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 293863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 2943347fa09STejun Heo 295bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 296bce90380STejun Heo 2974c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2984c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2994c16bd32STejun Heo 30068e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3012e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 3025bcab335STejun Heo 303e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30468e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3057d19c5ceSTejun Heo 306ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 307ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 308ef557180SMike Galbraith 309ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 310ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 311b05a7928SFrederic Weisbecker 312f303fccbSTejun Heo /* 313f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 314f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 315f303fccbSTejun Heo * to uncover usages which depend on it. 316f303fccbSTejun Heo */ 317f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 318f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 319f303fccbSTejun Heo #else 320f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 321f303fccbSTejun Heo #endif 322f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 323f303fccbSTejun Heo 3247d19c5ceSTejun Heo /* the per-cpu worker pools */ 32525528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3267d19c5ceSTejun Heo 32768e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3287d19c5ceSTejun Heo 32968e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 33029c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 33129c91e99STejun Heo 332c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33329c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33429c91e99STejun Heo 3358a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3368a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3378a2b7538STejun Heo 338d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 339ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 340044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3411aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 342044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 343d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 344044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 345f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 346044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 34724d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3480668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3490668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3500668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3510668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 352d320c038STejun Heo 3537d19c5ceSTejun Heo static int worker_thread(void *__worker); 3546ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3557d19c5ceSTejun Heo 35697bd2347STejun Heo #define CREATE_TRACE_POINTS 35797bd2347STejun Heo #include <trace/events/workqueue.h> 35897bd2347STejun Heo 35968e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 360f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 361f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36268e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3635bcab335STejun Heo 364b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 365f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 366f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 367b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 36876af4d93STejun Heo 3695b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 370f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 371f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 372f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3735b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3745b95e1afSLai Jiangshan 375f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 376f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 377f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3787a62c2c8STejun Heo (pool)++) 3794ce62e9eSTejun Heo 38049e3cf44STejun Heo /** 38117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38217116969STejun Heo * @pool: iteration cursor 383611c92a0STejun Heo * @pi: integer used for iteration 384fa1b54e6STejun Heo * 38568e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 38668e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38768e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 388fa1b54e6STejun Heo * 389fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 390fa1b54e6STejun Heo * ignored. 39117116969STejun Heo */ 392611c92a0STejun Heo #define for_each_pool(pool, pi) \ 393611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39468e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 395fa1b54e6STejun Heo else 39617116969STejun Heo 39717116969STejun Heo /** 398822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 399822d8405STejun Heo * @worker: iteration cursor 400822d8405STejun Heo * @pool: worker_pool to iterate workers of 401822d8405STejun Heo * 40292f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 403822d8405STejun Heo * 404822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 405822d8405STejun Heo * ignored. 406822d8405STejun Heo */ 407da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 408da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 40992f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 410822d8405STejun Heo else 411822d8405STejun Heo 412822d8405STejun Heo /** 41349e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41449e3cf44STejun Heo * @pwq: iteration cursor 41549e3cf44STejun Heo * @wq: the target workqueue 41676af4d93STejun Heo * 417b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 418794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 419794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 42076af4d93STejun Heo * 42176af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42276af4d93STejun Heo * ignored. 42349e3cf44STejun Heo */ 42449e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42576af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 426b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 42776af4d93STejun Heo else 428f3421797STejun Heo 429dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 430dc186ad7SThomas Gleixner 431dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 432dc186ad7SThomas Gleixner 43399777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43499777288SStanislaw Gruszka { 43599777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43699777288SStanislaw Gruszka } 43799777288SStanislaw Gruszka 438b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 439b9fdac7fSDu, Changbin { 440b9fdac7fSDu, Changbin struct work_struct *work = addr; 441b9fdac7fSDu, Changbin 442b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 443b9fdac7fSDu, Changbin } 444b9fdac7fSDu, Changbin 445dc186ad7SThomas Gleixner /* 446dc186ad7SThomas Gleixner * fixup_init is called when: 447dc186ad7SThomas Gleixner * - an active object is initialized 448dc186ad7SThomas Gleixner */ 44902a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 450dc186ad7SThomas Gleixner { 451dc186ad7SThomas Gleixner struct work_struct *work = addr; 452dc186ad7SThomas Gleixner 453dc186ad7SThomas Gleixner switch (state) { 454dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 455dc186ad7SThomas Gleixner cancel_work_sync(work); 456dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45702a982a6SDu, Changbin return true; 458dc186ad7SThomas Gleixner default: 45902a982a6SDu, Changbin return false; 460dc186ad7SThomas Gleixner } 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner /* 464dc186ad7SThomas Gleixner * fixup_free is called when: 465dc186ad7SThomas Gleixner * - an active object is freed 466dc186ad7SThomas Gleixner */ 46702a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 468dc186ad7SThomas Gleixner { 469dc186ad7SThomas Gleixner struct work_struct *work = addr; 470dc186ad7SThomas Gleixner 471dc186ad7SThomas Gleixner switch (state) { 472dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 473dc186ad7SThomas Gleixner cancel_work_sync(work); 474dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47502a982a6SDu, Changbin return true; 476dc186ad7SThomas Gleixner default: 47702a982a6SDu, Changbin return false; 478dc186ad7SThomas Gleixner } 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner 481dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 482dc186ad7SThomas Gleixner .name = "work_struct", 48399777288SStanislaw Gruszka .debug_hint = work_debug_hint, 484b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 485dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 486dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 487dc186ad7SThomas Gleixner }; 488dc186ad7SThomas Gleixner 489dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 490dc186ad7SThomas Gleixner { 491dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 492dc186ad7SThomas Gleixner } 493dc186ad7SThomas Gleixner 494dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 495dc186ad7SThomas Gleixner { 496dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner 499dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 500dc186ad7SThomas Gleixner { 501dc186ad7SThomas Gleixner if (onstack) 502dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 503dc186ad7SThomas Gleixner else 504dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 505dc186ad7SThomas Gleixner } 506dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 507dc186ad7SThomas Gleixner 508dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 511dc186ad7SThomas Gleixner } 512dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 513dc186ad7SThomas Gleixner 514ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 515ea2e64f2SThomas Gleixner { 516ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 517ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 518ea2e64f2SThomas Gleixner } 519ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 520ea2e64f2SThomas Gleixner 521dc186ad7SThomas Gleixner #else 522dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 523dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 524dc186ad7SThomas Gleixner #endif 525dc186ad7SThomas Gleixner 5264e8b22bdSLi Bin /** 5274e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5284e8b22bdSLi Bin * @pool: the pool pointer of interest 5294e8b22bdSLi Bin * 5304e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5314e8b22bdSLi Bin * successfully, -errno on failure. 5324e8b22bdSLi Bin */ 5339daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5349daf9e67STejun Heo { 5359daf9e67STejun Heo int ret; 5369daf9e67STejun Heo 53768e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5385bcab335STejun Heo 5394e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5404e8b22bdSLi Bin GFP_KERNEL); 541229641a6STejun Heo if (ret >= 0) { 542e68035fbSTejun Heo pool->id = ret; 543229641a6STejun Heo return 0; 544229641a6STejun Heo } 5459daf9e67STejun Heo return ret; 5469daf9e67STejun Heo } 5479daf9e67STejun Heo 54876af4d93STejun Heo /** 549df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 550df2d5ae4STejun Heo * @wq: the target workqueue 551df2d5ae4STejun Heo * @node: the node ID 552df2d5ae4STejun Heo * 5535b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5545b95e1afSLai Jiangshan * read locked. 555df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 556df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 557d185af30SYacine Belkadi * 558d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 559df2d5ae4STejun Heo */ 560df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 561df2d5ae4STejun Heo int node) 562df2d5ae4STejun Heo { 5635b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 564d6e022f1STejun Heo 565d6e022f1STejun Heo /* 566d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 567d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 568d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 569d6e022f1STejun Heo * happens, this workaround can be removed. 570d6e022f1STejun Heo */ 571d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 572d6e022f1STejun Heo return wq->dfl_pwq; 573d6e022f1STejun Heo 574df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 575df2d5ae4STejun Heo } 576df2d5ae4STejun Heo 57773f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57873f53c4aSTejun Heo { 57973f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 58073f53c4aSTejun Heo } 58173f53c4aSTejun Heo 58273f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58373f53c4aSTejun Heo { 58473f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58573f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58673f53c4aSTejun Heo } 58773f53c4aSTejun Heo 58873f53c4aSTejun Heo static int work_next_color(int color) 58973f53c4aSTejun Heo { 59073f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 591a848e3b6SOleg Nesterov } 592a848e3b6SOleg Nesterov 5934594bf15SDavid Howells /* 594112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 595112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5967c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5977a22ad75STejun Heo * 598112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 599112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 600bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 601bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6027a22ad75STejun Heo * 603112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6047c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 605112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6067c3eed5cSTejun Heo * available only while the work item is queued. 607bbb68dfaSTejun Heo * 608bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 609bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 610bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 611bbb68dfaSTejun Heo * try to steal the PENDING bit. 6124594bf15SDavid Howells */ 6137a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6147a22ad75STejun Heo unsigned long flags) 6157a22ad75STejun Heo { 6166183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6177a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6187a22ad75STejun Heo } 6197a22ad75STejun Heo 620112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6214690c4abSTejun Heo unsigned long extra_flags) 622365970a1SDavid Howells { 623112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 624112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 625365970a1SDavid Howells } 626365970a1SDavid Howells 6274468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6284468a00fSLai Jiangshan int pool_id) 6294468a00fSLai Jiangshan { 6304468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6314468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6324468a00fSLai Jiangshan } 6334468a00fSLai Jiangshan 6347c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6357c3eed5cSTejun Heo int pool_id) 6364d707b9fSOleg Nesterov { 63723657bb1STejun Heo /* 63823657bb1STejun Heo * The following wmb is paired with the implied mb in 63923657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 64023657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64123657bb1STejun Heo * owner. 64223657bb1STejun Heo */ 64323657bb1STejun Heo smp_wmb(); 6447c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 645346c09f8SRoman Pen /* 646346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 647346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 648346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 649346c09f8SRoman Pen * reordering can lead to a missed execution on attempt to qeueue 650346c09f8SRoman Pen * the same @work. E.g. consider this case: 651346c09f8SRoman Pen * 652346c09f8SRoman Pen * CPU#0 CPU#1 653346c09f8SRoman Pen * ---------------------------- -------------------------------- 654346c09f8SRoman Pen * 655346c09f8SRoman Pen * 1 STORE event_indicated 656346c09f8SRoman Pen * 2 queue_work_on() { 657346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 658346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 659346c09f8SRoman Pen * 5 set_work_data() # clear bit 660346c09f8SRoman Pen * 6 smp_mb() 661346c09f8SRoman Pen * 7 work->current_func() { 662346c09f8SRoman Pen * 8 LOAD event_indicated 663346c09f8SRoman Pen * } 664346c09f8SRoman Pen * 665346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 666346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 667346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 668346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 669346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 670346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 671346c09f8SRoman Pen * before actual STORE. 672346c09f8SRoman Pen */ 673346c09f8SRoman Pen smp_mb(); 6744d707b9fSOleg Nesterov } 6754d707b9fSOleg Nesterov 6767a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 677365970a1SDavid Howells { 6787c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6797c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6807a22ad75STejun Heo } 6817a22ad75STejun Heo 682112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6837a22ad75STejun Heo { 684e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6857a22ad75STejun Heo 686112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 687e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 688e120153dSTejun Heo else 689e120153dSTejun Heo return NULL; 6907a22ad75STejun Heo } 6917a22ad75STejun Heo 6927c3eed5cSTejun Heo /** 6937c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6947c3eed5cSTejun Heo * @work: the work item of interest 6957c3eed5cSTejun Heo * 69668e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69768e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 69868e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 699fa1b54e6STejun Heo * 700fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 701fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 702fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 703fa1b54e6STejun Heo * returned pool is and stays online. 704d185af30SYacine Belkadi * 705d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7067c3eed5cSTejun Heo */ 7077c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7087a22ad75STejun Heo { 709e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7107c3eed5cSTejun Heo int pool_id; 7117a22ad75STejun Heo 71268e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 713fa1b54e6STejun Heo 714112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 715112202d9STejun Heo return ((struct pool_workqueue *) 7167c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7177a22ad75STejun Heo 7187c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7197c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7207a22ad75STejun Heo return NULL; 7217a22ad75STejun Heo 722fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7237c3eed5cSTejun Heo } 7247c3eed5cSTejun Heo 7257c3eed5cSTejun Heo /** 7267c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7277c3eed5cSTejun Heo * @work: the work item of interest 7287c3eed5cSTejun Heo * 729d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7307c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7317c3eed5cSTejun Heo */ 7327c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7337c3eed5cSTejun Heo { 73454d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7357c3eed5cSTejun Heo 736112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 737112202d9STejun Heo return ((struct pool_workqueue *) 73854d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 73954d5b7d0SLai Jiangshan 74054d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7417c3eed5cSTejun Heo } 7427c3eed5cSTejun Heo 743bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 744bbb68dfaSTejun Heo { 7457c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 746bbb68dfaSTejun Heo 7477c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7487c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 749bbb68dfaSTejun Heo } 750bbb68dfaSTejun Heo 751bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 752bbb68dfaSTejun Heo { 753bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 754bbb68dfaSTejun Heo 755112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 756bbb68dfaSTejun Heo } 757bbb68dfaSTejun Heo 758e22bee78STejun Heo /* 7593270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7603270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 761d565ed63STejun Heo * they're being called with pool->lock held. 762e22bee78STejun Heo */ 763e22bee78STejun Heo 76463d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 765649027d7STejun Heo { 766e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 767649027d7STejun Heo } 768649027d7STejun Heo 769e22bee78STejun Heo /* 770e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 771e22bee78STejun Heo * running workers. 772974271c4STejun Heo * 773974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 774706026c2STejun Heo * function will always return %true for unbound pools as long as the 775974271c4STejun Heo * worklist isn't empty. 776e22bee78STejun Heo */ 77763d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 778e22bee78STejun Heo { 77963d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 780e22bee78STejun Heo } 781e22bee78STejun Heo 782e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78363d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 784e22bee78STejun Heo { 78563d95a91STejun Heo return pool->nr_idle; 786e22bee78STejun Heo } 787e22bee78STejun Heo 788e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 78963d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 790e22bee78STejun Heo { 791e19e397aSTejun Heo return !list_empty(&pool->worklist) && 792e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 793e22bee78STejun Heo } 794e22bee78STejun Heo 795e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79663d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 797e22bee78STejun Heo { 79863d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 799e22bee78STejun Heo } 800e22bee78STejun Heo 801e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80263d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 803e22bee78STejun Heo { 80434a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 80563d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80663d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 807e22bee78STejun Heo 808e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 809e22bee78STejun Heo } 810e22bee78STejun Heo 811e22bee78STejun Heo /* 812e22bee78STejun Heo * Wake up functions. 813e22bee78STejun Heo */ 814e22bee78STejun Heo 8151037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8161037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8177e11629dSTejun Heo { 81863d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8197e11629dSTejun Heo return NULL; 8207e11629dSTejun Heo 82163d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8227e11629dSTejun Heo } 8237e11629dSTejun Heo 8247e11629dSTejun Heo /** 8257e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82663d95a91STejun Heo * @pool: worker pool to wake worker from 8277e11629dSTejun Heo * 82863d95a91STejun Heo * Wake up the first idle worker of @pool. 8297e11629dSTejun Heo * 8307e11629dSTejun Heo * CONTEXT: 831d565ed63STejun Heo * spin_lock_irq(pool->lock). 8327e11629dSTejun Heo */ 83363d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8347e11629dSTejun Heo { 8351037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8367e11629dSTejun Heo 8377e11629dSTejun Heo if (likely(worker)) 8387e11629dSTejun Heo wake_up_process(worker->task); 8397e11629dSTejun Heo } 8407e11629dSTejun Heo 8414690c4abSTejun Heo /** 842e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 843e22bee78STejun Heo * @task: task waking up 844e22bee78STejun Heo * @cpu: CPU @task is waking up to 845e22bee78STejun Heo * 846e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 847e22bee78STejun Heo * being awoken. 848e22bee78STejun Heo * 849e22bee78STejun Heo * CONTEXT: 850e22bee78STejun Heo * spin_lock_irq(rq->lock) 851e22bee78STejun Heo */ 852d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 853e22bee78STejun Heo { 854e22bee78STejun Heo struct worker *worker = kthread_data(task); 855e22bee78STejun Heo 85636576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 857ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 858e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 859e22bee78STejun Heo } 86036576000SJoonsoo Kim } 861e22bee78STejun Heo 862e22bee78STejun Heo /** 863e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 864e22bee78STejun Heo * @task: task going to sleep 865e22bee78STejun Heo * 866e22bee78STejun Heo * This function is called during schedule() when a busy worker is 867e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 868e22bee78STejun Heo * returning pointer to its task. 869e22bee78STejun Heo * 870e22bee78STejun Heo * CONTEXT: 871e22bee78STejun Heo * spin_lock_irq(rq->lock) 872e22bee78STejun Heo * 873d185af30SYacine Belkadi * Return: 874e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 875e22bee78STejun Heo */ 8769b7f6597SAlexander Gordeev struct task_struct *wq_worker_sleeping(struct task_struct *task) 877e22bee78STejun Heo { 878e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 879111c225aSTejun Heo struct worker_pool *pool; 880e22bee78STejun Heo 881111c225aSTejun Heo /* 882111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 883111c225aSTejun Heo * workers, also reach here, let's not access anything before 884111c225aSTejun Heo * checking NOT_RUNNING. 885111c225aSTejun Heo */ 8862d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 887e22bee78STejun Heo return NULL; 888e22bee78STejun Heo 889111c225aSTejun Heo pool = worker->pool; 890111c225aSTejun Heo 891e22bee78STejun Heo /* this can only happen on the local cpu */ 8929b7f6597SAlexander Gordeev if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) 8936183c009STejun Heo return NULL; 894e22bee78STejun Heo 895e22bee78STejun Heo /* 896e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 897e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 898e22bee78STejun Heo * Please read comment there. 899e22bee78STejun Heo * 900628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 901628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 902628c78e7STejun Heo * disabled, which in turn means that none else could be 903d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 904628c78e7STejun Heo * lock is safe. 905e22bee78STejun Heo */ 906e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 907e19e397aSTejun Heo !list_empty(&pool->worklist)) 9081037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 909e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 910e22bee78STejun Heo } 911e22bee78STejun Heo 912e22bee78STejun Heo /** 913e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 914cb444766STejun Heo * @worker: self 915d302f017STejun Heo * @flags: flags to set 916d302f017STejun Heo * 917228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 918d302f017STejun Heo * 919cb444766STejun Heo * CONTEXT: 920d565ed63STejun Heo * spin_lock_irq(pool->lock) 921d302f017STejun Heo */ 922228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 923d302f017STejun Heo { 924bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 925e22bee78STejun Heo 926cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 927cb444766STejun Heo 928228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 929e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 930e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 931e19e397aSTejun Heo atomic_dec(&pool->nr_running); 932e22bee78STejun Heo } 933e22bee78STejun Heo 934d302f017STejun Heo worker->flags |= flags; 935d302f017STejun Heo } 936d302f017STejun Heo 937d302f017STejun Heo /** 938e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 939cb444766STejun Heo * @worker: self 940d302f017STejun Heo * @flags: flags to clear 941d302f017STejun Heo * 942e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 943d302f017STejun Heo * 944cb444766STejun Heo * CONTEXT: 945d565ed63STejun Heo * spin_lock_irq(pool->lock) 946d302f017STejun Heo */ 947d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 948d302f017STejun Heo { 94963d95a91STejun Heo struct worker_pool *pool = worker->pool; 950e22bee78STejun Heo unsigned int oflags = worker->flags; 951e22bee78STejun Heo 952cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 953cb444766STejun Heo 954d302f017STejun Heo worker->flags &= ~flags; 955e22bee78STejun Heo 95642c025f3STejun Heo /* 95742c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 95842c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 95942c025f3STejun Heo * of multiple flags, not a single flag. 96042c025f3STejun Heo */ 961e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 962e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 963e19e397aSTejun Heo atomic_inc(&pool->nr_running); 964d302f017STejun Heo } 965d302f017STejun Heo 966d302f017STejun Heo /** 9678cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 968c9e7cf27STejun Heo * @pool: pool of interest 9698cca0eeaSTejun Heo * @work: work to find worker for 9708cca0eeaSTejun Heo * 971c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 972c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 973a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 974a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 975a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 976a2c1c57bSTejun Heo * being executed. 