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 151bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 152bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 153ea1abd61SLai Jiangshan 154ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 155bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 156bd7bdd43STejun Heo 157bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 158bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 159bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 160bd7bdd43STejun Heo 161c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 162c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 163c9e7cf27STejun Heo /* L: hash of busy workers */ 164c9e7cf27STejun Heo 165bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16634a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 1672607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 16892f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 16992f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17060f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 171e19e397aSTejun Heo 1727cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 173e19e397aSTejun Heo 1747a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17568e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17668e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1777a4e344cSTejun Heo 178e19e397aSTejun Heo /* 179e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 180e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 181e19e397aSTejun Heo * cacheline. 182e19e397aSTejun Heo */ 183e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18429c91e99STejun Heo 18529c91e99STejun Heo /* 18629c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18729c91e99STejun Heo * from get_work_pool(). 18829c91e99STejun Heo */ 18929c91e99STejun Heo struct rcu_head rcu; 1908b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1918b03ae3cSTejun Heo 1928b03ae3cSTejun Heo /* 193112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 194112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 195112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 196112202d9STejun Heo * number of flag bits. 1971da177e4SLinus Torvalds */ 198112202d9STejun Heo struct pool_workqueue { 199bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2004690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20173f53c4aSTejun Heo int work_color; /* L: current color */ 20273f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2038864b4e5STejun Heo int refcnt; /* L: reference count */ 20473f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20573f53c4aSTejun Heo /* L: nr of in_flight works */ 2061e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 207a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2081e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2093c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2102e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2118864b4e5STejun Heo 2128864b4e5STejun Heo /* 2138864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2148864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2158864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 216b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2178864b4e5STejun Heo */ 2188864b4e5STejun Heo struct work_struct unbound_release_work; 2198864b4e5STejun Heo struct rcu_head rcu; 220e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2211da177e4SLinus Torvalds 2221da177e4SLinus Torvalds /* 22373f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22473f53c4aSTejun Heo */ 22573f53c4aSTejun Heo struct wq_flusher { 2263c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2273c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22873f53c4aSTejun Heo struct completion done; /* flush completion */ 22973f53c4aSTejun Heo }; 2301da177e4SLinus Torvalds 231226223abSTejun Heo struct wq_device; 232226223abSTejun Heo 23373f53c4aSTejun Heo /* 234c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 235c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2361da177e4SLinus Torvalds */ 2371da177e4SLinus Torvalds struct workqueue_struct { 2383c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 239e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 24073f53c4aSTejun Heo 2413c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2423c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2433c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 244112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2453c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2463c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2473c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24873f53c4aSTejun Heo 2492e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 250e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 251e22bee78STejun Heo 25287fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 253a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 254226223abSTejun Heo 2555b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2565b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2576029a918STejun Heo 258226223abSTejun Heo #ifdef CONFIG_SYSFS 259226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 260226223abSTejun Heo #endif 2614e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2624e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2634e6045f1SJohannes Berg #endif 264ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2652728fd2fSTejun Heo 266e2dca7adSTejun Heo /* 267e2dca7adSTejun Heo * Destruction of workqueue_struct is sched-RCU protected to allow 268e2dca7adSTejun Heo * walking the workqueues list without grabbing wq_pool_mutex. 269e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 270e2dca7adSTejun Heo */ 271e2dca7adSTejun Heo struct rcu_head rcu; 272e2dca7adSTejun Heo 2732728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2742728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2752728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2765b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2771da177e4SLinus Torvalds }; 2781da177e4SLinus Torvalds 279e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 280e904e6c2STejun Heo 281bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 282bce90380STejun Heo /* possible CPUs of each node */ 283bce90380STejun Heo 284d55262c4STejun Heo static bool wq_disable_numa; 285d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 286d55262c4STejun Heo 287cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 288552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 289cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 290cee22a15SViresh Kumar 291bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 292bce90380STejun Heo 2934c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2944c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2954c16bd32STejun Heo 29668e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2972e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2985bcab335STejun Heo 299e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30068e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3017d19c5ceSTejun Heo 302042f7df1SLai Jiangshan static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */ 303b05a7928SFrederic Weisbecker 3047d19c5ceSTejun Heo /* the per-cpu worker pools */ 3057d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 3067d19c5ceSTejun Heo cpu_worker_pools); 3077d19c5ceSTejun Heo 30868e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3097d19c5ceSTejun Heo 31068e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 31129c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 31229c91e99STejun Heo 313c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 31429c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 31529c91e99STejun Heo 3168a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3178a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3188a2b7538STejun Heo 319d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 320ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 321044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3221aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 323044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 324d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 325044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 326f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 327044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 32824d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3290668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3300668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3310668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3320668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 333d320c038STejun Heo 3347d19c5ceSTejun Heo static int worker_thread(void *__worker); 3356ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3367d19c5ceSTejun Heo 33797bd2347STejun Heo #define CREATE_TRACE_POINTS 33897bd2347STejun Heo #include <trace/events/workqueue.h> 33997bd2347STejun Heo 34068e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 341f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 342f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 34368e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3445bcab335STejun Heo 345b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 346f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 347f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 348b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 34976af4d93STejun Heo 3505b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 351f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 352f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 353f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3545b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3555b95e1afSLai Jiangshan 356f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 357f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 358f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3597a62c2c8STejun Heo (pool)++) 3604ce62e9eSTejun Heo 36149e3cf44STejun Heo /** 36217116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 36317116969STejun Heo * @pool: iteration cursor 364611c92a0STejun Heo * @pi: integer used for iteration 365fa1b54e6STejun Heo * 36668e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 36768e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 36868e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 369fa1b54e6STejun Heo * 370fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 371fa1b54e6STejun Heo * ignored. 37217116969STejun Heo */ 373611c92a0STejun Heo #define for_each_pool(pool, pi) \ 374611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 37568e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 376fa1b54e6STejun Heo else 37717116969STejun Heo 37817116969STejun Heo /** 379822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 380822d8405STejun Heo * @worker: iteration cursor 381822d8405STejun Heo * @pool: worker_pool to iterate workers of 382822d8405STejun Heo * 38392f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 384822d8405STejun Heo * 385822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 386822d8405STejun Heo * ignored. 387822d8405STejun Heo */ 388da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 389da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 39092f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 391822d8405STejun Heo else 392822d8405STejun Heo 393822d8405STejun Heo /** 39449e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 39549e3cf44STejun Heo * @pwq: iteration cursor 39649e3cf44STejun Heo * @wq: the target workqueue 39776af4d93STejun Heo * 398b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 399794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 400794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 40176af4d93STejun Heo * 40276af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 40376af4d93STejun Heo * ignored. 40449e3cf44STejun Heo */ 40549e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 40676af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 407b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 40876af4d93STejun Heo else 409f3421797STejun Heo 410dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 411dc186ad7SThomas Gleixner 412dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 413dc186ad7SThomas Gleixner 41499777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 41599777288SStanislaw Gruszka { 41699777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 41799777288SStanislaw Gruszka } 41899777288SStanislaw Gruszka 419dc186ad7SThomas Gleixner /* 420dc186ad7SThomas Gleixner * fixup_init is called when: 421dc186ad7SThomas Gleixner * - an active object is initialized 422dc186ad7SThomas Gleixner */ 423dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 424dc186ad7SThomas Gleixner { 425dc186ad7SThomas Gleixner struct work_struct *work = addr; 426dc186ad7SThomas Gleixner 427dc186ad7SThomas Gleixner switch (state) { 428dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 429dc186ad7SThomas Gleixner cancel_work_sync(work); 430dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 431dc186ad7SThomas Gleixner return 1; 432dc186ad7SThomas Gleixner default: 433dc186ad7SThomas Gleixner return 0; 434dc186ad7SThomas Gleixner } 435dc186ad7SThomas Gleixner } 436dc186ad7SThomas Gleixner 437dc186ad7SThomas Gleixner /* 438dc186ad7SThomas Gleixner * fixup_activate is called when: 439dc186ad7SThomas Gleixner * - an active object is activated 440dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 441dc186ad7SThomas Gleixner */ 442dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 443dc186ad7SThomas Gleixner { 444dc186ad7SThomas Gleixner struct work_struct *work = addr; 445dc186ad7SThomas Gleixner 446dc186ad7SThomas Gleixner switch (state) { 447dc186ad7SThomas Gleixner 448dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 449dc186ad7SThomas Gleixner /* 450dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 451dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 452dc186ad7SThomas Gleixner * is tracked in the object tracker. 453dc186ad7SThomas Gleixner */ 45422df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 455dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 456dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 457dc186ad7SThomas Gleixner return 0; 458dc186ad7SThomas Gleixner } 459dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 460dc186ad7SThomas Gleixner return 0; 461dc186ad7SThomas Gleixner 462dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 463dc186ad7SThomas Gleixner WARN_ON(1); 464dc186ad7SThomas Gleixner 465dc186ad7SThomas Gleixner default: 466dc186ad7SThomas Gleixner return 0; 467dc186ad7SThomas Gleixner } 468dc186ad7SThomas Gleixner } 469dc186ad7SThomas Gleixner 470dc186ad7SThomas Gleixner /* 471dc186ad7SThomas Gleixner * fixup_free is called when: 472dc186ad7SThomas Gleixner * - an active object is freed 473dc186ad7SThomas Gleixner */ 474dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 475dc186ad7SThomas Gleixner { 476dc186ad7SThomas Gleixner struct work_struct *work = addr; 477dc186ad7SThomas Gleixner 478dc186ad7SThomas Gleixner switch (state) { 479dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 480dc186ad7SThomas Gleixner cancel_work_sync(work); 481dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 482dc186ad7SThomas Gleixner return 1; 483dc186ad7SThomas Gleixner default: 484dc186ad7SThomas Gleixner return 0; 485dc186ad7SThomas Gleixner } 486dc186ad7SThomas Gleixner } 487dc186ad7SThomas Gleixner 488dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 489dc186ad7SThomas Gleixner .name = "work_struct", 49099777288SStanislaw Gruszka .debug_hint = work_debug_hint, 491dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 492dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 493dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 494dc186ad7SThomas Gleixner }; 495dc186ad7SThomas Gleixner 496dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 497dc186ad7SThomas Gleixner { 498dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 499dc186ad7SThomas Gleixner } 500dc186ad7SThomas Gleixner 501dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 502dc186ad7SThomas Gleixner { 503dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 504dc186ad7SThomas Gleixner } 505dc186ad7SThomas Gleixner 506dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 507dc186ad7SThomas Gleixner { 508dc186ad7SThomas Gleixner if (onstack) 509dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 510dc186ad7SThomas Gleixner else 511dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 512dc186ad7SThomas Gleixner } 513dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 514dc186ad7SThomas Gleixner 515dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 516dc186ad7SThomas Gleixner { 517dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 518dc186ad7SThomas Gleixner } 519dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 520dc186ad7SThomas Gleixner 521ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 522ea2e64f2SThomas Gleixner { 523ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 524ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 525ea2e64f2SThomas Gleixner } 526ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 527ea2e64f2SThomas Gleixner 528dc186ad7SThomas Gleixner #else 529dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 530dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 531dc186ad7SThomas Gleixner #endif 532dc186ad7SThomas Gleixner 5334e8b22bdSLi Bin /** 5344e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5354e8b22bdSLi Bin * @pool: the pool pointer of interest 5364e8b22bdSLi Bin * 5374e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5384e8b22bdSLi Bin * successfully, -errno on failure. 5394e8b22bdSLi Bin */ 5409daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5419daf9e67STejun Heo { 5429daf9e67STejun Heo int ret; 5439daf9e67STejun Heo 54468e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5455bcab335STejun Heo 5464e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5474e8b22bdSLi Bin GFP_KERNEL); 548229641a6STejun Heo if (ret >= 0) { 549e68035fbSTejun Heo pool->id = ret; 550229641a6STejun Heo return 0; 551229641a6STejun Heo } 5529daf9e67STejun Heo return ret; 5539daf9e67STejun Heo } 5549daf9e67STejun Heo 55576af4d93STejun Heo /** 556df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 557df2d5ae4STejun Heo * @wq: the target workqueue 558df2d5ae4STejun Heo * @node: the node ID 559df2d5ae4STejun Heo * 5605b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5615b95e1afSLai Jiangshan * read locked. 562df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 563df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 564d185af30SYacine Belkadi * 565d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 566df2d5ae4STejun Heo */ 567df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 568df2d5ae4STejun Heo int node) 569df2d5ae4STejun Heo { 5705b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 571df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 572df2d5ae4STejun Heo } 573df2d5ae4STejun Heo 57473f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57573f53c4aSTejun Heo { 57673f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 57773f53c4aSTejun Heo } 57873f53c4aSTejun Heo 57973f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58073f53c4aSTejun Heo { 58173f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58273f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58373f53c4aSTejun Heo } 58473f53c4aSTejun Heo 58573f53c4aSTejun Heo static int work_next_color(int color) 58673f53c4aSTejun Heo { 58773f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 588a848e3b6SOleg Nesterov } 589a848e3b6SOleg Nesterov 5904594bf15SDavid Howells /* 591112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 592112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5937c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5947a22ad75STejun Heo * 595112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 596112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 597bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 598bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5997a22ad75STejun Heo * 600112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6017c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 602112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6037c3eed5cSTejun Heo * available only while the work item is queued. 604bbb68dfaSTejun Heo * 605bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 606bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 607bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 608bbb68dfaSTejun Heo * try to steal the PENDING bit. 6094594bf15SDavid Howells */ 6107a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6117a22ad75STejun Heo unsigned long flags) 6127a22ad75STejun Heo { 6136183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6147a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6157a22ad75STejun Heo } 6167a22ad75STejun Heo 617112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6184690c4abSTejun Heo unsigned long extra_flags) 619365970a1SDavid Howells { 620112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 621112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 622365970a1SDavid Howells } 623365970a1SDavid Howells 6244468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6254468a00fSLai Jiangshan int pool_id) 6264468a00fSLai Jiangshan { 6274468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6284468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6294468a00fSLai Jiangshan } 6304468a00fSLai Jiangshan 6317c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6327c3eed5cSTejun Heo int pool_id) 6334d707b9fSOleg Nesterov { 63423657bb1STejun Heo /* 63523657bb1STejun Heo * The following wmb is paired with the implied mb in 63623657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 63723657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 63823657bb1STejun Heo * owner. 63923657bb1STejun Heo */ 64023657bb1STejun Heo smp_wmb(); 6417c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6424d707b9fSOleg Nesterov } 6434d707b9fSOleg Nesterov 6447a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 645365970a1SDavid Howells { 6467c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6477c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6487a22ad75STejun Heo } 6497a22ad75STejun Heo 650112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6517a22ad75STejun Heo { 652e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6537a22ad75STejun Heo 654112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 655e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 656e120153dSTejun Heo else 657e120153dSTejun Heo return NULL; 6587a22ad75STejun Heo } 6597a22ad75STejun Heo 6607c3eed5cSTejun Heo /** 6617c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6627c3eed5cSTejun Heo * @work: the work item of interest 6637c3eed5cSTejun Heo * 66468e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 66568e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 66668e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 667fa1b54e6STejun Heo * 668fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 669fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 670fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 671fa1b54e6STejun Heo * returned pool is and stays online. 672d185af30SYacine Belkadi * 673d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 6747c3eed5cSTejun Heo */ 6757c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6767a22ad75STejun Heo { 677e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6787c3eed5cSTejun Heo int pool_id; 6797a22ad75STejun Heo 68068e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 681fa1b54e6STejun Heo 682112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 683112202d9STejun Heo return ((struct pool_workqueue *) 6847c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6857a22ad75STejun Heo 6867c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6877c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6887a22ad75STejun Heo return NULL; 6897a22ad75STejun Heo 690fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6917c3eed5cSTejun Heo } 6927c3eed5cSTejun Heo 6937c3eed5cSTejun Heo /** 6947c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6957c3eed5cSTejun Heo * @work: the work item of interest 6967c3eed5cSTejun Heo * 697d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 6987c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6997c3eed5cSTejun Heo */ 7007c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7017c3eed5cSTejun Heo { 70254d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7037c3eed5cSTejun Heo 704112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 705112202d9STejun Heo return ((struct pool_workqueue *) 70654d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 70754d5b7d0SLai Jiangshan 70854d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7097c3eed5cSTejun Heo } 7107c3eed5cSTejun Heo 711bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 712bbb68dfaSTejun Heo { 7137c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 714bbb68dfaSTejun Heo 7157c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7167c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 717bbb68dfaSTejun Heo } 718bbb68dfaSTejun Heo 719bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 720bbb68dfaSTejun Heo { 721bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 722bbb68dfaSTejun Heo 723112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 724bbb68dfaSTejun Heo } 725bbb68dfaSTejun Heo 726e22bee78STejun Heo /* 7273270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7283270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 729d565ed63STejun Heo * they're being called with pool->lock held. 730e22bee78STejun Heo */ 731e22bee78STejun Heo 73263d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 733649027d7STejun Heo { 734e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 735649027d7STejun Heo } 736649027d7STejun Heo 737e22bee78STejun Heo /* 738e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 739e22bee78STejun Heo * running workers. 740974271c4STejun Heo * 741974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 742706026c2STejun Heo * function will always return %true for unbound pools as long as the 743974271c4STejun Heo * worklist isn't empty. 744e22bee78STejun Heo */ 74563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 746e22bee78STejun Heo { 74763d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 748e22bee78STejun Heo } 749e22bee78STejun Heo 750e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 75163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 752e22bee78STejun Heo { 75363d95a91STejun Heo return pool->nr_idle; 754e22bee78STejun Heo } 755e22bee78STejun Heo 756e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 75763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 758e22bee78STejun Heo { 759e19e397aSTejun Heo return !list_empty(&pool->worklist) && 760e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 761e22bee78STejun Heo } 762e22bee78STejun Heo 763e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 76463d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 765e22bee78STejun Heo { 76663d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 767e22bee78STejun Heo } 768e22bee78STejun Heo 769e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 77063d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 771e22bee78STejun Heo { 77234a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 77363d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 77463d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 775e22bee78STejun Heo 776e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 777e22bee78STejun Heo } 778e22bee78STejun Heo 779e22bee78STejun Heo /* 780e22bee78STejun Heo * Wake up functions. 781e22bee78STejun Heo */ 782e22bee78STejun Heo 7831037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 7841037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 7857e11629dSTejun Heo { 78663d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7877e11629dSTejun Heo return NULL; 7887e11629dSTejun Heo 78963d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7907e11629dSTejun Heo } 7917e11629dSTejun Heo 7927e11629dSTejun Heo /** 7937e11629dSTejun Heo * wake_up_worker - wake up an idle worker 79463d95a91STejun Heo * @pool: worker pool to wake worker from 7957e11629dSTejun Heo * 79663d95a91STejun Heo * Wake up the first idle worker of @pool. 7977e11629dSTejun Heo * 7987e11629dSTejun Heo * CONTEXT: 799d565ed63STejun Heo * spin_lock_irq(pool->lock). 