blob: 072cbc9b175dc1efbe95c14858f810f92db12130 [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 Lameter469340232006-10-11 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
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +080068 * attach_mutex to avoid changing binding state while
Lai Jiangshan4736cbf2014-05-20 17:46:35 +080069 * worker_attach_to_pool() is in progress.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070070 */
Tejun Heo24647572013-01-24 11:01:33 -080071 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020072
Tejun Heoc8e55f32010-06-29 10:07:12 +020073 /* worker flags */
Tejun Heoc8e55f32010-06-29 10:07:12 +020074 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
Tejun Heoe22bee72010-06-29 10:07:14 +020076 WORKER_PREP = 1 << 3, /* preparing to run works */
Tejun Heofb0e7be2010-06-29 10:07:15 +020077 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
Tejun Heof3421792010-07-02 10:03:51 +020078 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
Tejun Heoa9ab7752013-03-19 13:45:21 -070079 WORKER_REBOUND = 1 << 8, /* worker was rebound */
Tejun Heoe22bee72010-06-29 10:07:14 +020080
Tejun Heoa9ab7752013-03-19 13:45:21 -070081 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
Tejun Heodb7bccf2010-06-29 10:07:12 +020083
Tejun Heoe34cdddb2013-01-24 11:01:33 -080084 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
Tejun Heo4ce62e92012-07-13 22:16:44 -070085
Tejun Heo29c91e92013-03-12 11:30:03 -070086 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
Tejun Heoc8e55f32010-06-29 10:07:12 +020087 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020088
Tejun Heoe22bee72010-06-29 10:07:14 +020089 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91
Tejun Heo3233cdb2011-02-16 18:10:19 +010092 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
94 (min two ticks) */
Tejun Heoe22bee72010-06-29 10:07:14 +020095 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
Tejun Heoe22bee72010-06-29 10:07:14 +020097
98 /*
99 * Rescue workers are used only on emergencies and shared by
Dongsheng Yang8698a742014-03-11 18:09:12 +0800100 * all cpus. Give MIN_NICE.
Tejun Heoe22bee72010-06-29 10:07:14 +0200101 */
Dongsheng Yang8698a742014-03-11 18:09:12 +0800102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
Tejun Heoecf68812013-04-01 11:23:34 -0700104
105 WQ_NAME_LEN = 24,
Tejun Heoc8e55f32010-06-29 10:07:12 +0200106};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
108/*
Tejun Heo4690c4a2010-06-29 10:07:10 +0200109 * Structure fields follow one of the following exclusion rules.
110 *
Tejun Heoe41e7042010-08-24 14:22:47 +0200111 * I: Modifiable by initialization/destruction paths and read-only for
112 * everyone else.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200113 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
116 *
Tejun Heod565ed62013-01-24 11:01:33 -0800117 * L: pool->lock protected. Access with pool->lock held.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200118 *
Tejun Heod565ed62013-01-24 11:01:33 -0800119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
Tejun Heoe22bee72010-06-29 10:07:14 +0200123 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800124 * A: pool->attach_mutex protected.
Tejun Heo822d8402013-03-19 13:45:21 -0700125 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700126 * PL: wq_pool_mutex protected.
Tejun Heo76af4d92013-03-12 11:30:00 -0700127 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo5bcab332013-03-13 19:47:40 -0700129 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
131 *
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
134 *
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700135 * WQ: wq->mutex protected.
136 *
Lai Jiangshanb5927602013-03-25 16:57:19 -0700137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo2e109a22013-03-13 19:47:40 -0700138 *
139 * MD: wq_mayday_lock protected.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200140 */
141
Tejun Heo2eaebdb32013-01-18 14:05:55 -0800142/* struct worker is defined in workqueue_internal.h */
Tejun Heoc34056a2010-06-29 10:07:11 +0200143
Tejun Heobd7bdd42012-07-12 14:46:37 -0700144struct worker_pool {
Tejun Heod565ed62013-01-24 11:01:33 -0800145 spinlock_t lock; /* the pool lock */
Tejun Heod84ff052013-03-12 11:29:59 -0700146 int cpu; /* I: the associated cpu */
Tejun Heof3f90ad2013-04-01 11:23:34 -0700147 int node; /* I: the associated node ID */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800148 int id; /* I: pool ID */
Tejun Heo11ebea52012-07-12 14:46:37 -0700149 unsigned int flags; /* X: flags */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700150
Tejun Heo82607adc2015-12-08 11:28:04 -0500151 unsigned long watchdog_ts; /* L: watchdog timestamp */
152
Tejun Heobd7bdd42012-07-12 14:46:37 -0700153 struct list_head worklist; /* L: list of pending works */
154 int nr_workers; /* L: total number of workers */
Lai Jiangshanea1abd62012-09-18 09:59:22 -0700155
156 /* nr_idle includes the ones off idle_list for rebinding */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700157 int nr_idle; /* L: currently idle ones */
158
159 struct list_head idle_list; /* X: list of idle workers */
160 struct timer_list idle_timer; /* L: worker idle timeout */
161 struct timer_list mayday_timer; /* L: SOS timer for workers */
162
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700163 /* a workers is either on busy_hash or idle_list, or the manager */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800164 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165 /* L: hash of busy workers */
166
Tejun Heobc3a1af2013-03-13 19:47:39 -0700167 /* see manage_workers() for details on the two manager mutexes */
Tejun Heo34a06bd2013-03-12 11:30:00 -0700168 struct mutex manager_arb; /* manager arbitration */
Tejun Heo2607d7a2015-03-09 09:22:28 -0400169 struct worker *manager; /* L: purely informational */
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800170 struct mutex attach_mutex; /* attach/detach exclusion */
171 struct list_head workers; /* A: attached workers */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +0800172 struct completion *detach_completion; /* all workers detached */
Tejun Heoe19e3972013-01-24 11:39:44 -0800173
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +0800174 struct ida worker_ida; /* worker IDs for task name */
Tejun Heoe19e3972013-01-24 11:39:44 -0800175
Tejun Heo7a4e3442013-03-12 11:30:00 -0700176 struct workqueue_attrs *attrs; /* I: worker attributes */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
Tejun Heo7a4e3442013-03-12 11:30:00 -0700179
Tejun Heoe19e3972013-01-24 11:39:44 -0800180 /*
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
183 * cacheline.
184 */
185 atomic_t nr_running ____cacheline_aligned_in_smp;
Tejun Heo29c91e92013-03-12 11:30:03 -0700186
187 /*
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
190 */
191 struct rcu_head rcu;
Tejun Heo8b03ae32010-06-29 10:07:12 +0200192} ____cacheline_aligned_in_smp;
193
194/*
Tejun Heo112202d2013-02-13 19:29:12 -0800195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199 */
Tejun Heo112202d2013-02-13 19:29:12 -0800200struct pool_workqueue {
Tejun Heobd7bdd42012-07-12 14:46:37 -0700201 struct worker_pool *pool; /* I: the associated pool */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200202 struct workqueue_struct *wq; /* I: the owning workqueue */
Tejun Heo73f53c42010-06-29 10:07:11 +0200203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700205 int refcnt; /* L: reference count */
Tejun Heo73f53c42010-06-29 10:07:11 +0200206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
Tejun Heo1e19ffc62010-06-29 10:07:12 +0200208 int nr_active; /* L: nr of active works */
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200209 int max_active; /* L: max active works */
Tejun Heo1e19ffc62010-06-29 10:07:12 +0200210 struct list_head delayed_works; /* L: delayed works */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
Tejun Heo2e109a22013-03-13 19:47:40 -0700212 struct list_head mayday_node; /* MD: node on wq->maydays */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700213
214 /*
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700218 * determined without grabbing wq->mutex.
Tejun Heo8864b4e2013-03-12 11:30:04 -0700219 */
220 struct work_struct unbound_release_work;
221 struct rcu_head rcu;
Tejun Heoe904e6c2013-03-12 11:29:57 -0700222} __aligned(1 << WORK_STRUCT_FLAG_BITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224/*
Tejun Heo73f53c42010-06-29 10:07:11 +0200225 * Structure used to wait for workqueue flush.
226 */
227struct wq_flusher {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
Tejun Heo73f53c42010-06-29 10:07:11 +0200230 struct completion done; /* flush completion */
231};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232
Tejun Heo226223a2013-03-12 11:30:05 -0700233struct wq_device;
234
Tejun Heo73f53c42010-06-29 10:07:11 +0200235/*
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 */
239struct workqueue_struct {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700240 struct list_head pwqs; /* WR: all pwqs of this wq */
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400241 struct list_head list; /* PR: list of all workqueues */
Tejun Heo73f53c42010-06-29 10:07:11 +0200242
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
Tejun Heo112202d2013-02-13 19:29:12 -0800246 atomic_t nr_pwqs_to_flush; /* flush in progress */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
Tejun Heo73f53c42010-06-29 10:07:11 +0200250
Tejun Heo2e109a22013-03-13 19:47:40 -0700251 struct list_head maydays; /* MD: pwqs requesting rescue */
Tejun Heoe22bee72010-06-29 10:07:14 +0200252 struct worker *rescuer; /* I: rescue worker */
253
Lai Jiangshan87fc7412013-03-25 16:57:18 -0700254 int nr_drainers; /* WQ: drain in progress */
Lai Jiangshana357fc02013-03-25 16:57:19 -0700255 int saved_max_active; /* WQ: saved pwq max_active */
Tejun Heo226223a2013-03-12 11:30:05 -0700256
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
Tejun Heo6029a912013-04-01 11:23:34 -0700259
Tejun Heo226223a2013-03-12 11:30:05 -0700260#ifdef CONFIG_SYSFS
261 struct wq_device *wq_dev; /* I: for sysfs interface */
262#endif
Johannes Berg4e6045f2007-10-18 23:39:55 -0700263#ifdef CONFIG_LOCKDEP
Tejun Heo4690c4a2010-06-29 10:07:10 +0200264 struct lockdep_map lockdep_map;
Johannes Berg4e6045f2007-10-18 23:39:55 -0700265#endif
Tejun Heoecf68812013-04-01 11:23:34 -0700266 char name[WQ_NAME_LEN]; /* I: workqueue name */
Tejun Heo2728fd22013-04-01 11:23:35 -0700267
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400268 /*
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
272 */
273 struct rcu_head rcu;
274
Tejun Heo2728fd22013-04-01 11:23:35 -0700275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279};
280
Tejun Heoe904e6c2013-03-12 11:29:57 -0700281static struct kmem_cache *pwq_cache;
282
Tejun Heobce90382013-04-01 11:23:32 -0700283static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
285
Tejun Heod55262c2013-04-01 11:23:38 -0700286static bool wq_disable_numa;
287module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288
Viresh Kumarcee22a12013-04-08 16:45:40 +0530289/* see the comment above the definition of WQ_POWER_EFFICIENT */
Luis R. Rodriguez552f5302015-05-27 11:09:39 +0930290static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
Viresh Kumarcee22a12013-04-08 16:45:40 +0530291module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292
Tejun Heo863b7102016-09-16 15:49:34 -0400293static bool wq_online; /* can kworkers be created yet? */
Tejun Heo3347fa02016-09-16 15:49:32 -0400294
Tejun Heobce90382013-04-01 11:23:32 -0700295static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
296
Tejun Heo4c16bd32013-04-01 11:23:36 -0700297/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
299
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700300static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
Tejun Heo2e109a22013-03-13 19:47:40 -0700301static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
Tejun Heo5bcab332013-03-13 19:47:40 -0700302
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400303static LIST_HEAD(workqueues); /* PR: list of all workqueues */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700304static bool workqueue_freezing; /* PL: have wqs started freezing? */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700305
Mike Galbraithef5571802016-02-09 17:59:38 -0500306/* PL: allowable cpus for unbound wqs and work items */
307static cpumask_var_t wq_unbound_cpumask;
308
309/* CPU where unbound work was last round robin scheduled from this CPU */
310static DEFINE_PER_CPU(int, wq_rr_cpu_last);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +0800311
Tejun Heof303fccb2016-02-09 17:59:38 -0500312/*
313 * Local execution of unbound work items is no longer guaranteed. The
314 * following always forces round-robin CPU selection on unbound work items
315 * to uncover usages which depend on it.
316 */
317#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
318static bool wq_debug_force_rr_cpu = true;
319#else
320static bool wq_debug_force_rr_cpu = false;
321#endif
322module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
323
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700324/* the per-cpu worker pools */
Peter Zijlstra25528212016-03-15 14:52:49 -0700325static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700326
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700327static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700328
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700329/* PL: hash of all unbound pools keyed by pool->attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -0700330static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
331
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700332/* I: attributes used when instantiating standard unbound pools on demand */
Tejun Heo29c91e92013-03-12 11:30:03 -0700333static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
334
Tejun Heo8a2b7532013-09-05 12:30:04 -0400335/* I: attributes used when instantiating ordered pools on demand */
336static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
337
Tejun Heod320c032010-06-29 10:07:14 +0200338struct workqueue_struct *system_wq __read_mostly;
Marc Dionnead7b1f82013-05-06 17:44:55 -0400339EXPORT_SYMBOL(system_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300340struct workqueue_struct *system_highpri_wq __read_mostly;
Joonsoo Kim1aabe902012-08-15 23:25:39 +0900341EXPORT_SYMBOL_GPL(system_highpri_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300342struct workqueue_struct *system_long_wq __read_mostly;
Tejun Heod320c032010-06-29 10:07:14 +0200343EXPORT_SYMBOL_GPL(system_long_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300344struct workqueue_struct *system_unbound_wq __read_mostly;
Tejun Heof3421792010-07-02 10:03:51 +0200345EXPORT_SYMBOL_GPL(system_unbound_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300346struct workqueue_struct *system_freezable_wq __read_mostly;
Tejun Heo24d51ad2011-02-21 09:52:50 +0100347EXPORT_SYMBOL_GPL(system_freezable_wq);
Viresh Kumar06681062013-04-24 17:12:54 +0530348struct workqueue_struct *system_power_efficient_wq __read_mostly;
349EXPORT_SYMBOL_GPL(system_power_efficient_wq);
350struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
351EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
Tejun Heod320c032010-06-29 10:07:14 +0200352
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700353static int worker_thread(void *__worker);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +0800354static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700355
Tejun Heo97bd2342010-10-05 10:41:14 +0200356#define CREATE_TRACE_POINTS
357#include <trace/events/workqueue.h>
358
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700359#define assert_rcu_or_pool_mutex() \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700360 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
361 !lockdep_is_held(&wq_pool_mutex), \
362 "sched RCU or wq_pool_mutex should be held")
Tejun Heo5bcab332013-03-13 19:47:40 -0700363
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700364#define assert_rcu_or_wq_mutex(wq) \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700365 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
366 !lockdep_is_held(&wq->mutex), \
367 "sched RCU or wq->mutex should be held")
Tejun Heo76af4d92013-03-12 11:30:00 -0700368
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800369#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700370 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
371 !lockdep_is_held(&wq->mutex) && \
372 !lockdep_is_held(&wq_pool_mutex), \
373 "sched RCU, wq->mutex or wq_pool_mutex should be held")
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800374
Tejun Heof02ae732013-03-12 11:30:03 -0700375#define for_each_cpu_worker_pool(pool, cpu) \
376 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
377 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
Tejun Heo7a62c2c2013-03-12 11:30:03 -0700378 (pool)++)
Tejun Heo4ce62e92012-07-13 22:16:44 -0700379
Tejun Heo49e3cf42013-03-12 11:29:58 -0700380/**
Tejun Heo17116962013-03-12 11:29:58 -0700381 * for_each_pool - iterate through all worker_pools in the system
382 * @pool: iteration cursor
Tejun Heo611c92a2013-03-13 16:51:36 -0700383 * @pi: integer used for iteration
Tejun Heofa1b54e2013-03-12 11:30:00 -0700384 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700385 * This must be called either with wq_pool_mutex held or sched RCU read
386 * locked. If the pool needs to be used beyond the locking in effect, the
387 * caller is responsible for guaranteeing that the pool stays online.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700388 *
389 * The if/else clause exists only for the lockdep assertion and can be
390 * ignored.
Tejun Heo17116962013-03-12 11:29:58 -0700391 */
Tejun Heo611c92a2013-03-13 16:51:36 -0700392#define for_each_pool(pool, pi) \
393 idr_for_each_entry(&worker_pool_idr, pool, pi) \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700394 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
Tejun Heofa1b54e2013-03-12 11:30:00 -0700395 else
Tejun Heo17116962013-03-12 11:29:58 -0700396
397/**
Tejun Heo822d8402013-03-19 13:45:21 -0700398 * for_each_pool_worker - iterate through all workers of a worker_pool
399 * @worker: iteration cursor
Tejun Heo822d8402013-03-19 13:45:21 -0700400 * @pool: worker_pool to iterate workers of
401 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800402 * This must be called with @pool->attach_mutex.
Tejun Heo822d8402013-03-19 13:45:21 -0700403 *
404 * The if/else clause exists only for the lockdep assertion and can be
405 * ignored.
406 */
Lai Jiangshanda028462014-05-20 17:46:31 +0800407#define for_each_pool_worker(worker, pool) \
408 list_for_each_entry((worker), &(pool)->workers, node) \
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800409 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
Tejun Heo822d8402013-03-19 13:45:21 -0700410 else
411
412/**
Tejun Heo49e3cf42013-03-12 11:29:58 -0700413 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
414 * @pwq: iteration cursor
415 * @wq: the target workqueue
Tejun Heo76af4d92013-03-12 11:30:00 -0700416 *
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700417 * This must be called either with wq->mutex held or sched RCU read locked.
Tejun Heo794b18b2013-03-13 19:47:40 -0700418 * If the pwq needs to be used beyond the locking in effect, the caller is
419 * responsible for guaranteeing that the pwq stays online.
Tejun Heo76af4d92013-03-12 11:30:00 -0700420 *
421 * The if/else clause exists only for the lockdep assertion and can be
422 * ignored.
Tejun Heo49e3cf42013-03-12 11:29:58 -0700423 */
424#define for_each_pwq(pwq, wq) \
Tejun Heo76af4d92013-03-12 11:30:00 -0700425 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700426 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
Tejun Heo76af4d92013-03-12 11:30:00 -0700427 else
Tejun Heof3421792010-07-02 10:03:51 +0200428
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900429#ifdef CONFIG_DEBUG_OBJECTS_WORK
430
431static struct debug_obj_descr work_debug_descr;
432
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100433static void *work_debug_hint(void *addr)
434{
435 return ((struct work_struct *) addr)->func;
436}
437
Du, Changbinb9fdac7f2016-05-19 17:09:41 -0700438static bool work_is_static_object(void *addr)
439{
440 struct work_struct *work = addr;
441
442 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
443}
444
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900445/*
446 * fixup_init is called when:
447 * - an active object is initialized
448 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700449static bool work_fixup_init(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900450{
451 struct work_struct *work = addr;
452
453 switch (state) {
454 case ODEBUG_STATE_ACTIVE:
455 cancel_work_sync(work);
456 debug_object_init(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700457 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900458 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700459 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900460 }
461}
462
463/*
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900464 * fixup_free is called when:
465 * - an active object is freed
466 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700467static bool work_fixup_free(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900468{
469 struct work_struct *work = addr;
470
471 switch (state) {
472 case ODEBUG_STATE_ACTIVE:
473 cancel_work_sync(work);
474 debug_object_free(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700475 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900476 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700477 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900478 }
479}
480
481static struct debug_obj_descr work_debug_descr = {
482 .name = "work_struct",
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100483 .debug_hint = work_debug_hint,
Du, Changbinb9fdac7f2016-05-19 17:09:41 -0700484 .is_static_object = work_is_static_object,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900485 .fixup_init = work_fixup_init,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900486 .fixup_free = work_fixup_free,
487};
488
489static inline void debug_work_activate(struct work_struct *work)
490{
491 debug_object_activate(work, &work_debug_descr);
492}
493
494static inline void debug_work_deactivate(struct work_struct *work)
495{
496 debug_object_deactivate(work, &work_debug_descr);
497}
498
499void __init_work(struct work_struct *work, int onstack)
500{
501 if (onstack)
502 debug_object_init_on_stack(work, &work_debug_descr);
503 else
504 debug_object_init(work, &work_debug_descr);
505}
506EXPORT_SYMBOL_GPL(__init_work);
507
508void destroy_work_on_stack(struct work_struct *work)
509{
510 debug_object_free(work, &work_debug_descr);
511}
512EXPORT_SYMBOL_GPL(destroy_work_on_stack);
513
Thomas Gleixnerea2e64f2014-03-23 14:20:44 +0000514void destroy_delayed_work_on_stack(struct delayed_work *work)
515{
516 destroy_timer_on_stack(&work->timer);
517 debug_object_free(&work->work, &work_debug_descr);
518}
519EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
520
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900521#else
522static inline void debug_work_activate(struct work_struct *work) { }
523static inline void debug_work_deactivate(struct work_struct *work) { }
524#endif
525
Li Bin4e8b22b2013-09-10 09:52:35 +0800526/**
527 * worker_pool_assign_id - allocate ID and assing it to @pool
528 * @pool: the pool pointer of interest
529 *
530 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
531 * successfully, -errno on failure.
532 */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800533static int worker_pool_assign_id(struct worker_pool *pool)
534{
535 int ret;
536
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700537 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo5bcab332013-03-13 19:47:40 -0700538
Li Bin4e8b22b2013-09-10 09:52:35 +0800539 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
540 GFP_KERNEL);
Tejun Heo229641a2013-04-01 17:08:13 -0700541 if (ret >= 0) {
Tejun Heoe68035f2013-03-13 14:59:38 -0700542 pool->id = ret;
Tejun Heo229641a2013-04-01 17:08:13 -0700543 return 0;
544 }
Tejun Heo9daf9e62013-01-24 11:01:33 -0800545 return ret;
546}
547
Tejun Heo76af4d92013-03-12 11:30:00 -0700548/**
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700549 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
550 * @wq: the target workqueue
551 * @node: the node ID
552 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800553 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
554 * read locked.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700555 * If the pwq needs to be used beyond the locking in effect, the caller is
556 * responsible for guaranteeing that the pwq stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700557 *
558 * Return: The unbound pool_workqueue for @node.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700559 */
560static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
561 int node)
562{
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800563 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
Tejun Heod6e022f2016-02-03 13:54:25 -0500564
565 /*
566 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
567 * delayed item is pending. The plan is to keep CPU -> NODE
568 * mapping valid and stable across CPU on/offlines. Once that
569 * happens, this workaround can be removed.
570 */
571 if (unlikely(node == NUMA_NO_NODE))
572 return wq->dfl_pwq;
573
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700574 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
575}
576
Tejun Heo73f53c42010-06-29 10:07:11 +0200577static unsigned int work_color_to_flags(int color)
578{
579 return color << WORK_STRUCT_COLOR_SHIFT;
580}
581
582static int get_work_color(struct work_struct *work)
583{
584 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
585 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
586}
587
588static int work_next_color(int color)
589{
590 return (color + 1) % WORK_NR_COLORS;
Oleg Nesterova848e3b2007-05-09 02:34:17 -0700591}
592
David Howells4594bf12006-12-07 11:33:26 +0000593/*
Tejun Heo112202d2013-02-13 19:29:12 -0800594 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
595 * contain the pointer to the queued pwq. Once execution starts, the flag
Tejun Heo7c3eed52013-01-24 11:01:33 -0800596 * is cleared and the high bits contain OFFQ flags and pool ID.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200597 *
Tejun Heo112202d2013-02-13 19:29:12 -0800598 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
599 * and clear_work_data() can be used to set the pwq, pool or clear
Tejun Heobbb68df2012-08-03 10:30:46 -0700600 * work->data. These functions should only be called while the work is
601 * owned - ie. while the PENDING bit is set.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200602 *
Tejun Heo112202d2013-02-13 19:29:12 -0800603 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
Tejun Heo7c3eed52013-01-24 11:01:33 -0800604 * corresponding to a work. Pool is available once the work has been
Tejun Heo112202d2013-02-13 19:29:12 -0800605 * queued anywhere after initialization until it is sync canceled. pwq is
Tejun Heo7c3eed52013-01-24 11:01:33 -0800606 * available only while the work item is queued.
Tejun Heobbb68df2012-08-03 10:30:46 -0700607 *
608 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
609 * canceled. While being canceled, a work item may have its PENDING set
610 * but stay off timer and worklist for arbitrarily long and nobody should
611 * try to steal the PENDING bit.
David Howells4594bf12006-12-07 11:33:26 +0000612 */
Tejun Heo7a22ad72010-06-29 10:07:13 +0200613static inline void set_work_data(struct work_struct *work, unsigned long data,
614 unsigned long flags)
David Howells365970a2006-11-22 14:54:49 +0000615{
Tejun Heo6183c002013-03-12 11:29:57 -0700616 WARN_ON_ONCE(!work_pending(work));
Tejun Heo7a22ad72010-06-29 10:07:13 +0200617 atomic_long_set(&work->data, data | flags | work_static(work));
David Howells365970a2006-11-22 14:54:49 +0000618}
David Howells365970a2006-11-22 14:54:49 +0000619
Tejun Heo112202d2013-02-13 19:29:12 -0800620static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
Tejun Heo7a22ad72010-06-29 10:07:13 +0200621 unsigned long extra_flags)
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200622{
Tejun Heo112202d2013-02-13 19:29:12 -0800623 set_work_data(work, (unsigned long)pwq,
624 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200625}
626
Lai Jiangshan4468a002013-02-06 18:04:53 -0800627static void set_work_pool_and_keep_pending(struct work_struct *work,
628 int pool_id)
629{
630 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
631 WORK_STRUCT_PENDING);
632}
633
Tejun Heo7c3eed52013-01-24 11:01:33 -0800634static void set_work_pool_and_clear_pending(struct work_struct *work,
635 int pool_id)
David Howells365970a2006-11-22 14:54:49 +0000636{
Tejun Heo23657bb2012-08-13 17:08:19 -0700637 /*
638 * The following wmb is paired with the implied mb in
639 * test_and_set_bit(PENDING) and ensures all updates to @work made
640 * here are visible to and precede any updates by the next PENDING
641 * owner.
642 */
643 smp_wmb();
Tejun Heo7c3eed52013-01-24 11:01:33 -0800644 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
Roman Pen346c09f2016-04-26 13:15:35 +0200645 /*
646 * The following mb guarantees that previous clear of a PENDING bit
647 * will not be reordered with any speculative LOADS or STORES from
648 * work->current_func, which is executed afterwards. This possible
649 * reordering can lead to a missed execution on attempt to qeueue
650 * the same @work. E.g. consider this case:
651 *
652 * CPU#0 CPU#1
653 * ---------------------------- --------------------------------
654 *
655 * 1 STORE event_indicated
656 * 2 queue_work_on() {
657 * 3 test_and_set_bit(PENDING)
658 * 4 } set_..._and_clear_pending() {
659 * 5 set_work_data() # clear bit
660 * 6 smp_mb()
661 * 7 work->current_func() {
662 * 8 LOAD event_indicated
663 * }
664 *
665 * Without an explicit full barrier speculative LOAD on line 8 can
666 * be executed before CPU#0 does STORE on line 1. If that happens,
667 * CPU#0 observes the PENDING bit is still set and new execution of
668 * a @work is not queued in a hope, that CPU#1 will eventually
669 * finish the queued @work. Meanwhile CPU#1 does not see
670 * event_indicated is set, because speculative LOAD was executed
671 * before actual STORE.
672 */
673 smp_mb();
Tejun Heo7a22ad72010-06-29 10:07:13 +0200674}
675
676static void clear_work_data(struct work_struct *work)
677{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800678 smp_wmb(); /* see set_work_pool_and_clear_pending() */
679 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200680}
681
Tejun Heo112202d2013-02-13 19:29:12 -0800682static struct pool_workqueue *get_work_pwq(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200683{
Tejun Heoe1201532010-07-22 14:14:25 +0200684 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200685
Tejun Heo112202d2013-02-13 19:29:12 -0800686 if (data & WORK_STRUCT_PWQ)
Tejun Heoe1201532010-07-22 14:14:25 +0200687 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
688 else
689 return NULL;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200690}
691
Tejun Heo7c3eed52013-01-24 11:01:33 -0800692/**
693 * get_work_pool - return the worker_pool a given work was associated with
694 * @work: the work item of interest
695 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700696 * Pools are created and destroyed under wq_pool_mutex, and allows read
697 * access under sched-RCU read lock. As such, this function should be
698 * called under wq_pool_mutex or with preemption disabled.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700699 *
700 * All fields of the returned pool are accessible as long as the above
701 * mentioned locking is in effect. If the returned pool needs to be used
702 * beyond the critical section, the caller is responsible for ensuring the
703 * returned pool is and stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700704 *
705 * Return: The worker_pool @work was last associated with. %NULL if none.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800706 */
707static struct worker_pool *get_work_pool(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200708{
Tejun Heoe1201532010-07-22 14:14:25 +0200709 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800710 int pool_id;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200711
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700712 assert_rcu_or_pool_mutex();
Tejun Heofa1b54e2013-03-12 11:30:00 -0700713
Tejun Heo112202d2013-02-13 19:29:12 -0800714 if (data & WORK_STRUCT_PWQ)
715 return ((struct pool_workqueue *)
Tejun Heo7c3eed52013-01-24 11:01:33 -0800716 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200717
Tejun Heo7c3eed52013-01-24 11:01:33 -0800718 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
719 if (pool_id == WORK_OFFQ_POOL_NONE)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200720 return NULL;
721
Tejun Heofa1b54e2013-03-12 11:30:00 -0700722 return idr_find(&worker_pool_idr, pool_id);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800723}
724
725/**
726 * get_work_pool_id - return the worker pool ID a given work is associated with
727 * @work: the work item of interest
728 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700729 * Return: The worker_pool ID @work was last associated with.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800730 * %WORK_OFFQ_POOL_NONE if none.