977a2c1c57bSTejun Heo * 978a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 979a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 980a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 981a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 982a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 983a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 984a2c1c57bSTejun Heo * 985c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 986c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 987c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 988c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 989c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 990c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9918cca0eeaSTejun Heo * 9928cca0eeaSTejun Heo * CONTEXT: 993d565ed63STejun Heo * spin_lock_irq(pool->lock). 9948cca0eeaSTejun Heo * 995d185af30SYacine Belkadi * Return: 996d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9978cca0eeaSTejun Heo * otherwise. 9988cca0eeaSTejun Heo */ 999c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10008cca0eeaSTejun Heo struct work_struct *work) 10018cca0eeaSTejun Heo { 100242f8570fSSasha Levin struct worker *worker; 100342f8570fSSasha Levin 1004b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1005a2c1c57bSTejun Heo (unsigned long)work) 1006a2c1c57bSTejun Heo if (worker->current_work == work && 1007a2c1c57bSTejun Heo worker->current_func == work->func) 100842f8570fSSasha Levin return worker; 100942f8570fSSasha Levin 101042f8570fSSasha Levin return NULL; 10118cca0eeaSTejun Heo } 10128cca0eeaSTejun Heo 10138cca0eeaSTejun Heo /** 1014bf4ede01STejun Heo * move_linked_works - move linked works to a list 1015bf4ede01STejun Heo * @work: start of series of works to be scheduled 1016bf4ede01STejun Heo * @head: target list to append @work to 1017402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1018bf4ede01STejun Heo * 1019bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1020bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1021bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1022bf4ede01STejun Heo * 1023bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1024bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1025bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1026bf4ede01STejun Heo * 1027bf4ede01STejun Heo * CONTEXT: 1028d565ed63STejun Heo * spin_lock_irq(pool->lock). 1029bf4ede01STejun Heo */ 1030bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1031bf4ede01STejun Heo struct work_struct **nextp) 1032bf4ede01STejun Heo { 1033bf4ede01STejun Heo struct work_struct *n; 1034bf4ede01STejun Heo 1035bf4ede01STejun Heo /* 1036bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1037bf4ede01STejun Heo * use NULL for list head. 1038bf4ede01STejun Heo */ 1039bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1040bf4ede01STejun Heo list_move_tail(&work->entry, head); 1041bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1042bf4ede01STejun Heo break; 1043bf4ede01STejun Heo } 1044bf4ede01STejun Heo 1045bf4ede01STejun Heo /* 1046bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1047bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1048bf4ede01STejun Heo * needs to be updated. 1049bf4ede01STejun Heo */ 1050bf4ede01STejun Heo if (nextp) 1051bf4ede01STejun Heo *nextp = n; 1052bf4ede01STejun Heo } 1053bf4ede01STejun Heo 10548864b4e5STejun Heo /** 10558864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10568864b4e5STejun Heo * @pwq: pool_workqueue to get 10578864b4e5STejun Heo * 10588864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10598864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10608864b4e5STejun Heo */ 10618864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10628864b4e5STejun Heo { 10638864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10648864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10658864b4e5STejun Heo pwq->refcnt++; 10668864b4e5STejun Heo } 10678864b4e5STejun Heo 10688864b4e5STejun Heo /** 10698864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10708864b4e5STejun Heo * @pwq: pool_workqueue to put 10718864b4e5STejun Heo * 10728864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10738864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10748864b4e5STejun Heo */ 10758864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10768864b4e5STejun Heo { 10778864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10788864b4e5STejun Heo if (likely(--pwq->refcnt)) 10798864b4e5STejun Heo return; 10808864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10818864b4e5STejun Heo return; 10828864b4e5STejun Heo /* 10838864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10848864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10858864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10868864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10878864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10888864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10898864b4e5STejun Heo */ 10908864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10918864b4e5STejun Heo } 10928864b4e5STejun Heo 1093dce90d47STejun Heo /** 1094dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1095dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1096dce90d47STejun Heo * 1097dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1098dce90d47STejun Heo */ 1099dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1100dce90d47STejun Heo { 1101dce90d47STejun Heo if (pwq) { 1102dce90d47STejun Heo /* 1103dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1104dce90d47STejun Heo * following lock operations are safe. 1105dce90d47STejun Heo */ 1106dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1107dce90d47STejun Heo put_pwq(pwq); 1108dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1109dce90d47STejun Heo } 1110dce90d47STejun Heo } 1111dce90d47STejun Heo 1112112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1113bf4ede01STejun Heo { 1114112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1115bf4ede01STejun Heo 1116bf4ede01STejun Heo trace_workqueue_activate_work(work); 111782607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 111882607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1119112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1120bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1121112202d9STejun Heo pwq->nr_active++; 1122bf4ede01STejun Heo } 1123bf4ede01STejun Heo 1124112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11253aa62497SLai Jiangshan { 1126112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11273aa62497SLai Jiangshan struct work_struct, entry); 11283aa62497SLai Jiangshan 1129112202d9STejun Heo pwq_activate_delayed_work(work); 11303aa62497SLai Jiangshan } 11313aa62497SLai Jiangshan 1132bf4ede01STejun Heo /** 1133112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1134112202d9STejun Heo * @pwq: pwq of interest 1135bf4ede01STejun Heo * @color: color of work which left the queue 1136bf4ede01STejun Heo * 1137bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1138112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1139bf4ede01STejun Heo * 1140bf4ede01STejun Heo * CONTEXT: 1141d565ed63STejun Heo * spin_lock_irq(pool->lock). 1142bf4ede01STejun Heo */ 1143112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1144bf4ede01STejun Heo { 11458864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1146bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11478864b4e5STejun Heo goto out_put; 1148bf4ede01STejun Heo 1149112202d9STejun Heo pwq->nr_in_flight[color]--; 1150bf4ede01STejun Heo 1151112202d9STejun Heo pwq->nr_active--; 1152112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1153bf4ede01STejun Heo /* one down, submit a delayed one */ 1154112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1155112202d9STejun Heo pwq_activate_first_delayed(pwq); 1156bf4ede01STejun Heo } 1157bf4ede01STejun Heo 1158bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1159112202d9STejun Heo if (likely(pwq->flush_color != color)) 11608864b4e5STejun Heo goto out_put; 1161bf4ede01STejun Heo 1162bf4ede01STejun Heo /* are there still in-flight works? */ 1163112202d9STejun Heo if (pwq->nr_in_flight[color]) 11648864b4e5STejun Heo goto out_put; 1165bf4ede01STejun Heo 1166112202d9STejun Heo /* this pwq is done, clear flush_color */ 1167112202d9STejun Heo pwq->flush_color = -1; 1168bf4ede01STejun Heo 1169bf4ede01STejun Heo /* 1170112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1171bf4ede01STejun Heo * will handle the rest. 1172bf4ede01STejun Heo */ 1173112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1174112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11758864b4e5STejun Heo out_put: 11768864b4e5STejun Heo put_pwq(pwq); 1177bf4ede01STejun Heo } 1178bf4ede01STejun Heo 117936e227d2STejun Heo /** 1180bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 118136e227d2STejun Heo * @work: work item to steal 118236e227d2STejun Heo * @is_dwork: @work is a delayed_work 1183bbb68dfaSTejun Heo * @flags: place to store irq state 118436e227d2STejun Heo * 118536e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1186d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 118736e227d2STejun Heo * 1188d185af30SYacine Belkadi * Return: 118936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 119036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 119136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1192bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1193bbb68dfaSTejun Heo * for arbitrarily long 119436e227d2STejun Heo * 1195d185af30SYacine Belkadi * Note: 1196bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1197e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1198e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1199e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1200bbb68dfaSTejun Heo * 1201bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1202bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1203bbb68dfaSTejun Heo * 1204e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1205bf4ede01STejun Heo */ 1206bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1207bbb68dfaSTejun Heo unsigned long *flags) 1208bf4ede01STejun Heo { 1209d565ed63STejun Heo struct worker_pool *pool; 1210112202d9STejun Heo struct pool_workqueue *pwq; 1211bf4ede01STejun Heo 1212bbb68dfaSTejun Heo local_irq_save(*flags); 1213bbb68dfaSTejun Heo 121436e227d2STejun Heo /* try to steal the timer if it exists */ 121536e227d2STejun Heo if (is_dwork) { 121636e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 121736e227d2STejun Heo 1218e0aecdd8STejun Heo /* 1219e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1220e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1221e0aecdd8STejun Heo * running on the local CPU. 1222e0aecdd8STejun Heo */ 122336e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 122436e227d2STejun Heo return 1; 122536e227d2STejun Heo } 122636e227d2STejun Heo 122736e227d2STejun Heo /* try to claim PENDING the normal way */ 1228bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1229bf4ede01STejun Heo return 0; 1230bf4ede01STejun Heo 1231bf4ede01STejun Heo /* 1232bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1233bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1234bf4ede01STejun Heo */ 1235d565ed63STejun Heo pool = get_work_pool(work); 1236d565ed63STejun Heo if (!pool) 1237bbb68dfaSTejun Heo goto fail; 1238bf4ede01STejun Heo 1239d565ed63STejun Heo spin_lock(&pool->lock); 1240bf4ede01STejun Heo /* 1241112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1242112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1243112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1244112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1245112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12460b3dae68SLai Jiangshan * item is currently queued on that pool. 1247bf4ede01STejun Heo */ 1248112202d9STejun Heo pwq = get_work_pwq(work); 1249112202d9STejun Heo if (pwq && pwq->pool == pool) { 1250bf4ede01STejun Heo debug_work_deactivate(work); 12513aa62497SLai Jiangshan 12523aa62497SLai Jiangshan /* 125316062836STejun Heo * A delayed work item cannot be grabbed directly because 125416062836STejun Heo * it might have linked NO_COLOR work items which, if left 1255112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 125616062836STejun Heo * management later on and cause stall. Make sure the work 125716062836STejun Heo * item is activated before grabbing. 12583aa62497SLai Jiangshan */ 12593aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1260112202d9STejun Heo pwq_activate_delayed_work(work); 12613aa62497SLai Jiangshan 1262bf4ede01STejun Heo list_del_init(&work->entry); 12639c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 126436e227d2STejun Heo 1265112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12664468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12674468a00fSLai Jiangshan 1268d565ed63STejun Heo spin_unlock(&pool->lock); 126936e227d2STejun Heo return 1; 1270bf4ede01STejun Heo } 1271d565ed63STejun Heo spin_unlock(&pool->lock); 1272bbb68dfaSTejun Heo fail: 1273bbb68dfaSTejun Heo local_irq_restore(*flags); 1274bbb68dfaSTejun Heo if (work_is_canceling(work)) 1275bbb68dfaSTejun Heo return -ENOENT; 1276bbb68dfaSTejun Heo cpu_relax(); 127736e227d2STejun Heo return -EAGAIN; 1278bf4ede01STejun Heo } 1279bf4ede01STejun Heo 1280bf4ede01STejun Heo /** 1281706026c2STejun Heo * insert_work - insert a work into a pool 1282112202d9STejun Heo * @pwq: pwq @work belongs to 12834690c4abSTejun Heo * @work: work to insert 12844690c4abSTejun Heo * @head: insertion point 12854690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12864690c4abSTejun Heo * 1287112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1288706026c2STejun Heo * work_struct flags. 12894690c4abSTejun Heo * 12904690c4abSTejun Heo * CONTEXT: 1291d565ed63STejun Heo * spin_lock_irq(pool->lock). 1292365970a1SDavid Howells */ 1293112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1294112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1295b89deed3SOleg Nesterov { 1296112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1297e1d8aa9fSFrederic Weisbecker 12984690c4abSTejun Heo /* we own @work, set data and link */ 1299112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13001a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13018864b4e5STejun Heo get_pwq(pwq); 1302e22bee78STejun Heo 1303e22bee78STejun Heo /* 1304c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1305c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1306c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1307e22bee78STejun Heo */ 1308e22bee78STejun Heo smp_mb(); 1309e22bee78STejun Heo 131063d95a91STejun Heo if (__need_more_worker(pool)) 131163d95a91STejun Heo wake_up_worker(pool); 1312b89deed3SOleg Nesterov } 1313b89deed3SOleg Nesterov 1314c8efcc25STejun Heo /* 1315c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13168d03ecfeSTejun Heo * same workqueue. 1317c8efcc25STejun Heo */ 1318c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1319c8efcc25STejun Heo { 1320c8efcc25STejun Heo struct worker *worker; 1321c8efcc25STejun Heo 13228d03ecfeSTejun Heo worker = current_wq_worker(); 1323c8efcc25STejun Heo /* 13248d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13258d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1326c8efcc25STejun Heo */ 1327112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1328c8efcc25STejun Heo } 1329c8efcc25STejun Heo 1330ef557180SMike Galbraith /* 1331ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1332ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1333ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1334ef557180SMike Galbraith */ 1335ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1336ef557180SMike Galbraith { 1337f303fccbSTejun Heo static bool printed_dbg_warning; 1338ef557180SMike Galbraith int new_cpu; 1339ef557180SMike Galbraith 1340f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1341ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1342ef557180SMike Galbraith return cpu; 1343f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1344f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1345f303fccbSTejun Heo printed_dbg_warning = true; 1346f303fccbSTejun Heo } 1347f303fccbSTejun Heo 1348ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1349ef557180SMike Galbraith return cpu; 1350ef557180SMike Galbraith 1351ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1352ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1353ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1354ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1355ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1356ef557180SMike Galbraith return cpu; 1357ef557180SMike Galbraith } 1358ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1359ef557180SMike Galbraith 1360ef557180SMike Galbraith return new_cpu; 1361ef557180SMike Galbraith } 1362ef557180SMike Galbraith 1363d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13641da177e4SLinus Torvalds struct work_struct *work) 13651da177e4SLinus Torvalds { 1366112202d9STejun Heo struct pool_workqueue *pwq; 1367c9178087STejun Heo struct worker_pool *last_pool; 13681e19ffc6STejun Heo struct list_head *worklist; 13698a2e8e5dSTejun Heo unsigned int work_flags; 1370b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13718930cabaSTejun Heo 13728930cabaSTejun Heo /* 13738930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13748930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13758930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13768930cabaSTejun Heo * happen with IRQ disabled. 13778930cabaSTejun Heo */ 13788930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13791da177e4SLinus Torvalds 1380dc186ad7SThomas Gleixner debug_work_activate(work); 13811e19ffc6STejun Heo 13829ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1383618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1384c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1385e41e704bSTejun Heo return; 13869e8cd2f5STejun Heo retry: 1387df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1388ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1389df2d5ae4STejun Heo 1390df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1391df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1392c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1393df2d5ae4STejun Heo else 1394df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1395f3421797STejun Heo 139618aa9effSTejun Heo /* 1397c9178087STejun Heo * If @work was previously on a different pool, it might still be 1398c9178087STejun Heo * running there, in which case the work needs to be queued on that 1399c9178087STejun Heo * pool to guarantee non-reentrancy. 140018aa9effSTejun Heo */ 1401c9e7cf27STejun Heo last_pool = get_work_pool(work); 1402112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 140318aa9effSTejun Heo struct worker *worker; 140418aa9effSTejun Heo 1405d565ed63STejun Heo spin_lock(&last_pool->lock); 140618aa9effSTejun Heo 1407c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 140818aa9effSTejun Heo 1409112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1410c9178087STejun Heo pwq = worker->current_pwq; 14118594fadeSLai Jiangshan } else { 141218aa9effSTejun Heo /* meh... not running there, queue here */ 1413d565ed63STejun Heo spin_unlock(&last_pool->lock); 1414112202d9STejun Heo spin_lock(&pwq->pool->lock); 141518aa9effSTejun Heo } 14168930cabaSTejun Heo } else { 1417112202d9STejun Heo spin_lock(&pwq->pool->lock); 14188930cabaSTejun Heo } 1419502ca9d8STejun Heo 14209e8cd2f5STejun Heo /* 14219e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14229e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14239e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1424df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1425df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14269e8cd2f5STejun Heo * make forward-progress. 14279e8cd2f5STejun Heo */ 14289e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14299e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 14309e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 14319e8cd2f5STejun Heo cpu_relax(); 14329e8cd2f5STejun Heo goto retry; 14339e8cd2f5STejun Heo } 14349e8cd2f5STejun Heo /* oops */ 14359e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14369e8cd2f5STejun Heo wq->name, cpu); 14379e8cd2f5STejun Heo } 14389e8cd2f5STejun Heo 1439112202d9STejun Heo /* pwq determined, queue */ 1440112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1441502ca9d8STejun Heo 1442f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1443112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1444f5b2552bSDan Carpenter return; 1445f5b2552bSDan Carpenter } 14461e19ffc6STejun Heo 1447112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1448112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14491e19ffc6STejun Heo 1450112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1451cdadf009STejun Heo trace_workqueue_activate_work(work); 1452112202d9STejun Heo pwq->nr_active++; 1453112202d9STejun Heo worklist = &pwq->pool->worklist; 145482607adcSTejun Heo if (list_empty(worklist)) 145582607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14568a2e8e5dSTejun Heo } else { 14578a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1458112202d9STejun Heo worklist = &pwq->delayed_works; 14598a2e8e5dSTejun Heo } 14601e19ffc6STejun Heo 1461112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14621e19ffc6STejun Heo 1463112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14641da177e4SLinus Torvalds } 14651da177e4SLinus Torvalds 14660fcb78c2SRolf Eike Beer /** 1467c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1468c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1469c1a220e7SZhang Rui * @wq: workqueue to use 1470c1a220e7SZhang Rui * @work: work to queue 1471c1a220e7SZhang Rui * 1472c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1473c1a220e7SZhang Rui * can't go away. 1474d185af30SYacine Belkadi * 1475d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1476c1a220e7SZhang Rui */ 1477d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1478d4283e93STejun Heo struct work_struct *work) 1479c1a220e7SZhang Rui { 1480d4283e93STejun Heo bool ret = false; 14818930cabaSTejun Heo unsigned long flags; 14828930cabaSTejun Heo 14838930cabaSTejun Heo local_irq_save(flags); 1484c1a220e7SZhang Rui 148522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14864690c4abSTejun Heo __queue_work(cpu, wq, work); 1487d4283e93STejun Heo ret = true; 1488c1a220e7SZhang Rui } 14898930cabaSTejun Heo 14908930cabaSTejun Heo local_irq_restore(flags); 1491c1a220e7SZhang Rui return ret; 1492c1a220e7SZhang Rui } 1493ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1494c1a220e7SZhang Rui 1495d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14961da177e4SLinus Torvalds { 149752bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14981da177e4SLinus Torvalds 1499e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 150060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 15011da177e4SLinus Torvalds } 15021438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 15031da177e4SLinus Torvalds 15047beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 150552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15061da177e4SLinus Torvalds { 15077beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 15087beb2edfSTejun Heo struct work_struct *work = &dwork->work; 15091da177e4SLinus Torvalds 15107beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 15117beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1512fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1513fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 15147beb2edfSTejun Heo 15158852aac2STejun Heo /* 15168852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 15178852aac2STejun Heo * both optimization and correctness. The earliest @timer can 15188852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 15198852aac2STejun Heo * on that there's no such delay when @delay is 0. 15208852aac2STejun Heo */ 15218852aac2STejun Heo if (!delay) { 15228852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 15238852aac2STejun Heo return; 15248852aac2STejun Heo } 15258852aac2STejun Heo 15267beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 15277beb2edfSTejun Heo 152860c057bcSLai Jiangshan dwork->wq = wq; 15291265057fSTejun Heo dwork->cpu = cpu; 15307beb2edfSTejun Heo timer->expires = jiffies + delay; 15317beb2edfSTejun Heo 1532041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 15337beb2edfSTejun Heo add_timer_on(timer, cpu); 1534041bd12eSTejun Heo else 1535041bd12eSTejun Heo add_timer(timer); 15367beb2edfSTejun Heo } 15371da177e4SLinus Torvalds 15380fcb78c2SRolf Eike Beer /** 15390fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 15400fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 15410fcb78c2SRolf Eike Beer * @wq: workqueue to use 1542af9997e4SRandy Dunlap * @dwork: work to queue 15430fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 15440fcb78c2SRolf Eike Beer * 1545d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1546715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1547715f1300STejun Heo * execution. 15480fcb78c2SRolf Eike Beer */ 1549d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 155052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15517a6bc1cdSVenkatesh Pallipadi { 155252bad64dSDavid Howells struct work_struct *work = &dwork->work; 1553d4283e93STejun Heo bool ret = false; 15548930cabaSTejun Heo unsigned long flags; 15558930cabaSTejun Heo 15568930cabaSTejun Heo /* read the comment in __queue_work() */ 15578930cabaSTejun Heo local_irq_save(flags); 15587a6bc1cdSVenkatesh Pallipadi 155922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15607beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1561d4283e93STejun Heo ret = true; 15627a6bc1cdSVenkatesh Pallipadi } 15638930cabaSTejun Heo 15648930cabaSTejun Heo local_irq_restore(flags); 15657a6bc1cdSVenkatesh Pallipadi return ret; 15667a6bc1cdSVenkatesh Pallipadi } 1567ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15681da177e4SLinus Torvalds 1569c8e55f36STejun Heo /** 15708376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15718376fe22STejun Heo * @cpu: CPU number to execute work on 15728376fe22STejun Heo * @wq: workqueue to use 15738376fe22STejun Heo * @dwork: work to queue 15748376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15758376fe22STejun Heo * 15768376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15778376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15788376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15798376fe22STejun Heo * current state. 15808376fe22STejun Heo * 1581d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15828376fe22STejun Heo * pending and its timer was modified. 15838376fe22STejun Heo * 1584e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15858376fe22STejun Heo * See try_to_grab_pending() for details. 15868376fe22STejun Heo */ 15878376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15888376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15898376fe22STejun Heo { 15908376fe22STejun Heo unsigned long flags; 15918376fe22STejun Heo int ret; 15928376fe22STejun Heo 15938376fe22STejun Heo do { 15948376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15958376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15968376fe22STejun Heo 15978376fe22STejun Heo if (likely(ret >= 0)) { 15988376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15998376fe22STejun Heo local_irq_restore(flags); 16008376fe22STejun Heo } 16018376fe22STejun Heo 16028376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 16038376fe22STejun Heo return ret; 16048376fe22STejun Heo } 16058376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 16068376fe22STejun Heo 16078376fe22STejun Heo /** 1608c8e55f36STejun Heo * worker_enter_idle - enter idle state 1609c8e55f36STejun Heo * @worker: worker which is entering idle state 1610c8e55f36STejun Heo * 1611c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1612c8e55f36STejun Heo * necessary. 1613c8e55f36STejun Heo * 1614c8e55f36STejun Heo * LOCKING: 1615d565ed63STejun Heo * spin_lock_irq(pool->lock). 1616c8e55f36STejun Heo */ 1617c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 16181da177e4SLinus Torvalds { 1619bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1620c8e55f36STejun Heo 16216183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 16226183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 16236183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 16246183c009STejun Heo return; 1625c8e55f36STejun Heo 1626051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1627cb444766STejun Heo worker->flags |= WORKER_IDLE; 1628bd7bdd43STejun Heo pool->nr_idle++; 1629e22bee78STejun Heo worker->last_active = jiffies; 1630c8e55f36STejun Heo 1631c8e55f36STejun Heo /* idle_list is LIFO */ 1632bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1633db7bccf4STejun Heo 163463d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1635628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1636cb444766STejun Heo 1637544ecf31STejun Heo /* 1638706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1639d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1640628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1641628c78e7STejun Heo * unbind is not in progress. 