38758a2b7538STejun Heo return ret; 38769e8cd2f5STejun Heo } else { 38779e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 38789e8cd2f5STejun Heo } 38790f900049STejun Heo } 38800f900049STejun Heo 3881f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3882f3421797STejun Heo const char *name) 3883b71ab8c2STejun Heo { 3884f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3885f3421797STejun Heo 3886f3421797STejun Heo if (max_active < 1 || max_active > lim) 3887044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3888f3421797STejun Heo max_active, name, 1, lim); 3889b71ab8c2STejun Heo 3890f3421797STejun Heo return clamp_val(max_active, 1, lim); 3891b71ab8c2STejun Heo } 3892b71ab8c2STejun Heo 3893b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 389497e37d7bSTejun Heo unsigned int flags, 38951e19ffc6STejun Heo int max_active, 3896eb13ba87SJohannes Berg struct lock_class_key *key, 3897b196be89STejun Heo const char *lock_name, ...) 38983af24433SOleg Nesterov { 3899df2d5ae4STejun Heo size_t tbl_size = 0; 3900ecf6881fSTejun Heo va_list args; 39013af24433SOleg Nesterov struct workqueue_struct *wq; 390249e3cf44STejun Heo struct pool_workqueue *pwq; 3903b196be89STejun Heo 3904cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 3905cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 3906cee22a15SViresh Kumar flags |= WQ_UNBOUND; 3907cee22a15SViresh Kumar 3908ecf6881fSTejun Heo /* allocate wq and format name */ 3909df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 3910ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 3911df2d5ae4STejun Heo 3912df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 3913b196be89STejun Heo if (!wq) 3914d2c1d404STejun Heo return NULL; 3915b196be89STejun Heo 39166029a918STejun Heo if (flags & WQ_UNBOUND) { 39176029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39186029a918STejun Heo if (!wq->unbound_attrs) 39196029a918STejun Heo goto err_free_wq; 39206029a918STejun Heo } 39216029a918STejun Heo 3922ecf6881fSTejun Heo va_start(args, lock_name); 3923ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3924b196be89STejun Heo va_end(args); 39253af24433SOleg Nesterov 3926d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3927b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 39283af24433SOleg Nesterov 3929b196be89STejun Heo /* init wq */ 393097e37d7bSTejun Heo wq->flags = flags; 3931a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 39323c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3933112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 393430cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 393573f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 393673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3937493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 39383af24433SOleg Nesterov 3939eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3940cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 39413af24433SOleg Nesterov 394230cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3943d2c1d404STejun Heo goto err_free_wq; 39441537663fSTejun Heo 3945493008a8STejun Heo /* 3946493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3947493008a8STejun Heo * have a rescuer to guarantee forward progress. 3948493008a8STejun Heo */ 3949493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3950e22bee78STejun Heo struct worker *rescuer; 3951e22bee78STejun Heo 3952f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 3953e22bee78STejun Heo if (!rescuer) 3954d2c1d404STejun Heo goto err_destroy; 3955e22bee78STejun Heo 3956111c225aSTejun Heo rescuer->rescue_wq = wq; 3957111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3958b196be89STejun Heo wq->name); 3959d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3960d2c1d404STejun Heo kfree(rescuer); 3961d2c1d404STejun Heo goto err_destroy; 3962d2c1d404STejun Heo } 3963e22bee78STejun Heo 3964d2c1d404STejun Heo wq->rescuer = rescuer; 396525834c73SPeter Zijlstra kthread_bind_mask(rescuer->task, cpu_possible_mask); 3966e22bee78STejun Heo wake_up_process(rescuer->task); 39673af24433SOleg Nesterov } 39681537663fSTejun Heo 3969226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3970226223abSTejun Heo goto err_destroy; 3971226223abSTejun Heo 39723af24433SOleg Nesterov /* 397368e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 397468e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 397568e13a67SLai Jiangshan * list. 39763af24433SOleg Nesterov */ 397768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3978a0a1a5fdSTejun Heo 3979a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 398049e3cf44STejun Heo for_each_pwq(pwq, wq) 3981699ce097STejun Heo pwq_adjust_max_active(pwq); 3982a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3983a0a1a5fdSTejun Heo 3984e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 3985a0a1a5fdSTejun Heo 398668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 39873af24433SOleg Nesterov 39883af24433SOleg Nesterov return wq; 3989d2c1d404STejun Heo 3990d2c1d404STejun Heo err_free_wq: 39916029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 39924690c4abSTejun Heo kfree(wq); 3993d2c1d404STejun Heo return NULL; 3994d2c1d404STejun Heo err_destroy: 3995d2c1d404STejun Heo destroy_workqueue(wq); 39964690c4abSTejun Heo return NULL; 39971da177e4SLinus Torvalds } 3998d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 39991da177e4SLinus Torvalds 40003af24433SOleg Nesterov /** 40013af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 40023af24433SOleg Nesterov * @wq: target workqueue 40033af24433SOleg Nesterov * 40043af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 40053af24433SOleg Nesterov */ 40063af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 40073af24433SOleg Nesterov { 400849e3cf44STejun Heo struct pool_workqueue *pwq; 40094c16bd32STejun Heo int node; 40103af24433SOleg Nesterov 40119c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 40129c5a2ba7STejun Heo drain_workqueue(wq); 4013c8efcc25STejun Heo 40146183c009STejun Heo /* sanity checks */ 4015b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 401649e3cf44STejun Heo for_each_pwq(pwq, wq) { 40176183c009STejun Heo int i; 40186183c009STejun Heo 401976af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 402076af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4021b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 40226183c009STejun Heo return; 402376af4d93STejun Heo } 402476af4d93STejun Heo } 402576af4d93STejun Heo 40265c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 40278864b4e5STejun Heo WARN_ON(pwq->nr_active) || 402876af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4029b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 40306183c009STejun Heo return; 40316183c009STejun Heo } 403276af4d93STejun Heo } 4033b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 40346183c009STejun Heo 4035a0a1a5fdSTejun Heo /* 4036a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4037a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4038a0a1a5fdSTejun Heo */ 403968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4040e2dca7adSTejun Heo list_del_rcu(&wq->list); 404168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 40423af24433SOleg Nesterov 4043226223abSTejun Heo workqueue_sysfs_unregister(wq); 4044226223abSTejun Heo 4045e2dca7adSTejun Heo if (wq->rescuer) 4046e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4047e22bee78STejun Heo 40488864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 404929c91e99STejun Heo /* 40508864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4051e2dca7adSTejun Heo * schedule RCU free. 405229c91e99STejun Heo */ 4053e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 40548864b4e5STejun Heo } else { 40558864b4e5STejun Heo /* 40568864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 40574c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 40584c16bd32STejun Heo * @wq will be freed when the last pwq is released. 40598864b4e5STejun Heo */ 40604c16bd32STejun Heo for_each_node(node) { 40614c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 40624c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 40634c16bd32STejun Heo put_pwq_unlocked(pwq); 40644c16bd32STejun Heo } 40654c16bd32STejun Heo 40664c16bd32STejun Heo /* 40674c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 40684c16bd32STejun Heo * put. Don't access it afterwards. 40694c16bd32STejun Heo */ 40704c16bd32STejun Heo pwq = wq->dfl_pwq; 40714c16bd32STejun Heo wq->dfl_pwq = NULL; 4072dce90d47STejun Heo put_pwq_unlocked(pwq); 407329c91e99STejun Heo } 40743af24433SOleg Nesterov } 40753af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 40763af24433SOleg Nesterov 4077dcd989cbSTejun Heo /** 4078dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4079dcd989cbSTejun Heo * @wq: target workqueue 4080dcd989cbSTejun Heo * @max_active: new max_active value. 4081dcd989cbSTejun Heo * 4082dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4083dcd989cbSTejun Heo * 4084dcd989cbSTejun Heo * CONTEXT: 4085dcd989cbSTejun Heo * Don't call from IRQ context. 