731 */
732static int get_work_pool_id(struct work_struct *work)
733{
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800734 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800735
Tejun Heo112202d2013-02-13 19:29:12 -0800736 if (data & WORK_STRUCT_PWQ)
737 return ((struct pool_workqueue *)
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800738 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
739
740 return data >> WORK_OFFQ_POOL_SHIFT;
Tejun Heo7c3eed52013-01-24 11:01:33 -0800741}
742
Tejun Heobbb68df2012-08-03 10:30:46 -0700743static void mark_work_canceling(struct work_struct *work)
744{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800745 unsigned long pool_id = get_work_pool_id(work);
Tejun Heobbb68df2012-08-03 10:30:46 -0700746
Tejun Heo7c3eed52013-01-24 11:01:33 -0800747 pool_id <<= WORK_OFFQ_POOL_SHIFT;
748 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700749}
750
751static bool work_is_canceling(struct work_struct *work)
752{
753 unsigned long data = atomic_long_read(&work->data);
754
Tejun Heo112202d2013-02-13 19:29:12 -0800755 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700756}
757
David Howells365970a2006-11-22 14:54:49 +0000758/*
Tejun Heo32704762012-07-13 22:16:45 -0700759 * Policy functions. These define the policies on how the global worker
760 * pools are managed. Unless noted otherwise, these functions assume that
Tejun Heod565ed62013-01-24 11:01:33 -0800761 * they're being called with pool->lock held.
David Howells365970a2006-11-22 14:54:49 +0000762 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200763
Tejun Heo63d95a92012-07-12 14:46:37 -0700764static bool __need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000765{
Tejun Heoe19e3972013-01-24 11:39:44 -0800766 return !atomic_read(&pool->nr_running);
David Howells365970a2006-11-22 14:54:49 +0000767}
768
Tejun Heoe22bee72010-06-29 10:07:14 +0200769/*
770 * Need to wake up a worker? Called from anything but currently
771 * running workers.
Tejun Heo974271c2012-07-12 14:46:37 -0700772 *
773 * Note that, because unbound workers never contribute to nr_running, this
Tejun Heo706026c2013-01-24 11:01:34 -0800774 * function will always return %true for unbound pools as long as the
Tejun Heo974271c2012-07-12 14:46:37 -0700775 * worklist isn't empty.
Tejun Heoe22bee72010-06-29 10:07:14 +0200776 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700777static bool need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000778{
Tejun Heo63d95a92012-07-12 14:46:37 -0700779 return !list_empty(&pool->worklist) && __need_more_worker(pool);
David Howells365970a2006-11-22 14:54:49 +0000780}
781
Tejun Heoe22bee72010-06-29 10:07:14 +0200782/* Can I start working? Called from busy but !running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700783static bool may_start_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200784{
Tejun Heo63d95a92012-07-12 14:46:37 -0700785 return pool->nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200786}
787
788/* Do I need to keep working? Called from currently running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700789static bool keep_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200790{
Tejun Heoe19e3972013-01-24 11:39:44 -0800791 return !list_empty(&pool->worklist) &&
792 atomic_read(&pool->nr_running) <= 1;
Tejun Heoe22bee72010-06-29 10:07:14 +0200793}
794
795/* Do we need a new worker? Called from manager. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700796static bool need_to_create_worker(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200797{
Tejun Heo63d95a92012-07-12 14:46:37 -0700798 return need_more_worker(pool) && !may_start_working(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200799}
800
Tejun Heoe22bee72010-06-29 10:07:14 +0200801/* Do we have too many workers and should some go away? */
Tejun Heo63d95a92012-07-12 14:46:37 -0700802static bool too_many_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200803{
Tejun Heo34a06bd2013-03-12 11:30:00 -0700804 bool managing = mutex_is_locked(&pool->manager_arb);
Tejun Heo63d95a92012-07-12 14:46:37 -0700805 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
806 int nr_busy = pool->nr_workers - nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200807
808 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
809}
810
811/*
812 * Wake up functions.
813 */
814
Lai Jiangshan1037de32014-05-22 16:44:07 +0800815/* Return the first idle worker. Safe with preemption disabled */
816static struct worker *first_idle_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200817{
Tejun Heo63d95a92012-07-12 14:46:37 -0700818 if (unlikely(list_empty(&pool->idle_list)))
Tejun Heo7e116292010-06-29 10:07:13 +0200819 return NULL;
820
Tejun Heo63d95a92012-07-12 14:46:37 -0700821 return list_first_entry(&pool->idle_list, struct worker, entry);
Tejun Heo7e116292010-06-29 10:07:13 +0200822}
823
824/**
825 * wake_up_worker - wake up an idle worker
Tejun Heo63d95a92012-07-12 14:46:37 -0700826 * @pool: worker pool to wake worker from
Tejun Heo7e116292010-06-29 10:07:13 +0200827 *
Tejun Heo63d95a92012-07-12 14:46:37 -0700828 * Wake up the first idle worker of @pool.
Tejun Heo7e116292010-06-29 10:07:13 +0200829 *
830 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800831 * spin_lock_irq(pool->lock).
Tejun Heo7e116292010-06-29 10:07:13 +0200832 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700833static void wake_up_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200834{
Lai Jiangshan1037de32014-05-22 16:44:07 +0800835 struct worker *worker = first_idle_worker(pool);
Tejun Heo7e116292010-06-29 10:07:13 +0200836
837 if (likely(worker))
838 wake_up_process(worker->task);
839}
840
Tejun Heo4690c4a2010-06-29 10:07:10 +0200841/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200842 * wq_worker_waking_up - a worker is waking up
843 * @task: task waking up
844 * @cpu: CPU @task is waking up to
845 *
846 * This function is called during try_to_wake_up() when a worker is
847 * being awoken.
848 *
849 * CONTEXT:
850 * spin_lock_irq(rq->lock)
851 */
Tejun Heod84ff052013-03-12 11:29:59 -0700852void wq_worker_waking_up(struct task_struct *task, int cpu)
Tejun Heoe22bee72010-06-29 10:07:14 +0200853{
854 struct worker *worker = kthread_data(task);
855
Joonsoo Kim36576002012-10-26 23:03:49 +0900856 if (!(worker->flags & WORKER_NOT_RUNNING)) {
Tejun Heoec22ca52013-01-24 11:01:33 -0800857 WARN_ON_ONCE(worker->pool->cpu != cpu);
Tejun Heoe19e3972013-01-24 11:39:44 -0800858 atomic_inc(&worker->pool->nr_running);
Joonsoo Kim36576002012-10-26 23:03:49 +0900859 }
Tejun Heoe22bee72010-06-29 10:07:14 +0200860}
861
862/**
863 * wq_worker_sleeping - a worker is going to sleep
864 * @task: task going to sleep
Tejun Heoe22bee72010-06-29 10:07:14 +0200865 *
866 * This function is called during schedule() when a busy worker is
867 * going to sleep. Worker on the same cpu can be woken up by
868 * returning pointer to its task.
869 *
870 * CONTEXT:
871 * spin_lock_irq(rq->lock)
872 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700873 * Return:
Tejun Heoe22bee72010-06-29 10:07:14 +0200874 * Worker task on @cpu to wake up, %NULL if none.
875 */
Alexander Gordeev9b7f6592016-03-02 12:53:31 +0100876struct task_struct *wq_worker_sleeping(struct task_struct *task)
Tejun Heoe22bee72010-06-29 10:07:14 +0200877{
878 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
Tejun Heo111c2252013-01-17 17:16:24 -0800879 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200880
Tejun Heo111c2252013-01-17 17:16:24 -0800881 /*
882 * Rescuers, which may not have all the fields set up like normal
883 * workers, also reach here, let's not access anything before
884 * checking NOT_RUNNING.
885 */
Steven Rostedt2d646722010-12-03 23:12:33 -0500886 if (worker->flags & WORKER_NOT_RUNNING)
Tejun Heoe22bee72010-06-29 10:07:14 +0200887 return NULL;
888
Tejun Heo111c2252013-01-17 17:16:24 -0800889 pool = worker->pool;
Tejun Heo111c2252013-01-17 17:16:24 -0800890
Tejun Heoe22bee72010-06-29 10:07:14 +0200891 /* this can only happen on the local cpu */
Alexander Gordeev9b7f6592016-03-02 12:53:31 +0100892 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
Tejun Heo6183c002013-03-12 11:29:57 -0700893 return NULL;
Tejun Heoe22bee72010-06-29 10:07:14 +0200894
895 /*
896 * The counterpart of the following dec_and_test, implied mb,
897 * worklist not empty test sequence is in insert_work().
898 * Please read comment there.
899 *
Tejun Heo628c78e2012-07-17 12:39:27 -0700900 * NOT_RUNNING is clear. This means that we're bound to and
901 * running on the local cpu w/ rq lock held and preemption
902 * disabled, which in turn means that none else could be
Tejun Heod565ed62013-01-24 11:01:33 -0800903 * manipulating idle_list, so dereferencing idle_list without pool
Tejun Heo628c78e2012-07-17 12:39:27 -0700904 * lock is safe.
Tejun Heoe22bee72010-06-29 10:07:14 +0200905 */
Tejun Heoe19e3972013-01-24 11:39:44 -0800906 if (atomic_dec_and_test(&pool->nr_running) &&
907 !list_empty(&pool->worklist))
Lai Jiangshan1037de32014-05-22 16:44:07 +0800908 to_wakeup = first_idle_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200909 return to_wakeup ? to_wakeup->task : NULL;
910}
911
912/**
913 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200914 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200915 * @flags: flags to set
Tejun Heod302f012010-06-29 10:07:13 +0200916 *
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800917 * Set @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200918 *
Tejun Heocb444762010-07-02 10:03:50 +0200919 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800920 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200921 */
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800922static inline void worker_set_flags(struct worker *worker, unsigned int flags)
Tejun Heod302f012010-06-29 10:07:13 +0200923{
Tejun Heobd7bdd42012-07-12 14:46:37 -0700924 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200925
Tejun Heocb444762010-07-02 10:03:50 +0200926 WARN_ON_ONCE(worker->task != current);
927
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800928 /* If transitioning into NOT_RUNNING, adjust nr_running. */
Tejun Heoe22bee72010-06-29 10:07:14 +0200929 if ((flags & WORKER_NOT_RUNNING) &&
930 !(worker->flags & WORKER_NOT_RUNNING)) {
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800931 atomic_dec(&pool->nr_running);
Tejun Heoe22bee72010-06-29 10:07:14 +0200932 }
933
Tejun Heod302f012010-06-29 10:07:13 +0200934 worker->flags |= flags;
935}
936
937/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200938 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200939 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200940 * @flags: flags to clear
941 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200942 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200943 *
Tejun Heocb444762010-07-02 10:03:50 +0200944 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800945 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200946 */
947static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
948{
Tejun Heo63d95a92012-07-12 14:46:37 -0700949 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200950 unsigned int oflags = worker->flags;
951
Tejun Heocb444762010-07-02 10:03:50 +0200952 WARN_ON_ONCE(worker->task != current);
953
Tejun Heod302f012010-06-29 10:07:13 +0200954 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200955
Tejun Heo42c025f2011-01-11 15:58:49 +0100956 /*
957 * If transitioning out of NOT_RUNNING, increment nr_running. Note
958 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
959 * of multiple flags, not a single flag.
960 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200961 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
962 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800963 atomic_inc(&pool->nr_running);
Tejun Heod302f012010-06-29 10:07:13 +0200964}
965
966/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200967 * find_worker_executing_work - find worker which is executing a work
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800968 * @pool: pool of interest
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200969 * @work: work to find worker for
970 *
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800971 * Find a worker which is executing @work on @pool by searching
972 * @pool->busy_hash which is keyed by the address of @work. For a worker
Tejun Heoa2c1c572012-12-18 10:35:02 -0800973 * to match, its current execution should match the address of @work and
974 * its work function. This is to avoid unwanted dependency between
975 * unrelated work executions through a work item being recycled while still
976 * being executed.
977 *
978 * This is a bit tricky. A work item may be freed once its execution
979 * starts and nothing prevents the freed area from being recycled for
980 * another work item. If the same work item address ends up being reused
981 * before the original execution finishes, workqueue will identify the
982 * recycled work item as currently executing and make it wait until the
983 * current execution finishes, introducing an unwanted dependency.
984 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700985 * This function checks the work item address and work function to avoid
986 * false positives. Note that this isn't complete as one may construct a
987 * work function which can introduce dependency onto itself through a
988 * recycled work item. Well, if somebody wants to shoot oneself in the
989 * foot that badly, there's only so much we can do, and if such deadlock
990 * actually occurs, it should be easy to locate the culprit work function.
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200991 *
992 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800993 * spin_lock_irq(pool->lock).
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200994 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700995 * Return:
996 * Pointer to worker which is executing @work if found, %NULL
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200997 * otherwise.
998 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800999static struct worker *find_worker_executing_work(struct worker_pool *pool,
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001000 struct work_struct *work)
1001{
Sasha Levin42f85702012-12-17 10:01:23 -05001002 struct worker *worker;
Sasha Levin42f85702012-12-17 10:01:23 -05001003
Sasha Levinb67bfe02013-02-27 17:06:00 -08001004 hash_for_each_possible(pool->busy_hash, worker, hentry,
Tejun Heoa2c1c572012-12-18 10:35:02 -08001005 (unsigned long)work)
1006 if (worker->current_work == work &&
1007 worker->current_func == work->func)
Sasha Levin42f85702012-12-17 10:01:23 -05001008 return worker;
1009
1010 return NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001011}
1012
1013/**
Tejun Heobf4ede02012-08-03 10:30:46 -07001014 * move_linked_works - move linked works to a list
1015 * @work: start of series of works to be scheduled
1016 * @head: target list to append @work to
Shailendra Verma402dd892015-05-23 10:38:14 +05301017 * @nextp: out parameter for nested worklist walking
Tejun Heobf4ede02012-08-03 10:30:46 -07001018 *
1019 * Schedule linked works starting from @work to @head. Work series to
1020 * be scheduled starts at @work and includes any consecutive work with
1021 * WORK_STRUCT_LINKED set in its predecessor.
1022 *
1023 * If @nextp is not NULL, it's updated to point to the next work of
1024 * the last scheduled work. This allows move_linked_works() to be
1025 * nested inside outer list_for_each_entry_safe().
1026 *
1027 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001028 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001029 */
1030static void move_linked_works(struct work_struct *work, struct list_head *head,
1031 struct work_struct **nextp)
1032{
1033 struct work_struct *n;
1034
1035 /*
1036 * Linked worklist will always end before the end of the list,
1037 * use NULL for list head.
1038 */
1039 list_for_each_entry_safe_from(work, n, NULL, entry) {
1040 list_move_tail(&work->entry, head);
1041 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1042 break;
1043 }
1044
1045 /*
1046 * If we're already inside safe list traversal and have moved
1047 * multiple works to the scheduled queue, the next position
1048 * needs to be updated.
1049 */
1050 if (nextp)
1051 *nextp = n;
1052}
1053
Tejun Heo8864b4e2013-03-12 11:30:04 -07001054/**
1055 * get_pwq - get an extra reference on the specified pool_workqueue
1056 * @pwq: pool_workqueue to get
1057 *
1058 * Obtain an extra reference on @pwq. The caller should guarantee that
1059 * @pwq has positive refcnt and be holding the matching pool->lock.
1060 */
1061static void get_pwq(struct pool_workqueue *pwq)
1062{
1063 lockdep_assert_held(&pwq->pool->lock);
1064 WARN_ON_ONCE(pwq->refcnt <= 0);
1065 pwq->refcnt++;
1066}
1067
1068/**
1069 * put_pwq - put a pool_workqueue reference
1070 * @pwq: pool_workqueue to put
1071 *
1072 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1073 * destruction. The caller should be holding the matching pool->lock.
1074 */
1075static void put_pwq(struct pool_workqueue *pwq)
1076{
1077 lockdep_assert_held(&pwq->pool->lock);
1078 if (likely(--pwq->refcnt))
1079 return;
1080 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1081 return;
1082 /*
1083 * @pwq can't be released under pool->lock, bounce to
1084 * pwq_unbound_release_workfn(). This never recurses on the same
1085 * pool->lock as this path is taken only for unbound workqueues and
1086 * the release work item is scheduled on a per-cpu workqueue. To
1087 * avoid lockdep warning, unbound pool->locks are given lockdep
1088 * subclass of 1 in get_unbound_pool().
1089 */
1090 schedule_work(&pwq->unbound_release_work);
1091}
1092
Tejun Heodce90d42013-04-01 11:23:35 -07001093/**
1094 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1095 * @pwq: pool_workqueue to put (can be %NULL)
1096 *
1097 * put_pwq() with locking. This function also allows %NULL @pwq.
1098 */
1099static void put_pwq_unlocked(struct pool_workqueue *pwq)
1100{
1101 if (pwq) {
1102 /*
1103 * As both pwqs and pools are sched-RCU protected, the
1104 * following lock operations are safe.
1105 */
1106 spin_lock_irq(&pwq->pool->lock);
1107 put_pwq(pwq);
1108 spin_unlock_irq(&pwq->pool->lock);
1109 }
1110}
1111
Tejun Heo112202d2013-02-13 19:29:12 -08001112static void pwq_activate_delayed_work(struct work_struct *work)
Tejun Heobf4ede02012-08-03 10:30:46 -07001113{
Tejun Heo112202d2013-02-13 19:29:12 -08001114 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobf4ede02012-08-03 10:30:46 -07001115
1116 trace_workqueue_activate_work(work);
Tejun Heo82607adc2015-12-08 11:28:04 -05001117 if (list_empty(&pwq->pool->worklist))
1118 pwq->pool->watchdog_ts = jiffies;
Tejun Heo112202d2013-02-13 19:29:12 -08001119 move_linked_works(work, &pwq->pool->worklist, NULL);
Tejun Heobf4ede02012-08-03 10:30:46 -07001120 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo112202d2013-02-13 19:29:12 -08001121 pwq->nr_active++;
Tejun Heobf4ede02012-08-03 10:30:46 -07001122}
1123
Tejun Heo112202d2013-02-13 19:29:12 -08001124static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001125{
Tejun Heo112202d2013-02-13 19:29:12 -08001126 struct work_struct *work = list_first_entry(&pwq->delayed_works,
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001127 struct work_struct, entry);
1128
Tejun Heo112202d2013-02-13 19:29:12 -08001129 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001130}
1131
Tejun Heobf4ede02012-08-03 10:30:46 -07001132/**
Tejun Heo112202d2013-02-13 19:29:12 -08001133 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1134 * @pwq: pwq of interest
Tejun Heobf4ede02012-08-03 10:30:46 -07001135 * @color: color of work which left the queue
Tejun Heobf4ede02012-08-03 10:30:46 -07001136 *
1137 * A work either has completed or is removed from pending queue,
Tejun Heo112202d2013-02-13 19:29:12 -08001138 * decrement nr_in_flight of its pwq and handle workqueue flushing.
Tejun Heobf4ede02012-08-03 10:30:46 -07001139 *
1140 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001141 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001142 */
Tejun Heo112202d2013-02-13 19:29:12 -08001143static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
Tejun Heobf4ede02012-08-03 10:30:46 -07001144{
Tejun Heo8864b4e2013-03-12 11:30:04 -07001145 /* uncolored work items don't participate in flushing or nr_active */
Tejun Heobf4ede02012-08-03 10:30:46 -07001146 if (color == WORK_NO_COLOR)
Tejun Heo8864b4e2013-03-12 11:30:04 -07001147 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001148
Tejun Heo112202d2013-02-13 19:29:12 -08001149 pwq->nr_in_flight[color]--;
Tejun Heobf4ede02012-08-03 10:30:46 -07001150
Tejun Heo112202d2013-02-13 19:29:12 -08001151 pwq->nr_active--;
1152 if (!list_empty(&pwq->delayed_works)) {
Lai Jiangshanb3f9f402012-09-18 10:40:00 -07001153 /* one down, submit a delayed one */
Tejun Heo112202d2013-02-13 19:29:12 -08001154 if (pwq->nr_active < pwq->max_active)
1155 pwq_activate_first_delayed(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001156 }
1157
1158 /* is flush in progress and are we at the flushing tip? */
Tejun Heo112202d2013-02-13 19:29:12 -08001159 if (likely(pwq->flush_color != color))
Tejun Heo8864b4e2013-03-12 11:30:04 -07001160 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001161
1162 /* are there still in-flight works? */
Tejun Heo112202d2013-02-13 19:29:12 -08001163 if (pwq->nr_in_flight[color])
Tejun Heo8864b4e2013-03-12 11:30:04 -07001164 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001165
Tejun Heo112202d2013-02-13 19:29:12 -08001166 /* this pwq is done, clear flush_color */
1167 pwq->flush_color = -1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001168
1169 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001170 * If this was the last pwq, wake up the first flusher. It
Tejun Heobf4ede02012-08-03 10:30:46 -07001171 * will handle the rest.
1172 */
Tejun Heo112202d2013-02-13 19:29:12 -08001173 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1174 complete(&pwq->wq->first_flusher->done);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001175out_put:
1176 put_pwq(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001177}
1178
Tejun Heo36e227d2012-08-03 10:30:46 -07001179/**
Tejun Heobbb68df2012-08-03 10:30:46 -07001180 * try_to_grab_pending - steal work item from worklist and disable irq
Tejun Heo36e227d2012-08-03 10:30:46 -07001181 * @work: work item to steal
1182 * @is_dwork: @work is a delayed_work
Tejun Heobbb68df2012-08-03 10:30:46 -07001183 * @flags: place to store irq state
Tejun Heo36e227d2012-08-03 10:30:46 -07001184 *
1185 * Try to grab PENDING bit of @work. This function can handle @work in any
Yacine Belkadid185af32013-07-31 14:59:24 -07001186 * stable state - idle, on timer or on worklist.
Tejun Heo36e227d2012-08-03 10:30:46 -07001187 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001188 * Return:
Tejun Heo36e227d2012-08-03 10:30:46 -07001189 * 1 if @work was pending and we successfully stole PENDING
1190 * 0 if @work was idle and we claimed PENDING
1191 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
Tejun Heobbb68df2012-08-03 10:30:46 -07001192 * -ENOENT if someone else is canceling @work, this state may persist
1193 * for arbitrarily long
Tejun Heo36e227d2012-08-03 10:30:46 -07001194 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001195 * Note:
Tejun Heobbb68df2012-08-03 10:30:46 -07001196 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001197 * interrupted while holding PENDING and @work off queue, irq must be
1198 * disabled on entry. This, combined with delayed_work->timer being
1199 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
Tejun Heobbb68df2012-08-03 10:30:46 -07001200 *
1201 * On successful return, >= 0, irq is disabled and the caller is
1202 * responsible for releasing it using local_irq_restore(*@flags).
1203 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001204 * This function is safe to call from any context including IRQ handler.
Tejun Heobf4ede02012-08-03 10:30:46 -07001205 */
Tejun Heobbb68df2012-08-03 10:30:46 -07001206static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1207 unsigned long *flags)
Tejun Heobf4ede02012-08-03 10:30:46 -07001208{
Tejun Heod565ed62013-01-24 11:01:33 -08001209 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001210 struct pool_workqueue *pwq;
Tejun Heobf4ede02012-08-03 10:30:46 -07001211
Tejun Heobbb68df2012-08-03 10:30:46 -07001212 local_irq_save(*flags);
1213
Tejun Heo36e227d2012-08-03 10:30:46 -07001214 /* try to steal the timer if it exists */
1215 if (is_dwork) {
1216 struct delayed_work *dwork = to_delayed_work(work);
1217
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001218 /*
1219 * dwork->timer is irqsafe. If del_timer() fails, it's
1220 * guaranteed that the timer is not queued anywhere and not
1221 * running on the local CPU.
1222 */
Tejun Heo36e227d2012-08-03 10:30:46 -07001223 if (likely(del_timer(&dwork->timer)))
1224 return 1;
1225 }
1226
1227 /* try to claim PENDING the normal way */
Tejun Heobf4ede02012-08-03 10:30:46 -07001228 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1229 return 0;
1230
1231 /*
1232 * The queueing is in progress, or it is already queued. Try to
1233 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1234 */
Tejun Heod565ed62013-01-24 11:01:33 -08001235 pool = get_work_pool(work);
1236 if (!pool)
Tejun Heobbb68df2012-08-03 10:30:46 -07001237 goto fail;
Tejun Heobf4ede02012-08-03 10:30:46 -07001238
Tejun Heod565ed62013-01-24 11:01:33 -08001239 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001240 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001241 * work->data is guaranteed to point to pwq only while the work
1242 * item is queued on pwq->wq, and both updating work->data to point
1243 * to pwq on queueing and to pool on dequeueing are done under
1244 * pwq->pool->lock. This in turn guarantees that, if work->data
1245 * points to pwq which is associated with a locked pool, the work
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001246 * item is currently queued on that pool.
1247 */
Tejun Heo112202d2013-02-13 19:29:12 -08001248 pwq = get_work_pwq(work);
1249 if (pwq && pwq->pool == pool) {
Tejun Heo16062832013-02-06 18:04:53 -08001250 debug_work_deactivate(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001251
Tejun Heo16062832013-02-06 18:04:53 -08001252 /*
1253 * A delayed work item cannot be grabbed directly because
1254 * it might have linked NO_COLOR work items which, if left
Tejun Heo112202d2013-02-13 19:29:12 -08001255 * on the delayed_list, will confuse pwq->nr_active
Tejun Heo16062832013-02-06 18:04:53 -08001256 * management later on and cause stall. Make sure the work
1257 * item is activated before grabbing.
1258 */
1259 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
Tejun Heo112202d2013-02-13 19:29:12 -08001260 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001261
Tejun Heo16062832013-02-06 18:04:53 -08001262 list_del_init(&work->entry);
Lai Jiangshan9c34a702014-07-11 00:11:13 +08001263 pwq_dec_nr_in_flight(pwq, get_work_color(work));
Tejun Heo36e227d2012-08-03 10:30:46 -07001264
Tejun Heo112202d2013-02-13 19:29:12 -08001265 /* work->data points to pwq iff queued, point to pool */
Tejun Heo16062832013-02-06 18:04:53 -08001266 set_work_pool_and_keep_pending(work, pool->id);
Lai Jiangshan4468a002013-02-06 18:04:53 -08001267
Tejun Heo16062832013-02-06 18:04:53 -08001268 spin_unlock(&pool->lock);
1269 return 1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001270 }
Tejun Heod565ed62013-01-24 11:01:33 -08001271 spin_unlock(&pool->lock);
Tejun Heobbb68df2012-08-03 10:30:46 -07001272fail:
1273 local_irq_restore(*flags);
1274 if (work_is_canceling(work))
1275 return -ENOENT;
1276 cpu_relax();
Tejun Heo36e227d2012-08-03 10:30:46 -07001277 return -EAGAIN;
Tejun Heobf4ede02012-08-03 10:30:46 -07001278}
1279
1280/**
Tejun Heo706026c2013-01-24 11:01:34 -08001281 * insert_work - insert a work into a pool
Tejun Heo112202d2013-02-13 19:29:12 -08001282 * @pwq: pwq @work belongs to
Tejun Heo4690c4a2010-06-29 10:07:10 +02001283 * @work: work to insert
1284 * @head: insertion point
1285 * @extra_flags: extra WORK_STRUCT_* flags to set
1286 *
Tejun Heo112202d2013-02-13 19:29:12 -08001287 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
Tejun Heo706026c2013-01-24 11:01:34 -08001288 * work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +02001289 *
1290 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001291 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02001292 */
Tejun Heo112202d2013-02-13 19:29:12 -08001293static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1294 struct list_head *head, unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001295{
Tejun Heo112202d2013-02-13 19:29:12 -08001296 struct worker_pool *pool = pwq->pool;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01001297
Tejun Heo4690c4a2010-06-29 10:07:10 +02001298 /* we own @work, set data and link */
Tejun Heo112202d2013-02-13 19:29:12 -08001299 set_work_pwq(work, pwq, extra_flags);
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -07001300 list_add_tail(&work->entry, head);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001301 get_pwq(pwq);
Tejun Heoe22bee72010-06-29 10:07:14 +02001302
1303 /*
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001304 * Ensure either wq_worker_sleeping() sees the above
1305 * list_add_tail() or we see zero nr_running to avoid workers lying
1306 * around lazily while there are works to be processed.
Tejun Heoe22bee72010-06-29 10:07:14 +02001307 */
1308 smp_mb();
1309
Tejun Heo63d95a92012-07-12 14:46:37 -07001310 if (__need_more_worker(pool))
1311 wake_up_worker(pool);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001312}
1313
Tejun Heoc8efcc22010-12-20 19:32:04 +01001314/*
1315 * Test whether @work is being queued from another work executing on the
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001316 * same workqueue.
