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> 47bce90380STejun Heo #include <linux/nodemask.h> 484c16bd32STejun Heo #include <linux/moduleparam.h> 493d1cb205STejun Heo #include <linux/uaccess.h> 50e22bee78STejun Heo 51ea138446STejun Heo #include "workqueue_internal.h" 521da177e4SLinus Torvalds 53c8e55f36STejun Heo enum { 54bc2ae0f5STejun Heo /* 5524647570STejun Heo * worker_pool flags 56bc2ae0f5STejun Heo * 5724647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 58bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 59bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 60bc2ae0f5STejun Heo * is in effect. 61bc2ae0f5STejun Heo * 62bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 63bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6424647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 65bc2ae0f5STejun Heo * 66bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 67bc3a1afcSTejun Heo * manager_mutex to avoid changing binding state while 6824647570STejun Heo * create_worker() is in progress. 69bc2ae0f5STejun Heo */ 7011ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 7124647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 7235b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 73db7bccf4STejun Heo 74c8e55f36STejun Heo /* worker flags */ 75c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 76c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 77c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 78e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 79fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 80f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 81a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 82e22bee78STejun Heo 83a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 84a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 85db7bccf4STejun Heo 86e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 874ce62e9eSTejun Heo 8829c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 89c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 90db7bccf4STejun Heo 91e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 92e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 93e22bee78STejun Heo 943233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 953233cdbdSTejun Heo /* call for help after 10ms 963233cdbdSTejun Heo (min two ticks) */ 97e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 98e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 991da177e4SLinus Torvalds 1001da177e4SLinus Torvalds /* 101e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 102e22bee78STejun Heo * all cpus. Give -20. 103e22bee78STejun Heo */ 104e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1053270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 106ecf6881fSTejun Heo 107ecf6881fSTejun Heo WQ_NAME_LEN = 24, 108c8e55f36STejun Heo }; 109c8e55f36STejun Heo 1101da177e4SLinus Torvalds /* 1114690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1124690c4abSTejun Heo * 113e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 114e41e704bSTejun Heo * everyone else. 1154690c4abSTejun Heo * 116e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 117e22bee78STejun Heo * only be modified and accessed from the local cpu. 118e22bee78STejun Heo * 119d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1204690c4abSTejun Heo * 121d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 122d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 123d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 124d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 125e22bee78STejun Heo * 126822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 127822d8405STejun Heo * locks. Reads can happen under either lock. 128822d8405STejun Heo * 12968e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13076af4d93STejun Heo * 13168e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1325bcab335STejun Heo * 1333c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1343c25a55dSLai Jiangshan * 135b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1362e109a28STejun Heo * 1372e109a28STejun Heo * MD: wq_mayday_lock protected. 1384690c4abSTejun Heo */ 1394690c4abSTejun Heo 1402eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 141c34056a3STejun Heo 142bd7bdd43STejun Heo struct worker_pool { 143d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 144d84ff051STejun Heo int cpu; /* I: the associated cpu */ 145f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1469daf9e67STejun Heo int id; /* I: pool ID */ 14711ebea50STejun Heo unsigned int flags; /* X: flags */ 148bd7bdd43STejun Heo 149bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 150bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 151ea1abd61SLai Jiangshan 152ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 153bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 154bd7bdd43STejun Heo 155bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 156bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 157bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 158bd7bdd43STejun Heo 159c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 160c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 161c9e7cf27STejun Heo /* L: hash of busy workers */ 162c9e7cf27STejun Heo 163bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16434a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 165bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 166822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 167e19e397aSTejun Heo 1687a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16968e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17068e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1717a4e344cSTejun Heo 172e19e397aSTejun Heo /* 173e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 174e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 175e19e397aSTejun Heo * cacheline. 176e19e397aSTejun Heo */ 177e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17829c91e99STejun Heo 17929c91e99STejun Heo /* 18029c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18129c91e99STejun Heo * from get_work_pool(). 18229c91e99STejun Heo */ 18329c91e99STejun Heo struct rcu_head rcu; 1848b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1858b03ae3cSTejun Heo 1868b03ae3cSTejun Heo /* 187112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 188112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 189112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 190112202d9STejun Heo * number of flag bits. 1911da177e4SLinus Torvalds */ 192112202d9STejun Heo struct pool_workqueue { 193bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1944690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19573f53c4aSTejun Heo int work_color; /* L: current color */ 19673f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1978864b4e5STejun Heo int refcnt; /* L: reference count */ 19873f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19973f53c4aSTejun Heo /* L: nr of in_flight works */ 2001e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 201a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2021e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2033c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2042e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2058864b4e5STejun Heo 2068864b4e5STejun Heo /* 2078864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2088864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2098864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 210b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2118864b4e5STejun Heo */ 2128864b4e5STejun Heo struct work_struct unbound_release_work; 2138864b4e5STejun Heo struct rcu_head rcu; 214e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2151da177e4SLinus Torvalds 2161da177e4SLinus Torvalds /* 21773f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21873f53c4aSTejun Heo */ 21973f53c4aSTejun Heo struct wq_flusher { 2203c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2213c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22273f53c4aSTejun Heo struct completion done; /* flush completion */ 22373f53c4aSTejun Heo }; 2241da177e4SLinus Torvalds 225226223abSTejun Heo struct wq_device; 226226223abSTejun Heo 22773f53c4aSTejun Heo /* 228c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 229c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2301da177e4SLinus Torvalds */ 2311da177e4SLinus Torvalds struct workqueue_struct { 2323c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 23368e13a67SLai Jiangshan struct list_head list; /* PL: list of all workqueues */ 23473f53c4aSTejun Heo 2353c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2363c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2373c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 238112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2393c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2403c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2413c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24273f53c4aSTejun Heo 2432e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 244e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 245e22bee78STejun Heo 24687fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 247a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 248226223abSTejun Heo 2496029a918STejun Heo struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ 2504c16bd32STejun Heo struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ 2516029a918STejun Heo 252226223abSTejun Heo #ifdef CONFIG_SYSFS 253226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 254226223abSTejun Heo #endif 2554e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2564e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2574e6045f1SJohannes Berg #endif 258ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2592728fd2fSTejun Heo 2602728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2612728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2622728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 263df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2641da177e4SLinus Torvalds }; 2651da177e4SLinus Torvalds 266e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 267e904e6c2STejun Heo 268bce90380STejun Heo static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 269bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 270bce90380STejun Heo /* possible CPUs of each node */ 271bce90380STejun Heo 272d55262c4STejun Heo static bool wq_disable_numa; 273d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 274d55262c4STejun Heo 275cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 276cee22a15SViresh Kumar #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT 277cee22a15SViresh Kumar static bool wq_power_efficient = true; 278cee22a15SViresh Kumar #else 279cee22a15SViresh Kumar static bool wq_power_efficient; 280cee22a15SViresh Kumar #endif 281cee22a15SViresh Kumar 282cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 283cee22a15SViresh Kumar 284bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 285bce90380STejun Heo 2864c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2874c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2884c16bd32STejun Heo 28968e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2902e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2915bcab335STejun Heo 29268e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 29368e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2947d19c5ceSTejun Heo 2957d19c5ceSTejun Heo /* the per-cpu worker pools */ 2967d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2977d19c5ceSTejun Heo cpu_worker_pools); 2987d19c5ceSTejun Heo 29968e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3007d19c5ceSTejun Heo 30168e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 30229c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 30329c91e99STejun Heo 304c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 30529c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 30629c91e99STejun Heo 307d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 308ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 309044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3101aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 311044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 312d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 313044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 314f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 315044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 31624d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3170668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3180668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3190668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3200668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 321d320c038STejun Heo 3227d19c5ceSTejun Heo static int worker_thread(void *__worker); 3237d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3247d19c5ceSTejun Heo const struct workqueue_attrs *from); 3257d19c5ceSTejun Heo 32697bd2347STejun Heo #define CREATE_TRACE_POINTS 32797bd2347STejun Heo #include <trace/events/workqueue.h> 32897bd2347STejun Heo 32968e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3305bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 33168e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 33268e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3335bcab335STejun Heo 334b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 33576af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 336b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 337b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 33876af4d93STejun Heo 339822d8405STejun Heo #ifdef CONFIG_LOCKDEP 340822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 341519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 342519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 343822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 344822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 345822d8405STejun Heo #else 346822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 347822d8405STejun Heo #endif 348822d8405STejun Heo 349f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 350f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 351f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3527a62c2c8STejun Heo (pool)++) 3534ce62e9eSTejun Heo 35449e3cf44STejun Heo /** 35517116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 35617116969STejun Heo * @pool: iteration cursor 357611c92a0STejun Heo * @pi: integer used for iteration 358fa1b54e6STejun Heo * 35968e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 36068e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 36168e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 362fa1b54e6STejun Heo * 363fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 364fa1b54e6STejun Heo * ignored. 36517116969STejun Heo */ 366611c92a0STejun Heo #define for_each_pool(pool, pi) \ 367611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 36868e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 369fa1b54e6STejun Heo else 37017116969STejun Heo 37117116969STejun Heo /** 372822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 373822d8405STejun Heo * @worker: iteration cursor 374822d8405STejun Heo * @wi: integer used for iteration 375822d8405STejun Heo * @pool: worker_pool to iterate workers of 376822d8405STejun Heo * 377822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 378822d8405STejun Heo * 379822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 380822d8405STejun Heo * ignored. 381822d8405STejun Heo */ 382822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 383822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 384822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 385822d8405STejun Heo else 386822d8405STejun Heo 387822d8405STejun Heo /** 38849e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 38949e3cf44STejun Heo * @pwq: iteration cursor 39049e3cf44STejun Heo * @wq: the target workqueue 39176af4d93STejun Heo * 392b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 393794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 394794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 39576af4d93STejun Heo * 39676af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 39776af4d93STejun Heo * ignored. 39849e3cf44STejun Heo */ 39949e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 40076af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 401b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 40276af4d93STejun Heo else 403f3421797STejun Heo 404dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 405dc186ad7SThomas Gleixner 406dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 407dc186ad7SThomas Gleixner 40899777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 40999777288SStanislaw Gruszka { 41099777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 41199777288SStanislaw Gruszka } 41299777288SStanislaw Gruszka 413dc186ad7SThomas Gleixner /* 414dc186ad7SThomas Gleixner * fixup_init is called when: 415dc186ad7SThomas Gleixner * - an active object is initialized 416dc186ad7SThomas Gleixner */ 417dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 418dc186ad7SThomas Gleixner { 419dc186ad7SThomas Gleixner struct work_struct *work = addr; 420dc186ad7SThomas Gleixner 421dc186ad7SThomas Gleixner switch (state) { 422dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 423dc186ad7SThomas Gleixner cancel_work_sync(work); 424dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 425dc186ad7SThomas Gleixner return 1; 426dc186ad7SThomas Gleixner default: 427dc186ad7SThomas Gleixner return 0; 428dc186ad7SThomas Gleixner } 429dc186ad7SThomas Gleixner } 430dc186ad7SThomas Gleixner 431dc186ad7SThomas Gleixner /* 432dc186ad7SThomas Gleixner * fixup_activate is called when: 433dc186ad7SThomas Gleixner * - an active object is activated 434dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 435dc186ad7SThomas Gleixner */ 436dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 437dc186ad7SThomas Gleixner { 438dc186ad7SThomas Gleixner struct work_struct *work = addr; 439dc186ad7SThomas Gleixner 440dc186ad7SThomas Gleixner switch (state) { 441dc186ad7SThomas Gleixner 442dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 443dc186ad7SThomas Gleixner /* 444dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 445dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 446dc186ad7SThomas Gleixner * is tracked in the object tracker. 447dc186ad7SThomas Gleixner */ 44822df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 449dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 450dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 451dc186ad7SThomas Gleixner return 0; 452dc186ad7SThomas Gleixner } 453dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 454dc186ad7SThomas Gleixner return 0; 455dc186ad7SThomas Gleixner 456dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 457dc186ad7SThomas Gleixner WARN_ON(1); 458dc186ad7SThomas Gleixner 459dc186ad7SThomas Gleixner default: 460dc186ad7SThomas Gleixner return 0; 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner } 463dc186ad7SThomas Gleixner 464dc186ad7SThomas Gleixner /* 465dc186ad7SThomas Gleixner * fixup_free is called when: 466dc186ad7SThomas Gleixner * - an active object is freed 467dc186ad7SThomas Gleixner */ 468dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 469dc186ad7SThomas Gleixner { 470dc186ad7SThomas Gleixner struct work_struct *work = addr; 471dc186ad7SThomas Gleixner 472dc186ad7SThomas Gleixner switch (state) { 473dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 474dc186ad7SThomas Gleixner cancel_work_sync(work); 475dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 476dc186ad7SThomas Gleixner return 1; 477dc186ad7SThomas Gleixner default: 478dc186ad7SThomas Gleixner return 0; 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner } 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 483dc186ad7SThomas Gleixner .name = "work_struct", 48499777288SStanislaw Gruszka .debug_hint = work_debug_hint, 485dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 486dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 487dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 488dc186ad7SThomas Gleixner }; 489dc186ad7SThomas Gleixner 490dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 491dc186ad7SThomas Gleixner { 492dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 493dc186ad7SThomas Gleixner } 494dc186ad7SThomas Gleixner 495dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 496dc186ad7SThomas Gleixner { 497dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 498dc186ad7SThomas Gleixner } 499dc186ad7SThomas Gleixner 500dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 501dc186ad7SThomas Gleixner { 502dc186ad7SThomas Gleixner if (onstack) 503dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 504dc186ad7SThomas Gleixner else 505dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 506dc186ad7SThomas Gleixner } 507dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 508dc186ad7SThomas Gleixner 509dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 510dc186ad7SThomas Gleixner { 511dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 512dc186ad7SThomas Gleixner } 513dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 514dc186ad7SThomas Gleixner 515dc186ad7SThomas Gleixner #else 516dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 517dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 518dc186ad7SThomas Gleixner #endif 519dc186ad7SThomas Gleixner 5209daf9e67STejun Heo /* allocate ID and assign it to @pool */ 5219daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5229daf9e67STejun Heo { 5239daf9e67STejun Heo int ret; 5249daf9e67STejun Heo 52568e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5265bcab335STejun Heo 527e68035fbSTejun Heo ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL); 528229641a6STejun Heo if (ret >= 0) { 529e68035fbSTejun Heo pool->id = ret; 530229641a6STejun Heo return 0; 531229641a6STejun Heo } 5329daf9e67STejun Heo return ret; 5339daf9e67STejun Heo } 5349daf9e67STejun Heo 53576af4d93STejun Heo /** 536df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 537df2d5ae4STejun Heo * @wq: the target workqueue 538df2d5ae4STejun Heo * @node: the node ID 539df2d5ae4STejun Heo * 540df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 541df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 542df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 543df2d5ae4STejun Heo */ 544df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 545df2d5ae4STejun Heo int node) 546df2d5ae4STejun Heo { 547df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 548df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 549df2d5ae4STejun Heo } 550df2d5ae4STejun Heo 55173f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 55273f53c4aSTejun Heo { 55373f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 55473f53c4aSTejun Heo } 55573f53c4aSTejun Heo 55673f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 55773f53c4aSTejun Heo { 55873f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 55973f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 56073f53c4aSTejun Heo } 56173f53c4aSTejun Heo 56273f53c4aSTejun Heo static int work_next_color(int color) 56373f53c4aSTejun Heo { 56473f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 565a848e3b6SOleg Nesterov } 566a848e3b6SOleg Nesterov 5674594bf15SDavid Howells /* 568112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 569112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5707c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5717a22ad75STejun Heo * 572112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 573112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 574bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 575bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5767a22ad75STejun Heo * 577112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5787c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 579112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5807c3eed5cSTejun Heo * available only while the work item is queued. 581bbb68dfaSTejun Heo * 582bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 583bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 584bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 585bbb68dfaSTejun Heo * try to steal the PENDING bit. 5864594bf15SDavid Howells */ 5877a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5887a22ad75STejun Heo unsigned long flags) 5897a22ad75STejun Heo { 5906183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5917a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5927a22ad75STejun Heo } 5937a22ad75STejun Heo 594112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5954690c4abSTejun Heo unsigned long extra_flags) 596365970a1SDavid Howells { 597112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 598112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 599365970a1SDavid Howells } 600365970a1SDavid Howells 6014468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6024468a00fSLai Jiangshan int pool_id) 6034468a00fSLai Jiangshan { 6044468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6054468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6064468a00fSLai Jiangshan } 6074468a00fSLai Jiangshan 6087c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6097c3eed5cSTejun Heo int pool_id) 6104d707b9fSOleg Nesterov { 61123657bb1STejun Heo /* 61223657bb1STejun Heo * The following wmb is paired with the implied mb in 61323657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 61423657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 61523657bb1STejun Heo * owner. 61623657bb1STejun Heo */ 61723657bb1STejun Heo smp_wmb(); 6187c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6194d707b9fSOleg Nesterov } 6204d707b9fSOleg Nesterov 6217a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 622365970a1SDavid Howells { 6237c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6247c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6257a22ad75STejun Heo } 6267a22ad75STejun Heo 627112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6287a22ad75STejun Heo { 629e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6307a22ad75STejun Heo 631112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 632e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 633e120153dSTejun Heo else 634e120153dSTejun Heo return NULL; 6357a22ad75STejun Heo } 6367a22ad75STejun Heo 6377c3eed5cSTejun Heo /** 6387c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6397c3eed5cSTejun Heo * @work: the work item of interest 6407c3eed5cSTejun Heo * 6417c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 642fa1b54e6STejun Heo * 64368e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 64468e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 64568e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 646fa1b54e6STejun Heo * 647fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 648fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 649fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 650fa1b54e6STejun Heo * returned pool is and stays online. 