blob: 4114412649b850ba89f687bf3c44f46fb153a127 [file] [log] [blame]
Thomas Gleixner457c8992019-05-19 13:08:55 +01001// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
Tejun Heoc54fce62010-09-10 16:51:36 +02003 * kernel/workqueue.c - generic async execution with shared worker pool
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
Tejun Heoc54fce62010-09-10 16:51:36 +02005 * Copyright (C) 2002 Ingo Molnar
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
Tejun Heoc54fce62010-09-10 16:51:36 +02007 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
9 * Andrew Morton
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
Christoph Lameter89ada672005-10-30 15:01:59 -080012 *
Christoph Lametercde53532008-07-04 09:59:22 -070013 * Made to use alloc_percpu by Christoph Lameter.
Tejun Heoc54fce62010-09-10 16:51:36 +020014 *
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 *
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
Libinb11895c2013-08-21 08:50:39 +080020 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
Tejun Heoc54fce62010-09-10 16:51:36 +020024 *
Benjamin Peterson9a261492017-08-06 19:33:22 -070025 * Please read Documentation/core-api/workqueue.rst for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -070026 */
27
Paul Gortmaker9984de12011-05-23 14:51:41 -040028#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/kernel.h>
30#include <linux/sched.h>
31#include <linux/init.h>
32#include <linux/signal.h>
33#include <linux/completion.h>
34#include <linux/workqueue.h>
35#include <linux/slab.h>
36#include <linux/cpu.h>
37#include <linux/notifier.h>
38#include <linux/kthread.h>
James Bottomley1fa44ec2006-02-23 12:43:43 -060039#include <linux/hardirq.h>
Christoph Lameter469340232006-10-11 01:21:26 -070040#include <linux/mempolicy.h>
Rafael J. Wysocki341a5952006-12-06 20:34:49 -080041#include <linux/freezer.h>
Peter Zijlstrad5abe662006-12-06 20:37:26 -080042#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>
Tal Shorerc98a9802017-11-03 17:27:50 +020051#include <linux/sched/isolation.h>
Sergey Senozhatsky62635ea2018-01-11 09:53:35 +090052#include <linux/nmi.h>
Tejun Heoe22bee72010-06-29 10:07:14 +020053
Tejun Heoea138442013-01-18 14:05:55 -080054#include "workqueue_internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
Tejun Heoc8e55f32010-06-29 10:07:12 +020056enum {
Tejun Heobc2ae0f2012-07-17 12:39:27 -070057 /*
Tejun Heo24647572013-01-24 11:01:33 -080058 * worker_pool flags
Tejun Heobc2ae0f2012-07-17 12:39:27 -070059 *
Tejun Heo24647572013-01-24 11:01:33 -080060 * A bound pool is either associated or disassociated with its CPU.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070061 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * is in effect.
64 *
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
Tejun Heo24647572013-01-24 11:01:33 -080067 * be executing on any CPU. The pool behaves as an unbound one.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070068 *
Tejun Heobc3a1af2013-03-13 19:47:39 -070069 * Note that DISASSOCIATED should be flipped only while holding
Tejun Heo1258fae2018-05-18 08:47:13 -070070 * wq_pool_attach_mutex to avoid changing binding state while
Lai Jiangshan4736cbf2014-05-20 17:46:35 +080071 * worker_attach_to_pool() is in progress.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070072 */
Tejun Heo692b4822017-10-09 08:04:13 -070073 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
Tejun Heo24647572013-01-24 11:01:33 -080074 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020075
Tejun Heoc8e55f32010-06-29 10:07:12 +020076 /* worker flags */
Tejun Heoc8e55f32010-06-29 10:07:12 +020077 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
Tejun Heoe22bee72010-06-29 10:07:14 +020079 WORKER_PREP = 1 << 3, /* preparing to run works */
Tejun Heofb0e7be2010-06-29 10:07:15 +020080 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
Tejun Heof3421792010-07-02 10:03:51 +020081 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
Tejun Heoa9ab7752013-03-19 13:45:21 -070082 WORKER_REBOUND = 1 << 8, /* worker was rebound */
Tejun Heoe22bee72010-06-29 10:07:14 +020083
Tejun Heoa9ab7752013-03-19 13:45:21 -070084 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
Tejun Heodb7bccf2010-06-29 10:07:12 +020086
Tejun Heoe34cdddb2013-01-24 11:01:33 -080087 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
Tejun Heo4ce62e92012-07-13 22:16:44 -070088
Tejun Heo29c91e92013-03-12 11:30:03 -070089 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
Tejun Heoc8e55f32010-06-29 10:07:12 +020090 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020091
Tejun Heoe22bee72010-06-29 10:07:14 +020092 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94
Tejun Heo3233cdb2011-02-16 18:10:19 +010095 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
97 (min two ticks) */
Tejun Heoe22bee72010-06-29 10:07:14 +020098 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
Tejun Heoe22bee72010-06-29 10:07:14 +0200100
101 /*
102 * Rescue workers are used only on emergencies and shared by
Dongsheng Yang8698a742014-03-11 18:09:12 +0800103 * all cpus. Give MIN_NICE.
Tejun Heoe22bee72010-06-29 10:07:14 +0200104 */
Dongsheng Yang8698a742014-03-11 18:09:12 +0800105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
Tejun Heoecf68812013-04-01 11:23:34 -0700107
108 WQ_NAME_LEN = 24,
Tejun Heoc8e55f32010-06-29 10:07:12 +0200109};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111/*
Tejun Heo4690c4a2010-06-29 10:07:10 +0200112 * Structure fields follow one of the following exclusion rules.
113 *
Tejun Heoe41e7042010-08-24 14:22:47 +0200114 * I: Modifiable by initialization/destruction paths and read-only for
115 * everyone else.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200116 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
119 *
Tejun Heod565ed62013-01-24 11:01:33 -0800120 * L: pool->lock protected. Access with pool->lock held.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200121 *
Tejun Heod565ed62013-01-24 11:01:33 -0800122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
Tejun Heoe22bee72010-06-29 10:07:14 +0200126 *
Tejun Heo1258fae2018-05-18 08:47:13 -0700127 * A: wq_pool_attach_mutex protected.
Tejun Heo822d8402013-03-19 13:45:21 -0700128 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700129 * PL: wq_pool_mutex protected.
Tejun Heo76af4d92013-03-12 11:30:00 -0700130 *
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100131 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
Tejun Heo5bcab332013-03-13 19:47:40 -0700132 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
134 *
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100136 * RCU for reads.
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800137 *
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700138 * WQ: wq->mutex protected.
139 *
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100140 * WR: wq->mutex protected for writes. RCU protected for reads.