Tejun Heoc8efcc22010-12-20 19:32:04 +01001317 */
1318static bool is_chained_work(struct workqueue_struct *wq)
1319{
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001320 struct worker *worker;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001321
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001322 worker = current_wq_worker();
1323 /*
1324 * Return %true iff I'm a worker execuing a work item on @wq. If
1325 * I'm @worker, it's safe to dereference it without locking.
1326 */
Tejun Heo112202d2013-02-13 19:29:12 -08001327 return worker && worker->current_pwq->wq == wq;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001328}
1329
Mike Galbraithef5571802016-02-09 17:59:38 -05001330/*
1331 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1332 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1333 * avoid perturbing sensitive tasks.
1334 */
1335static int wq_select_unbound_cpu(int cpu)
1336{
Tejun Heof303fccb2016-02-09 17:59:38 -05001337 static bool printed_dbg_warning;
Mike Galbraithef5571802016-02-09 17:59:38 -05001338 int new_cpu;
1339
Tejun Heof303fccb2016-02-09 17:59:38 -05001340 if (likely(!wq_debug_force_rr_cpu)) {
1341 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1342 return cpu;
1343 } else if (!printed_dbg_warning) {
1344 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1345 printed_dbg_warning = true;
1346 }
1347
Mike Galbraithef5571802016-02-09 17:59:38 -05001348 if (cpumask_empty(wq_unbound_cpumask))
1349 return cpu;
1350
1351 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1352 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1353 if (unlikely(new_cpu >= nr_cpu_ids)) {
1354 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1355 if (unlikely(new_cpu >= nr_cpu_ids))
1356 return cpu;
1357 }
1358 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1359
1360 return new_cpu;
1361}
1362
Tejun Heod84ff052013-03-12 11:29:59 -07001363static void __queue_work(int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364 struct work_struct *work)
1365{
Tejun Heo112202d2013-02-13 19:29:12 -08001366 struct pool_workqueue *pwq;
Tejun Heoc9178082013-03-12 11:30:04 -07001367 struct worker_pool *last_pool;
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001368 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001369 unsigned int work_flags;
Joonsoo Kimb75cac92012-08-15 23:25:37 +09001370 unsigned int req_cpu = cpu;
Tejun Heo8930cab2012-08-03 10:30:45 -07001371
1372 /*
1373 * While a work item is PENDING && off queue, a task trying to
1374 * steal the PENDING will busy-loop waiting for it to either get
1375 * queued or lose PENDING. Grabbing PENDING and queueing should
1376 * happen with IRQ disabled.
1377 */
1378 WARN_ON_ONCE(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09001380 debug_work_activate(work);
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001381
Li Bin9ef28a72013-09-09 13:13:58 +08001382 /* if draining, only works from the same workqueue are allowed */
Tejun Heo618b01e2013-03-12 11:30:04 -07001383 if (unlikely(wq->flags & __WQ_DRAINING) &&
Tejun Heoc8efcc22010-12-20 19:32:04 +01001384 WARN_ON_ONCE(!is_chained_work(wq)))
Tejun Heoe41e7042010-08-24 14:22:47 +02001385 return;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001386retry:
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001387 if (req_cpu == WORK_CPU_UNBOUND)
Mike Galbraithef5571802016-02-09 17:59:38 -05001388 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001389
Tejun Heoc9178082013-03-12 11:30:04 -07001390 /* pwq which will be used unless @work is executing elsewhere */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001391 if (!(wq->flags & WQ_UNBOUND))
Tejun Heo7fb98ea2013-03-12 11:30:00 -07001392 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001393 else
1394 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodbf25762012-08-20 14:51:23 -07001395
Tejun Heoc9178082013-03-12 11:30:04 -07001396 /*
1397 * If @work was previously on a different pool, it might still be
1398 * running there, in which case the work needs to be queued on that
1399 * pool to guarantee non-reentrancy.
1400 */
1401 last_pool = get_work_pool(work);
1402 if (last_pool && last_pool != pwq->pool) {
1403 struct worker *worker;
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001404
Tejun Heoc9178082013-03-12 11:30:04 -07001405 spin_lock(&last_pool->lock);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001406
Tejun Heoc9178082013-03-12 11:30:04 -07001407 worker = find_worker_executing_work(last_pool, work);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001408
Tejun Heoc9178082013-03-12 11:30:04 -07001409 if (worker && worker->current_pwq->wq == wq) {
1410 pwq = worker->current_pwq;
Tejun Heo8930cab2012-08-03 10:30:45 -07001411 } else {
Tejun Heoc9178082013-03-12 11:30:04 -07001412 /* meh... not running there, queue here */
1413 spin_unlock(&last_pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08001414 spin_lock(&pwq->pool->lock);
Tejun Heo8930cab2012-08-03 10:30:45 -07001415 }
Tejun Heof3421792010-07-02 10:03:51 +02001416 } else {
Tejun Heo112202d2013-02-13 19:29:12 -08001417 spin_lock(&pwq->pool->lock);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001418 }
1419
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001420 /*
1421 * pwq is determined and locked. For unbound pools, we could have
1422 * raced with pwq release and it could already be dead. If its
1423 * refcnt is zero, repeat pwq selection. Note that pwqs never die
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001424 * without another pwq replacing it in the numa_pwq_tbl or while
1425 * work items are executing on it, so the retrying is guaranteed to
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001426 * make forward-progress.
1427 */
1428 if (unlikely(!pwq->refcnt)) {
1429 if (wq->flags & WQ_UNBOUND) {
1430 spin_unlock(&pwq->pool->lock);
1431 cpu_relax();
1432 goto retry;
1433 }
1434 /* oops */
1435 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1436 wq->name, cpu);
1437 }
1438
Tejun Heo112202d2013-02-13 19:29:12 -08001439 /* pwq determined, queue */
1440 trace_workqueue_queue_work(req_cpu, pwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001441
Dan Carpenterf5b25522012-04-13 22:06:58 +03001442 if (WARN_ON(!list_empty(&work->entry))) {
Tejun Heo112202d2013-02-13 19:29:12 -08001443 spin_unlock(&pwq->pool->lock);
Dan Carpenterf5b25522012-04-13 22:06:58 +03001444 return;
1445 }
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001446
Tejun Heo112202d2013-02-13 19:29:12 -08001447 pwq->nr_in_flight[pwq->work_color]++;
1448 work_flags = work_color_to_flags(pwq->work_color);
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001449
Tejun Heo112202d2013-02-13 19:29:12 -08001450 if (likely(pwq->nr_active < pwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001451 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001452 pwq->nr_active++;
1453 worklist = &pwq->pool->worklist;
Tejun Heo82607adc2015-12-08 11:28:04 -05001454 if (list_empty(worklist))
1455 pwq->pool->watchdog_ts = jiffies;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001456 } else {
1457 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo112202d2013-02-13 19:29:12 -08001458 worklist = &pwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001459 }
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001460
Tejun Heo112202d2013-02-13 19:29:12 -08001461 insert_work(pwq, work, worklist, work_flags);
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001462
Tejun Heo112202d2013-02-13 19:29:12 -08001463 spin_unlock(&pwq->pool->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464}
1465
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001466/**
Zhang Ruic1a220e2008-07-23 21:28:39 -07001467 * queue_work_on - queue work on specific cpu
1468 * @cpu: CPU number to execute work on
1469 * @wq: workqueue to use
1470 * @work: work to queue
1471 *
Zhang Ruic1a220e2008-07-23 21:28:39 -07001472 * We queue the work to a specific CPU, the caller must ensure it
1473 * can't go away.
Yacine Belkadid185af32013-07-31 14:59:24 -07001474 *
1475 * Return: %false if @work was already on a queue, %true otherwise.
Zhang Ruic1a220e2008-07-23 21:28:39 -07001476 */
Tejun Heod4283e92012-08-03 10:30:44 -07001477bool queue_work_on(int cpu, struct workqueue_struct *wq,
1478 struct work_struct *work)
Zhang Ruic1a220e2008-07-23 21:28:39 -07001479{
Tejun Heod4283e92012-08-03 10:30:44 -07001480 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001481 unsigned long flags;
1482
1483 local_irq_save(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001484
Tejun Heo22df02b2010-06-29 10:07:10 +02001485 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001486 __queue_work(cpu, wq, work);
Tejun Heod4283e92012-08-03 10:30:44 -07001487 ret = true;
Zhang Ruic1a220e2008-07-23 21:28:39 -07001488 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001489
1490 local_irq_restore(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001491 return ret;
1492}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001493EXPORT_SYMBOL(queue_work_on);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001494
Tejun Heod8e794d2012-08-03 10:30:45 -07001495void delayed_work_timer_fn(unsigned long __data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496{
David Howells52bad642006-11-22 14:54:01 +00001497 struct delayed_work *dwork = (struct delayed_work *)__data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001499 /* should have been called from irqsafe timer with irq already off */
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001500 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501}
Konstantin Khlebnikov1438ade52013-01-24 16:36:31 +04001502EXPORT_SYMBOL(delayed_work_timer_fn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001504static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1505 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506{
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001507 struct timer_list *timer = &dwork->timer;
1508 struct work_struct *work = &dwork->work;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001510 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1511 timer->data != (unsigned long)dwork);
Tejun Heofc4b5142012-12-04 07:40:39 -08001512 WARN_ON_ONCE(timer_pending(timer));
1513 WARN_ON_ONCE(!list_empty(&work->entry));
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001514
Tejun Heo8852aac2012-12-01 16:23:42 -08001515 /*
1516 * If @delay is 0, queue @dwork->work immediately. This is for
1517 * both optimization and correctness. The earliest @timer can
1518 * expire is on the closest next tick and delayed_work users depend
1519 * on that there's no such delay when @delay is 0.
1520 */
1521 if (!delay) {
1522 __queue_work(cpu, wq, &dwork->work);
1523 return;
1524 }
1525
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001526 dwork->wq = wq;
Tejun Heo12650572012-08-08 09:38:42 -07001527 dwork->cpu = cpu;
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001528 timer->expires = jiffies + delay;
1529
Tejun Heo041bd122016-02-09 16:11:26 -05001530 if (unlikely(cpu != WORK_CPU_UNBOUND))
1531 add_timer_on(timer, cpu);
1532 else
1533 add_timer(timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534}
1535
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001536/**
1537 * queue_delayed_work_on - queue work on specific CPU after delay
1538 * @cpu: CPU number to execute work on
1539 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001540 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001541 * @delay: number of jiffies to wait before queueing
1542 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001543 * Return: %false if @work was already on a queue, %true otherwise. If
Tejun Heo715f1302012-08-03 10:30:46 -07001544 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1545 * execution.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001546 */
Tejun Heod4283e92012-08-03 10:30:44 -07001547bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1548 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001549{
David Howells52bad642006-11-22 14:54:01 +00001550 struct work_struct *work = &dwork->work;
Tejun Heod4283e92012-08-03 10:30:44 -07001551 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001552 unsigned long flags;
1553
1554 /* read the comment in __queue_work() */
1555 local_irq_save(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001556
Tejun Heo22df02b2010-06-29 10:07:10 +02001557 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001558 __queue_delayed_work(cpu, wq, dwork, delay);
Tejun Heod4283e92012-08-03 10:30:44 -07001559 ret = true;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001560 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001561
1562 local_irq_restore(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001563 return ret;
1564}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001565EXPORT_SYMBOL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566
Tejun Heoc8e55f32010-06-29 10:07:12 +02001567/**
Tejun Heo8376fe22012-08-03 10:30:47 -07001568 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1569 * @cpu: CPU number to execute work on
1570 * @wq: workqueue to use
1571 * @dwork: work to queue
1572 * @delay: number of jiffies to wait before queueing
1573 *
1574 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1575 * modify @dwork's timer so that it expires after @delay. If @delay is
1576 * zero, @work is guaranteed to be scheduled immediately regardless of its
1577 * current state.
1578 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001579 * Return: %false if @dwork was idle and queued, %true if @dwork was
Tejun Heo8376fe22012-08-03 10:30:47 -07001580 * pending and its timer was modified.
1581 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001582 * This function is safe to call from any context including IRQ handler.
Tejun Heo8376fe22012-08-03 10:30:47 -07001583 * See try_to_grab_pending() for details.
1584 */
1585bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1586 struct delayed_work *dwork, unsigned long delay)
1587{
1588 unsigned long flags;
1589 int ret;
1590
1591 do {
1592 ret = try_to_grab_pending(&dwork->work, true, &flags);
1593 } while (unlikely(ret == -EAGAIN));
1594
1595 if (likely(ret >= 0)) {
1596 __queue_delayed_work(cpu, wq, dwork, delay);
1597 local_irq_restore(flags);
1598 }
1599
1600 /* -ENOENT from try_to_grab_pending() becomes %true */
1601 return ret;
1602}
1603EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1604
1605/**
Tejun Heoc8e55f32010-06-29 10:07:12 +02001606 * worker_enter_idle - enter idle state
1607 * @worker: worker which is entering idle state
1608 *
1609 * @worker is entering idle state. Update stats and idle timer if
1610 * necessary.
1611 *
1612 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001613 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001614 */
1615static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001617 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618
Tejun Heo6183c002013-03-12 11:29:57 -07001619 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1620 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1621 (worker->hentry.next || worker->hentry.pprev)))
1622 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
Lai Jiangshan051e1852014-07-22 13:03:02 +08001624 /* can't use worker_set_flags(), also called from create_worker() */
Tejun Heocb444762010-07-02 10:03:50 +02001625 worker->flags |= WORKER_IDLE;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001626 pool->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001627 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001628
Tejun Heoc8e55f32010-06-29 10:07:12 +02001629 /* idle_list is LIFO */
Tejun Heobd7bdd42012-07-12 14:46:37 -07001630 list_add(&worker->entry, &pool->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001631
Tejun Heo628c78e2012-07-17 12:39:27 -07001632 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1633 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
Tejun Heocb444762010-07-02 10:03:50 +02001634
Tejun Heo544ecf32012-05-14 15:04:50 -07001635 /*
Tejun Heo706026c2013-01-24 11:01:34 -08001636 * Sanity check nr_running. Because wq_unbind_fn() releases
Tejun Heod565ed62013-01-24 11:01:33 -08001637 * pool->lock between setting %WORKER_UNBOUND and zapping
Tejun Heo628c78e2012-07-17 12:39:27 -07001638 * nr_running, the warning may trigger spuriously. Check iff
1639 * unbind is not in progress.
Tejun Heo544ecf32012-05-14 15:04:50 -07001640 */
Tejun Heo24647572013-01-24 11:01:33 -08001641 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -07001642 pool->nr_workers == pool->nr_idle &&
Tejun Heoe19e3972013-01-24 11:39:44 -08001643 atomic_read(&pool->nr_running));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644}
1645
Tejun Heoc8e55f32010-06-29 10:07:12 +02001646/**
1647 * worker_leave_idle - leave idle state
1648 * @worker: worker which is leaving idle state
1649 *
1650 * @worker is leaving idle state. Update stats.
1651 *
1652 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001653 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001654 */
1655static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001657 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658
Tejun Heo6183c002013-03-12 11:29:57 -07001659 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1660 return;
Tejun Heod302f012010-06-29 10:07:13 +02001661 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001662 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001663 list_del_init(&worker->entry);
1664}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001666static struct worker *alloc_worker(int node)
Tejun Heoc34056a2010-06-29 10:07:11 +02001667{
1668 struct worker *worker;
1669
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001670 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001671 if (worker) {
1672 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001673 INIT_LIST_HEAD(&worker->scheduled);
Lai Jiangshanda028462014-05-20 17:46:31 +08001674 INIT_LIST_HEAD(&worker->node);
Tejun Heoe22bee72010-06-29 10:07:14 +02001675 /* on creation a worker is in !idle && prep state */
1676 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001677 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001678 return worker;
1679}
1680
1681/**
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001682 * worker_attach_to_pool() - attach a worker to a pool
1683 * @worker: worker to be attached
1684 * @pool: the target pool
1685 *
1686 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1687 * cpu-binding of @worker are kept coordinated with the pool across
1688 * cpu-[un]hotplugs.
1689 */
1690static void worker_attach_to_pool(struct worker *worker,
1691 struct worker_pool *pool)
1692{
1693 mutex_lock(&pool->attach_mutex);
1694
1695 /*
1696 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1697 * online CPUs. It'll be re-applied when any of the CPUs come up.
1698 */
1699 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1700
1701 /*
1702 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1703 * stable across this function. See the comments above the
1704 * flag definition for details.
1705 */
1706 if (pool->flags & POOL_DISASSOCIATED)
1707 worker->flags |= WORKER_UNBOUND;
1708
1709 list_add_tail(&worker->node, &pool->workers);
1710
1711 mutex_unlock(&pool->attach_mutex);
1712}
1713
1714/**
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001715 * worker_detach_from_pool() - detach a worker from its pool
1716 * @worker: worker which is attached to its pool
1717 * @pool: the pool @worker is attached to
1718 *
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001719 * Undo the attaching which had been done in worker_attach_to_pool(). The
1720 * caller worker shouldn't access to the pool after detached except it has
1721 * other reference to the pool.
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001722 */
1723static void worker_detach_from_pool(struct worker *worker,
1724 struct worker_pool *pool)
1725{
1726 struct completion *detach_completion = NULL;
1727
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001728 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08001729 list_del(&worker->node);
1730 if (list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001731 detach_completion = pool->detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001732 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001733
Lai Jiangshanb62c0752014-06-03 15:32:52 +08001734 /* clear leftover flags without pool->lock after it is detached */
1735 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1736
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001737 if (detach_completion)
1738 complete(detach_completion);
1739}
1740
1741/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001742 * create_worker - create a new workqueue worker
Tejun Heo63d95a92012-07-12 14:46:37 -07001743 * @pool: pool the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001744 *
Lai Jiangshan051e1852014-07-22 13:03:02 +08001745 * Create and start a new worker which is attached to @pool.
Tejun Heoc34056a2010-06-29 10:07:11 +02001746 *
1747 * CONTEXT:
1748 * Might sleep. Does GFP_KERNEL allocations.
1749 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001750 * Return:
Tejun Heoc34056a2010-06-29 10:07:11 +02001751 * Pointer to the newly created worker.
1752 */
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001753static struct worker *create_worker(struct worker_pool *pool)
Tejun Heoc34056a2010-06-29 10:07:11 +02001754{
Tejun Heoc34056a2010-06-29 10:07:11 +02001755 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001756 int id = -1;
Tejun Heoe3c916a2013-04-01 11:23:32 -07001757 char id_buf[16];
Tejun Heoc34056a2010-06-29 10:07:11 +02001758
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001759 /* ID is needed to determine kthread name */
1760 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
Tejun Heo822d8402013-03-19 13:45:21 -07001761 if (id < 0)
1762 goto fail;
Tejun Heoc34056a2010-06-29 10:07:11 +02001763
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001764 worker = alloc_worker(pool->node);
Tejun Heoc34056a2010-06-29 10:07:11 +02001765 if (!worker)
1766 goto fail;
1767
Tejun Heobd7bdd42012-07-12 14:46:37 -07001768 worker->pool = pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001769 worker->id = id;
1770
Tejun Heo29c91e92013-03-12 11:30:03 -07001771 if (pool->cpu >= 0)
Tejun Heoe3c916a2013-04-01 11:23:32 -07001772 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1773 pool->attrs->nice < 0 ? "H" : "");
Tejun Heof3421792010-07-02 10:03:51 +02001774 else
Tejun Heoe3c916a2013-04-01 11:23:32 -07001775 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1776
Tejun Heof3f90ad2013-04-01 11:23:34 -07001777 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
Tejun Heoe3c916a2013-04-01 11:23:32 -07001778 "kworker/%s", id_buf);
Tejun Heoc34056a2010-06-29 10:07:11 +02001779 if (IS_ERR(worker->task))
1780 goto fail;
1781
Oleg Nesterov91151222013-11-14 12:56:18 +01001782 set_user_nice(worker->task, pool->attrs->nice);
Peter Zijlstra25834c72015-05-15 17:43:34 +02001783 kthread_bind_mask(worker->task, pool->attrs->cpumask);
Oleg Nesterov91151222013-11-14 12:56:18 +01001784
Lai Jiangshanda028462014-05-20 17:46:31 +08001785 /* successful, attach the worker to the pool */
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001786 worker_attach_to_pool(worker, pool);
Tejun Heo822d8402013-03-19 13:45:21 -07001787
Lai Jiangshan051e1852014-07-22 13:03:02 +08001788 /* start the newly created worker */
1789 spin_lock_irq(&pool->lock);
1790 worker->pool->nr_workers++;
1791 worker_enter_idle(worker);
1792 wake_up_process(worker->task);
1793 spin_unlock_irq(&pool->lock);
1794
Tejun Heoc34056a2010-06-29 10:07:11 +02001795 return worker;
Tejun Heo822d8402013-03-19 13:45:21 -07001796
Tejun Heoc34056a2010-06-29 10:07:11 +02001797fail:
Lai Jiangshan9625ab12014-05-20 17:46:27 +08001798 if (id >= 0)
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001799 ida_simple_remove(&pool->worker_ida, id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001800 kfree(worker);
1801 return NULL;
1802}
1803
1804/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001805 * destroy_worker - destroy a workqueue worker
1806 * @worker: worker to be destroyed
1807 *
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001808 * Destroy @worker and adjust @pool stats accordingly. The worker should
1809 * be idle.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001810 *
1811 * CONTEXT:
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001812 * spin_lock_irq(pool->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001813 */
1814static void destroy_worker(struct worker *worker)
1815{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001816 struct worker_pool *pool = worker->pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001817
Tejun Heocd549682013-03-13 19:47:39 -07001818 lockdep_assert_held(&pool->lock);
1819
Tejun Heoc34056a2010-06-29 10:07:11 +02001820 /* sanity check frenzy */
Tejun Heo6183c002013-03-12 11:29:57 -07001821 if (WARN_ON(worker->current_work) ||
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001822 WARN_ON(!list_empty(&worker->scheduled)) ||
1823 WARN_ON(!(worker->flags & WORKER_IDLE)))
Tejun Heo6183c002013-03-12 11:29:57 -07001824 return;
Tejun Heoc34056a2010-06-29 10:07:11 +02001825
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001826 pool->nr_workers--;
1827 pool->nr_idle--;
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001828
Tejun Heoc8e55f32010-06-29 10:07:12 +02001829 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001830 worker->flags |= WORKER_DIE;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001831 wake_up_process(worker->task);
Tejun Heoc34056a2010-06-29 10:07:11 +02001832}
1833
Tejun Heo63d95a92012-07-12 14:46:37 -07001834static void idle_worker_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001835{
Tejun Heo63d95a92012-07-12 14:46:37 -07001836 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001837
Tejun Heod565ed62013-01-24 11:01:33 -08001838 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001839
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001840 while (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001841 struct worker *worker;
1842 unsigned long expires;
1843
1844 /* idle_list is kept in LIFO order, check the last one */
Tejun Heo63d95a92012-07-12 14:46:37 -07001845 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02001846 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1847
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001848 if (time_before(jiffies, expires)) {
Tejun Heo63d95a92012-07-12 14:46:37 -07001849 mod_timer(&pool->idle_timer, expires);
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001850 break;
Tejun Heoe22bee72010-06-29 10:07:14 +02001851 }
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001852
1853 destroy_worker(worker);
Tejun Heoe22bee72010-06-29 10:07:14 +02001854 }
1855
Tejun Heod565ed62013-01-24 11:01:33 -08001856 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001857}
1858
Tejun Heo493a1722013-03-12 11:29:59 -07001859static void send_mayday(struct work_struct *work)
Tejun Heoe22bee72010-06-29 10:07:14 +02001860{
Tejun Heo112202d2013-02-13 19:29:12 -08001861 struct pool_workqueue *pwq = get_work_pwq(work);
1862 struct workqueue_struct *wq = pwq->wq;
Tejun Heo493a1722013-03-12 11:29:59 -07001863
Tejun Heo2e109a22013-03-13 19:47:40 -07001864 lockdep_assert_held(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001865
Tejun Heo493008a2013-03-12 11:30:03 -07001866 if (!wq->rescuer)
Tejun Heo493a1722013-03-12 11:29:59 -07001867 return;
Tejun Heoe22bee72010-06-29 10:07:14 +02001868
1869 /* mayday mayday mayday */
Tejun Heo493a1722013-03-12 11:29:59 -07001870 if (list_empty(&pwq->mayday_node)) {
Lai Jiangshan77668c82014-04-18 11:04:16 -04001871 /*
1872 * If @pwq is for an unbound wq, its base ref may be put at
1873 * any time due to an attribute change. Pin @pwq until the
1874 * rescuer is done with it.
1875 */
1876 get_pwq(pwq);
Tejun Heo493a1722013-03-12 11:29:59 -07001877 list_add_tail(&pwq->mayday_node, &wq->maydays);
Tejun Heoe22bee72010-06-29 10:07:14 +02001878 wake_up_process(wq->rescuer->task);
Tejun Heo493a1722013-03-12 11:29:59 -07001879 }
Tejun Heoe22bee72010-06-29 10:07:14 +02001880}
1881
Tejun Heo706026c2013-01-24 11:01:34 -08001882static void pool_mayday_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001883{
Tejun Heo63d95a92012-07-12 14:46:37 -07001884 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001885 struct work_struct *work;
1886
Tejun Heob2d82902014-12-08 12:39:16 -05001887 spin_lock_irq(&pool->lock);
1888 spin_lock(&wq_mayday_lock); /* for wq->maydays */
Tejun Heoe22bee72010-06-29 10:07:14 +02001889
Tejun Heo63d95a92012-07-12 14:46:37 -07001890 if (need_to_create_worker(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001891 /*
1892 * We've been trying to create a new worker but
1893 * haven't been successful. We might be hitting an
1894 * allocation deadlock. Send distress signals to
1895 * rescuers.
1896 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001897 list_for_each_entry(work, &pool->worklist, entry)
Tejun Heoe22bee72010-06-29 10:07:14 +02001898 send_mayday(work);
1899 }
1900
Tejun Heob2d82902014-12-08 12:39:16 -05001901 spin_unlock(&wq_mayday_lock);
1902 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001903
Tejun Heo63d95a92012-07-12 14:46:37 -07001904 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
Tejun Heoe22bee72010-06-29 10:07:14 +02001905}
1906
1907/**
1908 * maybe_create_worker - create a new worker if necessary
Tejun Heo63d95a92012-07-12 14:46:37 -07001909 * @pool: pool to create a new worker for
Tejun Heoe22bee72010-06-29 10:07:14 +02001910 *
Tejun Heo63d95a92012-07-12 14:46:37 -07001911 * Create a new worker for @pool if necessary. @pool is guaranteed to
Tejun Heoe22bee72010-06-29 10:07:14 +02001912 * have at least one idle worker on return from this function. If
1913 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
Tejun Heo63d95a92012-07-12 14:46:37 -07001914 * sent to all rescuers with works scheduled on @pool to resolve
Tejun Heoe22bee72010-06-29 10:07:14 +02001915 * possible allocation deadlock.
1916 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001917 * On return, need_to_create_worker() is guaranteed to be %false and
1918 * may_start_working() %true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001919 *
1920 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001921 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001922 * multiple times. Does GFP_KERNEL allocations. Called only from
1923 * manager.
Tejun Heoe22bee72010-06-29 10:07:14 +02001924 */
Tejun Heo29187a92015-01-16 14:21:16 -05001925static void maybe_create_worker(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08001926__releases(&pool->lock)
1927__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001928{
Tejun Heoe22bee72010-06-29 10:07:14 +02001929restart:
Tejun Heod565ed62013-01-24 11:01:33 -08001930 spin_unlock_irq(&pool->lock);
Tejun Heo9f9c236442010-07-14 11:31:20 +02001931
Tejun Heoe22bee72010-06-29 10:07:14 +02001932 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
Tejun Heo63d95a92012-07-12 14:46:37 -07001933 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
Tejun Heoe22bee72010-06-29 10:07:14 +02001934
1935 while (true) {
Lai Jiangshan051e1852014-07-22 13:03:02 +08001936 if (create_worker(pool) || !need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001937 break;
1938
Lai Jiangshane212f362014-06-03 15:32:17 +08001939 schedule_timeout_interruptible(CREATE_COOLDOWN);
Tejun Heo9f9c236442010-07-14 11:31:20 +02001940
Tejun Heo63d95a92012-07-12 14:46:37 -07001941 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001942 break;
1943 }
1944
Tejun Heo63d95a92012-07-12 14:46:37 -07001945 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08001946 spin_lock_irq(&pool->lock);
Lai Jiangshan051e1852014-07-22 13:03:02 +08001947 /*
1948 * This is necessary even after a new worker was just successfully
1949 * created as @pool->lock was dropped and the new worker might have
1950 * already become busy.
1951 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001952 if (need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001953 goto restart;
Tejun Heoe22bee72010-06-29 10:07:14 +02001954}
1955
1956/**
Tejun Heoe22bee72010-06-29 10:07:14 +02001957 * manage_workers - manage worker pool
1958 * @worker: self
1959 *
Tejun Heo706026c2013-01-24 11:01:34 -08001960 * Assume the manager role and manage the worker pool @worker belongs
Tejun Heoe22bee72010-06-29 10:07:14 +02001961 * to. At any given time, there can be only zero or one manager per
Tejun Heo706026c2013-01-24 11:01:34 -08001962 * pool. The exclusion is handled automatically by this function.