1642544ecf31STejun Heo */ 164324647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1644bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1645e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1646c8e55f36STejun Heo } 1647c8e55f36STejun Heo 1648c8e55f36STejun Heo /** 1649c8e55f36STejun Heo * worker_leave_idle - leave idle state 1650c8e55f36STejun Heo * @worker: worker which is leaving idle state 1651c8e55f36STejun Heo * 1652c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1653c8e55f36STejun Heo * 1654c8e55f36STejun Heo * LOCKING: 1655d565ed63STejun Heo * spin_lock_irq(pool->lock). 1656c8e55f36STejun Heo */ 1657c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1658c8e55f36STejun Heo { 1659bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1660c8e55f36STejun Heo 16616183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16626183c009STejun Heo return; 1663d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1664bd7bdd43STejun Heo pool->nr_idle--; 1665c8e55f36STejun Heo list_del_init(&worker->entry); 1666c8e55f36STejun Heo } 1667c8e55f36STejun Heo 1668f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1669c34056a3STejun Heo { 1670c34056a3STejun Heo struct worker *worker; 1671c34056a3STejun Heo 1672f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1673c8e55f36STejun Heo if (worker) { 1674c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1675affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1676da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1677e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1678e22bee78STejun Heo worker->flags = WORKER_PREP; 1679c8e55f36STejun Heo } 1680c34056a3STejun Heo return worker; 1681c34056a3STejun Heo } 1682c34056a3STejun Heo 1683c34056a3STejun Heo /** 16844736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16854736cbf7SLai Jiangshan * @worker: worker to be attached 16864736cbf7SLai Jiangshan * @pool: the target pool 16874736cbf7SLai Jiangshan * 16884736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16894736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16904736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16914736cbf7SLai Jiangshan */ 16924736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16934736cbf7SLai Jiangshan struct worker_pool *pool) 16944736cbf7SLai Jiangshan { 16954736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16964736cbf7SLai Jiangshan 16974736cbf7SLai Jiangshan /* 16984736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16994736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 17004736cbf7SLai Jiangshan */ 17014736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17024736cbf7SLai Jiangshan 17034736cbf7SLai Jiangshan /* 17044736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 17054736cbf7SLai Jiangshan * stable across this function. See the comments above the 17064736cbf7SLai Jiangshan * flag definition for details. 17074736cbf7SLai Jiangshan */ 17084736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 17094736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 17104736cbf7SLai Jiangshan 17114736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 17124736cbf7SLai Jiangshan 17134736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 17144736cbf7SLai Jiangshan } 17154736cbf7SLai Jiangshan 17164736cbf7SLai Jiangshan /** 171760f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 171860f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 171960f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 172060f5a4bcSLai Jiangshan * 17214736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 17224736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 17234736cbf7SLai Jiangshan * other reference to the pool. 172460f5a4bcSLai Jiangshan */ 172560f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 172660f5a4bcSLai Jiangshan struct worker_pool *pool) 172760f5a4bcSLai Jiangshan { 172860f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 172960f5a4bcSLai Jiangshan 173092f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1731da028469SLai Jiangshan list_del(&worker->node); 1732da028469SLai Jiangshan if (list_empty(&pool->workers)) 173360f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 173492f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 173560f5a4bcSLai Jiangshan 1736b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1737b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1738b62c0751SLai Jiangshan 173960f5a4bcSLai Jiangshan if (detach_completion) 174060f5a4bcSLai Jiangshan complete(detach_completion); 174160f5a4bcSLai Jiangshan } 174260f5a4bcSLai Jiangshan 174360f5a4bcSLai Jiangshan /** 1744c34056a3STejun Heo * create_worker - create a new workqueue worker 174563d95a91STejun Heo * @pool: pool the new worker will belong to 1746c34056a3STejun Heo * 1747051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1748c34056a3STejun Heo * 1749c34056a3STejun Heo * CONTEXT: 1750c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1751c34056a3STejun Heo * 1752d185af30SYacine Belkadi * Return: 1753c34056a3STejun Heo * Pointer to the newly created worker. 1754c34056a3STejun Heo */ 1755bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1756c34056a3STejun Heo { 1757c34056a3STejun Heo struct worker *worker = NULL; 1758f3421797STejun Heo int id = -1; 1759e3c916a4STejun Heo char id_buf[16]; 1760c34056a3STejun Heo 17617cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17627cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1763822d8405STejun Heo if (id < 0) 1764c34056a3STejun Heo goto fail; 1765c34056a3STejun Heo 1766f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1767c34056a3STejun Heo if (!worker) 1768c34056a3STejun Heo goto fail; 1769c34056a3STejun Heo 1770bd7bdd43STejun Heo worker->pool = pool; 1771c34056a3STejun Heo worker->id = id; 1772c34056a3STejun Heo 177329c91e99STejun Heo if (pool->cpu >= 0) 1774e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1775e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1776f3421797STejun Heo else 1777e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1778e3c916a4STejun Heo 1779f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1780e3c916a4STejun Heo "kworker/%s", id_buf); 1781c34056a3STejun Heo if (IS_ERR(worker->task)) 1782c34056a3STejun Heo goto fail; 1783c34056a3STejun Heo 178491151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 178525834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 178691151228SOleg Nesterov 1787da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17884736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1789822d8405STejun Heo 1790051e1850SLai Jiangshan /* start the newly created worker */ 1791051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1792051e1850SLai Jiangshan worker->pool->nr_workers++; 1793051e1850SLai Jiangshan worker_enter_idle(worker); 1794051e1850SLai Jiangshan wake_up_process(worker->task); 1795051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1796051e1850SLai Jiangshan 1797c34056a3STejun Heo return worker; 1798822d8405STejun Heo 1799c34056a3STejun Heo fail: 18009625ab17SLai Jiangshan if (id >= 0) 18017cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1802c34056a3STejun Heo kfree(worker); 1803c34056a3STejun Heo return NULL; 1804c34056a3STejun Heo } 1805c34056a3STejun Heo 1806c34056a3STejun Heo /** 1807c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1808c34056a3STejun Heo * @worker: worker to be destroyed 1809c34056a3STejun Heo * 181073eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 181173eb7fe7SLai Jiangshan * be idle. 1812c8e55f36STejun Heo * 1813c8e55f36STejun Heo * CONTEXT: 181460f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1815c34056a3STejun Heo */ 1816c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1817c34056a3STejun Heo { 1818bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1819c34056a3STejun Heo 1820cd549687STejun Heo lockdep_assert_held(&pool->lock); 1821cd549687STejun Heo 1822c34056a3STejun Heo /* sanity check frenzy */ 18236183c009STejun Heo if (WARN_ON(worker->current_work) || 182473eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 182573eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 18266183c009STejun Heo return; 1827c34056a3STejun Heo 1828bd7bdd43STejun Heo pool->nr_workers--; 1829bd7bdd43STejun Heo pool->nr_idle--; 1830c8e55f36STejun Heo 1831c8e55f36STejun Heo list_del_init(&worker->entry); 1832cb444766STejun Heo worker->flags |= WORKER_DIE; 183360f5a4bcSLai Jiangshan wake_up_process(worker->task); 1834c34056a3STejun Heo } 1835c34056a3STejun Heo 183663d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1837e22bee78STejun Heo { 183863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1839e22bee78STejun Heo 1840d565ed63STejun Heo spin_lock_irq(&pool->lock); 1841e22bee78STejun Heo 18423347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1843e22bee78STejun Heo struct worker *worker; 1844e22bee78STejun Heo unsigned long expires; 1845e22bee78STejun Heo 1846e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 184763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1848e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1849e22bee78STejun Heo 18503347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 185163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18523347fc9fSLai Jiangshan break; 1853e22bee78STejun Heo } 18543347fc9fSLai Jiangshan 18553347fc9fSLai Jiangshan destroy_worker(worker); 1856e22bee78STejun Heo } 1857e22bee78STejun Heo 1858d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1859e22bee78STejun Heo } 1860e22bee78STejun Heo 1861493a1724STejun Heo static void send_mayday(struct work_struct *work) 1862e22bee78STejun Heo { 1863112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1864112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1865493a1724STejun Heo 18662e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1867e22bee78STejun Heo 1868493008a8STejun Heo if (!wq->rescuer) 1869493a1724STejun Heo return; 1870e22bee78STejun Heo 1871e22bee78STejun Heo /* mayday mayday mayday */ 1872493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 187377668c8bSLai Jiangshan /* 187477668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 187577668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 187677668c8bSLai Jiangshan * rescuer is done with it. 187777668c8bSLai Jiangshan */ 187877668c8bSLai Jiangshan get_pwq(pwq); 1879493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1880e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1881493a1724STejun Heo } 1882e22bee78STejun Heo } 1883e22bee78STejun Heo 1884706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1885e22bee78STejun Heo { 188663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1887e22bee78STejun Heo struct work_struct *work; 1888e22bee78STejun Heo 1889b2d82909STejun Heo spin_lock_irq(&pool->lock); 1890b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1891e22bee78STejun Heo 189263d95a91STejun Heo if (need_to_create_worker(pool)) { 1893e22bee78STejun Heo /* 1894e22bee78STejun Heo * We've been trying to create a new worker but 1895e22bee78STejun Heo * haven't been successful. We might be hitting an 1896e22bee78STejun Heo * allocation deadlock. Send distress signals to 1897e22bee78STejun Heo * rescuers. 1898e22bee78STejun Heo */ 189963d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1900e22bee78STejun Heo send_mayday(work); 1901e22bee78STejun Heo } 1902e22bee78STejun Heo 1903b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1904b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1905e22bee78STejun Heo 190663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1907e22bee78STejun Heo } 1908e22bee78STejun Heo 1909e22bee78STejun Heo /** 1910e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 191163d95a91STejun Heo * @pool: pool to create a new worker for 1912e22bee78STejun Heo * 191363d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1914e22bee78STejun Heo * have at least one idle worker on return from this function. If 1915e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 191663d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1917e22bee78STejun Heo * possible allocation deadlock. 1918e22bee78STejun Heo * 1919c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1920c5aa87bbSTejun Heo * may_start_working() %true. 1921e22bee78STejun Heo * 1922e22bee78STejun Heo * LOCKING: 1923d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1924e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1925e22bee78STejun Heo * manager. 1926e22bee78STejun Heo */ 192729187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1928d565ed63STejun Heo __releases(&pool->lock) 1929d565ed63STejun Heo __acquires(&pool->lock) 1930e22bee78STejun Heo { 1931e22bee78STejun Heo restart: 1932d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19339f9c2364STejun Heo 1934e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 193563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1936e22bee78STejun Heo 1937e22bee78STejun Heo while (true) { 1938051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1939e22bee78STejun Heo break; 1940e22bee78STejun Heo 1941e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19429f9c2364STejun Heo 194363d95a91STejun Heo if (!need_to_create_worker(pool)) 1944e22bee78STejun Heo break; 1945e22bee78STejun Heo } 1946e22bee78STejun Heo 194763d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1948d565ed63STejun Heo spin_lock_irq(&pool->lock); 1949051e1850SLai Jiangshan /* 1950051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1951051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1952051e1850SLai Jiangshan * already become busy. 1953051e1850SLai Jiangshan */ 195463d95a91STejun Heo if (need_to_create_worker(pool)) 1955e22bee78STejun Heo goto restart; 1956e22bee78STejun Heo } 1957e22bee78STejun Heo 1958e22bee78STejun Heo /** 1959e22bee78STejun Heo * manage_workers - manage worker pool 1960e22bee78STejun Heo * @worker: self 1961e22bee78STejun Heo * 1962706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1963e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1964706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1965e22bee78STejun Heo * 1966e22bee78STejun Heo * The caller can safely start processing works on false return. On 1967e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1968e22bee78STejun Heo * and may_start_working() is true. 1969e22bee78STejun Heo * 1970e22bee78STejun Heo * CONTEXT: 1971d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1972e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1973e22bee78STejun Heo * 1974d185af30SYacine Belkadi * Return: 197529187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 197629187a9eSTejun Heo * start processing works, %true if management function was performed and 197729187a9eSTejun Heo * the conditions that the caller verified before calling the function may 197829187a9eSTejun Heo * no longer be true. 1979e22bee78STejun Heo */ 1980e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1981e22bee78STejun Heo { 198263d95a91STejun Heo struct worker_pool *pool = worker->pool; 1983e22bee78STejun Heo 1984bc3a1afcSTejun Heo /* 1985bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1986bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1987bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1988bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1989bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1990bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1991bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1992bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1993bc3a1afcSTejun Heo */ 199434a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 199529187a9eSTejun Heo return false; 19962607d7a6STejun Heo pool->manager = worker; 1997e22bee78STejun Heo 199829187a9eSTejun Heo maybe_create_worker(pool); 1999e22bee78STejun Heo 20002607d7a6STejun Heo pool->manager = NULL; 200134a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 200229187a9eSTejun Heo return true; 2003e22bee78STejun Heo } 2004e22bee78STejun Heo 2005a62428c0STejun Heo /** 2006a62428c0STejun Heo * process_one_work - process single work 2007c34056a3STejun Heo * @worker: self 2008a62428c0STejun Heo * @work: work to process 2009a62428c0STejun Heo * 2010a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2011a62428c0STejun Heo * process a single work including synchronization against and 2012a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2013a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2014a62428c0STejun Heo * call this function to process a work. 2015a62428c0STejun Heo * 2016a62428c0STejun Heo * CONTEXT: 2017d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2018a62428c0STejun Heo */ 2019c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2020d565ed63STejun Heo __releases(&pool->lock) 2021d565ed63STejun Heo __acquires(&pool->lock) 20221da177e4SLinus Torvalds { 2023112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2024bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2025112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 202673f53c4aSTejun Heo int work_color; 20277e11629dSTejun Heo struct worker *collision; 20284e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20294e6045f1SJohannes Berg /* 2030a62428c0STejun Heo * It is permissible to free the struct work_struct from 2031a62428c0STejun Heo * inside the function that is called from it, this we need to 2032a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2033a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2034a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20354e6045f1SJohannes Berg */ 20364d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20374d82a1deSPeter Zijlstra 20384d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20394e6045f1SJohannes Berg #endif 2040807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 204185327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2042ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 204325511a47STejun Heo 20447e11629dSTejun Heo /* 20457e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20467e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20477e11629dSTejun Heo * already processing the work. If so, defer the work to the 20487e11629dSTejun Heo * currently executing one. 20497e11629dSTejun Heo */ 2050c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20517e11629dSTejun Heo if (unlikely(collision)) { 20527e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20537e11629dSTejun Heo return; 20547e11629dSTejun Heo } 20551da177e4SLinus Torvalds 20568930cabaSTejun Heo /* claim and dequeue */ 20571da177e4SLinus Torvalds debug_work_deactivate(work); 2058c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2059c34056a3STejun Heo worker->current_work = work; 2060a2c1c57bSTejun Heo worker->current_func = work->func; 2061112202d9STejun Heo worker->current_pwq = pwq; 206273f53c4aSTejun Heo work_color = get_work_color(work); 20637a22ad75STejun Heo 2064a62428c0STejun Heo list_del_init(&work->entry); 2065a62428c0STejun Heo 2066649027d7STejun Heo /* 2067228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2068228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2069228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2070228f1d00SLai Jiangshan * execution of the pending work items. 2071fb0e7bebSTejun Heo */ 2072fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2073228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2074fb0e7bebSTejun Heo 2075974271c4STejun Heo /* 2076a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2077a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2078a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2079a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2080228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2081974271c4STejun Heo */ 2082a489a03eSLai Jiangshan if (need_more_worker(pool)) 208363d95a91STejun Heo wake_up_worker(pool); 2084974271c4STejun Heo 20858930cabaSTejun Heo /* 20867c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2087d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 208823657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 208923657bb1STejun Heo * disabled. 20908930cabaSTejun Heo */ 20917c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20921da177e4SLinus Torvalds 2093d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2094365970a1SDavid Howells 2095112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20963295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2097e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2098a2c1c57bSTejun Heo worker->current_func(work); 2099e36c886aSArjan van de Ven /* 2100e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2101e36c886aSArjan van de Ven * point will only record its address. 2102e36c886aSArjan van de Ven */ 2103e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21043295f0efSIngo Molnar lock_map_release(&lockdep_map); 2105112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21061da177e4SLinus Torvalds 2107d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2108044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2109044c782cSValentin Ilie " last function: %pf\n", 2110a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2111a2c1c57bSTejun Heo worker->current_func); 2112d5abe669SPeter Zijlstra debug_show_held_locks(current); 2113d5abe669SPeter Zijlstra dump_stack(); 2114d5abe669SPeter Zijlstra } 2115d5abe669SPeter Zijlstra 2116b22ce278STejun Heo /* 2117b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2118b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2119b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2120b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2121789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2122789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2123b22ce278STejun Heo */ 21243e28e377SJoe Lawrence cond_resched_rcu_qs(); 2125b22ce278STejun Heo 2126d565ed63STejun Heo spin_lock_irq(&pool->lock); 2127a62428c0STejun Heo 2128fb0e7bebSTejun Heo /* clear cpu intensive status */ 2129fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2130fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2131fb0e7bebSTejun Heo 2132a62428c0STejun Heo /* we're done with it, release */ 213342f8570fSSasha Levin hash_del(&worker->hentry); 2134c34056a3STejun Heo worker->current_work = NULL; 2135a2c1c57bSTejun Heo worker->current_func = NULL; 2136112202d9STejun Heo worker->current_pwq = NULL; 21373d1cb205STejun Heo worker->desc_valid = false; 2138112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21391da177e4SLinus Torvalds } 21401da177e4SLinus Torvalds 2141affee4b2STejun Heo /** 2142affee4b2STejun Heo * process_scheduled_works - process scheduled works 2143affee4b2STejun Heo * @worker: self 2144affee4b2STejun Heo * 2145affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2146affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2147affee4b2STejun Heo * fetches a work from the top and executes it. 2148affee4b2STejun Heo * 2149affee4b2STejun Heo * CONTEXT: 2150d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2151affee4b2STejun Heo * multiple times. 2152affee4b2STejun Heo */ 2153affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21541da177e4SLinus Torvalds { 2155affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2156affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2157a62428c0STejun Heo struct work_struct, entry); 2158c34056a3STejun Heo process_one_work(worker, work); 2159a62428c0STejun Heo } 21601da177e4SLinus Torvalds } 21611da177e4SLinus Torvalds 21624690c4abSTejun Heo /** 21634690c4abSTejun Heo * worker_thread - the worker thread function 2164c34056a3STejun Heo * @__worker: self 21654690c4abSTejun Heo * 2166c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2167c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2168c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2169c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2170c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2171d185af30SYacine Belkadi * 2172d185af30SYacine Belkadi * Return: 0 21734690c4abSTejun Heo */ 2174c34056a3STejun Heo static int worker_thread(void *__worker) 21751da177e4SLinus Torvalds { 2176c34056a3STejun Heo struct worker *worker = __worker; 2177bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21781da177e4SLinus Torvalds 2179e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2180e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2181c8e55f36STejun Heo woke_up: 2182d565ed63STejun Heo spin_lock_irq(&pool->lock); 2183affee4b2STejun Heo 2184a9ab775bSTejun Heo /* am I supposed to die? */ 2185a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2186d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2187a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2188e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 218960f5a4bcSLai Jiangshan 219060f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21917cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 219260f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 219360f5a4bcSLai Jiangshan kfree(worker); 2194c8e55f36STejun Heo return 0; 2195c8e55f36STejun Heo } 2196c8e55f36STejun Heo 2197c8e55f36STejun Heo worker_leave_idle(worker); 2198db7bccf4STejun Heo recheck: 2199e22bee78STejun Heo /* no more worker necessary? */ 220063d95a91STejun Heo if (!need_more_worker(pool)) 2201e22bee78STejun Heo goto sleep; 2202e22bee78STejun Heo 2203e22bee78STejun Heo /* do we need to manage? */ 220463d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2205e22bee78STejun Heo goto recheck; 2206e22bee78STejun Heo 2207c8e55f36STejun Heo /* 2208c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2209c8e55f36STejun Heo * preparing to process a work or actually processing it. 2210c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2211c8e55f36STejun Heo */ 22126183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2213c8e55f36STejun Heo 2214e22bee78STejun Heo /* 2215a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2216a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2217a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2218a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2219a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2220e22bee78STejun Heo */ 2221a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2222e22bee78STejun Heo 2223e22bee78STejun Heo do { 2224affee4b2STejun Heo struct work_struct *work = 2225bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2226affee4b2STejun Heo struct work_struct, entry); 2227affee4b2STejun Heo 222882607adcSTejun Heo pool->watchdog_ts = jiffies; 222982607adcSTejun Heo 2230c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2231affee4b2STejun Heo /* optimization path, not strictly necessary */ 2232affee4b2STejun Heo process_one_work(worker, work); 2233affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2234affee4b2STejun Heo process_scheduled_works(worker); 2235affee4b2STejun Heo } else { 2236c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2237affee4b2STejun Heo process_scheduled_works(worker); 2238affee4b2STejun Heo } 223963d95a91STejun Heo } while (keep_working(pool)); 2240affee4b2STejun Heo 2241228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2242d313dd85STejun Heo sleep: 2243c8e55f36STejun Heo /* 2244d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2245d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2246d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2247d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2248d565ed63STejun Heo * event. 2249c8e55f36STejun Heo */ 2250c8e55f36STejun Heo worker_enter_idle(worker); 2251c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2252d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22531da177e4SLinus Torvalds schedule(); 2254c8e55f36STejun Heo goto woke_up; 22551da177e4SLinus Torvalds } 22561da177e4SLinus Torvalds 2257e22bee78STejun Heo /** 2258e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2259111c225aSTejun Heo * @__rescuer: self 2260e22bee78STejun Heo * 2261e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2262493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2263e22bee78STejun Heo * 2264706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2265e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2266e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2267e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2268e22bee78STejun Heo * the problem rescuer solves. 2269e22bee78STejun Heo * 2270706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2271706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2272e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2273e22bee78STejun Heo * 2274e22bee78STejun Heo * This should happen rarely. 2275d185af30SYacine Belkadi * 2276d185af30SYacine Belkadi * Return: 0 2277e22bee78STejun Heo */ 2278111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2279e22bee78STejun Heo { 2280111c225aSTejun Heo struct worker *rescuer = __rescuer; 2281111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2282e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22834d595b86SLai Jiangshan bool should_stop; 2284e22bee78STejun Heo 2285e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2286111c225aSTejun Heo 2287111c225aSTejun Heo /* 2288111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2289111c225aSTejun Heo * doesn't participate in concurrency management. 