Tejun Heo2e109a22013-03-13 19:47:40 -0700141 *
142 * MD: wq_mayday_lock protected.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200143 */
144
Tejun Heo2eaebdb32013-01-18 14:05:55 -0800145/* struct worker is defined in workqueue_internal.h */
Tejun Heoc34056a2010-06-29 10:07:11 +0200146
Tejun Heobd7bdd42012-07-12 14:46:37 -0700147struct worker_pool {
Tejun Heod565ed62013-01-24 11:01:33 -0800148 spinlock_t lock; /* the pool lock */
Tejun Heod84ff052013-03-12 11:29:59 -0700149 int cpu; /* I: the associated cpu */
Tejun Heof3f90ad2013-04-01 11:23:34 -0700150 int node; /* I: the associated node ID */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800151 int id; /* I: pool ID */
Tejun Heo11ebea52012-07-12 14:46:37 -0700152 unsigned int flags; /* X: flags */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700153
Tejun Heo82607adc2015-12-08 11:28:04 -0500154 unsigned long watchdog_ts; /* L: watchdog timestamp */
155
Tejun Heobd7bdd42012-07-12 14:46:37 -0700156 struct list_head worklist; /* L: list of pending works */
Lai Jiangshanea1abd62012-09-18 09:59:22 -0700157
Lai Jiangshan5826cc82018-03-20 17:24:05 +0800158 int nr_workers; /* L: total number of workers */
159 int nr_idle; /* L: currently idle workers */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700160
161 struct list_head idle_list; /* X: list of idle workers */
162 struct timer_list idle_timer; /* L: worker idle timeout */
163 struct timer_list mayday_timer; /* L: SOS timer for workers */
164
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700165 /* a workers is either on busy_hash or idle_list, or the manager */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800166 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
167 /* L: hash of busy workers */
168
Tejun Heo2607d7a2015-03-09 09:22:28 -0400169 struct worker *manager; /* L: purely informational */
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800170 struct list_head workers; /* A: attached workers */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +0800171 struct completion *detach_completion; /* all workers detached */
Tejun Heoe19e3972013-01-24 11:39:44 -0800172
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +0800173 struct ida worker_ida; /* worker IDs for task name */
Tejun Heoe19e3972013-01-24 11:39:44 -0800174
Tejun Heo7a4e3442013-03-12 11:30:00 -0700175 struct workqueue_attrs *attrs; /* I: worker attributes */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700176 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
177 int refcnt; /* PL: refcnt for unbound pools */
Tejun Heo7a4e3442013-03-12 11:30:00 -0700178
Tejun Heoe19e3972013-01-24 11:39:44 -0800179 /*
180 * The current concurrency level. As it's likely to be accessed
181 * from other CPUs during try_to_wake_up(), put it in a separate
182 * cacheline.
183 */
184 atomic_t nr_running ____cacheline_aligned_in_smp;
Tejun Heo29c91e92013-03-12 11:30:03 -0700185
186 /*
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100187 * Destruction of pool is RCU protected to allow dereferences
Tejun Heo29c91e92013-03-12 11:30:03 -0700188 * from get_work_pool().
189 */
190 struct rcu_head rcu;
Tejun Heo8b03ae32010-06-29 10:07:12 +0200191} ____cacheline_aligned_in_smp;
192
193/*
Tejun Heo112202d2013-02-13 19:29:12 -0800194 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
195 * of work_struct->data are used for flags and the remaining high bits
196 * point to the pwq; thus, pwqs need to be aligned at two's power of the
197 * number of flag bits.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 */
Tejun Heo112202d2013-02-13 19:29:12 -0800199struct pool_workqueue {
Tejun Heobd7bdd42012-07-12 14:46:37 -0700200 struct worker_pool *pool; /* I: the associated pool */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200201 struct workqueue_struct *wq; /* I: the owning workqueue */
Tejun Heo73f53c42010-06-29 10:07:11 +0200202 int work_color; /* L: current color */
203 int flush_color; /* L: flushing color */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700204 int refcnt; /* L: reference count */
Tejun Heo73f53c42010-06-29 10:07:11 +0200205 int nr_in_flight[WORK_NR_COLORS];
206 /* L: nr of in_flight works */
Tejun Heo1e19ffc62010-06-29 10:07:12 +0200207 int nr_active; /* L: nr of active works */
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200208 int max_active; /* L: max active works */
Tejun Heo1e19ffc62010-06-29 10:07:12 +0200209 struct list_head delayed_works; /* L: delayed works */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700210 struct list_head pwqs_node; /* WR: node on wq->pwqs */
Tejun Heo2e109a22013-03-13 19:47:40 -0700211 struct list_head mayday_node; /* MD: node on wq->maydays */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700212
213 /*
214 * Release of unbound pwq is punted to system_wq. See put_pwq()
215 * and pwq_unbound_release_workfn() for details. pool_workqueue
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100216 * itself is also RCU protected so that the first pwq can be
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700217 * determined without grabbing wq->mutex.
Tejun Heo8864b4e2013-03-12 11:30:04 -0700218 */
219 struct work_struct unbound_release_work;
220 struct rcu_head rcu;
Tejun Heoe904e6c2013-03-12 11:29:57 -0700221} __aligned(1 << WORK_STRUCT_FLAG_BITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223/*
Tejun Heo73f53c42010-06-29 10:07:11 +0200224 * Structure used to wait for workqueue flush.
225 */
226struct wq_flusher {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700227 struct list_head list; /* WQ: list of flushers */
228 int flush_color; /* WQ: flush color waiting for */
Tejun Heo73f53c42010-06-29 10:07:11 +0200229 struct completion done; /* flush completion */
230};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
Tejun Heo226223a2013-03-12 11:30:05 -0700232struct wq_device;
233
Tejun Heo73f53c42010-06-29 10:07:11 +0200234/*
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700235 * The externally visible workqueue. It relays the issued work items to
236 * the appropriate worker_pool through its pool_workqueues.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 */
238struct workqueue_struct {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700239 struct list_head pwqs; /* WR: all pwqs of this wq */
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400240 struct list_head list; /* PR: list of all workqueues */
Tejun Heo73f53c42010-06-29 10:07:11 +0200241
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700242 struct mutex mutex; /* protects this wq */
243 int work_color; /* WQ: current work color */
244 int flush_color; /* WQ: current flush color */
Tejun Heo112202d2013-02-13 19:29:12 -0800245 atomic_t nr_pwqs_to_flush; /* flush in progress */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700246 struct wq_flusher *first_flusher; /* WQ: first flusher */
247 struct list_head flusher_queue; /* WQ: flush waiters */
248 struct list_head flusher_overflow; /* WQ: flush overflow list */
Tejun Heo73f53c42010-06-29 10:07:11 +0200249
Tejun Heo2e109a22013-03-13 19:47:40 -0700250 struct list_head maydays; /* MD: pwqs requesting rescue */
Tejun Heo30ae2fc2019-09-20 14:09:14 -0700251 struct worker *rescuer; /* MD: rescue worker */
Tejun Heoe22bee72010-06-29 10:07:14 +0200252
Lai Jiangshan87fc7412013-03-25 16:57:18 -0700253 int nr_drainers; /* WQ: drain in progress */
Lai Jiangshana357fc02013-03-25 16:57:19 -0700254 int saved_max_active; /* WQ: saved pwq max_active */
Tejun Heo226223a2013-03-12 11:30:05 -0700255
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800256 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
257 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
Tejun Heo6029a912013-04-01 11:23:34 -0700258
Tejun Heo226223a2013-03-12 11:30:05 -0700259#ifdef CONFIG_SYSFS
260 struct wq_device *wq_dev; /* I: for sysfs interface */
261#endif
Johannes Berg4e6045f2007-10-18 23:39:55 -0700262#ifdef CONFIG_LOCKDEP
Bart Van Assche669de8b2019-02-14 15:00:54 -0800263 char *lock_name;
264 struct lock_class_key key;
Tejun Heo4690c4a2010-06-29 10:07:10 +0200265 struct lockdep_map lockdep_map;
Johannes Berg4e6045f2007-10-18 23:39:55 -0700266#endif
Tejun Heoecf68812013-04-01 11:23:34 -0700267 char name[WQ_NAME_LEN]; /* I: workqueue name */
Tejun Heo2728fd22013-04-01 11:23:35 -0700268
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400269 /*
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100270 * Destruction of workqueue_struct is RCU protected to allow walking
271 * the workqueues list without grabbing wq_pool_mutex.