Tejun Heoe22bee72010-06-29 10:07:14 +02001963 *
1964 * The caller can safely start processing works on false return. On
1965 * true return, it's guaranteed that need_to_create_worker() is false
1966 * and may_start_working() is true.
1967 *
1968 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001969 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001970 * multiple times. Does GFP_KERNEL allocations.
1971 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001972 * Return:
Tejun Heo29187a92015-01-16 14:21:16 -05001973 * %false if the pool doesn't need management and the caller can safely
1974 * start processing works, %true if management function was performed and
1975 * the conditions that the caller verified before calling the function may
1976 * no longer be true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001977 */
1978static bool manage_workers(struct worker *worker)
1979{
Tejun Heo63d95a92012-07-12 14:46:37 -07001980 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001981
Tejun Heobc3a1af2013-03-13 19:47:39 -07001982 /*
Tejun Heobc3a1af2013-03-13 19:47:39 -07001983 * Anyone who successfully grabs manager_arb wins the arbitration
1984 * and becomes the manager. mutex_trylock() on pool->manager_arb
1985 * failure while holding pool->lock reliably indicates that someone
1986 * else is managing the pool and the worker which failed trylock
1987 * can proceed to executing work items. This means that anyone
1988 * grabbing manager_arb is responsible for actually performing
1989 * manager duties. If manager_arb is grabbed and released without
1990 * actual management, the pool may stall indefinitely.
Tejun Heobc3a1af2013-03-13 19:47:39 -07001991 */
Tejun Heo34a06bd2013-03-12 11:30:00 -07001992 if (!mutex_trylock(&pool->manager_arb))
Tejun Heo29187a92015-01-16 14:21:16 -05001993 return false;
Tejun Heo2607d7a2015-03-09 09:22:28 -04001994 pool->manager = worker;
Tejun Heoe22bee72010-06-29 10:07:14 +02001995
Tejun Heo29187a92015-01-16 14:21:16 -05001996 maybe_create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02001997
Tejun Heo2607d7a2015-03-09 09:22:28 -04001998 pool->manager = NULL;
Tejun Heo34a06bd2013-03-12 11:30:00 -07001999 mutex_unlock(&pool->manager_arb);
Tejun Heo29187a92015-01-16 14:21:16 -05002000 return true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002001}
2002
Tejun Heoa62428c2010-06-29 10:07:10 +02002003/**
2004 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02002005 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02002006 * @work: work to process
2007 *
2008 * Process @work. This function contains all the logics necessary to
2009 * process a single work including synchronization against and
2010 * interaction with other workers on the same cpu, queueing and
2011 * flushing. As long as context requirement is met, any worker can
2012 * call this function to process a work.
2013 *
2014 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002015 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02002016 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002017static void process_one_work(struct worker *worker, struct work_struct *work)
Tejun Heod565ed62013-01-24 11:01:33 -08002018__releases(&pool->lock)
2019__acquires(&pool->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02002020{
Tejun Heo112202d2013-02-13 19:29:12 -08002021 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobd7bdd42012-07-12 14:46:37 -07002022 struct worker_pool *pool = worker->pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002023 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heo73f53c42010-06-29 10:07:11 +02002024 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02002025 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02002026#ifdef CONFIG_LOCKDEP
2027 /*
2028 * It is permissible to free the struct work_struct from
2029 * inside the function that is called from it, this we need to
2030 * take into account for lockdep too. To avoid bogus "held
2031 * lock freed" warnings as well as problems when looking into
2032 * work->lockdep_map, make a copy and use that here.
2033 */
Peter Zijlstra4d82a1d2012-05-15 08:06:19 -07002034 struct lockdep_map lockdep_map;
2035
2036 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002037#endif
Lai Jiangshan807407c2014-06-03 15:33:28 +08002038 /* ensure we're on the correct CPU */
Lai Jiangshan85327af2014-06-03 15:33:28 +08002039 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heoec22ca52013-01-24 11:01:33 -08002040 raw_smp_processor_id() != pool->cpu);
Tejun Heo25511a42012-07-17 12:39:27 -07002041
Tejun Heo7e116292010-06-29 10:07:13 +02002042 /*
2043 * A single work shouldn't be executed concurrently by
2044 * multiple workers on a single cpu. Check whether anyone is
2045 * already processing the work. If so, defer the work to the
2046 * currently executing one.
2047 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002048 collision = find_worker_executing_work(pool, work);
Tejun Heo7e116292010-06-29 10:07:13 +02002049 if (unlikely(collision)) {
2050 move_linked_works(work, &collision->scheduled, NULL);
2051 return;
2052 }
2053
Tejun Heo8930cab2012-08-03 10:30:45 -07002054 /* claim and dequeue */
Tejun Heoa62428c2010-06-29 10:07:10 +02002055 debug_work_deactivate(work);
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002056 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
Tejun Heoc34056a2010-06-29 10:07:11 +02002057 worker->current_work = work;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002058 worker->current_func = work->func;
Tejun Heo112202d2013-02-13 19:29:12 -08002059 worker->current_pwq = pwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002060 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002061
Tejun Heoa62428c2010-06-29 10:07:10 +02002062 list_del_init(&work->entry);
2063
Tejun Heo649027d2010-06-29 10:07:14 +02002064 /*
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002065 * CPU intensive works don't participate in concurrency management.
2066 * They're the scheduler's responsibility. This takes @worker out
2067 * of concurrency management and the next code block will chain
2068 * execution of the pending work items.
Tejun Heofb0e7be2010-06-29 10:07:15 +02002069 */
2070 if (unlikely(cpu_intensive))
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002071 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
Tejun Heofb0e7be2010-06-29 10:07:15 +02002072
Tejun Heo974271c2012-07-12 14:46:37 -07002073 /*
Lai Jiangshana489a032014-07-22 13:01:59 +08002074 * Wake up another worker if necessary. The condition is always
2075 * false for normal per-cpu workers since nr_running would always
2076 * be >= 1 at this point. This is used to chain execution of the
2077 * pending work items for WORKER_NOT_RUNNING workers such as the
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002078 * UNBOUND and CPU_INTENSIVE ones.
Tejun Heo974271c2012-07-12 14:46:37 -07002079 */
Lai Jiangshana489a032014-07-22 13:01:59 +08002080 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002081 wake_up_worker(pool);
Tejun Heo974271c2012-07-12 14:46:37 -07002082
Tejun Heo8930cab2012-08-03 10:30:45 -07002083 /*
Tejun Heo7c3eed52013-01-24 11:01:33 -08002084 * Record the last pool and clear PENDING which should be the last
Tejun Heod565ed62013-01-24 11:01:33 -08002085 * update to @work. Also, do this inside @pool->lock so that
Tejun Heo23657bb2012-08-13 17:08:19 -07002086 * PENDING and queued state changes happen together while IRQ is
2087 * disabled.
Tejun Heo8930cab2012-08-03 10:30:45 -07002088 */
Tejun Heo7c3eed52013-01-24 11:01:33 -08002089 set_work_pool_and_clear_pending(work, pool->id);
Tejun Heoa62428c2010-06-29 10:07:10 +02002090
Tejun Heod565ed62013-01-24 11:01:33 -08002091 spin_unlock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002092
Tejun Heo112202d2013-02-13 19:29:12 -08002093 lock_map_acquire_read(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002094 lock_map_acquire(&lockdep_map);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002095 trace_workqueue_execute_start(work);
Tejun Heoa2c1c572012-12-18 10:35:02 -08002096 worker->current_func(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002097 /*
2098 * While we must be careful to not use "work" after this, the trace
2099 * point will only record its address.
2100 */
2101 trace_workqueue_execute_end(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02002102 lock_map_release(&lockdep_map);
Tejun Heo112202d2013-02-13 19:29:12 -08002103 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002104
2105 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
Valentin Ilie044c7822012-08-19 00:52:42 +03002106 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2107 " last function: %pf\n",
Tejun Heoa2c1c572012-12-18 10:35:02 -08002108 current->comm, preempt_count(), task_pid_nr(current),
2109 worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002110 debug_show_held_locks(current);
2111 dump_stack();
2112 }
2113
Tejun Heob22ce272013-08-28 17:33:37 -04002114 /*
2115 * The following prevents a kworker from hogging CPU on !PREEMPT
2116 * kernels, where a requeueing work item waiting for something to
2117 * happen could deadlock with stop_machine as such work item could
2118 * indefinitely requeue itself while all other CPUs are trapped in
Joe Lawrence789cbbe2014-10-05 13:24:21 -04002119 * stop_machine. At the same time, report a quiescent RCU state so
2120 * the same condition doesn't freeze RCU.
Tejun Heob22ce272013-08-28 17:33:37 -04002121 */
Joe Lawrence3e28e372014-10-05 13:24:22 -04002122 cond_resched_rcu_qs();
Tejun Heob22ce272013-08-28 17:33:37 -04002123
Tejun Heod565ed62013-01-24 11:01:33 -08002124 spin_lock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002125
Tejun Heofb0e7be2010-06-29 10:07:15 +02002126 /* clear cpu intensive status */
2127 if (unlikely(cpu_intensive))
2128 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2129
Tejun Heoa62428c2010-06-29 10:07:10 +02002130 /* we're done with it, release */
Sasha Levin42f85702012-12-17 10:01:23 -05002131 hash_del(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002132 worker->current_work = NULL;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002133 worker->current_func = NULL;
Tejun Heo112202d2013-02-13 19:29:12 -08002134 worker->current_pwq = NULL;
Tejun Heo3d1cb202013-04-30 15:27:22 -07002135 worker->desc_valid = false;
Tejun Heo112202d2013-02-13 19:29:12 -08002136 pwq_dec_nr_in_flight(pwq, work_color);
Tejun Heoa62428c2010-06-29 10:07:10 +02002137}
2138
Tejun Heoaffee4b2010-06-29 10:07:12 +02002139/**
2140 * process_scheduled_works - process scheduled works
2141 * @worker: self
2142 *
2143 * Process all scheduled works. Please note that the scheduled list
2144 * may change while processing a work, so this function repeatedly
2145 * fetches a work from the top and executes it.
2146 *
2147 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002148 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02002149 * multiple times.
2150 */
2151static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002153 while (!list_empty(&worker->scheduled)) {
2154 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002156 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158}
2159
Tejun Heo4690c4a2010-06-29 10:07:10 +02002160/**
2161 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02002162 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02002163 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002164 * The worker thread function. All workers belong to a worker_pool -
2165 * either a per-cpu one or dynamic unbound one. These workers process all
2166 * work items regardless of their specific target workqueue. The only
2167 * exception is work items which belong to workqueues with a rescuer which
2168 * will be explained in rescuer_thread().
Yacine Belkadid185af32013-07-31 14:59:24 -07002169 *
2170 * Return: 0
Tejun Heo4690c4a2010-06-29 10:07:10 +02002171 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002172static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173{
Tejun Heoc34056a2010-06-29 10:07:11 +02002174 struct worker *worker = __worker;
Tejun Heobd7bdd42012-07-12 14:46:37 -07002175 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176
Tejun Heoe22bee72010-06-29 10:07:14 +02002177 /* tell the scheduler that this is a workqueue worker */
2178 worker->task->flags |= PF_WQ_WORKER;
Tejun Heoc8e55f32010-06-29 10:07:12 +02002179woke_up:
Tejun Heod565ed62013-01-24 11:01:33 -08002180 spin_lock_irq(&pool->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002181
Tejun Heoa9ab7752013-03-19 13:45:21 -07002182 /* am I supposed to die? */
2183 if (unlikely(worker->flags & WORKER_DIE)) {
Tejun Heod565ed62013-01-24 11:01:33 -08002184 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002185 WARN_ON_ONCE(!list_empty(&worker->entry));
2186 worker->task->flags &= ~PF_WQ_WORKER;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002187
2188 set_task_comm(worker->task, "kworker/dying");
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08002189 ida_simple_remove(&pool->worker_ida, worker->id);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002190 worker_detach_from_pool(worker, pool);
2191 kfree(worker);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002192 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07002194
Tejun Heoc8e55f32010-06-29 10:07:12 +02002195 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02002196recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02002197 /* no more worker necessary? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002198 if (!need_more_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002199 goto sleep;
2200
2201 /* do we need to manage? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002202 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002203 goto recheck;
2204
Tejun Heoc8e55f32010-06-29 10:07:12 +02002205 /*
2206 * ->scheduled list can only be filled while a worker is
2207 * preparing to process a work or actually processing it.
2208 * Make sure nobody diddled with it while I was sleeping.
2209 */
Tejun Heo6183c002013-03-12 11:29:57 -07002210 WARN_ON_ONCE(!list_empty(&worker->scheduled));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002211
Tejun Heoe22bee72010-06-29 10:07:14 +02002212 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07002213 * Finish PREP stage. We're guaranteed to have at least one idle
2214 * worker or that someone else has already assumed the manager
2215 * role. This is where @worker starts participating in concurrency
2216 * management if applicable and concurrency management is restored
2217 * after being rebound. See rebind_workers() for details.
Tejun Heoe22bee72010-06-29 10:07:14 +02002218 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07002219 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
Tejun Heoe22bee72010-06-29 10:07:14 +02002220
2221 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02002222 struct work_struct *work =
Tejun Heobd7bdd42012-07-12 14:46:37 -07002223 list_first_entry(&pool->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02002224 struct work_struct, entry);
2225
Tejun Heo82607adc2015-12-08 11:28:04 -05002226 pool->watchdog_ts = jiffies;
2227
Tejun Heoc8e55f32010-06-29 10:07:12 +02002228 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2229 /* optimization path, not strictly necessary */
2230 process_one_work(worker, work);
2231 if (unlikely(!list_empty(&worker->scheduled)))
2232 process_scheduled_works(worker);
2233 } else {
2234 move_linked_works(work, &worker->scheduled, NULL);
2235 process_scheduled_works(worker);
2236 }
Tejun Heo63d95a92012-07-12 14:46:37 -07002237 } while (keep_working(pool));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002238
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002239 worker_set_flags(worker, WORKER_PREP);
Tejun Heod313dd82010-07-02 10:03:51 +02002240sleep:
Tejun Heoc8e55f32010-06-29 10:07:12 +02002241 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002242 * pool->lock is held and there's no work to process and no need to
2243 * manage, sleep. Workers are woken up only while holding
2244 * pool->lock or from local cpu, so setting the current state
2245 * before releasing pool->lock is enough to prevent losing any
2246 * event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02002247 */
2248 worker_enter_idle(worker);
2249 __set_current_state(TASK_INTERRUPTIBLE);
Tejun Heod565ed62013-01-24 11:01:33 -08002250 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002251 schedule();
2252 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253}
2254
Tejun Heoe22bee72010-06-29 10:07:14 +02002255/**
2256 * rescuer_thread - the rescuer thread function
Tejun Heo111c2252013-01-17 17:16:24 -08002257 * @__rescuer: self
Tejun Heoe22bee72010-06-29 10:07:14 +02002258 *
2259 * Workqueue rescuer thread function. There's one rescuer for each
Tejun Heo493008a2013-03-12 11:30:03 -07002260 * workqueue which has WQ_MEM_RECLAIM set.
Tejun Heoe22bee72010-06-29 10:07:14 +02002261 *
Tejun Heo706026c2013-01-24 11:01:34 -08002262 * Regular work processing on a pool may block trying to create a new
Tejun Heoe22bee72010-06-29 10:07:14 +02002263 * worker which uses GFP_KERNEL allocation which has slight chance of
2264 * developing into deadlock if some works currently on the same queue
2265 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2266 * the problem rescuer solves.
2267 *
Tejun Heo706026c2013-01-24 11:01:34 -08002268 * When such condition is possible, the pool summons rescuers of all
2269 * workqueues which have works queued on the pool and let them process
Tejun Heoe22bee72010-06-29 10:07:14 +02002270 * those works so that forward progress can be guaranteed.
2271 *
2272 * This should happen rarely.
Yacine Belkadid185af32013-07-31 14:59:24 -07002273 *
2274 * Return: 0
Tejun Heoe22bee72010-06-29 10:07:14 +02002275 */
Tejun Heo111c2252013-01-17 17:16:24 -08002276static int rescuer_thread(void *__rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02002277{
Tejun Heo111c2252013-01-17 17:16:24 -08002278 struct worker *rescuer = __rescuer;
2279 struct workqueue_struct *wq = rescuer->rescue_wq;
Tejun Heoe22bee72010-06-29 10:07:14 +02002280 struct list_head *scheduled = &rescuer->scheduled;
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002281 bool should_stop;
Tejun Heoe22bee72010-06-29 10:07:14 +02002282
2283 set_user_nice(current, RESCUER_NICE_LEVEL);
Tejun Heo111c2252013-01-17 17:16:24 -08002284
2285 /*
2286 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2287 * doesn't participate in concurrency management.
2288 */
2289 rescuer->task->flags |= PF_WQ_WORKER;
Tejun Heoe22bee72010-06-29 10:07:14 +02002290repeat:
2291 set_current_state(TASK_INTERRUPTIBLE);
2292
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002293 /*
2294 * By the time the rescuer is requested to stop, the workqueue
2295 * shouldn't have any work pending, but @wq->maydays may still have
2296 * pwq(s) queued. This can happen by non-rescuer workers consuming
2297 * all the work items before the rescuer got to them. Go through
2298 * @wq->maydays processing before acting on should_stop so that the
2299 * list is always empty on exit.
2300 */
2301 should_stop = kthread_should_stop();
Tejun Heoe22bee72010-06-29 10:07:14 +02002302
Tejun Heo493a1722013-03-12 11:29:59 -07002303 /* see whether any pwq is asking for help */
Tejun Heo2e109a22013-03-13 19:47:40 -07002304 spin_lock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002305
2306 while (!list_empty(&wq->maydays)) {
2307 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2308 struct pool_workqueue, mayday_node);
Tejun Heo112202d2013-02-13 19:29:12 -08002309 struct worker_pool *pool = pwq->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002310 struct work_struct *work, *n;
Tejun Heo82607adc2015-12-08 11:28:04 -05002311 bool first = true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002312
2313 __set_current_state(TASK_RUNNING);
Tejun Heo493a1722013-03-12 11:29:59 -07002314 list_del_init(&pwq->mayday_node);
2315
Tejun Heo2e109a22013-03-13 19:47:40 -07002316 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002317
Lai Jiangshan51697d392014-05-20 17:46:36 +08002318 worker_attach_to_pool(rescuer, pool);
2319
2320 spin_lock_irq(&pool->lock);
Lai Jiangshanb3104102013-02-19 12:17:02 -08002321 rescuer->pool = pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002322
2323 /*
2324 * Slurp in all works issued via this workqueue and
2325 * process'em.
2326 */
Tejun Heo0479c8c2014-12-04 10:14:13 -05002327 WARN_ON_ONCE(!list_empty(scheduled));
Tejun Heo82607adc2015-12-08 11:28:04 -05002328 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2329 if (get_work_pwq(work) == pwq) {
2330 if (first)
2331 pool->watchdog_ts = jiffies;
Tejun Heoe22bee72010-06-29 10:07:14 +02002332 move_linked_works(work, scheduled, &n);
Tejun Heo82607adc2015-12-08 11:28:04 -05002333 }
2334 first = false;
2335 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002336
NeilBrown008847f2014-12-08 12:39:16 -05002337 if (!list_empty(scheduled)) {
2338 process_scheduled_works(rescuer);
2339
2340 /*
2341 * The above execution of rescued work items could
2342 * have created more to rescue through
2343 * pwq_activate_first_delayed() or chained
2344 * queueing. Let's put @pwq back on mayday list so
2345 * that such back-to-back work items, which may be
2346 * being used to relieve memory pressure, don't
2347 * incur MAYDAY_INTERVAL delay inbetween.
2348 */
2349 if (need_to_create_worker(pool)) {
2350 spin_lock(&wq_mayday_lock);
2351 get_pwq(pwq);
2352 list_move_tail(&pwq->mayday_node, &wq->maydays);
2353 spin_unlock(&wq_mayday_lock);
2354 }
2355 }
Tejun Heo75769582011-02-14 14:04:46 +01002356
2357 /*
Lai Jiangshan77668c82014-04-18 11:04:16 -04002358 * Put the reference grabbed by send_mayday(). @pool won't
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002359 * go away while we're still attached to it.
Lai Jiangshan77668c82014-04-18 11:04:16 -04002360 */
2361 put_pwq(pwq);
2362
2363 /*
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002364 * Leave this pool. If need_more_worker() is %true, notify a
Tejun Heo75769582011-02-14 14:04:46 +01002365 * regular worker; otherwise, we end up with 0 concurrency
2366 * and stalling the execution.
2367 */
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002368 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002369 wake_up_worker(pool);
Tejun Heo75769582011-02-14 14:04:46 +01002370
Lai Jiangshanb3104102013-02-19 12:17:02 -08002371 rescuer->pool = NULL;
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002372 spin_unlock_irq(&pool->lock);
2373
2374 worker_detach_from_pool(rescuer, pool);
2375
2376 spin_lock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002377 }
2378
Tejun Heo2e109a22013-03-13 19:47:40 -07002379 spin_unlock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002380
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002381 if (should_stop) {
2382 __set_current_state(TASK_RUNNING);
2383 rescuer->task->flags &= ~PF_WQ_WORKER;
2384 return 0;
2385 }
2386
Tejun Heo111c2252013-01-17 17:16:24 -08002387 /* rescuers should never participate in concurrency management */
2388 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
Tejun Heoe22bee72010-06-29 10:07:14 +02002389 schedule();
2390 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391}
2392
Tejun Heofca839c2015-12-07 10:58:57 -05002393/**
2394 * check_flush_dependency - check for flush dependency sanity
2395 * @target_wq: workqueue being flushed
2396 * @target_work: work item being flushed (NULL for workqueue flushes)
2397 *
2398 * %current is trying to flush the whole @target_wq or @target_work on it.
2399 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2400 * reclaiming memory or running on a workqueue which doesn't have
2401 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2402 * a deadlock.
2403 */
2404static void check_flush_dependency(struct workqueue_struct *target_wq,
2405 struct work_struct *target_work)
2406{
2407 work_func_t target_func = target_work ? target_work->func : NULL;
2408 struct worker *worker;
2409
2410 if (target_wq->flags & WQ_MEM_RECLAIM)
2411 return;
2412
2413 worker = current_wq_worker();
2414
2415 WARN_ONCE(current->flags & PF_MEMALLOC,
2416 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2417 current->pid, current->comm, target_wq->name, target_func);
Tejun Heo23d11a52016-01-29 05:59:46 -05002418 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2419 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
Tejun Heofca839c2015-12-07 10:58:57 -05002420 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2421 worker->current_pwq->wq->name, worker->current_func,
2422 target_wq->name, target_func);
2423}
2424
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002425struct wq_barrier {
2426 struct work_struct work;
2427 struct completion done;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002428 struct task_struct *task; /* purely informational */
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002429};
2430
2431static void wq_barrier_func(struct work_struct *work)
2432{
2433 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2434 complete(&barr->done);
2435}
2436
Tejun Heo4690c4a2010-06-29 10:07:10 +02002437/**
2438 * insert_wq_barrier - insert a barrier work
Tejun Heo112202d2013-02-13 19:29:12 -08002439 * @pwq: pwq to insert barrier into
Tejun Heo4690c4a2010-06-29 10:07:10 +02002440 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002441 * @target: target work to attach @barr to
2442 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002443 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002444 * @barr is linked to @target such that @barr is completed only after
2445 * @target finishes execution. Please note that the ordering
2446 * guarantee is observed only with respect to @target and on the local
2447 * cpu.
2448 *
2449 * Currently, a queued barrier can't be canceled. This is because
2450 * try_to_grab_pending() can't determine whether the work to be
2451 * grabbed is at the head of the queue and thus can't clear LINKED
2452 * flag of the previous work while there must be a valid next work
2453 * after a work with LINKED flag set.
2454 *
2455 * Note that when @worker is non-NULL, @target may be modified
Tejun Heo112202d2013-02-13 19:29:12 -08002456 * underneath us, so we can't reliably determine pwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002457 *
2458 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002459 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002460 */
Tejun Heo112202d2013-02-13 19:29:12 -08002461static void insert_wq_barrier(struct pool_workqueue *pwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002462 struct wq_barrier *barr,
2463 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002464{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002465 struct list_head *head;
2466 unsigned int linked = 0;
2467
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002468 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002469 * debugobject calls are safe here even with pool->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002470 * as we know for sure that this will not trigger any of the
2471 * checks and call back into the fixup functions where we
2472 * might deadlock.
2473 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002474 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002475 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002476 init_completion(&barr->done);
Tejun Heo2607d7a2015-03-09 09:22:28 -04002477 barr->task = current;
Oleg Nesterov83c22522007-05-09 02:33:54 -07002478
Tejun Heoaffee4b2010-06-29 10:07:12 +02002479 /*
2480 * If @target is currently being executed, schedule the
2481 * barrier to the worker; otherwise, put it after @target.
2482 */
2483 if (worker)
2484 head = worker->scheduled.next;
2485 else {
2486 unsigned long *bits = work_data_bits(target);
2487
2488 head = target->entry.next;
2489 /* there can already be other linked works, inherit and set */
2490 linked = *bits & WORK_STRUCT_LINKED;
2491 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2492 }
2493
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002494 debug_work_activate(&barr->work);
Tejun Heo112202d2013-02-13 19:29:12 -08002495 insert_work(pwq, &barr->work, head,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002496 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002497}
2498
Tejun Heo73f53c42010-06-29 10:07:11 +02002499/**
Tejun Heo112202d2013-02-13 19:29:12 -08002500 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
Tejun Heo73f53c42010-06-29 10:07:11 +02002501 * @wq: workqueue being flushed
2502 * @flush_color: new flush color, < 0 for no-op
2503 * @work_color: new work color, < 0 for no-op
2504 *
Tejun Heo112202d2013-02-13 19:29:12 -08002505 * Prepare pwqs for workqueue flushing.
Tejun Heo73f53c42010-06-29 10:07:11 +02002506 *
Tejun Heo112202d2013-02-13 19:29:12 -08002507 * If @flush_color is non-negative, flush_color on all pwqs should be
2508 * -1. If no pwq has in-flight commands at the specified color, all
2509 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2510 * has in flight commands, its pwq->flush_color is set to
2511 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
Tejun Heo73f53c42010-06-29 10:07:11 +02002512 * wakeup logic is armed and %true is returned.
2513 *
2514 * The caller should have initialized @wq->first_flusher prior to
2515 * calling this function with non-negative @flush_color. If
2516 * @flush_color is negative, no flush color update is done and %false
2517 * is returned.
2518 *
Tejun Heo112202d2013-02-13 19:29:12 -08002519 * If @work_color is non-negative, all pwqs should have the same
Tejun Heo73f53c42010-06-29 10:07:11 +02002520 * work_color which is previous to @work_color and all will be
2521 * advanced to @work_color.
2522 *
2523 * CONTEXT:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002524 * mutex_lock(wq->mutex).
Tejun Heo73f53c42010-06-29 10:07:11 +02002525 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002526 * Return:
Tejun Heo73f53c42010-06-29 10:07:11 +02002527 * %true if @flush_color >= 0 and there's something to flush. %false
2528 * otherwise.
2529 */
Tejun Heo112202d2013-02-13 19:29:12 -08002530static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
Tejun Heo73f53c42010-06-29 10:07:11 +02002531 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532{
Tejun Heo73f53c42010-06-29 10:07:11 +02002533 bool wait = false;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002534 struct pool_workqueue *pwq;
Oleg Nesterov14441962007-05-23 13:57:57 -07002535
Tejun Heo73f53c42010-06-29 10:07:11 +02002536 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002537 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
Tejun Heo112202d2013-02-13 19:29:12 -08002538 atomic_set(&wq->nr_pwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002539 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002540
Tejun Heo49e3cf42013-03-12 11:29:58 -07002541 for_each_pwq(pwq, wq) {
Tejun Heo112202d2013-02-13 19:29:12 -08002542 struct worker_pool *pool = pwq->pool;
Tejun Heo73f53c42010-06-29 10:07:11 +02002543
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002544 spin_lock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002545
2546 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002547 WARN_ON_ONCE(pwq->flush_color != -1);
Tejun Heo73f53c42010-06-29 10:07:11 +02002548
Tejun Heo112202d2013-02-13 19:29:12 -08002549 if (pwq->nr_in_flight[flush_color]) {
2550 pwq->flush_color = flush_color;
2551 atomic_inc(&wq->nr_pwqs_to_flush);
Tejun Heo73f53c42010-06-29 10:07:11 +02002552 wait = true;
2553 }
2554 }
2555
2556 if (work_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002557 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
Tejun Heo112202d2013-02-13 19:29:12 -08002558 pwq->work_color = work_color;
Tejun Heo73f53c42010-06-29 10:07:11 +02002559 }
2560
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002561 spin_unlock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002562 }
2563
Tejun Heo112202d2013-02-13 19:29:12 -08002564 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
Tejun Heo73f53c42010-06-29 10:07:11 +02002565 complete(&wq->first_flusher->done);
2566
2567 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568}
2569
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002570/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002572 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002574 * This function sleeps until all work items which were queued on entry
2575 * have finished execution, but it is not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002577void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578{
Tejun Heo73f53c42010-06-29 10:07:11 +02002579 struct wq_flusher this_flusher = {
2580 .list = LIST_HEAD_INIT(this_flusher.list),
2581 .flush_color = -1,
2582 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2583 };
2584 int next_color;
Oleg Nesterovb1f4ec172007-05-09 02:34:12 -07002585
Tejun Heo3347fa02016-09-16 15:49:32 -04002586 if (WARN_ON(!wq_online))
2587 return;
2588
Ingo Molnar3295f0e2008-08-11 10:30:30 +02002589 lock_map_acquire(&wq->lockdep_map);
2590 lock_map_release(&wq->lockdep_map);
Tejun Heo73f53c42010-06-29 10:07:11 +02002591
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002592 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002593
2594 /*
2595 * Start-to-wait phase
2596 */
2597 next_color = work_next_color(wq->work_color);
2598
2599 if (next_color != wq->flush_color) {
2600 /*
2601 * Color space is not full. The current work_color
2602 * becomes our flush_color and work_color is advanced
2603 * by one.