6517c3eed5cSTejun Heo */ 6527c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6537a22ad75STejun Heo { 654e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6557c3eed5cSTejun Heo int pool_id; 6567a22ad75STejun Heo 65768e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 658fa1b54e6STejun Heo 659112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 660112202d9STejun Heo return ((struct pool_workqueue *) 6617c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6627a22ad75STejun Heo 6637c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6647c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6657a22ad75STejun Heo return NULL; 6667a22ad75STejun Heo 667fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6687c3eed5cSTejun Heo } 6697c3eed5cSTejun Heo 6707c3eed5cSTejun Heo /** 6717c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6727c3eed5cSTejun Heo * @work: the work item of interest 6737c3eed5cSTejun Heo * 6747c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6757c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6767c3eed5cSTejun Heo */ 6777c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6787c3eed5cSTejun Heo { 67954d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6807c3eed5cSTejun Heo 681112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 682112202d9STejun Heo return ((struct pool_workqueue *) 68354d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 68454d5b7d0SLai Jiangshan 68554d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6867c3eed5cSTejun Heo } 6877c3eed5cSTejun Heo 688bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 689bbb68dfaSTejun Heo { 6907c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 691bbb68dfaSTejun Heo 6927c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6937c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 694bbb68dfaSTejun Heo } 695bbb68dfaSTejun Heo 696bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 697bbb68dfaSTejun Heo { 698bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 699bbb68dfaSTejun Heo 700112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 701bbb68dfaSTejun Heo } 702bbb68dfaSTejun Heo 703e22bee78STejun Heo /* 7043270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7053270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 706d565ed63STejun Heo * they're being called with pool->lock held. 707e22bee78STejun Heo */ 708e22bee78STejun Heo 70963d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 710649027d7STejun Heo { 711e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 712649027d7STejun Heo } 713649027d7STejun Heo 714e22bee78STejun Heo /* 715e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 716e22bee78STejun Heo * running workers. 717974271c4STejun Heo * 718974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 719706026c2STejun Heo * function will always return %true for unbound pools as long as the 720974271c4STejun Heo * worklist isn't empty. 721e22bee78STejun Heo */ 72263d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 723e22bee78STejun Heo { 72463d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 725e22bee78STejun Heo } 726e22bee78STejun Heo 727e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 72863d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 729e22bee78STejun Heo { 73063d95a91STejun Heo return pool->nr_idle; 731e22bee78STejun Heo } 732e22bee78STejun Heo 733e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 73463d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 735e22bee78STejun Heo { 736e19e397aSTejun Heo return !list_empty(&pool->worklist) && 737e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 738e22bee78STejun Heo } 739e22bee78STejun Heo 740e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 74163d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 742e22bee78STejun Heo { 74363d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 744e22bee78STejun Heo } 745e22bee78STejun Heo 746e22bee78STejun Heo /* Do I need to be the manager? */ 74763d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 748e22bee78STejun Heo { 74963d95a91STejun Heo return need_to_create_worker(pool) || 75011ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 751e22bee78STejun Heo } 752e22bee78STejun Heo 753e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 75463d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 755e22bee78STejun Heo { 75634a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 75763d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 75863d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 759e22bee78STejun Heo 760ea1abd61SLai Jiangshan /* 761ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 762ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 763ea1abd61SLai Jiangshan */ 764ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 765ea1abd61SLai Jiangshan return false; 766ea1abd61SLai Jiangshan 767e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 768e22bee78STejun Heo } 769e22bee78STejun Heo 770e22bee78STejun Heo /* 771e22bee78STejun Heo * Wake up functions. 772e22bee78STejun Heo */ 773e22bee78STejun Heo 7747e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 77563d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7767e11629dSTejun Heo { 77763d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7787e11629dSTejun Heo return NULL; 7797e11629dSTejun Heo 78063d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7817e11629dSTejun Heo } 7827e11629dSTejun Heo 7837e11629dSTejun Heo /** 7847e11629dSTejun Heo * wake_up_worker - wake up an idle worker 78563d95a91STejun Heo * @pool: worker pool to wake worker from 7867e11629dSTejun Heo * 78763d95a91STejun Heo * Wake up the first idle worker of @pool. 7887e11629dSTejun Heo * 7897e11629dSTejun Heo * CONTEXT: 790d565ed63STejun Heo * spin_lock_irq(pool->lock). 7917e11629dSTejun Heo */ 79263d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7937e11629dSTejun Heo { 79463d95a91STejun Heo struct worker *worker = first_worker(pool); 7957e11629dSTejun Heo 7967e11629dSTejun Heo if (likely(worker)) 7977e11629dSTejun Heo wake_up_process(worker->task); 7987e11629dSTejun Heo } 7997e11629dSTejun Heo 8004690c4abSTejun Heo /** 801e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 802e22bee78STejun Heo * @task: task waking up 803e22bee78STejun Heo * @cpu: CPU @task is waking up to 804e22bee78STejun Heo * 805e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 806e22bee78STejun Heo * being awoken. 807e22bee78STejun Heo * 808e22bee78STejun Heo * CONTEXT: 809e22bee78STejun Heo * spin_lock_irq(rq->lock) 810e22bee78STejun Heo */ 811d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 812e22bee78STejun Heo { 813e22bee78STejun Heo struct worker *worker = kthread_data(task); 814e22bee78STejun Heo 81536576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 816ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 817e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 818e22bee78STejun Heo } 81936576000SJoonsoo Kim } 820e22bee78STejun Heo 821e22bee78STejun Heo /** 822e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 823e22bee78STejun Heo * @task: task going to sleep 824e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 825e22bee78STejun Heo * 826e22bee78STejun Heo * This function is called during schedule() when a busy worker is 827e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 828e22bee78STejun Heo * returning pointer to its task. 829e22bee78STejun Heo * 830e22bee78STejun Heo * CONTEXT: 831e22bee78STejun Heo * spin_lock_irq(rq->lock) 832e22bee78STejun Heo * 833e22bee78STejun Heo * RETURNS: 834e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 835e22bee78STejun Heo */ 836d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 837e22bee78STejun Heo { 838e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 839111c225aSTejun Heo struct worker_pool *pool; 840e22bee78STejun Heo 841111c225aSTejun Heo /* 842111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 843111c225aSTejun Heo * workers, also reach here, let's not access anything before 844111c225aSTejun Heo * checking NOT_RUNNING. 845111c225aSTejun Heo */ 8462d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 847e22bee78STejun Heo return NULL; 848e22bee78STejun Heo 849111c225aSTejun Heo pool = worker->pool; 850111c225aSTejun Heo 851e22bee78STejun Heo /* this can only happen on the local cpu */ 8526183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8536183c009STejun Heo return NULL; 854e22bee78STejun Heo 855e22bee78STejun Heo /* 856e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 857e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 858e22bee78STejun Heo * Please read comment there. 859e22bee78STejun Heo * 860628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 861628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 862628c78e7STejun Heo * disabled, which in turn means that none else could be 863d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 864628c78e7STejun Heo * lock is safe. 865e22bee78STejun Heo */ 866e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 867e19e397aSTejun Heo !list_empty(&pool->worklist)) 86863d95a91STejun Heo to_wakeup = first_worker(pool); 869e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 870e22bee78STejun Heo } 871e22bee78STejun Heo 872e22bee78STejun Heo /** 873e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 874cb444766STejun Heo * @worker: self 875d302f017STejun Heo * @flags: flags to set 876d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 877d302f017STejun Heo * 878e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 879e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 880e22bee78STejun Heo * woken up. 881d302f017STejun Heo * 882cb444766STejun Heo * CONTEXT: 883d565ed63STejun Heo * spin_lock_irq(pool->lock) 884d302f017STejun Heo */ 885d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 886d302f017STejun Heo bool wakeup) 887d302f017STejun Heo { 888bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 889e22bee78STejun Heo 890cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 891cb444766STejun Heo 892e22bee78STejun Heo /* 893e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 894e22bee78STejun Heo * wake up an idle worker as necessary if requested by 895e22bee78STejun Heo * @wakeup. 896e22bee78STejun Heo */ 897e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 898e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 899e22bee78STejun Heo if (wakeup) { 900e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 901bd7bdd43STejun Heo !list_empty(&pool->worklist)) 90263d95a91STejun Heo wake_up_worker(pool); 903e22bee78STejun Heo } else 904e19e397aSTejun Heo atomic_dec(&pool->nr_running); 905e22bee78STejun Heo } 906e22bee78STejun Heo 907d302f017STejun Heo worker->flags |= flags; 908d302f017STejun Heo } 909d302f017STejun Heo 910d302f017STejun Heo /** 911e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 912cb444766STejun Heo * @worker: self 913d302f017STejun Heo * @flags: flags to clear 914d302f017STejun Heo * 915e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 916d302f017STejun Heo * 917cb444766STejun Heo * CONTEXT: 918d565ed63STejun Heo * spin_lock_irq(pool->lock) 919d302f017STejun Heo */ 920d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 921d302f017STejun Heo { 92263d95a91STejun Heo struct worker_pool *pool = worker->pool; 923e22bee78STejun Heo unsigned int oflags = worker->flags; 924e22bee78STejun Heo 925cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 926cb444766STejun Heo 927d302f017STejun Heo worker->flags &= ~flags; 928e22bee78STejun Heo 92942c025f3STejun Heo /* 93042c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 93142c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 93242c025f3STejun Heo * of multiple flags, not a single flag. 93342c025f3STejun Heo */ 934e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 935e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 936e19e397aSTejun Heo atomic_inc(&pool->nr_running); 937d302f017STejun Heo } 938d302f017STejun Heo 939d302f017STejun Heo /** 9408cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 941c9e7cf27STejun Heo * @pool: pool of interest 9428cca0eeaSTejun Heo * @work: work to find worker for 9438cca0eeaSTejun Heo * 944c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 945c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 946a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 947a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 948a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 949a2c1c57bSTejun Heo * being executed. 950a2c1c57bSTejun Heo * 951a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 952a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 953a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 954a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 955a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 956a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 957a2c1c57bSTejun Heo * 958c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 959c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 960c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 961c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 962c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 963c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9648cca0eeaSTejun Heo * 9658cca0eeaSTejun Heo * CONTEXT: 966d565ed63STejun Heo * spin_lock_irq(pool->lock). 9678cca0eeaSTejun Heo * 9688cca0eeaSTejun Heo * RETURNS: 9698cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9708cca0eeaSTejun Heo * otherwise. 9718cca0eeaSTejun Heo */ 972c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9738cca0eeaSTejun Heo struct work_struct *work) 9748cca0eeaSTejun Heo { 97542f8570fSSasha Levin struct worker *worker; 97642f8570fSSasha Levin 977b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 978a2c1c57bSTejun Heo (unsigned long)work) 979a2c1c57bSTejun Heo if (worker->current_work == work && 980a2c1c57bSTejun Heo worker->current_func == work->func) 98142f8570fSSasha Levin return worker; 98242f8570fSSasha Levin 98342f8570fSSasha Levin return NULL; 9848cca0eeaSTejun Heo } 9858cca0eeaSTejun Heo 9868cca0eeaSTejun Heo /** 987bf4ede01STejun Heo * move_linked_works - move linked works to a list 988bf4ede01STejun Heo * @work: start of series of works to be scheduled 989bf4ede01STejun Heo * @head: target list to append @work to 990bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 991bf4ede01STejun Heo * 992bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 993bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 994bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 995bf4ede01STejun Heo * 996bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 997bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 998bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 999bf4ede01STejun Heo * 1000bf4ede01STejun Heo * CONTEXT: 1001d565ed63STejun Heo * spin_lock_irq(pool->lock). 1002bf4ede01STejun Heo */ 1003bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1004bf4ede01STejun Heo struct work_struct **nextp) 1005bf4ede01STejun Heo { 1006bf4ede01STejun Heo struct work_struct *n; 1007bf4ede01STejun Heo 1008bf4ede01STejun Heo /* 1009bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1010bf4ede01STejun Heo * use NULL for list head. 1011bf4ede01STejun Heo */ 1012bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1013bf4ede01STejun Heo list_move_tail(&work->entry, head); 1014bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1015bf4ede01STejun Heo break; 1016bf4ede01STejun Heo } 1017bf4ede01STejun Heo 1018bf4ede01STejun Heo /* 1019bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1020bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1021bf4ede01STejun Heo * needs to be updated. 1022bf4ede01STejun Heo */ 1023bf4ede01STejun Heo if (nextp) 1024bf4ede01STejun Heo *nextp = n; 1025bf4ede01STejun Heo } 1026bf4ede01STejun Heo 10278864b4e5STejun Heo /** 10288864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10298864b4e5STejun Heo * @pwq: pool_workqueue to get 10308864b4e5STejun Heo * 10318864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10328864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10338864b4e5STejun Heo */ 10348864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10358864b4e5STejun Heo { 10368864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10378864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10388864b4e5STejun Heo pwq->refcnt++; 10398864b4e5STejun Heo } 10408864b4e5STejun Heo 10418864b4e5STejun Heo /** 10428864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10438864b4e5STejun Heo * @pwq: pool_workqueue to put 10448864b4e5STejun Heo * 10458864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10468864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10478864b4e5STejun Heo */ 10488864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10498864b4e5STejun Heo { 10508864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10518864b4e5STejun Heo if (likely(--pwq->refcnt)) 10528864b4e5STejun Heo return; 10538864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10548864b4e5STejun Heo return; 10558864b4e5STejun Heo /* 10568864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10578864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10588864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10598864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10608864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10618864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10628864b4e5STejun Heo */ 10638864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10648864b4e5STejun Heo } 10658864b4e5STejun Heo 1066dce90d47STejun Heo /** 1067dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1068dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1069dce90d47STejun Heo * 1070dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1071dce90d47STejun Heo */ 1072dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1073dce90d47STejun Heo { 1074dce90d47STejun Heo if (pwq) { 1075dce90d47STejun Heo /* 1076dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1077dce90d47STejun Heo * following lock operations are safe. 1078dce90d47STejun Heo */ 1079dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1080dce90d47STejun Heo put_pwq(pwq); 1081dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1082dce90d47STejun Heo } 1083dce90d47STejun Heo } 1084dce90d47STejun Heo 1085112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1086bf4ede01STejun Heo { 1087112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1088bf4ede01STejun Heo 1089bf4ede01STejun Heo trace_workqueue_activate_work(work); 1090112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1091bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1092112202d9STejun Heo pwq->nr_active++; 1093bf4ede01STejun Heo } 1094bf4ede01STejun Heo 1095112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10963aa62497SLai Jiangshan { 1097112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10983aa62497SLai Jiangshan struct work_struct, entry); 10993aa62497SLai Jiangshan 1100112202d9STejun Heo pwq_activate_delayed_work(work); 11013aa62497SLai Jiangshan } 11023aa62497SLai Jiangshan 1103bf4ede01STejun Heo /** 1104112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1105112202d9STejun Heo * @pwq: pwq of interest 1106bf4ede01STejun Heo * @color: color of work which left the queue 1107bf4ede01STejun Heo * 1108bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1109112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1110bf4ede01STejun Heo * 1111bf4ede01STejun Heo * CONTEXT: 1112d565ed63STejun Heo * spin_lock_irq(pool->lock). 1113bf4ede01STejun Heo */ 1114112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1115bf4ede01STejun Heo { 11168864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1117bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11188864b4e5STejun Heo goto out_put; 1119bf4ede01STejun Heo 1120112202d9STejun Heo pwq->nr_in_flight[color]--; 1121bf4ede01STejun Heo 1122112202d9STejun Heo pwq->nr_active--; 1123112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1124bf4ede01STejun Heo /* one down, submit a delayed one */ 1125112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1126112202d9STejun Heo pwq_activate_first_delayed(pwq); 1127bf4ede01STejun Heo } 1128bf4ede01STejun Heo 1129bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1130112202d9STejun Heo if (likely(pwq->flush_color != color)) 11318864b4e5STejun Heo goto out_put; 1132bf4ede01STejun Heo 1133bf4ede01STejun Heo /* are there still in-flight works? */ 1134112202d9STejun Heo if (pwq->nr_in_flight[color]) 11358864b4e5STejun Heo goto out_put; 1136bf4ede01STejun Heo 1137112202d9STejun Heo /* this pwq is done, clear flush_color */ 1138112202d9STejun Heo pwq->flush_color = -1; 1139bf4ede01STejun Heo 1140bf4ede01STejun Heo /* 1141112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1142bf4ede01STejun Heo * will handle the rest. 1143bf4ede01STejun Heo */ 1144112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1145112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11468864b4e5STejun Heo out_put: 11478864b4e5STejun Heo put_pwq(pwq); 1148bf4ede01STejun Heo } 1149bf4ede01STejun Heo 115036e227d2STejun Heo /** 1151bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 115236e227d2STejun Heo * @work: work item to steal 115336e227d2STejun Heo * @is_dwork: @work is a delayed_work 1154bbb68dfaSTejun Heo * @flags: place to store irq state 115536e227d2STejun Heo * 115636e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 115736e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 115836e227d2STejun Heo * 115936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 116036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 116136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1162bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1163bbb68dfaSTejun Heo * for arbitrarily long 116436e227d2STejun Heo * 1165bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1166e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1167e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1168e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1169bbb68dfaSTejun Heo * 1170bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1171bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1172bbb68dfaSTejun Heo * 1173e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1174bf4ede01STejun Heo */ 1175bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1176bbb68dfaSTejun Heo unsigned long *flags) 1177bf4ede01STejun Heo { 1178d565ed63STejun Heo struct worker_pool *pool; 1179112202d9STejun Heo struct pool_workqueue *pwq; 1180bf4ede01STejun Heo 1181bbb68dfaSTejun Heo local_irq_save(*flags); 1182bbb68dfaSTejun Heo 118336e227d2STejun Heo /* try to steal the timer if it exists */ 118436e227d2STejun Heo if (is_dwork) { 118536e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 118636e227d2STejun Heo 1187e0aecdd8STejun Heo /* 1188e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1189e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1190e0aecdd8STejun Heo * running on the local CPU. 1191e0aecdd8STejun Heo */ 119236e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 119336e227d2STejun Heo return 1; 119436e227d2STejun Heo } 119536e227d2STejun Heo 119636e227d2STejun Heo /* try to claim PENDING the normal way */ 1197bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1198bf4ede01STejun Heo return 0; 1199bf4ede01STejun Heo 1200bf4ede01STejun Heo /* 1201bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1202bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1203bf4ede01STejun Heo */ 1204d565ed63STejun Heo pool = get_work_pool(work); 1205d565ed63STejun Heo if (!pool) 1206bbb68dfaSTejun Heo goto fail; 1207bf4ede01STejun Heo 1208d565ed63STejun Heo spin_lock(&pool->lock); 1209bf4ede01STejun Heo /* 1210112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1211112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1212112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1213112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1214112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12150b3dae68SLai Jiangshan * item is currently queued on that pool. 1216bf4ede01STejun Heo */ 1217112202d9STejun Heo pwq = get_work_pwq(work); 1218112202d9STejun Heo if (pwq && pwq->pool == pool) { 1219bf4ede01STejun Heo debug_work_deactivate(work); 12203aa62497SLai Jiangshan 12213aa62497SLai Jiangshan /* 122216062836STejun Heo * A delayed work item cannot be grabbed directly because 122316062836STejun Heo * it might have linked NO_COLOR work items which, if left 1224112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 122516062836STejun Heo * management later on and cause stall. Make sure the work 122616062836STejun Heo * item is activated before grabbing. 12273aa62497SLai Jiangshan */ 12283aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1229112202d9STejun Heo pwq_activate_delayed_work(work); 12303aa62497SLai Jiangshan 1231bf4ede01STejun Heo list_del_init(&work->entry); 1232112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 123336e227d2STejun Heo 1234112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12354468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12364468a00fSLai Jiangshan 1237d565ed63STejun Heo spin_unlock(&pool->lock); 123836e227d2STejun Heo return 1; 1239bf4ede01STejun Heo } 1240d565ed63STejun Heo spin_unlock(&pool->lock); 1241bbb68dfaSTejun Heo fail: 1242bbb68dfaSTejun Heo local_irq_restore(*flags); 1243bbb68dfaSTejun Heo if (work_is_canceling(work)) 1244bbb68dfaSTejun Heo return -ENOENT; 1245bbb68dfaSTejun Heo cpu_relax(); 124636e227d2STejun Heo return -EAGAIN; 1247bf4ede01STejun Heo } 1248bf4ede01STejun Heo 1249bf4ede01STejun Heo /** 1250706026c2STejun Heo * insert_work - insert a work into a pool 1251112202d9STejun Heo * @pwq: pwq @work belongs to 12524690c4abSTejun Heo * @work: work to insert 12534690c4abSTejun Heo * @head: insertion point 12544690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12554690c4abSTejun Heo * 1256112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1257706026c2STejun Heo * work_struct flags. 12584690c4abSTejun Heo * 12594690c4abSTejun Heo * CONTEXT: 1260d565ed63STejun Heo * spin_lock_irq(pool->lock). 1261365970a1SDavid Howells */ 1262112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1263112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1264b89deed3SOleg Nesterov { 1265112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1266e1d8aa9fSFrederic Weisbecker 12674690c4abSTejun Heo /* we own @work, set data and link */ 1268112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12691a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12708864b4e5STejun Heo get_pwq(pwq); 1271e22bee78STejun Heo 1272e22bee78STejun Heo /* 1273c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1274c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1275c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1276e22bee78STejun Heo */ 1277e22bee78STejun Heo smp_mb(); 1278e22bee78STejun Heo 127963d95a91STejun Heo if (__need_more_worker(pool)) 128063d95a91STejun Heo wake_up_worker(pool); 1281b89deed3SOleg Nesterov } 1282b89deed3SOleg Nesterov 1283c8efcc25STejun Heo /* 1284c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12858d03ecfeSTejun Heo * same workqueue. 