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400272 * This is used to dump all workqueues from sysrq.
273 */
274 struct rcu_head rcu;
275
Tejun Heo2728fd22013-04-01 11:23:35 -0700276 /* hot fields used during command issue, aligned to cacheline */
277 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
278 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800279 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280};
281
Tejun Heoe904e6c2013-03-12 11:29:57 -0700282static struct kmem_cache *pwq_cache;
283
Tejun Heobce90382013-04-01 11:23:32 -0700284static cpumask_var_t *wq_numa_possible_cpumask;
285 /* possible CPUs of each node */
286
Tejun Heod55262c2013-04-01 11:23:38 -0700287static bool wq_disable_numa;
288module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289
Viresh Kumarcee22a12013-04-08 16:45:40 +0530290/* see the comment above the definition of WQ_POWER_EFFICIENT */
Luis R. Rodriguez552f5302015-05-27 11:09:39 +0930291static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
Viresh Kumarcee22a12013-04-08 16:45:40 +0530292module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293
Tejun Heo863b7102016-09-16 15:49:34 -0400294static bool wq_online; /* can kworkers be created yet? */
Tejun Heo3347fa02016-09-16 15:49:32 -0400295
Tejun Heobce90382013-04-01 11:23:32 -0700296static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297
Tejun Heo4c16bd32013-04-01 11:23:36 -0700298/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700301static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
Tejun Heo1258fae2018-05-18 08:47:13 -0700302static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
Tejun Heo2e109a22013-03-13 19:47:40 -0700303static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
Tejun Heo692b4822017-10-09 08:04:13 -0700304static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
Tejun Heo5bcab332013-03-13 19:47:40 -0700305
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400306static LIST_HEAD(workqueues); /* PR: list of all workqueues */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700307static bool workqueue_freezing; /* PL: have wqs started freezing? */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700308
Mike Galbraithef5571802016-02-09 17:59:38 -0500309/* PL: allowable cpus for unbound wqs and work items */
310static cpumask_var_t wq_unbound_cpumask;
311
312/* CPU where unbound work was last round robin scheduled from this CPU */
313static DEFINE_PER_CPU(int, wq_rr_cpu_last);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +0800314
Tejun Heof303fccb2016-02-09 17:59:38 -0500315/*
316 * Local execution of unbound work items is no longer guaranteed. The
317 * following always forces round-robin CPU selection on unbound work items
318 * to uncover usages which depend on it.
319 */
320#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
321static bool wq_debug_force_rr_cpu = true;
322#else
323static bool wq_debug_force_rr_cpu = false;
324#endif
325module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
326
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700327/* the per-cpu worker pools */
Peter Zijlstra25528212016-03-15 14:52:49 -0700328static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700329
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700330static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700331
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700332/* PL: hash of all unbound pools keyed by pool->attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -0700333static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
334
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700335/* I: attributes used when instantiating standard unbound pools on demand */
Tejun Heo29c91e92013-03-12 11:30:03 -0700336static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
337
Tejun Heo8a2b7532013-09-05 12:30:04 -0400338/* I: attributes used when instantiating ordered pools on demand */
339static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
340
Tejun Heod320c032010-06-29 10:07:14 +0200341struct workqueue_struct *system_wq __read_mostly;
Marc Dionnead7b1f82013-05-06 17:44:55 -0400342EXPORT_SYMBOL(system_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300343struct workqueue_struct *system_highpri_wq __read_mostly;
Joonsoo Kim1aabe902012-08-15 23:25:39 +0900344EXPORT_SYMBOL_GPL(system_highpri_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300345struct workqueue_struct *system_long_wq __read_mostly;
Tejun Heod320c032010-06-29 10:07:14 +0200346EXPORT_SYMBOL_GPL(system_long_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300347struct workqueue_struct *system_unbound_wq __read_mostly;
Tejun Heof3421792010-07-02 10:03:51 +0200348EXPORT_SYMBOL_GPL(system_unbound_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300349struct workqueue_struct *system_freezable_wq __read_mostly;
Tejun Heo24d51ad2011-02-21 09:52:50 +0100350EXPORT_SYMBOL_GPL(system_freezable_wq);
Viresh Kumar06681062013-04-24 17:12:54 +0530351struct workqueue_struct *system_power_efficient_wq __read_mostly;
352EXPORT_SYMBOL_GPL(system_power_efficient_wq);
353struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
354EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
Tejun Heod320c032010-06-29 10:07:14 +0200355
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700356static int worker_thread(void *__worker);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +0800357static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
Tejun Heoc29eb852019-09-23 11:08:58 -0700358static void show_pwq(struct pool_workqueue *pwq);
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700359
Tejun Heo97bd2342010-10-05 10:41:14 +0200360#define CREATE_TRACE_POINTS
361#include <trace/events/workqueue.h>
362
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700363#define assert_rcu_or_pool_mutex() \
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100364 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700365 !lockdep_is_held(&wq_pool_mutex), \
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100366 "RCU or wq_pool_mutex should be held")
Tejun Heo5bcab332013-03-13 19:47:40 -0700367
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800368#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100369 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700370 !lockdep_is_held(&wq->mutex) && \
371 !lockdep_is_held(&wq_pool_mutex), \
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100372 "RCU, wq->mutex or wq_pool_mutex should be held")
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800373
Tejun Heof02ae732013-03-12 11:30:03 -0700374#define for_each_cpu_worker_pool(pool, cpu) \
375 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
376 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
Tejun Heo7a62c2c2013-03-12 11:30:03 -0700377 (pool)++)
Tejun Heo4ce62e92012-07-13 22:16:44 -0700378
Tejun Heo49e3cf42013-03-12 11:29:58 -0700379/**
Tejun Heo17116962013-03-12 11:29:58 -0700380 * for_each_pool - iterate through all worker_pools in the system
381 * @pool: iteration cursor
Tejun Heo611c92a2013-03-13 16:51:36 -0700382 * @pi: integer used for iteration
Tejun Heofa1b54e2013-03-12 11:30:00 -0700383 *
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100384 * This must be called either with wq_pool_mutex held or RCU read
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700385 * locked. If the pool needs to be used beyond the locking in effect, the
386 * caller is responsible for guaranteeing that the pool stays online.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700387 *
388 * The if/else clause exists only for the lockdep assertion and can be
389 * ignored.