4086dcd989cbSTejun Heo */ 4087dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4088dcd989cbSTejun Heo { 408949e3cf44STejun Heo struct pool_workqueue *pwq; 4090dcd989cbSTejun Heo 40918719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 40928719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 40938719dceaSTejun Heo return; 40948719dceaSTejun Heo 4095f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4096dcd989cbSTejun Heo 4097a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4098dcd989cbSTejun Heo 4099dcd989cbSTejun Heo wq->saved_max_active = max_active; 4100dcd989cbSTejun Heo 4101699ce097STejun Heo for_each_pwq(pwq, wq) 4102699ce097STejun Heo pwq_adjust_max_active(pwq); 4103dcd989cbSTejun Heo 4104a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4105dcd989cbSTejun Heo } 4106dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4107dcd989cbSTejun Heo 4108dcd989cbSTejun Heo /** 4109e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4110e6267616STejun Heo * 4111e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4112e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4113d185af30SYacine Belkadi * 4114d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4115e6267616STejun Heo */ 4116e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4117e6267616STejun Heo { 4118e6267616STejun Heo struct worker *worker = current_wq_worker(); 4119e6267616STejun Heo 41206a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4121e6267616STejun Heo } 4122e6267616STejun Heo 4123e6267616STejun Heo /** 4124dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4125dcd989cbSTejun Heo * @cpu: CPU in question 4126dcd989cbSTejun Heo * @wq: target workqueue 4127dcd989cbSTejun Heo * 4128dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4129dcd989cbSTejun Heo * no synchronization around this function and the test result is 4130dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4131dcd989cbSTejun Heo * 4132d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4133d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4134d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4135d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4136d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4137d3251859STejun Heo * 4138d185af30SYacine Belkadi * Return: 4139dcd989cbSTejun Heo * %true if congested, %false otherwise. 4140dcd989cbSTejun Heo */ 4141d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4142dcd989cbSTejun Heo { 41437fb98ea7STejun Heo struct pool_workqueue *pwq; 414476af4d93STejun Heo bool ret; 414576af4d93STejun Heo 414688109453SLai Jiangshan rcu_read_lock_sched(); 41477fb98ea7STejun Heo 4148d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4149d3251859STejun Heo cpu = smp_processor_id(); 4150d3251859STejun Heo 41517fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 41527fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 41537fb98ea7STejun Heo else 4154df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4155dcd989cbSTejun Heo 415676af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 415788109453SLai Jiangshan rcu_read_unlock_sched(); 415876af4d93STejun Heo 415976af4d93STejun Heo return ret; 4160dcd989cbSTejun Heo } 4161dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4162dcd989cbSTejun Heo 4163dcd989cbSTejun Heo /** 4164dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4165dcd989cbSTejun Heo * @work: the work to be tested 4166dcd989cbSTejun Heo * 4167dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4168dcd989cbSTejun Heo * synchronization around this function and the test result is 4169dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4170dcd989cbSTejun Heo * 4171d185af30SYacine Belkadi * Return: 4172dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4173dcd989cbSTejun Heo */ 4174dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4175dcd989cbSTejun Heo { 4176fa1b54e6STejun Heo struct worker_pool *pool; 4177dcd989cbSTejun Heo unsigned long flags; 4178dcd989cbSTejun Heo unsigned int ret = 0; 4179dcd989cbSTejun Heo 4180dcd989cbSTejun Heo if (work_pending(work)) 4181dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4182038366c5SLai Jiangshan 4183fa1b54e6STejun Heo local_irq_save(flags); 4184fa1b54e6STejun Heo pool = get_work_pool(work); 4185038366c5SLai Jiangshan if (pool) { 4186fa1b54e6STejun Heo spin_lock(&pool->lock); 4187c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4188dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4189fa1b54e6STejun Heo spin_unlock(&pool->lock); 4190038366c5SLai Jiangshan } 4191fa1b54e6STejun Heo local_irq_restore(flags); 4192dcd989cbSTejun Heo 4193dcd989cbSTejun Heo return ret; 4194dcd989cbSTejun Heo } 4195dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4196dcd989cbSTejun Heo 41973d1cb205STejun Heo /** 41983d1cb205STejun Heo * set_worker_desc - set description for the current work item 41993d1cb205STejun Heo * @fmt: printf-style format string 42003d1cb205STejun Heo * @...: arguments for the format string 42013d1cb205STejun Heo * 42023d1cb205STejun Heo * This function can be called by a running work function to describe what 42033d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 42043d1cb205STejun Heo * information will be printed out together to help debugging. The 42053d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 42063d1cb205STejun Heo */ 42073d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 42083d1cb205STejun Heo { 42093d1cb205STejun Heo struct worker *worker = current_wq_worker(); 42103d1cb205STejun Heo va_list args; 42113d1cb205STejun Heo 42123d1cb205STejun Heo if (worker) { 42133d1cb205STejun Heo va_start(args, fmt); 42143d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 42153d1cb205STejun Heo va_end(args); 42163d1cb205STejun Heo worker->desc_valid = true; 42173d1cb205STejun Heo } 42183d1cb205STejun Heo } 42193d1cb205STejun Heo 42203d1cb205STejun Heo /** 42213d1cb205STejun Heo * print_worker_info - print out worker information and description 42223d1cb205STejun Heo * @log_lvl: the log level to use when printing 42233d1cb205STejun Heo * @task: target task 42243d1cb205STejun Heo * 42253d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 42263d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 42273d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 42283d1cb205STejun Heo * 42293d1cb205STejun Heo * This function can be safely called on any task as long as the 42303d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 42313d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 42323d1cb205STejun Heo */ 42333d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 42343d1cb205STejun Heo { 42353d1cb205STejun Heo work_func_t *fn = NULL; 42363d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 42373d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 42383d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 42393d1cb205STejun Heo struct workqueue_struct *wq = NULL; 42403d1cb205STejun Heo bool desc_valid = false; 42413d1cb205STejun Heo struct worker *worker; 42423d1cb205STejun Heo 42433d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 42443d1cb205STejun Heo return; 42453d1cb205STejun Heo 42463d1cb205STejun Heo /* 42473d1cb205STejun Heo * This function is called without any synchronization and @task 42483d1cb205STejun Heo * could be in any state. Be careful with dereferences. 42493d1cb205STejun Heo */ 42503d1cb205STejun Heo worker = probe_kthread_data(task); 42513d1cb205STejun Heo 42523d1cb205STejun Heo /* 42533d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 42543d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 42553d1cb205STejun Heo */ 42563d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 42573d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 42583d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 42593d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 42603d1cb205STejun Heo 42613d1cb205STejun Heo /* copy worker description */ 42623d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 42633d1cb205STejun Heo if (desc_valid) 42643d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 42653d1cb205STejun Heo 42663d1cb205STejun Heo if (fn || name[0] || desc[0]) { 42673d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 42683d1cb205STejun Heo if (desc[0]) 42693d1cb205STejun Heo pr_cont(" (%s)", desc); 42703d1cb205STejun Heo pr_cont("\n"); 42713d1cb205STejun Heo } 42723d1cb205STejun Heo } 42733d1cb205STejun Heo 42743494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 42753494fc30STejun Heo { 42763494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 42773494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 42783494fc30STejun Heo pr_cont(" node=%d", pool->node); 42793494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 42803494fc30STejun Heo } 42813494fc30STejun Heo 42823494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 42833494fc30STejun Heo { 42843494fc30STejun Heo if (work->func == wq_barrier_func) { 42853494fc30STejun Heo struct wq_barrier *barr; 42863494fc30STejun Heo 42873494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 42883494fc30STejun Heo 42893494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 42903494fc30STejun Heo task_pid_nr(barr->task)); 42913494fc30STejun Heo } else { 42923494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 42933494fc30STejun Heo } 42943494fc30STejun Heo } 42953494fc30STejun Heo 42963494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 42973494fc30STejun Heo { 42983494fc30STejun Heo struct worker_pool *pool = pwq->pool; 42993494fc30STejun Heo struct work_struct *work; 43003494fc30STejun Heo struct worker *worker; 43013494fc30STejun Heo bool has_in_flight = false, has_pending = false; 43023494fc30STejun Heo int bkt; 43033494fc30STejun Heo 43043494fc30STejun Heo pr_info(" pwq %d:", pool->id); 43053494fc30STejun Heo pr_cont_pool_info(pool); 43063494fc30STejun Heo 43073494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 43083494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 43093494fc30STejun Heo 43103494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43113494fc30STejun Heo if (worker->current_pwq == pwq) { 43123494fc30STejun Heo has_in_flight = true; 43133494fc30STejun Heo break; 43143494fc30STejun Heo } 43153494fc30STejun Heo } 43163494fc30STejun Heo if (has_in_flight) { 43173494fc30STejun Heo bool comma = false; 43183494fc30STejun Heo 43193494fc30STejun Heo pr_info(" in-flight:"); 43203494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43213494fc30STejun Heo if (worker->current_pwq != pwq) 43223494fc30STejun Heo continue; 43233494fc30STejun Heo 43243494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 43253494fc30STejun Heo task_pid_nr(worker->task), 43263494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 43273494fc30STejun Heo worker->current_func); 43283494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 43293494fc30STejun Heo pr_cont_work(false, work); 43303494fc30STejun Heo comma = true; 43313494fc30STejun Heo } 43323494fc30STejun Heo pr_cont("\n"); 43333494fc30STejun Heo } 43343494fc30STejun Heo 43353494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 43363494fc30STejun Heo if (get_work_pwq(work) == pwq) { 43373494fc30STejun Heo has_pending = true; 43383494fc30STejun Heo break; 43393494fc30STejun Heo } 43403494fc30STejun Heo } 43413494fc30STejun Heo if (has_pending) { 43423494fc30STejun Heo bool comma = false; 43433494fc30STejun Heo 43443494fc30STejun Heo pr_info(" pending:"); 43453494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 43463494fc30STejun Heo if (get_work_pwq(work) != pwq) 43473494fc30STejun Heo continue; 43483494fc30STejun Heo 43493494fc30STejun Heo pr_cont_work(comma, work); 43503494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 43513494fc30STejun Heo } 43523494fc30STejun Heo pr_cont("\n"); 43533494fc30STejun Heo } 43543494fc30STejun Heo 43553494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 43563494fc30STejun Heo bool comma = false; 43573494fc30STejun Heo 43583494fc30STejun Heo pr_info(" delayed:"); 43593494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 43603494fc30STejun Heo pr_cont_work(comma, work); 43613494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 43623494fc30STejun Heo } 43633494fc30STejun Heo pr_cont("\n"); 43643494fc30STejun Heo } 43653494fc30STejun Heo } 43663494fc30STejun Heo 43673494fc30STejun Heo /** 43683494fc30STejun Heo * show_workqueue_state - dump workqueue state 43693494fc30STejun Heo * 43703494fc30STejun Heo * Called from a sysrq handler and prints out all busy workqueues and 43713494fc30STejun Heo * pools. 43723494fc30STejun Heo */ 43733494fc30STejun Heo void show_workqueue_state(void) 43743494fc30STejun Heo { 43753494fc30STejun Heo struct workqueue_struct *wq; 43763494fc30STejun Heo struct worker_pool *pool; 43773494fc30STejun Heo unsigned long flags; 43783494fc30STejun Heo int pi; 43793494fc30STejun Heo 43803494fc30STejun Heo rcu_read_lock_sched(); 43813494fc30STejun Heo 43823494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 43833494fc30STejun Heo 43843494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 43853494fc30STejun Heo struct pool_workqueue *pwq; 43863494fc30STejun Heo bool idle = true; 43873494fc30STejun Heo 43883494fc30STejun Heo for_each_pwq(pwq, wq) { 43893494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 43903494fc30STejun Heo idle = false; 43913494fc30STejun Heo break; 43923494fc30STejun Heo } 43933494fc30STejun Heo } 43943494fc30STejun Heo if (idle) 43953494fc30STejun Heo continue; 43963494fc30STejun Heo 43973494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 43983494fc30STejun Heo 43993494fc30STejun Heo for_each_pwq(pwq, wq) { 44003494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 44013494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 44023494fc30STejun Heo show_pwq(pwq); 44033494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 44043494fc30STejun Heo } 44053494fc30STejun Heo } 44063494fc30STejun Heo 44073494fc30STejun Heo for_each_pool(pool, pi) { 44083494fc30STejun Heo struct worker *worker; 44093494fc30STejun Heo bool first = true; 44103494fc30STejun Heo 44113494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 44123494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 44133494fc30STejun Heo goto next_pool; 44143494fc30STejun Heo 44153494fc30STejun Heo pr_info("pool %d:", pool->id); 44163494fc30STejun Heo pr_cont_pool_info(pool); 441782607adcSTejun Heo pr_cont(" hung=%us workers=%d", 441882607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 441982607adcSTejun Heo pool->nr_workers); 44203494fc30STejun Heo if (pool->manager) 44213494fc30STejun Heo pr_cont(" manager: %d", 44223494fc30STejun Heo task_pid_nr(pool->manager->task)); 44233494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 44243494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 44253494fc30STejun Heo task_pid_nr(worker->task)); 44263494fc30STejun Heo first = false; 44273494fc30STejun Heo } 44283494fc30STejun Heo pr_cont("\n"); 44293494fc30STejun Heo next_pool: 44303494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 44313494fc30STejun Heo } 44323494fc30STejun Heo 44333494fc30STejun Heo rcu_read_unlock_sched(); 44343494fc30STejun Heo } 44353494fc30STejun Heo 4436db7bccf4STejun Heo /* 4437db7bccf4STejun Heo * CPU hotplug. 4438db7bccf4STejun Heo * 4439e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4440112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4441706026c2STejun Heo * pool which make migrating pending and scheduled works very 4442e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 444394cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4444e22bee78STejun Heo * blocked draining impractical. 4445e22bee78STejun Heo * 444624647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4447628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4448628c78e7STejun Heo * cpu comes back online. 4449db7bccf4STejun Heo */ 4450db7bccf4STejun Heo 4451706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4452db7bccf4STejun Heo { 445338db41d9STejun Heo int cpu = smp_processor_id(); 44544ce62e9eSTejun Heo struct worker_pool *pool; 4455db7bccf4STejun Heo struct worker *worker; 4456db7bccf4STejun Heo 4457f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 445892f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 445994cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4460e22bee78STejun Heo 4461f2d5a0eeSTejun Heo /* 446292f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 446394cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 446494cf58bbSTejun Heo * except for the ones which are still executing works from 446594cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 446694cf58bbSTejun Heo * this, they may become diasporas. 4467f2d5a0eeSTejun Heo */ 4468da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4469403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4470db7bccf4STejun Heo 447124647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4472f2d5a0eeSTejun Heo 447394cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 447492f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4475e22bee78STejun Heo 4476e22bee78STejun Heo /* 4477eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4478eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4479eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4480eb283428SLai Jiangshan * from other cpus. 4481628c78e7STejun Heo */ 4482628c78e7STejun Heo schedule(); 4483628c78e7STejun Heo 4484628c78e7STejun Heo /* 4485eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4486eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4487eb283428SLai Jiangshan * and keep_working() are always true as long as the 4488eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4489eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4490eb283428SLai Jiangshan * are served by workers tied to the pool. 4491e22bee78STejun Heo */ 4492e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4493eb283428SLai Jiangshan 4494eb283428SLai Jiangshan /* 4495eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4496eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4497eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4498eb283428SLai Jiangshan */ 4499eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4500eb283428SLai Jiangshan wake_up_worker(pool); 4501eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4502eb283428SLai Jiangshan } 4503db7bccf4STejun Heo } 4504db7bccf4STejun Heo 4505bd7c089eSTejun Heo /** 4506bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4507bd7c089eSTejun Heo * @pool: pool of interest 4508bd7c089eSTejun Heo * 4509a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4510bd7c089eSTejun Heo */ 4511bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4512bd7c089eSTejun Heo { 4513a9ab775bSTejun Heo struct worker *worker; 4514bd7c089eSTejun Heo 451592f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4516bd7c089eSTejun Heo 4517bd7c089eSTejun Heo /* 4518a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4519a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4520402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4521a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4522a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4523bd7c089eSTejun Heo */ 4524da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4525a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4526a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4527a9ab775bSTejun Heo 4528a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 45293de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4530a9ab775bSTejun Heo 4531da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4532a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4533a9ab775bSTejun Heo 4534a9ab775bSTejun Heo /* 4535a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4536a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4537a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4538a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4539a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4540a9ab775bSTejun Heo * be bound before @pool->lock is released. 