8007e11629dSTejun Heo */ 80163d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8027e11629dSTejun Heo { 8031037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8047e11629dSTejun Heo 8057e11629dSTejun Heo if (likely(worker)) 8067e11629dSTejun Heo wake_up_process(worker->task); 8077e11629dSTejun Heo } 8087e11629dSTejun Heo 8094690c4abSTejun Heo /** 810e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 811e22bee78STejun Heo * @task: task waking up 812e22bee78STejun Heo * @cpu: CPU @task is waking up to 813e22bee78STejun Heo * 814e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 815e22bee78STejun Heo * being awoken. 816e22bee78STejun Heo * 817e22bee78STejun Heo * CONTEXT: 818e22bee78STejun Heo * spin_lock_irq(rq->lock) 819e22bee78STejun Heo */ 820d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 821e22bee78STejun Heo { 822e22bee78STejun Heo struct worker *worker = kthread_data(task); 823e22bee78STejun Heo 82436576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 825ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 826e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 827e22bee78STejun Heo } 82836576000SJoonsoo Kim } 829e22bee78STejun Heo 830e22bee78STejun Heo /** 831e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 832e22bee78STejun Heo * @task: task going to sleep 833e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 834e22bee78STejun Heo * 835e22bee78STejun Heo * This function is called during schedule() when a busy worker is 836e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 837e22bee78STejun Heo * returning pointer to its task. 838e22bee78STejun Heo * 839e22bee78STejun Heo * CONTEXT: 840e22bee78STejun Heo * spin_lock_irq(rq->lock) 841e22bee78STejun Heo * 842d185af30SYacine Belkadi * Return: 843e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 844e22bee78STejun Heo */ 845d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 846e22bee78STejun Heo { 847e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 848111c225aSTejun Heo struct worker_pool *pool; 849e22bee78STejun Heo 850111c225aSTejun Heo /* 851111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 852111c225aSTejun Heo * workers, also reach here, let's not access anything before 853111c225aSTejun Heo * checking NOT_RUNNING. 854111c225aSTejun Heo */ 8552d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 856e22bee78STejun Heo return NULL; 857e22bee78STejun Heo 858111c225aSTejun Heo pool = worker->pool; 859111c225aSTejun Heo 860e22bee78STejun Heo /* this can only happen on the local cpu */ 86192b69f50SLai Jiangshan if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) 8626183c009STejun Heo return NULL; 863e22bee78STejun Heo 864e22bee78STejun Heo /* 865e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 866e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 867e22bee78STejun Heo * Please read comment there. 868e22bee78STejun Heo * 869628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 870628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 871628c78e7STejun Heo * disabled, which in turn means that none else could be 872d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 873628c78e7STejun Heo * lock is safe. 874e22bee78STejun Heo */ 875e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 876e19e397aSTejun Heo !list_empty(&pool->worklist)) 8771037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 878e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 879e22bee78STejun Heo } 880e22bee78STejun Heo 881e22bee78STejun Heo /** 882e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 883cb444766STejun Heo * @worker: self 884d302f017STejun Heo * @flags: flags to set 885d302f017STejun Heo * 886228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 887d302f017STejun Heo * 888cb444766STejun Heo * CONTEXT: 889d565ed63STejun Heo * spin_lock_irq(pool->lock) 890d302f017STejun Heo */ 891228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 892d302f017STejun Heo { 893bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 894e22bee78STejun Heo 895cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 896cb444766STejun Heo 897228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 898e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 899e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 900e19e397aSTejun Heo atomic_dec(&pool->nr_running); 901e22bee78STejun Heo } 902e22bee78STejun Heo 903d302f017STejun Heo worker->flags |= flags; 904d302f017STejun Heo } 905d302f017STejun Heo 906d302f017STejun Heo /** 907e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 908cb444766STejun Heo * @worker: self 909d302f017STejun Heo * @flags: flags to clear 910d302f017STejun Heo * 911e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 912d302f017STejun Heo * 913cb444766STejun Heo * CONTEXT: 914d565ed63STejun Heo * spin_lock_irq(pool->lock) 915d302f017STejun Heo */ 916d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 917d302f017STejun Heo { 91863d95a91STejun Heo struct worker_pool *pool = worker->pool; 919e22bee78STejun Heo unsigned int oflags = worker->flags; 920e22bee78STejun Heo 921cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 922cb444766STejun Heo 923d302f017STejun Heo worker->flags &= ~flags; 924e22bee78STejun Heo 92542c025f3STejun Heo /* 92642c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 92742c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 92842c025f3STejun Heo * of multiple flags, not a single flag. 92942c025f3STejun Heo */ 930e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 931e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 932e19e397aSTejun Heo atomic_inc(&pool->nr_running); 933d302f017STejun Heo } 934d302f017STejun Heo 935d302f017STejun Heo /** 9368cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 937c9e7cf27STejun Heo * @pool: pool of interest 9388cca0eeaSTejun Heo * @work: work to find worker for 9398cca0eeaSTejun Heo * 940c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 941c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 942a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 943a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 944a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 945a2c1c57bSTejun Heo * being executed. 946a2c1c57bSTejun Heo * 947a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 948a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 949a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 950a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 951a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 952a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 953a2c1c57bSTejun Heo * 954c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 955c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 956c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 957c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 958c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 959c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9608cca0eeaSTejun Heo * 9618cca0eeaSTejun Heo * CONTEXT: 962d565ed63STejun Heo * spin_lock_irq(pool->lock). 9638cca0eeaSTejun Heo * 964d185af30SYacine Belkadi * Return: 965d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9668cca0eeaSTejun Heo * otherwise. 9678cca0eeaSTejun Heo */ 968c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9698cca0eeaSTejun Heo struct work_struct *work) 9708cca0eeaSTejun Heo { 97142f8570fSSasha Levin struct worker *worker; 97242f8570fSSasha Levin 973b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 974a2c1c57bSTejun Heo (unsigned long)work) 975a2c1c57bSTejun Heo if (worker->current_work == work && 976a2c1c57bSTejun Heo worker->current_func == work->func) 97742f8570fSSasha Levin return worker; 97842f8570fSSasha Levin 97942f8570fSSasha Levin return NULL; 9808cca0eeaSTejun Heo } 9818cca0eeaSTejun Heo 9828cca0eeaSTejun Heo /** 983bf4ede01STejun Heo * move_linked_works - move linked works to a list 984bf4ede01STejun Heo * @work: start of series of works to be scheduled 985bf4ede01STejun Heo * @head: target list to append @work to 986402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 987bf4ede01STejun Heo * 988bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 989bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 990bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 991bf4ede01STejun Heo * 992bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 993bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 994bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 995bf4ede01STejun Heo * 996bf4ede01STejun Heo * CONTEXT: 997d565ed63STejun Heo * spin_lock_irq(pool->lock). 998bf4ede01STejun Heo */ 999bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1000bf4ede01STejun Heo struct work_struct **nextp) 1001bf4ede01STejun Heo { 1002bf4ede01STejun Heo struct work_struct *n; 1003bf4ede01STejun Heo 1004bf4ede01STejun Heo /* 1005bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1006bf4ede01STejun Heo * use NULL for list head. 1007bf4ede01STejun Heo */ 1008bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1009bf4ede01STejun Heo list_move_tail(&work->entry, head); 1010bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1011bf4ede01STejun Heo break; 1012bf4ede01STejun Heo } 1013bf4ede01STejun Heo 1014bf4ede01STejun Heo /* 1015bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1016bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1017bf4ede01STejun Heo * needs to be updated. 1018bf4ede01STejun Heo */ 1019bf4ede01STejun Heo if (nextp) 1020bf4ede01STejun Heo *nextp = n; 1021bf4ede01STejun Heo } 1022bf4ede01STejun Heo 10238864b4e5STejun Heo /** 10248864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10258864b4e5STejun Heo * @pwq: pool_workqueue to get 10268864b4e5STejun Heo * 10278864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10288864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10298864b4e5STejun Heo */ 10308864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10318864b4e5STejun Heo { 10328864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10338864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10348864b4e5STejun Heo pwq->refcnt++; 10358864b4e5STejun Heo } 10368864b4e5STejun Heo 10378864b4e5STejun Heo /** 10388864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10398864b4e5STejun Heo * @pwq: pool_workqueue to put 10408864b4e5STejun Heo * 10418864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10428864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10438864b4e5STejun Heo */ 10448864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10458864b4e5STejun Heo { 10468864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10478864b4e5STejun Heo if (likely(--pwq->refcnt)) 10488864b4e5STejun Heo return; 10498864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10508864b4e5STejun Heo return; 10518864b4e5STejun Heo /* 10528864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10538864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10548864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10558864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10568864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10578864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10588864b4e5STejun Heo */ 10598864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10608864b4e5STejun Heo } 10618864b4e5STejun Heo 1062dce90d47STejun Heo /** 1063dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1064dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1065dce90d47STejun Heo * 1066dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1067dce90d47STejun Heo */ 1068dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1069dce90d47STejun Heo { 1070dce90d47STejun Heo if (pwq) { 1071dce90d47STejun Heo /* 1072dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1073dce90d47STejun Heo * following lock operations are safe. 1074dce90d47STejun Heo */ 1075dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1076dce90d47STejun Heo put_pwq(pwq); 1077dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1078dce90d47STejun Heo } 1079dce90d47STejun Heo } 1080dce90d47STejun Heo 1081112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1082bf4ede01STejun Heo { 1083112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1084bf4ede01STejun Heo 1085bf4ede01STejun Heo trace_workqueue_activate_work(work); 1086112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1087bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1088112202d9STejun Heo pwq->nr_active++; 1089bf4ede01STejun Heo } 1090bf4ede01STejun Heo 1091112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10923aa62497SLai Jiangshan { 1093112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10943aa62497SLai Jiangshan struct work_struct, entry); 10953aa62497SLai Jiangshan 1096112202d9STejun Heo pwq_activate_delayed_work(work); 10973aa62497SLai Jiangshan } 10983aa62497SLai Jiangshan 1099bf4ede01STejun Heo /** 1100112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1101112202d9STejun Heo * @pwq: pwq of interest 1102bf4ede01STejun Heo * @color: color of work which left the queue 1103bf4ede01STejun Heo * 1104bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1105112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1106bf4ede01STejun Heo * 1107bf4ede01STejun Heo * CONTEXT: 1108d565ed63STejun Heo * spin_lock_irq(pool->lock). 1109bf4ede01STejun Heo */ 1110112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1111bf4ede01STejun Heo { 11128864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1113bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11148864b4e5STejun Heo goto out_put; 1115bf4ede01STejun Heo 1116112202d9STejun Heo pwq->nr_in_flight[color]--; 1117bf4ede01STejun Heo 1118112202d9STejun Heo pwq->nr_active--; 1119112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1120bf4ede01STejun Heo /* one down, submit a delayed one */ 1121112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1122112202d9STejun Heo pwq_activate_first_delayed(pwq); 1123bf4ede01STejun Heo } 1124bf4ede01STejun Heo 1125bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1126112202d9STejun Heo if (likely(pwq->flush_color != color)) 11278864b4e5STejun Heo goto out_put; 1128bf4ede01STejun Heo 1129bf4ede01STejun Heo /* are there still in-flight works? */ 1130112202d9STejun Heo if (pwq->nr_in_flight[color]) 11318864b4e5STejun Heo goto out_put; 1132bf4ede01STejun Heo 1133112202d9STejun Heo /* this pwq is done, clear flush_color */ 1134112202d9STejun Heo pwq->flush_color = -1; 1135bf4ede01STejun Heo 1136bf4ede01STejun Heo /* 1137112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1138bf4ede01STejun Heo * will handle the rest. 1139bf4ede01STejun Heo */ 1140112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1141112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11428864b4e5STejun Heo out_put: 11438864b4e5STejun Heo put_pwq(pwq); 1144bf4ede01STejun Heo } 1145bf4ede01STejun Heo 114636e227d2STejun Heo /** 1147bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 114836e227d2STejun Heo * @work: work item to steal 114936e227d2STejun Heo * @is_dwork: @work is a delayed_work 1150bbb68dfaSTejun Heo * @flags: place to store irq state 115136e227d2STejun Heo * 115236e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1153d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 115436e227d2STejun Heo * 1155d185af30SYacine Belkadi * Return: 115636e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 115736e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 115836e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1159bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1160bbb68dfaSTejun Heo * for arbitrarily long 116136e227d2STejun Heo * 1162d185af30SYacine Belkadi * Note: 1163bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1164e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1165e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1166e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1167bbb68dfaSTejun Heo * 1168bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1169bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1170bbb68dfaSTejun Heo * 1171e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1172bf4ede01STejun Heo */ 1173bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1174bbb68dfaSTejun Heo unsigned long *flags) 1175bf4ede01STejun Heo { 1176d565ed63STejun Heo struct worker_pool *pool; 1177112202d9STejun Heo struct pool_workqueue *pwq; 1178bf4ede01STejun Heo 1179bbb68dfaSTejun Heo local_irq_save(*flags); 1180bbb68dfaSTejun Heo 118136e227d2STejun Heo /* try to steal the timer if it exists */ 118236e227d2STejun Heo if (is_dwork) { 118336e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 118436e227d2STejun Heo 1185e0aecdd8STejun Heo /* 1186e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1187e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1188e0aecdd8STejun Heo * running on the local CPU. 1189e0aecdd8STejun Heo */ 119036e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 119136e227d2STejun Heo return 1; 119236e227d2STejun Heo } 119336e227d2STejun Heo 119436e227d2STejun Heo /* try to claim PENDING the normal way */ 1195bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1196bf4ede01STejun Heo return 0; 1197bf4ede01STejun Heo 1198bf4ede01STejun Heo /* 1199bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1200bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1201bf4ede01STejun Heo */ 1202d565ed63STejun Heo pool = get_work_pool(work); 1203d565ed63STejun Heo if (!pool) 1204bbb68dfaSTejun Heo goto fail; 1205bf4ede01STejun Heo 1206d565ed63STejun Heo spin_lock(&pool->lock); 1207bf4ede01STejun Heo /* 1208112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1209112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1210112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1211112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1212112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12130b3dae68SLai Jiangshan * item is currently queued on that pool. 1214bf4ede01STejun Heo */ 1215112202d9STejun Heo pwq = get_work_pwq(work); 1216112202d9STejun Heo if (pwq && pwq->pool == pool) { 1217bf4ede01STejun Heo debug_work_deactivate(work); 12183aa62497SLai Jiangshan 12193aa62497SLai Jiangshan /* 122016062836STejun Heo * A delayed work item cannot be grabbed directly because 122116062836STejun Heo * it might have linked NO_COLOR work items which, if left 1222112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 122316062836STejun Heo * management later on and cause stall. Make sure the work 122416062836STejun Heo * item is activated before grabbing. 12253aa62497SLai Jiangshan */ 12263aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1227112202d9STejun Heo pwq_activate_delayed_work(work); 12283aa62497SLai Jiangshan 1229bf4ede01STejun Heo list_del_init(&work->entry); 12309c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 123136e227d2STejun Heo 1232112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12334468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12344468a00fSLai Jiangshan 1235d565ed63STejun Heo spin_unlock(&pool->lock); 123636e227d2STejun Heo return 1; 1237bf4ede01STejun Heo } 1238d565ed63STejun Heo spin_unlock(&pool->lock); 1239bbb68dfaSTejun Heo fail: 1240bbb68dfaSTejun Heo local_irq_restore(*flags); 1241bbb68dfaSTejun Heo if (work_is_canceling(work)) 1242bbb68dfaSTejun Heo return -ENOENT; 1243bbb68dfaSTejun Heo cpu_relax(); 124436e227d2STejun Heo return -EAGAIN; 1245bf4ede01STejun Heo } 1246bf4ede01STejun Heo 1247bf4ede01STejun Heo /** 1248706026c2STejun Heo * insert_work - insert a work into a pool 1249112202d9STejun Heo * @pwq: pwq @work belongs to 12504690c4abSTejun Heo * @work: work to insert 12514690c4abSTejun Heo * @head: insertion point 12524690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12534690c4abSTejun Heo * 1254112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1255706026c2STejun Heo * work_struct flags. 12564690c4abSTejun Heo * 12574690c4abSTejun Heo * CONTEXT: 1258d565ed63STejun Heo * spin_lock_irq(pool->lock). 1259365970a1SDavid Howells */ 1260112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1261112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1262b89deed3SOleg Nesterov { 1263112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1264e1d8aa9fSFrederic Weisbecker 12654690c4abSTejun Heo /* we own @work, set data and link */ 1266112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12671a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12688864b4e5STejun Heo get_pwq(pwq); 1269e22bee78STejun Heo 1270e22bee78STejun Heo /* 1271c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1272c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1273c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1274e22bee78STejun Heo */ 1275e22bee78STejun Heo smp_mb(); 1276e22bee78STejun Heo 127763d95a91STejun Heo if (__need_more_worker(pool)) 127863d95a91STejun Heo wake_up_worker(pool); 1279b89deed3SOleg Nesterov } 1280b89deed3SOleg Nesterov 1281c8efcc25STejun Heo /* 1282c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12838d03ecfeSTejun Heo * same workqueue. 1284c8efcc25STejun Heo */ 1285c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1286c8efcc25STejun Heo { 1287c8efcc25STejun Heo struct worker *worker; 1288c8efcc25STejun Heo 12898d03ecfeSTejun Heo worker = current_wq_worker(); 1290c8efcc25STejun Heo /* 12918d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12928d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1293c8efcc25STejun Heo */ 1294112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1295c8efcc25STejun Heo } 1296c8efcc25STejun Heo 1297d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12981da177e4SLinus Torvalds struct work_struct *work) 12991da177e4SLinus Torvalds { 1300112202d9STejun Heo struct pool_workqueue *pwq; 1301c9178087STejun Heo struct worker_pool *last_pool; 13021e19ffc6STejun Heo struct list_head *worklist; 13038a2e8e5dSTejun Heo unsigned int work_flags; 1304b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13058930cabaSTejun Heo 13068930cabaSTejun Heo /* 13078930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13088930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13098930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13108930cabaSTejun Heo * happen with IRQ disabled. 13118930cabaSTejun Heo */ 13128930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13131da177e4SLinus Torvalds 1314dc186ad7SThomas Gleixner debug_work_activate(work); 13151e19ffc6STejun Heo 13169ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1317618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1318c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1319e41e704bSTejun Heo return; 13209e8cd2f5STejun Heo retry: 1321df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1322f3421797STejun Heo cpu = raw_smp_processor_id(); 1323df2d5ae4STejun Heo 1324df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1325df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1326c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1327df2d5ae4STejun Heo else 1328df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1329f3421797STejun Heo 133018aa9effSTejun Heo /* 1331c9178087STejun Heo * If @work was previously on a different pool, it might still be 1332c9178087STejun Heo * running there, in which case the work needs to be queued on that 1333c9178087STejun Heo * pool to guarantee non-reentrancy. 133418aa9effSTejun Heo */ 1335c9e7cf27STejun Heo last_pool = get_work_pool(work); 1336112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 133718aa9effSTejun Heo struct worker *worker; 133818aa9effSTejun Heo 1339d565ed63STejun Heo spin_lock(&last_pool->lock); 134018aa9effSTejun Heo 1341c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 134218aa9effSTejun Heo 1343112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1344c9178087STejun Heo pwq = worker->current_pwq; 13458594fadeSLai Jiangshan } else { 134618aa9effSTejun Heo /* meh... not running there, queue here */ 1347d565ed63STejun Heo spin_unlock(&last_pool->lock); 1348112202d9STejun Heo spin_lock(&pwq->pool->lock); 134918aa9effSTejun Heo } 13508930cabaSTejun Heo } else { 1351112202d9STejun Heo spin_lock(&pwq->pool->lock); 13528930cabaSTejun Heo } 1353502ca9d8STejun Heo 13549e8cd2f5STejun Heo /* 13559e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13569e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13579e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1358df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1359df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13609e8cd2f5STejun Heo * make forward-progress. 13619e8cd2f5STejun Heo */ 13629e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13639e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13649e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13659e8cd2f5STejun Heo cpu_relax(); 13669e8cd2f5STejun Heo goto retry; 13679e8cd2f5STejun Heo } 13689e8cd2f5STejun Heo /* oops */ 13699e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13709e8cd2f5STejun Heo wq->name, cpu); 13719e8cd2f5STejun Heo } 13729e8cd2f5STejun Heo 1373112202d9STejun Heo /* pwq determined, queue */ 1374112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1375502ca9d8STejun Heo 1376f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1377112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1378f5b2552bSDan Carpenter return; 1379f5b2552bSDan Carpenter } 13801e19ffc6STejun Heo 1381112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1382112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13831e19ffc6STejun Heo 1384112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1385cdadf009STejun Heo trace_workqueue_activate_work(work); 1386112202d9STejun Heo pwq->nr_active++; 1387112202d9STejun Heo worklist = &pwq->pool->worklist; 13888a2e8e5dSTejun Heo } else { 13898a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1390112202d9STejun Heo worklist = &pwq->delayed_works; 13918a2e8e5dSTejun Heo } 13921e19ffc6STejun Heo 1393112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13941e19ffc6STejun Heo 1395112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13961da177e4SLinus Torvalds } 13971da177e4SLinus Torvalds 13980fcb78c2SRolf Eike Beer /** 1399c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1400c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1401c1a220e7SZhang Rui * @wq: workqueue to use 1402c1a220e7SZhang Rui * @work: work to queue 1403c1a220e7SZhang Rui * 1404c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1405c1a220e7SZhang Rui * can't go away. 1406d185af30SYacine Belkadi * 1407d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1408c1a220e7SZhang Rui */ 1409d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1410d4283e93STejun Heo struct work_struct *work) 1411c1a220e7SZhang Rui { 1412d4283e93STejun Heo bool ret = false; 14138930cabaSTejun Heo unsigned long flags; 14148930cabaSTejun Heo 14158930cabaSTejun Heo local_irq_save(flags); 1416c1a220e7SZhang Rui 141722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14184690c4abSTejun Heo __queue_work(cpu, wq, work); 1419d4283e93STejun Heo ret = true; 1420c1a220e7SZhang Rui } 14218930cabaSTejun Heo 14228930cabaSTejun Heo local_irq_restore(flags); 1423c1a220e7SZhang Rui return ret; 1424c1a220e7SZhang Rui } 1425ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1426c1a220e7SZhang Rui 1427d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14281da177e4SLinus Torvalds { 142952bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14301da177e4SLinus Torvalds 1431e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 143260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14331da177e4SLinus Torvalds } 14341438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14351da177e4SLinus Torvalds 14367beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 143752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14381da177e4SLinus Torvalds { 14397beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14407beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14411da177e4SLinus Torvalds 14427beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14437beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1444fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1445fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14467beb2edfSTejun Heo 14478852aac2STejun Heo /* 14488852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14498852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14508852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14518852aac2STejun Heo * on that there's no such delay when @delay is 0. 