blob: c30c01b32ecea6d9fea573551bb1b71fb6e543d2 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Tejun Heoc54fce62010-09-10 16:51:36 +02002 * kernel/workqueue.c - generic async execution with shared worker pool
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
Tejun Heoc54fce62010-09-10 16:51:36 +02004 * Copyright (C) 2002 Ingo Molnar
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 *
Tejun Heoc54fce62010-09-10 16:51:36 +02006 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
Christoph Lameter89ada672005-10-30 15:01:59 -080011 *
Christoph Lametercde53532008-07-04 09:59:22 -070012 * Made to use alloc_percpu by Christoph Lameter.
Tejun Heoc54fce62010-09-10 16:51:36 +020013 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
Libinb11895c2013-08-21 08:50:39 +080019 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
Tejun Heoc54fce62010-09-10 16:51:36 +020023 *
24 * Please read Documentation/workqueue.txt for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -070025 */
26
Paul Gortmaker9984de12011-05-23 14:51:41 -040027#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
James Bottomley1fa44ec2006-02-23 12:43:43 -060038#include <linux/hardirq.h>
Christoph Lameter46934022006-10-11 01:21:26 -070039#include <linux/mempolicy.h>
Rafael J. Wysocki341a5952006-12-06 20:34:49 -080040#include <linux/freezer.h>
Peter Zijlstrad5abe662006-12-06 20:37:26 -080041#include <linux/kallsyms.h>
42#include <linux/debug_locks.h>
Johannes Berg4e6045f2007-10-18 23:39:55 -070043#include <linux/lockdep.h>
Tejun Heoc34056a2010-06-29 10:07:11 +020044#include <linux/idr.h>
Tejun Heo29c91e92013-03-12 11:30:03 -070045#include <linux/jhash.h>
Sasha Levin42f85702012-12-17 10:01:23 -050046#include <linux/hashtable.h>
Tejun Heo76af4d92013-03-12 11:30:00 -070047#include <linux/rculist.h>
Tejun Heobce90382013-04-01 11:23:32 -070048#include <linux/nodemask.h>
Tejun Heo4c16bd32013-04-01 11:23:36 -070049#include <linux/moduleparam.h>
Tejun Heo3d1cb202013-04-30 15:27:22 -070050#include <linux/uaccess.h>
Tejun Heoe22bee72010-06-29 10:07:14 +020051
Tejun Heoea138442013-01-18 14:05:55 -080052#include "workqueue_internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
Tejun Heoc8e55f32010-06-29 10:07:12 +020054enum {
Tejun Heobc2ae0f2012-07-17 12:39:27 -070055 /*
Tejun Heo24647572013-01-24 11:01:33 -080056 * worker_pool flags
Tejun Heobc2ae0f2012-07-17 12:39:27 -070057 *
Tejun Heo24647572013-01-24 11:01:33 -080058 * A bound pool is either associated or disassociated with its CPU.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070059 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
61 * is in effect.
62 *
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
Tejun Heo24647572013-01-24 11:01:33 -080065 * be executing on any CPU. The pool behaves as an unbound one.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070066 *
Tejun Heobc3a1af2013-03-13 19:47:39 -070067 * Note that DISASSOCIATED should be flipped only while holding
68 * manager_mutex to avoid changing binding state while
Tejun Heo24647572013-01-24 11:01:33 -080069 * create_worker() is in progress.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070070 */
Tejun Heo11ebea52012-07-12 14:46:37 -070071 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
Tejun Heo24647572013-01-24 11:01:33 -080072 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
Tejun Heo35b6bb62013-01-24 11:01:33 -080073 POOL_FREEZING = 1 << 3, /* freeze in progress */
Tejun Heodb7bccf2010-06-29 10:07:12 +020074
Tejun Heoc8e55f32010-06-29 10:07:12 +020075 /* worker flags */
76 WORKER_STARTED = 1 << 0, /* started */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
Tejun Heoe22bee72010-06-29 10:07:14 +020079 WORKER_PREP = 1 << 3, /* preparing to run works */
Tejun Heofb0e7be2010-06-29 10:07:15 +020080 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
Tejun Heof3421792010-07-02 10:03:51 +020081 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
Tejun Heoa9ab7752013-03-19 13:45:21 -070082 WORKER_REBOUND = 1 << 8, /* worker was rebound */
Tejun Heoe22bee72010-06-29 10:07:14 +020083
Tejun Heoa9ab7752013-03-19 13:45:21 -070084 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
Tejun Heodb7bccf2010-06-29 10:07:12 +020086
Tejun Heoe34cdddb2013-01-24 11:01:33 -080087 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
Tejun Heo4ce62e92012-07-13 22:16:44 -070088
Tejun Heo29c91e92013-03-12 11:30:03 -070089 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
Tejun Heoc8e55f32010-06-29 10:07:12 +020090 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020091
Tejun Heoe22bee72010-06-29 10:07:14 +020092 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94
Tejun Heo3233cdb2011-02-16 18:10:19 +010095 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
97 (min two ticks) */
Tejun Heoe22bee72010-06-29 10:07:14 +020098 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
Tejun Heoe22bee72010-06-29 10:07:14 +0200100
101 /*
102 * Rescue workers are used only on emergencies and shared by
Dongsheng Yang8698a742014-03-11 18:09:12 +0800103 * all cpus. Give MIN_NICE.
Tejun Heoe22bee72010-06-29 10:07:14 +0200104 */
Dongsheng Yang8698a742014-03-11 18:09:12 +0800105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
Tejun Heoecf68812013-04-01 11:23:34 -0700107
108 WQ_NAME_LEN = 24,
Tejun Heoc8e55f32010-06-29 10:07:12 +0200109};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111/*
Tejun Heo4690c4a2010-06-29 10:07:10 +0200112 * Structure fields follow one of the following exclusion rules.
113 *
Tejun Heoe41e7042010-08-24 14:22:47 +0200114 * I: Modifiable by initialization/destruction paths and read-only for
115 * everyone else.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200116 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
119 *
Tejun Heod565ed62013-01-24 11:01:33 -0800120 * L: pool->lock protected. Access with pool->lock held.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200121 *
Tejun Heod565ed62013-01-24 11:01:33 -0800122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
Tejun Heoe22bee72010-06-29 10:07:14 +0200126 *
Tejun Heo822d8402013-03-19 13:45:21 -0700127 * MG: pool->manager_mutex and pool->lock protected. Writes require both
128 * locks. Reads can happen under either lock.
129 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700130 * PL: wq_pool_mutex protected.
Tejun Heo76af4d92013-03-12 11:30:00 -0700131 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700132 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo5bcab332013-03-13 19:47:40 -0700133 *
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700134 * WQ: wq->mutex protected.
135 *
Lai Jiangshanb5927602013-03-25 16:57:19 -0700136 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo2e109a22013-03-13 19:47:40 -0700137 *
138 * MD: wq_mayday_lock protected.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200139 */
140
Tejun Heo2eaebdb2013-01-18 14:05:55 -0800141/* struct worker is defined in workqueue_internal.h */
Tejun Heoc34056a2010-06-29 10:07:11 +0200142
Tejun Heobd7bdd42012-07-12 14:46:37 -0700143struct worker_pool {
Tejun Heod565ed62013-01-24 11:01:33 -0800144 spinlock_t lock; /* the pool lock */
Tejun Heod84ff052013-03-12 11:29:59 -0700145 int cpu; /* I: the associated cpu */
Tejun Heof3f90ad2013-04-01 11:23:34 -0700146 int node; /* I: the associated node ID */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800147 int id; /* I: pool ID */
Tejun Heo11ebea52012-07-12 14:46:37 -0700148 unsigned int flags; /* X: flags */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700149
150 struct list_head worklist; /* L: list of pending works */
151 int nr_workers; /* L: total number of workers */
Lai Jiangshanea1abd62012-09-18 09:59:22 -0700152
153 /* nr_idle includes the ones off idle_list for rebinding */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700154 int nr_idle; /* L: currently idle ones */
155
156 struct list_head idle_list; /* X: list of idle workers */
157 struct timer_list idle_timer; /* L: worker idle timeout */
158 struct timer_list mayday_timer; /* L: SOS timer for workers */
159
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700160 /* a workers is either on busy_hash or idle_list, or the manager */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800161 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
162 /* L: hash of busy workers */
163
Tejun Heobc3a1af2013-03-13 19:47:39 -0700164 /* see manage_workers() for details on the two manager mutexes */
Tejun Heo34a06bd2013-03-12 11:30:00 -0700165 struct mutex manager_arb; /* manager arbitration */
Tejun Heobc3a1af2013-03-13 19:47:39 -0700166 struct mutex manager_mutex; /* manager exclusion */
Tejun Heo822d8402013-03-19 13:45:21 -0700167 struct idr worker_idr; /* MG: worker IDs and iteration */
Tejun Heoe19e3972013-01-24 11:39:44 -0800168
Tejun Heo7a4e3442013-03-12 11:30:00 -0700169 struct workqueue_attrs *attrs; /* I: worker attributes */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700170 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
171 int refcnt; /* PL: refcnt for unbound pools */
Tejun Heo7a4e3442013-03-12 11:30:00 -0700172
Tejun Heoe19e3972013-01-24 11:39:44 -0800173 /*
174 * The current concurrency level. As it's likely to be accessed
175 * from other CPUs during try_to_wake_up(), put it in a separate
176 * cacheline.
177 */
178 atomic_t nr_running ____cacheline_aligned_in_smp;
Tejun Heo29c91e92013-03-12 11:30:03 -0700179
180 /*
181 * Destruction of pool is sched-RCU protected to allow dereferences
182 * from get_work_pool().
183 */
184 struct rcu_head rcu;
Tejun Heo8b03ae32010-06-29 10:07:12 +0200185} ____cacheline_aligned_in_smp;
186
187/*
Tejun Heo112202d2013-02-13 19:29:12 -0800188 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
189 * of work_struct->data are used for flags and the remaining high bits
190 * point to the pwq; thus, pwqs need to be aligned at two's power of the
191 * number of flag bits.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 */
Tejun Heo112202d2013-02-13 19:29:12 -0800193struct pool_workqueue {
Tejun Heobd7bdd42012-07-12 14:46:37 -0700194 struct worker_pool *pool; /* I: the associated pool */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200195 struct workqueue_struct *wq; /* I: the owning workqueue */
Tejun Heo73f53c42010-06-29 10:07:11 +0200196 int work_color; /* L: current color */
197 int flush_color; /* L: flushing color */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700198 int refcnt; /* L: reference count */
Tejun Heo73f53c42010-06-29 10:07:11 +0200199 int nr_in_flight[WORK_NR_COLORS];
200 /* L: nr of in_flight works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200201 int nr_active; /* L: nr of active works */
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200202 int max_active; /* L: max active works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200203 struct list_head delayed_works; /* L: delayed works */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700204 struct list_head pwqs_node; /* WR: node on wq->pwqs */
Tejun Heo2e109a22013-03-13 19:47:40 -0700205 struct list_head mayday_node; /* MD: node on wq->maydays */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700206
207 /*
208 * Release of unbound pwq is punted to system_wq. See put_pwq()
209 * and pwq_unbound_release_workfn() for details. pool_workqueue
210 * itself is also sched-RCU protected so that the first pwq can be
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700211 * determined without grabbing wq->mutex.
Tejun Heo8864b4e2013-03-12 11:30:04 -0700212 */
213 struct work_struct unbound_release_work;
214 struct rcu_head rcu;
Tejun Heoe904e6c2013-03-12 11:29:57 -0700215} __aligned(1 << WORK_STRUCT_FLAG_BITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217/*
Tejun Heo73f53c42010-06-29 10:07:11 +0200218 * Structure used to wait for workqueue flush.
219 */
220struct wq_flusher {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700221 struct list_head list; /* WQ: list of flushers */
222 int flush_color; /* WQ: flush color waiting for */
Tejun Heo73f53c42010-06-29 10:07:11 +0200223 struct completion done; /* flush completion */
224};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225
Tejun Heo226223a2013-03-12 11:30:05 -0700226struct wq_device;
227
Tejun Heo73f53c42010-06-29 10:07:11 +0200228/*
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700229 * The externally visible workqueue. It relays the issued work items to
230 * the appropriate worker_pool through its pool_workqueues.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 */
232struct workqueue_struct {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700233 struct list_head pwqs; /* WR: all pwqs of this wq */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700234 struct list_head list; /* PL: list of all workqueues */
Tejun Heo73f53c42010-06-29 10:07:11 +0200235
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700236 struct mutex mutex; /* protects this wq */
237 int work_color; /* WQ: current work color */
238 int flush_color; /* WQ: current flush color */
Tejun Heo112202d2013-02-13 19:29:12 -0800239 atomic_t nr_pwqs_to_flush; /* flush in progress */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700240 struct wq_flusher *first_flusher; /* WQ: first flusher */
241 struct list_head flusher_queue; /* WQ: flush waiters */
242 struct list_head flusher_overflow; /* WQ: flush overflow list */
Tejun Heo73f53c42010-06-29 10:07:11 +0200243
Tejun Heo2e109a22013-03-13 19:47:40 -0700244 struct list_head maydays; /* MD: pwqs requesting rescue */
Tejun Heoe22bee72010-06-29 10:07:14 +0200245 struct worker *rescuer; /* I: rescue worker */
246
Lai Jiangshan87fc7412013-03-25 16:57:18 -0700247 int nr_drainers; /* WQ: drain in progress */
Lai Jiangshana357fc02013-03-25 16:57:19 -0700248 int saved_max_active; /* WQ: saved pwq max_active */
Tejun Heo226223a2013-03-12 11:30:05 -0700249
Tejun Heo6029a912013-04-01 11:23:34 -0700250 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
Tejun Heo4c16bd32013-04-01 11:23:36 -0700251 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
Tejun Heo6029a912013-04-01 11:23:34 -0700252
Tejun Heo226223a2013-03-12 11:30:05 -0700253#ifdef CONFIG_SYSFS
254 struct wq_device *wq_dev; /* I: for sysfs interface */
255#endif
Johannes Berg4e6045f2007-10-18 23:39:55 -0700256#ifdef CONFIG_LOCKDEP
Tejun Heo4690c4a2010-06-29 10:07:10 +0200257 struct lockdep_map lockdep_map;
Johannes Berg4e6045f2007-10-18 23:39:55 -0700258#endif
Tejun Heoecf68812013-04-01 11:23:34 -0700259 char name[WQ_NAME_LEN]; /* I: workqueue name */
Tejun Heo2728fd22013-04-01 11:23:35 -0700260
261 /* hot fields used during command issue, aligned to cacheline */
262 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
263 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700264 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265};
266
Tejun Heoe904e6c2013-03-12 11:29:57 -0700267static struct kmem_cache *pwq_cache;
268
Tejun Heobce90382013-04-01 11:23:32 -0700269static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
270static cpumask_var_t *wq_numa_possible_cpumask;
271 /* possible CPUs of each node */
272
Tejun Heod55262c2013-04-01 11:23:38 -0700273static bool wq_disable_numa;
274module_param_named(disable_numa, wq_disable_numa, bool, 0444);
275
Viresh Kumarcee22a12013-04-08 16:45:40 +0530276/* see the comment above the definition of WQ_POWER_EFFICIENT */
277#ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
278static bool wq_power_efficient = true;
279#else
280static bool wq_power_efficient;
281#endif
282
283module_param_named(power_efficient, wq_power_efficient, bool, 0444);
284
Tejun Heobce90382013-04-01 11:23:32 -0700285static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
286
Tejun Heo4c16bd32013-04-01 11:23:36 -0700287/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
288static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
289
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700290static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
Tejun Heo2e109a22013-03-13 19:47:40 -0700291static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
Tejun Heo5bcab332013-03-13 19:47:40 -0700292
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700293static LIST_HEAD(workqueues); /* PL: list of all workqueues */
294static bool workqueue_freezing; /* PL: have wqs started freezing? */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700295
296/* the per-cpu worker pools */
297static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
298 cpu_worker_pools);
299
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700300static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700301
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700302/* PL: hash of all unbound pools keyed by pool->attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -0700303static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
304
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700305/* I: attributes used when instantiating standard unbound pools on demand */
Tejun Heo29c91e92013-03-12 11:30:03 -0700306static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
307
Tejun Heo8a2b7532013-09-05 12:30:04 -0400308/* I: attributes used when instantiating ordered pools on demand */
309static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
310
Tejun Heod320c032010-06-29 10:07:14 +0200311struct workqueue_struct *system_wq __read_mostly;
Marc Dionnead7b1f82013-05-06 17:44:55 -0400312EXPORT_SYMBOL(system_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300313struct workqueue_struct *system_highpri_wq __read_mostly;
Joonsoo Kim1aabe902012-08-15 23:25:39 +0900314EXPORT_SYMBOL_GPL(system_highpri_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300315struct workqueue_struct *system_long_wq __read_mostly;
Tejun Heod320c032010-06-29 10:07:14 +0200316EXPORT_SYMBOL_GPL(system_long_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300317struct workqueue_struct *system_unbound_wq __read_mostly;
Tejun Heof3421792010-07-02 10:03:51 +0200318EXPORT_SYMBOL_GPL(system_unbound_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300319struct workqueue_struct *system_freezable_wq __read_mostly;
Tejun Heo24d51ad2011-02-21 09:52:50 +0100320EXPORT_SYMBOL_GPL(system_freezable_wq);
Viresh Kumar06681062013-04-24 17:12:54 +0530321struct workqueue_struct *system_power_efficient_wq __read_mostly;
322EXPORT_SYMBOL_GPL(system_power_efficient_wq);
323struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
324EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
Tejun Heod320c032010-06-29 10:07:14 +0200325
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700326static int worker_thread(void *__worker);
327static void copy_workqueue_attrs(struct workqueue_attrs *to,
328 const struct workqueue_attrs *from);
329
Tejun Heo97bd2342010-10-05 10:41:14 +0200330#define CREATE_TRACE_POINTS
331#include <trace/events/workqueue.h>
332
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700333#define assert_rcu_or_pool_mutex() \
Tejun Heo5bcab332013-03-13 19:47:40 -0700334 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700335 lockdep_is_held(&wq_pool_mutex), \
336 "sched RCU or wq_pool_mutex should be held")
Tejun Heo5bcab332013-03-13 19:47:40 -0700337
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700338#define assert_rcu_or_wq_mutex(wq) \
Tejun Heo76af4d92013-03-12 11:30:00 -0700339 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
Lai Jiangshanb5927602013-03-25 16:57:19 -0700340 lockdep_is_held(&wq->mutex), \
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700341 "sched RCU or wq->mutex should be held")
Tejun Heo76af4d92013-03-12 11:30:00 -0700342
Tejun Heo822d8402013-03-19 13:45:21 -0700343#ifdef CONFIG_LOCKDEP
344#define assert_manager_or_pool_lock(pool) \
Lai Jiangshan519e3c12013-03-20 03:28:21 +0800345 WARN_ONCE(debug_locks && \
346 !lockdep_is_held(&(pool)->manager_mutex) && \
Tejun Heo822d8402013-03-19 13:45:21 -0700347 !lockdep_is_held(&(pool)->lock), \
348 "pool->manager_mutex or ->lock should be held")
349#else
350#define assert_manager_or_pool_lock(pool) do { } while (0)
351#endif
352
Tejun Heof02ae732013-03-12 11:30:03 -0700353#define for_each_cpu_worker_pool(pool, cpu) \
354 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
355 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
Tejun Heo7a62c2c2013-03-12 11:30:03 -0700356 (pool)++)
Tejun Heo4ce62e92012-07-13 22:16:44 -0700357
Tejun Heo49e3cf42013-03-12 11:29:58 -0700358/**
Tejun Heo17116962013-03-12 11:29:58 -0700359 * for_each_pool - iterate through all worker_pools in the system
360 * @pool: iteration cursor
Tejun Heo611c92a2013-03-13 16:51:36 -0700361 * @pi: integer used for iteration
Tejun Heofa1b54e2013-03-12 11:30:00 -0700362 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700363 * This must be called either with wq_pool_mutex held or sched RCU read
364 * locked. If the pool needs to be used beyond the locking in effect, the
365 * caller is responsible for guaranteeing that the pool stays online.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700366 *
367 * The if/else clause exists only for the lockdep assertion and can be
368 * ignored.
Tejun Heo17116962013-03-12 11:29:58 -0700369 */
Tejun Heo611c92a2013-03-13 16:51:36 -0700370#define for_each_pool(pool, pi) \
371 idr_for_each_entry(&worker_pool_idr, pool, pi) \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700372 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
Tejun Heofa1b54e2013-03-12 11:30:00 -0700373 else
Tejun Heo17116962013-03-12 11:29:58 -0700374
375/**
Tejun Heo822d8402013-03-19 13:45:21 -0700376 * for_each_pool_worker - iterate through all workers of a worker_pool
377 * @worker: iteration cursor
378 * @wi: integer used for iteration
379 * @pool: worker_pool to iterate workers of
380 *
381 * This must be called with either @pool->manager_mutex or ->lock held.
382 *
383 * The if/else clause exists only for the lockdep assertion and can be
384 * ignored.
385 */
386#define for_each_pool_worker(worker, wi, pool) \
387 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
388 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
389 else
390
391/**
Tejun Heo49e3cf42013-03-12 11:29:58 -0700392 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
393 * @pwq: iteration cursor
394 * @wq: the target workqueue
Tejun Heo76af4d92013-03-12 11:30:00 -0700395 *
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700396 * This must be called either with wq->mutex held or sched RCU read locked.
Tejun Heo794b18b2013-03-13 19:47:40 -0700397 * If the pwq needs to be used beyond the locking in effect, the caller is
398 * responsible for guaranteeing that the pwq stays online.
Tejun Heo76af4d92013-03-12 11:30:00 -0700399 *
400 * The if/else clause exists only for the lockdep assertion and can be
401 * ignored.
Tejun Heo49e3cf42013-03-12 11:29:58 -0700402 */
403#define for_each_pwq(pwq, wq) \
Tejun Heo76af4d92013-03-12 11:30:00 -0700404 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700405 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
Tejun Heo76af4d92013-03-12 11:30:00 -0700406 else
Tejun Heof3421792010-07-02 10:03:51 +0200407
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900408#ifdef CONFIG_DEBUG_OBJECTS_WORK
409
410static struct debug_obj_descr work_debug_descr;
411
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100412static void *work_debug_hint(void *addr)
413{
414 return ((struct work_struct *) addr)->func;
415}
416
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900417/*
418 * fixup_init is called when:
419 * - an active object is initialized
420 */
421static int work_fixup_init(void *addr, enum debug_obj_state state)
422{
423 struct work_struct *work = addr;
424
425 switch (state) {
426 case ODEBUG_STATE_ACTIVE:
427 cancel_work_sync(work);
428 debug_object_init(work, &work_debug_descr);
429 return 1;
430 default:
431 return 0;
432 }
433}
434
435/*
436 * fixup_activate is called when:
437 * - an active object is activated
438 * - an unknown object is activated (might be a statically initialized object)
439 */
440static int work_fixup_activate(void *addr, enum debug_obj_state state)
441{
442 struct work_struct *work = addr;
443
444 switch (state) {
445
446 case ODEBUG_STATE_NOTAVAILABLE:
447 /*
448 * This is not really a fixup. The work struct was
449 * statically initialized. We just make sure that it
450 * is tracked in the object tracker.
451 */
Tejun Heo22df02b2010-06-29 10:07:10 +0200452 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900453 debug_object_init(work, &work_debug_descr);
454 debug_object_activate(work, &work_debug_descr);
455 return 0;
456 }
457 WARN_ON_ONCE(1);
458 return 0;
459
460 case ODEBUG_STATE_ACTIVE:
461 WARN_ON(1);
462
463 default:
464 return 0;
465 }
466}
467
468/*
469 * fixup_free is called when:
470 * - an active object is freed
471 */
472static int work_fixup_free(void *addr, enum debug_obj_state state)
473{
474 struct work_struct *work = addr;
475
476 switch (state) {
477 case ODEBUG_STATE_ACTIVE:
478 cancel_work_sync(work);
479 debug_object_free(work, &work_debug_descr);
480 return 1;
481 default:
482 return 0;
483 }
484}
485
486static struct debug_obj_descr work_debug_descr = {
487 .name = "work_struct",
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100488 .debug_hint = work_debug_hint,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900489 .fixup_init = work_fixup_init,
490 .fixup_activate = work_fixup_activate,
491 .fixup_free = work_fixup_free,
492};
493
494static inline void debug_work_activate(struct work_struct *work)
495{
496 debug_object_activate(work, &work_debug_descr);
497}
498
499static inline void debug_work_deactivate(struct work_struct *work)
500{
501 debug_object_deactivate(work, &work_debug_descr);
502}
503
504void __init_work(struct work_struct *work, int onstack)
505{
506 if (onstack)
507 debug_object_init_on_stack(work, &work_debug_descr);
508 else
509 debug_object_init(work, &work_debug_descr);
510}
511EXPORT_SYMBOL_GPL(__init_work);
512
513void destroy_work_on_stack(struct work_struct *work)
514{
515 debug_object_free(work, &work_debug_descr);
516}
517EXPORT_SYMBOL_GPL(destroy_work_on_stack);
518
Thomas Gleixnerea2e64f2014-03-23 14:20:44 +0000519void destroy_delayed_work_on_stack(struct delayed_work *work)
520{
521 destroy_timer_on_stack(&work->timer);
522 debug_object_free(&work->work, &work_debug_descr);
523}
524EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
525
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900526#else
527static inline void debug_work_activate(struct work_struct *work) { }
528static inline void debug_work_deactivate(struct work_struct *work) { }
529#endif
530
Li Bin4e8b22b2013-09-10 09:52:35 +0800531/**
532 * worker_pool_assign_id - allocate ID and assing it to @pool
533 * @pool: the pool pointer of interest
534 *
535 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
536 * successfully, -errno on failure.
537 */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800538static int worker_pool_assign_id(struct worker_pool *pool)
539{
540 int ret;
541
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700542 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo5bcab332013-03-13 19:47:40 -0700543
Li Bin4e8b22b2013-09-10 09:52:35 +0800544 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
545 GFP_KERNEL);
Tejun Heo229641a2013-04-01 17:08:13 -0700546 if (ret >= 0) {
Tejun Heoe68035f2013-03-13 14:59:38 -0700547 pool->id = ret;
Tejun Heo229641a2013-04-01 17:08:13 -0700548 return 0;
549 }
Tejun Heo9daf9e62013-01-24 11:01:33 -0800550 return ret;
551}
552
Tejun Heo76af4d92013-03-12 11:30:00 -0700553/**
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700554 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
555 * @wq: the target workqueue
556 * @node: the node ID
557 *
558 * This must be called either with pwq_lock held or sched RCU read locked.
559 * If the pwq needs to be used beyond the locking in effect, the caller is
560 * responsible for guaranteeing that the pwq stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700561 *
562 * Return: The unbound pool_workqueue for @node.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700563 */
564static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
565 int node)
566{
567 assert_rcu_or_wq_mutex(wq);
568 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
569}
570
Tejun Heo73f53c42010-06-29 10:07:11 +0200571static unsigned int work_color_to_flags(int color)
572{
573 return color << WORK_STRUCT_COLOR_SHIFT;
574}
575
576static int get_work_color(struct work_struct *work)
577{
578 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
579 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
580}
581
582static int work_next_color(int color)
583{
584 return (color + 1) % WORK_NR_COLORS;
Oleg Nesterova848e3b2007-05-09 02:34:17 -0700585}
586
David Howells4594bf12006-12-07 11:33:26 +0000587/*
Tejun Heo112202d2013-02-13 19:29:12 -0800588 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
589 * contain the pointer to the queued pwq. Once execution starts, the flag
Tejun Heo7c3eed52013-01-24 11:01:33 -0800590 * is cleared and the high bits contain OFFQ flags and pool ID.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200591 *
Tejun Heo112202d2013-02-13 19:29:12 -0800592 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
593 * and clear_work_data() can be used to set the pwq, pool or clear
Tejun Heobbb68df2012-08-03 10:30:46 -0700594 * work->data. These functions should only be called while the work is
595 * owned - ie. while the PENDING bit is set.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200596 *
Tejun Heo112202d2013-02-13 19:29:12 -0800597 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
Tejun Heo7c3eed52013-01-24 11:01:33 -0800598 * corresponding to a work. Pool is available once the work has been
Tejun Heo112202d2013-02-13 19:29:12 -0800599 * queued anywhere after initialization until it is sync canceled. pwq is
Tejun Heo7c3eed52013-01-24 11:01:33 -0800600 * available only while the work item is queued.
Tejun Heobbb68df2012-08-03 10:30:46 -0700601 *
602 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
603 * canceled. While being canceled, a work item may have its PENDING set
604 * but stay off timer and worklist for arbitrarily long and nobody should
605 * try to steal the PENDING bit.