4541a9ab775bSTejun Heo */ 4542a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4543bd7c089eSTejun Heo wake_up_process(worker->task); 4544bd7c089eSTejun Heo 4545bd7c089eSTejun Heo /* 4546a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4547a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4548a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4549a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4550a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4551a9ab775bSTejun Heo * concurrency management. Note that when or whether 4552a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4553a9ab775bSTejun Heo * 4554a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4555a9ab775bSTejun Heo * tested without holding any lock in 4556a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4557a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4558a9ab775bSTejun Heo * management operations. 4559bd7c089eSTejun Heo */ 4560a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4561a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4562a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4563a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4564bd7c089eSTejun Heo } 4565a9ab775bSTejun Heo 4566a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4567bd7c089eSTejun Heo } 4568bd7c089eSTejun Heo 45697dbc725eSTejun Heo /** 45707dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 45717dbc725eSTejun Heo * @pool: unbound pool of interest 45727dbc725eSTejun Heo * @cpu: the CPU which is coming up 45737dbc725eSTejun Heo * 45747dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 45757dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 45767dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 45777dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 45787dbc725eSTejun Heo */ 45797dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 45807dbc725eSTejun Heo { 45817dbc725eSTejun Heo static cpumask_t cpumask; 45827dbc725eSTejun Heo struct worker *worker; 45837dbc725eSTejun Heo 458492f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 45857dbc725eSTejun Heo 45867dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 45877dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 45887dbc725eSTejun Heo return; 45897dbc725eSTejun Heo 45907dbc725eSTejun Heo /* is @cpu the only online CPU? */ 45917dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 45927dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 45937dbc725eSTejun Heo return; 45947dbc725eSTejun Heo 45957dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4596da028469SLai Jiangshan for_each_pool_worker(worker, pool) 45977dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 45987dbc725eSTejun Heo pool->attrs->cpumask) < 0); 45997dbc725eSTejun Heo } 46007dbc725eSTejun Heo 46018db25e78STejun Heo /* 46028db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 46038db25e78STejun Heo * This will be registered high priority CPU notifier. 46048db25e78STejun Heo */ 46050db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 46061da177e4SLinus Torvalds unsigned long action, 46071da177e4SLinus Torvalds void *hcpu) 46081da177e4SLinus Torvalds { 4609d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46104ce62e9eSTejun Heo struct worker_pool *pool; 46114c16bd32STejun Heo struct workqueue_struct *wq; 46127dbc725eSTejun Heo int pi; 46131da177e4SLinus Torvalds 46148db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 46153af24433SOleg Nesterov case CPU_UP_PREPARE: 4616f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46173ce63377STejun Heo if (pool->nr_workers) 46183ce63377STejun Heo continue; 4619051e1850SLai Jiangshan if (!create_worker(pool)) 46203ce63377STejun Heo return NOTIFY_BAD; 46213af24433SOleg Nesterov } 46221da177e4SLinus Torvalds break; 46231da177e4SLinus Torvalds 462465758202STejun Heo case CPU_DOWN_FAILED: 462565758202STejun Heo case CPU_ONLINE: 462668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46277dbc725eSTejun Heo 46287dbc725eSTejun Heo for_each_pool(pool, pi) { 462992f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 463094cf58bbSTejun Heo 4631f05b558dSLai Jiangshan if (pool->cpu == cpu) 463294cf58bbSTejun Heo rebind_workers(pool); 4633f05b558dSLai Jiangshan else if (pool->cpu < 0) 46347dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 463594cf58bbSTejun Heo 463692f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 463794cf58bbSTejun Heo } 46387dbc725eSTejun Heo 46394c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46404c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46414c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 46424c16bd32STejun Heo 464368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46448db25e78STejun Heo break; 464565758202STejun Heo } 464665758202STejun Heo return NOTIFY_OK; 464765758202STejun Heo } 464865758202STejun Heo 464965758202STejun Heo /* 465065758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 465165758202STejun Heo * This will be registered as low priority CPU notifier. 465265758202STejun Heo */ 46530db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 465465758202STejun Heo unsigned long action, 465565758202STejun Heo void *hcpu) 465665758202STejun Heo { 4657d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46588db25e78STejun Heo struct work_struct unbind_work; 46594c16bd32STejun Heo struct workqueue_struct *wq; 46608db25e78STejun Heo 466165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 466265758202STejun Heo case CPU_DOWN_PREPARE: 46634c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4664706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46657635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46664c16bd32STejun Heo 46674c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46684c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46694c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46704c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 46714c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 46724c16bd32STejun Heo 46734c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 46748db25e78STejun Heo flush_work(&unbind_work); 4675440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 46768db25e78STejun Heo break; 467765758202STejun Heo } 467865758202STejun Heo return NOTIFY_OK; 467965758202STejun Heo } 468065758202STejun Heo 46812d3854a3SRusty Russell #ifdef CONFIG_SMP 46828ccad40dSRusty Russell 46832d3854a3SRusty Russell struct work_for_cpu { 4684ed48ece2STejun Heo struct work_struct work; 46852d3854a3SRusty Russell long (*fn)(void *); 46862d3854a3SRusty Russell void *arg; 46872d3854a3SRusty Russell long ret; 46882d3854a3SRusty Russell }; 46892d3854a3SRusty Russell 4690ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 46912d3854a3SRusty Russell { 4692ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4693ed48ece2STejun Heo 46942d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 46952d3854a3SRusty Russell } 46962d3854a3SRusty Russell 46972d3854a3SRusty Russell /** 46982d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 46992d3854a3SRusty Russell * @cpu: the cpu to run on 47002d3854a3SRusty Russell * @fn: the function to run 47012d3854a3SRusty Russell * @arg: the function arg 47022d3854a3SRusty Russell * 470331ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47046b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4705d185af30SYacine Belkadi * 4706d185af30SYacine Belkadi * Return: The value @fn returns. 47072d3854a3SRusty Russell */ 4708d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47092d3854a3SRusty Russell { 4710ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47112d3854a3SRusty Russell 4712ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4713ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 471412997d1aSBjorn Helgaas flush_work(&wfc.work); 4715440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 47162d3854a3SRusty Russell return wfc.ret; 47172d3854a3SRusty Russell } 47182d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47192d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47202d3854a3SRusty Russell 4721a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4722e7577c50SRusty Russell 4723a0a1a5fdSTejun Heo /** 4724a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4725a0a1a5fdSTejun Heo * 472658a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4727c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4728706026c2STejun Heo * pool->worklist. 