11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 20b11895c4SLibin * normal work items and the other for high priority ones) and some extra 21b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 22b11895c4SLibin * number of these backing pools is dynamic. 23c54fce6eSTejun Heo * 24c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 251da177e4SLinus Torvalds */ 261da177e4SLinus Torvalds 279984de1aSPaul Gortmaker #include <linux/export.h> 281da177e4SLinus Torvalds #include <linux/kernel.h> 291da177e4SLinus Torvalds #include <linux/sched.h> 301da177e4SLinus Torvalds #include <linux/init.h> 311da177e4SLinus Torvalds #include <linux/signal.h> 321da177e4SLinus Torvalds #include <linux/completion.h> 331da177e4SLinus Torvalds #include <linux/workqueue.h> 341da177e4SLinus Torvalds #include <linux/slab.h> 351da177e4SLinus Torvalds #include <linux/cpu.h> 361da177e4SLinus Torvalds #include <linux/notifier.h> 371da177e4SLinus Torvalds #include <linux/kthread.h> 381fa44ecaSJames Bottomley #include <linux/hardirq.h> 3946934023SChristoph Lameter #include <linux/mempolicy.h> 40341a5958SRafael J. Wysocki #include <linux/freezer.h> 41d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51e22bee78STejun Heo 52ea138446STejun Heo #include "workqueue_internal.h" 531da177e4SLinus Torvalds 54c8e55f36STejun Heo enum { 55bc2ae0f5STejun Heo /* 5624647570STejun Heo * worker_pool flags 57bc2ae0f5STejun Heo * 5824647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 59bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 60bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 61bc2ae0f5STejun Heo * is in effect. 62bc2ae0f5STejun Heo * 63bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 64bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6524647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 66bc2ae0f5STejun Heo * 67bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 6892f9c5c4SLai Jiangshan * attach_mutex to avoid changing binding state while 694736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 70bc2ae0f5STejun Heo */ 7124647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 72db7bccf4STejun Heo 73c8e55f36STejun Heo /* worker flags */ 74c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 75c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 76e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 77fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 78f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 79a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 80e22bee78STejun Heo 81a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 82a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 83db7bccf4STejun Heo 84e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 854ce62e9eSTejun Heo 8629c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 87c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 88db7bccf4STejun Heo 89e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 90e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 91e22bee78STejun Heo 923233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 933233cdbdSTejun Heo /* call for help after 10ms 943233cdbdSTejun Heo (min two ticks) */ 95e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 96e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds /* 99e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1008698a745SDongsheng Yang * all cpus. Give MIN_NICE. 101e22bee78STejun Heo */ 1028698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1038698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 104ecf6881fSTejun Heo 105ecf6881fSTejun Heo WQ_NAME_LEN = 24, 106c8e55f36STejun Heo }; 107c8e55f36STejun Heo 1081da177e4SLinus Torvalds /* 1094690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1104690c4abSTejun Heo * 111e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 112e41e704bSTejun Heo * everyone else. 1134690c4abSTejun Heo * 114e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 115e22bee78STejun Heo * only be modified and accessed from the local cpu. 116e22bee78STejun Heo * 117d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1184690c4abSTejun Heo * 119d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 120d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 121d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 122d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 123e22bee78STejun Heo * 12492f9c5c4SLai Jiangshan * A: pool->attach_mutex protected. 125822d8405STejun Heo * 12668e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12776af4d93STejun Heo * 12868e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1295bcab335STejun Heo * 1305b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1315b95e1afSLai Jiangshan * 1325b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 1335b95e1afSLai Jiangshan * sched-RCU for reads. 1345b95e1afSLai Jiangshan * 1353c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1363c25a55dSLai Jiangshan * 137b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1382e109a28STejun Heo * 1392e109a28STejun Heo * MD: wq_mayday_lock protected. 1404690c4abSTejun Heo */ 1414690c4abSTejun Heo 1422eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 143c34056a3STejun Heo 144bd7bdd43STejun Heo struct worker_pool { 145d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 146d84ff051STejun Heo int cpu; /* I: the associated cpu */ 147f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1489daf9e67STejun Heo int id; /* I: pool ID */ 14911ebea50STejun Heo unsigned int flags; /* X: flags */ 150bd7bdd43STejun Heo 15182607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15282607adcSTejun Heo 153bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 154bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 155ea1abd61SLai Jiangshan 156ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 157bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 158bd7bdd43STejun Heo 159bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 160bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 161bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 162bd7bdd43STejun Heo 163c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 164c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 165c9e7cf27STejun Heo /* L: hash of busy workers */ 166c9e7cf27STejun Heo 167bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16834a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 1692607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17092f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 17192f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17260f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 173e19e397aSTejun Heo 1747cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 175e19e397aSTejun Heo 1767a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17768e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17868e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1797a4e344cSTejun Heo 180e19e397aSTejun Heo /* 181e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 182e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 183e19e397aSTejun Heo * cacheline. 184e19e397aSTejun Heo */ 185e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18629c91e99STejun Heo 18729c91e99STejun Heo /* 18829c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18929c91e99STejun Heo * from get_work_pool(). 19029c91e99STejun Heo */ 19129c91e99STejun Heo struct rcu_head rcu; 1928b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1938b03ae3cSTejun Heo 1948b03ae3cSTejun Heo /* 195112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 196112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 197112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 198112202d9STejun Heo * number of flag bits. 1991da177e4SLinus Torvalds */ 200112202d9STejun Heo struct pool_workqueue { 201bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2024690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20373f53c4aSTejun Heo int work_color; /* L: current color */ 20473f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2058864b4e5STejun Heo int refcnt; /* L: reference count */ 20673f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20773f53c4aSTejun Heo /* L: nr of in_flight works */ 2081e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 209a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2101e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2113c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2122e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2138864b4e5STejun Heo 2148864b4e5STejun Heo /* 2158864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2168864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2178864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 218b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2198864b4e5STejun Heo */ 2208864b4e5STejun Heo struct work_struct unbound_release_work; 2218864b4e5STejun Heo struct rcu_head rcu; 222e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* 22573f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22673f53c4aSTejun Heo */ 22773f53c4aSTejun Heo struct wq_flusher { 2283c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2293c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 23073f53c4aSTejun Heo struct completion done; /* flush completion */ 23173f53c4aSTejun Heo }; 2321da177e4SLinus Torvalds 233226223abSTejun Heo struct wq_device; 234226223abSTejun Heo 23573f53c4aSTejun Heo /* 236c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 237c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2381da177e4SLinus Torvalds */ 2391da177e4SLinus Torvalds struct workqueue_struct { 2403c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 241e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 24273f53c4aSTejun Heo 2433c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2443c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2453c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 246112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2473c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2483c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2493c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 25073f53c4aSTejun Heo 2512e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 252e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 253e22bee78STejun Heo 25487fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 255a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 256226223abSTejun Heo 2575b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2585b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2596029a918STejun Heo 260226223abSTejun Heo #ifdef CONFIG_SYSFS 261226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 262226223abSTejun Heo #endif 2634e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2644e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2654e6045f1SJohannes Berg #endif 266ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2672728fd2fSTejun Heo 268e2dca7adSTejun Heo /* 269e2dca7adSTejun Heo * Destruction of workqueue_struct is sched-RCU protected to allow 270e2dca7adSTejun Heo * walking the workqueues list without grabbing wq_pool_mutex. 271e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 272e2dca7adSTejun Heo */ 273e2dca7adSTejun Heo struct rcu_head rcu; 274e2dca7adSTejun Heo 2752728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2762728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2772728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2785b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2791da177e4SLinus Torvalds }; 2801da177e4SLinus Torvalds 281e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 282e904e6c2STejun Heo 283bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 284bce90380STejun Heo /* possible CPUs of each node */ 285bce90380STejun Heo 286d55262c4STejun Heo static bool wq_disable_numa; 287d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 288d55262c4STejun Heo 289cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 290552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 291cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 292cee22a15SViresh Kumar 293bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 294bce90380STejun Heo 2954c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2964c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2974c16bd32STejun Heo 29868e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2992e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 3005bcab335STejun Heo 301e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30268e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3037d19c5ceSTejun Heo 304042f7df1SLai Jiangshan static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */ 305b05a7928SFrederic Weisbecker 3067d19c5ceSTejun Heo /* the per-cpu worker pools */ 3077d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 3087d19c5ceSTejun Heo cpu_worker_pools); 3097d19c5ceSTejun Heo 31068e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3117d19c5ceSTejun Heo 31268e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 31329c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 31429c91e99STejun Heo 315c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 31629c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 31729c91e99STejun Heo 3188a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3198a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3208a2b7538STejun Heo 321d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 322ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 323044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3241aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 325044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 326d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 327044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 328f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 329044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 33024d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3310668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3320668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3330668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3340668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 335d320c038STejun Heo 3367d19c5ceSTejun Heo static int worker_thread(void *__worker); 3376ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3387d19c5ceSTejun Heo 33997bd2347STejun Heo #define CREATE_TRACE_POINTS 34097bd2347STejun Heo #include <trace/events/workqueue.h> 34197bd2347STejun Heo 34268e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 343f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 344f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 34568e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3465bcab335STejun Heo 347b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 348f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 349f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 350b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 35176af4d93STejun Heo 3525b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 353f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 354f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 355f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3565b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3575b95e1afSLai Jiangshan 358f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 359f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 360f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3617a62c2c8STejun Heo (pool)++) 3624ce62e9eSTejun Heo 36349e3cf44STejun Heo /** 36417116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 36517116969STejun Heo * @pool: iteration cursor 366611c92a0STejun Heo * @pi: integer used for iteration 367fa1b54e6STejun Heo * 36868e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 36968e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 37068e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 371fa1b54e6STejun Heo * 372fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 373fa1b54e6STejun Heo * ignored. 37417116969STejun Heo */ 375611c92a0STejun Heo #define for_each_pool(pool, pi) \ 376611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 37768e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 378fa1b54e6STejun Heo else 37917116969STejun Heo 38017116969STejun Heo /** 381822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 382822d8405STejun Heo * @worker: iteration cursor 383822d8405STejun Heo * @pool: worker_pool to iterate workers of 384822d8405STejun Heo * 38592f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 386822d8405STejun Heo * 387822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 388822d8405STejun Heo * ignored. 389822d8405STejun Heo */ 390da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 391da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 39292f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 393822d8405STejun Heo else 394822d8405STejun Heo 395822d8405STejun Heo /** 39649e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 39749e3cf44STejun Heo * @pwq: iteration cursor 39849e3cf44STejun Heo * @wq: the target workqueue 39976af4d93STejun Heo * 400b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 401794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 402794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 40376af4d93STejun Heo * 40476af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 40576af4d93STejun Heo * ignored. 40649e3cf44STejun Heo */ 40749e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 40876af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 409b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 41076af4d93STejun Heo else 411f3421797STejun Heo 412dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 413dc186ad7SThomas Gleixner 414dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 415dc186ad7SThomas Gleixner 41699777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 41799777288SStanislaw Gruszka { 41899777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 41999777288SStanislaw Gruszka } 42099777288SStanislaw Gruszka 421dc186ad7SThomas Gleixner /* 422dc186ad7SThomas Gleixner * fixup_init is called when: 423dc186ad7SThomas Gleixner * - an active object is initialized 424dc186ad7SThomas Gleixner */ 425dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 426dc186ad7SThomas Gleixner { 427dc186ad7SThomas Gleixner struct work_struct *work = addr; 428dc186ad7SThomas Gleixner 429dc186ad7SThomas Gleixner switch (state) { 430dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 431dc186ad7SThomas Gleixner cancel_work_sync(work); 432dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 433dc186ad7SThomas Gleixner return 1; 434dc186ad7SThomas Gleixner default: 435dc186ad7SThomas Gleixner return 0; 436dc186ad7SThomas Gleixner } 437dc186ad7SThomas Gleixner } 438dc186ad7SThomas Gleixner 439dc186ad7SThomas Gleixner /* 440dc186ad7SThomas Gleixner * fixup_activate is called when: 441dc186ad7SThomas Gleixner * - an active object is activated 442dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 443dc186ad7SThomas Gleixner */ 444dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 445dc186ad7SThomas Gleixner { 446dc186ad7SThomas Gleixner struct work_struct *work = addr; 447dc186ad7SThomas Gleixner 448dc186ad7SThomas Gleixner switch (state) { 449dc186ad7SThomas Gleixner 450dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 451dc186ad7SThomas Gleixner /* 452dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 453dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 454dc186ad7SThomas Gleixner * is tracked in the object tracker. 455dc186ad7SThomas Gleixner */ 45622df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 457dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 458dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 459dc186ad7SThomas Gleixner return 0; 460dc186ad7SThomas Gleixner } 461dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 462dc186ad7SThomas Gleixner return 0; 463dc186ad7SThomas Gleixner 464dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 465dc186ad7SThomas Gleixner WARN_ON(1); 466dc186ad7SThomas Gleixner 467dc186ad7SThomas Gleixner default: 468dc186ad7SThomas Gleixner return 0; 469dc186ad7SThomas Gleixner } 470dc186ad7SThomas Gleixner } 471dc186ad7SThomas Gleixner 472dc186ad7SThomas Gleixner /* 473dc186ad7SThomas Gleixner * fixup_free is called when: 474dc186ad7SThomas Gleixner * - an active object is freed 475dc186ad7SThomas Gleixner */ 476dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 477dc186ad7SThomas Gleixner { 478dc186ad7SThomas Gleixner struct work_struct *work = addr; 479dc186ad7SThomas Gleixner 480dc186ad7SThomas Gleixner switch (state) { 481dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 482dc186ad7SThomas Gleixner cancel_work_sync(work); 483dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 484dc186ad7SThomas Gleixner return 1; 485dc186ad7SThomas Gleixner default: 486dc186ad7SThomas Gleixner return 0; 487dc186ad7SThomas Gleixner } 488dc186ad7SThomas Gleixner } 489dc186ad7SThomas Gleixner 490dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 491dc186ad7SThomas Gleixner .name = "work_struct", 49299777288SStanislaw Gruszka .debug_hint = work_debug_hint, 493dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 494dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 495dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 496dc186ad7SThomas Gleixner }; 497dc186ad7SThomas Gleixner 498dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 499dc186ad7SThomas Gleixner { 500dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 501dc186ad7SThomas Gleixner } 502dc186ad7SThomas Gleixner 503dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 504dc186ad7SThomas Gleixner { 505dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 506dc186ad7SThomas Gleixner } 507dc186ad7SThomas Gleixner 508dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner if (onstack) 511dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 512dc186ad7SThomas Gleixner else 513dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 514dc186ad7SThomas Gleixner } 515dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 516dc186ad7SThomas Gleixner 517dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 518dc186ad7SThomas Gleixner { 519dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 520dc186ad7SThomas Gleixner } 521dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 522dc186ad7SThomas Gleixner 523ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 524ea2e64f2SThomas Gleixner { 525ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 526ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 527ea2e64f2SThomas Gleixner } 528ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 529ea2e64f2SThomas Gleixner 530dc186ad7SThomas Gleixner #else 531dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 532dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 533dc186ad7SThomas Gleixner #endif 534dc186ad7SThomas Gleixner 5354e8b22bdSLi Bin /** 5364e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5374e8b22bdSLi Bin * @pool: the pool pointer of interest 5384e8b22bdSLi Bin * 5394e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5404e8b22bdSLi Bin * successfully, -errno on failure. 5414e8b22bdSLi Bin */ 5429daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5439daf9e67STejun Heo { 5449daf9e67STejun Heo int ret; 5459daf9e67STejun Heo 54668e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5475bcab335STejun Heo 5484e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5494e8b22bdSLi Bin GFP_KERNEL); 550229641a6STejun Heo if (ret >= 0) { 551e68035fbSTejun Heo pool->id = ret; 552229641a6STejun Heo return 0; 553229641a6STejun Heo } 5549daf9e67STejun Heo return ret; 5559daf9e67STejun Heo } 5569daf9e67STejun Heo 55776af4d93STejun Heo /** 558df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 559df2d5ae4STejun Heo * @wq: the target workqueue 560df2d5ae4STejun Heo * @node: the node ID 561df2d5ae4STejun Heo * 5625b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5635b95e1afSLai Jiangshan * read locked. 564df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 565df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 566d185af30SYacine Belkadi * 567d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 568df2d5ae4STejun Heo */ 569df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 570df2d5ae4STejun Heo int node) 571df2d5ae4STejun Heo { 5725b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 573df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 574df2d5ae4STejun Heo } 575df2d5ae4STejun Heo 57673f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57773f53c4aSTejun Heo { 57873f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 57973f53c4aSTejun Heo } 58073f53c4aSTejun Heo 58173f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58273f53c4aSTejun Heo { 58373f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58473f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58573f53c4aSTejun Heo } 58673f53c4aSTejun Heo 58773f53c4aSTejun Heo static int work_next_color(int color) 58873f53c4aSTejun Heo { 58973f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 590a848e3b6SOleg Nesterov } 591a848e3b6SOleg Nesterov 5924594bf15SDavid Howells /* 593112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 594112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5957c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5967a22ad75STejun Heo * 597112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 598112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 599bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 600bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6017a22ad75STejun Heo * 602112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6037c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 604112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6057c3eed5cSTejun Heo * available only while the work item is queued. 606bbb68dfaSTejun Heo * 607bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 608bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 609bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 610bbb68dfaSTejun Heo * try to steal the PENDING bit. 6114594bf15SDavid Howells */ 6127a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6137a22ad75STejun Heo unsigned long flags) 6147a22ad75STejun Heo { 6156183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6167a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6177a22ad75STejun Heo } 6187a22ad75STejun Heo 619112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6204690c4abSTejun Heo unsigned long extra_flags) 621365970a1SDavid Howells { 622112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 623112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 624365970a1SDavid Howells } 625365970a1SDavid Howells 6264468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6274468a00fSLai Jiangshan int pool_id) 6284468a00fSLai Jiangshan { 6294468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6304468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6314468a00fSLai Jiangshan } 6324468a00fSLai Jiangshan 6337c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6347c3eed5cSTejun Heo int pool_id) 6354d707b9fSOleg Nesterov { 63623657bb1STejun Heo /* 63723657bb1STejun Heo * The following wmb is paired with the implied mb in 63823657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 63923657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64023657bb1STejun Heo * owner. 64123657bb1STejun Heo */ 64223657bb1STejun Heo smp_wmb(); 6437c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6444d707b9fSOleg Nesterov } 6454d707b9fSOleg Nesterov 6467a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 647365970a1SDavid Howells { 6487c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6497c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6507a22ad75STejun Heo } 6517a22ad75STejun Heo 652112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6537a22ad75STejun Heo { 654e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6557a22ad75STejun Heo 656112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 657e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 658e120153dSTejun Heo else 659e120153dSTejun Heo return NULL; 6607a22ad75STejun Heo } 6617a22ad75STejun Heo 6627c3eed5cSTejun Heo /** 6637c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6647c3eed5cSTejun Heo * @work: the work item of interest 6657c3eed5cSTejun Heo * 66668e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 66768e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 66868e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 669fa1b54e6STejun Heo * 670fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 671fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 672fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 673fa1b54e6STejun Heo * returned pool is and stays online. 674d185af30SYacine Belkadi * 675d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6767c3eed5cSTejun Heo */ 6777c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6787a22ad75STejun Heo { 679e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6807c3eed5cSTejun Heo int pool_id; 6817a22ad75STejun Heo 68268e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 683fa1b54e6STejun Heo 684112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 685112202d9STejun Heo return ((struct pool_workqueue *) 6867c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6877a22ad75STejun Heo 6887c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6897c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6907a22ad75STejun Heo return NULL; 6917a22ad75STejun Heo 692fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6937c3eed5cSTejun Heo } 6947c3eed5cSTejun Heo 6957c3eed5cSTejun Heo /** 6967c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6977c3eed5cSTejun Heo * @work: the work item of interest 6987c3eed5cSTejun Heo * 699d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7007c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7017c3eed5cSTejun Heo */ 7027c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7037c3eed5cSTejun Heo { 70454d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7057c3eed5cSTejun Heo 706112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 707112202d9STejun Heo return ((struct pool_workqueue *) 70854d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 70954d5b7d0SLai Jiangshan 71054d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7117c3eed5cSTejun Heo } 7127c3eed5cSTejun Heo 713bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 714bbb68dfaSTejun Heo { 7157c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 716bbb68dfaSTejun Heo 7177c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7187c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 719bbb68dfaSTejun Heo } 720bbb68dfaSTejun Heo 721bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 722bbb68dfaSTejun Heo { 723bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 724bbb68dfaSTejun Heo 725112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 726bbb68dfaSTejun Heo } 727bbb68dfaSTejun Heo 728e22bee78STejun Heo /* 7293270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7303270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 731d565ed63STejun Heo * they're being called with pool->lock held. 732e22bee78STejun Heo */ 733e22bee78STejun Heo 73463d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 735649027d7STejun Heo { 736e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 737649027d7STejun Heo } 738649027d7STejun Heo 739e22bee78STejun Heo /* 740e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 741e22bee78STejun Heo * running workers. 