14528852aac2STejun Heo */ 14538852aac2STejun Heo if (!delay) { 14548852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14558852aac2STejun Heo return; 14568852aac2STejun Heo } 14578852aac2STejun Heo 14587beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14597beb2edfSTejun Heo 146060c057bcSLai Jiangshan dwork->wq = wq; 14611265057fSTejun Heo dwork->cpu = cpu; 14627beb2edfSTejun Heo timer->expires = jiffies + delay; 14637beb2edfSTejun Heo 14647beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14657beb2edfSTejun Heo add_timer_on(timer, cpu); 14667beb2edfSTejun Heo else 14677beb2edfSTejun Heo add_timer(timer); 14687beb2edfSTejun Heo } 14691da177e4SLinus Torvalds 14700fcb78c2SRolf Eike Beer /** 14710fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14720fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14730fcb78c2SRolf Eike Beer * @wq: workqueue to use 1474af9997e4SRandy Dunlap * @dwork: work to queue 14750fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14760fcb78c2SRolf Eike Beer * 1477d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1478715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1479715f1300STejun Heo * execution. 14800fcb78c2SRolf Eike Beer */ 1481d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 148252bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14837a6bc1cdSVenkatesh Pallipadi { 148452bad64dSDavid Howells struct work_struct *work = &dwork->work; 1485d4283e93STejun Heo bool ret = false; 14868930cabaSTejun Heo unsigned long flags; 14878930cabaSTejun Heo 14888930cabaSTejun Heo /* read the comment in __queue_work() */ 14898930cabaSTejun Heo local_irq_save(flags); 14907a6bc1cdSVenkatesh Pallipadi 149122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14927beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1493d4283e93STejun Heo ret = true; 14947a6bc1cdSVenkatesh Pallipadi } 14958930cabaSTejun Heo 14968930cabaSTejun Heo local_irq_restore(flags); 14977a6bc1cdSVenkatesh Pallipadi return ret; 14987a6bc1cdSVenkatesh Pallipadi } 1499ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15001da177e4SLinus Torvalds 1501c8e55f36STejun Heo /** 15028376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15038376fe22STejun Heo * @cpu: CPU number to execute work on 15048376fe22STejun Heo * @wq: workqueue to use 15058376fe22STejun Heo * @dwork: work to queue 15068376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15078376fe22STejun Heo * 15088376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15098376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15108376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15118376fe22STejun Heo * current state. 15128376fe22STejun Heo * 1513d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15148376fe22STejun Heo * pending and its timer was modified. 15158376fe22STejun Heo * 1516e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15178376fe22STejun Heo * See try_to_grab_pending() for details. 15188376fe22STejun Heo */ 15198376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15208376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15218376fe22STejun Heo { 15228376fe22STejun Heo unsigned long flags; 15238376fe22STejun Heo int ret; 15248376fe22STejun Heo 15258376fe22STejun Heo do { 15268376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15278376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15288376fe22STejun Heo 15298376fe22STejun Heo if (likely(ret >= 0)) { 15308376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15318376fe22STejun Heo local_irq_restore(flags); 15328376fe22STejun Heo } 15338376fe22STejun Heo 15348376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15358376fe22STejun Heo return ret; 15368376fe22STejun Heo } 15378376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15388376fe22STejun Heo 15398376fe22STejun Heo /** 1540c8e55f36STejun Heo * worker_enter_idle - enter idle state 1541c8e55f36STejun Heo * @worker: worker which is entering idle state 1542c8e55f36STejun Heo * 1543c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1544c8e55f36STejun Heo * necessary. 1545c8e55f36STejun Heo * 1546c8e55f36STejun Heo * LOCKING: 1547d565ed63STejun Heo * spin_lock_irq(pool->lock). 1548c8e55f36STejun Heo */ 1549c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15501da177e4SLinus Torvalds { 1551bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1552c8e55f36STejun Heo 15536183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15546183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15556183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15566183c009STejun Heo return; 1557c8e55f36STejun Heo 1558051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1559cb444766STejun Heo worker->flags |= WORKER_IDLE; 1560bd7bdd43STejun Heo pool->nr_idle++; 1561e22bee78STejun Heo worker->last_active = jiffies; 1562c8e55f36STejun Heo 1563c8e55f36STejun Heo /* idle_list is LIFO */ 1564bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1565db7bccf4STejun Heo 156663d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1567628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1568cb444766STejun Heo 1569544ecf31STejun Heo /* 1570706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1571d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1572628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1573628c78e7STejun Heo * unbind is not in progress. 1574544ecf31STejun Heo */ 157524647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1576bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1577e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1578c8e55f36STejun Heo } 1579c8e55f36STejun Heo 1580c8e55f36STejun Heo /** 1581c8e55f36STejun Heo * worker_leave_idle - leave idle state 1582c8e55f36STejun Heo * @worker: worker which is leaving idle state 1583c8e55f36STejun Heo * 1584c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1585c8e55f36STejun Heo * 1586c8e55f36STejun Heo * LOCKING: 1587d565ed63STejun Heo * spin_lock_irq(pool->lock). 1588c8e55f36STejun Heo */ 1589c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1590c8e55f36STejun Heo { 1591bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1592c8e55f36STejun Heo 15936183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15946183c009STejun Heo return; 1595d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1596bd7bdd43STejun Heo pool->nr_idle--; 1597c8e55f36STejun Heo list_del_init(&worker->entry); 1598c8e55f36STejun Heo } 1599c8e55f36STejun Heo 1600f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1601c34056a3STejun Heo { 1602c34056a3STejun Heo struct worker *worker; 1603c34056a3STejun Heo 1604f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1605c8e55f36STejun Heo if (worker) { 1606c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1607affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1608da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1609e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1610e22bee78STejun Heo worker->flags = WORKER_PREP; 1611c8e55f36STejun Heo } 1612c34056a3STejun Heo return worker; 1613c34056a3STejun Heo } 1614c34056a3STejun Heo 1615c34056a3STejun Heo /** 16164736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 16174736cbf7SLai Jiangshan * @worker: worker to be attached 16184736cbf7SLai Jiangshan * @pool: the target pool 16194736cbf7SLai Jiangshan * 16204736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 16214736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 16224736cbf7SLai Jiangshan * cpu-[un]hotplugs. 16234736cbf7SLai Jiangshan */ 16244736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 16254736cbf7SLai Jiangshan struct worker_pool *pool) 16264736cbf7SLai Jiangshan { 16274736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 16284736cbf7SLai Jiangshan 16294736cbf7SLai Jiangshan /* 16304736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 16314736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 16324736cbf7SLai Jiangshan */ 16334736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16344736cbf7SLai Jiangshan 16354736cbf7SLai Jiangshan /* 16364736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 16374736cbf7SLai Jiangshan * stable across this function. See the comments above the 16384736cbf7SLai Jiangshan * flag definition for details. 16394736cbf7SLai Jiangshan */ 16404736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 16414736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 16424736cbf7SLai Jiangshan 16434736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 16444736cbf7SLai Jiangshan 16454736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 16464736cbf7SLai Jiangshan } 16474736cbf7SLai Jiangshan 16484736cbf7SLai Jiangshan /** 164960f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 165060f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 165160f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 165260f5a4bcSLai Jiangshan * 16534736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 16544736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 16554736cbf7SLai Jiangshan * other reference to the pool. 165660f5a4bcSLai Jiangshan */ 165760f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 165860f5a4bcSLai Jiangshan struct worker_pool *pool) 165960f5a4bcSLai Jiangshan { 166060f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 166160f5a4bcSLai Jiangshan 166292f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1663da028469SLai Jiangshan list_del(&worker->node); 1664da028469SLai Jiangshan if (list_empty(&pool->workers)) 166560f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 166692f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 166760f5a4bcSLai Jiangshan 1668b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1669b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1670b62c0751SLai Jiangshan 167160f5a4bcSLai Jiangshan if (detach_completion) 167260f5a4bcSLai Jiangshan complete(detach_completion); 167360f5a4bcSLai Jiangshan } 167460f5a4bcSLai Jiangshan 167560f5a4bcSLai Jiangshan /** 1676c34056a3STejun Heo * create_worker - create a new workqueue worker 167763d95a91STejun Heo * @pool: pool the new worker will belong to 1678c34056a3STejun Heo * 1679051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1680c34056a3STejun Heo * 1681c34056a3STejun Heo * CONTEXT: 1682c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1683c34056a3STejun Heo * 1684d185af30SYacine Belkadi * Return: 1685c34056a3STejun Heo * Pointer to the newly created worker. 1686c34056a3STejun Heo */ 1687bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1688c34056a3STejun Heo { 1689c34056a3STejun Heo struct worker *worker = NULL; 1690f3421797STejun Heo int id = -1; 1691e3c916a4STejun Heo char id_buf[16]; 1692c34056a3STejun Heo 16937cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 16947cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1695822d8405STejun Heo if (id < 0) 1696c34056a3STejun Heo goto fail; 1697c34056a3STejun Heo 1698f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1699c34056a3STejun Heo if (!worker) 1700c34056a3STejun Heo goto fail; 1701c34056a3STejun Heo 1702bd7bdd43STejun Heo worker->pool = pool; 1703c34056a3STejun Heo worker->id = id; 1704c34056a3STejun Heo 170529c91e99STejun Heo if (pool->cpu >= 0) 1706e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1707e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1708f3421797STejun Heo else 1709e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1710e3c916a4STejun Heo 1711f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1712e3c916a4STejun Heo "kworker/%s", id_buf); 1713c34056a3STejun Heo if (IS_ERR(worker->task)) 1714c34056a3STejun Heo goto fail; 1715c34056a3STejun Heo 171691151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 171791151228SOleg Nesterov 171891151228SOleg Nesterov /* prevent userland from meddling with cpumask of workqueue workers */ 171991151228SOleg Nesterov worker->task->flags |= PF_NO_SETAFFINITY; 172091151228SOleg Nesterov 1721da028469SLai Jiangshan /* successful, attach the worker to the pool */ 17224736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1723822d8405STejun Heo 1724051e1850SLai Jiangshan /* start the newly created worker */ 1725051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1726051e1850SLai Jiangshan worker->pool->nr_workers++; 1727051e1850SLai Jiangshan worker_enter_idle(worker); 1728051e1850SLai Jiangshan wake_up_process(worker->task); 1729051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1730051e1850SLai Jiangshan 1731c34056a3STejun Heo return worker; 1732822d8405STejun Heo 1733c34056a3STejun Heo fail: 17349625ab17SLai Jiangshan if (id >= 0) 17357cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1736c34056a3STejun Heo kfree(worker); 1737c34056a3STejun Heo return NULL; 1738c34056a3STejun Heo } 1739c34056a3STejun Heo 1740c34056a3STejun Heo /** 1741c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1742c34056a3STejun Heo * @worker: worker to be destroyed 1743c34056a3STejun Heo * 174473eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 174573eb7fe7SLai Jiangshan * be idle. 1746c8e55f36STejun Heo * 1747c8e55f36STejun Heo * CONTEXT: 174860f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1749c34056a3STejun Heo */ 1750c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1751c34056a3STejun Heo { 1752bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1753c34056a3STejun Heo 1754cd549687STejun Heo lockdep_assert_held(&pool->lock); 1755cd549687STejun Heo 1756c34056a3STejun Heo /* sanity check frenzy */ 17576183c009STejun Heo if (WARN_ON(worker->current_work) || 175873eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 175973eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 17606183c009STejun Heo return; 1761c34056a3STejun Heo 1762bd7bdd43STejun Heo pool->nr_workers--; 1763bd7bdd43STejun Heo pool->nr_idle--; 1764c8e55f36STejun Heo 1765c8e55f36STejun Heo list_del_init(&worker->entry); 1766cb444766STejun Heo worker->flags |= WORKER_DIE; 176760f5a4bcSLai Jiangshan wake_up_process(worker->task); 1768c34056a3STejun Heo } 1769c34056a3STejun Heo 177063d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1771e22bee78STejun Heo { 177263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1773e22bee78STejun Heo 1774d565ed63STejun Heo spin_lock_irq(&pool->lock); 1775e22bee78STejun Heo 17763347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1777e22bee78STejun Heo struct worker *worker; 1778e22bee78STejun Heo unsigned long expires; 1779e22bee78STejun Heo 1780e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 178163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1782e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1783e22bee78STejun Heo 17843347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 178563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 17863347fc9fSLai Jiangshan break; 1787e22bee78STejun Heo } 17883347fc9fSLai Jiangshan 17893347fc9fSLai Jiangshan destroy_worker(worker); 1790e22bee78STejun Heo } 1791e22bee78STejun Heo 1792d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1793e22bee78STejun Heo } 1794e22bee78STejun Heo 1795493a1724STejun Heo static void send_mayday(struct work_struct *work) 1796e22bee78STejun Heo { 1797112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1798112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1799493a1724STejun Heo 18002e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1801e22bee78STejun Heo 1802493008a8STejun Heo if (!wq->rescuer) 1803493a1724STejun Heo return; 1804e22bee78STejun Heo 1805e22bee78STejun Heo /* mayday mayday mayday */ 1806493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 180777668c8bSLai Jiangshan /* 180877668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 180977668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 181077668c8bSLai Jiangshan * rescuer is done with it. 181177668c8bSLai Jiangshan */ 181277668c8bSLai Jiangshan get_pwq(pwq); 1813493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1814e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1815493a1724STejun Heo } 1816e22bee78STejun Heo } 1817e22bee78STejun Heo 1818706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1819e22bee78STejun Heo { 182063d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1821e22bee78STejun Heo struct work_struct *work; 1822e22bee78STejun Heo 1823b2d82909STejun Heo spin_lock_irq(&pool->lock); 1824b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1825e22bee78STejun Heo 182663d95a91STejun Heo if (need_to_create_worker(pool)) { 1827e22bee78STejun Heo /* 1828e22bee78STejun Heo * We've been trying to create a new worker but 1829e22bee78STejun Heo * haven't been successful. We might be hitting an 1830e22bee78STejun Heo * allocation deadlock. Send distress signals to 1831e22bee78STejun Heo * rescuers. 1832e22bee78STejun Heo */ 183363d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1834e22bee78STejun Heo send_mayday(work); 1835e22bee78STejun Heo } 1836e22bee78STejun Heo 1837b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1838b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1839e22bee78STejun Heo 184063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1841e22bee78STejun Heo } 1842e22bee78STejun Heo 1843e22bee78STejun Heo /** 1844e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 184563d95a91STejun Heo * @pool: pool to create a new worker for 1846e22bee78STejun Heo * 184763d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1848e22bee78STejun Heo * have at least one idle worker on return from this function. If 1849e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 185063d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1851e22bee78STejun Heo * possible allocation deadlock. 1852e22bee78STejun Heo * 1853c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1854c5aa87bbSTejun Heo * may_start_working() %true. 1855e22bee78STejun Heo * 1856e22bee78STejun Heo * LOCKING: 1857d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1858e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1859e22bee78STejun Heo * manager. 1860e22bee78STejun Heo */ 186129187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1862d565ed63STejun Heo __releases(&pool->lock) 1863d565ed63STejun Heo __acquires(&pool->lock) 1864e22bee78STejun Heo { 1865e22bee78STejun Heo restart: 1866d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18679f9c2364STejun Heo 1868e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 186963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1870e22bee78STejun Heo 1871e22bee78STejun Heo while (true) { 1872051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1873e22bee78STejun Heo break; 1874e22bee78STejun Heo 1875e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 18769f9c2364STejun Heo 187763d95a91STejun Heo if (!need_to_create_worker(pool)) 1878e22bee78STejun Heo break; 1879e22bee78STejun Heo } 1880e22bee78STejun Heo 188163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1882d565ed63STejun Heo spin_lock_irq(&pool->lock); 1883051e1850SLai Jiangshan /* 1884051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1885051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1886051e1850SLai Jiangshan * already become busy. 1887051e1850SLai Jiangshan */ 188863d95a91STejun Heo if (need_to_create_worker(pool)) 1889e22bee78STejun Heo goto restart; 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 1892e22bee78STejun Heo /** 1893e22bee78STejun Heo * manage_workers - manage worker pool 1894e22bee78STejun Heo * @worker: self 1895e22bee78STejun Heo * 1896706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1897e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1898706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1899e22bee78STejun Heo * 1900e22bee78STejun Heo * The caller can safely start processing works on false return. On 1901e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1902e22bee78STejun Heo * and may_start_working() is true. 1903e22bee78STejun Heo * 1904e22bee78STejun Heo * CONTEXT: 1905d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1906e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1907e22bee78STejun Heo * 1908d185af30SYacine Belkadi * Return: 190929187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 191029187a9eSTejun Heo * start processing works, %true if management function was performed and 191129187a9eSTejun Heo * the conditions that the caller verified before calling the function may 191229187a9eSTejun Heo * no longer be true. 1913e22bee78STejun Heo */ 1914e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1915e22bee78STejun Heo { 191663d95a91STejun Heo struct worker_pool *pool = worker->pool; 1917e22bee78STejun Heo 1918bc3a1afcSTejun Heo /* 1919bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1920bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1921bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1922bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1923bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1924bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1925bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1926bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1927bc3a1afcSTejun Heo */ 192834a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 192929187a9eSTejun Heo return false; 19302607d7a6STejun Heo pool->manager = worker; 1931e22bee78STejun Heo 193229187a9eSTejun Heo maybe_create_worker(pool); 1933e22bee78STejun Heo 19342607d7a6STejun Heo pool->manager = NULL; 193534a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 193629187a9eSTejun Heo return true; 1937e22bee78STejun Heo } 1938e22bee78STejun Heo 1939a62428c0STejun Heo /** 1940a62428c0STejun Heo * process_one_work - process single work 1941c34056a3STejun Heo * @worker: self 1942a62428c0STejun Heo * @work: work to process 1943a62428c0STejun Heo * 1944a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1945a62428c0STejun Heo * process a single work including synchronization against and 1946a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1947a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1948a62428c0STejun Heo * call this function to process a work. 1949a62428c0STejun Heo * 1950a62428c0STejun Heo * CONTEXT: 1951d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1952a62428c0STejun Heo */ 1953c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 1954d565ed63STejun Heo __releases(&pool->lock) 1955d565ed63STejun Heo __acquires(&pool->lock) 19561da177e4SLinus Torvalds { 1957112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1958bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1959112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 196073f53c4aSTejun Heo int work_color; 19617e11629dSTejun Heo struct worker *collision; 19624e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 19634e6045f1SJohannes Berg /* 1964a62428c0STejun Heo * It is permissible to free the struct work_struct from 1965a62428c0STejun Heo * inside the function that is called from it, this we need to 1966a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1967a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1968a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 19694e6045f1SJohannes Berg */ 19704d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 19714d82a1deSPeter Zijlstra 19724d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 19734e6045f1SJohannes Berg #endif 1974807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 197585327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1976ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 197725511a47STejun Heo 19787e11629dSTejun Heo /* 19797e11629dSTejun Heo * A single work shouldn't be executed concurrently by 19807e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 19817e11629dSTejun Heo * already processing the work. If so, defer the work to the 19827e11629dSTejun Heo * currently executing one. 19837e11629dSTejun Heo */ 1984c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 19857e11629dSTejun Heo if (unlikely(collision)) { 19867e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 19877e11629dSTejun Heo return; 19887e11629dSTejun Heo } 19891da177e4SLinus Torvalds 19908930cabaSTejun Heo /* claim and dequeue */ 19911da177e4SLinus Torvalds debug_work_deactivate(work); 1992c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 1993c34056a3STejun Heo worker->current_work = work; 1994a2c1c57bSTejun Heo worker->current_func = work->func; 1995112202d9STejun Heo worker->current_pwq = pwq; 199673f53c4aSTejun Heo work_color = get_work_color(work); 19977a22ad75STejun Heo 1998a62428c0STejun Heo list_del_init(&work->entry); 1999a62428c0STejun Heo 2000649027d7STejun Heo /* 2001228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2002228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2003228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2004228f1d00SLai Jiangshan * execution of the pending work items. 2005fb0e7bebSTejun Heo */ 2006fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2007228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2008fb0e7bebSTejun Heo 2009974271c4STejun Heo /* 2010a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2011a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2012a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2013a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2014228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2015974271c4STejun Heo */ 2016a489a03eSLai Jiangshan if (need_more_worker(pool)) 201763d95a91STejun Heo wake_up_worker(pool); 2018974271c4STejun Heo 20198930cabaSTejun Heo /* 20207c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2021d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 202223657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 202323657bb1STejun Heo * disabled. 20248930cabaSTejun Heo */ 20257c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 20261da177e4SLinus Torvalds 2027d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2028365970a1SDavid Howells 2029112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 20303295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2031e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2032a2c1c57bSTejun Heo worker->current_func(work); 2033e36c886aSArjan van de Ven /* 2034e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2035e36c886aSArjan van de Ven * point will only record its address. 2036e36c886aSArjan van de Ven */ 2037e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20383295f0efSIngo Molnar lock_map_release(&lockdep_map); 2039112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 20401da177e4SLinus Torvalds 2041d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2042044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2043044c782cSValentin Ilie " last function: %pf\n", 2044a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2045a2c1c57bSTejun Heo worker->current_func); 2046d5abe669SPeter Zijlstra debug_show_held_locks(current); 2047d5abe669SPeter Zijlstra dump_stack(); 2048d5abe669SPeter Zijlstra } 2049d5abe669SPeter Zijlstra 2050b22ce278STejun Heo /* 2051b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2052b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2053b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2054b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2055789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2056789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2057b22ce278STejun Heo */ 20583e28e377SJoe Lawrence cond_resched_rcu_qs(); 2059b22ce278STejun Heo 2060d565ed63STejun Heo spin_lock_irq(&pool->lock); 2061a62428c0STejun Heo 2062fb0e7bebSTejun Heo /* clear cpu intensive status */ 2063fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2064fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2065fb0e7bebSTejun Heo 2066a62428c0STejun Heo /* we're done with it, release */ 206742f8570fSSasha Levin hash_del(&worker->hentry); 2068c34056a3STejun Heo worker->current_work = NULL; 2069a2c1c57bSTejun Heo worker->current_func = NULL; 2070112202d9STejun Heo worker->current_pwq = NULL; 20713d1cb205STejun Heo worker->desc_valid = false; 2072112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 20731da177e4SLinus Torvalds } 20741da177e4SLinus Torvalds 2075affee4b2STejun Heo /** 2076affee4b2STejun Heo * process_scheduled_works - process scheduled works 2077affee4b2STejun Heo * @worker: self 2078affee4b2STejun Heo * 2079affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2080affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2081affee4b2STejun Heo * fetches a work from the top and executes it. 2082affee4b2STejun Heo * 2083affee4b2STejun Heo * CONTEXT: 2084d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2085affee4b2STejun Heo * multiple times. 