1286c8efcc25STejun Heo */ 1287c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1288c8efcc25STejun Heo { 1289c8efcc25STejun Heo struct worker *worker; 1290c8efcc25STejun Heo 12918d03ecfeSTejun Heo worker = current_wq_worker(); 1292c8efcc25STejun Heo /* 12938d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12948d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1295c8efcc25STejun Heo */ 1296112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1297c8efcc25STejun Heo } 1298c8efcc25STejun Heo 1299d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13001da177e4SLinus Torvalds struct work_struct *work) 13011da177e4SLinus Torvalds { 1302112202d9STejun Heo struct pool_workqueue *pwq; 1303c9178087STejun Heo struct worker_pool *last_pool; 13041e19ffc6STejun Heo struct list_head *worklist; 13058a2e8e5dSTejun Heo unsigned int work_flags; 1306b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13078930cabaSTejun Heo 13088930cabaSTejun Heo /* 13098930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13108930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13118930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13128930cabaSTejun Heo * happen with IRQ disabled. 13138930cabaSTejun Heo */ 13148930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13151da177e4SLinus Torvalds 1316dc186ad7SThomas Gleixner debug_work_activate(work); 13171e19ffc6STejun Heo 1318c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1319618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1320c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1321e41e704bSTejun Heo return; 13229e8cd2f5STejun Heo retry: 1323df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1324f3421797STejun Heo cpu = raw_smp_processor_id(); 1325df2d5ae4STejun Heo 1326df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1327df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1328c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1329df2d5ae4STejun Heo else 1330df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1331f3421797STejun Heo 133218aa9effSTejun Heo /* 1333c9178087STejun Heo * If @work was previously on a different pool, it might still be 1334c9178087STejun Heo * running there, in which case the work needs to be queued on that 1335c9178087STejun Heo * pool to guarantee non-reentrancy. 133618aa9effSTejun Heo */ 1337c9e7cf27STejun Heo last_pool = get_work_pool(work); 1338112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 133918aa9effSTejun Heo struct worker *worker; 134018aa9effSTejun Heo 1341d565ed63STejun Heo spin_lock(&last_pool->lock); 134218aa9effSTejun Heo 1343c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 134418aa9effSTejun Heo 1345112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1346c9178087STejun Heo pwq = worker->current_pwq; 13478594fadeSLai Jiangshan } else { 134818aa9effSTejun Heo /* meh... not running there, queue here */ 1349d565ed63STejun Heo spin_unlock(&last_pool->lock); 1350112202d9STejun Heo spin_lock(&pwq->pool->lock); 135118aa9effSTejun Heo } 13528930cabaSTejun Heo } else { 1353112202d9STejun Heo spin_lock(&pwq->pool->lock); 13548930cabaSTejun Heo } 1355502ca9d8STejun Heo 13569e8cd2f5STejun Heo /* 13579e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13589e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13599e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1360df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1361df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13629e8cd2f5STejun Heo * make forward-progress. 13639e8cd2f5STejun Heo */ 13649e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13659e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13669e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13679e8cd2f5STejun Heo cpu_relax(); 13689e8cd2f5STejun Heo goto retry; 13699e8cd2f5STejun Heo } 13709e8cd2f5STejun Heo /* oops */ 13719e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13729e8cd2f5STejun Heo wq->name, cpu); 13739e8cd2f5STejun Heo } 13749e8cd2f5STejun Heo 1375112202d9STejun Heo /* pwq determined, queue */ 1376112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1377502ca9d8STejun Heo 1378f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1379112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1380f5b2552bSDan Carpenter return; 1381f5b2552bSDan Carpenter } 13821e19ffc6STejun Heo 1383112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1384112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13851e19ffc6STejun Heo 1386112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1387cdadf009STejun Heo trace_workqueue_activate_work(work); 1388112202d9STejun Heo pwq->nr_active++; 1389112202d9STejun Heo worklist = &pwq->pool->worklist; 13908a2e8e5dSTejun Heo } else { 13918a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1392112202d9STejun Heo worklist = &pwq->delayed_works; 13938a2e8e5dSTejun Heo } 13941e19ffc6STejun Heo 1395112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13961e19ffc6STejun Heo 1397112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13981da177e4SLinus Torvalds } 13991da177e4SLinus Torvalds 14000fcb78c2SRolf Eike Beer /** 1401c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1402c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1403c1a220e7SZhang Rui * @wq: workqueue to use 1404c1a220e7SZhang Rui * @work: work to queue 1405c1a220e7SZhang Rui * 1406d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1407c1a220e7SZhang Rui * 1408c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1409c1a220e7SZhang Rui * can't go away. 1410c1a220e7SZhang Rui */ 1411d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1412d4283e93STejun Heo struct work_struct *work) 1413c1a220e7SZhang Rui { 1414d4283e93STejun Heo bool ret = false; 14158930cabaSTejun Heo unsigned long flags; 14168930cabaSTejun Heo 14178930cabaSTejun Heo local_irq_save(flags); 1418c1a220e7SZhang Rui 141922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14204690c4abSTejun Heo __queue_work(cpu, wq, work); 1421d4283e93STejun Heo ret = true; 1422c1a220e7SZhang Rui } 14238930cabaSTejun Heo 14248930cabaSTejun Heo local_irq_restore(flags); 1425c1a220e7SZhang Rui return ret; 1426c1a220e7SZhang Rui } 1427ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1428c1a220e7SZhang Rui 1429d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14301da177e4SLinus Torvalds { 143152bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14321da177e4SLinus Torvalds 1433e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 143460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14351da177e4SLinus Torvalds } 14361438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14371da177e4SLinus Torvalds 14387beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 143952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14401da177e4SLinus Torvalds { 14417beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14427beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14431da177e4SLinus Torvalds 14447beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14457beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1446fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1447fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14487beb2edfSTejun Heo 14498852aac2STejun Heo /* 14508852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14518852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14528852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14538852aac2STejun Heo * on that there's no such delay when @delay is 0. 14548852aac2STejun Heo */ 14558852aac2STejun Heo if (!delay) { 14568852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14578852aac2STejun Heo return; 14588852aac2STejun Heo } 14598852aac2STejun Heo 14607beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14617beb2edfSTejun Heo 146260c057bcSLai Jiangshan dwork->wq = wq; 14631265057fSTejun Heo dwork->cpu = cpu; 14647beb2edfSTejun Heo timer->expires = jiffies + delay; 14657beb2edfSTejun Heo 14667beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14677beb2edfSTejun Heo add_timer_on(timer, cpu); 14687beb2edfSTejun Heo else 14697beb2edfSTejun Heo add_timer(timer); 14707beb2edfSTejun Heo } 14711da177e4SLinus Torvalds 14720fcb78c2SRolf Eike Beer /** 14730fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14740fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14750fcb78c2SRolf Eike Beer * @wq: workqueue to use 1476af9997e4SRandy Dunlap * @dwork: work to queue 14770fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14780fcb78c2SRolf Eike Beer * 1479715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1480715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1481715f1300STejun Heo * execution. 14820fcb78c2SRolf Eike Beer */ 1483d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 148452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14857a6bc1cdSVenkatesh Pallipadi { 148652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1487d4283e93STejun Heo bool ret = false; 14888930cabaSTejun Heo unsigned long flags; 14898930cabaSTejun Heo 14908930cabaSTejun Heo /* read the comment in __queue_work() */ 14918930cabaSTejun Heo local_irq_save(flags); 14927a6bc1cdSVenkatesh Pallipadi 149322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14947beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1495d4283e93STejun Heo ret = true; 14967a6bc1cdSVenkatesh Pallipadi } 14978930cabaSTejun Heo 14988930cabaSTejun Heo local_irq_restore(flags); 14997a6bc1cdSVenkatesh Pallipadi return ret; 15007a6bc1cdSVenkatesh Pallipadi } 1501ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15021da177e4SLinus Torvalds 1503c8e55f36STejun Heo /** 15048376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15058376fe22STejun Heo * @cpu: CPU number to execute work on 15068376fe22STejun Heo * @wq: workqueue to use 15078376fe22STejun Heo * @dwork: work to queue 15088376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15098376fe22STejun Heo * 15108376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15118376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15128376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15138376fe22STejun Heo * current state. 15148376fe22STejun Heo * 15158376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 15168376fe22STejun Heo * pending and its timer was modified. 15178376fe22STejun Heo * 1518e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15198376fe22STejun Heo * See try_to_grab_pending() for details. 15208376fe22STejun Heo */ 15218376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15228376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15238376fe22STejun Heo { 15248376fe22STejun Heo unsigned long flags; 15258376fe22STejun Heo int ret; 15268376fe22STejun Heo 15278376fe22STejun Heo do { 15288376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15298376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15308376fe22STejun Heo 15318376fe22STejun Heo if (likely(ret >= 0)) { 15328376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15338376fe22STejun Heo local_irq_restore(flags); 15348376fe22STejun Heo } 15358376fe22STejun Heo 15368376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15378376fe22STejun Heo return ret; 15388376fe22STejun Heo } 15398376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15408376fe22STejun Heo 15418376fe22STejun Heo /** 1542c8e55f36STejun Heo * worker_enter_idle - enter idle state 1543c8e55f36STejun Heo * @worker: worker which is entering idle state 1544c8e55f36STejun Heo * 1545c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1546c8e55f36STejun Heo * necessary. 1547c8e55f36STejun Heo * 1548c8e55f36STejun Heo * LOCKING: 1549d565ed63STejun Heo * spin_lock_irq(pool->lock). 1550c8e55f36STejun Heo */ 1551c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15521da177e4SLinus Torvalds { 1553bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1554c8e55f36STejun Heo 15556183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15566183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15576183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15586183c009STejun Heo return; 1559c8e55f36STejun Heo 1560cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1561cb444766STejun Heo worker->flags |= WORKER_IDLE; 1562bd7bdd43STejun Heo pool->nr_idle++; 1563e22bee78STejun Heo worker->last_active = jiffies; 1564c8e55f36STejun Heo 1565c8e55f36STejun Heo /* idle_list is LIFO */ 1566bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1567db7bccf4STejun Heo 156863d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1569628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1570cb444766STejun Heo 1571544ecf31STejun Heo /* 1572706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1573d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1574628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1575628c78e7STejun Heo * unbind is not in progress. 1576544ecf31STejun Heo */ 157724647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1578bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1579e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1580c8e55f36STejun Heo } 1581c8e55f36STejun Heo 1582c8e55f36STejun Heo /** 1583c8e55f36STejun Heo * worker_leave_idle - leave idle state 1584c8e55f36STejun Heo * @worker: worker which is leaving idle state 1585c8e55f36STejun Heo * 1586c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1587c8e55f36STejun Heo * 1588c8e55f36STejun Heo * LOCKING: 1589d565ed63STejun Heo * spin_lock_irq(pool->lock). 1590c8e55f36STejun Heo */ 1591c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1592c8e55f36STejun Heo { 1593bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1594c8e55f36STejun Heo 15956183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15966183c009STejun Heo return; 1597d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1598bd7bdd43STejun Heo pool->nr_idle--; 1599c8e55f36STejun Heo list_del_init(&worker->entry); 1600c8e55f36STejun Heo } 1601c8e55f36STejun Heo 1602e22bee78STejun Heo /** 1603f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1604f36dc67bSLai Jiangshan * @pool: target worker_pool 1605f36dc67bSLai Jiangshan * 1606f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1607e22bee78STejun Heo * 1608e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1609e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1610e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1611e22bee78STejun Heo * guaranteed to execute on the cpu. 1612e22bee78STejun Heo * 1613f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1614e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1615e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1616e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1617706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1618e22bee78STejun Heo * [dis]associated in the meantime. 1619e22bee78STejun Heo * 1620706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 162124647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1622f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1623f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1624f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1625e22bee78STejun Heo * 1626e22bee78STejun Heo * CONTEXT: 1627d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1628e22bee78STejun Heo * held. 1629e22bee78STejun Heo * 1630e22bee78STejun Heo * RETURNS: 1631706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1632e22bee78STejun Heo * bound), %false if offline. 1633e22bee78STejun Heo */ 1634f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1635d565ed63STejun Heo __acquires(&pool->lock) 1636e22bee78STejun Heo { 1637e22bee78STejun Heo while (true) { 1638e22bee78STejun Heo /* 1639e22bee78STejun Heo * The following call may fail, succeed or succeed 1640e22bee78STejun Heo * without actually migrating the task to the cpu if 1641e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 164224647570STejun Heo * against POOL_DISASSOCIATED. 1643e22bee78STejun Heo */ 164424647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16457a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1646e22bee78STejun Heo 1647d565ed63STejun Heo spin_lock_irq(&pool->lock); 164824647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1649e22bee78STejun Heo return false; 1650f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16517a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1652e22bee78STejun Heo return true; 1653d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1654e22bee78STejun Heo 16555035b20fSTejun Heo /* 16565035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16575035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16585035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16595035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16605035b20fSTejun Heo */ 1661e22bee78STejun Heo cpu_relax(); 16625035b20fSTejun Heo cond_resched(); 1663e22bee78STejun Heo } 1664e22bee78STejun Heo } 1665e22bee78STejun Heo 1666c34056a3STejun Heo static struct worker *alloc_worker(void) 1667c34056a3STejun Heo { 1668c34056a3STejun Heo struct worker *worker; 1669c34056a3STejun Heo 1670c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1671c8e55f36STejun Heo if (worker) { 1672c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1673affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1674e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1675e22bee78STejun Heo worker->flags = WORKER_PREP; 1676c8e55f36STejun Heo } 1677c34056a3STejun Heo return worker; 1678c34056a3STejun Heo } 1679c34056a3STejun Heo 1680c34056a3STejun Heo /** 1681c34056a3STejun Heo * create_worker - create a new workqueue worker 168263d95a91STejun Heo * @pool: pool the new worker will belong to 1683c34056a3STejun Heo * 168463d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1685c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1686c34056a3STejun Heo * destroy_worker(). 1687c34056a3STejun Heo * 1688c34056a3STejun Heo * CONTEXT: 1689c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1690c34056a3STejun Heo * 1691c34056a3STejun Heo * RETURNS: 1692c34056a3STejun Heo * Pointer to the newly created worker. 1693c34056a3STejun Heo */ 1694bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1695c34056a3STejun Heo { 1696c34056a3STejun Heo struct worker *worker = NULL; 1697f3421797STejun Heo int id = -1; 1698e3c916a4STejun Heo char id_buf[16]; 1699c34056a3STejun Heo 1700cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1701cd549687STejun Heo 1702822d8405STejun Heo /* 1703822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1704822d8405STejun Heo * without installing the pointer. 1705822d8405STejun Heo */ 1706822d8405STejun Heo idr_preload(GFP_KERNEL); 1707d565ed63STejun Heo spin_lock_irq(&pool->lock); 1708822d8405STejun Heo 1709822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1710822d8405STejun Heo 1711d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1712822d8405STejun Heo idr_preload_end(); 1713822d8405STejun Heo if (id < 0) 1714c34056a3STejun Heo goto fail; 1715c34056a3STejun Heo 1716c34056a3STejun Heo worker = alloc_worker(); 1717c34056a3STejun Heo if (!worker) 1718c34056a3STejun Heo goto fail; 1719c34056a3STejun Heo 1720bd7bdd43STejun Heo worker->pool = pool; 1721c34056a3STejun Heo worker->id = id; 1722c34056a3STejun Heo 172329c91e99STejun Heo if (pool->cpu >= 0) 1724e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1725e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1726f3421797STejun Heo else 1727e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1728e3c916a4STejun Heo 1729f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1730e3c916a4STejun Heo "kworker/%s", id_buf); 1731c34056a3STejun Heo if (IS_ERR(worker->task)) 1732c34056a3STejun Heo goto fail; 1733c34056a3STejun Heo 1734c5aa87bbSTejun Heo /* 1735c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1736c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1737c5aa87bbSTejun Heo */ 17387a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 17397a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17403270476aSTejun Heo 174114a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 174214a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 17437a4e344cSTejun Heo 17447a4e344cSTejun Heo /* 17457a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17467a4e344cSTejun Heo * remains stable across this function. See the comments above the 17477a4e344cSTejun Heo * flag definition for details. 17487a4e344cSTejun Heo */ 17497a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1750f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1751c34056a3STejun Heo 1752822d8405STejun Heo /* successful, commit the pointer to idr */ 1753822d8405STejun Heo spin_lock_irq(&pool->lock); 1754822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1755822d8405STejun Heo spin_unlock_irq(&pool->lock); 1756822d8405STejun Heo 1757c34056a3STejun Heo return worker; 1758822d8405STejun Heo 1759c34056a3STejun Heo fail: 1760c34056a3STejun Heo if (id >= 0) { 1761d565ed63STejun Heo spin_lock_irq(&pool->lock); 1762822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1763d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1764c34056a3STejun Heo } 1765c34056a3STejun Heo kfree(worker); 1766c34056a3STejun Heo return NULL; 1767c34056a3STejun Heo } 1768c34056a3STejun Heo 1769c34056a3STejun Heo /** 1770c34056a3STejun Heo * start_worker - start a newly created worker 1771c34056a3STejun Heo * @worker: worker to start 1772c34056a3STejun Heo * 1773706026c2STejun Heo * Make the pool aware of @worker and start it. 1774c34056a3STejun Heo * 1775c34056a3STejun Heo * CONTEXT: 1776d565ed63STejun Heo * spin_lock_irq(pool->lock). 1777c34056a3STejun Heo */ 1778c34056a3STejun Heo static void start_worker(struct worker *worker) 1779c34056a3STejun Heo { 1780cb444766STejun Heo worker->flags |= WORKER_STARTED; 1781bd7bdd43STejun Heo worker->pool->nr_workers++; 1782c8e55f36STejun Heo worker_enter_idle(worker); 1783c34056a3STejun Heo wake_up_process(worker->task); 1784c34056a3STejun Heo } 1785c34056a3STejun Heo 1786c34056a3STejun Heo /** 1787ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1788ebf44d16STejun Heo * @pool: the target pool 1789ebf44d16STejun Heo * 1790cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1791ebf44d16STejun Heo */ 1792ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1793ebf44d16STejun Heo { 1794ebf44d16STejun Heo struct worker *worker; 1795ebf44d16STejun Heo 1796cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1797cd549687STejun Heo 1798ebf44d16STejun Heo worker = create_worker(pool); 1799ebf44d16STejun Heo if (worker) { 1800ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1801ebf44d16STejun Heo start_worker(worker); 1802ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1803ebf44d16STejun Heo } 1804ebf44d16STejun Heo 1805cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1806cd549687STejun Heo 1807ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1808ebf44d16STejun Heo } 1809ebf44d16STejun Heo 1810ebf44d16STejun Heo /** 1811c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1812c34056a3STejun Heo * @worker: worker to be destroyed 1813c34056a3STejun Heo * 1814706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1815c8e55f36STejun Heo * 1816c8e55f36STejun Heo * CONTEXT: 1817d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1818c34056a3STejun Heo */ 1819c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1820c34056a3STejun Heo { 1821bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1822c34056a3STejun Heo 1823cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1824cd549687STejun Heo lockdep_assert_held(&pool->lock); 1825cd549687STejun Heo 1826c34056a3STejun Heo /* sanity check frenzy */ 18276183c009STejun Heo if (WARN_ON(worker->current_work) || 18286183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18296183c009STejun Heo return; 1830c34056a3STejun Heo 1831c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1832bd7bdd43STejun Heo pool->nr_workers--; 1833c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1834bd7bdd43STejun Heo pool->nr_idle--; 1835c8e55f36STejun Heo 1836c8e55f36STejun Heo list_del_init(&worker->entry); 1837cb444766STejun Heo worker->flags |= WORKER_DIE; 1838c8e55f36STejun Heo 1839822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1840822d8405STejun Heo 1841d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1842c8e55f36STejun Heo 1843c34056a3STejun Heo kthread_stop(worker->task); 1844c34056a3STejun Heo kfree(worker); 1845c34056a3STejun Heo 1846d565ed63STejun Heo spin_lock_irq(&pool->lock); 1847c34056a3STejun Heo } 1848c34056a3STejun Heo 184963d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1850e22bee78STejun Heo { 185163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1852e22bee78STejun Heo 1853d565ed63STejun Heo spin_lock_irq(&pool->lock); 1854e22bee78STejun Heo 185563d95a91STejun Heo if (too_many_workers(pool)) { 1856e22bee78STejun Heo struct worker *worker; 1857e22bee78STejun Heo unsigned long expires; 1858e22bee78STejun Heo 1859e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 186063d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1861e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1862e22bee78STejun Heo 1863e22bee78STejun Heo if (time_before(jiffies, expires)) 186463d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1865e22bee78STejun Heo else { 1866e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 186711ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 186863d95a91STejun Heo wake_up_worker(pool); 1869e22bee78STejun Heo } 1870e22bee78STejun Heo } 1871e22bee78STejun Heo 1872d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1873e22bee78STejun Heo } 1874e22bee78STejun Heo 1875493a1724STejun Heo static void send_mayday(struct work_struct *work) 1876e22bee78STejun Heo { 1877112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1878112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1879493a1724STejun Heo 18802e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1881e22bee78STejun Heo 1882493008a8STejun Heo if (!wq->rescuer) 1883493a1724STejun Heo return; 1884e22bee78STejun Heo 1885e22bee78STejun Heo /* mayday mayday mayday */ 1886493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1887493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1888e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1889493a1724STejun Heo } 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 1892706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1893e22bee78STejun Heo { 189463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1895e22bee78STejun Heo struct work_struct *work; 1896e22bee78STejun Heo 18972e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1898493a1724STejun Heo spin_lock(&pool->lock); 1899e22bee78STejun Heo 190063d95a91STejun Heo if (need_to_create_worker(pool)) { 1901e22bee78STejun Heo /* 1902e22bee78STejun Heo * We've been trying to create a new worker but 1903e22bee78STejun Heo * haven't been successful. We might be hitting an 1904e22bee78STejun Heo * allocation deadlock. Send distress signals to 1905e22bee78STejun Heo * rescuers. 1906e22bee78STejun Heo */ 190763d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1908e22bee78STejun Heo send_mayday(work); 1909e22bee78STejun Heo } 1910e22bee78STejun Heo 1911493a1724STejun Heo spin_unlock(&pool->lock); 19122e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1913e22bee78STejun Heo 191463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1915e22bee78STejun Heo } 1916e22bee78STejun Heo 1917e22bee78STejun Heo /** 1918e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 191963d95a91STejun Heo * @pool: pool to create a new worker for 1920e22bee78STejun Heo * 192163d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1922e22bee78STejun Heo * have at least one idle worker on return from this function. If 1923e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 192463d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1925e22bee78STejun Heo * possible allocation deadlock. 