742974271c4STejun Heo * 743974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 744706026c2STejun Heo * function will always return %true for unbound pools as long as the 745974271c4STejun Heo * worklist isn't empty. 746e22bee78STejun Heo */ 74763d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 748e22bee78STejun Heo { 74963d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 750e22bee78STejun Heo } 751e22bee78STejun Heo 752e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 75363d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 754e22bee78STejun Heo { 75563d95a91STejun Heo return pool->nr_idle; 756e22bee78STejun Heo } 757e22bee78STejun Heo 758e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 75963d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 760e22bee78STejun Heo { 761e19e397aSTejun Heo return !list_empty(&pool->worklist) && 762e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 763e22bee78STejun Heo } 764e22bee78STejun Heo 765e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 76663d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 767e22bee78STejun Heo { 76863d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 769e22bee78STejun Heo } 770e22bee78STejun Heo 771e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 77263d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 773e22bee78STejun Heo { 77434a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 77563d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 77663d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 777e22bee78STejun Heo 778e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 779e22bee78STejun Heo } 780e22bee78STejun Heo 781e22bee78STejun Heo /* 782e22bee78STejun Heo * Wake up functions. 783e22bee78STejun Heo */ 784e22bee78STejun Heo 7851037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 7861037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 7877e11629dSTejun Heo { 78863d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7897e11629dSTejun Heo return NULL; 7907e11629dSTejun Heo 79163d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7927e11629dSTejun Heo } 7937e11629dSTejun Heo 7947e11629dSTejun Heo /** 7957e11629dSTejun Heo * wake_up_worker - wake up an idle worker 79663d95a91STejun Heo * @pool: worker pool to wake worker from 7977e11629dSTejun Heo * 79863d95a91STejun Heo * Wake up the first idle worker of @pool. 7997e11629dSTejun Heo * 8007e11629dSTejun Heo * CONTEXT: 801d565ed63STejun Heo * spin_lock_irq(pool->lock). 8027e11629dSTejun Heo */ 80363d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8047e11629dSTejun Heo { 8051037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8067e11629dSTejun Heo 8077e11629dSTejun Heo if (likely(worker)) 8087e11629dSTejun Heo wake_up_process(worker->task); 8097e11629dSTejun Heo } 8107e11629dSTejun Heo 8114690c4abSTejun Heo /** 812e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 813e22bee78STejun Heo * @task: task waking up 814e22bee78STejun Heo * @cpu: CPU @task is waking up to 815e22bee78STejun Heo * 816e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 817e22bee78STejun Heo * being awoken. 818e22bee78STejun Heo * 819e22bee78STejun Heo * CONTEXT: 820e22bee78STejun Heo * spin_lock_irq(rq->lock) 821e22bee78STejun Heo */ 822d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 823e22bee78STejun Heo { 824e22bee78STejun Heo struct worker *worker = kthread_data(task); 825e22bee78STejun Heo 82636576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 827ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 828e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 829e22bee78STejun Heo } 83036576000SJoonsoo Kim } 831e22bee78STejun Heo 832e22bee78STejun Heo /** 833e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 834e22bee78STejun Heo * @task: task going to sleep 835e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 836e22bee78STejun Heo * 837e22bee78STejun Heo * This function is called during schedule() when a busy worker is 838e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 839e22bee78STejun Heo * returning pointer to its task. 840e22bee78STejun Heo * 841e22bee78STejun Heo * CONTEXT: 842e22bee78STejun Heo * spin_lock_irq(rq->lock) 843e22bee78STejun Heo * 844d185af30SYacine Belkadi * Return: 845e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 846e22bee78STejun Heo */ 847d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 848e22bee78STejun Heo { 849e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 850111c225aSTejun Heo struct worker_pool *pool; 851e22bee78STejun Heo 852111c225aSTejun Heo /* 853111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 854111c225aSTejun Heo * workers, also reach here, let's not access anything before 855111c225aSTejun Heo * checking NOT_RUNNING. 856111c225aSTejun Heo */ 8572d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 858e22bee78STejun Heo return NULL; 859e22bee78STejun Heo 860111c225aSTejun Heo pool = worker->pool; 861111c225aSTejun Heo 862e22bee78STejun Heo /* this can only happen on the local cpu */ 86392b69f50SLai Jiangshan if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) 8646183c009STejun Heo return NULL; 865e22bee78STejun Heo 866e22bee78STejun Heo /* 867e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 868e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 869e22bee78STejun Heo * Please read comment there. 870e22bee78STejun Heo * 871628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 872628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 873628c78e7STejun Heo * disabled, which in turn means that none else could be 874d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 875628c78e7STejun Heo * lock is safe. 876e22bee78STejun Heo */ 877e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 878e19e397aSTejun Heo !list_empty(&pool->worklist)) 8791037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 880e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 881e22bee78STejun Heo } 882e22bee78STejun Heo 883e22bee78STejun Heo /** 884e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 885cb444766STejun Heo * @worker: self 886d302f017STejun Heo * @flags: flags to set 887d302f017STejun Heo * 888228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 889d302f017STejun Heo * 890cb444766STejun Heo * CONTEXT: 891d565ed63STejun Heo * spin_lock_irq(pool->lock) 892d302f017STejun Heo */ 893228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 894d302f017STejun Heo { 895bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 896e22bee78STejun Heo 897cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 898cb444766STejun Heo 899228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 900e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 901e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 902e19e397aSTejun Heo atomic_dec(&pool->nr_running); 903e22bee78STejun Heo } 904e22bee78STejun Heo 905d302f017STejun Heo worker->flags |= flags; 906d302f017STejun Heo } 907d302f017STejun Heo 908d302f017STejun Heo /** 909e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 910cb444766STejun Heo * @worker: self 911d302f017STejun Heo * @flags: flags to clear 912d302f017STejun Heo * 913e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 914d302f017STejun Heo * 915cb444766STejun Heo * CONTEXT: 916d565ed63STejun Heo * spin_lock_irq(pool->lock) 917d302f017STejun Heo */ 918d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 919d302f017STejun Heo { 92063d95a91STejun Heo struct worker_pool *pool = worker->pool; 921e22bee78STejun Heo unsigned int oflags = worker->flags; 922e22bee78STejun Heo 923cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 924cb444766STejun Heo 925d302f017STejun Heo worker->flags &= ~flags; 926e22bee78STejun Heo 92742c025f3STejun Heo /* 92842c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 92942c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 93042c025f3STejun Heo * of multiple flags, not a single flag. 93142c025f3STejun Heo */ 932e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 933e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 934e19e397aSTejun Heo atomic_inc(&pool->nr_running); 935d302f017STejun Heo } 936d302f017STejun Heo 937d302f017STejun Heo /** 9388cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 939c9e7cf27STejun Heo * @pool: pool of interest 9408cca0eeaSTejun Heo * @work: work to find worker for 9418cca0eeaSTejun Heo * 942c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 943c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 944a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 945a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 946a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 947a2c1c57bSTejun Heo * being executed. 948a2c1c57bSTejun Heo * 949a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 950a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 951a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 952a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 953a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 954a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 955a2c1c57bSTejun Heo * 956c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 957c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 958c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 959c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 960c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 961c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9628cca0eeaSTejun Heo * 9638cca0eeaSTejun Heo * CONTEXT: 964d565ed63STejun Heo * spin_lock_irq(pool->lock). 9658cca0eeaSTejun Heo * 966d185af30SYacine Belkadi * Return: 967d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9688cca0eeaSTejun Heo * otherwise. 9698cca0eeaSTejun Heo */ 970c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9718cca0eeaSTejun Heo struct work_struct *work) 9728cca0eeaSTejun Heo { 97342f8570fSSasha Levin struct worker *worker; 97442f8570fSSasha Levin 975b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 976a2c1c57bSTejun Heo (unsigned long)work) 977a2c1c57bSTejun Heo if (worker->current_work == work && 978a2c1c57bSTejun Heo worker->current_func == work->func) 97942f8570fSSasha Levin return worker; 98042f8570fSSasha Levin 98142f8570fSSasha Levin return NULL; 9828cca0eeaSTejun Heo } 9838cca0eeaSTejun Heo 9848cca0eeaSTejun Heo /** 985bf4ede01STejun Heo * move_linked_works - move linked works to a list 986bf4ede01STejun Heo * @work: start of series of works to be scheduled 987bf4ede01STejun Heo * @head: target list to append @work to 988402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 989bf4ede01STejun Heo * 990bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 991bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 992bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 993bf4ede01STejun Heo * 994bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 995bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 996bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 997bf4ede01STejun Heo * 998bf4ede01STejun Heo * CONTEXT: 999d565ed63STejun Heo * spin_lock_irq(pool->lock). 1000bf4ede01STejun Heo */ 1001bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1002bf4ede01STejun Heo struct work_struct **nextp) 1003bf4ede01STejun Heo { 1004bf4ede01STejun Heo struct work_struct *n; 1005bf4ede01STejun Heo 1006bf4ede01STejun Heo /* 1007bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1008bf4ede01STejun Heo * use NULL for list head. 1009bf4ede01STejun Heo */ 1010bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1011bf4ede01STejun Heo list_move_tail(&work->entry, head); 1012bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1013bf4ede01STejun Heo break; 1014bf4ede01STejun Heo } 1015bf4ede01STejun Heo 1016bf4ede01STejun Heo /* 1017bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1018bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1019bf4ede01STejun Heo * needs to be updated. 1020bf4ede01STejun Heo */ 1021bf4ede01STejun Heo if (nextp) 1022bf4ede01STejun Heo *nextp = n; 1023bf4ede01STejun Heo } 1024bf4ede01STejun Heo 10258864b4e5STejun Heo /** 10268864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10278864b4e5STejun Heo * @pwq: pool_workqueue to get 10288864b4e5STejun Heo * 10298864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10308864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10318864b4e5STejun Heo */ 10328864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10338864b4e5STejun Heo { 10348864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10358864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10368864b4e5STejun Heo pwq->refcnt++; 10378864b4e5STejun Heo } 10388864b4e5STejun Heo 10398864b4e5STejun Heo /** 10408864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10418864b4e5STejun Heo * @pwq: pool_workqueue to put 10428864b4e5STejun Heo * 10438864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10448864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10458864b4e5STejun Heo */ 10468864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10478864b4e5STejun Heo { 10488864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10498864b4e5STejun Heo if (likely(--pwq->refcnt)) 10508864b4e5STejun Heo return; 10518864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10528864b4e5STejun Heo return; 10538864b4e5STejun Heo /* 10548864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10558864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10568864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10578864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10588864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10598864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10608864b4e5STejun Heo */ 10618864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10628864b4e5STejun Heo } 10638864b4e5STejun Heo 1064dce90d47STejun Heo /** 1065dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1066dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1067dce90d47STejun Heo * 1068dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1069dce90d47STejun Heo */ 1070dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1071dce90d47STejun Heo { 1072dce90d47STejun Heo if (pwq) { 1073dce90d47STejun Heo /* 1074dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1075dce90d47STejun Heo * following lock operations are safe. 1076dce90d47STejun Heo */ 1077dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1078dce90d47STejun Heo put_pwq(pwq); 1079dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1080dce90d47STejun Heo } 1081dce90d47STejun Heo } 1082dce90d47STejun Heo 1083112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1084bf4ede01STejun Heo { 1085112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1086bf4ede01STejun Heo 1087bf4ede01STejun Heo trace_workqueue_activate_work(work); 108882607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 108982607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1090112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1091bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1092112202d9STejun Heo pwq->nr_active++; 1093bf4ede01STejun Heo } 1094bf4ede01STejun Heo 1095112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10963aa62497SLai Jiangshan { 1097112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10983aa62497SLai Jiangshan struct work_struct, entry); 10993aa62497SLai Jiangshan 1100112202d9STejun Heo pwq_activate_delayed_work(work); 11013aa62497SLai Jiangshan } 11023aa62497SLai Jiangshan 1103bf4ede01STejun Heo /** 1104112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1105112202d9STejun Heo * @pwq: pwq of interest 1106bf4ede01STejun Heo * @color: color of work which left the queue 1107bf4ede01STejun Heo * 1108bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1109112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1110bf4ede01STejun Heo * 1111bf4ede01STejun Heo * CONTEXT: 1112d565ed63STejun Heo * spin_lock_irq(pool->lock). 1113bf4ede01STejun Heo */ 1114112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1115bf4ede01STejun Heo { 11168864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1117bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11188864b4e5STejun Heo goto out_put; 1119bf4ede01STejun Heo 1120112202d9STejun Heo pwq->nr_in_flight[color]--; 1121bf4ede01STejun Heo 1122112202d9STejun Heo pwq->nr_active--; 1123112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1124bf4ede01STejun Heo /* one down, submit a delayed one */ 1125112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1126112202d9STejun Heo pwq_activate_first_delayed(pwq); 1127bf4ede01STejun Heo } 1128bf4ede01STejun Heo 1129bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1130112202d9STejun Heo if (likely(pwq->flush_color != color)) 11318864b4e5STejun Heo goto out_put; 1132bf4ede01STejun Heo 1133bf4ede01STejun Heo /* are there still in-flight works? */ 1134112202d9STejun Heo if (pwq->nr_in_flight[color]) 11358864b4e5STejun Heo goto out_put; 1136bf4ede01STejun Heo 1137112202d9STejun Heo /* this pwq is done, clear flush_color */ 1138112202d9STejun Heo pwq->flush_color = -1; 1139bf4ede01STejun Heo 1140bf4ede01STejun Heo /* 1141112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1142bf4ede01STejun Heo * will handle the rest. 1143bf4ede01STejun Heo */ 1144112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1145112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11468864b4e5STejun Heo out_put: 11478864b4e5STejun Heo put_pwq(pwq); 1148bf4ede01STejun Heo } 1149bf4ede01STejun Heo 115036e227d2STejun Heo /** 1151bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 115236e227d2STejun Heo * @work: work item to steal 115336e227d2STejun Heo * @is_dwork: @work is a delayed_work 1154bbb68dfaSTejun Heo * @flags: place to store irq state 115536e227d2STejun Heo * 115636e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1157d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 115836e227d2STejun Heo * 1159d185af30SYacine Belkadi * Return: 116036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 116136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 116236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1163bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1164bbb68dfaSTejun Heo * for arbitrarily long 116536e227d2STejun Heo * 1166d185af30SYacine Belkadi * Note: 1167bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1168e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1169e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1170e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1171bbb68dfaSTejun Heo * 1172bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1173bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1174bbb68dfaSTejun Heo * 1175e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1176bf4ede01STejun Heo */ 1177bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1178bbb68dfaSTejun Heo unsigned long *flags) 1179bf4ede01STejun Heo { 1180d565ed63STejun Heo struct worker_pool *pool; 1181112202d9STejun Heo struct pool_workqueue *pwq; 1182bf4ede01STejun Heo 1183bbb68dfaSTejun Heo local_irq_save(*flags); 1184bbb68dfaSTejun Heo 118536e227d2STejun Heo /* try to steal the timer if it exists */ 118636e227d2STejun Heo if (is_dwork) { 118736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 118836e227d2STejun Heo 1189e0aecdd8STejun Heo /* 1190e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1191e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1192e0aecdd8STejun Heo * running on the local CPU. 1193e0aecdd8STejun Heo */ 119436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 119536e227d2STejun Heo return 1; 119636e227d2STejun Heo } 119736e227d2STejun Heo 119836e227d2STejun Heo /* try to claim PENDING the normal way */ 1199bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1200bf4ede01STejun Heo return 0; 1201bf4ede01STejun Heo 1202bf4ede01STejun Heo /* 1203bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1204bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1205bf4ede01STejun Heo */ 1206d565ed63STejun Heo pool = get_work_pool(work); 1207d565ed63STejun Heo if (!pool) 1208bbb68dfaSTejun Heo goto fail; 1209bf4ede01STejun Heo 1210d565ed63STejun Heo spin_lock(&pool->lock); 1211bf4ede01STejun Heo /* 1212112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1213112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1214112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1215112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1216112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12170b3dae68SLai Jiangshan * item is currently queued on that pool. 1218bf4ede01STejun Heo */ 1219112202d9STejun Heo pwq = get_work_pwq(work); 1220112202d9STejun Heo if (pwq && pwq->pool == pool) { 1221bf4ede01STejun Heo debug_work_deactivate(work); 12223aa62497SLai Jiangshan 12233aa62497SLai Jiangshan /* 122416062836STejun Heo * A delayed work item cannot be grabbed directly because 122516062836STejun Heo * it might have linked NO_COLOR work items which, if left 1226112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 122716062836STejun Heo * management later on and cause stall. Make sure the work 122816062836STejun Heo * item is activated before grabbing. 12293aa62497SLai Jiangshan */ 12303aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1231112202d9STejun Heo pwq_activate_delayed_work(work); 12323aa62497SLai Jiangshan 1233bf4ede01STejun Heo list_del_init(&work->entry); 12349c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 123536e227d2STejun Heo 1236112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12374468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12384468a00fSLai Jiangshan 1239d565ed63STejun Heo spin_unlock(&pool->lock); 124036e227d2STejun Heo return 1; 1241bf4ede01STejun Heo } 1242d565ed63STejun Heo spin_unlock(&pool->lock); 1243bbb68dfaSTejun Heo fail: 1244bbb68dfaSTejun Heo local_irq_restore(*flags); 1245bbb68dfaSTejun Heo if (work_is_canceling(work)) 1246bbb68dfaSTejun Heo return -ENOENT; 1247bbb68dfaSTejun Heo cpu_relax(); 124836e227d2STejun Heo return -EAGAIN; 1249bf4ede01STejun Heo } 1250bf4ede01STejun Heo 1251bf4ede01STejun Heo /** 1252706026c2STejun Heo * insert_work - insert a work into a pool 1253112202d9STejun Heo * @pwq: pwq @work belongs to 12544690c4abSTejun Heo * @work: work to insert 12554690c4abSTejun Heo * @head: insertion point 12564690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12574690c4abSTejun Heo * 1258112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1259706026c2STejun Heo * work_struct flags. 12604690c4abSTejun Heo * 12614690c4abSTejun Heo * CONTEXT: 1262d565ed63STejun Heo * spin_lock_irq(pool->lock). 1263365970a1SDavid Howells */ 1264112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1265112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1266b89deed3SOleg Nesterov { 1267112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1268e1d8aa9fSFrederic Weisbecker 12694690c4abSTejun Heo /* we own @work, set data and link */ 1270112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12711a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12728864b4e5STejun Heo get_pwq(pwq); 1273e22bee78STejun Heo 1274e22bee78STejun Heo /* 1275c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1276c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1277c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1278e22bee78STejun Heo */ 1279e22bee78STejun Heo smp_mb(); 1280e22bee78STejun Heo 128163d95a91STejun Heo if (__need_more_worker(pool)) 128263d95a91STejun Heo wake_up_worker(pool); 1283b89deed3SOleg Nesterov } 1284b89deed3SOleg Nesterov 1285c8efcc25STejun Heo /* 1286c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12878d03ecfeSTejun Heo * same workqueue. 1288c8efcc25STejun Heo */ 1289c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1290c8efcc25STejun Heo { 1291c8efcc25STejun Heo struct worker *worker; 1292c8efcc25STejun Heo 12938d03ecfeSTejun Heo worker = current_wq_worker(); 1294c8efcc25STejun Heo /* 12958d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12968d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1297c8efcc25STejun Heo */ 1298112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1299c8efcc25STejun Heo } 1300c8efcc25STejun Heo 1301d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13021da177e4SLinus Torvalds struct work_struct *work) 13031da177e4SLinus Torvalds { 1304112202d9STejun Heo struct pool_workqueue *pwq; 1305c9178087STejun Heo struct worker_pool *last_pool; 13061e19ffc6STejun Heo struct list_head *worklist; 13078a2e8e5dSTejun Heo unsigned int work_flags; 1308b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13098930cabaSTejun Heo 13108930cabaSTejun Heo /* 13118930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13128930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13138930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13148930cabaSTejun Heo * happen with IRQ disabled. 13158930cabaSTejun Heo */ 13168930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13171da177e4SLinus Torvalds 1318dc186ad7SThomas Gleixner debug_work_activate(work); 13191e19ffc6STejun Heo 13209ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1321618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1322c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1323e41e704bSTejun Heo return; 13249e8cd2f5STejun Heo retry: 1325df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1326f3421797STejun Heo cpu = raw_smp_processor_id(); 1327df2d5ae4STejun Heo 1328df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1329df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1330c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1331df2d5ae4STejun Heo else 1332df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1333f3421797STejun Heo 133418aa9effSTejun Heo /* 1335c9178087STejun Heo * If @work was previously on a different pool, it might still be 1336c9178087STejun Heo * running there, in which case the work needs to be queued on that 1337c9178087STejun Heo * pool to guarantee non-reentrancy. 133818aa9effSTejun Heo */ 1339c9e7cf27STejun Heo last_pool = get_work_pool(work); 1340112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 134118aa9effSTejun Heo struct worker *worker; 134218aa9effSTejun Heo 1343d565ed63STejun Heo spin_lock(&last_pool->lock); 134418aa9effSTejun Heo 1345c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 134618aa9effSTejun Heo 1347112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1348c9178087STejun Heo pwq = worker->current_pwq; 13498594fadeSLai Jiangshan } else { 135018aa9effSTejun Heo /* meh... not running there, queue here */ 1351d565ed63STejun Heo spin_unlock(&last_pool->lock); 1352112202d9STejun Heo spin_lock(&pwq->pool->lock); 135318aa9effSTejun Heo } 13548930cabaSTejun Heo } else { 1355112202d9STejun Heo spin_lock(&pwq->pool->lock); 13568930cabaSTejun Heo } 1357502ca9d8STejun Heo 13589e8cd2f5STejun Heo /* 13599e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13609e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13619e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1362df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1363df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13649e8cd2f5STejun Heo * make forward-progress. 