David Howells4594bf12006-12-07 11:33:26 +0000606 */
Tejun Heo7a22ad72010-06-29 10:07:13 +0200607static inline void set_work_data(struct work_struct *work, unsigned long data,
608 unsigned long flags)
David Howells365970a2006-11-22 14:54:49 +0000609{
Tejun Heo6183c002013-03-12 11:29:57 -0700610 WARN_ON_ONCE(!work_pending(work));
Tejun Heo7a22ad72010-06-29 10:07:13 +0200611 atomic_long_set(&work->data, data | flags | work_static(work));
David Howells365970a2006-11-22 14:54:49 +0000612}
David Howells365970a2006-11-22 14:54:49 +0000613
Tejun Heo112202d2013-02-13 19:29:12 -0800614static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
Tejun Heo7a22ad72010-06-29 10:07:13 +0200615 unsigned long extra_flags)
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200616{
Tejun Heo112202d2013-02-13 19:29:12 -0800617 set_work_data(work, (unsigned long)pwq,
618 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200619}
620
Lai Jiangshan4468a002013-02-06 18:04:53 -0800621static void set_work_pool_and_keep_pending(struct work_struct *work,
622 int pool_id)
623{
624 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
625 WORK_STRUCT_PENDING);
626}
627
Tejun Heo7c3eed52013-01-24 11:01:33 -0800628static void set_work_pool_and_clear_pending(struct work_struct *work,
629 int pool_id)
David Howells365970a2006-11-22 14:54:49 +0000630{
Tejun Heo23657bb2012-08-13 17:08:19 -0700631 /*
632 * The following wmb is paired with the implied mb in
633 * test_and_set_bit(PENDING) and ensures all updates to @work made
634 * here are visible to and precede any updates by the next PENDING
635 * owner.
636 */
637 smp_wmb();
Tejun Heo7c3eed52013-01-24 11:01:33 -0800638 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200639}
640
641static void clear_work_data(struct work_struct *work)
642{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800643 smp_wmb(); /* see set_work_pool_and_clear_pending() */
644 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200645}
646
Tejun Heo112202d2013-02-13 19:29:12 -0800647static struct pool_workqueue *get_work_pwq(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200648{
Tejun Heoe1201532010-07-22 14:14:25 +0200649 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200650
Tejun Heo112202d2013-02-13 19:29:12 -0800651 if (data & WORK_STRUCT_PWQ)
Tejun Heoe1201532010-07-22 14:14:25 +0200652 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
653 else
654 return NULL;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200655}
656
Tejun Heo7c3eed52013-01-24 11:01:33 -0800657/**
658 * get_work_pool - return the worker_pool a given work was associated with
659 * @work: the work item of interest
660 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700661 * Pools are created and destroyed under wq_pool_mutex, and allows read
662 * access under sched-RCU read lock. As such, this function should be
663 * called under wq_pool_mutex or with preemption disabled.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700664 *
665 * All fields of the returned pool are accessible as long as the above
666 * mentioned locking is in effect. If the returned pool needs to be used
667 * beyond the critical section, the caller is responsible for ensuring the
668 * returned pool is and stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700669 *
670 * Return: The worker_pool @work was last associated with. %NULL if none.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800671 */
672static struct worker_pool *get_work_pool(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200673{
Tejun Heoe1201532010-07-22 14:14:25 +0200674 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800675 int pool_id;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200676
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700677 assert_rcu_or_pool_mutex();
Tejun Heofa1b54e2013-03-12 11:30:00 -0700678
Tejun Heo112202d2013-02-13 19:29:12 -0800679 if (data & WORK_STRUCT_PWQ)
680 return ((struct pool_workqueue *)
Tejun Heo7c3eed52013-01-24 11:01:33 -0800681 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200682
Tejun Heo7c3eed52013-01-24 11:01:33 -0800683 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
684 if (pool_id == WORK_OFFQ_POOL_NONE)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200685 return NULL;
686
Tejun Heofa1b54e2013-03-12 11:30:00 -0700687 return idr_find(&worker_pool_idr, pool_id);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800688}
689
690/**
691 * get_work_pool_id - return the worker pool ID a given work is associated with
692 * @work: the work item of interest
693 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700694 * Return: The worker_pool ID @work was last associated with.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800695 * %WORK_OFFQ_POOL_NONE if none.
696 */
697static int get_work_pool_id(struct work_struct *work)
698{
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800699 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800700
Tejun Heo112202d2013-02-13 19:29:12 -0800701 if (data & WORK_STRUCT_PWQ)
702 return ((struct pool_workqueue *)
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800703 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
704
705 return data >> WORK_OFFQ_POOL_SHIFT;
Tejun Heo7c3eed52013-01-24 11:01:33 -0800706}
707
Tejun Heobbb68df2012-08-03 10:30:46 -0700708static void mark_work_canceling(struct work_struct *work)
709{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800710 unsigned long pool_id = get_work_pool_id(work);
Tejun Heobbb68df2012-08-03 10:30:46 -0700711
Tejun Heo7c3eed52013-01-24 11:01:33 -0800712 pool_id <<= WORK_OFFQ_POOL_SHIFT;
713 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700714}
715
716static bool work_is_canceling(struct work_struct *work)
717{
718 unsigned long data = atomic_long_read(&work->data);
719
Tejun Heo112202d2013-02-13 19:29:12 -0800720 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700721}
722
David Howells365970a2006-11-22 14:54:49 +0000723/*
Tejun Heo32704762012-07-13 22:16:45 -0700724 * Policy functions. These define the policies on how the global worker
725 * pools are managed. Unless noted otherwise, these functions assume that
Tejun Heod565ed62013-01-24 11:01:33 -0800726 * they're being called with pool->lock held.
David Howells365970a2006-11-22 14:54:49 +0000727 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200728
Tejun Heo63d95a92012-07-12 14:46:37 -0700729static bool __need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000730{
Tejun Heoe19e3972013-01-24 11:39:44 -0800731 return !atomic_read(&pool->nr_running);
David Howells365970a2006-11-22 14:54:49 +0000732}
733
Tejun Heoe22bee72010-06-29 10:07:14 +0200734/*
735 * Need to wake up a worker? Called from anything but currently
736 * running workers.
Tejun Heo974271c2012-07-12 14:46:37 -0700737 *
738 * Note that, because unbound workers never contribute to nr_running, this
Tejun Heo706026c2013-01-24 11:01:34 -0800739 * function will always return %true for unbound pools as long as the
Tejun Heo974271c2012-07-12 14:46:37 -0700740 * worklist isn't empty.
Tejun Heoe22bee72010-06-29 10:07:14 +0200741 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700742static bool need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000743{
Tejun Heo63d95a92012-07-12 14:46:37 -0700744 return !list_empty(&pool->worklist) && __need_more_worker(pool);
David Howells365970a2006-11-22 14:54:49 +0000745}
746
Tejun Heoe22bee72010-06-29 10:07:14 +0200747/* Can I start working? Called from busy but !running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700748static bool may_start_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200749{
Tejun Heo63d95a92012-07-12 14:46:37 -0700750 return pool->nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200751}
752
753/* Do I need to keep working? Called from currently running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700754static bool keep_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200755{
Tejun Heoe19e3972013-01-24 11:39:44 -0800756 return !list_empty(&pool->worklist) &&
757 atomic_read(&pool->nr_running) <= 1;
Tejun Heoe22bee72010-06-29 10:07:14 +0200758}
759
760/* Do we need a new worker? Called from manager. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700761static bool need_to_create_worker(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200762{
Tejun Heo63d95a92012-07-12 14:46:37 -0700763 return need_more_worker(pool) && !may_start_working(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200764}
765
766/* Do I need to be the manager? */
Tejun Heo63d95a92012-07-12 14:46:37 -0700767static bool need_to_manage_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200768{
Tejun Heo63d95a92012-07-12 14:46:37 -0700769 return need_to_create_worker(pool) ||
Tejun Heo11ebea52012-07-12 14:46:37 -0700770 (pool->flags & POOL_MANAGE_WORKERS);
Tejun Heoe22bee72010-06-29 10:07:14 +0200771}
772
773/* Do we have too many workers and should some go away? */
Tejun Heo63d95a92012-07-12 14:46:37 -0700774static bool too_many_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200775{
Tejun Heo34a06bd2013-03-12 11:30:00 -0700776 bool managing = mutex_is_locked(&pool->manager_arb);
Tejun Heo63d95a92012-07-12 14:46:37 -0700777 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
778 int nr_busy = pool->nr_workers - nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200779
Lai Jiangshanea1abd62012-09-18 09:59:22 -0700780 /*
781 * nr_idle and idle_list may disagree if idle rebinding is in
782 * progress. Never return %true if idle_list is empty.
783 */
784 if (list_empty(&pool->idle_list))
785 return false;
786
Tejun Heoe22bee72010-06-29 10:07:14 +0200787 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
788}
789
790/*
791 * Wake up functions.
792 */
793
Tejun Heo7e116292010-06-29 10:07:13 +0200794/* Return the first worker. Safe with preemption disabled */
Tejun Heo63d95a92012-07-12 14:46:37 -0700795static struct worker *first_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200796{
Tejun Heo63d95a92012-07-12 14:46:37 -0700797 if (unlikely(list_empty(&pool->idle_list)))
Tejun Heo7e116292010-06-29 10:07:13 +0200798 return NULL;
799
Tejun Heo63d95a92012-07-12 14:46:37 -0700800 return list_first_entry(&pool->idle_list, struct worker, entry);
Tejun Heo7e116292010-06-29 10:07:13 +0200801}
802
803/**
804 * wake_up_worker - wake up an idle worker
Tejun Heo63d95a92012-07-12 14:46:37 -0700805 * @pool: worker pool to wake worker from
Tejun Heo7e116292010-06-29 10:07:13 +0200806 *
Tejun Heo63d95a92012-07-12 14:46:37 -0700807 * Wake up the first idle worker of @pool.
Tejun Heo7e116292010-06-29 10:07:13 +0200808 *
809 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800810 * spin_lock_irq(pool->lock).
Tejun Heo7e116292010-06-29 10:07:13 +0200811 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700812static void wake_up_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200813{
Tejun Heo63d95a92012-07-12 14:46:37 -0700814 struct worker *worker = first_worker(pool);
Tejun Heo7e116292010-06-29 10:07:13 +0200815
816 if (likely(worker))
817 wake_up_process(worker->task);
818}
819
Tejun Heo4690c4a2010-06-29 10:07:10 +0200820/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200821 * wq_worker_waking_up - a worker is waking up
822 * @task: task waking up
823 * @cpu: CPU @task is waking up to
824 *
825 * This function is called during try_to_wake_up() when a worker is
826 * being awoken.
827 *
828 * CONTEXT:
829 * spin_lock_irq(rq->lock)
830 */
Tejun Heod84ff052013-03-12 11:29:59 -0700831void wq_worker_waking_up(struct task_struct *task, int cpu)
Tejun Heoe22bee72010-06-29 10:07:14 +0200832{
833 struct worker *worker = kthread_data(task);
834
Joonsoo Kim36576002012-10-26 23:03:49 +0900835 if (!(worker->flags & WORKER_NOT_RUNNING)) {
Tejun Heoec22ca52013-01-24 11:01:33 -0800836 WARN_ON_ONCE(worker->pool->cpu != cpu);
Tejun Heoe19e3972013-01-24 11:39:44 -0800837 atomic_inc(&worker->pool->nr_running);
Joonsoo Kim36576002012-10-26 23:03:49 +0900838 }
Tejun Heoe22bee72010-06-29 10:07:14 +0200839}
840
841/**
842 * wq_worker_sleeping - a worker is going to sleep
843 * @task: task going to sleep
844 * @cpu: CPU in question, must be the current CPU number
845 *
846 * This function is called during schedule() when a busy worker is
847 * going to sleep. Worker on the same cpu can be woken up by
848 * returning pointer to its task.
849 *
850 * CONTEXT:
851 * spin_lock_irq(rq->lock)
852 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700853 * Return:
Tejun Heoe22bee72010-06-29 10:07:14 +0200854 * Worker task on @cpu to wake up, %NULL if none.
855 */
Tejun Heod84ff052013-03-12 11:29:59 -0700856struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
Tejun Heoe22bee72010-06-29 10:07:14 +0200857{
858 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
Tejun Heo111c2252013-01-17 17:16:24 -0800859 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200860
Tejun Heo111c2252013-01-17 17:16:24 -0800861 /*
862 * Rescuers, which may not have all the fields set up like normal
863 * workers, also reach here, let's not access anything before
864 * checking NOT_RUNNING.
865 */
Steven Rostedt2d646722010-12-03 23:12:33 -0500866 if (worker->flags & WORKER_NOT_RUNNING)
Tejun Heoe22bee72010-06-29 10:07:14 +0200867 return NULL;
868
Tejun Heo111c2252013-01-17 17:16:24 -0800869 pool = worker->pool;
Tejun Heo111c2252013-01-17 17:16:24 -0800870
Tejun Heoe22bee72010-06-29 10:07:14 +0200871 /* this can only happen on the local cpu */
Tejun Heo6183c002013-03-12 11:29:57 -0700872 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
873 return NULL;
Tejun Heoe22bee72010-06-29 10:07:14 +0200874
875 /*
876 * The counterpart of the following dec_and_test, implied mb,
877 * worklist not empty test sequence is in insert_work().
878 * Please read comment there.
879 *
Tejun Heo628c78e2012-07-17 12:39:27 -0700880 * NOT_RUNNING is clear. This means that we're bound to and
881 * running on the local cpu w/ rq lock held and preemption
882 * disabled, which in turn means that none else could be
Tejun Heod565ed62013-01-24 11:01:33 -0800883 * manipulating idle_list, so dereferencing idle_list without pool
Tejun Heo628c78e2012-07-17 12:39:27 -0700884 * lock is safe.
Tejun Heoe22bee72010-06-29 10:07:14 +0200885 */
Tejun Heoe19e3972013-01-24 11:39:44 -0800886 if (atomic_dec_and_test(&pool->nr_running) &&
887 !list_empty(&pool->worklist))
Tejun Heo63d95a92012-07-12 14:46:37 -0700888 to_wakeup = first_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200889 return to_wakeup ? to_wakeup->task : NULL;
890}
891
892/**
893 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200894 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200895 * @flags: flags to set
896 * @wakeup: wakeup an idle worker if necessary
897 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200898 * Set @flags in @worker->flags and adjust nr_running accordingly. If
899 * nr_running becomes zero and @wakeup is %true, an idle worker is
900 * woken up.
Tejun Heod302f012010-06-29 10:07:13 +0200901 *
Tejun Heocb444762010-07-02 10:03:50 +0200902 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800903 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200904 */
905static inline void worker_set_flags(struct worker *worker, unsigned int flags,
906 bool wakeup)
907{
Tejun Heobd7bdd42012-07-12 14:46:37 -0700908 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200909
Tejun Heocb444762010-07-02 10:03:50 +0200910 WARN_ON_ONCE(worker->task != current);
911
Tejun Heoe22bee72010-06-29 10:07:14 +0200912 /*
913 * If transitioning into NOT_RUNNING, adjust nr_running and
914 * wake up an idle worker as necessary if requested by
915 * @wakeup.
916 */
917 if ((flags & WORKER_NOT_RUNNING) &&
918 !(worker->flags & WORKER_NOT_RUNNING)) {
Tejun Heoe22bee72010-06-29 10:07:14 +0200919 if (wakeup) {
Tejun Heoe19e3972013-01-24 11:39:44 -0800920 if (atomic_dec_and_test(&pool->nr_running) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -0700921 !list_empty(&pool->worklist))
Tejun Heo63d95a92012-07-12 14:46:37 -0700922 wake_up_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200923 } else
Tejun Heoe19e3972013-01-24 11:39:44 -0800924 atomic_dec(&pool->nr_running);
Tejun Heoe22bee72010-06-29 10:07:14 +0200925 }
926
Tejun Heod302f012010-06-29 10:07:13 +0200927 worker->flags |= flags;
928}
929
930/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200931 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200932 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200933 * @flags: flags to clear
934 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200935 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200936 *
Tejun Heocb444762010-07-02 10:03:50 +0200937 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800938 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200939 */
940static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
941{
Tejun Heo63d95a92012-07-12 14:46:37 -0700942 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200943 unsigned int oflags = worker->flags;
944
Tejun Heocb444762010-07-02 10:03:50 +0200945 WARN_ON_ONCE(worker->task != current);
946
Tejun Heod302f012010-06-29 10:07:13 +0200947 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200948
Tejun Heo42c025f2011-01-11 15:58:49 +0100949 /*
950 * If transitioning out of NOT_RUNNING, increment nr_running. Note
951 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
952 * of multiple flags, not a single flag.
953 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200954 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
955 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800956 atomic_inc(&pool->nr_running);
Tejun Heod302f012010-06-29 10:07:13 +0200957}
958
959/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200960 * find_worker_executing_work - find worker which is executing a work
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800961 * @pool: pool of interest
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200962 * @work: work to find worker for
963 *
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800964 * Find a worker which is executing @work on @pool by searching
965 * @pool->busy_hash which is keyed by the address of @work. For a worker
Tejun Heoa2c1c572012-12-18 10:35:02 -0800966 * to match, its current execution should match the address of @work and
967 * its work function. This is to avoid unwanted dependency between
968 * unrelated work executions through a work item being recycled while still
969 * being executed.
970 *
971 * This is a bit tricky. A work item may be freed once its execution
972 * starts and nothing prevents the freed area from being recycled for
973 * another work item. If the same work item address ends up being reused
974 * before the original execution finishes, workqueue will identify the
975 * recycled work item as currently executing and make it wait until the
976 * current execution finishes, introducing an unwanted dependency.
977 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700978 * This function checks the work item address and work function to avoid
979 * false positives. Note that this isn't complete as one may construct a
980 * work function which can introduce dependency onto itself through a
981 * recycled work item. Well, if somebody wants to shoot oneself in the
982 * foot that badly, there's only so much we can do, and if such deadlock
983 * actually occurs, it should be easy to locate the culprit work function.
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200984 *
985 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800986 * spin_lock_irq(pool->lock).
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200987 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700988 * Return:
989 * Pointer to worker which is executing @work if found, %NULL
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200990 * otherwise.
991 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800992static struct worker *find_worker_executing_work(struct worker_pool *pool,
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200993 struct work_struct *work)
994{
Sasha Levin42f85702012-12-17 10:01:23 -0500995 struct worker *worker;
Sasha Levin42f85702012-12-17 10:01:23 -0500996
Sasha Levinb67bfe02013-02-27 17:06:00 -0800997 hash_for_each_possible(pool->busy_hash, worker, hentry,
Tejun Heoa2c1c572012-12-18 10:35:02 -0800998 (unsigned long)work)
999 if (worker->current_work == work &&
1000 worker->current_func == work->func)
Sasha Levin42f85702012-12-17 10:01:23 -05001001 return worker;
1002
1003 return NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001004}
1005
1006/**
Tejun Heobf4ede02012-08-03 10:30:46 -07001007 * move_linked_works - move linked works to a list
1008 * @work: start of series of works to be scheduled
1009 * @head: target list to append @work to
1010 * @nextp: out paramter for nested worklist walking
1011 *
1012 * Schedule linked works starting from @work to @head. Work series to
1013 * be scheduled starts at @work and includes any consecutive work with
1014 * WORK_STRUCT_LINKED set in its predecessor.
1015 *
1016 * If @nextp is not NULL, it's updated to point to the next work of
1017 * the last scheduled work. This allows move_linked_works() to be
1018 * nested inside outer list_for_each_entry_safe().
1019 *
1020 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001021 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001022 */
1023static void move_linked_works(struct work_struct *work, struct list_head *head,
1024 struct work_struct **nextp)
1025{
1026 struct work_struct *n;
1027
1028 /*
1029 * Linked worklist will always end before the end of the list,
1030 * use NULL for list head.
1031 */
1032 list_for_each_entry_safe_from(work, n, NULL, entry) {
1033 list_move_tail(&work->entry, head);
1034 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1035 break;
1036 }
1037
1038 /*
1039 * If we're already inside safe list traversal and have moved
1040 * multiple works to the scheduled queue, the next position
1041 * needs to be updated.
1042 */
1043 if (nextp)
1044 *nextp = n;
1045}
1046
Tejun Heo8864b4e2013-03-12 11:30:04 -07001047/**
1048 * get_pwq - get an extra reference on the specified pool_workqueue
1049 * @pwq: pool_workqueue to get
1050 *
1051 * Obtain an extra reference on @pwq. The caller should guarantee that
1052 * @pwq has positive refcnt and be holding the matching pool->lock.
1053 */
1054static void get_pwq(struct pool_workqueue *pwq)
1055{
1056 lockdep_assert_held(&pwq->pool->lock);
1057 WARN_ON_ONCE(pwq->refcnt <= 0);
1058 pwq->refcnt++;
1059}
1060
1061/**
1062 * put_pwq - put a pool_workqueue reference
1063 * @pwq: pool_workqueue to put
1064 *
1065 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1066 * destruction. The caller should be holding the matching pool->lock.
1067 */
1068static void put_pwq(struct pool_workqueue *pwq)
1069{
1070 lockdep_assert_held(&pwq->pool->lock);
1071 if (likely(--pwq->refcnt))
1072 return;
1073 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1074 return;
1075 /*
1076 * @pwq can't be released under pool->lock, bounce to
1077 * pwq_unbound_release_workfn(). This never recurses on the same
1078 * pool->lock as this path is taken only for unbound workqueues and
1079 * the release work item is scheduled on a per-cpu workqueue. To
1080 * avoid lockdep warning, unbound pool->locks are given lockdep
1081 * subclass of 1 in get_unbound_pool().
1082 */
1083 schedule_work(&pwq->unbound_release_work);
1084}
1085
Tejun Heodce90d42013-04-01 11:23:35 -07001086/**
1087 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1088 * @pwq: pool_workqueue to put (can be %NULL)
1089 *
1090 * put_pwq() with locking. This function also allows %NULL @pwq.
1091 */
1092static void put_pwq_unlocked(struct pool_workqueue *pwq)
1093{
1094 if (pwq) {
1095 /*
1096 * As both pwqs and pools are sched-RCU protected, the
1097 * following lock operations are safe.
1098 */
1099 spin_lock_irq(&pwq->pool->lock);
1100 put_pwq(pwq);
1101 spin_unlock_irq(&pwq->pool->lock);
1102 }
1103}
1104
Tejun Heo112202d2013-02-13 19:29:12 -08001105static void pwq_activate_delayed_work(struct work_struct *work)
Tejun Heobf4ede02012-08-03 10:30:46 -07001106{
Tejun Heo112202d2013-02-13 19:29:12 -08001107 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobf4ede02012-08-03 10:30:46 -07001108
1109 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001110 move_linked_works(work, &pwq->pool->worklist, NULL);
Tejun Heobf4ede02012-08-03 10:30:46 -07001111 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo112202d2013-02-13 19:29:12 -08001112 pwq->nr_active++;
Tejun Heobf4ede02012-08-03 10:30:46 -07001113}
1114
Tejun Heo112202d2013-02-13 19:29:12 -08001115static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001116{
Tejun Heo112202d2013-02-13 19:29:12 -08001117 struct work_struct *work = list_first_entry(&pwq->delayed_works,
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001118 struct work_struct, entry);
1119
Tejun Heo112202d2013-02-13 19:29:12 -08001120 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001121}
1122
Tejun Heobf4ede02012-08-03 10:30:46 -07001123/**
Tejun Heo112202d2013-02-13 19:29:12 -08001124 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1125 * @pwq: pwq of interest
Tejun Heobf4ede02012-08-03 10:30:46 -07001126 * @color: color of work which left the queue
Tejun Heobf4ede02012-08-03 10:30:46 -07001127 *
1128 * A work either has completed or is removed from pending queue,
Tejun Heo112202d2013-02-13 19:29:12 -08001129 * decrement nr_in_flight of its pwq and handle workqueue flushing.
Tejun Heobf4ede02012-08-03 10:30:46 -07001130 *
1131 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001132 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001133 */
Tejun Heo112202d2013-02-13 19:29:12 -08001134static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
Tejun Heobf4ede02012-08-03 10:30:46 -07001135{
Tejun Heo8864b4e2013-03-12 11:30:04 -07001136 /* uncolored work items don't participate in flushing or nr_active */
Tejun Heobf4ede02012-08-03 10:30:46 -07001137 if (color == WORK_NO_COLOR)
Tejun Heo8864b4e2013-03-12 11:30:04 -07001138 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001139
Tejun Heo112202d2013-02-13 19:29:12 -08001140 pwq->nr_in_flight[color]--;
Tejun Heobf4ede02012-08-03 10:30:46 -07001141
Tejun Heo112202d2013-02-13 19:29:12 -08001142 pwq->nr_active--;
1143 if (!list_empty(&pwq->delayed_works)) {
Lai Jiangshanb3f9f402012-09-18 10:40:00 -07001144 /* one down, submit a delayed one */
Tejun Heo112202d2013-02-13 19:29:12 -08001145 if (pwq->nr_active < pwq->max_active)
1146 pwq_activate_first_delayed(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001147 }
1148
1149 /* is flush in progress and are we at the flushing tip? */
Tejun Heo112202d2013-02-13 19:29:12 -08001150 if (likely(pwq->flush_color != color))
Tejun Heo8864b4e2013-03-12 11:30:04 -07001151 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001152
1153 /* are there still in-flight works? */
Tejun Heo112202d2013-02-13 19:29:12 -08001154 if (pwq->nr_in_flight[color])
Tejun Heo8864b4e2013-03-12 11:30:04 -07001155 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001156
Tejun Heo112202d2013-02-13 19:29:12 -08001157 /* this pwq is done, clear flush_color */
1158 pwq->flush_color = -1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001159
1160 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001161 * If this was the last pwq, wake up the first flusher. It
Tejun Heobf4ede02012-08-03 10:30:46 -07001162 * will handle the rest.
1163 */
Tejun Heo112202d2013-02-13 19:29:12 -08001164 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1165 complete(&pwq->wq->first_flusher->done);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001166out_put:
1167 put_pwq(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001168}
1169
Tejun Heo36e227d2012-08-03 10:30:46 -07001170/**
Tejun Heobbb68df2012-08-03 10:30:46 -07001171 * try_to_grab_pending - steal work item from worklist and disable irq
Tejun Heo36e227d2012-08-03 10:30:46 -07001172 * @work: work item to steal
1173 * @is_dwork: @work is a delayed_work
Tejun Heobbb68df2012-08-03 10:30:46 -07001174 * @flags: place to store irq state
Tejun Heo36e227d2012-08-03 10:30:46 -07001175 *
1176 * Try to grab PENDING bit of @work. This function can handle @work in any
Yacine Belkadid185af32013-07-31 14:59:24 -07001177 * stable state - idle, on timer or on worklist.
Tejun Heo36e227d2012-08-03 10:30:46 -07001178 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001179 * Return:
Tejun Heo36e227d2012-08-03 10:30:46 -07001180 * 1 if @work was pending and we successfully stole PENDING
1181 * 0 if @work was idle and we claimed PENDING
1182 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
Tejun Heobbb68df2012-08-03 10:30:46 -07001183 * -ENOENT if someone else is canceling @work, this state may persist
1184 * for arbitrarily long
Tejun Heo36e227d2012-08-03 10:30:46 -07001185 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001186 * Note:
Tejun Heobbb68df2012-08-03 10:30:46 -07001187 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001188 * interrupted while holding PENDING and @work off queue, irq must be
1189 * disabled on entry. This, combined with delayed_work->timer being
1190 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
Tejun Heobbb68df2012-08-03 10:30:46 -07001191 *
1192 * On successful return, >= 0, irq is disabled and the caller is
1193 * responsible for releasing it using local_irq_restore(*@flags).
1194 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001195 * This function is safe to call from any context including IRQ handler.
Tejun Heobf4ede02012-08-03 10:30:46 -07001196 */
Tejun Heobbb68df2012-08-03 10:30:46 -07001197static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1198 unsigned long *flags)
Tejun Heobf4ede02012-08-03 10:30:46 -07001199{
Tejun Heod565ed62013-01-24 11:01:33 -08001200 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001201 struct pool_workqueue *pwq;
Tejun Heobf4ede02012-08-03 10:30:46 -07001202
Tejun Heobbb68df2012-08-03 10:30:46 -07001203 local_irq_save(*flags);
1204
Tejun Heo36e227d2012-08-03 10:30:46 -07001205 /* try to steal the timer if it exists */
1206 if (is_dwork) {
1207 struct delayed_work *dwork = to_delayed_work(work);
1208
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001209 /*
1210 * dwork->timer is irqsafe. If del_timer() fails, it's
1211 * guaranteed that the timer is not queued anywhere and not
1212 * running on the local CPU.
