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 19c54fce6eSTejun Heo * automatically managed. There is one worker pool for each CPU and 20c54fce6eSTejun Heo * one extra for works which are better served by workers which are 21c54fce6eSTejun Heo * not bound to any specific CPU. 22c54fce6eSTejun Heo * 23c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 241da177e4SLinus Torvalds */ 251da177e4SLinus Torvalds 269984de1aSPaul Gortmaker #include <linux/export.h> 271da177e4SLinus Torvalds #include <linux/kernel.h> 281da177e4SLinus Torvalds #include <linux/sched.h> 291da177e4SLinus Torvalds #include <linux/init.h> 301da177e4SLinus Torvalds #include <linux/signal.h> 311da177e4SLinus Torvalds #include <linux/completion.h> 321da177e4SLinus Torvalds #include <linux/workqueue.h> 331da177e4SLinus Torvalds #include <linux/slab.h> 341da177e4SLinus Torvalds #include <linux/cpu.h> 351da177e4SLinus Torvalds #include <linux/notifier.h> 361da177e4SLinus Torvalds #include <linux/kthread.h> 371fa44ecaSJames Bottomley #include <linux/hardirq.h> 3846934023SChristoph Lameter #include <linux/mempolicy.h> 39341a5958SRafael J. Wysocki #include <linux/freezer.h> 40d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4429c91e99STejun Heo #include <linux/jhash.h> 4542f8570fSSasha Levin #include <linux/hashtable.h> 4676af4d93STejun Heo #include <linux/rculist.h> 47e22bee78STejun Heo 48ea138446STejun Heo #include "workqueue_internal.h" 491da177e4SLinus Torvalds 50c8e55f36STejun Heo enum { 51bc2ae0f5STejun Heo /* 5224647570STejun Heo * worker_pool flags 53bc2ae0f5STejun Heo * 5424647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 55bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 56bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 57bc2ae0f5STejun Heo * is in effect. 58bc2ae0f5STejun Heo * 59bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 60bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6124647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 62bc2ae0f5STejun Heo * 63bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 64bc3a1afcSTejun Heo * manager_mutex to avoid changing binding state while 6524647570STejun Heo * create_worker() is in progress. 66bc2ae0f5STejun Heo */ 6711ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 6824647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 6935b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 70db7bccf4STejun Heo 71c8e55f36STejun Heo /* worker flags */ 72c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 73c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 74c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 75e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 76fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 77f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 78e22bee78STejun Heo 795f7dabfdSLai Jiangshan WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND | 80403c821dSTejun Heo WORKER_CPU_INTENSIVE, 81db7bccf4STejun Heo 82e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 834ce62e9eSTejun Heo 8429c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 85c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 86db7bccf4STejun Heo 87e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 88e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 89e22bee78STejun Heo 903233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 913233cdbdSTejun Heo /* call for help after 10ms 923233cdbdSTejun Heo (min two ticks) */ 93e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 94e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 951da177e4SLinus Torvalds 961da177e4SLinus Torvalds /* 97e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 98e22bee78STejun Heo * all cpus. Give -20. 99e22bee78STejun Heo */ 100e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1013270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 102c8e55f36STejun Heo }; 103c8e55f36STejun Heo 1041da177e4SLinus Torvalds /* 1054690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1064690c4abSTejun Heo * 107e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 108e41e704bSTejun Heo * everyone else. 1094690c4abSTejun Heo * 110e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 111e22bee78STejun Heo * only be modified and accessed from the local cpu. 112e22bee78STejun Heo * 113d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1144690c4abSTejun Heo * 115d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 116d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 117d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 118d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 119e22bee78STejun Heo * 12073f53c4aSTejun Heo * F: wq->flush_mutex protected. 12173f53c4aSTejun Heo * 1225bcab335STejun Heo * WQ: wq_mutex protected. 12376af4d93STejun Heo * 1245bcab335STejun Heo * WR: wq_mutex protected for writes. Sched-RCU protected for reads. 1255bcab335STejun Heo * 126794b18bcSTejun Heo * PW: pwq_lock protected. 127794b18bcSTejun Heo * 128794b18bcSTejun Heo * FR: wq->flush_mutex and pwq_lock protected for writes. Sched-RCU 12975ccf595STejun Heo * protected for reads. 1302e109a28STejun Heo * 1312e109a28STejun Heo * MD: wq_mayday_lock protected. 1324690c4abSTejun Heo */ 1334690c4abSTejun Heo 1342eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 135c34056a3STejun Heo 136bd7bdd43STejun Heo struct worker_pool { 137d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 138d84ff051STejun Heo int cpu; /* I: the associated cpu */ 1399daf9e67STejun Heo int id; /* I: pool ID */ 14011ebea50STejun Heo unsigned int flags; /* X: flags */ 141bd7bdd43STejun Heo 142bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 143bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 144ea1abd61SLai Jiangshan 145ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 146bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 147bd7bdd43STejun Heo 148bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 149bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 150bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 151bd7bdd43STejun Heo 152c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 153c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 154c9e7cf27STejun Heo /* L: hash of busy workers */ 155c9e7cf27STejun Heo 156bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 15734a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 158bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 159bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 160e19e397aSTejun Heo 1617a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 1625bcab335STejun Heo struct hlist_node hash_node; /* WQ: unbound_pool_hash node */ 1635bcab335STejun Heo int refcnt; /* WQ: refcnt for unbound pools */ 1647a4e344cSTejun Heo 165e19e397aSTejun Heo /* 166e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 167e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 168e19e397aSTejun Heo * cacheline. 169e19e397aSTejun Heo */ 170e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17129c91e99STejun Heo 17229c91e99STejun Heo /* 17329c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 17429c91e99STejun Heo * from get_work_pool(). 17529c91e99STejun Heo */ 17629c91e99STejun Heo struct rcu_head rcu; 1778b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1788b03ae3cSTejun Heo 1798b03ae3cSTejun Heo /* 180112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 181112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 182112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 183112202d9STejun Heo * number of flag bits. 1841da177e4SLinus Torvalds */ 185112202d9STejun Heo struct pool_workqueue { 186bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1874690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 18873f53c4aSTejun Heo int work_color; /* L: current color */ 18973f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1908864b4e5STejun Heo int refcnt; /* L: reference count */ 19173f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19273f53c4aSTejun Heo /* L: nr of in_flight works */ 1931e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 194a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1951e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 19675ccf595STejun Heo struct list_head pwqs_node; /* FR: node on wq->pwqs */ 1972e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 1988864b4e5STejun Heo 1998864b4e5STejun Heo /* 2008864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2018864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2028864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 203794b18bcSTejun Heo * determined without grabbing pwq_lock. 2048864b4e5STejun Heo */ 2058864b4e5STejun Heo struct work_struct unbound_release_work; 2068864b4e5STejun Heo struct rcu_head rcu; 207e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2081da177e4SLinus Torvalds 2091da177e4SLinus Torvalds /* 21073f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21173f53c4aSTejun Heo */ 21273f53c4aSTejun Heo struct wq_flusher { 21373f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 21473f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 21573f53c4aSTejun Heo struct completion done; /* flush completion */ 21673f53c4aSTejun Heo }; 2171da177e4SLinus Torvalds 218226223abSTejun Heo struct wq_device; 219226223abSTejun Heo 22073f53c4aSTejun Heo /* 221c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 222c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2231da177e4SLinus Torvalds */ 2241da177e4SLinus Torvalds struct workqueue_struct { 2255bcab335STejun Heo unsigned int flags; /* WQ: WQ_* flags */ 226420c0ddbSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */ 22775ccf595STejun Heo struct list_head pwqs; /* FR: all pwqs of this wq */ 2285bcab335STejun Heo struct list_head list; /* WQ: list of all workqueues */ 22973f53c4aSTejun Heo 23073f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 23173f53c4aSTejun Heo int work_color; /* F: current work color */ 23273f53c4aSTejun Heo int flush_color; /* F: current flush color */ 233112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 23473f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 23573f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 23673f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 23773f53c4aSTejun Heo 2382e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 239e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 240e22bee78STejun Heo 2415bcab335STejun Heo int nr_drainers; /* WQ: drain in progress */ 242794b18bcSTejun Heo int saved_max_active; /* PW: saved pwq max_active */ 243226223abSTejun Heo 244226223abSTejun Heo #ifdef CONFIG_SYSFS 245226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 246226223abSTejun Heo #endif 2474e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2484e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2494e6045f1SJohannes Berg #endif 250b196be89STejun Heo char name[]; /* I: workqueue name */ 2511da177e4SLinus Torvalds }; 2521da177e4SLinus Torvalds 253e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 254e904e6c2STejun Heo 2555bcab335STejun Heo static DEFINE_MUTEX(wq_mutex); /* protects workqueues and pools */ 256794b18bcSTejun Heo static DEFINE_SPINLOCK(pwq_lock); /* protects pool_workqueues */ 2572e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2585bcab335STejun Heo 2595bcab335STejun Heo static LIST_HEAD(workqueues); /* WQ: list of all workqueues */ 2605bcab335STejun Heo static bool workqueue_freezing; /* WQ: have wqs started freezing? */ 2617d19c5ceSTejun Heo 2627d19c5ceSTejun Heo /* the per-cpu worker pools */ 2637d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2647d19c5ceSTejun Heo cpu_worker_pools); 2657d19c5ceSTejun Heo 2665bcab335STejun Heo static DEFINE_IDR(worker_pool_idr); /* WR: idr of all pools */ 2677d19c5ceSTejun Heo 2685bcab335STejun Heo /* WQ: hash of all unbound pools keyed by pool->attrs */ 26929c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 27029c91e99STejun Heo 271c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 27229c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 27329c91e99STejun Heo 274d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 275d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 276044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2771aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 278044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 279d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 280044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 281f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 282044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 28324d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 284d320c038STejun Heo 2857d19c5ceSTejun Heo static int worker_thread(void *__worker); 2867d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 2877d19c5ceSTejun Heo const struct workqueue_attrs *from); 2887d19c5ceSTejun Heo 28997bd2347STejun Heo #define CREATE_TRACE_POINTS 29097bd2347STejun Heo #include <trace/events/workqueue.h> 29197bd2347STejun Heo 2925bcab335STejun Heo #define assert_rcu_or_wq_mutex() \ 2935bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 2945bcab335STejun Heo lockdep_is_held(&wq_mutex), \ 2955bcab335STejun Heo "sched RCU or wq_mutex should be held") 2965bcab335STejun Heo 297794b18bcSTejun Heo #define assert_rcu_or_pwq_lock() \ 29876af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 299794b18bcSTejun Heo lockdep_is_held(&pwq_lock), \ 300794b18bcSTejun Heo "sched RCU or pwq_lock should be held") 30176af4d93STejun Heo 302f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 303f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 304f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3057a62c2c8STejun Heo (pool)++) 3064ce62e9eSTejun Heo 307b67bfe0dSSasha Levin #define for_each_busy_worker(worker, i, pool) \ 308b67bfe0dSSasha Levin hash_for_each(pool->busy_hash, i, worker, hentry) 309db7bccf4STejun Heo 31049e3cf44STejun Heo /** 31117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 31217116969STejun Heo * @pool: iteration cursor 313611c92a0STejun Heo * @pi: integer used for iteration 314fa1b54e6STejun Heo * 3155bcab335STejun Heo * This must be called either with wq_mutex held or sched RCU read locked. 3165bcab335STejun Heo * If the pool needs to be used beyond the locking in effect, the caller is 3175bcab335STejun Heo * responsible for guaranteeing that the pool stays online. 318fa1b54e6STejun Heo * 319fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 320fa1b54e6STejun Heo * ignored. 32117116969STejun Heo */ 322611c92a0STejun Heo #define for_each_pool(pool, pi) \ 323611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 3245bcab335STejun Heo if (({ assert_rcu_or_wq_mutex(); false; })) { } \ 325fa1b54e6STejun Heo else 32617116969STejun Heo 32717116969STejun Heo /** 32849e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 32949e3cf44STejun Heo * @pwq: iteration cursor 33049e3cf44STejun Heo * @wq: the target workqueue 33176af4d93STejun Heo * 332794b18bcSTejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 333794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 334794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 33576af4d93STejun Heo * 33676af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 33776af4d93STejun Heo * ignored. 33849e3cf44STejun Heo */ 33949e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 34076af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 341794b18bcSTejun Heo if (({ assert_rcu_or_pwq_lock(); false; })) { } \ 34276af4d93STejun Heo else 343f3421797STejun Heo 344dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 345dc186ad7SThomas Gleixner 346dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 347dc186ad7SThomas Gleixner 34899777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 34999777288SStanislaw Gruszka { 35099777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 35199777288SStanislaw Gruszka } 35299777288SStanislaw Gruszka 353dc186ad7SThomas Gleixner /* 354dc186ad7SThomas Gleixner * fixup_init is called when: 355dc186ad7SThomas Gleixner * - an active object is initialized 356dc186ad7SThomas Gleixner */ 357dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 358dc186ad7SThomas Gleixner { 359dc186ad7SThomas Gleixner struct work_struct *work = addr; 360dc186ad7SThomas Gleixner 361dc186ad7SThomas Gleixner switch (state) { 362dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 363dc186ad7SThomas Gleixner cancel_work_sync(work); 364dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 365dc186ad7SThomas Gleixner return 1; 366dc186ad7SThomas Gleixner default: 367dc186ad7SThomas Gleixner return 0; 368dc186ad7SThomas Gleixner } 369dc186ad7SThomas Gleixner } 370dc186ad7SThomas Gleixner 371dc186ad7SThomas Gleixner /* 372dc186ad7SThomas Gleixner * fixup_activate is called when: 373dc186ad7SThomas Gleixner * - an active object is activated 374dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 375dc186ad7SThomas Gleixner */ 376dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 377dc186ad7SThomas Gleixner { 378dc186ad7SThomas Gleixner struct work_struct *work = addr; 379dc186ad7SThomas Gleixner 380dc186ad7SThomas Gleixner switch (state) { 381dc186ad7SThomas Gleixner 382dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 383dc186ad7SThomas Gleixner /* 384dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 385dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 386dc186ad7SThomas Gleixner * is tracked in the object tracker. 387dc186ad7SThomas Gleixner */ 38822df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 389dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 390dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 391dc186ad7SThomas Gleixner return 0; 392dc186ad7SThomas Gleixner } 393dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 394dc186ad7SThomas Gleixner return 0; 395dc186ad7SThomas Gleixner 396dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 397dc186ad7SThomas Gleixner WARN_ON(1); 398dc186ad7SThomas Gleixner 399dc186ad7SThomas Gleixner default: 400dc186ad7SThomas Gleixner return 0; 401dc186ad7SThomas Gleixner } 402dc186ad7SThomas Gleixner } 403dc186ad7SThomas Gleixner 404dc186ad7SThomas Gleixner /* 405dc186ad7SThomas Gleixner * fixup_free is called when: 406dc186ad7SThomas Gleixner * - an active object is freed 407dc186ad7SThomas Gleixner */ 408dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 409dc186ad7SThomas Gleixner { 410dc186ad7SThomas Gleixner struct work_struct *work = addr; 411dc186ad7SThomas Gleixner 412dc186ad7SThomas Gleixner switch (state) { 413dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 414dc186ad7SThomas Gleixner cancel_work_sync(work); 415dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 416dc186ad7SThomas Gleixner return 1; 417dc186ad7SThomas Gleixner default: 418dc186ad7SThomas Gleixner return 0; 419dc186ad7SThomas Gleixner } 420dc186ad7SThomas Gleixner } 421dc186ad7SThomas Gleixner 422dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 423dc186ad7SThomas Gleixner .name = "work_struct", 42499777288SStanislaw Gruszka .debug_hint = work_debug_hint, 425dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 426dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 427dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 428dc186ad7SThomas Gleixner }; 429dc186ad7SThomas Gleixner 430dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 431dc186ad7SThomas Gleixner { 432dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 433dc186ad7SThomas Gleixner } 434dc186ad7SThomas Gleixner 435dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 436dc186ad7SThomas Gleixner { 437dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 438dc186ad7SThomas Gleixner } 439dc186ad7SThomas Gleixner 440dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 441dc186ad7SThomas Gleixner { 442dc186ad7SThomas Gleixner if (onstack) 443dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 444dc186ad7SThomas Gleixner else 445dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 446dc186ad7SThomas Gleixner } 447dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 448dc186ad7SThomas Gleixner 449dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 450dc186ad7SThomas Gleixner { 451dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 452dc186ad7SThomas Gleixner } 453dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 454dc186ad7SThomas Gleixner 455dc186ad7SThomas Gleixner #else 456dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 457dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 458dc186ad7SThomas Gleixner #endif 459dc186ad7SThomas Gleixner 4609daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4619daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4629daf9e67STejun Heo { 4639daf9e67STejun Heo int ret; 4649daf9e67STejun Heo 4655bcab335STejun Heo lockdep_assert_held(&wq_mutex); 4665bcab335STejun Heo 467fa1b54e6STejun Heo do { 468fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 469fa1b54e6STejun Heo return -ENOMEM; 4709daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 471fa1b54e6STejun Heo } while (ret == -EAGAIN); 4729daf9e67STejun Heo 4739daf9e67STejun Heo return ret; 4749daf9e67STejun Heo } 4759daf9e67STejun Heo 47676af4d93STejun Heo /** 47776af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 47876af4d93STejun Heo * @wq: the target workqueue 47976af4d93STejun Heo * 480794b18bcSTejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 481794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 482794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 48376af4d93STejun Heo */ 4847fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 485a848e3b6SOleg Nesterov { 486794b18bcSTejun Heo assert_rcu_or_pwq_lock(); 48776af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 48876af4d93STejun Heo pwqs_node); 489f3421797STejun Heo } 490a848e3b6SOleg Nesterov 49173f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 49273f53c4aSTejun Heo { 49373f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 49473f53c4aSTejun Heo } 49573f53c4aSTejun Heo 49673f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 49773f53c4aSTejun Heo { 49873f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 49973f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 50073f53c4aSTejun Heo } 50173f53c4aSTejun Heo 50273f53c4aSTejun Heo static int work_next_color(int color) 50373f53c4aSTejun Heo { 50473f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 505b1f4ec17SOleg Nesterov } 506b1f4ec17SOleg Nesterov 5074594bf15SDavid Howells /* 508112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 509112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5107c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5117a22ad75STejun Heo * 512112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 513112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 514bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 515bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5167a22ad75STejun Heo * 517112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5187c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 519112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5207c3eed5cSTejun Heo * available only while the work item is queued. 521bbb68dfaSTejun Heo * 522bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 523bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 524bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 525bbb68dfaSTejun Heo * try to steal the PENDING bit. 5264594bf15SDavid Howells */ 5277a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5287a22ad75STejun Heo unsigned long flags) 5297a22ad75STejun Heo { 5306183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5317a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5327a22ad75STejun Heo } 5337a22ad75STejun Heo 534112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5354690c4abSTejun Heo unsigned long extra_flags) 536365970a1SDavid Howells { 537112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 538112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 539365970a1SDavid Howells } 540365970a1SDavid Howells 5414468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5424468a00fSLai Jiangshan int pool_id) 5434468a00fSLai Jiangshan { 5444468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5454468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5464468a00fSLai Jiangshan } 5474468a00fSLai Jiangshan 5487c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5497c3eed5cSTejun Heo int pool_id) 5504d707b9fSOleg Nesterov { 55123657bb1STejun Heo /* 55223657bb1STejun Heo * The following wmb is paired with the implied mb in 55323657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 55423657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 55523657bb1STejun Heo * owner. 