2290111c225aSTejun Heo */ 2291111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2292e22bee78STejun Heo repeat: 2293e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22941da177e4SLinus Torvalds 22954d595b86SLai Jiangshan /* 22964d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22974d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22984d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22994d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 23004d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 23014d595b86SLai Jiangshan * list is always empty on exit. 23024d595b86SLai Jiangshan */ 23034d595b86SLai Jiangshan should_stop = kthread_should_stop(); 23041da177e4SLinus Torvalds 2305493a1724STejun Heo /* see whether any pwq is asking for help */ 23062e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2307493a1724STejun Heo 2308493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2309493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2310493a1724STejun Heo struct pool_workqueue, mayday_node); 2311112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2312e22bee78STejun Heo struct work_struct *work, *n; 231382607adcSTejun Heo bool first = true; 2314e22bee78STejun Heo 2315e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2316493a1724STejun Heo list_del_init(&pwq->mayday_node); 2317493a1724STejun Heo 23182e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2319e22bee78STejun Heo 232051697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 232151697d39SLai Jiangshan 232251697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2323b3104104SLai Jiangshan rescuer->pool = pool; 2324e22bee78STejun Heo 2325e22bee78STejun Heo /* 2326e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2327e22bee78STejun Heo * process'em. 2328e22bee78STejun Heo */ 23290479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 233082607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 233182607adcSTejun Heo if (get_work_pwq(work) == pwq) { 233282607adcSTejun Heo if (first) 233382607adcSTejun Heo pool->watchdog_ts = jiffies; 2334e22bee78STejun Heo move_linked_works(work, scheduled, &n); 233582607adcSTejun Heo } 233682607adcSTejun Heo first = false; 233782607adcSTejun Heo } 2338e22bee78STejun Heo 2339008847f6SNeilBrown if (!list_empty(scheduled)) { 2340e22bee78STejun Heo process_scheduled_works(rescuer); 23417576958aSTejun Heo 23427576958aSTejun Heo /* 2343008847f6SNeilBrown * The above execution of rescued work items could 2344008847f6SNeilBrown * have created more to rescue through 2345008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2346008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2347008847f6SNeilBrown * that such back-to-back work items, which may be 2348008847f6SNeilBrown * being used to relieve memory pressure, don't 2349008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2350008847f6SNeilBrown */ 2351008847f6SNeilBrown if (need_to_create_worker(pool)) { 2352008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2353008847f6SNeilBrown get_pwq(pwq); 2354008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2355008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2356008847f6SNeilBrown } 2357008847f6SNeilBrown } 2358008847f6SNeilBrown 2359008847f6SNeilBrown /* 236077668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 236113b1d625SLai Jiangshan * go away while we're still attached to it. 236277668c8bSLai Jiangshan */ 236377668c8bSLai Jiangshan put_pwq(pwq); 236477668c8bSLai Jiangshan 236577668c8bSLai Jiangshan /* 2366d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 23677576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23687576958aSTejun Heo * and stalling the execution. 23697576958aSTejun Heo */ 2370d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 237163d95a91STejun Heo wake_up_worker(pool); 23727576958aSTejun Heo 2373b3104104SLai Jiangshan rescuer->pool = NULL; 237413b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 237513b1d625SLai Jiangshan 237613b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 237713b1d625SLai Jiangshan 237813b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 23791da177e4SLinus Torvalds } 23801da177e4SLinus Torvalds 23812e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2382493a1724STejun Heo 23834d595b86SLai Jiangshan if (should_stop) { 23844d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23854d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23864d595b86SLai Jiangshan return 0; 23874d595b86SLai Jiangshan } 23884d595b86SLai Jiangshan 2389111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2390111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2391e22bee78STejun Heo schedule(); 2392e22bee78STejun Heo goto repeat; 23931da177e4SLinus Torvalds } 23941da177e4SLinus Torvalds 2395fca839c0STejun Heo /** 2396fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2397fca839c0STejun Heo * @target_wq: workqueue being flushed 2398fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2399fca839c0STejun Heo * 2400fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2401fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2402fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2403fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2404fca839c0STejun Heo * a deadlock. 2405fca839c0STejun Heo */ 2406fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2407fca839c0STejun Heo struct work_struct *target_work) 2408fca839c0STejun Heo { 2409fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2410fca839c0STejun Heo struct worker *worker; 2411fca839c0STejun Heo 2412fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2413fca839c0STejun Heo return; 2414fca839c0STejun Heo 2415fca839c0STejun Heo worker = current_wq_worker(); 2416fca839c0STejun Heo 2417fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2418fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2419fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 242023d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 242123d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2422fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2423fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2424fca839c0STejun Heo target_wq->name, target_func); 2425fca839c0STejun Heo } 2426fca839c0STejun Heo 2427fc2e4d70SOleg Nesterov struct wq_barrier { 2428fc2e4d70SOleg Nesterov struct work_struct work; 2429fc2e4d70SOleg Nesterov struct completion done; 24302607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2431fc2e4d70SOleg Nesterov }; 2432fc2e4d70SOleg Nesterov 2433fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2434fc2e4d70SOleg Nesterov { 2435fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2436fc2e4d70SOleg Nesterov complete(&barr->done); 2437fc2e4d70SOleg Nesterov } 2438fc2e4d70SOleg Nesterov 24394690c4abSTejun Heo /** 24404690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2441112202d9STejun Heo * @pwq: pwq to insert barrier into 24424690c4abSTejun Heo * @barr: wq_barrier to insert 2443affee4b2STejun Heo * @target: target work to attach @barr to 2444affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24454690c4abSTejun Heo * 2446affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2447affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2448affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2449affee4b2STejun Heo * cpu. 2450affee4b2STejun Heo * 2451affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2452affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2453affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2454affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2455affee4b2STejun Heo * after a work with LINKED flag set. 2456affee4b2STejun Heo * 2457affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2458112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24594690c4abSTejun Heo * 24604690c4abSTejun Heo * CONTEXT: 2461d565ed63STejun Heo * spin_lock_irq(pool->lock). 24624690c4abSTejun Heo */ 2463112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2464affee4b2STejun Heo struct wq_barrier *barr, 2465affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2466fc2e4d70SOleg Nesterov { 2467affee4b2STejun Heo struct list_head *head; 2468affee4b2STejun Heo unsigned int linked = 0; 2469affee4b2STejun Heo 2470dc186ad7SThomas Gleixner /* 2471d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2472dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2473dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2474dc186ad7SThomas Gleixner * might deadlock. 2475dc186ad7SThomas Gleixner */ 2476ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 247722df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2478fc2e4d70SOleg Nesterov init_completion(&barr->done); 24792607d7a6STejun Heo barr->task = current; 248083c22520SOleg Nesterov 2481affee4b2STejun Heo /* 2482affee4b2STejun Heo * If @target is currently being executed, schedule the 2483affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2484affee4b2STejun Heo */ 2485affee4b2STejun Heo if (worker) 2486affee4b2STejun Heo head = worker->scheduled.next; 2487affee4b2STejun Heo else { 2488affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2489affee4b2STejun Heo 2490affee4b2STejun Heo head = target->entry.next; 2491affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2492affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2493affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2494affee4b2STejun Heo } 2495affee4b2STejun Heo 2496dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2497112202d9STejun Heo insert_work(pwq, &barr->work, head, 2498affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2499fc2e4d70SOleg Nesterov } 2500fc2e4d70SOleg Nesterov 250173f53c4aSTejun Heo /** 2502112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 250373f53c4aSTejun Heo * @wq: workqueue being flushed 250473f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 250573f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 250673f53c4aSTejun Heo * 2507112202d9STejun Heo * Prepare pwqs for workqueue flushing. 250873f53c4aSTejun Heo * 2509112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2510112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2511112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2512112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2513112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 251473f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 251573f53c4aSTejun Heo * 251673f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 251773f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 251873f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 251973f53c4aSTejun Heo * is returned. 252073f53c4aSTejun Heo * 2521112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 252273f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 252373f53c4aSTejun Heo * advanced to @work_color. 252473f53c4aSTejun Heo * 252573f53c4aSTejun Heo * CONTEXT: 25263c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 252773f53c4aSTejun Heo * 2528d185af30SYacine Belkadi * Return: 252973f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 253073f53c4aSTejun Heo * otherwise. 253173f53c4aSTejun Heo */ 2532112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 253373f53c4aSTejun Heo int flush_color, int work_color) 25341da177e4SLinus Torvalds { 253573f53c4aSTejun Heo bool wait = false; 253649e3cf44STejun Heo struct pool_workqueue *pwq; 25371da177e4SLinus Torvalds 253873f53c4aSTejun Heo if (flush_color >= 0) { 25396183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2540112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2541dc186ad7SThomas Gleixner } 254214441960SOleg Nesterov 254349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2544112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25451da177e4SLinus Torvalds 2546b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 254773f53c4aSTejun Heo 254873f53c4aSTejun Heo if (flush_color >= 0) { 25496183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 255073f53c4aSTejun Heo 2551112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2552112202d9STejun Heo pwq->flush_color = flush_color; 2553112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 255473f53c4aSTejun Heo wait = true; 25551da177e4SLinus Torvalds } 255673f53c4aSTejun Heo } 255773f53c4aSTejun Heo 255873f53c4aSTejun Heo if (work_color >= 0) { 25596183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2560112202d9STejun Heo pwq->work_color = work_color; 256173f53c4aSTejun Heo } 256273f53c4aSTejun Heo 2563b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25641da177e4SLinus Torvalds } 25651da177e4SLinus Torvalds 2566112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 256773f53c4aSTejun Heo complete(&wq->first_flusher->done); 256873f53c4aSTejun Heo 256973f53c4aSTejun Heo return wait; 257083c22520SOleg Nesterov } 25711da177e4SLinus Torvalds 25720fcb78c2SRolf Eike Beer /** 25731da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25740fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25751da177e4SLinus Torvalds * 2576c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2577c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25781da177e4SLinus Torvalds */ 25797ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25801da177e4SLinus Torvalds { 258173f53c4aSTejun Heo struct wq_flusher this_flusher = { 258273f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 258373f53c4aSTejun Heo .flush_color = -1, 258473f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 258573f53c4aSTejun Heo }; 258673f53c4aSTejun Heo int next_color; 2587b1f4ec17SOleg Nesterov 25883347fa09STejun Heo if (WARN_ON(!wq_online)) 25893347fa09STejun Heo return; 25903347fa09STejun Heo 25913295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25923295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 259373f53c4aSTejun Heo 25943c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 259573f53c4aSTejun Heo 259673f53c4aSTejun Heo /* 259773f53c4aSTejun Heo * Start-to-wait phase 259873f53c4aSTejun Heo */ 259973f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 260073f53c4aSTejun Heo 260173f53c4aSTejun Heo if (next_color != wq->flush_color) { 260273f53c4aSTejun Heo /* 260373f53c4aSTejun Heo * Color space is not full. The current work_color 260473f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 260573f53c4aSTejun Heo * by one. 260673f53c4aSTejun Heo */ 26076183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 260873f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 260973f53c4aSTejun Heo wq->work_color = next_color; 261073f53c4aSTejun Heo 261173f53c4aSTejun Heo if (!wq->first_flusher) { 261273f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26136183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 261473f53c4aSTejun Heo 261573f53c4aSTejun Heo wq->first_flusher = &this_flusher; 261673f53c4aSTejun Heo 2617112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 261873f53c4aSTejun Heo wq->work_color)) { 261973f53c4aSTejun Heo /* nothing to flush, done */ 262073f53c4aSTejun Heo wq->flush_color = next_color; 262173f53c4aSTejun Heo wq->first_flusher = NULL; 262273f53c4aSTejun Heo goto out_unlock; 262373f53c4aSTejun Heo } 262473f53c4aSTejun Heo } else { 262573f53c4aSTejun Heo /* wait in queue */ 26266183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 262773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2628112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 262973f53c4aSTejun Heo } 263073f53c4aSTejun Heo } else { 263173f53c4aSTejun Heo /* 263273f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 263373f53c4aSTejun Heo * The next flush completion will assign us 263473f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 263573f53c4aSTejun Heo */ 263673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 263773f53c4aSTejun Heo } 263873f53c4aSTejun Heo 2639fca839c0STejun Heo check_flush_dependency(wq, NULL); 2640fca839c0STejun Heo 26413c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 264273f53c4aSTejun Heo 264373f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 264473f53c4aSTejun Heo 264573f53c4aSTejun Heo /* 264673f53c4aSTejun Heo * Wake-up-and-cascade phase 264773f53c4aSTejun Heo * 264873f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 264973f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 265073f53c4aSTejun Heo */ 265173f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 265273f53c4aSTejun Heo return; 265373f53c4aSTejun Heo 26543c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 265573f53c4aSTejun Heo 26564ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26574ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26584ce48b37STejun Heo goto out_unlock; 26594ce48b37STejun Heo 266073f53c4aSTejun Heo wq->first_flusher = NULL; 266173f53c4aSTejun Heo 26626183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26636183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 266473f53c4aSTejun Heo 266573f53c4aSTejun Heo while (true) { 266673f53c4aSTejun Heo struct wq_flusher *next, *tmp; 266773f53c4aSTejun Heo 266873f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 266973f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 267073f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 267173f53c4aSTejun Heo break; 267273f53c4aSTejun Heo list_del_init(&next->list); 267373f53c4aSTejun Heo complete(&next->done); 267473f53c4aSTejun Heo } 267573f53c4aSTejun Heo 26766183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 267773f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 267873f53c4aSTejun Heo 267973f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 268073f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 268173f53c4aSTejun Heo 268273f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 268373f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 268473f53c4aSTejun Heo /* 268573f53c4aSTejun Heo * Assign the same color to all overflowed 268673f53c4aSTejun Heo * flushers, advance work_color and append to 268773f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 268873f53c4aSTejun Heo * phase for these overflowed flushers. 268973f53c4aSTejun Heo */ 269073f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 269173f53c4aSTejun Heo tmp->flush_color = wq->work_color; 269273f53c4aSTejun Heo 269373f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 269473f53c4aSTejun Heo 269573f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 269673f53c4aSTejun Heo &wq->flusher_queue); 2697112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 269873f53c4aSTejun Heo } 269973f53c4aSTejun Heo 270073f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27016183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 270273f53c4aSTejun Heo break; 270373f53c4aSTejun Heo } 270473f53c4aSTejun Heo 270573f53c4aSTejun Heo /* 270673f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2707112202d9STejun Heo * the new first flusher and arm pwqs. 270873f53c4aSTejun Heo */ 27096183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27106183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 271173f53c4aSTejun Heo 271273f53c4aSTejun Heo list_del_init(&next->list); 271373f53c4aSTejun Heo wq->first_flusher = next; 271473f53c4aSTejun Heo 2715112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 271673f53c4aSTejun Heo break; 271773f53c4aSTejun Heo 271873f53c4aSTejun Heo /* 271973f53c4aSTejun Heo * Meh... this color is already done, clear first 272073f53c4aSTejun Heo * flusher and repeat cascading. 272173f53c4aSTejun Heo */ 272273f53c4aSTejun Heo wq->first_flusher = NULL; 272373f53c4aSTejun Heo } 272473f53c4aSTejun Heo 272573f53c4aSTejun Heo out_unlock: 27263c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27271da177e4SLinus Torvalds } 27281dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 27291da177e4SLinus Torvalds 27309c5a2ba7STejun Heo /** 27319c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27329c5a2ba7STejun Heo * @wq: workqueue to drain 27339c5a2ba7STejun Heo * 27349c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27359c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27369c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2737b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 27389c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27399c5a2ba7STejun Heo * takes too long. 27409c5a2ba7STejun Heo */ 27419c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27429c5a2ba7STejun Heo { 27439c5a2ba7STejun Heo unsigned int flush_cnt = 0; 274449e3cf44STejun Heo struct pool_workqueue *pwq; 27459c5a2ba7STejun Heo 27469c5a2ba7STejun Heo /* 27479c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27489c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2749618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27509c5a2ba7STejun Heo */ 275187fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27529c5a2ba7STejun Heo if (!wq->nr_drainers++) 2753618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 275487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27559c5a2ba7STejun Heo reflush: 27569c5a2ba7STejun Heo flush_workqueue(wq); 27579c5a2ba7STejun Heo 2758b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 275976af4d93STejun Heo 276049e3cf44STejun Heo for_each_pwq(pwq, wq) { 2761fa2563e4SThomas Tuttle bool drained; 27629c5a2ba7STejun Heo 2763b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2764112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2765b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2766fa2563e4SThomas Tuttle 2767fa2563e4SThomas Tuttle if (drained) 27689c5a2ba7STejun Heo continue; 27699c5a2ba7STejun Heo 27709c5a2ba7STejun Heo if (++flush_cnt == 10 || 27719c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2772c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27739c5a2ba7STejun Heo wq->name, flush_cnt); 277476af4d93STejun Heo 2775b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27769c5a2ba7STejun Heo goto reflush; 27779c5a2ba7STejun Heo } 27789c5a2ba7STejun Heo 27799c5a2ba7STejun Heo if (!--wq->nr_drainers) 2780618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 278187fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27829c5a2ba7STejun Heo } 27839c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27849c5a2ba7STejun Heo 2785606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2786baf59022STejun Heo { 2787baf59022STejun Heo struct worker *worker = NULL; 2788c9e7cf27STejun Heo struct worker_pool *pool; 2789112202d9STejun Heo struct pool_workqueue *pwq; 2790baf59022STejun Heo 2791baf59022STejun Heo might_sleep(); 2792baf59022STejun Heo 2793fa1b54e6STejun Heo local_irq_disable(); 2794fa1b54e6STejun Heo pool = get_work_pool(work); 2795fa1b54e6STejun Heo if (!pool) { 2796fa1b54e6STejun Heo local_irq_enable(); 2797fa1b54e6STejun Heo return false; 2798fa1b54e6STejun Heo } 2799fa1b54e6STejun Heo 2800fa1b54e6STejun Heo spin_lock(&pool->lock); 28010b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2802112202d9STejun Heo pwq = get_work_pwq(work); 2803112202d9STejun Heo if (pwq) { 2804112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2805baf59022STejun Heo goto already_gone; 2806606a5020STejun Heo } else { 2807c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2808baf59022STejun Heo if (!worker) 2809baf59022STejun Heo goto already_gone; 2810112202d9STejun Heo pwq = worker->current_pwq; 2811606a5020STejun Heo } 2812baf59022STejun Heo 2813fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2814fca839c0STejun Heo 2815112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2816d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2817baf59022STejun Heo 2818e159489bSTejun Heo /* 2819e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2820e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2821e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2822e159489bSTejun Heo * access. 2823e159489bSTejun Heo */ 2824493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2825112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2826e159489bSTejun Heo else 2827112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2828112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2829e159489bSTejun Heo 2830baf59022STejun Heo return true; 2831baf59022STejun Heo already_gone: 2832d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2833baf59022STejun Heo return false; 2834baf59022STejun Heo } 2835baf59022STejun Heo 2836db700897SOleg Nesterov /** 2837401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2838401a8d04STejun Heo * @work: the work to flush 2839db700897SOleg Nesterov * 2840606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2841606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2842401a8d04STejun Heo * 2843d185af30SYacine Belkadi * Return: 2844401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2845401a8d04STejun Heo * %false if it was already idle. 2846db700897SOleg Nesterov */ 2847401a8d04STejun Heo bool flush_work(struct work_struct *work) 2848db700897SOleg Nesterov { 284912997d1aSBjorn Helgaas struct wq_barrier barr; 285012997d1aSBjorn Helgaas 28513347fa09STejun Heo if (WARN_ON(!wq_online)) 28523347fa09STejun Heo return false; 28533347fa09STejun Heo 28540976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28550976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28560976dfc1SStephen Boyd 285712997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 285812997d1aSBjorn Helgaas wait_for_completion(&barr.done); 285912997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 286012997d1aSBjorn Helgaas return true; 286112997d1aSBjorn Helgaas } else { 286212997d1aSBjorn Helgaas return false; 286312997d1aSBjorn Helgaas } 2864606a5020STejun Heo } 2865db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2866db700897SOleg Nesterov 28678603e1b3STejun Heo struct cwt_wait { 28688603e1b3STejun Heo wait_queue_t wait; 28698603e1b3STejun Heo struct work_struct *work; 28708603e1b3STejun Heo }; 28718603e1b3STejun Heo 28728603e1b3STejun Heo static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) 28738603e1b3STejun Heo { 28748603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 28758603e1b3STejun Heo 28768603e1b3STejun Heo if (cwait->work != key) 28778603e1b3STejun Heo return 0; 28788603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 28798603e1b3STejun Heo } 28808603e1b3STejun Heo 288136e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2882401a8d04STejun Heo { 28838603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2884bbb68dfaSTejun Heo unsigned long flags; 28851f1f642eSOleg Nesterov int ret; 28861f1f642eSOleg Nesterov 28871f1f642eSOleg Nesterov do { 2888bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2889bbb68dfaSTejun Heo /* 28908603e1b3STejun Heo * If someone else is already canceling, wait for it to 28918603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 28928603e1b3STejun Heo * because we may get scheduled between @work's completion 28938603e1b3STejun Heo * and the other canceling task resuming and clearing 28948603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 28958603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 28968603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 28978603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 28988603e1b3STejun Heo * we're hogging the CPU. 28998603e1b3STejun Heo * 29008603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 29018603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 29028603e1b3STejun Heo * wake function which matches @work along with exclusive 29038603e1b3STejun Heo * wait and wakeup. 2904bbb68dfaSTejun Heo */ 29058603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 29068603e1b3STejun Heo struct cwt_wait cwait; 29078603e1b3STejun Heo 29088603e1b3STejun Heo init_wait(&cwait.wait); 29098603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 29108603e1b3STejun Heo cwait.work = work; 29118603e1b3STejun Heo 29128603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 29138603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 29148603e1b3STejun Heo if (work_is_canceling(work)) 29158603e1b3STejun Heo schedule(); 29168603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 29178603e1b3STejun Heo } 29181f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29191f1f642eSOleg Nesterov 2920bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2921bbb68dfaSTejun Heo mark_work_canceling(work); 2922bbb68dfaSTejun Heo local_irq_restore(flags); 2923bbb68dfaSTejun Heo 29243347fa09STejun Heo /* 29253347fa09STejun Heo * This allows canceling during early boot. We know that @work 29263347fa09STejun Heo * isn't executing. 29273347fa09STejun Heo */ 29283347fa09STejun Heo if (wq_online) 2929606a5020STejun Heo flush_work(work); 29303347fa09STejun Heo 29317a22ad75STejun Heo clear_work_data(work); 29328603e1b3STejun Heo 29338603e1b3STejun Heo /* 29348603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 29358603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 29368603e1b3STejun Heo * visible there. 