1213 */
Tejun Heo36e227d2012-08-03 10:30:46 -07001214 if (likely(del_timer(&dwork->timer)))
1215 return 1;
1216 }
1217
1218 /* try to claim PENDING the normal way */
Tejun Heobf4ede02012-08-03 10:30:46 -07001219 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1220 return 0;
1221
1222 /*
1223 * The queueing is in progress, or it is already queued. Try to
1224 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1225 */
Tejun Heod565ed62013-01-24 11:01:33 -08001226 pool = get_work_pool(work);
1227 if (!pool)
Tejun Heobbb68df2012-08-03 10:30:46 -07001228 goto fail;
Tejun Heobf4ede02012-08-03 10:30:46 -07001229
Tejun Heod565ed62013-01-24 11:01:33 -08001230 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001231 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001232 * work->data is guaranteed to point to pwq only while the work
1233 * item is queued on pwq->wq, and both updating work->data to point
1234 * to pwq on queueing and to pool on dequeueing are done under
1235 * pwq->pool->lock. This in turn guarantees that, if work->data
1236 * points to pwq which is associated with a locked pool, the work
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001237 * item is currently queued on that pool.
1238 */
Tejun Heo112202d2013-02-13 19:29:12 -08001239 pwq = get_work_pwq(work);
1240 if (pwq && pwq->pool == pool) {
Tejun Heo16062832013-02-06 18:04:53 -08001241 debug_work_deactivate(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001242
Tejun Heo16062832013-02-06 18:04:53 -08001243 /*
1244 * A delayed work item cannot be grabbed directly because
1245 * it might have linked NO_COLOR work items which, if left
Tejun Heo112202d2013-02-13 19:29:12 -08001246 * on the delayed_list, will confuse pwq->nr_active
Tejun Heo16062832013-02-06 18:04:53 -08001247 * management later on and cause stall. Make sure the work
1248 * item is activated before grabbing.
1249 */
1250 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
Tejun Heo112202d2013-02-13 19:29:12 -08001251 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001252
Tejun Heo16062832013-02-06 18:04:53 -08001253 list_del_init(&work->entry);
Tejun Heo112202d2013-02-13 19:29:12 -08001254 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
Tejun Heo36e227d2012-08-03 10:30:46 -07001255
Tejun Heo112202d2013-02-13 19:29:12 -08001256 /* work->data points to pwq iff queued, point to pool */
Tejun Heo16062832013-02-06 18:04:53 -08001257 set_work_pool_and_keep_pending(work, pool->id);
Lai Jiangshan4468a002013-02-06 18:04:53 -08001258
Tejun Heo16062832013-02-06 18:04:53 -08001259 spin_unlock(&pool->lock);
1260 return 1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001261 }
Tejun Heod565ed62013-01-24 11:01:33 -08001262 spin_unlock(&pool->lock);
Tejun Heobbb68df2012-08-03 10:30:46 -07001263fail:
1264 local_irq_restore(*flags);
1265 if (work_is_canceling(work))
1266 return -ENOENT;
1267 cpu_relax();
Tejun Heo36e227d2012-08-03 10:30:46 -07001268 return -EAGAIN;
Tejun Heobf4ede02012-08-03 10:30:46 -07001269}
1270
1271/**
Tejun Heo706026c2013-01-24 11:01:34 -08001272 * insert_work - insert a work into a pool
Tejun Heo112202d2013-02-13 19:29:12 -08001273 * @pwq: pwq @work belongs to
Tejun Heo4690c4a2010-06-29 10:07:10 +02001274 * @work: work to insert
1275 * @head: insertion point
1276 * @extra_flags: extra WORK_STRUCT_* flags to set
1277 *
Tejun Heo112202d2013-02-13 19:29:12 -08001278 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
Tejun Heo706026c2013-01-24 11:01:34 -08001279 * work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +02001280 *
1281 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001282 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02001283 */
Tejun Heo112202d2013-02-13 19:29:12 -08001284static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1285 struct list_head *head, unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001286{
Tejun Heo112202d2013-02-13 19:29:12 -08001287 struct worker_pool *pool = pwq->pool;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01001288
Tejun Heo4690c4a2010-06-29 10:07:10 +02001289 /* we own @work, set data and link */
Tejun Heo112202d2013-02-13 19:29:12 -08001290 set_work_pwq(work, pwq, extra_flags);
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -07001291 list_add_tail(&work->entry, head);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001292 get_pwq(pwq);
Tejun Heoe22bee72010-06-29 10:07:14 +02001293
1294 /*
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001295 * Ensure either wq_worker_sleeping() sees the above
1296 * list_add_tail() or we see zero nr_running to avoid workers lying
1297 * around lazily while there are works to be processed.
Tejun Heoe22bee72010-06-29 10:07:14 +02001298 */
1299 smp_mb();
1300
Tejun Heo63d95a92012-07-12 14:46:37 -07001301 if (__need_more_worker(pool))
1302 wake_up_worker(pool);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001303}
1304
Tejun Heoc8efcc22010-12-20 19:32:04 +01001305/*
1306 * Test whether @work is being queued from another work executing on the
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001307 * same workqueue.
Tejun Heoc8efcc22010-12-20 19:32:04 +01001308 */
1309static bool is_chained_work(struct workqueue_struct *wq)
1310{
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001311 struct worker *worker;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001312
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001313 worker = current_wq_worker();
1314 /*
1315 * Return %true iff I'm a worker execuing a work item on @wq. If
1316 * I'm @worker, it's safe to dereference it without locking.
1317 */
Tejun Heo112202d2013-02-13 19:29:12 -08001318 return worker && worker->current_pwq->wq == wq;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001319}
1320
Tejun Heod84ff052013-03-12 11:29:59 -07001321static void __queue_work(int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 struct work_struct *work)
1323{
Tejun Heo112202d2013-02-13 19:29:12 -08001324 struct pool_workqueue *pwq;
Tejun Heoc9178082013-03-12 11:30:04 -07001325 struct worker_pool *last_pool;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001326 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001327 unsigned int work_flags;
Joonsoo Kimb75cac92012-08-15 23:25:37 +09001328 unsigned int req_cpu = cpu;
Tejun Heo8930cab2012-08-03 10:30:45 -07001329
1330 /*
1331 * While a work item is PENDING && off queue, a task trying to
1332 * steal the PENDING will busy-loop waiting for it to either get
1333 * queued or lose PENDING. Grabbing PENDING and queueing should
1334 * happen with IRQ disabled.
1335 */
1336 WARN_ON_ONCE(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09001338 debug_work_activate(work);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001339
Li Bin9ef28a72013-09-09 13:13:58 +08001340 /* if draining, only works from the same workqueue are allowed */
Tejun Heo618b01e2013-03-12 11:30:04 -07001341 if (unlikely(wq->flags & __WQ_DRAINING) &&
Tejun Heoc8efcc22010-12-20 19:32:04 +01001342 WARN_ON_ONCE(!is_chained_work(wq)))
Tejun Heoe41e7042010-08-24 14:22:47 +02001343 return;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001344retry:
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001345 if (req_cpu == WORK_CPU_UNBOUND)
1346 cpu = raw_smp_processor_id();
1347
Tejun Heoc9178082013-03-12 11:30:04 -07001348 /* pwq which will be used unless @work is executing elsewhere */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001349 if (!(wq->flags & WQ_UNBOUND))
Tejun Heo7fb98ea2013-03-12 11:30:00 -07001350 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001351 else
1352 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodbf25762012-08-20 14:51:23 -07001353
Tejun Heoc9178082013-03-12 11:30:04 -07001354 /*
1355 * If @work was previously on a different pool, it might still be
1356 * running there, in which case the work needs to be queued on that
1357 * pool to guarantee non-reentrancy.
1358 */
1359 last_pool = get_work_pool(work);
1360 if (last_pool && last_pool != pwq->pool) {
1361 struct worker *worker;
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001362
Tejun Heoc9178082013-03-12 11:30:04 -07001363 spin_lock(&last_pool->lock);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001364
Tejun Heoc9178082013-03-12 11:30:04 -07001365 worker = find_worker_executing_work(last_pool, work);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001366
Tejun Heoc9178082013-03-12 11:30:04 -07001367 if (worker && worker->current_pwq->wq == wq) {
1368 pwq = worker->current_pwq;
Tejun Heo8930cab2012-08-03 10:30:45 -07001369 } else {
Tejun Heoc9178082013-03-12 11:30:04 -07001370 /* meh... not running there, queue here */
1371 spin_unlock(&last_pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08001372 spin_lock(&pwq->pool->lock);
Tejun Heo8930cab2012-08-03 10:30:45 -07001373 }
Tejun Heof3421792010-07-02 10:03:51 +02001374 } else {
Tejun Heo112202d2013-02-13 19:29:12 -08001375 spin_lock(&pwq->pool->lock);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001376 }
1377
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001378 /*
1379 * pwq is determined and locked. For unbound pools, we could have
1380 * raced with pwq release and it could already be dead. If its
1381 * refcnt is zero, repeat pwq selection. Note that pwqs never die
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001382 * without another pwq replacing it in the numa_pwq_tbl or while
1383 * work items are executing on it, so the retrying is guaranteed to
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001384 * make forward-progress.
1385 */
1386 if (unlikely(!pwq->refcnt)) {
1387 if (wq->flags & WQ_UNBOUND) {
1388 spin_unlock(&pwq->pool->lock);
1389 cpu_relax();
1390 goto retry;
1391 }
1392 /* oops */
1393 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1394 wq->name, cpu);
1395 }
1396
Tejun Heo112202d2013-02-13 19:29:12 -08001397 /* pwq determined, queue */
1398 trace_workqueue_queue_work(req_cpu, pwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001399
Dan Carpenterf5b25522012-04-13 22:06:58 +03001400 if (WARN_ON(!list_empty(&work->entry))) {
Tejun Heo112202d2013-02-13 19:29:12 -08001401 spin_unlock(&pwq->pool->lock);
Dan Carpenterf5b25522012-04-13 22:06:58 +03001402 return;
1403 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001404
Tejun Heo112202d2013-02-13 19:29:12 -08001405 pwq->nr_in_flight[pwq->work_color]++;
1406 work_flags = work_color_to_flags(pwq->work_color);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001407
Tejun Heo112202d2013-02-13 19:29:12 -08001408 if (likely(pwq->nr_active < pwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001409 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001410 pwq->nr_active++;
1411 worklist = &pwq->pool->worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001412 } else {
1413 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo112202d2013-02-13 19:29:12 -08001414 worklist = &pwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001415 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001416
Tejun Heo112202d2013-02-13 19:29:12 -08001417 insert_work(pwq, work, worklist, work_flags);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001418
Tejun Heo112202d2013-02-13 19:29:12 -08001419 spin_unlock(&pwq->pool->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420}
1421
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001422/**
Zhang Ruic1a220e2008-07-23 21:28:39 -07001423 * queue_work_on - queue work on specific cpu
1424 * @cpu: CPU number to execute work on
1425 * @wq: workqueue to use
1426 * @work: work to queue
1427 *
Zhang Ruic1a220e2008-07-23 21:28:39 -07001428 * We queue the work to a specific CPU, the caller must ensure it
1429 * can't go away.
Yacine Belkadid185af32013-07-31 14:59:24 -07001430 *
1431 * Return: %false if @work was already on a queue, %true otherwise.
Zhang Ruic1a220e2008-07-23 21:28:39 -07001432 */
Tejun Heod4283e92012-08-03 10:30:44 -07001433bool queue_work_on(int cpu, struct workqueue_struct *wq,
1434 struct work_struct *work)
Zhang Ruic1a220e2008-07-23 21:28:39 -07001435{
Tejun Heod4283e92012-08-03 10:30:44 -07001436 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001437 unsigned long flags;
1438
1439 local_irq_save(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001440
Tejun Heo22df02b2010-06-29 10:07:10 +02001441 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001442 __queue_work(cpu, wq, work);
Tejun Heod4283e92012-08-03 10:30:44 -07001443 ret = true;
Zhang Ruic1a220e2008-07-23 21:28:39 -07001444 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001445
1446 local_irq_restore(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001447 return ret;
1448}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001449EXPORT_SYMBOL(queue_work_on);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001450
Tejun Heod8e794d2012-08-03 10:30:45 -07001451void delayed_work_timer_fn(unsigned long __data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452{
David Howells52bad642006-11-22 14:54:01 +00001453 struct delayed_work *dwork = (struct delayed_work *)__data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001455 /* should have been called from irqsafe timer with irq already off */
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001456 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457}
Konstantin Khlebnikov1438ade52013-01-24 16:36:31 +04001458EXPORT_SYMBOL(delayed_work_timer_fn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001460static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1461 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462{
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001463 struct timer_list *timer = &dwork->timer;
1464 struct work_struct *work = &dwork->work;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001466 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1467 timer->data != (unsigned long)dwork);
Tejun Heofc4b5142012-12-04 07:40:39 -08001468 WARN_ON_ONCE(timer_pending(timer));
1469 WARN_ON_ONCE(!list_empty(&work->entry));
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001470
Tejun Heo8852aac2012-12-01 16:23:42 -08001471 /*
1472 * If @delay is 0, queue @dwork->work immediately. This is for
1473 * both optimization and correctness. The earliest @timer can
1474 * expire is on the closest next tick and delayed_work users depend
1475 * on that there's no such delay when @delay is 0.
1476 */
1477 if (!delay) {
1478 __queue_work(cpu, wq, &dwork->work);
1479 return;
1480 }
1481
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001482 timer_stats_timer_set_start_info(&dwork->timer);
1483
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001484 dwork->wq = wq;
Tejun Heo12650572012-08-08 09:38:42 -07001485 dwork->cpu = cpu;
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001486 timer->expires = jiffies + delay;
1487
1488 if (unlikely(cpu != WORK_CPU_UNBOUND))
1489 add_timer_on(timer, cpu);
1490 else
1491 add_timer(timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492}
1493
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001494/**
1495 * queue_delayed_work_on - queue work on specific CPU after delay
1496 * @cpu: CPU number to execute work on
1497 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001498 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001499 * @delay: number of jiffies to wait before queueing
1500 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001501 * Return: %false if @work was already on a queue, %true otherwise. If
Tejun Heo715f1302012-08-03 10:30:46 -07001502 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1503 * execution.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001504 */
Tejun Heod4283e92012-08-03 10:30:44 -07001505bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1506 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001507{
David Howells52bad642006-11-22 14:54:01 +00001508 struct work_struct *work = &dwork->work;
Tejun Heod4283e92012-08-03 10:30:44 -07001509 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001510 unsigned long flags;
1511
1512 /* read the comment in __queue_work() */
1513 local_irq_save(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001514
Tejun Heo22df02b2010-06-29 10:07:10 +02001515 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001516 __queue_delayed_work(cpu, wq, dwork, delay);
Tejun Heod4283e92012-08-03 10:30:44 -07001517 ret = true;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001518 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001519
1520 local_irq_restore(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001521 return ret;
1522}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001523EXPORT_SYMBOL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524
Tejun Heoc8e55f32010-06-29 10:07:12 +02001525/**
Tejun Heo8376fe22012-08-03 10:30:47 -07001526 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1527 * @cpu: CPU number to execute work on
1528 * @wq: workqueue to use
1529 * @dwork: work to queue
1530 * @delay: number of jiffies to wait before queueing
1531 *
1532 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1533 * modify @dwork's timer so that it expires after @delay. If @delay is
1534 * zero, @work is guaranteed to be scheduled immediately regardless of its
1535 * current state.
1536 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001537 * Return: %false if @dwork was idle and queued, %true if @dwork was
Tejun Heo8376fe22012-08-03 10:30:47 -07001538 * pending and its timer was modified.
1539 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001540 * This function is safe to call from any context including IRQ handler.
Tejun Heo8376fe22012-08-03 10:30:47 -07001541 * See try_to_grab_pending() for details.
1542 */
1543bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1544 struct delayed_work *dwork, unsigned long delay)
1545{
1546 unsigned long flags;
1547 int ret;
1548
1549 do {
1550 ret = try_to_grab_pending(&dwork->work, true, &flags);
1551 } while (unlikely(ret == -EAGAIN));
1552
1553 if (likely(ret >= 0)) {
1554 __queue_delayed_work(cpu, wq, dwork, delay);
1555 local_irq_restore(flags);
1556 }
1557
1558 /* -ENOENT from try_to_grab_pending() becomes %true */
1559 return ret;
1560}
1561EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1562
1563/**
Tejun Heoc8e55f32010-06-29 10:07:12 +02001564 * worker_enter_idle - enter idle state
1565 * @worker: worker which is entering idle state
1566 *
1567 * @worker is entering idle state. Update stats and idle timer if
1568 * necessary.
1569 *
1570 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001571 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001572 */
1573static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001575 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576
Tejun Heo6183c002013-03-12 11:29:57 -07001577 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1578 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1579 (worker->hentry.next || worker->hentry.pprev)))
1580 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581
Tejun Heocb444762010-07-02 10:03:50 +02001582 /* can't use worker_set_flags(), also called from start_worker() */
1583 worker->flags |= WORKER_IDLE;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001584 pool->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001585 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001586
Tejun Heoc8e55f32010-06-29 10:07:12 +02001587 /* idle_list is LIFO */
Tejun Heobd7bdd42012-07-12 14:46:37 -07001588 list_add(&worker->entry, &pool->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001589
Tejun Heo628c78e2012-07-17 12:39:27 -07001590 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1591 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
Tejun Heocb444762010-07-02 10:03:50 +02001592
Tejun Heo544ecf32012-05-14 15:04:50 -07001593 /*
Tejun Heo706026c2013-01-24 11:01:34 -08001594 * Sanity check nr_running. Because wq_unbind_fn() releases
Tejun Heod565ed62013-01-24 11:01:33 -08001595 * pool->lock between setting %WORKER_UNBOUND and zapping
Tejun Heo628c78e2012-07-17 12:39:27 -07001596 * nr_running, the warning may trigger spuriously. Check iff
1597 * unbind is not in progress.
Tejun Heo544ecf32012-05-14 15:04:50 -07001598 */
Tejun Heo24647572013-01-24 11:01:33 -08001599 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -07001600 pool->nr_workers == pool->nr_idle &&
Tejun Heoe19e3972013-01-24 11:39:44 -08001601 atomic_read(&pool->nr_running));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602}
1603
Tejun Heoc8e55f32010-06-29 10:07:12 +02001604/**
1605 * worker_leave_idle - leave idle state
1606 * @worker: worker which is leaving idle state
1607 *
1608 * @worker is leaving idle state. Update stats.
1609 *
1610 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001611 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001612 */
1613static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001615 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616
Tejun Heo6183c002013-03-12 11:29:57 -07001617 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1618 return;
Tejun Heod302f012010-06-29 10:07:13 +02001619 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001620 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001621 list_del_init(&worker->entry);
1622}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
Tejun Heoe22bee72010-06-29 10:07:14 +02001624/**
Lai Jiangshanf36dc672013-02-19 12:17:02 -08001625 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1626 * @pool: target worker_pool
1627 *
1628 * Bind %current to the cpu of @pool if it is associated and lock @pool.
Tejun Heoe22bee72010-06-29 10:07:14 +02001629 *
1630 * Works which are scheduled while the cpu is online must at least be
1631 * scheduled to a worker which is bound to the cpu so that if they are
1632 * flushed from cpu callbacks while cpu is going down, they are
1633 * guaranteed to execute on the cpu.
1634 *
Lai Jiangshanf5faa072013-02-19 12:17:02 -08001635 * This function is to be used by unbound workers and rescuers to bind
Tejun Heoe22bee72010-06-29 10:07:14 +02001636 * themselves to the target cpu and may race with cpu going down or
1637 * coming online. kthread_bind() can't be used because it may put the
1638 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
Tejun Heo706026c2013-01-24 11:01:34 -08001639 * verbatim as it's best effort and blocking and pool may be
Tejun Heoe22bee72010-06-29 10:07:14 +02001640 * [dis]associated in the meantime.
1641 *
Tejun Heo706026c2013-01-24 11:01:34 -08001642 * This function tries set_cpus_allowed() and locks pool and verifies the
Tejun Heo24647572013-01-24 11:01:33 -08001643 * binding against %POOL_DISASSOCIATED which is set during
Tejun Heof2d5a0e2012-07-17 12:39:26 -07001644 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1645 * enters idle state or fetches works without dropping lock, it can
1646 * guarantee the scheduling requirement described in the first paragraph.
Tejun Heoe22bee72010-06-29 10:07:14 +02001647 *
1648 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001649 * Might sleep. Called without any lock but returns with pool->lock
Tejun Heoe22bee72010-06-29 10:07:14 +02001650 * held.
1651 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001652 * Return:
Tejun Heo706026c2013-01-24 11:01:34 -08001653 * %true if the associated pool is online (@worker is successfully
Tejun Heoe22bee72010-06-29 10:07:14 +02001654 * bound), %false if offline.
1655 */
Lai Jiangshanf36dc672013-02-19 12:17:02 -08001656static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08001657__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001658{
Tejun Heoe22bee72010-06-29 10:07:14 +02001659 while (true) {
1660 /*
1661 * The following call may fail, succeed or succeed
1662 * without actually migrating the task to the cpu if
1663 * it races with cpu hotunplug operation. Verify
Tejun Heo24647572013-01-24 11:01:33 -08001664 * against POOL_DISASSOCIATED.
Tejun Heoe22bee72010-06-29 10:07:14 +02001665 */
Tejun Heo24647572013-01-24 11:01:33 -08001666 if (!(pool->flags & POOL_DISASSOCIATED))
Tejun Heo7a4e3442013-03-12 11:30:00 -07001667 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
Oleg Nesterov85f41862007-05-09 02:34:20 -07001668
Tejun Heod565ed62013-01-24 11:01:33 -08001669 spin_lock_irq(&pool->lock);
Tejun Heo24647572013-01-24 11:01:33 -08001670 if (pool->flags & POOL_DISASSOCIATED)
Tejun Heoe22bee72010-06-29 10:07:14 +02001671 return false;
Lai Jiangshanf5faa072013-02-19 12:17:02 -08001672 if (task_cpu(current) == pool->cpu &&
Tejun Heo7a4e3442013-03-12 11:30:00 -07001673 cpumask_equal(&current->cpus_allowed, pool->attrs->cpumask))
Tejun Heoe22bee72010-06-29 10:07:14 +02001674 return true;
Tejun Heod565ed62013-01-24 11:01:33 -08001675 spin_unlock_irq(&pool->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07001676
Tejun Heo5035b202011-04-29 18:08:37 +02001677 /*
1678 * We've raced with CPU hot[un]plug. Give it a breather
1679 * and retry migration. cond_resched() is required here;
1680 * otherwise, we might deadlock against cpu_stop trying to
1681 * bring down the CPU on non-preemptive kernel.
1682 */
Tejun Heoe22bee72010-06-29 10:07:14 +02001683 cpu_relax();
Tejun Heo5035b202011-04-29 18:08:37 +02001684 cond_resched();
Tejun Heoe22bee72010-06-29 10:07:14 +02001685 }
1686}
1687
Tejun Heoc34056a2010-06-29 10:07:11 +02001688static struct worker *alloc_worker(void)
1689{
1690 struct worker *worker;
1691
1692 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001693 if (worker) {
1694 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001695 INIT_LIST_HEAD(&worker->scheduled);
Tejun Heoe22bee72010-06-29 10:07:14 +02001696 /* on creation a worker is in !idle && prep state */
1697 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001698 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001699 return worker;
1700}
1701
1702/**
1703 * create_worker - create a new workqueue worker
Tejun Heo63d95a92012-07-12 14:46:37 -07001704 * @pool: pool the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001705 *
Tejun Heo63d95a92012-07-12 14:46:37 -07001706 * Create a new worker which is bound to @pool. The returned worker
Tejun Heoc34056a2010-06-29 10:07:11 +02001707 * can be started by calling start_worker() or destroyed using
1708 * destroy_worker().
1709 *
1710 * CONTEXT:
1711 * Might sleep. Does GFP_KERNEL allocations.
1712 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001713 * Return:
Tejun Heoc34056a2010-06-29 10:07:11 +02001714 * Pointer to the newly created worker.
1715 */
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001716static struct worker *create_worker(struct worker_pool *pool)
Tejun Heoc34056a2010-06-29 10:07:11 +02001717{
Tejun Heoc34056a2010-06-29 10:07:11 +02001718 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001719 int id = -1;
Tejun Heoe3c916a2013-04-01 11:23:32 -07001720 char id_buf[16];
Tejun Heoc34056a2010-06-29 10:07:11 +02001721
Tejun Heocd549682013-03-13 19:47:39 -07001722 lockdep_assert_held(&pool->manager_mutex);
1723
Tejun Heo822d8402013-03-19 13:45:21 -07001724 /*
1725 * ID is needed to determine kthread name. Allocate ID first
1726 * without installing the pointer.
1727 */
1728 idr_preload(GFP_KERNEL);
Tejun Heod565ed62013-01-24 11:01:33 -08001729 spin_lock_irq(&pool->lock);
Tejun Heo822d8402013-03-19 13:45:21 -07001730
1731 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1732
Tejun Heod565ed62013-01-24 11:01:33 -08001733 spin_unlock_irq(&pool->lock);
Tejun Heo822d8402013-03-19 13:45:21 -07001734 idr_preload_end();
1735 if (id < 0)
1736 goto fail;
Tejun Heoc34056a2010-06-29 10:07:11 +02001737
1738 worker = alloc_worker();
1739 if (!worker)
1740 goto fail;
1741
Tejun Heobd7bdd42012-07-12 14:46:37 -07001742 worker->pool = pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001743 worker->id = id;
1744
Tejun Heo29c91e92013-03-12 11:30:03 -07001745 if (pool->cpu >= 0)
Tejun Heoe3c916a2013-04-01 11:23:32 -07001746 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1747 pool->attrs->nice < 0 ? "H" : "");
Tejun Heof3421792010-07-02 10:03:51 +02001748 else
Tejun Heoe3c916a2013-04-01 11:23:32 -07001749 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1750
Tejun Heof3f90ad2013-04-01 11:23:34 -07001751 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
Tejun Heoe3c916a2013-04-01 11:23:32 -07001752 "kworker/%s", id_buf);
Tejun Heoc34056a2010-06-29 10:07:11 +02001753 if (IS_ERR(worker->task))
1754 goto fail;
1755
Oleg Nesterov91151222013-11-14 12:56:18 +01001756 set_user_nice(worker->task, pool->attrs->nice);
1757
1758 /* prevent userland from meddling with cpumask of workqueue workers */
1759 worker->task->flags |= PF_NO_SETAFFINITY;
1760
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001761 /*
1762 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1763 * online CPUs. It'll be re-applied when any of the CPUs come up.
1764 */
Tejun Heo7a4e3442013-03-12 11:30:00 -07001765 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
Tejun Heo32704762012-07-13 22:16:45 -07001766
Tejun Heo7a4e3442013-03-12 11:30:00 -07001767 /*
1768 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1769 * remains stable across this function. See the comments above the
1770 * flag definition for details.
1771 */
1772 if (pool->flags & POOL_DISASSOCIATED)
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001773 worker->flags |= WORKER_UNBOUND;
Oleg Nesterov3af244332007-05-09 02:34:09 -07001774
Tejun Heo822d8402013-03-19 13:45:21 -07001775 /* successful, commit the pointer to idr */
1776 spin_lock_irq(&pool->lock);
1777 idr_replace(&pool->worker_idr, worker, worker->id);
1778 spin_unlock_irq(&pool->lock);
1779
Tejun Heoc34056a2010-06-29 10:07:11 +02001780 return worker;
Tejun Heo822d8402013-03-19 13:45:21 -07001781
Tejun Heoc34056a2010-06-29 10:07:11 +02001782fail:
1783 if (id >= 0) {
Tejun Heod565ed62013-01-24 11:01:33 -08001784 spin_lock_irq(&pool->lock);
Tejun Heo822d8402013-03-19 13:45:21 -07001785 idr_remove(&pool->worker_idr, id);
Tejun Heod565ed62013-01-24 11:01:33 -08001786 spin_unlock_irq(&pool->lock);
Tejun Heoc34056a2010-06-29 10:07:11 +02001787 }
1788 kfree(worker);
1789 return NULL;
1790}
1791
1792/**
1793 * start_worker - start a newly created worker
1794 * @worker: worker to start
1795 *
Tejun Heo706026c2013-01-24 11:01:34 -08001796 * Make the pool aware of @worker and start it.
Tejun Heoc34056a2010-06-29 10:07:11 +02001797 *
1798 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001799 * spin_lock_irq(pool->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001800 */
1801static void start_worker(struct worker *worker)
1802{
Tejun Heocb444762010-07-02 10:03:50 +02001803 worker->flags |= WORKER_STARTED;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001804 worker->pool->nr_workers++;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001805 worker_enter_idle(worker);
Tejun Heoc34056a2010-06-29 10:07:11 +02001806 wake_up_process(worker->task);
1807}
1808
1809/**
Tejun Heoebf44d12013-03-13 19:47:39 -07001810 * create_and_start_worker - create and start a worker for a pool
1811 * @pool: the target pool
1812 *
Tejun Heocd549682013-03-13 19:47:39 -07001813 * Grab the managership of @pool and create and start a new worker for it.