2604 */
Tejun Heo6183c002013-03-12 11:29:57 -07002605 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
Tejun Heo73f53c42010-06-29 10:07:11 +02002606 this_flusher.flush_color = wq->work_color;
2607 wq->work_color = next_color;
2608
2609 if (!wq->first_flusher) {
2610 /* no flush in progress, become the first flusher */
Tejun Heo6183c002013-03-12 11:29:57 -07002611 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002612
2613 wq->first_flusher = &this_flusher;
2614
Tejun Heo112202d2013-02-13 19:29:12 -08002615 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
Tejun Heo73f53c42010-06-29 10:07:11 +02002616 wq->work_color)) {
2617 /* nothing to flush, done */
2618 wq->flush_color = next_color;
2619 wq->first_flusher = NULL;
2620 goto out_unlock;
2621 }
2622 } else {
2623 /* wait in queue */
Tejun Heo6183c002013-03-12 11:29:57 -07002624 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002625 list_add_tail(&this_flusher.list, &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002626 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002627 }
2628 } else {
2629 /*
2630 * Oops, color space is full, wait on overflow queue.
2631 * The next flush completion will assign us
2632 * flush_color and transfer to flusher_queue.
2633 */
2634 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2635 }
2636
Tejun Heofca839c2015-12-07 10:58:57 -05002637 check_flush_dependency(wq, NULL);
2638
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002639 mutex_unlock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002640
2641 wait_for_completion(&this_flusher.done);
2642
2643 /*
2644 * Wake-up-and-cascade phase
2645 *
2646 * First flushers are responsible for cascading flushes and
2647 * handling overflow. Non-first flushers can simply return.
2648 */
2649 if (wq->first_flusher != &this_flusher)
2650 return;
2651
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002652 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002653
Tejun Heo4ce48b32010-07-02 10:03:51 +02002654 /* we might have raced, check again with mutex held */
2655 if (wq->first_flusher != &this_flusher)
2656 goto out_unlock;
2657
Tejun Heo73f53c42010-06-29 10:07:11 +02002658 wq->first_flusher = NULL;
2659
Tejun Heo6183c002013-03-12 11:29:57 -07002660 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2661 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002662
2663 while (true) {
2664 struct wq_flusher *next, *tmp;
2665
2666 /* complete all the flushers sharing the current flush color */
2667 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2668 if (next->flush_color != wq->flush_color)
2669 break;
2670 list_del_init(&next->list);
2671 complete(&next->done);
2672 }
2673
Tejun Heo6183c002013-03-12 11:29:57 -07002674 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2675 wq->flush_color != work_next_color(wq->work_color));
Tejun Heo73f53c42010-06-29 10:07:11 +02002676
2677 /* this flush_color is finished, advance by one */
2678 wq->flush_color = work_next_color(wq->flush_color);
2679
2680 /* one color has been freed, handle overflow queue */
2681 if (!list_empty(&wq->flusher_overflow)) {
2682 /*
2683 * Assign the same color to all overflowed
2684 * flushers, advance work_color and append to
2685 * flusher_queue. This is the start-to-wait
2686 * phase for these overflowed flushers.
2687 */
2688 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2689 tmp->flush_color = wq->work_color;
2690
2691 wq->work_color = work_next_color(wq->work_color);
2692
2693 list_splice_tail_init(&wq->flusher_overflow,
2694 &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002695 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002696 }
2697
2698 if (list_empty(&wq->flusher_queue)) {
Tejun Heo6183c002013-03-12 11:29:57 -07002699 WARN_ON_ONCE(wq->flush_color != wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002700 break;
2701 }
2702
2703 /*
2704 * Need to flush more colors. Make the next flusher
Tejun Heo112202d2013-02-13 19:29:12 -08002705 * the new first flusher and arm pwqs.
Tejun Heo73f53c42010-06-29 10:07:11 +02002706 */
Tejun Heo6183c002013-03-12 11:29:57 -07002707 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2708 WARN_ON_ONCE(wq->flush_color != next->flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002709
2710 list_del_init(&next->list);
2711 wq->first_flusher = next;
2712
Tejun Heo112202d2013-02-13 19:29:12 -08002713 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
Tejun Heo73f53c42010-06-29 10:07:11 +02002714 break;
2715
2716 /*
2717 * Meh... this color is already done, clear first
2718 * flusher and repeat cascading.
2719 */
2720 wq->first_flusher = NULL;
2721 }
2722
2723out_unlock:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002724 mutex_unlock(&wq->mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725}
Tim Gardner1dadafa2015-08-04 11:26:04 -06002726EXPORT_SYMBOL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002728/**
2729 * drain_workqueue - drain a workqueue
2730 * @wq: workqueue to drain
2731 *
2732 * Wait until the workqueue becomes empty. While draining is in progress,
2733 * only chain queueing is allowed. IOW, only currently pending or running
2734 * work items on @wq can queue further work items on it. @wq is flushed
Chen Hanxiaob749b1b2015-05-13 06:10:05 -04002735 * repeatedly until it becomes empty. The number of flushing is determined
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002736 * by the depth of chaining and should be relatively short. Whine if it
2737 * takes too long.
2738 */
2739void drain_workqueue(struct workqueue_struct *wq)
2740{
2741 unsigned int flush_cnt = 0;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002742 struct pool_workqueue *pwq;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002743
2744 /*
2745 * __queue_work() needs to test whether there are drainers, is much
2746 * hotter than drain_workqueue() and already looks at @wq->flags.
Tejun Heo618b01e2013-03-12 11:30:04 -07002747 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002748 */
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002749 mutex_lock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002750 if (!wq->nr_drainers++)
Tejun Heo618b01e2013-03-12 11:30:04 -07002751 wq->flags |= __WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002752 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002753reflush:
2754 flush_workqueue(wq);
2755
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002756 mutex_lock(&wq->mutex);
Tejun Heo76af4d92013-03-12 11:30:00 -07002757
Tejun Heo49e3cf42013-03-12 11:29:58 -07002758 for_each_pwq(pwq, wq) {
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002759 bool drained;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002760
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002761 spin_lock_irq(&pwq->pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08002762 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002763 spin_unlock_irq(&pwq->pool->lock);
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002764
2765 if (drained)
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002766 continue;
2767
2768 if (++flush_cnt == 10 ||
2769 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002770 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
Valentin Ilie044c7822012-08-19 00:52:42 +03002771 wq->name, flush_cnt);
Tejun Heo76af4d92013-03-12 11:30:00 -07002772
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002773 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002774 goto reflush;
2775 }
2776
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002777 if (!--wq->nr_drainers)
Tejun Heo618b01e2013-03-12 11:30:04 -07002778 wq->flags &= ~__WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002779 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002780}
2781EXPORT_SYMBOL_GPL(drain_workqueue);
2782
Tejun Heo606a5022012-08-20 14:51:23 -07002783static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
Tejun Heobaf59022010-09-16 10:42:16 +02002784{
2785 struct worker *worker = NULL;
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002786 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002787 struct pool_workqueue *pwq;
Tejun Heobaf59022010-09-16 10:42:16 +02002788
2789 might_sleep();
Tejun Heobaf59022010-09-16 10:42:16 +02002790
Tejun Heofa1b54e2013-03-12 11:30:00 -07002791 local_irq_disable();
2792 pool = get_work_pool(work);
2793 if (!pool) {
2794 local_irq_enable();
2795 return false;
2796 }
2797
2798 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08002799 /* see the comment in try_to_grab_pending() with the same code */
Tejun Heo112202d2013-02-13 19:29:12 -08002800 pwq = get_work_pwq(work);
2801 if (pwq) {
2802 if (unlikely(pwq->pool != pool))
Tejun Heobaf59022010-09-16 10:42:16 +02002803 goto already_gone;
Tejun Heo606a5022012-08-20 14:51:23 -07002804 } else {
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002805 worker = find_worker_executing_work(pool, work);
Tejun Heobaf59022010-09-16 10:42:16 +02002806 if (!worker)
2807 goto already_gone;
Tejun Heo112202d2013-02-13 19:29:12 -08002808 pwq = worker->current_pwq;
Tejun Heo606a5022012-08-20 14:51:23 -07002809 }
Tejun Heobaf59022010-09-16 10:42:16 +02002810
Tejun Heofca839c2015-12-07 10:58:57 -05002811 check_flush_dependency(pwq->wq, work);
2812
Tejun Heo112202d2013-02-13 19:29:12 -08002813 insert_wq_barrier(pwq, barr, work, worker);
Tejun Heod565ed62013-01-24 11:01:33 -08002814 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002815
Tejun Heoe159489b2011-01-09 23:32:15 +01002816 /*
2817 * If @max_active is 1 or rescuer is in use, flushing another work
2818 * item on the same workqueue may lead to deadlock. Make sure the
2819 * flusher is not running on the same workqueue by verifying write
2820 * access.
2821 */
Tejun Heo493008a2013-03-12 11:30:03 -07002822 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
Tejun Heo112202d2013-02-13 19:29:12 -08002823 lock_map_acquire(&pwq->wq->lockdep_map);
Tejun Heoe159489b2011-01-09 23:32:15 +01002824 else
Tejun Heo112202d2013-02-13 19:29:12 -08002825 lock_map_acquire_read(&pwq->wq->lockdep_map);
2826 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoe159489b2011-01-09 23:32:15 +01002827
Tejun Heobaf59022010-09-16 10:42:16 +02002828 return true;
2829already_gone:
Tejun Heod565ed62013-01-24 11:01:33 -08002830 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002831 return false;
2832}
2833
Oleg Nesterovdb700892008-07-25 01:47:49 -07002834/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002835 * flush_work - wait for a work to finish executing the last queueing instance
2836 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07002837 *
Tejun Heo606a5022012-08-20 14:51:23 -07002838 * Wait until @work has finished execution. @work is guaranteed to be idle
2839 * on return if it hasn't been requeued since flush started.
Tejun Heo401a8d02010-09-16 10:36:00 +02002840 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002841 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002842 * %true if flush_work() waited for the work to finish execution,
2843 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07002844 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002845bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07002846{
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002847 struct wq_barrier barr;
2848
Tejun Heo3347fa02016-09-16 15:49:32 -04002849 if (WARN_ON(!wq_online))
2850 return false;
2851
Stephen Boyd0976dfc2012-04-20 17:28:50 -07002852 lock_map_acquire(&work->lockdep_map);
2853 lock_map_release(&work->lockdep_map);
2854
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002855 if (start_flush_work(work, &barr)) {
2856 wait_for_completion(&barr.done);
2857 destroy_work_on_stack(&barr.work);
2858 return true;
2859 } else {
2860 return false;
2861 }
Oleg Nesterovdb700892008-07-25 01:47:49 -07002862}
2863EXPORT_SYMBOL_GPL(flush_work);
2864
Tejun Heo8603e1b32015-03-05 08:04:13 -05002865struct cwt_wait {
2866 wait_queue_t wait;
2867 struct work_struct *work;
2868};
2869
2870static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2871{
2872 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2873
2874 if (cwait->work != key)
2875 return 0;
2876 return autoremove_wake_function(wait, mode, sync, key);
2877}
2878
Tejun Heo36e227d2012-08-03 10:30:46 -07002879static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
Tejun Heo401a8d02010-09-16 10:36:00 +02002880{
Tejun Heo8603e1b32015-03-05 08:04:13 -05002881 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
Tejun Heobbb68df2012-08-03 10:30:46 -07002882 unsigned long flags;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002883 int ret;
2884
2885 do {
Tejun Heobbb68df2012-08-03 10:30:46 -07002886 ret = try_to_grab_pending(work, is_dwork, &flags);
2887 /*
Tejun Heo8603e1b32015-03-05 08:04:13 -05002888 * If someone else is already canceling, wait for it to
2889 * finish. flush_work() doesn't work for PREEMPT_NONE
2890 * because we may get scheduled between @work's completion
2891 * and the other canceling task resuming and clearing
2892 * CANCELING - flush_work() will return false immediately
2893 * as @work is no longer busy, try_to_grab_pending() will
2894 * return -ENOENT as @work is still being canceled and the
2895 * other canceling task won't be able to clear CANCELING as
2896 * we're hogging the CPU.
2897 *
2898 * Let's wait for completion using a waitqueue. As this
2899 * may lead to the thundering herd problem, use a custom
2900 * wake function which matches @work along with exclusive
2901 * wait and wakeup.
Tejun Heobbb68df2012-08-03 10:30:46 -07002902 */
Tejun Heo8603e1b32015-03-05 08:04:13 -05002903 if (unlikely(ret == -ENOENT)) {
2904 struct cwt_wait cwait;
2905
2906 init_wait(&cwait.wait);
2907 cwait.wait.func = cwt_wakefn;
2908 cwait.work = work;
2909
2910 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2911 TASK_UNINTERRUPTIBLE);
2912 if (work_is_canceling(work))
2913 schedule();
2914 finish_wait(&cancel_waitq, &cwait.wait);
2915 }
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002916 } while (unlikely(ret < 0));
2917
Tejun Heobbb68df2012-08-03 10:30:46 -07002918 /* tell other tasks trying to grab @work to back off */
2919 mark_work_canceling(work);
2920 local_irq_restore(flags);
2921
Tejun Heo3347fa02016-09-16 15:49:32 -04002922 /*
2923 * This allows canceling during early boot. We know that @work
2924 * isn't executing.
2925 */
2926 if (wq_online)
2927 flush_work(work);
2928
Tejun Heo7a22ad72010-06-29 10:07:13 +02002929 clear_work_data(work);
Tejun Heo8603e1b32015-03-05 08:04:13 -05002930
2931 /*
2932 * Paired with prepare_to_wait() above so that either
2933 * waitqueue_active() is visible here or !work_is_canceling() is
2934 * visible there.
2935 */
2936 smp_mb();
2937 if (waitqueue_active(&cancel_waitq))
2938 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2939
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002940 return ret;
2941}
2942
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002943/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002944 * cancel_work_sync - cancel a work and wait for it to finish
2945 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002946 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002947 * Cancel @work and wait for its execution to finish. This function
2948 * can be used even if the work re-queues itself or migrates to
2949 * another workqueue. On return from this function, @work is
2950 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002951 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002952 * cancel_work_sync(&delayed_work->work) must not be used for
2953 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002954 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002955 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002956 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02002957 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002958 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002959 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002960 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002961bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002962{
Tejun Heo36e227d2012-08-03 10:30:46 -07002963 return __cancel_work_timer(work, false);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002964}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07002965EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002966
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002967/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002968 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2969 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002970 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002971 * Delayed timer is cancelled and the pending work is queued for
2972 * immediate execution. Like flush_work(), this function only
2973 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002974 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002975 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002976 * %true if flush_work() waited for the work to finish execution,
2977 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002978 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002979bool flush_delayed_work(struct delayed_work *dwork)
2980{
Tejun Heo8930cab2012-08-03 10:30:45 -07002981 local_irq_disable();
Tejun Heo401a8d02010-09-16 10:36:00 +02002982 if (del_timer_sync(&dwork->timer))
Lai Jiangshan60c057b2013-02-06 18:04:53 -08002983 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Tejun Heo8930cab2012-08-03 10:30:45 -07002984 local_irq_enable();
Tejun Heo401a8d02010-09-16 10:36:00 +02002985 return flush_work(&dwork->work);
2986}
2987EXPORT_SYMBOL(flush_delayed_work);
2988
Jens Axboef72b8792016-08-24 15:51:50 -06002989static bool __cancel_work(struct work_struct *work, bool is_dwork)
2990{
2991 unsigned long flags;
2992 int ret;
2993
2994 do {
2995 ret = try_to_grab_pending(work, is_dwork, &flags);
2996 } while (unlikely(ret == -EAGAIN));
2997
2998 if (unlikely(ret < 0))
2999 return false;
3000
3001 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3002 local_irq_restore(flags);
3003 return ret;
3004}
3005
3006/*
3007 * See cancel_delayed_work()
3008 */
3009bool cancel_work(struct work_struct *work)
3010{
3011 return __cancel_work(work, false);
3012}
3013
Tejun Heo401a8d02010-09-16 10:36:00 +02003014/**
Tejun Heo57b30ae2012-08-21 13:18:24 -07003015 * cancel_delayed_work - cancel a delayed work
3016 * @dwork: delayed_work to cancel
Tejun Heo09383492010-09-16 10:48:29 +02003017 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003018 * Kill off a pending delayed_work.
3019 *
3020 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3021 * pending.
3022 *
3023 * Note:
3024 * The work callback function may still be running on return, unless
3025 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3026 * use cancel_delayed_work_sync() to wait on it.
Tejun Heo09383492010-09-16 10:48:29 +02003027 *
Tejun Heo57b30ae2012-08-21 13:18:24 -07003028 * This function is safe to call from any context including IRQ handler.
Tejun Heo09383492010-09-16 10:48:29 +02003029 */
Tejun Heo57b30ae2012-08-21 13:18:24 -07003030bool cancel_delayed_work(struct delayed_work *dwork)
Tejun Heo09383492010-09-16 10:48:29 +02003031{
Jens Axboef72b8792016-08-24 15:51:50 -06003032 return __cancel_work(&dwork->work, true);
Tejun Heo09383492010-09-16 10:48:29 +02003033}
Tejun Heo57b30ae2012-08-21 13:18:24 -07003034EXPORT_SYMBOL(cancel_delayed_work);
Tejun Heo09383492010-09-16 10:48:29 +02003035
3036/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003037 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3038 * @dwork: the delayed work cancel
3039 *
3040 * This is cancel_work_sync() for delayed works.
3041 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003042 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003043 * %true if @dwork was pending, %false otherwise.
3044 */
3045bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003046{
Tejun Heo36e227d2012-08-03 10:30:46 -07003047 return __cancel_work_timer(&dwork->work, true);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003048}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07003049EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07003051/**
Tejun Heo31ddd872010-10-19 11:14:49 +02003052 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07003053 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07003054 *
Tejun Heo31ddd872010-10-19 11:14:49 +02003055 * schedule_on_each_cpu() executes @func on each online CPU using the
3056 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07003057 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02003058 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003059 * Return:
Tejun Heo31ddd872010-10-19 11:14:49 +02003060 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07003061 */
David Howells65f27f32006-11-22 14:55:48 +00003062int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003063{
3064 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02003065 struct work_struct __percpu *works;
Christoph Lameter15316ba2006-01-08 01:00:43 -08003066
Andrew Mortonb6136772006-06-25 05:47:49 -07003067 works = alloc_percpu(struct work_struct);
3068 if (!works)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003069 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07003070
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003071 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08003072
Christoph Lameter15316ba2006-01-08 01:00:43 -08003073 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01003074 struct work_struct *work = per_cpu_ptr(works, cpu);
3075
3076 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003077 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02003078 }
Tejun Heo93981802009-11-17 14:06:20 -08003079
3080 for_each_online_cpu(cpu)
3081 flush_work(per_cpu_ptr(works, cpu));
3082
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003083 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07003084 free_percpu(works);
Christoph Lameter15316ba2006-01-08 01:00:43 -08003085 return 0;
3086}
3087
Alan Sterneef6a7d2010-02-12 17:39:21 +09003088/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06003089 * execute_in_process_context - reliably execute the routine with user context
3090 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06003091 * @ew: guaranteed storage for the execute work structure (must
3092 * be available when the work executes)
3093 *
3094 * Executes the function immediately if process context is available,
3095 * otherwise schedules the function for delayed execution.
3096 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003097 * Return: 0 - function was executed
James Bottomley1fa44ec2006-02-23 12:43:43 -06003098 * 1 - function was scheduled for execution
3099 */
David Howells65f27f32006-11-22 14:55:48 +00003100int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06003101{
3102 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00003103 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003104 return 0;
3105 }
3106
David Howells65f27f32006-11-22 14:55:48 +00003107 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003108 schedule_work(&ew->work);
3109
3110 return 1;
3111}
3112EXPORT_SYMBOL_GPL(execute_in_process_context);
3113
Tejun Heo7a4e3442013-03-12 11:30:00 -07003114/**
3115 * free_workqueue_attrs - free a workqueue_attrs
3116 * @attrs: workqueue_attrs to free
3117 *
3118 * Undo alloc_workqueue_attrs().
3119 */
3120void free_workqueue_attrs(struct workqueue_attrs *attrs)
3121{
3122 if (attrs) {
3123 free_cpumask_var(attrs->cpumask);
3124 kfree(attrs);
3125 }
3126}
3127
3128/**
3129 * alloc_workqueue_attrs - allocate a workqueue_attrs
3130 * @gfp_mask: allocation mask to use
3131 *
3132 * Allocate a new workqueue_attrs, initialize with default settings and
Yacine Belkadid185af32013-07-31 14:59:24 -07003133 * return it.
3134 *
3135 * Return: The allocated new workqueue_attr on success. %NULL on failure.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003136 */
3137struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3138{
3139 struct workqueue_attrs *attrs;
3140
3141 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3142 if (!attrs)
3143 goto fail;
3144 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3145 goto fail;
3146
Tejun Heo13e2e552013-04-01 11:23:31 -07003147 cpumask_copy(attrs->cpumask, cpu_possible_mask);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003148 return attrs;
3149fail:
3150 free_workqueue_attrs(attrs);
3151 return NULL;
3152}
3153
Tejun Heo29c91e92013-03-12 11:30:03 -07003154static void copy_workqueue_attrs(struct workqueue_attrs *to,
3155 const struct workqueue_attrs *from)
3156{
3157 to->nice = from->nice;
3158 cpumask_copy(to->cpumask, from->cpumask);
Shaohua Li2865a8f2013-08-01 09:56:36 +08003159 /*
3160 * Unlike hash and equality test, this function doesn't ignore
3161 * ->no_numa as it is used for both pool and wq attrs. Instead,
3162 * get_unbound_pool() explicitly clears ->no_numa after copying.
3163 */
3164 to->no_numa = from->no_numa;
Tejun Heo29c91e92013-03-12 11:30:03 -07003165}
3166
Tejun Heo29c91e92013-03-12 11:30:03 -07003167/* hash value of the content of @attr */
3168static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3169{
3170 u32 hash = 0;
3171
3172 hash = jhash_1word(attrs->nice, hash);
Tejun Heo13e2e552013-04-01 11:23:31 -07003173 hash = jhash(cpumask_bits(attrs->cpumask),
3174 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
Tejun Heo29c91e92013-03-12 11:30:03 -07003175 return hash;
3176}
3177
3178/* content equality test */
3179static bool wqattrs_equal(const struct workqueue_attrs *a,
3180 const struct workqueue_attrs *b)
3181{
3182 if (a->nice != b->nice)
3183 return false;
3184 if (!cpumask_equal(a->cpumask, b->cpumask))
3185 return false;
3186 return true;
3187}
3188
Tejun Heo7a4e3442013-03-12 11:30:00 -07003189/**
3190 * init_worker_pool - initialize a newly zalloc'd worker_pool
3191 * @pool: worker_pool to initialize
3192 *
Shailendra Verma402dd892015-05-23 10:38:14 +05303193 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
Yacine Belkadid185af32013-07-31 14:59:24 -07003194 *
3195 * Return: 0 on success, -errno on failure. Even on failure, all fields
Tejun Heo29c91e92013-03-12 11:30:03 -07003196 * inside @pool proper are initialized and put_unbound_pool() can be called
3197 * on @pool safely to release it.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003198 */
3199static int init_worker_pool(struct worker_pool *pool)
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003200{
3201 spin_lock_init(&pool->lock);
Tejun Heo29c91e92013-03-12 11:30:03 -07003202 pool->id = -1;
3203 pool->cpu = -1;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003204 pool->node = NUMA_NO_NODE;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003205 pool->flags |= POOL_DISASSOCIATED;
Tejun Heo82607adc2015-12-08 11:28:04 -05003206 pool->watchdog_ts = jiffies;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003207 INIT_LIST_HEAD(&pool->worklist);
3208 INIT_LIST_HEAD(&pool->idle_list);
3209 hash_init(pool->busy_hash);
3210
3211 init_timer_deferrable(&pool->idle_timer);
3212 pool->idle_timer.function = idle_worker_timeout;
3213 pool->idle_timer.data = (unsigned long)pool;
3214
3215 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3216 (unsigned long)pool);
3217
3218 mutex_init(&pool->manager_arb);
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003219 mutex_init(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003220 INIT_LIST_HEAD(&pool->workers);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003221
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003222 ida_init(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003223 INIT_HLIST_NODE(&pool->hash_node);
3224 pool->refcnt = 1;
3225
3226 /* shouldn't fail above this point */
Tejun Heo7a4e3442013-03-12 11:30:00 -07003227 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3228 if (!pool->attrs)
3229 return -ENOMEM;
3230 return 0;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003231}
3232
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003233static void rcu_free_wq(struct rcu_head *rcu)
3234{
3235 struct workqueue_struct *wq =
3236 container_of(rcu, struct workqueue_struct, rcu);
3237
3238 if (!(wq->flags & WQ_UNBOUND))
3239 free_percpu(wq->cpu_pwqs);
3240 else
3241 free_workqueue_attrs(wq->unbound_attrs);
3242
3243 kfree(wq->rescuer);
3244 kfree(wq);
3245}
3246
Tejun Heo29c91e92013-03-12 11:30:03 -07003247static void rcu_free_pool(struct rcu_head *rcu)
3248{
3249 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3250
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003251 ida_destroy(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003252 free_workqueue_attrs(pool->attrs);
3253 kfree(pool);
3254}
3255
3256/**
3257 * put_unbound_pool - put a worker_pool
3258 * @pool: worker_pool to put
3259 *
3260 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003261 * safe manner. get_unbound_pool() calls this function on its failure path
3262 * and this function should be able to release pools which went through,
3263 * successfully or not, init_worker_pool().
Tejun Heoa892cac2013-04-01 11:23:32 -07003264 *
3265 * Should be called with wq_pool_mutex held.
Tejun Heo29c91e92013-03-12 11:30:03 -07003266 */
3267static void put_unbound_pool(struct worker_pool *pool)
3268{
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003269 DECLARE_COMPLETION_ONSTACK(detach_completion);
Tejun Heo29c91e92013-03-12 11:30:03 -07003270 struct worker *worker;
3271
Tejun Heoa892cac2013-04-01 11:23:32 -07003272 lockdep_assert_held(&wq_pool_mutex);
3273
3274 if (--pool->refcnt)
Tejun Heo29c91e92013-03-12 11:30:03 -07003275 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003276
3277 /* sanity checks */
Lai Jiangshan61d0fbb2014-06-03 15:31:45 +08003278 if (WARN_ON(!(pool->cpu < 0)) ||
Tejun Heoa892cac2013-04-01 11:23:32 -07003279 WARN_ON(!list_empty(&pool->worklist)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003280 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003281
3282 /* release id and unhash */
3283 if (pool->id >= 0)
3284 idr_remove(&worker_pool_idr, pool->id);
3285 hash_del(&pool->hash_node);
3286
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003287 /*
3288 * Become the manager and destroy all workers. Grabbing
3289 * manager_arb prevents @pool's workers from blocking on
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003290 * attach_mutex.
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003291 */
Tejun Heo29c91e92013-03-12 11:30:03 -07003292 mutex_lock(&pool->manager_arb);
Tejun Heo29c91e92013-03-12 11:30:03 -07003293
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003294 spin_lock_irq(&pool->lock);
Lai Jiangshan1037de32014-05-22 16:44:07 +08003295 while ((worker = first_idle_worker(pool)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003296 destroy_worker(worker);
3297 WARN_ON(pool->nr_workers || pool->nr_idle);
Tejun Heo29c91e92013-03-12 11:30:03 -07003298 spin_unlock_irq(&pool->lock);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003299
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003300 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003301 if (!list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003302 pool->detach_completion = &detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003303 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003304
3305 if (pool->detach_completion)
3306 wait_for_completion(pool->detach_completion);
3307
Tejun Heo29c91e92013-03-12 11:30:03 -07003308 mutex_unlock(&pool->manager_arb);
3309
3310 /* shut down the timers */
3311 del_timer_sync(&pool->idle_timer);
3312 del_timer_sync(&pool->mayday_timer);
3313
3314 /* sched-RCU protected to allow dereferences from get_work_pool() */
3315 call_rcu_sched(&pool->rcu, rcu_free_pool);
3316}
3317
3318/**
3319 * get_unbound_pool - get a worker_pool with the specified attributes
3320 * @attrs: the attributes of the worker_pool to get
3321 *
3322 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3323 * reference count and return it. If there already is a matching
3324 * worker_pool, it will be used; otherwise, this function attempts to
Yacine Belkadid185af32013-07-31 14:59:24 -07003325 * create a new one.