Tejun Heo17116962013-03-12 11:29:58 -0700390 */
Tejun Heo611c92a2013-03-13 16:51:36 -0700391#define for_each_pool(pool, pi) \
392 idr_for_each_entry(&worker_pool_idr, pool, pi) \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700393 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
Tejun Heofa1b54e2013-03-12 11:30:00 -0700394 else
Tejun Heo17116962013-03-12 11:29:58 -0700395
396/**
Tejun Heo822d8402013-03-19 13:45:21 -0700397 * for_each_pool_worker - iterate through all workers of a worker_pool
398 * @worker: iteration cursor
Tejun Heo822d8402013-03-19 13:45:21 -0700399 * @pool: worker_pool to iterate workers of
400 *
Tejun Heo1258fae2018-05-18 08:47:13 -0700401 * This must be called with wq_pool_attach_mutex.
Tejun Heo822d8402013-03-19 13:45:21 -0700402 *
403 * The if/else clause exists only for the lockdep assertion and can be
404 * ignored.
405 */
Lai Jiangshanda028462014-05-20 17:46:31 +0800406#define for_each_pool_worker(worker, pool) \
407 list_for_each_entry((worker), &(pool)->workers, node) \
Tejun Heo1258fae2018-05-18 08:47:13 -0700408 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
Tejun Heo822d8402013-03-19 13:45:21 -0700409 else
410
411/**
Tejun Heo49e3cf42013-03-12 11:29:58 -0700412 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
413 * @pwq: iteration cursor
414 * @wq: the target workqueue
Tejun Heo76af4d92013-03-12 11:30:00 -0700415 *
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100416 * This must be called either with wq->mutex held or RCU read locked.
Tejun Heo794b18b2013-03-13 19:47:40 -0700417 * If the pwq needs to be used beyond the locking in effect, the caller is
418 * responsible for guaranteeing that the pwq stays online.
Tejun Heo76af4d92013-03-12 11:30:00 -0700419 *
420 * The if/else clause exists only for the lockdep assertion and can be
421 * ignored.
Tejun Heo49e3cf42013-03-12 11:29:58 -0700422 */
423#define for_each_pwq(pwq, wq) \
Sebastian Andrzej Siewior49e9d1a2019-11-15 19:01:25 +0100424 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
Joel Fernandes (Google)5a644662019-08-15 10:18:42 -0400425 lockdep_is_held(&(wq->mutex)))
Tejun Heof3421792010-07-02 10:03:51 +0200426
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900427#ifdef CONFIG_DEBUG_OBJECTS_WORK
428
429static struct debug_obj_descr work_debug_descr;
430
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100431static void *work_debug_hint(void *addr)
432{
433 return ((struct work_struct *) addr)->func;
434}
435
Du, Changbinb9fdac7f2016-05-19 17:09:41 -0700436static bool work_is_static_object(void *addr)
437{
438 struct work_struct *work = addr;
439
440 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
441}
442
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900443/*
444 * fixup_init is called when:
445 * - an active object is initialized
446 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700447static bool work_fixup_init(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900448{
449 struct work_struct *work = addr;
450
451 switch (state) {
452 case ODEBUG_STATE_ACTIVE:
453 cancel_work_sync(work);
454 debug_object_init(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700455 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900456 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700457 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900458 }
459}
460
461/*
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900462 * fixup_free is called when:
463 * - an active object is freed
464 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700465static bool work_fixup_free(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900466{
467 struct work_struct *work = addr;
468
469 switch (state) {
470 case ODEBUG_STATE_ACTIVE:
471 cancel_work_sync(work);
472 debug_object_free(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700473 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900474 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700475 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900476 }
477}
478
479static struct debug_obj_descr work_debug_descr = {
480 .name = "work_struct",
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100481 .debug_hint = work_debug_hint,
Du, Changbinb9fdac7f2016-05-19 17:09:41 -0700482 .is_static_object = work_is_static_object,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900483 .fixup_init = work_fixup_init,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900484 .fixup_free = work_fixup_free,
485};
486
487static inline void debug_work_activate(struct work_struct *work)
488{
489 debug_object_activate(work, &work_debug_descr);
490}
491
492static inline void debug_work_deactivate(struct work_struct *work)
493{
494 debug_object_deactivate(work, &work_debug_descr);
495}
496
497void __init_work(struct work_struct *work, int onstack)
498{
499 if (onstack)
500 debug_object_init_on_stack(work, &work_debug_descr);
501 else
502 debug_object_init(work, &work_debug_descr);
503}
504EXPORT_SYMBOL_GPL(__init_work);
505
506void destroy_work_on_stack(struct work_struct *work)
507{
508 debug_object_free(work, &work_debug_descr);
509}
510EXPORT_SYMBOL_GPL(destroy_work_on_stack);
511
Thomas Gleixnerea2e64f2014-03-23 14:20:44 +0000512void destroy_delayed_work_on_stack(struct delayed_work *work)
513{
514 destroy_timer_on_stack(&work->timer);
515 debug_object_free(&work->work, &work_debug_descr);
516}
517EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
518
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900519#else
520static inline void debug_work_activate(struct work_struct *work) { }
521static inline void debug_work_deactivate(struct work_struct *work) { }
522#endif
523
Li Bin4e8b22b2013-09-10 09:52:35 +0800524/**
525 * worker_pool_assign_id - allocate ID and assing it to @pool
526 * @pool: the pool pointer of interest
527 *
528 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
529 * successfully, -errno on failure.
530 */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800531static int worker_pool_assign_id(struct worker_pool *pool)
532{
533 int ret;
534
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700535 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo5bcab332013-03-13 19:47:40 -0700536
Li Bin4e8b22b2013-09-10 09:52:35 +0800537 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
538 GFP_KERNEL);
Tejun Heo229641a2013-04-01 17:08:13 -0700539 if (ret >= 0) {
Tejun Heoe68035f2013-03-13 14:59:38 -0700540 pool->id = ret;
Tejun Heo229641a2013-04-01 17:08:13 -0700541 return 0;
542 }
Tejun Heo9daf9e62013-01-24 11:01:33 -0800543 return ret;
544}
545
Tejun Heo76af4d92013-03-12 11:30:00 -0700546/**
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700547 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
548 * @wq: the target workqueue
549 * @node: the node ID
550 *
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100551 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800552 * read locked.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700553 * If the pwq needs to be used beyond the locking in effect, the caller is
554 * responsible for guaranteeing that the pwq stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700555 *
556 * Return: The unbound pool_workqueue for @node.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700557 */
558static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
559 int node)
560{
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800561 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
Tejun Heod6e022f2016-02-03 13:54:25 -0500562
563 /*
564 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
565 * delayed item is pending. The plan is to keep CPU -> NODE
566 * mapping valid and stable across CPU on/offlines. Once that
567 * happens, this workaround can be removed.