Yacine Belkadid185af32013-07-31 14:59:24 -07001814 *
1815 * Return: 0 on success. A negative error code otherwise.
Tejun Heoebf44d12013-03-13 19:47:39 -07001816 */
1817static int create_and_start_worker(struct worker_pool *pool)
1818{
1819 struct worker *worker;
1820
Tejun Heocd549682013-03-13 19:47:39 -07001821 mutex_lock(&pool->manager_mutex);
1822
Tejun Heoebf44d12013-03-13 19:47:39 -07001823 worker = create_worker(pool);
1824 if (worker) {
1825 spin_lock_irq(&pool->lock);
1826 start_worker(worker);
1827 spin_unlock_irq(&pool->lock);
1828 }
1829
Tejun Heocd549682013-03-13 19:47:39 -07001830 mutex_unlock(&pool->manager_mutex);
1831
Tejun Heoebf44d12013-03-13 19:47:39 -07001832 return worker ? 0 : -ENOMEM;
1833}
1834
1835/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001836 * destroy_worker - destroy a workqueue worker
1837 * @worker: worker to be destroyed
1838 *
Tejun Heo706026c2013-01-24 11:01:34 -08001839 * Destroy @worker and adjust @pool stats accordingly.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001840 *
1841 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001842 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoc34056a2010-06-29 10:07:11 +02001843 */
1844static void destroy_worker(struct worker *worker)
1845{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001846 struct worker_pool *pool = worker->pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001847
Tejun Heocd549682013-03-13 19:47:39 -07001848 lockdep_assert_held(&pool->manager_mutex);
1849 lockdep_assert_held(&pool->lock);
1850
Tejun Heoc34056a2010-06-29 10:07:11 +02001851 /* sanity check frenzy */
Tejun Heo6183c002013-03-12 11:29:57 -07001852 if (WARN_ON(worker->current_work) ||
1853 WARN_ON(!list_empty(&worker->scheduled)))
1854 return;
Tejun Heoc34056a2010-06-29 10:07:11 +02001855
Tejun Heoc8e55f32010-06-29 10:07:12 +02001856 if (worker->flags & WORKER_STARTED)
Tejun Heobd7bdd42012-07-12 14:46:37 -07001857 pool->nr_workers--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001858 if (worker->flags & WORKER_IDLE)
Tejun Heobd7bdd42012-07-12 14:46:37 -07001859 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001860
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001861 /*
1862 * Once WORKER_DIE is set, the kworker may destroy itself at any
1863 * point. Pin to ensure the task stays until we're done with it.
1864 */
1865 get_task_struct(worker->task);
1866
Tejun Heoc8e55f32010-06-29 10:07:12 +02001867 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001868 worker->flags |= WORKER_DIE;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001869
Tejun Heo822d8402013-03-19 13:45:21 -07001870 idr_remove(&pool->worker_idr, worker->id);
1871
Tejun Heod565ed62013-01-24 11:01:33 -08001872 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001873
Tejun Heoc34056a2010-06-29 10:07:11 +02001874 kthread_stop(worker->task);
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001875 put_task_struct(worker->task);
Tejun Heoc34056a2010-06-29 10:07:11 +02001876 kfree(worker);
1877
Tejun Heod565ed62013-01-24 11:01:33 -08001878 spin_lock_irq(&pool->lock);
Tejun Heoc34056a2010-06-29 10:07:11 +02001879}
1880
Tejun Heo63d95a92012-07-12 14:46:37 -07001881static void idle_worker_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001882{
Tejun Heo63d95a92012-07-12 14:46:37 -07001883 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001884
Tejun Heod565ed62013-01-24 11:01:33 -08001885 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001886
Tejun Heo63d95a92012-07-12 14:46:37 -07001887 if (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001888 struct worker *worker;
1889 unsigned long expires;
1890
1891 /* idle_list is kept in LIFO order, check the last one */
Tejun Heo63d95a92012-07-12 14:46:37 -07001892 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02001893 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1894
1895 if (time_before(jiffies, expires))
Tejun Heo63d95a92012-07-12 14:46:37 -07001896 mod_timer(&pool->idle_timer, expires);
Tejun Heoe22bee72010-06-29 10:07:14 +02001897 else {
1898 /* it's been idle for too long, wake up manager */
Tejun Heo11ebea52012-07-12 14:46:37 -07001899 pool->flags |= POOL_MANAGE_WORKERS;
Tejun Heo63d95a92012-07-12 14:46:37 -07001900 wake_up_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02001901 }
1902 }
1903
Tejun Heod565ed62013-01-24 11:01:33 -08001904 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001905}
1906
Tejun Heo493a1722013-03-12 11:29:59 -07001907static void send_mayday(struct work_struct *work)
Tejun Heoe22bee72010-06-29 10:07:14 +02001908{
Tejun Heo112202d2013-02-13 19:29:12 -08001909 struct pool_workqueue *pwq = get_work_pwq(work);
1910 struct workqueue_struct *wq = pwq->wq;
Tejun Heo493a1722013-03-12 11:29:59 -07001911
Tejun Heo2e109a22013-03-13 19:47:40 -07001912 lockdep_assert_held(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001913
Tejun Heo493008a2013-03-12 11:30:03 -07001914 if (!wq->rescuer)
Tejun Heo493a1722013-03-12 11:29:59 -07001915 return;
Tejun Heoe22bee72010-06-29 10:07:14 +02001916
1917 /* mayday mayday mayday */
Tejun Heo493a1722013-03-12 11:29:59 -07001918 if (list_empty(&pwq->mayday_node)) {
1919 list_add_tail(&pwq->mayday_node, &wq->maydays);
Tejun Heoe22bee72010-06-29 10:07:14 +02001920 wake_up_process(wq->rescuer->task);
Tejun Heo493a1722013-03-12 11:29:59 -07001921 }
Tejun Heoe22bee72010-06-29 10:07:14 +02001922}
1923
Tejun Heo706026c2013-01-24 11:01:34 -08001924static void pool_mayday_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001925{
Tejun Heo63d95a92012-07-12 14:46:37 -07001926 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001927 struct work_struct *work;
1928
Tejun Heo2e109a22013-03-13 19:47:40 -07001929 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
Tejun Heo493a1722013-03-12 11:29:59 -07001930 spin_lock(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001931
Tejun Heo63d95a92012-07-12 14:46:37 -07001932 if (need_to_create_worker(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001933 /*
1934 * We've been trying to create a new worker but
1935 * haven't been successful. We might be hitting an
1936 * allocation deadlock. Send distress signals to
1937 * rescuers.
1938 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001939 list_for_each_entry(work, &pool->worklist, entry)
Tejun Heoe22bee72010-06-29 10:07:14 +02001940 send_mayday(work);
1941 }
1942
Tejun Heo493a1722013-03-12 11:29:59 -07001943 spin_unlock(&pool->lock);
Tejun Heo2e109a22013-03-13 19:47:40 -07001944 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001945
Tejun Heo63d95a92012-07-12 14:46:37 -07001946 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
Tejun Heoe22bee72010-06-29 10:07:14 +02001947}
1948
1949/**
1950 * maybe_create_worker - create a new worker if necessary
Tejun Heo63d95a92012-07-12 14:46:37 -07001951 * @pool: pool to create a new worker for
Tejun Heoe22bee72010-06-29 10:07:14 +02001952 *
Tejun Heo63d95a92012-07-12 14:46:37 -07001953 * Create a new worker for @pool if necessary. @pool is guaranteed to
Tejun Heoe22bee72010-06-29 10:07:14 +02001954 * have at least one idle worker on return from this function. If
1955 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
Tejun Heo63d95a92012-07-12 14:46:37 -07001956 * sent to all rescuers with works scheduled on @pool to resolve
Tejun Heoe22bee72010-06-29 10:07:14 +02001957 * possible allocation deadlock.
1958 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001959 * On return, need_to_create_worker() is guaranteed to be %false and
1960 * may_start_working() %true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001961 *
1962 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001963 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001964 * multiple times. Does GFP_KERNEL allocations. Called only from
1965 * manager.
1966 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001967 * Return:
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001968 * %false if no action was taken and pool->lock stayed locked, %true
Tejun Heoe22bee72010-06-29 10:07:14 +02001969 * otherwise.
1970 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001971static bool maybe_create_worker(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08001972__releases(&pool->lock)
1973__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001974{
Tejun Heo63d95a92012-07-12 14:46:37 -07001975 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001976 return false;
1977restart:
Tejun Heod565ed62013-01-24 11:01:33 -08001978 spin_unlock_irq(&pool->lock);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001979
Tejun Heoe22bee72010-06-29 10:07:14 +02001980 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
Tejun Heo63d95a92012-07-12 14:46:37 -07001981 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
Tejun Heoe22bee72010-06-29 10:07:14 +02001982
1983 while (true) {
1984 struct worker *worker;
1985
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001986 worker = create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02001987 if (worker) {
Tejun Heo63d95a92012-07-12 14:46:37 -07001988 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08001989 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001990 start_worker(worker);
Tejun Heo6183c002013-03-12 11:29:57 -07001991 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1992 goto restart;
Tejun Heoe22bee72010-06-29 10:07:14 +02001993 return true;
1994 }
1995
Tejun Heo63d95a92012-07-12 14:46:37 -07001996 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001997 break;
1998
Tejun Heoe22bee72010-06-29 10:07:14 +02001999 __set_current_state(TASK_INTERRUPTIBLE);
2000 schedule_timeout(CREATE_COOLDOWN);
Tejun Heo9f9c2362010-07-14 11:31:20 +02002001
Tejun Heo63d95a92012-07-12 14:46:37 -07002002 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002003 break;
2004 }
2005
Tejun Heo63d95a92012-07-12 14:46:37 -07002006 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08002007 spin_lock_irq(&pool->lock);
Tejun Heo63d95a92012-07-12 14:46:37 -07002008 if (need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002009 goto restart;
2010 return true;
2011}
2012
2013/**
2014 * maybe_destroy_worker - destroy workers which have been idle for a while
Tejun Heo63d95a92012-07-12 14:46:37 -07002015 * @pool: pool to destroy workers for
Tejun Heoe22bee72010-06-29 10:07:14 +02002016 *
Tejun Heo63d95a92012-07-12 14:46:37 -07002017 * Destroy @pool workers which have been idle for longer than
Tejun Heoe22bee72010-06-29 10:07:14 +02002018 * IDLE_WORKER_TIMEOUT.
2019 *
2020 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08002021 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02002022 * multiple times. Called only from manager.
2023 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002024 * Return:
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002025 * %false if no action was taken and pool->lock stayed locked, %true
Tejun Heoe22bee72010-06-29 10:07:14 +02002026 * otherwise.
2027 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002028static bool maybe_destroy_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02002029{
2030 bool ret = false;
2031
Tejun Heo63d95a92012-07-12 14:46:37 -07002032 while (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02002033 struct worker *worker;
2034 unsigned long expires;
2035
Tejun Heo63d95a92012-07-12 14:46:37 -07002036 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02002037 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2038
2039 if (time_before(jiffies, expires)) {
Tejun Heo63d95a92012-07-12 14:46:37 -07002040 mod_timer(&pool->idle_timer, expires);
Tejun Heoe22bee72010-06-29 10:07:14 +02002041 break;
2042 }
2043
2044 destroy_worker(worker);
2045 ret = true;
2046 }
2047
2048 return ret;
2049}
2050
2051/**
2052 * manage_workers - manage worker pool
2053 * @worker: self
2054 *
Tejun Heo706026c2013-01-24 11:01:34 -08002055 * Assume the manager role and manage the worker pool @worker belongs
Tejun Heoe22bee72010-06-29 10:07:14 +02002056 * to. At any given time, there can be only zero or one manager per
Tejun Heo706026c2013-01-24 11:01:34 -08002057 * pool. The exclusion is handled automatically by this function.
Tejun Heoe22bee72010-06-29 10:07:14 +02002058 *
2059 * The caller can safely start processing works on false return. On
2060 * true return, it's guaranteed that need_to_create_worker() is false
2061 * and may_start_working() is true.
2062 *
2063 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002064 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02002065 * multiple times. Does GFP_KERNEL allocations.
2066 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002067 * Return:
Libin2d498db2013-08-21 08:50:40 +08002068 * %false if the pool don't need management and the caller can safely start
2069 * processing works, %true indicates that the function released pool->lock
2070 * and reacquired it to perform some management function and that the
2071 * conditions that the caller verified while holding the lock before
2072 * calling the function might no longer be true.
Tejun Heoe22bee72010-06-29 10:07:14 +02002073 */
2074static bool manage_workers(struct worker *worker)
2075{
Tejun Heo63d95a92012-07-12 14:46:37 -07002076 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002077 bool ret = false;
2078
Tejun Heobc3a1af2013-03-13 19:47:39 -07002079 /*
2080 * Managership is governed by two mutexes - manager_arb and
2081 * manager_mutex. manager_arb handles arbitration of manager role.
2082 * Anyone who successfully grabs manager_arb wins the arbitration
2083 * and becomes the manager. mutex_trylock() on pool->manager_arb
2084 * failure while holding pool->lock reliably indicates that someone
2085 * else is managing the pool and the worker which failed trylock
2086 * can proceed to executing work items. This means that anyone
2087 * grabbing manager_arb is responsible for actually performing
2088 * manager duties. If manager_arb is grabbed and released without
2089 * actual management, the pool may stall indefinitely.
2090 *
2091 * manager_mutex is used for exclusion of actual management
2092 * operations. The holder of manager_mutex can be sure that none
2093 * of management operations, including creation and destruction of
2094 * workers, won't take place until the mutex is released. Because
2095 * manager_mutex doesn't interfere with manager role arbitration,
2096 * it is guaranteed that the pool's management, while may be
2097 * delayed, won't be disturbed by someone else grabbing
2098 * manager_mutex.
2099 */
Tejun Heo34a06bd2013-03-12 11:30:00 -07002100 if (!mutex_trylock(&pool->manager_arb))
Tejun Heoe22bee72010-06-29 10:07:14 +02002101 return ret;
2102
Lai Jiangshanee378aa2012-09-10 10:03:44 -07002103 /*
Tejun Heobc3a1af2013-03-13 19:47:39 -07002104 * With manager arbitration won, manager_mutex would be free in
2105 * most cases. trylock first without dropping @pool->lock.
Lai Jiangshanee378aa2012-09-10 10:03:44 -07002106 */
Tejun Heobc3a1af2013-03-13 19:47:39 -07002107 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
Tejun Heod565ed62013-01-24 11:01:33 -08002108 spin_unlock_irq(&pool->lock);
Tejun Heobc3a1af2013-03-13 19:47:39 -07002109 mutex_lock(&pool->manager_mutex);
Joonsoo Kim8f174b12013-05-01 00:07:00 +09002110 spin_lock_irq(&pool->lock);
Lai Jiangshanee378aa2012-09-10 10:03:44 -07002111 ret = true;
2112 }
2113
Tejun Heo11ebea52012-07-12 14:46:37 -07002114 pool->flags &= ~POOL_MANAGE_WORKERS;
Tejun Heoe22bee72010-06-29 10:07:14 +02002115
2116 /*
2117 * Destroy and then create so that may_start_working() is true
2118 * on return.
2119 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002120 ret |= maybe_destroy_workers(pool);
2121 ret |= maybe_create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02002122
Tejun Heobc3a1af2013-03-13 19:47:39 -07002123 mutex_unlock(&pool->manager_mutex);
Tejun Heo34a06bd2013-03-12 11:30:00 -07002124 mutex_unlock(&pool->manager_arb);
Tejun Heoe22bee72010-06-29 10:07:14 +02002125 return ret;
2126}
2127
Tejun Heoa62428c2010-06-29 10:07:10 +02002128/**
2129 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02002130 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02002131 * @work: work to process
2132 *
2133 * Process @work. This function contains all the logics necessary to
2134 * process a single work including synchronization against and
2135 * interaction with other workers on the same cpu, queueing and
2136 * flushing. As long as context requirement is met, any worker can
2137 * call this function to process a work.
2138 *
2139 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002140 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02002141 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002142static void process_one_work(struct worker *worker, struct work_struct *work)
Tejun Heod565ed62013-01-24 11:01:33 -08002143__releases(&pool->lock)
2144__acquires(&pool->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02002145{
Tejun Heo112202d2013-02-13 19:29:12 -08002146 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobd7bdd42012-07-12 14:46:37 -07002147 struct worker_pool *pool = worker->pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002148 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heo73f53c42010-06-29 10:07:11 +02002149 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02002150 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02002151#ifdef CONFIG_LOCKDEP
2152 /*
2153 * It is permissible to free the struct work_struct from
2154 * inside the function that is called from it, this we need to
2155 * take into account for lockdep too. To avoid bogus "held
2156 * lock freed" warnings as well as problems when looking into
2157 * work->lockdep_map, make a copy and use that here.
2158 */
Peter Zijlstra4d82a1d2012-05-15 08:06:19 -07002159 struct lockdep_map lockdep_map;
2160
2161 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002162#endif
Tejun Heo6fec10a2012-07-22 10:16:34 -07002163 /*
2164 * Ensure we're on the correct CPU. DISASSOCIATED test is
2165 * necessary to avoid spurious warnings from rescuers servicing the
Tejun Heo24647572013-01-24 11:01:33 -08002166 * unbound or a disassociated pool.
Tejun Heo6fec10a2012-07-22 10:16:34 -07002167 */
Lai Jiangshan5f7dabf2012-09-18 09:59:23 -07002168 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
Tejun Heo24647572013-01-24 11:01:33 -08002169 !(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heoec22ca52013-01-24 11:01:33 -08002170 raw_smp_processor_id() != pool->cpu);
Tejun Heo25511a42012-07-17 12:39:27 -07002171
Tejun Heo7e116292010-06-29 10:07:13 +02002172 /*
2173 * A single work shouldn't be executed concurrently by
2174 * multiple workers on a single cpu. Check whether anyone is
2175 * already processing the work. If so, defer the work to the
2176 * currently executing one.
2177 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002178 collision = find_worker_executing_work(pool, work);
Tejun Heo7e116292010-06-29 10:07:13 +02002179 if (unlikely(collision)) {
2180 move_linked_works(work, &collision->scheduled, NULL);
2181 return;
2182 }
2183
Tejun Heo8930cab2012-08-03 10:30:45 -07002184 /* claim and dequeue */
Tejun Heoa62428c2010-06-29 10:07:10 +02002185 debug_work_deactivate(work);
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002186 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
Tejun Heoc34056a2010-06-29 10:07:11 +02002187 worker->current_work = work;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002188 worker->current_func = work->func;
Tejun Heo112202d2013-02-13 19:29:12 -08002189 worker->current_pwq = pwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002190 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002191
Tejun Heoa62428c2010-06-29 10:07:10 +02002192 list_del_init(&work->entry);
2193
Tejun Heo649027d2010-06-29 10:07:14 +02002194 /*
Tejun Heofb0e7be2010-06-29 10:07:15 +02002195 * CPU intensive works don't participate in concurrency
2196 * management. They're the scheduler's responsibility.
2197 */
2198 if (unlikely(cpu_intensive))
2199 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2200
Tejun Heo974271c2012-07-12 14:46:37 -07002201 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002202 * Unbound pool isn't concurrency managed and work items should be
Tejun Heo974271c2012-07-12 14:46:37 -07002203 * executed ASAP. Wake up another worker if necessary.
2204 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002205 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2206 wake_up_worker(pool);
Tejun Heo974271c2012-07-12 14:46:37 -07002207
Tejun Heo8930cab2012-08-03 10:30:45 -07002208 /*
Tejun Heo7c3eed52013-01-24 11:01:33 -08002209 * Record the last pool and clear PENDING which should be the last
Tejun Heod565ed62013-01-24 11:01:33 -08002210 * update to @work. Also, do this inside @pool->lock so that
Tejun Heo23657bb2012-08-13 17:08:19 -07002211 * PENDING and queued state changes happen together while IRQ is
2212 * disabled.
Tejun Heo8930cab2012-08-03 10:30:45 -07002213 */
Tejun Heo7c3eed52013-01-24 11:01:33 -08002214 set_work_pool_and_clear_pending(work, pool->id);
Tejun Heoa62428c2010-06-29 10:07:10 +02002215
Tejun Heod565ed62013-01-24 11:01:33 -08002216 spin_unlock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002217
Tejun Heo112202d2013-02-13 19:29:12 -08002218 lock_map_acquire_read(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002219 lock_map_acquire(&lockdep_map);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002220 trace_workqueue_execute_start(work);
Tejun Heoa2c1c572012-12-18 10:35:02 -08002221 worker->current_func(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002222 /*
2223 * While we must be careful to not use "work" after this, the trace
2224 * point will only record its address.
2225 */
2226 trace_workqueue_execute_end(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02002227 lock_map_release(&lockdep_map);
Tejun Heo112202d2013-02-13 19:29:12 -08002228 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002229
2230 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
Valentin Ilie044c7822012-08-19 00:52:42 +03002231 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2232 " last function: %pf\n",
Tejun Heoa2c1c572012-12-18 10:35:02 -08002233 current->comm, preempt_count(), task_pid_nr(current),
2234 worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002235 debug_show_held_locks(current);
2236 dump_stack();
2237 }
2238
Tejun Heob22ce272013-08-28 17:33:37 -04002239 /*
2240 * The following prevents a kworker from hogging CPU on !PREEMPT
2241 * kernels, where a requeueing work item waiting for something to
2242 * happen could deadlock with stop_machine as such work item could
2243 * indefinitely requeue itself while all other CPUs are trapped in
2244 * stop_machine.
2245 */
2246 cond_resched();
2247
Tejun Heod565ed62013-01-24 11:01:33 -08002248 spin_lock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002249
Tejun Heofb0e7be2010-06-29 10:07:15 +02002250 /* clear cpu intensive status */
2251 if (unlikely(cpu_intensive))
2252 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2253
Tejun Heoa62428c2010-06-29 10:07:10 +02002254 /* we're done with it, release */
Sasha Levin42f85702012-12-17 10:01:23 -05002255 hash_del(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002256 worker->current_work = NULL;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002257 worker->current_func = NULL;
Tejun Heo112202d2013-02-13 19:29:12 -08002258 worker->current_pwq = NULL;
Tejun Heo3d1cb202013-04-30 15:27:22 -07002259 worker->desc_valid = false;
Tejun Heo112202d2013-02-13 19:29:12 -08002260 pwq_dec_nr_in_flight(pwq, work_color);
Tejun Heoa62428c2010-06-29 10:07:10 +02002261}
2262
Tejun Heoaffee4b2010-06-29 10:07:12 +02002263/**
2264 * process_scheduled_works - process scheduled works
2265 * @worker: self
2266 *
2267 * Process all scheduled works. Please note that the scheduled list
2268 * may change while processing a work, so this function repeatedly
2269 * fetches a work from the top and executes it.
2270 *
2271 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002272 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02002273 * multiple times.
2274 */
2275static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002277 while (!list_empty(&worker->scheduled)) {
2278 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002280 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282}
2283
Tejun Heo4690c4a2010-06-29 10:07:10 +02002284/**
2285 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02002286 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02002287 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002288 * The worker thread function. All workers belong to a worker_pool -
2289 * either a per-cpu one or dynamic unbound one. These workers process all
2290 * work items regardless of their specific target workqueue. The only
2291 * exception is work items which belong to workqueues with a rescuer which
2292 * will be explained in rescuer_thread().
Yacine Belkadid185af32013-07-31 14:59:24 -07002293 *
2294 * Return: 0
Tejun Heo4690c4a2010-06-29 10:07:10 +02002295 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002296static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297{
Tejun Heoc34056a2010-06-29 10:07:11 +02002298 struct worker *worker = __worker;
Tejun Heobd7bdd42012-07-12 14:46:37 -07002299 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300
Tejun Heoe22bee72010-06-29 10:07:14 +02002301 /* tell the scheduler that this is a workqueue worker */
2302 worker->task->flags |= PF_WQ_WORKER;
Tejun Heoc8e55f32010-06-29 10:07:12 +02002303woke_up:
Tejun Heod565ed62013-01-24 11:01:33 -08002304 spin_lock_irq(&pool->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305
Tejun Heoa9ab7752013-03-19 13:45:21 -07002306 /* am I supposed to die? */
2307 if (unlikely(worker->flags & WORKER_DIE)) {
Tejun Heod565ed62013-01-24 11:01:33 -08002308 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002309 WARN_ON_ONCE(!list_empty(&worker->entry));
2310 worker->task->flags &= ~PF_WQ_WORKER;
2311 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 }
2313
Tejun Heoc8e55f32010-06-29 10:07:12 +02002314 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02002315recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02002316 /* no more worker necessary? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002317 if (!need_more_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002318 goto sleep;
2319
2320 /* do we need to manage? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002321 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002322 goto recheck;
2323
Tejun Heoc8e55f32010-06-29 10:07:12 +02002324 /*
2325 * ->scheduled list can only be filled while a worker is
2326 * preparing to process a work or actually processing it.
2327 * Make sure nobody diddled with it while I was sleeping.
2328 */
Tejun Heo6183c002013-03-12 11:29:57 -07002329 WARN_ON_ONCE(!list_empty(&worker->scheduled));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002330
Tejun Heoe22bee72010-06-29 10:07:14 +02002331 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07002332 * Finish PREP stage. We're guaranteed to have at least one idle
2333 * worker or that someone else has already assumed the manager
2334 * role. This is where @worker starts participating in concurrency
2335 * management if applicable and concurrency management is restored
2336 * after being rebound. See rebind_workers() for details.
Tejun Heoe22bee72010-06-29 10:07:14 +02002337 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07002338 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
Tejun Heoe22bee72010-06-29 10:07:14 +02002339
2340 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02002341 struct work_struct *work =
Tejun Heobd7bdd42012-07-12 14:46:37 -07002342 list_first_entry(&pool->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02002343 struct work_struct, entry);
2344
2345 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2346 /* optimization path, not strictly necessary */
2347 process_one_work(worker, work);
2348 if (unlikely(!list_empty(&worker->scheduled)))
2349 process_scheduled_works(worker);
2350 } else {
2351 move_linked_works(work, &worker->scheduled, NULL);
2352 process_scheduled_works(worker);
2353 }
Tejun Heo63d95a92012-07-12 14:46:37 -07002354 } while (keep_working(pool));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002355
Tejun Heoe22bee72010-06-29 10:07:14 +02002356 worker_set_flags(worker, WORKER_PREP, false);
Tejun Heod313dd82010-07-02 10:03:51 +02002357sleep:
Tejun Heo63d95a92012-07-12 14:46:37 -07002358 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002359 goto recheck;
Tejun Heod313dd82010-07-02 10:03:51 +02002360
Tejun Heoc8e55f32010-06-29 10:07:12 +02002361 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002362 * pool->lock is held and there's no work to process and no need to
2363 * manage, sleep. Workers are woken up only while holding
2364 * pool->lock or from local cpu, so setting the current state
2365 * before releasing pool->lock is enough to prevent losing any
2366 * event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02002367 */
2368 worker_enter_idle(worker);
2369 __set_current_state(TASK_INTERRUPTIBLE);
Tejun Heod565ed62013-01-24 11:01:33 -08002370 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002371 schedule();
2372 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373}
2374
Tejun Heoe22bee72010-06-29 10:07:14 +02002375/**
2376 * rescuer_thread - the rescuer thread function
Tejun Heo111c2252013-01-17 17:16:24 -08002377 * @__rescuer: self
Tejun Heoe22bee72010-06-29 10:07:14 +02002378 *
2379 * Workqueue rescuer thread function. There's one rescuer for each
Tejun Heo493008a2013-03-12 11:30:03 -07002380 * workqueue which has WQ_MEM_RECLAIM set.
Tejun Heoe22bee72010-06-29 10:07:14 +02002381 *
Tejun Heo706026c2013-01-24 11:01:34 -08002382 * Regular work processing on a pool may block trying to create a new
Tejun Heoe22bee72010-06-29 10:07:14 +02002383 * worker which uses GFP_KERNEL allocation which has slight chance of
2384 * developing into deadlock if some works currently on the same queue
2385 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2386 * the problem rescuer solves.
2387 *
Tejun Heo706026c2013-01-24 11:01:34 -08002388 * When such condition is possible, the pool summons rescuers of all
2389 * workqueues which have works queued on the pool and let them process
Tejun Heoe22bee72010-06-29 10:07:14 +02002390 * those works so that forward progress can be guaranteed.
2391 *
2392 * This should happen rarely.
Yacine Belkadid185af32013-07-31 14:59:24 -07002393 *
2394 * Return: 0
Tejun Heoe22bee72010-06-29 10:07:14 +02002395 */
Tejun Heo111c2252013-01-17 17:16:24 -08002396static int rescuer_thread(void *__rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02002397{
Tejun Heo111c2252013-01-17 17:16:24 -08002398 struct worker *rescuer = __rescuer;
2399 struct workqueue_struct *wq = rescuer->rescue_wq;
Tejun Heoe22bee72010-06-29 10:07:14 +02002400 struct list_head *scheduled = &rescuer->scheduled;
Tejun Heoe22bee72010-06-29 10:07:14 +02002401
2402 set_user_nice(current, RESCUER_NICE_LEVEL);
Tejun Heo111c2252013-01-17 17:16:24 -08002403
2404 /*
2405 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2406 * doesn't participate in concurrency management.