55623657bb1STejun Heo */ 55723657bb1STejun Heo smp_wmb(); 5587c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5594d707b9fSOleg Nesterov } 5604d707b9fSOleg Nesterov 5617a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 562365970a1SDavid Howells { 5637c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5647c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5657a22ad75STejun Heo } 5667a22ad75STejun Heo 567112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5687a22ad75STejun Heo { 569e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5707a22ad75STejun Heo 571112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 572e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 573e120153dSTejun Heo else 574e120153dSTejun Heo return NULL; 5757a22ad75STejun Heo } 5767a22ad75STejun Heo 5777c3eed5cSTejun Heo /** 5787c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 5797c3eed5cSTejun Heo * @work: the work item of interest 5807c3eed5cSTejun Heo * 5817c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 582fa1b54e6STejun Heo * 5835bcab335STejun Heo * Pools are created and destroyed under wq_mutex, and allows read access 5845bcab335STejun Heo * under sched-RCU read lock. As such, this function should be called 5855bcab335STejun Heo * under wq_mutex or with preemption disabled. 586fa1b54e6STejun Heo * 587fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 588fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 589fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 590fa1b54e6STejun Heo * returned pool is and stays online. 5917c3eed5cSTejun Heo */ 5927c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 5937a22ad75STejun Heo { 594e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5957c3eed5cSTejun Heo int pool_id; 5967a22ad75STejun Heo 5975bcab335STejun Heo assert_rcu_or_wq_mutex(); 598fa1b54e6STejun Heo 599112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 600112202d9STejun Heo return ((struct pool_workqueue *) 6017c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6027a22ad75STejun Heo 6037c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6047c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6057a22ad75STejun Heo return NULL; 6067a22ad75STejun Heo 607fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6087c3eed5cSTejun Heo } 6097c3eed5cSTejun Heo 6107c3eed5cSTejun Heo /** 6117c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6127c3eed5cSTejun Heo * @work: the work item of interest 6137c3eed5cSTejun Heo * 6147c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6157c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6167c3eed5cSTejun Heo */ 6177c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6187c3eed5cSTejun Heo { 61954d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6207c3eed5cSTejun Heo 621112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 622112202d9STejun Heo return ((struct pool_workqueue *) 62354d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 62454d5b7d0SLai Jiangshan 62554d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6267c3eed5cSTejun Heo } 6277c3eed5cSTejun Heo 628bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 629bbb68dfaSTejun Heo { 6307c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 631bbb68dfaSTejun Heo 6327c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6337c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 634bbb68dfaSTejun Heo } 635bbb68dfaSTejun Heo 636bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 637bbb68dfaSTejun Heo { 638bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 639bbb68dfaSTejun Heo 640112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 641bbb68dfaSTejun Heo } 642bbb68dfaSTejun Heo 643e22bee78STejun Heo /* 6443270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6453270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 646d565ed63STejun Heo * they're being called with pool->lock held. 647e22bee78STejun Heo */ 648e22bee78STejun Heo 64963d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 650649027d7STejun Heo { 651e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 652649027d7STejun Heo } 653649027d7STejun Heo 654e22bee78STejun Heo /* 655e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 656e22bee78STejun Heo * running workers. 657974271c4STejun Heo * 658974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 659706026c2STejun Heo * function will always return %true for unbound pools as long as the 660974271c4STejun Heo * worklist isn't empty. 661e22bee78STejun Heo */ 66263d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 663e22bee78STejun Heo { 66463d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 665e22bee78STejun Heo } 666e22bee78STejun Heo 667e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 66863d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 669e22bee78STejun Heo { 67063d95a91STejun Heo return pool->nr_idle; 671e22bee78STejun Heo } 672e22bee78STejun Heo 673e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 67463d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 675e22bee78STejun Heo { 676e19e397aSTejun Heo return !list_empty(&pool->worklist) && 677e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 678e22bee78STejun Heo } 679e22bee78STejun Heo 680e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 68163d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 682e22bee78STejun Heo { 68363d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 684e22bee78STejun Heo } 685e22bee78STejun Heo 686e22bee78STejun Heo /* Do I need to be the manager? */ 68763d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 688e22bee78STejun Heo { 68963d95a91STejun Heo return need_to_create_worker(pool) || 69011ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 691e22bee78STejun Heo } 692e22bee78STejun Heo 693e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 69463d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 695e22bee78STejun Heo { 69634a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 69763d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 69863d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 699e22bee78STejun Heo 700ea1abd61SLai Jiangshan /* 701ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 702ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 703ea1abd61SLai Jiangshan */ 704ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 705ea1abd61SLai Jiangshan return false; 706ea1abd61SLai Jiangshan 707e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 708e22bee78STejun Heo } 709e22bee78STejun Heo 710e22bee78STejun Heo /* 711e22bee78STejun Heo * Wake up functions. 712e22bee78STejun Heo */ 713e22bee78STejun Heo 7147e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 71563d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7167e11629dSTejun Heo { 71763d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7187e11629dSTejun Heo return NULL; 7197e11629dSTejun Heo 72063d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7217e11629dSTejun Heo } 7227e11629dSTejun Heo 7237e11629dSTejun Heo /** 7247e11629dSTejun Heo * wake_up_worker - wake up an idle worker 72563d95a91STejun Heo * @pool: worker pool to wake worker from 7267e11629dSTejun Heo * 72763d95a91STejun Heo * Wake up the first idle worker of @pool. 7287e11629dSTejun Heo * 7297e11629dSTejun Heo * CONTEXT: 730d565ed63STejun Heo * spin_lock_irq(pool->lock). 7317e11629dSTejun Heo */ 73263d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7337e11629dSTejun Heo { 73463d95a91STejun Heo struct worker *worker = first_worker(pool); 7357e11629dSTejun Heo 7367e11629dSTejun Heo if (likely(worker)) 7377e11629dSTejun Heo wake_up_process(worker->task); 7387e11629dSTejun Heo } 7397e11629dSTejun Heo 7404690c4abSTejun Heo /** 741e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 742e22bee78STejun Heo * @task: task waking up 743e22bee78STejun Heo * @cpu: CPU @task is waking up to 744e22bee78STejun Heo * 745e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 746e22bee78STejun Heo * being awoken. 747e22bee78STejun Heo * 748e22bee78STejun Heo * CONTEXT: 749e22bee78STejun Heo * spin_lock_irq(rq->lock) 750e22bee78STejun Heo */ 751d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 752e22bee78STejun Heo { 753e22bee78STejun Heo struct worker *worker = kthread_data(task); 754e22bee78STejun Heo 75536576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 756ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 757e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 758e22bee78STejun Heo } 75936576000SJoonsoo Kim } 760e22bee78STejun Heo 761e22bee78STejun Heo /** 762e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 763e22bee78STejun Heo * @task: task going to sleep 764e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 765e22bee78STejun Heo * 766e22bee78STejun Heo * This function is called during schedule() when a busy worker is 767e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 768e22bee78STejun Heo * returning pointer to its task. 769e22bee78STejun Heo * 770e22bee78STejun Heo * CONTEXT: 771e22bee78STejun Heo * spin_lock_irq(rq->lock) 772e22bee78STejun Heo * 773e22bee78STejun Heo * RETURNS: 774e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 775e22bee78STejun Heo */ 776d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 777e22bee78STejun Heo { 778e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 779111c225aSTejun Heo struct worker_pool *pool; 780e22bee78STejun Heo 781111c225aSTejun Heo /* 782111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 783111c225aSTejun Heo * workers, also reach here, let's not access anything before 784111c225aSTejun Heo * checking NOT_RUNNING. 785111c225aSTejun Heo */ 7862d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 787e22bee78STejun Heo return NULL; 788e22bee78STejun Heo 789111c225aSTejun Heo pool = worker->pool; 790111c225aSTejun Heo 791e22bee78STejun Heo /* this can only happen on the local cpu */ 7926183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 7936183c009STejun Heo return NULL; 794e22bee78STejun Heo 795e22bee78STejun Heo /* 796e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 797e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 798e22bee78STejun Heo * Please read comment there. 799e22bee78STejun Heo * 800628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 801628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 802628c78e7STejun Heo * disabled, which in turn means that none else could be 803d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 804628c78e7STejun Heo * lock is safe. 805e22bee78STejun Heo */ 806e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 807e19e397aSTejun Heo !list_empty(&pool->worklist)) 80863d95a91STejun Heo to_wakeup = first_worker(pool); 809e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 810e22bee78STejun Heo } 811e22bee78STejun Heo 812e22bee78STejun Heo /** 813e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 814cb444766STejun Heo * @worker: self 815d302f017STejun Heo * @flags: flags to set 816d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 817d302f017STejun Heo * 818e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 819e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 820e22bee78STejun Heo * woken up. 821d302f017STejun Heo * 822cb444766STejun Heo * CONTEXT: 823d565ed63STejun Heo * spin_lock_irq(pool->lock) 824d302f017STejun Heo */ 825d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 826d302f017STejun Heo bool wakeup) 827d302f017STejun Heo { 828bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 829e22bee78STejun Heo 830cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 831cb444766STejun Heo 832e22bee78STejun Heo /* 833e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 834e22bee78STejun Heo * wake up an idle worker as necessary if requested by 835e22bee78STejun Heo * @wakeup. 836e22bee78STejun Heo */ 837e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 838e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 839e22bee78STejun Heo if (wakeup) { 840e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 841bd7bdd43STejun Heo !list_empty(&pool->worklist)) 84263d95a91STejun Heo wake_up_worker(pool); 843e22bee78STejun Heo } else 844e19e397aSTejun Heo atomic_dec(&pool->nr_running); 845e22bee78STejun Heo } 846e22bee78STejun Heo 847d302f017STejun Heo worker->flags |= flags; 848d302f017STejun Heo } 849d302f017STejun Heo 850d302f017STejun Heo /** 851e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 852cb444766STejun Heo * @worker: self 853d302f017STejun Heo * @flags: flags to clear 854d302f017STejun Heo * 855e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 856d302f017STejun Heo * 857cb444766STejun Heo * CONTEXT: 858d565ed63STejun Heo * spin_lock_irq(pool->lock) 859d302f017STejun Heo */ 860d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 861d302f017STejun Heo { 86263d95a91STejun Heo struct worker_pool *pool = worker->pool; 863e22bee78STejun Heo unsigned int oflags = worker->flags; 864e22bee78STejun Heo 865cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 866cb444766STejun Heo 867d302f017STejun Heo worker->flags &= ~flags; 868e22bee78STejun Heo 86942c025f3STejun Heo /* 87042c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 87142c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 87242c025f3STejun Heo * of multiple flags, not a single flag. 87342c025f3STejun Heo */ 874e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 875e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 876e19e397aSTejun Heo atomic_inc(&pool->nr_running); 877d302f017STejun Heo } 878d302f017STejun Heo 879d302f017STejun Heo /** 8808cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 881c9e7cf27STejun Heo * @pool: pool of interest 8828cca0eeaSTejun Heo * @work: work to find worker for 8838cca0eeaSTejun Heo * 884c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 885c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 886a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 887a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 888a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 889a2c1c57bSTejun Heo * being executed. 890a2c1c57bSTejun Heo * 891a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 892a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 893a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 894a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 895a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 896a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 897a2c1c57bSTejun Heo * 898c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 899c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 900c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 901c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 902c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 903c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9048cca0eeaSTejun Heo * 9058cca0eeaSTejun Heo * CONTEXT: 906d565ed63STejun Heo * spin_lock_irq(pool->lock). 9078cca0eeaSTejun Heo * 9088cca0eeaSTejun Heo * RETURNS: 9098cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9108cca0eeaSTejun Heo * otherwise. 9118cca0eeaSTejun Heo */ 912c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9138cca0eeaSTejun Heo struct work_struct *work) 9148cca0eeaSTejun Heo { 91542f8570fSSasha Levin struct worker *worker; 91642f8570fSSasha Levin 917b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 918a2c1c57bSTejun Heo (unsigned long)work) 919a2c1c57bSTejun Heo if (worker->current_work == work && 920a2c1c57bSTejun Heo worker->current_func == work->func) 92142f8570fSSasha Levin return worker; 92242f8570fSSasha Levin 92342f8570fSSasha Levin return NULL; 9248cca0eeaSTejun Heo } 9258cca0eeaSTejun Heo 9268cca0eeaSTejun Heo /** 927bf4ede01STejun Heo * move_linked_works - move linked works to a list 928bf4ede01STejun Heo * @work: start of series of works to be scheduled 929bf4ede01STejun Heo * @head: target list to append @work to 930bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 931bf4ede01STejun Heo * 932bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 933bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 934bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 935bf4ede01STejun Heo * 936bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 937bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 938bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 939bf4ede01STejun Heo * 940bf4ede01STejun Heo * CONTEXT: 941d565ed63STejun Heo * spin_lock_irq(pool->lock). 942bf4ede01STejun Heo */ 943bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 944bf4ede01STejun Heo struct work_struct **nextp) 945bf4ede01STejun Heo { 946bf4ede01STejun Heo struct work_struct *n; 947bf4ede01STejun Heo 948bf4ede01STejun Heo /* 949bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 950bf4ede01STejun Heo * use NULL for list head. 951bf4ede01STejun Heo */ 952bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 953bf4ede01STejun Heo list_move_tail(&work->entry, head); 954bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 955bf4ede01STejun Heo break; 956bf4ede01STejun Heo } 957bf4ede01STejun Heo 958bf4ede01STejun Heo /* 959bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 960bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 961bf4ede01STejun Heo * needs to be updated. 962bf4ede01STejun Heo */ 963bf4ede01STejun Heo if (nextp) 964bf4ede01STejun Heo *nextp = n; 965bf4ede01STejun Heo } 966bf4ede01STejun Heo 9678864b4e5STejun Heo /** 9688864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 9698864b4e5STejun Heo * @pwq: pool_workqueue to get 9708864b4e5STejun Heo * 9718864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 9728864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 9738864b4e5STejun Heo */ 9748864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 9758864b4e5STejun Heo { 9768864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 9778864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 9788864b4e5STejun Heo pwq->refcnt++; 9798864b4e5STejun Heo } 9808864b4e5STejun Heo 9818864b4e5STejun Heo /** 9828864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 9838864b4e5STejun Heo * @pwq: pool_workqueue to put 9848864b4e5STejun Heo * 9858864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 9868864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 9878864b4e5STejun Heo */ 9888864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 9898864b4e5STejun Heo { 9908864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 9918864b4e5STejun Heo if (likely(--pwq->refcnt)) 9928864b4e5STejun Heo return; 9938864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 9948864b4e5STejun Heo return; 9958864b4e5STejun Heo /* 9968864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 9978864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 9988864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 9998864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10008864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10018864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10028864b4e5STejun Heo */ 10038864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10048864b4e5STejun Heo } 10058864b4e5STejun Heo 1006112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1007bf4ede01STejun Heo { 1008112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1009bf4ede01STejun Heo 1010bf4ede01STejun Heo trace_workqueue_activate_work(work); 1011112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1012bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1013112202d9STejun Heo pwq->nr_active++; 1014bf4ede01STejun Heo } 1015bf4ede01STejun Heo 1016112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10173aa62497SLai Jiangshan { 1018112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10193aa62497SLai Jiangshan struct work_struct, entry); 10203aa62497SLai Jiangshan 1021112202d9STejun Heo pwq_activate_delayed_work(work); 10223aa62497SLai Jiangshan } 10233aa62497SLai Jiangshan 1024bf4ede01STejun Heo /** 1025112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1026112202d9STejun Heo * @pwq: pwq of interest 1027bf4ede01STejun Heo * @color: color of work which left the queue 1028bf4ede01STejun Heo * 1029bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1030112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1031bf4ede01STejun Heo * 1032bf4ede01STejun Heo * CONTEXT: 1033d565ed63STejun Heo * spin_lock_irq(pool->lock). 1034bf4ede01STejun Heo */ 1035112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1036bf4ede01STejun Heo { 10378864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1038bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10398864b4e5STejun Heo goto out_put; 1040bf4ede01STejun Heo 1041112202d9STejun Heo pwq->nr_in_flight[color]--; 1042bf4ede01STejun Heo 1043112202d9STejun Heo pwq->nr_active--; 1044112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1045bf4ede01STejun Heo /* one down, submit a delayed one */ 1046112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1047112202d9STejun Heo pwq_activate_first_delayed(pwq); 1048bf4ede01STejun Heo } 1049bf4ede01STejun Heo 1050bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1051112202d9STejun Heo if (likely(pwq->flush_color != color)) 10528864b4e5STejun Heo goto out_put; 1053bf4ede01STejun Heo 1054bf4ede01STejun Heo /* are there still in-flight works? */ 1055112202d9STejun Heo if (pwq->nr_in_flight[color]) 10568864b4e5STejun Heo goto out_put; 1057bf4ede01STejun Heo 1058112202d9STejun Heo /* this pwq is done, clear flush_color */ 1059112202d9STejun Heo pwq->flush_color = -1; 1060bf4ede01STejun Heo 1061bf4ede01STejun Heo /* 1062112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1063bf4ede01STejun Heo * will handle the rest. 1064bf4ede01STejun Heo */ 1065112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1066112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 10678864b4e5STejun Heo out_put: 10688864b4e5STejun Heo put_pwq(pwq); 1069bf4ede01STejun Heo } 1070bf4ede01STejun Heo 107136e227d2STejun Heo /** 1072bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 107336e227d2STejun Heo * @work: work item to steal 107436e227d2STejun Heo * @is_dwork: @work is a delayed_work 1075bbb68dfaSTejun Heo * @flags: place to store irq state 107636e227d2STejun Heo * 107736e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 107836e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 107936e227d2STejun Heo * 108036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 108136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 108236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1083bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1084bbb68dfaSTejun Heo * for arbitrarily long 108536e227d2STejun Heo * 1086bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1087e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1088e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1089e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1090bbb68dfaSTejun Heo * 1091bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1092bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1093bbb68dfaSTejun Heo * 1094e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1095bf4ede01STejun Heo */ 1096bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1097bbb68dfaSTejun Heo unsigned long *flags) 1098bf4ede01STejun Heo { 1099d565ed63STejun Heo struct worker_pool *pool; 1100112202d9STejun Heo struct pool_workqueue *pwq; 1101bf4ede01STejun Heo 1102bbb68dfaSTejun Heo local_irq_save(*flags); 1103bbb68dfaSTejun Heo 110436e227d2STejun Heo /* try to steal the timer if it exists */ 110536e227d2STejun Heo if (is_dwork) { 110636e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 110736e227d2STejun Heo 1108e0aecdd8STejun Heo /* 1109e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1110e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1111e0aecdd8STejun Heo * running on the local CPU. 1112e0aecdd8STejun Heo */ 111336e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 111436e227d2STejun Heo return 1; 111536e227d2STejun Heo } 111636e227d2STejun Heo 111736e227d2STejun Heo /* try to claim PENDING the normal way */ 1118bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1119bf4ede01STejun Heo return 0; 1120bf4ede01STejun Heo 1121bf4ede01STejun Heo /* 1122bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1123bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1124bf4ede01STejun Heo */ 1125d565ed63STejun Heo pool = get_work_pool(work); 1126d565ed63STejun Heo if (!pool) 1127bbb68dfaSTejun Heo goto fail; 1128bf4ede01STejun Heo 1129d565ed63STejun Heo spin_lock(&pool->lock); 1130bf4ede01STejun Heo /* 1131112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1132112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1133112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1134112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1135112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11360b3dae68SLai Jiangshan * item is currently queued on that pool. 1137bf4ede01STejun Heo */ 1138112202d9STejun Heo pwq = get_work_pwq(work); 1139112202d9STejun Heo if (pwq && pwq->pool == pool) { 1140bf4ede01STejun Heo debug_work_deactivate(work); 11413aa62497SLai Jiangshan 11423aa62497SLai Jiangshan /* 114316062836STejun Heo * A delayed work item cannot be grabbed directly because 114416062836STejun Heo * it might have linked NO_COLOR work items which, if left 1145112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 114616062836STejun Heo * management later on and cause stall. Make sure the work 114716062836STejun Heo * item is activated before grabbing. 