Tejun Heoa892cac2013-04-01 11:23:32 -07003326 *
3327 * Should be called with wq_pool_mutex held.
Yacine Belkadid185af32013-07-31 14:59:24 -07003328 *
3329 * Return: On success, a worker_pool with the same attributes as @attrs.
3330 * On failure, %NULL.
Tejun Heo29c91e92013-03-12 11:30:03 -07003331 */
3332static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3333{
Tejun Heo29c91e92013-03-12 11:30:03 -07003334 u32 hash = wqattrs_hash(attrs);
3335 struct worker_pool *pool;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003336 int node;
Xunlei Pange22735842015-10-09 11:53:12 +08003337 int target_node = NUMA_NO_NODE;
Tejun Heo29c91e92013-03-12 11:30:03 -07003338
Tejun Heoa892cac2013-04-01 11:23:32 -07003339 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003340
3341 /* do we already have a matching pool? */
Tejun Heo29c91e92013-03-12 11:30:03 -07003342 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3343 if (wqattrs_equal(pool->attrs, attrs)) {
3344 pool->refcnt++;
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003345 return pool;
Tejun Heo29c91e92013-03-12 11:30:03 -07003346 }
3347 }
Tejun Heo29c91e92013-03-12 11:30:03 -07003348
Xunlei Pange22735842015-10-09 11:53:12 +08003349 /* if cpumask is contained inside a NUMA node, we belong to that node */
3350 if (wq_numa_enabled) {
3351 for_each_node(node) {
3352 if (cpumask_subset(attrs->cpumask,
3353 wq_numa_possible_cpumask[node])) {
3354 target_node = node;
3355 break;
3356 }
3357 }
3358 }
3359
Tejun Heo29c91e92013-03-12 11:30:03 -07003360 /* nope, create a new one */
Xunlei Pange22735842015-10-09 11:53:12 +08003361 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
Tejun Heo29c91e92013-03-12 11:30:03 -07003362 if (!pool || init_worker_pool(pool) < 0)
3363 goto fail;
3364
Tejun Heo8864b4e2013-03-12 11:30:04 -07003365 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
Tejun Heo29c91e92013-03-12 11:30:03 -07003366 copy_workqueue_attrs(pool->attrs, attrs);
Xunlei Pange22735842015-10-09 11:53:12 +08003367 pool->node = target_node;
Tejun Heo29c91e92013-03-12 11:30:03 -07003368
Shaohua Li2865a8f2013-08-01 09:56:36 +08003369 /*
3370 * no_numa isn't a worker_pool attribute, always clear it. See
3371 * 'struct workqueue_attrs' comments for detail.
3372 */
3373 pool->attrs->no_numa = false;
3374
Tejun Heo29c91e92013-03-12 11:30:03 -07003375 if (worker_pool_assign_id(pool) < 0)
3376 goto fail;
3377
3378 /* create and start the initial worker */
Tejun Heo3347fa02016-09-16 15:49:32 -04003379 if (wq_online && !create_worker(pool))
Tejun Heo29c91e92013-03-12 11:30:03 -07003380 goto fail;
3381
Tejun Heo29c91e92013-03-12 11:30:03 -07003382 /* install */
Tejun Heo29c91e92013-03-12 11:30:03 -07003383 hash_add(unbound_pool_hash, &pool->hash_node, hash);
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003384
Tejun Heo29c91e92013-03-12 11:30:03 -07003385 return pool;
3386fail:
Tejun Heo29c91e92013-03-12 11:30:03 -07003387 if (pool)
3388 put_unbound_pool(pool);
3389 return NULL;
3390}
3391
Tejun Heo8864b4e2013-03-12 11:30:04 -07003392static void rcu_free_pwq(struct rcu_head *rcu)
3393{
3394 kmem_cache_free(pwq_cache,
3395 container_of(rcu, struct pool_workqueue, rcu));
3396}
3397
3398/*
3399 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3400 * and needs to be destroyed.
3401 */
3402static void pwq_unbound_release_workfn(struct work_struct *work)
3403{
3404 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3405 unbound_release_work);
3406 struct workqueue_struct *wq = pwq->wq;
3407 struct worker_pool *pool = pwq->pool;
Tejun Heobc0caf02013-04-01 11:23:31 -07003408 bool is_last;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003409
3410 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3411 return;
3412
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003413 mutex_lock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003414 list_del_rcu(&pwq->pwqs_node);
Tejun Heobc0caf02013-04-01 11:23:31 -07003415 is_last = list_empty(&wq->pwqs);
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003416 mutex_unlock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003417
Tejun Heoa892cac2013-04-01 11:23:32 -07003418 mutex_lock(&wq_pool_mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003419 put_unbound_pool(pool);
Tejun Heoa892cac2013-04-01 11:23:32 -07003420 mutex_unlock(&wq_pool_mutex);
3421
Tejun Heo8864b4e2013-03-12 11:30:04 -07003422 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3423
3424 /*
3425 * If we're the last pwq going away, @wq is already dead and no one
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003426 * is gonna access it anymore. Schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003427 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003428 if (is_last)
3429 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003430}
3431
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003432/**
Tejun Heo699ce092013-03-13 16:51:35 -07003433 * pwq_adjust_max_active - update a pwq's max_active to the current setting
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003434 * @pwq: target pool_workqueue
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003435 *
Tejun Heo699ce092013-03-13 16:51:35 -07003436 * If @pwq isn't freezing, set @pwq->max_active to the associated
3437 * workqueue's saved_max_active and activate delayed work items
3438 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003439 */
Tejun Heo699ce092013-03-13 16:51:35 -07003440static void pwq_adjust_max_active(struct pool_workqueue *pwq)
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003441{
Tejun Heo699ce092013-03-13 16:51:35 -07003442 struct workqueue_struct *wq = pwq->wq;
3443 bool freezable = wq->flags & WQ_FREEZABLE;
Tejun Heo3347fa02016-09-16 15:49:32 -04003444 unsigned long flags;
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003445
Tejun Heo699ce092013-03-13 16:51:35 -07003446 /* for @wq->saved_max_active */
Lai Jiangshana357fc02013-03-25 16:57:19 -07003447 lockdep_assert_held(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003448
3449 /* fast exit for non-freezable wqs */
3450 if (!freezable && pwq->max_active == wq->saved_max_active)
3451 return;
3452
Tejun Heo3347fa02016-09-16 15:49:32 -04003453 /* this function can be called during early boot w/ irq disabled */
3454 spin_lock_irqsave(&pwq->pool->lock, flags);
Tejun Heo699ce092013-03-13 16:51:35 -07003455
Lai Jiangshan74b414e2014-05-22 19:01:16 +08003456 /*
3457 * During [un]freezing, the caller is responsible for ensuring that
3458 * this function is called at least once after @workqueue_freezing
3459 * is updated and visible.
3460 */
3461 if (!freezable || !workqueue_freezing) {
Tejun Heo699ce092013-03-13 16:51:35 -07003462 pwq->max_active = wq->saved_max_active;
3463
3464 while (!list_empty(&pwq->delayed_works) &&
3465 pwq->nr_active < pwq->max_active)
3466 pwq_activate_first_delayed(pwq);
Lai Jiangshan951a0782013-03-20 10:52:30 -07003467
3468 /*
3469 * Need to kick a worker after thawed or an unbound wq's
3470 * max_active is bumped. It's a slow path. Do it always.
3471 */
3472 wake_up_worker(pwq->pool);
Tejun Heo699ce092013-03-13 16:51:35 -07003473 } else {
3474 pwq->max_active = 0;
3475 }
3476
Tejun Heo3347fa02016-09-16 15:49:32 -04003477 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003478}
3479
Tejun Heoe50aba92013-04-01 11:23:35 -07003480/* initialize newly alloced @pwq which is associated with @wq and @pool */
Tejun Heof147f292013-04-01 11:23:35 -07003481static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3482 struct worker_pool *pool)
Tejun Heod2c1d402013-03-12 11:30:04 -07003483{
3484 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3485
Tejun Heoe50aba92013-04-01 11:23:35 -07003486 memset(pwq, 0, sizeof(*pwq));
3487
Tejun Heod2c1d402013-03-12 11:30:04 -07003488 pwq->pool = pool;
3489 pwq->wq = wq;
3490 pwq->flush_color = -1;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003491 pwq->refcnt = 1;
Tejun Heod2c1d402013-03-12 11:30:04 -07003492 INIT_LIST_HEAD(&pwq->delayed_works);
Tejun Heo1befcf32013-04-01 11:23:35 -07003493 INIT_LIST_HEAD(&pwq->pwqs_node);
Tejun Heod2c1d402013-03-12 11:30:04 -07003494 INIT_LIST_HEAD(&pwq->mayday_node);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003495 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
Tejun Heof147f292013-04-01 11:23:35 -07003496}
Tejun Heod2c1d402013-03-12 11:30:04 -07003497
Tejun Heof147f292013-04-01 11:23:35 -07003498/* sync @pwq with the current state of its associated wq and link it */
Tejun Heo1befcf32013-04-01 11:23:35 -07003499static void link_pwq(struct pool_workqueue *pwq)
Tejun Heof147f292013-04-01 11:23:35 -07003500{
3501 struct workqueue_struct *wq = pwq->wq;
3502
3503 lockdep_assert_held(&wq->mutex);
Tejun Heo75ccf592013-03-12 11:30:04 -07003504
Tejun Heo1befcf32013-04-01 11:23:35 -07003505 /* may be called multiple times, ignore if already linked */
3506 if (!list_empty(&pwq->pwqs_node))
3507 return;
3508
Lai Jiangshan29b1cb42014-07-22 13:04:27 +08003509 /* set the matching work_color */
Tejun Heo75ccf592013-03-12 11:30:04 -07003510 pwq->work_color = wq->work_color;
Tejun Heo983ca252013-03-13 16:51:35 -07003511
3512 /* sync max_active to the current setting */
3513 pwq_adjust_max_active(pwq);
3514
3515 /* link in @pwq */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003516 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
Tejun Heof147f292013-04-01 11:23:35 -07003517}
Lai Jiangshana357fc02013-03-25 16:57:19 -07003518
Tejun Heof147f292013-04-01 11:23:35 -07003519/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3520static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3521 const struct workqueue_attrs *attrs)
3522{
3523 struct worker_pool *pool;
3524 struct pool_workqueue *pwq;
3525
3526 lockdep_assert_held(&wq_pool_mutex);
3527
3528 pool = get_unbound_pool(attrs);
3529 if (!pool)
3530 return NULL;
3531
Tejun Heoe50aba92013-04-01 11:23:35 -07003532 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
Tejun Heof147f292013-04-01 11:23:35 -07003533 if (!pwq) {
3534 put_unbound_pool(pool);
3535 return NULL;
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003536 }
Tejun Heo6029a912013-04-01 11:23:34 -07003537
Tejun Heof147f292013-04-01 11:23:35 -07003538 init_pwq(pwq, wq, pool);
3539 return pwq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003540}
3541
Tejun Heo4c16bd32013-04-01 11:23:36 -07003542/**
Gong Zhaogang30186c62015-05-11 11:02:47 -04003543 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003544 * @attrs: the wq_attrs of the default pwq of the target workqueue
Tejun Heo4c16bd32013-04-01 11:23:36 -07003545 * @node: the target NUMA node
3546 * @cpu_going_down: if >= 0, the CPU to consider as offline
3547 * @cpumask: outarg, the resulting cpumask
3548 *
3549 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3550 * @cpu_going_down is >= 0, that cpu is considered offline during
Yacine Belkadid185af32013-07-31 14:59:24 -07003551 * calculation. The result is stored in @cpumask.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003552 *
3553 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3554 * enabled and @node has online CPUs requested by @attrs, the returned
3555 * cpumask is the intersection of the possible CPUs of @node and
3556 * @attrs->cpumask.
3557 *
3558 * The caller is responsible for ensuring that the cpumask of @node stays
3559 * stable.
Yacine Belkadid185af32013-07-31 14:59:24 -07003560 *
3561 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3562 * %false if equal.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003563 */
3564static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3565 int cpu_going_down, cpumask_t *cpumask)
3566{
Tejun Heod55262c2013-04-01 11:23:38 -07003567 if (!wq_numa_enabled || attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003568 goto use_dfl;
3569
3570 /* does @node have any online CPUs @attrs wants? */
3571 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3572 if (cpu_going_down >= 0)
3573 cpumask_clear_cpu(cpu_going_down, cpumask);
3574
3575 if (cpumask_empty(cpumask))
3576 goto use_dfl;
3577
3578 /* yeap, return possible CPUs in @node that @attrs wants */
3579 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3580 return !cpumask_equal(cpumask, attrs->cpumask);
3581
3582use_dfl:
3583 cpumask_copy(cpumask, attrs->cpumask);
3584 return false;
3585}
3586
Tejun Heo1befcf32013-04-01 11:23:35 -07003587/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3588static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3589 int node,
3590 struct pool_workqueue *pwq)
3591{
3592 struct pool_workqueue *old_pwq;
3593
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +08003594 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003595 lockdep_assert_held(&wq->mutex);
3596
3597 /* link_pwq() can handle duplicate calls */
3598 link_pwq(pwq);
3599
3600 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3601 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3602 return old_pwq;
3603}
3604
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003605/* context to store the prepared attrs & pwqs before applying */
3606struct apply_wqattrs_ctx {
3607 struct workqueue_struct *wq; /* target workqueue */
3608 struct workqueue_attrs *attrs; /* attrs to apply */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003609 struct list_head list; /* queued for batching commit */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003610 struct pool_workqueue *dfl_pwq;
3611 struct pool_workqueue *pwq_tbl[];
3612};
3613
3614/* free the resources after success or abort */
3615static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3616{
3617 if (ctx) {
3618 int node;
3619
3620 for_each_node(node)
3621 put_pwq_unlocked(ctx->pwq_tbl[node]);
3622 put_pwq_unlocked(ctx->dfl_pwq);
3623
3624 free_workqueue_attrs(ctx->attrs);
3625
3626 kfree(ctx);
3627 }
3628}
3629
3630/* allocate the attrs and pwqs for later installation */
3631static struct apply_wqattrs_ctx *
3632apply_wqattrs_prepare(struct workqueue_struct *wq,
3633 const struct workqueue_attrs *attrs)
3634{
3635 struct apply_wqattrs_ctx *ctx;
3636 struct workqueue_attrs *new_attrs, *tmp_attrs;
3637 int node;
3638
3639 lockdep_assert_held(&wq_pool_mutex);
3640
3641 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3642 GFP_KERNEL);
3643
3644 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3645 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3646 if (!ctx || !new_attrs || !tmp_attrs)
3647 goto out_free;
3648
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003649 /*
3650 * Calculate the attrs of the default pwq.
3651 * If the user configured cpumask doesn't overlap with the
3652 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3653 */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003654 copy_workqueue_attrs(new_attrs, attrs);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08003655 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003656 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3657 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003658
3659 /*
3660 * We may create multiple pwqs with differing cpumasks. Make a
3661 * copy of @new_attrs which will be modified and used to obtain
3662 * pools.
3663 */
3664 copy_workqueue_attrs(tmp_attrs, new_attrs);
3665
3666 /*
3667 * If something goes wrong during CPU up/down, we'll fall back to
3668 * the default pwq covering whole @attrs->cpumask. Always create
3669 * it even if we don't use it immediately.
3670 */
3671 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3672 if (!ctx->dfl_pwq)
3673 goto out_free;
3674
3675 for_each_node(node) {
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003676 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003677 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3678 if (!ctx->pwq_tbl[node])
3679 goto out_free;
3680 } else {
3681 ctx->dfl_pwq->refcnt++;
3682 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3683 }
3684 }
3685
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003686 /* save the user configured attrs and sanitize it. */
3687 copy_workqueue_attrs(new_attrs, attrs);
3688 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003689 ctx->attrs = new_attrs;
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003690
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003691 ctx->wq = wq;
3692 free_workqueue_attrs(tmp_attrs);
3693 return ctx;
3694
3695out_free:
3696 free_workqueue_attrs(tmp_attrs);
3697 free_workqueue_attrs(new_attrs);
3698 apply_wqattrs_cleanup(ctx);
3699 return NULL;
3700}
3701
3702/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3703static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3704{
3705 int node;
3706
3707 /* all pwqs have been created successfully, let's install'em */
3708 mutex_lock(&ctx->wq->mutex);
3709
3710 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3711
3712 /* save the previous pwq and install the new one */
3713 for_each_node(node)
3714 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3715 ctx->pwq_tbl[node]);
3716
3717 /* @dfl_pwq might not have been used, ensure it's linked */
3718 link_pwq(ctx->dfl_pwq);
3719 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3720
3721 mutex_unlock(&ctx->wq->mutex);
3722}
3723
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003724static void apply_wqattrs_lock(void)
3725{
3726 /* CPUs should stay stable across pwq creations and installations */
3727 get_online_cpus();
3728 mutex_lock(&wq_pool_mutex);
3729}
3730
3731static void apply_wqattrs_unlock(void)
3732{
3733 mutex_unlock(&wq_pool_mutex);
3734 put_online_cpus();
3735}
3736
3737static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3738 const struct workqueue_attrs *attrs)
3739{
3740 struct apply_wqattrs_ctx *ctx;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003741
3742 /* only unbound workqueues can change attributes */
3743 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3744 return -EINVAL;
3745
3746 /* creating multiple pwqs breaks ordering guarantee */
3747 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3748 return -EINVAL;
3749
3750 ctx = apply_wqattrs_prepare(wq, attrs);
wanghaibin62011712016-01-07 20:38:59 +08003751 if (!ctx)
3752 return -ENOMEM;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003753
3754 /* the ctx has been prepared successfully, let's commit it */
wanghaibin62011712016-01-07 20:38:59 +08003755 apply_wqattrs_commit(ctx);
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003756 apply_wqattrs_cleanup(ctx);
3757
wanghaibin62011712016-01-07 20:38:59 +08003758 return 0;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003759}
3760
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003761/**
3762 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3763 * @wq: the target workqueue
3764 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3765 *
Tejun Heo4c16bd32013-04-01 11:23:36 -07003766 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3767 * machines, this function maps a separate pwq to each NUMA node with
3768 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3769 * NUMA node it was issued on. Older pwqs are released as in-flight work
3770 * items finish. Note that a work item which repeatedly requeues itself
3771 * back-to-back will stay on its current pwq.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003772 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003773 * Performs GFP_KERNEL allocations.
3774 *
3775 * Return: 0 on success and -errno on failure.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003776 */
3777int apply_workqueue_attrs(struct workqueue_struct *wq,
3778 const struct workqueue_attrs *attrs)
3779{
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003780 int ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003781
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003782 apply_wqattrs_lock();
3783 ret = apply_workqueue_attrs_locked(wq, attrs);
3784 apply_wqattrs_unlock();
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003785
Tejun Heo48621252013-04-01 11:23:31 -07003786 return ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003787}
3788
Tejun Heo4c16bd32013-04-01 11:23:36 -07003789/**
3790 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3791 * @wq: the target workqueue
3792 * @cpu: the CPU coming up or going down
3793 * @online: whether @cpu is coming up or going down
3794 *
3795 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3796 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3797 * @wq accordingly.
3798 *
3799 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3800 * falls back to @wq->dfl_pwq which may not be optimal but is always
3801 * correct.
3802 *
3803 * Note that when the last allowed CPU of a NUMA node goes offline for a
3804 * workqueue with a cpumask spanning multiple nodes, the workers which were
3805 * already executing the work items for the workqueue will lose their CPU
3806 * affinity and may execute on any CPU. This is similar to how per-cpu
3807 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3808 * affinity, it's the user's responsibility to flush the work item from
3809 * CPU_DOWN_PREPARE.
3810 */
3811static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3812 bool online)
3813{
3814 int node = cpu_to_node(cpu);
3815 int cpu_off = online ? -1 : cpu;
3816 struct pool_workqueue *old_pwq = NULL, *pwq;
3817 struct workqueue_attrs *target_attrs;
3818 cpumask_t *cpumask;
3819
3820 lockdep_assert_held(&wq_pool_mutex);
3821
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003822 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3823 wq->unbound_attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003824 return;
3825
3826 /*
3827 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3828 * Let's use a preallocated one. The following buf is protected by
3829 * CPU hotplug exclusion.
3830 */
3831 target_attrs = wq_update_unbound_numa_attrs_buf;
3832 cpumask = target_attrs->cpumask;
3833
Tejun Heo4c16bd32013-04-01 11:23:36 -07003834 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3835 pwq = unbound_pwq_by_node(wq, node);
3836
3837 /*
3838 * Let's determine what needs to be done. If the target cpumask is
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003839 * different from the default pwq's, we need to compare it to @pwq's
3840 * and create a new one if they don't match. If the target cpumask
3841 * equals the default pwq's, the default pwq should be used.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003842 */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003843 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
Tejun Heo4c16bd32013-04-01 11:23:36 -07003844 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003845 return;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003846 } else {
Daeseok Youn534a3fb2014-04-18 09:08:14 +09003847 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003848 }
3849
Tejun Heo4c16bd32013-04-01 11:23:36 -07003850 /* create a new pwq */
3851 pwq = alloc_unbound_pwq(wq, target_attrs);
3852 if (!pwq) {
Fabian Frederick2d916032014-05-12 13:59:35 -04003853 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3854 wq->name);
Daeseok Youn77f300b2014-04-16 14:32:29 +09003855 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003856 }
3857
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003858 /* Install the new pwq. */
Tejun Heo4c16bd32013-04-01 11:23:36 -07003859 mutex_lock(&wq->mutex);
3860 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3861 goto out_unlock;
3862
3863use_dfl_pwq:
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003864 mutex_lock(&wq->mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07003865 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3866 get_pwq(wq->dfl_pwq);
3867 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3868 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3869out_unlock:
3870 mutex_unlock(&wq->mutex);
3871 put_pwq_unlocked(old_pwq);
3872}
3873
Tejun Heo30cdf2492013-03-12 11:29:57 -07003874static int alloc_and_link_pwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875{
Tejun Heo49e3cf42013-03-12 11:29:58 -07003876 bool highpri = wq->flags & WQ_HIGHPRI;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003877 int cpu, ret;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01003878
Tejun Heo30cdf2492013-03-12 11:29:57 -07003879 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo420c0dd2013-03-12 11:29:59 -07003880 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3881 if (!wq->cpu_pwqs)
Tejun Heo30cdf2492013-03-12 11:29:57 -07003882 return -ENOMEM;
3883
3884 for_each_possible_cpu(cpu) {
Tejun Heo7fb98ea2013-03-12 11:30:00 -07003885 struct pool_workqueue *pwq =
3886 per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heo7a62c2c2013-03-12 11:30:03 -07003887 struct worker_pool *cpu_pools =
Tejun Heof02ae732013-03-12 11:30:03 -07003888 per_cpu(cpu_worker_pools, cpu);
Tejun Heo30cdf2492013-03-12 11:29:57 -07003889
Tejun Heof147f292013-04-01 11:23:35 -07003890 init_pwq(pwq, wq, &cpu_pools[highpri]);
3891
3892 mutex_lock(&wq->mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003893 link_pwq(pwq);
Tejun Heof147f292013-04-01 11:23:35 -07003894 mutex_unlock(&wq->mutex);
Tejun Heo30cdf2492013-03-12 11:29:57 -07003895 }
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003896 return 0;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003897 } else if (wq->flags & __WQ_ORDERED) {
3898 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3899 /* there should only be single pwq for ordering guarantee */
3900 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3901 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3902 "ordering guarantee broken for workqueue %s\n", wq->name);
3903 return ret;
Tejun Heo30cdf2492013-03-12 11:29:57 -07003904 } else {
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003905 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
Tejun Heo30cdf2492013-03-12 11:29:57 -07003906 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07003907}
3908
Tejun Heof3421792010-07-02 10:03:51 +02003909static int wq_clamp_max_active(int max_active, unsigned int flags,
3910 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02003911{
Tejun Heof3421792010-07-02 10:03:51 +02003912 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3913
3914 if (max_active < 1 || max_active > lim)
Valentin Ilie044c7822012-08-19 00:52:42 +03003915 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3916 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003917
Tejun Heof3421792010-07-02 10:03:51 +02003918 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003919}
3920
Tejun Heob196be82012-01-10 15:11:35 -08003921struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
Tejun Heod320c032010-06-29 10:07:14 +02003922 unsigned int flags,
3923 int max_active,
3924 struct lock_class_key *key,
Tejun Heob196be82012-01-10 15:11:35 -08003925 const char *lock_name, ...)
Oleg Nesterov3af244332007-05-09 02:34:09 -07003926{
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003927 size_t tbl_size = 0;
Tejun Heoecf68812013-04-01 11:23:34 -07003928 va_list args;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003929 struct workqueue_struct *wq;
Tejun Heo49e3cf42013-03-12 11:29:58 -07003930 struct pool_workqueue *pwq;
Tejun Heob196be82012-01-10 15:11:35 -08003931
Viresh Kumarcee22a12013-04-08 16:45:40 +05303932 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3933 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3934 flags |= WQ_UNBOUND;
3935
Tejun Heoecf68812013-04-01 11:23:34 -07003936 /* allocate wq and format name */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003937 if (flags & WQ_UNBOUND)
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08003938 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003939
3940 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
Tejun Heob196be82012-01-10 15:11:35 -08003941 if (!wq)
Tejun Heod2c1d402013-03-12 11:30:04 -07003942 return NULL;
Tejun Heob196be82012-01-10 15:11:35 -08003943
Tejun Heo6029a912013-04-01 11:23:34 -07003944 if (flags & WQ_UNBOUND) {
3945 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3946 if (!wq->unbound_attrs)
3947 goto err_free_wq;
3948 }
3949
Tejun Heoecf68812013-04-01 11:23:34 -07003950 va_start(args, lock_name);
3951 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
Tejun Heob196be82012-01-10 15:11:35 -08003952 va_end(args);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003953
Tejun Heod320c032010-06-29 10:07:14 +02003954 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heob196be82012-01-10 15:11:35 -08003955 max_active = wq_clamp_max_active(max_active, flags, wq->name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003956
Tejun Heob196be82012-01-10 15:11:35 -08003957 /* init wq */
Tejun Heo97e37d72010-06-29 10:07:10 +02003958 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003959 wq->saved_max_active = max_active;
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003960 mutex_init(&wq->mutex);
Tejun Heo112202d2013-02-13 19:29:12 -08003961 atomic_set(&wq->nr_pwqs_to_flush, 0);
Tejun Heo30cdf2492013-03-12 11:29:57 -07003962 INIT_LIST_HEAD(&wq->pwqs);
Tejun Heo73f53c42010-06-29 10:07:11 +02003963 INIT_LIST_HEAD(&wq->flusher_queue);
3964 INIT_LIST_HEAD(&wq->flusher_overflow);
Tejun Heo493a1722013-03-12 11:29:59 -07003965 INIT_LIST_HEAD(&wq->maydays);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003966
Johannes Bergeb13ba82008-01-16 09:51:58 +01003967 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07003968 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003969
Tejun Heo30cdf2492013-03-12 11:29:57 -07003970 if (alloc_and_link_pwqs(wq) < 0)
Tejun Heod2c1d402013-03-12 11:30:04 -07003971 goto err_free_wq;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003972
Tejun Heo493008a2013-03-12 11:30:03 -07003973 /*
3974 * Workqueues which may be used during memory reclaim should
3975 * have a rescuer to guarantee forward progress.
3976 */
3977 if (flags & WQ_MEM_RECLAIM) {
Tejun Heoe22bee72010-06-29 10:07:14 +02003978 struct worker *rescuer;
3979
Lai Jiangshanf7537df2014-07-15 17:24:15 +08003980 rescuer = alloc_worker(NUMA_NO_NODE);
Tejun Heoe22bee72010-06-29 10:07:14 +02003981 if (!rescuer)
Tejun Heod2c1d402013-03-12 11:30:04 -07003982 goto err_destroy;
Tejun Heoe22bee72010-06-29 10:07:14 +02003983
Tejun Heo111c2252013-01-17 17:16:24 -08003984 rescuer->rescue_wq = wq;
3985 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
Tejun Heob196be82012-01-10 15:11:35 -08003986 wq->name);
Tejun Heod2c1d402013-03-12 11:30:04 -07003987 if (IS_ERR(rescuer->task)) {
3988 kfree(rescuer);
3989 goto err_destroy;
3990 }
Tejun Heoe22bee72010-06-29 10:07:14 +02003991
Tejun Heod2c1d402013-03-12 11:30:04 -07003992 wq->rescuer = rescuer;
Peter Zijlstra25834c72015-05-15 17:43:34 +02003993 kthread_bind_mask(rescuer->task, cpu_possible_mask);
Tejun Heoe22bee72010-06-29 10:07:14 +02003994 wake_up_process(rescuer->task);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003995 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07003996
Tejun Heo226223a2013-03-12 11:30:05 -07003997 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3998 goto err_destroy;
3999
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004000 /*
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004001 * wq_pool_mutex protects global freeze state and workqueues list.