13659e8cd2f5STejun Heo */ 13669e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13679e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13689e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13699e8cd2f5STejun Heo cpu_relax(); 13709e8cd2f5STejun Heo goto retry; 13719e8cd2f5STejun Heo } 13729e8cd2f5STejun Heo /* oops */ 13739e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13749e8cd2f5STejun Heo wq->name, cpu); 13759e8cd2f5STejun Heo } 13769e8cd2f5STejun Heo 1377112202d9STejun Heo /* pwq determined, queue */ 1378112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1379502ca9d8STejun Heo 1380f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1381112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1382f5b2552bSDan Carpenter return; 1383f5b2552bSDan Carpenter } 13841e19ffc6STejun Heo 1385112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1386112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13871e19ffc6STejun Heo 1388112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1389cdadf009STejun Heo trace_workqueue_activate_work(work); 1390112202d9STejun Heo pwq->nr_active++; 1391112202d9STejun Heo worklist = &pwq->pool->worklist; 139282607adcSTejun Heo if (list_empty(worklist)) 139382607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 13948a2e8e5dSTejun Heo } else { 13958a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1396112202d9STejun Heo worklist = &pwq->delayed_works; 13978a2e8e5dSTejun Heo } 13981e19ffc6STejun Heo 1399112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14001e19ffc6STejun Heo 1401112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14021da177e4SLinus Torvalds } 14031da177e4SLinus Torvalds 14040fcb78c2SRolf Eike Beer /** 1405c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1406c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1407c1a220e7SZhang Rui * @wq: workqueue to use 1408c1a220e7SZhang Rui * @work: work to queue 1409c1a220e7SZhang Rui * 1410c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1411c1a220e7SZhang Rui * can't go away. 1412d185af30SYacine Belkadi * 1413d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1414c1a220e7SZhang Rui */ 1415d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1416d4283e93STejun Heo struct work_struct *work) 1417c1a220e7SZhang Rui { 1418d4283e93STejun Heo bool ret = false; 14198930cabaSTejun Heo unsigned long flags; 14208930cabaSTejun Heo 14218930cabaSTejun Heo local_irq_save(flags); 1422c1a220e7SZhang Rui 142322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14244690c4abSTejun Heo __queue_work(cpu, wq, work); 1425d4283e93STejun Heo ret = true; 1426c1a220e7SZhang Rui } 14278930cabaSTejun Heo 14288930cabaSTejun Heo local_irq_restore(flags); 1429c1a220e7SZhang Rui return ret; 1430c1a220e7SZhang Rui } 1431ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1432c1a220e7SZhang Rui 1433d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14341da177e4SLinus Torvalds { 143552bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14361da177e4SLinus Torvalds 1437e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 143860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14391da177e4SLinus Torvalds } 14401438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14411da177e4SLinus Torvalds 14427beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 144352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14441da177e4SLinus Torvalds { 14457beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14467beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14471da177e4SLinus Torvalds 14487beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14497beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1450fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1451fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14527beb2edfSTejun Heo 14538852aac2STejun Heo /* 14548852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14558852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14568852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14578852aac2STejun Heo * on that there's no such delay when @delay is 0. 14588852aac2STejun Heo */ 14598852aac2STejun Heo if (!delay) { 14608852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14618852aac2STejun Heo return; 14628852aac2STejun Heo } 14638852aac2STejun Heo 14647beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14657beb2edfSTejun Heo 146660c057bcSLai Jiangshan dwork->wq = wq; 1467874bbfe6SShaohua Li /* timer isn't guaranteed to run in this cpu, record earlier */ 1468874bbfe6SShaohua Li if (cpu == WORK_CPU_UNBOUND) 1469874bbfe6SShaohua Li cpu = raw_smp_processor_id(); 14701265057fSTejun Heo dwork->cpu = cpu; 14717beb2edfSTejun Heo timer->expires = jiffies + delay; 14727beb2edfSTejun Heo 14737beb2edfSTejun Heo add_timer_on(timer, cpu); 14747beb2edfSTejun Heo } 14751da177e4SLinus Torvalds 14760fcb78c2SRolf Eike Beer /** 14770fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14780fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14790fcb78c2SRolf Eike Beer * @wq: workqueue to use 1480af9997e4SRandy Dunlap * @dwork: work to queue 14810fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14820fcb78c2SRolf Eike Beer * 1483d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1484715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1485715f1300STejun Heo * execution. 14860fcb78c2SRolf Eike Beer */ 1487d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 148852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14897a6bc1cdSVenkatesh Pallipadi { 149052bad64dSDavid Howells struct work_struct *work = &dwork->work; 1491d4283e93STejun Heo bool ret = false; 14928930cabaSTejun Heo unsigned long flags; 14938930cabaSTejun Heo 14948930cabaSTejun Heo /* read the comment in __queue_work() */ 14958930cabaSTejun Heo local_irq_save(flags); 14967a6bc1cdSVenkatesh Pallipadi 149722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14987beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1499d4283e93STejun Heo ret = true; 15007a6bc1cdSVenkatesh Pallipadi } 15018930cabaSTejun Heo 15028930cabaSTejun Heo local_irq_restore(flags); 15037a6bc1cdSVenkatesh Pallipadi return ret; 15047a6bc1cdSVenkatesh Pallipadi } 1505ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15061da177e4SLinus Torvalds 1507c8e55f36STejun Heo /** 15088376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15098376fe22STejun Heo * @cpu: CPU number to execute work on 15108376fe22STejun Heo * @wq: workqueue to use 15118376fe22STejun Heo * @dwork: work to queue 15128376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15138376fe22STejun Heo * 15148376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15158376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15168376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15178376fe22STejun Heo * current state. 15188376fe22STejun Heo * 1519d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15208376fe22STejun Heo * pending and its timer was modified. 15218376fe22STejun Heo * 1522e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15238376fe22STejun Heo * See try_to_grab_pending() for details. 15248376fe22STejun Heo */ 15258376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15268376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15278376fe22STejun Heo { 15288376fe22STejun Heo unsigned long flags; 15298376fe22STejun Heo int ret; 15308376fe22STejun Heo 15318376fe22STejun Heo do { 15328376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15338376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15348376fe22STejun Heo 15358376fe22STejun Heo if (likely(ret >= 0)) { 15368376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15378376fe22STejun Heo local_irq_restore(flags); 15388376fe22STejun Heo } 15398376fe22STejun Heo 15408376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15418376fe22STejun Heo return ret; 15428376fe22STejun Heo } 15438376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15448376fe22STejun Heo 15458376fe22STejun Heo /** 1546c8e55f36STejun Heo * worker_enter_idle - enter idle state 1547c8e55f36STejun Heo * @worker: worker which is entering idle state 1548c8e55f36STejun Heo * 1549c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1550c8e55f36STejun Heo * necessary. 1551c8e55f36STejun Heo * 1552c8e55f36STejun Heo * LOCKING: 1553d565ed63STejun Heo * spin_lock_irq(pool->lock). 1554c8e55f36STejun Heo */ 1555c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15561da177e4SLinus Torvalds { 1557bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1558c8e55f36STejun Heo 15596183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15606183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15616183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15626183c009STejun Heo return; 1563c8e55f36STejun Heo 1564051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1565cb444766STejun Heo worker->flags |= WORKER_IDLE; 1566bd7bdd43STejun Heo pool->nr_idle++; 1567e22bee78STejun Heo worker->last_active = jiffies; 1568c8e55f36STejun Heo 1569c8e55f36STejun Heo /* idle_list is LIFO */ 1570bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1571db7bccf4STejun Heo 157263d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1573628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1574cb444766STejun Heo 1575544ecf31STejun Heo /* 1576706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1577d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1578628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1579628c78e7STejun Heo * unbind is not in progress. 1580544ecf31STejun Heo */ 158124647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1582bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1583e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1584c8e55f36STejun Heo } 1585c8e55f36STejun Heo 1586c8e55f36STejun Heo /** 1587c8e55f36STejun Heo * worker_leave_idle - leave idle state 1588c8e55f36STejun Heo * @worker: worker which is leaving idle state 1589c8e55f36STejun Heo * 1590c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1591c8e55f36STejun Heo * 1592c8e55f36STejun Heo * LOCKING: 1593d565ed63STejun Heo * spin_lock_irq(pool->lock). 1594c8e55f36STejun Heo */ 1595c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1596c8e55f36STejun Heo { 1597bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1598c8e55f36STejun Heo 15996183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16006183c009STejun Heo return; 1601d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1602bd7bdd43STejun Heo pool->nr_idle--; 1603c8e55f36STejun Heo list_del_init(&worker->entry); 1604c8e55f36STejun Heo } 1605c8e55f36STejun Heo 1606f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1607c34056a3STejun Heo { 1608c34056a3STejun Heo struct worker *worker; 1609c34056a3STejun Heo 1610f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1611c8e55f36STejun Heo if (worker) { 1612c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1613affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1614da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1615e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1616e22bee78STejun Heo worker->flags = WORKER_PREP; 1617c8e55f36STejun Heo } 1618c34056a3STejun Heo return worker; 1619c34056a3STejun Heo } 1620c34056a3STejun Heo 1621c34056a3STejun Heo /** 16224736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16234736cbf7SLai Jiangshan * @worker: worker to be attached 16244736cbf7SLai Jiangshan * @pool: the target pool 16254736cbf7SLai Jiangshan * 16264736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16274736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16284736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16294736cbf7SLai Jiangshan */ 16304736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16314736cbf7SLai Jiangshan struct worker_pool *pool) 16324736cbf7SLai Jiangshan { 16334736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16344736cbf7SLai Jiangshan 16354736cbf7SLai Jiangshan /* 16364736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16374736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16384736cbf7SLai Jiangshan */ 16394736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16404736cbf7SLai Jiangshan 16414736cbf7SLai Jiangshan /* 16424736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 16434736cbf7SLai Jiangshan * stable across this function. See the comments above the 16444736cbf7SLai Jiangshan * flag definition for details. 16454736cbf7SLai Jiangshan */ 16464736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 16474736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 16484736cbf7SLai Jiangshan 16494736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 16504736cbf7SLai Jiangshan 16514736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 16524736cbf7SLai Jiangshan } 16534736cbf7SLai Jiangshan 16544736cbf7SLai Jiangshan /** 165560f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 165660f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 165760f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 165860f5a4bcSLai Jiangshan * 16594736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 16604736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 16614736cbf7SLai Jiangshan * other reference to the pool. 166260f5a4bcSLai Jiangshan */ 166360f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 166460f5a4bcSLai Jiangshan struct worker_pool *pool) 166560f5a4bcSLai Jiangshan { 166660f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 166760f5a4bcSLai Jiangshan 166892f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1669da028469SLai Jiangshan list_del(&worker->node); 1670da028469SLai Jiangshan if (list_empty(&pool->workers)) 167160f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 167292f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 167360f5a4bcSLai Jiangshan 1674b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1675b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1676b62c0751SLai Jiangshan 167760f5a4bcSLai Jiangshan if (detach_completion) 167860f5a4bcSLai Jiangshan complete(detach_completion); 167960f5a4bcSLai Jiangshan } 168060f5a4bcSLai Jiangshan 168160f5a4bcSLai Jiangshan /** 1682c34056a3STejun Heo * create_worker - create a new workqueue worker 168363d95a91STejun Heo * @pool: pool the new worker will belong to 1684c34056a3STejun Heo * 1685051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1686c34056a3STejun Heo * 1687c34056a3STejun Heo * CONTEXT: 1688c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1689c34056a3STejun Heo * 1690d185af30SYacine Belkadi * Return: 1691c34056a3STejun Heo * Pointer to the newly created worker. 1692c34056a3STejun Heo */ 1693bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1694c34056a3STejun Heo { 1695c34056a3STejun Heo struct worker *worker = NULL; 1696f3421797STejun Heo int id = -1; 1697e3c916a4STejun Heo char id_buf[16]; 1698c34056a3STejun Heo 16997cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17007cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1701822d8405STejun Heo if (id < 0) 1702c34056a3STejun Heo goto fail; 1703c34056a3STejun Heo 1704f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1705c34056a3STejun Heo if (!worker) 1706c34056a3STejun Heo goto fail; 1707c34056a3STejun Heo 1708bd7bdd43STejun Heo worker->pool = pool; 1709c34056a3STejun Heo worker->id = id; 1710c34056a3STejun Heo 171129c91e99STejun Heo if (pool->cpu >= 0) 1712e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1713e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1714f3421797STejun Heo else 1715e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1716e3c916a4STejun Heo 1717f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1718e3c916a4STejun Heo "kworker/%s", id_buf); 1719c34056a3STejun Heo if (IS_ERR(worker->task)) 1720c34056a3STejun Heo goto fail; 1721c34056a3STejun Heo 172291151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 172325834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 172491151228SOleg Nesterov 1725da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17264736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1727822d8405STejun Heo 1728051e1850SLai Jiangshan /* start the newly created worker */ 1729051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1730051e1850SLai Jiangshan worker->pool->nr_workers++; 1731051e1850SLai Jiangshan worker_enter_idle(worker); 1732051e1850SLai Jiangshan wake_up_process(worker->task); 1733051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1734051e1850SLai Jiangshan 1735c34056a3STejun Heo return worker; 1736822d8405STejun Heo 1737c34056a3STejun Heo fail: 17389625ab17SLai Jiangshan if (id >= 0) 17397cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1740c34056a3STejun Heo kfree(worker); 1741c34056a3STejun Heo return NULL; 1742c34056a3STejun Heo } 1743c34056a3STejun Heo 1744c34056a3STejun Heo /** 1745c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1746c34056a3STejun Heo * @worker: worker to be destroyed 1747c34056a3STejun Heo * 174873eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 174973eb7fe7SLai Jiangshan * be idle. 1750c8e55f36STejun Heo * 1751c8e55f36STejun Heo * CONTEXT: 175260f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1753c34056a3STejun Heo */ 1754c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1755c34056a3STejun Heo { 1756bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1757c34056a3STejun Heo 1758cd549687STejun Heo lockdep_assert_held(&pool->lock); 1759cd549687STejun Heo 1760c34056a3STejun Heo /* sanity check frenzy */ 17616183c009STejun Heo if (WARN_ON(worker->current_work) || 176273eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 176373eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 17646183c009STejun Heo return; 1765c34056a3STejun Heo 1766bd7bdd43STejun Heo pool->nr_workers--; 1767bd7bdd43STejun Heo pool->nr_idle--; 1768c8e55f36STejun Heo 1769c8e55f36STejun Heo list_del_init(&worker->entry); 1770cb444766STejun Heo worker->flags |= WORKER_DIE; 177160f5a4bcSLai Jiangshan wake_up_process(worker->task); 1772c34056a3STejun Heo } 1773c34056a3STejun Heo 177463d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1775e22bee78STejun Heo { 177663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1777e22bee78STejun Heo 1778d565ed63STejun Heo spin_lock_irq(&pool->lock); 1779e22bee78STejun Heo 17803347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1781e22bee78STejun Heo struct worker *worker; 1782e22bee78STejun Heo unsigned long expires; 1783e22bee78STejun Heo 1784e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 178563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1786e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1787e22bee78STejun Heo 17883347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 178963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 17903347fc9fSLai Jiangshan break; 1791e22bee78STejun Heo } 17923347fc9fSLai Jiangshan 17933347fc9fSLai Jiangshan destroy_worker(worker); 1794e22bee78STejun Heo } 1795e22bee78STejun Heo 1796d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1797e22bee78STejun Heo } 1798e22bee78STejun Heo 1799493a1724STejun Heo static void send_mayday(struct work_struct *work) 1800e22bee78STejun Heo { 1801112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1802112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1803493a1724STejun Heo 18042e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1805e22bee78STejun Heo 1806493008a8STejun Heo if (!wq->rescuer) 1807493a1724STejun Heo return; 1808e22bee78STejun Heo 1809e22bee78STejun Heo /* mayday mayday mayday */ 1810493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 181177668c8bSLai Jiangshan /* 181277668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 181377668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 181477668c8bSLai Jiangshan * rescuer is done with it. 181577668c8bSLai Jiangshan */ 181677668c8bSLai Jiangshan get_pwq(pwq); 1817493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1818e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1819493a1724STejun Heo } 1820e22bee78STejun Heo } 1821e22bee78STejun Heo 1822706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1823e22bee78STejun Heo { 182463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1825e22bee78STejun Heo struct work_struct *work; 1826e22bee78STejun Heo 1827b2d82909STejun Heo spin_lock_irq(&pool->lock); 1828b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1829e22bee78STejun Heo 183063d95a91STejun Heo if (need_to_create_worker(pool)) { 1831e22bee78STejun Heo /* 1832e22bee78STejun Heo * We've been trying to create a new worker but 1833e22bee78STejun Heo * haven't been successful. We might be hitting an 1834e22bee78STejun Heo * allocation deadlock. Send distress signals to 1835e22bee78STejun Heo * rescuers. 1836e22bee78STejun Heo */ 183763d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1838e22bee78STejun Heo send_mayday(work); 1839e22bee78STejun Heo } 1840e22bee78STejun Heo 1841b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1842b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1843e22bee78STejun Heo 184463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1845e22bee78STejun Heo } 1846e22bee78STejun Heo 1847e22bee78STejun Heo /** 1848e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 184963d95a91STejun Heo * @pool: pool to create a new worker for 1850e22bee78STejun Heo * 185163d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1852e22bee78STejun Heo * have at least one idle worker on return from this function. If 1853e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 185463d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1855e22bee78STejun Heo * possible allocation deadlock. 1856e22bee78STejun Heo * 1857c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1858c5aa87bbSTejun Heo * may_start_working() %true. 1859e22bee78STejun Heo * 1860e22bee78STejun Heo * LOCKING: 1861d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1862e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1863e22bee78STejun Heo * manager. 1864e22bee78STejun Heo */ 186529187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1866d565ed63STejun Heo __releases(&pool->lock) 1867d565ed63STejun Heo __acquires(&pool->lock) 1868e22bee78STejun Heo { 1869e22bee78STejun Heo restart: 1870d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18719f9c2364STejun Heo 1872e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 187363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1874e22bee78STejun Heo 1875e22bee78STejun Heo while (true) { 1876051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1877e22bee78STejun Heo break; 1878e22bee78STejun Heo 1879e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 18809f9c2364STejun Heo 188163d95a91STejun Heo if (!need_to_create_worker(pool)) 1882e22bee78STejun Heo break; 1883e22bee78STejun Heo } 1884e22bee78STejun Heo 188563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1886d565ed63STejun Heo spin_lock_irq(&pool->lock); 1887051e1850SLai Jiangshan /* 1888051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1889051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1890051e1850SLai Jiangshan * already become busy. 1891051e1850SLai Jiangshan */ 189263d95a91STejun Heo if (need_to_create_worker(pool)) 1893e22bee78STejun Heo goto restart; 1894e22bee78STejun Heo } 1895e22bee78STejun Heo 1896e22bee78STejun Heo /** 1897e22bee78STejun Heo * manage_workers - manage worker pool 1898e22bee78STejun Heo * @worker: self 1899e22bee78STejun Heo * 1900706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1901e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1902706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1903e22bee78STejun Heo * 1904e22bee78STejun Heo * The caller can safely start processing works on false return. On 1905e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1906e22bee78STejun Heo * and may_start_working() is true. 1907e22bee78STejun Heo * 1908e22bee78STejun Heo * CONTEXT: 1909d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1910e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1911e22bee78STejun Heo * 1912d185af30SYacine Belkadi * Return: 191329187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 191429187a9eSTejun Heo * start processing works, %true if management function was performed and 191529187a9eSTejun Heo * the conditions that the caller verified before calling the function may 191629187a9eSTejun Heo * no longer be true. 1917e22bee78STejun Heo */ 1918e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1919e22bee78STejun Heo { 192063d95a91STejun Heo struct worker_pool *pool = worker->pool; 1921e22bee78STejun Heo 1922bc3a1afcSTejun Heo /* 1923bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1924bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1925bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1926bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1927bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1928bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1929bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1930bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1931bc3a1afcSTejun Heo */ 193234a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 193329187a9eSTejun Heo return false; 19342607d7a6STejun Heo pool->manager = worker; 1935e22bee78STejun Heo 193629187a9eSTejun Heo maybe_create_worker(pool); 1937e22bee78STejun Heo 19382607d7a6STejun Heo pool->manager = NULL; 193934a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 194029187a9eSTejun Heo return true; 1941e22bee78STejun Heo } 1942e22bee78STejun Heo 1943a62428c0STejun Heo /** 1944a62428c0STejun Heo * process_one_work - process single work 1945c34056a3STejun Heo * @worker: self 1946a62428c0STejun Heo * @work: work to process 1947a62428c0STejun Heo * 1948a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1949a62428c0STejun Heo * process a single work including synchronization against and 1950a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1951a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1952a62428c0STejun Heo * call this function to process a work. 1953a62428c0STejun Heo * 1954a62428c0STejun Heo * CONTEXT: 1955d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1956a62428c0STejun Heo */ 1957c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 1958d565ed63STejun Heo __releases(&pool->lock) 1959d565ed63STejun Heo __acquires(&pool->lock) 19601da177e4SLinus Torvalds { 1961112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1962bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1963112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 196473f53c4aSTejun Heo int work_color; 19657e11629dSTejun Heo struct worker *collision; 19664e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 19674e6045f1SJohannes Berg /* 1968a62428c0STejun Heo * It is permissible to free the struct work_struct from 1969a62428c0STejun Heo * inside the function that is called from it, this we need to 1970a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1971a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1972a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 19734e6045f1SJohannes Berg */ 19744d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 19754d82a1deSPeter Zijlstra 19764d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 19774e6045f1SJohannes Berg #endif 1978807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 197985327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1980ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 198125511a47STejun Heo 19827e11629dSTejun Heo /* 19837e11629dSTejun Heo * A single work shouldn't be executed concurrently by 19847e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 19857e11629dSTejun Heo * already processing the work. If so, defer the work to the 19867e11629dSTejun Heo * currently executing one. 19877e11629dSTejun Heo */ 1988c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 19897e11629dSTejun Heo if (unlikely(collision)) { 19907e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 19917e11629dSTejun Heo return; 19927e11629dSTejun Heo } 19931da177e4SLinus Torvalds 19948930cabaSTejun Heo /* claim and dequeue */ 19951da177e4SLinus Torvalds debug_work_deactivate(work); 1996c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 1997c34056a3STejun Heo worker->current_work = work; 1998a2c1c57bSTejun Heo worker->current_func = work->func; 1999112202d9STejun Heo worker->current_pwq = pwq; 200073f53c4aSTejun Heo work_color = get_work_color(work); 20017a22ad75STejun Heo 2002a62428c0STejun Heo list_del_init(&work->entry); 2003a62428c0STejun Heo 2004649027d7STejun Heo /* 2005228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2006228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2007228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2008228f1d00SLai Jiangshan * execution of the pending work items. 2009fb0e7bebSTejun Heo */ 2010fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2011228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2012fb0e7bebSTejun Heo 2013974271c4STejun Heo /* 2014a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2015a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2016a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2017a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2018228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2019974271c4STejun Heo */ 2020a489a03eSLai Jiangshan if (need_more_worker(pool)) 202163d95a91STejun Heo wake_up_worker(pool); 2022974271c4STejun Heo 20238930cabaSTejun Heo /* 20247c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2025d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 202623657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 202723657bb1STejun Heo * disabled. 20288930cabaSTejun Heo */ 20297c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20301da177e4SLinus Torvalds 2031d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2032365970a1SDavid Howells 2033112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20343295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2035e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2036a2c1c57bSTejun Heo worker->current_func(work); 2037e36c886aSArjan van de Ven /* 2038e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2039e36c886aSArjan van de Ven * point will only record its address. 2040e36c886aSArjan van de Ven */ 2041e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20423295f0efSIngo Molnar lock_map_release(&lockdep_map); 2043112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 20441da177e4SLinus Torvalds 2045d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2046044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2047044c782cSValentin Ilie " last function: %pf\n", 2048a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2049a2c1c57bSTejun Heo worker->current_func); 2050d5abe669SPeter Zijlstra debug_show_held_locks(current); 2051d5abe669SPeter Zijlstra dump_stack(); 2052d5abe669SPeter Zijlstra } 2053d5abe669SPeter Zijlstra 2054b22ce278STejun Heo /* 2055b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2056b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2057b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2058b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2059789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2060789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2061b22ce278STejun Heo */ 20623e28e377SJoe Lawrence cond_resched_rcu_qs(); 2063b22ce278STejun Heo 2064d565ed63STejun Heo spin_lock_irq(&pool->lock); 2065a62428c0STejun Heo 2066fb0e7bebSTejun Heo /* clear cpu intensive status */ 2067fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2068fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2069fb0e7bebSTejun Heo 2070a62428c0STejun Heo /* we're done with it, release */ 207142f8570fSSasha Levin hash_del(&worker->hentry); 2072c34056a3STejun Heo worker->current_work = NULL; 2073a2c1c57bSTejun Heo worker->current_func = NULL; 2074112202d9STejun Heo worker->current_pwq = NULL; 20753d1cb205STejun Heo worker->desc_valid = false; 2076112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 20771da177e4SLinus Torvalds } 20781da177e4SLinus Torvalds 2079affee4b2STejun Heo /** 2080affee4b2STejun Heo * process_scheduled_works - process scheduled works 2081affee4b2STejun Heo * @worker: self 2082affee4b2STejun Heo * 2083affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2084affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2085affee4b2STejun Heo * fetches a work from the top and executes it. 2086affee4b2STejun Heo * 2087affee4b2STejun Heo * CONTEXT: 2088d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2089affee4b2STejun Heo * multiple times. 