1926e22bee78STejun Heo * 1927c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1928c5aa87bbSTejun Heo * may_start_working() %true. 1929e22bee78STejun Heo * 1930e22bee78STejun Heo * LOCKING: 1931d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1932e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1933e22bee78STejun Heo * manager. 1934e22bee78STejun Heo * 1935e22bee78STejun Heo * RETURNS: 1936c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1937e22bee78STejun Heo * otherwise. 1938e22bee78STejun Heo */ 193963d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1940d565ed63STejun Heo __releases(&pool->lock) 1941d565ed63STejun Heo __acquires(&pool->lock) 1942e22bee78STejun Heo { 194363d95a91STejun Heo if (!need_to_create_worker(pool)) 1944e22bee78STejun Heo return false; 1945e22bee78STejun Heo restart: 1946d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19479f9c2364STejun Heo 1948e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 194963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1950e22bee78STejun Heo 1951e22bee78STejun Heo while (true) { 1952e22bee78STejun Heo struct worker *worker; 1953e22bee78STejun Heo 1954bc2ae0f5STejun Heo worker = create_worker(pool); 1955e22bee78STejun Heo if (worker) { 195663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1957d565ed63STejun Heo spin_lock_irq(&pool->lock); 1958e22bee78STejun Heo start_worker(worker); 19596183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19606183c009STejun Heo goto restart; 1961e22bee78STejun Heo return true; 1962e22bee78STejun Heo } 1963e22bee78STejun Heo 196463d95a91STejun Heo if (!need_to_create_worker(pool)) 1965e22bee78STejun Heo break; 1966e22bee78STejun Heo 1967e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1968e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19699f9c2364STejun Heo 197063d95a91STejun Heo if (!need_to_create_worker(pool)) 1971e22bee78STejun Heo break; 1972e22bee78STejun Heo } 1973e22bee78STejun Heo 197463d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1975d565ed63STejun Heo spin_lock_irq(&pool->lock); 197663d95a91STejun Heo if (need_to_create_worker(pool)) 1977e22bee78STejun Heo goto restart; 1978e22bee78STejun Heo return true; 1979e22bee78STejun Heo } 1980e22bee78STejun Heo 1981e22bee78STejun Heo /** 1982e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 198363d95a91STejun Heo * @pool: pool to destroy workers for 1984e22bee78STejun Heo * 198563d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1986e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1987e22bee78STejun Heo * 1988e22bee78STejun Heo * LOCKING: 1989d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1990e22bee78STejun Heo * multiple times. Called only from manager. 1991e22bee78STejun Heo * 1992e22bee78STejun Heo * RETURNS: 1993c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1994e22bee78STejun Heo * otherwise. 1995e22bee78STejun Heo */ 199663d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1997e22bee78STejun Heo { 1998e22bee78STejun Heo bool ret = false; 1999e22bee78STejun Heo 200063d95a91STejun Heo while (too_many_workers(pool)) { 2001e22bee78STejun Heo struct worker *worker; 2002e22bee78STejun Heo unsigned long expires; 2003e22bee78STejun Heo 200463d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2005e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2006e22bee78STejun Heo 2007e22bee78STejun Heo if (time_before(jiffies, expires)) { 200863d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2009e22bee78STejun Heo break; 2010e22bee78STejun Heo } 2011e22bee78STejun Heo 2012e22bee78STejun Heo destroy_worker(worker); 2013e22bee78STejun Heo ret = true; 2014e22bee78STejun Heo } 2015e22bee78STejun Heo 2016e22bee78STejun Heo return ret; 2017e22bee78STejun Heo } 2018e22bee78STejun Heo 2019e22bee78STejun Heo /** 2020e22bee78STejun Heo * manage_workers - manage worker pool 2021e22bee78STejun Heo * @worker: self 2022e22bee78STejun Heo * 2023706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2024e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2025706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2026e22bee78STejun Heo * 2027e22bee78STejun Heo * The caller can safely start processing works on false return. On 2028e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2029e22bee78STejun Heo * and may_start_working() is true. 2030e22bee78STejun Heo * 2031e22bee78STejun Heo * CONTEXT: 2032d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2033e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2034e22bee78STejun Heo * 2035e22bee78STejun Heo * RETURNS: 2036d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2037d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2038e22bee78STejun Heo */ 2039e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2040e22bee78STejun Heo { 204163d95a91STejun Heo struct worker_pool *pool = worker->pool; 2042e22bee78STejun Heo bool ret = false; 2043e22bee78STejun Heo 2044bc3a1afcSTejun Heo /* 2045bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2046bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2047bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2048bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2049bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2050bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2051bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2052bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2053bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2054bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2055bc3a1afcSTejun Heo * 2056bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2057bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2058bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2059bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2060bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2061bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2062bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2063bc3a1afcSTejun Heo * manager_mutex. 2064bc3a1afcSTejun Heo */ 206534a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2066e22bee78STejun Heo return ret; 2067e22bee78STejun Heo 2068ee378aa4SLai Jiangshan /* 2069bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2070bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2071ee378aa4SLai Jiangshan */ 2072bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2073d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2074bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 20758f174b11SJoonsoo Kim spin_lock_irq(&pool->lock); 2076ee378aa4SLai Jiangshan ret = true; 2077ee378aa4SLai Jiangshan } 2078ee378aa4SLai Jiangshan 207911ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2080e22bee78STejun Heo 2081e22bee78STejun Heo /* 2082e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2083e22bee78STejun Heo * on return. 2084e22bee78STejun Heo */ 208563d95a91STejun Heo ret |= maybe_destroy_workers(pool); 208663d95a91STejun Heo ret |= maybe_create_worker(pool); 2087e22bee78STejun Heo 2088bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 208934a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2090e22bee78STejun Heo return ret; 2091e22bee78STejun Heo } 2092e22bee78STejun Heo 2093a62428c0STejun Heo /** 2094a62428c0STejun Heo * process_one_work - process single work 2095c34056a3STejun Heo * @worker: self 2096a62428c0STejun Heo * @work: work to process 2097a62428c0STejun Heo * 2098a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2099a62428c0STejun Heo * process a single work including synchronization against and 2100a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2101a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2102a62428c0STejun Heo * call this function to process a work. 2103a62428c0STejun Heo * 2104a62428c0STejun Heo * CONTEXT: 2105d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2106a62428c0STejun Heo */ 2107c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2108d565ed63STejun Heo __releases(&pool->lock) 2109d565ed63STejun Heo __acquires(&pool->lock) 21101da177e4SLinus Torvalds { 2111112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2112bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2113112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 211473f53c4aSTejun Heo int work_color; 21157e11629dSTejun Heo struct worker *collision; 21164e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21174e6045f1SJohannes Berg /* 2118a62428c0STejun Heo * It is permissible to free the struct work_struct from 2119a62428c0STejun Heo * inside the function that is called from it, this we need to 2120a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2121a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2122a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21234e6045f1SJohannes Berg */ 21244d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21254d82a1deSPeter Zijlstra 21264d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21274e6045f1SJohannes Berg #endif 21286fec10a1STejun Heo /* 21296fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21306fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 213124647570STejun Heo * unbound or a disassociated pool. 21326fec10a1STejun Heo */ 21335f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 213424647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2135ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 213625511a47STejun Heo 21377e11629dSTejun Heo /* 21387e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21397e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21407e11629dSTejun Heo * already processing the work. If so, defer the work to the 21417e11629dSTejun Heo * currently executing one. 21427e11629dSTejun Heo */ 2143c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21447e11629dSTejun Heo if (unlikely(collision)) { 21457e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21467e11629dSTejun Heo return; 21477e11629dSTejun Heo } 21481da177e4SLinus Torvalds 21498930cabaSTejun Heo /* claim and dequeue */ 21501da177e4SLinus Torvalds debug_work_deactivate(work); 2151c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2152c34056a3STejun Heo worker->current_work = work; 2153a2c1c57bSTejun Heo worker->current_func = work->func; 2154112202d9STejun Heo worker->current_pwq = pwq; 215573f53c4aSTejun Heo work_color = get_work_color(work); 21567a22ad75STejun Heo 2157a62428c0STejun Heo list_del_init(&work->entry); 2158a62428c0STejun Heo 2159649027d7STejun Heo /* 2160fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2161fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2162fb0e7bebSTejun Heo */ 2163fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2164fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2165fb0e7bebSTejun Heo 2166974271c4STejun Heo /* 2167d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2168974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2169974271c4STejun Heo */ 217063d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 217163d95a91STejun Heo wake_up_worker(pool); 2172974271c4STejun Heo 21738930cabaSTejun Heo /* 21747c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2175d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 217623657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 217723657bb1STejun Heo * disabled. 21788930cabaSTejun Heo */ 21797c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21801da177e4SLinus Torvalds 2181d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2182365970a1SDavid Howells 2183112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21843295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2185e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2186a2c1c57bSTejun Heo worker->current_func(work); 2187e36c886aSArjan van de Ven /* 2188e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2189e36c886aSArjan van de Ven * point will only record its address. 2190e36c886aSArjan van de Ven */ 2191e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21923295f0efSIngo Molnar lock_map_release(&lockdep_map); 2193112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21941da177e4SLinus Torvalds 2195d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2196044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2197044c782cSValentin Ilie " last function: %pf\n", 2198a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2199a2c1c57bSTejun Heo worker->current_func); 2200d5abe669SPeter Zijlstra debug_show_held_locks(current); 2201d5abe669SPeter Zijlstra dump_stack(); 2202d5abe669SPeter Zijlstra } 2203d5abe669SPeter Zijlstra 2204b22ce278STejun Heo /* 2205b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2206b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2207b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2208b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2209b22ce278STejun Heo * stop_machine. 2210b22ce278STejun Heo */ 2211b22ce278STejun Heo cond_resched(); 2212b22ce278STejun Heo 2213d565ed63STejun Heo spin_lock_irq(&pool->lock); 2214a62428c0STejun Heo 2215fb0e7bebSTejun Heo /* clear cpu intensive status */ 2216fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2217fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2218fb0e7bebSTejun Heo 2219a62428c0STejun Heo /* we're done with it, release */ 222042f8570fSSasha Levin hash_del(&worker->hentry); 2221c34056a3STejun Heo worker->current_work = NULL; 2222a2c1c57bSTejun Heo worker->current_func = NULL; 2223112202d9STejun Heo worker->current_pwq = NULL; 22243d1cb205STejun Heo worker->desc_valid = false; 2225112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 22261da177e4SLinus Torvalds } 22271da177e4SLinus Torvalds 2228affee4b2STejun Heo /** 2229affee4b2STejun Heo * process_scheduled_works - process scheduled works 2230affee4b2STejun Heo * @worker: self 2231affee4b2STejun Heo * 2232affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2233affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2234affee4b2STejun Heo * fetches a work from the top and executes it. 2235affee4b2STejun Heo * 2236affee4b2STejun Heo * CONTEXT: 2237d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2238affee4b2STejun Heo * multiple times. 2239affee4b2STejun Heo */ 2240affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22411da177e4SLinus Torvalds { 2242affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2243affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2244a62428c0STejun Heo struct work_struct, entry); 2245c34056a3STejun Heo process_one_work(worker, work); 2246a62428c0STejun Heo } 22471da177e4SLinus Torvalds } 22481da177e4SLinus Torvalds 22494690c4abSTejun Heo /** 22504690c4abSTejun Heo * worker_thread - the worker thread function 2251c34056a3STejun Heo * @__worker: self 22524690c4abSTejun Heo * 2253c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2254c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2255c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2256c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2257c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 22584690c4abSTejun Heo */ 2259c34056a3STejun Heo static int worker_thread(void *__worker) 22601da177e4SLinus Torvalds { 2261c34056a3STejun Heo struct worker *worker = __worker; 2262bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22631da177e4SLinus Torvalds 2264e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2265e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2266c8e55f36STejun Heo woke_up: 2267d565ed63STejun Heo spin_lock_irq(&pool->lock); 2268affee4b2STejun Heo 2269a9ab775bSTejun Heo /* am I supposed to die? */ 2270a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2271d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2272a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2273e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2274c8e55f36STejun Heo return 0; 2275c8e55f36STejun Heo } 2276c8e55f36STejun Heo 2277c8e55f36STejun Heo worker_leave_idle(worker); 2278db7bccf4STejun Heo recheck: 2279e22bee78STejun Heo /* no more worker necessary? */ 228063d95a91STejun Heo if (!need_more_worker(pool)) 2281e22bee78STejun Heo goto sleep; 2282e22bee78STejun Heo 2283e22bee78STejun Heo /* do we need to manage? */ 228463d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2285e22bee78STejun Heo goto recheck; 2286e22bee78STejun Heo 2287c8e55f36STejun Heo /* 2288c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2289c8e55f36STejun Heo * preparing to process a work or actually processing it. 2290c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2291c8e55f36STejun Heo */ 22926183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2293c8e55f36STejun Heo 2294e22bee78STejun Heo /* 2295a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2296a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2297a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2298a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2299a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2300e22bee78STejun Heo */ 2301a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2302e22bee78STejun Heo 2303e22bee78STejun Heo do { 2304affee4b2STejun Heo struct work_struct *work = 2305bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2306affee4b2STejun Heo struct work_struct, entry); 2307affee4b2STejun Heo 2308c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2309affee4b2STejun Heo /* optimization path, not strictly necessary */ 2310affee4b2STejun Heo process_one_work(worker, work); 2311affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2312affee4b2STejun Heo process_scheduled_works(worker); 2313affee4b2STejun Heo } else { 2314c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2315affee4b2STejun Heo process_scheduled_works(worker); 2316affee4b2STejun Heo } 231763d95a91STejun Heo } while (keep_working(pool)); 2318affee4b2STejun Heo 2319e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2320d313dd85STejun Heo sleep: 232163d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2322e22bee78STejun Heo goto recheck; 2323d313dd85STejun Heo 2324c8e55f36STejun Heo /* 2325d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2326d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2327d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2328d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2329d565ed63STejun Heo * event. 2330c8e55f36STejun Heo */ 2331c8e55f36STejun Heo worker_enter_idle(worker); 2332c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2333d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23341da177e4SLinus Torvalds schedule(); 2335c8e55f36STejun Heo goto woke_up; 23361da177e4SLinus Torvalds } 23371da177e4SLinus Torvalds 2338e22bee78STejun Heo /** 2339e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2340111c225aSTejun Heo * @__rescuer: self 2341e22bee78STejun Heo * 2342e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2343493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2344e22bee78STejun Heo * 2345706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2346e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2347e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2348e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2349e22bee78STejun Heo * the problem rescuer solves. 2350e22bee78STejun Heo * 2351706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2352706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2353e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2354e22bee78STejun Heo * 2355e22bee78STejun Heo * This should happen rarely. 2356e22bee78STejun Heo */ 2357111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2358e22bee78STejun Heo { 2359111c225aSTejun Heo struct worker *rescuer = __rescuer; 2360111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2361e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2362e22bee78STejun Heo 2363e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2364111c225aSTejun Heo 2365111c225aSTejun Heo /* 2366111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2367111c225aSTejun Heo * doesn't participate in concurrency management. 2368111c225aSTejun Heo */ 2369111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2370e22bee78STejun Heo repeat: 2371e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23721da177e4SLinus Torvalds 2373412d32e6SMike Galbraith if (kthread_should_stop()) { 2374412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2375111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2376e22bee78STejun Heo return 0; 2377412d32e6SMike Galbraith } 23781da177e4SLinus Torvalds 2379493a1724STejun Heo /* see whether any pwq is asking for help */ 23802e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2381493a1724STejun Heo 2382493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2383493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2384493a1724STejun Heo struct pool_workqueue, mayday_node); 2385112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2386e22bee78STejun Heo struct work_struct *work, *n; 2387e22bee78STejun Heo 2388e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2389493a1724STejun Heo list_del_init(&pwq->mayday_node); 2390493a1724STejun Heo 23912e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2392e22bee78STejun Heo 2393e22bee78STejun Heo /* migrate to the target cpu if possible */ 2394f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2395b3104104SLai Jiangshan rescuer->pool = pool; 2396e22bee78STejun Heo 2397e22bee78STejun Heo /* 2398e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2399e22bee78STejun Heo * process'em. 2400e22bee78STejun Heo */ 24016183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2402bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2403112202d9STejun Heo if (get_work_pwq(work) == pwq) 2404e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2405e22bee78STejun Heo 2406e22bee78STejun Heo process_scheduled_works(rescuer); 24077576958aSTejun Heo 24087576958aSTejun Heo /* 2409d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 24107576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24117576958aSTejun Heo * and stalling the execution. 24127576958aSTejun Heo */ 241363d95a91STejun Heo if (keep_working(pool)) 241463d95a91STejun Heo wake_up_worker(pool); 24157576958aSTejun Heo 2416b3104104SLai Jiangshan rescuer->pool = NULL; 2417493a1724STejun Heo spin_unlock(&pool->lock); 24182e109a28STejun Heo spin_lock(&wq_mayday_lock); 24191da177e4SLinus Torvalds } 24201da177e4SLinus Torvalds 24212e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2422493a1724STejun Heo 2423111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2424111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2425e22bee78STejun Heo schedule(); 2426e22bee78STejun Heo goto repeat; 24271da177e4SLinus Torvalds } 24281da177e4SLinus Torvalds 2429fc2e4d70SOleg Nesterov struct wq_barrier { 2430fc2e4d70SOleg Nesterov struct work_struct work; 2431fc2e4d70SOleg Nesterov struct completion done; 2432fc2e4d70SOleg Nesterov }; 2433fc2e4d70SOleg Nesterov 2434fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2435fc2e4d70SOleg Nesterov { 2436fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2437fc2e4d70SOleg Nesterov complete(&barr->done); 2438fc2e4d70SOleg Nesterov } 2439fc2e4d70SOleg Nesterov 24404690c4abSTejun Heo /** 24414690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2442112202d9STejun Heo * @pwq: pwq to insert barrier into 24434690c4abSTejun Heo * @barr: wq_barrier to insert 2444affee4b2STejun Heo * @target: target work to attach @barr to 2445affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24464690c4abSTejun Heo * 2447affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2448affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2449affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2450affee4b2STejun Heo * cpu. 2451affee4b2STejun Heo * 2452affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2453affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2454affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2455affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2456affee4b2STejun Heo * after a work with LINKED flag set. 2457affee4b2STejun Heo * 2458affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2459112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24604690c4abSTejun Heo * 24614690c4abSTejun Heo * CONTEXT: 2462d565ed63STejun Heo * spin_lock_irq(pool->lock). 24634690c4abSTejun Heo */ 2464112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2465affee4b2STejun Heo struct wq_barrier *barr, 2466affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2467fc2e4d70SOleg Nesterov { 2468affee4b2STejun Heo struct list_head *head; 2469affee4b2STejun Heo unsigned int linked = 0; 2470affee4b2STejun Heo 2471dc186ad7SThomas Gleixner /* 2472d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2473dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2474dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2475dc186ad7SThomas Gleixner * might deadlock. 2476dc186ad7SThomas Gleixner */ 2477ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 247822df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2479fc2e4d70SOleg Nesterov init_completion(&barr->done); 248083c22520SOleg Nesterov 2481affee4b2STejun Heo /* 2482affee4b2STejun Heo * If @target is currently being executed, schedule the 2483affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2484affee4b2STejun Heo */ 2485affee4b2STejun Heo if (worker) 2486affee4b2STejun Heo head = worker->scheduled.next; 2487affee4b2STejun Heo else { 2488affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2489affee4b2STejun Heo 2490affee4b2STejun Heo head = target->entry.next; 2491affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2492affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2493affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2494affee4b2STejun Heo } 2495affee4b2STejun Heo 2496dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2497112202d9STejun Heo insert_work(pwq, &barr->work, head, 2498affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2499fc2e4d70SOleg Nesterov } 2500fc2e4d70SOleg Nesterov 250173f53c4aSTejun Heo /** 2502112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 250373f53c4aSTejun Heo * @wq: workqueue being flushed 250473f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 250573f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 250673f53c4aSTejun Heo * 2507112202d9STejun Heo * Prepare pwqs for workqueue flushing. 250873f53c4aSTejun Heo * 2509112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2510112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2511112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2512112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2513112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 251473f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 251573f53c4aSTejun Heo * 251673f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 251773f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 251873f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 251973f53c4aSTejun Heo * is returned. 252073f53c4aSTejun Heo * 2521112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 252273f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 252373f53c4aSTejun Heo * advanced to @work_color. 252473f53c4aSTejun Heo * 252573f53c4aSTejun Heo * CONTEXT: 25263c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 252773f53c4aSTejun Heo * 252873f53c4aSTejun Heo * RETURNS: 252973f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 253073f53c4aSTejun Heo * otherwise. 