29378603e1b3STejun Heo */ 29388603e1b3STejun Heo smp_mb(); 29398603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 29408603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 29418603e1b3STejun Heo 29421f1f642eSOleg Nesterov return ret; 29431f1f642eSOleg Nesterov } 29441f1f642eSOleg Nesterov 29456e84d644SOleg Nesterov /** 2946401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2947401a8d04STejun Heo * @work: the work to cancel 29486e84d644SOleg Nesterov * 2949401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2950401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2951401a8d04STejun Heo * another workqueue. On return from this function, @work is 2952401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29531f1f642eSOleg Nesterov * 2954401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2955401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29566e84d644SOleg Nesterov * 2957401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29586e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2959401a8d04STejun Heo * 2960d185af30SYacine Belkadi * Return: 2961401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29626e84d644SOleg Nesterov */ 2963401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29646e84d644SOleg Nesterov { 296536e227d2STejun Heo return __cancel_work_timer(work, false); 2966b89deed3SOleg Nesterov } 296728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2968b89deed3SOleg Nesterov 29696e84d644SOleg Nesterov /** 2970401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2971401a8d04STejun Heo * @dwork: the delayed work to flush 29726e84d644SOleg Nesterov * 2973401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2974401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2975401a8d04STejun Heo * considers the last queueing instance of @dwork. 29761f1f642eSOleg Nesterov * 2977d185af30SYacine Belkadi * Return: 2978401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2979401a8d04STejun Heo * %false if it was already idle. 29806e84d644SOleg Nesterov */ 2981401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2982401a8d04STejun Heo { 29838930cabaSTejun Heo local_irq_disable(); 2984401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 298560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29868930cabaSTejun Heo local_irq_enable(); 2987401a8d04STejun Heo return flush_work(&dwork->work); 2988401a8d04STejun Heo } 2989401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2990401a8d04STejun Heo 2991401a8d04STejun Heo /** 299257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 299357b30ae7STejun Heo * @dwork: delayed_work to cancel 299409383498STejun Heo * 2995d185af30SYacine Belkadi * Kill off a pending delayed_work. 2996d185af30SYacine Belkadi * 2997d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2998d185af30SYacine Belkadi * pending. 2999d185af30SYacine Belkadi * 3000d185af30SYacine Belkadi * Note: 3001d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3002d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3003d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 300409383498STejun Heo * 300557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 300609383498STejun Heo */ 300757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 300809383498STejun Heo { 300957b30ae7STejun Heo unsigned long flags; 301057b30ae7STejun Heo int ret; 301157b30ae7STejun Heo 301257b30ae7STejun Heo do { 301357b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 301457b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 301557b30ae7STejun Heo 301657b30ae7STejun Heo if (unlikely(ret < 0)) 301757b30ae7STejun Heo return false; 301857b30ae7STejun Heo 30197c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 30207c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 302157b30ae7STejun Heo local_irq_restore(flags); 3022c0158ca6SDan Magenheimer return ret; 302309383498STejun Heo } 302457b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 302509383498STejun Heo 302609383498STejun Heo /** 3027401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3028401a8d04STejun Heo * @dwork: the delayed work cancel 3029401a8d04STejun Heo * 3030401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3031401a8d04STejun Heo * 3032d185af30SYacine Belkadi * Return: 3033401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3034401a8d04STejun Heo */ 3035401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30366e84d644SOleg Nesterov { 303736e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30386e84d644SOleg Nesterov } 3039f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30401da177e4SLinus Torvalds 30410fcb78c2SRolf Eike Beer /** 304231ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3043b6136773SAndrew Morton * @func: the function to call 3044b6136773SAndrew Morton * 304531ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 304631ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3047b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 304831ddd871STejun Heo * 3049d185af30SYacine Belkadi * Return: 305031ddd871STejun Heo * 0 on success, -errno on failure. 3051b6136773SAndrew Morton */ 305265f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 305315316ba8SChristoph Lameter { 305415316ba8SChristoph Lameter int cpu; 305538f51568SNamhyung Kim struct work_struct __percpu *works; 305615316ba8SChristoph Lameter 3057b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3058b6136773SAndrew Morton if (!works) 305915316ba8SChristoph Lameter return -ENOMEM; 3060b6136773SAndrew Morton 306195402b38SGautham R Shenoy get_online_cpus(); 306293981800STejun Heo 306315316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30649bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30659bfb1839SIngo Molnar 30669bfb1839SIngo Molnar INIT_WORK(work, func); 30678de6d308SOleg Nesterov schedule_work_on(cpu, work); 306815316ba8SChristoph Lameter } 306993981800STejun Heo 307093981800STejun Heo for_each_online_cpu(cpu) 30718616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 307293981800STejun Heo 307395402b38SGautham R Shenoy put_online_cpus(); 3074b6136773SAndrew Morton free_percpu(works); 307515316ba8SChristoph Lameter return 0; 307615316ba8SChristoph Lameter } 307715316ba8SChristoph Lameter 3078eef6a7d5SAlan Stern /** 30791fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30801fa44ecaSJames Bottomley * @fn: the function to execute 30811fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30821fa44ecaSJames Bottomley * be available when the work executes) 30831fa44ecaSJames Bottomley * 30841fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30851fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30861fa44ecaSJames Bottomley * 3087d185af30SYacine Belkadi * Return: 0 - function was executed 30881fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30891fa44ecaSJames Bottomley */ 309065f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30911fa44ecaSJames Bottomley { 30921fa44ecaSJames Bottomley if (!in_interrupt()) { 309365f27f38SDavid Howells fn(&ew->work); 30941fa44ecaSJames Bottomley return 0; 30951fa44ecaSJames Bottomley } 30961fa44ecaSJames Bottomley 309765f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30981fa44ecaSJames Bottomley schedule_work(&ew->work); 30991fa44ecaSJames Bottomley 31001fa44ecaSJames Bottomley return 1; 31011fa44ecaSJames Bottomley } 31021fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31031fa44ecaSJames Bottomley 31047a4e344cSTejun Heo /** 31057a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 31067a4e344cSTejun Heo * @attrs: workqueue_attrs to free 31077a4e344cSTejun Heo * 31087a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 31097a4e344cSTejun Heo */ 31107a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 31117a4e344cSTejun Heo { 31127a4e344cSTejun Heo if (attrs) { 31137a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 31147a4e344cSTejun Heo kfree(attrs); 31157a4e344cSTejun Heo } 31167a4e344cSTejun Heo } 31177a4e344cSTejun Heo 31187a4e344cSTejun Heo /** 31197a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 31207a4e344cSTejun Heo * @gfp_mask: allocation mask to use 31217a4e344cSTejun Heo * 31227a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3123d185af30SYacine Belkadi * return it. 3124d185af30SYacine Belkadi * 3125d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 31267a4e344cSTejun Heo */ 31277a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 31287a4e344cSTejun Heo { 31297a4e344cSTejun Heo struct workqueue_attrs *attrs; 31307a4e344cSTejun Heo 31317a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 31327a4e344cSTejun Heo if (!attrs) 31337a4e344cSTejun Heo goto fail; 31347a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 31357a4e344cSTejun Heo goto fail; 31367a4e344cSTejun Heo 313713e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 31387a4e344cSTejun Heo return attrs; 31397a4e344cSTejun Heo fail: 31407a4e344cSTejun Heo free_workqueue_attrs(attrs); 31417a4e344cSTejun Heo return NULL; 31427a4e344cSTejun Heo } 31437a4e344cSTejun Heo 314429c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 314529c91e99STejun Heo const struct workqueue_attrs *from) 314629c91e99STejun Heo { 314729c91e99STejun Heo to->nice = from->nice; 314829c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 31492865a8fbSShaohua Li /* 31502865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 31512865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 31522865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 31532865a8fbSShaohua Li */ 31542865a8fbSShaohua Li to->no_numa = from->no_numa; 315529c91e99STejun Heo } 315629c91e99STejun Heo 315729c91e99STejun Heo /* hash value of the content of @attr */ 315829c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 315929c91e99STejun Heo { 316029c91e99STejun Heo u32 hash = 0; 316129c91e99STejun Heo 316229c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 316313e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 316413e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 316529c91e99STejun Heo return hash; 316629c91e99STejun Heo } 316729c91e99STejun Heo 316829c91e99STejun Heo /* content equality test */ 316929c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 317029c91e99STejun Heo const struct workqueue_attrs *b) 317129c91e99STejun Heo { 317229c91e99STejun Heo if (a->nice != b->nice) 317329c91e99STejun Heo return false; 317429c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 317529c91e99STejun Heo return false; 317629c91e99STejun Heo return true; 317729c91e99STejun Heo } 317829c91e99STejun Heo 31797a4e344cSTejun Heo /** 31807a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 31817a4e344cSTejun Heo * @pool: worker_pool to initialize 31827a4e344cSTejun Heo * 3183402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3184d185af30SYacine Belkadi * 3185d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 318629c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 318729c91e99STejun Heo * on @pool safely to release it. 31887a4e344cSTejun Heo */ 31897a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 31904e1a1f9aSTejun Heo { 31914e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 319229c91e99STejun Heo pool->id = -1; 319329c91e99STejun Heo pool->cpu = -1; 3194f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 31954e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 319682607adcSTejun Heo pool->watchdog_ts = jiffies; 31974e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 31984e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 31994e1a1f9aSTejun Heo hash_init(pool->busy_hash); 32004e1a1f9aSTejun Heo 32014e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 32024e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 32034e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 32044e1a1f9aSTejun Heo 32054e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 32064e1a1f9aSTejun Heo (unsigned long)pool); 32074e1a1f9aSTejun Heo 32084e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 320992f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3210da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 32117a4e344cSTejun Heo 32127cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 321329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 321429c91e99STejun Heo pool->refcnt = 1; 321529c91e99STejun Heo 321629c91e99STejun Heo /* shouldn't fail above this point */ 32177a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 32187a4e344cSTejun Heo if (!pool->attrs) 32197a4e344cSTejun Heo return -ENOMEM; 32207a4e344cSTejun Heo return 0; 32214e1a1f9aSTejun Heo } 32224e1a1f9aSTejun Heo 3223e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3224e2dca7adSTejun Heo { 3225e2dca7adSTejun Heo struct workqueue_struct *wq = 3226e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3227e2dca7adSTejun Heo 3228e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3229e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3230e2dca7adSTejun Heo else 3231e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3232e2dca7adSTejun Heo 3233e2dca7adSTejun Heo kfree(wq->rescuer); 3234e2dca7adSTejun Heo kfree(wq); 3235e2dca7adSTejun Heo } 3236e2dca7adSTejun Heo 323729c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 323829c91e99STejun Heo { 323929c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 324029c91e99STejun Heo 32417cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 324229c91e99STejun Heo free_workqueue_attrs(pool->attrs); 324329c91e99STejun Heo kfree(pool); 324429c91e99STejun Heo } 324529c91e99STejun Heo 324629c91e99STejun Heo /** 324729c91e99STejun Heo * put_unbound_pool - put a worker_pool 324829c91e99STejun Heo * @pool: worker_pool to put 324929c91e99STejun Heo * 325029c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3251c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3252c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3253c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3254a892caccSTejun Heo * 3255a892caccSTejun Heo * Should be called with wq_pool_mutex held. 325629c91e99STejun Heo */ 325729c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 325829c91e99STejun Heo { 325960f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 326029c91e99STejun Heo struct worker *worker; 326129c91e99STejun Heo 3262a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3263a892caccSTejun Heo 3264a892caccSTejun Heo if (--pool->refcnt) 326529c91e99STejun Heo return; 326629c91e99STejun Heo 326729c91e99STejun Heo /* sanity checks */ 326861d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3269a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 327029c91e99STejun Heo return; 327129c91e99STejun Heo 327229c91e99STejun Heo /* release id and unhash */ 327329c91e99STejun Heo if (pool->id >= 0) 327429c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 327529c91e99STejun Heo hash_del(&pool->hash_node); 327629c91e99STejun Heo 3277c5aa87bbSTejun Heo /* 3278c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3279c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 328092f9c5c4SLai Jiangshan * attach_mutex. 3281c5aa87bbSTejun Heo */ 328229c91e99STejun Heo mutex_lock(&pool->manager_arb); 328329c91e99STejun Heo 328460f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 32851037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 328629c91e99STejun Heo destroy_worker(worker); 328729c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 328829c91e99STejun Heo spin_unlock_irq(&pool->lock); 328960f5a4bcSLai Jiangshan 329092f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3291da028469SLai Jiangshan if (!list_empty(&pool->workers)) 329260f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 329392f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 329460f5a4bcSLai Jiangshan 329560f5a4bcSLai Jiangshan if (pool->detach_completion) 329660f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 329760f5a4bcSLai Jiangshan 329829c91e99STejun Heo mutex_unlock(&pool->manager_arb); 329929c91e99STejun Heo 330029c91e99STejun Heo /* shut down the timers */ 330129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 330229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 330329c91e99STejun Heo 330429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 330529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 330629c91e99STejun Heo } 330729c91e99STejun Heo 330829c91e99STejun Heo /** 330929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 331029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 331129c91e99STejun Heo * 331229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 331329c91e99STejun Heo * reference count and return it. If there already is a matching 331429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3315d185af30SYacine Belkadi * create a new one. 3316a892caccSTejun Heo * 3317a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3318d185af30SYacine Belkadi * 3319d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3320d185af30SYacine Belkadi * On failure, %NULL. 332129c91e99STejun Heo */ 332229c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 332329c91e99STejun Heo { 332429c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 332529c91e99STejun Heo struct worker_pool *pool; 3326f3f90ad4STejun Heo int node; 3327e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 332829c91e99STejun Heo 3329a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 333029c91e99STejun Heo 333129c91e99STejun Heo /* do we already have a matching pool? */ 333229c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 333329c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 333429c91e99STejun Heo pool->refcnt++; 33353fb1823cSLai Jiangshan return pool; 333629c91e99STejun Heo } 333729c91e99STejun Heo } 333829c91e99STejun Heo 3339e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3340e2273584SXunlei Pang if (wq_numa_enabled) { 3341e2273584SXunlei Pang for_each_node(node) { 3342e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3343e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3344e2273584SXunlei Pang target_node = node; 3345e2273584SXunlei Pang break; 3346e2273584SXunlei Pang } 3347e2273584SXunlei Pang } 3348e2273584SXunlei Pang } 3349e2273584SXunlei Pang 335029c91e99STejun Heo /* nope, create a new one */ 3351e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 335229c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 335329c91e99STejun Heo goto fail; 335429c91e99STejun Heo 33558864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 335629c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3357e2273584SXunlei Pang pool->node = target_node; 335829c91e99STejun Heo 33592865a8fbSShaohua Li /* 33602865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 33612865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 33622865a8fbSShaohua Li */ 33632865a8fbSShaohua Li pool->attrs->no_numa = false; 33642865a8fbSShaohua Li 336529c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 336629c91e99STejun Heo goto fail; 336729c91e99STejun Heo 336829c91e99STejun Heo /* create and start the initial worker */ 33693347fa09STejun Heo if (wq_online && !create_worker(pool)) 337029c91e99STejun Heo goto fail; 337129c91e99STejun Heo 337229c91e99STejun Heo /* install */ 337329c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 33743fb1823cSLai Jiangshan 337529c91e99STejun Heo return pool; 337629c91e99STejun Heo fail: 337729c91e99STejun Heo if (pool) 337829c91e99STejun Heo put_unbound_pool(pool); 337929c91e99STejun Heo return NULL; 338029c91e99STejun Heo } 338129c91e99STejun Heo 33828864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 33838864b4e5STejun Heo { 33848864b4e5STejun Heo kmem_cache_free(pwq_cache, 33858864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 33868864b4e5STejun Heo } 33878864b4e5STejun Heo 33888864b4e5STejun Heo /* 33898864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 33908864b4e5STejun Heo * and needs to be destroyed. 33918864b4e5STejun Heo */ 33928864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 33938864b4e5STejun Heo { 33948864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 33958864b4e5STejun Heo unbound_release_work); 33968864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 33978864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3398bc0caf09STejun Heo bool is_last; 33998864b4e5STejun Heo 34008864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 34018864b4e5STejun Heo return; 34028864b4e5STejun Heo 34033c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 34048864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3405bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 34063c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 34078864b4e5STejun Heo 3408a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 34098864b4e5STejun Heo put_unbound_pool(pool); 3410a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3411a892caccSTejun Heo 34128864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 34138864b4e5STejun Heo 34148864b4e5STejun Heo /* 34158864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3416e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 34178864b4e5STejun Heo */ 3418e2dca7adSTejun Heo if (is_last) 3419e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 34206029a918STejun Heo } 34218864b4e5STejun Heo 34220fbd95aaSTejun Heo /** 3423699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 34240fbd95aaSTejun Heo * @pwq: target pool_workqueue 34250fbd95aaSTejun Heo * 3426699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3427699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3428699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 34290fbd95aaSTejun Heo */ 3430699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 34310fbd95aaSTejun Heo { 3432699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3433699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 34343347fa09STejun Heo unsigned long flags; 3435699ce097STejun Heo 3436699ce097STejun Heo /* for @wq->saved_max_active */ 3437a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3438699ce097STejun Heo 3439699ce097STejun Heo /* fast exit for non-freezable wqs */ 3440699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3441699ce097STejun Heo return; 3442699ce097STejun Heo 34433347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 34443347fa09STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 3445699ce097STejun Heo 344674b414eaSLai Jiangshan /* 344774b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 344874b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 344974b414eaSLai Jiangshan * is updated and visible. 345074b414eaSLai Jiangshan */ 345174b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3452699ce097STejun Heo pwq->max_active = wq->saved_max_active; 34530fbd95aaSTejun Heo 34540fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 34550fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 34560fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3457951a078aSLai Jiangshan 3458951a078aSLai Jiangshan /* 3459951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3460951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3461951a078aSLai Jiangshan */ 3462951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3463699ce097STejun Heo } else { 3464699ce097STejun Heo pwq->max_active = 0; 3465699ce097STejun Heo } 3466699ce097STejun Heo 34673347fa09STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 34680fbd95aaSTejun Heo } 34690fbd95aaSTejun Heo 3470e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3471f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3472f147f29eSTejun Heo struct worker_pool *pool) 3473d2c1d404STejun Heo { 3474d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3475d2c1d404STejun Heo 3476e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3477e50aba9aSTejun Heo 3478d2c1d404STejun Heo pwq->pool = pool; 3479d2c1d404STejun Heo pwq->wq = wq; 3480d2c1d404STejun Heo pwq->flush_color = -1; 34818864b4e5STejun Heo pwq->refcnt = 1; 3482d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 34831befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3484d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 34858864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3486f147f29eSTejun Heo } 3487d2c1d404STejun Heo 3488f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 34891befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3490f147f29eSTejun Heo { 3491f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3492f147f29eSTejun Heo 3493f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 349475ccf595STejun Heo 34951befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 34961befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 34971befcf30STejun Heo return; 34981befcf30STejun Heo 349929b1cb41SLai Jiangshan /* set the matching work_color */ 350075ccf595STejun Heo pwq->work_color = wq->work_color; 3501983ca25eSTejun Heo 3502983ca25eSTejun Heo /* sync max_active to the current setting */ 3503983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3504983ca25eSTejun Heo 3505983ca25eSTejun Heo /* link in @pwq */ 35069e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3507df2d5ae4STejun Heo } 35086029a918STejun Heo 3509f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3510f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3511f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3512f147f29eSTejun Heo { 3513f147f29eSTejun Heo struct worker_pool *pool; 3514f147f29eSTejun Heo struct pool_workqueue *pwq; 3515f147f29eSTejun Heo 3516f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3517f147f29eSTejun Heo 3518f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3519f147f29eSTejun Heo if (!pool) 3520f147f29eSTejun Heo return NULL; 3521f147f29eSTejun Heo 3522e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3523f147f29eSTejun Heo if (!pwq) { 3524f147f29eSTejun Heo put_unbound_pool(pool); 3525f147f29eSTejun Heo return NULL; 3526f147f29eSTejun Heo } 3527f147f29eSTejun Heo 3528f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3529f147f29eSTejun Heo return pwq; 3530d2c1d404STejun Heo } 3531d2c1d404STejun Heo 35324c16bd32STejun Heo /** 353330186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3534042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 35354c16bd32STejun Heo * @node: the target NUMA node 35364c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 35374c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 35384c16bd32STejun Heo * 35394c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 35404c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3541d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 35424c16bd32STejun Heo * 35434c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 35444c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 35454c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 35464c16bd32STejun Heo * @attrs->cpumask. 35474c16bd32STejun Heo * 35484c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 35494c16bd32STejun Heo * stable. 3550d185af30SYacine Belkadi * 3551d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3552d185af30SYacine Belkadi * %false if equal. 35534c16bd32STejun Heo */ 35544c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 35554c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 35564c16bd32STejun Heo { 3557d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 35584c16bd32STejun Heo goto use_dfl; 35594c16bd32STejun Heo 35604c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 35614c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 35624c16bd32STejun Heo if (cpu_going_down >= 0) 35634c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 35644c16bd32STejun Heo 35654c16bd32STejun Heo if (cpumask_empty(cpumask)) 35664c16bd32STejun Heo goto use_dfl; 35674c16bd32STejun Heo 35684c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 35694c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 35704c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 35714c16bd32STejun Heo 35724c16bd32STejun Heo use_dfl: 35734c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 35744c16bd32STejun Heo return false; 35754c16bd32STejun Heo } 35764c16bd32STejun Heo 35771befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 35781befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 35791befcf30STejun Heo int node, 35801befcf30STejun Heo struct pool_workqueue *pwq) 35811befcf30STejun Heo { 35821befcf30STejun Heo struct pool_workqueue *old_pwq; 35831befcf30STejun Heo 35845b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 35851befcf30STejun Heo lockdep_assert_held(&wq->mutex); 35861befcf30STejun Heo 35871befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 35881befcf30STejun Heo link_pwq(pwq); 35891befcf30STejun Heo 35901befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 35911befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 35921befcf30STejun Heo return old_pwq; 35931befcf30STejun Heo } 35941befcf30STejun Heo 35952d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 35962d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 35972d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 35982d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3599042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 36002d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 36012d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 36022d5f0764SLai Jiangshan }; 36032d5f0764SLai Jiangshan 36042d5f0764SLai Jiangshan /* free the resources after success or abort */ 36052d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 36062d5f0764SLai Jiangshan { 36072d5f0764SLai Jiangshan if (ctx) { 36082d5f0764SLai Jiangshan int node; 36092d5f0764SLai Jiangshan 36102d5f0764SLai Jiangshan for_each_node(node) 36112d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 36122d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 36132d5f0764SLai Jiangshan 36142d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 36152d5f0764SLai Jiangshan 36162d5f0764SLai Jiangshan kfree(ctx); 36172d5f0764SLai Jiangshan } 36182d5f0764SLai Jiangshan } 36192d5f0764SLai Jiangshan 36202d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 36212d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 36222d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 36232d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 36242d5f0764SLai Jiangshan { 36252d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 36262d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 36272d5f0764SLai Jiangshan int node; 36282d5f0764SLai Jiangshan 36292d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36302d5f0764SLai Jiangshan 36312d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 36322d5f0764SLai Jiangshan GFP_KERNEL); 36332d5f0764SLai Jiangshan 36342d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36352d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 36362d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 36372d5f0764SLai Jiangshan goto out_free; 36382d5f0764SLai Jiangshan 3639042f7df1SLai Jiangshan /* 3640042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3641042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3642042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3643042f7df1SLai Jiangshan */ 36442d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3645b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3646042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3647042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 36482d5f0764SLai Jiangshan 36492d5f0764SLai Jiangshan /* 36502d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 36512d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 36522d5f0764SLai Jiangshan * pools. 