2090affee4b2STejun Heo */ 2091affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 20921da177e4SLinus Torvalds { 2093affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2094affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2095a62428c0STejun Heo struct work_struct, entry); 2096c34056a3STejun Heo process_one_work(worker, work); 2097a62428c0STejun Heo } 20981da177e4SLinus Torvalds } 20991da177e4SLinus Torvalds 21004690c4abSTejun Heo /** 21014690c4abSTejun Heo * worker_thread - the worker thread function 2102c34056a3STejun Heo * @__worker: self 21034690c4abSTejun Heo * 2104c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2105c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2106c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2107c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2108c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2109d185af30SYacine Belkadi * 2110d185af30SYacine Belkadi * Return: 0 21114690c4abSTejun Heo */ 2112c34056a3STejun Heo static int worker_thread(void *__worker) 21131da177e4SLinus Torvalds { 2114c34056a3STejun Heo struct worker *worker = __worker; 2115bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21161da177e4SLinus Torvalds 2117e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2118e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2119c8e55f36STejun Heo woke_up: 2120d565ed63STejun Heo spin_lock_irq(&pool->lock); 2121affee4b2STejun Heo 2122a9ab775bSTejun Heo /* am I supposed to die? */ 2123a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2124d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2125a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2126e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 212760f5a4bcSLai Jiangshan 212860f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21297cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 213060f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 213160f5a4bcSLai Jiangshan kfree(worker); 2132c8e55f36STejun Heo return 0; 2133c8e55f36STejun Heo } 2134c8e55f36STejun Heo 2135c8e55f36STejun Heo worker_leave_idle(worker); 2136db7bccf4STejun Heo recheck: 2137e22bee78STejun Heo /* no more worker necessary? */ 213863d95a91STejun Heo if (!need_more_worker(pool)) 2139e22bee78STejun Heo goto sleep; 2140e22bee78STejun Heo 2141e22bee78STejun Heo /* do we need to manage? */ 214263d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2143e22bee78STejun Heo goto recheck; 2144e22bee78STejun Heo 2145c8e55f36STejun Heo /* 2146c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2147c8e55f36STejun Heo * preparing to process a work or actually processing it. 2148c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2149c8e55f36STejun Heo */ 21506183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2151c8e55f36STejun Heo 2152e22bee78STejun Heo /* 2153a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2154a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2155a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2156a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2157a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2158e22bee78STejun Heo */ 2159a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2160e22bee78STejun Heo 2161e22bee78STejun Heo do { 2162affee4b2STejun Heo struct work_struct *work = 2163bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2164affee4b2STejun Heo struct work_struct, entry); 2165affee4b2STejun Heo 216682607adcSTejun Heo pool->watchdog_ts = jiffies; 216782607adcSTejun Heo 2168c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2169affee4b2STejun Heo /* optimization path, not strictly necessary */ 2170affee4b2STejun Heo process_one_work(worker, work); 2171affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2172affee4b2STejun Heo process_scheduled_works(worker); 2173affee4b2STejun Heo } else { 2174c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2175affee4b2STejun Heo process_scheduled_works(worker); 2176affee4b2STejun Heo } 217763d95a91STejun Heo } while (keep_working(pool)); 2178affee4b2STejun Heo 2179228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2180d313dd85STejun Heo sleep: 2181c8e55f36STejun Heo /* 2182d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2183d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2184d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2185d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2186d565ed63STejun Heo * event. 2187c8e55f36STejun Heo */ 2188c8e55f36STejun Heo worker_enter_idle(worker); 2189c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2190d565ed63STejun Heo spin_unlock_irq(&pool->lock); 21911da177e4SLinus Torvalds schedule(); 2192c8e55f36STejun Heo goto woke_up; 21931da177e4SLinus Torvalds } 21941da177e4SLinus Torvalds 2195e22bee78STejun Heo /** 2196e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2197111c225aSTejun Heo * @__rescuer: self 2198e22bee78STejun Heo * 2199e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2200493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2201e22bee78STejun Heo * 2202706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2203e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2204e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2205e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2206e22bee78STejun Heo * the problem rescuer solves. 2207e22bee78STejun Heo * 2208706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2209706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2210e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2211e22bee78STejun Heo * 2212e22bee78STejun Heo * This should happen rarely. 2213d185af30SYacine Belkadi * 2214d185af30SYacine Belkadi * Return: 0 2215e22bee78STejun Heo */ 2216111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2217e22bee78STejun Heo { 2218111c225aSTejun Heo struct worker *rescuer = __rescuer; 2219111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2220e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22214d595b86SLai Jiangshan bool should_stop; 2222e22bee78STejun Heo 2223e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2224111c225aSTejun Heo 2225111c225aSTejun Heo /* 2226111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2227111c225aSTejun Heo * doesn't participate in concurrency management. 2228111c225aSTejun Heo */ 2229111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2230e22bee78STejun Heo repeat: 2231e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22321da177e4SLinus Torvalds 22334d595b86SLai Jiangshan /* 22344d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22354d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22364d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22374d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22384d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22394d595b86SLai Jiangshan * list is always empty on exit. 22404d595b86SLai Jiangshan */ 22414d595b86SLai Jiangshan should_stop = kthread_should_stop(); 22421da177e4SLinus Torvalds 2243493a1724STejun Heo /* see whether any pwq is asking for help */ 22442e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2245493a1724STejun Heo 2246493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2247493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2248493a1724STejun Heo struct pool_workqueue, mayday_node); 2249112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2250e22bee78STejun Heo struct work_struct *work, *n; 225182607adcSTejun Heo bool first = true; 2252e22bee78STejun Heo 2253e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2254493a1724STejun Heo list_del_init(&pwq->mayday_node); 2255493a1724STejun Heo 22562e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2257e22bee78STejun Heo 225851697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 225951697d39SLai Jiangshan 226051697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2261b3104104SLai Jiangshan rescuer->pool = pool; 2262e22bee78STejun Heo 2263e22bee78STejun Heo /* 2264e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2265e22bee78STejun Heo * process'em. 2266e22bee78STejun Heo */ 22670479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 226882607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 226982607adcSTejun Heo if (get_work_pwq(work) == pwq) { 227082607adcSTejun Heo if (first) 227182607adcSTejun Heo pool->watchdog_ts = jiffies; 2272e22bee78STejun Heo move_linked_works(work, scheduled, &n); 227382607adcSTejun Heo } 227482607adcSTejun Heo first = false; 227582607adcSTejun Heo } 2276e22bee78STejun Heo 2277008847f6SNeilBrown if (!list_empty(scheduled)) { 2278e22bee78STejun Heo process_scheduled_works(rescuer); 22797576958aSTejun Heo 22807576958aSTejun Heo /* 2281008847f6SNeilBrown * The above execution of rescued work items could 2282008847f6SNeilBrown * have created more to rescue through 2283008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2284008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2285008847f6SNeilBrown * that such back-to-back work items, which may be 2286008847f6SNeilBrown * being used to relieve memory pressure, don't 2287008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2288008847f6SNeilBrown */ 2289008847f6SNeilBrown if (need_to_create_worker(pool)) { 2290008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2291008847f6SNeilBrown get_pwq(pwq); 2292008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2293008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2294008847f6SNeilBrown } 2295008847f6SNeilBrown } 2296008847f6SNeilBrown 2297008847f6SNeilBrown /* 229877668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 229913b1d625SLai Jiangshan * go away while we're still attached to it. 230077668c8bSLai Jiangshan */ 230177668c8bSLai Jiangshan put_pwq(pwq); 230277668c8bSLai Jiangshan 230377668c8bSLai Jiangshan /* 2304d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 23057576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23067576958aSTejun Heo * and stalling the execution. 23077576958aSTejun Heo */ 2308d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 230963d95a91STejun Heo wake_up_worker(pool); 23107576958aSTejun Heo 2311b3104104SLai Jiangshan rescuer->pool = NULL; 231213b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 231313b1d625SLai Jiangshan 231413b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 231513b1d625SLai Jiangshan 231613b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 23171da177e4SLinus Torvalds } 23181da177e4SLinus Torvalds 23192e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2320493a1724STejun Heo 23214d595b86SLai Jiangshan if (should_stop) { 23224d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23234d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23244d595b86SLai Jiangshan return 0; 23254d595b86SLai Jiangshan } 23264d595b86SLai Jiangshan 2327111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2328111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2329e22bee78STejun Heo schedule(); 2330e22bee78STejun Heo goto repeat; 23311da177e4SLinus Torvalds } 23321da177e4SLinus Torvalds 2333fca839c0STejun Heo /** 2334fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2335fca839c0STejun Heo * @target_wq: workqueue being flushed 2336fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2337fca839c0STejun Heo * 2338fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2339fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2340fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2341fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2342fca839c0STejun Heo * a deadlock. 2343fca839c0STejun Heo */ 2344fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2345fca839c0STejun Heo struct work_struct *target_work) 2346fca839c0STejun Heo { 2347fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2348fca839c0STejun Heo struct worker *worker; 2349fca839c0STejun Heo 2350fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2351fca839c0STejun Heo return; 2352fca839c0STejun Heo 2353fca839c0STejun Heo worker = current_wq_worker(); 2354fca839c0STejun Heo 2355fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2356fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2357fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 2358fca839c0STejun Heo WARN_ONCE(worker && (worker->current_pwq->wq->flags & WQ_MEM_RECLAIM), 2359fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2360fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2361fca839c0STejun Heo target_wq->name, target_func); 2362fca839c0STejun Heo } 2363fca839c0STejun Heo 2364fc2e4d70SOleg Nesterov struct wq_barrier { 2365fc2e4d70SOleg Nesterov struct work_struct work; 2366fc2e4d70SOleg Nesterov struct completion done; 23672607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2368fc2e4d70SOleg Nesterov }; 2369fc2e4d70SOleg Nesterov 2370fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2371fc2e4d70SOleg Nesterov { 2372fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2373fc2e4d70SOleg Nesterov complete(&barr->done); 2374fc2e4d70SOleg Nesterov } 2375fc2e4d70SOleg Nesterov 23764690c4abSTejun Heo /** 23774690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2378112202d9STejun Heo * @pwq: pwq to insert barrier into 23794690c4abSTejun Heo * @barr: wq_barrier to insert 2380affee4b2STejun Heo * @target: target work to attach @barr to 2381affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23824690c4abSTejun Heo * 2383affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2384affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2385affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2386affee4b2STejun Heo * cpu. 2387affee4b2STejun Heo * 2388affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2389affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2390affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2391affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2392affee4b2STejun Heo * after a work with LINKED flag set. 2393affee4b2STejun Heo * 2394affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2395112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23964690c4abSTejun Heo * 23974690c4abSTejun Heo * CONTEXT: 2398d565ed63STejun Heo * spin_lock_irq(pool->lock). 23994690c4abSTejun Heo */ 2400112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2401affee4b2STejun Heo struct wq_barrier *barr, 2402affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2403fc2e4d70SOleg Nesterov { 2404affee4b2STejun Heo struct list_head *head; 2405affee4b2STejun Heo unsigned int linked = 0; 2406affee4b2STejun Heo 2407dc186ad7SThomas Gleixner /* 2408d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2409dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2410dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2411dc186ad7SThomas Gleixner * might deadlock. 2412dc186ad7SThomas Gleixner */ 2413ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 241422df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2415fc2e4d70SOleg Nesterov init_completion(&barr->done); 24162607d7a6STejun Heo barr->task = current; 241783c22520SOleg Nesterov 2418affee4b2STejun Heo /* 2419affee4b2STejun Heo * If @target is currently being executed, schedule the 2420affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2421affee4b2STejun Heo */ 2422affee4b2STejun Heo if (worker) 2423affee4b2STejun Heo head = worker->scheduled.next; 2424affee4b2STejun Heo else { 2425affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2426affee4b2STejun Heo 2427affee4b2STejun Heo head = target->entry.next; 2428affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2429affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2430affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2431affee4b2STejun Heo } 2432affee4b2STejun Heo 2433dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2434112202d9STejun Heo insert_work(pwq, &barr->work, head, 2435affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2436fc2e4d70SOleg Nesterov } 2437fc2e4d70SOleg Nesterov 243873f53c4aSTejun Heo /** 2439112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 244073f53c4aSTejun Heo * @wq: workqueue being flushed 244173f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 244273f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 244373f53c4aSTejun Heo * 2444112202d9STejun Heo * Prepare pwqs for workqueue flushing. 244573f53c4aSTejun Heo * 2446112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2447112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2448112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2449112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2450112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 245173f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 245273f53c4aSTejun Heo * 245373f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 245473f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 245573f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 245673f53c4aSTejun Heo * is returned. 245773f53c4aSTejun Heo * 2458112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 245973f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 246073f53c4aSTejun Heo * advanced to @work_color. 246173f53c4aSTejun Heo * 246273f53c4aSTejun Heo * CONTEXT: 24633c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 246473f53c4aSTejun Heo * 2465d185af30SYacine Belkadi * Return: 246673f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 246773f53c4aSTejun Heo * otherwise. 246873f53c4aSTejun Heo */ 2469112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 247073f53c4aSTejun Heo int flush_color, int work_color) 24711da177e4SLinus Torvalds { 247273f53c4aSTejun Heo bool wait = false; 247349e3cf44STejun Heo struct pool_workqueue *pwq; 24741da177e4SLinus Torvalds 247573f53c4aSTejun Heo if (flush_color >= 0) { 24766183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2477112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2478dc186ad7SThomas Gleixner } 247914441960SOleg Nesterov 248049e3cf44STejun Heo for_each_pwq(pwq, wq) { 2481112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24821da177e4SLinus Torvalds 2483b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 248473f53c4aSTejun Heo 248573f53c4aSTejun Heo if (flush_color >= 0) { 24866183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 248773f53c4aSTejun Heo 2488112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2489112202d9STejun Heo pwq->flush_color = flush_color; 2490112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 249173f53c4aSTejun Heo wait = true; 24921da177e4SLinus Torvalds } 249373f53c4aSTejun Heo } 249473f53c4aSTejun Heo 249573f53c4aSTejun Heo if (work_color >= 0) { 24966183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2497112202d9STejun Heo pwq->work_color = work_color; 249873f53c4aSTejun Heo } 249973f53c4aSTejun Heo 2500b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25011da177e4SLinus Torvalds } 25021da177e4SLinus Torvalds 2503112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 250473f53c4aSTejun Heo complete(&wq->first_flusher->done); 250573f53c4aSTejun Heo 250673f53c4aSTejun Heo return wait; 250783c22520SOleg Nesterov } 25081da177e4SLinus Torvalds 25090fcb78c2SRolf Eike Beer /** 25101da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25110fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25121da177e4SLinus Torvalds * 2513c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2514c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25151da177e4SLinus Torvalds */ 25167ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25171da177e4SLinus Torvalds { 251873f53c4aSTejun Heo struct wq_flusher this_flusher = { 251973f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 252073f53c4aSTejun Heo .flush_color = -1, 252173f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 252273f53c4aSTejun Heo }; 252373f53c4aSTejun Heo int next_color; 2524b1f4ec17SOleg Nesterov 25253295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25263295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 252773f53c4aSTejun Heo 25283c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 252973f53c4aSTejun Heo 253073f53c4aSTejun Heo /* 253173f53c4aSTejun Heo * Start-to-wait phase 253273f53c4aSTejun Heo */ 253373f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 253473f53c4aSTejun Heo 253573f53c4aSTejun Heo if (next_color != wq->flush_color) { 253673f53c4aSTejun Heo /* 253773f53c4aSTejun Heo * Color space is not full. The current work_color 253873f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 253973f53c4aSTejun Heo * by one. 254073f53c4aSTejun Heo */ 25416183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 254273f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 254373f53c4aSTejun Heo wq->work_color = next_color; 254473f53c4aSTejun Heo 254573f53c4aSTejun Heo if (!wq->first_flusher) { 254673f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25476183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 254873f53c4aSTejun Heo 254973f53c4aSTejun Heo wq->first_flusher = &this_flusher; 255073f53c4aSTejun Heo 2551112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 255273f53c4aSTejun Heo wq->work_color)) { 255373f53c4aSTejun Heo /* nothing to flush, done */ 255473f53c4aSTejun Heo wq->flush_color = next_color; 255573f53c4aSTejun Heo wq->first_flusher = NULL; 255673f53c4aSTejun Heo goto out_unlock; 255773f53c4aSTejun Heo } 255873f53c4aSTejun Heo } else { 255973f53c4aSTejun Heo /* wait in queue */ 25606183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 256173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2562112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 256373f53c4aSTejun Heo } 256473f53c4aSTejun Heo } else { 256573f53c4aSTejun Heo /* 256673f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 256773f53c4aSTejun Heo * The next flush completion will assign us 256873f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 256973f53c4aSTejun Heo */ 257073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 257173f53c4aSTejun Heo } 257273f53c4aSTejun Heo 2573fca839c0STejun Heo check_flush_dependency(wq, NULL); 2574fca839c0STejun Heo 25753c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 257673f53c4aSTejun Heo 257773f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 257873f53c4aSTejun Heo 257973f53c4aSTejun Heo /* 258073f53c4aSTejun Heo * Wake-up-and-cascade phase 258173f53c4aSTejun Heo * 258273f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 258373f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 258473f53c4aSTejun Heo */ 258573f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 258673f53c4aSTejun Heo return; 258773f53c4aSTejun Heo 25883c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 258973f53c4aSTejun Heo 25904ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25914ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25924ce48b37STejun Heo goto out_unlock; 25934ce48b37STejun Heo 259473f53c4aSTejun Heo wq->first_flusher = NULL; 259573f53c4aSTejun Heo 25966183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25976183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 259873f53c4aSTejun Heo 259973f53c4aSTejun Heo while (true) { 260073f53c4aSTejun Heo struct wq_flusher *next, *tmp; 260173f53c4aSTejun Heo 260273f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 260373f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 260473f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 260573f53c4aSTejun Heo break; 260673f53c4aSTejun Heo list_del_init(&next->list); 260773f53c4aSTejun Heo complete(&next->done); 260873f53c4aSTejun Heo } 260973f53c4aSTejun Heo 26106183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 261173f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 261273f53c4aSTejun Heo 261373f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 261473f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 261573f53c4aSTejun Heo 261673f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 261773f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 261873f53c4aSTejun Heo /* 261973f53c4aSTejun Heo * Assign the same color to all overflowed 262073f53c4aSTejun Heo * flushers, advance work_color and append to 262173f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 262273f53c4aSTejun Heo * phase for these overflowed flushers. 262373f53c4aSTejun Heo */ 262473f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 262573f53c4aSTejun Heo tmp->flush_color = wq->work_color; 262673f53c4aSTejun Heo 262773f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 262873f53c4aSTejun Heo 262973f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 263073f53c4aSTejun Heo &wq->flusher_queue); 2631112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 263273f53c4aSTejun Heo } 263373f53c4aSTejun Heo 263473f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26356183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 263673f53c4aSTejun Heo break; 263773f53c4aSTejun Heo } 263873f53c4aSTejun Heo 263973f53c4aSTejun Heo /* 264073f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2641112202d9STejun Heo * the new first flusher and arm pwqs. 264273f53c4aSTejun Heo */ 26436183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26446183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 264573f53c4aSTejun Heo 264673f53c4aSTejun Heo list_del_init(&next->list); 264773f53c4aSTejun Heo wq->first_flusher = next; 264873f53c4aSTejun Heo 2649112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 265073f53c4aSTejun Heo break; 265173f53c4aSTejun Heo 265273f53c4aSTejun Heo /* 265373f53c4aSTejun Heo * Meh... this color is already done, clear first 265473f53c4aSTejun Heo * flusher and repeat cascading. 265573f53c4aSTejun Heo */ 265673f53c4aSTejun Heo wq->first_flusher = NULL; 265773f53c4aSTejun Heo } 265873f53c4aSTejun Heo 265973f53c4aSTejun Heo out_unlock: 26603c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26611da177e4SLinus Torvalds } 26621dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 26631da177e4SLinus Torvalds 26649c5a2ba7STejun Heo /** 26659c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26669c5a2ba7STejun Heo * @wq: workqueue to drain 26679c5a2ba7STejun Heo * 26689c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26699c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26709c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2671b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 26729c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26739c5a2ba7STejun Heo * takes too long. 26749c5a2ba7STejun Heo */ 26759c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26769c5a2ba7STejun Heo { 26779c5a2ba7STejun Heo unsigned int flush_cnt = 0; 267849e3cf44STejun Heo struct pool_workqueue *pwq; 26799c5a2ba7STejun Heo 26809c5a2ba7STejun Heo /* 26819c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26829c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2683618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26849c5a2ba7STejun Heo */ 268587fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26869c5a2ba7STejun Heo if (!wq->nr_drainers++) 2687618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 268887fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26899c5a2ba7STejun Heo reflush: 26909c5a2ba7STejun Heo flush_workqueue(wq); 26919c5a2ba7STejun Heo 2692b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 269376af4d93STejun Heo 269449e3cf44STejun Heo for_each_pwq(pwq, wq) { 2695fa2563e4SThomas Tuttle bool drained; 26969c5a2ba7STejun Heo 2697b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2698112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2699b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2700fa2563e4SThomas Tuttle 2701fa2563e4SThomas Tuttle if (drained) 27029c5a2ba7STejun Heo continue; 27039c5a2ba7STejun Heo 27049c5a2ba7STejun Heo if (++flush_cnt == 10 || 27059c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2706c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27079c5a2ba7STejun Heo wq->name, flush_cnt); 270876af4d93STejun Heo 2709b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27109c5a2ba7STejun Heo goto reflush; 27119c5a2ba7STejun Heo } 27129c5a2ba7STejun Heo 27139c5a2ba7STejun Heo if (!--wq->nr_drainers) 2714618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 271587fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27169c5a2ba7STejun Heo } 27179c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27189c5a2ba7STejun Heo 2719606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2720baf59022STejun Heo { 2721baf59022STejun Heo struct worker *worker = NULL; 2722c9e7cf27STejun Heo struct worker_pool *pool; 2723112202d9STejun Heo struct pool_workqueue *pwq; 2724baf59022STejun Heo 2725baf59022STejun Heo might_sleep(); 2726baf59022STejun Heo 2727fa1b54e6STejun Heo local_irq_disable(); 2728fa1b54e6STejun Heo pool = get_work_pool(work); 2729fa1b54e6STejun Heo if (!pool) { 2730fa1b54e6STejun Heo local_irq_enable(); 2731fa1b54e6STejun Heo return false; 2732fa1b54e6STejun Heo } 2733fa1b54e6STejun Heo 2734fa1b54e6STejun Heo spin_lock(&pool->lock); 27350b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2736112202d9STejun Heo pwq = get_work_pwq(work); 2737112202d9STejun Heo if (pwq) { 2738112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2739baf59022STejun Heo goto already_gone; 2740606a5020STejun Heo } else { 2741c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2742baf59022STejun Heo if (!worker) 2743baf59022STejun Heo goto already_gone; 2744112202d9STejun Heo pwq = worker->current_pwq; 2745606a5020STejun Heo } 2746baf59022STejun Heo 2747fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2748fca839c0STejun Heo 2749112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2750d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2751baf59022STejun Heo 2752e159489bSTejun Heo /* 2753e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2754e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2755e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2756e159489bSTejun Heo * access. 