568 */
569 if (unlikely(node == NUMA_NO_NODE))
570 return wq->dfl_pwq;
571
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700572 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
573}
574
Tejun Heo73f53c42010-06-29 10:07:11 +0200575static unsigned int work_color_to_flags(int color)
576{
577 return color << WORK_STRUCT_COLOR_SHIFT;
578}
579
580static int get_work_color(struct work_struct *work)
581{
582 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
583 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
584}
585
586static int work_next_color(int color)
587{
588 return (color + 1) % WORK_NR_COLORS;
Oleg Nesterova848e3b2007-05-09 02:34:17 -0700589}
590
David Howells4594bf12006-12-07 11:33:26 +0000591/*
Tejun Heo112202d2013-02-13 19:29:12 -0800592 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
593 * contain the pointer to the queued pwq. Once execution starts, the flag
Tejun Heo7c3eed52013-01-24 11:01:33 -0800594 * is cleared and the high bits contain OFFQ flags and pool ID.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200595 *
Tejun Heo112202d2013-02-13 19:29:12 -0800596 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
597 * and clear_work_data() can be used to set the pwq, pool or clear
Tejun Heobbb68df2012-08-03 10:30:46 -0700598 * work->data. These functions should only be called while the work is
599 * owned - ie. while the PENDING bit is set.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200600 *
Tejun Heo112202d2013-02-13 19:29:12 -0800601 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
Tejun Heo7c3eed52013-01-24 11:01:33 -0800602 * corresponding to a work. Pool is available once the work has been
Tejun Heo112202d2013-02-13 19:29:12 -0800603 * queued anywhere after initialization until it is sync canceled. pwq is
Tejun Heo7c3eed52013-01-24 11:01:33 -0800604 * available only while the work item is queued.
Tejun Heobbb68df2012-08-03 10:30:46 -0700605 *
606 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
607 * canceled. While being canceled, a work item may have its PENDING set
608 * but stay off timer and worklist for arbitrarily long and nobody should
609 * try to steal the PENDING bit.
David Howells4594bf12006-12-07 11:33:26 +0000610 */
Tejun Heo7a22ad72010-06-29 10:07:13 +0200611static inline void set_work_data(struct work_struct *work, unsigned long data,
612 unsigned long flags)
David Howells365970a2006-11-22 14:54:49 +0000613{
Tejun Heo6183c002013-03-12 11:29:57 -0700614 WARN_ON_ONCE(!work_pending(work));
Tejun Heo7a22ad72010-06-29 10:07:13 +0200615 atomic_long_set(&work->data, data | flags | work_static(work));
David Howells365970a2006-11-22 14:54:49 +0000616}
David Howells365970a2006-11-22 14:54:49 +0000617
Tejun Heo112202d2013-02-13 19:29:12 -0800618static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
Tejun Heo7a22ad72010-06-29 10:07:13 +0200619 unsigned long extra_flags)
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200620{
Tejun Heo112202d2013-02-13 19:29:12 -0800621 set_work_data(work, (unsigned long)pwq,
622 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200623}
624
Lai Jiangshan4468a002013-02-06 18:04:53 -0800625static void set_work_pool_and_keep_pending(struct work_struct *work,
626 int pool_id)
627{
628 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
629 WORK_STRUCT_PENDING);
630}
631
Tejun Heo7c3eed52013-01-24 11:01:33 -0800632static void set_work_pool_and_clear_pending(struct work_struct *work,
633 int pool_id)
David Howells365970a2006-11-22 14:54:49 +0000634{
Tejun Heo23657bb2012-08-13 17:08:19 -0700635 /*
636 * The following wmb is paired with the implied mb in
637 * test_and_set_bit(PENDING) and ensures all updates to @work made
638 * here are visible to and precede any updates by the next PENDING
639 * owner.
640 */
641 smp_wmb();
Tejun Heo7c3eed52013-01-24 11:01:33 -0800642 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
Roman Pen346c09f2016-04-26 13:15:35 +0200643 /*
644 * The following mb guarantees that previous clear of a PENDING bit
645 * will not be reordered with any speculative LOADS or STORES from
646 * work->current_func, which is executed afterwards. This possible
Liu Song8bdc6202019-02-19 23:53:27 +0800647 * reordering can lead to a missed execution on attempt to queue
Roman Pen346c09f2016-04-26 13:15:35 +0200648 * the same @work. E.g. consider this case:
649 *
650 * CPU#0 CPU#1
651 * ---------------------------- --------------------------------
652 *
653 * 1 STORE event_indicated
654 * 2 queue_work_on() {
655 * 3 test_and_set_bit(PENDING)
656 * 4 } set_..._and_clear_pending() {
657 * 5 set_work_data() # clear bit
658 * 6 smp_mb()
659 * 7 work->current_func() {
660 * 8 LOAD event_indicated
661 * }
662 *
663 * Without an explicit full barrier speculative LOAD on line 8 can
664 * be executed before CPU#0 does STORE on line 1. If that happens,
665 * CPU#0 observes the PENDING bit is still set and new execution of
666 * a @work is not queued in a hope, that CPU#1 will eventually
667 * finish the queued @work. Meanwhile CPU#1 does not see
668 * event_indicated is set, because speculative LOAD was executed
669 * before actual STORE.
670 */
671 smp_mb();
Tejun Heo7a22ad72010-06-29 10:07:13 +0200672}
673
674static void clear_work_data(struct work_struct *work)
675{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800676 smp_wmb(); /* see set_work_pool_and_clear_pending() */
677 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200678}
679
Tejun Heo112202d2013-02-13 19:29:12 -0800680static struct pool_workqueue *get_work_pwq(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200681{
Tejun Heoe1201532010-07-22 14:14:25 +0200682 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200683
Tejun Heo112202d2013-02-13 19:29:12 -0800684 if (data & WORK_STRUCT_PWQ)
Tejun Heoe1201532010-07-22 14:14:25 +0200685 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
686 else
687 return NULL;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200688}
689
Tejun Heo7c3eed52013-01-24 11:01:33 -0800690/**
691 * get_work_pool - return the worker_pool a given work was associated with
692 * @work: the work item of interest
693 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700694 * Pools are created and destroyed under wq_pool_mutex, and allows read
Thomas Gleixner24acfb72019-03-13 17:55:47 +0100695 * access under RCU read lock. As such, this function should be
696 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700697 *
698 * All fields of the returned pool are accessible as long as the above
699 * mentioned locking is in effect. If the returned pool needs to be used
700 * beyond the critical section, the caller is responsible for ensuring the
701 * returned pool is and stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700702 *
703 * Return: The worker_pool @work was last associated with. %NULL if none.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800704 */
705static struct worker_pool *get_work_pool(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200706{
Tejun Heoe1201532010-07-22 14:14:25 +0200707 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800708 int pool_id;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200709
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700710 assert_rcu_or_pool_mutex();
Tejun Heofa1b54e2013-03-12 11:30:00 -0700711
Tejun Heo112202d2013-02-13 19:29:12 -0800712 if (data & WORK_STRUCT_PWQ)
713 return ((struct pool_workqueue *)
Tejun Heo7c3eed52013-01-24 11:01:33 -0800714 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200715
Tejun Heo7c3eed52013-01-24 11:01:33 -0800716 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
717 if (pool_id == WORK_OFFQ_POOL_NONE)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200718 return NULL;
719
Tejun Heofa1b54e2013-03-12 11:30:00 -0700720 return idr_find(&worker_pool_idr, pool_id);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800721}
722
723/**
724 * get_work_pool_id - return the worker pool ID a given work is associated with
725 * @work: the work item of interest
726 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700727 * Return: The worker_pool ID @work was last associated with.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800728 * %WORK_OFFQ_POOL_NONE if none.