11483aa62497SLai Jiangshan */ 11493aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1150112202d9STejun Heo pwq_activate_delayed_work(work); 11513aa62497SLai Jiangshan 1152bf4ede01STejun Heo list_del_init(&work->entry); 1153112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 115436e227d2STejun Heo 1155112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11564468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11574468a00fSLai Jiangshan 1158d565ed63STejun Heo spin_unlock(&pool->lock); 115936e227d2STejun Heo return 1; 1160bf4ede01STejun Heo } 1161d565ed63STejun Heo spin_unlock(&pool->lock); 1162bbb68dfaSTejun Heo fail: 1163bbb68dfaSTejun Heo local_irq_restore(*flags); 1164bbb68dfaSTejun Heo if (work_is_canceling(work)) 1165bbb68dfaSTejun Heo return -ENOENT; 1166bbb68dfaSTejun Heo cpu_relax(); 116736e227d2STejun Heo return -EAGAIN; 1168bf4ede01STejun Heo } 1169bf4ede01STejun Heo 1170bf4ede01STejun Heo /** 1171706026c2STejun Heo * insert_work - insert a work into a pool 1172112202d9STejun Heo * @pwq: pwq @work belongs to 11734690c4abSTejun Heo * @work: work to insert 11744690c4abSTejun Heo * @head: insertion point 11754690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11764690c4abSTejun Heo * 1177112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1178706026c2STejun Heo * work_struct flags. 11794690c4abSTejun Heo * 11804690c4abSTejun Heo * CONTEXT: 1181d565ed63STejun Heo * spin_lock_irq(pool->lock). 1182365970a1SDavid Howells */ 1183112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1184112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1185b89deed3SOleg Nesterov { 1186112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1187e1d8aa9fSFrederic Weisbecker 11884690c4abSTejun Heo /* we own @work, set data and link */ 1189112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 11901a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 11918864b4e5STejun Heo get_pwq(pwq); 1192e22bee78STejun Heo 1193e22bee78STejun Heo /* 1194c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1195c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1196c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1197e22bee78STejun Heo */ 1198e22bee78STejun Heo smp_mb(); 1199e22bee78STejun Heo 120063d95a91STejun Heo if (__need_more_worker(pool)) 120163d95a91STejun Heo wake_up_worker(pool); 1202b89deed3SOleg Nesterov } 1203b89deed3SOleg Nesterov 1204c8efcc25STejun Heo /* 1205c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12068d03ecfeSTejun Heo * same workqueue. 1207c8efcc25STejun Heo */ 1208c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1209c8efcc25STejun Heo { 1210c8efcc25STejun Heo struct worker *worker; 1211c8efcc25STejun Heo 12128d03ecfeSTejun Heo worker = current_wq_worker(); 1213c8efcc25STejun Heo /* 12148d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12158d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1216c8efcc25STejun Heo */ 1217112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1218c8efcc25STejun Heo } 1219c8efcc25STejun Heo 1220d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12211da177e4SLinus Torvalds struct work_struct *work) 12221da177e4SLinus Torvalds { 1223112202d9STejun Heo struct pool_workqueue *pwq; 1224c9178087STejun Heo struct worker_pool *last_pool; 12251e19ffc6STejun Heo struct list_head *worklist; 12268a2e8e5dSTejun Heo unsigned int work_flags; 1227b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12288930cabaSTejun Heo 12298930cabaSTejun Heo /* 12308930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12318930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12328930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12338930cabaSTejun Heo * happen with IRQ disabled. 12348930cabaSTejun Heo */ 12358930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12361da177e4SLinus Torvalds 1237dc186ad7SThomas Gleixner debug_work_activate(work); 12381e19ffc6STejun Heo 1239c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1240618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1241c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1242e41e704bSTejun Heo return; 12439e8cd2f5STejun Heo retry: 1244c9178087STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1245c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 124657469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1247f3421797STejun Heo cpu = raw_smp_processor_id(); 1248c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1249c9178087STejun Heo } else { 1250c9178087STejun Heo pwq = first_pwq(wq); 1251c9178087STejun Heo } 1252f3421797STejun Heo 125318aa9effSTejun Heo /* 1254c9178087STejun Heo * If @work was previously on a different pool, it might still be 1255c9178087STejun Heo * running there, in which case the work needs to be queued on that 1256c9178087STejun Heo * pool to guarantee non-reentrancy. 125718aa9effSTejun Heo */ 1258c9e7cf27STejun Heo last_pool = get_work_pool(work); 1259112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 126018aa9effSTejun Heo struct worker *worker; 126118aa9effSTejun Heo 1262d565ed63STejun Heo spin_lock(&last_pool->lock); 126318aa9effSTejun Heo 1264c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 126518aa9effSTejun Heo 1266112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1267c9178087STejun Heo pwq = worker->current_pwq; 12688594fadeSLai Jiangshan } else { 126918aa9effSTejun Heo /* meh... not running there, queue here */ 1270d565ed63STejun Heo spin_unlock(&last_pool->lock); 1271112202d9STejun Heo spin_lock(&pwq->pool->lock); 127218aa9effSTejun Heo } 12738930cabaSTejun Heo } else { 1274112202d9STejun Heo spin_lock(&pwq->pool->lock); 12758930cabaSTejun Heo } 1276502ca9d8STejun Heo 12779e8cd2f5STejun Heo /* 12789e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 12799e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 12809e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 12819e8cd2f5STejun Heo * without another pwq replacing it as the first pwq or while a 12829e8cd2f5STejun Heo * work item is executing on it, so the retying is guaranteed to 12839e8cd2f5STejun Heo * make forward-progress. 12849e8cd2f5STejun Heo */ 12859e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 12869e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 12879e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 12889e8cd2f5STejun Heo cpu_relax(); 12899e8cd2f5STejun Heo goto retry; 12909e8cd2f5STejun Heo } 12919e8cd2f5STejun Heo /* oops */ 12929e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 12939e8cd2f5STejun Heo wq->name, cpu); 12949e8cd2f5STejun Heo } 12959e8cd2f5STejun Heo 1296112202d9STejun Heo /* pwq determined, queue */ 1297112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1298502ca9d8STejun Heo 1299f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1300112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1301f5b2552bSDan Carpenter return; 1302f5b2552bSDan Carpenter } 13031e19ffc6STejun Heo 1304112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1305112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13061e19ffc6STejun Heo 1307112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1308cdadf009STejun Heo trace_workqueue_activate_work(work); 1309112202d9STejun Heo pwq->nr_active++; 1310112202d9STejun Heo worklist = &pwq->pool->worklist; 13118a2e8e5dSTejun Heo } else { 13128a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1313112202d9STejun Heo worklist = &pwq->delayed_works; 13148a2e8e5dSTejun Heo } 13151e19ffc6STejun Heo 1316112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13171e19ffc6STejun Heo 1318112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13191da177e4SLinus Torvalds } 13201da177e4SLinus Torvalds 13210fcb78c2SRolf Eike Beer /** 1322c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1323c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1324c1a220e7SZhang Rui * @wq: workqueue to use 1325c1a220e7SZhang Rui * @work: work to queue 1326c1a220e7SZhang Rui * 1327d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1328c1a220e7SZhang Rui * 1329c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1330c1a220e7SZhang Rui * can't go away. 1331c1a220e7SZhang Rui */ 1332d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1333d4283e93STejun Heo struct work_struct *work) 1334c1a220e7SZhang Rui { 1335d4283e93STejun Heo bool ret = false; 13368930cabaSTejun Heo unsigned long flags; 13378930cabaSTejun Heo 13388930cabaSTejun Heo local_irq_save(flags); 1339c1a220e7SZhang Rui 134022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13414690c4abSTejun Heo __queue_work(cpu, wq, work); 1342d4283e93STejun Heo ret = true; 1343c1a220e7SZhang Rui } 13448930cabaSTejun Heo 13458930cabaSTejun Heo local_irq_restore(flags); 1346c1a220e7SZhang Rui return ret; 1347c1a220e7SZhang Rui } 1348c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1349c1a220e7SZhang Rui 1350d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13511da177e4SLinus Torvalds { 135252bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13531da177e4SLinus Torvalds 1354e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 135560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13561da177e4SLinus Torvalds } 13571438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13581da177e4SLinus Torvalds 13597beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 136052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13611da177e4SLinus Torvalds { 13627beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13637beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13641da177e4SLinus Torvalds 13657beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13667beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1367fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1368fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13697beb2edfSTejun Heo 13708852aac2STejun Heo /* 13718852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13728852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13738852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13748852aac2STejun Heo * on that there's no such delay when @delay is 0. 13758852aac2STejun Heo */ 13768852aac2STejun Heo if (!delay) { 13778852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13788852aac2STejun Heo return; 13798852aac2STejun Heo } 13808852aac2STejun Heo 13817beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13827beb2edfSTejun Heo 138360c057bcSLai Jiangshan dwork->wq = wq; 13841265057fSTejun Heo dwork->cpu = cpu; 13857beb2edfSTejun Heo timer->expires = jiffies + delay; 13867beb2edfSTejun Heo 13877beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13887beb2edfSTejun Heo add_timer_on(timer, cpu); 13897beb2edfSTejun Heo else 13907beb2edfSTejun Heo add_timer(timer); 13917beb2edfSTejun Heo } 13921da177e4SLinus Torvalds 13930fcb78c2SRolf Eike Beer /** 13940fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13950fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13960fcb78c2SRolf Eike Beer * @wq: workqueue to use 1397af9997e4SRandy Dunlap * @dwork: work to queue 13980fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13990fcb78c2SRolf Eike Beer * 1400715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1401715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1402715f1300STejun Heo * execution. 14030fcb78c2SRolf Eike Beer */ 1404d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 140552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14067a6bc1cdSVenkatesh Pallipadi { 140752bad64dSDavid Howells struct work_struct *work = &dwork->work; 1408d4283e93STejun Heo bool ret = false; 14098930cabaSTejun Heo unsigned long flags; 14108930cabaSTejun Heo 14118930cabaSTejun Heo /* read the comment in __queue_work() */ 14128930cabaSTejun Heo local_irq_save(flags); 14137a6bc1cdSVenkatesh Pallipadi 141422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14157beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1416d4283e93STejun Heo ret = true; 14177a6bc1cdSVenkatesh Pallipadi } 14188930cabaSTejun Heo 14198930cabaSTejun Heo local_irq_restore(flags); 14207a6bc1cdSVenkatesh Pallipadi return ret; 14217a6bc1cdSVenkatesh Pallipadi } 1422ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14231da177e4SLinus Torvalds 1424c8e55f36STejun Heo /** 14258376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14268376fe22STejun Heo * @cpu: CPU number to execute work on 14278376fe22STejun Heo * @wq: workqueue to use 14288376fe22STejun Heo * @dwork: work to queue 14298376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14308376fe22STejun Heo * 14318376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14328376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14338376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14348376fe22STejun Heo * current state. 14358376fe22STejun Heo * 14368376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14378376fe22STejun Heo * pending and its timer was modified. 14388376fe22STejun Heo * 1439e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14408376fe22STejun Heo * See try_to_grab_pending() for details. 14418376fe22STejun Heo */ 14428376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14438376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14448376fe22STejun Heo { 14458376fe22STejun Heo unsigned long flags; 14468376fe22STejun Heo int ret; 14478376fe22STejun Heo 14488376fe22STejun Heo do { 14498376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14508376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14518376fe22STejun Heo 14528376fe22STejun Heo if (likely(ret >= 0)) { 14538376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14548376fe22STejun Heo local_irq_restore(flags); 14558376fe22STejun Heo } 14568376fe22STejun Heo 14578376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14588376fe22STejun Heo return ret; 14598376fe22STejun Heo } 14608376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14618376fe22STejun Heo 14628376fe22STejun Heo /** 1463c8e55f36STejun Heo * worker_enter_idle - enter idle state 1464c8e55f36STejun Heo * @worker: worker which is entering idle state 1465c8e55f36STejun Heo * 1466c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1467c8e55f36STejun Heo * necessary. 1468c8e55f36STejun Heo * 1469c8e55f36STejun Heo * LOCKING: 1470d565ed63STejun Heo * spin_lock_irq(pool->lock). 1471c8e55f36STejun Heo */ 1472c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14731da177e4SLinus Torvalds { 1474bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1475c8e55f36STejun Heo 14766183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 14776183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 14786183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 14796183c009STejun Heo return; 1480c8e55f36STejun Heo 1481cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1482cb444766STejun Heo worker->flags |= WORKER_IDLE; 1483bd7bdd43STejun Heo pool->nr_idle++; 1484e22bee78STejun Heo worker->last_active = jiffies; 1485c8e55f36STejun Heo 1486c8e55f36STejun Heo /* idle_list is LIFO */ 1487bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1488db7bccf4STejun Heo 148963d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1490628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1491cb444766STejun Heo 1492544ecf31STejun Heo /* 1493706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1494d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1495628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1496628c78e7STejun Heo * unbind is not in progress. 1497544ecf31STejun Heo */ 149824647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1499bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1500e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1501c8e55f36STejun Heo } 1502c8e55f36STejun Heo 1503c8e55f36STejun Heo /** 1504c8e55f36STejun Heo * worker_leave_idle - leave idle state 1505c8e55f36STejun Heo * @worker: worker which is leaving idle state 1506c8e55f36STejun Heo * 1507c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1508c8e55f36STejun Heo * 1509c8e55f36STejun Heo * LOCKING: 1510d565ed63STejun Heo * spin_lock_irq(pool->lock). 1511c8e55f36STejun Heo */ 1512c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1513c8e55f36STejun Heo { 1514bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1515c8e55f36STejun Heo 15166183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15176183c009STejun Heo return; 1518d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1519bd7bdd43STejun Heo pool->nr_idle--; 1520c8e55f36STejun Heo list_del_init(&worker->entry); 1521c8e55f36STejun Heo } 1522c8e55f36STejun Heo 1523e22bee78STejun Heo /** 1524f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1525f36dc67bSLai Jiangshan * @pool: target worker_pool 1526f36dc67bSLai Jiangshan * 1527f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1528e22bee78STejun Heo * 1529e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1530e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1531e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1532e22bee78STejun Heo * guaranteed to execute on the cpu. 1533e22bee78STejun Heo * 1534f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1535e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1536e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1537e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1538706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1539e22bee78STejun Heo * [dis]associated in the meantime. 1540e22bee78STejun Heo * 1541706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 154224647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1543f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1544f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1545f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1546e22bee78STejun Heo * 1547e22bee78STejun Heo * CONTEXT: 1548d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1549e22bee78STejun Heo * held. 1550e22bee78STejun Heo * 1551e22bee78STejun Heo * RETURNS: 1552706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1553e22bee78STejun Heo * bound), %false if offline. 1554e22bee78STejun Heo */ 1555f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1556d565ed63STejun Heo __acquires(&pool->lock) 1557e22bee78STejun Heo { 1558e22bee78STejun Heo while (true) { 1559e22bee78STejun Heo /* 1560e22bee78STejun Heo * The following call may fail, succeed or succeed 1561e22bee78STejun Heo * without actually migrating the task to the cpu if 1562e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 156324647570STejun Heo * against POOL_DISASSOCIATED. 1564e22bee78STejun Heo */ 156524647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 15667a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1567e22bee78STejun Heo 1568d565ed63STejun Heo spin_lock_irq(&pool->lock); 156924647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1570e22bee78STejun Heo return false; 1571f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 15727a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1573e22bee78STejun Heo return true; 1574d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1575e22bee78STejun Heo 15765035b20fSTejun Heo /* 15775035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 15785035b20fSTejun Heo * and retry migration. cond_resched() is required here; 15795035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 15805035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 15815035b20fSTejun Heo */ 1582e22bee78STejun Heo cpu_relax(); 15835035b20fSTejun Heo cond_resched(); 1584e22bee78STejun Heo } 1585e22bee78STejun Heo } 1586e22bee78STejun Heo 1587e22bee78STejun Heo /* 1588ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 15895f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 159025511a47STejun Heo */ 159125511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 159225511a47STejun Heo { 15935f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 1594f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 15955f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 159625511a47STejun Heo 1597ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1598ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1599d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 160025511a47STejun Heo } 160125511a47STejun Heo 160225511a47STejun Heo /* 160325511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1604403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1605403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1606403c821dSTejun Heo * executed twice without intervening cpu down. 1607e22bee78STejun Heo */ 160825511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1609e22bee78STejun Heo { 1610e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1611e22bee78STejun Heo 1612f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 1613eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1614e22bee78STejun Heo 1615d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 1616e22bee78STejun Heo } 1617e22bee78STejun Heo 161825511a47STejun Heo /** 161994cf58bbSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 162094cf58bbSTejun Heo * @pool: pool of interest 162125511a47STejun Heo * 162294cf58bbSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding 162325511a47STejun Heo * is different for idle and busy ones. 162425511a47STejun Heo * 1625ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1626ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1627ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1628ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 162925511a47STejun Heo * 1630ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1631ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1632ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1633ea1abd61SLai Jiangshan * rebind. 163425511a47STejun Heo * 1635ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1636ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1637ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1638ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 163925511a47STejun Heo */ 164094cf58bbSTejun Heo static void rebind_workers(struct worker_pool *pool) 164125511a47STejun Heo { 1642ea1abd61SLai Jiangshan struct worker *worker, *n; 164325511a47STejun Heo int i; 164425511a47STejun Heo 1645bc3a1afcSTejun Heo lockdep_assert_held(&pool->manager_mutex); 1646d565ed63STejun Heo lockdep_assert_held(&pool->lock); 164725511a47STejun Heo 16485f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 1649ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1650ea1abd61SLai Jiangshan /* 165194cf58bbSTejun Heo * idle workers should be off @pool->idle_list until rebind 165294cf58bbSTejun Heo * is complete to avoid receiving premature local wake-ups. 1653ea1abd61SLai Jiangshan */ 1654ea1abd61SLai Jiangshan list_del_init(&worker->entry); 165525511a47STejun Heo 165625511a47STejun Heo /* 165794cf58bbSTejun Heo * worker_thread() will see the above dequeuing and call 165894cf58bbSTejun Heo * idle_worker_rebind(). 