253173f53c4aSTejun Heo */ 2532112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 253373f53c4aSTejun Heo int flush_color, int work_color) 25341da177e4SLinus Torvalds { 253573f53c4aSTejun Heo bool wait = false; 253649e3cf44STejun Heo struct pool_workqueue *pwq; 25371da177e4SLinus Torvalds 253873f53c4aSTejun Heo if (flush_color >= 0) { 25396183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2540112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2541dc186ad7SThomas Gleixner } 254214441960SOleg Nesterov 254349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2544112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25451da177e4SLinus Torvalds 2546b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 254773f53c4aSTejun Heo 254873f53c4aSTejun Heo if (flush_color >= 0) { 25496183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 255073f53c4aSTejun Heo 2551112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2552112202d9STejun Heo pwq->flush_color = flush_color; 2553112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 255473f53c4aSTejun Heo wait = true; 25551da177e4SLinus Torvalds } 255673f53c4aSTejun Heo } 255773f53c4aSTejun Heo 255873f53c4aSTejun Heo if (work_color >= 0) { 25596183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2560112202d9STejun Heo pwq->work_color = work_color; 256173f53c4aSTejun Heo } 256273f53c4aSTejun Heo 2563b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25641da177e4SLinus Torvalds } 25651da177e4SLinus Torvalds 2566112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 256773f53c4aSTejun Heo complete(&wq->first_flusher->done); 256873f53c4aSTejun Heo 256973f53c4aSTejun Heo return wait; 257083c22520SOleg Nesterov } 25711da177e4SLinus Torvalds 25720fcb78c2SRolf Eike Beer /** 25731da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25740fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25751da177e4SLinus Torvalds * 2576c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2577c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25781da177e4SLinus Torvalds */ 25797ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25801da177e4SLinus Torvalds { 258173f53c4aSTejun Heo struct wq_flusher this_flusher = { 258273f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 258373f53c4aSTejun Heo .flush_color = -1, 258473f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 258573f53c4aSTejun Heo }; 258673f53c4aSTejun Heo int next_color; 2587b1f4ec17SOleg Nesterov 25883295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25893295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 259073f53c4aSTejun Heo 25913c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 259273f53c4aSTejun Heo 259373f53c4aSTejun Heo /* 259473f53c4aSTejun Heo * Start-to-wait phase 259573f53c4aSTejun Heo */ 259673f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 259773f53c4aSTejun Heo 259873f53c4aSTejun Heo if (next_color != wq->flush_color) { 259973f53c4aSTejun Heo /* 260073f53c4aSTejun Heo * Color space is not full. The current work_color 260173f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 260273f53c4aSTejun Heo * by one. 260373f53c4aSTejun Heo */ 26046183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 260573f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 260673f53c4aSTejun Heo wq->work_color = next_color; 260773f53c4aSTejun Heo 260873f53c4aSTejun Heo if (!wq->first_flusher) { 260973f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26106183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 261173f53c4aSTejun Heo 261273f53c4aSTejun Heo wq->first_flusher = &this_flusher; 261373f53c4aSTejun Heo 2614112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 261573f53c4aSTejun Heo wq->work_color)) { 261673f53c4aSTejun Heo /* nothing to flush, done */ 261773f53c4aSTejun Heo wq->flush_color = next_color; 261873f53c4aSTejun Heo wq->first_flusher = NULL; 261973f53c4aSTejun Heo goto out_unlock; 262073f53c4aSTejun Heo } 262173f53c4aSTejun Heo } else { 262273f53c4aSTejun Heo /* wait in queue */ 26236183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 262473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2625112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 262673f53c4aSTejun Heo } 262773f53c4aSTejun Heo } else { 262873f53c4aSTejun Heo /* 262973f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 263073f53c4aSTejun Heo * The next flush completion will assign us 263173f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 263273f53c4aSTejun Heo */ 263373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 263473f53c4aSTejun Heo } 263573f53c4aSTejun Heo 26363c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 263973f53c4aSTejun Heo 264073f53c4aSTejun Heo /* 264173f53c4aSTejun Heo * Wake-up-and-cascade phase 264273f53c4aSTejun Heo * 264373f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 264473f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 264573f53c4aSTejun Heo */ 264673f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 264773f53c4aSTejun Heo return; 264873f53c4aSTejun Heo 26493c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 265073f53c4aSTejun Heo 26514ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26524ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26534ce48b37STejun Heo goto out_unlock; 26544ce48b37STejun Heo 265573f53c4aSTejun Heo wq->first_flusher = NULL; 265673f53c4aSTejun Heo 26576183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26586183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 265973f53c4aSTejun Heo 266073f53c4aSTejun Heo while (true) { 266173f53c4aSTejun Heo struct wq_flusher *next, *tmp; 266273f53c4aSTejun Heo 266373f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 266473f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 266573f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 266673f53c4aSTejun Heo break; 266773f53c4aSTejun Heo list_del_init(&next->list); 266873f53c4aSTejun Heo complete(&next->done); 266973f53c4aSTejun Heo } 267073f53c4aSTejun Heo 26716183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 267273f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 267373f53c4aSTejun Heo 267473f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 267573f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 267673f53c4aSTejun Heo 267773f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 267873f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 267973f53c4aSTejun Heo /* 268073f53c4aSTejun Heo * Assign the same color to all overflowed 268173f53c4aSTejun Heo * flushers, advance work_color and append to 268273f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 268373f53c4aSTejun Heo * phase for these overflowed flushers. 268473f53c4aSTejun Heo */ 268573f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 268673f53c4aSTejun Heo tmp->flush_color = wq->work_color; 268773f53c4aSTejun Heo 268873f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 268973f53c4aSTejun Heo 269073f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 269173f53c4aSTejun Heo &wq->flusher_queue); 2692112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 269373f53c4aSTejun Heo } 269473f53c4aSTejun Heo 269573f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26966183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 269773f53c4aSTejun Heo break; 269873f53c4aSTejun Heo } 269973f53c4aSTejun Heo 270073f53c4aSTejun Heo /* 270173f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2702112202d9STejun Heo * the new first flusher and arm pwqs. 270373f53c4aSTejun Heo */ 27046183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27056183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 270673f53c4aSTejun Heo 270773f53c4aSTejun Heo list_del_init(&next->list); 270873f53c4aSTejun Heo wq->first_flusher = next; 270973f53c4aSTejun Heo 2710112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 271173f53c4aSTejun Heo break; 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo /* 271473f53c4aSTejun Heo * Meh... this color is already done, clear first 271573f53c4aSTejun Heo * flusher and repeat cascading. 271673f53c4aSTejun Heo */ 271773f53c4aSTejun Heo wq->first_flusher = NULL; 271873f53c4aSTejun Heo } 271973f53c4aSTejun Heo 272073f53c4aSTejun Heo out_unlock: 27213c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27221da177e4SLinus Torvalds } 2723ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27241da177e4SLinus Torvalds 27259c5a2ba7STejun Heo /** 27269c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27279c5a2ba7STejun Heo * @wq: workqueue to drain 27289c5a2ba7STejun Heo * 27299c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27309c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27319c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27329c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27339c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27349c5a2ba7STejun Heo * takes too long. 27359c5a2ba7STejun Heo */ 27369c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27379c5a2ba7STejun Heo { 27389c5a2ba7STejun Heo unsigned int flush_cnt = 0; 273949e3cf44STejun Heo struct pool_workqueue *pwq; 27409c5a2ba7STejun Heo 27419c5a2ba7STejun Heo /* 27429c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27439c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2744618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27459c5a2ba7STejun Heo */ 274687fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27479c5a2ba7STejun Heo if (!wq->nr_drainers++) 2748618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 274987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27509c5a2ba7STejun Heo reflush: 27519c5a2ba7STejun Heo flush_workqueue(wq); 27529c5a2ba7STejun Heo 2753b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 275476af4d93STejun Heo 275549e3cf44STejun Heo for_each_pwq(pwq, wq) { 2756fa2563e4SThomas Tuttle bool drained; 27579c5a2ba7STejun Heo 2758b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2759112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2760b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2761fa2563e4SThomas Tuttle 2762fa2563e4SThomas Tuttle if (drained) 27639c5a2ba7STejun Heo continue; 27649c5a2ba7STejun Heo 27659c5a2ba7STejun Heo if (++flush_cnt == 10 || 27669c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2767c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27689c5a2ba7STejun Heo wq->name, flush_cnt); 276976af4d93STejun Heo 2770b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27719c5a2ba7STejun Heo goto reflush; 27729c5a2ba7STejun Heo } 27739c5a2ba7STejun Heo 27749c5a2ba7STejun Heo if (!--wq->nr_drainers) 2775618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 277687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27779c5a2ba7STejun Heo } 27789c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27799c5a2ba7STejun Heo 2780606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2781baf59022STejun Heo { 2782baf59022STejun Heo struct worker *worker = NULL; 2783c9e7cf27STejun Heo struct worker_pool *pool; 2784112202d9STejun Heo struct pool_workqueue *pwq; 2785baf59022STejun Heo 2786baf59022STejun Heo might_sleep(); 2787baf59022STejun Heo 2788fa1b54e6STejun Heo local_irq_disable(); 2789fa1b54e6STejun Heo pool = get_work_pool(work); 2790fa1b54e6STejun Heo if (!pool) { 2791fa1b54e6STejun Heo local_irq_enable(); 2792fa1b54e6STejun Heo return false; 2793fa1b54e6STejun Heo } 2794fa1b54e6STejun Heo 2795fa1b54e6STejun Heo spin_lock(&pool->lock); 27960b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2797112202d9STejun Heo pwq = get_work_pwq(work); 2798112202d9STejun Heo if (pwq) { 2799112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2800baf59022STejun Heo goto already_gone; 2801606a5020STejun Heo } else { 2802c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2803baf59022STejun Heo if (!worker) 2804baf59022STejun Heo goto already_gone; 2805112202d9STejun Heo pwq = worker->current_pwq; 2806606a5020STejun Heo } 2807baf59022STejun Heo 2808112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2809d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2810baf59022STejun Heo 2811e159489bSTejun Heo /* 2812e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2813e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2814e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2815e159489bSTejun Heo * access. 2816e159489bSTejun Heo */ 2817493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2818112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2819e159489bSTejun Heo else 2820112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2821112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2822e159489bSTejun Heo 2823baf59022STejun Heo return true; 2824baf59022STejun Heo already_gone: 2825d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2826baf59022STejun Heo return false; 2827baf59022STejun Heo } 2828baf59022STejun Heo 2829c2fda509SLai Jiangshan static bool __flush_work(struct work_struct *work) 2830c2fda509SLai Jiangshan { 2831c2fda509SLai Jiangshan struct wq_barrier barr; 2832c2fda509SLai Jiangshan 2833c2fda509SLai Jiangshan if (start_flush_work(work, &barr)) { 2834c2fda509SLai Jiangshan wait_for_completion(&barr.done); 2835c2fda509SLai Jiangshan destroy_work_on_stack(&barr.work); 2836c2fda509SLai Jiangshan return true; 2837c2fda509SLai Jiangshan } else { 2838c2fda509SLai Jiangshan return false; 2839c2fda509SLai Jiangshan } 2840c2fda509SLai Jiangshan } 2841c2fda509SLai Jiangshan 2842db700897SOleg Nesterov /** 2843401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2844401a8d04STejun Heo * @work: the work to flush 2845db700897SOleg Nesterov * 2846606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2847606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2848401a8d04STejun Heo * 2849401a8d04STejun Heo * RETURNS: 2850401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2851401a8d04STejun Heo * %false if it was already idle. 2852db700897SOleg Nesterov */ 2853401a8d04STejun Heo bool flush_work(struct work_struct *work) 2854db700897SOleg Nesterov { 28550976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28560976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28570976dfc1SStephen Boyd 2858c2fda509SLai Jiangshan return __flush_work(work); 2859606a5020STejun Heo } 2860db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2861db700897SOleg Nesterov 286236e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2863401a8d04STejun Heo { 2864bbb68dfaSTejun Heo unsigned long flags; 28651f1f642eSOleg Nesterov int ret; 28661f1f642eSOleg Nesterov 28671f1f642eSOleg Nesterov do { 2868bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2869bbb68dfaSTejun Heo /* 2870bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2871bbb68dfaSTejun Heo * would be waiting for before retrying. 2872bbb68dfaSTejun Heo */ 2873bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2874606a5020STejun Heo flush_work(work); 28751f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28761f1f642eSOleg Nesterov 2877bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2878bbb68dfaSTejun Heo mark_work_canceling(work); 2879bbb68dfaSTejun Heo local_irq_restore(flags); 2880bbb68dfaSTejun Heo 2881606a5020STejun Heo flush_work(work); 28827a22ad75STejun Heo clear_work_data(work); 28831f1f642eSOleg Nesterov return ret; 28841f1f642eSOleg Nesterov } 28851f1f642eSOleg Nesterov 28866e84d644SOleg Nesterov /** 2887401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2888401a8d04STejun Heo * @work: the work to cancel 28896e84d644SOleg Nesterov * 2890401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2891401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2892401a8d04STejun Heo * another workqueue. On return from this function, @work is 2893401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28941f1f642eSOleg Nesterov * 2895401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2896401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28976e84d644SOleg Nesterov * 2898401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28996e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2900401a8d04STejun Heo * 2901401a8d04STejun Heo * RETURNS: 2902401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29036e84d644SOleg Nesterov */ 2904401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29056e84d644SOleg Nesterov { 290636e227d2STejun Heo return __cancel_work_timer(work, false); 2907b89deed3SOleg Nesterov } 290828e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2909b89deed3SOleg Nesterov 29106e84d644SOleg Nesterov /** 2911401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2912401a8d04STejun Heo * @dwork: the delayed work to flush 29136e84d644SOleg Nesterov * 2914401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2915401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2916401a8d04STejun Heo * considers the last queueing instance of @dwork. 29171f1f642eSOleg Nesterov * 2918401a8d04STejun Heo * RETURNS: 2919401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2920401a8d04STejun Heo * %false if it was already idle. 29216e84d644SOleg Nesterov */ 2922401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2923401a8d04STejun Heo { 29248930cabaSTejun Heo local_irq_disable(); 2925401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 292660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29278930cabaSTejun Heo local_irq_enable(); 2928401a8d04STejun Heo return flush_work(&dwork->work); 2929401a8d04STejun Heo } 2930401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2931401a8d04STejun Heo 2932401a8d04STejun Heo /** 293357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 293457b30ae7STejun Heo * @dwork: delayed_work to cancel 293509383498STejun Heo * 293657b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 293757b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 293857b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 293957b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 294057b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 294109383498STejun Heo * 294257b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 294309383498STejun Heo */ 294457b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 294509383498STejun Heo { 294657b30ae7STejun Heo unsigned long flags; 294757b30ae7STejun Heo int ret; 294857b30ae7STejun Heo 294957b30ae7STejun Heo do { 295057b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 295157b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 295257b30ae7STejun Heo 295357b30ae7STejun Heo if (unlikely(ret < 0)) 295457b30ae7STejun Heo return false; 295557b30ae7STejun Heo 29567c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29577c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 295857b30ae7STejun Heo local_irq_restore(flags); 2959c0158ca6SDan Magenheimer return ret; 296009383498STejun Heo } 296157b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 296209383498STejun Heo 296309383498STejun Heo /** 2964401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2965401a8d04STejun Heo * @dwork: the delayed work cancel 2966401a8d04STejun Heo * 2967401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2968401a8d04STejun Heo * 2969401a8d04STejun Heo * RETURNS: 2970401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2971401a8d04STejun Heo */ 2972401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29736e84d644SOleg Nesterov { 297436e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29756e84d644SOleg Nesterov } 2976f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29771da177e4SLinus Torvalds 29780fcb78c2SRolf Eike Beer /** 297931ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2980b6136773SAndrew Morton * @func: the function to call 2981b6136773SAndrew Morton * 298231ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 298331ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2984b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 298531ddd871STejun Heo * 298631ddd871STejun Heo * RETURNS: 298731ddd871STejun Heo * 0 on success, -errno on failure. 2988b6136773SAndrew Morton */ 298965f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 299015316ba8SChristoph Lameter { 299115316ba8SChristoph Lameter int cpu; 299238f51568SNamhyung Kim struct work_struct __percpu *works; 299315316ba8SChristoph Lameter 2994b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2995b6136773SAndrew Morton if (!works) 299615316ba8SChristoph Lameter return -ENOMEM; 2997b6136773SAndrew Morton 299895402b38SGautham R Shenoy get_online_cpus(); 299993981800STejun Heo 300015316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30019bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30029bfb1839SIngo Molnar 30039bfb1839SIngo Molnar INIT_WORK(work, func); 30048de6d308SOleg Nesterov schedule_work_on(cpu, work); 300515316ba8SChristoph Lameter } 300693981800STejun Heo 300793981800STejun Heo for_each_online_cpu(cpu) 30088616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 300993981800STejun Heo 301095402b38SGautham R Shenoy put_online_cpus(); 3011b6136773SAndrew Morton free_percpu(works); 301215316ba8SChristoph Lameter return 0; 301315316ba8SChristoph Lameter } 301415316ba8SChristoph Lameter 3015eef6a7d5SAlan Stern /** 3016eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3017eef6a7d5SAlan Stern * 3018eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3019eef6a7d5SAlan Stern * completion. 3020eef6a7d5SAlan Stern * 3021eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3022eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3023eef6a7d5SAlan Stern * will lead to deadlock: 3024eef6a7d5SAlan Stern * 3025eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3026eef6a7d5SAlan Stern * a lock held by your code or its caller. 3027eef6a7d5SAlan Stern * 3028eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3029eef6a7d5SAlan Stern * 3030eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3031eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3032eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3033eef6a7d5SAlan Stern * 3034eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3035eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3036eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3037eef6a7d5SAlan Stern * cancel_work_sync() instead. 3038eef6a7d5SAlan Stern */ 30391da177e4SLinus Torvalds void flush_scheduled_work(void) 30401da177e4SLinus Torvalds { 3041d320c038STejun Heo flush_workqueue(system_wq); 30421da177e4SLinus Torvalds } 3043ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30441da177e4SLinus Torvalds 30451da177e4SLinus Torvalds /** 30461fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30471fa44ecaSJames Bottomley * @fn: the function to execute 30481fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30491fa44ecaSJames Bottomley * be available when the work executes) 30501fa44ecaSJames Bottomley * 30511fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30521fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30531fa44ecaSJames Bottomley * 30541fa44ecaSJames Bottomley * Returns: 0 - function was executed 30551fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30561fa44ecaSJames Bottomley */ 305765f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30581fa44ecaSJames Bottomley { 30591fa44ecaSJames Bottomley if (!in_interrupt()) { 306065f27f38SDavid Howells fn(&ew->work); 30611fa44ecaSJames Bottomley return 0; 30621fa44ecaSJames Bottomley } 30631fa44ecaSJames Bottomley 306465f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30651fa44ecaSJames Bottomley schedule_work(&ew->work); 30661fa44ecaSJames Bottomley 30671fa44ecaSJames Bottomley return 1; 30681fa44ecaSJames Bottomley } 30691fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30701fa44ecaSJames Bottomley 3071226223abSTejun Heo #ifdef CONFIG_SYSFS 3072226223abSTejun Heo /* 3073226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3074226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3075226223abSTejun Heo * following attributes. 