4729a0a1a5fdSTejun Heo * 4730a0a1a5fdSTejun Heo * CONTEXT: 4731a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4732a0a1a5fdSTejun Heo */ 4733a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4734a0a1a5fdSTejun Heo { 473524b8a847STejun Heo struct workqueue_struct *wq; 473624b8a847STejun Heo struct pool_workqueue *pwq; 4737a0a1a5fdSTejun Heo 473868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4739a0a1a5fdSTejun Heo 47406183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4741a0a1a5fdSTejun Heo workqueue_freezing = true; 4742a0a1a5fdSTejun Heo 474324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4744a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4745699ce097STejun Heo for_each_pwq(pwq, wq) 4746699ce097STejun Heo pwq_adjust_max_active(pwq); 4747a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4748a1056305STejun Heo } 47495bcab335STejun Heo 475068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4751a0a1a5fdSTejun Heo } 4752a0a1a5fdSTejun Heo 4753a0a1a5fdSTejun Heo /** 475458a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4755a0a1a5fdSTejun Heo * 4756a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4757a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4758a0a1a5fdSTejun Heo * 4759a0a1a5fdSTejun Heo * CONTEXT: 476068e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4761a0a1a5fdSTejun Heo * 4762d185af30SYacine Belkadi * Return: 476358a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 476458a69cb4STejun Heo * is complete. 4765a0a1a5fdSTejun Heo */ 4766a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4767a0a1a5fdSTejun Heo { 4768a0a1a5fdSTejun Heo bool busy = false; 476924b8a847STejun Heo struct workqueue_struct *wq; 477024b8a847STejun Heo struct pool_workqueue *pwq; 4771a0a1a5fdSTejun Heo 477268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4773a0a1a5fdSTejun Heo 47746183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4775a0a1a5fdSTejun Heo 477624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 477724b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 477824b8a847STejun Heo continue; 4779a0a1a5fdSTejun Heo /* 4780a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4781a0a1a5fdSTejun Heo * to peek without lock. 4782a0a1a5fdSTejun Heo */ 478388109453SLai Jiangshan rcu_read_lock_sched(); 478424b8a847STejun Heo for_each_pwq(pwq, wq) { 47856183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4786112202d9STejun Heo if (pwq->nr_active) { 4787a0a1a5fdSTejun Heo busy = true; 478888109453SLai Jiangshan rcu_read_unlock_sched(); 4789a0a1a5fdSTejun Heo goto out_unlock; 4790a0a1a5fdSTejun Heo } 4791a0a1a5fdSTejun Heo } 479288109453SLai Jiangshan rcu_read_unlock_sched(); 4793a0a1a5fdSTejun Heo } 4794a0a1a5fdSTejun Heo out_unlock: 479568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4796a0a1a5fdSTejun Heo return busy; 4797a0a1a5fdSTejun Heo } 4798a0a1a5fdSTejun Heo 4799a0a1a5fdSTejun Heo /** 4800a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4801a0a1a5fdSTejun Heo * 4802a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4803706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4804a0a1a5fdSTejun Heo * 4805a0a1a5fdSTejun Heo * CONTEXT: 4806a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4807a0a1a5fdSTejun Heo */ 4808a0a1a5fdSTejun Heo void thaw_workqueues(void) 4809a0a1a5fdSTejun Heo { 481024b8a847STejun Heo struct workqueue_struct *wq; 481124b8a847STejun Heo struct pool_workqueue *pwq; 4812a0a1a5fdSTejun Heo 481368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4814a0a1a5fdSTejun Heo 4815a0a1a5fdSTejun Heo if (!workqueue_freezing) 4816a0a1a5fdSTejun Heo goto out_unlock; 4817a0a1a5fdSTejun Heo 481874b414eaSLai Jiangshan workqueue_freezing = false; 481924b8a847STejun Heo 482024b8a847STejun Heo /* restore max_active and repopulate worklist */ 482124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4822a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4823699ce097STejun Heo for_each_pwq(pwq, wq) 4824699ce097STejun Heo pwq_adjust_max_active(pwq); 4825a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 482624b8a847STejun Heo } 482724b8a847STejun Heo 4828a0a1a5fdSTejun Heo out_unlock: 482968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4830a0a1a5fdSTejun Heo } 4831a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4832a0a1a5fdSTejun Heo 4833042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4834042f7df1SLai Jiangshan { 4835042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4836042f7df1SLai Jiangshan int ret = 0; 4837042f7df1SLai Jiangshan struct workqueue_struct *wq; 4838042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4839042f7df1SLai Jiangshan 4840042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4841042f7df1SLai Jiangshan 4842042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4843042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4844042f7df1SLai Jiangshan continue; 4845042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 4846042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 4847042f7df1SLai Jiangshan continue; 4848042f7df1SLai Jiangshan 4849042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 4850042f7df1SLai Jiangshan if (!ctx) { 4851042f7df1SLai Jiangshan ret = -ENOMEM; 4852042f7df1SLai Jiangshan break; 4853042f7df1SLai Jiangshan } 4854042f7df1SLai Jiangshan 4855042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 4856042f7df1SLai Jiangshan } 4857042f7df1SLai Jiangshan 4858042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 4859042f7df1SLai Jiangshan if (!ret) 4860042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 4861042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 4862042f7df1SLai Jiangshan } 4863042f7df1SLai Jiangshan 4864042f7df1SLai Jiangshan return ret; 4865042f7df1SLai Jiangshan } 4866042f7df1SLai Jiangshan 4867042f7df1SLai Jiangshan /** 4868042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 4869042f7df1SLai Jiangshan * @cpumask: the cpumask to set 4870042f7df1SLai Jiangshan * 4871042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 4872042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 4873042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 4874042f7df1SLai Jiangshan * 4875042f7df1SLai Jiangshan * Retun: 0 - Success 4876042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 4877042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 4878042f7df1SLai Jiangshan */ 4879042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 4880042f7df1SLai Jiangshan { 4881042f7df1SLai Jiangshan int ret = -EINVAL; 4882042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 4883042f7df1SLai Jiangshan 4884042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 4885042f7df1SLai Jiangshan return -ENOMEM; 4886042f7df1SLai Jiangshan 4887042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 4888042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 4889a0111cf6SLai Jiangshan apply_wqattrs_lock(); 4890042f7df1SLai Jiangshan 4891042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 4892042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 4893042f7df1SLai Jiangshan 4894042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 4895042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 4896042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 4897042f7df1SLai Jiangshan 4898042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 4899042f7df1SLai Jiangshan if (ret < 0) 4900042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 4901042f7df1SLai Jiangshan 4902a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 4903042f7df1SLai Jiangshan } 4904042f7df1SLai Jiangshan 4905042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 4906042f7df1SLai Jiangshan return ret; 4907042f7df1SLai Jiangshan } 4908042f7df1SLai Jiangshan 49096ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 49106ba94429SFrederic Weisbecker /* 49116ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 49126ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 49136ba94429SFrederic Weisbecker * following attributes. 