729 */
730static int get_work_pool_id(struct work_struct *work)
731{
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800732 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800733
Tejun Heo112202d2013-02-13 19:29:12 -0800734 if (data & WORK_STRUCT_PWQ)
735 return ((struct pool_workqueue *)
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800736 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
737
738 return data >> WORK_OFFQ_POOL_SHIFT;
Tejun Heo7c3eed52013-01-24 11:01:33 -0800739}
740
Tejun Heobbb68df2012-08-03 10:30:46 -0700741static void mark_work_canceling(struct work_struct *work)
742{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800743 unsigned long pool_id = get_work_pool_id(work);
Tejun Heobbb68df2012-08-03 10:30:46 -0700744
Tejun Heo7c3eed52013-01-24 11:01:33 -0800745 pool_id <<= WORK_OFFQ_POOL_SHIFT;
746 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700747}
748
749static bool work_is_canceling(struct work_struct *work)
750{
751 unsigned long data = atomic_long_read(&work->data);
752
Tejun Heo112202d2013-02-13 19:29:12 -0800753 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700754}
755
David Howells365970a2006-11-22 14:54:49 +0000756/*
Tejun Heo32704762012-07-13 22:16:45 -0700757 * Policy functions. These define the policies on how the global worker
758 * pools are managed. Unless noted otherwise, these functions assume that
Tejun Heod565ed62013-01-24 11:01:33 -0800759 * they're being called with pool->lock held.
David Howells365970a2006-11-22 14:54:49 +0000760 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200761
Tejun Heo63d95a92012-07-12 14:46:37 -0700762static bool __need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000763{
Tejun Heoe19e3972013-01-24 11:39:44 -0800764 return !atomic_read(&pool->nr_running);
David Howells365970a2006-11-22 14:54:49 +0000765}
766
Tejun Heoe22bee72010-06-29 10:07:14 +0200767/*
768 * Need to wake up a worker? Called from anything but currently
769 * running workers.
Tejun Heo974271c2012-07-12 14:46:37 -0700770 *
771 * Note that, because unbound workers never contribute to nr_running, this
Tejun Heo706026c2013-01-24 11:01:34 -0800772 * function will always return %true for unbound pools as long as the
Tejun Heo974271c2012-07-12 14:46:37 -0700773 * worklist isn't empty.
Tejun Heoe22bee72010-06-29 10:07:14 +0200774 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700775static bool need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000776{
Tejun Heo63d95a92012-07-12 14:46:37 -0700777 return !list_empty(&pool->worklist) && __need_more_worker(pool);
David Howells365970a2006-11-22 14:54:49 +0000778}
779
Tejun Heoe22bee72010-06-29 10:07:14 +0200780/* Can I start working? Called from busy but !running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700781static bool may_start_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200782{
Tejun Heo63d95a92012-07-12 14:46:37 -0700783 return pool->nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200784}
785
786/* Do I need to keep working? Called from currently running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700787static bool keep_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200788{
Tejun Heoe19e3972013-01-24 11:39:44 -0800789 return !list_empty(&pool->worklist) &&
790 atomic_read(&pool->nr_running) <= 1;
Tejun Heoe22bee72010-06-29 10:07:14 +0200791}
792
793/* Do we need a new worker? Called from manager. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700794static bool need_to_create_worker(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200795{
Tejun Heo63d95a92012-07-12 14:46:37 -0700796 return need_more_worker(pool) && !may_start_working(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200797}
798
Tejun Heoe22bee72010-06-29 10:07:14 +0200799/* Do we have too many workers and should some go away? */
Tejun Heo63d95a92012-07-12 14:46:37 -0700800static bool too_many_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200801{
Tejun Heo692b4822017-10-09 08:04:13 -0700802 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
Tejun Heo63d95a92012-07-12 14:46:37 -0700803 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
804 int nr_busy = pool->nr_workers - nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200805
806 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
807}
808
809/*
810 * Wake up functions.
811 */
812
Lai Jiangshan1037de32014-05-22 16:44:07 +0800813/* Return the first idle worker. Safe with preemption disabled */
814static struct worker *first_idle_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200815{
Tejun Heo63d95a92012-07-12 14:46:37 -0700816 if (unlikely(list_empty(&pool->idle_list)))
Tejun Heo7e116292010-06-29 10:07:13 +0200817 return NULL;
818
Tejun Heo63d95a92012-07-12 14:46:37 -0700819 return list_first_entry(&pool->idle_list, struct worker, entry);
Tejun Heo7e116292010-06-29 10:07:13 +0200820}
821
822/**
823 * wake_up_worker - wake up an idle worker
Tejun Heo63d95a92012-07-12 14:46:37 -0700824 * @pool: worker pool to wake worker from
Tejun Heo7e116292010-06-29 10:07:13 +0200825 *
Tejun Heo63d95a92012-07-12 14:46:37 -0700826 * Wake up the first idle worker of @pool.
Tejun Heo7e116292010-06-29 10:07:13 +0200827 *
828 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800829 * spin_lock_irq(pool->lock).
Tejun Heo7e116292010-06-29 10:07:13 +0200830 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700831static void wake_up_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200832{
Lai Jiangshan1037de32014-05-22 16:44:07 +0800833 struct worker *worker = first_idle_worker(pool);
Tejun Heo7e116292010-06-29 10:07:13 +0200834
835 if (likely(worker))
836 wake_up_process(worker->task);
837}
838
Tejun Heo4690c4a2010-06-29 10:07:10 +0200839/**
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100840 * wq_worker_running - a worker is running again
Tejun Heoe22bee72010-06-29 10:07:14 +0200841 * @task: task waking up
Tejun Heoe22bee72010-06-29 10:07:14 +0200842 *
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100843 * This function is called when a worker returns from schedule()
Tejun Heoe22bee72010-06-29 10:07:14 +0200844 */
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100845void wq_worker_running(struct task_struct *task)
Tejun Heoe22bee72010-06-29 10:07:14 +0200846{
847 struct worker *worker = kthread_data(task);
848
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100849 if (!worker->sleeping)
850 return;
851 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800852 atomic_inc(&worker->pool->nr_running);
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100853 worker->sleeping = 0;
Tejun Heoe22bee72010-06-29 10:07:14 +0200854}
855
856/**
857 * wq_worker_sleeping - a worker is going to sleep
858 * @task: task going to sleep
Tejun Heoe22bee72010-06-29 10:07:14 +0200859 *
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100860 * This function is called from schedule() when a busy worker is
861 * going to sleep.