4002 * Grab it, adjust max_active and add the new @wq to workqueues
4003 * list.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004004 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004005 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004006
Lai Jiangshana357fc02013-03-25 16:57:19 -07004007 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004008 for_each_pwq(pwq, wq)
4009 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004010 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004011
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004012 list_add_tail_rcu(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004013
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004014 mutex_unlock(&wq_pool_mutex);
Tejun Heo15376632010-06-29 10:07:11 +02004015
Oleg Nesterov3af244332007-05-09 02:34:09 -07004016 return wq;
Tejun Heod2c1d402013-03-12 11:30:04 -07004017
4018err_free_wq:
Tejun Heo6029a912013-04-01 11:23:34 -07004019 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heod2c1d402013-03-12 11:30:04 -07004020 kfree(wq);
4021 return NULL;
4022err_destroy:
4023 destroy_workqueue(wq);
Tejun Heo4690c4a2010-06-29 10:07:10 +02004024 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004025}
Tejun Heod320c032010-06-29 10:07:14 +02004026EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004027
4028/**
4029 * destroy_workqueue - safely terminate a workqueue
4030 * @wq: target workqueue
4031 *
4032 * Safely destroy a workqueue. All work currently pending will be done first.
4033 */
4034void destroy_workqueue(struct workqueue_struct *wq)
4035{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004036 struct pool_workqueue *pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004037 int node;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004038
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02004039 /* drain it before proceeding with destruction */
4040 drain_workqueue(wq);
Tejun Heoc8efcc22010-12-20 19:32:04 +01004041
Tejun Heo6183c002013-03-12 11:29:57 -07004042 /* sanity checks */
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004043 mutex_lock(&wq->mutex);
Tejun Heo49e3cf42013-03-12 11:29:58 -07004044 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004045 int i;
4046
Tejun Heo76af4d92013-03-12 11:30:00 -07004047 for (i = 0; i < WORK_NR_COLORS; i++) {
4048 if (WARN_ON(pwq->nr_in_flight[i])) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004049 mutex_unlock(&wq->mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004050 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004051 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004052 }
4053 }
4054
Lai Jiangshan5c529592013-04-04 10:05:38 +08004055 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
Tejun Heo8864b4e2013-03-12 11:30:04 -07004056 WARN_ON(pwq->nr_active) ||
Tejun Heo76af4d92013-03-12 11:30:00 -07004057 WARN_ON(!list_empty(&pwq->delayed_works))) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004058 mutex_unlock(&wq->mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004059 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004060 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004061 }
Tejun Heo6183c002013-03-12 11:29:57 -07004062 }
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004063 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004064
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004065 /*
4066 * wq list is used to freeze wq, remove from list after
4067 * flushing is complete in case freeze races us.
4068 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004069 mutex_lock(&wq_pool_mutex);
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004070 list_del_rcu(&wq->list);
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004071 mutex_unlock(&wq_pool_mutex);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004072
Tejun Heo226223a2013-03-12 11:30:05 -07004073 workqueue_sysfs_unregister(wq);
4074
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004075 if (wq->rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02004076 kthread_stop(wq->rescuer->task);
Tejun Heoe22bee72010-06-29 10:07:14 +02004077
Tejun Heo8864b4e2013-03-12 11:30:04 -07004078 if (!(wq->flags & WQ_UNBOUND)) {
4079 /*
4080 * The base ref is never dropped on per-cpu pwqs. Directly
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004081 * schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004082 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004083 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07004084 } else {
4085 /*
4086 * We're the sole accessor of @wq at this point. Directly
Tejun Heo4c16bd32013-04-01 11:23:36 -07004087 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4088 * @wq will be freed when the last pwq is released.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004089 */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004090 for_each_node(node) {
4091 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4092 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4093 put_pwq_unlocked(pwq);
4094 }
4095
4096 /*
4097 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4098 * put. Don't access it afterwards.
4099 */
4100 pwq = wq->dfl_pwq;
4101 wq->dfl_pwq = NULL;
Tejun Heodce90d42013-04-01 11:23:35 -07004102 put_pwq_unlocked(pwq);
Tejun Heo29c91e92013-03-12 11:30:03 -07004103 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004104}
4105EXPORT_SYMBOL_GPL(destroy_workqueue);
4106
Tejun Heodcd989c2010-06-29 10:07:14 +02004107/**
4108 * workqueue_set_max_active - adjust max_active of a workqueue
4109 * @wq: target workqueue
4110 * @max_active: new max_active value.
4111 *
4112 * Set max_active of @wq to @max_active.
4113 *
4114 * CONTEXT:
4115 * Don't call from IRQ context.
4116 */
4117void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4118{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004119 struct pool_workqueue *pwq;
Tejun Heodcd989c2010-06-29 10:07:14 +02004120
Tejun Heo8719dce2013-03-12 11:30:04 -07004121 /* disallow meddling with max_active for ordered workqueues */
4122 if (WARN_ON(wq->flags & __WQ_ORDERED))
4123 return;
4124
Tejun Heof3421792010-07-02 10:03:51 +02004125 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02004126
Lai Jiangshana357fc02013-03-25 16:57:19 -07004127 mutex_lock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004128
4129 wq->saved_max_active = max_active;
4130
Tejun Heo699ce092013-03-13 16:51:35 -07004131 for_each_pwq(pwq, wq)
4132 pwq_adjust_max_active(pwq);
Tejun Heodcd989c2010-06-29 10:07:14 +02004133
Lai Jiangshana357fc02013-03-25 16:57:19 -07004134 mutex_unlock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004135}
4136EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4137
4138/**
Tejun Heoe62676162013-03-12 17:41:37 -07004139 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4140 *
4141 * Determine whether %current is a workqueue rescuer. Can be used from
4142 * work functions to determine whether it's being run off the rescuer task.
Yacine Belkadid185af32013-07-31 14:59:24 -07004143 *
4144 * Return: %true if %current is a workqueue rescuer. %false otherwise.
Tejun Heoe62676162013-03-12 17:41:37 -07004145 */
4146bool current_is_workqueue_rescuer(void)
4147{
4148 struct worker *worker = current_wq_worker();
4149
Lai Jiangshan6a092df2013-03-20 03:28:03 +08004150 return worker && worker->rescue_wq;
Tejun Heoe62676162013-03-12 17:41:37 -07004151}
4152
4153/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004154 * workqueue_congested - test whether a workqueue is congested
4155 * @cpu: CPU in question
4156 * @wq: target workqueue
4157 *
4158 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4159 * no synchronization around this function and the test result is
4160 * unreliable and only useful as advisory hints or for debugging.
4161 *
Tejun Heod3251852013-05-10 11:10:17 -07004162 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4163 * Note that both per-cpu and unbound workqueues may be associated with
4164 * multiple pool_workqueues which have separate congested states. A
4165 * workqueue being congested on one CPU doesn't mean the workqueue is also
4166 * contested on other CPUs / NUMA nodes.
4167 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004168 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004169 * %true if congested, %false otherwise.
4170 */
Tejun Heod84ff052013-03-12 11:29:59 -07004171bool workqueue_congested(int cpu, struct workqueue_struct *wq)
Tejun Heodcd989c2010-06-29 10:07:14 +02004172{
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004173 struct pool_workqueue *pwq;
Tejun Heo76af4d92013-03-12 11:30:00 -07004174 bool ret;
4175
Lai Jiangshan88109452013-03-20 03:28:10 +08004176 rcu_read_lock_sched();
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004177
Tejun Heod3251852013-05-10 11:10:17 -07004178 if (cpu == WORK_CPU_UNBOUND)
4179 cpu = smp_processor_id();
4180
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004181 if (!(wq->flags & WQ_UNBOUND))
4182 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4183 else
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004184 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodcd989c2010-06-29 10:07:14 +02004185
Tejun Heo76af4d92013-03-12 11:30:00 -07004186 ret = !list_empty(&pwq->delayed_works);
Lai Jiangshan88109452013-03-20 03:28:10 +08004187 rcu_read_unlock_sched();
Tejun Heo76af4d92013-03-12 11:30:00 -07004188
4189 return ret;
Tejun Heodcd989c2010-06-29 10:07:14 +02004190}
4191EXPORT_SYMBOL_GPL(workqueue_congested);
4192
4193/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004194 * work_busy - test whether a work is currently pending or running
4195 * @work: the work to be tested
4196 *
4197 * Test whether @work is currently pending or running. There is no
4198 * synchronization around this function and the test result is
4199 * unreliable and only useful as advisory hints or for debugging.
Tejun Heodcd989c2010-06-29 10:07:14 +02004200 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004201 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004202 * OR'd bitmask of WORK_BUSY_* bits.
4203 */
4204unsigned int work_busy(struct work_struct *work)
4205{
Tejun Heofa1b54e2013-03-12 11:30:00 -07004206 struct worker_pool *pool;
Tejun Heodcd989c2010-06-29 10:07:14 +02004207 unsigned long flags;
4208 unsigned int ret = 0;
4209
Tejun Heodcd989c2010-06-29 10:07:14 +02004210 if (work_pending(work))
4211 ret |= WORK_BUSY_PENDING;
Tejun Heodcd989c2010-06-29 10:07:14 +02004212
Tejun Heofa1b54e2013-03-12 11:30:00 -07004213 local_irq_save(flags);
4214 pool = get_work_pool(work);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004215 if (pool) {
Tejun Heofa1b54e2013-03-12 11:30:00 -07004216 spin_lock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004217 if (find_worker_executing_work(pool, work))
4218 ret |= WORK_BUSY_RUNNING;
Tejun Heofa1b54e2013-03-12 11:30:00 -07004219 spin_unlock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004220 }
Tejun Heofa1b54e2013-03-12 11:30:00 -07004221 local_irq_restore(flags);
Tejun Heodcd989c2010-06-29 10:07:14 +02004222
4223 return ret;
4224}
4225EXPORT_SYMBOL_GPL(work_busy);
4226
Tejun Heo3d1cb202013-04-30 15:27:22 -07004227/**
4228 * set_worker_desc - set description for the current work item
4229 * @fmt: printf-style format string
4230 * @...: arguments for the format string
4231 *
4232 * This function can be called by a running work function to describe what
4233 * the work item is about. If the worker task gets dumped, this
4234 * information will be printed out together to help debugging. The
4235 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4236 */
4237void set_worker_desc(const char *fmt, ...)
4238{
4239 struct worker *worker = current_wq_worker();
4240 va_list args;
4241
4242 if (worker) {
4243 va_start(args, fmt);
4244 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4245 va_end(args);
4246 worker->desc_valid = true;
4247 }
4248}
4249
4250/**
4251 * print_worker_info - print out worker information and description
4252 * @log_lvl: the log level to use when printing
4253 * @task: target task
4254 *
4255 * If @task is a worker and currently executing a work item, print out the
4256 * name of the workqueue being serviced and worker description set with
4257 * set_worker_desc() by the currently executing work item.
4258 *
4259 * This function can be safely called on any task as long as the
4260 * task_struct itself is accessible. While safe, this function isn't
4261 * synchronized and may print out mixups or garbages of limited length.
4262 */
4263void print_worker_info(const char *log_lvl, struct task_struct *task)
4264{
4265 work_func_t *fn = NULL;
4266 char name[WQ_NAME_LEN] = { };
4267 char desc[WORKER_DESC_LEN] = { };
4268 struct pool_workqueue *pwq = NULL;
4269 struct workqueue_struct *wq = NULL;
4270 bool desc_valid = false;
4271 struct worker *worker;
4272
4273 if (!(task->flags & PF_WQ_WORKER))
4274 return;
4275
4276 /*
4277 * This function is called without any synchronization and @task
4278 * could be in any state. Be careful with dereferences.
4279 */
Petr Mladeke7005912016-10-11 13:55:17 -07004280 worker = kthread_probe_data(task);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004281
4282 /*
4283 * Carefully copy the associated workqueue's workfn and name. Keep
4284 * the original last '\0' in case the original contains garbage.
4285 */
4286 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4287 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4288 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4289 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4290
4291 /* copy worker description */
4292 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4293 if (desc_valid)
4294 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4295
4296 if (fn || name[0] || desc[0]) {
4297 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4298 if (desc[0])
4299 pr_cont(" (%s)", desc);
4300 pr_cont("\n");
4301 }
4302}
4303
Tejun Heo3494fc32015-03-09 09:22:28 -04004304static void pr_cont_pool_info(struct worker_pool *pool)
4305{
4306 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4307 if (pool->node != NUMA_NO_NODE)
4308 pr_cont(" node=%d", pool->node);
4309 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4310}
4311
4312static void pr_cont_work(bool comma, struct work_struct *work)
4313{
4314 if (work->func == wq_barrier_func) {
4315 struct wq_barrier *barr;
4316
4317 barr = container_of(work, struct wq_barrier, work);
4318
4319 pr_cont("%s BAR(%d)", comma ? "," : "",
4320 task_pid_nr(barr->task));
4321 } else {
4322 pr_cont("%s %pf", comma ? "," : "", work->func);
4323 }
4324}
4325
4326static void show_pwq(struct pool_workqueue *pwq)
4327{
4328 struct worker_pool *pool = pwq->pool;
4329 struct work_struct *work;
4330 struct worker *worker;
4331 bool has_in_flight = false, has_pending = false;
4332 int bkt;
4333
4334 pr_info(" pwq %d:", pool->id);
4335 pr_cont_pool_info(pool);
4336
4337 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4338 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4339
4340 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4341 if (worker->current_pwq == pwq) {
4342 has_in_flight = true;
4343 break;
4344 }
4345 }
4346 if (has_in_flight) {
4347 bool comma = false;
4348
4349 pr_info(" in-flight:");
4350 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4351 if (worker->current_pwq != pwq)
4352 continue;
4353
4354 pr_cont("%s %d%s:%pf", comma ? "," : "",
4355 task_pid_nr(worker->task),
4356 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4357 worker->current_func);
4358 list_for_each_entry(work, &worker->scheduled, entry)
4359 pr_cont_work(false, work);
4360 comma = true;
4361 }
4362 pr_cont("\n");
4363 }
4364
4365 list_for_each_entry(work, &pool->worklist, entry) {
4366 if (get_work_pwq(work) == pwq) {
4367 has_pending = true;
4368 break;
4369 }
4370 }
4371 if (has_pending) {
4372 bool comma = false;
4373
4374 pr_info(" pending:");
4375 list_for_each_entry(work, &pool->worklist, entry) {
4376 if (get_work_pwq(work) != pwq)
4377 continue;
4378
4379 pr_cont_work(comma, work);
4380 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4381 }
4382 pr_cont("\n");
4383 }
4384
4385 if (!list_empty(&pwq->delayed_works)) {
4386 bool comma = false;
4387
4388 pr_info(" delayed:");
4389 list_for_each_entry(work, &pwq->delayed_works, entry) {
4390 pr_cont_work(comma, work);
4391 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4392 }
4393 pr_cont("\n");
4394 }
4395}
4396
4397/**
4398 * show_workqueue_state - dump workqueue state
4399 *
Roger Lu7b776af2016-07-01 11:05:02 +08004400 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4401 * all busy workqueues and pools.
Tejun Heo3494fc32015-03-09 09:22:28 -04004402 */
4403void show_workqueue_state(void)
4404{
4405 struct workqueue_struct *wq;
4406 struct worker_pool *pool;
4407 unsigned long flags;
4408 int pi;
4409
4410 rcu_read_lock_sched();
4411
4412 pr_info("Showing busy workqueues and worker pools:\n");
4413
4414 list_for_each_entry_rcu(wq, &workqueues, list) {
4415 struct pool_workqueue *pwq;
4416 bool idle = true;
4417
4418 for_each_pwq(pwq, wq) {
4419 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4420 idle = false;
4421 break;
4422 }
4423 }
4424 if (idle)
4425 continue;
4426
4427 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4428
4429 for_each_pwq(pwq, wq) {
4430 spin_lock_irqsave(&pwq->pool->lock, flags);
4431 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4432 show_pwq(pwq);
4433 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4434 }
4435 }
4436
4437 for_each_pool(pool, pi) {
4438 struct worker *worker;
4439 bool first = true;
4440
4441 spin_lock_irqsave(&pool->lock, flags);
4442 if (pool->nr_workers == pool->nr_idle)
4443 goto next_pool;
4444
4445 pr_info("pool %d:", pool->id);
4446 pr_cont_pool_info(pool);
Tejun Heo82607adc2015-12-08 11:28:04 -05004447 pr_cont(" hung=%us workers=%d",
4448 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4449 pool->nr_workers);
Tejun Heo3494fc32015-03-09 09:22:28 -04004450 if (pool->manager)
4451 pr_cont(" manager: %d",
4452 task_pid_nr(pool->manager->task));
4453 list_for_each_entry(worker, &pool->idle_list, entry) {
4454 pr_cont(" %s%d", first ? "idle: " : "",
4455 task_pid_nr(worker->task));
4456 first = false;
4457 }
4458 pr_cont("\n");
4459 next_pool:
4460 spin_unlock_irqrestore(&pool->lock, flags);
4461 }
4462
4463 rcu_read_unlock_sched();
4464}
4465
Tejun Heodb7bccf2010-06-29 10:07:12 +02004466/*
4467 * CPU hotplug.
4468 *
Tejun Heoe22bee72010-06-29 10:07:14 +02004469 * There are two challenges in supporting CPU hotplug. Firstly, there
Tejun Heo112202d2013-02-13 19:29:12 -08004470 * are a lot of assumptions on strong associations among work, pwq and
Tejun Heo706026c2013-01-24 11:01:34 -08004471 * pool which make migrating pending and scheduled works very
Tejun Heoe22bee72010-06-29 10:07:14 +02004472 * difficult to implement without impacting hot paths. Secondly,
Tejun Heo94cf58b2013-01-24 11:01:33 -08004473 * worker pools serve mix of short, long and very long running works making
Tejun Heoe22bee72010-06-29 10:07:14 +02004474 * blocked draining impractical.
4475 *
Tejun Heo24647572013-01-24 11:01:33 -08004476 * This is solved by allowing the pools to be disassociated from the CPU
Tejun Heo628c78e2012-07-17 12:39:27 -07004477 * running as an unbound one and allowing it to be reattached later if the
4478 * cpu comes back online.
Tejun Heodb7bccf2010-06-29 10:07:12 +02004479 */
4480
Tejun Heo706026c2013-01-24 11:01:34 -08004481static void wq_unbind_fn(struct work_struct *work)
Tejun Heodb7bccf2010-06-29 10:07:12 +02004482{
Tejun Heo38db41d2013-01-24 11:01:34 -08004483 int cpu = smp_processor_id();
Tejun Heo4ce62e92012-07-13 22:16:44 -07004484 struct worker_pool *pool;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004485 struct worker *worker;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004486
Tejun Heof02ae732013-03-12 11:30:03 -07004487 for_each_cpu_worker_pool(pool, cpu) {
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004488 mutex_lock(&pool->attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004489 spin_lock_irq(&pool->lock);
4490
4491 /*
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004492 * We've blocked all attach/detach operations. Make all workers
Tejun Heo94cf58b2013-01-24 11:01:33 -08004493 * unbound and set DISASSOCIATED. Before this, all workers
4494 * except for the ones which are still executing works from
4495 * before the last CPU down must be on the cpu. After
4496 * this, they may become diasporas.
4497 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004498 for_each_pool_worker(worker, pool)
Tejun Heoc9e7cf22013-01-24 11:01:33 -08004499 worker->flags |= WORKER_UNBOUND;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004500
Tejun Heo24647572013-01-24 11:01:33 -08004501 pool->flags |= POOL_DISASSOCIATED;
Tejun Heof2d5a0e2012-07-17 12:39:26 -07004502
Tejun Heo94cf58b2013-01-24 11:01:33 -08004503 spin_unlock_irq(&pool->lock);
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004504 mutex_unlock(&pool->attach_mutex);
Tejun Heoe22bee72010-06-29 10:07:14 +02004505
Lai Jiangshaneb283422013-03-08 15:18:28 -08004506 /*
4507 * Call schedule() so that we cross rq->lock and thus can
4508 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4509 * This is necessary as scheduler callbacks may be invoked
4510 * from other cpus.
4511 */
4512 schedule();
Tejun Heo628c78e2012-07-17 12:39:27 -07004513
Lai Jiangshaneb283422013-03-08 15:18:28 -08004514 /*
4515 * Sched callbacks are disabled now. Zap nr_running.
4516 * After this, nr_running stays zero and need_more_worker()
4517 * and keep_working() are always true as long as the
4518 * worklist is not empty. This pool now behaves as an
4519 * unbound (in terms of concurrency management) pool which
4520 * are served by workers tied to the pool.
4521 */
Tejun Heoe19e3972013-01-24 11:39:44 -08004522 atomic_set(&pool->nr_running, 0);
Lai Jiangshaneb283422013-03-08 15:18:28 -08004523
4524 /*
4525 * With concurrency management just turned off, a busy
4526 * worker blocking could lead to lengthy stalls. Kick off
4527 * unbound chain execution of currently pending work items.
4528 */
4529 spin_lock_irq(&pool->lock);
4530 wake_up_worker(pool);
4531 spin_unlock_irq(&pool->lock);
4532 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004533}
4534
Tejun Heobd7c0892013-03-19 13:45:21 -07004535/**
4536 * rebind_workers - rebind all workers of a pool to the associated CPU
4537 * @pool: pool of interest
4538 *
Tejun Heoa9ab7752013-03-19 13:45:21 -07004539 * @pool->cpu is coming online. Rebind all workers to the CPU.
Tejun Heobd7c0892013-03-19 13:45:21 -07004540 */
4541static void rebind_workers(struct worker_pool *pool)
4542{
Tejun Heoa9ab7752013-03-19 13:45:21 -07004543 struct worker *worker;
Tejun Heobd7c0892013-03-19 13:45:21 -07004544
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004545 lockdep_assert_held(&pool->attach_mutex);
Tejun Heobd7c0892013-03-19 13:45:21 -07004546
Tejun Heoa9ab7752013-03-19 13:45:21 -07004547 /*
4548 * Restore CPU affinity of all workers. As all idle workers should
4549 * be on the run-queue of the associated CPU before any local
Shailendra Verma402dd892015-05-23 10:38:14 +05304550 * wake-ups for concurrency management happen, restore CPU affinity
Tejun Heoa9ab7752013-03-19 13:45:21 -07004551 * of all workers first and then clear UNBOUND. As we're called
4552 * from CPU_ONLINE, the following shouldn't fail.
4553 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004554 for_each_pool_worker(worker, pool)
Tejun Heoa9ab7752013-03-19 13:45:21 -07004555 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4556 pool->attrs->cpumask) < 0);
4557
4558 spin_lock_irq(&pool->lock);
Wanpeng Lif7c17d22016-05-11 17:55:18 +08004559
4560 /*
4561 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4562 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4563 * being reworked and this can go away in time.
4564 */
4565 if (!(pool->flags & POOL_DISASSOCIATED)) {
4566 spin_unlock_irq(&pool->lock);
4567 return;
4568 }
4569
Lai Jiangshan3de5e882014-06-03 15:33:27 +08004570 pool->flags &= ~POOL_DISASSOCIATED;
Tejun Heoa9ab7752013-03-19 13:45:21 -07004571
Lai Jiangshanda028462014-05-20 17:46:31 +08004572 for_each_pool_worker(worker, pool) {
Tejun Heoa9ab7752013-03-19 13:45:21 -07004573 unsigned int worker_flags = worker->flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004574
4575 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07004576 * A bound idle worker should actually be on the runqueue
4577 * of the associated CPU for local wake-ups targeting it to
4578 * work. Kick all idle workers so that they migrate to the
4579 * associated CPU. Doing this in the same loop as
4580 * replacing UNBOUND with REBOUND is safe as no worker will
4581 * be bound before @pool->lock is released.
Tejun Heobd7c0892013-03-19 13:45:21 -07004582 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004583 if (worker_flags & WORKER_IDLE)
4584 wake_up_process(worker->task);
4585
4586 /*
4587 * We want to clear UNBOUND but can't directly call
4588 * worker_clr_flags() or adjust nr_running. Atomically
4589 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4590 * @worker will clear REBOUND using worker_clr_flags() when
4591 * it initiates the next execution cycle thus restoring
4592 * concurrency management. Note that when or whether
4593 * @worker clears REBOUND doesn't affect correctness.
4594 *
4595 * ACCESS_ONCE() is necessary because @worker->flags may be
4596 * tested without holding any lock in
4597 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4598 * fail incorrectly leading to premature concurrency
4599 * management operations.
4600 */
4601 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4602 worker_flags |= WORKER_REBOUND;
4603 worker_flags &= ~WORKER_UNBOUND;
4604 ACCESS_ONCE(worker->flags) = worker_flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004605 }
4606
Tejun Heoa9ab7752013-03-19 13:45:21 -07004607 spin_unlock_irq(&pool->lock);
Tejun Heobd7c0892013-03-19 13:45:21 -07004608}
4609
Tejun Heo7dbc7252013-03-19 13:45:21 -07004610/**
4611 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4612 * @pool: unbound pool of interest
4613 * @cpu: the CPU which is coming up
4614 *
4615 * An unbound pool may end up with a cpumask which doesn't have any online
4616 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4617 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4618 * online CPU before, cpus_allowed of all its workers should be restored.
4619 */
4620static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4621{
4622 static cpumask_t cpumask;
4623 struct worker *worker;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004624
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004625 lockdep_assert_held(&pool->attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004626
4627 /* is @cpu allowed for @pool? */
4628 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4629 return;
4630
Tejun Heo7dbc7252013-03-19 13:45:21 -07004631 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004632
4633 /* as we're called from CPU_ONLINE, the following shouldn't fail */
Lai Jiangshanda028462014-05-20 17:46:31 +08004634 for_each_pool_worker(worker, pool)
Peter Zijlstrad945b5e92016-06-16 14:38:42 +02004635 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004636}
4637
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004638int workqueue_prepare_cpu(unsigned int cpu)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004639{
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004640 struct worker_pool *pool;
4641
4642 for_each_cpu_worker_pool(pool, cpu) {
4643 if (pool->nr_workers)
4644 continue;
4645 if (!create_worker(pool))
4646 return -ENOMEM;
4647 }
4648 return 0;
4649}
4650
4651int workqueue_online_cpu(unsigned int cpu)
4652{
Tejun Heo4ce62e92012-07-13 22:16:44 -07004653 struct worker_pool *pool;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004654 struct workqueue_struct *wq;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004655 int pi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004656
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004657 mutex_lock(&wq_pool_mutex);
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07004658
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004659 for_each_pool(pool, pi) {
4660 mutex_lock(&pool->attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004661
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004662 if (pool->cpu == cpu)
4663 rebind_workers(pool);
4664 else if (pool->cpu < 0)
4665 restore_unbound_workers_cpumask(pool, cpu);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004666
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004667 mutex_unlock(&pool->attach_mutex);
Tejun Heo65758202012-07-17 12:39:26 -07004668 }
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004669
4670 /* update NUMA affinity of unbound workqueues */
4671 list_for_each_entry(wq, &workqueues, list)
4672 wq_update_unbound_numa(wq, cpu, true);
4673
4674 mutex_unlock(&wq_pool_mutex);
4675 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07004676}
4677
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004678int workqueue_offline_cpu(unsigned int cpu)
Tejun Heo65758202012-07-17 12:39:26 -07004679{
Tejun Heo8db25e72012-07-17 12:39:28 -07004680 struct work_struct unbind_work;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004681 struct workqueue_struct *wq;
Tejun Heo8db25e72012-07-17 12:39:28 -07004682
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004683 /* unbinding per-cpu workers should happen on the local CPU */
4684 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4685 queue_work_on(cpu, system_highpri_wq, &unbind_work);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004686
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004687 /* update NUMA affinity of unbound workqueues */
4688 mutex_lock(&wq_pool_mutex);
4689 list_for_each_entry(wq, &workqueues, list)
4690 wq_update_unbound_numa(wq, cpu, false);
4691 mutex_unlock(&wq_pool_mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004692
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004693 /* wait for per-cpu unbinding to finish */
4694 flush_work(&unbind_work);
4695 destroy_work_on_stack(&unbind_work);
4696 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07004697}
4698
Rusty Russell2d3854a2008-11-05 13:39:10 +11004699#ifdef CONFIG_SMP
Rusty Russell8ccad402009-01-16 15:31:15 -08004700
Rusty Russell2d3854a2008-11-05 13:39:10 +11004701struct work_for_cpu {
Tejun Heoed48ece2012-09-18 12:48:43 -07004702 struct work_struct work;
Rusty Russell2d3854a2008-11-05 13:39:10 +11004703 long (*fn)(void *);
4704 void *arg;
4705 long ret;
4706};
4707
Tejun Heoed48ece2012-09-18 12:48:43 -07004708static void work_for_cpu_fn(struct work_struct *work)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004709{
Tejun Heoed48ece2012-09-18 12:48:43 -07004710 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4711
Rusty Russell2d3854a2008-11-05 13:39:10 +11004712 wfc->ret = wfc->fn(wfc->arg);
4713}
4714
4715/**
Anna-Maria Gleixner22aceb32016-03-10 12:07:38 +01004716 * work_on_cpu - run a function in thread context on a particular cpu
Rusty Russell2d3854a2008-11-05 13:39:10 +11004717 * @cpu: the cpu to run on
4718 * @fn: the function to run
4719 * @arg: the function arg
4720 *
Rusty Russell31ad9082009-01-16 15:31:15 -08004721 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06004722 * The caller must not hold any locks which would prevent @fn from completing.