2407 */
2408 rescuer->task->flags |= PF_WQ_WORKER;
Tejun Heoe22bee72010-06-29 10:07:14 +02002409repeat:
2410 set_current_state(TASK_INTERRUPTIBLE);
2411
Mike Galbraith412d32e2012-11-28 07:17:18 +01002412 if (kthread_should_stop()) {
2413 __set_current_state(TASK_RUNNING);
Tejun Heo111c2252013-01-17 17:16:24 -08002414 rescuer->task->flags &= ~PF_WQ_WORKER;
Tejun Heoe22bee72010-06-29 10:07:14 +02002415 return 0;
Mike Galbraith412d32e2012-11-28 07:17:18 +01002416 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002417
Tejun Heo493a1722013-03-12 11:29:59 -07002418 /* see whether any pwq is asking for help */
Tejun Heo2e109a22013-03-13 19:47:40 -07002419 spin_lock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002420
2421 while (!list_empty(&wq->maydays)) {
2422 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2423 struct pool_workqueue, mayday_node);
Tejun Heo112202d2013-02-13 19:29:12 -08002424 struct worker_pool *pool = pwq->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002425 struct work_struct *work, *n;
2426
2427 __set_current_state(TASK_RUNNING);
Tejun Heo493a1722013-03-12 11:29:59 -07002428 list_del_init(&pwq->mayday_node);
2429
Tejun Heo2e109a22013-03-13 19:47:40 -07002430 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002431
2432 /* migrate to the target cpu if possible */
Lai Jiangshanf36dc672013-02-19 12:17:02 -08002433 worker_maybe_bind_and_lock(pool);
Lai Jiangshanb3104102013-02-19 12:17:02 -08002434 rescuer->pool = pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002435
2436 /*
2437 * Slurp in all works issued via this workqueue and
2438 * process'em.
2439 */
Tejun Heo6183c002013-03-12 11:29:57 -07002440 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
Tejun Heobd7bdd42012-07-12 14:46:37 -07002441 list_for_each_entry_safe(work, n, &pool->worklist, entry)
Tejun Heo112202d2013-02-13 19:29:12 -08002442 if (get_work_pwq(work) == pwq)
Tejun Heoe22bee72010-06-29 10:07:14 +02002443 move_linked_works(work, scheduled, &n);
2444
2445 process_scheduled_works(rescuer);
Tejun Heo75769582011-02-14 14:04:46 +01002446
2447 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002448 * Leave this pool. If keep_working() is %true, notify a
Tejun Heo75769582011-02-14 14:04:46 +01002449 * regular worker; otherwise, we end up with 0 concurrency
2450 * and stalling the execution.
2451 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002452 if (keep_working(pool))
2453 wake_up_worker(pool);
Tejun Heo75769582011-02-14 14:04:46 +01002454
Lai Jiangshanb3104102013-02-19 12:17:02 -08002455 rescuer->pool = NULL;
Tejun Heo493a1722013-03-12 11:29:59 -07002456 spin_unlock(&pool->lock);
Tejun Heo2e109a22013-03-13 19:47:40 -07002457 spin_lock(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002458 }
2459
Tejun Heo2e109a22013-03-13 19:47:40 -07002460 spin_unlock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002461
Tejun Heo111c2252013-01-17 17:16:24 -08002462 /* rescuers should never participate in concurrency management */
2463 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
Tejun Heoe22bee72010-06-29 10:07:14 +02002464 schedule();
2465 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466}
2467
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002468struct wq_barrier {
2469 struct work_struct work;
2470 struct completion done;
2471};
2472
2473static void wq_barrier_func(struct work_struct *work)
2474{
2475 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2476 complete(&barr->done);
2477}
2478
Tejun Heo4690c4a2010-06-29 10:07:10 +02002479/**
2480 * insert_wq_barrier - insert a barrier work
Tejun Heo112202d2013-02-13 19:29:12 -08002481 * @pwq: pwq to insert barrier into
Tejun Heo4690c4a2010-06-29 10:07:10 +02002482 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002483 * @target: target work to attach @barr to
2484 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002485 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002486 * @barr is linked to @target such that @barr is completed only after
2487 * @target finishes execution. Please note that the ordering
2488 * guarantee is observed only with respect to @target and on the local
2489 * cpu.
2490 *
2491 * Currently, a queued barrier can't be canceled. This is because
2492 * try_to_grab_pending() can't determine whether the work to be
2493 * grabbed is at the head of the queue and thus can't clear LINKED
2494 * flag of the previous work while there must be a valid next work
2495 * after a work with LINKED flag set.
2496 *
2497 * Note that when @worker is non-NULL, @target may be modified
Tejun Heo112202d2013-02-13 19:29:12 -08002498 * underneath us, so we can't reliably determine pwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002499 *
2500 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002501 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002502 */
Tejun Heo112202d2013-02-13 19:29:12 -08002503static void insert_wq_barrier(struct pool_workqueue *pwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002504 struct wq_barrier *barr,
2505 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002506{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002507 struct list_head *head;
2508 unsigned int linked = 0;
2509
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002510 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002511 * debugobject calls are safe here even with pool->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002512 * as we know for sure that this will not trigger any of the
2513 * checks and call back into the fixup functions where we
2514 * might deadlock.
2515 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002516 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002517 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002518 init_completion(&barr->done);
Oleg Nesterov83c22522007-05-09 02:33:54 -07002519
Tejun Heoaffee4b2010-06-29 10:07:12 +02002520 /*
2521 * If @target is currently being executed, schedule the
2522 * barrier to the worker; otherwise, put it after @target.
2523 */
2524 if (worker)
2525 head = worker->scheduled.next;
2526 else {
2527 unsigned long *bits = work_data_bits(target);
2528
2529 head = target->entry.next;
2530 /* there can already be other linked works, inherit and set */
2531 linked = *bits & WORK_STRUCT_LINKED;
2532 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2533 }
2534
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002535 debug_work_activate(&barr->work);
Tejun Heo112202d2013-02-13 19:29:12 -08002536 insert_work(pwq, &barr->work, head,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002537 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002538}
2539
Tejun Heo73f53c42010-06-29 10:07:11 +02002540/**
Tejun Heo112202d2013-02-13 19:29:12 -08002541 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
Tejun Heo73f53c42010-06-29 10:07:11 +02002542 * @wq: workqueue being flushed
2543 * @flush_color: new flush color, < 0 for no-op
2544 * @work_color: new work color, < 0 for no-op
2545 *
Tejun Heo112202d2013-02-13 19:29:12 -08002546 * Prepare pwqs for workqueue flushing.
Tejun Heo73f53c42010-06-29 10:07:11 +02002547 *
Tejun Heo112202d2013-02-13 19:29:12 -08002548 * If @flush_color is non-negative, flush_color on all pwqs should be
2549 * -1. If no pwq has in-flight commands at the specified color, all
2550 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2551 * has in flight commands, its pwq->flush_color is set to
2552 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
Tejun Heo73f53c42010-06-29 10:07:11 +02002553 * wakeup logic is armed and %true is returned.
2554 *
2555 * The caller should have initialized @wq->first_flusher prior to
2556 * calling this function with non-negative @flush_color. If
2557 * @flush_color is negative, no flush color update is done and %false
2558 * is returned.
2559 *
Tejun Heo112202d2013-02-13 19:29:12 -08002560 * If @work_color is non-negative, all pwqs should have the same
Tejun Heo73f53c42010-06-29 10:07:11 +02002561 * work_color which is previous to @work_color and all will be
2562 * advanced to @work_color.
2563 *
2564 * CONTEXT:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002565 * mutex_lock(wq->mutex).
Tejun Heo73f53c42010-06-29 10:07:11 +02002566 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002567 * Return:
Tejun Heo73f53c42010-06-29 10:07:11 +02002568 * %true if @flush_color >= 0 and there's something to flush. %false
2569 * otherwise.
2570 */
Tejun Heo112202d2013-02-13 19:29:12 -08002571static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
Tejun Heo73f53c42010-06-29 10:07:11 +02002572 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573{
Tejun Heo73f53c42010-06-29 10:07:11 +02002574 bool wait = false;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002575 struct pool_workqueue *pwq;
Oleg Nesterov14441962007-05-23 13:57:57 -07002576
Tejun Heo73f53c42010-06-29 10:07:11 +02002577 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002578 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
Tejun Heo112202d2013-02-13 19:29:12 -08002579 atomic_set(&wq->nr_pwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002580 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002581
Tejun Heo49e3cf42013-03-12 11:29:58 -07002582 for_each_pwq(pwq, wq) {
Tejun Heo112202d2013-02-13 19:29:12 -08002583 struct worker_pool *pool = pwq->pool;
Tejun Heo73f53c42010-06-29 10:07:11 +02002584
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002585 spin_lock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002586
2587 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002588 WARN_ON_ONCE(pwq->flush_color != -1);
Tejun Heo73f53c42010-06-29 10:07:11 +02002589
Tejun Heo112202d2013-02-13 19:29:12 -08002590 if (pwq->nr_in_flight[flush_color]) {
2591 pwq->flush_color = flush_color;
2592 atomic_inc(&wq->nr_pwqs_to_flush);
Tejun Heo73f53c42010-06-29 10:07:11 +02002593 wait = true;
2594 }
2595 }
2596
2597 if (work_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002598 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
Tejun Heo112202d2013-02-13 19:29:12 -08002599 pwq->work_color = work_color;
Tejun Heo73f53c42010-06-29 10:07:11 +02002600 }
2601
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002602 spin_unlock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002603 }
2604
Tejun Heo112202d2013-02-13 19:29:12 -08002605 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
Tejun Heo73f53c42010-06-29 10:07:11 +02002606 complete(&wq->first_flusher->done);
2607
2608 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609}
2610
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002611/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002613 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002615 * This function sleeps until all work items which were queued on entry
2616 * have finished execution, but it is not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002618void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619{
Tejun Heo73f53c42010-06-29 10:07:11 +02002620 struct wq_flusher this_flusher = {
2621 .list = LIST_HEAD_INIT(this_flusher.list),
2622 .flush_color = -1,
2623 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2624 };
2625 int next_color;
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -07002626
Ingo Molnar3295f0e2008-08-11 10:30:30 +02002627 lock_map_acquire(&wq->lockdep_map);
2628 lock_map_release(&wq->lockdep_map);
Tejun Heo73f53c42010-06-29 10:07:11 +02002629
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002630 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002631
2632 /*
2633 * Start-to-wait phase
2634 */
2635 next_color = work_next_color(wq->work_color);
2636
2637 if (next_color != wq->flush_color) {
2638 /*
2639 * Color space is not full. The current work_color
2640 * becomes our flush_color and work_color is advanced
2641 * by one.
2642 */
Tejun Heo6183c002013-03-12 11:29:57 -07002643 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
Tejun Heo73f53c42010-06-29 10:07:11 +02002644 this_flusher.flush_color = wq->work_color;
2645 wq->work_color = next_color;
2646
2647 if (!wq->first_flusher) {
2648 /* no flush in progress, become the first flusher */
Tejun Heo6183c002013-03-12 11:29:57 -07002649 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002650
2651 wq->first_flusher = &this_flusher;
2652
Tejun Heo112202d2013-02-13 19:29:12 -08002653 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
Tejun Heo73f53c42010-06-29 10:07:11 +02002654 wq->work_color)) {
2655 /* nothing to flush, done */
2656 wq->flush_color = next_color;
2657 wq->first_flusher = NULL;
2658 goto out_unlock;
2659 }
2660 } else {
2661 /* wait in queue */
Tejun Heo6183c002013-03-12 11:29:57 -07002662 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002663 list_add_tail(&this_flusher.list, &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002664 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002665 }
2666 } else {
2667 /*
2668 * Oops, color space is full, wait on overflow queue.
2669 * The next flush completion will assign us
2670 * flush_color and transfer to flusher_queue.
2671 */
2672 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2673 }
2674
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002675 mutex_unlock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002676
2677 wait_for_completion(&this_flusher.done);
2678
2679 /*
2680 * Wake-up-and-cascade phase
2681 *
2682 * First flushers are responsible for cascading flushes and
2683 * handling overflow. Non-first flushers can simply return.
2684 */
2685 if (wq->first_flusher != &this_flusher)
2686 return;
2687
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002688 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002689
Tejun Heo4ce48b32010-07-02 10:03:51 +02002690 /* we might have raced, check again with mutex held */
2691 if (wq->first_flusher != &this_flusher)
2692 goto out_unlock;
2693
Tejun Heo73f53c42010-06-29 10:07:11 +02002694 wq->first_flusher = NULL;
2695
Tejun Heo6183c002013-03-12 11:29:57 -07002696 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2697 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002698
2699 while (true) {
2700 struct wq_flusher *next, *tmp;
2701
2702 /* complete all the flushers sharing the current flush color */
2703 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2704 if (next->flush_color != wq->flush_color)
2705 break;
2706 list_del_init(&next->list);
2707 complete(&next->done);
2708 }
2709
Tejun Heo6183c002013-03-12 11:29:57 -07002710 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2711 wq->flush_color != work_next_color(wq->work_color));
Tejun Heo73f53c42010-06-29 10:07:11 +02002712
2713 /* this flush_color is finished, advance by one */
2714 wq->flush_color = work_next_color(wq->flush_color);
2715
2716 /* one color has been freed, handle overflow queue */
2717 if (!list_empty(&wq->flusher_overflow)) {
2718 /*
2719 * Assign the same color to all overflowed
2720 * flushers, advance work_color and append to
2721 * flusher_queue. This is the start-to-wait
2722 * phase for these overflowed flushers.
2723 */
2724 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2725 tmp->flush_color = wq->work_color;
2726
2727 wq->work_color = work_next_color(wq->work_color);
2728
2729 list_splice_tail_init(&wq->flusher_overflow,
2730 &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002731 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002732 }
2733
2734 if (list_empty(&wq->flusher_queue)) {
Tejun Heo6183c002013-03-12 11:29:57 -07002735 WARN_ON_ONCE(wq->flush_color != wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002736 break;
2737 }
2738
2739 /*
2740 * Need to flush more colors. Make the next flusher
Tejun Heo112202d2013-02-13 19:29:12 -08002741 * the new first flusher and arm pwqs.
Tejun Heo73f53c42010-06-29 10:07:11 +02002742 */
Tejun Heo6183c002013-03-12 11:29:57 -07002743 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2744 WARN_ON_ONCE(wq->flush_color != next->flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002745
2746 list_del_init(&next->list);
2747 wq->first_flusher = next;
2748
Tejun Heo112202d2013-02-13 19:29:12 -08002749 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
Tejun Heo73f53c42010-06-29 10:07:11 +02002750 break;
2751
2752 /*
2753 * Meh... this color is already done, clear first
2754 * flusher and repeat cascading.
2755 */
2756 wq->first_flusher = NULL;
2757 }
2758
2759out_unlock:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002760 mutex_unlock(&wq->mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761}
Dave Jonesae90dd52006-06-30 01:40:45 -04002762EXPORT_SYMBOL_GPL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002764/**
2765 * drain_workqueue - drain a workqueue
2766 * @wq: workqueue to drain
2767 *
2768 * Wait until the workqueue becomes empty. While draining is in progress,
2769 * only chain queueing is allowed. IOW, only currently pending or running
2770 * work items on @wq can queue further work items on it. @wq is flushed
2771 * repeatedly until it becomes empty. The number of flushing is detemined
2772 * by the depth of chaining and should be relatively short. Whine if it
2773 * takes too long.
2774 */
2775void drain_workqueue(struct workqueue_struct *wq)
2776{
2777 unsigned int flush_cnt = 0;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002778 struct pool_workqueue *pwq;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002779
2780 /*
2781 * __queue_work() needs to test whether there are drainers, is much
2782 * hotter than drain_workqueue() and already looks at @wq->flags.
Tejun Heo618b01e2013-03-12 11:30:04 -07002783 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002784 */
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002785 mutex_lock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002786 if (!wq->nr_drainers++)
Tejun Heo618b01e2013-03-12 11:30:04 -07002787 wq->flags |= __WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002788 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002789reflush:
2790 flush_workqueue(wq);
2791
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002792 mutex_lock(&wq->mutex);
Tejun Heo76af4d92013-03-12 11:30:00 -07002793
Tejun Heo49e3cf42013-03-12 11:29:58 -07002794 for_each_pwq(pwq, wq) {
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002795 bool drained;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002796
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002797 spin_lock_irq(&pwq->pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08002798 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002799 spin_unlock_irq(&pwq->pool->lock);
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002800
2801 if (drained)
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002802 continue;
2803
2804 if (++flush_cnt == 10 ||
2805 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002806 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
Valentin Ilie044c7822012-08-19 00:52:42 +03002807 wq->name, flush_cnt);
Tejun Heo76af4d92013-03-12 11:30:00 -07002808
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002809 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002810 goto reflush;
2811 }
2812
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002813 if (!--wq->nr_drainers)
Tejun Heo618b01e2013-03-12 11:30:04 -07002814 wq->flags &= ~__WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002815 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002816}
2817EXPORT_SYMBOL_GPL(drain_workqueue);
2818
Tejun Heo606a5022012-08-20 14:51:23 -07002819static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
Tejun Heobaf59022010-09-16 10:42:16 +02002820{
2821 struct worker *worker = NULL;
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002822 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002823 struct pool_workqueue *pwq;
Tejun Heobaf59022010-09-16 10:42:16 +02002824
2825 might_sleep();
Tejun Heobaf59022010-09-16 10:42:16 +02002826
Tejun Heofa1b54e2013-03-12 11:30:00 -07002827 local_irq_disable();
2828 pool = get_work_pool(work);
2829 if (!pool) {
2830 local_irq_enable();
2831 return false;
2832 }
2833
2834 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08002835 /* see the comment in try_to_grab_pending() with the same code */
Tejun Heo112202d2013-02-13 19:29:12 -08002836 pwq = get_work_pwq(work);
2837 if (pwq) {
2838 if (unlikely(pwq->pool != pool))
Tejun Heobaf59022010-09-16 10:42:16 +02002839 goto already_gone;
Tejun Heo606a5022012-08-20 14:51:23 -07002840 } else {
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002841 worker = find_worker_executing_work(pool, work);
Tejun Heobaf59022010-09-16 10:42:16 +02002842 if (!worker)
2843 goto already_gone;
Tejun Heo112202d2013-02-13 19:29:12 -08002844 pwq = worker->current_pwq;
Tejun Heo606a5022012-08-20 14:51:23 -07002845 }
Tejun Heobaf59022010-09-16 10:42:16 +02002846
Tejun Heo112202d2013-02-13 19:29:12 -08002847 insert_wq_barrier(pwq, barr, work, worker);
Tejun Heod565ed62013-01-24 11:01:33 -08002848 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002849
Tejun Heoe1594892011-01-09 23:32:15 +01002850 /*
2851 * If @max_active is 1 or rescuer is in use, flushing another work
2852 * item on the same workqueue may lead to deadlock. Make sure the
2853 * flusher is not running on the same workqueue by verifying write
2854 * access.
2855 */
Tejun Heo493008a2013-03-12 11:30:03 -07002856 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
Tejun Heo112202d2013-02-13 19:29:12 -08002857 lock_map_acquire(&pwq->wq->lockdep_map);
Tejun Heoe1594892011-01-09 23:32:15 +01002858 else
Tejun Heo112202d2013-02-13 19:29:12 -08002859 lock_map_acquire_read(&pwq->wq->lockdep_map);
2860 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoe1594892011-01-09 23:32:15 +01002861
Tejun Heobaf59022010-09-16 10:42:16 +02002862 return true;
2863already_gone:
Tejun Heod565ed62013-01-24 11:01:33 -08002864 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002865 return false;
2866}
2867
Oleg Nesterovdb700892008-07-25 01:47:49 -07002868/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002869 * flush_work - wait for a work to finish executing the last queueing instance
2870 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07002871 *
Tejun Heo606a5022012-08-20 14:51:23 -07002872 * Wait until @work has finished execution. @work is guaranteed to be idle
2873 * on return if it hasn't been requeued since flush started.
Tejun Heo401a8d02010-09-16 10:36:00 +02002874 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002875 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002876 * %true if flush_work() waited for the work to finish execution,
2877 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07002878 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002879bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07002880{
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002881 struct wq_barrier barr;
2882
Stephen Boyd0976dfc2012-04-20 17:28:50 -07002883 lock_map_acquire(&work->lockdep_map);
2884 lock_map_release(&work->lockdep_map);
2885
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002886 if (start_flush_work(work, &barr)) {
2887 wait_for_completion(&barr.done);
2888 destroy_work_on_stack(&barr.work);
2889 return true;
2890 } else {
2891 return false;
2892 }
Oleg Nesterovdb700892008-07-25 01:47:49 -07002893}
2894EXPORT_SYMBOL_GPL(flush_work);
2895
Tejun Heo36e227d2012-08-03 10:30:46 -07002896static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
Tejun Heo401a8d02010-09-16 10:36:00 +02002897{
Tejun Heobbb68df2012-08-03 10:30:46 -07002898 unsigned long flags;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002899 int ret;
2900
2901 do {
Tejun Heobbb68df2012-08-03 10:30:46 -07002902 ret = try_to_grab_pending(work, is_dwork, &flags);
2903 /*
2904 * If someone else is canceling, wait for the same event it
2905 * would be waiting for before retrying.
2906 */
2907 if (unlikely(ret == -ENOENT))
Tejun Heo606a5022012-08-20 14:51:23 -07002908 flush_work(work);
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002909 } while (unlikely(ret < 0));
2910
Tejun Heobbb68df2012-08-03 10:30:46 -07002911 /* tell other tasks trying to grab @work to back off */
2912 mark_work_canceling(work);
2913 local_irq_restore(flags);
2914
Tejun Heo606a5022012-08-20 14:51:23 -07002915 flush_work(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002916 clear_work_data(work);
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002917 return ret;
2918}
2919
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002920/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002921 * cancel_work_sync - cancel a work and wait for it to finish
2922 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002923 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002924 * Cancel @work and wait for its execution to finish. This function
2925 * can be used even if the work re-queues itself or migrates to
2926 * another workqueue. On return from this function, @work is
2927 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002928 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002929 * cancel_work_sync(&delayed_work->work) must not be used for
2930 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002931 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002932 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002933 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02002934 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002935 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002936 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002937 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002938bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002939{
Tejun Heo36e227d2012-08-03 10:30:46 -07002940 return __cancel_work_timer(work, false);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002941}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07002942EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002943
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002944/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002945 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2946 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002947 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002948 * Delayed timer is cancelled and the pending work is queued for
2949 * immediate execution. Like flush_work(), this function only
2950 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002951 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002952 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002953 * %true if flush_work() waited for the work to finish execution,
2954 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002955 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002956bool flush_delayed_work(struct delayed_work *dwork)
2957{
Tejun Heo8930cab2012-08-03 10:30:45 -07002958 local_irq_disable();
Tejun Heo401a8d02010-09-16 10:36:00 +02002959 if (del_timer_sync(&dwork->timer))
Lai Jiangshan60c057b2013-02-06 18:04:53 -08002960 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Tejun Heo8930cab2012-08-03 10:30:45 -07002961 local_irq_enable();
Tejun Heo401a8d02010-09-16 10:36:00 +02002962 return flush_work(&dwork->work);
2963}
2964EXPORT_SYMBOL(flush_delayed_work);
2965
2966/**
Tejun Heo57b30ae2012-08-21 13:18:24 -07002967 * cancel_delayed_work - cancel a delayed work
2968 * @dwork: delayed_work to cancel
Tejun Heo09383492010-09-16 10:48:29 +02002969 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002970 * Kill off a pending delayed_work.
2971 *
2972 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2973 * pending.
2974 *
2975 * Note:
2976 * The work callback function may still be running on return, unless
2977 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2978 * use cancel_delayed_work_sync() to wait on it.
Tejun Heo09383492010-09-16 10:48:29 +02002979 *
Tejun Heo57b30ae2012-08-21 13:18:24 -07002980 * This function is safe to call from any context including IRQ handler.
Tejun Heo09383492010-09-16 10:48:29 +02002981 */
Tejun Heo57b30ae2012-08-21 13:18:24 -07002982bool cancel_delayed_work(struct delayed_work *dwork)
Tejun Heo09383492010-09-16 10:48:29 +02002983{
Tejun Heo57b30ae2012-08-21 13:18:24 -07002984 unsigned long flags;
2985 int ret;
2986
2987 do {
2988 ret = try_to_grab_pending(&dwork->work, true, &flags);
2989 } while (unlikely(ret == -EAGAIN));
2990
2991 if (unlikely(ret < 0))
2992 return false;
2993
Tejun Heo7c3eed52013-01-24 11:01:33 -08002994 set_work_pool_and_clear_pending(&dwork->work,
2995 get_work_pool_id(&dwork->work));
Tejun Heo57b30ae2012-08-21 13:18:24 -07002996 local_irq_restore(flags);
Dan Magenheimerc0158ca2012-10-18 16:31:37 -07002997 return ret;
Tejun Heo09383492010-09-16 10:48:29 +02002998}
Tejun Heo57b30ae2012-08-21 13:18:24 -07002999EXPORT_SYMBOL(cancel_delayed_work);
Tejun Heo09383492010-09-16 10:48:29 +02003000
3001/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003002 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3003 * @dwork: the delayed work cancel
3004 *
3005 * This is cancel_work_sync() for delayed works.
3006 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003007 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003008 * %true if @dwork was pending, %false otherwise.
3009 */
3010bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003011{
Tejun Heo36e227d2012-08-03 10:30:46 -07003012 return __cancel_work_timer(&dwork->work, true);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003013}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07003014EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07003016/**
Tejun Heo31ddd872010-10-19 11:14:49 +02003017 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07003018 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07003019 *
Tejun Heo31ddd872010-10-19 11:14:49 +02003020 * schedule_on_each_cpu() executes @func on each online CPU using the
3021 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07003022 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02003023 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003024 * Return:
Tejun Heo31ddd872010-10-19 11:14:49 +02003025 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07003026 */
David Howells65f27f32006-11-22 14:55:48 +00003027int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003028{
3029 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02003030 struct work_struct __percpu *works;
Christoph Lameter15316ba2006-01-08 01:00:43 -08003031
Andrew Mortonb6136772006-06-25 05:47:49 -07003032 works = alloc_percpu(struct work_struct);
3033 if (!works)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003034 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07003035
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003036 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08003037
Christoph Lameter15316ba2006-01-08 01:00:43 -08003038 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01003039 struct work_struct *work = per_cpu_ptr(works, cpu);
3040
3041 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003042 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02003043 }
Tejun Heo93981802009-11-17 14:06:20 -08003044
3045 for_each_online_cpu(cpu)
3046 flush_work(per_cpu_ptr(works, cpu));
3047
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003048 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07003049 free_percpu(works);
Christoph Lameter15316ba2006-01-08 01:00:43 -08003050 return 0;
3051}
3052
Alan Sterneef6a7d2010-02-12 17:39:21 +09003053/**
3054 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3055 *
3056 * Forces execution of the kernel-global workqueue and blocks until its
3057 * completion.
3058 *
3059 * Think twice before calling this function! It's very easy to get into
3060 * trouble if you don't take great care. Either of the following situations
3061 * will lead to deadlock:
3062 *
3063 * One of the work items currently on the workqueue needs to acquire
3064 * a lock held by your code or its caller.
3065 *
3066 * Your code is running in the context of a work routine.
3067 *
3068 * They will be detected by lockdep when they occur, but the first might not
3069 * occur very often. It depends on what work items are on the workqueue and
3070 * what locks they need, which you have no control over.
3071 *
3072 * In most situations flushing the entire workqueue is overkill; you merely
3073 * need to know that a particular work item isn't queued and isn't running.
3074 * In such cases you should use cancel_delayed_work_sync() or
3075 * cancel_work_sync() instead.
3076 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077void flush_scheduled_work(void)
3078{
Tejun Heod320c032010-06-29 10:07:14 +02003079 flush_workqueue(system_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080}
Dave Jonesae90dd52006-06-30 01:40:45 -04003081EXPORT_SYMBOL(flush_scheduled_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082
3083/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06003084 * execute_in_process_context - reliably execute the routine with user context
3085 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06003086 * @ew: guaranteed storage for the execute work structure (must
3087 * be available when the work executes)
3088 *
3089 * Executes the function immediately if process context is available,
3090 * otherwise schedules the function for delayed execution.