Tejun Heoe22bee72010-06-29 10:07:14 +0200862 */
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100863void wq_worker_sleeping(struct task_struct *task)
Tejun Heoe22bee72010-06-29 10:07:14 +0200864{
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100865 struct worker *next, *worker = kthread_data(task);
Tejun Heo111c2252013-01-17 17:16:24 -0800866 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200867
Tejun Heo111c2252013-01-17 17:16:24 -0800868 /*
869 * Rescuers, which may not have all the fields set up like normal
870 * workers, also reach here, let's not access anything before
871 * checking NOT_RUNNING.
872 */
Steven Rostedt2d646722010-12-03 23:12:33 -0500873 if (worker->flags & WORKER_NOT_RUNNING)
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100874 return;
Tejun Heoe22bee72010-06-29 10:07:14 +0200875
Tejun Heo111c2252013-01-17 17:16:24 -0800876 pool = worker->pool;
Tejun Heo111c2252013-01-17 17:16:24 -0800877
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100878 if (WARN_ON_ONCE(worker->sleeping))
879 return;
880
881 worker->sleeping = 1;
882 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +0200883
884 /*
885 * The counterpart of the following dec_and_test, implied mb,
886 * worklist not empty test sequence is in insert_work().
887 * Please read comment there.
888 *
Tejun Heo628c78e2012-07-17 12:39:27 -0700889 * NOT_RUNNING is clear. This means that we're bound to and
890 * running on the local cpu w/ rq lock held and preemption
891 * disabled, which in turn means that none else could be
Tejun Heod565ed62013-01-24 11:01:33 -0800892 * manipulating idle_list, so dereferencing idle_list without pool
Tejun Heo628c78e2012-07-17 12:39:27 -0700893 * lock is safe.
Tejun Heoe22bee72010-06-29 10:07:14 +0200894 */
Tejun Heoe19e3972013-01-24 11:39:44 -0800895 if (atomic_dec_and_test(&pool->nr_running) &&
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100896 !list_empty(&pool->worklist)) {
897 next = first_idle_worker(pool);
898 if (next)
899 wake_up_process(next->task);
900 }
901 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +0200902}
903
904/**
Johannes Weiner1b69ac62019-02-01 14:20:42 -0800905 * wq_worker_last_func - retrieve worker's last work function
Bart Van Assche8194fe92019-03-19 10:45:09 -0700906 * @task: Task to retrieve last work function of.
Johannes Weiner1b69ac62019-02-01 14:20:42 -0800907 *
908 * Determine the last function a worker executed. This is called from
909 * the scheduler to get a worker's last known identity.
910 *
911 * CONTEXT:
912 * spin_lock_irq(rq->lock)
913 *
Johannes Weiner4b047002019-03-07 16:29:30 -0800914 * This function is called during schedule() when a kworker is going
915 * to sleep. It's used by psi to identify aggregation workers during
916 * dequeuing, to allow periodic aggregation to shut-off when that
917 * worker is the last task in the system or cgroup to go to sleep.
918 *
919 * As this function doesn't involve any workqueue-related locking, it
920 * only returns stable values when called from inside the scheduler's
921 * queuing and dequeuing paths, when @task, which must be a kworker,
922 * is guaranteed to not be processing any works.
923 *
Johannes Weiner1b69ac62019-02-01 14:20:42 -0800924 * Return:
925 * The last work function %current executed as a worker, NULL if it
926 * hasn't executed any work yet.
927 */
928work_func_t wq_worker_last_func(struct task_struct *task)
929{
930 struct worker *worker = kthread_data(task);
931
932 return worker->last_func;
933}
934
935/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200936 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200937 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200938 * @flags: flags to set
Tejun Heod302f012010-06-29 10:07:13 +0200939 *
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800940 * Set @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200941 *
Tejun Heocb444762010-07-02 10:03:50 +0200942 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800943 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200944 */
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800945static inline void worker_set_flags(struct worker *worker, unsigned int flags)
Tejun Heod302f012010-06-29 10:07:13 +0200946{
Tejun Heobd7bdd42012-07-12 14:46:37 -0700947 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200948
Tejun Heocb444762010-07-02 10:03:50 +0200949 WARN_ON_ONCE(worker->task != current);
950
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800951 /* If transitioning into NOT_RUNNING, adjust nr_running. */
Tejun Heoe22bee72010-06-29 10:07:14 +0200952 if ((flags & WORKER_NOT_RUNNING) &&
953 !(worker->flags & WORKER_NOT_RUNNING)) {
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800954 atomic_dec(&pool->nr_running);
Tejun Heoe22bee72010-06-29 10:07:14 +0200955 }
956
Tejun Heod302f012010-06-29 10:07:13 +0200957 worker->flags |= flags;
958}
959
960/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200961 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200962 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200963 * @flags: flags to clear
964 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200965 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200966 *
Tejun Heocb444762010-07-02 10:03:50 +0200967 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800968 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200969 */
970static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
971{
Tejun Heo63d95a92012-07-12 14:46:37 -0700972 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200973 unsigned int oflags = worker->flags;
974
Tejun Heocb444762010-07-02 10:03:50 +0200975 WARN_ON_ONCE(worker->task != current);
976
Tejun Heod302f012010-06-29 10:07:13 +0200977 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200978
Tejun Heo42c025f2011-01-11 15:58:49 +0100979 /*
980 * If transitioning out of NOT_RUNNING, increment nr_running. Note
981 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
982 * of multiple flags, not a single flag.
983 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200984 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
985 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800986 atomic_inc(&pool->nr_running);
Tejun Heod302f012010-06-29 10:07:13 +0200987}
988
989/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200990 * find_worker_executing_work - find worker which is executing a work
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800991 * @pool: pool of interest
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200992 * @work: work to find worker for
993 *
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800994 * Find a worker which is executing @work on @pool by searching
995 * @pool->busy_hash which is keyed by the address of @work. For a worker
Tejun Heoa2c1c572012-12-18 10:35:02 -0800996 * to match, its current execution should match the address of @work and
997 * its work function. This is to avoid unwanted dependency between
998 * unrelated work executions through a work item being recycled while still
999 * being executed.
1000 *
1001 * This is a bit tricky. A work item may be freed once its execution
1002 * starts and nothing prevents the freed area from being recycled for
1003 * another work item. If the same work item address ends up being reused
1004 * before the original execution finishes, workqueue will identify the
1005 * recycled work item as currently executing and make it wait until the
1006 * current execution finishes, introducing an unwanted dependency.
1007 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001008 * This function checks the work item address and work function to avoid
1009 * false positives. Note that this isn't complete as one may construct a
1010 * work function which can introduce dependency onto itself through a
1011 * recycled work item. Well, if somebody wants to shoot oneself in the
1012 * foot that badly, there's only so much we can do, and if such deadlock
1013 * actually occurs, it should be easy to locate the culprit work function.