165925511a47STejun Heo */ 166025511a47STejun Heo wake_up_process(worker->task); 166125511a47STejun Heo } 166225511a47STejun Heo 1663ea1abd61SLai Jiangshan /* rebind busy workers */ 1664b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) { 166525511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1666e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 166725511a47STejun Heo 166825511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 166925511a47STejun Heo work_data_bits(rebind_work))) 167025511a47STejun Heo continue; 167125511a47STejun Heo 167225511a47STejun Heo debug_work_activate(rebind_work); 167390beca5dSTejun Heo 167490beca5dSTejun Heo /* 167594cf58bbSTejun Heo * wq doesn't really matter but let's keep @worker->pool 1676112202d9STejun Heo * and @pwq->pool consistent for sanity. 167790beca5dSTejun Heo */ 16787a4e344cSTejun Heo if (worker->pool->attrs->nice < 0) 1679e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1680e2b6a6d5SJoonsoo Kim else 1681e2b6a6d5SJoonsoo Kim wq = system_wq; 1682ec58815aSTejun Heo 16837fb98ea7STejun Heo insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work, 168425511a47STejun Heo worker->scheduled.next, 168525511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1686ec58815aSTejun Heo } 168725511a47STejun Heo } 168825511a47STejun Heo 1689c34056a3STejun Heo static struct worker *alloc_worker(void) 1690c34056a3STejun Heo { 1691c34056a3STejun Heo struct worker *worker; 1692c34056a3STejun Heo 1693c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1694c8e55f36STejun Heo if (worker) { 1695c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1696affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 169725511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1698e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1699e22bee78STejun Heo worker->flags = WORKER_PREP; 1700c8e55f36STejun Heo } 1701c34056a3STejun Heo return worker; 1702c34056a3STejun Heo } 1703c34056a3STejun Heo 1704c34056a3STejun Heo /** 1705c34056a3STejun Heo * create_worker - create a new workqueue worker 170663d95a91STejun Heo * @pool: pool the new worker will belong to 1707c34056a3STejun Heo * 170863d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1709c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1710c34056a3STejun Heo * destroy_worker(). 1711c34056a3STejun Heo * 1712c34056a3STejun Heo * CONTEXT: 1713c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1714c34056a3STejun Heo * 1715c34056a3STejun Heo * RETURNS: 1716c34056a3STejun Heo * Pointer to the newly created worker. 1717c34056a3STejun Heo */ 1718bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1719c34056a3STejun Heo { 17207a4e344cSTejun Heo const char *pri = pool->attrs->nice < 0 ? "H" : ""; 1721c34056a3STejun Heo struct worker *worker = NULL; 1722f3421797STejun Heo int id = -1; 1723c34056a3STejun Heo 1724cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1725cd549687STejun Heo 1726d565ed63STejun Heo spin_lock_irq(&pool->lock); 1727bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 1728d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1729bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1730c34056a3STejun Heo goto fail; 1731d565ed63STejun Heo spin_lock_irq(&pool->lock); 1732c34056a3STejun Heo } 1733d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1734c34056a3STejun Heo 1735c34056a3STejun Heo worker = alloc_worker(); 1736c34056a3STejun Heo if (!worker) 1737c34056a3STejun Heo goto fail; 1738c34056a3STejun Heo 1739bd7bdd43STejun Heo worker->pool = pool; 1740c34056a3STejun Heo worker->id = id; 1741c34056a3STejun Heo 174229c91e99STejun Heo if (pool->cpu >= 0) 174394dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1744ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1745d84ff051STejun Heo "kworker/%d:%d%s", pool->cpu, id, pri); 1746f3421797STejun Heo else 1747f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1748ac6104cdSTejun Heo "kworker/u%d:%d%s", 1749ac6104cdSTejun Heo pool->id, id, pri); 1750c34056a3STejun Heo if (IS_ERR(worker->task)) 1751c34056a3STejun Heo goto fail; 1752c34056a3STejun Heo 1753c5aa87bbSTejun Heo /* 1754c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1755c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1756c5aa87bbSTejun Heo */ 17577a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 17587a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17593270476aSTejun Heo 1760db7bccf4STejun Heo /* 17617a4e344cSTejun Heo * %PF_THREAD_BOUND is used to prevent userland from meddling with 17627a4e344cSTejun Heo * cpumask of workqueue workers. This is an abuse. We need 17637a4e344cSTejun Heo * %PF_NO_SETAFFINITY. 1764db7bccf4STejun Heo */ 1765db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 17667a4e344cSTejun Heo 17677a4e344cSTejun Heo /* 17687a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17697a4e344cSTejun Heo * remains stable across this function. See the comments above the 17707a4e344cSTejun Heo * flag definition for details. 17717a4e344cSTejun Heo */ 17727a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1773f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1774c34056a3STejun Heo 1775c34056a3STejun Heo return worker; 1776c34056a3STejun Heo fail: 1777c34056a3STejun Heo if (id >= 0) { 1778d565ed63STejun Heo spin_lock_irq(&pool->lock); 1779bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1780d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1781c34056a3STejun Heo } 1782c34056a3STejun Heo kfree(worker); 1783c34056a3STejun Heo return NULL; 1784c34056a3STejun Heo } 1785c34056a3STejun Heo 1786c34056a3STejun Heo /** 1787c34056a3STejun Heo * start_worker - start a newly created worker 1788c34056a3STejun Heo * @worker: worker to start 1789c34056a3STejun Heo * 1790706026c2STejun Heo * Make the pool aware of @worker and start it. 1791c34056a3STejun Heo * 1792c34056a3STejun Heo * CONTEXT: 1793d565ed63STejun Heo * spin_lock_irq(pool->lock). 1794c34056a3STejun Heo */ 1795c34056a3STejun Heo static void start_worker(struct worker *worker) 1796c34056a3STejun Heo { 1797cb444766STejun Heo worker->flags |= WORKER_STARTED; 1798bd7bdd43STejun Heo worker->pool->nr_workers++; 1799c8e55f36STejun Heo worker_enter_idle(worker); 1800c34056a3STejun Heo wake_up_process(worker->task); 1801c34056a3STejun Heo } 1802c34056a3STejun Heo 1803c34056a3STejun Heo /** 1804ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1805ebf44d16STejun Heo * @pool: the target pool 1806ebf44d16STejun Heo * 1807cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1808ebf44d16STejun Heo */ 1809ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1810ebf44d16STejun Heo { 1811ebf44d16STejun Heo struct worker *worker; 1812ebf44d16STejun Heo 1813cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1814cd549687STejun Heo 1815ebf44d16STejun Heo worker = create_worker(pool); 1816ebf44d16STejun Heo if (worker) { 1817ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1818ebf44d16STejun Heo start_worker(worker); 1819ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1820ebf44d16STejun Heo } 1821ebf44d16STejun Heo 1822cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1823cd549687STejun Heo 1824ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1825ebf44d16STejun Heo } 1826ebf44d16STejun Heo 1827ebf44d16STejun Heo /** 1828c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1829c34056a3STejun Heo * @worker: worker to be destroyed 1830c34056a3STejun Heo * 1831706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1832c8e55f36STejun Heo * 1833c8e55f36STejun Heo * CONTEXT: 1834d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1835c34056a3STejun Heo */ 1836c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1837c34056a3STejun Heo { 1838bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1839c34056a3STejun Heo int id = worker->id; 1840c34056a3STejun Heo 1841cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1842cd549687STejun Heo lockdep_assert_held(&pool->lock); 1843cd549687STejun Heo 1844c34056a3STejun Heo /* sanity check frenzy */ 18456183c009STejun Heo if (WARN_ON(worker->current_work) || 18466183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18476183c009STejun Heo return; 1848c34056a3STejun Heo 1849c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1850bd7bdd43STejun Heo pool->nr_workers--; 1851c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1852bd7bdd43STejun Heo pool->nr_idle--; 1853c8e55f36STejun Heo 1854c8e55f36STejun Heo list_del_init(&worker->entry); 1855cb444766STejun Heo worker->flags |= WORKER_DIE; 1856c8e55f36STejun Heo 1857d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1858c8e55f36STejun Heo 1859c34056a3STejun Heo kthread_stop(worker->task); 1860c34056a3STejun Heo kfree(worker); 1861c34056a3STejun Heo 1862d565ed63STejun Heo spin_lock_irq(&pool->lock); 1863bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1864c34056a3STejun Heo } 1865c34056a3STejun Heo 186663d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1867e22bee78STejun Heo { 186863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1869e22bee78STejun Heo 1870d565ed63STejun Heo spin_lock_irq(&pool->lock); 1871e22bee78STejun Heo 187263d95a91STejun Heo if (too_many_workers(pool)) { 1873e22bee78STejun Heo struct worker *worker; 1874e22bee78STejun Heo unsigned long expires; 1875e22bee78STejun Heo 1876e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 187763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1878e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1879e22bee78STejun Heo 1880e22bee78STejun Heo if (time_before(jiffies, expires)) 188163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1882e22bee78STejun Heo else { 1883e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 188411ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 188563d95a91STejun Heo wake_up_worker(pool); 1886e22bee78STejun Heo } 1887e22bee78STejun Heo } 1888e22bee78STejun Heo 1889d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 1892493a1724STejun Heo static void send_mayday(struct work_struct *work) 1893e22bee78STejun Heo { 1894112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1895112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1896493a1724STejun Heo 18972e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1898e22bee78STejun Heo 1899493008a8STejun Heo if (!wq->rescuer) 1900493a1724STejun Heo return; 1901e22bee78STejun Heo 1902e22bee78STejun Heo /* mayday mayday mayday */ 1903493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1904493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1905e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1906493a1724STejun Heo } 1907e22bee78STejun Heo } 1908e22bee78STejun Heo 1909706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1910e22bee78STejun Heo { 191163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1912e22bee78STejun Heo struct work_struct *work; 1913e22bee78STejun Heo 19142e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1915493a1724STejun Heo spin_lock(&pool->lock); 1916e22bee78STejun Heo 191763d95a91STejun Heo if (need_to_create_worker(pool)) { 1918e22bee78STejun Heo /* 1919e22bee78STejun Heo * We've been trying to create a new worker but 1920e22bee78STejun Heo * haven't been successful. We might be hitting an 1921e22bee78STejun Heo * allocation deadlock. Send distress signals to 1922e22bee78STejun Heo * rescuers. 1923e22bee78STejun Heo */ 192463d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1925e22bee78STejun Heo send_mayday(work); 1926e22bee78STejun Heo } 1927e22bee78STejun Heo 1928493a1724STejun Heo spin_unlock(&pool->lock); 19292e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1930e22bee78STejun Heo 193163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1932e22bee78STejun Heo } 1933e22bee78STejun Heo 1934e22bee78STejun Heo /** 1935e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 193663d95a91STejun Heo * @pool: pool to create a new worker for 1937e22bee78STejun Heo * 193863d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1939e22bee78STejun Heo * have at least one idle worker on return from this function. If 1940e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 194163d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1942e22bee78STejun Heo * possible allocation deadlock. 1943e22bee78STejun Heo * 1944c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1945c5aa87bbSTejun Heo * may_start_working() %true. 1946e22bee78STejun Heo * 1947e22bee78STejun Heo * LOCKING: 1948d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1949e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1950e22bee78STejun Heo * manager. 1951e22bee78STejun Heo * 1952e22bee78STejun Heo * RETURNS: 1953c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1954e22bee78STejun Heo * otherwise. 1955e22bee78STejun Heo */ 195663d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1957d565ed63STejun Heo __releases(&pool->lock) 1958d565ed63STejun Heo __acquires(&pool->lock) 1959e22bee78STejun Heo { 196063d95a91STejun Heo if (!need_to_create_worker(pool)) 1961e22bee78STejun Heo return false; 1962e22bee78STejun Heo restart: 1963d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19649f9c2364STejun Heo 1965e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 196663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1967e22bee78STejun Heo 1968e22bee78STejun Heo while (true) { 1969e22bee78STejun Heo struct worker *worker; 1970e22bee78STejun Heo 1971bc2ae0f5STejun Heo worker = create_worker(pool); 1972e22bee78STejun Heo if (worker) { 197363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1974d565ed63STejun Heo spin_lock_irq(&pool->lock); 1975e22bee78STejun Heo start_worker(worker); 19766183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19776183c009STejun Heo goto restart; 1978e22bee78STejun Heo return true; 1979e22bee78STejun Heo } 1980e22bee78STejun Heo 198163d95a91STejun Heo if (!need_to_create_worker(pool)) 1982e22bee78STejun Heo break; 1983e22bee78STejun Heo 1984e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1985e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19869f9c2364STejun Heo 198763d95a91STejun Heo if (!need_to_create_worker(pool)) 1988e22bee78STejun Heo break; 1989e22bee78STejun Heo } 1990e22bee78STejun Heo 199163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1992d565ed63STejun Heo spin_lock_irq(&pool->lock); 199363d95a91STejun Heo if (need_to_create_worker(pool)) 1994e22bee78STejun Heo goto restart; 1995e22bee78STejun Heo return true; 1996e22bee78STejun Heo } 1997e22bee78STejun Heo 1998e22bee78STejun Heo /** 1999e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 200063d95a91STejun Heo * @pool: pool to destroy workers for 2001e22bee78STejun Heo * 200263d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 2003e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 2004e22bee78STejun Heo * 2005e22bee78STejun Heo * LOCKING: 2006d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2007e22bee78STejun Heo * multiple times. Called only from manager. 2008e22bee78STejun Heo * 2009e22bee78STejun Heo * RETURNS: 2010c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 2011e22bee78STejun Heo * otherwise. 2012e22bee78STejun Heo */ 201363d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 2014e22bee78STejun Heo { 2015e22bee78STejun Heo bool ret = false; 2016e22bee78STejun Heo 201763d95a91STejun Heo while (too_many_workers(pool)) { 2018e22bee78STejun Heo struct worker *worker; 2019e22bee78STejun Heo unsigned long expires; 2020e22bee78STejun Heo 202163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2022e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2023e22bee78STejun Heo 2024e22bee78STejun Heo if (time_before(jiffies, expires)) { 202563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2026e22bee78STejun Heo break; 2027e22bee78STejun Heo } 2028e22bee78STejun Heo 2029e22bee78STejun Heo destroy_worker(worker); 2030e22bee78STejun Heo ret = true; 2031e22bee78STejun Heo } 2032e22bee78STejun Heo 2033e22bee78STejun Heo return ret; 2034e22bee78STejun Heo } 2035e22bee78STejun Heo 2036e22bee78STejun Heo /** 2037e22bee78STejun Heo * manage_workers - manage worker pool 2038e22bee78STejun Heo * @worker: self 2039e22bee78STejun Heo * 2040706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2041e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2042706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2043e22bee78STejun Heo * 2044e22bee78STejun Heo * The caller can safely start processing works on false return. On 2045e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2046e22bee78STejun Heo * and may_start_working() is true. 2047e22bee78STejun Heo * 2048e22bee78STejun Heo * CONTEXT: 2049d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2050e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2051e22bee78STejun Heo * 2052e22bee78STejun Heo * RETURNS: 2053d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2054d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2055e22bee78STejun Heo */ 2056e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2057e22bee78STejun Heo { 205863d95a91STejun Heo struct worker_pool *pool = worker->pool; 2059e22bee78STejun Heo bool ret = false; 2060e22bee78STejun Heo 2061bc3a1afcSTejun Heo /* 2062bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2063bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2064bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2065bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2066bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2067bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2068bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2069bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2070bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2071bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2072bc3a1afcSTejun Heo * 2073bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2074bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2075bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2076bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2077bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2078bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2079bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2080bc3a1afcSTejun Heo * manager_mutex. 2081bc3a1afcSTejun Heo */ 208234a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2083e22bee78STejun Heo return ret; 2084e22bee78STejun Heo 2085ee378aa4SLai Jiangshan /* 2086bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2087bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2088ee378aa4SLai Jiangshan */ 2089bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2090d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2091bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2092ee378aa4SLai Jiangshan /* 2093ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2094b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2095ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2096706026c2STejun Heo * @pool's state and ours could have deviated. 2097ee378aa4SLai Jiangshan * 2098bc3a1afcSTejun Heo * As hotplug is now excluded via manager_mutex, we can 2099ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2100706026c2STejun Heo * on @pool's current state. Try it and adjust 2101ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2102ee378aa4SLai Jiangshan */ 2103f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(pool)) 2104ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2105ee378aa4SLai Jiangshan else 2106ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2107ee378aa4SLai Jiangshan 2108ee378aa4SLai Jiangshan ret = true; 2109ee378aa4SLai Jiangshan } 2110ee378aa4SLai Jiangshan 211111ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2112e22bee78STejun Heo 2113e22bee78STejun Heo /* 2114e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2115e22bee78STejun Heo * on return. 2116e22bee78STejun Heo */ 211763d95a91STejun Heo ret |= maybe_destroy_workers(pool); 211863d95a91STejun Heo ret |= maybe_create_worker(pool); 2119e22bee78STejun Heo 2120bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 212134a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2122e22bee78STejun Heo return ret; 2123e22bee78STejun Heo } 2124e22bee78STejun Heo 2125a62428c0STejun Heo /** 2126a62428c0STejun Heo * process_one_work - process single work 2127c34056a3STejun Heo * @worker: self 2128a62428c0STejun Heo * @work: work to process 2129a62428c0STejun Heo * 2130a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2131a62428c0STejun Heo * process a single work including synchronization against and 2132a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2133a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2134a62428c0STejun Heo * call this function to process a work. 2135a62428c0STejun Heo * 2136a62428c0STejun Heo * CONTEXT: 2137d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2138a62428c0STejun Heo */ 2139c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2140d565ed63STejun Heo __releases(&pool->lock) 2141d565ed63STejun Heo __acquires(&pool->lock) 21421da177e4SLinus Torvalds { 2143112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2144bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2145112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 214673f53c4aSTejun Heo int work_color; 21477e11629dSTejun Heo struct worker *collision; 21484e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21494e6045f1SJohannes Berg /* 2150a62428c0STejun Heo * It is permissible to free the struct work_struct from 2151a62428c0STejun Heo * inside the function that is called from it, this we need to 2152a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2153a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2154a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21554e6045f1SJohannes Berg */ 21564d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21574d82a1deSPeter Zijlstra 21584d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21594e6045f1SJohannes Berg #endif 21606fec10a1STejun Heo /* 21616fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21626fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 216324647570STejun Heo * unbound or a disassociated pool. 21646fec10a1STejun Heo */ 21655f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 216624647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2167ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 216825511a47STejun Heo 21697e11629dSTejun Heo /* 21707e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21717e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21727e11629dSTejun Heo * already processing the work. If so, defer the work to the 21737e11629dSTejun Heo * currently executing one. 21747e11629dSTejun Heo */ 2175c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21767e11629dSTejun Heo if (unlikely(collision)) { 21777e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21787e11629dSTejun Heo return; 21797e11629dSTejun Heo } 21801da177e4SLinus Torvalds 21818930cabaSTejun Heo /* claim and dequeue */ 21821da177e4SLinus Torvalds debug_work_deactivate(work); 2183c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2184c34056a3STejun Heo worker->current_work = work; 2185a2c1c57bSTejun Heo worker->current_func = work->func; 2186112202d9STejun Heo worker->current_pwq = pwq; 218773f53c4aSTejun Heo work_color = get_work_color(work); 21887a22ad75STejun Heo 2189a62428c0STejun Heo list_del_init(&work->entry); 2190a62428c0STejun Heo 2191649027d7STejun Heo /* 2192fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2193fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2194fb0e7bebSTejun Heo */ 2195fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2196fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2197fb0e7bebSTejun Heo 2198974271c4STejun Heo /* 2199d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2200974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2201974271c4STejun Heo */ 220263d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 220363d95a91STejun Heo wake_up_worker(pool); 2204974271c4STejun Heo 22058930cabaSTejun Heo /* 22067c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2207d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 220823657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 220923657bb1STejun Heo * disabled. 22108930cabaSTejun Heo */ 22117c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22121da177e4SLinus Torvalds 2213d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2214365970a1SDavid Howells 2215112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 22163295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2217e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2218a2c1c57bSTejun Heo worker->current_func(work); 2219e36c886aSArjan van de Ven /* 2220e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2221e36c886aSArjan van de Ven * point will only record its address. 