3091 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003092 * Return: 0 - function was executed
James Bottomley1fa44ec2006-02-23 12:43:43 -06003093 * 1 - function was scheduled for execution
3094 */
David Howells65f27f32006-11-22 14:55:48 +00003095int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06003096{
3097 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00003098 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003099 return 0;
3100 }
3101
David Howells65f27f32006-11-22 14:55:48 +00003102 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003103 schedule_work(&ew->work);
3104
3105 return 1;
3106}
3107EXPORT_SYMBOL_GPL(execute_in_process_context);
3108
Tejun Heo226223a2013-03-12 11:30:05 -07003109#ifdef CONFIG_SYSFS
3110/*
3111 * Workqueues with WQ_SYSFS flag set is visible to userland via
3112 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3113 * following attributes.
3114 *
3115 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3116 * max_active RW int : maximum number of in-flight work items
3117 *
3118 * Unbound workqueues have the following extra attributes.
3119 *
3120 * id RO int : the associated pool ID
3121 * nice RW int : nice value of the workers
3122 * cpumask RW mask : bitmask of allowed CPUs for the workers
3123 */
3124struct wq_device {
3125 struct workqueue_struct *wq;
3126 struct device dev;
3127};
3128
3129static struct workqueue_struct *dev_to_wq(struct device *dev)
3130{
3131 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3132
3133 return wq_dev->wq;
3134}
3135
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003136static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3137 char *buf)
Tejun Heo226223a2013-03-12 11:30:05 -07003138{
3139 struct workqueue_struct *wq = dev_to_wq(dev);
3140
3141 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3142}
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003143static DEVICE_ATTR_RO(per_cpu);
Tejun Heo226223a2013-03-12 11:30:05 -07003144
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003145static ssize_t max_active_show(struct device *dev,
3146 struct device_attribute *attr, char *buf)
Tejun Heo226223a2013-03-12 11:30:05 -07003147{
3148 struct workqueue_struct *wq = dev_to_wq(dev);
3149
3150 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3151}
3152
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003153static ssize_t max_active_store(struct device *dev,
3154 struct device_attribute *attr, const char *buf,
3155 size_t count)
Tejun Heo226223a2013-03-12 11:30:05 -07003156{
3157 struct workqueue_struct *wq = dev_to_wq(dev);
3158 int val;
3159
3160 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3161 return -EINVAL;
3162
3163 workqueue_set_max_active(wq, val);
3164 return count;
3165}
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003166static DEVICE_ATTR_RW(max_active);
Tejun Heo226223a2013-03-12 11:30:05 -07003167
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003168static struct attribute *wq_sysfs_attrs[] = {
3169 &dev_attr_per_cpu.attr,
3170 &dev_attr_max_active.attr,
3171 NULL,
Tejun Heo226223a2013-03-12 11:30:05 -07003172};
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003173ATTRIBUTE_GROUPS(wq_sysfs);
Tejun Heo226223a2013-03-12 11:30:05 -07003174
Tejun Heod55262c2013-04-01 11:23:38 -07003175static ssize_t wq_pool_ids_show(struct device *dev,
3176 struct device_attribute *attr, char *buf)
Tejun Heo226223a2013-03-12 11:30:05 -07003177{
3178 struct workqueue_struct *wq = dev_to_wq(dev);
Tejun Heod55262c2013-04-01 11:23:38 -07003179 const char *delim = "";
3180 int node, written = 0;
Tejun Heo226223a2013-03-12 11:30:05 -07003181
3182 rcu_read_lock_sched();
Tejun Heod55262c2013-04-01 11:23:38 -07003183 for_each_node(node) {
3184 written += scnprintf(buf + written, PAGE_SIZE - written,
3185 "%s%d:%d", delim, node,
3186 unbound_pwq_by_node(wq, node)->pool->id);
3187 delim = " ";
3188 }
3189 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
Tejun Heo226223a2013-03-12 11:30:05 -07003190 rcu_read_unlock_sched();
3191
3192 return written;
3193}
3194
3195static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3196 char *buf)
3197{
3198 struct workqueue_struct *wq = dev_to_wq(dev);
3199 int written;
3200
Tejun Heo6029a912013-04-01 11:23:34 -07003201 mutex_lock(&wq->mutex);
3202 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3203 mutex_unlock(&wq->mutex);
Tejun Heo226223a2013-03-12 11:30:05 -07003204
3205 return written;
3206}
3207
3208/* prepare workqueue_attrs for sysfs store operations */
3209static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3210{
3211 struct workqueue_attrs *attrs;
3212
3213 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3214 if (!attrs)
3215 return NULL;
3216
Tejun Heo6029a912013-04-01 11:23:34 -07003217 mutex_lock(&wq->mutex);
3218 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3219 mutex_unlock(&wq->mutex);
Tejun Heo226223a2013-03-12 11:30:05 -07003220 return attrs;
3221}
3222
3223static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3224 const char *buf, size_t count)
3225{
3226 struct workqueue_struct *wq = dev_to_wq(dev);
3227 struct workqueue_attrs *attrs;
3228 int ret;
3229
3230 attrs = wq_sysfs_prep_attrs(wq);
3231 if (!attrs)
3232 return -ENOMEM;
3233
3234 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
Dongsheng Yang14481842014-02-11 15:34:52 +08003235 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
Tejun Heo226223a2013-03-12 11:30:05 -07003236 ret = apply_workqueue_attrs(wq, attrs);
3237 else
3238 ret = -EINVAL;
3239
3240 free_workqueue_attrs(attrs);
3241 return ret ?: count;
3242}
3243
3244static ssize_t wq_cpumask_show(struct device *dev,
3245 struct device_attribute *attr, char *buf)
3246{
3247 struct workqueue_struct *wq = dev_to_wq(dev);
3248 int written;
3249
Tejun Heo6029a912013-04-01 11:23:34 -07003250 mutex_lock(&wq->mutex);
3251 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3252 mutex_unlock(&wq->mutex);
Tejun Heo226223a2013-03-12 11:30:05 -07003253
3254 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3255 return written;
3256}
3257
3258static ssize_t wq_cpumask_store(struct device *dev,
3259 struct device_attribute *attr,
3260 const char *buf, size_t count)
3261{
3262 struct workqueue_struct *wq = dev_to_wq(dev);
3263 struct workqueue_attrs *attrs;
3264 int ret;
3265
3266 attrs = wq_sysfs_prep_attrs(wq);
3267 if (!attrs)
3268 return -ENOMEM;
3269
3270 ret = cpumask_parse(buf, attrs->cpumask);
3271 if (!ret)
3272 ret = apply_workqueue_attrs(wq, attrs);
3273
3274 free_workqueue_attrs(attrs);
3275 return ret ?: count;
3276}
3277
Tejun Heod55262c2013-04-01 11:23:38 -07003278static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3279 char *buf)
3280{
3281 struct workqueue_struct *wq = dev_to_wq(dev);
3282 int written;
3283
3284 mutex_lock(&wq->mutex);
3285 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3286 !wq->unbound_attrs->no_numa);
3287 mutex_unlock(&wq->mutex);
3288
3289 return written;
3290}
3291
3292static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3293 const char *buf, size_t count)
3294{
3295 struct workqueue_struct *wq = dev_to_wq(dev);
3296 struct workqueue_attrs *attrs;
3297 int v, ret;
3298
3299 attrs = wq_sysfs_prep_attrs(wq);
3300 if (!attrs)
3301 return -ENOMEM;
3302
3303 ret = -EINVAL;
3304 if (sscanf(buf, "%d", &v) == 1) {
3305 attrs->no_numa = !v;
3306 ret = apply_workqueue_attrs(wq, attrs);
3307 }
3308
3309 free_workqueue_attrs(attrs);
3310 return ret ?: count;
3311}
3312
Tejun Heo226223a2013-03-12 11:30:05 -07003313static struct device_attribute wq_sysfs_unbound_attrs[] = {
Tejun Heod55262c2013-04-01 11:23:38 -07003314 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
Tejun Heo226223a2013-03-12 11:30:05 -07003315 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3316 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
Tejun Heod55262c2013-04-01 11:23:38 -07003317 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
Tejun Heo226223a2013-03-12 11:30:05 -07003318 __ATTR_NULL,
3319};
3320
3321static struct bus_type wq_subsys = {
3322 .name = "workqueue",
Greg Kroah-Hartman1a6661d2013-08-23 14:24:41 -07003323 .dev_groups = wq_sysfs_groups,
Tejun Heo226223a2013-03-12 11:30:05 -07003324};
3325
3326static int __init wq_sysfs_init(void)
3327{
3328 return subsys_virtual_register(&wq_subsys, NULL);
3329}
3330core_initcall(wq_sysfs_init);
3331
3332static void wq_device_release(struct device *dev)
3333{
3334 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3335
3336 kfree(wq_dev);
3337}
3338
3339/**
3340 * workqueue_sysfs_register - make a workqueue visible in sysfs
3341 * @wq: the workqueue to register
3342 *
3343 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3344 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3345 * which is the preferred method.
3346 *
3347 * Workqueue user should use this function directly iff it wants to apply
3348 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3349 * apply_workqueue_attrs() may race against userland updating the
3350 * attributes.
3351 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003352 * Return: 0 on success, -errno on failure.
Tejun Heo226223a2013-03-12 11:30:05 -07003353 */
3354int workqueue_sysfs_register(struct workqueue_struct *wq)
3355{
3356 struct wq_device *wq_dev;
3357 int ret;
3358
3359 /*
3360 * Adjusting max_active or creating new pwqs by applyting
3361 * attributes breaks ordering guarantee. Disallow exposing ordered
3362 * workqueues.
3363 */
3364 if (WARN_ON(wq->flags & __WQ_ORDERED))
3365 return -EINVAL;
3366
3367 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3368 if (!wq_dev)
3369 return -ENOMEM;
3370
3371 wq_dev->wq = wq;
3372 wq_dev->dev.bus = &wq_subsys;
3373 wq_dev->dev.init_name = wq->name;
3374 wq_dev->dev.release = wq_device_release;
3375
3376 /*
3377 * unbound_attrs are created separately. Suppress uevent until
3378 * everything is ready.
3379 */
3380 dev_set_uevent_suppress(&wq_dev->dev, true);
3381
3382 ret = device_register(&wq_dev->dev);
3383 if (ret) {
3384 kfree(wq_dev);
3385 wq->wq_dev = NULL;
3386 return ret;
3387 }
3388
3389 if (wq->flags & WQ_UNBOUND) {
3390 struct device_attribute *attr;
3391
3392 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3393 ret = device_create_file(&wq_dev->dev, attr);
3394 if (ret) {
3395 device_unregister(&wq_dev->dev);
3396 wq->wq_dev = NULL;
3397 return ret;
3398 }
3399 }
3400 }
3401
3402 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3403 return 0;
3404}
3405
3406/**
3407 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3408 * @wq: the workqueue to unregister
3409 *
3410 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3411 */
3412static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3413{
3414 struct wq_device *wq_dev = wq->wq_dev;
3415
3416 if (!wq->wq_dev)
3417 return;
3418
3419 wq->wq_dev = NULL;
3420 device_unregister(&wq_dev->dev);
3421}
3422#else /* CONFIG_SYSFS */
3423static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3424#endif /* CONFIG_SYSFS */
3425
Tejun Heo7a4e3442013-03-12 11:30:00 -07003426/**
3427 * free_workqueue_attrs - free a workqueue_attrs
3428 * @attrs: workqueue_attrs to free
3429 *
3430 * Undo alloc_workqueue_attrs().
3431 */
3432void free_workqueue_attrs(struct workqueue_attrs *attrs)
3433{
3434 if (attrs) {
3435 free_cpumask_var(attrs->cpumask);
3436 kfree(attrs);
3437 }
3438}
3439
3440/**
3441 * alloc_workqueue_attrs - allocate a workqueue_attrs
3442 * @gfp_mask: allocation mask to use
3443 *
3444 * Allocate a new workqueue_attrs, initialize with default settings and
Yacine Belkadid185af32013-07-31 14:59:24 -07003445 * return it.
3446 *
3447 * Return: The allocated new workqueue_attr on success. %NULL on failure.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003448 */
3449struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3450{
3451 struct workqueue_attrs *attrs;
3452
3453 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3454 if (!attrs)
3455 goto fail;
3456 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3457 goto fail;
3458
Tejun Heo13e2e552013-04-01 11:23:31 -07003459 cpumask_copy(attrs->cpumask, cpu_possible_mask);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003460 return attrs;
3461fail:
3462 free_workqueue_attrs(attrs);
3463 return NULL;
3464}
3465
Tejun Heo29c91e92013-03-12 11:30:03 -07003466static void copy_workqueue_attrs(struct workqueue_attrs *to,
3467 const struct workqueue_attrs *from)
3468{
3469 to->nice = from->nice;
3470 cpumask_copy(to->cpumask, from->cpumask);
Shaohua Li2865a8f2013-08-01 09:56:36 +08003471 /*
3472 * Unlike hash and equality test, this function doesn't ignore
3473 * ->no_numa as it is used for both pool and wq attrs. Instead,
3474 * get_unbound_pool() explicitly clears ->no_numa after copying.
3475 */
3476 to->no_numa = from->no_numa;
Tejun Heo29c91e92013-03-12 11:30:03 -07003477}
3478
Tejun Heo29c91e92013-03-12 11:30:03 -07003479/* hash value of the content of @attr */
3480static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3481{
3482 u32 hash = 0;
3483
3484 hash = jhash_1word(attrs->nice, hash);
Tejun Heo13e2e552013-04-01 11:23:31 -07003485 hash = jhash(cpumask_bits(attrs->cpumask),
3486 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
Tejun Heo29c91e92013-03-12 11:30:03 -07003487 return hash;
3488}
3489
3490/* content equality test */
3491static bool wqattrs_equal(const struct workqueue_attrs *a,
3492 const struct workqueue_attrs *b)
3493{
3494 if (a->nice != b->nice)
3495 return false;
3496 if (!cpumask_equal(a->cpumask, b->cpumask))
3497 return false;
3498 return true;
3499}
3500
Tejun Heo7a4e3442013-03-12 11:30:00 -07003501/**
3502 * init_worker_pool - initialize a newly zalloc'd worker_pool
3503 * @pool: worker_pool to initialize
3504 *
3505 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
Yacine Belkadid185af32013-07-31 14:59:24 -07003506 *
3507 * Return: 0 on success, -errno on failure. Even on failure, all fields
Tejun Heo29c91e92013-03-12 11:30:03 -07003508 * inside @pool proper are initialized and put_unbound_pool() can be called
3509 * on @pool safely to release it.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003510 */
3511static int init_worker_pool(struct worker_pool *pool)
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003512{
3513 spin_lock_init(&pool->lock);
Tejun Heo29c91e92013-03-12 11:30:03 -07003514 pool->id = -1;
3515 pool->cpu = -1;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003516 pool->node = NUMA_NO_NODE;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003517 pool->flags |= POOL_DISASSOCIATED;
3518 INIT_LIST_HEAD(&pool->worklist);
3519 INIT_LIST_HEAD(&pool->idle_list);
3520 hash_init(pool->busy_hash);
3521
3522 init_timer_deferrable(&pool->idle_timer);
3523 pool->idle_timer.function = idle_worker_timeout;
3524 pool->idle_timer.data = (unsigned long)pool;
3525
3526 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3527 (unsigned long)pool);
3528
3529 mutex_init(&pool->manager_arb);
Tejun Heobc3a1af2013-03-13 19:47:39 -07003530 mutex_init(&pool->manager_mutex);
Tejun Heo822d8402013-03-19 13:45:21 -07003531 idr_init(&pool->worker_idr);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003532
Tejun Heo29c91e92013-03-12 11:30:03 -07003533 INIT_HLIST_NODE(&pool->hash_node);
3534 pool->refcnt = 1;
3535
3536 /* shouldn't fail above this point */
Tejun Heo7a4e3442013-03-12 11:30:00 -07003537 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3538 if (!pool->attrs)
3539 return -ENOMEM;
3540 return 0;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003541}
3542
Tejun Heo29c91e92013-03-12 11:30:03 -07003543static void rcu_free_pool(struct rcu_head *rcu)
3544{
3545 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3546
Tejun Heo822d8402013-03-19 13:45:21 -07003547 idr_destroy(&pool->worker_idr);
Tejun Heo29c91e92013-03-12 11:30:03 -07003548 free_workqueue_attrs(pool->attrs);
3549 kfree(pool);
3550}
3551
3552/**
3553 * put_unbound_pool - put a worker_pool
3554 * @pool: worker_pool to put
3555 *
3556 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003557 * safe manner. get_unbound_pool() calls this function on its failure path
3558 * and this function should be able to release pools which went through,
3559 * successfully or not, init_worker_pool().
Tejun Heoa892cac2013-04-01 11:23:32 -07003560 *
3561 * Should be called with wq_pool_mutex held.
Tejun Heo29c91e92013-03-12 11:30:03 -07003562 */
3563static void put_unbound_pool(struct worker_pool *pool)
3564{
3565 struct worker *worker;
3566
Tejun Heoa892cac2013-04-01 11:23:32 -07003567 lockdep_assert_held(&wq_pool_mutex);
3568
3569 if (--pool->refcnt)
Tejun Heo29c91e92013-03-12 11:30:03 -07003570 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003571
3572 /* sanity checks */
3573 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
Tejun Heoa892cac2013-04-01 11:23:32 -07003574 WARN_ON(!list_empty(&pool->worklist)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003575 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003576
3577 /* release id and unhash */
3578 if (pool->id >= 0)
3579 idr_remove(&worker_pool_idr, pool->id);
3580 hash_del(&pool->hash_node);
3581
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003582 /*
3583 * Become the manager and destroy all workers. Grabbing
3584 * manager_arb prevents @pool's workers from blocking on
3585 * manager_mutex.
3586 */
Tejun Heo29c91e92013-03-12 11:30:03 -07003587 mutex_lock(&pool->manager_arb);
Tejun Heocd549682013-03-13 19:47:39 -07003588 mutex_lock(&pool->manager_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003589 spin_lock_irq(&pool->lock);
3590
3591 while ((worker = first_worker(pool)))
3592 destroy_worker(worker);
3593 WARN_ON(pool->nr_workers || pool->nr_idle);
3594
3595 spin_unlock_irq(&pool->lock);
Tejun Heocd549682013-03-13 19:47:39 -07003596 mutex_unlock(&pool->manager_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003597 mutex_unlock(&pool->manager_arb);
3598
3599 /* shut down the timers */
3600 del_timer_sync(&pool->idle_timer);
3601 del_timer_sync(&pool->mayday_timer);
3602
3603 /* sched-RCU protected to allow dereferences from get_work_pool() */
3604 call_rcu_sched(&pool->rcu, rcu_free_pool);
3605}
3606
3607/**
3608 * get_unbound_pool - get a worker_pool with the specified attributes
3609 * @attrs: the attributes of the worker_pool to get
3610 *
3611 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3612 * reference count and return it. If there already is a matching
3613 * worker_pool, it will be used; otherwise, this function attempts to
Yacine Belkadid185af32013-07-31 14:59:24 -07003614 * create a new one.
Tejun Heoa892cac2013-04-01 11:23:32 -07003615 *
3616 * Should be called with wq_pool_mutex held.
Yacine Belkadid185af32013-07-31 14:59:24 -07003617 *
3618 * Return: On success, a worker_pool with the same attributes as @attrs.
3619 * On failure, %NULL.
Tejun Heo29c91e92013-03-12 11:30:03 -07003620 */
3621static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3622{
Tejun Heo29c91e92013-03-12 11:30:03 -07003623 u32 hash = wqattrs_hash(attrs);
3624 struct worker_pool *pool;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003625 int node;
Tejun Heo29c91e92013-03-12 11:30:03 -07003626
Tejun Heoa892cac2013-04-01 11:23:32 -07003627 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003628
3629 /* do we already have a matching pool? */
Tejun Heo29c91e92013-03-12 11:30:03 -07003630 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3631 if (wqattrs_equal(pool->attrs, attrs)) {
3632 pool->refcnt++;
3633 goto out_unlock;
3634 }
3635 }
Tejun Heo29c91e92013-03-12 11:30:03 -07003636
3637 /* nope, create a new one */
3638 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3639 if (!pool || init_worker_pool(pool) < 0)
3640 goto fail;
3641
Lai Jiangshan12ee4fc2013-03-20 03:28:01 +08003642 if (workqueue_freezing)
3643 pool->flags |= POOL_FREEZING;
3644
Tejun Heo8864b4e2013-03-12 11:30:04 -07003645 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
Tejun Heo29c91e92013-03-12 11:30:03 -07003646 copy_workqueue_attrs(pool->attrs, attrs);
3647
Shaohua Li2865a8f2013-08-01 09:56:36 +08003648 /*
3649 * no_numa isn't a worker_pool attribute, always clear it. See
3650 * 'struct workqueue_attrs' comments for detail.
3651 */
3652 pool->attrs->no_numa = false;
3653
Tejun Heof3f90ad2013-04-01 11:23:34 -07003654 /* if cpumask is contained inside a NUMA node, we belong to that node */
3655 if (wq_numa_enabled) {
3656 for_each_node(node) {
3657 if (cpumask_subset(pool->attrs->cpumask,
3658 wq_numa_possible_cpumask[node])) {
3659 pool->node = node;
3660 break;
3661 }
3662 }
3663 }
3664
Tejun Heo29c91e92013-03-12 11:30:03 -07003665 if (worker_pool_assign_id(pool) < 0)
3666 goto fail;
3667
3668 /* create and start the initial worker */
Tejun Heoebf44d12013-03-13 19:47:39 -07003669 if (create_and_start_worker(pool) < 0)
Tejun Heo29c91e92013-03-12 11:30:03 -07003670 goto fail;
3671
Tejun Heo29c91e92013-03-12 11:30:03 -07003672 /* install */
Tejun Heo29c91e92013-03-12 11:30:03 -07003673 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3674out_unlock:
Tejun Heo29c91e92013-03-12 11:30:03 -07003675 return pool;
3676fail:
Tejun Heo29c91e92013-03-12 11:30:03 -07003677 if (pool)
3678 put_unbound_pool(pool);
3679 return NULL;
3680}
3681
Tejun Heo8864b4e2013-03-12 11:30:04 -07003682static void rcu_free_pwq(struct rcu_head *rcu)
3683{
3684 kmem_cache_free(pwq_cache,
3685 container_of(rcu, struct pool_workqueue, rcu));
3686}
3687
3688/*
3689 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3690 * and needs to be destroyed.
3691 */
3692static void pwq_unbound_release_workfn(struct work_struct *work)
3693{
3694 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3695 unbound_release_work);
3696 struct workqueue_struct *wq = pwq->wq;
3697 struct worker_pool *pool = pwq->pool;
Tejun Heobc0caf02013-04-01 11:23:31 -07003698 bool is_last;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003699
3700 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3701 return;
3702
Tejun Heo75ccf592013-03-12 11:30:04 -07003703 /*
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003704 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
Tejun Heo75ccf592013-03-12 11:30:04 -07003705 * necessary on release but do it anyway. It's easier to verify
3706 * and consistent with the linking path.
3707 */
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003708 mutex_lock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003709 list_del_rcu(&pwq->pwqs_node);
Tejun Heobc0caf02013-04-01 11:23:31 -07003710 is_last = list_empty(&wq->pwqs);
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003711 mutex_unlock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003712
Tejun Heoa892cac2013-04-01 11:23:32 -07003713 mutex_lock(&wq_pool_mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003714 put_unbound_pool(pool);
Tejun Heoa892cac2013-04-01 11:23:32 -07003715 mutex_unlock(&wq_pool_mutex);
3716
Tejun Heo8864b4e2013-03-12 11:30:04 -07003717 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3718
3719 /*
3720 * If we're the last pwq going away, @wq is already dead and no one
3721 * is gonna access it anymore. Free it.
3722 */
Tejun Heo6029a912013-04-01 11:23:34 -07003723 if (is_last) {
3724 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003725 kfree(wq);
Tejun Heo6029a912013-04-01 11:23:34 -07003726 }
Tejun Heo8864b4e2013-03-12 11:30:04 -07003727}
3728
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003729/**
Tejun Heo699ce092013-03-13 16:51:35 -07003730 * pwq_adjust_max_active - update a pwq's max_active to the current setting
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003731 * @pwq: target pool_workqueue
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003732 *
Tejun Heo699ce092013-03-13 16:51:35 -07003733 * If @pwq isn't freezing, set @pwq->max_active to the associated
3734 * workqueue's saved_max_active and activate delayed work items
3735 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003736 */
Tejun Heo699ce092013-03-13 16:51:35 -07003737static void pwq_adjust_max_active(struct pool_workqueue *pwq)
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003738{
Tejun Heo699ce092013-03-13 16:51:35 -07003739 struct workqueue_struct *wq = pwq->wq;
3740 bool freezable = wq->flags & WQ_FREEZABLE;
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003741
Tejun Heo699ce092013-03-13 16:51:35 -07003742 /* for @wq->saved_max_active */
Lai Jiangshana357fc02013-03-25 16:57:19 -07003743 lockdep_assert_held(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003744
3745 /* fast exit for non-freezable wqs */
3746 if (!freezable && pwq->max_active == wq->saved_max_active)
3747 return;
3748
Lai Jiangshana357fc02013-03-25 16:57:19 -07003749 spin_lock_irq(&pwq->pool->lock);
Tejun Heo699ce092013-03-13 16:51:35 -07003750
3751 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3752 pwq->max_active = wq->saved_max_active;
3753
3754 while (!list_empty(&pwq->delayed_works) &&
3755 pwq->nr_active < pwq->max_active)
3756 pwq_activate_first_delayed(pwq);
Lai Jiangshan951a0782013-03-20 10:52:30 -07003757
3758 /*
3759 * Need to kick a worker after thawed or an unbound wq's
3760 * max_active is bumped. It's a slow path. Do it always.
3761 */
3762 wake_up_worker(pwq->pool);
Tejun Heo699ce092013-03-13 16:51:35 -07003763 } else {
3764 pwq->max_active = 0;
3765 }
3766
Lai Jiangshana357fc02013-03-25 16:57:19 -07003767 spin_unlock_irq(&pwq->pool->lock);
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003768}
3769
Tejun Heoe50aba92013-04-01 11:23:35 -07003770/* initialize newly alloced @pwq which is associated with @wq and @pool */
Tejun Heof147f292013-04-01 11:23:35 -07003771static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3772 struct worker_pool *pool)
Tejun Heod2c1d402013-03-12 11:30:04 -07003773{
3774 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3775
Tejun Heoe50aba92013-04-01 11:23:35 -07003776 memset(pwq, 0, sizeof(*pwq));
3777
Tejun Heod2c1d402013-03-12 11:30:04 -07003778 pwq->pool = pool;
3779 pwq->wq = wq;
3780 pwq->flush_color = -1;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003781 pwq->refcnt = 1;
Tejun Heod2c1d402013-03-12 11:30:04 -07003782 INIT_LIST_HEAD(&pwq->delayed_works);
Tejun Heo1befcf32013-04-01 11:23:35 -07003783 INIT_LIST_HEAD(&pwq->pwqs_node);
Tejun Heod2c1d402013-03-12 11:30:04 -07003784 INIT_LIST_HEAD(&pwq->mayday_node);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003785 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
Tejun Heof147f292013-04-01 11:23:35 -07003786}
Tejun Heod2c1d402013-03-12 11:30:04 -07003787
Tejun Heof147f292013-04-01 11:23:35 -07003788/* sync @pwq with the current state of its associated wq and link it */
Tejun Heo1befcf32013-04-01 11:23:35 -07003789static void link_pwq(struct pool_workqueue *pwq)
Tejun Heof147f292013-04-01 11:23:35 -07003790{
3791 struct workqueue_struct *wq = pwq->wq;
3792
3793 lockdep_assert_held(&wq->mutex);
Tejun Heo75ccf592013-03-12 11:30:04 -07003794
Tejun Heo1befcf32013-04-01 11:23:35 -07003795 /* may be called multiple times, ignore if already linked */
3796 if (!list_empty(&pwq->pwqs_node))
3797 return;
3798
Tejun Heo983ca252013-03-13 16:51:35 -07003799 /*
3800 * Set the matching work_color. This is synchronized with
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003801 * wq->mutex to avoid confusing flush_workqueue().