3076226223abSTejun Heo * 3077226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3078226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3079226223abSTejun Heo * 3080226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3081226223abSTejun Heo * 3082226223abSTejun Heo * id RO int : the associated pool ID 3083226223abSTejun Heo * nice RW int : nice value of the workers 3084226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3085226223abSTejun Heo */ 3086226223abSTejun Heo struct wq_device { 3087226223abSTejun Heo struct workqueue_struct *wq; 3088226223abSTejun Heo struct device dev; 3089226223abSTejun Heo }; 3090226223abSTejun Heo 3091226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3092226223abSTejun Heo { 3093226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3094226223abSTejun Heo 3095226223abSTejun Heo return wq_dev->wq; 3096226223abSTejun Heo } 3097226223abSTejun Heo 3098226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3099226223abSTejun Heo struct device_attribute *attr, char *buf) 3100226223abSTejun Heo { 3101226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3102226223abSTejun Heo 3103226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3104226223abSTejun Heo } 3105226223abSTejun Heo 3106226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3107226223abSTejun Heo struct device_attribute *attr, char *buf) 3108226223abSTejun Heo { 3109226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3110226223abSTejun Heo 3111226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3112226223abSTejun Heo } 3113226223abSTejun Heo 3114226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3115226223abSTejun Heo struct device_attribute *attr, 3116226223abSTejun Heo const char *buf, size_t count) 3117226223abSTejun Heo { 3118226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3119226223abSTejun Heo int val; 3120226223abSTejun Heo 3121226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3122226223abSTejun Heo return -EINVAL; 3123226223abSTejun Heo 3124226223abSTejun Heo workqueue_set_max_active(wq, val); 3125226223abSTejun Heo return count; 3126226223abSTejun Heo } 3127226223abSTejun Heo 3128226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3129226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3130226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3131226223abSTejun Heo __ATTR_NULL, 3132226223abSTejun Heo }; 3133226223abSTejun Heo 3134d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3135226223abSTejun Heo struct device_attribute *attr, char *buf) 3136226223abSTejun Heo { 3137226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3138d55262c4STejun Heo const char *delim = ""; 3139d55262c4STejun Heo int node, written = 0; 3140226223abSTejun Heo 3141226223abSTejun Heo rcu_read_lock_sched(); 3142d55262c4STejun Heo for_each_node(node) { 3143d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3144d55262c4STejun Heo "%s%d:%d", delim, node, 3145d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3146d55262c4STejun Heo delim = " "; 3147d55262c4STejun Heo } 3148d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3149226223abSTejun Heo rcu_read_unlock_sched(); 3150226223abSTejun Heo 3151226223abSTejun Heo return written; 3152226223abSTejun Heo } 3153226223abSTejun Heo 3154226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3155226223abSTejun Heo char *buf) 3156226223abSTejun Heo { 3157226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3158226223abSTejun Heo int written; 3159226223abSTejun Heo 31606029a918STejun Heo mutex_lock(&wq->mutex); 31616029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 31626029a918STejun Heo mutex_unlock(&wq->mutex); 3163226223abSTejun Heo 3164226223abSTejun Heo return written; 3165226223abSTejun Heo } 3166226223abSTejun Heo 3167226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3168226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3169226223abSTejun Heo { 3170226223abSTejun Heo struct workqueue_attrs *attrs; 3171226223abSTejun Heo 3172226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3173226223abSTejun Heo if (!attrs) 3174226223abSTejun Heo return NULL; 3175226223abSTejun Heo 31766029a918STejun Heo mutex_lock(&wq->mutex); 31776029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 31786029a918STejun Heo mutex_unlock(&wq->mutex); 3179226223abSTejun Heo return attrs; 3180226223abSTejun Heo } 3181226223abSTejun Heo 3182226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3183226223abSTejun Heo const char *buf, size_t count) 3184226223abSTejun Heo { 3185226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3186226223abSTejun Heo struct workqueue_attrs *attrs; 3187226223abSTejun Heo int ret; 3188226223abSTejun Heo 3189226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3190226223abSTejun Heo if (!attrs) 3191226223abSTejun Heo return -ENOMEM; 3192226223abSTejun Heo 3193226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3194226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3195226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3196226223abSTejun Heo else 3197226223abSTejun Heo ret = -EINVAL; 3198226223abSTejun Heo 3199226223abSTejun Heo free_workqueue_attrs(attrs); 3200226223abSTejun Heo return ret ?: count; 3201226223abSTejun Heo } 3202226223abSTejun Heo 3203226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3204226223abSTejun Heo struct device_attribute *attr, char *buf) 3205226223abSTejun Heo { 3206226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3207226223abSTejun Heo int written; 3208226223abSTejun Heo 32096029a918STejun Heo mutex_lock(&wq->mutex); 32106029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 32116029a918STejun Heo mutex_unlock(&wq->mutex); 3212226223abSTejun Heo 3213226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3214226223abSTejun Heo return written; 3215226223abSTejun Heo } 3216226223abSTejun Heo 3217226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3218226223abSTejun Heo struct device_attribute *attr, 3219226223abSTejun Heo const char *buf, size_t count) 3220226223abSTejun Heo { 3221226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3222226223abSTejun Heo struct workqueue_attrs *attrs; 3223226223abSTejun Heo int ret; 3224226223abSTejun Heo 3225226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3226226223abSTejun Heo if (!attrs) 3227226223abSTejun Heo return -ENOMEM; 3228226223abSTejun Heo 3229226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3230226223abSTejun Heo if (!ret) 3231226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3232226223abSTejun Heo 3233226223abSTejun Heo free_workqueue_attrs(attrs); 3234226223abSTejun Heo return ret ?: count; 3235226223abSTejun Heo } 3236226223abSTejun Heo 3237d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3238d55262c4STejun Heo char *buf) 3239d55262c4STejun Heo { 3240d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3241d55262c4STejun Heo int written; 3242d55262c4STejun Heo 3243d55262c4STejun Heo mutex_lock(&wq->mutex); 3244d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3245d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3246d55262c4STejun Heo mutex_unlock(&wq->mutex); 3247d55262c4STejun Heo 3248d55262c4STejun Heo return written; 3249d55262c4STejun Heo } 3250d55262c4STejun Heo 3251d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3252d55262c4STejun Heo const char *buf, size_t count) 3253d55262c4STejun Heo { 3254d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3255d55262c4STejun Heo struct workqueue_attrs *attrs; 3256d55262c4STejun Heo int v, ret; 3257d55262c4STejun Heo 3258d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3259d55262c4STejun Heo if (!attrs) 3260d55262c4STejun Heo return -ENOMEM; 3261d55262c4STejun Heo 3262d55262c4STejun Heo ret = -EINVAL; 3263d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3264d55262c4STejun Heo attrs->no_numa = !v; 3265d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3266d55262c4STejun Heo } 3267d55262c4STejun Heo 3268d55262c4STejun Heo free_workqueue_attrs(attrs); 3269d55262c4STejun Heo return ret ?: count; 3270d55262c4STejun Heo } 3271d55262c4STejun Heo 3272226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3273d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3274226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3275226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3276d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3277226223abSTejun Heo __ATTR_NULL, 3278226223abSTejun Heo }; 3279226223abSTejun Heo 3280226223abSTejun Heo static struct bus_type wq_subsys = { 3281226223abSTejun Heo .name = "workqueue", 3282226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3283226223abSTejun Heo }; 3284226223abSTejun Heo 3285226223abSTejun Heo static int __init wq_sysfs_init(void) 3286226223abSTejun Heo { 3287226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3288226223abSTejun Heo } 3289226223abSTejun Heo core_initcall(wq_sysfs_init); 3290226223abSTejun Heo 3291226223abSTejun Heo static void wq_device_release(struct device *dev) 3292226223abSTejun Heo { 3293226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3294226223abSTejun Heo 3295226223abSTejun Heo kfree(wq_dev); 3296226223abSTejun Heo } 3297226223abSTejun Heo 3298226223abSTejun Heo /** 3299226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3300226223abSTejun Heo * @wq: the workqueue to register 3301226223abSTejun Heo * 3302226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3303226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3304226223abSTejun Heo * which is the preferred method. 3305226223abSTejun Heo * 3306226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3307226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3308226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3309226223abSTejun Heo * attributes. 3310226223abSTejun Heo * 3311226223abSTejun Heo * Returns 0 on success, -errno on failure. 3312226223abSTejun Heo */ 3313226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3314226223abSTejun Heo { 3315226223abSTejun Heo struct wq_device *wq_dev; 3316226223abSTejun Heo int ret; 3317226223abSTejun Heo 3318226223abSTejun Heo /* 3319226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3320226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3321226223abSTejun Heo * workqueues. 3322226223abSTejun Heo */ 3323226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3324226223abSTejun Heo return -EINVAL; 3325226223abSTejun Heo 3326226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3327226223abSTejun Heo if (!wq_dev) 3328226223abSTejun Heo return -ENOMEM; 3329226223abSTejun Heo 3330226223abSTejun Heo wq_dev->wq = wq; 3331226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3332226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3333226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3334226223abSTejun Heo 3335226223abSTejun Heo /* 3336226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3337226223abSTejun Heo * everything is ready. 3338226223abSTejun Heo */ 3339226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3340226223abSTejun Heo 3341226223abSTejun Heo ret = device_register(&wq_dev->dev); 3342226223abSTejun Heo if (ret) { 3343226223abSTejun Heo kfree(wq_dev); 3344226223abSTejun Heo wq->wq_dev = NULL; 3345226223abSTejun Heo return ret; 3346226223abSTejun Heo } 3347226223abSTejun Heo 3348226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3349226223abSTejun Heo struct device_attribute *attr; 3350226223abSTejun Heo 3351226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3352226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3353226223abSTejun Heo if (ret) { 3354226223abSTejun Heo device_unregister(&wq_dev->dev); 3355226223abSTejun Heo wq->wq_dev = NULL; 3356226223abSTejun Heo return ret; 3357226223abSTejun Heo } 3358226223abSTejun Heo } 3359226223abSTejun Heo } 3360226223abSTejun Heo 3361226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3362226223abSTejun Heo return 0; 3363226223abSTejun Heo } 3364226223abSTejun Heo 3365226223abSTejun Heo /** 3366226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3367226223abSTejun Heo * @wq: the workqueue to unregister 3368226223abSTejun Heo * 3369226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3370226223abSTejun Heo */ 3371226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3372226223abSTejun Heo { 3373226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3374226223abSTejun Heo 3375226223abSTejun Heo if (!wq->wq_dev) 3376226223abSTejun Heo return; 3377226223abSTejun Heo 3378226223abSTejun Heo wq->wq_dev = NULL; 3379226223abSTejun Heo device_unregister(&wq_dev->dev); 3380226223abSTejun Heo } 3381226223abSTejun Heo #else /* CONFIG_SYSFS */ 3382226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3383226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3384226223abSTejun Heo 33857a4e344cSTejun Heo /** 33867a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33877a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33887a4e344cSTejun Heo * 33897a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33907a4e344cSTejun Heo */ 33917a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33927a4e344cSTejun Heo { 33937a4e344cSTejun Heo if (attrs) { 33947a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33957a4e344cSTejun Heo kfree(attrs); 33967a4e344cSTejun Heo } 33977a4e344cSTejun Heo } 33987a4e344cSTejun Heo 33997a4e344cSTejun Heo /** 34007a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 34017a4e344cSTejun Heo * @gfp_mask: allocation mask to use 34027a4e344cSTejun Heo * 34037a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 34047a4e344cSTejun Heo * return it. Returns NULL on failure. 34057a4e344cSTejun Heo */ 34067a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 34077a4e344cSTejun Heo { 34087a4e344cSTejun Heo struct workqueue_attrs *attrs; 34097a4e344cSTejun Heo 34107a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 34117a4e344cSTejun Heo if (!attrs) 34127a4e344cSTejun Heo goto fail; 34137a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 34147a4e344cSTejun Heo goto fail; 34157a4e344cSTejun Heo 341613e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 34177a4e344cSTejun Heo return attrs; 34187a4e344cSTejun Heo fail: 34197a4e344cSTejun Heo free_workqueue_attrs(attrs); 34207a4e344cSTejun Heo return NULL; 34217a4e344cSTejun Heo } 34227a4e344cSTejun Heo 342329c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 342429c91e99STejun Heo const struct workqueue_attrs *from) 342529c91e99STejun Heo { 342629c91e99STejun Heo to->nice = from->nice; 342729c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 34282865a8fbSShaohua Li /* 34292865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 34302865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 34312865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 34322865a8fbSShaohua Li */ 34332865a8fbSShaohua Li to->no_numa = from->no_numa; 343429c91e99STejun Heo } 343529c91e99STejun Heo 343629c91e99STejun Heo /* hash value of the content of @attr */ 343729c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 343829c91e99STejun Heo { 343929c91e99STejun Heo u32 hash = 0; 344029c91e99STejun Heo 344129c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 344213e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 344313e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 344429c91e99STejun Heo return hash; 344529c91e99STejun Heo } 344629c91e99STejun Heo 344729c91e99STejun Heo /* content equality test */ 344829c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 344929c91e99STejun Heo const struct workqueue_attrs *b) 345029c91e99STejun Heo { 345129c91e99STejun Heo if (a->nice != b->nice) 345229c91e99STejun Heo return false; 345329c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 345429c91e99STejun Heo return false; 345529c91e99STejun Heo return true; 345629c91e99STejun Heo } 345729c91e99STejun Heo 34587a4e344cSTejun Heo /** 34597a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34607a4e344cSTejun Heo * @pool: worker_pool to initialize 34617a4e344cSTejun Heo * 34627a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 346329c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 346429c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 346529c91e99STejun Heo * on @pool safely to release it. 34667a4e344cSTejun Heo */ 34677a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34684e1a1f9aSTejun Heo { 34694e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 347029c91e99STejun Heo pool->id = -1; 347129c91e99STejun Heo pool->cpu = -1; 3472f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34734e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 34744e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34754e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34764e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34774e1a1f9aSTejun Heo 34784e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 34794e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 34804e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 34814e1a1f9aSTejun Heo 34824e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 34834e1a1f9aSTejun Heo (unsigned long)pool); 34844e1a1f9aSTejun Heo 34854e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3486bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3487822d8405STejun Heo idr_init(&pool->worker_idr); 34887a4e344cSTejun Heo 348929c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 349029c91e99STejun Heo pool->refcnt = 1; 349129c91e99STejun Heo 349229c91e99STejun Heo /* shouldn't fail above this point */ 34937a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34947a4e344cSTejun Heo if (!pool->attrs) 34957a4e344cSTejun Heo return -ENOMEM; 34967a4e344cSTejun Heo return 0; 34974e1a1f9aSTejun Heo } 34984e1a1f9aSTejun Heo 349929c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 350029c91e99STejun Heo { 350129c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 350229c91e99STejun Heo 3503822d8405STejun Heo idr_destroy(&pool->worker_idr); 350429c91e99STejun Heo free_workqueue_attrs(pool->attrs); 350529c91e99STejun Heo kfree(pool); 350629c91e99STejun Heo } 350729c91e99STejun Heo 350829c91e99STejun Heo /** 350929c91e99STejun Heo * put_unbound_pool - put a worker_pool 351029c91e99STejun Heo * @pool: worker_pool to put 351129c91e99STejun Heo * 351229c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3513c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3514c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3515c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3516a892caccSTejun Heo * 3517a892caccSTejun Heo * Should be called with wq_pool_mutex held. 351829c91e99STejun Heo */ 351929c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 352029c91e99STejun Heo { 352129c91e99STejun Heo struct worker *worker; 352229c91e99STejun Heo 3523a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3524a892caccSTejun Heo 3525a892caccSTejun Heo if (--pool->refcnt) 352629c91e99STejun Heo return; 352729c91e99STejun Heo 352829c91e99STejun Heo /* sanity checks */ 352929c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3530a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 353129c91e99STejun Heo return; 353229c91e99STejun Heo 353329c91e99STejun Heo /* release id and unhash */ 353429c91e99STejun Heo if (pool->id >= 0) 353529c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 353629c91e99STejun Heo hash_del(&pool->hash_node); 353729c91e99STejun Heo 3538c5aa87bbSTejun Heo /* 3539c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3540c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3541c5aa87bbSTejun Heo * manager_mutex. 3542c5aa87bbSTejun Heo */ 354329c91e99STejun Heo mutex_lock(&pool->manager_arb); 3544cd549687STejun Heo mutex_lock(&pool->manager_mutex); 354529c91e99STejun Heo spin_lock_irq(&pool->lock); 354629c91e99STejun Heo 354729c91e99STejun Heo while ((worker = first_worker(pool))) 354829c91e99STejun Heo destroy_worker(worker); 354929c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 355029c91e99STejun Heo 355129c91e99STejun Heo spin_unlock_irq(&pool->lock); 3552cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 355329c91e99STejun Heo mutex_unlock(&pool->manager_arb); 355429c91e99STejun Heo 355529c91e99STejun Heo /* shut down the timers */ 355629c91e99STejun Heo del_timer_sync(&pool->idle_timer); 355729c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 355829c91e99STejun Heo 355929c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 356029c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 356129c91e99STejun Heo } 356229c91e99STejun Heo 356329c91e99STejun Heo /** 356429c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 356529c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 356629c91e99STejun Heo * 356729c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 356829c91e99STejun Heo * reference count and return it. If there already is a matching 356929c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 357029c91e99STejun Heo * create a new one. On failure, returns NULL. 3571a892caccSTejun Heo * 3572a892caccSTejun Heo * Should be called with wq_pool_mutex held. 357329c91e99STejun Heo */ 357429c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 357529c91e99STejun Heo { 357629c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 357729c91e99STejun Heo struct worker_pool *pool; 3578f3f90ad4STejun Heo int node; 357929c91e99STejun Heo 3580a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 358129c91e99STejun Heo 358229c91e99STejun Heo /* do we already have a matching pool? */ 358329c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 358429c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 358529c91e99STejun Heo pool->refcnt++; 358629c91e99STejun Heo goto out_unlock; 358729c91e99STejun Heo } 358829c91e99STejun Heo } 358929c91e99STejun Heo 359029c91e99STejun Heo /* nope, create a new one */ 359129c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 359229c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 359329c91e99STejun Heo goto fail; 359429c91e99STejun Heo 359512ee4fc6SLai Jiangshan if (workqueue_freezing) 359612ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 359712ee4fc6SLai Jiangshan 35988864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 359929c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 360029c91e99STejun Heo 36012865a8fbSShaohua Li /* 36022865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36032865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36042865a8fbSShaohua Li */ 36052865a8fbSShaohua Li pool->attrs->no_numa = false; 36062865a8fbSShaohua Li 3607f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3608f3f90ad4STejun Heo if (wq_numa_enabled) { 3609f3f90ad4STejun Heo for_each_node(node) { 3610f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3611f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3612f3f90ad4STejun Heo pool->node = node; 3613f3f90ad4STejun Heo break; 3614f3f90ad4STejun Heo } 3615f3f90ad4STejun Heo } 3616f3f90ad4STejun Heo } 3617f3f90ad4STejun Heo 361829c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 361929c91e99STejun Heo goto fail; 362029c91e99STejun Heo 362129c91e99STejun Heo /* create and start the initial worker */ 3622ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 362329c91e99STejun Heo goto fail; 362429c91e99STejun Heo 362529c91e99STejun Heo /* install */ 362629c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 362729c91e99STejun Heo out_unlock: 362829c91e99STejun Heo return pool; 362929c91e99STejun Heo fail: 363029c91e99STejun Heo if (pool) 363129c91e99STejun Heo put_unbound_pool(pool); 363229c91e99STejun Heo return NULL; 363329c91e99STejun Heo } 363429c91e99STejun Heo 36358864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36368864b4e5STejun Heo { 36378864b4e5STejun Heo kmem_cache_free(pwq_cache, 36388864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36398864b4e5STejun Heo } 36408864b4e5STejun Heo 36418864b4e5STejun Heo /* 36428864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36438864b4e5STejun Heo * and needs to be destroyed. 36448864b4e5STejun Heo */ 36458864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36468864b4e5STejun Heo { 36478864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36488864b4e5STejun Heo unbound_release_work); 36498864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36508864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3651bc0caf09STejun Heo bool is_last; 36528864b4e5STejun Heo 36538864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36548864b4e5STejun Heo return; 36558864b4e5STejun Heo 365675ccf595STejun Heo /* 36573c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 365875ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 365975ccf595STejun Heo * and consistent with the linking path. 366075ccf595STejun Heo */ 36613c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 36628864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3663bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 36643c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 36658864b4e5STejun Heo 3666a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 36678864b4e5STejun Heo put_unbound_pool(pool); 3668a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3669a892caccSTejun Heo 36708864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 36718864b4e5STejun Heo 36728864b4e5STejun Heo /* 36738864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 36748864b4e5STejun Heo * is gonna access it anymore. Free it. 36758864b4e5STejun Heo */ 36766029a918STejun Heo if (is_last) { 36776029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 36788864b4e5STejun Heo kfree(wq); 36798864b4e5STejun Heo } 36806029a918STejun Heo } 36818864b4e5STejun Heo 36820fbd95aaSTejun Heo /** 3683699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36840fbd95aaSTejun Heo * @pwq: target pool_workqueue 36850fbd95aaSTejun Heo * 3686699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3687699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3688699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36890fbd95aaSTejun Heo */ 3690699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36910fbd95aaSTejun Heo { 3692699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3693699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3694699ce097STejun Heo 3695699ce097STejun Heo /* for @wq->saved_max_active */ 3696a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3697699ce097STejun Heo 3698699ce097STejun Heo /* fast exit for non-freezable wqs */ 3699699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3700699ce097STejun Heo return; 3701699ce097STejun Heo 3702a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3703699ce097STejun Heo 3704699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3705699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37060fbd95aaSTejun Heo 37070fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 37080fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 37090fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3710951a078aSLai Jiangshan 3711951a078aSLai Jiangshan /* 3712951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3713951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3714951a078aSLai Jiangshan */ 3715951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3716699ce097STejun Heo } else { 3717699ce097STejun Heo pwq->max_active = 0; 3718699ce097STejun Heo } 3719699ce097STejun Heo 3720a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 37210fbd95aaSTejun Heo } 37220fbd95aaSTejun Heo 3723e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3724f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3725f147f29eSTejun Heo struct worker_pool *pool) 3726d2c1d404STejun Heo { 3727d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3728d2c1d404STejun Heo 3729e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3730e50aba9aSTejun Heo 3731d2c1d404STejun Heo pwq->pool = pool; 3732d2c1d404STejun Heo pwq->wq = wq; 3733d2c1d404STejun Heo pwq->flush_color = -1; 37348864b4e5STejun Heo pwq->refcnt = 1; 3735d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 37361befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3737d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37388864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3739f147f29eSTejun Heo } 3740d2c1d404STejun Heo 3741f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37421befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3743f147f29eSTejun Heo { 3744f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3745f147f29eSTejun Heo 3746f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 374775ccf595STejun Heo 37481befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37491befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37501befcf30STejun Heo return; 37511befcf30STejun Heo 3752983ca25eSTejun Heo /* 3753983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 37543c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3755983ca25eSTejun Heo */ 375675ccf595STejun Heo pwq->work_color = wq->work_color; 3757983ca25eSTejun Heo 3758983ca25eSTejun Heo /* sync max_active to the current setting */ 3759983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3760983ca25eSTejun Heo 3761983ca25eSTejun Heo /* link in @pwq */ 37629e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3763df2d5ae4STejun Heo } 37646029a918STejun Heo 3765f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3766f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3767f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3768f147f29eSTejun Heo { 3769f147f29eSTejun Heo struct worker_pool *pool; 3770f147f29eSTejun Heo struct pool_workqueue *pwq; 3771f147f29eSTejun Heo 3772f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3773f147f29eSTejun Heo 3774f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3775f147f29eSTejun Heo if (!pool) 3776f147f29eSTejun Heo return NULL; 3777f147f29eSTejun Heo 3778e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3779f147f29eSTejun Heo if (!pwq) { 3780f147f29eSTejun Heo put_unbound_pool(pool); 3781f147f29eSTejun Heo return NULL; 3782f147f29eSTejun Heo } 3783f147f29eSTejun Heo 3784f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3785f147f29eSTejun Heo return pwq; 3786d2c1d404STejun Heo } 3787d2c1d404STejun Heo 37884c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 37894c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 37904c16bd32STejun Heo { 37914c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37924c16bd32STejun Heo 37934c16bd32STejun Heo if (pwq) { 37944c16bd32STejun Heo put_unbound_pool(pwq->pool); 3795cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 37964c16bd32STejun Heo } 37974c16bd32STejun Heo } 37984c16bd32STejun Heo 37994c16bd32STejun Heo /** 38004c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 38014c16bd32STejun Heo * @attrs: the wq_attrs of interest 38024c16bd32STejun Heo * @node: the target NUMA node 38034c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38044c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38054c16bd32STejun Heo * 38064c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38074c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 38084c16bd32STejun Heo * calculation. The result is stored in @cpumask. This function returns 38094c16bd32STejun Heo * %true if the resulting @cpumask is different from @attrs->cpumask, 38104c16bd32STejun Heo * %false if equal. 