Yacine Belkadid185af32013-07-31 14:59:24 -07004723 *
4724 * Return: The value @fn returns.
Rusty Russell2d3854a2008-11-05 13:39:10 +11004725 */
Tejun Heod84ff052013-03-12 11:29:59 -07004726long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004727{
Tejun Heoed48ece2012-09-18 12:48:43 -07004728 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
Rusty Russell2d3854a2008-11-05 13:39:10 +11004729
Tejun Heoed48ece2012-09-18 12:48:43 -07004730 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4731 schedule_work_on(cpu, &wfc.work);
Bjorn Helgaas12997d12013-11-18 11:00:29 -07004732 flush_work(&wfc.work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05004733 destroy_work_on_stack(&wfc.work);
Rusty Russell2d3854a2008-11-05 13:39:10 +11004734 return wfc.ret;
4735}
4736EXPORT_SYMBOL_GPL(work_on_cpu);
4737#endif /* CONFIG_SMP */
4738
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004739#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10304740
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004741/**
4742 * freeze_workqueues_begin - begin freezing workqueues
4743 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01004744 * Start freezing workqueues. After this function returns, all freezable
Tejun Heoc5aa87b2013-03-13 16:51:36 -07004745 * workqueues will queue new works to their delayed_works list instead of
Tejun Heo706026c2013-01-24 11:01:34 -08004746 * pool->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004747 *
4748 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004749 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004750 */
4751void freeze_workqueues_begin(void)
4752{
Tejun Heo24b8a842013-03-12 11:29:58 -07004753 struct workqueue_struct *wq;
4754 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004755
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004756 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004757
Tejun Heo6183c002013-03-12 11:29:57 -07004758 WARN_ON_ONCE(workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004759 workqueue_freezing = true;
4760
Tejun Heo24b8a842013-03-12 11:29:58 -07004761 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004762 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004763 for_each_pwq(pwq, wq)
4764 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004765 mutex_unlock(&wq->mutex);
Tejun Heo24b8a842013-03-12 11:29:58 -07004766 }
Tejun Heo5bcab332013-03-13 19:47:40 -07004767
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004768 mutex_unlock(&wq_pool_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004769}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004770
4771/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01004772 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004773 *
4774 * Check whether freezing is complete. This function must be called
4775 * between freeze_workqueues_begin() and thaw_workqueues().
4776 *
4777 * CONTEXT:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004778 * Grabs and releases wq_pool_mutex.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004779 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004780 * Return:
Tejun Heo58a69cb2011-02-16 09:25:31 +01004781 * %true if some freezable workqueues are still busy. %false if freezing
4782 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004783 */
4784bool freeze_workqueues_busy(void)
4785{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004786 bool busy = false;
Tejun Heo24b8a842013-03-12 11:29:58 -07004787 struct workqueue_struct *wq;
4788 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004789
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004790 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004791
Tejun Heo6183c002013-03-12 11:29:57 -07004792 WARN_ON_ONCE(!workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004793
Tejun Heo24b8a842013-03-12 11:29:58 -07004794 list_for_each_entry(wq, &workqueues, list) {
4795 if (!(wq->flags & WQ_FREEZABLE))
4796 continue;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004797 /*
4798 * nr_active is monotonically decreasing. It's safe
4799 * to peek without lock.
4800 */
Lai Jiangshan88109452013-03-20 03:28:10 +08004801 rcu_read_lock_sched();
Tejun Heo24b8a842013-03-12 11:29:58 -07004802 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004803 WARN_ON_ONCE(pwq->nr_active < 0);
Tejun Heo112202d2013-02-13 19:29:12 -08004804 if (pwq->nr_active) {
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004805 busy = true;
Lai Jiangshan88109452013-03-20 03:28:10 +08004806 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004807 goto out_unlock;
4808 }
4809 }
Lai Jiangshan88109452013-03-20 03:28:10 +08004810 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004811 }
4812out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004813 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004814 return busy;
4815}
4816
4817/**
4818 * thaw_workqueues - thaw workqueues
4819 *
4820 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo706026c2013-01-24 11:01:34 -08004821 * frozen works are transferred to their respective pool worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004822 *
4823 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004824 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004825 */
4826void thaw_workqueues(void)
4827{
Tejun Heo24b8a842013-03-12 11:29:58 -07004828 struct workqueue_struct *wq;
4829 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004830
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004831 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004832
4833 if (!workqueue_freezing)
4834 goto out_unlock;
4835
Lai Jiangshan74b414e2014-05-22 19:01:16 +08004836 workqueue_freezing = false;
Tejun Heo8b03ae32010-06-29 10:07:12 +02004837
Tejun Heo24b8a842013-03-12 11:29:58 -07004838 /* restore max_active and repopulate worklist */
4839 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004840 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004841 for_each_pwq(pwq, wq)
4842 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004843 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004844 }
4845
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004846out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004847 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004848}
4849#endif /* CONFIG_FREEZER */
4850
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004851static int workqueue_apply_unbound_cpumask(void)
4852{
4853 LIST_HEAD(ctxs);
4854 int ret = 0;
4855 struct workqueue_struct *wq;
4856 struct apply_wqattrs_ctx *ctx, *n;
4857
4858 lockdep_assert_held(&wq_pool_mutex);
4859
4860 list_for_each_entry(wq, &workqueues, list) {
4861 if (!(wq->flags & WQ_UNBOUND))
4862 continue;
4863 /* creating multiple pwqs breaks ordering guarantee */
4864 if (wq->flags & __WQ_ORDERED)
4865 continue;
4866
4867 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4868 if (!ctx) {
4869 ret = -ENOMEM;
4870 break;
4871 }
4872
4873 list_add_tail(&ctx->list, &ctxs);
4874 }
4875
4876 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4877 if (!ret)
4878 apply_wqattrs_commit(ctx);
4879 apply_wqattrs_cleanup(ctx);
4880 }
4881
4882 return ret;
4883}
4884
4885/**
4886 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4887 * @cpumask: the cpumask to set
4888 *
4889 * The low-level workqueues cpumask is a global cpumask that limits
4890 * the affinity of all unbound workqueues. This function check the @cpumask
4891 * and apply it to all unbound workqueues and updates all pwqs of them.
4892 *
4893 * Retun: 0 - Success
4894 * -EINVAL - Invalid @cpumask
4895 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4896 */
4897int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4898{
4899 int ret = -EINVAL;
4900 cpumask_var_t saved_cpumask;
4901
4902 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4903 return -ENOMEM;
4904
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004905 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4906 if (!cpumask_empty(cpumask)) {
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004907 apply_wqattrs_lock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004908
4909 /* save the old wq_unbound_cpumask. */
4910 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4911
4912 /* update wq_unbound_cpumask at first and apply it to wqs. */
4913 cpumask_copy(wq_unbound_cpumask, cpumask);
4914 ret = workqueue_apply_unbound_cpumask();
4915
4916 /* restore the wq_unbound_cpumask when failed. */
4917 if (ret < 0)
4918 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4919
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004920 apply_wqattrs_unlock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004921 }
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004922
4923 free_cpumask_var(saved_cpumask);
4924 return ret;
4925}
4926
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004927#ifdef CONFIG_SYSFS
4928/*
4929 * Workqueues with WQ_SYSFS flag set is visible to userland via
4930 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4931 * following attributes.
4932 *
4933 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4934 * max_active RW int : maximum number of in-flight work items
4935 *
4936 * Unbound workqueues have the following extra attributes.
4937 *
4938 * id RO int : the associated pool ID
4939 * nice RW int : nice value of the workers
4940 * cpumask RW mask : bitmask of allowed CPUs for the workers
4941 */
4942struct wq_device {
4943 struct workqueue_struct *wq;
4944 struct device dev;
4945};
4946
4947static struct workqueue_struct *dev_to_wq(struct device *dev)
4948{
4949 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4950
4951 return wq_dev->wq;
4952}
4953
4954static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4955 char *buf)
4956{
4957 struct workqueue_struct *wq = dev_to_wq(dev);
4958
4959 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4960}
4961static DEVICE_ATTR_RO(per_cpu);
4962
4963static ssize_t max_active_show(struct device *dev,
4964 struct device_attribute *attr, char *buf)
4965{
4966 struct workqueue_struct *wq = dev_to_wq(dev);
4967
4968 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4969}
4970
4971static ssize_t max_active_store(struct device *dev,
4972 struct device_attribute *attr, const char *buf,
4973 size_t count)
4974{
4975 struct workqueue_struct *wq = dev_to_wq(dev);
4976 int val;
4977
4978 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4979 return -EINVAL;
4980
4981 workqueue_set_max_active(wq, val);
4982 return count;
4983}
4984static DEVICE_ATTR_RW(max_active);
4985
4986static struct attribute *wq_sysfs_attrs[] = {
4987 &dev_attr_per_cpu.attr,
4988 &dev_attr_max_active.attr,
4989 NULL,
4990};
4991ATTRIBUTE_GROUPS(wq_sysfs);
4992
4993static ssize_t wq_pool_ids_show(struct device *dev,
4994 struct device_attribute *attr, char *buf)
4995{
4996 struct workqueue_struct *wq = dev_to_wq(dev);
4997 const char *delim = "";
4998 int node, written = 0;
4999
5000 rcu_read_lock_sched();
5001 for_each_node(node) {
5002 written += scnprintf(buf + written, PAGE_SIZE - written,
5003 "%s%d:%d", delim, node,
5004 unbound_pwq_by_node(wq, node)->pool->id);
5005 delim = " ";
5006 }
5007 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5008 rcu_read_unlock_sched();
5009
5010 return written;
5011}
5012
5013static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5014 char *buf)
5015{
5016 struct workqueue_struct *wq = dev_to_wq(dev);
5017 int written;
5018
5019 mutex_lock(&wq->mutex);
5020 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5021 mutex_unlock(&wq->mutex);
5022
5023 return written;
5024}
5025
5026/* prepare workqueue_attrs for sysfs store operations */
5027static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5028{
5029 struct workqueue_attrs *attrs;
5030
Lai Jiangshan899a94f2015-05-20 14:41:18 +08005031 lockdep_assert_held(&wq_pool_mutex);
5032
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005033 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5034 if (!attrs)
5035 return NULL;
5036
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005037 copy_workqueue_attrs(attrs, wq->unbound_attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005038 return attrs;
5039}
5040
5041static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5042 const char *buf, size_t count)
5043{
5044 struct workqueue_struct *wq = dev_to_wq(dev);
5045 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005046 int ret = -ENOMEM;
5047
5048 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005049
5050 attrs = wq_sysfs_prep_attrs(wq);
5051 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005052 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005053
5054 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5055 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005056 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005057 else
5058 ret = -EINVAL;
5059
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005060out_unlock:
5061 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005062 free_workqueue_attrs(attrs);
5063 return ret ?: count;
5064}
5065
5066static ssize_t wq_cpumask_show(struct device *dev,
5067 struct device_attribute *attr, char *buf)
5068{
5069 struct workqueue_struct *wq = dev_to_wq(dev);
5070 int written;
5071
5072 mutex_lock(&wq->mutex);
5073 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5074 cpumask_pr_args(wq->unbound_attrs->cpumask));
5075 mutex_unlock(&wq->mutex);
5076 return written;
5077}
5078
5079static ssize_t wq_cpumask_store(struct device *dev,
5080 struct device_attribute *attr,
5081 const char *buf, size_t count)
5082{
5083 struct workqueue_struct *wq = dev_to_wq(dev);
5084 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005085 int ret = -ENOMEM;
5086
5087 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005088
5089 attrs = wq_sysfs_prep_attrs(wq);
5090 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005091 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005092
5093 ret = cpumask_parse(buf, attrs->cpumask);
5094 if (!ret)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005095 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005096
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005097out_unlock:
5098 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005099 free_workqueue_attrs(attrs);
5100 return ret ?: count;
5101}
5102
5103static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5104 char *buf)
5105{
5106 struct workqueue_struct *wq = dev_to_wq(dev);
5107 int written;
5108
5109 mutex_lock(&wq->mutex);
5110 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5111 !wq->unbound_attrs->no_numa);
5112 mutex_unlock(&wq->mutex);
5113
5114 return written;
5115}
5116
5117static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5118 const char *buf, size_t count)
5119{
5120 struct workqueue_struct *wq = dev_to_wq(dev);
5121 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005122 int v, ret = -ENOMEM;
5123
5124 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005125
5126 attrs = wq_sysfs_prep_attrs(wq);
5127 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005128 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005129
5130 ret = -EINVAL;
5131 if (sscanf(buf, "%d", &v) == 1) {
5132 attrs->no_numa = !v;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005133 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005134 }
5135
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005136out_unlock:
5137 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005138 free_workqueue_attrs(attrs);
5139 return ret ?: count;
5140}
5141
5142static struct device_attribute wq_sysfs_unbound_attrs[] = {
5143 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5144 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5145 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5146 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5147 __ATTR_NULL,
5148};
5149
5150static struct bus_type wq_subsys = {
5151 .name = "workqueue",
5152 .dev_groups = wq_sysfs_groups,
5153};
5154
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005155static ssize_t wq_unbound_cpumask_show(struct device *dev,
5156 struct device_attribute *attr, char *buf)
5157{
5158 int written;
5159
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005160 mutex_lock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005161 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5162 cpumask_pr_args(wq_unbound_cpumask));
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005163 mutex_unlock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005164
5165 return written;
5166}
5167
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005168static ssize_t wq_unbound_cpumask_store(struct device *dev,
5169 struct device_attribute *attr, const char *buf, size_t count)
5170{
5171 cpumask_var_t cpumask;
5172 int ret;
5173
5174 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5175 return -ENOMEM;
5176
5177 ret = cpumask_parse(buf, cpumask);
5178 if (!ret)
5179 ret = workqueue_set_unbound_cpumask(cpumask);
5180
5181 free_cpumask_var(cpumask);
5182 return ret ? ret : count;
5183}
5184
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005185static struct device_attribute wq_sysfs_cpumask_attr =
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005186 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5187 wq_unbound_cpumask_store);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005188
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005189static int __init wq_sysfs_init(void)
5190{
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005191 int err;
5192
5193 err = subsys_virtual_register(&wq_subsys, NULL);
5194 if (err)
5195 return err;
5196
5197 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005198}
5199core_initcall(wq_sysfs_init);
5200
5201static void wq_device_release(struct device *dev)
5202{
5203 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5204
5205 kfree(wq_dev);
5206}
5207
5208/**
5209 * workqueue_sysfs_register - make a workqueue visible in sysfs
5210 * @wq: the workqueue to register
5211 *
5212 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5213 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5214 * which is the preferred method.
5215 *
5216 * Workqueue user should use this function directly iff it wants to apply
5217 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5218 * apply_workqueue_attrs() may race against userland updating the
5219 * attributes.
5220 *
5221 * Return: 0 on success, -errno on failure.
5222 */
5223int workqueue_sysfs_register(struct workqueue_struct *wq)
5224{
5225 struct wq_device *wq_dev;
5226 int ret;
5227
5228 /*
Shailendra Verma402dd892015-05-23 10:38:14 +05305229 * Adjusting max_active or creating new pwqs by applying
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005230 * attributes breaks ordering guarantee. Disallow exposing ordered
5231 * workqueues.
5232 */
5233 if (WARN_ON(wq->flags & __WQ_ORDERED))
5234 return -EINVAL;
5235
5236 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5237 if (!wq_dev)
5238 return -ENOMEM;
5239
5240 wq_dev->wq = wq;
5241 wq_dev->dev.bus = &wq_subsys;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005242 wq_dev->dev.release = wq_device_release;
Lars-Peter Clausen23217b42016-02-17 21:04:41 +01005243 dev_set_name(&wq_dev->dev, "%s", wq->name);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005244
5245 /*
5246 * unbound_attrs are created separately. Suppress uevent until
5247 * everything is ready.
5248 */
5249 dev_set_uevent_suppress(&wq_dev->dev, true);
5250
5251 ret = device_register(&wq_dev->dev);
5252 if (ret) {
5253 kfree(wq_dev);
5254 wq->wq_dev = NULL;
5255 return ret;
5256 }
5257
5258 if (wq->flags & WQ_UNBOUND) {
5259 struct device_attribute *attr;
5260
5261 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5262 ret = device_create_file(&wq_dev->dev, attr);
5263 if (ret) {
5264 device_unregister(&wq_dev->dev);
5265 wq->wq_dev = NULL;
5266 return ret;
5267 }
5268 }
5269 }
5270
5271 dev_set_uevent_suppress(&wq_dev->dev, false);
5272 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5273 return 0;
5274}
5275
5276/**
5277 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5278 * @wq: the workqueue to unregister
5279 *
5280 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5281 */
5282static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5283{
5284 struct wq_device *wq_dev = wq->wq_dev;
5285
5286 if (!wq->wq_dev)
5287 return;
5288
5289 wq->wq_dev = NULL;
5290 device_unregister(&wq_dev->dev);
5291}
5292#else /* CONFIG_SYSFS */
5293static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5294#endif /* CONFIG_SYSFS */
5295
Tejun Heo82607adc2015-12-08 11:28:04 -05005296/*
5297 * Workqueue watchdog.
5298 *
5299 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5300 * flush dependency, a concurrency managed work item which stays RUNNING
5301 * indefinitely. Workqueue stalls can be very difficult to debug as the
5302 * usual warning mechanisms don't trigger and internal workqueue state is
5303 * largely opaque.
5304 *
5305 * Workqueue watchdog monitors all worker pools periodically and dumps
5306 * state if some pools failed to make forward progress for a while where
5307 * forward progress is defined as the first item on ->worklist changing.
5308 *
5309 * This mechanism is controlled through the kernel parameter
5310 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5311 * corresponding sysfs parameter file.
5312 */
5313#ifdef CONFIG_WQ_WATCHDOG
5314
5315static void wq_watchdog_timer_fn(unsigned long data);
5316
5317static unsigned long wq_watchdog_thresh = 30;
5318static struct timer_list wq_watchdog_timer =
5319 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5320
5321static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5322static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5323
5324static void wq_watchdog_reset_touched(void)
5325{
5326 int cpu;
5327
5328 wq_watchdog_touched = jiffies;
5329 for_each_possible_cpu(cpu)
5330 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5331}
5332
5333static void wq_watchdog_timer_fn(unsigned long data)
5334{
5335 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5336 bool lockup_detected = false;
5337 struct worker_pool *pool;
5338 int pi;
5339
5340 if (!thresh)
5341 return;
5342
5343 rcu_read_lock();
5344
5345 for_each_pool(pool, pi) {
5346 unsigned long pool_ts, touched, ts;
5347
5348 if (list_empty(&pool->worklist))
5349 continue;
5350
5351 /* get the latest of pool and touched timestamps */
5352 pool_ts = READ_ONCE(pool->watchdog_ts);
5353 touched = READ_ONCE(wq_watchdog_touched);
5354
5355 if (time_after(pool_ts, touched))
5356 ts = pool_ts;
5357 else
5358 ts = touched;
5359
5360 if (pool->cpu >= 0) {
5361 unsigned long cpu_touched =
5362 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5363 pool->cpu));
5364 if (time_after(cpu_touched, ts))
5365 ts = cpu_touched;
5366 }
5367
5368 /* did we stall? */
5369 if (time_after(jiffies, ts + thresh)) {
5370 lockup_detected = true;
5371 pr_emerg("BUG: workqueue lockup - pool");
5372 pr_cont_pool_info(pool);
5373 pr_cont(" stuck for %us!\n",
5374 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5375 }
5376 }
5377
5378 rcu_read_unlock();
5379
5380 if (lockup_detected)
5381 show_workqueue_state();
5382
5383 wq_watchdog_reset_touched();
5384 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5385}
5386
5387void wq_watchdog_touch(int cpu)
5388{
5389 if (cpu >= 0)
5390 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5391 else
5392 wq_watchdog_touched = jiffies;
5393}
5394
5395static void wq_watchdog_set_thresh(unsigned long thresh)
5396{
5397 wq_watchdog_thresh = 0;
5398 del_timer_sync(&wq_watchdog_timer);
5399
5400 if (thresh) {
5401 wq_watchdog_thresh = thresh;
5402 wq_watchdog_reset_touched();
5403 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5404 }
5405}
5406
5407static int wq_watchdog_param_set_thresh(const char *val,
5408 const struct kernel_param *kp)
5409{
5410 unsigned long thresh;
5411 int ret;
5412
5413 ret = kstrtoul(val, 0, &thresh);
5414 if (ret)
5415 return ret;
5416
5417 if (system_wq)
5418 wq_watchdog_set_thresh(thresh);
5419 else
5420 wq_watchdog_thresh = thresh;
5421
5422 return 0;
5423}
5424
5425static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5426 .set = wq_watchdog_param_set_thresh,
5427 .get = param_get_ulong,
5428};
5429
5430module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5431 0644);
5432
5433static void wq_watchdog_init(void)
5434{
5435 wq_watchdog_set_thresh(wq_watchdog_thresh);
5436}
5437
5438#else /* CONFIG_WQ_WATCHDOG */
5439
5440static inline void wq_watchdog_init(void) { }
5441
5442#endif /* CONFIG_WQ_WATCHDOG */
5443
Tejun Heobce90382013-04-01 11:23:32 -07005444static void __init wq_numa_init(void)
5445{
5446 cpumask_var_t *tbl;
5447 int node, cpu;
5448
Tejun Heobce90382013-04-01 11:23:32 -07005449 if (num_possible_nodes() <= 1)
5450 return;
5451
Tejun Heod55262c2013-04-01 11:23:38 -07005452 if (wq_disable_numa) {
5453 pr_info("workqueue: NUMA affinity support disabled\n");
5454 return;
5455 }
5456
Tejun Heo4c16bd32013-04-01 11:23:36 -07005457 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5458 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5459
Tejun Heobce90382013-04-01 11:23:32 -07005460 /*
5461 * We want masks of possible CPUs of each node which isn't readily
5462 * available. Build one from cpu_to_node() which should have been
5463 * fully initialized by now.
5464 */
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08005465 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
Tejun Heobce90382013-04-01 11:23:32 -07005466 BUG_ON(!tbl);
5467
5468 for_each_node(node)
Yasuaki Ishimatsu5a6024f2014-07-07 09:56:48 -04005469 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
Tejun Heo1be0c252013-05-15 14:24:24 -07005470 node_online(node) ? node : NUMA_NO_NODE));
Tejun Heobce90382013-04-01 11:23:32 -07005471
5472 for_each_possible_cpu(cpu) {
5473 node = cpu_to_node(cpu);
5474 if (WARN_ON(node == NUMA_NO_NODE)) {
5475 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5476 /* happens iff arch is bonkers, let's just proceed */
5477 return;
5478 }
5479 cpumask_set_cpu(cpu, tbl[node]);
5480 }
5481
5482 wq_numa_possible_cpumask = tbl;
5483 wq_numa_enabled = true;
5484}
5485
Tejun Heo3347fa02016-09-16 15:49:32 -04005486/**
5487 * workqueue_init_early - early init for workqueue subsystem
5488 *
5489 * This is the first half of two-staged workqueue subsystem initialization
5490 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5491 * idr are up. It sets up all the data structures and system workqueues
5492 * and allows early boot code to create workqueues and queue/cancel work
5493 * items. Actual work item execution starts only after kthreads can be
5494 * created and scheduled right before early initcalls.
5495 */
5496int __init workqueue_init_early(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005497{
Tejun Heo7a4e3442013-03-12 11:30:00 -07005498 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5499 int i, cpu;
Tejun Heoc34056a2010-06-29 10:07:11 +02005500
Tejun Heoe904e6c2013-03-12 11:29:57 -07005501 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5502
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005503 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5504 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5505
Tejun Heoe904e6c2013-03-12 11:29:57 -07005506 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5507
Tejun Heo706026c2013-01-24 11:01:34 -08005508 /* initialize CPU pools */
Tejun Heo29c91e92013-03-12 11:30:03 -07005509 for_each_possible_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005510 struct worker_pool *pool;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005511
Tejun Heo7a4e3442013-03-12 11:30:00 -07005512 i = 0;
Tejun Heof02ae732013-03-12 11:30:03 -07005513 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo7a4e3442013-03-12 11:30:00 -07005514 BUG_ON(init_worker_pool(pool));
Tejun Heoec22ca52013-01-24 11:01:33 -08005515 pool->cpu = cpu;
Tejun Heo29c91e92013-03-12 11:30:03 -07005516 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
Tejun Heo7a4e3442013-03-12 11:30:00 -07005517 pool->attrs->nice = std_nice[i++];
Tejun Heof3f90ad2013-04-01 11:23:34 -07005518 pool->node = cpu_to_node(cpu);
Tejun Heo7a4e3442013-03-12 11:30:00 -07005519
Tejun Heo9daf9e62013-01-24 11:01:33 -08005520 /* alloc pool ID */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005521 mutex_lock(&wq_pool_mutex);
Tejun Heo9daf9e62013-01-24 11:01:33 -08005522 BUG_ON(worker_pool_assign_id(pool));
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005523 mutex_unlock(&wq_pool_mutex);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005524 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02005525 }
5526
Tejun Heo8a2b7532013-09-05 12:30:04 -04005527 /* create default unbound and ordered wq attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -07005528 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5529 struct workqueue_attrs *attrs;
5530
5531 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
Tejun Heo29c91e92013-03-12 11:30:03 -07005532 attrs->nice = std_nice[i];
Tejun Heo29c91e92013-03-12 11:30:03 -07005533 unbound_std_wq_attrs[i] = attrs;
Tejun Heo8a2b7532013-09-05 12:30:04 -04005534
5535 /*
5536 * An ordered wq should have only one pwq as ordering is
5537 * guaranteed by max_active which is enforced by pwqs.
5538 * Turn off NUMA so that dfl_pwq is used for all nodes.
5539 */
5540 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5541 attrs->nice = std_nice[i];
5542 attrs->no_numa = true;
5543 ordered_wq_attrs[i] = attrs;
Tejun Heo29c91e92013-03-12 11:30:03 -07005544 }
5545
Tejun Heod320c032010-06-29 10:07:14 +02005546 system_wq = alloc_workqueue("events", 0, 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005547 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
Tejun Heod320c032010-06-29 10:07:14 +02005548 system_long_wq = alloc_workqueue("events_long", 0, 0);
Tejun Heof3421792010-07-02 10:03:51 +02005549 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5550 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01005551 system_freezable_wq = alloc_workqueue("events_freezable",
5552 WQ_FREEZABLE, 0);
Viresh Kumar06681062013-04-24 17:12:54 +05305553 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5554 WQ_POWER_EFFICIENT, 0);
5555 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5556 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5557 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005558 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
Viresh Kumar06681062013-04-24 17:12:54 +05305559 !system_unbound_wq || !system_freezable_wq ||
5560 !system_power_efficient_wq ||
5561 !system_freezable_power_efficient_wq);
Tejun Heo82607adc2015-12-08 11:28:04 -05005562
Tejun Heo3347fa02016-09-16 15:49:32 -04005563 return 0;
5564}
5565
5566/**
5567 * workqueue_init - bring workqueue subsystem fully online
5568 *
5569 * This is the latter half of two-staged workqueue subsystem initialization
5570 * and invoked as soon as kthreads can be created and scheduled.
5571 * Workqueues have been created and work items queued on them, but there
5572 * are no kworkers executing the work items yet. Populate the worker pools
5573 * with the initial workers and enable future kworker creations.
5574 */
5575int __init workqueue_init(void)
5576{
Tejun Heo2186d9f2016-10-19 12:01:27 -04005577 struct workqueue_struct *wq;
Tejun Heo3347fa02016-09-16 15:49:32 -04005578 struct worker_pool *pool;
5579 int cpu, bkt;
5580
Tejun Heo2186d9f2016-10-19 12:01:27 -04005581 /*
5582 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5583 * CPU to node mapping may not be available that early on some
5584 * archs such as power and arm64. As per-cpu pools created
5585 * previously could be missing node hint and unbound pools NUMA
5586 * affinity, fix them up.
5587 */
5588 wq_numa_init();
5589
5590 mutex_lock(&wq_pool_mutex);
5591
5592 for_each_possible_cpu(cpu) {
5593 for_each_cpu_worker_pool(pool, cpu) {
5594 pool->node = cpu_to_node(cpu);
5595 }
5596 }
5597
5598 list_for_each_entry(wq, &workqueues, list)
5599 wq_update_unbound_numa(wq, smp_processor_id(), true);
5600
5601 mutex_unlock(&wq_pool_mutex);
5602
Tejun Heo3347fa02016-09-16 15:49:32 -04005603 /* create the initial workers */
5604 for_each_online_cpu(cpu) {
5605 for_each_cpu_worker_pool(pool, cpu) {
5606 pool->flags &= ~POOL_DISASSOCIATED;
5607 BUG_ON(!create_worker(pool));
5608 }
5609 }
5610
5611 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5612 BUG_ON(!create_worker(pool));
5613
5614 wq_online = true;
Tejun Heo82607adc2015-12-08 11:28:04 -05005615 wq_watchdog_init();
5616
Suresh Siddha6ee05782010-07-30 14:57:37 -07005617 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005618}