49146ba94429SFrederic Weisbecker * 49156ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 49166ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 49176ba94429SFrederic Weisbecker * 49186ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 49196ba94429SFrederic Weisbecker * 49206ba94429SFrederic Weisbecker * id RO int : the associated pool ID 49216ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 49226ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 49236ba94429SFrederic Weisbecker */ 49246ba94429SFrederic Weisbecker struct wq_device { 49256ba94429SFrederic Weisbecker struct workqueue_struct *wq; 49266ba94429SFrederic Weisbecker struct device dev; 49276ba94429SFrederic Weisbecker }; 49286ba94429SFrederic Weisbecker 49296ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 49306ba94429SFrederic Weisbecker { 49316ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 49326ba94429SFrederic Weisbecker 49336ba94429SFrederic Weisbecker return wq_dev->wq; 49346ba94429SFrederic Weisbecker } 49356ba94429SFrederic Weisbecker 49366ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 49376ba94429SFrederic Weisbecker char *buf) 49386ba94429SFrederic Weisbecker { 49396ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49406ba94429SFrederic Weisbecker 49416ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 49426ba94429SFrederic Weisbecker } 49436ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 49446ba94429SFrederic Weisbecker 49456ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 49466ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49476ba94429SFrederic Weisbecker { 49486ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49496ba94429SFrederic Weisbecker 49506ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 49516ba94429SFrederic Weisbecker } 49526ba94429SFrederic Weisbecker 49536ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 49546ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 49556ba94429SFrederic Weisbecker size_t count) 49566ba94429SFrederic Weisbecker { 49576ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49586ba94429SFrederic Weisbecker int val; 49596ba94429SFrederic Weisbecker 49606ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 49616ba94429SFrederic Weisbecker return -EINVAL; 49626ba94429SFrederic Weisbecker 49636ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 49646ba94429SFrederic Weisbecker return count; 49656ba94429SFrederic Weisbecker } 49666ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 49676ba94429SFrederic Weisbecker 49686ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 49696ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 49706ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 49716ba94429SFrederic Weisbecker NULL, 49726ba94429SFrederic Weisbecker }; 49736ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 49746ba94429SFrederic Weisbecker 49756ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 49766ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49776ba94429SFrederic Weisbecker { 49786ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49796ba94429SFrederic Weisbecker const char *delim = ""; 49806ba94429SFrederic Weisbecker int node, written = 0; 49816ba94429SFrederic Weisbecker 49826ba94429SFrederic Weisbecker rcu_read_lock_sched(); 49836ba94429SFrederic Weisbecker for_each_node(node) { 49846ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 49856ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 49866ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 49876ba94429SFrederic Weisbecker delim = " "; 49886ba94429SFrederic Weisbecker } 49896ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 49906ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 49916ba94429SFrederic Weisbecker 49926ba94429SFrederic Weisbecker return written; 49936ba94429SFrederic Weisbecker } 49946ba94429SFrederic Weisbecker 49956ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 49966ba94429SFrederic Weisbecker char *buf) 49976ba94429SFrederic Weisbecker { 49986ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49996ba94429SFrederic Weisbecker int written; 50006ba94429SFrederic Weisbecker 50016ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50026ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 50036ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50046ba94429SFrederic Weisbecker 50056ba94429SFrederic Weisbecker return written; 50066ba94429SFrederic Weisbecker } 50076ba94429SFrederic Weisbecker 50086ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 50096ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 50106ba94429SFrederic Weisbecker { 50116ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 50126ba94429SFrederic Weisbecker 5013899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5014899a94feSLai Jiangshan 50156ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 50166ba94429SFrederic Weisbecker if (!attrs) 50176ba94429SFrederic Weisbecker return NULL; 50186ba94429SFrederic Weisbecker 50196ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 50206ba94429SFrederic Weisbecker return attrs; 50216ba94429SFrederic Weisbecker } 50226ba94429SFrederic Weisbecker 50236ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 50246ba94429SFrederic Weisbecker const char *buf, size_t count) 50256ba94429SFrederic Weisbecker { 50266ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50276ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5028d4d3e257SLai Jiangshan int ret = -ENOMEM; 5029d4d3e257SLai Jiangshan 5030d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50316ba94429SFrederic Weisbecker 50326ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50336ba94429SFrederic Weisbecker if (!attrs) 5034d4d3e257SLai Jiangshan goto out_unlock; 50356ba94429SFrederic Weisbecker 50366ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 50376ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5038d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50396ba94429SFrederic Weisbecker else 50406ba94429SFrederic Weisbecker ret = -EINVAL; 50416ba94429SFrederic Weisbecker 5042d4d3e257SLai Jiangshan out_unlock: 5043d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50446ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50456ba94429SFrederic Weisbecker return ret ?: count; 50466ba94429SFrederic Weisbecker } 50476ba94429SFrederic Weisbecker 50486ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 50496ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 50506ba94429SFrederic Weisbecker { 50516ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50526ba94429SFrederic Weisbecker int written; 50536ba94429SFrederic Weisbecker 50546ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50556ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 50566ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 50576ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50586ba94429SFrederic Weisbecker return written; 50596ba94429SFrederic Weisbecker } 50606ba94429SFrederic Weisbecker 50616ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 50626ba94429SFrederic Weisbecker struct device_attribute *attr, 50636ba94429SFrederic Weisbecker const char *buf, size_t count) 50646ba94429SFrederic Weisbecker { 50656ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50666ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5067d4d3e257SLai Jiangshan int ret = -ENOMEM; 5068d4d3e257SLai Jiangshan 5069d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50706ba94429SFrederic Weisbecker 50716ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50726ba94429SFrederic Weisbecker if (!attrs) 5073d4d3e257SLai Jiangshan goto out_unlock; 50746ba94429SFrederic Weisbecker 50756ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 50766ba94429SFrederic Weisbecker if (!ret) 5077d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50786ba94429SFrederic Weisbecker 5079d4d3e257SLai Jiangshan out_unlock: 5080d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50816ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50826ba94429SFrederic Weisbecker return ret ?: count; 50836ba94429SFrederic Weisbecker } 50846ba94429SFrederic Weisbecker 50856ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 50866ba94429SFrederic Weisbecker char *buf) 50876ba94429SFrederic Weisbecker { 50886ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50896ba94429SFrederic Weisbecker int written; 50906ba94429SFrederic Weisbecker 50916ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50926ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 50936ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 50946ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50956ba94429SFrederic Weisbecker 50966ba94429SFrederic Weisbecker return written; 50976ba94429SFrederic Weisbecker } 50986ba94429SFrederic Weisbecker 50996ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 51006ba94429SFrederic Weisbecker const char *buf, size_t count) 51016ba94429SFrederic Weisbecker { 51026ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51036ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5104d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5105d4d3e257SLai Jiangshan 5106d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51076ba94429SFrederic Weisbecker 51086ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51096ba94429SFrederic Weisbecker if (!attrs) 5110d4d3e257SLai Jiangshan goto out_unlock; 51116ba94429SFrederic Weisbecker 51126ba94429SFrederic Weisbecker ret = -EINVAL; 51136ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 51146ba94429SFrederic Weisbecker attrs->no_numa = !v; 5115d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51166ba94429SFrederic Weisbecker } 51176ba94429SFrederic Weisbecker 5118d4d3e257SLai Jiangshan out_unlock: 5119d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 51206ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 51216ba94429SFrederic Weisbecker return ret ?: count; 51226ba94429SFrederic Weisbecker } 51236ba94429SFrederic Weisbecker 51246ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 51256ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 51266ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 51276ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 51286ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 51296ba94429SFrederic Weisbecker __ATTR_NULL, 51306ba94429SFrederic Weisbecker }; 51316ba94429SFrederic Weisbecker 51326ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 51336ba94429SFrederic Weisbecker .name = "workqueue", 51346ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 51356ba94429SFrederic Weisbecker }; 