36532d5f0764SLai Jiangshan */ 36542d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 36552d5f0764SLai Jiangshan 36562d5f0764SLai Jiangshan /* 36572d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 36582d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 36592d5f0764SLai Jiangshan * it even if we don't use it immediately. 36602d5f0764SLai Jiangshan */ 36612d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 36622d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 36632d5f0764SLai Jiangshan goto out_free; 36642d5f0764SLai Jiangshan 36652d5f0764SLai Jiangshan for_each_node(node) { 3666042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 36672d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 36682d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 36692d5f0764SLai Jiangshan goto out_free; 36702d5f0764SLai Jiangshan } else { 36712d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 36722d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 36732d5f0764SLai Jiangshan } 36742d5f0764SLai Jiangshan } 36752d5f0764SLai Jiangshan 3676042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3677042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3678042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 36792d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3680042f7df1SLai Jiangshan 36812d5f0764SLai Jiangshan ctx->wq = wq; 36822d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36832d5f0764SLai Jiangshan return ctx; 36842d5f0764SLai Jiangshan 36852d5f0764SLai Jiangshan out_free: 36862d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36872d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 36882d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 36892d5f0764SLai Jiangshan return NULL; 36902d5f0764SLai Jiangshan } 36912d5f0764SLai Jiangshan 36922d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 36932d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 36942d5f0764SLai Jiangshan { 36952d5f0764SLai Jiangshan int node; 36962d5f0764SLai Jiangshan 36972d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 36982d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 36992d5f0764SLai Jiangshan 37002d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 37012d5f0764SLai Jiangshan 37022d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 37032d5f0764SLai Jiangshan for_each_node(node) 37042d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 37052d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 37062d5f0764SLai Jiangshan 37072d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 37082d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 37092d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 37102d5f0764SLai Jiangshan 37112d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 37122d5f0764SLai Jiangshan } 37132d5f0764SLai Jiangshan 3714a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3715a0111cf6SLai Jiangshan { 3716a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3717a0111cf6SLai Jiangshan get_online_cpus(); 3718a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3719a0111cf6SLai Jiangshan } 3720a0111cf6SLai Jiangshan 3721a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3722a0111cf6SLai Jiangshan { 3723a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3724a0111cf6SLai Jiangshan put_online_cpus(); 3725a0111cf6SLai Jiangshan } 3726a0111cf6SLai Jiangshan 3727a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3728a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3729a0111cf6SLai Jiangshan { 3730a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3731a0111cf6SLai Jiangshan 3732a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3733a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3734a0111cf6SLai Jiangshan return -EINVAL; 3735a0111cf6SLai Jiangshan 3736a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 3737a0111cf6SLai Jiangshan if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 3738a0111cf6SLai Jiangshan return -EINVAL; 3739a0111cf6SLai Jiangshan 3740a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 37416201171eSwanghaibin if (!ctx) 37426201171eSwanghaibin return -ENOMEM; 3743a0111cf6SLai Jiangshan 3744a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3745a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3746a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3747a0111cf6SLai Jiangshan 37486201171eSwanghaibin return 0; 3749a0111cf6SLai Jiangshan } 3750a0111cf6SLai Jiangshan 37519e8cd2f5STejun Heo /** 37529e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37539e8cd2f5STejun Heo * @wq: the target workqueue 37549e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37559e8cd2f5STejun Heo * 37564c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 37574c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 37584c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 37594c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 37604c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 37614c16bd32STejun Heo * back-to-back will stay on its current pwq. 37629e8cd2f5STejun Heo * 3763d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3764d185af30SYacine Belkadi * 3765d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 37669e8cd2f5STejun Heo */ 37679e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 37689e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 37699e8cd2f5STejun Heo { 3770a0111cf6SLai Jiangshan int ret; 37719e8cd2f5STejun Heo 3772a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3773a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3774a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 37752d5f0764SLai Jiangshan 37762d5f0764SLai Jiangshan return ret; 37779e8cd2f5STejun Heo } 37789e8cd2f5STejun Heo 37794c16bd32STejun Heo /** 37804c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 37814c16bd32STejun Heo * @wq: the target workqueue 37824c16bd32STejun Heo * @cpu: the CPU coming up or going down 37834c16bd32STejun Heo * @online: whether @cpu is coming up or going down 37844c16bd32STejun Heo * 37854c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 37864c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 37874c16bd32STejun Heo * @wq accordingly. 37884c16bd32STejun Heo * 37894c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 37904c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 37914c16bd32STejun Heo * correct. 37924c16bd32STejun Heo * 37934c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 37944c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 37954c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 37964c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 37974c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 37984c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 37994c16bd32STejun Heo * CPU_DOWN_PREPARE. 38004c16bd32STejun Heo */ 38014c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 38024c16bd32STejun Heo bool online) 38034c16bd32STejun Heo { 38044c16bd32STejun Heo int node = cpu_to_node(cpu); 38054c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 38064c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 38074c16bd32STejun Heo struct workqueue_attrs *target_attrs; 38084c16bd32STejun Heo cpumask_t *cpumask; 38094c16bd32STejun Heo 38104c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38114c16bd32STejun Heo 3812f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3813f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 38144c16bd32STejun Heo return; 38154c16bd32STejun Heo 38164c16bd32STejun Heo /* 38174c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 38184c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 38194c16bd32STejun Heo * CPU hotplug exclusion. 38204c16bd32STejun Heo */ 38214c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 38224c16bd32STejun Heo cpumask = target_attrs->cpumask; 38234c16bd32STejun Heo 38244c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 38254c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 38264c16bd32STejun Heo 38274c16bd32STejun Heo /* 38284c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3829042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3830042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3831042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 38324c16bd32STejun Heo */ 3833042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 38344c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3835f7142ed4SLai Jiangshan return; 38364c16bd32STejun Heo } else { 38374c16bd32STejun Heo goto use_dfl_pwq; 38384c16bd32STejun Heo } 38394c16bd32STejun Heo 38404c16bd32STejun Heo /* create a new pwq */ 38414c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 38424c16bd32STejun Heo if (!pwq) { 38432d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 38444c16bd32STejun Heo wq->name); 384577f300b1SDaeseok Youn goto use_dfl_pwq; 38464c16bd32STejun Heo } 38474c16bd32STejun Heo 3848f7142ed4SLai Jiangshan /* Install the new pwq. */ 38494c16bd32STejun Heo mutex_lock(&wq->mutex); 38504c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 38514c16bd32STejun Heo goto out_unlock; 38524c16bd32STejun Heo 38534c16bd32STejun Heo use_dfl_pwq: 3854f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 38554c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 38564c16bd32STejun Heo get_pwq(wq->dfl_pwq); 38574c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 38584c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 38594c16bd32STejun Heo out_unlock: 38604c16bd32STejun Heo mutex_unlock(&wq->mutex); 38614c16bd32STejun Heo put_pwq_unlocked(old_pwq); 38624c16bd32STejun Heo } 38634c16bd32STejun Heo 386430cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 38651da177e4SLinus Torvalds { 386649e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 38678a2b7538STejun Heo int cpu, ret; 3868e1d8aa9fSFrederic Weisbecker 386930cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3870420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3871420c0ddbSTejun Heo if (!wq->cpu_pwqs) 387230cdf249STejun Heo return -ENOMEM; 387330cdf249STejun Heo 387430cdf249STejun Heo for_each_possible_cpu(cpu) { 38757fb98ea7STejun Heo struct pool_workqueue *pwq = 38767fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 38777a62c2c8STejun Heo struct worker_pool *cpu_pools = 3878f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 387930cdf249STejun Heo 3880f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3881f147f29eSTejun Heo 3882f147f29eSTejun Heo mutex_lock(&wq->mutex); 38831befcf30STejun Heo link_pwq(pwq); 3884f147f29eSTejun Heo mutex_unlock(&wq->mutex); 388530cdf249STejun Heo } 388630cdf249STejun Heo return 0; 38878a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 38888a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 38898a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 38908a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 38918a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 38928a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 38938a2b7538STejun Heo return ret; 38949e8cd2f5STejun Heo } else { 38959e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 38969e8cd2f5STejun Heo } 38970f900049STejun Heo } 38980f900049STejun Heo 3899f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3900f3421797STejun Heo const char *name) 3901b71ab8c2STejun Heo { 3902f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3903f3421797STejun Heo 3904f3421797STejun Heo if (max_active < 1 || max_active > lim) 3905044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3906f3421797STejun Heo max_active, name, 1, lim); 3907b71ab8c2STejun Heo 3908f3421797STejun Heo return clamp_val(max_active, 1, lim); 3909b71ab8c2STejun Heo } 3910b71ab8c2STejun Heo 3911b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 391297e37d7bSTejun Heo unsigned int flags, 39131e19ffc6STejun Heo int max_active, 3914eb13ba87SJohannes Berg struct lock_class_key *key, 3915b196be89STejun Heo const char *lock_name, ...) 39163af24433SOleg Nesterov { 3917df2d5ae4STejun Heo size_t tbl_size = 0; 3918ecf6881fSTejun Heo va_list args; 39193af24433SOleg Nesterov struct workqueue_struct *wq; 392049e3cf44STejun Heo struct pool_workqueue *pwq; 3921b196be89STejun Heo 3922cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 3923cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 3924cee22a15SViresh Kumar flags |= WQ_UNBOUND; 3925cee22a15SViresh Kumar 3926ecf6881fSTejun Heo /* allocate wq and format name */ 3927df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 3928ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 3929df2d5ae4STejun Heo 3930df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 3931b196be89STejun Heo if (!wq) 3932d2c1d404STejun Heo return NULL; 3933b196be89STejun Heo 39346029a918STejun Heo if (flags & WQ_UNBOUND) { 39356029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 39366029a918STejun Heo if (!wq->unbound_attrs) 39376029a918STejun Heo goto err_free_wq; 39386029a918STejun Heo } 39396029a918STejun Heo 3940ecf6881fSTejun Heo va_start(args, lock_name); 3941ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3942b196be89STejun Heo va_end(args); 39433af24433SOleg Nesterov 3944d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3945b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 39463af24433SOleg Nesterov 3947b196be89STejun Heo /* init wq */ 394897e37d7bSTejun Heo wq->flags = flags; 3949a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 39503c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3951112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 395230cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 395373f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 395473f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3955493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 39563af24433SOleg Nesterov 3957eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3958cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 39593af24433SOleg Nesterov 396030cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3961d2c1d404STejun Heo goto err_free_wq; 39621537663fSTejun Heo 3963493008a8STejun Heo /* 3964493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3965493008a8STejun Heo * have a rescuer to guarantee forward progress. 3966493008a8STejun Heo */ 3967493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3968e22bee78STejun Heo struct worker *rescuer; 3969e22bee78STejun Heo 3970f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 3971e22bee78STejun Heo if (!rescuer) 3972d2c1d404STejun Heo goto err_destroy; 3973e22bee78STejun Heo 3974111c225aSTejun Heo rescuer->rescue_wq = wq; 3975111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3976b196be89STejun Heo wq->name); 3977d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3978d2c1d404STejun Heo kfree(rescuer); 3979d2c1d404STejun Heo goto err_destroy; 3980d2c1d404STejun Heo } 3981e22bee78STejun Heo 3982d2c1d404STejun Heo wq->rescuer = rescuer; 398325834c73SPeter Zijlstra kthread_bind_mask(rescuer->task, cpu_possible_mask); 3984e22bee78STejun Heo wake_up_process(rescuer->task); 39853af24433SOleg Nesterov } 39861537663fSTejun Heo 3987226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3988226223abSTejun Heo goto err_destroy; 3989226223abSTejun Heo 39903af24433SOleg Nesterov /* 399168e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 399268e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 399368e13a67SLai Jiangshan * list. 39943af24433SOleg Nesterov */ 399568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3996a0a1a5fdSTejun Heo 3997a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 399849e3cf44STejun Heo for_each_pwq(pwq, wq) 3999699ce097STejun Heo pwq_adjust_max_active(pwq); 4000a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4001a0a1a5fdSTejun Heo 4002e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4003a0a1a5fdSTejun Heo 400468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 40053af24433SOleg Nesterov 40063af24433SOleg Nesterov return wq; 4007d2c1d404STejun Heo 4008d2c1d404STejun Heo err_free_wq: 40096029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 40104690c4abSTejun Heo kfree(wq); 4011d2c1d404STejun Heo return NULL; 4012d2c1d404STejun Heo err_destroy: 4013d2c1d404STejun Heo destroy_workqueue(wq); 40144690c4abSTejun Heo return NULL; 40151da177e4SLinus Torvalds } 4016d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 40171da177e4SLinus Torvalds 40183af24433SOleg Nesterov /** 40193af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 40203af24433SOleg Nesterov * @wq: target workqueue 40213af24433SOleg Nesterov * 40223af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 40233af24433SOleg Nesterov */ 40243af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 40253af24433SOleg Nesterov { 402649e3cf44STejun Heo struct pool_workqueue *pwq; 40274c16bd32STejun Heo int node; 40283af24433SOleg Nesterov 40299c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 40309c5a2ba7STejun Heo drain_workqueue(wq); 4031c8efcc25STejun Heo 40326183c009STejun Heo /* sanity checks */ 4033b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 403449e3cf44STejun Heo for_each_pwq(pwq, wq) { 40356183c009STejun Heo int i; 40366183c009STejun Heo 403776af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 403876af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4039b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4040fa07fb6aSTejun Heo show_workqueue_state(); 40416183c009STejun Heo return; 404276af4d93STejun Heo } 404376af4d93STejun Heo } 404476af4d93STejun Heo 40455c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 40468864b4e5STejun Heo WARN_ON(pwq->nr_active) || 404776af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4048b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4049fa07fb6aSTejun Heo show_workqueue_state(); 40506183c009STejun Heo return; 40516183c009STejun Heo } 405276af4d93STejun Heo } 4053b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 40546183c009STejun Heo 4055a0a1a5fdSTejun Heo /* 4056a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4057a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4058a0a1a5fdSTejun Heo */ 405968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4060e2dca7adSTejun Heo list_del_rcu(&wq->list); 406168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 40623af24433SOleg Nesterov 4063226223abSTejun Heo workqueue_sysfs_unregister(wq); 4064226223abSTejun Heo 4065e2dca7adSTejun Heo if (wq->rescuer) 4066e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4067e22bee78STejun Heo 40688864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 406929c91e99STejun Heo /* 40708864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4071e2dca7adSTejun Heo * schedule RCU free. 407229c91e99STejun Heo */ 4073e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 40748864b4e5STejun Heo } else { 40758864b4e5STejun Heo /* 40768864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 40774c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 40784c16bd32STejun Heo * @wq will be freed when the last pwq is released. 40798864b4e5STejun Heo */ 40804c16bd32STejun Heo for_each_node(node) { 40814c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 40824c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 40834c16bd32STejun Heo put_pwq_unlocked(pwq); 40844c16bd32STejun Heo } 40854c16bd32STejun Heo 40864c16bd32STejun Heo /* 40874c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 40884c16bd32STejun Heo * put. Don't access it afterwards. 40894c16bd32STejun Heo */ 40904c16bd32STejun Heo pwq = wq->dfl_pwq; 40914c16bd32STejun Heo wq->dfl_pwq = NULL; 4092dce90d47STejun Heo put_pwq_unlocked(pwq); 409329c91e99STejun Heo } 40943af24433SOleg Nesterov } 40953af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 40963af24433SOleg Nesterov 4097dcd989cbSTejun Heo /** 4098dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4099dcd989cbSTejun Heo * @wq: target workqueue 4100dcd989cbSTejun Heo * @max_active: new max_active value. 4101dcd989cbSTejun Heo * 4102dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4103dcd989cbSTejun Heo * 4104dcd989cbSTejun Heo * CONTEXT: 4105dcd989cbSTejun Heo * Don't call from IRQ context. 4106dcd989cbSTejun Heo */ 4107dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4108dcd989cbSTejun Heo { 410949e3cf44STejun Heo struct pool_workqueue *pwq; 4110dcd989cbSTejun Heo 41118719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 41128719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 41138719dceaSTejun Heo return; 41148719dceaSTejun Heo 4115f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4116dcd989cbSTejun Heo 4117a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4118dcd989cbSTejun Heo 4119dcd989cbSTejun Heo wq->saved_max_active = max_active; 4120dcd989cbSTejun Heo 4121699ce097STejun Heo for_each_pwq(pwq, wq) 4122699ce097STejun Heo pwq_adjust_max_active(pwq); 4123dcd989cbSTejun Heo 4124a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4125dcd989cbSTejun Heo } 4126dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4127dcd989cbSTejun Heo 4128dcd989cbSTejun Heo /** 4129e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4130e6267616STejun Heo * 4131e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4132e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4133d185af30SYacine Belkadi * 4134d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4135e6267616STejun Heo */ 4136e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4137e6267616STejun Heo { 4138e6267616STejun Heo struct worker *worker = current_wq_worker(); 4139e6267616STejun Heo 41406a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4141e6267616STejun Heo } 4142e6267616STejun Heo 4143e6267616STejun Heo /** 4144dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4145dcd989cbSTejun Heo * @cpu: CPU in question 4146dcd989cbSTejun Heo * @wq: target workqueue 4147dcd989cbSTejun Heo * 4148dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4149dcd989cbSTejun Heo * no synchronization around this function and the test result is 4150dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4151dcd989cbSTejun Heo * 4152d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4153d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4154d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4155d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4156d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4157d3251859STejun Heo * 4158d185af30SYacine Belkadi * Return: 4159dcd989cbSTejun Heo * %true if congested, %false otherwise. 4160dcd989cbSTejun Heo */ 4161d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4162dcd989cbSTejun Heo { 41637fb98ea7STejun Heo struct pool_workqueue *pwq; 416476af4d93STejun Heo bool ret; 416576af4d93STejun Heo 416688109453SLai Jiangshan rcu_read_lock_sched(); 41677fb98ea7STejun Heo 4168d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4169d3251859STejun Heo cpu = smp_processor_id(); 4170d3251859STejun Heo 41717fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 41727fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 41737fb98ea7STejun Heo else 4174df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4175dcd989cbSTejun Heo 417676af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 417788109453SLai Jiangshan rcu_read_unlock_sched(); 417876af4d93STejun Heo 417976af4d93STejun Heo return ret; 4180dcd989cbSTejun Heo } 4181dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4182dcd989cbSTejun Heo 4183dcd989cbSTejun Heo /** 4184dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4185dcd989cbSTejun Heo * @work: the work to be tested 4186dcd989cbSTejun Heo * 4187dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4188dcd989cbSTejun Heo * synchronization around this function and the test result is 4189dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4190dcd989cbSTejun Heo * 4191d185af30SYacine Belkadi * Return: 4192dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4193dcd989cbSTejun Heo */ 4194dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4195dcd989cbSTejun Heo { 4196fa1b54e6STejun Heo struct worker_pool *pool; 4197dcd989cbSTejun Heo unsigned long flags; 4198dcd989cbSTejun Heo unsigned int ret = 0; 4199dcd989cbSTejun Heo 4200dcd989cbSTejun Heo if (work_pending(work)) 4201dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4202038366c5SLai Jiangshan 4203fa1b54e6STejun Heo local_irq_save(flags); 4204fa1b54e6STejun Heo pool = get_work_pool(work); 4205038366c5SLai Jiangshan if (pool) { 4206fa1b54e6STejun Heo spin_lock(&pool->lock); 4207c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4208dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4209fa1b54e6STejun Heo spin_unlock(&pool->lock); 4210038366c5SLai Jiangshan } 4211fa1b54e6STejun Heo local_irq_restore(flags); 4212dcd989cbSTejun Heo 4213dcd989cbSTejun Heo return ret; 4214dcd989cbSTejun Heo } 4215dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4216dcd989cbSTejun Heo 42173d1cb205STejun Heo /** 42183d1cb205STejun Heo * set_worker_desc - set description for the current work item 42193d1cb205STejun Heo * @fmt: printf-style format string 42203d1cb205STejun Heo * @...: arguments for the format string 42213d1cb205STejun Heo * 42223d1cb205STejun Heo * This function can be called by a running work function to describe what 42233d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 42243d1cb205STejun Heo * information will be printed out together to help debugging. The 42253d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 42263d1cb205STejun Heo */ 42273d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 42283d1cb205STejun Heo { 42293d1cb205STejun Heo struct worker *worker = current_wq_worker(); 42303d1cb205STejun Heo va_list args; 42313d1cb205STejun Heo 42323d1cb205STejun Heo if (worker) { 42333d1cb205STejun Heo va_start(args, fmt); 42343d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 42353d1cb205STejun Heo va_end(args); 42363d1cb205STejun Heo worker->desc_valid = true; 42373d1cb205STejun Heo } 42383d1cb205STejun Heo } 42393d1cb205STejun Heo 42403d1cb205STejun Heo /** 42413d1cb205STejun Heo * print_worker_info - print out worker information and description 42423d1cb205STejun Heo * @log_lvl: the log level to use when printing 42433d1cb205STejun Heo * @task: target task 42443d1cb205STejun Heo * 42453d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 42463d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 42473d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 42483d1cb205STejun Heo * 42493d1cb205STejun Heo * This function can be safely called on any task as long as the 42503d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 42513d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 42523d1cb205STejun Heo */ 42533d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 42543d1cb205STejun Heo { 42553d1cb205STejun Heo work_func_t *fn = NULL; 42563d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 42573d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 42583d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 42593d1cb205STejun Heo struct workqueue_struct *wq = NULL; 42603d1cb205STejun Heo bool desc_valid = false; 42613d1cb205STejun Heo struct worker *worker; 42623d1cb205STejun Heo 42633d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 42643d1cb205STejun Heo return; 42653d1cb205STejun Heo 42663d1cb205STejun Heo /* 42673d1cb205STejun Heo * This function is called without any synchronization and @task 42683d1cb205STejun Heo * could be in any state. Be careful with dereferences. 