2086affee4b2STejun Heo */ 2087affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 20881da177e4SLinus Torvalds { 2089affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2090affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2091a62428c0STejun Heo struct work_struct, entry); 2092c34056a3STejun Heo process_one_work(worker, work); 2093a62428c0STejun Heo } 20941da177e4SLinus Torvalds } 20951da177e4SLinus Torvalds 20964690c4abSTejun Heo /** 20974690c4abSTejun Heo * worker_thread - the worker thread function 2098c34056a3STejun Heo * @__worker: self 20994690c4abSTejun Heo * 2100c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2101c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2102c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2103c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2104c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2105d185af30SYacine Belkadi * 2106d185af30SYacine Belkadi * Return: 0 21074690c4abSTejun Heo */ 2108c34056a3STejun Heo static int worker_thread(void *__worker) 21091da177e4SLinus Torvalds { 2110c34056a3STejun Heo struct worker *worker = __worker; 2111bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21121da177e4SLinus Torvalds 2113e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2114e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2115c8e55f36STejun Heo woke_up: 2116d565ed63STejun Heo spin_lock_irq(&pool->lock); 2117affee4b2STejun Heo 2118a9ab775bSTejun Heo /* am I supposed to die? */ 2119a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2120d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2121a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2122e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 212360f5a4bcSLai Jiangshan 212460f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 21257cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 212660f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 212760f5a4bcSLai Jiangshan kfree(worker); 2128c8e55f36STejun Heo return 0; 2129c8e55f36STejun Heo } 2130c8e55f36STejun Heo 2131c8e55f36STejun Heo worker_leave_idle(worker); 2132db7bccf4STejun Heo recheck: 2133e22bee78STejun Heo /* no more worker necessary? */ 213463d95a91STejun Heo if (!need_more_worker(pool)) 2135e22bee78STejun Heo goto sleep; 2136e22bee78STejun Heo 2137e22bee78STejun Heo /* do we need to manage? */ 213863d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2139e22bee78STejun Heo goto recheck; 2140e22bee78STejun Heo 2141c8e55f36STejun Heo /* 2142c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2143c8e55f36STejun Heo * preparing to process a work or actually processing it. 2144c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2145c8e55f36STejun Heo */ 21466183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2147c8e55f36STejun Heo 2148e22bee78STejun Heo /* 2149a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2150a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2151a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2152a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2153a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2154e22bee78STejun Heo */ 2155a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2156e22bee78STejun Heo 2157e22bee78STejun Heo do { 2158affee4b2STejun Heo struct work_struct *work = 2159bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2160affee4b2STejun Heo struct work_struct, entry); 2161affee4b2STejun Heo 2162c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2163affee4b2STejun Heo /* optimization path, not strictly necessary */ 2164affee4b2STejun Heo process_one_work(worker, work); 2165affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2166affee4b2STejun Heo process_scheduled_works(worker); 2167affee4b2STejun Heo } else { 2168c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2169affee4b2STejun Heo process_scheduled_works(worker); 2170affee4b2STejun Heo } 217163d95a91STejun Heo } while (keep_working(pool)); 2172affee4b2STejun Heo 2173228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2174d313dd85STejun Heo sleep: 2175c8e55f36STejun Heo /* 2176d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2177d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2178d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2179d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2180d565ed63STejun Heo * event. 2181c8e55f36STejun Heo */ 2182c8e55f36STejun Heo worker_enter_idle(worker); 2183c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2184d565ed63STejun Heo spin_unlock_irq(&pool->lock); 21851da177e4SLinus Torvalds schedule(); 2186c8e55f36STejun Heo goto woke_up; 21871da177e4SLinus Torvalds } 21881da177e4SLinus Torvalds 2189e22bee78STejun Heo /** 2190e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2191111c225aSTejun Heo * @__rescuer: self 2192e22bee78STejun Heo * 2193e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2194493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2195e22bee78STejun Heo * 2196706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2197e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2198e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2199e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2200e22bee78STejun Heo * the problem rescuer solves. 2201e22bee78STejun Heo * 2202706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2203706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2204e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2205e22bee78STejun Heo * 2206e22bee78STejun Heo * This should happen rarely. 2207d185af30SYacine Belkadi * 2208d185af30SYacine Belkadi * Return: 0 2209e22bee78STejun Heo */ 2210111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2211e22bee78STejun Heo { 2212111c225aSTejun Heo struct worker *rescuer = __rescuer; 2213111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2214e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 22154d595b86SLai Jiangshan bool should_stop; 2216e22bee78STejun Heo 2217e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2218111c225aSTejun Heo 2219111c225aSTejun Heo /* 2220111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2221111c225aSTejun Heo * doesn't participate in concurrency management. 2222111c225aSTejun Heo */ 2223111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2224e22bee78STejun Heo repeat: 2225e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 22261da177e4SLinus Torvalds 22274d595b86SLai Jiangshan /* 22284d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 22294d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 22304d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 22314d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 22324d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 22334d595b86SLai Jiangshan * list is always empty on exit. 22344d595b86SLai Jiangshan */ 22354d595b86SLai Jiangshan should_stop = kthread_should_stop(); 22361da177e4SLinus Torvalds 2237493a1724STejun Heo /* see whether any pwq is asking for help */ 22382e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2239493a1724STejun Heo 2240493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2241493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2242493a1724STejun Heo struct pool_workqueue, mayday_node); 2243112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2244e22bee78STejun Heo struct work_struct *work, *n; 2245e22bee78STejun Heo 2246e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2247493a1724STejun Heo list_del_init(&pwq->mayday_node); 2248493a1724STejun Heo 22492e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2250e22bee78STejun Heo 225151697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 225251697d39SLai Jiangshan 225351697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2254b3104104SLai Jiangshan rescuer->pool = pool; 2255e22bee78STejun Heo 2256e22bee78STejun Heo /* 2257e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2258e22bee78STejun Heo * process'em. 2259e22bee78STejun Heo */ 22600479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 2261bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2262112202d9STejun Heo if (get_work_pwq(work) == pwq) 2263e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2264e22bee78STejun Heo 2265008847f6SNeilBrown if (!list_empty(scheduled)) { 2266e22bee78STejun Heo process_scheduled_works(rescuer); 22677576958aSTejun Heo 22687576958aSTejun Heo /* 2269008847f6SNeilBrown * The above execution of rescued work items could 2270008847f6SNeilBrown * have created more to rescue through 2271008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2272008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2273008847f6SNeilBrown * that such back-to-back work items, which may be 2274008847f6SNeilBrown * being used to relieve memory pressure, don't 2275008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2276008847f6SNeilBrown */ 2277008847f6SNeilBrown if (need_to_create_worker(pool)) { 2278008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2279008847f6SNeilBrown get_pwq(pwq); 2280008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2281008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2282008847f6SNeilBrown } 2283008847f6SNeilBrown } 2284008847f6SNeilBrown 2285008847f6SNeilBrown /* 228677668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 228713b1d625SLai Jiangshan * go away while we're still attached to it. 228877668c8bSLai Jiangshan */ 228977668c8bSLai Jiangshan put_pwq(pwq); 229077668c8bSLai Jiangshan 229177668c8bSLai Jiangshan /* 2292d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 22937576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 22947576958aSTejun Heo * and stalling the execution. 22957576958aSTejun Heo */ 2296d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 229763d95a91STejun Heo wake_up_worker(pool); 22987576958aSTejun Heo 2299b3104104SLai Jiangshan rescuer->pool = NULL; 230013b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 230113b1d625SLai Jiangshan 230213b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 230313b1d625SLai Jiangshan 230413b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 23051da177e4SLinus Torvalds } 23061da177e4SLinus Torvalds 23072e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2308493a1724STejun Heo 23094d595b86SLai Jiangshan if (should_stop) { 23104d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 23114d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 23124d595b86SLai Jiangshan return 0; 23134d595b86SLai Jiangshan } 23144d595b86SLai Jiangshan 2315111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2316111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2317e22bee78STejun Heo schedule(); 2318e22bee78STejun Heo goto repeat; 23191da177e4SLinus Torvalds } 23201da177e4SLinus Torvalds 2321fc2e4d70SOleg Nesterov struct wq_barrier { 2322fc2e4d70SOleg Nesterov struct work_struct work; 2323fc2e4d70SOleg Nesterov struct completion done; 23242607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2325fc2e4d70SOleg Nesterov }; 2326fc2e4d70SOleg Nesterov 2327fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2328fc2e4d70SOleg Nesterov { 2329fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2330fc2e4d70SOleg Nesterov complete(&barr->done); 2331fc2e4d70SOleg Nesterov } 2332fc2e4d70SOleg Nesterov 23334690c4abSTejun Heo /** 23344690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2335112202d9STejun Heo * @pwq: pwq to insert barrier into 23364690c4abSTejun Heo * @barr: wq_barrier to insert 2337affee4b2STejun Heo * @target: target work to attach @barr to 2338affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23394690c4abSTejun Heo * 2340affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2341affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2342affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2343affee4b2STejun Heo * cpu. 2344affee4b2STejun Heo * 2345affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2346affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2347affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2348affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2349affee4b2STejun Heo * after a work with LINKED flag set. 2350affee4b2STejun Heo * 2351affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2352112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23534690c4abSTejun Heo * 23544690c4abSTejun Heo * CONTEXT: 2355d565ed63STejun Heo * spin_lock_irq(pool->lock). 23564690c4abSTejun Heo */ 2357112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2358affee4b2STejun Heo struct wq_barrier *barr, 2359affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2360fc2e4d70SOleg Nesterov { 2361affee4b2STejun Heo struct list_head *head; 2362affee4b2STejun Heo unsigned int linked = 0; 2363affee4b2STejun Heo 2364dc186ad7SThomas Gleixner /* 2365d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2366dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2367dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2368dc186ad7SThomas Gleixner * might deadlock. 2369dc186ad7SThomas Gleixner */ 2370ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 237122df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2372fc2e4d70SOleg Nesterov init_completion(&barr->done); 23732607d7a6STejun Heo barr->task = current; 237483c22520SOleg Nesterov 2375affee4b2STejun Heo /* 2376affee4b2STejun Heo * If @target is currently being executed, schedule the 2377affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2378affee4b2STejun Heo */ 2379affee4b2STejun Heo if (worker) 2380affee4b2STejun Heo head = worker->scheduled.next; 2381affee4b2STejun Heo else { 2382affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2383affee4b2STejun Heo 2384affee4b2STejun Heo head = target->entry.next; 2385affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2386affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2387affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2388affee4b2STejun Heo } 2389affee4b2STejun Heo 2390dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2391112202d9STejun Heo insert_work(pwq, &barr->work, head, 2392affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2393fc2e4d70SOleg Nesterov } 2394fc2e4d70SOleg Nesterov 239573f53c4aSTejun Heo /** 2396112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 239773f53c4aSTejun Heo * @wq: workqueue being flushed 239873f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 239973f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 240073f53c4aSTejun Heo * 2401112202d9STejun Heo * Prepare pwqs for workqueue flushing. 240273f53c4aSTejun Heo * 2403112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2404112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2405112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2406112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2407112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 240873f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 240973f53c4aSTejun Heo * 241073f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 241173f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 241273f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 241373f53c4aSTejun Heo * is returned. 241473f53c4aSTejun Heo * 2415112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 241673f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 241773f53c4aSTejun Heo * advanced to @work_color. 241873f53c4aSTejun Heo * 241973f53c4aSTejun Heo * CONTEXT: 24203c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 242173f53c4aSTejun Heo * 2422d185af30SYacine Belkadi * Return: 242373f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 242473f53c4aSTejun Heo * otherwise. 242573f53c4aSTejun Heo */ 2426112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 242773f53c4aSTejun Heo int flush_color, int work_color) 24281da177e4SLinus Torvalds { 242973f53c4aSTejun Heo bool wait = false; 243049e3cf44STejun Heo struct pool_workqueue *pwq; 24311da177e4SLinus Torvalds 243273f53c4aSTejun Heo if (flush_color >= 0) { 24336183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2434112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2435dc186ad7SThomas Gleixner } 243614441960SOleg Nesterov 243749e3cf44STejun Heo for_each_pwq(pwq, wq) { 2438112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24391da177e4SLinus Torvalds 2440b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 244173f53c4aSTejun Heo 244273f53c4aSTejun Heo if (flush_color >= 0) { 24436183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 244473f53c4aSTejun Heo 2445112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2446112202d9STejun Heo pwq->flush_color = flush_color; 2447112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 244873f53c4aSTejun Heo wait = true; 24491da177e4SLinus Torvalds } 245073f53c4aSTejun Heo } 245173f53c4aSTejun Heo 245273f53c4aSTejun Heo if (work_color >= 0) { 24536183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2454112202d9STejun Heo pwq->work_color = work_color; 245573f53c4aSTejun Heo } 245673f53c4aSTejun Heo 2457b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 24581da177e4SLinus Torvalds } 24591da177e4SLinus Torvalds 2460112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 246173f53c4aSTejun Heo complete(&wq->first_flusher->done); 246273f53c4aSTejun Heo 246373f53c4aSTejun Heo return wait; 246483c22520SOleg Nesterov } 24651da177e4SLinus Torvalds 24660fcb78c2SRolf Eike Beer /** 24671da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 24680fcb78c2SRolf Eike Beer * @wq: workqueue to flush 24691da177e4SLinus Torvalds * 2470c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2471c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 24721da177e4SLinus Torvalds */ 24737ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 24741da177e4SLinus Torvalds { 247573f53c4aSTejun Heo struct wq_flusher this_flusher = { 247673f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 247773f53c4aSTejun Heo .flush_color = -1, 247873f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 247973f53c4aSTejun Heo }; 248073f53c4aSTejun Heo int next_color; 2481b1f4ec17SOleg Nesterov 24823295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 24833295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 248473f53c4aSTejun Heo 24853c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 248673f53c4aSTejun Heo 248773f53c4aSTejun Heo /* 248873f53c4aSTejun Heo * Start-to-wait phase 248973f53c4aSTejun Heo */ 249073f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 249173f53c4aSTejun Heo 249273f53c4aSTejun Heo if (next_color != wq->flush_color) { 249373f53c4aSTejun Heo /* 249473f53c4aSTejun Heo * Color space is not full. The current work_color 249573f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 249673f53c4aSTejun Heo * by one. 249773f53c4aSTejun Heo */ 24986183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 249973f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 250073f53c4aSTejun Heo wq->work_color = next_color; 250173f53c4aSTejun Heo 250273f53c4aSTejun Heo if (!wq->first_flusher) { 250373f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25046183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 250573f53c4aSTejun Heo 250673f53c4aSTejun Heo wq->first_flusher = &this_flusher; 250773f53c4aSTejun Heo 2508112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 250973f53c4aSTejun Heo wq->work_color)) { 251073f53c4aSTejun Heo /* nothing to flush, done */ 251173f53c4aSTejun Heo wq->flush_color = next_color; 251273f53c4aSTejun Heo wq->first_flusher = NULL; 251373f53c4aSTejun Heo goto out_unlock; 251473f53c4aSTejun Heo } 251573f53c4aSTejun Heo } else { 251673f53c4aSTejun Heo /* wait in queue */ 25176183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 251873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2519112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 252073f53c4aSTejun Heo } 252173f53c4aSTejun Heo } else { 252273f53c4aSTejun Heo /* 252373f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 252473f53c4aSTejun Heo * The next flush completion will assign us 252573f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 252673f53c4aSTejun Heo */ 252773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 252873f53c4aSTejun Heo } 252973f53c4aSTejun Heo 25303c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 253173f53c4aSTejun Heo 253273f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 253373f53c4aSTejun Heo 253473f53c4aSTejun Heo /* 253573f53c4aSTejun Heo * Wake-up-and-cascade phase 253673f53c4aSTejun Heo * 253773f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 253873f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 253973f53c4aSTejun Heo */ 254073f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 254173f53c4aSTejun Heo return; 254273f53c4aSTejun Heo 25433c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 254473f53c4aSTejun Heo 25454ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25464ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25474ce48b37STejun Heo goto out_unlock; 25484ce48b37STejun Heo 254973f53c4aSTejun Heo wq->first_flusher = NULL; 255073f53c4aSTejun Heo 25516183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25526183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 255373f53c4aSTejun Heo 255473f53c4aSTejun Heo while (true) { 255573f53c4aSTejun Heo struct wq_flusher *next, *tmp; 255673f53c4aSTejun Heo 255773f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 255873f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 255973f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 256073f53c4aSTejun Heo break; 256173f53c4aSTejun Heo list_del_init(&next->list); 256273f53c4aSTejun Heo complete(&next->done); 256373f53c4aSTejun Heo } 256473f53c4aSTejun Heo 25656183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 256673f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 256773f53c4aSTejun Heo 256873f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 256973f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 257073f53c4aSTejun Heo 257173f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 257273f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 257373f53c4aSTejun Heo /* 257473f53c4aSTejun Heo * Assign the same color to all overflowed 257573f53c4aSTejun Heo * flushers, advance work_color and append to 257673f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 257773f53c4aSTejun Heo * phase for these overflowed flushers. 257873f53c4aSTejun Heo */ 257973f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 258073f53c4aSTejun Heo tmp->flush_color = wq->work_color; 258173f53c4aSTejun Heo 258273f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 258373f53c4aSTejun Heo 258473f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 258573f53c4aSTejun Heo &wq->flusher_queue); 2586112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 258773f53c4aSTejun Heo } 258873f53c4aSTejun Heo 258973f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 25906183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 259173f53c4aSTejun Heo break; 259273f53c4aSTejun Heo } 259373f53c4aSTejun Heo 259473f53c4aSTejun Heo /* 259573f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2596112202d9STejun Heo * the new first flusher and arm pwqs. 259773f53c4aSTejun Heo */ 25986183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 25996183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 260073f53c4aSTejun Heo 260173f53c4aSTejun Heo list_del_init(&next->list); 260273f53c4aSTejun Heo wq->first_flusher = next; 260373f53c4aSTejun Heo 2604112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 260573f53c4aSTejun Heo break; 260673f53c4aSTejun Heo 260773f53c4aSTejun Heo /* 260873f53c4aSTejun Heo * Meh... this color is already done, clear first 260973f53c4aSTejun Heo * flusher and repeat cascading. 261073f53c4aSTejun Heo */ 261173f53c4aSTejun Heo wq->first_flusher = NULL; 261273f53c4aSTejun Heo } 261373f53c4aSTejun Heo 261473f53c4aSTejun Heo out_unlock: 26153c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26161da177e4SLinus Torvalds } 2617ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26181da177e4SLinus Torvalds 26199c5a2ba7STejun Heo /** 26209c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26219c5a2ba7STejun Heo * @wq: workqueue to drain 26229c5a2ba7STejun Heo * 26239c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26249c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26259c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2626b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 26279c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26289c5a2ba7STejun Heo * takes too long. 26299c5a2ba7STejun Heo */ 26309c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26319c5a2ba7STejun Heo { 26329c5a2ba7STejun Heo unsigned int flush_cnt = 0; 263349e3cf44STejun Heo struct pool_workqueue *pwq; 26349c5a2ba7STejun Heo 26359c5a2ba7STejun Heo /* 26369c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26379c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2638618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26399c5a2ba7STejun Heo */ 264087fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26419c5a2ba7STejun Heo if (!wq->nr_drainers++) 2642618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 264387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26449c5a2ba7STejun Heo reflush: 26459c5a2ba7STejun Heo flush_workqueue(wq); 26469c5a2ba7STejun Heo 2647b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 264876af4d93STejun Heo 264949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2650fa2563e4SThomas Tuttle bool drained; 26519c5a2ba7STejun Heo 2652b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2653112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2654b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2655fa2563e4SThomas Tuttle 2656fa2563e4SThomas Tuttle if (drained) 26579c5a2ba7STejun Heo continue; 26589c5a2ba7STejun Heo 26599c5a2ba7STejun Heo if (++flush_cnt == 10 || 26609c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2661c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 26629c5a2ba7STejun Heo wq->name, flush_cnt); 266376af4d93STejun Heo 2664b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 26659c5a2ba7STejun Heo goto reflush; 26669c5a2ba7STejun Heo } 26679c5a2ba7STejun Heo 26689c5a2ba7STejun Heo if (!--wq->nr_drainers) 2669618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 267087fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26719c5a2ba7STejun Heo } 26729c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 26739c5a2ba7STejun Heo 2674606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2675baf59022STejun Heo { 2676baf59022STejun Heo struct worker *worker = NULL; 2677c9e7cf27STejun Heo struct worker_pool *pool; 2678112202d9STejun Heo struct pool_workqueue *pwq; 2679baf59022STejun Heo 2680baf59022STejun Heo might_sleep(); 2681baf59022STejun Heo 2682fa1b54e6STejun Heo local_irq_disable(); 2683fa1b54e6STejun Heo pool = get_work_pool(work); 2684fa1b54e6STejun Heo if (!pool) { 2685fa1b54e6STejun Heo local_irq_enable(); 2686fa1b54e6STejun Heo return false; 2687fa1b54e6STejun Heo } 2688fa1b54e6STejun Heo 2689fa1b54e6STejun Heo spin_lock(&pool->lock); 26900b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2691112202d9STejun Heo pwq = get_work_pwq(work); 2692112202d9STejun Heo if (pwq) { 2693112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2694baf59022STejun Heo goto already_gone; 2695606a5020STejun Heo } else { 2696c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2697baf59022STejun Heo if (!worker) 2698baf59022STejun Heo goto already_gone; 2699112202d9STejun Heo pwq = worker->current_pwq; 2700606a5020STejun Heo } 2701baf59022STejun Heo 2702112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2703d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2704baf59022STejun Heo 2705e159489bSTejun Heo /* 2706e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2707e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2708e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2709e159489bSTejun Heo * access. 