2757e159489bSTejun Heo */ 2758493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2759112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2760e159489bSTejun Heo else 2761112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2762112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2763e159489bSTejun Heo 2764baf59022STejun Heo return true; 2765baf59022STejun Heo already_gone: 2766d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2767baf59022STejun Heo return false; 2768baf59022STejun Heo } 2769baf59022STejun Heo 2770db700897SOleg Nesterov /** 2771401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2772401a8d04STejun Heo * @work: the work to flush 2773db700897SOleg Nesterov * 2774606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2775606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2776401a8d04STejun Heo * 2777d185af30SYacine Belkadi * Return: 2778401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2779401a8d04STejun Heo * %false if it was already idle. 2780db700897SOleg Nesterov */ 2781401a8d04STejun Heo bool flush_work(struct work_struct *work) 2782db700897SOleg Nesterov { 278312997d1aSBjorn Helgaas struct wq_barrier barr; 278412997d1aSBjorn Helgaas 27850976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27860976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27870976dfc1SStephen Boyd 278812997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 278912997d1aSBjorn Helgaas wait_for_completion(&barr.done); 279012997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 279112997d1aSBjorn Helgaas return true; 279212997d1aSBjorn Helgaas } else { 279312997d1aSBjorn Helgaas return false; 279412997d1aSBjorn Helgaas } 2795606a5020STejun Heo } 2796db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2797db700897SOleg Nesterov 27988603e1b3STejun Heo struct cwt_wait { 27998603e1b3STejun Heo wait_queue_t wait; 28008603e1b3STejun Heo struct work_struct *work; 28018603e1b3STejun Heo }; 28028603e1b3STejun Heo 28038603e1b3STejun Heo static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) 28048603e1b3STejun Heo { 28058603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 28068603e1b3STejun Heo 28078603e1b3STejun Heo if (cwait->work != key) 28088603e1b3STejun Heo return 0; 28098603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 28108603e1b3STejun Heo } 28118603e1b3STejun Heo 281236e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2813401a8d04STejun Heo { 28148603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2815bbb68dfaSTejun Heo unsigned long flags; 28161f1f642eSOleg Nesterov int ret; 28171f1f642eSOleg Nesterov 28181f1f642eSOleg Nesterov do { 2819bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2820bbb68dfaSTejun Heo /* 28218603e1b3STejun Heo * If someone else is already canceling, wait for it to 28228603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 28238603e1b3STejun Heo * because we may get scheduled between @work's completion 28248603e1b3STejun Heo * and the other canceling task resuming and clearing 28258603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 28268603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 28278603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 28288603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 28298603e1b3STejun Heo * we're hogging the CPU. 28308603e1b3STejun Heo * 28318603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 28328603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 28338603e1b3STejun Heo * wake function which matches @work along with exclusive 28348603e1b3STejun Heo * wait and wakeup. 2835bbb68dfaSTejun Heo */ 28368603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 28378603e1b3STejun Heo struct cwt_wait cwait; 28388603e1b3STejun Heo 28398603e1b3STejun Heo init_wait(&cwait.wait); 28408603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 28418603e1b3STejun Heo cwait.work = work; 28428603e1b3STejun Heo 28438603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 28448603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 28458603e1b3STejun Heo if (work_is_canceling(work)) 28468603e1b3STejun Heo schedule(); 28478603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 28488603e1b3STejun Heo } 28491f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28501f1f642eSOleg Nesterov 2851bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2852bbb68dfaSTejun Heo mark_work_canceling(work); 2853bbb68dfaSTejun Heo local_irq_restore(flags); 2854bbb68dfaSTejun Heo 2855606a5020STejun Heo flush_work(work); 28567a22ad75STejun Heo clear_work_data(work); 28578603e1b3STejun Heo 28588603e1b3STejun Heo /* 28598603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 28608603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 28618603e1b3STejun Heo * visible there. 28628603e1b3STejun Heo */ 28638603e1b3STejun Heo smp_mb(); 28648603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 28658603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 28668603e1b3STejun Heo 28671f1f642eSOleg Nesterov return ret; 28681f1f642eSOleg Nesterov } 28691f1f642eSOleg Nesterov 28706e84d644SOleg Nesterov /** 2871401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2872401a8d04STejun Heo * @work: the work to cancel 28736e84d644SOleg Nesterov * 2874401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2875401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2876401a8d04STejun Heo * another workqueue. On return from this function, @work is 2877401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28781f1f642eSOleg Nesterov * 2879401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2880401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28816e84d644SOleg Nesterov * 2882401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28836e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2884401a8d04STejun Heo * 2885d185af30SYacine Belkadi * Return: 2886401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28876e84d644SOleg Nesterov */ 2888401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28896e84d644SOleg Nesterov { 289036e227d2STejun Heo return __cancel_work_timer(work, false); 2891b89deed3SOleg Nesterov } 289228e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2893b89deed3SOleg Nesterov 28946e84d644SOleg Nesterov /** 2895401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2896401a8d04STejun Heo * @dwork: the delayed work to flush 28976e84d644SOleg Nesterov * 2898401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2899401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2900401a8d04STejun Heo * considers the last queueing instance of @dwork. 29011f1f642eSOleg Nesterov * 2902d185af30SYacine Belkadi * Return: 2903401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2904401a8d04STejun Heo * %false if it was already idle. 29056e84d644SOleg Nesterov */ 2906401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2907401a8d04STejun Heo { 29088930cabaSTejun Heo local_irq_disable(); 2909401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 291060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29118930cabaSTejun Heo local_irq_enable(); 2912401a8d04STejun Heo return flush_work(&dwork->work); 2913401a8d04STejun Heo } 2914401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2915401a8d04STejun Heo 2916401a8d04STejun Heo /** 291757b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 291857b30ae7STejun Heo * @dwork: delayed_work to cancel 291909383498STejun Heo * 2920d185af30SYacine Belkadi * Kill off a pending delayed_work. 2921d185af30SYacine Belkadi * 2922d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2923d185af30SYacine Belkadi * pending. 2924d185af30SYacine Belkadi * 2925d185af30SYacine Belkadi * Note: 2926d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2927d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2928d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 292909383498STejun Heo * 293057b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 293109383498STejun Heo */ 293257b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 293309383498STejun Heo { 293457b30ae7STejun Heo unsigned long flags; 293557b30ae7STejun Heo int ret; 293657b30ae7STejun Heo 293757b30ae7STejun Heo do { 293857b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 293957b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 294057b30ae7STejun Heo 294157b30ae7STejun Heo if (unlikely(ret < 0)) 294257b30ae7STejun Heo return false; 294357b30ae7STejun Heo 29447c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29457c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 294657b30ae7STejun Heo local_irq_restore(flags); 2947c0158ca6SDan Magenheimer return ret; 294809383498STejun Heo } 294957b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 295009383498STejun Heo 295109383498STejun Heo /** 2952401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2953401a8d04STejun Heo * @dwork: the delayed work cancel 2954401a8d04STejun Heo * 2955401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2956401a8d04STejun Heo * 2957d185af30SYacine Belkadi * Return: 2958401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2959401a8d04STejun Heo */ 2960401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29616e84d644SOleg Nesterov { 296236e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29636e84d644SOleg Nesterov } 2964f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29651da177e4SLinus Torvalds 29660fcb78c2SRolf Eike Beer /** 296731ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2968b6136773SAndrew Morton * @func: the function to call 2969b6136773SAndrew Morton * 297031ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 297131ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2972b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 297331ddd871STejun Heo * 2974d185af30SYacine Belkadi * Return: 297531ddd871STejun Heo * 0 on success, -errno on failure. 2976b6136773SAndrew Morton */ 297765f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 297815316ba8SChristoph Lameter { 297915316ba8SChristoph Lameter int cpu; 298038f51568SNamhyung Kim struct work_struct __percpu *works; 298115316ba8SChristoph Lameter 2982b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2983b6136773SAndrew Morton if (!works) 298415316ba8SChristoph Lameter return -ENOMEM; 2985b6136773SAndrew Morton 298695402b38SGautham R Shenoy get_online_cpus(); 298793981800STejun Heo 298815316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29899bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29909bfb1839SIngo Molnar 29919bfb1839SIngo Molnar INIT_WORK(work, func); 29928de6d308SOleg Nesterov schedule_work_on(cpu, work); 299315316ba8SChristoph Lameter } 299493981800STejun Heo 299593981800STejun Heo for_each_online_cpu(cpu) 29968616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 299793981800STejun Heo 299895402b38SGautham R Shenoy put_online_cpus(); 2999b6136773SAndrew Morton free_percpu(works); 300015316ba8SChristoph Lameter return 0; 300115316ba8SChristoph Lameter } 300215316ba8SChristoph Lameter 3003eef6a7d5SAlan Stern /** 30041fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30051fa44ecaSJames Bottomley * @fn: the function to execute 30061fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30071fa44ecaSJames Bottomley * be available when the work executes) 30081fa44ecaSJames Bottomley * 30091fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30101fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30111fa44ecaSJames Bottomley * 3012d185af30SYacine Belkadi * Return: 0 - function was executed 30131fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30141fa44ecaSJames Bottomley */ 301565f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30161fa44ecaSJames Bottomley { 30171fa44ecaSJames Bottomley if (!in_interrupt()) { 301865f27f38SDavid Howells fn(&ew->work); 30191fa44ecaSJames Bottomley return 0; 30201fa44ecaSJames Bottomley } 30211fa44ecaSJames Bottomley 302265f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30231fa44ecaSJames Bottomley schedule_work(&ew->work); 30241fa44ecaSJames Bottomley 30251fa44ecaSJames Bottomley return 1; 30261fa44ecaSJames Bottomley } 30271fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30281fa44ecaSJames Bottomley 30297a4e344cSTejun Heo /** 30307a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 30317a4e344cSTejun Heo * @attrs: workqueue_attrs to free 30327a4e344cSTejun Heo * 30337a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 30347a4e344cSTejun Heo */ 30357a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 30367a4e344cSTejun Heo { 30377a4e344cSTejun Heo if (attrs) { 30387a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 30397a4e344cSTejun Heo kfree(attrs); 30407a4e344cSTejun Heo } 30417a4e344cSTejun Heo } 30427a4e344cSTejun Heo 30437a4e344cSTejun Heo /** 30447a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 30457a4e344cSTejun Heo * @gfp_mask: allocation mask to use 30467a4e344cSTejun Heo * 30477a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3048d185af30SYacine Belkadi * return it. 3049d185af30SYacine Belkadi * 3050d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 30517a4e344cSTejun Heo */ 30527a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 30537a4e344cSTejun Heo { 30547a4e344cSTejun Heo struct workqueue_attrs *attrs; 30557a4e344cSTejun Heo 30567a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 30577a4e344cSTejun Heo if (!attrs) 30587a4e344cSTejun Heo goto fail; 30597a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 30607a4e344cSTejun Heo goto fail; 30617a4e344cSTejun Heo 306213e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 30637a4e344cSTejun Heo return attrs; 30647a4e344cSTejun Heo fail: 30657a4e344cSTejun Heo free_workqueue_attrs(attrs); 30667a4e344cSTejun Heo return NULL; 30677a4e344cSTejun Heo } 30687a4e344cSTejun Heo 306929c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 307029c91e99STejun Heo const struct workqueue_attrs *from) 307129c91e99STejun Heo { 307229c91e99STejun Heo to->nice = from->nice; 307329c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 30742865a8fbSShaohua Li /* 30752865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 30762865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 30772865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 30782865a8fbSShaohua Li */ 30792865a8fbSShaohua Li to->no_numa = from->no_numa; 308029c91e99STejun Heo } 308129c91e99STejun Heo 308229c91e99STejun Heo /* hash value of the content of @attr */ 308329c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 308429c91e99STejun Heo { 308529c91e99STejun Heo u32 hash = 0; 308629c91e99STejun Heo 308729c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 308813e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 308913e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 309029c91e99STejun Heo return hash; 309129c91e99STejun Heo } 309229c91e99STejun Heo 309329c91e99STejun Heo /* content equality test */ 309429c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 309529c91e99STejun Heo const struct workqueue_attrs *b) 309629c91e99STejun Heo { 309729c91e99STejun Heo if (a->nice != b->nice) 309829c91e99STejun Heo return false; 309929c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 310029c91e99STejun Heo return false; 310129c91e99STejun Heo return true; 310229c91e99STejun Heo } 310329c91e99STejun Heo 31047a4e344cSTejun Heo /** 31057a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 31067a4e344cSTejun Heo * @pool: worker_pool to initialize 31077a4e344cSTejun Heo * 3108402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3109d185af30SYacine Belkadi * 3110d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 311129c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 311229c91e99STejun Heo * on @pool safely to release it. 31137a4e344cSTejun Heo */ 31147a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 31154e1a1f9aSTejun Heo { 31164e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 311729c91e99STejun Heo pool->id = -1; 311829c91e99STejun Heo pool->cpu = -1; 3119f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 31204e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 312182607adcSTejun Heo pool->watchdog_ts = jiffies; 31224e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 31234e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 31244e1a1f9aSTejun Heo hash_init(pool->busy_hash); 31254e1a1f9aSTejun Heo 31264e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 31274e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 31284e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 31294e1a1f9aSTejun Heo 31304e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 31314e1a1f9aSTejun Heo (unsigned long)pool); 31324e1a1f9aSTejun Heo 31334e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 313492f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3135da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 31367a4e344cSTejun Heo 31377cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 313829c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 313929c91e99STejun Heo pool->refcnt = 1; 314029c91e99STejun Heo 314129c91e99STejun Heo /* shouldn't fail above this point */ 31427a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 31437a4e344cSTejun Heo if (!pool->attrs) 31447a4e344cSTejun Heo return -ENOMEM; 31457a4e344cSTejun Heo return 0; 31464e1a1f9aSTejun Heo } 31474e1a1f9aSTejun Heo 3148e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3149e2dca7adSTejun Heo { 3150e2dca7adSTejun Heo struct workqueue_struct *wq = 3151e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3152e2dca7adSTejun Heo 3153e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3154e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3155e2dca7adSTejun Heo else 3156e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3157e2dca7adSTejun Heo 3158e2dca7adSTejun Heo kfree(wq->rescuer); 3159e2dca7adSTejun Heo kfree(wq); 3160e2dca7adSTejun Heo } 3161e2dca7adSTejun Heo 316229c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 316329c91e99STejun Heo { 316429c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 316529c91e99STejun Heo 31667cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 316729c91e99STejun Heo free_workqueue_attrs(pool->attrs); 316829c91e99STejun Heo kfree(pool); 316929c91e99STejun Heo } 317029c91e99STejun Heo 317129c91e99STejun Heo /** 317229c91e99STejun Heo * put_unbound_pool - put a worker_pool 317329c91e99STejun Heo * @pool: worker_pool to put 317429c91e99STejun Heo * 317529c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3176c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3177c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3178c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3179a892caccSTejun Heo * 3180a892caccSTejun Heo * Should be called with wq_pool_mutex held. 318129c91e99STejun Heo */ 318229c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 318329c91e99STejun Heo { 318460f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 318529c91e99STejun Heo struct worker *worker; 318629c91e99STejun Heo 3187a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3188a892caccSTejun Heo 3189a892caccSTejun Heo if (--pool->refcnt) 319029c91e99STejun Heo return; 319129c91e99STejun Heo 319229c91e99STejun Heo /* sanity checks */ 319361d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3194a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 319529c91e99STejun Heo return; 319629c91e99STejun Heo 319729c91e99STejun Heo /* release id and unhash */ 319829c91e99STejun Heo if (pool->id >= 0) 319929c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 320029c91e99STejun Heo hash_del(&pool->hash_node); 320129c91e99STejun Heo 3202c5aa87bbSTejun Heo /* 3203c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3204c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 320592f9c5c4SLai Jiangshan * attach_mutex. 3206c5aa87bbSTejun Heo */ 320729c91e99STejun Heo mutex_lock(&pool->manager_arb); 320829c91e99STejun Heo 320960f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 32101037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 321129c91e99STejun Heo destroy_worker(worker); 321229c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 321329c91e99STejun Heo spin_unlock_irq(&pool->lock); 321460f5a4bcSLai Jiangshan 321592f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3216da028469SLai Jiangshan if (!list_empty(&pool->workers)) 321760f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 321892f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 321960f5a4bcSLai Jiangshan 322060f5a4bcSLai Jiangshan if (pool->detach_completion) 322160f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 322260f5a4bcSLai Jiangshan 322329c91e99STejun Heo mutex_unlock(&pool->manager_arb); 322429c91e99STejun Heo 322529c91e99STejun Heo /* shut down the timers */ 322629c91e99STejun Heo del_timer_sync(&pool->idle_timer); 322729c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 322829c91e99STejun Heo 322929c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 323029c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 323129c91e99STejun Heo } 323229c91e99STejun Heo 323329c91e99STejun Heo /** 323429c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 323529c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 323629c91e99STejun Heo * 323729c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 323829c91e99STejun Heo * reference count and return it. If there already is a matching 323929c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3240d185af30SYacine Belkadi * create a new one. 3241a892caccSTejun Heo * 3242a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3243d185af30SYacine Belkadi * 3244d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3245d185af30SYacine Belkadi * On failure, %NULL. 324629c91e99STejun Heo */ 324729c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 324829c91e99STejun Heo { 324929c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 325029c91e99STejun Heo struct worker_pool *pool; 3251f3f90ad4STejun Heo int node; 3252e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 325329c91e99STejun Heo 3254a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 325529c91e99STejun Heo 325629c91e99STejun Heo /* do we already have a matching pool? */ 325729c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 325829c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 325929c91e99STejun Heo pool->refcnt++; 32603fb1823cSLai Jiangshan return pool; 326129c91e99STejun Heo } 326229c91e99STejun Heo } 326329c91e99STejun Heo 3264e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3265e2273584SXunlei Pang if (wq_numa_enabled) { 3266e2273584SXunlei Pang for_each_node(node) { 3267e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3268e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3269e2273584SXunlei Pang target_node = node; 3270e2273584SXunlei Pang break; 3271e2273584SXunlei Pang } 3272e2273584SXunlei Pang } 3273e2273584SXunlei Pang } 3274e2273584SXunlei Pang 327529c91e99STejun Heo /* nope, create a new one */ 3276e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 327729c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 327829c91e99STejun Heo goto fail; 327929c91e99STejun Heo 32808864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 328129c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3282e2273584SXunlei Pang pool->node = target_node; 328329c91e99STejun Heo 32842865a8fbSShaohua Li /* 32852865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 32862865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 32872865a8fbSShaohua Li */ 32882865a8fbSShaohua Li pool->attrs->no_numa = false; 32892865a8fbSShaohua Li 329029c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 329129c91e99STejun Heo goto fail; 329229c91e99STejun Heo 329329c91e99STejun Heo /* create and start the initial worker */ 3294051e1850SLai Jiangshan if (!create_worker(pool)) 329529c91e99STejun Heo goto fail; 329629c91e99STejun Heo 329729c91e99STejun Heo /* install */ 329829c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 32993fb1823cSLai Jiangshan 330029c91e99STejun Heo return pool; 330129c91e99STejun Heo fail: 330229c91e99STejun Heo if (pool) 330329c91e99STejun Heo put_unbound_pool(pool); 330429c91e99STejun Heo return NULL; 330529c91e99STejun Heo } 330629c91e99STejun Heo 33078864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 33088864b4e5STejun Heo { 33098864b4e5STejun Heo kmem_cache_free(pwq_cache, 33108864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 33118864b4e5STejun Heo } 33128864b4e5STejun Heo 33138864b4e5STejun Heo /* 33148864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 33158864b4e5STejun Heo * and needs to be destroyed. 33168864b4e5STejun Heo */ 33178864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 33188864b4e5STejun Heo { 33198864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 33208864b4e5STejun Heo unbound_release_work); 33218864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 33228864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3323bc0caf09STejun Heo bool is_last; 33248864b4e5STejun Heo 33258864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 33268864b4e5STejun Heo return; 33278864b4e5STejun Heo 33283c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 33298864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3330bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 33313c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 33328864b4e5STejun Heo 3333a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 33348864b4e5STejun Heo put_unbound_pool(pool); 3335a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3336a892caccSTejun Heo 33378864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 33388864b4e5STejun Heo 33398864b4e5STejun Heo /* 33408864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3341e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 33428864b4e5STejun Heo */ 3343e2dca7adSTejun Heo if (is_last) 3344e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 33456029a918STejun Heo } 33468864b4e5STejun Heo 33470fbd95aaSTejun Heo /** 3348699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 33490fbd95aaSTejun Heo * @pwq: target pool_workqueue 33500fbd95aaSTejun Heo * 3351699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3352699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3353699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 33540fbd95aaSTejun Heo */ 3355699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 33560fbd95aaSTejun Heo { 3357699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3358699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3359699ce097STejun Heo 3360699ce097STejun Heo /* for @wq->saved_max_active */ 3361a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3362699ce097STejun Heo 3363699ce097STejun Heo /* fast exit for non-freezable wqs */ 3364699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3365699ce097STejun Heo return; 3366699ce097STejun Heo 3367a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3368699ce097STejun Heo 336974b414eaSLai Jiangshan /* 337074b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 337174b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 337274b414eaSLai Jiangshan * is updated and visible. 337374b414eaSLai Jiangshan */ 337474b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3375699ce097STejun Heo pwq->max_active = wq->saved_max_active; 33760fbd95aaSTejun Heo 33770fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 33780fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 33790fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3380951a078aSLai Jiangshan 3381951a078aSLai Jiangshan /* 3382951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3383951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3384951a078aSLai Jiangshan */ 3385951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3386699ce097STejun Heo } else { 3387699ce097STejun Heo pwq->max_active = 0; 3388699ce097STejun Heo } 3389699ce097STejun Heo 3390a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 33910fbd95aaSTejun Heo } 33920fbd95aaSTejun Heo 3393e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3394f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3395f147f29eSTejun Heo struct worker_pool *pool) 3396d2c1d404STejun Heo { 3397d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3398d2c1d404STejun Heo 3399e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3400e50aba9aSTejun Heo 3401d2c1d404STejun Heo pwq->pool = pool; 3402d2c1d404STejun Heo pwq->wq = wq; 3403d2c1d404STejun Heo pwq->flush_color = -1; 34048864b4e5STejun Heo pwq->refcnt = 1; 3405d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 34061befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3407d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 34088864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3409f147f29eSTejun Heo } 3410d2c1d404STejun Heo 3411f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 34121befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3413f147f29eSTejun Heo { 3414f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3415f147f29eSTejun Heo 3416f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 341775ccf595STejun Heo 34181befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 34191befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 34201befcf30STejun Heo return; 34211befcf30STejun Heo 342229b1cb41SLai Jiangshan /* set the matching work_color */ 342375ccf595STejun Heo pwq->work_color = wq->work_color; 3424983ca25eSTejun Heo 3425983ca25eSTejun Heo /* sync max_active to the current setting */ 3426983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3427983ca25eSTejun Heo 3428983ca25eSTejun Heo /* link in @pwq */ 34299e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3430df2d5ae4STejun Heo } 34316029a918STejun Heo 3432f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3433f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3434f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3435f147f29eSTejun Heo { 3436f147f29eSTejun Heo struct worker_pool *pool; 3437f147f29eSTejun Heo struct pool_workqueue *pwq; 3438f147f29eSTejun Heo 3439f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3440f147f29eSTejun Heo 3441f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3442f147f29eSTejun Heo if (!pool) 3443f147f29eSTejun Heo return NULL; 3444f147f29eSTejun Heo 3445e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3446f147f29eSTejun Heo if (!pwq) { 3447f147f29eSTejun Heo put_unbound_pool(pool); 3448f147f29eSTejun Heo return NULL; 3449f147f29eSTejun Heo } 3450f147f29eSTejun Heo 3451f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3452f147f29eSTejun Heo return pwq; 3453d2c1d404STejun Heo } 3454d2c1d404STejun Heo 34554c16bd32STejun Heo /** 345630186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3457042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 34584c16bd32STejun Heo * @node: the target NUMA node 34594c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 34604c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 34614c16bd32STejun Heo * 34624c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 34634c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3464d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 34654c16bd32STejun Heo * 34664c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 34674c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 34684c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 34694c16bd32STejun Heo * @attrs->cpumask. 