38114c16bd32STejun Heo * 38124c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38134c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38144c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38154c16bd32STejun Heo * @attrs->cpumask. 38164c16bd32STejun Heo * 38174c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38184c16bd32STejun Heo * stable. 38194c16bd32STejun Heo */ 38204c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38214c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38224c16bd32STejun Heo { 3823d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38244c16bd32STejun Heo goto use_dfl; 38254c16bd32STejun Heo 38264c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38274c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38284c16bd32STejun Heo if (cpu_going_down >= 0) 38294c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38304c16bd32STejun Heo 38314c16bd32STejun Heo if (cpumask_empty(cpumask)) 38324c16bd32STejun Heo goto use_dfl; 38334c16bd32STejun Heo 38344c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38354c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38364c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 38374c16bd32STejun Heo 38384c16bd32STejun Heo use_dfl: 38394c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38404c16bd32STejun Heo return false; 38414c16bd32STejun Heo } 38424c16bd32STejun Heo 38431befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38441befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38451befcf30STejun Heo int node, 38461befcf30STejun Heo struct pool_workqueue *pwq) 38471befcf30STejun Heo { 38481befcf30STejun Heo struct pool_workqueue *old_pwq; 38491befcf30STejun Heo 38501befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38511befcf30STejun Heo 38521befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 38531befcf30STejun Heo link_pwq(pwq); 38541befcf30STejun Heo 38551befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 38561befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 38571befcf30STejun Heo return old_pwq; 38581befcf30STejun Heo } 38591befcf30STejun Heo 38609e8cd2f5STejun Heo /** 38619e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38629e8cd2f5STejun Heo * @wq: the target workqueue 38639e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38649e8cd2f5STejun Heo * 38654c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 38664c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 38674c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38684c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38694c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38704c16bd32STejun Heo * back-to-back will stay on its current pwq. 38719e8cd2f5STejun Heo * 38729e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 38739e8cd2f5STejun Heo * failure. 38749e8cd2f5STejun Heo */ 38759e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38769e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38779e8cd2f5STejun Heo { 38784c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 38794c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3880f147f29eSTejun Heo int node, ret; 38819e8cd2f5STejun Heo 38828719dceaSTejun Heo /* only unbound workqueues can change attributes */ 38839e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 38849e8cd2f5STejun Heo return -EINVAL; 38859e8cd2f5STejun Heo 38868719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 38878719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 38888719dceaSTejun Heo return -EINVAL; 38898719dceaSTejun Heo 38904c16bd32STejun Heo pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); 389113e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38924c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38934c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 389413e2e556STejun Heo goto enomem; 389513e2e556STejun Heo 38964c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 389713e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 389813e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 389913e2e556STejun Heo 39004c16bd32STejun Heo /* 39014c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 39024c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 39034c16bd32STejun Heo * pools. 39044c16bd32STejun Heo */ 39054c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 39069e8cd2f5STejun Heo 39074c16bd32STejun Heo /* 39084c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 39094c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 39104c16bd32STejun Heo * pwqs accordingly. 39114c16bd32STejun Heo */ 39124c16bd32STejun Heo get_online_cpus(); 39134c16bd32STejun Heo 39144c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 39154c16bd32STejun Heo 39164c16bd32STejun Heo /* 39174c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 39184c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 39194c16bd32STejun Heo * it even if we don't use it immediately. 39204c16bd32STejun Heo */ 39214c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39224c16bd32STejun Heo if (!dfl_pwq) 39234c16bd32STejun Heo goto enomem_pwq; 39244c16bd32STejun Heo 39254c16bd32STejun Heo for_each_node(node) { 39264c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 39274c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39284c16bd32STejun Heo if (!pwq_tbl[node]) 39294c16bd32STejun Heo goto enomem_pwq; 39304c16bd32STejun Heo } else { 39314c16bd32STejun Heo dfl_pwq->refcnt++; 39324c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 39334c16bd32STejun Heo } 39344c16bd32STejun Heo } 39354c16bd32STejun Heo 39364c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39374c16bd32STejun Heo 39384c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3939f147f29eSTejun Heo mutex_lock(&wq->mutex); 3940a892caccSTejun Heo 3941f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 39424c16bd32STejun Heo 39434c16bd32STejun Heo /* save the previous pwq and install the new one */ 3944f147f29eSTejun Heo for_each_node(node) 39454c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 39464c16bd32STejun Heo 39474c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 39484c16bd32STejun Heo link_pwq(dfl_pwq); 39494c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3950f147f29eSTejun Heo 3951f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3952f147f29eSTejun Heo 39534c16bd32STejun Heo /* put the old pwqs */ 39544c16bd32STejun Heo for_each_node(node) 39554c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 39564c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 39574c16bd32STejun Heo 39584c16bd32STejun Heo put_online_cpus(); 39594862125bSTejun Heo ret = 0; 39604862125bSTejun Heo /* fall through */ 39614862125bSTejun Heo out_free: 39624c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 39634862125bSTejun Heo free_workqueue_attrs(new_attrs); 39644c16bd32STejun Heo kfree(pwq_tbl); 39654862125bSTejun Heo return ret; 396613e2e556STejun Heo 39674c16bd32STejun Heo enomem_pwq: 39684c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 39694c16bd32STejun Heo for_each_node(node) 39704c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 39714c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 39724c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39734c16bd32STejun Heo put_online_cpus(); 397413e2e556STejun Heo enomem: 39754862125bSTejun Heo ret = -ENOMEM; 39764862125bSTejun Heo goto out_free; 39779e8cd2f5STejun Heo } 39789e8cd2f5STejun Heo 39794c16bd32STejun Heo /** 39804c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 39814c16bd32STejun Heo * @wq: the target workqueue 39824c16bd32STejun Heo * @cpu: the CPU coming up or going down 39834c16bd32STejun Heo * @online: whether @cpu is coming up or going down 39844c16bd32STejun Heo * 39854c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 39864c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 39874c16bd32STejun Heo * @wq accordingly. 39884c16bd32STejun Heo * 39894c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 39904c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 39914c16bd32STejun Heo * correct. 39924c16bd32STejun Heo * 39934c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 39944c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 39954c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 39964c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 39974c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 39984c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 39994c16bd32STejun Heo * CPU_DOWN_PREPARE. 40004c16bd32STejun Heo */ 40014c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 40024c16bd32STejun Heo bool online) 40034c16bd32STejun Heo { 40044c16bd32STejun Heo int node = cpu_to_node(cpu); 40054c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 40064c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 40074c16bd32STejun Heo struct workqueue_attrs *target_attrs; 40084c16bd32STejun Heo cpumask_t *cpumask; 40094c16bd32STejun Heo 40104c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 40114c16bd32STejun Heo 40124c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 40134c16bd32STejun Heo return; 40144c16bd32STejun Heo 40154c16bd32STejun Heo /* 40164c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 40174c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 40184c16bd32STejun Heo * CPU hotplug exclusion. 40194c16bd32STejun Heo */ 40204c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 40214c16bd32STejun Heo cpumask = target_attrs->cpumask; 40224c16bd32STejun Heo 40234c16bd32STejun Heo mutex_lock(&wq->mutex); 4024d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 4025d55262c4STejun Heo goto out_unlock; 40264c16bd32STejun Heo 40274c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 40284c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 40294c16bd32STejun Heo 40304c16bd32STejun Heo /* 40314c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 40324c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 40334c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 40344c16bd32STejun Heo * wq's, the default pwq should be used. If @pwq is already the 40354c16bd32STejun Heo * default one, nothing to do; otherwise, install the default one. 40364c16bd32STejun Heo */ 40374c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 40384c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 40394c16bd32STejun Heo goto out_unlock; 40404c16bd32STejun Heo } else { 40414c16bd32STejun Heo if (pwq == wq->dfl_pwq) 40424c16bd32STejun Heo goto out_unlock; 40434c16bd32STejun Heo else 40444c16bd32STejun Heo goto use_dfl_pwq; 40454c16bd32STejun Heo } 40464c16bd32STejun Heo 40474c16bd32STejun Heo mutex_unlock(&wq->mutex); 40484c16bd32STejun Heo 40494c16bd32STejun Heo /* create a new pwq */ 40504c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 40514c16bd32STejun Heo if (!pwq) { 40524c16bd32STejun Heo pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 40534c16bd32STejun Heo wq->name); 40544c16bd32STejun Heo goto out_unlock; 40554c16bd32STejun Heo } 40564c16bd32STejun Heo 40574c16bd32STejun Heo /* 40584c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 40594c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 40604c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 40614c16bd32STejun Heo * inbetween. 40624c16bd32STejun Heo */ 40634c16bd32STejun Heo mutex_lock(&wq->mutex); 40644c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 40654c16bd32STejun Heo goto out_unlock; 40664c16bd32STejun Heo 40674c16bd32STejun Heo use_dfl_pwq: 40684c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 40694c16bd32STejun Heo get_pwq(wq->dfl_pwq); 40704c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 40714c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 40724c16bd32STejun Heo out_unlock: 40734c16bd32STejun Heo mutex_unlock(&wq->mutex); 40744c16bd32STejun Heo put_pwq_unlocked(old_pwq); 40754c16bd32STejun Heo } 40764c16bd32STejun Heo 407730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 40781da177e4SLinus Torvalds { 407949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 408030cdf249STejun Heo int cpu; 4081e1d8aa9fSFrederic Weisbecker 408230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4083420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4084420c0ddbSTejun Heo if (!wq->cpu_pwqs) 408530cdf249STejun Heo return -ENOMEM; 408630cdf249STejun Heo 408730cdf249STejun Heo for_each_possible_cpu(cpu) { 40887fb98ea7STejun Heo struct pool_workqueue *pwq = 40897fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 40907a62c2c8STejun Heo struct worker_pool *cpu_pools = 4091f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 409230cdf249STejun Heo 4093f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4094f147f29eSTejun Heo 4095f147f29eSTejun Heo mutex_lock(&wq->mutex); 40961befcf30STejun Heo link_pwq(pwq); 4097f147f29eSTejun Heo mutex_unlock(&wq->mutex); 409830cdf249STejun Heo } 409930cdf249STejun Heo return 0; 41009e8cd2f5STejun Heo } else { 41019e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 41029e8cd2f5STejun Heo } 41030f900049STejun Heo } 41040f900049STejun Heo 4105f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4106f3421797STejun Heo const char *name) 4107b71ab8c2STejun Heo { 4108f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4109f3421797STejun Heo 4110f3421797STejun Heo if (max_active < 1 || max_active > lim) 4111044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4112f3421797STejun Heo max_active, name, 1, lim); 4113b71ab8c2STejun Heo 4114f3421797STejun Heo return clamp_val(max_active, 1, lim); 4115b71ab8c2STejun Heo } 4116b71ab8c2STejun Heo 4117b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 411897e37d7bSTejun Heo unsigned int flags, 41191e19ffc6STejun Heo int max_active, 4120eb13ba87SJohannes Berg struct lock_class_key *key, 4121b196be89STejun Heo const char *lock_name, ...) 41223af24433SOleg Nesterov { 4123df2d5ae4STejun Heo size_t tbl_size = 0; 4124ecf6881fSTejun Heo va_list args; 41253af24433SOleg Nesterov struct workqueue_struct *wq; 412649e3cf44STejun Heo struct pool_workqueue *pwq; 4127b196be89STejun Heo 4128cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4129cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4130cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4131cee22a15SViresh Kumar 4132ecf6881fSTejun Heo /* allocate wq and format name */ 4133df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4134df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 4135df2d5ae4STejun Heo 4136df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4137b196be89STejun Heo if (!wq) 4138d2c1d404STejun Heo return NULL; 4139b196be89STejun Heo 41406029a918STejun Heo if (flags & WQ_UNBOUND) { 41416029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 41426029a918STejun Heo if (!wq->unbound_attrs) 41436029a918STejun Heo goto err_free_wq; 41446029a918STejun Heo } 41456029a918STejun Heo 4146ecf6881fSTejun Heo va_start(args, lock_name); 4147ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4148b196be89STejun Heo va_end(args); 41493af24433SOleg Nesterov 4150d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4151b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 41523af24433SOleg Nesterov 4153b196be89STejun Heo /* init wq */ 415497e37d7bSTejun Heo wq->flags = flags; 4155a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 41563c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4157112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 415830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 415973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 416073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4161493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 41623af24433SOleg Nesterov 4163eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4164cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 41653af24433SOleg Nesterov 416630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4167d2c1d404STejun Heo goto err_free_wq; 41681537663fSTejun Heo 4169493008a8STejun Heo /* 4170493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4171493008a8STejun Heo * have a rescuer to guarantee forward progress. 4172493008a8STejun Heo */ 4173493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4174e22bee78STejun Heo struct worker *rescuer; 4175e22bee78STejun Heo 4176d2c1d404STejun Heo rescuer = alloc_worker(); 4177e22bee78STejun Heo if (!rescuer) 4178d2c1d404STejun Heo goto err_destroy; 4179e22bee78STejun Heo 4180111c225aSTejun Heo rescuer->rescue_wq = wq; 4181111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4182b196be89STejun Heo wq->name); 4183d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4184d2c1d404STejun Heo kfree(rescuer); 4185d2c1d404STejun Heo goto err_destroy; 4186d2c1d404STejun Heo } 4187e22bee78STejun Heo 4188d2c1d404STejun Heo wq->rescuer = rescuer; 418914a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4190e22bee78STejun Heo wake_up_process(rescuer->task); 41913af24433SOleg Nesterov } 41921537663fSTejun Heo 4193226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4194226223abSTejun Heo goto err_destroy; 4195226223abSTejun Heo 41963af24433SOleg Nesterov /* 419768e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 419868e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 419968e13a67SLai Jiangshan * list. 42003af24433SOleg Nesterov */ 420168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4202a0a1a5fdSTejun Heo 4203a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 420449e3cf44STejun Heo for_each_pwq(pwq, wq) 4205699ce097STejun Heo pwq_adjust_max_active(pwq); 4206a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4207a0a1a5fdSTejun Heo 42083af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4209a0a1a5fdSTejun Heo 421068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42113af24433SOleg Nesterov 42123af24433SOleg Nesterov return wq; 4213d2c1d404STejun Heo 4214d2c1d404STejun Heo err_free_wq: 42156029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 42164690c4abSTejun Heo kfree(wq); 4217d2c1d404STejun Heo return NULL; 4218d2c1d404STejun Heo err_destroy: 4219d2c1d404STejun Heo destroy_workqueue(wq); 42204690c4abSTejun Heo return NULL; 42211da177e4SLinus Torvalds } 4222d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 42231da177e4SLinus Torvalds 42243af24433SOleg Nesterov /** 42253af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 42263af24433SOleg Nesterov * @wq: target workqueue 42273af24433SOleg Nesterov * 42283af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 42293af24433SOleg Nesterov */ 42303af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 42313af24433SOleg Nesterov { 423249e3cf44STejun Heo struct pool_workqueue *pwq; 42334c16bd32STejun Heo int node; 42343af24433SOleg Nesterov 42359c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 42369c5a2ba7STejun Heo drain_workqueue(wq); 4237c8efcc25STejun Heo 42386183c009STejun Heo /* sanity checks */ 4239b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 424049e3cf44STejun Heo for_each_pwq(pwq, wq) { 42416183c009STejun Heo int i; 42426183c009STejun Heo 424376af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 424476af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4245b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42466183c009STejun Heo return; 424776af4d93STejun Heo } 424876af4d93STejun Heo } 424976af4d93STejun Heo 42505c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 42518864b4e5STejun Heo WARN_ON(pwq->nr_active) || 425276af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4253b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42546183c009STejun Heo return; 42556183c009STejun Heo } 425676af4d93STejun Heo } 4257b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42586183c009STejun Heo 4259a0a1a5fdSTejun Heo /* 4260a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4261a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4262a0a1a5fdSTejun Heo */ 426368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4264d2c1d404STejun Heo list_del_init(&wq->list); 426568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42663af24433SOleg Nesterov 4267226223abSTejun Heo workqueue_sysfs_unregister(wq); 4268226223abSTejun Heo 4269493008a8STejun Heo if (wq->rescuer) { 4270e22bee78STejun Heo kthread_stop(wq->rescuer->task); 42718d9df9f0SXiaotian Feng kfree(wq->rescuer); 4272493008a8STejun Heo wq->rescuer = NULL; 4273e22bee78STejun Heo } 4274e22bee78STejun Heo 42758864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 427629c91e99STejun Heo /* 42778864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 42788864b4e5STejun Heo * free the pwqs and wq. 427929c91e99STejun Heo */ 42808864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 42818864b4e5STejun Heo kfree(wq); 42828864b4e5STejun Heo } else { 42838864b4e5STejun Heo /* 42848864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 42854c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 42864c16bd32STejun Heo * @wq will be freed when the last pwq is released. 42878864b4e5STejun Heo */ 42884c16bd32STejun Heo for_each_node(node) { 42894c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42904c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 42914c16bd32STejun Heo put_pwq_unlocked(pwq); 42924c16bd32STejun Heo } 42934c16bd32STejun Heo 42944c16bd32STejun Heo /* 42954c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 42964c16bd32STejun Heo * put. Don't access it afterwards. 42974c16bd32STejun Heo */ 42984c16bd32STejun Heo pwq = wq->dfl_pwq; 42994c16bd32STejun Heo wq->dfl_pwq = NULL; 4300dce90d47STejun Heo put_pwq_unlocked(pwq); 430129c91e99STejun Heo } 43023af24433SOleg Nesterov } 43033af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 43043af24433SOleg Nesterov 4305dcd989cbSTejun Heo /** 4306dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4307dcd989cbSTejun Heo * @wq: target workqueue 4308dcd989cbSTejun Heo * @max_active: new max_active value. 4309dcd989cbSTejun Heo * 4310dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4311dcd989cbSTejun Heo * 4312dcd989cbSTejun Heo * CONTEXT: 4313dcd989cbSTejun Heo * Don't call from IRQ context. 4314dcd989cbSTejun Heo */ 4315dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4316dcd989cbSTejun Heo { 431749e3cf44STejun Heo struct pool_workqueue *pwq; 4318dcd989cbSTejun Heo 43198719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 43208719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 43218719dceaSTejun Heo return; 43228719dceaSTejun Heo 4323f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4324dcd989cbSTejun Heo 4325a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4326dcd989cbSTejun Heo 4327dcd989cbSTejun Heo wq->saved_max_active = max_active; 4328dcd989cbSTejun Heo 4329699ce097STejun Heo for_each_pwq(pwq, wq) 4330699ce097STejun Heo pwq_adjust_max_active(pwq); 4331dcd989cbSTejun Heo 4332a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4333dcd989cbSTejun Heo } 4334dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4335dcd989cbSTejun Heo 4336dcd989cbSTejun Heo /** 4337e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4338e6267616STejun Heo * 4339e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4340e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4341e6267616STejun Heo */ 4342e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4343e6267616STejun Heo { 4344e6267616STejun Heo struct worker *worker = current_wq_worker(); 4345e6267616STejun Heo 43466a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4347e6267616STejun Heo } 4348e6267616STejun Heo 4349e6267616STejun Heo /** 4350dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4351dcd989cbSTejun Heo * @cpu: CPU in question 4352dcd989cbSTejun Heo * @wq: target workqueue 4353dcd989cbSTejun Heo * 4354dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4355dcd989cbSTejun Heo * no synchronization around this function and the test result is 4356dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4357dcd989cbSTejun Heo * 4358d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4359d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4360d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4361d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4362d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4363d3251859STejun Heo * 4364dcd989cbSTejun Heo * RETURNS: 4365dcd989cbSTejun Heo * %true if congested, %false otherwise. 4366dcd989cbSTejun Heo */ 4367d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4368dcd989cbSTejun Heo { 43697fb98ea7STejun Heo struct pool_workqueue *pwq; 437076af4d93STejun Heo bool ret; 437176af4d93STejun Heo 437288109453SLai Jiangshan rcu_read_lock_sched(); 43737fb98ea7STejun Heo 4374d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4375d3251859STejun Heo cpu = smp_processor_id(); 4376d3251859STejun Heo 43777fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 43787fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 43797fb98ea7STejun Heo else 4380df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4381dcd989cbSTejun Heo 438276af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 438388109453SLai Jiangshan rcu_read_unlock_sched(); 438476af4d93STejun Heo 438576af4d93STejun Heo return ret; 4386dcd989cbSTejun Heo } 4387dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4388dcd989cbSTejun Heo 4389dcd989cbSTejun Heo /** 4390dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4391dcd989cbSTejun Heo * @work: the work to be tested 4392dcd989cbSTejun Heo * 4393dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4394dcd989cbSTejun Heo * synchronization around this function and the test result is 4395dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4396dcd989cbSTejun Heo * 4397dcd989cbSTejun Heo * RETURNS: 4398dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4399dcd989cbSTejun Heo */ 4400dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4401dcd989cbSTejun Heo { 4402fa1b54e6STejun Heo struct worker_pool *pool; 4403dcd989cbSTejun Heo unsigned long flags; 4404dcd989cbSTejun Heo unsigned int ret = 0; 4405dcd989cbSTejun Heo 4406dcd989cbSTejun Heo if (work_pending(work)) 4407dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4408038366c5SLai Jiangshan 4409fa1b54e6STejun Heo local_irq_save(flags); 4410fa1b54e6STejun Heo pool = get_work_pool(work); 4411038366c5SLai Jiangshan if (pool) { 4412fa1b54e6STejun Heo spin_lock(&pool->lock); 4413c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4414dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4415fa1b54e6STejun Heo spin_unlock(&pool->lock); 4416038366c5SLai Jiangshan } 4417fa1b54e6STejun Heo local_irq_restore(flags); 4418dcd989cbSTejun Heo 4419dcd989cbSTejun Heo return ret; 4420dcd989cbSTejun Heo } 4421dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4422dcd989cbSTejun Heo 44233d1cb205STejun Heo /** 44243d1cb205STejun Heo * set_worker_desc - set description for the current work item 44253d1cb205STejun Heo * @fmt: printf-style format string 44263d1cb205STejun Heo * @...: arguments for the format string 44273d1cb205STejun Heo * 44283d1cb205STejun Heo * This function can be called by a running work function to describe what 44293d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 44303d1cb205STejun Heo * information will be printed out together to help debugging. The 44313d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 44323d1cb205STejun Heo */ 44333d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 44343d1cb205STejun Heo { 44353d1cb205STejun Heo struct worker *worker = current_wq_worker(); 44363d1cb205STejun Heo va_list args; 44373d1cb205STejun Heo 44383d1cb205STejun Heo if (worker) { 44393d1cb205STejun Heo va_start(args, fmt); 44403d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 44413d1cb205STejun Heo va_end(args); 44423d1cb205STejun Heo worker->desc_valid = true; 44433d1cb205STejun Heo } 44443d1cb205STejun Heo } 44453d1cb205STejun Heo 44463d1cb205STejun Heo /** 44473d1cb205STejun Heo * print_worker_info - print out worker information and description 44483d1cb205STejun Heo * @log_lvl: the log level to use when printing 44493d1cb205STejun Heo * @task: target task 44503d1cb205STejun Heo * 44513d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 44523d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 44533d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 44543d1cb205STejun Heo * 44553d1cb205STejun Heo * This function can be safely called on any task as long as the 44563d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 44573d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 44583d1cb205STejun Heo */ 44593d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 44603d1cb205STejun Heo { 44613d1cb205STejun Heo work_func_t *fn = NULL; 44623d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 44633d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 44643d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 44653d1cb205STejun Heo struct workqueue_struct *wq = NULL; 44663d1cb205STejun Heo bool desc_valid = false; 44673d1cb205STejun Heo struct worker *worker; 44683d1cb205STejun Heo 44693d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 44703d1cb205STejun Heo return; 44713d1cb205STejun Heo 44723d1cb205STejun Heo /* 44733d1cb205STejun Heo * This function is called without any synchronization and @task 44743d1cb205STejun Heo * could be in any state. Be careful with dereferences. 