51366ba94429SFrederic Weisbecker 5137b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5138b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5139b05a7928SFrederic Weisbecker { 5140b05a7928SFrederic Weisbecker int written; 5141b05a7928SFrederic Weisbecker 5142042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5143b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5144b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5145042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5146b05a7928SFrederic Weisbecker 5147b05a7928SFrederic Weisbecker return written; 5148b05a7928SFrederic Weisbecker } 5149b05a7928SFrederic Weisbecker 5150042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5151042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5152042f7df1SLai Jiangshan { 5153042f7df1SLai Jiangshan cpumask_var_t cpumask; 5154042f7df1SLai Jiangshan int ret; 5155042f7df1SLai Jiangshan 5156042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5157042f7df1SLai Jiangshan return -ENOMEM; 5158042f7df1SLai Jiangshan 5159042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5160042f7df1SLai Jiangshan if (!ret) 5161042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5162042f7df1SLai Jiangshan 5163042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5164042f7df1SLai Jiangshan return ret ? ret : count; 5165042f7df1SLai Jiangshan } 5166042f7df1SLai Jiangshan 5167b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5168042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5169042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5170b05a7928SFrederic Weisbecker 51716ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 51726ba94429SFrederic Weisbecker { 5173b05a7928SFrederic Weisbecker int err; 5174b05a7928SFrederic Weisbecker 5175b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5176b05a7928SFrederic Weisbecker if (err) 5177b05a7928SFrederic Weisbecker return err; 5178b05a7928SFrederic Weisbecker 5179b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 51806ba94429SFrederic Weisbecker } 51816ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 51826ba94429SFrederic Weisbecker 51836ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 51846ba94429SFrederic Weisbecker { 51856ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 51866ba94429SFrederic Weisbecker 51876ba94429SFrederic Weisbecker kfree(wq_dev); 51886ba94429SFrederic Weisbecker } 51896ba94429SFrederic Weisbecker 51906ba94429SFrederic Weisbecker /** 51916ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 51926ba94429SFrederic Weisbecker * @wq: the workqueue to register 51936ba94429SFrederic Weisbecker * 51946ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 51956ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 51966ba94429SFrederic Weisbecker * which is the preferred method. 51976ba94429SFrederic Weisbecker * 51986ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 51996ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 52006ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 52016ba94429SFrederic Weisbecker * attributes. 52026ba94429SFrederic Weisbecker * 52036ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 52046ba94429SFrederic Weisbecker */ 52056ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 52066ba94429SFrederic Weisbecker { 52076ba94429SFrederic Weisbecker struct wq_device *wq_dev; 52086ba94429SFrederic Weisbecker int ret; 52096ba94429SFrederic Weisbecker 52106ba94429SFrederic Weisbecker /* 5211402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 52126ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 52136ba94429SFrederic Weisbecker * workqueues. 52146ba94429SFrederic Weisbecker */ 52156ba94429SFrederic Weisbecker if (WARN_ON(wq->flags & __WQ_ORDERED)) 52166ba94429SFrederic Weisbecker return -EINVAL; 52176ba94429SFrederic Weisbecker 52186ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 52196ba94429SFrederic Weisbecker if (!wq_dev) 52206ba94429SFrederic Weisbecker return -ENOMEM; 52216ba94429SFrederic Weisbecker 52226ba94429SFrederic Weisbecker wq_dev->wq = wq; 52236ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 52246ba94429SFrederic Weisbecker wq_dev->dev.init_name = wq->name; 52256ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 52266ba94429SFrederic Weisbecker 52276ba94429SFrederic Weisbecker /* 52286ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 52296ba94429SFrederic Weisbecker * everything is ready. 52306ba94429SFrederic Weisbecker */ 52316ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 52326ba94429SFrederic Weisbecker 52336ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 52346ba94429SFrederic Weisbecker if (ret) { 52356ba94429SFrederic Weisbecker kfree(wq_dev); 52366ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52376ba94429SFrederic Weisbecker return ret; 52386ba94429SFrederic Weisbecker } 52396ba94429SFrederic Weisbecker 52406ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 52416ba94429SFrederic Weisbecker struct device_attribute *attr; 52426ba94429SFrederic Weisbecker 52436ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 52446ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 52456ba94429SFrederic Weisbecker if (ret) { 52466ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 52476ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52486ba94429SFrederic Weisbecker return ret; 52496ba94429SFrederic Weisbecker } 52506ba94429SFrederic Weisbecker } 52516ba94429SFrederic Weisbecker } 52526ba94429SFrederic Weisbecker 52536ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 52546ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 52556ba94429SFrederic Weisbecker return 0; 52566ba94429SFrederic Weisbecker } 52576ba94429SFrederic Weisbecker 52586ba94429SFrederic Weisbecker /** 52596ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 52606ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 52616ba94429SFrederic Weisbecker * 52626ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 52636ba94429SFrederic Weisbecker */ 52646ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 52656ba94429SFrederic Weisbecker { 52666ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 52676ba94429SFrederic Weisbecker 52686ba94429SFrederic Weisbecker if (!wq->wq_dev) 52696ba94429SFrederic Weisbecker return; 52706ba94429SFrederic Weisbecker 52716ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52726ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 52736ba94429SFrederic Weisbecker } 52746ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 52756ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 52766ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 52776ba94429SFrederic Weisbecker 527882607adcSTejun Heo /* 527982607adcSTejun Heo * Workqueue watchdog. 528082607adcSTejun Heo * 528182607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 528282607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 528382607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 528482607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 528582607adcSTejun Heo * largely opaque. 528682607adcSTejun Heo * 528782607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 528882607adcSTejun Heo * state if some pools failed to make forward progress for a while where 528982607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 529082607adcSTejun Heo * 529182607adcSTejun Heo * This mechanism is controlled through the kernel parameter 529282607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 529382607adcSTejun Heo * corresponding sysfs parameter file. 529482607adcSTejun Heo */ 529582607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 529682607adcSTejun Heo 529782607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data); 529882607adcSTejun Heo 529982607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 530082607adcSTejun Heo static struct timer_list wq_watchdog_timer = 530182607adcSTejun Heo TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0); 530282607adcSTejun Heo 530382607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 530482607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 530582607adcSTejun Heo 530682607adcSTejun Heo static void wq_watchdog_reset_touched(void) 530782607adcSTejun Heo { 530882607adcSTejun Heo int cpu; 530982607adcSTejun Heo 531082607adcSTejun Heo wq_watchdog_touched = jiffies; 531182607adcSTejun Heo for_each_possible_cpu(cpu) 531282607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 531382607adcSTejun Heo } 531482607adcSTejun Heo 531582607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data) 531682607adcSTejun Heo { 531782607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 531882607adcSTejun Heo bool lockup_detected = false; 531982607adcSTejun Heo struct worker_pool *pool; 532082607adcSTejun Heo int pi; 532182607adcSTejun Heo 532282607adcSTejun Heo if (!thresh) 532382607adcSTejun Heo return; 532482607adcSTejun Heo 532582607adcSTejun Heo rcu_read_lock(); 532682607adcSTejun Heo 532782607adcSTejun Heo for_each_pool(pool, pi) { 532882607adcSTejun Heo unsigned long pool_ts, touched, ts; 532982607adcSTejun Heo 533082607adcSTejun Heo if (list_empty(&pool->worklist)) 533182607adcSTejun Heo continue; 533282607adcSTejun Heo 533382607adcSTejun Heo /* get the latest of pool and touched timestamps */ 533482607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 533582607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 533682607adcSTejun Heo 533782607adcSTejun Heo if (time_after(pool_ts, touched)) 533882607adcSTejun Heo ts = pool_ts; 533982607adcSTejun Heo else 534082607adcSTejun Heo ts = touched; 534182607adcSTejun Heo 534282607adcSTejun Heo if (pool->cpu >= 0) { 534382607adcSTejun Heo unsigned long cpu_touched = 534482607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 534582607adcSTejun Heo pool->cpu)); 534682607adcSTejun Heo if (time_after(cpu_touched, ts)) 534782607adcSTejun Heo ts = cpu_touched; 534882607adcSTejun Heo } 534982607adcSTejun Heo 535082607adcSTejun Heo /* did we stall? */ 535182607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 535282607adcSTejun Heo lockup_detected = true; 535382607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 535482607adcSTejun Heo pr_cont_pool_info(pool); 535582607adcSTejun Heo pr_cont(" stuck for %us!