34704c16bd32STejun Heo * 34714c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 34724c16bd32STejun Heo * stable. 3473d185af30SYacine Belkadi * 3474d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3475d185af30SYacine Belkadi * %false if equal. 34764c16bd32STejun Heo */ 34774c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 34784c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 34794c16bd32STejun Heo { 3480d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 34814c16bd32STejun Heo goto use_dfl; 34824c16bd32STejun Heo 34834c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 34844c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 34854c16bd32STejun Heo if (cpu_going_down >= 0) 34864c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 34874c16bd32STejun Heo 34884c16bd32STejun Heo if (cpumask_empty(cpumask)) 34894c16bd32STejun Heo goto use_dfl; 34904c16bd32STejun Heo 34914c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 34924c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 34934c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 34944c16bd32STejun Heo 34954c16bd32STejun Heo use_dfl: 34964c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 34974c16bd32STejun Heo return false; 34984c16bd32STejun Heo } 34994c16bd32STejun Heo 35001befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 35011befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 35021befcf30STejun Heo int node, 35031befcf30STejun Heo struct pool_workqueue *pwq) 35041befcf30STejun Heo { 35051befcf30STejun Heo struct pool_workqueue *old_pwq; 35061befcf30STejun Heo 35075b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 35081befcf30STejun Heo lockdep_assert_held(&wq->mutex); 35091befcf30STejun Heo 35101befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 35111befcf30STejun Heo link_pwq(pwq); 35121befcf30STejun Heo 35131befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 35141befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 35151befcf30STejun Heo return old_pwq; 35161befcf30STejun Heo } 35171befcf30STejun Heo 35182d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 35192d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 35202d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 35212d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3522042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 35232d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 35242d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 35252d5f0764SLai Jiangshan }; 35262d5f0764SLai Jiangshan 35272d5f0764SLai Jiangshan /* free the resources after success or abort */ 35282d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 35292d5f0764SLai Jiangshan { 35302d5f0764SLai Jiangshan if (ctx) { 35312d5f0764SLai Jiangshan int node; 35322d5f0764SLai Jiangshan 35332d5f0764SLai Jiangshan for_each_node(node) 35342d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 35352d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 35362d5f0764SLai Jiangshan 35372d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 35382d5f0764SLai Jiangshan 35392d5f0764SLai Jiangshan kfree(ctx); 35402d5f0764SLai Jiangshan } 35412d5f0764SLai Jiangshan } 35422d5f0764SLai Jiangshan 35432d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 35442d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 35452d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 35462d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 35472d5f0764SLai Jiangshan { 35482d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 35492d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 35502d5f0764SLai Jiangshan int node; 35512d5f0764SLai Jiangshan 35522d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 35532d5f0764SLai Jiangshan 35542d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 35552d5f0764SLai Jiangshan GFP_KERNEL); 35562d5f0764SLai Jiangshan 35572d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 35582d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 35592d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 35602d5f0764SLai Jiangshan goto out_free; 35612d5f0764SLai Jiangshan 3562042f7df1SLai Jiangshan /* 3563042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3564042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3565042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3566042f7df1SLai Jiangshan */ 35672d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3568b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3569042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3570042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 35712d5f0764SLai Jiangshan 35722d5f0764SLai Jiangshan /* 35732d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 35742d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 35752d5f0764SLai Jiangshan * pools. 35762d5f0764SLai Jiangshan */ 35772d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 35782d5f0764SLai Jiangshan 35792d5f0764SLai Jiangshan /* 35802d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 35812d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 35822d5f0764SLai Jiangshan * it even if we don't use it immediately. 35832d5f0764SLai Jiangshan */ 35842d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 35852d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 35862d5f0764SLai Jiangshan goto out_free; 35872d5f0764SLai Jiangshan 35882d5f0764SLai Jiangshan for_each_node(node) { 3589042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 35902d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 35912d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 35922d5f0764SLai Jiangshan goto out_free; 35932d5f0764SLai Jiangshan } else { 35942d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 35952d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 35962d5f0764SLai Jiangshan } 35972d5f0764SLai Jiangshan } 35982d5f0764SLai Jiangshan 3599042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3600042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3601042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 36022d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3603042f7df1SLai Jiangshan 36042d5f0764SLai Jiangshan ctx->wq = wq; 36052d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36062d5f0764SLai Jiangshan return ctx; 36072d5f0764SLai Jiangshan 36082d5f0764SLai Jiangshan out_free: 36092d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 36102d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 36112d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 36122d5f0764SLai Jiangshan return NULL; 36132d5f0764SLai Jiangshan } 36142d5f0764SLai Jiangshan 36152d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 36162d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 36172d5f0764SLai Jiangshan { 36182d5f0764SLai Jiangshan int node; 36192d5f0764SLai Jiangshan 36202d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 36212d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 36222d5f0764SLai Jiangshan 36232d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 36242d5f0764SLai Jiangshan 36252d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 36262d5f0764SLai Jiangshan for_each_node(node) 36272d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 36282d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 36292d5f0764SLai Jiangshan 36302d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 36312d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 36322d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 36332d5f0764SLai Jiangshan 36342d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 36352d5f0764SLai Jiangshan } 36362d5f0764SLai Jiangshan 3637a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3638a0111cf6SLai Jiangshan { 3639a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3640a0111cf6SLai Jiangshan get_online_cpus(); 3641a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3642a0111cf6SLai Jiangshan } 3643a0111cf6SLai Jiangshan 3644a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3645a0111cf6SLai Jiangshan { 3646a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3647a0111cf6SLai Jiangshan put_online_cpus(); 3648a0111cf6SLai Jiangshan } 3649a0111cf6SLai Jiangshan 3650a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3651a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3652a0111cf6SLai Jiangshan { 3653a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3654a0111cf6SLai Jiangshan 3655a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3656a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3657a0111cf6SLai Jiangshan return -EINVAL; 3658a0111cf6SLai Jiangshan 3659a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 3660a0111cf6SLai Jiangshan if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 3661a0111cf6SLai Jiangshan return -EINVAL; 3662a0111cf6SLai Jiangshan 3663a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 36646201171eSwanghaibin if (!ctx) 36656201171eSwanghaibin return -ENOMEM; 3666a0111cf6SLai Jiangshan 3667a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3668a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3669a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3670a0111cf6SLai Jiangshan 36716201171eSwanghaibin return 0; 3672a0111cf6SLai Jiangshan } 3673a0111cf6SLai Jiangshan 36749e8cd2f5STejun Heo /** 36759e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36769e8cd2f5STejun Heo * @wq: the target workqueue 36779e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36789e8cd2f5STejun Heo * 36794c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 36804c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 36814c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 36824c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 36834c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 36844c16bd32STejun Heo * back-to-back will stay on its current pwq. 36859e8cd2f5STejun Heo * 3686d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3687d185af30SYacine Belkadi * 3688d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 36899e8cd2f5STejun Heo */ 36909e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36919e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36929e8cd2f5STejun Heo { 3693a0111cf6SLai Jiangshan int ret; 36949e8cd2f5STejun Heo 3695a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3696a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3697a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 36982d5f0764SLai Jiangshan 36992d5f0764SLai Jiangshan return ret; 37009e8cd2f5STejun Heo } 37019e8cd2f5STejun Heo 37024c16bd32STejun Heo /** 37034c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 37044c16bd32STejun Heo * @wq: the target workqueue 37054c16bd32STejun Heo * @cpu: the CPU coming up or going down 37064c16bd32STejun Heo * @online: whether @cpu is coming up or going down 37074c16bd32STejun Heo * 37084c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 37094c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 37104c16bd32STejun Heo * @wq accordingly. 37114c16bd32STejun Heo * 37124c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 37134c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 37144c16bd32STejun Heo * correct. 37154c16bd32STejun Heo * 37164c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 37174c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 37184c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 37194c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 37204c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 37214c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 37224c16bd32STejun Heo * CPU_DOWN_PREPARE. 37234c16bd32STejun Heo */ 37244c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 37254c16bd32STejun Heo bool online) 37264c16bd32STejun Heo { 37274c16bd32STejun Heo int node = cpu_to_node(cpu); 37284c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 37294c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 37304c16bd32STejun Heo struct workqueue_attrs *target_attrs; 37314c16bd32STejun Heo cpumask_t *cpumask; 37324c16bd32STejun Heo 37334c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37344c16bd32STejun Heo 3735f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3736f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 37374c16bd32STejun Heo return; 37384c16bd32STejun Heo 37394c16bd32STejun Heo /* 37404c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 37414c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 37424c16bd32STejun Heo * CPU hotplug exclusion. 37434c16bd32STejun Heo */ 37444c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 37454c16bd32STejun Heo cpumask = target_attrs->cpumask; 37464c16bd32STejun Heo 37474c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 37484c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 37494c16bd32STejun Heo 37504c16bd32STejun Heo /* 37514c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3752042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3753042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3754042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 37554c16bd32STejun Heo */ 3756042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 37574c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3758f7142ed4SLai Jiangshan return; 37594c16bd32STejun Heo } else { 37604c16bd32STejun Heo goto use_dfl_pwq; 37614c16bd32STejun Heo } 37624c16bd32STejun Heo 37634c16bd32STejun Heo /* create a new pwq */ 37644c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 37654c16bd32STejun Heo if (!pwq) { 37662d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 37674c16bd32STejun Heo wq->name); 376877f300b1SDaeseok Youn goto use_dfl_pwq; 37694c16bd32STejun Heo } 37704c16bd32STejun Heo 3771f7142ed4SLai Jiangshan /* Install the new pwq. */ 37724c16bd32STejun Heo mutex_lock(&wq->mutex); 37734c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 37744c16bd32STejun Heo goto out_unlock; 37754c16bd32STejun Heo 37764c16bd32STejun Heo use_dfl_pwq: 3777f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 37784c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 37794c16bd32STejun Heo get_pwq(wq->dfl_pwq); 37804c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 37814c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 37824c16bd32STejun Heo out_unlock: 37834c16bd32STejun Heo mutex_unlock(&wq->mutex); 37844c16bd32STejun Heo put_pwq_unlocked(old_pwq); 37854c16bd32STejun Heo } 37864c16bd32STejun Heo 378730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37881da177e4SLinus Torvalds { 378949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 37908a2b7538STejun Heo int cpu, ret; 3791e1d8aa9fSFrederic Weisbecker 379230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3793420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3794420c0ddbSTejun Heo if (!wq->cpu_pwqs) 379530cdf249STejun Heo return -ENOMEM; 379630cdf249STejun Heo 379730cdf249STejun Heo for_each_possible_cpu(cpu) { 37987fb98ea7STejun Heo struct pool_workqueue *pwq = 37997fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 38007a62c2c8STejun Heo struct worker_pool *cpu_pools = 3801f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 380230cdf249STejun Heo 3803f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3804f147f29eSTejun Heo 3805f147f29eSTejun Heo mutex_lock(&wq->mutex); 38061befcf30STejun Heo link_pwq(pwq); 3807f147f29eSTejun Heo mutex_unlock(&wq->mutex); 380830cdf249STejun Heo } 380930cdf249STejun Heo return 0; 38108a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 38118a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 38128a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 38138a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 38148a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 38158a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 38168a2b7538STejun Heo return ret; 38179e8cd2f5STejun Heo } else { 38189e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 38199e8cd2f5STejun Heo } 38200f900049STejun Heo } 38210f900049STejun Heo 3822f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3823f3421797STejun Heo const char *name) 3824b71ab8c2STejun Heo { 3825f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3826f3421797STejun Heo 3827f3421797STejun Heo if (max_active < 1 || max_active > lim) 3828044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3829f3421797STejun Heo max_active, name, 1, lim); 3830b71ab8c2STejun Heo 3831f3421797STejun Heo return clamp_val(max_active, 1, lim); 3832b71ab8c2STejun Heo } 3833b71ab8c2STejun Heo 3834b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 383597e37d7bSTejun Heo unsigned int flags, 38361e19ffc6STejun Heo int max_active, 3837eb13ba87SJohannes Berg struct lock_class_key *key, 3838b196be89STejun Heo const char *lock_name, ...) 38393af24433SOleg Nesterov { 3840df2d5ae4STejun Heo size_t tbl_size = 0; 3841ecf6881fSTejun Heo va_list args; 38423af24433SOleg Nesterov struct workqueue_struct *wq; 384349e3cf44STejun Heo struct pool_workqueue *pwq; 3844b196be89STejun Heo 3845cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 3846cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 3847cee22a15SViresh Kumar flags |= WQ_UNBOUND; 3848cee22a15SViresh Kumar 3849ecf6881fSTejun Heo /* allocate wq and format name */ 3850df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 3851ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 3852df2d5ae4STejun Heo 3853df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 3854b196be89STejun Heo if (!wq) 3855d2c1d404STejun Heo return NULL; 3856b196be89STejun Heo 38576029a918STejun Heo if (flags & WQ_UNBOUND) { 38586029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38596029a918STejun Heo if (!wq->unbound_attrs) 38606029a918STejun Heo goto err_free_wq; 38616029a918STejun Heo } 38626029a918STejun Heo 3863ecf6881fSTejun Heo va_start(args, lock_name); 3864ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3865b196be89STejun Heo va_end(args); 38663af24433SOleg Nesterov 3867d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3868b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 38693af24433SOleg Nesterov 3870b196be89STejun Heo /* init wq */ 387197e37d7bSTejun Heo wq->flags = flags; 3872a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 38733c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3874112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 387530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 387673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 387773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3878493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 38793af24433SOleg Nesterov 3880eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3881cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 38823af24433SOleg Nesterov 388330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3884d2c1d404STejun Heo goto err_free_wq; 38851537663fSTejun Heo 3886493008a8STejun Heo /* 3887493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3888493008a8STejun Heo * have a rescuer to guarantee forward progress. 3889493008a8STejun Heo */ 3890493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3891e22bee78STejun Heo struct worker *rescuer; 3892e22bee78STejun Heo 3893f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 3894e22bee78STejun Heo if (!rescuer) 3895d2c1d404STejun Heo goto err_destroy; 3896e22bee78STejun Heo 3897111c225aSTejun Heo rescuer->rescue_wq = wq; 3898111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3899b196be89STejun Heo wq->name); 3900d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3901d2c1d404STejun Heo kfree(rescuer); 3902d2c1d404STejun Heo goto err_destroy; 3903d2c1d404STejun Heo } 3904e22bee78STejun Heo 3905d2c1d404STejun Heo wq->rescuer = rescuer; 390625834c73SPeter Zijlstra kthread_bind_mask(rescuer->task, cpu_possible_mask); 3907e22bee78STejun Heo wake_up_process(rescuer->task); 39083af24433SOleg Nesterov } 39091537663fSTejun Heo 3910226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3911226223abSTejun Heo goto err_destroy; 3912226223abSTejun Heo 39133af24433SOleg Nesterov /* 391468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 391568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 391668e13a67SLai Jiangshan * list. 39173af24433SOleg Nesterov */ 391868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3919a0a1a5fdSTejun Heo 3920a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 392149e3cf44STejun Heo for_each_pwq(pwq, wq) 3922699ce097STejun Heo pwq_adjust_max_active(pwq); 3923a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3924a0a1a5fdSTejun Heo 3925e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 3926a0a1a5fdSTejun Heo 392768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 39283af24433SOleg Nesterov 39293af24433SOleg Nesterov return wq; 3930d2c1d404STejun Heo 3931d2c1d404STejun Heo err_free_wq: 39326029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 39334690c4abSTejun Heo kfree(wq); 3934d2c1d404STejun Heo return NULL; 3935d2c1d404STejun Heo err_destroy: 3936d2c1d404STejun Heo destroy_workqueue(wq); 39374690c4abSTejun Heo return NULL; 39381da177e4SLinus Torvalds } 3939d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 39401da177e4SLinus Torvalds 39413af24433SOleg Nesterov /** 39423af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 39433af24433SOleg Nesterov * @wq: target workqueue 39443af24433SOleg Nesterov * 39453af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 39463af24433SOleg Nesterov */ 39473af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 39483af24433SOleg Nesterov { 394949e3cf44STejun Heo struct pool_workqueue *pwq; 39504c16bd32STejun Heo int node; 39513af24433SOleg Nesterov 39529c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 39539c5a2ba7STejun Heo drain_workqueue(wq); 3954c8efcc25STejun Heo 39556183c009STejun Heo /* sanity checks */ 3956b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 395749e3cf44STejun Heo for_each_pwq(pwq, wq) { 39586183c009STejun Heo int i; 39596183c009STejun Heo 396076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 396176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3962b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39636183c009STejun Heo return; 396476af4d93STejun Heo } 396576af4d93STejun Heo } 396676af4d93STejun Heo 39675c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 39688864b4e5STejun Heo WARN_ON(pwq->nr_active) || 396976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3970b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39716183c009STejun Heo return; 39726183c009STejun Heo } 397376af4d93STejun Heo } 3974b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39756183c009STejun Heo 3976a0a1a5fdSTejun Heo /* 3977a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3978a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3979a0a1a5fdSTejun Heo */ 398068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3981e2dca7adSTejun Heo list_del_rcu(&wq->list); 398268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 39833af24433SOleg Nesterov 3984226223abSTejun Heo workqueue_sysfs_unregister(wq); 3985226223abSTejun Heo 3986e2dca7adSTejun Heo if (wq->rescuer) 3987e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3988e22bee78STejun Heo 39898864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 399029c91e99STejun Heo /* 39918864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 3992e2dca7adSTejun Heo * schedule RCU free. 399329c91e99STejun Heo */ 3994e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 39958864b4e5STejun Heo } else { 39968864b4e5STejun Heo /* 39978864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 39984c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 39994c16bd32STejun Heo * @wq will be freed when the last pwq is released. 40008864b4e5STejun Heo */ 40014c16bd32STejun Heo for_each_node(node) { 40024c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 40034c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 40044c16bd32STejun Heo put_pwq_unlocked(pwq); 40054c16bd32STejun Heo } 40064c16bd32STejun Heo 40074c16bd32STejun Heo /* 40084c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 40094c16bd32STejun Heo * put. Don't access it afterwards. 40104c16bd32STejun Heo */ 40114c16bd32STejun Heo pwq = wq->dfl_pwq; 40124c16bd32STejun Heo wq->dfl_pwq = NULL; 4013dce90d47STejun Heo put_pwq_unlocked(pwq); 401429c91e99STejun Heo } 40153af24433SOleg Nesterov } 40163af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 40173af24433SOleg Nesterov 4018dcd989cbSTejun Heo /** 4019dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4020dcd989cbSTejun Heo * @wq: target workqueue 4021dcd989cbSTejun Heo * @max_active: new max_active value. 4022dcd989cbSTejun Heo * 4023dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4024dcd989cbSTejun Heo * 4025dcd989cbSTejun Heo * CONTEXT: 4026dcd989cbSTejun Heo * Don't call from IRQ context. 4027dcd989cbSTejun Heo */ 4028dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4029dcd989cbSTejun Heo { 403049e3cf44STejun Heo struct pool_workqueue *pwq; 4031dcd989cbSTejun Heo 40328719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 40338719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 40348719dceaSTejun Heo return; 40358719dceaSTejun Heo 4036f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4037dcd989cbSTejun Heo 4038a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4039dcd989cbSTejun Heo 4040dcd989cbSTejun Heo wq->saved_max_active = max_active; 4041dcd989cbSTejun Heo 4042699ce097STejun Heo for_each_pwq(pwq, wq) 4043699ce097STejun Heo pwq_adjust_max_active(pwq); 4044dcd989cbSTejun Heo 4045a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4046dcd989cbSTejun Heo } 4047dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4048dcd989cbSTejun Heo 4049dcd989cbSTejun Heo /** 4050e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4051e6267616STejun Heo * 4052e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4053e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4054d185af30SYacine Belkadi * 4055d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4056e6267616STejun Heo */ 4057e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4058e6267616STejun Heo { 4059e6267616STejun Heo struct worker *worker = current_wq_worker(); 4060e6267616STejun Heo 40616a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4062e6267616STejun Heo } 4063e6267616STejun Heo 4064e6267616STejun Heo /** 4065dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4066dcd989cbSTejun Heo * @cpu: CPU in question 4067dcd989cbSTejun Heo * @wq: target workqueue 4068dcd989cbSTejun Heo * 4069dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4070dcd989cbSTejun Heo * no synchronization around this function and the test result is 4071dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4072dcd989cbSTejun Heo * 4073d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4074d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4075d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4076d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4077d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4078d3251859STejun Heo * 4079d185af30SYacine Belkadi * Return: 4080dcd989cbSTejun Heo * %true if congested, %false otherwise. 4081dcd989cbSTejun Heo */ 4082d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4083dcd989cbSTejun Heo { 40847fb98ea7STejun Heo struct pool_workqueue *pwq; 408576af4d93STejun Heo bool ret; 408676af4d93STejun Heo 408788109453SLai Jiangshan rcu_read_lock_sched(); 40887fb98ea7STejun Heo 4089d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4090d3251859STejun Heo cpu = smp_processor_id(); 4091d3251859STejun Heo 40927fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 40937fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 40947fb98ea7STejun Heo else 4095df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4096dcd989cbSTejun Heo 409776af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 409888109453SLai Jiangshan rcu_read_unlock_sched(); 409976af4d93STejun Heo 410076af4d93STejun Heo return ret; 4101dcd989cbSTejun Heo } 4102dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4103dcd989cbSTejun Heo 4104dcd989cbSTejun Heo /** 4105dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4106dcd989cbSTejun Heo * @work: the work to be tested 4107dcd989cbSTejun Heo * 4108dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4109dcd989cbSTejun Heo * synchronization around this function and the test result is 4110dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4111dcd989cbSTejun Heo * 4112d185af30SYacine Belkadi * Return: 4113dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4114dcd989cbSTejun Heo */ 4115dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4116dcd989cbSTejun Heo { 4117fa1b54e6STejun Heo struct worker_pool *pool; 4118dcd989cbSTejun Heo unsigned long flags; 4119dcd989cbSTejun Heo unsigned int ret = 0; 4120dcd989cbSTejun Heo 4121dcd989cbSTejun Heo if (work_pending(work)) 4122dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4123038366c5SLai Jiangshan 4124fa1b54e6STejun Heo local_irq_save(flags); 4125fa1b54e6STejun Heo pool = get_work_pool(work); 4126038366c5SLai Jiangshan if (pool) { 4127fa1b54e6STejun Heo spin_lock(&pool->lock); 4128c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4129dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4130fa1b54e6STejun Heo spin_unlock(&pool->lock); 4131038366c5SLai Jiangshan } 4132fa1b54e6STejun Heo local_irq_restore(flags); 4133dcd989cbSTejun Heo 4134dcd989cbSTejun Heo return ret; 4135dcd989cbSTejun Heo } 4136dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4137dcd989cbSTejun Heo 41383d1cb205STejun Heo /** 41393d1cb205STejun Heo * set_worker_desc - set description for the current work item 41403d1cb205STejun Heo * @fmt: printf-style format string 41413d1cb205STejun Heo * @...: arguments for the format string 41423d1cb205STejun Heo * 41433d1cb205STejun Heo * This function can be called by a running work function to describe what 41443d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 41453d1cb205STejun Heo * information will be printed out together to help debugging. The 41463d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 41473d1cb205STejun Heo */ 41483d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 41493d1cb205STejun Heo { 41503d1cb205STejun Heo struct worker *worker = current_wq_worker(); 41513d1cb205STejun Heo va_list args; 41523d1cb205STejun Heo 41533d1cb205STejun Heo if (worker) { 41543d1cb205STejun Heo va_start(args, fmt); 41553d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 41563d1cb205STejun Heo va_end(args); 41573d1cb205STejun Heo worker->desc_valid = true; 41583d1cb205STejun Heo } 41593d1cb205STejun Heo } 41603d1cb205STejun Heo 41613d1cb205STejun Heo /** 41623d1cb205STejun Heo * print_worker_info - print out worker information and description 41633d1cb205STejun Heo * @log_lvl: the log level to use when printing 41643d1cb205STejun Heo * @task: target task 41653d1cb205STejun Heo * 41663d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 41673d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 41683d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 41693d1cb205STejun Heo * 41703d1cb205STejun Heo * This function can be safely called on any task as long as the 41713d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 41723d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 41733d1cb205STejun Heo */ 41743d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 41753d1cb205STejun Heo { 41763d1cb205STejun Heo work_func_t *fn = NULL; 41773d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 41783d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 41793d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 41803d1cb205STejun Heo struct workqueue_struct *wq = NULL; 41813d1cb205STejun Heo bool desc_valid = false; 41823d1cb205STejun Heo struct worker *worker; 41833d1cb205STejun Heo 41843d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 41853d1cb205STejun Heo return; 41863d1cb205STejun Heo 41873d1cb205STejun Heo /* 41883d1cb205STejun Heo * This function is called without any synchronization and @task 41893d1cb205STejun Heo * could be in any state. Be careful with dereferences. 