2222e36c886aSArjan van de Ven */ 2223e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22243295f0efSIngo Molnar lock_map_release(&lockdep_map); 2225112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22261da177e4SLinus Torvalds 2227d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2228044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2229044c782cSValentin Ilie " last function: %pf\n", 2230a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2231a2c1c57bSTejun Heo worker->current_func); 2232d5abe669SPeter Zijlstra debug_show_held_locks(current); 2233d5abe669SPeter Zijlstra dump_stack(); 2234d5abe669SPeter Zijlstra } 2235d5abe669SPeter Zijlstra 2236d565ed63STejun Heo spin_lock_irq(&pool->lock); 2237a62428c0STejun Heo 2238fb0e7bebSTejun Heo /* clear cpu intensive status */ 2239fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2240fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2241fb0e7bebSTejun Heo 2242a62428c0STejun Heo /* we're done with it, release */ 224342f8570fSSasha Levin hash_del(&worker->hentry); 2244c34056a3STejun Heo worker->current_work = NULL; 2245a2c1c57bSTejun Heo worker->current_func = NULL; 2246112202d9STejun Heo worker->current_pwq = NULL; 2247112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 22481da177e4SLinus Torvalds } 22491da177e4SLinus Torvalds 2250affee4b2STejun Heo /** 2251affee4b2STejun Heo * process_scheduled_works - process scheduled works 2252affee4b2STejun Heo * @worker: self 2253affee4b2STejun Heo * 2254affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2255affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2256affee4b2STejun Heo * fetches a work from the top and executes it. 2257affee4b2STejun Heo * 2258affee4b2STejun Heo * CONTEXT: 2259d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2260affee4b2STejun Heo * multiple times. 2261affee4b2STejun Heo */ 2262affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22631da177e4SLinus Torvalds { 2264affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2265affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2266a62428c0STejun Heo struct work_struct, entry); 2267c34056a3STejun Heo process_one_work(worker, work); 2268a62428c0STejun Heo } 22691da177e4SLinus Torvalds } 22701da177e4SLinus Torvalds 22714690c4abSTejun Heo /** 22724690c4abSTejun Heo * worker_thread - the worker thread function 2273c34056a3STejun Heo * @__worker: self 22744690c4abSTejun Heo * 2275c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2276c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2277c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2278c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2279c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 22804690c4abSTejun Heo */ 2281c34056a3STejun Heo static int worker_thread(void *__worker) 22821da177e4SLinus Torvalds { 2283c34056a3STejun Heo struct worker *worker = __worker; 2284bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22851da177e4SLinus Torvalds 2286e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2287e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2288c8e55f36STejun Heo woke_up: 2289d565ed63STejun Heo spin_lock_irq(&pool->lock); 2290affee4b2STejun Heo 22915f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22925f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2293d565ed63STejun Heo spin_unlock_irq(&pool->lock); 229425511a47STejun Heo 22955f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 229625511a47STejun Heo if (worker->flags & WORKER_DIE) { 2297e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2298c8e55f36STejun Heo return 0; 2299c8e55f36STejun Heo } 2300c8e55f36STejun Heo 23015f7dabfdSLai Jiangshan /* otherwise, rebind */ 230225511a47STejun Heo idle_worker_rebind(worker); 230325511a47STejun Heo goto woke_up; 230425511a47STejun Heo } 230525511a47STejun Heo 2306c8e55f36STejun Heo worker_leave_idle(worker); 2307db7bccf4STejun Heo recheck: 2308e22bee78STejun Heo /* no more worker necessary? */ 230963d95a91STejun Heo if (!need_more_worker(pool)) 2310e22bee78STejun Heo goto sleep; 2311e22bee78STejun Heo 2312e22bee78STejun Heo /* do we need to manage? */ 231363d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2314e22bee78STejun Heo goto recheck; 2315e22bee78STejun Heo 2316c8e55f36STejun Heo /* 2317c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2318c8e55f36STejun Heo * preparing to process a work or actually processing it. 2319c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2320c8e55f36STejun Heo */ 23216183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2322c8e55f36STejun Heo 2323e22bee78STejun Heo /* 2324e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2325e22bee78STejun Heo * at least one idle worker or that someone else has already 2326e22bee78STejun Heo * assumed the manager role. 2327e22bee78STejun Heo */ 2328e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2329e22bee78STejun Heo 2330e22bee78STejun Heo do { 2331affee4b2STejun Heo struct work_struct *work = 2332bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2333affee4b2STejun Heo struct work_struct, entry); 2334affee4b2STejun Heo 2335c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2336affee4b2STejun Heo /* optimization path, not strictly necessary */ 2337affee4b2STejun Heo process_one_work(worker, work); 2338affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2339affee4b2STejun Heo process_scheduled_works(worker); 2340affee4b2STejun Heo } else { 2341c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2342affee4b2STejun Heo process_scheduled_works(worker); 2343affee4b2STejun Heo } 234463d95a91STejun Heo } while (keep_working(pool)); 2345affee4b2STejun Heo 2346e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2347d313dd85STejun Heo sleep: 234863d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2349e22bee78STejun Heo goto recheck; 2350d313dd85STejun Heo 2351c8e55f36STejun Heo /* 2352d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2353d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2354d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2355d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2356d565ed63STejun Heo * event. 2357c8e55f36STejun Heo */ 2358c8e55f36STejun Heo worker_enter_idle(worker); 2359c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2360d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23611da177e4SLinus Torvalds schedule(); 2362c8e55f36STejun Heo goto woke_up; 23631da177e4SLinus Torvalds } 23641da177e4SLinus Torvalds 2365e22bee78STejun Heo /** 2366e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2367111c225aSTejun Heo * @__rescuer: self 2368e22bee78STejun Heo * 2369e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2370493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2371e22bee78STejun Heo * 2372706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2373e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2374e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2375e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2376e22bee78STejun Heo * the problem rescuer solves. 2377e22bee78STejun Heo * 2378706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2379706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2380e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2381e22bee78STejun Heo * 2382e22bee78STejun Heo * This should happen rarely. 2383e22bee78STejun Heo */ 2384111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2385e22bee78STejun Heo { 2386111c225aSTejun Heo struct worker *rescuer = __rescuer; 2387111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2388e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2389e22bee78STejun Heo 2390e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2391111c225aSTejun Heo 2392111c225aSTejun Heo /* 2393111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2394111c225aSTejun Heo * doesn't participate in concurrency management. 2395111c225aSTejun Heo */ 2396111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2397e22bee78STejun Heo repeat: 2398e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23991da177e4SLinus Torvalds 2400412d32e6SMike Galbraith if (kthread_should_stop()) { 2401412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2402111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2403e22bee78STejun Heo return 0; 2404412d32e6SMike Galbraith } 24051da177e4SLinus Torvalds 2406493a1724STejun Heo /* see whether any pwq is asking for help */ 24072e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2408493a1724STejun Heo 2409493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2410493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2411493a1724STejun Heo struct pool_workqueue, mayday_node); 2412112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2413e22bee78STejun Heo struct work_struct *work, *n; 2414e22bee78STejun Heo 2415e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2416493a1724STejun Heo list_del_init(&pwq->mayday_node); 2417493a1724STejun Heo 24182e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2419e22bee78STejun Heo 2420e22bee78STejun Heo /* migrate to the target cpu if possible */ 2421f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2422b3104104SLai Jiangshan rescuer->pool = pool; 2423e22bee78STejun Heo 2424e22bee78STejun Heo /* 2425e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2426e22bee78STejun Heo * process'em. 2427e22bee78STejun Heo */ 24286183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2429bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2430112202d9STejun Heo if (get_work_pwq(work) == pwq) 2431e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2432e22bee78STejun Heo 2433e22bee78STejun Heo process_scheduled_works(rescuer); 24347576958aSTejun Heo 24357576958aSTejun Heo /* 2436d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 24377576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24387576958aSTejun Heo * and stalling the execution. 24397576958aSTejun Heo */ 244063d95a91STejun Heo if (keep_working(pool)) 244163d95a91STejun Heo wake_up_worker(pool); 24427576958aSTejun Heo 2443b3104104SLai Jiangshan rescuer->pool = NULL; 2444493a1724STejun Heo spin_unlock(&pool->lock); 24452e109a28STejun Heo spin_lock(&wq_mayday_lock); 24461da177e4SLinus Torvalds } 24471da177e4SLinus Torvalds 24482e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2449493a1724STejun Heo 2450111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2451111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2452e22bee78STejun Heo schedule(); 2453e22bee78STejun Heo goto repeat; 24541da177e4SLinus Torvalds } 24551da177e4SLinus Torvalds 2456fc2e4d70SOleg Nesterov struct wq_barrier { 2457fc2e4d70SOleg Nesterov struct work_struct work; 2458fc2e4d70SOleg Nesterov struct completion done; 2459fc2e4d70SOleg Nesterov }; 2460fc2e4d70SOleg Nesterov 2461fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2462fc2e4d70SOleg Nesterov { 2463fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2464fc2e4d70SOleg Nesterov complete(&barr->done); 2465fc2e4d70SOleg Nesterov } 2466fc2e4d70SOleg Nesterov 24674690c4abSTejun Heo /** 24684690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2469112202d9STejun Heo * @pwq: pwq to insert barrier into 24704690c4abSTejun Heo * @barr: wq_barrier to insert 2471affee4b2STejun Heo * @target: target work to attach @barr to 2472affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24734690c4abSTejun Heo * 2474affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2475affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2476affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2477affee4b2STejun Heo * cpu. 2478affee4b2STejun Heo * 2479affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2480affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2481affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2482affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2483affee4b2STejun Heo * after a work with LINKED flag set. 2484affee4b2STejun Heo * 2485affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2486112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24874690c4abSTejun Heo * 24884690c4abSTejun Heo * CONTEXT: 2489d565ed63STejun Heo * spin_lock_irq(pool->lock). 24904690c4abSTejun Heo */ 2491112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2492affee4b2STejun Heo struct wq_barrier *barr, 2493affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2494fc2e4d70SOleg Nesterov { 2495affee4b2STejun Heo struct list_head *head; 2496affee4b2STejun Heo unsigned int linked = 0; 2497affee4b2STejun Heo 2498dc186ad7SThomas Gleixner /* 2499d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2500dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2501dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2502dc186ad7SThomas Gleixner * might deadlock. 2503dc186ad7SThomas Gleixner */ 2504ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 250522df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2506fc2e4d70SOleg Nesterov init_completion(&barr->done); 250783c22520SOleg Nesterov 2508affee4b2STejun Heo /* 2509affee4b2STejun Heo * If @target is currently being executed, schedule the 2510affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2511affee4b2STejun Heo */ 2512affee4b2STejun Heo if (worker) 2513affee4b2STejun Heo head = worker->scheduled.next; 2514affee4b2STejun Heo else { 2515affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2516affee4b2STejun Heo 2517affee4b2STejun Heo head = target->entry.next; 2518affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2519affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2520affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2521affee4b2STejun Heo } 2522affee4b2STejun Heo 2523dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2524112202d9STejun Heo insert_work(pwq, &barr->work, head, 2525affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2526fc2e4d70SOleg Nesterov } 2527fc2e4d70SOleg Nesterov 252873f53c4aSTejun Heo /** 2529112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 253073f53c4aSTejun Heo * @wq: workqueue being flushed 253173f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 253273f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 253373f53c4aSTejun Heo * 2534112202d9STejun Heo * Prepare pwqs for workqueue flushing. 253573f53c4aSTejun Heo * 2536112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2537112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2538112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2539112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2540112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 254173f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 254273f53c4aSTejun Heo * 254373f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 254473f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 254573f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 254673f53c4aSTejun Heo * is returned. 254773f53c4aSTejun Heo * 2548112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 254973f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 255073f53c4aSTejun Heo * advanced to @work_color. 255173f53c4aSTejun Heo * 255273f53c4aSTejun Heo * CONTEXT: 255373f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 255473f53c4aSTejun Heo * 255573f53c4aSTejun Heo * RETURNS: 255673f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 255773f53c4aSTejun Heo * otherwise. 255873f53c4aSTejun Heo */ 2559112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 256073f53c4aSTejun Heo int flush_color, int work_color) 25611da177e4SLinus Torvalds { 256273f53c4aSTejun Heo bool wait = false; 256349e3cf44STejun Heo struct pool_workqueue *pwq; 25641da177e4SLinus Torvalds 256573f53c4aSTejun Heo if (flush_color >= 0) { 25666183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2567112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2568dc186ad7SThomas Gleixner } 256914441960SOleg Nesterov 257076af4d93STejun Heo local_irq_disable(); 257176af4d93STejun Heo 257249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2573112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25741da177e4SLinus Torvalds 257576af4d93STejun Heo spin_lock(&pool->lock); 257673f53c4aSTejun Heo 257773f53c4aSTejun Heo if (flush_color >= 0) { 25786183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 257973f53c4aSTejun Heo 2580112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2581112202d9STejun Heo pwq->flush_color = flush_color; 2582112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 258373f53c4aSTejun Heo wait = true; 25841da177e4SLinus Torvalds } 258573f53c4aSTejun Heo } 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo if (work_color >= 0) { 25886183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2589112202d9STejun Heo pwq->work_color = work_color; 259073f53c4aSTejun Heo } 259173f53c4aSTejun Heo 259276af4d93STejun Heo spin_unlock(&pool->lock); 25931da177e4SLinus Torvalds } 25941da177e4SLinus Torvalds 259576af4d93STejun Heo local_irq_enable(); 259676af4d93STejun Heo 2597112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 259873f53c4aSTejun Heo complete(&wq->first_flusher->done); 259973f53c4aSTejun Heo 260073f53c4aSTejun Heo return wait; 260183c22520SOleg Nesterov } 26021da177e4SLinus Torvalds 26030fcb78c2SRolf Eike Beer /** 26041da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26050fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26061da177e4SLinus Torvalds * 2607c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2608c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 26091da177e4SLinus Torvalds */ 26107ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26111da177e4SLinus Torvalds { 261273f53c4aSTejun Heo struct wq_flusher this_flusher = { 261373f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 261473f53c4aSTejun Heo .flush_color = -1, 261573f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 261673f53c4aSTejun Heo }; 261773f53c4aSTejun Heo int next_color; 2618b1f4ec17SOleg Nesterov 26193295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 26203295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 262173f53c4aSTejun Heo 262273f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 262373f53c4aSTejun Heo 262473f53c4aSTejun Heo /* 262573f53c4aSTejun Heo * Start-to-wait phase 262673f53c4aSTejun Heo */ 262773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 262873f53c4aSTejun Heo 262973f53c4aSTejun Heo if (next_color != wq->flush_color) { 263073f53c4aSTejun Heo /* 263173f53c4aSTejun Heo * Color space is not full. The current work_color 263273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 263373f53c4aSTejun Heo * by one. 263473f53c4aSTejun Heo */ 26356183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 263673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 263773f53c4aSTejun Heo wq->work_color = next_color; 263873f53c4aSTejun Heo 263973f53c4aSTejun Heo if (!wq->first_flusher) { 264073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26416183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 264273f53c4aSTejun Heo 264373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 264473f53c4aSTejun Heo 2645112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 264673f53c4aSTejun Heo wq->work_color)) { 264773f53c4aSTejun Heo /* nothing to flush, done */ 264873f53c4aSTejun Heo wq->flush_color = next_color; 264973f53c4aSTejun Heo wq->first_flusher = NULL; 265073f53c4aSTejun Heo goto out_unlock; 265173f53c4aSTejun Heo } 265273f53c4aSTejun Heo } else { 265373f53c4aSTejun Heo /* wait in queue */ 26546183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 265573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2656112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 265773f53c4aSTejun Heo } 265873f53c4aSTejun Heo } else { 265973f53c4aSTejun Heo /* 266073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 266173f53c4aSTejun Heo * The next flush completion will assign us 266273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 266373f53c4aSTejun Heo */ 266473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 266573f53c4aSTejun Heo } 266673f53c4aSTejun Heo 266773f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 266873f53c4aSTejun Heo 266973f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 267073f53c4aSTejun Heo 267173f53c4aSTejun Heo /* 267273f53c4aSTejun Heo * Wake-up-and-cascade phase 267373f53c4aSTejun Heo * 267473f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 267573f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 267673f53c4aSTejun Heo */ 267773f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 267873f53c4aSTejun Heo return; 267973f53c4aSTejun Heo 268073f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 268173f53c4aSTejun Heo 26824ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26834ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26844ce48b37STejun Heo goto out_unlock; 26854ce48b37STejun Heo 268673f53c4aSTejun Heo wq->first_flusher = NULL; 268773f53c4aSTejun Heo 26886183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26896183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 269073f53c4aSTejun Heo 269173f53c4aSTejun Heo while (true) { 269273f53c4aSTejun Heo struct wq_flusher *next, *tmp; 269373f53c4aSTejun Heo 269473f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 269573f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 269673f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 269773f53c4aSTejun Heo break; 269873f53c4aSTejun Heo list_del_init(&next->list); 269973f53c4aSTejun Heo complete(&next->done); 270073f53c4aSTejun Heo } 270173f53c4aSTejun Heo 27026183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 270373f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 270473f53c4aSTejun Heo 270573f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 270673f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 270773f53c4aSTejun Heo 270873f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 270973f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 271073f53c4aSTejun Heo /* 271173f53c4aSTejun Heo * Assign the same color to all overflowed 271273f53c4aSTejun Heo * flushers, advance work_color and append to 271373f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 271473f53c4aSTejun Heo * phase for these overflowed flushers. 271573f53c4aSTejun Heo */ 271673f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 271773f53c4aSTejun Heo tmp->flush_color = wq->work_color; 271873f53c4aSTejun Heo 271973f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 272073f53c4aSTejun Heo 272173f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 272273f53c4aSTejun Heo &wq->flusher_queue); 2723112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 272473f53c4aSTejun Heo } 272573f53c4aSTejun Heo 272673f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27276183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 272873f53c4aSTejun Heo break; 272973f53c4aSTejun Heo } 273073f53c4aSTejun Heo 273173f53c4aSTejun Heo /* 273273f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2733112202d9STejun Heo * the new first flusher and arm pwqs. 273473f53c4aSTejun Heo */ 27356183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27366183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 273773f53c4aSTejun Heo 273873f53c4aSTejun Heo list_del_init(&next->list); 273973f53c4aSTejun Heo wq->first_flusher = next; 274073f53c4aSTejun Heo 2741112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 274273f53c4aSTejun Heo break; 274373f53c4aSTejun Heo 274473f53c4aSTejun Heo /* 274573f53c4aSTejun Heo * Meh... this color is already done, clear first 274673f53c4aSTejun Heo * flusher and repeat cascading. 