2710e159489bSTejun Heo */ 2711493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2712112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2713e159489bSTejun Heo else 2714112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2715112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2716e159489bSTejun Heo 2717baf59022STejun Heo return true; 2718baf59022STejun Heo already_gone: 2719d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2720baf59022STejun Heo return false; 2721baf59022STejun Heo } 2722baf59022STejun Heo 2723db700897SOleg Nesterov /** 2724401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2725401a8d04STejun Heo * @work: the work to flush 2726db700897SOleg Nesterov * 2727606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2728606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2729401a8d04STejun Heo * 2730d185af30SYacine Belkadi * Return: 2731401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2732401a8d04STejun Heo * %false if it was already idle. 2733db700897SOleg Nesterov */ 2734401a8d04STejun Heo bool flush_work(struct work_struct *work) 2735db700897SOleg Nesterov { 273612997d1aSBjorn Helgaas struct wq_barrier barr; 273712997d1aSBjorn Helgaas 27380976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27390976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27400976dfc1SStephen Boyd 274112997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 274212997d1aSBjorn Helgaas wait_for_completion(&barr.done); 274312997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 274412997d1aSBjorn Helgaas return true; 274512997d1aSBjorn Helgaas } else { 274612997d1aSBjorn Helgaas return false; 274712997d1aSBjorn Helgaas } 2748606a5020STejun Heo } 2749db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2750db700897SOleg Nesterov 27518603e1b3STejun Heo struct cwt_wait { 27528603e1b3STejun Heo wait_queue_t wait; 27538603e1b3STejun Heo struct work_struct *work; 27548603e1b3STejun Heo }; 27558603e1b3STejun Heo 27568603e1b3STejun Heo static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) 27578603e1b3STejun Heo { 27588603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 27598603e1b3STejun Heo 27608603e1b3STejun Heo if (cwait->work != key) 27618603e1b3STejun Heo return 0; 27628603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 27638603e1b3STejun Heo } 27648603e1b3STejun Heo 276536e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2766401a8d04STejun Heo { 27678603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2768bbb68dfaSTejun Heo unsigned long flags; 27691f1f642eSOleg Nesterov int ret; 27701f1f642eSOleg Nesterov 27711f1f642eSOleg Nesterov do { 2772bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2773bbb68dfaSTejun Heo /* 27748603e1b3STejun Heo * If someone else is already canceling, wait for it to 27758603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 27768603e1b3STejun Heo * because we may get scheduled between @work's completion 27778603e1b3STejun Heo * and the other canceling task resuming and clearing 27788603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 27798603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 27808603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 27818603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 27828603e1b3STejun Heo * we're hogging the CPU. 27838603e1b3STejun Heo * 27848603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 27858603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 27868603e1b3STejun Heo * wake function which matches @work along with exclusive 27878603e1b3STejun Heo * wait and wakeup. 2788bbb68dfaSTejun Heo */ 27898603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 27908603e1b3STejun Heo struct cwt_wait cwait; 27918603e1b3STejun Heo 27928603e1b3STejun Heo init_wait(&cwait.wait); 27938603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 27948603e1b3STejun Heo cwait.work = work; 27958603e1b3STejun Heo 27968603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 27978603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 27988603e1b3STejun Heo if (work_is_canceling(work)) 27998603e1b3STejun Heo schedule(); 28008603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 28018603e1b3STejun Heo } 28021f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28031f1f642eSOleg Nesterov 2804bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2805bbb68dfaSTejun Heo mark_work_canceling(work); 2806bbb68dfaSTejun Heo local_irq_restore(flags); 2807bbb68dfaSTejun Heo 2808606a5020STejun Heo flush_work(work); 28097a22ad75STejun Heo clear_work_data(work); 28108603e1b3STejun Heo 28118603e1b3STejun Heo /* 28128603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 28138603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 28148603e1b3STejun Heo * visible there. 28158603e1b3STejun Heo */ 28168603e1b3STejun Heo smp_mb(); 28178603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 28188603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 28198603e1b3STejun Heo 28201f1f642eSOleg Nesterov return ret; 28211f1f642eSOleg Nesterov } 28221f1f642eSOleg Nesterov 28236e84d644SOleg Nesterov /** 2824401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2825401a8d04STejun Heo * @work: the work to cancel 28266e84d644SOleg Nesterov * 2827401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2828401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2829401a8d04STejun Heo * another workqueue. On return from this function, @work is 2830401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28311f1f642eSOleg Nesterov * 2832401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2833401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28346e84d644SOleg Nesterov * 2835401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28366e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2837401a8d04STejun Heo * 2838d185af30SYacine Belkadi * Return: 2839401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28406e84d644SOleg Nesterov */ 2841401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28426e84d644SOleg Nesterov { 284336e227d2STejun Heo return __cancel_work_timer(work, false); 2844b89deed3SOleg Nesterov } 284528e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2846b89deed3SOleg Nesterov 28476e84d644SOleg Nesterov /** 2848401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2849401a8d04STejun Heo * @dwork: the delayed work to flush 28506e84d644SOleg Nesterov * 2851401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2852401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2853401a8d04STejun Heo * considers the last queueing instance of @dwork. 28541f1f642eSOleg Nesterov * 2855d185af30SYacine Belkadi * Return: 2856401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2857401a8d04STejun Heo * %false if it was already idle. 28586e84d644SOleg Nesterov */ 2859401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2860401a8d04STejun Heo { 28618930cabaSTejun Heo local_irq_disable(); 2862401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 286360c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28648930cabaSTejun Heo local_irq_enable(); 2865401a8d04STejun Heo return flush_work(&dwork->work); 2866401a8d04STejun Heo } 2867401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2868401a8d04STejun Heo 2869401a8d04STejun Heo /** 287057b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 287157b30ae7STejun Heo * @dwork: delayed_work to cancel 287209383498STejun Heo * 2873d185af30SYacine Belkadi * Kill off a pending delayed_work. 2874d185af30SYacine Belkadi * 2875d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 2876d185af30SYacine Belkadi * pending. 2877d185af30SYacine Belkadi * 2878d185af30SYacine Belkadi * Note: 2879d185af30SYacine Belkadi * The work callback function may still be running on return, unless 2880d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 2881d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 288209383498STejun Heo * 288357b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 288409383498STejun Heo */ 288557b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 288609383498STejun Heo { 288757b30ae7STejun Heo unsigned long flags; 288857b30ae7STejun Heo int ret; 288957b30ae7STejun Heo 289057b30ae7STejun Heo do { 289157b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 289257b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 289357b30ae7STejun Heo 289457b30ae7STejun Heo if (unlikely(ret < 0)) 289557b30ae7STejun Heo return false; 289657b30ae7STejun Heo 28977c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28987c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 289957b30ae7STejun Heo local_irq_restore(flags); 2900c0158ca6SDan Magenheimer return ret; 290109383498STejun Heo } 290257b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 290309383498STejun Heo 290409383498STejun Heo /** 2905401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2906401a8d04STejun Heo * @dwork: the delayed work cancel 2907401a8d04STejun Heo * 2908401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2909401a8d04STejun Heo * 2910d185af30SYacine Belkadi * Return: 2911401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2912401a8d04STejun Heo */ 2913401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29146e84d644SOleg Nesterov { 291536e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29166e84d644SOleg Nesterov } 2917f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29181da177e4SLinus Torvalds 29190fcb78c2SRolf Eike Beer /** 292031ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2921b6136773SAndrew Morton * @func: the function to call 2922b6136773SAndrew Morton * 292331ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 292431ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2925b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 292631ddd871STejun Heo * 2927d185af30SYacine Belkadi * Return: 292831ddd871STejun Heo * 0 on success, -errno on failure. 2929b6136773SAndrew Morton */ 293065f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 293115316ba8SChristoph Lameter { 293215316ba8SChristoph Lameter int cpu; 293338f51568SNamhyung Kim struct work_struct __percpu *works; 293415316ba8SChristoph Lameter 2935b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2936b6136773SAndrew Morton if (!works) 293715316ba8SChristoph Lameter return -ENOMEM; 2938b6136773SAndrew Morton 293995402b38SGautham R Shenoy get_online_cpus(); 294093981800STejun Heo 294115316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29429bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29439bfb1839SIngo Molnar 29449bfb1839SIngo Molnar INIT_WORK(work, func); 29458de6d308SOleg Nesterov schedule_work_on(cpu, work); 294615316ba8SChristoph Lameter } 294793981800STejun Heo 294893981800STejun Heo for_each_online_cpu(cpu) 29498616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 295093981800STejun Heo 295195402b38SGautham R Shenoy put_online_cpus(); 2952b6136773SAndrew Morton free_percpu(works); 295315316ba8SChristoph Lameter return 0; 295415316ba8SChristoph Lameter } 295515316ba8SChristoph Lameter 2956eef6a7d5SAlan Stern /** 29571fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29581fa44ecaSJames Bottomley * @fn: the function to execute 29591fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29601fa44ecaSJames Bottomley * be available when the work executes) 29611fa44ecaSJames Bottomley * 29621fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29631fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29641fa44ecaSJames Bottomley * 2965d185af30SYacine Belkadi * Return: 0 - function was executed 29661fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29671fa44ecaSJames Bottomley */ 296865f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29691fa44ecaSJames Bottomley { 29701fa44ecaSJames Bottomley if (!in_interrupt()) { 297165f27f38SDavid Howells fn(&ew->work); 29721fa44ecaSJames Bottomley return 0; 29731fa44ecaSJames Bottomley } 29741fa44ecaSJames Bottomley 297565f27f38SDavid Howells INIT_WORK(&ew->work, fn); 29761fa44ecaSJames Bottomley schedule_work(&ew->work); 29771fa44ecaSJames Bottomley 29781fa44ecaSJames Bottomley return 1; 29791fa44ecaSJames Bottomley } 29801fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 29811fa44ecaSJames Bottomley 29827a4e344cSTejun Heo /** 29837a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 29847a4e344cSTejun Heo * @attrs: workqueue_attrs to free 29857a4e344cSTejun Heo * 29867a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 29877a4e344cSTejun Heo */ 29887a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 29897a4e344cSTejun Heo { 29907a4e344cSTejun Heo if (attrs) { 29917a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 29927a4e344cSTejun Heo kfree(attrs); 29937a4e344cSTejun Heo } 29947a4e344cSTejun Heo } 29957a4e344cSTejun Heo 29967a4e344cSTejun Heo /** 29977a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 29987a4e344cSTejun Heo * @gfp_mask: allocation mask to use 29997a4e344cSTejun Heo * 30007a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3001d185af30SYacine Belkadi * return it. 3002d185af30SYacine Belkadi * 3003d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 30047a4e344cSTejun Heo */ 30057a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 30067a4e344cSTejun Heo { 30077a4e344cSTejun Heo struct workqueue_attrs *attrs; 30087a4e344cSTejun Heo 30097a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 30107a4e344cSTejun Heo if (!attrs) 30117a4e344cSTejun Heo goto fail; 30127a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 30137a4e344cSTejun Heo goto fail; 30147a4e344cSTejun Heo 301513e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 30167a4e344cSTejun Heo return attrs; 30177a4e344cSTejun Heo fail: 30187a4e344cSTejun Heo free_workqueue_attrs(attrs); 30197a4e344cSTejun Heo return NULL; 30207a4e344cSTejun Heo } 30217a4e344cSTejun Heo 302229c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 302329c91e99STejun Heo const struct workqueue_attrs *from) 302429c91e99STejun Heo { 302529c91e99STejun Heo to->nice = from->nice; 302629c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 30272865a8fbSShaohua Li /* 30282865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 30292865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 30302865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 30312865a8fbSShaohua Li */ 30322865a8fbSShaohua Li to->no_numa = from->no_numa; 303329c91e99STejun Heo } 303429c91e99STejun Heo 303529c91e99STejun Heo /* hash value of the content of @attr */ 303629c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 303729c91e99STejun Heo { 303829c91e99STejun Heo u32 hash = 0; 303929c91e99STejun Heo 304029c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 304113e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 304213e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 304329c91e99STejun Heo return hash; 304429c91e99STejun Heo } 304529c91e99STejun Heo 304629c91e99STejun Heo /* content equality test */ 304729c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 304829c91e99STejun Heo const struct workqueue_attrs *b) 304929c91e99STejun Heo { 305029c91e99STejun Heo if (a->nice != b->nice) 305129c91e99STejun Heo return false; 305229c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 305329c91e99STejun Heo return false; 305429c91e99STejun Heo return true; 305529c91e99STejun Heo } 305629c91e99STejun Heo 30577a4e344cSTejun Heo /** 30587a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 30597a4e344cSTejun Heo * @pool: worker_pool to initialize 30607a4e344cSTejun Heo * 3061402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3062d185af30SYacine Belkadi * 3063d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 306429c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 306529c91e99STejun Heo * on @pool safely to release it. 30667a4e344cSTejun Heo */ 30677a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 30684e1a1f9aSTejun Heo { 30694e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 307029c91e99STejun Heo pool->id = -1; 307129c91e99STejun Heo pool->cpu = -1; 3072f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 30734e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 30744e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 30754e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 30764e1a1f9aSTejun Heo hash_init(pool->busy_hash); 30774e1a1f9aSTejun Heo 30784e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 30794e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 30804e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 30814e1a1f9aSTejun Heo 30824e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 30834e1a1f9aSTejun Heo (unsigned long)pool); 30844e1a1f9aSTejun Heo 30854e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 308692f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3087da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 30887a4e344cSTejun Heo 30897cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 309029c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 309129c91e99STejun Heo pool->refcnt = 1; 309229c91e99STejun Heo 309329c91e99STejun Heo /* shouldn't fail above this point */ 30947a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 30957a4e344cSTejun Heo if (!pool->attrs) 30967a4e344cSTejun Heo return -ENOMEM; 30977a4e344cSTejun Heo return 0; 30984e1a1f9aSTejun Heo } 30994e1a1f9aSTejun Heo 3100e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3101e2dca7adSTejun Heo { 3102e2dca7adSTejun Heo struct workqueue_struct *wq = 3103e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3104e2dca7adSTejun Heo 3105e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3106e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3107e2dca7adSTejun Heo else 3108e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3109e2dca7adSTejun Heo 3110e2dca7adSTejun Heo kfree(wq->rescuer); 3111e2dca7adSTejun Heo kfree(wq); 3112e2dca7adSTejun Heo } 3113e2dca7adSTejun Heo 311429c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 311529c91e99STejun Heo { 311629c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 311729c91e99STejun Heo 31187cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 311929c91e99STejun Heo free_workqueue_attrs(pool->attrs); 312029c91e99STejun Heo kfree(pool); 312129c91e99STejun Heo } 312229c91e99STejun Heo 312329c91e99STejun Heo /** 312429c91e99STejun Heo * put_unbound_pool - put a worker_pool 312529c91e99STejun Heo * @pool: worker_pool to put 312629c91e99STejun Heo * 312729c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3128c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3129c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3130c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3131a892caccSTejun Heo * 3132a892caccSTejun Heo * Should be called with wq_pool_mutex held. 313329c91e99STejun Heo */ 313429c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 313529c91e99STejun Heo { 313660f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 313729c91e99STejun Heo struct worker *worker; 313829c91e99STejun Heo 3139a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3140a892caccSTejun Heo 3141a892caccSTejun Heo if (--pool->refcnt) 314229c91e99STejun Heo return; 314329c91e99STejun Heo 314429c91e99STejun Heo /* sanity checks */ 314561d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3146a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 314729c91e99STejun Heo return; 314829c91e99STejun Heo 314929c91e99STejun Heo /* release id and unhash */ 315029c91e99STejun Heo if (pool->id >= 0) 315129c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 315229c91e99STejun Heo hash_del(&pool->hash_node); 315329c91e99STejun Heo 3154c5aa87bbSTejun Heo /* 3155c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3156c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 315792f9c5c4SLai Jiangshan * attach_mutex. 3158c5aa87bbSTejun Heo */ 315929c91e99STejun Heo mutex_lock(&pool->manager_arb); 316029c91e99STejun Heo 316160f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 31621037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 316329c91e99STejun Heo destroy_worker(worker); 316429c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 316529c91e99STejun Heo spin_unlock_irq(&pool->lock); 316660f5a4bcSLai Jiangshan 316792f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3168da028469SLai Jiangshan if (!list_empty(&pool->workers)) 316960f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 317092f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 317160f5a4bcSLai Jiangshan 317260f5a4bcSLai Jiangshan if (pool->detach_completion) 317360f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 317460f5a4bcSLai Jiangshan 317529c91e99STejun Heo mutex_unlock(&pool->manager_arb); 317629c91e99STejun Heo 317729c91e99STejun Heo /* shut down the timers */ 317829c91e99STejun Heo del_timer_sync(&pool->idle_timer); 317929c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 318029c91e99STejun Heo 318129c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 318229c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 318329c91e99STejun Heo } 318429c91e99STejun Heo 318529c91e99STejun Heo /** 318629c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 318729c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 318829c91e99STejun Heo * 318929c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 319029c91e99STejun Heo * reference count and return it. If there already is a matching 319129c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3192d185af30SYacine Belkadi * create a new one. 3193a892caccSTejun Heo * 3194a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3195d185af30SYacine Belkadi * 3196d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3197d185af30SYacine Belkadi * On failure, %NULL. 319829c91e99STejun Heo */ 319929c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 320029c91e99STejun Heo { 320129c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 320229c91e99STejun Heo struct worker_pool *pool; 3203f3f90ad4STejun Heo int node; 320429c91e99STejun Heo 3205a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 320629c91e99STejun Heo 320729c91e99STejun Heo /* do we already have a matching pool? */ 320829c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 320929c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 321029c91e99STejun Heo pool->refcnt++; 32113fb1823cSLai Jiangshan return pool; 321229c91e99STejun Heo } 321329c91e99STejun Heo } 321429c91e99STejun Heo 321529c91e99STejun Heo /* nope, create a new one */ 321629c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 321729c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 321829c91e99STejun Heo goto fail; 321929c91e99STejun Heo 32208864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 322129c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 322229c91e99STejun Heo 32232865a8fbSShaohua Li /* 32242865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 32252865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 32262865a8fbSShaohua Li */ 32272865a8fbSShaohua Li pool->attrs->no_numa = false; 32282865a8fbSShaohua Li 3229f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3230f3f90ad4STejun Heo if (wq_numa_enabled) { 3231f3f90ad4STejun Heo for_each_node(node) { 3232f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3233f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3234f3f90ad4STejun Heo pool->node = node; 3235f3f90ad4STejun Heo break; 3236f3f90ad4STejun Heo } 3237f3f90ad4STejun Heo } 3238f3f90ad4STejun Heo } 3239f3f90ad4STejun Heo 324029c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 324129c91e99STejun Heo goto fail; 324229c91e99STejun Heo 324329c91e99STejun Heo /* create and start the initial worker */ 3244051e1850SLai Jiangshan if (!create_worker(pool)) 324529c91e99STejun Heo goto fail; 324629c91e99STejun Heo 324729c91e99STejun Heo /* install */ 324829c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 32493fb1823cSLai Jiangshan 325029c91e99STejun Heo return pool; 325129c91e99STejun Heo fail: 325229c91e99STejun Heo if (pool) 325329c91e99STejun Heo put_unbound_pool(pool); 325429c91e99STejun Heo return NULL; 325529c91e99STejun Heo } 325629c91e99STejun Heo 32578864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 32588864b4e5STejun Heo { 32598864b4e5STejun Heo kmem_cache_free(pwq_cache, 32608864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 32618864b4e5STejun Heo } 32628864b4e5STejun Heo 32638864b4e5STejun Heo /* 32648864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 32658864b4e5STejun Heo * and needs to be destroyed. 32668864b4e5STejun Heo */ 32678864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 32688864b4e5STejun Heo { 32698864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 32708864b4e5STejun Heo unbound_release_work); 32718864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 32728864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3273bc0caf09STejun Heo bool is_last; 32748864b4e5STejun Heo 32758864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 32768864b4e5STejun Heo return; 32778864b4e5STejun Heo 32783c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 32798864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3280bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 32813c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32828864b4e5STejun Heo 3283a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 32848864b4e5STejun Heo put_unbound_pool(pool); 3285a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3286a892caccSTejun Heo 32878864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 32888864b4e5STejun Heo 32898864b4e5STejun Heo /* 32908864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3291e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 32928864b4e5STejun Heo */ 3293e2dca7adSTejun Heo if (is_last) 3294e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 32956029a918STejun Heo } 32968864b4e5STejun Heo 32970fbd95aaSTejun Heo /** 3298699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 32990fbd95aaSTejun Heo * @pwq: target pool_workqueue 33000fbd95aaSTejun Heo * 3301699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3302699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3303699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 33040fbd95aaSTejun Heo */ 3305699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 33060fbd95aaSTejun Heo { 3307699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3308699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3309699ce097STejun Heo 3310699ce097STejun Heo /* for @wq->saved_max_active */ 3311a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3312699ce097STejun Heo 3313699ce097STejun Heo /* fast exit for non-freezable wqs */ 3314699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3315699ce097STejun Heo return; 3316699ce097STejun Heo 3317a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3318699ce097STejun Heo 331974b414eaSLai Jiangshan /* 332074b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 332174b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 332274b414eaSLai Jiangshan * is updated and visible. 332374b414eaSLai Jiangshan */ 332474b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3325699ce097STejun Heo pwq->max_active = wq->saved_max_active; 33260fbd95aaSTejun Heo 33270fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 33280fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 33290fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3330951a078aSLai Jiangshan 3331951a078aSLai Jiangshan /* 3332951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3333951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3334951a078aSLai Jiangshan */ 3335951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3336699ce097STejun Heo } else { 3337699ce097STejun Heo pwq->max_active = 0; 3338699ce097STejun Heo } 3339699ce097STejun Heo 3340a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 33410fbd95aaSTejun Heo } 33420fbd95aaSTejun Heo 3343e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3344f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3345f147f29eSTejun Heo struct worker_pool *pool) 3346d2c1d404STejun Heo { 3347d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3348d2c1d404STejun Heo 3349e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3350e50aba9aSTejun Heo 3351d2c1d404STejun Heo pwq->pool = pool; 3352d2c1d404STejun Heo pwq->wq = wq; 3353d2c1d404STejun Heo pwq->flush_color = -1; 33548864b4e5STejun Heo pwq->refcnt = 1; 3355d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 33561befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3357d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 33588864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3359f147f29eSTejun Heo } 3360d2c1d404STejun Heo 3361f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 33621befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3363f147f29eSTejun Heo { 3364f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3365f147f29eSTejun Heo 3366f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 336775ccf595STejun Heo 33681befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 33691befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 33701befcf30STejun Heo return; 33711befcf30STejun Heo 337229b1cb41SLai Jiangshan /* set the matching work_color */ 337375ccf595STejun Heo pwq->work_color = wq->work_color; 3374983ca25eSTejun Heo 3375983ca25eSTejun Heo /* sync max_active to the current setting */ 3376983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3377983ca25eSTejun Heo 3378983ca25eSTejun Heo /* link in @pwq */ 33799e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3380df2d5ae4STejun Heo } 33816029a918STejun Heo 3382f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3383f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3384f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3385f147f29eSTejun Heo { 3386f147f29eSTejun Heo struct worker_pool *pool; 3387f147f29eSTejun Heo struct pool_workqueue *pwq; 3388f147f29eSTejun Heo 3389f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3390f147f29eSTejun Heo 3391f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3392f147f29eSTejun Heo if (!pool) 3393f147f29eSTejun Heo return NULL; 3394f147f29eSTejun Heo 3395e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3396f147f29eSTejun Heo if (!pwq) { 3397f147f29eSTejun Heo put_unbound_pool(pool); 3398f147f29eSTejun Heo return NULL; 3399f147f29eSTejun Heo } 3400f147f29eSTejun Heo 3401f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3402f147f29eSTejun Heo return pwq; 3403d2c1d404STejun Heo } 3404d2c1d404STejun Heo 34054c16bd32STejun Heo /** 340630186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3407042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 34084c16bd32STejun Heo * @node: the target NUMA node 34094c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 34104c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 34114c16bd32STejun Heo * 34124c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 34134c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3414d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 34154c16bd32STejun Heo * 34164c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 34174c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 34184c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 34194c16bd32STejun Heo * @attrs->cpumask. 