42693d1cb205STejun Heo */ 42703d1cb205STejun Heo worker = probe_kthread_data(task); 42713d1cb205STejun Heo 42723d1cb205STejun Heo /* 42733d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 42743d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 42753d1cb205STejun Heo */ 42763d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 42773d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 42783d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 42793d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 42803d1cb205STejun Heo 42813d1cb205STejun Heo /* copy worker description */ 42823d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 42833d1cb205STejun Heo if (desc_valid) 42843d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 42853d1cb205STejun Heo 42863d1cb205STejun Heo if (fn || name[0] || desc[0]) { 42873d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 42883d1cb205STejun Heo if (desc[0]) 42893d1cb205STejun Heo pr_cont(" (%s)", desc); 42903d1cb205STejun Heo pr_cont("\n"); 42913d1cb205STejun Heo } 42923d1cb205STejun Heo } 42933d1cb205STejun Heo 42943494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 42953494fc30STejun Heo { 42963494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 42973494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 42983494fc30STejun Heo pr_cont(" node=%d", pool->node); 42993494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 43003494fc30STejun Heo } 43013494fc30STejun Heo 43023494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 43033494fc30STejun Heo { 43043494fc30STejun Heo if (work->func == wq_barrier_func) { 43053494fc30STejun Heo struct wq_barrier *barr; 43063494fc30STejun Heo 43073494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 43083494fc30STejun Heo 43093494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 43103494fc30STejun Heo task_pid_nr(barr->task)); 43113494fc30STejun Heo } else { 43123494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 43133494fc30STejun Heo } 43143494fc30STejun Heo } 43153494fc30STejun Heo 43163494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 43173494fc30STejun Heo { 43183494fc30STejun Heo struct worker_pool *pool = pwq->pool; 43193494fc30STejun Heo struct work_struct *work; 43203494fc30STejun Heo struct worker *worker; 43213494fc30STejun Heo bool has_in_flight = false, has_pending = false; 43223494fc30STejun Heo int bkt; 43233494fc30STejun Heo 43243494fc30STejun Heo pr_info(" pwq %d:", pool->id); 43253494fc30STejun Heo pr_cont_pool_info(pool); 43263494fc30STejun Heo 43273494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 43283494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 43293494fc30STejun Heo 43303494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43313494fc30STejun Heo if (worker->current_pwq == pwq) { 43323494fc30STejun Heo has_in_flight = true; 43333494fc30STejun Heo break; 43343494fc30STejun Heo } 43353494fc30STejun Heo } 43363494fc30STejun Heo if (has_in_flight) { 43373494fc30STejun Heo bool comma = false; 43383494fc30STejun Heo 43393494fc30STejun Heo pr_info(" in-flight:"); 43403494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 43413494fc30STejun Heo if (worker->current_pwq != pwq) 43423494fc30STejun Heo continue; 43433494fc30STejun Heo 43443494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 43453494fc30STejun Heo task_pid_nr(worker->task), 43463494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 43473494fc30STejun Heo worker->current_func); 43483494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 43493494fc30STejun Heo pr_cont_work(false, work); 43503494fc30STejun Heo comma = true; 43513494fc30STejun Heo } 43523494fc30STejun Heo pr_cont("\n"); 43533494fc30STejun Heo } 43543494fc30STejun Heo 43553494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 43563494fc30STejun Heo if (get_work_pwq(work) == pwq) { 43573494fc30STejun Heo has_pending = true; 43583494fc30STejun Heo break; 43593494fc30STejun Heo } 43603494fc30STejun Heo } 43613494fc30STejun Heo if (has_pending) { 43623494fc30STejun Heo bool comma = false; 43633494fc30STejun Heo 43643494fc30STejun Heo pr_info(" pending:"); 43653494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 43663494fc30STejun Heo if (get_work_pwq(work) != pwq) 43673494fc30STejun Heo continue; 43683494fc30STejun Heo 43693494fc30STejun Heo pr_cont_work(comma, work); 43703494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 43713494fc30STejun Heo } 43723494fc30STejun Heo pr_cont("\n"); 43733494fc30STejun Heo } 43743494fc30STejun Heo 43753494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 43763494fc30STejun Heo bool comma = false; 43773494fc30STejun Heo 43783494fc30STejun Heo pr_info(" delayed:"); 43793494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 43803494fc30STejun Heo pr_cont_work(comma, work); 43813494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 43823494fc30STejun Heo } 43833494fc30STejun Heo pr_cont("\n"); 43843494fc30STejun Heo } 43853494fc30STejun Heo } 43863494fc30STejun Heo 43873494fc30STejun Heo /** 43883494fc30STejun Heo * show_workqueue_state - dump workqueue state 43893494fc30STejun Heo * 43907b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 43917b776af6SRoger Lu * all busy workqueues and pools. 43923494fc30STejun Heo */ 43933494fc30STejun Heo void show_workqueue_state(void) 43943494fc30STejun Heo { 43953494fc30STejun Heo struct workqueue_struct *wq; 43963494fc30STejun Heo struct worker_pool *pool; 43973494fc30STejun Heo unsigned long flags; 43983494fc30STejun Heo int pi; 43993494fc30STejun Heo 44003494fc30STejun Heo rcu_read_lock_sched(); 44013494fc30STejun Heo 44023494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 44033494fc30STejun Heo 44043494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 44053494fc30STejun Heo struct pool_workqueue *pwq; 44063494fc30STejun Heo bool idle = true; 44073494fc30STejun Heo 44083494fc30STejun Heo for_each_pwq(pwq, wq) { 44093494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 44103494fc30STejun Heo idle = false; 44113494fc30STejun Heo break; 44123494fc30STejun Heo } 44133494fc30STejun Heo } 44143494fc30STejun Heo if (idle) 44153494fc30STejun Heo continue; 44163494fc30STejun Heo 44173494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 44183494fc30STejun Heo 44193494fc30STejun Heo for_each_pwq(pwq, wq) { 44203494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 44213494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 44223494fc30STejun Heo show_pwq(pwq); 44233494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 44243494fc30STejun Heo } 44253494fc30STejun Heo } 44263494fc30STejun Heo 44273494fc30STejun Heo for_each_pool(pool, pi) { 44283494fc30STejun Heo struct worker *worker; 44293494fc30STejun Heo bool first = true; 44303494fc30STejun Heo 44313494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 44323494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 44333494fc30STejun Heo goto next_pool; 44343494fc30STejun Heo 44353494fc30STejun Heo pr_info("pool %d:", pool->id); 44363494fc30STejun Heo pr_cont_pool_info(pool); 443782607adcSTejun Heo pr_cont(" hung=%us workers=%d", 443882607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 443982607adcSTejun Heo pool->nr_workers); 44403494fc30STejun Heo if (pool->manager) 44413494fc30STejun Heo pr_cont(" manager: %d", 44423494fc30STejun Heo task_pid_nr(pool->manager->task)); 44433494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 44443494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 44453494fc30STejun Heo task_pid_nr(worker->task)); 44463494fc30STejun Heo first = false; 44473494fc30STejun Heo } 44483494fc30STejun Heo pr_cont("\n"); 44493494fc30STejun Heo next_pool: 44503494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 44513494fc30STejun Heo } 44523494fc30STejun Heo 44533494fc30STejun Heo rcu_read_unlock_sched(); 44543494fc30STejun Heo } 44553494fc30STejun Heo 4456db7bccf4STejun Heo /* 4457db7bccf4STejun Heo * CPU hotplug. 4458db7bccf4STejun Heo * 4459e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4460112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4461706026c2STejun Heo * pool which make migrating pending and scheduled works very 4462e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 446394cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4464e22bee78STejun Heo * blocked draining impractical. 4465e22bee78STejun Heo * 446624647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4467628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4468628c78e7STejun Heo * cpu comes back online. 4469db7bccf4STejun Heo */ 4470db7bccf4STejun Heo 4471706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4472db7bccf4STejun Heo { 447338db41d9STejun Heo int cpu = smp_processor_id(); 44744ce62e9eSTejun Heo struct worker_pool *pool; 4475db7bccf4STejun Heo struct worker *worker; 4476db7bccf4STejun Heo 4477f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 447892f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 447994cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4480e22bee78STejun Heo 4481f2d5a0eeSTejun Heo /* 448292f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 448394cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 448494cf58bbSTejun Heo * except for the ones which are still executing works from 448594cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 448694cf58bbSTejun Heo * this, they may become diasporas. 4487f2d5a0eeSTejun Heo */ 4488da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4489403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4490db7bccf4STejun Heo 449124647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4492f2d5a0eeSTejun Heo 449394cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 449492f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4495e22bee78STejun Heo 4496e22bee78STejun Heo /* 4497eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4498eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4499eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4500eb283428SLai Jiangshan * from other cpus. 4501628c78e7STejun Heo */ 4502628c78e7STejun Heo schedule(); 4503628c78e7STejun Heo 4504628c78e7STejun Heo /* 4505eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4506eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4507eb283428SLai Jiangshan * and keep_working() are always true as long as the 4508eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4509eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4510eb283428SLai Jiangshan * are served by workers tied to the pool. 4511e22bee78STejun Heo */ 4512e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4513eb283428SLai Jiangshan 4514eb283428SLai Jiangshan /* 4515eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4516eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4517eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4518eb283428SLai Jiangshan */ 4519eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4520eb283428SLai Jiangshan wake_up_worker(pool); 4521eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4522eb283428SLai Jiangshan } 4523db7bccf4STejun Heo } 4524db7bccf4STejun Heo 4525bd7c089eSTejun Heo /** 4526bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4527bd7c089eSTejun Heo * @pool: pool of interest 4528bd7c089eSTejun Heo * 4529a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4530bd7c089eSTejun Heo */ 4531bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4532bd7c089eSTejun Heo { 4533a9ab775bSTejun Heo struct worker *worker; 4534bd7c089eSTejun Heo 453592f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4536bd7c089eSTejun Heo 4537bd7c089eSTejun Heo /* 4538a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4539a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4540402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4541a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4542a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4543bd7c089eSTejun Heo */ 4544da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4545a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4546a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4547a9ab775bSTejun Heo 4548a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4549f7c17d26SWanpeng Li 4550f7c17d26SWanpeng Li /* 4551f7c17d26SWanpeng Li * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED 4552f7c17d26SWanpeng Li * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is 4553f7c17d26SWanpeng Li * being reworked and this can go away in time. 4554f7c17d26SWanpeng Li */ 4555f7c17d26SWanpeng Li if (!(pool->flags & POOL_DISASSOCIATED)) { 4556f7c17d26SWanpeng Li spin_unlock_irq(&pool->lock); 4557f7c17d26SWanpeng Li return; 4558f7c17d26SWanpeng Li } 4559f7c17d26SWanpeng Li 45603de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4561a9ab775bSTejun Heo 4562da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4563a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4564a9ab775bSTejun Heo 4565a9ab775bSTejun Heo /* 4566a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4567a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4568a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4569a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4570a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4571a9ab775bSTejun Heo * be bound before @pool->lock is released. 4572a9ab775bSTejun Heo */ 4573a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4574bd7c089eSTejun Heo wake_up_process(worker->task); 4575bd7c089eSTejun Heo 4576bd7c089eSTejun Heo /* 4577a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4578a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4579a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4580a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4581a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4582a9ab775bSTejun Heo * concurrency management. Note that when or whether 4583a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4584a9ab775bSTejun Heo * 4585a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4586a9ab775bSTejun Heo * tested without holding any lock in 4587a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4588a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4589a9ab775bSTejun Heo * management operations. 4590bd7c089eSTejun Heo */ 4591a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4592a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4593a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4594a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4595bd7c089eSTejun Heo } 4596a9ab775bSTejun Heo 4597a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4598bd7c089eSTejun Heo } 4599bd7c089eSTejun Heo 46007dbc725eSTejun Heo /** 46017dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 46027dbc725eSTejun Heo * @pool: unbound pool of interest 46037dbc725eSTejun Heo * @cpu: the CPU which is coming up 46047dbc725eSTejun Heo * 46057dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 46067dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 46077dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 46087dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 46097dbc725eSTejun Heo */ 46107dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 46117dbc725eSTejun Heo { 46127dbc725eSTejun Heo static cpumask_t cpumask; 46137dbc725eSTejun Heo struct worker *worker; 46147dbc725eSTejun Heo 461592f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 46167dbc725eSTejun Heo 46177dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 46187dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 46197dbc725eSTejun Heo return; 46207dbc725eSTejun Heo 46217dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 46227dbc725eSTejun Heo 46237dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4624da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4625d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 46267dbc725eSTejun Heo } 46277dbc725eSTejun Heo 46287ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 46291da177e4SLinus Torvalds { 46304ce62e9eSTejun Heo struct worker_pool *pool; 46311da177e4SLinus Torvalds 4632f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46333ce63377STejun Heo if (pool->nr_workers) 46343ce63377STejun Heo continue; 4635051e1850SLai Jiangshan if (!create_worker(pool)) 46367ee681b2SThomas Gleixner return -ENOMEM; 46373af24433SOleg Nesterov } 46387ee681b2SThomas Gleixner return 0; 46397ee681b2SThomas Gleixner } 46401da177e4SLinus Torvalds 46417ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 46427ee681b2SThomas Gleixner { 46437ee681b2SThomas Gleixner struct worker_pool *pool; 46447ee681b2SThomas Gleixner struct workqueue_struct *wq; 46457ee681b2SThomas Gleixner int pi; 46467ee681b2SThomas Gleixner 464768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46487dbc725eSTejun Heo 46497dbc725eSTejun Heo for_each_pool(pool, pi) { 465092f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 465194cf58bbSTejun Heo 4652f05b558dSLai Jiangshan if (pool->cpu == cpu) 465394cf58bbSTejun Heo rebind_workers(pool); 4654f05b558dSLai Jiangshan else if (pool->cpu < 0) 46557dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 465694cf58bbSTejun Heo 465792f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 465894cf58bbSTejun Heo } 46597dbc725eSTejun Heo 46604c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46614c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46624c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 46634c16bd32STejun Heo 466468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46657ee681b2SThomas Gleixner return 0; 466665758202STejun Heo } 466765758202STejun Heo 46687ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 466965758202STejun Heo { 46708db25e78STejun Heo struct work_struct unbind_work; 46714c16bd32STejun Heo struct workqueue_struct *wq; 46728db25e78STejun Heo 46734c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4674706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46757635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46764c16bd32STejun Heo 46774c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46784c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46794c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46804c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 46814c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 46824c16bd32STejun Heo 46834c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 46848db25e78STejun Heo flush_work(&unbind_work); 4685440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 46867ee681b2SThomas Gleixner return 0; 468765758202STejun Heo } 468865758202STejun Heo 46892d3854a3SRusty Russell #ifdef CONFIG_SMP 46908ccad40dSRusty Russell 46912d3854a3SRusty Russell struct work_for_cpu { 4692ed48ece2STejun Heo struct work_struct work; 46932d3854a3SRusty Russell long (*fn)(void *); 46942d3854a3SRusty Russell void *arg; 46952d3854a3SRusty Russell long ret; 46962d3854a3SRusty Russell }; 46972d3854a3SRusty Russell 4698ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 46992d3854a3SRusty Russell { 4700ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4701ed48ece2STejun Heo 47022d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 47032d3854a3SRusty Russell } 47042d3854a3SRusty Russell 47052d3854a3SRusty Russell /** 470622aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 47072d3854a3SRusty Russell * @cpu: the cpu to run on 47082d3854a3SRusty Russell * @fn: the function to run 47092d3854a3SRusty Russell * @arg: the function arg 47102d3854a3SRusty Russell * 471131ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47126b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4713d185af30SYacine Belkadi * 4714d185af30SYacine Belkadi * Return: The value @fn returns. 47152d3854a3SRusty Russell */ 4716d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47172d3854a3SRusty Russell { 4718ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47192d3854a3SRusty Russell 4720ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4721ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 472212997d1aSBjorn Helgaas flush_work(&wfc.work); 4723440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 47242d3854a3SRusty Russell return wfc.ret; 47252d3854a3SRusty Russell } 47262d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47272d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47282d3854a3SRusty Russell 4729a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4730e7577c50SRusty Russell 4731a0a1a5fdSTejun Heo /** 4732a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4733a0a1a5fdSTejun Heo * 473458a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4735c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4736706026c2STejun Heo * pool->worklist. 4737a0a1a5fdSTejun Heo * 4738a0a1a5fdSTejun Heo * CONTEXT: 4739a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4740a0a1a5fdSTejun Heo */ 4741a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4742a0a1a5fdSTejun Heo { 474324b8a847STejun Heo struct workqueue_struct *wq; 474424b8a847STejun Heo struct pool_workqueue *pwq; 4745a0a1a5fdSTejun Heo 474668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4747a0a1a5fdSTejun Heo 47486183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4749a0a1a5fdSTejun Heo workqueue_freezing = true; 4750a0a1a5fdSTejun Heo 475124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4752a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4753699ce097STejun Heo for_each_pwq(pwq, wq) 4754699ce097STejun Heo pwq_adjust_max_active(pwq); 4755a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4756a1056305STejun Heo } 47575bcab335STejun Heo 475868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4759a0a1a5fdSTejun Heo } 4760a0a1a5fdSTejun Heo 4761a0a1a5fdSTejun Heo /** 476258a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4763a0a1a5fdSTejun Heo * 4764a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4765a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4766a0a1a5fdSTejun Heo * 4767a0a1a5fdSTejun Heo * CONTEXT: 476868e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4769a0a1a5fdSTejun Heo * 4770d185af30SYacine Belkadi * Return: 477158a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 477258a69cb4STejun Heo * is complete. 4773a0a1a5fdSTejun Heo */ 4774a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4775a0a1a5fdSTejun Heo { 4776a0a1a5fdSTejun Heo bool busy = false; 477724b8a847STejun Heo struct workqueue_struct *wq; 477824b8a847STejun Heo struct pool_workqueue *pwq; 4779a0a1a5fdSTejun Heo 478068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4781a0a1a5fdSTejun Heo 47826183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4783a0a1a5fdSTejun Heo 478424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 478524b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 478624b8a847STejun Heo continue; 4787a0a1a5fdSTejun Heo /* 4788a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4789a0a1a5fdSTejun Heo * to peek without lock. 4790a0a1a5fdSTejun Heo */ 479188109453SLai Jiangshan rcu_read_lock_sched(); 479224b8a847STejun Heo for_each_pwq(pwq, wq) { 47936183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4794112202d9STejun Heo if (pwq->nr_active) { 4795a0a1a5fdSTejun Heo busy = true; 479688109453SLai Jiangshan rcu_read_unlock_sched(); 4797a0a1a5fdSTejun Heo goto out_unlock; 4798a0a1a5fdSTejun Heo } 4799a0a1a5fdSTejun Heo } 480088109453SLai Jiangshan rcu_read_unlock_sched(); 4801a0a1a5fdSTejun Heo } 4802a0a1a5fdSTejun Heo out_unlock: 480368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4804a0a1a5fdSTejun Heo return busy; 4805a0a1a5fdSTejun Heo } 4806a0a1a5fdSTejun Heo 4807a0a1a5fdSTejun Heo /** 4808a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4809a0a1a5fdSTejun Heo * 4810a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4811706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4812a0a1a5fdSTejun Heo * 4813a0a1a5fdSTejun Heo * CONTEXT: 4814a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4815a0a1a5fdSTejun Heo */ 4816a0a1a5fdSTejun Heo void thaw_workqueues(void) 4817a0a1a5fdSTejun Heo { 481824b8a847STejun Heo struct workqueue_struct *wq; 481924b8a847STejun Heo struct pool_workqueue *pwq; 4820a0a1a5fdSTejun Heo 482168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4822a0a1a5fdSTejun Heo 4823a0a1a5fdSTejun Heo if (!workqueue_freezing) 4824a0a1a5fdSTejun Heo goto out_unlock; 4825a0a1a5fdSTejun Heo 482674b414eaSLai Jiangshan workqueue_freezing = false; 482724b8a847STejun Heo 482824b8a847STejun Heo /* restore max_active and repopulate worklist */ 482924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4830a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4831699ce097STejun Heo for_each_pwq(pwq, wq) 4832699ce097STejun Heo pwq_adjust_max_active(pwq); 4833a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 483424b8a847STejun Heo } 483524b8a847STejun Heo 4836a0a1a5fdSTejun Heo out_unlock: 483768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4838a0a1a5fdSTejun Heo } 4839a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4840a0a1a5fdSTejun Heo 4841042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4842042f7df1SLai Jiangshan { 4843042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4844042f7df1SLai Jiangshan int ret = 0; 4845042f7df1SLai Jiangshan struct workqueue_struct *wq; 4846042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4847042f7df1SLai Jiangshan 4848042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4849042f7df1SLai Jiangshan 4850042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4851042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4852042f7df1SLai Jiangshan continue; 4853042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 4854042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 4855042f7df1SLai Jiangshan continue; 4856042f7df1SLai Jiangshan 4857042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 4858042f7df1SLai Jiangshan if (!ctx) { 4859042f7df1SLai Jiangshan ret = -ENOMEM; 4860042f7df1SLai Jiangshan break; 4861042f7df1SLai Jiangshan } 4862042f7df1SLai Jiangshan 4863042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 4864042f7df1SLai Jiangshan } 4865042f7df1SLai Jiangshan 4866042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 4867042f7df1SLai Jiangshan if (!ret) 4868042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 4869042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 4870042f7df1SLai Jiangshan } 4871042f7df1SLai Jiangshan 4872042f7df1SLai Jiangshan return ret; 4873042f7df1SLai Jiangshan } 4874042f7df1SLai Jiangshan 4875042f7df1SLai Jiangshan /** 4876042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 4877042f7df1SLai Jiangshan * @cpumask: the cpumask to set 4878042f7df1SLai Jiangshan * 4879042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 4880042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 4881042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 4882042f7df1SLai Jiangshan * 4883042f7df1SLai Jiangshan * Retun: 0 - Success 4884042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 4885042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 4886042f7df1SLai Jiangshan */ 4887042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 4888042f7df1SLai Jiangshan { 4889042f7df1SLai Jiangshan int ret = -EINVAL; 4890042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 4891042f7df1SLai Jiangshan 4892042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 4893042f7df1SLai Jiangshan return -ENOMEM; 4894042f7df1SLai Jiangshan 4895042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 4896042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 4897a0111cf6SLai Jiangshan apply_wqattrs_lock(); 4898042f7df1SLai Jiangshan 4899042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 4900042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 4901042f7df1SLai Jiangshan 4902042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 4903042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 4904042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 4905042f7df1SLai Jiangshan 4906042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 4907042f7df1SLai Jiangshan if (ret < 0) 4908042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 4909042f7df1SLai Jiangshan 4910a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 4911042f7df1SLai Jiangshan } 4912042f7df1SLai Jiangshan 4913042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 4914042f7df1SLai Jiangshan return ret; 4915042f7df1SLai Jiangshan } 4916042f7df1SLai Jiangshan 49176ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 49186ba94429SFrederic Weisbecker /* 49196ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 49206ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 49216ba94429SFrederic Weisbecker * following attributes. 