Tejun Heo983ca252013-03-13 16:51:35 -07003802 */
Tejun Heo75ccf592013-03-12 11:30:04 -07003803 pwq->work_color = wq->work_color;
Tejun Heo983ca252013-03-13 16:51:35 -07003804
3805 /* sync max_active to the current setting */
3806 pwq_adjust_max_active(pwq);
3807
3808 /* link in @pwq */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003809 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
Tejun Heof147f292013-04-01 11:23:35 -07003810}
Lai Jiangshana357fc02013-03-25 16:57:19 -07003811
Tejun Heof147f292013-04-01 11:23:35 -07003812/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3813static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3814 const struct workqueue_attrs *attrs)
3815{
3816 struct worker_pool *pool;
3817 struct pool_workqueue *pwq;
3818
3819 lockdep_assert_held(&wq_pool_mutex);
3820
3821 pool = get_unbound_pool(attrs);
3822 if (!pool)
3823 return NULL;
3824
Tejun Heoe50aba92013-04-01 11:23:35 -07003825 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
Tejun Heof147f292013-04-01 11:23:35 -07003826 if (!pwq) {
3827 put_unbound_pool(pool);
3828 return NULL;
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003829 }
Tejun Heo6029a912013-04-01 11:23:34 -07003830
Tejun Heof147f292013-04-01 11:23:35 -07003831 init_pwq(pwq, wq, pool);
3832 return pwq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003833}
3834
Tejun Heo4c16bd32013-04-01 11:23:36 -07003835/* undo alloc_unbound_pwq(), used only in the error path */
3836static void free_unbound_pwq(struct pool_workqueue *pwq)
3837{
3838 lockdep_assert_held(&wq_pool_mutex);
3839
3840 if (pwq) {
3841 put_unbound_pool(pwq->pool);
Wei Yongjuncece95d2013-04-09 14:29:11 +08003842 kmem_cache_free(pwq_cache, pwq);
Tejun Heo4c16bd32013-04-01 11:23:36 -07003843 }
3844}
3845
3846/**
3847 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3848 * @attrs: the wq_attrs of interest
3849 * @node: the target NUMA node
3850 * @cpu_going_down: if >= 0, the CPU to consider as offline
3851 * @cpumask: outarg, the resulting cpumask
3852 *
3853 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3854 * @cpu_going_down is >= 0, that cpu is considered offline during
Yacine Belkadid185af32013-07-31 14:59:24 -07003855 * calculation. The result is stored in @cpumask.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003856 *
3857 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3858 * enabled and @node has online CPUs requested by @attrs, the returned
3859 * cpumask is the intersection of the possible CPUs of @node and
3860 * @attrs->cpumask.
3861 *
3862 * The caller is responsible for ensuring that the cpumask of @node stays
3863 * stable.
Yacine Belkadid185af32013-07-31 14:59:24 -07003864 *
3865 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3866 * %false if equal.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003867 */
3868static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3869 int cpu_going_down, cpumask_t *cpumask)
3870{
Tejun Heod55262c2013-04-01 11:23:38 -07003871 if (!wq_numa_enabled || attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003872 goto use_dfl;
3873
3874 /* does @node have any online CPUs @attrs wants? */
3875 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3876 if (cpu_going_down >= 0)
3877 cpumask_clear_cpu(cpu_going_down, cpumask);
3878
3879 if (cpumask_empty(cpumask))
3880 goto use_dfl;
3881
3882 /* yeap, return possible CPUs in @node that @attrs wants */
3883 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3884 return !cpumask_equal(cpumask, attrs->cpumask);
3885
3886use_dfl:
3887 cpumask_copy(cpumask, attrs->cpumask);
3888 return false;
3889}
3890
Tejun Heo1befcf32013-04-01 11:23:35 -07003891/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3892static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3893 int node,
3894 struct pool_workqueue *pwq)
3895{
3896 struct pool_workqueue *old_pwq;
3897
3898 lockdep_assert_held(&wq->mutex);
3899
3900 /* link_pwq() can handle duplicate calls */
3901 link_pwq(pwq);
3902
3903 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3904 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3905 return old_pwq;
3906}
3907
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003908/**
3909 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3910 * @wq: the target workqueue
3911 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3912 *
Tejun Heo4c16bd32013-04-01 11:23:36 -07003913 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3914 * machines, this function maps a separate pwq to each NUMA node with
3915 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3916 * NUMA node it was issued on. Older pwqs are released as in-flight work
3917 * items finish. Note that a work item which repeatedly requeues itself
3918 * back-to-back will stay on its current pwq.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003919 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003920 * Performs GFP_KERNEL allocations.
3921 *
3922 * Return: 0 on success and -errno on failure.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003923 */
3924int apply_workqueue_attrs(struct workqueue_struct *wq,
3925 const struct workqueue_attrs *attrs)
3926{
Tejun Heo4c16bd32013-04-01 11:23:36 -07003927 struct workqueue_attrs *new_attrs, *tmp_attrs;
3928 struct pool_workqueue **pwq_tbl, *dfl_pwq;
Tejun Heof147f292013-04-01 11:23:35 -07003929 int node, ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003930
Tejun Heo8719dce2013-03-12 11:30:04 -07003931 /* only unbound workqueues can change attributes */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003932 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3933 return -EINVAL;
3934
Tejun Heo8719dce2013-03-12 11:30:04 -07003935 /* creating multiple pwqs breaks ordering guarantee */
3936 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3937 return -EINVAL;
3938
Tejun Heo4c16bd32013-04-01 11:23:36 -07003939 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
Tejun Heo13e2e552013-04-01 11:23:31 -07003940 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
Tejun Heo4c16bd32013-04-01 11:23:36 -07003941 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3942 if (!pwq_tbl || !new_attrs || !tmp_attrs)
Tejun Heo13e2e552013-04-01 11:23:31 -07003943 goto enomem;
3944
Tejun Heo4c16bd32013-04-01 11:23:36 -07003945 /* make a copy of @attrs and sanitize it */
Tejun Heo13e2e552013-04-01 11:23:31 -07003946 copy_workqueue_attrs(new_attrs, attrs);
3947 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3948
Tejun Heo4c16bd32013-04-01 11:23:36 -07003949 /*
3950 * We may create multiple pwqs with differing cpumasks. Make a
3951 * copy of @new_attrs which will be modified and used to obtain
3952 * pools.
3953 */
3954 copy_workqueue_attrs(tmp_attrs, new_attrs);
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003955
Tejun Heo4c16bd32013-04-01 11:23:36 -07003956 /*
3957 * CPUs should stay stable across pwq creations and installations.
3958 * Pin CPUs, determine the target cpumask for each node and create
3959 * pwqs accordingly.
3960 */
3961 get_online_cpus();
3962
3963 mutex_lock(&wq_pool_mutex);
3964
3965 /*
3966 * If something goes wrong during CPU up/down, we'll fall back to
3967 * the default pwq covering whole @attrs->cpumask. Always create
3968 * it even if we don't use it immediately.
3969 */
3970 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3971 if (!dfl_pwq)
3972 goto enomem_pwq;
3973
3974 for_each_node(node) {
3975 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3976 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3977 if (!pwq_tbl[node])
3978 goto enomem_pwq;
3979 } else {
3980 dfl_pwq->refcnt++;
3981 pwq_tbl[node] = dfl_pwq;
3982 }
3983 }
3984
3985 mutex_unlock(&wq_pool_mutex);
3986
3987 /* all pwqs have been created successfully, let's install'em */
Tejun Heof147f292013-04-01 11:23:35 -07003988 mutex_lock(&wq->mutex);
3989
Tejun Heof147f292013-04-01 11:23:35 -07003990 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
Tejun Heo4c16bd32013-04-01 11:23:36 -07003991
3992 /* save the previous pwq and install the new one */
Tejun Heof147f292013-04-01 11:23:35 -07003993 for_each_node(node)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003994 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3995
3996 /* @dfl_pwq might not have been used, ensure it's linked */
3997 link_pwq(dfl_pwq);
3998 swap(wq->dfl_pwq, dfl_pwq);
Tejun Heof147f292013-04-01 11:23:35 -07003999
4000 mutex_unlock(&wq->mutex);
4001
Tejun Heo4c16bd32013-04-01 11:23:36 -07004002 /* put the old pwqs */
4003 for_each_node(node)
4004 put_pwq_unlocked(pwq_tbl[node]);
4005 put_pwq_unlocked(dfl_pwq);
4006
4007 put_online_cpus();
Tejun Heo48621252013-04-01 11:23:31 -07004008 ret = 0;
4009 /* fall through */
4010out_free:
Tejun Heo4c16bd32013-04-01 11:23:36 -07004011 free_workqueue_attrs(tmp_attrs);
Tejun Heo48621252013-04-01 11:23:31 -07004012 free_workqueue_attrs(new_attrs);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004013 kfree(pwq_tbl);
Tejun Heo48621252013-04-01 11:23:31 -07004014 return ret;
Tejun Heo13e2e552013-04-01 11:23:31 -07004015
Tejun Heo4c16bd32013-04-01 11:23:36 -07004016enomem_pwq:
4017 free_unbound_pwq(dfl_pwq);
4018 for_each_node(node)
4019 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
4020 free_unbound_pwq(pwq_tbl[node]);
4021 mutex_unlock(&wq_pool_mutex);
4022 put_online_cpus();
Tejun Heo13e2e552013-04-01 11:23:31 -07004023enomem:
Tejun Heo48621252013-04-01 11:23:31 -07004024 ret = -ENOMEM;
4025 goto out_free;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004026}
4027
Tejun Heo4c16bd32013-04-01 11:23:36 -07004028/**
4029 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4030 * @wq: the target workqueue
4031 * @cpu: the CPU coming up or going down
4032 * @online: whether @cpu is coming up or going down
4033 *
4034 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4035 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4036 * @wq accordingly.
4037 *
4038 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4039 * falls back to @wq->dfl_pwq which may not be optimal but is always
4040 * correct.
4041 *
4042 * Note that when the last allowed CPU of a NUMA node goes offline for a
4043 * workqueue with a cpumask spanning multiple nodes, the workers which were
4044 * already executing the work items for the workqueue will lose their CPU
4045 * affinity and may execute on any CPU. This is similar to how per-cpu
4046 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4047 * affinity, it's the user's responsibility to flush the work item from
4048 * CPU_DOWN_PREPARE.
4049 */
4050static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4051 bool online)
4052{
4053 int node = cpu_to_node(cpu);
4054 int cpu_off = online ? -1 : cpu;
4055 struct pool_workqueue *old_pwq = NULL, *pwq;
4056 struct workqueue_attrs *target_attrs;
4057 cpumask_t *cpumask;
4058
4059 lockdep_assert_held(&wq_pool_mutex);
4060
4061 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4062 return;
4063
4064 /*
4065 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4066 * Let's use a preallocated one. The following buf is protected by
4067 * CPU hotplug exclusion.
4068 */
4069 target_attrs = wq_update_unbound_numa_attrs_buf;
4070 cpumask = target_attrs->cpumask;
4071
4072 mutex_lock(&wq->mutex);
Tejun Heod55262c2013-04-01 11:23:38 -07004073 if (wq->unbound_attrs->no_numa)
4074 goto out_unlock;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004075
4076 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4077 pwq = unbound_pwq_by_node(wq, node);
4078
4079 /*
4080 * Let's determine what needs to be done. If the target cpumask is
4081 * different from wq's, we need to compare it to @pwq's and create
4082 * a new one if they don't match. If the target cpumask equals
4083 * wq's, the default pwq should be used. If @pwq is already the
4084 * default one, nothing to do; otherwise, install the default one.
4085 */
4086 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4087 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4088 goto out_unlock;
4089 } else {
4090 if (pwq == wq->dfl_pwq)
4091 goto out_unlock;
4092 else
4093 goto use_dfl_pwq;
4094 }
4095
4096 mutex_unlock(&wq->mutex);
4097
4098 /* create a new pwq */
4099 pwq = alloc_unbound_pwq(wq, target_attrs);
4100 if (!pwq) {
4101 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4102 wq->name);
4103 goto out_unlock;
4104 }
4105
4106 /*
4107 * Install the new pwq. As this function is called only from CPU
4108 * hotplug callbacks and applying a new attrs is wrapped with
4109 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4110 * inbetween.
4111 */
4112 mutex_lock(&wq->mutex);
4113 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4114 goto out_unlock;
4115
4116use_dfl_pwq:
4117 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4118 get_pwq(wq->dfl_pwq);
4119 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4120 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4121out_unlock:
4122 mutex_unlock(&wq->mutex);
4123 put_pwq_unlocked(old_pwq);
4124}
4125
Tejun Heo30cdf242013-03-12 11:29:57 -07004126static int alloc_and_link_pwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004128 bool highpri = wq->flags & WQ_HIGHPRI;
Tejun Heo8a2b7532013-09-05 12:30:04 -04004129 int cpu, ret;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01004130
Tejun Heo30cdf242013-03-12 11:29:57 -07004131 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo420c0dd2013-03-12 11:29:59 -07004132 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4133 if (!wq->cpu_pwqs)
Tejun Heo30cdf242013-03-12 11:29:57 -07004134 return -ENOMEM;
4135
4136 for_each_possible_cpu(cpu) {
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004137 struct pool_workqueue *pwq =
4138 per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heo7a62c2c2013-03-12 11:30:03 -07004139 struct worker_pool *cpu_pools =
Tejun Heof02ae732013-03-12 11:30:03 -07004140 per_cpu(cpu_worker_pools, cpu);
Tejun Heo30cdf242013-03-12 11:29:57 -07004141
Tejun Heof147f292013-04-01 11:23:35 -07004142 init_pwq(pwq, wq, &cpu_pools[highpri]);
4143
4144 mutex_lock(&wq->mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07004145 link_pwq(pwq);
Tejun Heof147f292013-04-01 11:23:35 -07004146 mutex_unlock(&wq->mutex);
Tejun Heo30cdf242013-03-12 11:29:57 -07004147 }
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004148 return 0;
Tejun Heo8a2b7532013-09-05 12:30:04 -04004149 } else if (wq->flags & __WQ_ORDERED) {
4150 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4151 /* there should only be single pwq for ordering guarantee */
4152 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4153 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4154 "ordering guarantee broken for workqueue %s\n", wq->name);
4155 return ret;
Tejun Heo30cdf242013-03-12 11:29:57 -07004156 } else {
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004157 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
Tejun Heo30cdf242013-03-12 11:29:57 -07004158 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004159}
4160
Tejun Heof3421792010-07-02 10:03:51 +02004161static int wq_clamp_max_active(int max_active, unsigned int flags,
4162 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02004163{
Tejun Heof3421792010-07-02 10:03:51 +02004164 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4165
4166 if (max_active < 1 || max_active > lim)
Valentin Ilie044c7822012-08-19 00:52:42 +03004167 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4168 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02004169
Tejun Heof3421792010-07-02 10:03:51 +02004170 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02004171}
4172
Tejun Heob196be82012-01-10 15:11:35 -08004173struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
Tejun Heod320c032010-06-29 10:07:14 +02004174 unsigned int flags,
4175 int max_active,
4176 struct lock_class_key *key,
Tejun Heob196be82012-01-10 15:11:35 -08004177 const char *lock_name, ...)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004178{
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004179 size_t tbl_size = 0;
Tejun Heoecf68812013-04-01 11:23:34 -07004180 va_list args;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004181 struct workqueue_struct *wq;
Tejun Heo49e3cf42013-03-12 11:29:58 -07004182 struct pool_workqueue *pwq;
Tejun Heob196be82012-01-10 15:11:35 -08004183
Viresh Kumarcee22a12013-04-08 16:45:40 +05304184 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4185 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4186 flags |= WQ_UNBOUND;
4187
Tejun Heoecf68812013-04-01 11:23:34 -07004188 /* allocate wq and format name */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004189 if (flags & WQ_UNBOUND)
4190 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4191
4192 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
Tejun Heob196be82012-01-10 15:11:35 -08004193 if (!wq)
Tejun Heod2c1d402013-03-12 11:30:04 -07004194 return NULL;
Tejun Heob196be82012-01-10 15:11:35 -08004195
Tejun Heo6029a912013-04-01 11:23:34 -07004196 if (flags & WQ_UNBOUND) {
4197 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4198 if (!wq->unbound_attrs)
4199 goto err_free_wq;
4200 }
4201
Tejun Heoecf68812013-04-01 11:23:34 -07004202 va_start(args, lock_name);
4203 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
Tejun Heob196be82012-01-10 15:11:35 -08004204 va_end(args);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004205
Tejun Heod320c032010-06-29 10:07:14 +02004206 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heob196be82012-01-10 15:11:35 -08004207 max_active = wq_clamp_max_active(max_active, flags, wq->name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004208
Tejun Heob196be82012-01-10 15:11:35 -08004209 /* init wq */
Tejun Heo97e37d72010-06-29 10:07:10 +02004210 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004211 wq->saved_max_active = max_active;
Lai Jiangshan3c25a552013-03-25 16:57:17 -07004212 mutex_init(&wq->mutex);
Tejun Heo112202d2013-02-13 19:29:12 -08004213 atomic_set(&wq->nr_pwqs_to_flush, 0);
Tejun Heo30cdf242013-03-12 11:29:57 -07004214 INIT_LIST_HEAD(&wq->pwqs);
Tejun Heo73f53c42010-06-29 10:07:11 +02004215 INIT_LIST_HEAD(&wq->flusher_queue);
4216 INIT_LIST_HEAD(&wq->flusher_overflow);
Tejun Heo493a1722013-03-12 11:29:59 -07004217 INIT_LIST_HEAD(&wq->maydays);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004218
Johannes Bergeb13ba82008-01-16 09:51:58 +01004219 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07004220 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004221
Tejun Heo30cdf242013-03-12 11:29:57 -07004222 if (alloc_and_link_pwqs(wq) < 0)
Tejun Heod2c1d402013-03-12 11:30:04 -07004223 goto err_free_wq;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004224
Tejun Heo493008a2013-03-12 11:30:03 -07004225 /*
4226 * Workqueues which may be used during memory reclaim should
4227 * have a rescuer to guarantee forward progress.
4228 */
4229 if (flags & WQ_MEM_RECLAIM) {
Tejun Heoe22bee72010-06-29 10:07:14 +02004230 struct worker *rescuer;
4231
Tejun Heod2c1d402013-03-12 11:30:04 -07004232 rescuer = alloc_worker();
Tejun Heoe22bee72010-06-29 10:07:14 +02004233 if (!rescuer)
Tejun Heod2c1d402013-03-12 11:30:04 -07004234 goto err_destroy;
Tejun Heoe22bee72010-06-29 10:07:14 +02004235
Tejun Heo111c2252013-01-17 17:16:24 -08004236 rescuer->rescue_wq = wq;
4237 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
Tejun Heob196be82012-01-10 15:11:35 -08004238 wq->name);
Tejun Heod2c1d402013-03-12 11:30:04 -07004239 if (IS_ERR(rescuer->task)) {
4240 kfree(rescuer);
4241 goto err_destroy;
4242 }
Tejun Heoe22bee72010-06-29 10:07:14 +02004243
Tejun Heod2c1d402013-03-12 11:30:04 -07004244 wq->rescuer = rescuer;
Tejun Heo14a40ff2013-03-19 13:45:20 -07004245 rescuer->task->flags |= PF_NO_SETAFFINITY;
Tejun Heoe22bee72010-06-29 10:07:14 +02004246 wake_up_process(rescuer->task);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004247 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004248
Tejun Heo226223a2013-03-12 11:30:05 -07004249 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4250 goto err_destroy;
4251
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004252 /*
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004253 * wq_pool_mutex protects global freeze state and workqueues list.
4254 * Grab it, adjust max_active and add the new @wq to workqueues
4255 * list.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004256 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004257 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004258
Lai Jiangshana357fc02013-03-25 16:57:19 -07004259 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004260 for_each_pwq(pwq, wq)
4261 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004262 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004263
Tejun Heo15376632010-06-29 10:07:11 +02004264 list_add(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004265
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004266 mutex_unlock(&wq_pool_mutex);
Tejun Heo15376632010-06-29 10:07:11 +02004267
Oleg Nesterov3af244332007-05-09 02:34:09 -07004268 return wq;
Tejun Heod2c1d402013-03-12 11:30:04 -07004269
4270err_free_wq:
Tejun Heo6029a912013-04-01 11:23:34 -07004271 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heod2c1d402013-03-12 11:30:04 -07004272 kfree(wq);
4273 return NULL;
4274err_destroy:
4275 destroy_workqueue(wq);
Tejun Heo4690c4a2010-06-29 10:07:10 +02004276 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004277}
Tejun Heod320c032010-06-29 10:07:14 +02004278EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004279
4280/**
4281 * destroy_workqueue - safely terminate a workqueue
4282 * @wq: target workqueue
4283 *
4284 * Safely destroy a workqueue. All work currently pending will be done first.
4285 */
4286void destroy_workqueue(struct workqueue_struct *wq)
4287{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004288 struct pool_workqueue *pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004289 int node;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004290
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02004291 /* drain it before proceeding with destruction */
4292 drain_workqueue(wq);
Tejun Heoc8efcc22010-12-20 19:32:04 +01004293
Tejun Heo6183c002013-03-12 11:29:57 -07004294 /* sanity checks */
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004295 mutex_lock(&wq->mutex);
Tejun Heo49e3cf42013-03-12 11:29:58 -07004296 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004297 int i;
4298
Tejun Heo76af4d92013-03-12 11:30:00 -07004299 for (i = 0; i < WORK_NR_COLORS; i++) {
4300 if (WARN_ON(pwq->nr_in_flight[i])) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004301 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004302 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004303 }
4304 }
4305
Lai Jiangshan5c529592013-04-04 10:05:38 +08004306 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
Tejun Heo8864b4e2013-03-12 11:30:04 -07004307 WARN_ON(pwq->nr_active) ||
Tejun Heo76af4d92013-03-12 11:30:00 -07004308 WARN_ON(!list_empty(&pwq->delayed_works))) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004309 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004310 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004311 }
Tejun Heo6183c002013-03-12 11:29:57 -07004312 }
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004313 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004314
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004315 /*
4316 * wq list is used to freeze wq, remove from list after
4317 * flushing is complete in case freeze races us.
4318 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004319 mutex_lock(&wq_pool_mutex);
Tejun Heod2c1d402013-03-12 11:30:04 -07004320 list_del_init(&wq->list);
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004321 mutex_unlock(&wq_pool_mutex);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004322
Tejun Heo226223a2013-03-12 11:30:05 -07004323 workqueue_sysfs_unregister(wq);
4324
Tejun Heo493008a2013-03-12 11:30:03 -07004325 if (wq->rescuer) {
Tejun Heoe22bee72010-06-29 10:07:14 +02004326 kthread_stop(wq->rescuer->task);
Xiaotian Feng8d9df9f2010-08-16 09:54:28 +02004327 kfree(wq->rescuer);
Tejun Heo493008a2013-03-12 11:30:03 -07004328 wq->rescuer = NULL;
Tejun Heoe22bee72010-06-29 10:07:14 +02004329 }
4330
Tejun Heo8864b4e2013-03-12 11:30:04 -07004331 if (!(wq->flags & WQ_UNBOUND)) {
4332 /*
4333 * The base ref is never dropped on per-cpu pwqs. Directly
4334 * free the pwqs and wq.
4335 */
4336 free_percpu(wq->cpu_pwqs);
4337 kfree(wq);
4338 } else {
4339 /*
4340 * We're the sole accessor of @wq at this point. Directly
Tejun Heo4c16bd32013-04-01 11:23:36 -07004341 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4342 * @wq will be freed when the last pwq is released.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004343 */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004344 for_each_node(node) {
4345 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4346 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4347 put_pwq_unlocked(pwq);
4348 }
4349
4350 /*
4351 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4352 * put. Don't access it afterwards.
4353 */
4354 pwq = wq->dfl_pwq;
4355 wq->dfl_pwq = NULL;
Tejun Heodce90d42013-04-01 11:23:35 -07004356 put_pwq_unlocked(pwq);
Tejun Heo29c91e92013-03-12 11:30:03 -07004357 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004358}
4359EXPORT_SYMBOL_GPL(destroy_workqueue);
4360
Tejun Heodcd989c2010-06-29 10:07:14 +02004361/**
4362 * workqueue_set_max_active - adjust max_active of a workqueue
4363 * @wq: target workqueue
4364 * @max_active: new max_active value.
4365 *
4366 * Set max_active of @wq to @max_active.
4367 *
4368 * CONTEXT:
4369 * Don't call from IRQ context.
4370 */
4371void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4372{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004373 struct pool_workqueue *pwq;
Tejun Heodcd989c2010-06-29 10:07:14 +02004374
Tejun Heo8719dce2013-03-12 11:30:04 -07004375 /* disallow meddling with max_active for ordered workqueues */
4376 if (WARN_ON(wq->flags & __WQ_ORDERED))
4377 return;
4378
Tejun Heof3421792010-07-02 10:03:51 +02004379 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02004380
Lai Jiangshana357fc02013-03-25 16:57:19 -07004381 mutex_lock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004382
4383 wq->saved_max_active = max_active;
4384
Tejun Heo699ce092013-03-13 16:51:35 -07004385 for_each_pwq(pwq, wq)
4386 pwq_adjust_max_active(pwq);
Tejun Heodcd989c2010-06-29 10:07:14 +02004387
Lai Jiangshana357fc02013-03-25 16:57:19 -07004388 mutex_unlock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004389}
4390EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4391
4392/**
Tejun Heoe62676162013-03-12 17:41:37 -07004393 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4394 *
4395 * Determine whether %current is a workqueue rescuer. Can be used from
4396 * work functions to determine whether it's being run off the rescuer task.
Yacine Belkadid185af32013-07-31 14:59:24 -07004397 *
4398 * Return: %true if %current is a workqueue rescuer. %false otherwise.
Tejun Heoe62676162013-03-12 17:41:37 -07004399 */
4400bool current_is_workqueue_rescuer(void)
4401{
4402 struct worker *worker = current_wq_worker();
4403
Lai Jiangshan6a092df2013-03-20 03:28:03 +08004404 return worker && worker->rescue_wq;
Tejun Heoe62676162013-03-12 17:41:37 -07004405}
4406
4407/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004408 * workqueue_congested - test whether a workqueue is congested
4409 * @cpu: CPU in question
4410 * @wq: target workqueue
4411 *
4412 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4413 * no synchronization around this function and the test result is
4414 * unreliable and only useful as advisory hints or for debugging.
4415 *
Tejun Heod3251852013-05-10 11:10:17 -07004416 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4417 * Note that both per-cpu and unbound workqueues may be associated with
4418 * multiple pool_workqueues which have separate congested states. A
4419 * workqueue being congested on one CPU doesn't mean the workqueue is also
4420 * contested on other CPUs / NUMA nodes.
4421 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004422 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004423 * %true if congested, %false otherwise.
4424 */
Tejun Heod84ff052013-03-12 11:29:59 -07004425bool workqueue_congested(int cpu, struct workqueue_struct *wq)
Tejun Heodcd989c2010-06-29 10:07:14 +02004426{
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004427 struct pool_workqueue *pwq;
Tejun Heo76af4d92013-03-12 11:30:00 -07004428 bool ret;
4429
Lai Jiangshan88109452013-03-20 03:28:10 +08004430 rcu_read_lock_sched();
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004431
Tejun Heod3251852013-05-10 11:10:17 -07004432 if (cpu == WORK_CPU_UNBOUND)
4433 cpu = smp_processor_id();
4434
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004435 if (!(wq->flags & WQ_UNBOUND))
4436 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4437 else
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004438 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodcd989c2010-06-29 10:07:14 +02004439
Tejun Heo76af4d92013-03-12 11:30:00 -07004440 ret = !list_empty(&pwq->delayed_works);
Lai Jiangshan88109452013-03-20 03:28:10 +08004441 rcu_read_unlock_sched();
Tejun Heo76af4d92013-03-12 11:30:00 -07004442
4443 return ret;
Tejun Heodcd989c2010-06-29 10:07:14 +02004444}
4445EXPORT_SYMBOL_GPL(workqueue_congested);
4446
4447/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004448 * work_busy - test whether a work is currently pending or running
4449 * @work: the work to be tested
4450 *
4451 * Test whether @work is currently pending or running. There is no
4452 * synchronization around this function and the test result is
4453 * unreliable and only useful as advisory hints or for debugging.
Tejun Heodcd989c2010-06-29 10:07:14 +02004454 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004455 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004456 * OR'd bitmask of WORK_BUSY_* bits.
4457 */
4458unsigned int work_busy(struct work_struct *work)
4459{
Tejun Heofa1b54e2013-03-12 11:30:00 -07004460 struct worker_pool *pool;
Tejun Heodcd989c2010-06-29 10:07:14 +02004461 unsigned long flags;
4462 unsigned int ret = 0;
4463
Tejun Heodcd989c2010-06-29 10:07:14 +02004464 if (work_pending(work))
4465 ret |= WORK_BUSY_PENDING;
Tejun Heodcd989c2010-06-29 10:07:14 +02004466
Tejun Heofa1b54e2013-03-12 11:30:00 -07004467 local_irq_save(flags);
4468 pool = get_work_pool(work);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004469 if (pool) {
Tejun Heofa1b54e2013-03-12 11:30:00 -07004470 spin_lock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004471 if (find_worker_executing_work(pool, work))
4472 ret |= WORK_BUSY_RUNNING;
Tejun Heofa1b54e2013-03-12 11:30:00 -07004473 spin_unlock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004474 }
Tejun Heofa1b54e2013-03-12 11:30:00 -07004475 local_irq_restore(flags);
Tejun Heodcd989c2010-06-29 10:07:14 +02004476
4477 return ret;
4478}
4479EXPORT_SYMBOL_GPL(work_busy);
4480
Tejun Heo3d1cb202013-04-30 15:27:22 -07004481/**
4482 * set_worker_desc - set description for the current work item
4483 * @fmt: printf-style format string
4484 * @...: arguments for the format string
4485 *
4486 * This function can be called by a running work function to describe what
4487 * the work item is about. If the worker task gets dumped, this
4488 * information will be printed out together to help debugging. The
4489 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4490 */
4491void set_worker_desc(const char *fmt, ...)