34204c16bd32STejun Heo * 34214c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 34224c16bd32STejun Heo * stable. 3423d185af30SYacine Belkadi * 3424d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3425d185af30SYacine Belkadi * %false if equal. 34264c16bd32STejun Heo */ 34274c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 34284c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 34294c16bd32STejun Heo { 3430d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 34314c16bd32STejun Heo goto use_dfl; 34324c16bd32STejun Heo 34334c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 34344c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 34354c16bd32STejun Heo if (cpu_going_down >= 0) 34364c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 34374c16bd32STejun Heo 34384c16bd32STejun Heo if (cpumask_empty(cpumask)) 34394c16bd32STejun Heo goto use_dfl; 34404c16bd32STejun Heo 34414c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 34424c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 34434c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 34444c16bd32STejun Heo 34454c16bd32STejun Heo use_dfl: 34464c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 34474c16bd32STejun Heo return false; 34484c16bd32STejun Heo } 34494c16bd32STejun Heo 34501befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 34511befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 34521befcf30STejun Heo int node, 34531befcf30STejun Heo struct pool_workqueue *pwq) 34541befcf30STejun Heo { 34551befcf30STejun Heo struct pool_workqueue *old_pwq; 34561befcf30STejun Heo 34575b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 34581befcf30STejun Heo lockdep_assert_held(&wq->mutex); 34591befcf30STejun Heo 34601befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 34611befcf30STejun Heo link_pwq(pwq); 34621befcf30STejun Heo 34631befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 34641befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 34651befcf30STejun Heo return old_pwq; 34661befcf30STejun Heo } 34671befcf30STejun Heo 34682d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 34692d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 34702d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 34712d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3472042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 34732d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 34742d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 34752d5f0764SLai Jiangshan }; 34762d5f0764SLai Jiangshan 34772d5f0764SLai Jiangshan /* free the resources after success or abort */ 34782d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 34792d5f0764SLai Jiangshan { 34802d5f0764SLai Jiangshan if (ctx) { 34812d5f0764SLai Jiangshan int node; 34822d5f0764SLai Jiangshan 34832d5f0764SLai Jiangshan for_each_node(node) 34842d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 34852d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 34862d5f0764SLai Jiangshan 34872d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 34882d5f0764SLai Jiangshan 34892d5f0764SLai Jiangshan kfree(ctx); 34902d5f0764SLai Jiangshan } 34912d5f0764SLai Jiangshan } 34922d5f0764SLai Jiangshan 34932d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 34942d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 34952d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 34962d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 34972d5f0764SLai Jiangshan { 34982d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 34992d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 35002d5f0764SLai Jiangshan int node; 35012d5f0764SLai Jiangshan 35022d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 35032d5f0764SLai Jiangshan 35042d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 35052d5f0764SLai Jiangshan GFP_KERNEL); 35062d5f0764SLai Jiangshan 35072d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 35082d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 35092d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 35102d5f0764SLai Jiangshan goto out_free; 35112d5f0764SLai Jiangshan 3512042f7df1SLai Jiangshan /* 3513042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3514042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3515042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3516042f7df1SLai Jiangshan */ 35172d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3518b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3519042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3520042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 35212d5f0764SLai Jiangshan 35222d5f0764SLai Jiangshan /* 35232d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 35242d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 35252d5f0764SLai Jiangshan * pools. 35262d5f0764SLai Jiangshan */ 35272d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 35282d5f0764SLai Jiangshan 35292d5f0764SLai Jiangshan /* 35302d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 35312d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 35322d5f0764SLai Jiangshan * it even if we don't use it immediately. 35332d5f0764SLai Jiangshan */ 35342d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 35352d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 35362d5f0764SLai Jiangshan goto out_free; 35372d5f0764SLai Jiangshan 35382d5f0764SLai Jiangshan for_each_node(node) { 3539042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 35402d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 35412d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 35422d5f0764SLai Jiangshan goto out_free; 35432d5f0764SLai Jiangshan } else { 35442d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 35452d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 35462d5f0764SLai Jiangshan } 35472d5f0764SLai Jiangshan } 35482d5f0764SLai Jiangshan 3549042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3550042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3551042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 35522d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3553042f7df1SLai Jiangshan 35542d5f0764SLai Jiangshan ctx->wq = wq; 35552d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 35562d5f0764SLai Jiangshan return ctx; 35572d5f0764SLai Jiangshan 35582d5f0764SLai Jiangshan out_free: 35592d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 35602d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 35612d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 35622d5f0764SLai Jiangshan return NULL; 35632d5f0764SLai Jiangshan } 35642d5f0764SLai Jiangshan 35652d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 35662d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 35672d5f0764SLai Jiangshan { 35682d5f0764SLai Jiangshan int node; 35692d5f0764SLai Jiangshan 35702d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 35712d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 35722d5f0764SLai Jiangshan 35732d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 35742d5f0764SLai Jiangshan 35752d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 35762d5f0764SLai Jiangshan for_each_node(node) 35772d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 35782d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 35792d5f0764SLai Jiangshan 35802d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 35812d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 35822d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 35832d5f0764SLai Jiangshan 35842d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 35852d5f0764SLai Jiangshan } 35862d5f0764SLai Jiangshan 3587a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3588a0111cf6SLai Jiangshan { 3589a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3590a0111cf6SLai Jiangshan get_online_cpus(); 3591a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3592a0111cf6SLai Jiangshan } 3593a0111cf6SLai Jiangshan 3594a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3595a0111cf6SLai Jiangshan { 3596a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3597a0111cf6SLai Jiangshan put_online_cpus(); 3598a0111cf6SLai Jiangshan } 3599a0111cf6SLai Jiangshan 3600a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3601a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3602a0111cf6SLai Jiangshan { 3603a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3604a0111cf6SLai Jiangshan int ret = -ENOMEM; 3605a0111cf6SLai Jiangshan 3606a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3607a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3608a0111cf6SLai Jiangshan return -EINVAL; 3609a0111cf6SLai Jiangshan 3610a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 3611a0111cf6SLai Jiangshan if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 3612a0111cf6SLai Jiangshan return -EINVAL; 3613a0111cf6SLai Jiangshan 3614a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 3615a0111cf6SLai Jiangshan 3616a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3617a0111cf6SLai Jiangshan if (ctx) { 3618a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3619a0111cf6SLai Jiangshan ret = 0; 3620a0111cf6SLai Jiangshan } 3621a0111cf6SLai Jiangshan 3622a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3623a0111cf6SLai Jiangshan 3624a0111cf6SLai Jiangshan return ret; 3625a0111cf6SLai Jiangshan } 3626a0111cf6SLai Jiangshan 36279e8cd2f5STejun Heo /** 36289e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36299e8cd2f5STejun Heo * @wq: the target workqueue 36309e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36319e8cd2f5STejun Heo * 36324c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 36334c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 36344c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 36354c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 36364c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 36374c16bd32STejun Heo * back-to-back will stay on its current pwq. 36389e8cd2f5STejun Heo * 3639d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3640d185af30SYacine Belkadi * 3641d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 36429e8cd2f5STejun Heo */ 36439e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36449e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36459e8cd2f5STejun Heo { 3646a0111cf6SLai Jiangshan int ret; 36479e8cd2f5STejun Heo 3648a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3649a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3650a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 36512d5f0764SLai Jiangshan 36522d5f0764SLai Jiangshan return ret; 36539e8cd2f5STejun Heo } 36549e8cd2f5STejun Heo 36554c16bd32STejun Heo /** 36564c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 36574c16bd32STejun Heo * @wq: the target workqueue 36584c16bd32STejun Heo * @cpu: the CPU coming up or going down 36594c16bd32STejun Heo * @online: whether @cpu is coming up or going down 36604c16bd32STejun Heo * 36614c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 36624c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 36634c16bd32STejun Heo * @wq accordingly. 36644c16bd32STejun Heo * 36654c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 36664c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 36674c16bd32STejun Heo * correct. 36684c16bd32STejun Heo * 36694c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 36704c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 36714c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 36724c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 36734c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 36744c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 36754c16bd32STejun Heo * CPU_DOWN_PREPARE. 36764c16bd32STejun Heo */ 36774c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 36784c16bd32STejun Heo bool online) 36794c16bd32STejun Heo { 36804c16bd32STejun Heo int node = cpu_to_node(cpu); 36814c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 36824c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 36834c16bd32STejun Heo struct workqueue_attrs *target_attrs; 36844c16bd32STejun Heo cpumask_t *cpumask; 36854c16bd32STejun Heo 36864c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 36874c16bd32STejun Heo 3688f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3689f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 36904c16bd32STejun Heo return; 36914c16bd32STejun Heo 36924c16bd32STejun Heo /* 36934c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 36944c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 36954c16bd32STejun Heo * CPU hotplug exclusion. 36964c16bd32STejun Heo */ 36974c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 36984c16bd32STejun Heo cpumask = target_attrs->cpumask; 36994c16bd32STejun Heo 37004c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 37014c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 37024c16bd32STejun Heo 37034c16bd32STejun Heo /* 37044c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3705042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3706042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3707042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 37084c16bd32STejun Heo */ 3709042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 37104c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3711f7142ed4SLai Jiangshan return; 37124c16bd32STejun Heo } else { 37134c16bd32STejun Heo goto use_dfl_pwq; 37144c16bd32STejun Heo } 37154c16bd32STejun Heo 37164c16bd32STejun Heo /* create a new pwq */ 37174c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 37184c16bd32STejun Heo if (!pwq) { 37192d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 37204c16bd32STejun Heo wq->name); 372177f300b1SDaeseok Youn goto use_dfl_pwq; 37224c16bd32STejun Heo } 37234c16bd32STejun Heo 3724f7142ed4SLai Jiangshan /* Install the new pwq. */ 37254c16bd32STejun Heo mutex_lock(&wq->mutex); 37264c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 37274c16bd32STejun Heo goto out_unlock; 37284c16bd32STejun Heo 37294c16bd32STejun Heo use_dfl_pwq: 3730f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 37314c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 37324c16bd32STejun Heo get_pwq(wq->dfl_pwq); 37334c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 37344c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 37354c16bd32STejun Heo out_unlock: 37364c16bd32STejun Heo mutex_unlock(&wq->mutex); 37374c16bd32STejun Heo put_pwq_unlocked(old_pwq); 37384c16bd32STejun Heo } 37394c16bd32STejun Heo 374030cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37411da177e4SLinus Torvalds { 374249e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 37438a2b7538STejun Heo int cpu, ret; 3744e1d8aa9fSFrederic Weisbecker 374530cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3746420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3747420c0ddbSTejun Heo if (!wq->cpu_pwqs) 374830cdf249STejun Heo return -ENOMEM; 374930cdf249STejun Heo 375030cdf249STejun Heo for_each_possible_cpu(cpu) { 37517fb98ea7STejun Heo struct pool_workqueue *pwq = 37527fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37537a62c2c8STejun Heo struct worker_pool *cpu_pools = 3754f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 375530cdf249STejun Heo 3756f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3757f147f29eSTejun Heo 3758f147f29eSTejun Heo mutex_lock(&wq->mutex); 37591befcf30STejun Heo link_pwq(pwq); 3760f147f29eSTejun Heo mutex_unlock(&wq->mutex); 376130cdf249STejun Heo } 376230cdf249STejun Heo return 0; 37638a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 37648a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 37658a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 37668a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 37678a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 37688a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 37698a2b7538STejun Heo return ret; 37709e8cd2f5STejun Heo } else { 37719e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 37729e8cd2f5STejun Heo } 37730f900049STejun Heo } 37740f900049STejun Heo 3775f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3776f3421797STejun Heo const char *name) 3777b71ab8c2STejun Heo { 3778f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3779f3421797STejun Heo 3780f3421797STejun Heo if (max_active < 1 || max_active > lim) 3781044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3782f3421797STejun Heo max_active, name, 1, lim); 3783b71ab8c2STejun Heo 3784f3421797STejun Heo return clamp_val(max_active, 1, lim); 3785b71ab8c2STejun Heo } 3786b71ab8c2STejun Heo 3787b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 378897e37d7bSTejun Heo unsigned int flags, 37891e19ffc6STejun Heo int max_active, 3790eb13ba87SJohannes Berg struct lock_class_key *key, 3791b196be89STejun Heo const char *lock_name, ...) 37923af24433SOleg Nesterov { 3793df2d5ae4STejun Heo size_t tbl_size = 0; 3794ecf6881fSTejun Heo va_list args; 37953af24433SOleg Nesterov struct workqueue_struct *wq; 379649e3cf44STejun Heo struct pool_workqueue *pwq; 3797b196be89STejun Heo 3798cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 3799cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 3800cee22a15SViresh Kumar flags |= WQ_UNBOUND; 3801cee22a15SViresh Kumar 3802ecf6881fSTejun Heo /* allocate wq and format name */ 3803df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 3804ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 3805df2d5ae4STejun Heo 3806df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 3807b196be89STejun Heo if (!wq) 3808d2c1d404STejun Heo return NULL; 3809b196be89STejun Heo 38106029a918STejun Heo if (flags & WQ_UNBOUND) { 38116029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38126029a918STejun Heo if (!wq->unbound_attrs) 38136029a918STejun Heo goto err_free_wq; 38146029a918STejun Heo } 38156029a918STejun Heo 3816ecf6881fSTejun Heo va_start(args, lock_name); 3817ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3818b196be89STejun Heo va_end(args); 38193af24433SOleg Nesterov 3820d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3821b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 38223af24433SOleg Nesterov 3823b196be89STejun Heo /* init wq */ 382497e37d7bSTejun Heo wq->flags = flags; 3825a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 38263c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3827112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 382830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 382973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 383073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3831493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 38323af24433SOleg Nesterov 3833eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3834cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 38353af24433SOleg Nesterov 383630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3837d2c1d404STejun Heo goto err_free_wq; 38381537663fSTejun Heo 3839493008a8STejun Heo /* 3840493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3841493008a8STejun Heo * have a rescuer to guarantee forward progress. 3842493008a8STejun Heo */ 3843493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3844e22bee78STejun Heo struct worker *rescuer; 3845e22bee78STejun Heo 3846f7537df5SLai Jiangshan rescuer = alloc_worker(NUMA_NO_NODE); 3847e22bee78STejun Heo if (!rescuer) 3848d2c1d404STejun Heo goto err_destroy; 3849e22bee78STejun Heo 3850111c225aSTejun Heo rescuer->rescue_wq = wq; 3851111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3852b196be89STejun Heo wq->name); 3853d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3854d2c1d404STejun Heo kfree(rescuer); 3855d2c1d404STejun Heo goto err_destroy; 3856d2c1d404STejun Heo } 3857e22bee78STejun Heo 3858d2c1d404STejun Heo wq->rescuer = rescuer; 385914a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3860e22bee78STejun Heo wake_up_process(rescuer->task); 38613af24433SOleg Nesterov } 38621537663fSTejun Heo 3863226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3864226223abSTejun Heo goto err_destroy; 3865226223abSTejun Heo 38663af24433SOleg Nesterov /* 386768e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 386868e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 386968e13a67SLai Jiangshan * list. 38703af24433SOleg Nesterov */ 387168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3872a0a1a5fdSTejun Heo 3873a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 387449e3cf44STejun Heo for_each_pwq(pwq, wq) 3875699ce097STejun Heo pwq_adjust_max_active(pwq); 3876a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3877a0a1a5fdSTejun Heo 3878e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 3879a0a1a5fdSTejun Heo 388068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38813af24433SOleg Nesterov 38823af24433SOleg Nesterov return wq; 3883d2c1d404STejun Heo 3884d2c1d404STejun Heo err_free_wq: 38856029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 38864690c4abSTejun Heo kfree(wq); 3887d2c1d404STejun Heo return NULL; 3888d2c1d404STejun Heo err_destroy: 3889d2c1d404STejun Heo destroy_workqueue(wq); 38904690c4abSTejun Heo return NULL; 38911da177e4SLinus Torvalds } 3892d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 38931da177e4SLinus Torvalds 38943af24433SOleg Nesterov /** 38953af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 38963af24433SOleg Nesterov * @wq: target workqueue 38973af24433SOleg Nesterov * 38983af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 38993af24433SOleg Nesterov */ 39003af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 39013af24433SOleg Nesterov { 390249e3cf44STejun Heo struct pool_workqueue *pwq; 39034c16bd32STejun Heo int node; 39043af24433SOleg Nesterov 39059c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 39069c5a2ba7STejun Heo drain_workqueue(wq); 3907c8efcc25STejun Heo 39086183c009STejun Heo /* sanity checks */ 3909b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 391049e3cf44STejun Heo for_each_pwq(pwq, wq) { 39116183c009STejun Heo int i; 39126183c009STejun Heo 391376af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 391476af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3915b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39166183c009STejun Heo return; 391776af4d93STejun Heo } 391876af4d93STejun Heo } 391976af4d93STejun Heo 39205c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 39218864b4e5STejun Heo WARN_ON(pwq->nr_active) || 392276af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3923b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39246183c009STejun Heo return; 39256183c009STejun Heo } 392676af4d93STejun Heo } 3927b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39286183c009STejun Heo 3929a0a1a5fdSTejun Heo /* 3930a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3931a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3932a0a1a5fdSTejun Heo */ 393368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3934e2dca7adSTejun Heo list_del_rcu(&wq->list); 393568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 39363af24433SOleg Nesterov 3937226223abSTejun Heo workqueue_sysfs_unregister(wq); 3938226223abSTejun Heo 3939e2dca7adSTejun Heo if (wq->rescuer) 3940e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3941e22bee78STejun Heo 39428864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 394329c91e99STejun Heo /* 39448864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 3945e2dca7adSTejun Heo * schedule RCU free. 394629c91e99STejun Heo */ 3947e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 39488864b4e5STejun Heo } else { 39498864b4e5STejun Heo /* 39508864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 39514c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 39524c16bd32STejun Heo * @wq will be freed when the last pwq is released. 39538864b4e5STejun Heo */ 39544c16bd32STejun Heo for_each_node(node) { 39554c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 39564c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 39574c16bd32STejun Heo put_pwq_unlocked(pwq); 39584c16bd32STejun Heo } 39594c16bd32STejun Heo 39604c16bd32STejun Heo /* 39614c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 39624c16bd32STejun Heo * put. Don't access it afterwards. 39634c16bd32STejun Heo */ 39644c16bd32STejun Heo pwq = wq->dfl_pwq; 39654c16bd32STejun Heo wq->dfl_pwq = NULL; 3966dce90d47STejun Heo put_pwq_unlocked(pwq); 396729c91e99STejun Heo } 39683af24433SOleg Nesterov } 39693af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 39703af24433SOleg Nesterov 3971dcd989cbSTejun Heo /** 3972dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3973dcd989cbSTejun Heo * @wq: target workqueue 3974dcd989cbSTejun Heo * @max_active: new max_active value. 3975dcd989cbSTejun Heo * 3976dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3977dcd989cbSTejun Heo * 3978dcd989cbSTejun Heo * CONTEXT: 3979dcd989cbSTejun Heo * Don't call from IRQ context. 3980dcd989cbSTejun Heo */ 3981dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3982dcd989cbSTejun Heo { 398349e3cf44STejun Heo struct pool_workqueue *pwq; 3984dcd989cbSTejun Heo 39858719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 39868719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 39878719dceaSTejun Heo return; 39888719dceaSTejun Heo 3989f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3990dcd989cbSTejun Heo 3991a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 3992dcd989cbSTejun Heo 3993dcd989cbSTejun Heo wq->saved_max_active = max_active; 3994dcd989cbSTejun Heo 3995699ce097STejun Heo for_each_pwq(pwq, wq) 3996699ce097STejun Heo pwq_adjust_max_active(pwq); 3997dcd989cbSTejun Heo 3998a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3999dcd989cbSTejun Heo } 4000dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4001dcd989cbSTejun Heo 4002dcd989cbSTejun Heo /** 4003e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4004e6267616STejun Heo * 4005e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4006e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4007d185af30SYacine Belkadi * 4008d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4009e6267616STejun Heo */ 4010e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4011e6267616STejun Heo { 4012e6267616STejun Heo struct worker *worker = current_wq_worker(); 4013e6267616STejun Heo 40146a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4015e6267616STejun Heo } 4016e6267616STejun Heo 4017e6267616STejun Heo /** 4018dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4019dcd989cbSTejun Heo * @cpu: CPU in question 4020dcd989cbSTejun Heo * @wq: target workqueue 4021dcd989cbSTejun Heo * 4022dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4023dcd989cbSTejun Heo * no synchronization around this function and the test result is 4024dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4025dcd989cbSTejun Heo * 4026d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4027d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4028d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4029d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4030d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4031d3251859STejun Heo * 4032d185af30SYacine Belkadi * Return: 4033dcd989cbSTejun Heo * %true if congested, %false otherwise. 