49226ba94429SFrederic Weisbecker * 49236ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 49246ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 49256ba94429SFrederic Weisbecker * 49266ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 49276ba94429SFrederic Weisbecker * 49286ba94429SFrederic Weisbecker * id RO int : the associated pool ID 49296ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 49306ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 49316ba94429SFrederic Weisbecker */ 49326ba94429SFrederic Weisbecker struct wq_device { 49336ba94429SFrederic Weisbecker struct workqueue_struct *wq; 49346ba94429SFrederic Weisbecker struct device dev; 49356ba94429SFrederic Weisbecker }; 49366ba94429SFrederic Weisbecker 49376ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 49386ba94429SFrederic Weisbecker { 49396ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 49406ba94429SFrederic Weisbecker 49416ba94429SFrederic Weisbecker return wq_dev->wq; 49426ba94429SFrederic Weisbecker } 49436ba94429SFrederic Weisbecker 49446ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 49456ba94429SFrederic Weisbecker char *buf) 49466ba94429SFrederic Weisbecker { 49476ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49486ba94429SFrederic Weisbecker 49496ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 49506ba94429SFrederic Weisbecker } 49516ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 49526ba94429SFrederic Weisbecker 49536ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 49546ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49556ba94429SFrederic Weisbecker { 49566ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49576ba94429SFrederic Weisbecker 49586ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 49596ba94429SFrederic Weisbecker } 49606ba94429SFrederic Weisbecker 49616ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 49626ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 49636ba94429SFrederic Weisbecker size_t count) 49646ba94429SFrederic Weisbecker { 49656ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49666ba94429SFrederic Weisbecker int val; 49676ba94429SFrederic Weisbecker 49686ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 49696ba94429SFrederic Weisbecker return -EINVAL; 49706ba94429SFrederic Weisbecker 49716ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 49726ba94429SFrederic Weisbecker return count; 49736ba94429SFrederic Weisbecker } 49746ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 49756ba94429SFrederic Weisbecker 49766ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 49776ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 49786ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 49796ba94429SFrederic Weisbecker NULL, 49806ba94429SFrederic Weisbecker }; 49816ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 49826ba94429SFrederic Weisbecker 49836ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 49846ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49856ba94429SFrederic Weisbecker { 49866ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49876ba94429SFrederic Weisbecker const char *delim = ""; 49886ba94429SFrederic Weisbecker int node, written = 0; 49896ba94429SFrederic Weisbecker 49906ba94429SFrederic Weisbecker rcu_read_lock_sched(); 49916ba94429SFrederic Weisbecker for_each_node(node) { 49926ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 49936ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 49946ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 49956ba94429SFrederic Weisbecker delim = " "; 49966ba94429SFrederic Weisbecker } 49976ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 49986ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 49996ba94429SFrederic Weisbecker 50006ba94429SFrederic Weisbecker return written; 50016ba94429SFrederic Weisbecker } 50026ba94429SFrederic Weisbecker 50036ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 50046ba94429SFrederic Weisbecker char *buf) 50056ba94429SFrederic Weisbecker { 50066ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50076ba94429SFrederic Weisbecker int written; 50086ba94429SFrederic Weisbecker 50096ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50106ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 50116ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50126ba94429SFrederic Weisbecker 50136ba94429SFrederic Weisbecker return written; 50146ba94429SFrederic Weisbecker } 50156ba94429SFrederic Weisbecker 50166ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 50176ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 50186ba94429SFrederic Weisbecker { 50196ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 50206ba94429SFrederic Weisbecker 5021899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5022899a94feSLai Jiangshan 50236ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 50246ba94429SFrederic Weisbecker if (!attrs) 50256ba94429SFrederic Weisbecker return NULL; 50266ba94429SFrederic Weisbecker 50276ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 50286ba94429SFrederic Weisbecker return attrs; 50296ba94429SFrederic Weisbecker } 50306ba94429SFrederic Weisbecker 50316ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 50326ba94429SFrederic Weisbecker const char *buf, size_t count) 50336ba94429SFrederic Weisbecker { 50346ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50356ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5036d4d3e257SLai Jiangshan int ret = -ENOMEM; 5037d4d3e257SLai Jiangshan 5038d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50396ba94429SFrederic Weisbecker 50406ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50416ba94429SFrederic Weisbecker if (!attrs) 5042d4d3e257SLai Jiangshan goto out_unlock; 50436ba94429SFrederic Weisbecker 50446ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 50456ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5046d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50476ba94429SFrederic Weisbecker else 50486ba94429SFrederic Weisbecker ret = -EINVAL; 50496ba94429SFrederic Weisbecker 5050d4d3e257SLai Jiangshan out_unlock: 5051d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50526ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50536ba94429SFrederic Weisbecker return ret ?: count; 50546ba94429SFrederic Weisbecker } 50556ba94429SFrederic Weisbecker 50566ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 50576ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 50586ba94429SFrederic Weisbecker { 50596ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50606ba94429SFrederic Weisbecker int written; 50616ba94429SFrederic Weisbecker 50626ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50636ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 50646ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 50656ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50666ba94429SFrederic Weisbecker return written; 50676ba94429SFrederic Weisbecker } 50686ba94429SFrederic Weisbecker 50696ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 50706ba94429SFrederic Weisbecker struct device_attribute *attr, 50716ba94429SFrederic Weisbecker const char *buf, size_t count) 50726ba94429SFrederic Weisbecker { 50736ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50746ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5075d4d3e257SLai Jiangshan int ret = -ENOMEM; 5076d4d3e257SLai Jiangshan 5077d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50786ba94429SFrederic Weisbecker 50796ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50806ba94429SFrederic Weisbecker if (!attrs) 5081d4d3e257SLai Jiangshan goto out_unlock; 50826ba94429SFrederic Weisbecker 50836ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 50846ba94429SFrederic Weisbecker if (!ret) 5085d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50866ba94429SFrederic Weisbecker 5087d4d3e257SLai Jiangshan out_unlock: 5088d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50896ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50906ba94429SFrederic Weisbecker return ret ?: count; 50916ba94429SFrederic Weisbecker } 50926ba94429SFrederic Weisbecker 50936ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 50946ba94429SFrederic Weisbecker char *buf) 50956ba94429SFrederic Weisbecker { 50966ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50976ba94429SFrederic Weisbecker int written; 50986ba94429SFrederic Weisbecker 50996ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 51006ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 51016ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 51026ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 51036ba94429SFrederic Weisbecker 51046ba94429SFrederic Weisbecker return written; 51056ba94429SFrederic Weisbecker } 51066ba94429SFrederic Weisbecker 51076ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 51086ba94429SFrederic Weisbecker const char *buf, size_t count) 51096ba94429SFrederic Weisbecker { 51106ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51116ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5112d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5113d4d3e257SLai Jiangshan 5114d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51156ba94429SFrederic Weisbecker 51166ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51176ba94429SFrederic Weisbecker if (!attrs) 5118d4d3e257SLai Jiangshan goto out_unlock; 51196ba94429SFrederic Weisbecker 51206ba94429SFrederic Weisbecker ret = -EINVAL; 51216ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 51226ba94429SFrederic Weisbecker attrs->no_numa = !v; 5123d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51246ba94429SFrederic Weisbecker } 51256ba94429SFrederic Weisbecker 5126d4d3e257SLai Jiangshan out_unlock: 5127d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 51286ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 51296ba94429SFrederic Weisbecker return ret ?: count; 51306ba94429SFrederic Weisbecker } 51316ba94429SFrederic Weisbecker 51326ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 51336ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 51346ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 51356ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 51366ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 51376ba94429SFrederic Weisbecker __ATTR_NULL, 51386ba94429SFrederic Weisbecker }; 51396ba94429SFrederic Weisbecker 51406ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 51416ba94429SFrederic Weisbecker .name = "workqueue", 51426ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 51436ba94429SFrederic Weisbecker }; 51446ba94429SFrederic Weisbecker 5145b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5146b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5147b05a7928SFrederic Weisbecker { 5148b05a7928SFrederic Weisbecker int written; 5149b05a7928SFrederic Weisbecker 5150042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5151b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5152b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5153042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5154b05a7928SFrederic Weisbecker 5155b05a7928SFrederic Weisbecker return written; 5156b05a7928SFrederic Weisbecker } 5157b05a7928SFrederic Weisbecker 5158042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5159042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5160042f7df1SLai Jiangshan { 5161042f7df1SLai Jiangshan cpumask_var_t cpumask; 5162042f7df1SLai Jiangshan int ret; 5163042f7df1SLai Jiangshan 5164042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5165042f7df1SLai Jiangshan return -ENOMEM; 5166042f7df1SLai Jiangshan 5167042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5168042f7df1SLai Jiangshan if (!ret) 5169042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5170042f7df1SLai Jiangshan 5171042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5172042f7df1SLai Jiangshan return ret ? ret : count; 5173042f7df1SLai Jiangshan } 5174042f7df1SLai Jiangshan 5175b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5176042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5177042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5178b05a7928SFrederic Weisbecker 51796ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 51806ba94429SFrederic Weisbecker { 5181b05a7928SFrederic Weisbecker int err; 5182b05a7928SFrederic Weisbecker 5183b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5184b05a7928SFrederic Weisbecker if (err) 5185b05a7928SFrederic Weisbecker return err; 5186b05a7928SFrederic Weisbecker 5187b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 51886ba94429SFrederic Weisbecker } 51896ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 51906ba94429SFrederic Weisbecker 51916ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 51926ba94429SFrederic Weisbecker { 51936ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 51946ba94429SFrederic Weisbecker 51956ba94429SFrederic Weisbecker kfree(wq_dev); 51966ba94429SFrederic Weisbecker } 51976ba94429SFrederic Weisbecker 51986ba94429SFrederic Weisbecker /** 51996ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 52006ba94429SFrederic Weisbecker * @wq: the workqueue to register 52016ba94429SFrederic Weisbecker * 52026ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 52036ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 52046ba94429SFrederic Weisbecker * which is the preferred method. 52056ba94429SFrederic Weisbecker * 52066ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 52076ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 52086ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 52096ba94429SFrederic Weisbecker * attributes. 52106ba94429SFrederic Weisbecker * 52116ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 52126ba94429SFrederic Weisbecker */ 52136ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 52146ba94429SFrederic Weisbecker { 52156ba94429SFrederic Weisbecker struct wq_device *wq_dev; 52166ba94429SFrederic Weisbecker int ret; 52176ba94429SFrederic Weisbecker 52186ba94429SFrederic Weisbecker /* 5219402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 52206ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 52216ba94429SFrederic Weisbecker * workqueues. 52226ba94429SFrederic Weisbecker */ 52236ba94429SFrederic Weisbecker if (WARN_ON(wq->flags & __WQ_ORDERED)) 52246ba94429SFrederic Weisbecker return -EINVAL; 52256ba94429SFrederic Weisbecker 52266ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 52276ba94429SFrederic Weisbecker if (!wq_dev) 52286ba94429SFrederic Weisbecker return -ENOMEM; 52296ba94429SFrederic Weisbecker 52306ba94429SFrederic Weisbecker wq_dev->wq = wq; 52316ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 52326ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 523323217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 52346ba94429SFrederic Weisbecker 52356ba94429SFrederic Weisbecker /* 52366ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 52376ba94429SFrederic Weisbecker * everything is ready. 52386ba94429SFrederic Weisbecker */ 52396ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 52406ba94429SFrederic Weisbecker 52416ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 52426ba94429SFrederic Weisbecker if (ret) { 52436ba94429SFrederic Weisbecker kfree(wq_dev); 52446ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52456ba94429SFrederic Weisbecker return ret; 52466ba94429SFrederic Weisbecker } 52476ba94429SFrederic Weisbecker 52486ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 52496ba94429SFrederic Weisbecker struct device_attribute *attr; 52506ba94429SFrederic Weisbecker 52516ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 52526ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 52536ba94429SFrederic Weisbecker if (ret) { 52546ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 52556ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52566ba94429SFrederic Weisbecker return ret; 52576ba94429SFrederic Weisbecker } 52586ba94429SFrederic Weisbecker } 52596ba94429SFrederic Weisbecker } 52606ba94429SFrederic Weisbecker 52616ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 52626ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 52636ba94429SFrederic Weisbecker return 0; 52646ba94429SFrederic Weisbecker } 52656ba94429SFrederic Weisbecker 52666ba94429SFrederic Weisbecker /** 52676ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 52686ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 52696ba94429SFrederic Weisbecker * 52706ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 52716ba94429SFrederic Weisbecker */ 52726ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 52736ba94429SFrederic Weisbecker { 52746ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 52756ba94429SFrederic Weisbecker 52766ba94429SFrederic Weisbecker if (!wq->wq_dev) 52776ba94429SFrederic Weisbecker return; 52786ba94429SFrederic Weisbecker 52796ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52806ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 52816ba94429SFrederic Weisbecker } 52826ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 52836ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 52846ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 52856ba94429SFrederic Weisbecker 528682607adcSTejun Heo /* 528782607adcSTejun Heo * Workqueue watchdog. 528882607adcSTejun Heo * 528982607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 529082607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 529182607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 529282607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 529382607adcSTejun Heo * largely opaque. 529482607adcSTejun Heo * 529582607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 529682607adcSTejun Heo * state if some pools failed to make forward progress for a while where 529782607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 529882607adcSTejun Heo * 529982607adcSTejun Heo * This mechanism is controlled through the kernel parameter 530082607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 530182607adcSTejun Heo * corresponding sysfs parameter file. 530282607adcSTejun Heo */ 530382607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 530482607adcSTejun Heo 530582607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data); 530682607adcSTejun Heo 530782607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 530882607adcSTejun Heo static struct timer_list wq_watchdog_timer = 530982607adcSTejun Heo TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0); 531082607adcSTejun Heo 531182607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 531282607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 531382607adcSTejun Heo 531482607adcSTejun Heo static void wq_watchdog_reset_touched(void) 531582607adcSTejun Heo { 531682607adcSTejun Heo int cpu; 531782607adcSTejun Heo 531882607adcSTejun Heo wq_watchdog_touched = jiffies; 531982607adcSTejun Heo for_each_possible_cpu(cpu) 532082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 532182607adcSTejun Heo } 532282607adcSTejun Heo 532382607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data) 532482607adcSTejun Heo { 532582607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 532682607adcSTejun Heo bool lockup_detected = false; 532782607adcSTejun Heo struct worker_pool *pool; 532882607adcSTejun Heo int pi; 532982607adcSTejun Heo 533082607adcSTejun Heo if (!thresh) 533182607adcSTejun Heo return; 533282607adcSTejun Heo 533382607adcSTejun Heo rcu_read_lock(); 533482607adcSTejun Heo 533582607adcSTejun Heo for_each_pool(pool, pi) { 533682607adcSTejun Heo unsigned long pool_ts, touched, ts; 533782607adcSTejun Heo 533882607adcSTejun Heo if (list_empty(&pool->worklist)) 533982607adcSTejun Heo continue; 534082607adcSTejun Heo 534182607adcSTejun Heo /* get the latest of pool and touched timestamps */ 534282607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 534382607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 534482607adcSTejun Heo 534582607adcSTejun Heo if (time_after(pool_ts, touched)) 534682607adcSTejun Heo ts = pool_ts; 534782607adcSTejun Heo else 534882607adcSTejun Heo ts = touched; 534982607adcSTejun Heo 535082607adcSTejun Heo if (pool->cpu >= 0) { 535182607adcSTejun Heo unsigned long cpu_touched = 535282607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 535382607adcSTejun Heo pool->cpu)); 535482607adcSTejun Heo if (time_after(cpu_touched, ts)) 535582607adcSTejun Heo ts = cpu_touched; 535682607adcSTejun Heo } 535782607adcSTejun Heo 535882607adcSTejun Heo /* did we stall? */ 535982607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 536082607adcSTejun Heo lockup_detected = true; 536182607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 536282607adcSTejun Heo pr_cont_pool_info(pool); 536382607adcSTejun Heo pr_cont(" stuck for %us!\n", 536482607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 536582607adcSTejun Heo } 536682607adcSTejun Heo } 536782607adcSTejun Heo 536882607adcSTejun Heo rcu_read_unlock(); 536982607adcSTejun Heo 537082607adcSTejun Heo if (lockup_detected) 537182607adcSTejun Heo show_workqueue_state(); 537282607adcSTejun Heo 537382607adcSTejun Heo wq_watchdog_reset_touched(); 537482607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 537582607adcSTejun Heo } 537682607adcSTejun Heo 537782607adcSTejun Heo void wq_watchdog_touch(int cpu) 537882607adcSTejun Heo { 537982607adcSTejun Heo if (cpu >= 0) 538082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 538182607adcSTejun Heo else 538282607adcSTejun Heo wq_watchdog_touched = jiffies; 538382607adcSTejun Heo } 538482607adcSTejun Heo 538582607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 538682607adcSTejun Heo { 538782607adcSTejun Heo wq_watchdog_thresh = 0; 538882607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 538982607adcSTejun Heo 539082607adcSTejun Heo if (thresh) { 539182607adcSTejun Heo wq_watchdog_thresh = thresh; 539282607adcSTejun Heo wq_watchdog_reset_touched(); 539382607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 539482607adcSTejun Heo } 539582607adcSTejun Heo } 539682607adcSTejun Heo 539782607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 539882607adcSTejun Heo const struct kernel_param *kp) 539982607adcSTejun Heo { 540082607adcSTejun Heo unsigned long thresh; 540182607adcSTejun Heo int ret; 540282607adcSTejun Heo 540382607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 540482607adcSTejun Heo if (ret) 540582607adcSTejun Heo return ret; 540682607adcSTejun Heo 540782607adcSTejun Heo if (system_wq) 540882607adcSTejun Heo wq_watchdog_set_thresh(thresh); 540982607adcSTejun Heo else 541082607adcSTejun Heo wq_watchdog_thresh = thresh; 541182607adcSTejun Heo 541282607adcSTejun Heo return 0; 541382607adcSTejun Heo } 541482607adcSTejun Heo 541582607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 541682607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 541782607adcSTejun Heo .get = param_get_ulong, 541882607adcSTejun Heo }; 541982607adcSTejun Heo 542082607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 542182607adcSTejun Heo 0644); 542282607adcSTejun Heo 542382607adcSTejun Heo static void wq_watchdog_init(void) 542482607adcSTejun Heo { 542582607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 542682607adcSTejun Heo } 542782607adcSTejun Heo 542882607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 542982607adcSTejun Heo 543082607adcSTejun Heo static inline void wq_watchdog_init(void) { } 543182607adcSTejun Heo 543282607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 543382607adcSTejun Heo 5434bce90380STejun Heo static void __init wq_numa_init(void) 5435bce90380STejun Heo { 5436bce90380STejun Heo cpumask_var_t *tbl; 5437bce90380STejun Heo int node, cpu; 5438bce90380STejun Heo 5439bce90380STejun Heo if (num_possible_nodes() <= 1) 5440bce90380STejun Heo return; 5441bce90380STejun Heo 5442d55262c4STejun Heo if (wq_disable_numa) { 5443d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5444d55262c4STejun Heo return; 5445d55262c4STejun Heo } 5446d55262c4STejun Heo 54474c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 54484c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 54494c16bd32STejun Heo 5450bce90380STejun Heo /* 5451bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5452bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5453bce90380STejun Heo * fully initialized by now. 5454bce90380STejun Heo */ 5455ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5456bce90380STejun Heo BUG_ON(!tbl); 5457bce90380STejun Heo 5458bce90380STejun Heo for_each_node(node) 54595a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 54601be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5461bce90380STejun Heo 5462bce90380STejun Heo for_each_possible_cpu(cpu) { 5463bce90380STejun Heo node = cpu_to_node(cpu); 5464bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5465bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5466bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5467bce90380STejun Heo return; 5468bce90380STejun Heo } 5469bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5470bce90380STejun Heo } 5471bce90380STejun Heo 5472bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5473bce90380STejun Heo wq_numa_enabled = true; 5474bce90380STejun Heo } 5475bce90380STejun Heo 54763347fa09STejun Heo /** 54773347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 54783347fa09STejun Heo * 54793347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 54803347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 54813347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 54823347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 54833347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 54843347fa09STejun Heo * created and scheduled right before early initcalls. 54853347fa09STejun Heo */ 54863347fa09STejun Heo int __init workqueue_init_early(void) 54871da177e4SLinus Torvalds { 54887a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 54897a4e344cSTejun Heo int i, cpu; 5490c34056a3STejun Heo 5491e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5492e904e6c2STejun Heo 5493b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5494b05a7928SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 5495b05a7928SFrederic Weisbecker 5496e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5497e904e6c2STejun Heo 5498706026c2STejun Heo /* initialize CPU pools */ 549929c91e99STejun Heo for_each_possible_cpu(cpu) { 55004ce62e9eSTejun Heo struct worker_pool *pool; 55018b03ae3cSTejun Heo 55027a4e344cSTejun Heo i = 0; 5503f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 55047a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5505ec22ca5eSTejun Heo pool->cpu = cpu; 55067a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 55077a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5508f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 55097a4e344cSTejun Heo 55109daf9e67STejun Heo /* alloc pool ID */ 551168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 55129daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 551368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55144ce62e9eSTejun Heo } 55158b03ae3cSTejun Heo } 55168b03ae3cSTejun Heo 55178a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 551829c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 551929c91e99STejun Heo struct workqueue_attrs *attrs; 552029c91e99STejun Heo 552129c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 552229c91e99STejun Heo attrs->nice = std_nice[i]; 552329c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 55248a2b7538STejun Heo 55258a2b7538STejun Heo /* 55268a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 55278a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 55288a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 55298a2b7538STejun Heo */ 55308a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 55318a2b7538STejun Heo attrs->nice = std_nice[i]; 55328a2b7538STejun Heo attrs->no_numa = true; 55338a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 553429c91e99STejun Heo } 553529c91e99STejun Heo 5536d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 55371aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5538d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5539f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5540f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 554124d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 554224d51addSTejun Heo WQ_FREEZABLE, 0); 55430668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 55440668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 55450668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 55460668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 55470668106cSViresh Kumar 0); 55481aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 55490668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 55500668106cSViresh Kumar !system_power_efficient_wq || 55510668106cSViresh Kumar !system_freezable_power_efficient_wq); 555282607adcSTejun Heo 55533347fa09STejun Heo return 0; 55543347fa09STejun Heo } 55553347fa09STejun Heo 55563347fa09STejun Heo /** 55573347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 55583347fa09STejun Heo * 55593347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 55603347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 55613347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 55623347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 55633347fa09STejun Heo * with the initial workers and enable future kworker creations. 55643347fa09STejun Heo */ 55653347fa09STejun Heo int __init workqueue_init(void) 55663347fa09STejun Heo { 55672186d9f9STejun Heo struct workqueue_struct *wq; 55683347fa09STejun Heo struct worker_pool *pool; 55693347fa09STejun Heo int cpu, bkt; 55703347fa09STejun Heo 55712186d9f9STejun Heo /* 55722186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 55732186d9f9STejun Heo * CPU to node mapping may not be available that early on some 55742186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 55752186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 55762186d9f9STejun Heo * affinity, fix them up. 55772186d9f9STejun Heo */ 55782186d9f9STejun Heo wq_numa_init(); 55792186d9f9STejun Heo 55802186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 55812186d9f9STejun Heo 55822186d9f9STejun Heo for_each_possible_cpu(cpu) { 55832186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 55842186d9f9STejun Heo pool->node = cpu_to_node(cpu); 55852186d9f9STejun Heo } 55862186d9f9STejun Heo } 55872186d9f9STejun Heo 55882186d9f9STejun Heo list_for_each_entry(wq, &workqueues, list) 55892186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 55902186d9f9STejun Heo 55912186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 55922186d9f9STejun Heo 55933347fa09STejun Heo /* create the initial workers */ 55943347fa09STejun Heo for_each_online_cpu(cpu) { 55953347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 55963347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 55973347fa09STejun Heo BUG_ON(!create_worker(pool)); 55983347fa09STejun Heo } 55993347fa09STejun Heo } 56003347fa09STejun Heo 56013347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 56023347fa09STejun Heo BUG_ON(!create_worker(pool)); 56033347fa09STejun Heo 56043347fa09STejun Heo wq_online = true; 560582607adcSTejun Heo wq_watchdog_init(); 560682607adcSTejun Heo 56076ee0578bSSuresh Siddha return 0; 56081da177e4SLinus Torvalds } 5609