274773f53c4aSTejun Heo */ 274873f53c4aSTejun Heo wq->first_flusher = NULL; 274973f53c4aSTejun Heo } 275073f53c4aSTejun Heo 275173f53c4aSTejun Heo out_unlock: 275273f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 27531da177e4SLinus Torvalds } 2754ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27551da177e4SLinus Torvalds 27569c5a2ba7STejun Heo /** 27579c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27589c5a2ba7STejun Heo * @wq: workqueue to drain 27599c5a2ba7STejun Heo * 27609c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27619c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27629c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27639c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27649c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27659c5a2ba7STejun Heo * takes too long. 27669c5a2ba7STejun Heo */ 27679c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27689c5a2ba7STejun Heo { 27699c5a2ba7STejun Heo unsigned int flush_cnt = 0; 277049e3cf44STejun Heo struct pool_workqueue *pwq; 27719c5a2ba7STejun Heo 27729c5a2ba7STejun Heo /* 27739c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27749c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2775618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27769c5a2ba7STejun Heo */ 27775bcab335STejun Heo mutex_lock(&wq_mutex); 27789c5a2ba7STejun Heo if (!wq->nr_drainers++) 2779618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 27805bcab335STejun Heo mutex_unlock(&wq_mutex); 27819c5a2ba7STejun Heo reflush: 27829c5a2ba7STejun Heo flush_workqueue(wq); 27839c5a2ba7STejun Heo 278476af4d93STejun Heo local_irq_disable(); 278576af4d93STejun Heo 278649e3cf44STejun Heo for_each_pwq(pwq, wq) { 2787fa2563e4SThomas Tuttle bool drained; 27889c5a2ba7STejun Heo 278976af4d93STejun Heo spin_lock(&pwq->pool->lock); 2790112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 279176af4d93STejun Heo spin_unlock(&pwq->pool->lock); 2792fa2563e4SThomas Tuttle 2793fa2563e4SThomas Tuttle if (drained) 27949c5a2ba7STejun Heo continue; 27959c5a2ba7STejun Heo 27969c5a2ba7STejun Heo if (++flush_cnt == 10 || 27979c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2798c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27999c5a2ba7STejun Heo wq->name, flush_cnt); 280076af4d93STejun Heo 280176af4d93STejun Heo local_irq_enable(); 28029c5a2ba7STejun Heo goto reflush; 28039c5a2ba7STejun Heo } 28049c5a2ba7STejun Heo 28055bcab335STejun Heo local_irq_enable(); 28065bcab335STejun Heo 28075bcab335STejun Heo mutex_lock(&wq_mutex); 28089c5a2ba7STejun Heo if (!--wq->nr_drainers) 2809618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 28105bcab335STejun Heo mutex_unlock(&wq_mutex); 28119c5a2ba7STejun Heo } 28129c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28139c5a2ba7STejun Heo 2814606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2815baf59022STejun Heo { 2816baf59022STejun Heo struct worker *worker = NULL; 2817c9e7cf27STejun Heo struct worker_pool *pool; 2818112202d9STejun Heo struct pool_workqueue *pwq; 2819baf59022STejun Heo 2820baf59022STejun Heo might_sleep(); 2821baf59022STejun Heo 2822fa1b54e6STejun Heo local_irq_disable(); 2823fa1b54e6STejun Heo pool = get_work_pool(work); 2824fa1b54e6STejun Heo if (!pool) { 2825fa1b54e6STejun Heo local_irq_enable(); 2826fa1b54e6STejun Heo return false; 2827fa1b54e6STejun Heo } 2828fa1b54e6STejun Heo 2829fa1b54e6STejun Heo spin_lock(&pool->lock); 28300b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2831112202d9STejun Heo pwq = get_work_pwq(work); 2832112202d9STejun Heo if (pwq) { 2833112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2834baf59022STejun Heo goto already_gone; 2835606a5020STejun Heo } else { 2836c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2837baf59022STejun Heo if (!worker) 2838baf59022STejun Heo goto already_gone; 2839112202d9STejun Heo pwq = worker->current_pwq; 2840606a5020STejun Heo } 2841baf59022STejun Heo 2842112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2843d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2844baf59022STejun Heo 2845e159489bSTejun Heo /* 2846e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2847e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2848e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2849e159489bSTejun Heo * access. 2850e159489bSTejun Heo */ 2851493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2852112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2853e159489bSTejun Heo else 2854112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2855112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2856e159489bSTejun Heo 2857baf59022STejun Heo return true; 2858baf59022STejun Heo already_gone: 2859d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2860baf59022STejun Heo return false; 2861baf59022STejun Heo } 2862baf59022STejun Heo 2863db700897SOleg Nesterov /** 2864401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2865401a8d04STejun Heo * @work: the work to flush 2866db700897SOleg Nesterov * 2867606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2868606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2869401a8d04STejun Heo * 2870401a8d04STejun Heo * RETURNS: 2871401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2872401a8d04STejun Heo * %false if it was already idle. 2873db700897SOleg Nesterov */ 2874401a8d04STejun Heo bool flush_work(struct work_struct *work) 2875db700897SOleg Nesterov { 2876db700897SOleg Nesterov struct wq_barrier barr; 2877db700897SOleg Nesterov 28780976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28790976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28800976dfc1SStephen Boyd 2881606a5020STejun Heo if (start_flush_work(work, &barr)) { 2882db700897SOleg Nesterov wait_for_completion(&barr.done); 2883dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2884401a8d04STejun Heo return true; 2885606a5020STejun Heo } else { 2886401a8d04STejun Heo return false; 2887db700897SOleg Nesterov } 2888606a5020STejun Heo } 2889db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2890db700897SOleg Nesterov 289136e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2892401a8d04STejun Heo { 2893bbb68dfaSTejun Heo unsigned long flags; 28941f1f642eSOleg Nesterov int ret; 28951f1f642eSOleg Nesterov 28961f1f642eSOleg Nesterov do { 2897bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2898bbb68dfaSTejun Heo /* 2899bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2900bbb68dfaSTejun Heo * would be waiting for before retrying. 2901bbb68dfaSTejun Heo */ 2902bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2903606a5020STejun Heo flush_work(work); 29041f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29051f1f642eSOleg Nesterov 2906bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2907bbb68dfaSTejun Heo mark_work_canceling(work); 2908bbb68dfaSTejun Heo local_irq_restore(flags); 2909bbb68dfaSTejun Heo 2910606a5020STejun Heo flush_work(work); 29117a22ad75STejun Heo clear_work_data(work); 29121f1f642eSOleg Nesterov return ret; 29131f1f642eSOleg Nesterov } 29141f1f642eSOleg Nesterov 29156e84d644SOleg Nesterov /** 2916401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2917401a8d04STejun Heo * @work: the work to cancel 29186e84d644SOleg Nesterov * 2919401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2920401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2921401a8d04STejun Heo * another workqueue. On return from this function, @work is 2922401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29231f1f642eSOleg Nesterov * 2924401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2925401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29266e84d644SOleg Nesterov * 2927401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29286e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2929401a8d04STejun Heo * 2930401a8d04STejun Heo * RETURNS: 2931401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29326e84d644SOleg Nesterov */ 2933401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29346e84d644SOleg Nesterov { 293536e227d2STejun Heo return __cancel_work_timer(work, false); 2936b89deed3SOleg Nesterov } 293728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2938b89deed3SOleg Nesterov 29396e84d644SOleg Nesterov /** 2940401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2941401a8d04STejun Heo * @dwork: the delayed work to flush 29426e84d644SOleg Nesterov * 2943401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2944401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2945401a8d04STejun Heo * considers the last queueing instance of @dwork. 29461f1f642eSOleg Nesterov * 2947401a8d04STejun Heo * RETURNS: 2948401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2949401a8d04STejun Heo * %false if it was already idle. 29506e84d644SOleg Nesterov */ 2951401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2952401a8d04STejun Heo { 29538930cabaSTejun Heo local_irq_disable(); 2954401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 295560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29568930cabaSTejun Heo local_irq_enable(); 2957401a8d04STejun Heo return flush_work(&dwork->work); 2958401a8d04STejun Heo } 2959401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2960401a8d04STejun Heo 2961401a8d04STejun Heo /** 296257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 296357b30ae7STejun Heo * @dwork: delayed_work to cancel 296409383498STejun Heo * 296557b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 296657b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 296757b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 296857b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 296957b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 297009383498STejun Heo * 297157b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 297209383498STejun Heo */ 297357b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 297409383498STejun Heo { 297557b30ae7STejun Heo unsigned long flags; 297657b30ae7STejun Heo int ret; 297757b30ae7STejun Heo 297857b30ae7STejun Heo do { 297957b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 298057b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 298157b30ae7STejun Heo 298257b30ae7STejun Heo if (unlikely(ret < 0)) 298357b30ae7STejun Heo return false; 298457b30ae7STejun Heo 29857c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29867c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 298757b30ae7STejun Heo local_irq_restore(flags); 2988c0158ca6SDan Magenheimer return ret; 298909383498STejun Heo } 299057b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 299109383498STejun Heo 299209383498STejun Heo /** 2993401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2994401a8d04STejun Heo * @dwork: the delayed work cancel 2995401a8d04STejun Heo * 2996401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2997401a8d04STejun Heo * 2998401a8d04STejun Heo * RETURNS: 2999401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3000401a8d04STejun Heo */ 3001401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30026e84d644SOleg Nesterov { 300336e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30046e84d644SOleg Nesterov } 3005f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30061da177e4SLinus Torvalds 30070fcb78c2SRolf Eike Beer /** 300831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3009b6136773SAndrew Morton * @func: the function to call 3010b6136773SAndrew Morton * 301131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 301231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3013b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 301431ddd871STejun Heo * 301531ddd871STejun Heo * RETURNS: 301631ddd871STejun Heo * 0 on success, -errno on failure. 3017b6136773SAndrew Morton */ 301865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 301915316ba8SChristoph Lameter { 302015316ba8SChristoph Lameter int cpu; 302138f51568SNamhyung Kim struct work_struct __percpu *works; 302215316ba8SChristoph Lameter 3023b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3024b6136773SAndrew Morton if (!works) 302515316ba8SChristoph Lameter return -ENOMEM; 3026b6136773SAndrew Morton 302795402b38SGautham R Shenoy get_online_cpus(); 302893981800STejun Heo 302915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30309bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30319bfb1839SIngo Molnar 30329bfb1839SIngo Molnar INIT_WORK(work, func); 30338de6d308SOleg Nesterov schedule_work_on(cpu, work); 303415316ba8SChristoph Lameter } 303593981800STejun Heo 303693981800STejun Heo for_each_online_cpu(cpu) 30378616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 303893981800STejun Heo 303995402b38SGautham R Shenoy put_online_cpus(); 3040b6136773SAndrew Morton free_percpu(works); 304115316ba8SChristoph Lameter return 0; 304215316ba8SChristoph Lameter } 304315316ba8SChristoph Lameter 3044eef6a7d5SAlan Stern /** 3045eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3046eef6a7d5SAlan Stern * 3047eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3048eef6a7d5SAlan Stern * completion. 3049eef6a7d5SAlan Stern * 3050eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3051eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3052eef6a7d5SAlan Stern * will lead to deadlock: 3053eef6a7d5SAlan Stern * 3054eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3055eef6a7d5SAlan Stern * a lock held by your code or its caller. 3056eef6a7d5SAlan Stern * 3057eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3058eef6a7d5SAlan Stern * 3059eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3060eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3061eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3062eef6a7d5SAlan Stern * 3063eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3064eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3065eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3066eef6a7d5SAlan Stern * cancel_work_sync() instead. 3067eef6a7d5SAlan Stern */ 30681da177e4SLinus Torvalds void flush_scheduled_work(void) 30691da177e4SLinus Torvalds { 3070d320c038STejun Heo flush_workqueue(system_wq); 30711da177e4SLinus Torvalds } 3072ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30731da177e4SLinus Torvalds 30741da177e4SLinus Torvalds /** 30751fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30761fa44ecaSJames Bottomley * @fn: the function to execute 30771fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30781fa44ecaSJames Bottomley * be available when the work executes) 30791fa44ecaSJames Bottomley * 30801fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30811fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30821fa44ecaSJames Bottomley * 30831fa44ecaSJames Bottomley * Returns: 0 - function was executed 30841fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30851fa44ecaSJames Bottomley */ 308665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30871fa44ecaSJames Bottomley { 30881fa44ecaSJames Bottomley if (!in_interrupt()) { 308965f27f38SDavid Howells fn(&ew->work); 30901fa44ecaSJames Bottomley return 0; 30911fa44ecaSJames Bottomley } 30921fa44ecaSJames Bottomley 309365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30941fa44ecaSJames Bottomley schedule_work(&ew->work); 30951fa44ecaSJames Bottomley 30961fa44ecaSJames Bottomley return 1; 30971fa44ecaSJames Bottomley } 30981fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30991fa44ecaSJames Bottomley 3100226223abSTejun Heo #ifdef CONFIG_SYSFS 3101226223abSTejun Heo /* 3102226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3103226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3104226223abSTejun Heo * following attributes. 3105226223abSTejun Heo * 3106226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3107226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3108226223abSTejun Heo * 3109226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3110226223abSTejun Heo * 3111226223abSTejun Heo * id RO int : the associated pool ID 3112226223abSTejun Heo * nice RW int : nice value of the workers 3113226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3114226223abSTejun Heo */ 3115226223abSTejun Heo struct wq_device { 3116226223abSTejun Heo struct workqueue_struct *wq; 3117226223abSTejun Heo struct device dev; 3118226223abSTejun Heo }; 3119226223abSTejun Heo 3120226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3121226223abSTejun Heo { 3122226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3123226223abSTejun Heo 3124226223abSTejun Heo return wq_dev->wq; 3125226223abSTejun Heo } 3126226223abSTejun Heo 3127226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3128226223abSTejun Heo struct device_attribute *attr, char *buf) 3129226223abSTejun Heo { 3130226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3131226223abSTejun Heo 3132226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3133226223abSTejun Heo } 3134226223abSTejun Heo 3135226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3136226223abSTejun Heo struct device_attribute *attr, char *buf) 3137226223abSTejun Heo { 3138226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3139226223abSTejun Heo 3140226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3141226223abSTejun Heo } 3142226223abSTejun Heo 3143226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3144226223abSTejun Heo struct device_attribute *attr, 3145226223abSTejun Heo const char *buf, size_t count) 3146226223abSTejun Heo { 3147226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3148226223abSTejun Heo int val; 3149226223abSTejun Heo 3150226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3151226223abSTejun Heo return -EINVAL; 3152226223abSTejun Heo 3153226223abSTejun Heo workqueue_set_max_active(wq, val); 3154226223abSTejun Heo return count; 3155226223abSTejun Heo } 3156226223abSTejun Heo 3157226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3158226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3159226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3160226223abSTejun Heo __ATTR_NULL, 3161226223abSTejun Heo }; 3162226223abSTejun Heo 3163226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3164226223abSTejun Heo struct device_attribute *attr, char *buf) 3165226223abSTejun Heo { 3166226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3167226223abSTejun Heo struct worker_pool *pool; 3168226223abSTejun Heo int written; 3169226223abSTejun Heo 3170226223abSTejun Heo rcu_read_lock_sched(); 3171226223abSTejun Heo pool = first_pwq(wq)->pool; 3172226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3173226223abSTejun Heo rcu_read_unlock_sched(); 3174226223abSTejun Heo 3175226223abSTejun Heo return written; 3176226223abSTejun Heo } 3177226223abSTejun Heo 3178226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3179226223abSTejun Heo char *buf) 3180226223abSTejun Heo { 3181226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3182226223abSTejun Heo int written; 3183226223abSTejun Heo 3184226223abSTejun Heo rcu_read_lock_sched(); 3185226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3186226223abSTejun Heo first_pwq(wq)->pool->attrs->nice); 3187226223abSTejun Heo rcu_read_unlock_sched(); 3188226223abSTejun Heo 3189226223abSTejun Heo return written; 3190226223abSTejun Heo } 3191226223abSTejun Heo 3192226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3193226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3194226223abSTejun Heo { 3195226223abSTejun Heo struct workqueue_attrs *attrs; 3196226223abSTejun Heo 3197226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3198226223abSTejun Heo if (!attrs) 3199226223abSTejun Heo return NULL; 3200226223abSTejun Heo 3201226223abSTejun Heo rcu_read_lock_sched(); 3202226223abSTejun Heo copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs); 3203226223abSTejun Heo rcu_read_unlock_sched(); 3204226223abSTejun Heo return attrs; 3205226223abSTejun Heo } 3206226223abSTejun Heo 3207226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3208226223abSTejun Heo const char *buf, size_t count) 3209226223abSTejun Heo { 3210226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3211226223abSTejun Heo struct workqueue_attrs *attrs; 3212226223abSTejun Heo int ret; 3213226223abSTejun Heo 3214226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3215226223abSTejun Heo if (!attrs) 3216226223abSTejun Heo return -ENOMEM; 3217226223abSTejun Heo 3218226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3219226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3220226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3221226223abSTejun Heo else 3222226223abSTejun Heo ret = -EINVAL; 3223226223abSTejun Heo 3224226223abSTejun Heo free_workqueue_attrs(attrs); 3225226223abSTejun Heo return ret ?: count; 3226226223abSTejun Heo } 3227226223abSTejun Heo 3228226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3229226223abSTejun Heo struct device_attribute *attr, char *buf) 3230226223abSTejun Heo { 3231226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3232226223abSTejun Heo int written; 3233226223abSTejun Heo 3234226223abSTejun Heo rcu_read_lock_sched(); 3235226223abSTejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, 3236226223abSTejun Heo first_pwq(wq)->pool->attrs->cpumask); 3237226223abSTejun Heo rcu_read_unlock_sched(); 3238226223abSTejun Heo 3239226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3240226223abSTejun Heo return written; 3241226223abSTejun Heo } 3242226223abSTejun Heo 3243226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3244226223abSTejun Heo struct device_attribute *attr, 3245226223abSTejun Heo const char *buf, size_t count) 3246226223abSTejun Heo { 3247226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3248226223abSTejun Heo struct workqueue_attrs *attrs; 3249226223abSTejun Heo int ret; 3250226223abSTejun Heo 3251226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3252226223abSTejun Heo if (!attrs) 3253226223abSTejun Heo return -ENOMEM; 3254226223abSTejun Heo 3255226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3256226223abSTejun Heo if (!ret) 3257226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3258226223abSTejun Heo 3259226223abSTejun Heo free_workqueue_attrs(attrs); 3260226223abSTejun Heo return ret ?: count; 3261226223abSTejun Heo } 3262226223abSTejun Heo 3263226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3264226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3265226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3266226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3267226223abSTejun Heo __ATTR_NULL, 3268226223abSTejun Heo }; 3269226223abSTejun Heo 3270226223abSTejun Heo static struct bus_type wq_subsys = { 3271226223abSTejun Heo .name = "workqueue", 3272226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3273226223abSTejun Heo }; 3274226223abSTejun Heo 3275226223abSTejun Heo static int __init wq_sysfs_init(void) 3276226223abSTejun Heo { 3277226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3278226223abSTejun Heo } 3279226223abSTejun Heo core_initcall(wq_sysfs_init); 3280226223abSTejun Heo 3281226223abSTejun Heo static void wq_device_release(struct device *dev) 3282226223abSTejun Heo { 3283226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3284226223abSTejun Heo 3285226223abSTejun Heo kfree(wq_dev); 3286226223abSTejun Heo } 3287226223abSTejun Heo 3288226223abSTejun Heo /** 3289226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3290226223abSTejun Heo * @wq: the workqueue to register 3291226223abSTejun Heo * 3292226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3293226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3294226223abSTejun Heo * which is the preferred method. 3295226223abSTejun Heo * 3296226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3297226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3298226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3299226223abSTejun Heo * attributes. 3300226223abSTejun Heo * 3301226223abSTejun Heo * Returns 0 on success, -errno on failure. 3302226223abSTejun Heo */ 3303226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3304226223abSTejun Heo { 3305226223abSTejun Heo struct wq_device *wq_dev; 3306226223abSTejun Heo int ret; 3307226223abSTejun Heo 3308226223abSTejun Heo /* 3309226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3310226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3311226223abSTejun Heo * workqueues. 3312226223abSTejun Heo */ 3313226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3314226223abSTejun Heo return -EINVAL; 3315226223abSTejun Heo 3316226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3317226223abSTejun Heo if (!wq_dev) 3318226223abSTejun Heo return -ENOMEM; 3319226223abSTejun Heo 3320226223abSTejun Heo wq_dev->wq = wq; 3321226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3322226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3323226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3324226223abSTejun Heo 3325226223abSTejun Heo /* 3326226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3327226223abSTejun Heo * everything is ready. 