\n", 535682607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 535782607adcSTejun Heo } 535882607adcSTejun Heo } 535982607adcSTejun Heo 536082607adcSTejun Heo rcu_read_unlock(); 536182607adcSTejun Heo 536282607adcSTejun Heo if (lockup_detected) 536382607adcSTejun Heo show_workqueue_state(); 536482607adcSTejun Heo 536582607adcSTejun Heo wq_watchdog_reset_touched(); 536682607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 536782607adcSTejun Heo } 536882607adcSTejun Heo 536982607adcSTejun Heo void wq_watchdog_touch(int cpu) 537082607adcSTejun Heo { 537182607adcSTejun Heo if (cpu >= 0) 537282607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 537382607adcSTejun Heo else 537482607adcSTejun Heo wq_watchdog_touched = jiffies; 537582607adcSTejun Heo } 537682607adcSTejun Heo 537782607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 537882607adcSTejun Heo { 537982607adcSTejun Heo wq_watchdog_thresh = 0; 538082607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 538182607adcSTejun Heo 538282607adcSTejun Heo if (thresh) { 538382607adcSTejun Heo wq_watchdog_thresh = thresh; 538482607adcSTejun Heo wq_watchdog_reset_touched(); 538582607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 538682607adcSTejun Heo } 538782607adcSTejun Heo } 538882607adcSTejun Heo 538982607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 539082607adcSTejun Heo const struct kernel_param *kp) 539182607adcSTejun Heo { 539282607adcSTejun Heo unsigned long thresh; 539382607adcSTejun Heo int ret; 539482607adcSTejun Heo 539582607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 539682607adcSTejun Heo if (ret) 539782607adcSTejun Heo return ret; 539882607adcSTejun Heo 539982607adcSTejun Heo if (system_wq) 540082607adcSTejun Heo wq_watchdog_set_thresh(thresh); 540182607adcSTejun Heo else 540282607adcSTejun Heo wq_watchdog_thresh = thresh; 540382607adcSTejun Heo 540482607adcSTejun Heo return 0; 540582607adcSTejun Heo } 540682607adcSTejun Heo 540782607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 540882607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 540982607adcSTejun Heo .get = param_get_ulong, 541082607adcSTejun Heo }; 541182607adcSTejun Heo 541282607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 541382607adcSTejun Heo 0644); 541482607adcSTejun Heo 541582607adcSTejun Heo static void wq_watchdog_init(void) 541682607adcSTejun Heo { 541782607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 541882607adcSTejun Heo } 541982607adcSTejun Heo 542082607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 542182607adcSTejun Heo 542282607adcSTejun Heo static inline void wq_watchdog_init(void) { } 542382607adcSTejun Heo 542482607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 542582607adcSTejun Heo 5426bce90380STejun Heo static void __init wq_numa_init(void) 5427bce90380STejun Heo { 5428bce90380STejun Heo cpumask_var_t *tbl; 5429bce90380STejun Heo int node, cpu; 5430bce90380STejun Heo 5431bce90380STejun Heo if (num_possible_nodes() <= 1) 5432bce90380STejun Heo return; 5433bce90380STejun Heo 5434d55262c4STejun Heo if (wq_disable_numa) { 5435d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5436d55262c4STejun Heo return; 5437d55262c4STejun Heo } 5438d55262c4STejun Heo 54394c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 54404c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 54414c16bd32STejun Heo 5442bce90380STejun Heo /* 5443bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5444bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5445bce90380STejun Heo * fully initialized by now. 5446bce90380STejun Heo */ 5447ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5448bce90380STejun Heo BUG_ON(!tbl); 5449bce90380STejun Heo 5450bce90380STejun Heo for_each_node(node) 54515a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 54521be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5453bce90380STejun Heo 5454bce90380STejun Heo for_each_possible_cpu(cpu) { 5455bce90380STejun Heo node = cpu_to_node(cpu); 5456bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5457bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5458bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5459bce90380STejun Heo return; 5460bce90380STejun Heo } 5461bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5462bce90380STejun Heo } 5463bce90380STejun Heo 5464bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5465bce90380STejun Heo wq_numa_enabled = true; 5466bce90380STejun Heo } 5467bce90380STejun Heo 54686ee0578bSSuresh Siddha static int __init init_workqueues(void) 54691da177e4SLinus Torvalds { 54707a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 54717a4e344cSTejun Heo int i, cpu; 5472c34056a3STejun Heo 5473e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5474e904e6c2STejun Heo 5475b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5476b05a7928SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 5477b05a7928SFrederic Weisbecker 5478e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5479e904e6c2STejun Heo 548065758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 5481a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 54828b03ae3cSTejun Heo 5483bce90380STejun Heo wq_numa_init(); 5484bce90380STejun Heo 5485706026c2STejun Heo /* initialize CPU pools */ 548629c91e99STejun Heo for_each_possible_cpu(cpu) { 54874ce62e9eSTejun Heo struct worker_pool *pool; 54888b03ae3cSTejun Heo 54897a4e344cSTejun Heo i = 0; 5490f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54917a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5492ec22ca5eSTejun Heo pool->cpu = cpu; 54937a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 54947a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5495f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 54967a4e344cSTejun Heo 54979daf9e67STejun Heo /* alloc pool ID */ 549868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 54999daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 550068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55014ce62e9eSTejun Heo } 55028b03ae3cSTejun Heo } 55038b03ae3cSTejun Heo 5504e22bee78STejun Heo /* create the initial worker */ 550529c91e99STejun Heo for_each_online_cpu(cpu) { 55064ce62e9eSTejun Heo struct worker_pool *pool; 5507e22bee78STejun Heo 5508f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 550924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5510051e1850SLai Jiangshan BUG_ON(!create_worker(pool)); 5511e22bee78STejun Heo } 55124ce62e9eSTejun Heo } 5513e22bee78STejun Heo 55148a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 551529c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 551629c91e99STejun Heo struct workqueue_attrs *attrs; 551729c91e99STejun Heo 551829c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 551929c91e99STejun Heo attrs->nice = std_nice[i]; 552029c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 55218a2b7538STejun Heo 55228a2b7538STejun Heo /* 55238a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 55248a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 55258a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 55268a2b7538STejun Heo */ 55278a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 55288a2b7538STejun Heo attrs->nice = std_nice[i]; 55298a2b7538STejun Heo attrs->no_numa = true; 55308a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 553129c91e99STejun Heo } 553229c91e99STejun Heo 5533d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 55341aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5535d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5536f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5537f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 553824d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 553924d51addSTejun Heo WQ_FREEZABLE, 0); 55400668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 55410668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 55420668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 55430668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 55440668106cSViresh Kumar 0); 55451aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 55460668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 55470668106cSViresh Kumar !system_power_efficient_wq || 55480668106cSViresh Kumar !system_freezable_power_efficient_wq); 554982607adcSTejun Heo 555082607adcSTejun Heo wq_watchdog_init(); 555182607adcSTejun Heo 55526ee0578bSSuresh Siddha return 0; 55531da177e4SLinus Torvalds } 55546ee0578bSSuresh Siddha early_initcall(init_workqueues); 5555