41903d1cb205STejun Heo */ 41913d1cb205STejun Heo worker = probe_kthread_data(task); 41923d1cb205STejun Heo 41933d1cb205STejun Heo /* 41943d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 41953d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 41963d1cb205STejun Heo */ 41973d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 41983d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 41993d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 42003d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 42013d1cb205STejun Heo 42023d1cb205STejun Heo /* copy worker description */ 42033d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 42043d1cb205STejun Heo if (desc_valid) 42053d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 42063d1cb205STejun Heo 42073d1cb205STejun Heo if (fn || name[0] || desc[0]) { 42083d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 42093d1cb205STejun Heo if (desc[0]) 42103d1cb205STejun Heo pr_cont(" (%s)", desc); 42113d1cb205STejun Heo pr_cont("\n"); 42123d1cb205STejun Heo } 42133d1cb205STejun Heo } 42143d1cb205STejun Heo 42153494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 42163494fc30STejun Heo { 42173494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 42183494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 42193494fc30STejun Heo pr_cont(" node=%d", pool->node); 42203494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 42213494fc30STejun Heo } 42223494fc30STejun Heo 42233494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 42243494fc30STejun Heo { 42253494fc30STejun Heo if (work->func == wq_barrier_func) { 42263494fc30STejun Heo struct wq_barrier *barr; 42273494fc30STejun Heo 42283494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 42293494fc30STejun Heo 42303494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 42313494fc30STejun Heo task_pid_nr(barr->task)); 42323494fc30STejun Heo } else { 42333494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 42343494fc30STejun Heo } 42353494fc30STejun Heo } 42363494fc30STejun Heo 42373494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 42383494fc30STejun Heo { 42393494fc30STejun Heo struct worker_pool *pool = pwq->pool; 42403494fc30STejun Heo struct work_struct *work; 42413494fc30STejun Heo struct worker *worker; 42423494fc30STejun Heo bool has_in_flight = false, has_pending = false; 42433494fc30STejun Heo int bkt; 42443494fc30STejun Heo 42453494fc30STejun Heo pr_info(" pwq %d:", pool->id); 42463494fc30STejun Heo pr_cont_pool_info(pool); 42473494fc30STejun Heo 42483494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 42493494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 42503494fc30STejun Heo 42513494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 42523494fc30STejun Heo if (worker->current_pwq == pwq) { 42533494fc30STejun Heo has_in_flight = true; 42543494fc30STejun Heo break; 42553494fc30STejun Heo } 42563494fc30STejun Heo } 42573494fc30STejun Heo if (has_in_flight) { 42583494fc30STejun Heo bool comma = false; 42593494fc30STejun Heo 42603494fc30STejun Heo pr_info(" in-flight:"); 42613494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 42623494fc30STejun Heo if (worker->current_pwq != pwq) 42633494fc30STejun Heo continue; 42643494fc30STejun Heo 42653494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 42663494fc30STejun Heo task_pid_nr(worker->task), 42673494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 42683494fc30STejun Heo worker->current_func); 42693494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 42703494fc30STejun Heo pr_cont_work(false, work); 42713494fc30STejun Heo comma = true; 42723494fc30STejun Heo } 42733494fc30STejun Heo pr_cont("\n"); 42743494fc30STejun Heo } 42753494fc30STejun Heo 42763494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 42773494fc30STejun Heo if (get_work_pwq(work) == pwq) { 42783494fc30STejun Heo has_pending = true; 42793494fc30STejun Heo break; 42803494fc30STejun Heo } 42813494fc30STejun Heo } 42823494fc30STejun Heo if (has_pending) { 42833494fc30STejun Heo bool comma = false; 42843494fc30STejun Heo 42853494fc30STejun Heo pr_info(" pending:"); 42863494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 42873494fc30STejun Heo if (get_work_pwq(work) != pwq) 42883494fc30STejun Heo continue; 42893494fc30STejun Heo 42903494fc30STejun Heo pr_cont_work(comma, work); 42913494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 42923494fc30STejun Heo } 42933494fc30STejun Heo pr_cont("\n"); 42943494fc30STejun Heo } 42953494fc30STejun Heo 42963494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 42973494fc30STejun Heo bool comma = false; 42983494fc30STejun Heo 42993494fc30STejun Heo pr_info(" delayed:"); 43003494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 43013494fc30STejun Heo pr_cont_work(comma, work); 43023494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 43033494fc30STejun Heo } 43043494fc30STejun Heo pr_cont("\n"); 43053494fc30STejun Heo } 43063494fc30STejun Heo } 43073494fc30STejun Heo 43083494fc30STejun Heo /** 43093494fc30STejun Heo * show_workqueue_state - dump workqueue state 43103494fc30STejun Heo * 43113494fc30STejun Heo * Called from a sysrq handler and prints out all busy workqueues and 43123494fc30STejun Heo * pools. 43133494fc30STejun Heo */ 43143494fc30STejun Heo void show_workqueue_state(void) 43153494fc30STejun Heo { 43163494fc30STejun Heo struct workqueue_struct *wq; 43173494fc30STejun Heo struct worker_pool *pool; 43183494fc30STejun Heo unsigned long flags; 43193494fc30STejun Heo int pi; 43203494fc30STejun Heo 43213494fc30STejun Heo rcu_read_lock_sched(); 43223494fc30STejun Heo 43233494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 43243494fc30STejun Heo 43253494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 43263494fc30STejun Heo struct pool_workqueue *pwq; 43273494fc30STejun Heo bool idle = true; 43283494fc30STejun Heo 43293494fc30STejun Heo for_each_pwq(pwq, wq) { 43303494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 43313494fc30STejun Heo idle = false; 43323494fc30STejun Heo break; 43333494fc30STejun Heo } 43343494fc30STejun Heo } 43353494fc30STejun Heo if (idle) 43363494fc30STejun Heo continue; 43373494fc30STejun Heo 43383494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 43393494fc30STejun Heo 43403494fc30STejun Heo for_each_pwq(pwq, wq) { 43413494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 43423494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 43433494fc30STejun Heo show_pwq(pwq); 43443494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 43453494fc30STejun Heo } 43463494fc30STejun Heo } 43473494fc30STejun Heo 43483494fc30STejun Heo for_each_pool(pool, pi) { 43493494fc30STejun Heo struct worker *worker; 43503494fc30STejun Heo bool first = true; 43513494fc30STejun Heo 43523494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 43533494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 43543494fc30STejun Heo goto next_pool; 43553494fc30STejun Heo 43563494fc30STejun Heo pr_info("pool %d:", pool->id); 43573494fc30STejun Heo pr_cont_pool_info(pool); 435882607adcSTejun Heo pr_cont(" hung=%us workers=%d", 435982607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 436082607adcSTejun Heo pool->nr_workers); 43613494fc30STejun Heo if (pool->manager) 43623494fc30STejun Heo pr_cont(" manager: %d", 43633494fc30STejun Heo task_pid_nr(pool->manager->task)); 43643494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 43653494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 43663494fc30STejun Heo task_pid_nr(worker->task)); 43673494fc30STejun Heo first = false; 43683494fc30STejun Heo } 43693494fc30STejun Heo pr_cont("\n"); 43703494fc30STejun Heo next_pool: 43713494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 43723494fc30STejun Heo } 43733494fc30STejun Heo 43743494fc30STejun Heo rcu_read_unlock_sched(); 43753494fc30STejun Heo } 43763494fc30STejun Heo 4377db7bccf4STejun Heo /* 4378db7bccf4STejun Heo * CPU hotplug. 4379db7bccf4STejun Heo * 4380e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4381112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4382706026c2STejun Heo * pool which make migrating pending and scheduled works very 4383e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 438494cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4385e22bee78STejun Heo * blocked draining impractical. 4386e22bee78STejun Heo * 438724647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4388628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4389628c78e7STejun Heo * cpu comes back online. 4390db7bccf4STejun Heo */ 4391db7bccf4STejun Heo 4392706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4393db7bccf4STejun Heo { 439438db41d9STejun Heo int cpu = smp_processor_id(); 43954ce62e9eSTejun Heo struct worker_pool *pool; 4396db7bccf4STejun Heo struct worker *worker; 4397db7bccf4STejun Heo 4398f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 439992f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 440094cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4401e22bee78STejun Heo 4402f2d5a0eeSTejun Heo /* 440392f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 440494cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 440594cf58bbSTejun Heo * except for the ones which are still executing works from 440694cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 440794cf58bbSTejun Heo * this, they may become diasporas. 4408f2d5a0eeSTejun Heo */ 4409da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4410403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4411db7bccf4STejun Heo 441224647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4413f2d5a0eeSTejun Heo 441494cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 441592f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4416e22bee78STejun Heo 4417e22bee78STejun Heo /* 4418eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4419eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4420eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4421eb283428SLai Jiangshan * from other cpus. 4422628c78e7STejun Heo */ 4423628c78e7STejun Heo schedule(); 4424628c78e7STejun Heo 4425628c78e7STejun Heo /* 4426eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4427eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4428eb283428SLai Jiangshan * and keep_working() are always true as long as the 4429eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4430eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4431eb283428SLai Jiangshan * are served by workers tied to the pool. 4432e22bee78STejun Heo */ 4433e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4434eb283428SLai Jiangshan 4435eb283428SLai Jiangshan /* 4436eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4437eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4438eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4439eb283428SLai Jiangshan */ 4440eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4441eb283428SLai Jiangshan wake_up_worker(pool); 4442eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4443eb283428SLai Jiangshan } 4444db7bccf4STejun Heo } 4445db7bccf4STejun Heo 4446bd7c089eSTejun Heo /** 4447bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4448bd7c089eSTejun Heo * @pool: pool of interest 4449bd7c089eSTejun Heo * 4450a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4451bd7c089eSTejun Heo */ 4452bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4453bd7c089eSTejun Heo { 4454a9ab775bSTejun Heo struct worker *worker; 4455bd7c089eSTejun Heo 445692f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4457bd7c089eSTejun Heo 4458bd7c089eSTejun Heo /* 4459a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4460a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4461402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4462a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4463a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4464bd7c089eSTejun Heo */ 4465da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4466a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4467a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4468a9ab775bSTejun Heo 4469a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 44703de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4471a9ab775bSTejun Heo 4472da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4473a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4474a9ab775bSTejun Heo 4475a9ab775bSTejun Heo /* 4476a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4477a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4478a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4479a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4480a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4481a9ab775bSTejun Heo * be bound before @pool->lock is released. 4482a9ab775bSTejun Heo */ 4483a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4484bd7c089eSTejun Heo wake_up_process(worker->task); 4485bd7c089eSTejun Heo 4486bd7c089eSTejun Heo /* 4487a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4488a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4489a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4490a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4491a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4492a9ab775bSTejun Heo * concurrency management. Note that when or whether 4493a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4494a9ab775bSTejun Heo * 4495a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4496a9ab775bSTejun Heo * tested without holding any lock in 4497a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4498a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4499a9ab775bSTejun Heo * management operations. 4500bd7c089eSTejun Heo */ 4501a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4502a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4503a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4504a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4505bd7c089eSTejun Heo } 4506a9ab775bSTejun Heo 4507a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4508bd7c089eSTejun Heo } 4509bd7c089eSTejun Heo 45107dbc725eSTejun Heo /** 45117dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 45127dbc725eSTejun Heo * @pool: unbound pool of interest 45137dbc725eSTejun Heo * @cpu: the CPU which is coming up 45147dbc725eSTejun Heo * 45157dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 45167dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 45177dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 45187dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 45197dbc725eSTejun Heo */ 45207dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 45217dbc725eSTejun Heo { 45227dbc725eSTejun Heo static cpumask_t cpumask; 45237dbc725eSTejun Heo struct worker *worker; 45247dbc725eSTejun Heo 452592f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 45267dbc725eSTejun Heo 45277dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 45287dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 45297dbc725eSTejun Heo return; 45307dbc725eSTejun Heo 45317dbc725eSTejun Heo /* is @cpu the only online CPU? */ 45327dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 45337dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 45347dbc725eSTejun Heo return; 45357dbc725eSTejun Heo 45367dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4537da028469SLai Jiangshan for_each_pool_worker(worker, pool) 45387dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 45397dbc725eSTejun Heo pool->attrs->cpumask) < 0); 45407dbc725eSTejun Heo } 45417dbc725eSTejun Heo 45428db25e78STejun Heo /* 45438db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 45448db25e78STejun Heo * This will be registered high priority CPU notifier. 45458db25e78STejun Heo */ 45460db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 45471da177e4SLinus Torvalds unsigned long action, 45481da177e4SLinus Torvalds void *hcpu) 45491da177e4SLinus Torvalds { 4550d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45514ce62e9eSTejun Heo struct worker_pool *pool; 45524c16bd32STejun Heo struct workqueue_struct *wq; 45537dbc725eSTejun Heo int pi; 45541da177e4SLinus Torvalds 45558db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 45563af24433SOleg Nesterov case CPU_UP_PREPARE: 4557f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45583ce63377STejun Heo if (pool->nr_workers) 45593ce63377STejun Heo continue; 4560051e1850SLai Jiangshan if (!create_worker(pool)) 45613ce63377STejun Heo return NOTIFY_BAD; 45623af24433SOleg Nesterov } 45631da177e4SLinus Torvalds break; 45641da177e4SLinus Torvalds 456565758202STejun Heo case CPU_DOWN_FAILED: 456665758202STejun Heo case CPU_ONLINE: 456768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45687dbc725eSTejun Heo 45697dbc725eSTejun Heo for_each_pool(pool, pi) { 457092f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 457194cf58bbSTejun Heo 4572f05b558dSLai Jiangshan if (pool->cpu == cpu) 457394cf58bbSTejun Heo rebind_workers(pool); 4574f05b558dSLai Jiangshan else if (pool->cpu < 0) 45757dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 457694cf58bbSTejun Heo 457792f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 457894cf58bbSTejun Heo } 45797dbc725eSTejun Heo 45804c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45814c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45824c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 45834c16bd32STejun Heo 458468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45858db25e78STejun Heo break; 458665758202STejun Heo } 458765758202STejun Heo return NOTIFY_OK; 458865758202STejun Heo } 458965758202STejun Heo 459065758202STejun Heo /* 459165758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 459265758202STejun Heo * This will be registered as low priority CPU notifier. 459365758202STejun Heo */ 45940db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 459565758202STejun Heo unsigned long action, 459665758202STejun Heo void *hcpu) 459765758202STejun Heo { 4598d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45998db25e78STejun Heo struct work_struct unbind_work; 46004c16bd32STejun Heo struct workqueue_struct *wq; 46018db25e78STejun Heo 460265758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 460365758202STejun Heo case CPU_DOWN_PREPARE: 46044c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4605706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46067635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46074c16bd32STejun Heo 46084c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46094c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46104c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46114c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 46124c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 46134c16bd32STejun Heo 46144c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 46158db25e78STejun Heo flush_work(&unbind_work); 4616440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 46178db25e78STejun Heo break; 461865758202STejun Heo } 461965758202STejun Heo return NOTIFY_OK; 462065758202STejun Heo } 462165758202STejun Heo 46222d3854a3SRusty Russell #ifdef CONFIG_SMP 46238ccad40dSRusty Russell 46242d3854a3SRusty Russell struct work_for_cpu { 4625ed48ece2STejun Heo struct work_struct work; 46262d3854a3SRusty Russell long (*fn)(void *); 46272d3854a3SRusty Russell void *arg; 46282d3854a3SRusty Russell long ret; 46292d3854a3SRusty Russell }; 46302d3854a3SRusty Russell 4631ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 46322d3854a3SRusty Russell { 4633ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4634ed48ece2STejun Heo 46352d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 46362d3854a3SRusty Russell } 46372d3854a3SRusty Russell 46382d3854a3SRusty Russell /** 46392d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 46402d3854a3SRusty Russell * @cpu: the cpu to run on 46412d3854a3SRusty Russell * @fn: the function to run 46422d3854a3SRusty Russell * @arg: the function arg 46432d3854a3SRusty Russell * 464431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 46456b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4646d185af30SYacine Belkadi * 4647d185af30SYacine Belkadi * Return: The value @fn returns. 46482d3854a3SRusty Russell */ 4649d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 46502d3854a3SRusty Russell { 4651ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 46522d3854a3SRusty Russell 4653ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4654ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 465512997d1aSBjorn Helgaas flush_work(&wfc.work); 4656440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 46572d3854a3SRusty Russell return wfc.ret; 46582d3854a3SRusty Russell } 46592d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 46602d3854a3SRusty Russell #endif /* CONFIG_SMP */ 46612d3854a3SRusty Russell 4662a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4663e7577c50SRusty Russell 4664a0a1a5fdSTejun Heo /** 4665a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4666a0a1a5fdSTejun Heo * 466758a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4668c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4669706026c2STejun Heo * pool->worklist. 4670a0a1a5fdSTejun Heo * 4671a0a1a5fdSTejun Heo * CONTEXT: 4672a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4673a0a1a5fdSTejun Heo */ 4674a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4675a0a1a5fdSTejun Heo { 467624b8a847STejun Heo struct workqueue_struct *wq; 467724b8a847STejun Heo struct pool_workqueue *pwq; 4678a0a1a5fdSTejun Heo 467968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4680a0a1a5fdSTejun Heo 46816183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4682a0a1a5fdSTejun Heo workqueue_freezing = true; 4683a0a1a5fdSTejun Heo 468424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4685a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4686699ce097STejun Heo for_each_pwq(pwq, wq) 4687699ce097STejun Heo pwq_adjust_max_active(pwq); 4688a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4689a1056305STejun Heo } 46905bcab335STejun Heo 469168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4692a0a1a5fdSTejun Heo } 4693a0a1a5fdSTejun Heo 4694a0a1a5fdSTejun Heo /** 469558a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4696a0a1a5fdSTejun Heo * 4697a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4698a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4699a0a1a5fdSTejun Heo * 4700a0a1a5fdSTejun Heo * CONTEXT: 470168e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4702a0a1a5fdSTejun Heo * 4703d185af30SYacine Belkadi * Return: 470458a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 470558a69cb4STejun Heo * is complete. 4706a0a1a5fdSTejun Heo */ 4707a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4708a0a1a5fdSTejun Heo { 4709a0a1a5fdSTejun Heo bool busy = false; 471024b8a847STejun Heo struct workqueue_struct *wq; 471124b8a847STejun Heo struct pool_workqueue *pwq; 4712a0a1a5fdSTejun Heo 471368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4714a0a1a5fdSTejun Heo 47156183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4716a0a1a5fdSTejun Heo 471724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 471824b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 471924b8a847STejun Heo continue; 4720a0a1a5fdSTejun Heo /* 4721a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4722a0a1a5fdSTejun Heo * to peek without lock. 4723a0a1a5fdSTejun Heo */ 472488109453SLai Jiangshan rcu_read_lock_sched(); 472524b8a847STejun Heo for_each_pwq(pwq, wq) { 47266183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4727112202d9STejun Heo if (pwq->nr_active) { 4728a0a1a5fdSTejun Heo busy = true; 472988109453SLai Jiangshan rcu_read_unlock_sched(); 4730a0a1a5fdSTejun Heo goto out_unlock; 4731a0a1a5fdSTejun Heo } 4732a0a1a5fdSTejun Heo } 473388109453SLai Jiangshan rcu_read_unlock_sched(); 4734a0a1a5fdSTejun Heo } 4735a0a1a5fdSTejun Heo out_unlock: 473668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4737a0a1a5fdSTejun Heo return busy; 4738a0a1a5fdSTejun Heo } 4739a0a1a5fdSTejun Heo 4740a0a1a5fdSTejun Heo /** 4741a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4742a0a1a5fdSTejun Heo * 4743a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4744706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4745a0a1a5fdSTejun Heo * 4746a0a1a5fdSTejun Heo * CONTEXT: 4747a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4748a0a1a5fdSTejun Heo */ 4749a0a1a5fdSTejun Heo void thaw_workqueues(void) 4750a0a1a5fdSTejun Heo { 475124b8a847STejun Heo struct workqueue_struct *wq; 475224b8a847STejun Heo struct pool_workqueue *pwq; 4753a0a1a5fdSTejun Heo 475468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4755a0a1a5fdSTejun Heo 4756a0a1a5fdSTejun Heo if (!workqueue_freezing) 4757a0a1a5fdSTejun Heo goto out_unlock; 4758a0a1a5fdSTejun Heo 475974b414eaSLai Jiangshan workqueue_freezing = false; 476024b8a847STejun Heo 476124b8a847STejun Heo /* restore max_active and repopulate worklist */ 476224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4763a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4764699ce097STejun Heo for_each_pwq(pwq, wq) 4765699ce097STejun Heo pwq_adjust_max_active(pwq); 4766a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 476724b8a847STejun Heo } 476824b8a847STejun Heo 4769a0a1a5fdSTejun Heo out_unlock: 477068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4771a0a1a5fdSTejun Heo } 4772a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4773a0a1a5fdSTejun Heo 4774042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4775042f7df1SLai Jiangshan { 4776042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4777042f7df1SLai Jiangshan int ret = 0; 4778042f7df1SLai Jiangshan struct workqueue_struct *wq; 4779042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4780042f7df1SLai Jiangshan 4781042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4782042f7df1SLai Jiangshan 4783042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4784042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4785042f7df1SLai Jiangshan continue; 4786042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 4787042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 4788042f7df1SLai Jiangshan continue; 4789042f7df1SLai Jiangshan 4790042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 4791042f7df1SLai Jiangshan if (!ctx) { 4792042f7df1SLai Jiangshan ret = -ENOMEM; 4793042f7df1SLai Jiangshan break; 4794042f7df1SLai Jiangshan } 4795042f7df1SLai Jiangshan 4796042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 4797042f7df1SLai Jiangshan } 4798042f7df1SLai Jiangshan 4799042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 4800042f7df1SLai Jiangshan if (!ret) 4801042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 4802042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 4803042f7df1SLai Jiangshan } 4804042f7df1SLai Jiangshan 4805042f7df1SLai Jiangshan return ret; 4806042f7df1SLai Jiangshan } 4807042f7df1SLai Jiangshan 4808042f7df1SLai Jiangshan /** 4809042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 4810042f7df1SLai Jiangshan * @cpumask: the cpumask to set 4811042f7df1SLai Jiangshan * 4812042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 4813042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 4814042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 4815042f7df1SLai Jiangshan * 4816042f7df1SLai Jiangshan * Retun: 0 - Success 4817042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 4818042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 4819042f7df1SLai Jiangshan */ 4820042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 4821042f7df1SLai Jiangshan { 4822042f7df1SLai Jiangshan int ret = -EINVAL; 4823042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 4824042f7df1SLai Jiangshan 4825042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 4826042f7df1SLai Jiangshan return -ENOMEM; 4827042f7df1SLai Jiangshan 4828042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 4829042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 4830a0111cf6SLai Jiangshan apply_wqattrs_lock(); 4831042f7df1SLai Jiangshan 4832042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 4833042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 4834042f7df1SLai Jiangshan 4835042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 4836042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 4837042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 4838042f7df1SLai Jiangshan 4839042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 4840042f7df1SLai Jiangshan if (ret < 0) 4841042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 4842042f7df1SLai Jiangshan 4843a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 4844042f7df1SLai Jiangshan } 4845042f7df1SLai Jiangshan 4846042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 4847042f7df1SLai Jiangshan return ret; 4848042f7df1SLai Jiangshan } 4849042f7df1SLai Jiangshan 48506ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 48516ba94429SFrederic Weisbecker /* 48526ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 48536ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 48546ba94429SFrederic Weisbecker * following attributes. 