3328226223abSTejun Heo */ 3329226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3330226223abSTejun Heo 3331226223abSTejun Heo ret = device_register(&wq_dev->dev); 3332226223abSTejun Heo if (ret) { 3333226223abSTejun Heo kfree(wq_dev); 3334226223abSTejun Heo wq->wq_dev = NULL; 3335226223abSTejun Heo return ret; 3336226223abSTejun Heo } 3337226223abSTejun Heo 3338226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3339226223abSTejun Heo struct device_attribute *attr; 3340226223abSTejun Heo 3341226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3342226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3343226223abSTejun Heo if (ret) { 3344226223abSTejun Heo device_unregister(&wq_dev->dev); 3345226223abSTejun Heo wq->wq_dev = NULL; 3346226223abSTejun Heo return ret; 3347226223abSTejun Heo } 3348226223abSTejun Heo } 3349226223abSTejun Heo } 3350226223abSTejun Heo 3351226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3352226223abSTejun Heo return 0; 3353226223abSTejun Heo } 3354226223abSTejun Heo 3355226223abSTejun Heo /** 3356226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3357226223abSTejun Heo * @wq: the workqueue to unregister 3358226223abSTejun Heo * 3359226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3360226223abSTejun Heo */ 3361226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3362226223abSTejun Heo { 3363226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3364226223abSTejun Heo 3365226223abSTejun Heo if (!wq->wq_dev) 3366226223abSTejun Heo return; 3367226223abSTejun Heo 3368226223abSTejun Heo wq->wq_dev = NULL; 3369226223abSTejun Heo device_unregister(&wq_dev->dev); 3370226223abSTejun Heo } 3371226223abSTejun Heo #else /* CONFIG_SYSFS */ 3372226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3373226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3374226223abSTejun Heo 33757a4e344cSTejun Heo /** 33767a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33777a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33787a4e344cSTejun Heo * 33797a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33807a4e344cSTejun Heo */ 33817a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33827a4e344cSTejun Heo { 33837a4e344cSTejun Heo if (attrs) { 33847a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33857a4e344cSTejun Heo kfree(attrs); 33867a4e344cSTejun Heo } 33877a4e344cSTejun Heo } 33887a4e344cSTejun Heo 33897a4e344cSTejun Heo /** 33907a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33917a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33927a4e344cSTejun Heo * 33937a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33947a4e344cSTejun Heo * return it. Returns NULL on failure. 33957a4e344cSTejun Heo */ 33967a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33977a4e344cSTejun Heo { 33987a4e344cSTejun Heo struct workqueue_attrs *attrs; 33997a4e344cSTejun Heo 34007a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 34017a4e344cSTejun Heo if (!attrs) 34027a4e344cSTejun Heo goto fail; 34037a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 34047a4e344cSTejun Heo goto fail; 34057a4e344cSTejun Heo 34067a4e344cSTejun Heo cpumask_setall(attrs->cpumask); 34077a4e344cSTejun Heo return attrs; 34087a4e344cSTejun Heo fail: 34097a4e344cSTejun Heo free_workqueue_attrs(attrs); 34107a4e344cSTejun Heo return NULL; 34117a4e344cSTejun Heo } 34127a4e344cSTejun Heo 341329c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 341429c91e99STejun Heo const struct workqueue_attrs *from) 341529c91e99STejun Heo { 341629c91e99STejun Heo to->nice = from->nice; 341729c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 341829c91e99STejun Heo } 341929c91e99STejun Heo 342029c91e99STejun Heo /* 342129c91e99STejun Heo * Hacky implementation of jhash of bitmaps which only considers the 342229c91e99STejun Heo * specified number of bits. We probably want a proper implementation in 342329c91e99STejun Heo * include/linux/jhash.h. 342429c91e99STejun Heo */ 342529c91e99STejun Heo static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash) 342629c91e99STejun Heo { 342729c91e99STejun Heo int nr_longs = bits / BITS_PER_LONG; 342829c91e99STejun Heo int nr_leftover = bits % BITS_PER_LONG; 342929c91e99STejun Heo unsigned long leftover = 0; 343029c91e99STejun Heo 343129c91e99STejun Heo if (nr_longs) 343229c91e99STejun Heo hash = jhash(bitmap, nr_longs * sizeof(long), hash); 343329c91e99STejun Heo if (nr_leftover) { 343429c91e99STejun Heo bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover); 343529c91e99STejun Heo hash = jhash(&leftover, sizeof(long), hash); 343629c91e99STejun Heo } 343729c91e99STejun Heo return hash; 343829c91e99STejun Heo } 343929c91e99STejun Heo 344029c91e99STejun Heo /* hash value of the content of @attr */ 344129c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 344229c91e99STejun Heo { 344329c91e99STejun Heo u32 hash = 0; 344429c91e99STejun Heo 344529c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 344629c91e99STejun Heo hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash); 344729c91e99STejun Heo return hash; 344829c91e99STejun Heo } 344929c91e99STejun Heo 345029c91e99STejun Heo /* content equality test */ 345129c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 345229c91e99STejun Heo const struct workqueue_attrs *b) 345329c91e99STejun Heo { 345429c91e99STejun Heo if (a->nice != b->nice) 345529c91e99STejun Heo return false; 345629c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 345729c91e99STejun Heo return false; 345829c91e99STejun Heo return true; 345929c91e99STejun Heo } 346029c91e99STejun Heo 34617a4e344cSTejun Heo /** 34627a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34637a4e344cSTejun Heo * @pool: worker_pool to initialize 34647a4e344cSTejun Heo * 34657a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 346629c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 346729c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 346829c91e99STejun Heo * on @pool safely to release it. 34697a4e344cSTejun Heo */ 34707a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34714e1a1f9aSTejun Heo { 34724e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 347329c91e99STejun Heo pool->id = -1; 347429c91e99STejun Heo pool->cpu = -1; 34754e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 34764e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34774e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34784e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34794e1a1f9aSTejun Heo 34804e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 34814e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 34824e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 34834e1a1f9aSTejun Heo 34844e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 34854e1a1f9aSTejun Heo (unsigned long)pool); 34864e1a1f9aSTejun Heo 34874e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3488bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 34894e1a1f9aSTejun Heo ida_init(&pool->worker_ida); 34907a4e344cSTejun Heo 349129c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 349229c91e99STejun Heo pool->refcnt = 1; 349329c91e99STejun Heo 349429c91e99STejun Heo /* shouldn't fail above this point */ 34957a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34967a4e344cSTejun Heo if (!pool->attrs) 34977a4e344cSTejun Heo return -ENOMEM; 34987a4e344cSTejun Heo return 0; 34994e1a1f9aSTejun Heo } 35004e1a1f9aSTejun Heo 350129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 350229c91e99STejun Heo { 350329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 350429c91e99STejun Heo 350529c91e99STejun Heo ida_destroy(&pool->worker_ida); 350629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 350729c91e99STejun Heo kfree(pool); 350829c91e99STejun Heo } 350929c91e99STejun Heo 351029c91e99STejun Heo /** 351129c91e99STejun Heo * put_unbound_pool - put a worker_pool 351229c91e99STejun Heo * @pool: worker_pool to put 351329c91e99STejun Heo * 351429c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3515c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3516c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3517c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 351829c91e99STejun Heo */ 351929c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 352029c91e99STejun Heo { 352129c91e99STejun Heo struct worker *worker; 352229c91e99STejun Heo 35235bcab335STejun Heo mutex_lock(&wq_mutex); 352429c91e99STejun Heo if (--pool->refcnt) { 35255bcab335STejun Heo mutex_unlock(&wq_mutex); 352629c91e99STejun Heo return; 352729c91e99STejun Heo } 352829c91e99STejun Heo 352929c91e99STejun Heo /* sanity checks */ 353029c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 353129c91e99STejun Heo WARN_ON(!list_empty(&pool->worklist))) { 35325bcab335STejun Heo mutex_unlock(&wq_mutex); 353329c91e99STejun Heo return; 353429c91e99STejun Heo } 353529c91e99STejun Heo 353629c91e99STejun Heo /* release id and unhash */ 353729c91e99STejun Heo if (pool->id >= 0) 353829c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 353929c91e99STejun Heo hash_del(&pool->hash_node); 354029c91e99STejun Heo 35415bcab335STejun Heo mutex_unlock(&wq_mutex); 354229c91e99STejun Heo 3543c5aa87bbSTejun Heo /* 3544c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3545c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3546c5aa87bbSTejun Heo * manager_mutex. 3547c5aa87bbSTejun Heo */ 354829c91e99STejun Heo mutex_lock(&pool->manager_arb); 3549cd549687STejun Heo mutex_lock(&pool->manager_mutex); 355029c91e99STejun Heo spin_lock_irq(&pool->lock); 355129c91e99STejun Heo 355229c91e99STejun Heo while ((worker = first_worker(pool))) 355329c91e99STejun Heo destroy_worker(worker); 355429c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 355529c91e99STejun Heo 355629c91e99STejun Heo spin_unlock_irq(&pool->lock); 3557cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 355829c91e99STejun Heo mutex_unlock(&pool->manager_arb); 355929c91e99STejun Heo 356029c91e99STejun Heo /* shut down the timers */ 356129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 356229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 356329c91e99STejun Heo 356429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 356529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 356629c91e99STejun Heo } 356729c91e99STejun Heo 356829c91e99STejun Heo /** 356929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 357029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 357129c91e99STejun Heo * 357229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 357329c91e99STejun Heo * reference count and return it. If there already is a matching 357429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 357529c91e99STejun Heo * create a new one. On failure, returns NULL. 357629c91e99STejun Heo */ 357729c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 357829c91e99STejun Heo { 357929c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 358029c91e99STejun Heo struct worker_pool *pool; 358129c91e99STejun Heo 35825bcab335STejun Heo mutex_lock(&wq_mutex); 358329c91e99STejun Heo 358429c91e99STejun Heo /* do we already have a matching pool? */ 358529c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 358629c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 358729c91e99STejun Heo pool->refcnt++; 358829c91e99STejun Heo goto out_unlock; 358929c91e99STejun Heo } 359029c91e99STejun Heo } 359129c91e99STejun Heo 359229c91e99STejun Heo /* nope, create a new one */ 359329c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 359429c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 359529c91e99STejun Heo goto fail; 359629c91e99STejun Heo 35978864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 359829c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 359929c91e99STejun Heo 360029c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 360129c91e99STejun Heo goto fail; 360229c91e99STejun Heo 360329c91e99STejun Heo /* create and start the initial worker */ 3604ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 360529c91e99STejun Heo goto fail; 360629c91e99STejun Heo 360729c91e99STejun Heo /* install */ 360829c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 360929c91e99STejun Heo out_unlock: 36105bcab335STejun Heo mutex_unlock(&wq_mutex); 361129c91e99STejun Heo return pool; 361229c91e99STejun Heo fail: 36135bcab335STejun Heo mutex_unlock(&wq_mutex); 361429c91e99STejun Heo if (pool) 361529c91e99STejun Heo put_unbound_pool(pool); 361629c91e99STejun Heo return NULL; 361729c91e99STejun Heo } 361829c91e99STejun Heo 36198864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36208864b4e5STejun Heo { 36218864b4e5STejun Heo kmem_cache_free(pwq_cache, 36228864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36238864b4e5STejun Heo } 36248864b4e5STejun Heo 36258864b4e5STejun Heo /* 36268864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36278864b4e5STejun Heo * and needs to be destroyed. 36288864b4e5STejun Heo */ 36298864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36308864b4e5STejun Heo { 36318864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36328864b4e5STejun Heo unbound_release_work); 36338864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36348864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 36358864b4e5STejun Heo 36368864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36378864b4e5STejun Heo return; 36388864b4e5STejun Heo 363975ccf595STejun Heo /* 364075ccf595STejun Heo * Unlink @pwq. Synchronization against flush_mutex isn't strictly 364175ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 364275ccf595STejun Heo * and consistent with the linking path. 364375ccf595STejun Heo */ 364475ccf595STejun Heo mutex_lock(&wq->flush_mutex); 3645794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 36468864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3647794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 364875ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 36498864b4e5STejun Heo 36508864b4e5STejun Heo put_unbound_pool(pool); 36518864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 36528864b4e5STejun Heo 36538864b4e5STejun Heo /* 36548864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 36558864b4e5STejun Heo * is gonna access it anymore. Free it. 36568864b4e5STejun Heo */ 36578864b4e5STejun Heo if (list_empty(&wq->pwqs)) 36588864b4e5STejun Heo kfree(wq); 36598864b4e5STejun Heo } 36608864b4e5STejun Heo 36610fbd95aaSTejun Heo /** 3662699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36630fbd95aaSTejun Heo * @pwq: target pool_workqueue 36640fbd95aaSTejun Heo * 3665699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3666699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3667699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36680fbd95aaSTejun Heo */ 3669699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36700fbd95aaSTejun Heo { 3671699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3672699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3673699ce097STejun Heo 3674699ce097STejun Heo /* for @wq->saved_max_active */ 3675794b18bcSTejun Heo lockdep_assert_held(&pwq_lock); 3676699ce097STejun Heo 3677699ce097STejun Heo /* fast exit for non-freezable wqs */ 3678699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3679699ce097STejun Heo return; 3680699ce097STejun Heo 3681699ce097STejun Heo spin_lock(&pwq->pool->lock); 3682699ce097STejun Heo 3683699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3684699ce097STejun Heo pwq->max_active = wq->saved_max_active; 36850fbd95aaSTejun Heo 36860fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36870fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36880fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3689699ce097STejun Heo } else { 3690699ce097STejun Heo pwq->max_active = 0; 3691699ce097STejun Heo } 3692699ce097STejun Heo 3693699ce097STejun Heo spin_unlock(&pwq->pool->lock); 36940fbd95aaSTejun Heo } 36950fbd95aaSTejun Heo 3696d2c1d404STejun Heo static void init_and_link_pwq(struct pool_workqueue *pwq, 3697d2c1d404STejun Heo struct workqueue_struct *wq, 36989e8cd2f5STejun Heo struct worker_pool *pool, 36999e8cd2f5STejun Heo struct pool_workqueue **p_last_pwq) 3700d2c1d404STejun Heo { 3701d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3702d2c1d404STejun Heo 3703d2c1d404STejun Heo pwq->pool = pool; 3704d2c1d404STejun Heo pwq->wq = wq; 3705d2c1d404STejun Heo pwq->flush_color = -1; 37068864b4e5STejun Heo pwq->refcnt = 1; 3707d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3708d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37098864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3710d2c1d404STejun Heo 371175ccf595STejun Heo mutex_lock(&wq->flush_mutex); 3712794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 371375ccf595STejun Heo 3714983ca25eSTejun Heo /* 3715983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 3716983ca25eSTejun Heo * flush_mutex to avoid confusing flush_workqueue(). 3717983ca25eSTejun Heo */ 37189e8cd2f5STejun Heo if (p_last_pwq) 37199e8cd2f5STejun Heo *p_last_pwq = first_pwq(wq); 372075ccf595STejun Heo pwq->work_color = wq->work_color; 3721983ca25eSTejun Heo 3722983ca25eSTejun Heo /* sync max_active to the current setting */ 3723983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3724983ca25eSTejun Heo 3725983ca25eSTejun Heo /* link in @pwq */ 37269e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 372775ccf595STejun Heo 3728794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 372975ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 3730d2c1d404STejun Heo } 3731d2c1d404STejun Heo 37329e8cd2f5STejun Heo /** 37339e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37349e8cd2f5STejun Heo * @wq: the target workqueue 37359e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37369e8cd2f5STejun Heo * 37379e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 37389e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 37399e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 37409e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 37419e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 37429e8cd2f5STejun Heo * 37439e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 37449e8cd2f5STejun Heo * failure. 37459e8cd2f5STejun Heo */ 37469e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 37479e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 37489e8cd2f5STejun Heo { 37499e8cd2f5STejun Heo struct pool_workqueue *pwq, *last_pwq; 37509e8cd2f5STejun Heo struct worker_pool *pool; 37519e8cd2f5STejun Heo 37528719dceaSTejun Heo /* only unbound workqueues can change attributes */ 37539e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 37549e8cd2f5STejun Heo return -EINVAL; 37559e8cd2f5STejun Heo 37568719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 37578719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 37588719dceaSTejun Heo return -EINVAL; 37598719dceaSTejun Heo 37609e8cd2f5STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 37619e8cd2f5STejun Heo if (!pwq) 37629e8cd2f5STejun Heo return -ENOMEM; 37639e8cd2f5STejun Heo 37649e8cd2f5STejun Heo pool = get_unbound_pool(attrs); 37659e8cd2f5STejun Heo if (!pool) { 37669e8cd2f5STejun Heo kmem_cache_free(pwq_cache, pwq); 37679e8cd2f5STejun Heo return -ENOMEM; 37689e8cd2f5STejun Heo } 37699e8cd2f5STejun Heo 37709e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, pool, &last_pwq); 37719e8cd2f5STejun Heo if (last_pwq) { 37729e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 37739e8cd2f5STejun Heo put_pwq(last_pwq); 37749e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 37759e8cd2f5STejun Heo } 37769e8cd2f5STejun Heo 37779e8cd2f5STejun Heo return 0; 37789e8cd2f5STejun Heo } 37799e8cd2f5STejun Heo 378030cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37811da177e4SLinus Torvalds { 378249e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 378330cdf249STejun Heo int cpu; 3784e1d8aa9fSFrederic Weisbecker 378530cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3786420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3787420c0ddbSTejun Heo if (!wq->cpu_pwqs) 378830cdf249STejun Heo return -ENOMEM; 378930cdf249STejun Heo 379030cdf249STejun Heo for_each_possible_cpu(cpu) { 37917fb98ea7STejun Heo struct pool_workqueue *pwq = 37927fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37937a62c2c8STejun Heo struct worker_pool *cpu_pools = 3794f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 379530cdf249STejun Heo 37969e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL); 379730cdf249STejun Heo } 379830cdf249STejun Heo return 0; 37999e8cd2f5STejun Heo } else { 38009e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 38019e8cd2f5STejun Heo } 38020f900049STejun Heo } 38030f900049STejun Heo 3804f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3805f3421797STejun Heo const char *name) 3806b71ab8c2STejun Heo { 3807f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3808f3421797STejun Heo 3809f3421797STejun Heo if (max_active < 1 || max_active > lim) 3810044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3811f3421797STejun Heo max_active, name, 1, lim); 3812b71ab8c2STejun Heo 3813f3421797STejun Heo return clamp_val(max_active, 1, lim); 3814b71ab8c2STejun Heo } 3815b71ab8c2STejun Heo 3816b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 381797e37d7bSTejun Heo unsigned int flags, 38181e19ffc6STejun Heo int max_active, 3819eb13ba87SJohannes Berg struct lock_class_key *key, 3820b196be89STejun Heo const char *lock_name, ...) 38213af24433SOleg Nesterov { 3822b196be89STejun Heo va_list args, args1; 38233af24433SOleg Nesterov struct workqueue_struct *wq; 382449e3cf44STejun Heo struct pool_workqueue *pwq; 3825b196be89STejun Heo size_t namelen; 3826b196be89STejun Heo 3827b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3828b196be89STejun Heo va_start(args, lock_name); 3829b196be89STejun Heo va_copy(args1, args); 3830b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3831b196be89STejun Heo 3832b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3833b196be89STejun Heo if (!wq) 3834d2c1d404STejun Heo return NULL; 3835b196be89STejun Heo 3836b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3837b196be89STejun Heo va_end(args); 3838b196be89STejun Heo va_end(args1); 38393af24433SOleg Nesterov 3840d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3841b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 38423af24433SOleg Nesterov 3843b196be89STejun Heo /* init wq */ 384497e37d7bSTejun Heo wq->flags = flags; 3845a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 384673f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 3847112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 384830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 384973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 385073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3851493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 38523af24433SOleg Nesterov 3853eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3854cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 38553af24433SOleg Nesterov 385630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3857d2c1d404STejun Heo goto err_free_wq; 38581537663fSTejun Heo 3859493008a8STejun Heo /* 3860493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3861493008a8STejun Heo * have a rescuer to guarantee forward progress. 