4034dcd989cbSTejun Heo */ 4035d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4036dcd989cbSTejun Heo { 40377fb98ea7STejun Heo struct pool_workqueue *pwq; 403876af4d93STejun Heo bool ret; 403976af4d93STejun Heo 404088109453SLai Jiangshan rcu_read_lock_sched(); 40417fb98ea7STejun Heo 4042d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4043d3251859STejun Heo cpu = smp_processor_id(); 4044d3251859STejun Heo 40457fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 40467fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 40477fb98ea7STejun Heo else 4048df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4049dcd989cbSTejun Heo 405076af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 405188109453SLai Jiangshan rcu_read_unlock_sched(); 405276af4d93STejun Heo 405376af4d93STejun Heo return ret; 4054dcd989cbSTejun Heo } 4055dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4056dcd989cbSTejun Heo 4057dcd989cbSTejun Heo /** 4058dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4059dcd989cbSTejun Heo * @work: the work to be tested 4060dcd989cbSTejun Heo * 4061dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4062dcd989cbSTejun Heo * synchronization around this function and the test result is 4063dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4064dcd989cbSTejun Heo * 4065d185af30SYacine Belkadi * Return: 4066dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4067dcd989cbSTejun Heo */ 4068dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4069dcd989cbSTejun Heo { 4070fa1b54e6STejun Heo struct worker_pool *pool; 4071dcd989cbSTejun Heo unsigned long flags; 4072dcd989cbSTejun Heo unsigned int ret = 0; 4073dcd989cbSTejun Heo 4074dcd989cbSTejun Heo if (work_pending(work)) 4075dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4076038366c5SLai Jiangshan 4077fa1b54e6STejun Heo local_irq_save(flags); 4078fa1b54e6STejun Heo pool = get_work_pool(work); 4079038366c5SLai Jiangshan if (pool) { 4080fa1b54e6STejun Heo spin_lock(&pool->lock); 4081c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4082dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4083fa1b54e6STejun Heo spin_unlock(&pool->lock); 4084038366c5SLai Jiangshan } 4085fa1b54e6STejun Heo local_irq_restore(flags); 4086dcd989cbSTejun Heo 4087dcd989cbSTejun Heo return ret; 4088dcd989cbSTejun Heo } 4089dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4090dcd989cbSTejun Heo 40913d1cb205STejun Heo /** 40923d1cb205STejun Heo * set_worker_desc - set description for the current work item 40933d1cb205STejun Heo * @fmt: printf-style format string 40943d1cb205STejun Heo * @...: arguments for the format string 40953d1cb205STejun Heo * 40963d1cb205STejun Heo * This function can be called by a running work function to describe what 40973d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 40983d1cb205STejun Heo * information will be printed out together to help debugging. The 40993d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 41003d1cb205STejun Heo */ 41013d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 41023d1cb205STejun Heo { 41033d1cb205STejun Heo struct worker *worker = current_wq_worker(); 41043d1cb205STejun Heo va_list args; 41053d1cb205STejun Heo 41063d1cb205STejun Heo if (worker) { 41073d1cb205STejun Heo va_start(args, fmt); 41083d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 41093d1cb205STejun Heo va_end(args); 41103d1cb205STejun Heo worker->desc_valid = true; 41113d1cb205STejun Heo } 41123d1cb205STejun Heo } 41133d1cb205STejun Heo 41143d1cb205STejun Heo /** 41153d1cb205STejun Heo * print_worker_info - print out worker information and description 41163d1cb205STejun Heo * @log_lvl: the log level to use when printing 41173d1cb205STejun Heo * @task: target task 41183d1cb205STejun Heo * 41193d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 41203d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 41213d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 41223d1cb205STejun Heo * 41233d1cb205STejun Heo * This function can be safely called on any task as long as the 41243d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 41253d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 41263d1cb205STejun Heo */ 41273d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 41283d1cb205STejun Heo { 41293d1cb205STejun Heo work_func_t *fn = NULL; 41303d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 41313d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 41323d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 41333d1cb205STejun Heo struct workqueue_struct *wq = NULL; 41343d1cb205STejun Heo bool desc_valid = false; 41353d1cb205STejun Heo struct worker *worker; 41363d1cb205STejun Heo 41373d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 41383d1cb205STejun Heo return; 41393d1cb205STejun Heo 41403d1cb205STejun Heo /* 41413d1cb205STejun Heo * This function is called without any synchronization and @task 41423d1cb205STejun Heo * could be in any state. Be careful with dereferences. 41433d1cb205STejun Heo */ 41443d1cb205STejun Heo worker = probe_kthread_data(task); 41453d1cb205STejun Heo 41463d1cb205STejun Heo /* 41473d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 41483d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 41493d1cb205STejun Heo */ 41503d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 41513d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 41523d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 41533d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 41543d1cb205STejun Heo 41553d1cb205STejun Heo /* copy worker description */ 41563d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 41573d1cb205STejun Heo if (desc_valid) 41583d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 41593d1cb205STejun Heo 41603d1cb205STejun Heo if (fn || name[0] || desc[0]) { 41613d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 41623d1cb205STejun Heo if (desc[0]) 41633d1cb205STejun Heo pr_cont(" (%s)", desc); 41643d1cb205STejun Heo pr_cont("\n"); 41653d1cb205STejun Heo } 41663d1cb205STejun Heo } 41673d1cb205STejun Heo 41683494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 41693494fc30STejun Heo { 41703494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 41713494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 41723494fc30STejun Heo pr_cont(" node=%d", pool->node); 41733494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 41743494fc30STejun Heo } 41753494fc30STejun Heo 41763494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 41773494fc30STejun Heo { 41783494fc30STejun Heo if (work->func == wq_barrier_func) { 41793494fc30STejun Heo struct wq_barrier *barr; 41803494fc30STejun Heo 41813494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 41823494fc30STejun Heo 41833494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 41843494fc30STejun Heo task_pid_nr(barr->task)); 41853494fc30STejun Heo } else { 41863494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 41873494fc30STejun Heo } 41883494fc30STejun Heo } 41893494fc30STejun Heo 41903494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 41913494fc30STejun Heo { 41923494fc30STejun Heo struct worker_pool *pool = pwq->pool; 41933494fc30STejun Heo struct work_struct *work; 41943494fc30STejun Heo struct worker *worker; 41953494fc30STejun Heo bool has_in_flight = false, has_pending = false; 41963494fc30STejun Heo int bkt; 41973494fc30STejun Heo 41983494fc30STejun Heo pr_info(" pwq %d:", pool->id); 41993494fc30STejun Heo pr_cont_pool_info(pool); 42003494fc30STejun Heo 42013494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 42023494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 42033494fc30STejun Heo 42043494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 42053494fc30STejun Heo if (worker->current_pwq == pwq) { 42063494fc30STejun Heo has_in_flight = true; 42073494fc30STejun Heo break; 42083494fc30STejun Heo } 42093494fc30STejun Heo } 42103494fc30STejun Heo if (has_in_flight) { 42113494fc30STejun Heo bool comma = false; 42123494fc30STejun Heo 42133494fc30STejun Heo pr_info(" in-flight:"); 42143494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 42153494fc30STejun Heo if (worker->current_pwq != pwq) 42163494fc30STejun Heo continue; 42173494fc30STejun Heo 42183494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 42193494fc30STejun Heo task_pid_nr(worker->task), 42203494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 42213494fc30STejun Heo worker->current_func); 42223494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 42233494fc30STejun Heo pr_cont_work(false, work); 42243494fc30STejun Heo comma = true; 42253494fc30STejun Heo } 42263494fc30STejun Heo pr_cont("\n"); 42273494fc30STejun Heo } 42283494fc30STejun Heo 42293494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 42303494fc30STejun Heo if (get_work_pwq(work) == pwq) { 42313494fc30STejun Heo has_pending = true; 42323494fc30STejun Heo break; 42333494fc30STejun Heo } 42343494fc30STejun Heo } 42353494fc30STejun Heo if (has_pending) { 42363494fc30STejun Heo bool comma = false; 42373494fc30STejun Heo 42383494fc30STejun Heo pr_info(" pending:"); 42393494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 42403494fc30STejun Heo if (get_work_pwq(work) != pwq) 42413494fc30STejun Heo continue; 42423494fc30STejun Heo 42433494fc30STejun Heo pr_cont_work(comma, work); 42443494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 42453494fc30STejun Heo } 42463494fc30STejun Heo pr_cont("\n"); 42473494fc30STejun Heo } 42483494fc30STejun Heo 42493494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 42503494fc30STejun Heo bool comma = false; 42513494fc30STejun Heo 42523494fc30STejun Heo pr_info(" delayed:"); 42533494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 42543494fc30STejun Heo pr_cont_work(comma, work); 42553494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 42563494fc30STejun Heo } 42573494fc30STejun Heo pr_cont("\n"); 42583494fc30STejun Heo } 42593494fc30STejun Heo } 42603494fc30STejun Heo 42613494fc30STejun Heo /** 42623494fc30STejun Heo * show_workqueue_state - dump workqueue state 42633494fc30STejun Heo * 42643494fc30STejun Heo * Called from a sysrq handler and prints out all busy workqueues and 42653494fc30STejun Heo * pools. 42663494fc30STejun Heo */ 42673494fc30STejun Heo void show_workqueue_state(void) 42683494fc30STejun Heo { 42693494fc30STejun Heo struct workqueue_struct *wq; 42703494fc30STejun Heo struct worker_pool *pool; 42713494fc30STejun Heo unsigned long flags; 42723494fc30STejun Heo int pi; 42733494fc30STejun Heo 42743494fc30STejun Heo rcu_read_lock_sched(); 42753494fc30STejun Heo 42763494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 42773494fc30STejun Heo 42783494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 42793494fc30STejun Heo struct pool_workqueue *pwq; 42803494fc30STejun Heo bool idle = true; 42813494fc30STejun Heo 42823494fc30STejun Heo for_each_pwq(pwq, wq) { 42833494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 42843494fc30STejun Heo idle = false; 42853494fc30STejun Heo break; 42863494fc30STejun Heo } 42873494fc30STejun Heo } 42883494fc30STejun Heo if (idle) 42893494fc30STejun Heo continue; 42903494fc30STejun Heo 42913494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 42923494fc30STejun Heo 42933494fc30STejun Heo for_each_pwq(pwq, wq) { 42943494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 42953494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 42963494fc30STejun Heo show_pwq(pwq); 42973494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 42983494fc30STejun Heo } 42993494fc30STejun Heo } 43003494fc30STejun Heo 43013494fc30STejun Heo for_each_pool(pool, pi) { 43023494fc30STejun Heo struct worker *worker; 43033494fc30STejun Heo bool first = true; 43043494fc30STejun Heo 43053494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 43063494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 43073494fc30STejun Heo goto next_pool; 43083494fc30STejun Heo 43093494fc30STejun Heo pr_info("pool %d:", pool->id); 43103494fc30STejun Heo pr_cont_pool_info(pool); 43113494fc30STejun Heo pr_cont(" workers=%d", pool->nr_workers); 43123494fc30STejun Heo if (pool->manager) 43133494fc30STejun Heo pr_cont(" manager: %d", 43143494fc30STejun Heo task_pid_nr(pool->manager->task)); 43153494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 43163494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 43173494fc30STejun Heo task_pid_nr(worker->task)); 43183494fc30STejun Heo first = false; 43193494fc30STejun Heo } 43203494fc30STejun Heo pr_cont("\n"); 43213494fc30STejun Heo next_pool: 43223494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 43233494fc30STejun Heo } 43243494fc30STejun Heo 43253494fc30STejun Heo rcu_read_unlock_sched(); 43263494fc30STejun Heo } 43273494fc30STejun Heo 4328db7bccf4STejun Heo /* 4329db7bccf4STejun Heo * CPU hotplug. 4330db7bccf4STejun Heo * 4331e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4332112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4333706026c2STejun Heo * pool which make migrating pending and scheduled works very 4334e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 433594cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4336e22bee78STejun Heo * blocked draining impractical. 4337e22bee78STejun Heo * 433824647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4339628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4340628c78e7STejun Heo * cpu comes back online. 4341db7bccf4STejun Heo */ 4342db7bccf4STejun Heo 4343706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4344db7bccf4STejun Heo { 434538db41d9STejun Heo int cpu = smp_processor_id(); 43464ce62e9eSTejun Heo struct worker_pool *pool; 4347db7bccf4STejun Heo struct worker *worker; 4348db7bccf4STejun Heo 4349f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 435092f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 435194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4352e22bee78STejun Heo 4353f2d5a0eeSTejun Heo /* 435492f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 435594cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 435694cf58bbSTejun Heo * except for the ones which are still executing works from 435794cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 435894cf58bbSTejun Heo * this, they may become diasporas. 4359f2d5a0eeSTejun Heo */ 4360da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4361403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4362db7bccf4STejun Heo 436324647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4364f2d5a0eeSTejun Heo 436594cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 436692f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4367e22bee78STejun Heo 4368e22bee78STejun Heo /* 4369eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4370eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4371eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4372eb283428SLai Jiangshan * from other cpus. 4373628c78e7STejun Heo */ 4374628c78e7STejun Heo schedule(); 4375628c78e7STejun Heo 4376628c78e7STejun Heo /* 4377eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4378eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4379eb283428SLai Jiangshan * and keep_working() are always true as long as the 4380eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4381eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4382eb283428SLai Jiangshan * are served by workers tied to the pool. 4383e22bee78STejun Heo */ 4384e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4385eb283428SLai Jiangshan 4386eb283428SLai Jiangshan /* 4387eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4388eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4389eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4390eb283428SLai Jiangshan */ 4391eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4392eb283428SLai Jiangshan wake_up_worker(pool); 4393eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4394eb283428SLai Jiangshan } 4395db7bccf4STejun Heo } 4396db7bccf4STejun Heo 4397bd7c089eSTejun Heo /** 4398bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4399bd7c089eSTejun Heo * @pool: pool of interest 4400bd7c089eSTejun Heo * 4401a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4402bd7c089eSTejun Heo */ 4403bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4404bd7c089eSTejun Heo { 4405a9ab775bSTejun Heo struct worker *worker; 4406bd7c089eSTejun Heo 440792f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4408bd7c089eSTejun Heo 4409bd7c089eSTejun Heo /* 4410a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4411a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4412402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4413a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4414a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4415bd7c089eSTejun Heo */ 4416da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4417a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4418a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4419a9ab775bSTejun Heo 4420a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 44213de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4422a9ab775bSTejun Heo 4423da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4424a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4425a9ab775bSTejun Heo 4426a9ab775bSTejun Heo /* 4427a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4428a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4429a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4430a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4431a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4432a9ab775bSTejun Heo * be bound before @pool->lock is released. 4433a9ab775bSTejun Heo */ 4434a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4435bd7c089eSTejun Heo wake_up_process(worker->task); 4436bd7c089eSTejun Heo 4437bd7c089eSTejun Heo /* 4438a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4439a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4440a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4441a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4442a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4443a9ab775bSTejun Heo * concurrency management. Note that when or whether 4444a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4445a9ab775bSTejun Heo * 4446a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4447a9ab775bSTejun Heo * tested without holding any lock in 4448a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4449a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4450a9ab775bSTejun Heo * management operations. 4451bd7c089eSTejun Heo */ 4452a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4453a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4454a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4455a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4456bd7c089eSTejun Heo } 4457a9ab775bSTejun Heo 4458a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4459bd7c089eSTejun Heo } 4460bd7c089eSTejun Heo 44617dbc725eSTejun Heo /** 44627dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 44637dbc725eSTejun Heo * @pool: unbound pool of interest 44647dbc725eSTejun Heo * @cpu: the CPU which is coming up 44657dbc725eSTejun Heo * 44667dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 44677dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 44687dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 44697dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 44707dbc725eSTejun Heo */ 44717dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 44727dbc725eSTejun Heo { 44737dbc725eSTejun Heo static cpumask_t cpumask; 44747dbc725eSTejun Heo struct worker *worker; 44757dbc725eSTejun Heo 447692f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 44777dbc725eSTejun Heo 44787dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 44797dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 44807dbc725eSTejun Heo return; 44817dbc725eSTejun Heo 44827dbc725eSTejun Heo /* is @cpu the only online CPU? */ 44837dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 44847dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 44857dbc725eSTejun Heo return; 44867dbc725eSTejun Heo 44877dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4488da028469SLai Jiangshan for_each_pool_worker(worker, pool) 44897dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 44907dbc725eSTejun Heo pool->attrs->cpumask) < 0); 44917dbc725eSTejun Heo } 44927dbc725eSTejun Heo 44938db25e78STejun Heo /* 44948db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 44958db25e78STejun Heo * This will be registered high priority CPU notifier. 44968db25e78STejun Heo */ 44970db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 44981da177e4SLinus Torvalds unsigned long action, 44991da177e4SLinus Torvalds void *hcpu) 45001da177e4SLinus Torvalds { 4501d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45024ce62e9eSTejun Heo struct worker_pool *pool; 45034c16bd32STejun Heo struct workqueue_struct *wq; 45047dbc725eSTejun Heo int pi; 45051da177e4SLinus Torvalds 45068db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 45073af24433SOleg Nesterov case CPU_UP_PREPARE: 4508f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45093ce63377STejun Heo if (pool->nr_workers) 45103ce63377STejun Heo continue; 4511051e1850SLai Jiangshan if (!create_worker(pool)) 45123ce63377STejun Heo return NOTIFY_BAD; 45133af24433SOleg Nesterov } 45141da177e4SLinus Torvalds break; 45151da177e4SLinus Torvalds 451665758202STejun Heo case CPU_DOWN_FAILED: 451765758202STejun Heo case CPU_ONLINE: 451868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45197dbc725eSTejun Heo 45207dbc725eSTejun Heo for_each_pool(pool, pi) { 452192f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 452294cf58bbSTejun Heo 4523f05b558dSLai Jiangshan if (pool->cpu == cpu) 452494cf58bbSTejun Heo rebind_workers(pool); 4525f05b558dSLai Jiangshan else if (pool->cpu < 0) 45267dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 452794cf58bbSTejun Heo 452892f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 452994cf58bbSTejun Heo } 45307dbc725eSTejun Heo 45314c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45324c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45334c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 45344c16bd32STejun Heo 453568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45368db25e78STejun Heo break; 453765758202STejun Heo } 453865758202STejun Heo return NOTIFY_OK; 453965758202STejun Heo } 454065758202STejun Heo 454165758202STejun Heo /* 454265758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 454365758202STejun Heo * This will be registered as low priority CPU notifier. 454465758202STejun Heo */ 45450db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 454665758202STejun Heo unsigned long action, 454765758202STejun Heo void *hcpu) 454865758202STejun Heo { 4549d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45508db25e78STejun Heo struct work_struct unbind_work; 45514c16bd32STejun Heo struct workqueue_struct *wq; 45528db25e78STejun Heo 455365758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 455465758202STejun Heo case CPU_DOWN_PREPARE: 45554c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4556706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 45577635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 45584c16bd32STejun Heo 45594c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45604c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 45614c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45624c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 45634c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 45644c16bd32STejun Heo 45654c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 45668db25e78STejun Heo flush_work(&unbind_work); 4567440a1136SChuansheng Liu destroy_work_on_stack(&unbind_work); 45688db25e78STejun Heo break; 456965758202STejun Heo } 457065758202STejun Heo return NOTIFY_OK; 457165758202STejun Heo } 457265758202STejun Heo 45732d3854a3SRusty Russell #ifdef CONFIG_SMP 45748ccad40dSRusty Russell 45752d3854a3SRusty Russell struct work_for_cpu { 4576ed48ece2STejun Heo struct work_struct work; 45772d3854a3SRusty Russell long (*fn)(void *); 45782d3854a3SRusty Russell void *arg; 45792d3854a3SRusty Russell long ret; 45802d3854a3SRusty Russell }; 45812d3854a3SRusty Russell 4582ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 45832d3854a3SRusty Russell { 4584ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4585ed48ece2STejun Heo 45862d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 45872d3854a3SRusty Russell } 45882d3854a3SRusty Russell 45892d3854a3SRusty Russell /** 45902d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 45912d3854a3SRusty Russell * @cpu: the cpu to run on 45922d3854a3SRusty Russell * @fn: the function to run 45932d3854a3SRusty Russell * @arg: the function arg 45942d3854a3SRusty Russell * 459531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 45966b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4597d185af30SYacine Belkadi * 4598d185af30SYacine Belkadi * Return: The value @fn returns. 45992d3854a3SRusty Russell */ 4600d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 46012d3854a3SRusty Russell { 4602ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 46032d3854a3SRusty Russell 4604ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4605ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 460612997d1aSBjorn Helgaas flush_work(&wfc.work); 4607440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 46082d3854a3SRusty Russell return wfc.ret; 46092d3854a3SRusty Russell } 46102d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 46112d3854a3SRusty Russell #endif /* CONFIG_SMP */ 46122d3854a3SRusty Russell 4613a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4614e7577c50SRusty Russell 4615a0a1a5fdSTejun Heo /** 4616a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4617a0a1a5fdSTejun Heo * 461858a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4619c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4620706026c2STejun Heo * pool->worklist. 4621a0a1a5fdSTejun Heo * 4622a0a1a5fdSTejun Heo * CONTEXT: 4623a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4624a0a1a5fdSTejun Heo */ 4625a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4626a0a1a5fdSTejun Heo { 462724b8a847STejun Heo struct workqueue_struct *wq; 462824b8a847STejun Heo struct pool_workqueue *pwq; 4629a0a1a5fdSTejun Heo 463068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4631a0a1a5fdSTejun Heo 46326183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4633a0a1a5fdSTejun Heo workqueue_freezing = true; 4634a0a1a5fdSTejun Heo 463524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4636a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4637699ce097STejun Heo for_each_pwq(pwq, wq) 4638699ce097STejun Heo pwq_adjust_max_active(pwq); 4639a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4640a1056305STejun Heo } 46415bcab335STejun Heo 464268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4643a0a1a5fdSTejun Heo } 4644a0a1a5fdSTejun Heo 4645a0a1a5fdSTejun Heo /** 464658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4647a0a1a5fdSTejun Heo * 4648a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4649a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4650a0a1a5fdSTejun Heo * 4651a0a1a5fdSTejun Heo * CONTEXT: 465268e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4653a0a1a5fdSTejun Heo * 4654d185af30SYacine Belkadi * Return: 465558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 465658a69cb4STejun Heo * is complete. 