48556ba94429SFrederic Weisbecker * 48566ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 48576ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 48586ba94429SFrederic Weisbecker * 48596ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 48606ba94429SFrederic Weisbecker * 48616ba94429SFrederic Weisbecker * id RO int : the associated pool ID 48626ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 48636ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 48646ba94429SFrederic Weisbecker */ 48656ba94429SFrederic Weisbecker struct wq_device { 48666ba94429SFrederic Weisbecker struct workqueue_struct *wq; 48676ba94429SFrederic Weisbecker struct device dev; 48686ba94429SFrederic Weisbecker }; 48696ba94429SFrederic Weisbecker 48706ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 48716ba94429SFrederic Weisbecker { 48726ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 48736ba94429SFrederic Weisbecker 48746ba94429SFrederic Weisbecker return wq_dev->wq; 48756ba94429SFrederic Weisbecker } 48766ba94429SFrederic Weisbecker 48776ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 48786ba94429SFrederic Weisbecker char *buf) 48796ba94429SFrederic Weisbecker { 48806ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48816ba94429SFrederic Weisbecker 48826ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 48836ba94429SFrederic Weisbecker } 48846ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 48856ba94429SFrederic Weisbecker 48866ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 48876ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 48886ba94429SFrederic Weisbecker { 48896ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48906ba94429SFrederic Weisbecker 48916ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 48926ba94429SFrederic Weisbecker } 48936ba94429SFrederic Weisbecker 48946ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 48956ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 48966ba94429SFrederic Weisbecker size_t count) 48976ba94429SFrederic Weisbecker { 48986ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48996ba94429SFrederic Weisbecker int val; 49006ba94429SFrederic Weisbecker 49016ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 49026ba94429SFrederic Weisbecker return -EINVAL; 49036ba94429SFrederic Weisbecker 49046ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 49056ba94429SFrederic Weisbecker return count; 49066ba94429SFrederic Weisbecker } 49076ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 49086ba94429SFrederic Weisbecker 49096ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 49106ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 49116ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 49126ba94429SFrederic Weisbecker NULL, 49136ba94429SFrederic Weisbecker }; 49146ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 49156ba94429SFrederic Weisbecker 49166ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 49176ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49186ba94429SFrederic Weisbecker { 49196ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49206ba94429SFrederic Weisbecker const char *delim = ""; 49216ba94429SFrederic Weisbecker int node, written = 0; 49226ba94429SFrederic Weisbecker 49236ba94429SFrederic Weisbecker rcu_read_lock_sched(); 49246ba94429SFrederic Weisbecker for_each_node(node) { 49256ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 49266ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 49276ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 49286ba94429SFrederic Weisbecker delim = " "; 49296ba94429SFrederic Weisbecker } 49306ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 49316ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 49326ba94429SFrederic Weisbecker 49336ba94429SFrederic Weisbecker return written; 49346ba94429SFrederic Weisbecker } 49356ba94429SFrederic Weisbecker 49366ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 49376ba94429SFrederic Weisbecker char *buf) 49386ba94429SFrederic Weisbecker { 49396ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49406ba94429SFrederic Weisbecker int written; 49416ba94429SFrederic Weisbecker 49426ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 49436ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 49446ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 49456ba94429SFrederic Weisbecker 49466ba94429SFrederic Weisbecker return written; 49476ba94429SFrederic Weisbecker } 49486ba94429SFrederic Weisbecker 49496ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 49506ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 49516ba94429SFrederic Weisbecker { 49526ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 49536ba94429SFrederic Weisbecker 4954899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4955899a94feSLai Jiangshan 49566ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 49576ba94429SFrederic Weisbecker if (!attrs) 49586ba94429SFrederic Weisbecker return NULL; 49596ba94429SFrederic Weisbecker 49606ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 49616ba94429SFrederic Weisbecker return attrs; 49626ba94429SFrederic Weisbecker } 49636ba94429SFrederic Weisbecker 49646ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 49656ba94429SFrederic Weisbecker const char *buf, size_t count) 49666ba94429SFrederic Weisbecker { 49676ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49686ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 4969d4d3e257SLai Jiangshan int ret = -ENOMEM; 4970d4d3e257SLai Jiangshan 4971d4d3e257SLai Jiangshan apply_wqattrs_lock(); 49726ba94429SFrederic Weisbecker 49736ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 49746ba94429SFrederic Weisbecker if (!attrs) 4975d4d3e257SLai Jiangshan goto out_unlock; 49766ba94429SFrederic Weisbecker 49776ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 49786ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 4979d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 49806ba94429SFrederic Weisbecker else 49816ba94429SFrederic Weisbecker ret = -EINVAL; 49826ba94429SFrederic Weisbecker 4983d4d3e257SLai Jiangshan out_unlock: 4984d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 49856ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 49866ba94429SFrederic Weisbecker return ret ?: count; 49876ba94429SFrederic Weisbecker } 49886ba94429SFrederic Weisbecker 49896ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 49906ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49916ba94429SFrederic Weisbecker { 49926ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49936ba94429SFrederic Weisbecker int written; 49946ba94429SFrederic Weisbecker 49956ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 49966ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 49976ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 49986ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 49996ba94429SFrederic Weisbecker return written; 50006ba94429SFrederic Weisbecker } 50016ba94429SFrederic Weisbecker 50026ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 50036ba94429SFrederic Weisbecker struct device_attribute *attr, 50046ba94429SFrederic Weisbecker const char *buf, size_t count) 50056ba94429SFrederic Weisbecker { 50066ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50076ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5008d4d3e257SLai Jiangshan int ret = -ENOMEM; 5009d4d3e257SLai Jiangshan 5010d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50116ba94429SFrederic Weisbecker 50126ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50136ba94429SFrederic Weisbecker if (!attrs) 5014d4d3e257SLai Jiangshan goto out_unlock; 50156ba94429SFrederic Weisbecker 50166ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 50176ba94429SFrederic Weisbecker if (!ret) 5018d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50196ba94429SFrederic Weisbecker 5020d4d3e257SLai Jiangshan out_unlock: 5021d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50226ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50236ba94429SFrederic Weisbecker return ret ?: count; 50246ba94429SFrederic Weisbecker } 50256ba94429SFrederic Weisbecker 50266ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 50276ba94429SFrederic Weisbecker char *buf) 50286ba94429SFrederic Weisbecker { 50296ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50306ba94429SFrederic Weisbecker int written; 50316ba94429SFrederic Weisbecker 50326ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 50336ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 50346ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 50356ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 50366ba94429SFrederic Weisbecker 50376ba94429SFrederic Weisbecker return written; 50386ba94429SFrederic Weisbecker } 50396ba94429SFrederic Weisbecker 50406ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 50416ba94429SFrederic Weisbecker const char *buf, size_t count) 50426ba94429SFrederic Weisbecker { 50436ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50446ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5045d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5046d4d3e257SLai Jiangshan 5047d4d3e257SLai Jiangshan apply_wqattrs_lock(); 50486ba94429SFrederic Weisbecker 50496ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50506ba94429SFrederic Weisbecker if (!attrs) 5051d4d3e257SLai Jiangshan goto out_unlock; 50526ba94429SFrederic Weisbecker 50536ba94429SFrederic Weisbecker ret = -EINVAL; 50546ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 50556ba94429SFrederic Weisbecker attrs->no_numa = !v; 5056d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50576ba94429SFrederic Weisbecker } 50586ba94429SFrederic Weisbecker 5059d4d3e257SLai Jiangshan out_unlock: 5060d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50616ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50626ba94429SFrederic Weisbecker return ret ?: count; 50636ba94429SFrederic Weisbecker } 50646ba94429SFrederic Weisbecker 50656ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 50666ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 50676ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 50686ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 50696ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 50706ba94429SFrederic Weisbecker __ATTR_NULL, 50716ba94429SFrederic Weisbecker }; 50726ba94429SFrederic Weisbecker 50736ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 50746ba94429SFrederic Weisbecker .name = "workqueue", 50756ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 50766ba94429SFrederic Weisbecker }; 50776ba94429SFrederic Weisbecker 5078b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5079b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5080b05a7928SFrederic Weisbecker { 5081b05a7928SFrederic Weisbecker int written; 5082b05a7928SFrederic Weisbecker 5083042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5084b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5085b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5086042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5087b05a7928SFrederic Weisbecker 5088b05a7928SFrederic Weisbecker return written; 5089b05a7928SFrederic Weisbecker } 5090b05a7928SFrederic Weisbecker 5091042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5092042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5093042f7df1SLai Jiangshan { 5094042f7df1SLai Jiangshan cpumask_var_t cpumask; 5095042f7df1SLai Jiangshan int ret; 5096042f7df1SLai Jiangshan 5097042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5098042f7df1SLai Jiangshan return -ENOMEM; 5099042f7df1SLai Jiangshan 5100042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5101042f7df1SLai Jiangshan if (!ret) 5102042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5103042f7df1SLai Jiangshan 5104042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5105042f7df1SLai Jiangshan return ret ? ret : count; 5106042f7df1SLai Jiangshan } 5107042f7df1SLai Jiangshan 5108b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5109042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5110042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5111b05a7928SFrederic Weisbecker 51126ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 51136ba94429SFrederic Weisbecker { 5114b05a7928SFrederic Weisbecker int err; 5115b05a7928SFrederic Weisbecker 5116b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5117b05a7928SFrederic Weisbecker if (err) 5118b05a7928SFrederic Weisbecker return err; 5119b05a7928SFrederic Weisbecker 5120b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 51216ba94429SFrederic Weisbecker } 51226ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 51236ba94429SFrederic Weisbecker 51246ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 51256ba94429SFrederic Weisbecker { 51266ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 51276ba94429SFrederic Weisbecker 51286ba94429SFrederic Weisbecker kfree(wq_dev); 51296ba94429SFrederic Weisbecker } 51306ba94429SFrederic Weisbecker 51316ba94429SFrederic Weisbecker /** 51326ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 51336ba94429SFrederic Weisbecker * @wq: the workqueue to register 51346ba94429SFrederic Weisbecker * 51356ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 51366ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 51376ba94429SFrederic Weisbecker * which is the preferred method. 51386ba94429SFrederic Weisbecker * 51396ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 51406ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 51416ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 51426ba94429SFrederic Weisbecker * attributes. 51436ba94429SFrederic Weisbecker * 51446ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 51456ba94429SFrederic Weisbecker */ 51466ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 51476ba94429SFrederic Weisbecker { 51486ba94429SFrederic Weisbecker struct wq_device *wq_dev; 51496ba94429SFrederic Weisbecker int ret; 51506ba94429SFrederic Weisbecker 51516ba94429SFrederic Weisbecker /* 5152402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 51536ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 51546ba94429SFrederic Weisbecker * workqueues. 51556ba94429SFrederic Weisbecker */ 51566ba94429SFrederic Weisbecker if (WARN_ON(wq->flags & __WQ_ORDERED)) 51576ba94429SFrederic Weisbecker return -EINVAL; 51586ba94429SFrederic Weisbecker 51596ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 51606ba94429SFrederic Weisbecker if (!wq_dev) 51616ba94429SFrederic Weisbecker return -ENOMEM; 51626ba94429SFrederic Weisbecker 51636ba94429SFrederic Weisbecker wq_dev->wq = wq; 51646ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 51656ba94429SFrederic Weisbecker wq_dev->dev.init_name = wq->name; 51666ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 51676ba94429SFrederic Weisbecker 51686ba94429SFrederic Weisbecker /* 51696ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 51706ba94429SFrederic Weisbecker * everything is ready. 51716ba94429SFrederic Weisbecker */ 51726ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 51736ba94429SFrederic Weisbecker 51746ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 51756ba94429SFrederic Weisbecker if (ret) { 51766ba94429SFrederic Weisbecker kfree(wq_dev); 51776ba94429SFrederic Weisbecker wq->wq_dev = NULL; 51786ba94429SFrederic Weisbecker return ret; 51796ba94429SFrederic Weisbecker } 51806ba94429SFrederic Weisbecker 51816ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 51826ba94429SFrederic Weisbecker struct device_attribute *attr; 51836ba94429SFrederic Weisbecker 51846ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 51856ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 51866ba94429SFrederic Weisbecker if (ret) { 51876ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 51886ba94429SFrederic Weisbecker wq->wq_dev = NULL; 51896ba94429SFrederic Weisbecker return ret; 51906ba94429SFrederic Weisbecker } 51916ba94429SFrederic Weisbecker } 51926ba94429SFrederic Weisbecker } 51936ba94429SFrederic Weisbecker 51946ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 51956ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 51966ba94429SFrederic Weisbecker return 0; 51976ba94429SFrederic Weisbecker } 51986ba94429SFrederic Weisbecker 51996ba94429SFrederic Weisbecker /** 52006ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 52016ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 52026ba94429SFrederic Weisbecker * 52036ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 52046ba94429SFrederic Weisbecker */ 52056ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 52066ba94429SFrederic Weisbecker { 52076ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 52086ba94429SFrederic Weisbecker 52096ba94429SFrederic Weisbecker if (!wq->wq_dev) 52106ba94429SFrederic Weisbecker return; 52116ba94429SFrederic Weisbecker 52126ba94429SFrederic Weisbecker wq->wq_dev = NULL; 52136ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 52146ba94429SFrederic Weisbecker } 52156ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 52166ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 52176ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 52186ba94429SFrederic Weisbecker 521982607adcSTejun Heo /* 522082607adcSTejun Heo * Workqueue watchdog. 522182607adcSTejun Heo * 522282607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 522382607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 522482607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 522582607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 522682607adcSTejun Heo * largely opaque. 522782607adcSTejun Heo * 522882607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 522982607adcSTejun Heo * state if some pools failed to make forward progress for a while where 523082607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 523182607adcSTejun Heo * 523282607adcSTejun Heo * This mechanism is controlled through the kernel parameter 523382607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 523482607adcSTejun Heo * corresponding sysfs parameter file. 523582607adcSTejun Heo */ 523682607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 523782607adcSTejun Heo 523882607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data); 523982607adcSTejun Heo 524082607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 524182607adcSTejun Heo static struct timer_list wq_watchdog_timer = 524282607adcSTejun Heo TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0); 524382607adcSTejun Heo 524482607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 524582607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 524682607adcSTejun Heo 524782607adcSTejun Heo static void wq_watchdog_reset_touched(void) 524882607adcSTejun Heo { 524982607adcSTejun Heo int cpu; 525082607adcSTejun Heo 525182607adcSTejun Heo wq_watchdog_touched = jiffies; 525282607adcSTejun Heo for_each_possible_cpu(cpu) 525382607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 525482607adcSTejun Heo } 525582607adcSTejun Heo 525682607adcSTejun Heo static void wq_watchdog_timer_fn(unsigned long data) 525782607adcSTejun Heo { 525882607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 525982607adcSTejun Heo bool lockup_detected = false; 526082607adcSTejun Heo struct worker_pool *pool; 526182607adcSTejun Heo int pi; 526282607adcSTejun Heo 526382607adcSTejun Heo if (!thresh) 526482607adcSTejun Heo return; 526582607adcSTejun Heo 526682607adcSTejun Heo rcu_read_lock(); 526782607adcSTejun Heo 526882607adcSTejun Heo for_each_pool(pool, pi) { 526982607adcSTejun Heo unsigned long pool_ts, touched, ts; 527082607adcSTejun Heo 527182607adcSTejun Heo if (list_empty(&pool->worklist)) 527282607adcSTejun Heo continue; 527382607adcSTejun Heo 527482607adcSTejun Heo /* get the latest of pool and touched timestamps */ 527582607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 527682607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 527782607adcSTejun Heo 527882607adcSTejun Heo if (time_after(pool_ts, touched)) 527982607adcSTejun Heo ts = pool_ts; 528082607adcSTejun Heo else 528182607adcSTejun Heo ts = touched; 528282607adcSTejun Heo 528382607adcSTejun Heo if (pool->cpu >= 0) { 528482607adcSTejun Heo unsigned long cpu_touched = 528582607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 528682607adcSTejun Heo pool->cpu)); 528782607adcSTejun Heo if (time_after(cpu_touched, ts)) 528882607adcSTejun Heo ts = cpu_touched; 528982607adcSTejun Heo } 529082607adcSTejun Heo 529182607adcSTejun Heo /* did we stall? */ 529282607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 529382607adcSTejun Heo lockup_detected = true; 529482607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 529582607adcSTejun Heo pr_cont_pool_info(pool); 529682607adcSTejun Heo pr_cont(" stuck for %us!\n", 529782607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 529882607adcSTejun Heo } 529982607adcSTejun Heo } 530082607adcSTejun Heo 530182607adcSTejun Heo rcu_read_unlock(); 530282607adcSTejun Heo 530382607adcSTejun Heo if (lockup_detected) 530482607adcSTejun Heo show_workqueue_state(); 530582607adcSTejun Heo 530682607adcSTejun Heo wq_watchdog_reset_touched(); 530782607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 530882607adcSTejun Heo } 530982607adcSTejun Heo 531082607adcSTejun Heo void wq_watchdog_touch(int cpu) 531182607adcSTejun Heo { 531282607adcSTejun Heo if (cpu >= 0) 531382607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 531482607adcSTejun Heo else 531582607adcSTejun Heo wq_watchdog_touched = jiffies; 531682607adcSTejun Heo } 531782607adcSTejun Heo 531882607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 531982607adcSTejun Heo { 532082607adcSTejun Heo wq_watchdog_thresh = 0; 532182607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 532282607adcSTejun Heo 532382607adcSTejun Heo if (thresh) { 532482607adcSTejun Heo wq_watchdog_thresh = thresh; 532582607adcSTejun Heo wq_watchdog_reset_touched(); 532682607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 532782607adcSTejun Heo } 532882607adcSTejun Heo } 532982607adcSTejun Heo 533082607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 533182607adcSTejun Heo const struct kernel_param *kp) 533282607adcSTejun Heo { 533382607adcSTejun Heo unsigned long thresh; 533482607adcSTejun Heo int ret; 533582607adcSTejun Heo 533682607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 533782607adcSTejun Heo if (ret) 533882607adcSTejun Heo return ret; 533982607adcSTejun Heo 534082607adcSTejun Heo if (system_wq) 534182607adcSTejun Heo wq_watchdog_set_thresh(thresh); 534282607adcSTejun Heo else 534382607adcSTejun Heo wq_watchdog_thresh = thresh; 534482607adcSTejun Heo 534582607adcSTejun Heo return 0; 534682607adcSTejun Heo } 534782607adcSTejun Heo 534882607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 534982607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 535082607adcSTejun Heo .get = param_get_ulong, 535182607adcSTejun Heo }; 535282607adcSTejun Heo 535382607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 535482607adcSTejun Heo 0644); 535582607adcSTejun Heo 535682607adcSTejun Heo static void wq_watchdog_init(void) 535782607adcSTejun Heo { 535882607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 535982607adcSTejun Heo } 536082607adcSTejun Heo 536182607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 536282607adcSTejun Heo 536382607adcSTejun Heo static inline void wq_watchdog_init(void) { } 536482607adcSTejun Heo 536582607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 536682607adcSTejun Heo 5367bce90380STejun Heo static void __init wq_numa_init(void) 5368bce90380STejun Heo { 5369bce90380STejun Heo cpumask_var_t *tbl; 5370bce90380STejun Heo int node, cpu; 5371bce90380STejun Heo 5372bce90380STejun Heo if (num_possible_nodes() <= 1) 5373bce90380STejun Heo return; 5374bce90380STejun Heo 5375d55262c4STejun Heo if (wq_disable_numa) { 5376d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5377d55262c4STejun Heo return; 5378d55262c4STejun Heo } 5379d55262c4STejun Heo 53804c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 53814c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 53824c16bd32STejun Heo 5383bce90380STejun Heo /* 5384bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5385bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5386bce90380STejun Heo * fully initialized by now. 5387bce90380STejun Heo */ 5388ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5389bce90380STejun Heo BUG_ON(!tbl); 5390bce90380STejun Heo 5391bce90380STejun Heo for_each_node(node) 53925a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 53931be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5394bce90380STejun Heo 5395bce90380STejun Heo for_each_possible_cpu(cpu) { 5396bce90380STejun Heo node = cpu_to_node(cpu); 5397bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5398bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5399bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5400bce90380STejun Heo return; 5401bce90380STejun Heo } 5402bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5403bce90380STejun Heo } 5404bce90380STejun Heo 5405bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5406bce90380STejun Heo wq_numa_enabled = true; 5407bce90380STejun Heo } 5408bce90380STejun Heo 54096ee0578bSSuresh Siddha static int __init init_workqueues(void) 54101da177e4SLinus Torvalds { 54117a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 54127a4e344cSTejun Heo int i, cpu; 5413c34056a3STejun Heo 5414e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5415e904e6c2STejun Heo 5416b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5417b05a7928SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 5418b05a7928SFrederic Weisbecker 5419e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5420e904e6c2STejun Heo 542165758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 5422a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 54238b03ae3cSTejun Heo 5424bce90380STejun Heo wq_numa_init(); 5425bce90380STejun Heo 5426706026c2STejun Heo /* initialize CPU pools */ 542729c91e99STejun Heo for_each_possible_cpu(cpu) { 54284ce62e9eSTejun Heo struct worker_pool *pool; 54298b03ae3cSTejun Heo 54307a4e344cSTejun Heo i = 0; 5431f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54327a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5433ec22ca5eSTejun Heo pool->cpu = cpu; 54347a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 54357a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5436f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 54377a4e344cSTejun Heo 54389daf9e67STejun Heo /* alloc pool ID */ 543968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 54409daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 544168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 54424ce62e9eSTejun Heo } 54438b03ae3cSTejun Heo } 54448b03ae3cSTejun Heo 5445e22bee78STejun Heo /* create the initial worker */ 544629c91e99STejun Heo for_each_online_cpu(cpu) { 54474ce62e9eSTejun Heo struct worker_pool *pool; 5448e22bee78STejun Heo 5449f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 545024647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5451051e1850SLai Jiangshan BUG_ON(!create_worker(pool)); 5452e22bee78STejun Heo } 54534ce62e9eSTejun Heo } 5454e22bee78STejun Heo 54558a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 545629c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 545729c91e99STejun Heo struct workqueue_attrs *attrs; 545829c91e99STejun Heo 545929c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 546029c91e99STejun Heo attrs->nice = std_nice[i]; 546129c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 54628a2b7538STejun Heo 54638a2b7538STejun Heo /* 54648a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 54658a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 54668a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 54678a2b7538STejun Heo */ 54688a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 54698a2b7538STejun Heo attrs->nice = std_nice[i]; 54708a2b7538STejun Heo attrs->no_numa = true; 54718a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 547229c91e99STejun Heo } 547329c91e99STejun Heo 5474d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 54751aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5476d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5477f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5478f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 547924d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 548024d51addSTejun Heo WQ_FREEZABLE, 0); 54810668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 54820668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 54830668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 54840668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 54850668106cSViresh Kumar 0); 54861aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 54870668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 54880668106cSViresh Kumar !system_power_efficient_wq || 54890668106cSViresh Kumar !system_freezable_power_efficient_wq); 549082607adcSTejun Heo 549182607adcSTejun Heo wq_watchdog_init(); 549282607adcSTejun Heo 54936ee0578bSSuresh Siddha return 0; 54941da177e4SLinus Torvalds } 54956ee0578bSSuresh Siddha early_initcall(init_workqueues); 5496