44753d1cb205STejun Heo */ 44763d1cb205STejun Heo worker = probe_kthread_data(task); 44773d1cb205STejun Heo 44783d1cb205STejun Heo /* 44793d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 44803d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 44813d1cb205STejun Heo */ 44823d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 44833d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 44843d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 44853d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 44863d1cb205STejun Heo 44873d1cb205STejun Heo /* copy worker description */ 44883d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 44893d1cb205STejun Heo if (desc_valid) 44903d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 44913d1cb205STejun Heo 44923d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44933d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 44943d1cb205STejun Heo if (desc[0]) 44953d1cb205STejun Heo pr_cont(" (%s)", desc); 44963d1cb205STejun Heo pr_cont("\n"); 44973d1cb205STejun Heo } 44983d1cb205STejun Heo } 44993d1cb205STejun Heo 4500db7bccf4STejun Heo /* 4501db7bccf4STejun Heo * CPU hotplug. 4502db7bccf4STejun Heo * 4503e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4504112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4505706026c2STejun Heo * pool which make migrating pending and scheduled works very 4506e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 450794cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4508e22bee78STejun Heo * blocked draining impractical. 4509e22bee78STejun Heo * 451024647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4511628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4512628c78e7STejun Heo * cpu comes back online. 4513db7bccf4STejun Heo */ 4514db7bccf4STejun Heo 4515706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4516db7bccf4STejun Heo { 451738db41d9STejun Heo int cpu = smp_processor_id(); 45184ce62e9eSTejun Heo struct worker_pool *pool; 4519db7bccf4STejun Heo struct worker *worker; 4520a9ab775bSTejun Heo int wi; 4521db7bccf4STejun Heo 4522f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45236183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4524db7bccf4STejun Heo 4525bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 452694cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4527e22bee78STejun Heo 4528f2d5a0eeSTejun Heo /* 4529bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 453094cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 453194cf58bbSTejun Heo * except for the ones which are still executing works from 453294cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 453394cf58bbSTejun Heo * this, they may become diasporas. 4534f2d5a0eeSTejun Heo */ 4535a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4536403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4537db7bccf4STejun Heo 453824647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4539f2d5a0eeSTejun Heo 454094cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4541bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 4542e22bee78STejun Heo 4543e22bee78STejun Heo /* 4544eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4545eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4546eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4547eb283428SLai Jiangshan * from other cpus. 4548628c78e7STejun Heo */ 4549628c78e7STejun Heo schedule(); 4550628c78e7STejun Heo 4551628c78e7STejun Heo /* 4552eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4553eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4554eb283428SLai Jiangshan * and keep_working() are always true as long as the 4555eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4556eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4557eb283428SLai Jiangshan * are served by workers tied to the pool. 4558e22bee78STejun Heo */ 4559e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4560eb283428SLai Jiangshan 4561eb283428SLai Jiangshan /* 4562eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4563eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4564eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4565eb283428SLai Jiangshan */ 4566eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4567eb283428SLai Jiangshan wake_up_worker(pool); 4568eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4569eb283428SLai Jiangshan } 4570db7bccf4STejun Heo } 4571db7bccf4STejun Heo 4572bd7c089eSTejun Heo /** 4573bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4574bd7c089eSTejun Heo * @pool: pool of interest 4575bd7c089eSTejun Heo * 4576a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4577bd7c089eSTejun Heo */ 4578bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4579bd7c089eSTejun Heo { 4580a9ab775bSTejun Heo struct worker *worker; 4581a9ab775bSTejun Heo int wi; 4582bd7c089eSTejun Heo 4583bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4584bd7c089eSTejun Heo 4585bd7c089eSTejun Heo /* 4586a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4587a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4588a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4589a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4590a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4591bd7c089eSTejun Heo */ 4592a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4593a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4594a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4595a9ab775bSTejun Heo 4596a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4597a9ab775bSTejun Heo 4598a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4599a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4600a9ab775bSTejun Heo 4601a9ab775bSTejun Heo /* 4602a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4603a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4604a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4605a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4606a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4607a9ab775bSTejun Heo * be bound before @pool->lock is released. 4608a9ab775bSTejun Heo */ 4609a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4610bd7c089eSTejun Heo wake_up_process(worker->task); 4611bd7c089eSTejun Heo 4612bd7c089eSTejun Heo /* 4613a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4614a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4615a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4616a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4617a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4618a9ab775bSTejun Heo * concurrency management. Note that when or whether 4619a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4620a9ab775bSTejun Heo * 4621a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4622a9ab775bSTejun Heo * tested without holding any lock in 4623a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4624a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4625a9ab775bSTejun Heo * management operations. 4626bd7c089eSTejun Heo */ 4627a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4628a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4629a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4630a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4631bd7c089eSTejun Heo } 4632a9ab775bSTejun Heo 4633a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4634bd7c089eSTejun Heo } 4635bd7c089eSTejun Heo 46367dbc725eSTejun Heo /** 46377dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 46387dbc725eSTejun Heo * @pool: unbound pool of interest 46397dbc725eSTejun Heo * @cpu: the CPU which is coming up 46407dbc725eSTejun Heo * 46417dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 46427dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 46437dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 46447dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 46457dbc725eSTejun Heo */ 46467dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 46477dbc725eSTejun Heo { 46487dbc725eSTejun Heo static cpumask_t cpumask; 46497dbc725eSTejun Heo struct worker *worker; 46507dbc725eSTejun Heo int wi; 46517dbc725eSTejun Heo 46527dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 46537dbc725eSTejun Heo 46547dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 46557dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 46567dbc725eSTejun Heo return; 46577dbc725eSTejun Heo 46587dbc725eSTejun Heo /* is @cpu the only online CPU? */ 46597dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 46607dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 46617dbc725eSTejun Heo return; 46627dbc725eSTejun Heo 46637dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 46647dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 46657dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 46667dbc725eSTejun Heo pool->attrs->cpumask) < 0); 46677dbc725eSTejun Heo } 46687dbc725eSTejun Heo 46698db25e78STejun Heo /* 46708db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 46718db25e78STejun Heo * This will be registered high priority CPU notifier. 46728db25e78STejun Heo */ 46730db0628dSPaul Gortmaker static int workqueue_cpu_up_callback(struct notifier_block *nfb, 46741da177e4SLinus Torvalds unsigned long action, 46751da177e4SLinus Torvalds void *hcpu) 46761da177e4SLinus Torvalds { 4677d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46784ce62e9eSTejun Heo struct worker_pool *pool; 46794c16bd32STejun Heo struct workqueue_struct *wq; 46807dbc725eSTejun Heo int pi; 46811da177e4SLinus Torvalds 46828db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 46833af24433SOleg Nesterov case CPU_UP_PREPARE: 4684f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46853ce63377STejun Heo if (pool->nr_workers) 46863ce63377STejun Heo continue; 4687ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 46883ce63377STejun Heo return NOTIFY_BAD; 46893af24433SOleg Nesterov } 46901da177e4SLinus Torvalds break; 46911da177e4SLinus Torvalds 469265758202STejun Heo case CPU_DOWN_FAILED: 469365758202STejun Heo case CPU_ONLINE: 469468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46957dbc725eSTejun Heo 46967dbc725eSTejun Heo for_each_pool(pool, pi) { 4697bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 469894cf58bbSTejun Heo 46997dbc725eSTejun Heo if (pool->cpu == cpu) { 4700a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 470124647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4702a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4703a9ab775bSTejun Heo 470494cf58bbSTejun Heo rebind_workers(pool); 47057dbc725eSTejun Heo } else if (pool->cpu < 0) { 47067dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 47077dbc725eSTejun Heo } 470894cf58bbSTejun Heo 4709bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 471094cf58bbSTejun Heo } 47117dbc725eSTejun Heo 47124c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47134c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47144c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 47154c16bd32STejun Heo 471668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47178db25e78STejun Heo break; 471865758202STejun Heo } 471965758202STejun Heo return NOTIFY_OK; 472065758202STejun Heo } 472165758202STejun Heo 472265758202STejun Heo /* 472365758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 472465758202STejun Heo * This will be registered as low priority CPU notifier. 472565758202STejun Heo */ 47260db0628dSPaul Gortmaker static int workqueue_cpu_down_callback(struct notifier_block *nfb, 472765758202STejun Heo unsigned long action, 472865758202STejun Heo void *hcpu) 472965758202STejun Heo { 4730d84ff051STejun Heo int cpu = (unsigned long)hcpu; 47318db25e78STejun Heo struct work_struct unbind_work; 47324c16bd32STejun Heo struct workqueue_struct *wq; 47338db25e78STejun Heo 473465758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 473565758202STejun Heo case CPU_DOWN_PREPARE: 47364c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4737706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 47387635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 47394c16bd32STejun Heo 47404c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47414c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 47424c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47434c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 47444c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 47454c16bd32STejun Heo 47464c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 47478db25e78STejun Heo flush_work(&unbind_work); 47488db25e78STejun Heo break; 474965758202STejun Heo } 475065758202STejun Heo return NOTIFY_OK; 475165758202STejun Heo } 475265758202STejun Heo 47532d3854a3SRusty Russell #ifdef CONFIG_SMP 47548ccad40dSRusty Russell 47552d3854a3SRusty Russell struct work_for_cpu { 4756ed48ece2STejun Heo struct work_struct work; 47572d3854a3SRusty Russell long (*fn)(void *); 47582d3854a3SRusty Russell void *arg; 47592d3854a3SRusty Russell long ret; 47602d3854a3SRusty Russell }; 47612d3854a3SRusty Russell 4762ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 47632d3854a3SRusty Russell { 4764ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4765ed48ece2STejun Heo 47662d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 47672d3854a3SRusty Russell } 47682d3854a3SRusty Russell 47692d3854a3SRusty Russell /** 47702d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 47712d3854a3SRusty Russell * @cpu: the cpu to run on 47722d3854a3SRusty Russell * @fn: the function to run 47732d3854a3SRusty Russell * @arg: the function arg 47742d3854a3SRusty Russell * 477531ad9081SRusty Russell * This will return the value @fn returns. 477631ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47776b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 47782d3854a3SRusty Russell */ 4779d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47802d3854a3SRusty Russell { 4781ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47822d3854a3SRusty Russell 4783ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4784ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4785c2fda509SLai Jiangshan 4786c2fda509SLai Jiangshan /* 4787c2fda509SLai Jiangshan * The work item is on-stack and can't lead to deadlock through 4788c2fda509SLai Jiangshan * flushing. Use __flush_work() to avoid spurious lockdep warnings 4789c2fda509SLai Jiangshan * when work_on_cpu()s are nested. 4790c2fda509SLai Jiangshan */ 4791c2fda509SLai Jiangshan __flush_work(&wfc.work); 4792c2fda509SLai Jiangshan 47932d3854a3SRusty Russell return wfc.ret; 47942d3854a3SRusty Russell } 47952d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47962d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47972d3854a3SRusty Russell 4798a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4799e7577c50SRusty Russell 4800a0a1a5fdSTejun Heo /** 4801a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4802a0a1a5fdSTejun Heo * 480358a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4804c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4805706026c2STejun Heo * pool->worklist. 4806a0a1a5fdSTejun Heo * 4807a0a1a5fdSTejun Heo * CONTEXT: 4808a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4809a0a1a5fdSTejun Heo */ 4810a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4811a0a1a5fdSTejun Heo { 481217116969STejun Heo struct worker_pool *pool; 481324b8a847STejun Heo struct workqueue_struct *wq; 481424b8a847STejun Heo struct pool_workqueue *pwq; 4815611c92a0STejun Heo int pi; 4816a0a1a5fdSTejun Heo 481768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4818a0a1a5fdSTejun Heo 48196183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4820a0a1a5fdSTejun Heo workqueue_freezing = true; 4821a0a1a5fdSTejun Heo 482224b8a847STejun Heo /* set FREEZING */ 4823611c92a0STejun Heo for_each_pool(pool, pi) { 48245bcab335STejun Heo spin_lock_irq(&pool->lock); 482535b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 482635b6bb63STejun Heo pool->flags |= POOL_FREEZING; 48275bcab335STejun Heo spin_unlock_irq(&pool->lock); 48281da177e4SLinus Torvalds } 48298b03ae3cSTejun Heo 483024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4831a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4832699ce097STejun Heo for_each_pwq(pwq, wq) 4833699ce097STejun Heo pwq_adjust_max_active(pwq); 4834a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4835a1056305STejun Heo } 48365bcab335STejun Heo 483768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4838a0a1a5fdSTejun Heo } 4839a0a1a5fdSTejun Heo 4840a0a1a5fdSTejun Heo /** 484158a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4842a0a1a5fdSTejun Heo * 4843a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4844a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4845a0a1a5fdSTejun Heo * 4846a0a1a5fdSTejun Heo * CONTEXT: 484768e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4848a0a1a5fdSTejun Heo * 4849a0a1a5fdSTejun Heo * RETURNS: 485058a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 485158a69cb4STejun Heo * is complete. 4852a0a1a5fdSTejun Heo */ 4853a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4854a0a1a5fdSTejun Heo { 4855a0a1a5fdSTejun Heo bool busy = false; 485624b8a847STejun Heo struct workqueue_struct *wq; 485724b8a847STejun Heo struct pool_workqueue *pwq; 4858a0a1a5fdSTejun Heo 485968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4860a0a1a5fdSTejun Heo 48616183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4862a0a1a5fdSTejun Heo 486324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 486424b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 486524b8a847STejun Heo continue; 4866a0a1a5fdSTejun Heo /* 4867a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4868a0a1a5fdSTejun Heo * to peek without lock. 4869a0a1a5fdSTejun Heo */ 487088109453SLai Jiangshan rcu_read_lock_sched(); 487124b8a847STejun Heo for_each_pwq(pwq, wq) { 48726183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4873112202d9STejun Heo if (pwq->nr_active) { 4874a0a1a5fdSTejun Heo busy = true; 487588109453SLai Jiangshan rcu_read_unlock_sched(); 4876a0a1a5fdSTejun Heo goto out_unlock; 4877a0a1a5fdSTejun Heo } 4878a0a1a5fdSTejun Heo } 487988109453SLai Jiangshan rcu_read_unlock_sched(); 4880a0a1a5fdSTejun Heo } 4881a0a1a5fdSTejun Heo out_unlock: 488268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4883a0a1a5fdSTejun Heo return busy; 4884a0a1a5fdSTejun Heo } 4885a0a1a5fdSTejun Heo 4886a0a1a5fdSTejun Heo /** 4887a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4888a0a1a5fdSTejun Heo * 4889a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4890706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4891a0a1a5fdSTejun Heo * 4892a0a1a5fdSTejun Heo * CONTEXT: 4893a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4894a0a1a5fdSTejun Heo */ 4895a0a1a5fdSTejun Heo void thaw_workqueues(void) 4896a0a1a5fdSTejun Heo { 489724b8a847STejun Heo struct workqueue_struct *wq; 489824b8a847STejun Heo struct pool_workqueue *pwq; 489924b8a847STejun Heo struct worker_pool *pool; 4900611c92a0STejun Heo int pi; 4901a0a1a5fdSTejun Heo 490268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4903a0a1a5fdSTejun Heo 4904a0a1a5fdSTejun Heo if (!workqueue_freezing) 4905a0a1a5fdSTejun Heo goto out_unlock; 4906a0a1a5fdSTejun Heo 490724b8a847STejun Heo /* clear FREEZING */ 4908611c92a0STejun Heo for_each_pool(pool, pi) { 49095bcab335STejun Heo spin_lock_irq(&pool->lock); 491035b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 491135b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 49125bcab335STejun Heo spin_unlock_irq(&pool->lock); 4913d565ed63STejun Heo } 491424b8a847STejun Heo 491524b8a847STejun Heo /* restore max_active and repopulate worklist */ 491624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4917a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4918699ce097STejun Heo for_each_pwq(pwq, wq) 4919699ce097STejun Heo pwq_adjust_max_active(pwq); 4920a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 492124b8a847STejun Heo } 492224b8a847STejun Heo 4923a0a1a5fdSTejun Heo workqueue_freezing = false; 4924a0a1a5fdSTejun Heo out_unlock: 492568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4926a0a1a5fdSTejun Heo } 4927a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4928a0a1a5fdSTejun Heo 4929bce90380STejun Heo static void __init wq_numa_init(void) 4930bce90380STejun Heo { 4931bce90380STejun Heo cpumask_var_t *tbl; 4932bce90380STejun Heo int node, cpu; 4933bce90380STejun Heo 4934bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4935bce90380STejun Heo for_each_node(node) 4936bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4937bce90380STejun Heo 4938bce90380STejun Heo if (num_possible_nodes() <= 1) 4939bce90380STejun Heo return; 4940bce90380STejun Heo 4941d55262c4STejun Heo if (wq_disable_numa) { 4942d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 4943d55262c4STejun Heo return; 4944d55262c4STejun Heo } 4945d55262c4STejun Heo 49464c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 49474c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 49484c16bd32STejun Heo 4949bce90380STejun Heo /* 4950bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4951bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4952bce90380STejun Heo * fully initialized by now. 4953bce90380STejun Heo */ 4954bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4955bce90380STejun Heo BUG_ON(!tbl); 4956bce90380STejun Heo 4957bce90380STejun Heo for_each_node(node) 49581be0c25dSTejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 49591be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 4960bce90380STejun Heo 4961bce90380STejun Heo for_each_possible_cpu(cpu) { 4962bce90380STejun Heo node = cpu_to_node(cpu); 4963bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4964bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4965bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4966bce90380STejun Heo return; 4967bce90380STejun Heo } 4968bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4969bce90380STejun Heo } 4970bce90380STejun Heo 4971bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4972bce90380STejun Heo wq_numa_enabled = true; 4973bce90380STejun Heo } 4974bce90380STejun Heo 49756ee0578bSSuresh Siddha static int __init init_workqueues(void) 49761da177e4SLinus Torvalds { 49777a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 49787a4e344cSTejun Heo int i, cpu; 4979c34056a3STejun Heo 49807c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 49817c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 49826be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4983b5490077STejun Heo 4984e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4985e904e6c2STejun Heo 4986e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4987e904e6c2STejun Heo 498865758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4989a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 49908b03ae3cSTejun Heo 4991bce90380STejun Heo wq_numa_init(); 4992bce90380STejun Heo 4993706026c2STejun Heo /* initialize CPU pools */ 499429c91e99STejun Heo for_each_possible_cpu(cpu) { 49954ce62e9eSTejun Heo struct worker_pool *pool; 49968b03ae3cSTejun Heo 49977a4e344cSTejun Heo i = 0; 4998f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49997a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5000ec22ca5eSTejun Heo pool->cpu = cpu; 50017a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 50027a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5003f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 50047a4e344cSTejun Heo 50059daf9e67STejun Heo /* alloc pool ID */ 500668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 50079daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 500868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 50094ce62e9eSTejun Heo } 50108b03ae3cSTejun Heo } 50118b03ae3cSTejun Heo 5012e22bee78STejun Heo /* create the initial worker */ 501329c91e99STejun Heo for_each_online_cpu(cpu) { 50144ce62e9eSTejun Heo struct worker_pool *pool; 5015e22bee78STejun Heo 5016f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 501724647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 5018ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 5019e22bee78STejun Heo } 50204ce62e9eSTejun Heo } 5021e22bee78STejun Heo 502229c91e99STejun Heo /* create default unbound wq attrs */ 502329c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 502429c91e99STejun Heo struct workqueue_attrs *attrs; 502529c91e99STejun Heo 502629c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 502729c91e99STejun Heo attrs->nice = std_nice[i]; 502829c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 502929c91e99STejun Heo } 503029c91e99STejun Heo 5031d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 50321aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5033d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5034f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5035f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 503624d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 503724d51addSTejun Heo WQ_FREEZABLE, 0); 50380668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 50390668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 50400668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 50410668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 50420668106cSViresh Kumar 0); 50431aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 50440668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 50450668106cSViresh Kumar !system_power_efficient_wq || 50460668106cSViresh Kumar !system_freezable_power_efficient_wq); 50476ee0578bSSuresh Siddha return 0; 50481da177e4SLinus Torvalds } 50496ee0578bSSuresh Siddha early_initcall(init_workqueues); 5050