4657a0a1a5fdSTejun Heo */ 4658a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4659a0a1a5fdSTejun Heo { 4660a0a1a5fdSTejun Heo bool busy = false; 466124b8a847STejun Heo struct workqueue_struct *wq; 466224b8a847STejun Heo struct pool_workqueue *pwq; 4663a0a1a5fdSTejun Heo 466468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4665a0a1a5fdSTejun Heo 46666183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4667a0a1a5fdSTejun Heo 466824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 466924b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 467024b8a847STejun Heo continue; 4671a0a1a5fdSTejun Heo /* 4672a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4673a0a1a5fdSTejun Heo * to peek without lock. 4674a0a1a5fdSTejun Heo */ 467588109453SLai Jiangshan rcu_read_lock_sched(); 467624b8a847STejun Heo for_each_pwq(pwq, wq) { 46776183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4678112202d9STejun Heo if (pwq->nr_active) { 4679a0a1a5fdSTejun Heo busy = true; 468088109453SLai Jiangshan rcu_read_unlock_sched(); 4681a0a1a5fdSTejun Heo goto out_unlock; 4682a0a1a5fdSTejun Heo } 4683a0a1a5fdSTejun Heo } 468488109453SLai Jiangshan rcu_read_unlock_sched(); 4685a0a1a5fdSTejun Heo } 4686a0a1a5fdSTejun Heo out_unlock: 468768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4688a0a1a5fdSTejun Heo return busy; 4689a0a1a5fdSTejun Heo } 4690a0a1a5fdSTejun Heo 4691a0a1a5fdSTejun Heo /** 4692a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4693a0a1a5fdSTejun Heo * 4694a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4695706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4696a0a1a5fdSTejun Heo * 4697a0a1a5fdSTejun Heo * CONTEXT: 4698a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4699a0a1a5fdSTejun Heo */ 4700a0a1a5fdSTejun Heo void thaw_workqueues(void) 4701a0a1a5fdSTejun Heo { 470224b8a847STejun Heo struct workqueue_struct *wq; 470324b8a847STejun Heo struct pool_workqueue *pwq; 4704a0a1a5fdSTejun Heo 470568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4706a0a1a5fdSTejun Heo 4707a0a1a5fdSTejun Heo if (!workqueue_freezing) 4708a0a1a5fdSTejun Heo goto out_unlock; 4709a0a1a5fdSTejun Heo 471074b414eaSLai Jiangshan workqueue_freezing = false; 471124b8a847STejun Heo 471224b8a847STejun Heo /* restore max_active and repopulate worklist */ 471324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4714a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4715699ce097STejun Heo for_each_pwq(pwq, wq) 4716699ce097STejun Heo pwq_adjust_max_active(pwq); 4717a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 471824b8a847STejun Heo } 471924b8a847STejun Heo 4720a0a1a5fdSTejun Heo out_unlock: 472168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4722a0a1a5fdSTejun Heo } 4723a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4724a0a1a5fdSTejun Heo 4725042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4726042f7df1SLai Jiangshan { 4727042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4728042f7df1SLai Jiangshan int ret = 0; 4729042f7df1SLai Jiangshan struct workqueue_struct *wq; 4730042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4731042f7df1SLai Jiangshan 4732042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4733042f7df1SLai Jiangshan 4734042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4735042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4736042f7df1SLai Jiangshan continue; 4737042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 4738042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 4739042f7df1SLai Jiangshan continue; 4740042f7df1SLai Jiangshan 4741042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 4742042f7df1SLai Jiangshan if (!ctx) { 4743042f7df1SLai Jiangshan ret = -ENOMEM; 4744042f7df1SLai Jiangshan break; 4745042f7df1SLai Jiangshan } 4746042f7df1SLai Jiangshan 4747042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 4748042f7df1SLai Jiangshan } 4749042f7df1SLai Jiangshan 4750042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 4751042f7df1SLai Jiangshan if (!ret) 4752042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 4753042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 4754042f7df1SLai Jiangshan } 4755042f7df1SLai Jiangshan 4756042f7df1SLai Jiangshan return ret; 4757042f7df1SLai Jiangshan } 4758042f7df1SLai Jiangshan 4759042f7df1SLai Jiangshan /** 4760042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 4761042f7df1SLai Jiangshan * @cpumask: the cpumask to set 4762042f7df1SLai Jiangshan * 4763042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 4764042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 4765042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 4766042f7df1SLai Jiangshan * 4767042f7df1SLai Jiangshan * Retun: 0 - Success 4768042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 4769042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 4770042f7df1SLai Jiangshan */ 4771042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 4772042f7df1SLai Jiangshan { 4773042f7df1SLai Jiangshan int ret = -EINVAL; 4774042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 4775042f7df1SLai Jiangshan 4776042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 4777042f7df1SLai Jiangshan return -ENOMEM; 4778042f7df1SLai Jiangshan 4779042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 4780042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 4781a0111cf6SLai Jiangshan apply_wqattrs_lock(); 4782042f7df1SLai Jiangshan 4783042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 4784042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 4785042f7df1SLai Jiangshan 4786042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 4787042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 4788042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 4789042f7df1SLai Jiangshan 4790042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 4791042f7df1SLai Jiangshan if (ret < 0) 4792042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 4793042f7df1SLai Jiangshan 4794a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 4795042f7df1SLai Jiangshan } 4796042f7df1SLai Jiangshan 4797042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 4798042f7df1SLai Jiangshan return ret; 4799042f7df1SLai Jiangshan } 4800042f7df1SLai Jiangshan 48016ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 48026ba94429SFrederic Weisbecker /* 48036ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 48046ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 48056ba94429SFrederic Weisbecker * following attributes. 48066ba94429SFrederic Weisbecker * 48076ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 48086ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 48096ba94429SFrederic Weisbecker * 48106ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 48116ba94429SFrederic Weisbecker * 48126ba94429SFrederic Weisbecker * id RO int : the associated pool ID 48136ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 48146ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 48156ba94429SFrederic Weisbecker */ 48166ba94429SFrederic Weisbecker struct wq_device { 48176ba94429SFrederic Weisbecker struct workqueue_struct *wq; 48186ba94429SFrederic Weisbecker struct device dev; 48196ba94429SFrederic Weisbecker }; 48206ba94429SFrederic Weisbecker 48216ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 48226ba94429SFrederic Weisbecker { 48236ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 48246ba94429SFrederic Weisbecker 48256ba94429SFrederic Weisbecker return wq_dev->wq; 48266ba94429SFrederic Weisbecker } 48276ba94429SFrederic Weisbecker 48286ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 48296ba94429SFrederic Weisbecker char *buf) 48306ba94429SFrederic Weisbecker { 48316ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48326ba94429SFrederic Weisbecker 48336ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 48346ba94429SFrederic Weisbecker } 48356ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 48366ba94429SFrederic Weisbecker 48376ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 48386ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 48396ba94429SFrederic Weisbecker { 48406ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48416ba94429SFrederic Weisbecker 48426ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 48436ba94429SFrederic Weisbecker } 48446ba94429SFrederic Weisbecker 48456ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 48466ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 48476ba94429SFrederic Weisbecker size_t count) 48486ba94429SFrederic Weisbecker { 48496ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48506ba94429SFrederic Weisbecker int val; 48516ba94429SFrederic Weisbecker 48526ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 48536ba94429SFrederic Weisbecker return -EINVAL; 48546ba94429SFrederic Weisbecker 48556ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 48566ba94429SFrederic Weisbecker return count; 48576ba94429SFrederic Weisbecker } 48586ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 48596ba94429SFrederic Weisbecker 48606ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 48616ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 48626ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 48636ba94429SFrederic Weisbecker NULL, 48646ba94429SFrederic Weisbecker }; 48656ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 48666ba94429SFrederic Weisbecker 48676ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 48686ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 48696ba94429SFrederic Weisbecker { 48706ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48716ba94429SFrederic Weisbecker const char *delim = ""; 48726ba94429SFrederic Weisbecker int node, written = 0; 48736ba94429SFrederic Weisbecker 48746ba94429SFrederic Weisbecker rcu_read_lock_sched(); 48756ba94429SFrederic Weisbecker for_each_node(node) { 48766ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 48776ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 48786ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 48796ba94429SFrederic Weisbecker delim = " "; 48806ba94429SFrederic Weisbecker } 48816ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 48826ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 48836ba94429SFrederic Weisbecker 48846ba94429SFrederic Weisbecker return written; 48856ba94429SFrederic Weisbecker } 48866ba94429SFrederic Weisbecker 48876ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 48886ba94429SFrederic Weisbecker char *buf) 48896ba94429SFrederic Weisbecker { 48906ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 48916ba94429SFrederic Weisbecker int written; 48926ba94429SFrederic Weisbecker 48936ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 48946ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 48956ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 48966ba94429SFrederic Weisbecker 48976ba94429SFrederic Weisbecker return written; 48986ba94429SFrederic Weisbecker } 48996ba94429SFrederic Weisbecker 49006ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 49016ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 49026ba94429SFrederic Weisbecker { 49036ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 49046ba94429SFrederic Weisbecker 4905899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4906899a94feSLai Jiangshan 49076ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 49086ba94429SFrederic Weisbecker if (!attrs) 49096ba94429SFrederic Weisbecker return NULL; 49106ba94429SFrederic Weisbecker 49116ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 49126ba94429SFrederic Weisbecker return attrs; 49136ba94429SFrederic Weisbecker } 49146ba94429SFrederic Weisbecker 49156ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 49166ba94429SFrederic Weisbecker const char *buf, size_t count) 49176ba94429SFrederic Weisbecker { 49186ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49196ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 4920d4d3e257SLai Jiangshan int ret = -ENOMEM; 4921d4d3e257SLai Jiangshan 4922d4d3e257SLai Jiangshan apply_wqattrs_lock(); 49236ba94429SFrederic Weisbecker 49246ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 49256ba94429SFrederic Weisbecker if (!attrs) 4926d4d3e257SLai Jiangshan goto out_unlock; 49276ba94429SFrederic Weisbecker 49286ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 49296ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 4930d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 49316ba94429SFrederic Weisbecker else 49326ba94429SFrederic Weisbecker ret = -EINVAL; 49336ba94429SFrederic Weisbecker 4934d4d3e257SLai Jiangshan out_unlock: 4935d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 49366ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 49376ba94429SFrederic Weisbecker return ret ?: count; 49386ba94429SFrederic Weisbecker } 49396ba94429SFrederic Weisbecker 49406ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 49416ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 49426ba94429SFrederic Weisbecker { 49436ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49446ba94429SFrederic Weisbecker int written; 49456ba94429SFrederic Weisbecker 49466ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 49476ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 49486ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 49496ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 49506ba94429SFrederic Weisbecker return written; 49516ba94429SFrederic Weisbecker } 49526ba94429SFrederic Weisbecker 49536ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 49546ba94429SFrederic Weisbecker struct device_attribute *attr, 49556ba94429SFrederic Weisbecker const char *buf, size_t count) 49566ba94429SFrederic Weisbecker { 49576ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49586ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 4959d4d3e257SLai Jiangshan int ret = -ENOMEM; 4960d4d3e257SLai Jiangshan 4961d4d3e257SLai Jiangshan apply_wqattrs_lock(); 49626ba94429SFrederic Weisbecker 49636ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 49646ba94429SFrederic Weisbecker if (!attrs) 4965d4d3e257SLai Jiangshan goto out_unlock; 49666ba94429SFrederic Weisbecker 49676ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 49686ba94429SFrederic Weisbecker if (!ret) 4969d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 49706ba94429SFrederic Weisbecker 4971d4d3e257SLai Jiangshan out_unlock: 4972d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 49736ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 49746ba94429SFrederic Weisbecker return ret ?: count; 49756ba94429SFrederic Weisbecker } 49766ba94429SFrederic Weisbecker 49776ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 49786ba94429SFrederic Weisbecker char *buf) 49796ba94429SFrederic Weisbecker { 49806ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49816ba94429SFrederic Weisbecker int written; 49826ba94429SFrederic Weisbecker 49836ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 49846ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 49856ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 49866ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 49876ba94429SFrederic Weisbecker 49886ba94429SFrederic Weisbecker return written; 49896ba94429SFrederic Weisbecker } 49906ba94429SFrederic Weisbecker 49916ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 49926ba94429SFrederic Weisbecker const char *buf, size_t count) 49936ba94429SFrederic Weisbecker { 49946ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 49956ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 4996d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 4997d4d3e257SLai Jiangshan 4998d4d3e257SLai Jiangshan apply_wqattrs_lock(); 49996ba94429SFrederic Weisbecker 50006ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 50016ba94429SFrederic Weisbecker if (!attrs) 5002d4d3e257SLai Jiangshan goto out_unlock; 50036ba94429SFrederic Weisbecker 50046ba94429SFrederic Weisbecker ret = -EINVAL; 50056ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 50066ba94429SFrederic Weisbecker attrs->no_numa = !v; 5007d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 50086ba94429SFrederic Weisbecker } 50096ba94429SFrederic Weisbecker 5010d4d3e257SLai Jiangshan out_unlock: 5011d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 50126ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 50136ba94429SFrederic Weisbecker return ret ?: count; 50146ba94429SFrederic Weisbecker } 50156ba94429SFrederic Weisbecker 50166ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 50176ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 50186ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 50196ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 50206ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 50216ba94429SFrederic Weisbecker __ATTR_NULL, 50226ba94429SFrederic Weisbecker }; 50236ba94429SFrederic Weisbecker 50246ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 50256ba94429SFrederic Weisbecker .name = "workqueue", 50266ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 50276ba94429SFrederic Weisbecker }; 50286ba94429SFrederic Weisbecker 5029b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5030b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5031b05a7928SFrederic Weisbecker { 5032b05a7928SFrederic Weisbecker int written; 5033b05a7928SFrederic Weisbecker 5034042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5035b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5036b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5037042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5038b05a7928SFrederic Weisbecker 5039b05a7928SFrederic Weisbecker return written; 5040b05a7928SFrederic Weisbecker } 5041b05a7928SFrederic Weisbecker 5042042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5043042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5044042f7df1SLai Jiangshan { 5045042f7df1SLai Jiangshan cpumask_var_t cpumask; 5046042f7df1SLai Jiangshan int ret; 5047042f7df1SLai Jiangshan 5048042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5049042f7df1SLai Jiangshan return -ENOMEM; 5050042f7df1SLai Jiangshan 5051042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5052042f7df1SLai Jiangshan if (!ret) 5053042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5054042f7df1SLai Jiangshan 5055042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5056042f7df1SLai Jiangshan return ret ? ret : count; 5057042f7df1SLai Jiangshan } 5058042f7df1SLai Jiangshan 5059b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5060042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5061042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5062b05a7928SFrederic Weisbecker 50636ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 50646ba94429SFrederic Weisbecker { 5065b05a7928SFrederic Weisbecker int err; 5066b05a7928SFrederic Weisbecker 5067b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5068b05a7928SFrederic Weisbecker if (err) 5069b05a7928SFrederic Weisbecker return err; 5070b05a7928SFrederic Weisbecker 5071b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 50726ba94429SFrederic Weisbecker } 50736ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 50746ba94429SFrederic Weisbecker 50756ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 50766ba94429SFrederic Weisbecker { 50776ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 50786ba94429SFrederic Weisbecker 50796ba94429SFrederic Weisbecker kfree(wq_dev); 50806ba94429SFrederic Weisbecker } 50816ba94429SFrederic Weisbecker 50826ba94429SFrederic Weisbecker /** 50836ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 50846ba94429SFrederic Weisbecker * @wq: the workqueue to register 50856ba94429SFrederic Weisbecker * 50866ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 50876ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 50886ba94429SFrederic Weisbecker * which is the preferred method. 50896ba94429SFrederic Weisbecker * 50906ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 50916ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 50926ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 50936ba94429SFrederic Weisbecker * attributes. 50946ba94429SFrederic Weisbecker * 50956ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 50966ba94429SFrederic Weisbecker */ 50976ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 50986ba94429SFrederic Weisbecker { 50996ba94429SFrederic Weisbecker struct wq_device *wq_dev; 51006ba94429SFrederic Weisbecker int ret; 51016ba94429SFrederic Weisbecker 51026ba94429SFrederic Weisbecker /* 5103402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 51046ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 51056ba94429SFrederic Weisbecker * workqueues. 51066ba94429SFrederic Weisbecker */ 51076ba94429SFrederic Weisbecker if (WARN_ON(wq->flags & __WQ_ORDERED)) 51086ba94429SFrederic Weisbecker return -EINVAL; 51096ba94429SFrederic Weisbecker 51106ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 51116ba94429SFrederic Weisbecker if (!wq_dev) 51126ba94429SFrederic Weisbecker return -ENOMEM; 51136ba94429SFrederic Weisbecker 51146ba94429SFrederic Weisbecker wq_dev->wq = wq; 51156ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 51166ba94429SFrederic Weisbecker wq_dev->dev.init_name = wq->name; 51176ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 51186ba94429SFrederic Weisbecker 51196ba94429SFrederic Weisbecker /* 51206ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 51216ba94429SFrederic Weisbecker * everything is ready. 51226ba94429SFrederic Weisbecker */ 51236ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 51246ba94429SFrederic Weisbecker 51256ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 51266ba94429SFrederic Weisbecker if (ret) { 51276ba94429SFrederic Weisbecker kfree(wq_dev); 51286ba94429SFrederic Weisbecker wq->wq_dev = NULL; 51296ba94429SFrederic Weisbecker return ret; 51306ba94429SFrederic Weisbecker } 51316ba94429SFrederic Weisbecker 51326ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 51336ba94429SFrederic Weisbecker struct device_attribute *attr; 51346ba94429SFrederic Weisbecker 51356ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 51366ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 51376ba94429SFrederic Weisbecker if (ret) { 51386ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 51396ba94429SFrederic Weisbecker wq->wq_dev = NULL; 51406ba94429SFrederic Weisbecker return ret; 51416ba94429SFrederic Weisbecker } 51426ba94429SFrederic Weisbecker } 51436ba94429SFrederic Weisbecker } 51446ba94429SFrederic Weisbecker 51456ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 51466ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 51476ba94429SFrederic Weisbecker return 0; 51486ba94429SFrederic Weisbecker } 51496ba94429SFrederic Weisbecker 51506ba94429SFrederic Weisbecker /** 51516ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 51526ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 51536ba94429SFrederic Weisbecker * 51546ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 51556ba94429SFrederic Weisbecker */ 51566ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 51576ba94429SFrederic Weisbecker { 51586ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 51596ba94429SFrederic Weisbecker 51606ba94429SFrederic Weisbecker if (!wq->wq_dev) 51616ba94429SFrederic Weisbecker return; 51626ba94429SFrederic Weisbecker 51636ba94429SFrederic Weisbecker wq->wq_dev = NULL; 51646ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 51656ba94429SFrederic Weisbecker } 51666ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 51676ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 51686ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 51696ba94429SFrederic Weisbecker 5170bce90380STejun Heo static void __init wq_numa_init(void) 5171bce90380STejun Heo { 5172bce90380STejun Heo cpumask_var_t *tbl; 5173bce90380STejun Heo int node, cpu; 5174bce90380STejun Heo 5175bce90380STejun Heo if (num_possible_nodes() <= 1) 5176bce90380STejun Heo return; 5177bce90380STejun Heo 5178d55262c4STejun Heo if (wq_disable_numa) { 5179d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5180d55262c4STejun Heo return; 5181d55262c4STejun Heo } 5182d55262c4STejun Heo 51834c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 51844c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 51854c16bd32STejun Heo 5186bce90380STejun Heo /* 5187bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5188bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5189bce90380STejun Heo * fully initialized by now. 5190bce90380STejun Heo */ 5191ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5192bce90380STejun Heo BUG_ON(!tbl); 5193bce90380STejun Heo 5194bce90380STejun Heo for_each_node(node) 51955a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 51961be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5197bce90380STejun Heo 5198bce90380STejun Heo for_each_possible_cpu(cpu) { 5199bce90380STejun Heo node = cpu_to_node(cpu); 5200bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5201bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5202bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5203bce90380STejun Heo return; 5204bce90380STejun Heo } 5205bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5206bce90380STejun Heo } 5207bce90380STejun Heo 5208bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5209bce90380STejun Heo wq_numa_enabled = true; 5210bce90380STejun Heo } 5211bce90380STejun Heo 52126ee0578bSSuresh Siddha static int __init init_workqueues(void) 52131da177e4SLinus Torvalds { 52147a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 52157a4e344cSTejun Heo int i, cpu; 5216c34056a3STejun Heo 5217e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5218e904e6c2STejun Heo 5219b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5220b05a7928SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 5221b05a7928SFrederic Weisbecker 5222e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5223e904e6c2STejun Heo 522465758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 5225a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 52268b03ae3cSTejun Heo 5227bce90380STejun Heo wq_numa_init(); 5228bce90380STejun Heo 5229706026c2STejun Heo /* initialize CPU pools */ 523029c91e99STejun Heo for_each_possible_cpu(cpu) { 52314ce62e9eSTejun Heo struct worker_pool *pool; 52328b03ae3cSTejun Heo 52337a4e344cSTejun Heo i = 0; 5234f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 52357a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5236ec22ca5eSTejun Heo pool->cpu = cpu; 52377a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 52387a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5239f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 52407a4e344cSTejun Heo 52419daf9e67STejun Heo /* alloc pool ID */ 524268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 52439daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 524468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 52454ce62e9eSTejun Heo } 52468b03ae3cSTejun Heo } 52478b03ae3cSTejun Heo 5248e22bee78STejun Heo /* create the initial worker */ 524929c91e99STejun Heo for_each_online_cpu(cpu) { 52504ce62e9eSTejun Heo struct worker_pool *pool; 5251e22bee78STejun Heo 5252f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 525324647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5254051e1850SLai Jiangshan BUG_ON(!create_worker(pool)); 5255e22bee78STejun Heo } 52564ce62e9eSTejun Heo } 5257e22bee78STejun Heo 52588a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 525929c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 526029c91e99STejun Heo struct workqueue_attrs *attrs; 526129c91e99STejun Heo 526229c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 526329c91e99STejun Heo attrs->nice = std_nice[i]; 526429c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 52658a2b7538STejun Heo 52668a2b7538STejun Heo /* 52678a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 52688a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 52698a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 52708a2b7538STejun Heo */ 52718a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 52728a2b7538STejun Heo attrs->nice = std_nice[i]; 52738a2b7538STejun Heo attrs->no_numa = true; 52748a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 527529c91e99STejun Heo } 527629c91e99STejun Heo 5277d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 52781aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5279d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5280f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5281f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 528224d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 528324d51addSTejun Heo WQ_FREEZABLE, 0); 52840668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 52850668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 52860668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 52870668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 52880668106cSViresh Kumar 0); 52891aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 52900668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 52910668106cSViresh Kumar !system_power_efficient_wq || 52920668106cSViresh Kumar !system_freezable_power_efficient_wq); 52936ee0578bSSuresh Siddha return 0; 52941da177e4SLinus Torvalds } 52956ee0578bSSuresh Siddha early_initcall(init_workqueues); 5296