3862493008a8STejun Heo */ 3863493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3864e22bee78STejun Heo struct worker *rescuer; 3865e22bee78STejun Heo 3866d2c1d404STejun Heo rescuer = alloc_worker(); 3867e22bee78STejun Heo if (!rescuer) 3868d2c1d404STejun Heo goto err_destroy; 3869e22bee78STejun Heo 3870111c225aSTejun Heo rescuer->rescue_wq = wq; 3871111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3872b196be89STejun Heo wq->name); 3873d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3874d2c1d404STejun Heo kfree(rescuer); 3875d2c1d404STejun Heo goto err_destroy; 3876d2c1d404STejun Heo } 3877e22bee78STejun Heo 3878d2c1d404STejun Heo wq->rescuer = rescuer; 3879e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3880e22bee78STejun Heo wake_up_process(rescuer->task); 38813af24433SOleg Nesterov } 38821537663fSTejun Heo 3883226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3884226223abSTejun Heo goto err_destroy; 3885226223abSTejun Heo 38863af24433SOleg Nesterov /* 38875bcab335STejun Heo * wq_mutex protects global freeze state and workqueues list. Grab 38885bcab335STejun Heo * it, adjust max_active and add the new @wq to workqueues list. 38893af24433SOleg Nesterov */ 38905bcab335STejun Heo mutex_lock(&wq_mutex); 3891a0a1a5fdSTejun Heo 3892794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 389349e3cf44STejun Heo for_each_pwq(pwq, wq) 3894699ce097STejun Heo pwq_adjust_max_active(pwq); 3895794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 3896a0a1a5fdSTejun Heo 38973af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3898a0a1a5fdSTejun Heo 38995bcab335STejun Heo mutex_unlock(&wq_mutex); 39003af24433SOleg Nesterov 39013af24433SOleg Nesterov return wq; 3902d2c1d404STejun Heo 3903d2c1d404STejun Heo err_free_wq: 39044690c4abSTejun Heo kfree(wq); 3905d2c1d404STejun Heo return NULL; 3906d2c1d404STejun Heo err_destroy: 3907d2c1d404STejun Heo destroy_workqueue(wq); 39084690c4abSTejun Heo return NULL; 39091da177e4SLinus Torvalds } 3910d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 39111da177e4SLinus Torvalds 39123af24433SOleg Nesterov /** 39133af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 39143af24433SOleg Nesterov * @wq: target workqueue 39153af24433SOleg Nesterov * 39163af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 39173af24433SOleg Nesterov */ 39183af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 39193af24433SOleg Nesterov { 392049e3cf44STejun Heo struct pool_workqueue *pwq; 39213af24433SOleg Nesterov 39229c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 39239c5a2ba7STejun Heo drain_workqueue(wq); 3924c8efcc25STejun Heo 39256183c009STejun Heo /* sanity checks */ 3926794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 392749e3cf44STejun Heo for_each_pwq(pwq, wq) { 39286183c009STejun Heo int i; 39296183c009STejun Heo 393076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 393176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3932794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 39336183c009STejun Heo return; 393476af4d93STejun Heo } 393576af4d93STejun Heo } 393676af4d93STejun Heo 39378864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 39388864b4e5STejun Heo WARN_ON(pwq->nr_active) || 393976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3940794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 39416183c009STejun Heo return; 39426183c009STejun Heo } 394376af4d93STejun Heo } 3944794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 39456183c009STejun Heo 3946a0a1a5fdSTejun Heo /* 3947a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3948a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3949a0a1a5fdSTejun Heo */ 39505bcab335STejun Heo mutex_lock(&wq_mutex); 3951d2c1d404STejun Heo list_del_init(&wq->list); 39525bcab335STejun Heo mutex_unlock(&wq_mutex); 39533af24433SOleg Nesterov 3954226223abSTejun Heo workqueue_sysfs_unregister(wq); 3955226223abSTejun Heo 3956493008a8STejun Heo if (wq->rescuer) { 3957e22bee78STejun Heo kthread_stop(wq->rescuer->task); 39588d9df9f0SXiaotian Feng kfree(wq->rescuer); 3959493008a8STejun Heo wq->rescuer = NULL; 3960e22bee78STejun Heo } 3961e22bee78STejun Heo 39628864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 396329c91e99STejun Heo /* 39648864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 39658864b4e5STejun Heo * free the pwqs and wq. 396629c91e99STejun Heo */ 39678864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 39688864b4e5STejun Heo kfree(wq); 39698864b4e5STejun Heo } else { 39708864b4e5STejun Heo /* 39718864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 39728864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 39738864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 39748864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 39758864b4e5STejun Heo * released. 39768864b4e5STejun Heo */ 397729c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 397829c91e99STejun Heo pwqs_node); 39798864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 39808864b4e5STejun Heo put_pwq(pwq); 39818864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 398229c91e99STejun Heo } 39833af24433SOleg Nesterov } 39843af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 39853af24433SOleg Nesterov 3986dcd989cbSTejun Heo /** 3987dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3988dcd989cbSTejun Heo * @wq: target workqueue 3989dcd989cbSTejun Heo * @max_active: new max_active value. 3990dcd989cbSTejun Heo * 3991dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3992dcd989cbSTejun Heo * 3993dcd989cbSTejun Heo * CONTEXT: 3994dcd989cbSTejun Heo * Don't call from IRQ context. 3995dcd989cbSTejun Heo */ 3996dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3997dcd989cbSTejun Heo { 399849e3cf44STejun Heo struct pool_workqueue *pwq; 3999dcd989cbSTejun Heo 40008719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 40018719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 40028719dceaSTejun Heo return; 40038719dceaSTejun Heo 4004f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4005dcd989cbSTejun Heo 4006794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 4007dcd989cbSTejun Heo 4008dcd989cbSTejun Heo wq->saved_max_active = max_active; 4009dcd989cbSTejun Heo 4010699ce097STejun Heo for_each_pwq(pwq, wq) 4011699ce097STejun Heo pwq_adjust_max_active(pwq); 4012dcd989cbSTejun Heo 4013794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 4014dcd989cbSTejun Heo } 4015dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4016dcd989cbSTejun Heo 4017dcd989cbSTejun Heo /** 4018e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4019e6267616STejun Heo * 4020e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4021e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4022e6267616STejun Heo */ 4023e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4024e6267616STejun Heo { 4025e6267616STejun Heo struct worker *worker = current_wq_worker(); 4026e6267616STejun Heo 4027e6267616STejun Heo return worker && worker == worker->current_pwq->wq->rescuer; 4028e6267616STejun Heo } 4029e6267616STejun Heo 4030e6267616STejun Heo /** 4031dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4032dcd989cbSTejun Heo * @cpu: CPU in question 4033dcd989cbSTejun Heo * @wq: target workqueue 4034dcd989cbSTejun Heo * 4035dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4036dcd989cbSTejun Heo * no synchronization around this function and the test result is 4037dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4038dcd989cbSTejun Heo * 4039dcd989cbSTejun Heo * RETURNS: 4040dcd989cbSTejun Heo * %true if congested, %false otherwise. 4041dcd989cbSTejun Heo */ 4042d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4043dcd989cbSTejun Heo { 40447fb98ea7STejun Heo struct pool_workqueue *pwq; 404576af4d93STejun Heo bool ret; 404676af4d93STejun Heo 404776af4d93STejun Heo preempt_disable(); 40487fb98ea7STejun Heo 40497fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 40507fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 40517fb98ea7STejun Heo else 40527fb98ea7STejun Heo pwq = first_pwq(wq); 4053dcd989cbSTejun Heo 405476af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 405576af4d93STejun Heo preempt_enable(); 405676af4d93STejun Heo 405776af4d93STejun Heo return ret; 4058dcd989cbSTejun Heo } 4059dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4060dcd989cbSTejun Heo 4061dcd989cbSTejun Heo /** 4062dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4063dcd989cbSTejun Heo * @work: the work to be tested 4064dcd989cbSTejun Heo * 4065dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4066dcd989cbSTejun Heo * synchronization around this function and the test result is 4067dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4068dcd989cbSTejun Heo * 4069dcd989cbSTejun Heo * RETURNS: 4070dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4071dcd989cbSTejun Heo */ 4072dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4073dcd989cbSTejun Heo { 4074fa1b54e6STejun Heo struct worker_pool *pool; 4075dcd989cbSTejun Heo unsigned long flags; 4076dcd989cbSTejun Heo unsigned int ret = 0; 4077dcd989cbSTejun Heo 4078dcd989cbSTejun Heo if (work_pending(work)) 4079dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4080038366c5SLai Jiangshan 4081fa1b54e6STejun Heo local_irq_save(flags); 4082fa1b54e6STejun Heo pool = get_work_pool(work); 4083038366c5SLai Jiangshan if (pool) { 4084fa1b54e6STejun Heo spin_lock(&pool->lock); 4085c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4086dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4087fa1b54e6STejun Heo spin_unlock(&pool->lock); 4088038366c5SLai Jiangshan } 4089fa1b54e6STejun Heo local_irq_restore(flags); 4090dcd989cbSTejun Heo 4091dcd989cbSTejun Heo return ret; 4092dcd989cbSTejun Heo } 4093dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4094dcd989cbSTejun Heo 4095db7bccf4STejun Heo /* 4096db7bccf4STejun Heo * CPU hotplug. 4097db7bccf4STejun Heo * 4098e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4099112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4100706026c2STejun Heo * pool which make migrating pending and scheduled works very 4101e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 410294cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4103e22bee78STejun Heo * blocked draining impractical. 4104e22bee78STejun Heo * 410524647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4106628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4107628c78e7STejun Heo * cpu comes back online. 4108db7bccf4STejun Heo */ 4109db7bccf4STejun Heo 4110706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4111db7bccf4STejun Heo { 411238db41d9STejun Heo int cpu = smp_processor_id(); 41134ce62e9eSTejun Heo struct worker_pool *pool; 4114db7bccf4STejun Heo struct worker *worker; 4115db7bccf4STejun Heo int i; 4116db7bccf4STejun Heo 4117f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 41186183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4119db7bccf4STejun Heo 4120bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 412194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4122e22bee78STejun Heo 4123f2d5a0eeSTejun Heo /* 4124bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 412594cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 412694cf58bbSTejun Heo * except for the ones which are still executing works from 412794cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 412894cf58bbSTejun Heo * this, they may become diasporas. 4129f2d5a0eeSTejun Heo */ 41304ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 4131403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4132db7bccf4STejun Heo 4133b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) 4134403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4135db7bccf4STejun Heo 413624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4137f2d5a0eeSTejun Heo 413894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4139bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 414094cf58bbSTejun Heo } 4141e22bee78STejun Heo 4142e22bee78STejun Heo /* 4143628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4144628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4145628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4146628c78e7STejun Heo */ 4147628c78e7STejun Heo schedule(); 4148628c78e7STejun Heo 4149628c78e7STejun Heo /* 4150628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4151628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 415238db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 415338db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 415438db41d9STejun Heo * pools which are served by workers tied to the CPU. 4155628c78e7STejun Heo * 4156628c78e7STejun Heo * On return from this function, the current worker would trigger 4157628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4158628c78e7STejun Heo * didn't already. 4159e22bee78STejun Heo */ 4160f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4161e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4162db7bccf4STejun Heo } 4163db7bccf4STejun Heo 41648db25e78STejun Heo /* 41658db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 41668db25e78STejun Heo * This will be registered high priority CPU notifier. 41678db25e78STejun Heo */ 41689fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 41691da177e4SLinus Torvalds unsigned long action, 41701da177e4SLinus Torvalds void *hcpu) 41711da177e4SLinus Torvalds { 4172d84ff051STejun Heo int cpu = (unsigned long)hcpu; 41734ce62e9eSTejun Heo struct worker_pool *pool; 41741da177e4SLinus Torvalds 41758db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 41763af24433SOleg Nesterov case CPU_UP_PREPARE: 4177f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 41783ce63377STejun Heo if (pool->nr_workers) 41793ce63377STejun Heo continue; 4180ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 41813ce63377STejun Heo return NOTIFY_BAD; 41823af24433SOleg Nesterov } 41831da177e4SLinus Torvalds break; 41841da177e4SLinus Torvalds 418565758202STejun Heo case CPU_DOWN_FAILED: 418665758202STejun Heo case CPU_ONLINE: 4187f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 4188bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 418994cf58bbSTejun Heo spin_lock_irq(&pool->lock); 419094cf58bbSTejun Heo 419124647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 419294cf58bbSTejun Heo rebind_workers(pool); 419394cf58bbSTejun Heo 419494cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4195bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 419694cf58bbSTejun Heo } 41978db25e78STejun Heo break; 419865758202STejun Heo } 419965758202STejun Heo return NOTIFY_OK; 420065758202STejun Heo } 420165758202STejun Heo 420265758202STejun Heo /* 420365758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 420465758202STejun Heo * This will be registered as low priority CPU notifier. 420565758202STejun Heo */ 42069fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 420765758202STejun Heo unsigned long action, 420865758202STejun Heo void *hcpu) 420965758202STejun Heo { 4210d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42118db25e78STejun Heo struct work_struct unbind_work; 42128db25e78STejun Heo 421365758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 421465758202STejun Heo case CPU_DOWN_PREPARE: 42158db25e78STejun Heo /* unbinding should happen on the local CPU */ 4216706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 42177635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 42188db25e78STejun Heo flush_work(&unbind_work); 42198db25e78STejun Heo break; 422065758202STejun Heo } 422165758202STejun Heo return NOTIFY_OK; 422265758202STejun Heo } 422365758202STejun Heo 42242d3854a3SRusty Russell #ifdef CONFIG_SMP 42258ccad40dSRusty Russell 42262d3854a3SRusty Russell struct work_for_cpu { 4227ed48ece2STejun Heo struct work_struct work; 42282d3854a3SRusty Russell long (*fn)(void *); 42292d3854a3SRusty Russell void *arg; 42302d3854a3SRusty Russell long ret; 42312d3854a3SRusty Russell }; 42322d3854a3SRusty Russell 4233ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 42342d3854a3SRusty Russell { 4235ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4236ed48ece2STejun Heo 42372d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 42382d3854a3SRusty Russell } 42392d3854a3SRusty Russell 42402d3854a3SRusty Russell /** 42412d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 42422d3854a3SRusty Russell * @cpu: the cpu to run on 42432d3854a3SRusty Russell * @fn: the function to run 42442d3854a3SRusty Russell * @arg: the function arg 42452d3854a3SRusty Russell * 424631ad9081SRusty Russell * This will return the value @fn returns. 424731ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 42486b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 42492d3854a3SRusty Russell */ 4250d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 42512d3854a3SRusty Russell { 4252ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 42532d3854a3SRusty Russell 4254ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4255ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4256ed48ece2STejun Heo flush_work(&wfc.work); 42572d3854a3SRusty Russell return wfc.ret; 42582d3854a3SRusty Russell } 42592d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 42602d3854a3SRusty Russell #endif /* CONFIG_SMP */ 42612d3854a3SRusty Russell 4262a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4263e7577c50SRusty Russell 4264a0a1a5fdSTejun Heo /** 4265a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4266a0a1a5fdSTejun Heo * 426758a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4268c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4269706026c2STejun Heo * pool->worklist. 4270a0a1a5fdSTejun Heo * 4271a0a1a5fdSTejun Heo * CONTEXT: 4272794b18bcSTejun Heo * Grabs and releases wq_mutex, pwq_lock and pool->lock's. 4273a0a1a5fdSTejun Heo */ 4274a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4275a0a1a5fdSTejun Heo { 427617116969STejun Heo struct worker_pool *pool; 427724b8a847STejun Heo struct workqueue_struct *wq; 427824b8a847STejun Heo struct pool_workqueue *pwq; 4279611c92a0STejun Heo int pi; 4280a0a1a5fdSTejun Heo 42815bcab335STejun Heo mutex_lock(&wq_mutex); 4282a0a1a5fdSTejun Heo 42836183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4284a0a1a5fdSTejun Heo workqueue_freezing = true; 4285a0a1a5fdSTejun Heo 428624b8a847STejun Heo /* set FREEZING */ 4287611c92a0STejun Heo for_each_pool(pool, pi) { 42885bcab335STejun Heo spin_lock_irq(&pool->lock); 428935b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 429035b6bb63STejun Heo pool->flags |= POOL_FREEZING; 42915bcab335STejun Heo spin_unlock_irq(&pool->lock); 42921da177e4SLinus Torvalds } 42938b03ae3cSTejun Heo 429424b8a847STejun Heo /* suppress further executions by setting max_active to zero */ 4295794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 429624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4297699ce097STejun Heo for_each_pwq(pwq, wq) 4298699ce097STejun Heo pwq_adjust_max_active(pwq); 4299a1056305STejun Heo } 4300794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 43015bcab335STejun Heo 43025bcab335STejun Heo mutex_unlock(&wq_mutex); 4303a0a1a5fdSTejun Heo } 4304a0a1a5fdSTejun Heo 4305a0a1a5fdSTejun Heo /** 430658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4307a0a1a5fdSTejun Heo * 4308a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4309a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4310a0a1a5fdSTejun Heo * 4311a0a1a5fdSTejun Heo * CONTEXT: 43125bcab335STejun Heo * Grabs and releases wq_mutex. 4313a0a1a5fdSTejun Heo * 4314a0a1a5fdSTejun Heo * RETURNS: 431558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 431658a69cb4STejun Heo * is complete. 4317a0a1a5fdSTejun Heo */ 4318a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4319a0a1a5fdSTejun Heo { 4320a0a1a5fdSTejun Heo bool busy = false; 432124b8a847STejun Heo struct workqueue_struct *wq; 432224b8a847STejun Heo struct pool_workqueue *pwq; 4323a0a1a5fdSTejun Heo 43245bcab335STejun Heo mutex_lock(&wq_mutex); 4325a0a1a5fdSTejun Heo 43266183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4327a0a1a5fdSTejun Heo 432824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 432924b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 433024b8a847STejun Heo continue; 4331a0a1a5fdSTejun Heo /* 4332a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4333a0a1a5fdSTejun Heo * to peek without lock. 4334a0a1a5fdSTejun Heo */ 43355bcab335STejun Heo preempt_disable(); 433624b8a847STejun Heo for_each_pwq(pwq, wq) { 43376183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4338112202d9STejun Heo if (pwq->nr_active) { 4339a0a1a5fdSTejun Heo busy = true; 43405bcab335STejun Heo preempt_enable(); 4341a0a1a5fdSTejun Heo goto out_unlock; 4342a0a1a5fdSTejun Heo } 4343a0a1a5fdSTejun Heo } 43445bcab335STejun Heo preempt_enable(); 4345a0a1a5fdSTejun Heo } 4346a0a1a5fdSTejun Heo out_unlock: 43475bcab335STejun Heo mutex_unlock(&wq_mutex); 4348a0a1a5fdSTejun Heo return busy; 4349a0a1a5fdSTejun Heo } 4350a0a1a5fdSTejun Heo 4351a0a1a5fdSTejun Heo /** 4352a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4353a0a1a5fdSTejun Heo * 4354a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4355706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4356a0a1a5fdSTejun Heo * 4357a0a1a5fdSTejun Heo * CONTEXT: 4358794b18bcSTejun Heo * Grabs and releases wq_mutex, pwq_lock and pool->lock's. 4359a0a1a5fdSTejun Heo */ 4360a0a1a5fdSTejun Heo void thaw_workqueues(void) 4361a0a1a5fdSTejun Heo { 436224b8a847STejun Heo struct workqueue_struct *wq; 436324b8a847STejun Heo struct pool_workqueue *pwq; 436424b8a847STejun Heo struct worker_pool *pool; 4365611c92a0STejun Heo int pi; 4366a0a1a5fdSTejun Heo 43675bcab335STejun Heo mutex_lock(&wq_mutex); 4368a0a1a5fdSTejun Heo 4369a0a1a5fdSTejun Heo if (!workqueue_freezing) 4370a0a1a5fdSTejun Heo goto out_unlock; 4371a0a1a5fdSTejun Heo 437224b8a847STejun Heo /* clear FREEZING */ 4373611c92a0STejun Heo for_each_pool(pool, pi) { 43745bcab335STejun Heo spin_lock_irq(&pool->lock); 437535b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 437635b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 43775bcab335STejun Heo spin_unlock_irq(&pool->lock); 4378d565ed63STejun Heo } 437924b8a847STejun Heo 438024b8a847STejun Heo /* restore max_active and repopulate worklist */ 4381794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 438224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4383699ce097STejun Heo for_each_pwq(pwq, wq) 4384699ce097STejun Heo pwq_adjust_max_active(pwq); 438524b8a847STejun Heo } 4386794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 438724b8a847STejun Heo 438824b8a847STejun Heo /* kick workers */ 4389611c92a0STejun Heo for_each_pool(pool, pi) { 43905bcab335STejun Heo spin_lock_irq(&pool->lock); 439124b8a847STejun Heo wake_up_worker(pool); 43925bcab335STejun Heo spin_unlock_irq(&pool->lock); 4393a0a1a5fdSTejun Heo } 4394a0a1a5fdSTejun Heo 4395a0a1a5fdSTejun Heo workqueue_freezing = false; 4396a0a1a5fdSTejun Heo out_unlock: 43975bcab335STejun Heo mutex_unlock(&wq_mutex); 4398a0a1a5fdSTejun Heo } 4399a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4400a0a1a5fdSTejun Heo 44016ee0578bSSuresh Siddha static int __init init_workqueues(void) 44021da177e4SLinus Torvalds { 44037a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 44047a4e344cSTejun Heo int i, cpu; 4405c34056a3STejun Heo 44067c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 44077c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 44086be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4409b5490077STejun Heo 4410e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4411e904e6c2STejun Heo 4412e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4413e904e6c2STejun Heo 441465758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4415a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 44168b03ae3cSTejun Heo 4417706026c2STejun Heo /* initialize CPU pools */ 441829c91e99STejun Heo for_each_possible_cpu(cpu) { 44194ce62e9eSTejun Heo struct worker_pool *pool; 44208b03ae3cSTejun Heo 44217a4e344cSTejun Heo i = 0; 4422f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44237a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4424ec22ca5eSTejun Heo pool->cpu = cpu; 44257a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 44267a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 44277a4e344cSTejun Heo 44289daf9e67STejun Heo /* alloc pool ID */ 44295bcab335STejun Heo mutex_lock(&wq_mutex); 44309daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 44315bcab335STejun Heo mutex_unlock(&wq_mutex); 44324ce62e9eSTejun Heo } 44338b03ae3cSTejun Heo } 44348b03ae3cSTejun Heo 4435e22bee78STejun Heo /* create the initial worker */ 443629c91e99STejun Heo for_each_online_cpu(cpu) { 44374ce62e9eSTejun Heo struct worker_pool *pool; 4438e22bee78STejun Heo 4439f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 444024647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4441ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4442e22bee78STejun Heo } 44434ce62e9eSTejun Heo } 4444e22bee78STejun Heo 444529c91e99STejun Heo /* create default unbound wq attrs */ 444629c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 444729c91e99STejun Heo struct workqueue_attrs *attrs; 444829c91e99STejun Heo 444929c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 445029c91e99STejun Heo 445129c91e99STejun Heo attrs->nice = std_nice[i]; 445229c91e99STejun Heo cpumask_setall(attrs->cpumask); 445329c91e99STejun Heo 445429c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 445529c91e99STejun Heo } 445629c91e99STejun Heo 4457d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 44581aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4459d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4460f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4461f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 446224d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 446324d51addSTejun Heo WQ_FREEZABLE, 0); 44641aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4465ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 44666ee0578bSSuresh Siddha return 0; 44671da177e4SLinus Torvalds } 44686ee0578bSSuresh Siddha early_initcall(init_workqueues); 4469