4492{
4493 struct worker *worker = current_wq_worker();
4494 va_list args;
4495
4496 if (worker) {
4497 va_start(args, fmt);
4498 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4499 va_end(args);
4500 worker->desc_valid = true;
4501 }
4502}
4503
4504/**
4505 * print_worker_info - print out worker information and description
4506 * @log_lvl: the log level to use when printing
4507 * @task: target task
4508 *
4509 * If @task is a worker and currently executing a work item, print out the
4510 * name of the workqueue being serviced and worker description set with
4511 * set_worker_desc() by the currently executing work item.
4512 *
4513 * This function can be safely called on any task as long as the
4514 * task_struct itself is accessible. While safe, this function isn't
4515 * synchronized and may print out mixups or garbages of limited length.
4516 */
4517void print_worker_info(const char *log_lvl, struct task_struct *task)
4518{
4519 work_func_t *fn = NULL;
4520 char name[WQ_NAME_LEN] = { };
4521 char desc[WORKER_DESC_LEN] = { };
4522 struct pool_workqueue *pwq = NULL;
4523 struct workqueue_struct *wq = NULL;
4524 bool desc_valid = false;
4525 struct worker *worker;
4526
4527 if (!(task->flags & PF_WQ_WORKER))
4528 return;
4529
4530 /*
4531 * This function is called without any synchronization and @task
4532 * could be in any state. Be careful with dereferences.
4533 */
4534 worker = probe_kthread_data(task);
4535
4536 /*
4537 * Carefully copy the associated workqueue's workfn and name. Keep
4538 * the original last '\0' in case the original contains garbage.
4539 */
4540 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4541 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4542 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4543 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4544
4545 /* copy worker description */
4546 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4547 if (desc_valid)
4548 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4549
4550 if (fn || name[0] || desc[0]) {
4551 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4552 if (desc[0])
4553 pr_cont(" (%s)", desc);
4554 pr_cont("\n");
4555 }
4556}
4557
Tejun Heodb7bccf2010-06-29 10:07:12 +02004558/*
4559 * CPU hotplug.
4560 *
Tejun Heoe22bee72010-06-29 10:07:14 +02004561 * There are two challenges in supporting CPU hotplug. Firstly, there
Tejun Heo112202d2013-02-13 19:29:12 -08004562 * are a lot of assumptions on strong associations among work, pwq and
Tejun Heo706026c2013-01-24 11:01:34 -08004563 * pool which make migrating pending and scheduled works very
Tejun Heoe22bee72010-06-29 10:07:14 +02004564 * difficult to implement without impacting hot paths. Secondly,
Tejun Heo94cf58b2013-01-24 11:01:33 -08004565 * worker pools serve mix of short, long and very long running works making
Tejun Heoe22bee72010-06-29 10:07:14 +02004566 * blocked draining impractical.
4567 *
Tejun Heo24647572013-01-24 11:01:33 -08004568 * This is solved by allowing the pools to be disassociated from the CPU
Tejun Heo628c78e2012-07-17 12:39:27 -07004569 * running as an unbound one and allowing it to be reattached later if the
4570 * cpu comes back online.
Tejun Heodb7bccf2010-06-29 10:07:12 +02004571 */
4572
Tejun Heo706026c2013-01-24 11:01:34 -08004573static void wq_unbind_fn(struct work_struct *work)
Tejun Heodb7bccf2010-06-29 10:07:12 +02004574{
Tejun Heo38db41d2013-01-24 11:01:34 -08004575 int cpu = smp_processor_id();
Tejun Heo4ce62e92012-07-13 22:16:44 -07004576 struct worker_pool *pool;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004577 struct worker *worker;
Tejun Heoa9ab7752013-03-19 13:45:21 -07004578 int wi;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004579
Tejun Heof02ae732013-03-12 11:30:03 -07004580 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo6183c002013-03-12 11:29:57 -07004581 WARN_ON_ONCE(cpu != smp_processor_id());
Tejun Heo94cf58b2013-01-24 11:01:33 -08004582
Tejun Heobc3a1af2013-03-13 19:47:39 -07004583 mutex_lock(&pool->manager_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004584 spin_lock_irq(&pool->lock);
4585
4586 /*
Tejun Heobc3a1af2013-03-13 19:47:39 -07004587 * We've blocked all manager operations. Make all workers
Tejun Heo94cf58b2013-01-24 11:01:33 -08004588 * unbound and set DISASSOCIATED. Before this, all workers
4589 * except for the ones which are still executing works from
4590 * before the last CPU down must be on the cpu. After
4591 * this, they may become diasporas.
4592 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004593 for_each_pool_worker(worker, wi, pool)
Tejun Heoc9e7cf22013-01-24 11:01:33 -08004594 worker->flags |= WORKER_UNBOUND;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004595
Tejun Heo24647572013-01-24 11:01:33 -08004596 pool->flags |= POOL_DISASSOCIATED;
Tejun Heof2d5a0e2012-07-17 12:39:26 -07004597
Tejun Heo94cf58b2013-01-24 11:01:33 -08004598 spin_unlock_irq(&pool->lock);
Tejun Heobc3a1af2013-03-13 19:47:39 -07004599 mutex_unlock(&pool->manager_mutex);
Tejun Heoe22bee72010-06-29 10:07:14 +02004600
Lai Jiangshaneb283422013-03-08 15:18:28 -08004601 /*
4602 * Call schedule() so that we cross rq->lock and thus can
4603 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4604 * This is necessary as scheduler callbacks may be invoked
4605 * from other cpus.
4606 */
4607 schedule();
Tejun Heo628c78e2012-07-17 12:39:27 -07004608
Lai Jiangshaneb283422013-03-08 15:18:28 -08004609 /*
4610 * Sched callbacks are disabled now. Zap nr_running.
4611 * After this, nr_running stays zero and need_more_worker()
4612 * and keep_working() are always true as long as the
4613 * worklist is not empty. This pool now behaves as an
4614 * unbound (in terms of concurrency management) pool which
4615 * are served by workers tied to the pool.
4616 */
Tejun Heoe19e3972013-01-24 11:39:44 -08004617 atomic_set(&pool->nr_running, 0);
Lai Jiangshaneb283422013-03-08 15:18:28 -08004618
4619 /*
4620 * With concurrency management just turned off, a busy
4621 * worker blocking could lead to lengthy stalls. Kick off
4622 * unbound chain execution of currently pending work items.
4623 */
4624 spin_lock_irq(&pool->lock);
4625 wake_up_worker(pool);
4626 spin_unlock_irq(&pool->lock);
4627 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004628}
4629
Tejun Heobd7c0892013-03-19 13:45:21 -07004630/**
4631 * rebind_workers - rebind all workers of a pool to the associated CPU
4632 * @pool: pool of interest
4633 *
Tejun Heoa9ab7752013-03-19 13:45:21 -07004634 * @pool->cpu is coming online. Rebind all workers to the CPU.
Tejun Heobd7c0892013-03-19 13:45:21 -07004635 */
4636static void rebind_workers(struct worker_pool *pool)
4637{
Tejun Heoa9ab7752013-03-19 13:45:21 -07004638 struct worker *worker;
4639 int wi;
Tejun Heobd7c0892013-03-19 13:45:21 -07004640
4641 lockdep_assert_held(&pool->manager_mutex);
Tejun Heobd7c0892013-03-19 13:45:21 -07004642
Tejun Heoa9ab7752013-03-19 13:45:21 -07004643 /*
4644 * Restore CPU affinity of all workers. As all idle workers should
4645 * be on the run-queue of the associated CPU before any local
4646 * wake-ups for concurrency management happen, restore CPU affinty
4647 * of all workers first and then clear UNBOUND. As we're called
4648 * from CPU_ONLINE, the following shouldn't fail.
4649 */
4650 for_each_pool_worker(worker, wi, pool)
4651 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4652 pool->attrs->cpumask) < 0);
4653
4654 spin_lock_irq(&pool->lock);
4655
4656 for_each_pool_worker(worker, wi, pool) {
4657 unsigned int worker_flags = worker->flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004658
4659 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07004660 * A bound idle worker should actually be on the runqueue
4661 * of the associated CPU for local wake-ups targeting it to
4662 * work. Kick all idle workers so that they migrate to the
4663 * associated CPU. Doing this in the same loop as
4664 * replacing UNBOUND with REBOUND is safe as no worker will
4665 * be bound before @pool->lock is released.
Tejun Heobd7c0892013-03-19 13:45:21 -07004666 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004667 if (worker_flags & WORKER_IDLE)
4668 wake_up_process(worker->task);
4669
4670 /*
4671 * We want to clear UNBOUND but can't directly call
4672 * worker_clr_flags() or adjust nr_running. Atomically
4673 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4674 * @worker will clear REBOUND using worker_clr_flags() when
4675 * it initiates the next execution cycle thus restoring
4676 * concurrency management. Note that when or whether
4677 * @worker clears REBOUND doesn't affect correctness.
4678 *
4679 * ACCESS_ONCE() is necessary because @worker->flags may be
4680 * tested without holding any lock in
4681 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4682 * fail incorrectly leading to premature concurrency
4683 * management operations.
4684 */
4685 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4686 worker_flags |= WORKER_REBOUND;
4687 worker_flags &= ~WORKER_UNBOUND;
4688 ACCESS_ONCE(worker->flags) = worker_flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004689 }
4690
Tejun Heoa9ab7752013-03-19 13:45:21 -07004691 spin_unlock_irq(&pool->lock);
Tejun Heobd7c0892013-03-19 13:45:21 -07004692}
4693
Tejun Heo7dbc7252013-03-19 13:45:21 -07004694/**
4695 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4696 * @pool: unbound pool of interest
4697 * @cpu: the CPU which is coming up
4698 *
4699 * An unbound pool may end up with a cpumask which doesn't have any online
4700 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4701 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4702 * online CPU before, cpus_allowed of all its workers should be restored.
4703 */
4704static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4705{
4706 static cpumask_t cpumask;
4707 struct worker *worker;
4708 int wi;
4709
4710 lockdep_assert_held(&pool->manager_mutex);
4711
4712 /* is @cpu allowed for @pool? */
4713 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4714 return;
4715
4716 /* is @cpu the only online CPU? */
4717 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4718 if (cpumask_weight(&cpumask) != 1)
4719 return;
4720
4721 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4722 for_each_pool_worker(worker, wi, pool)
4723 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4724 pool->attrs->cpumask) < 0);
4725}
4726
Tejun Heo8db25e72012-07-17 12:39:28 -07004727/*
4728 * Workqueues should be brought up before normal priority CPU notifiers.
4729 * This will be registered high priority CPU notifier.
4730 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04004731static int workqueue_cpu_up_callback(struct notifier_block *nfb,
Tejun Heo8db25e72012-07-17 12:39:28 -07004732 unsigned long action,
4733 void *hcpu)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004734{
Tejun Heod84ff052013-03-12 11:29:59 -07004735 int cpu = (unsigned long)hcpu;
Tejun Heo4ce62e92012-07-13 22:16:44 -07004736 struct worker_pool *pool;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004737 struct workqueue_struct *wq;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004738 int pi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004739
Tejun Heo8db25e72012-07-17 12:39:28 -07004740 switch (action & ~CPU_TASKS_FROZEN) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004741 case CPU_UP_PREPARE:
Tejun Heof02ae732013-03-12 11:30:03 -07004742 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo3ce63372012-07-17 12:39:27 -07004743 if (pool->nr_workers)
4744 continue;
Tejun Heoebf44d12013-03-13 19:47:39 -07004745 if (create_and_start_worker(pool) < 0)
Tejun Heo3ce63372012-07-17 12:39:27 -07004746 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004747 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004748 break;
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07004749
Tejun Heo65758202012-07-17 12:39:26 -07004750 case CPU_DOWN_FAILED:
4751 case CPU_ONLINE:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004752 mutex_lock(&wq_pool_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004753
4754 for_each_pool(pool, pi) {
Tejun Heobc3a1af2013-03-13 19:47:39 -07004755 mutex_lock(&pool->manager_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004756
Tejun Heo7dbc7252013-03-19 13:45:21 -07004757 if (pool->cpu == cpu) {
4758 spin_lock_irq(&pool->lock);
4759 pool->flags &= ~POOL_DISASSOCIATED;
4760 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07004761
Tejun Heo7dbc7252013-03-19 13:45:21 -07004762 rebind_workers(pool);
4763 } else if (pool->cpu < 0) {
4764 restore_unbound_workers_cpumask(pool, cpu);
4765 }
Tejun Heo94cf58b2013-01-24 11:01:33 -08004766
Tejun Heobc3a1af2013-03-13 19:47:39 -07004767 mutex_unlock(&pool->manager_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004768 }
Tejun Heo7dbc7252013-03-19 13:45:21 -07004769
Tejun Heo4c16bd32013-04-01 11:23:36 -07004770 /* update NUMA affinity of unbound workqueues */
4771 list_for_each_entry(wq, &workqueues, list)
4772 wq_update_unbound_numa(wq, cpu, true);
4773
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004774 mutex_unlock(&wq_pool_mutex);
Tejun Heo8db25e72012-07-17 12:39:28 -07004775 break;
Tejun Heo65758202012-07-17 12:39:26 -07004776 }
4777 return NOTIFY_OK;
4778}
4779
4780/*
4781 * Workqueues should be brought down after normal priority CPU notifiers.
4782 * This will be registered as low priority CPU notifier.
4783 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04004784static int workqueue_cpu_down_callback(struct notifier_block *nfb,
Tejun Heo65758202012-07-17 12:39:26 -07004785 unsigned long action,
4786 void *hcpu)
4787{
Tejun Heod84ff052013-03-12 11:29:59 -07004788 int cpu = (unsigned long)hcpu;
Tejun Heo8db25e72012-07-17 12:39:28 -07004789 struct work_struct unbind_work;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004790 struct workqueue_struct *wq;
Tejun Heo8db25e72012-07-17 12:39:28 -07004791
Tejun Heo65758202012-07-17 12:39:26 -07004792 switch (action & ~CPU_TASKS_FROZEN) {
4793 case CPU_DOWN_PREPARE:
Tejun Heo4c16bd32013-04-01 11:23:36 -07004794 /* unbinding per-cpu workers should happen on the local CPU */
Tejun Heo706026c2013-01-24 11:01:34 -08004795 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
Joonsoo Kim7635d2f2012-08-15 23:25:41 +09004796 queue_work_on(cpu, system_highpri_wq, &unbind_work);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004797
4798 /* update NUMA affinity of unbound workqueues */
4799 mutex_lock(&wq_pool_mutex);
4800 list_for_each_entry(wq, &workqueues, list)
4801 wq_update_unbound_numa(wq, cpu, false);
4802 mutex_unlock(&wq_pool_mutex);
4803
4804 /* wait for per-cpu unbinding to finish */
Tejun Heo8db25e72012-07-17 12:39:28 -07004805 flush_work(&unbind_work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05004806 destroy_work_on_stack(&unbind_work);
Tejun Heo8db25e72012-07-17 12:39:28 -07004807 break;
Tejun Heo65758202012-07-17 12:39:26 -07004808 }
4809 return NOTIFY_OK;
4810}
4811
Rusty Russell2d3854a2008-11-05 13:39:10 +11004812#ifdef CONFIG_SMP
Rusty Russell8ccad402009-01-16 15:31:15 -08004813
Rusty Russell2d3854a2008-11-05 13:39:10 +11004814struct work_for_cpu {
Tejun Heoed48ece2012-09-18 12:48:43 -07004815 struct work_struct work;
Rusty Russell2d3854a2008-11-05 13:39:10 +11004816 long (*fn)(void *);
4817 void *arg;
4818 long ret;
4819};
4820
Tejun Heoed48ece2012-09-18 12:48:43 -07004821static void work_for_cpu_fn(struct work_struct *work)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004822{
Tejun Heoed48ece2012-09-18 12:48:43 -07004823 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4824
Rusty Russell2d3854a2008-11-05 13:39:10 +11004825 wfc->ret = wfc->fn(wfc->arg);
4826}
4827
4828/**
4829 * work_on_cpu - run a function in user context on a particular cpu
4830 * @cpu: the cpu to run on
4831 * @fn: the function to run
4832 * @arg: the function arg
4833 *
Rusty Russell31ad9082009-01-16 15:31:15 -08004834 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06004835 * The caller must not hold any locks which would prevent @fn from completing.
Yacine Belkadid185af32013-07-31 14:59:24 -07004836 *
4837 * Return: The value @fn returns.
Rusty Russell2d3854a2008-11-05 13:39:10 +11004838 */
Tejun Heod84ff052013-03-12 11:29:59 -07004839long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004840{
Tejun Heoed48ece2012-09-18 12:48:43 -07004841 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
Rusty Russell2d3854a2008-11-05 13:39:10 +11004842
Tejun Heoed48ece2012-09-18 12:48:43 -07004843 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4844 schedule_work_on(cpu, &wfc.work);
Bjorn Helgaas12997d12013-11-18 11:00:29 -07004845 flush_work(&wfc.work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05004846 destroy_work_on_stack(&wfc.work);
Rusty Russell2d3854a2008-11-05 13:39:10 +11004847 return wfc.ret;
4848}
4849EXPORT_SYMBOL_GPL(work_on_cpu);
4850#endif /* CONFIG_SMP */
4851
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004852#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10304853
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004854/**
4855 * freeze_workqueues_begin - begin freezing workqueues
4856 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01004857 * Start freezing workqueues. After this function returns, all freezable
Tejun Heoc5aa87b2013-03-13 16:51:36 -07004858 * workqueues will queue new works to their delayed_works list instead of
Tejun Heo706026c2013-01-24 11:01:34 -08004859 * pool->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004860 *
4861 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004862 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004863 */
4864void freeze_workqueues_begin(void)
4865{
Tejun Heo17116962013-03-12 11:29:58 -07004866 struct worker_pool *pool;
Tejun Heo24b8a842013-03-12 11:29:58 -07004867 struct workqueue_struct *wq;
4868 struct pool_workqueue *pwq;
Tejun Heo611c92a2013-03-13 16:51:36 -07004869 int pi;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004870
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004871 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004872
Tejun Heo6183c002013-03-12 11:29:57 -07004873 WARN_ON_ONCE(workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004874 workqueue_freezing = true;
4875
Tejun Heo24b8a842013-03-12 11:29:58 -07004876 /* set FREEZING */
Tejun Heo611c92a2013-03-13 16:51:36 -07004877 for_each_pool(pool, pi) {
Tejun Heo5bcab332013-03-13 19:47:40 -07004878 spin_lock_irq(&pool->lock);
Tejun Heo17116962013-03-12 11:29:58 -07004879 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4880 pool->flags |= POOL_FREEZING;
Tejun Heo5bcab332013-03-13 19:47:40 -07004881 spin_unlock_irq(&pool->lock);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004882 }
4883
Tejun Heo24b8a842013-03-12 11:29:58 -07004884 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004885 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004886 for_each_pwq(pwq, wq)
4887 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004888 mutex_unlock(&wq->mutex);
Tejun Heo24b8a842013-03-12 11:29:58 -07004889 }
Tejun Heo5bcab332013-03-13 19:47:40 -07004890
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004891 mutex_unlock(&wq_pool_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004892}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004893
4894/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01004895 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004896 *
4897 * Check whether freezing is complete. This function must be called
4898 * between freeze_workqueues_begin() and thaw_workqueues().
4899 *
4900 * CONTEXT:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004901 * Grabs and releases wq_pool_mutex.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004902 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004903 * Return:
Tejun Heo58a69cb2011-02-16 09:25:31 +01004904 * %true if some freezable workqueues are still busy. %false if freezing
4905 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004906 */
4907bool freeze_workqueues_busy(void)
4908{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004909 bool busy = false;
Tejun Heo24b8a842013-03-12 11:29:58 -07004910 struct workqueue_struct *wq;
4911 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004912
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004913 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004914
Tejun Heo6183c002013-03-12 11:29:57 -07004915 WARN_ON_ONCE(!workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004916
Tejun Heo24b8a842013-03-12 11:29:58 -07004917 list_for_each_entry(wq, &workqueues, list) {
4918 if (!(wq->flags & WQ_FREEZABLE))
4919 continue;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004920 /*
4921 * nr_active is monotonically decreasing. It's safe
4922 * to peek without lock.
4923 */
Lai Jiangshan88109452013-03-20 03:28:10 +08004924 rcu_read_lock_sched();
Tejun Heo24b8a842013-03-12 11:29:58 -07004925 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004926 WARN_ON_ONCE(pwq->nr_active < 0);
Tejun Heo112202d2013-02-13 19:29:12 -08004927 if (pwq->nr_active) {
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004928 busy = true;
Lai Jiangshan88109452013-03-20 03:28:10 +08004929 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004930 goto out_unlock;
4931 }
4932 }
Lai Jiangshan88109452013-03-20 03:28:10 +08004933 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004934 }
4935out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004936 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004937 return busy;
4938}
4939
4940/**
4941 * thaw_workqueues - thaw workqueues
4942 *
4943 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo706026c2013-01-24 11:01:34 -08004944 * frozen works are transferred to their respective pool worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004945 *
4946 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004947 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004948 */
4949void thaw_workqueues(void)
4950{
Tejun Heo24b8a842013-03-12 11:29:58 -07004951 struct workqueue_struct *wq;
4952 struct pool_workqueue *pwq;
4953 struct worker_pool *pool;
Tejun Heo611c92a2013-03-13 16:51:36 -07004954 int pi;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004955
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004956 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004957
4958 if (!workqueue_freezing)
4959 goto out_unlock;
4960
Tejun Heo24b8a842013-03-12 11:29:58 -07004961 /* clear FREEZING */
Tejun Heo611c92a2013-03-13 16:51:36 -07004962 for_each_pool(pool, pi) {
Tejun Heo5bcab332013-03-13 19:47:40 -07004963 spin_lock_irq(&pool->lock);
Tejun Heo24b8a842013-03-12 11:29:58 -07004964 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4965 pool->flags &= ~POOL_FREEZING;
Tejun Heo5bcab332013-03-13 19:47:40 -07004966 spin_unlock_irq(&pool->lock);
Tejun Heo24b8a842013-03-12 11:29:58 -07004967 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02004968
Tejun Heo24b8a842013-03-12 11:29:58 -07004969 /* restore max_active and repopulate worklist */
4970 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004971 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004972 for_each_pwq(pwq, wq)
4973 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004974 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004975 }
4976
4977 workqueue_freezing = false;
4978out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004979 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004980}
4981#endif /* CONFIG_FREEZER */
4982
Tejun Heobce90382013-04-01 11:23:32 -07004983static void __init wq_numa_init(void)
4984{
4985 cpumask_var_t *tbl;
4986 int node, cpu;
4987
4988 /* determine NUMA pwq table len - highest node id + 1 */
4989 for_each_node(node)
4990 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4991
4992 if (num_possible_nodes() <= 1)
4993 return;
4994
Tejun Heod55262c2013-04-01 11:23:38 -07004995 if (wq_disable_numa) {
4996 pr_info("workqueue: NUMA affinity support disabled\n");
4997 return;
4998 }
4999
Tejun Heo4c16bd32013-04-01 11:23:36 -07005000 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5001 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5002
Tejun Heobce90382013-04-01 11:23:32 -07005003 /*
5004 * We want masks of possible CPUs of each node which isn't readily
5005 * available. Build one from cpu_to_node() which should have been
5006 * fully initialized by now.
5007 */
5008 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
5009 BUG_ON(!tbl);
5010
5011 for_each_node(node)
Tejun Heo1be0c252013-05-15 14:24:24 -07005012 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5013 node_online(node) ? node : NUMA_NO_NODE));
Tejun Heobce90382013-04-01 11:23:32 -07005014
5015 for_each_possible_cpu(cpu) {
5016 node = cpu_to_node(cpu);
5017 if (WARN_ON(node == NUMA_NO_NODE)) {
5018 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5019 /* happens iff arch is bonkers, let's just proceed */
5020 return;
5021 }
5022 cpumask_set_cpu(cpu, tbl[node]);
5023 }
5024
5025 wq_numa_possible_cpumask = tbl;
5026 wq_numa_enabled = true;
5027}
5028
Suresh Siddha6ee05782010-07-30 14:57:37 -07005029static int __init init_workqueues(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005030{
Tejun Heo7a4e3442013-03-12 11:30:00 -07005031 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5032 int i, cpu;
Tejun Heoc34056a2010-06-29 10:07:11 +02005033
Tejun Heoe904e6c2013-03-12 11:29:57 -07005034 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5035
5036 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5037
Tejun Heo65758202012-07-17 12:39:26 -07005038 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
Lai Jiangshana5b4e572012-09-18 09:59:23 -07005039 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
Tejun Heo8b03ae32010-06-29 10:07:12 +02005040
Tejun Heobce90382013-04-01 11:23:32 -07005041 wq_numa_init();
5042
Tejun Heo706026c2013-01-24 11:01:34 -08005043 /* initialize CPU pools */
Tejun Heo29c91e92013-03-12 11:30:03 -07005044 for_each_possible_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005045 struct worker_pool *pool;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005046
Tejun Heo7a4e3442013-03-12 11:30:00 -07005047 i = 0;
Tejun Heof02ae732013-03-12 11:30:03 -07005048 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo7a4e3442013-03-12 11:30:00 -07005049 BUG_ON(init_worker_pool(pool));
Tejun Heoec22ca52013-01-24 11:01:33 -08005050 pool->cpu = cpu;
Tejun Heo29c91e92013-03-12 11:30:03 -07005051 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
Tejun Heo7a4e3442013-03-12 11:30:00 -07005052 pool->attrs->nice = std_nice[i++];
Tejun Heof3f90ad2013-04-01 11:23:34 -07005053 pool->node = cpu_to_node(cpu);
Tejun Heo7a4e3442013-03-12 11:30:00 -07005054
Tejun Heo9daf9e62013-01-24 11:01:33 -08005055 /* alloc pool ID */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005056 mutex_lock(&wq_pool_mutex);
Tejun Heo9daf9e62013-01-24 11:01:33 -08005057 BUG_ON(worker_pool_assign_id(pool));
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005058 mutex_unlock(&wq_pool_mutex);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005059 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02005060 }
5061
Tejun Heoe22bee72010-06-29 10:07:14 +02005062 /* create the initial worker */
Tejun Heo29c91e92013-03-12 11:30:03 -07005063 for_each_online_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005064 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02005065
Tejun Heof02ae732013-03-12 11:30:03 -07005066 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo29c91e92013-03-12 11:30:03 -07005067 pool->flags &= ~POOL_DISASSOCIATED;
Tejun Heoebf44d12013-03-13 19:47:39 -07005068 BUG_ON(create_and_start_worker(pool) < 0);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005069 }
Tejun Heoe22bee72010-06-29 10:07:14 +02005070 }
5071
Tejun Heo8a2b7532013-09-05 12:30:04 -04005072 /* create default unbound and ordered wq attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -07005073 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5074 struct workqueue_attrs *attrs;
5075
5076 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
Tejun Heo29c91e92013-03-12 11:30:03 -07005077 attrs->nice = std_nice[i];
Tejun Heo29c91e92013-03-12 11:30:03 -07005078 unbound_std_wq_attrs[i] = attrs;
Tejun Heo8a2b7532013-09-05 12:30:04 -04005079
5080 /*
5081 * An ordered wq should have only one pwq as ordering is
5082 * guaranteed by max_active which is enforced by pwqs.
5083 * Turn off NUMA so that dfl_pwq is used for all nodes.
5084 */
5085 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5086 attrs->nice = std_nice[i];
5087 attrs->no_numa = true;
5088 ordered_wq_attrs[i] = attrs;
Tejun Heo29c91e92013-03-12 11:30:03 -07005089 }
5090
Tejun Heod320c032010-06-29 10:07:14 +02005091 system_wq = alloc_workqueue("events", 0, 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005092 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
Tejun Heod320c032010-06-29 10:07:14 +02005093 system_long_wq = alloc_workqueue("events_long", 0, 0);
Tejun Heof3421792010-07-02 10:03:51 +02005094 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5095 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01005096 system_freezable_wq = alloc_workqueue("events_freezable",
5097 WQ_FREEZABLE, 0);
Viresh Kumar06681062013-04-24 17:12:54 +05305098 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5099 WQ_POWER_EFFICIENT, 0);
5100 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5101 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5102 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005103 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
Viresh Kumar06681062013-04-24 17:12:54 +05305104 !system_unbound_wq || !system_freezable_wq ||
5105 !system_power_efficient_wq ||
5106 !system_freezable_power_efficient_wq);
Suresh Siddha6ee05782010-07-30 14:57:37 -07005107 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005108}
Suresh Siddha6ee05782010-07-30 14:57:37 -07005109early_initcall(init_workqueues);