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001014 *
1015 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001016 * spin_lock_irq(pool->lock).
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001017 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001018 * Return:
1019 * Pointer to worker which is executing @work if found, %NULL
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001020 * otherwise.
1021 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08001022static struct worker *find_worker_executing_work(struct worker_pool *pool,
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001023 struct work_struct *work)
1024{
Sasha Levin42f85702012-12-17 10:01:23 -05001025 struct worker *worker;
Sasha Levin42f85702012-12-17 10:01:23 -05001026
Sasha Levinb67bfe02013-02-27 17:06:00 -08001027 hash_for_each_possible(pool->busy_hash, worker, hentry,
Tejun Heoa2c1c572012-12-18 10:35:02 -08001028 (unsigned long)work)
1029 if (worker->current_work == work &&
1030 worker->current_func == work->func)
Sasha Levin42f85702012-12-17 10:01:23 -05001031 return worker;
1032
1033 return NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001034}
1035
1036/**
Tejun Heobf4ede02012-08-03 10:30:46 -07001037 * move_linked_works - move linked works to a list
1038 * @work: start of series of works to be scheduled
1039 * @head: target list to append @work to
Shailendra Verma402dd892015-05-23 10:38:14 +05301040 * @nextp: out parameter for nested worklist walking
Tejun Heobf4ede02012-08-03 10:30:46 -07001041 *
1042 * Schedule linked works starting from @work to @head. Work series to
1043 * be scheduled starts at @work and includes any consecutive work with
1044 * WORK_STRUCT_LINKED set in its predecessor.
1045 *
1046 * If @nextp is not NULL, it's updated to point to the next work of
1047 * the last scheduled work. This allows move_linked_works() to be
1048 * nested inside outer list_for_each_entry_safe().
1049 *
1050 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001051 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001052 */
1053static void move_linked_works(struct work_struct *work, struct list_head *head,
1054 struct work_struct **nextp)
1055{
1056 struct work_struct *n;
1057
1058 /*
1059 * Linked worklist will always end before the end of the list,
1060 * use NULL for list head.
1061 */
1062 list_for_each_entry_safe_from(work, n, NULL, entry) {
1063 list_move_tail(&work->entry, head);
1064 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1065 break;
1066 }
1067
1068 /*
1069 * If we're already inside safe list traversal and have moved
1070 * multiple works to the scheduled queue, the next position
1071 * needs to be updated.
1072 */
1073 if (nextp)
1074 *nextp = n;
1075}
1076
Tejun Heo8864b4e2013-03-12 11:30:04 -07001077/**
1078 * get_pwq - get an extra reference on the specified pool_workqueue
1079 * @pwq: pool_workqueue to get
1080 *
1081 * Obtain an extra reference on @pwq. The caller should guarantee that
1082 * @pwq has positive refcnt and be holding the matching pool->lock.
1083 */
1084static void get_pwq(struct pool_workqueue *pwq)
1085{
1086 lockdep_assert_held(&pwq->pool->lock);
1087 WARN_ON_ONCE(pwq->refcnt <= 0);
1088 pwq->refcnt++;
1089}
1090
1091/**
1092 * put_pwq - put a pool_workqueue reference
1093 * @pwq: pool_workqueue to put
1094 *
1095 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1096 * destruction. The caller should be holding the matching pool->lock.
1097 */
1098static void put_pwq(struct pool_workqueue *pwq)
1099{
1100 lockdep_assert_held(&pwq->pool->lock);
1101 if (likely(--pwq->refcnt))
1102 return;
1103 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1104 return;
1105 /*
1106 * @pwq can't be released under pool->lock, bounce to
1107 * pwq_unbound_release_workfn(). This never recurses on the same
1108 * pool->lock as this path is taken only for unbound workqueues and
1109 * the release work item is scheduled on a per-cpu workqueue. To
1110 * avoid lockdep warning, unbound pool->locks are given lockdep
1111 * subclass of 1 in get_unbound_pool().
1112 */
1113 schedule_work(&pwq->unbound_release_work);
1114}
1115
Tejun Heodce90d42013-04-01 11:23:35 -07001116/**
1117 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1118 * @pwq: pool_workqueue to put (can be %NULL)
1119 *
1120 * put_pwq() with locking. This function also allows %NULL @pwq.
1121 */
1122static void put_pwq_unlocked(struct pool_workqueue *pwq)
1123{
1124 if (pwq) {
1125 /*
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001126 * As both pwqs and pools are RCU protected, the
Tejun Heodce90d42013-04-01 11:23:35 -07001127 * following lock operations are safe.
1128 */
1129 spin_lock_irq(&pwq->pool->lock);
1130 put_pwq(pwq);
1131 spin_unlock_irq(&pwq->pool->lock);
1132 }
1133}
1134
Tejun Heo112202d2013-02-13 19:29:12 -08001135static void pwq_activate_delayed_work(struct work_struct *work)
Tejun Heobf4ede02012-08-03 10:30:46 -07001136{
Tejun Heo112202d2013-02-13 19:29:12 -08001137 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobf4ede02012-08-03 10:30:46 -07001138
1139 trace_workqueue_activate_work(work);
Tejun Heo82607adc2015-12-08 11:28:04 -05001140 if (list_empty(&pwq->pool->worklist))
1141 pwq->pool->watchdog_ts = jiffies;
Tejun Heo112202d2013-02-13 19:29:12 -08001142 move_linked_works(work, &pwq->pool->worklist, NULL);
Tejun Heobf4ede02012-08-03 10:30:46 -07001143 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo112202d2013-02-13 19:29:12 -08001144 pwq->nr_active++;
Tejun Heobf4ede02012-08-03 10:30:46 -07001145}
1146
Tejun Heo112202d2013-02-13 19:29:12 -08001147static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001148{
Tejun Heo112202d2013-02-13 19:29:12 -08001149 struct work_struct *work = list_first_entry(&pwq->delayed_works,
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001150 struct work_struct, entry);
1151
Tejun Heo112202d2013-02-13 19:29:12 -08001152 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001153}
1154
Tejun Heobf4ede02012-08-03 10:30:46 -07001155/**
Tejun Heo112202d2013-02-13 19:29:12 -08001156 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1157 * @pwq: pwq of interest
Tejun Heobf4ede02012-08-03 10:30:46 -07001158 * @color: color of work which left the queue
Tejun Heobf4ede02012-08-03 10:30:46 -07001159 *
1160 * A work either has completed or is removed from pending queue,
Tejun Heo112202d2013-02-13 19:29:12 -08001161 * decrement nr_in_flight of its pwq and handle workqueue flushing.
Tejun Heobf4ede02012-08-03 10:30:46 -07001162 *
1163 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001164 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001165 */
Tejun Heo112202d2013-02-13 19:29:12 -08001166static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
Tejun Heobf4ede02012-08-03 10:30:46 -07001167{
Tejun Heo8864b4e2013-03-12 11:30:04 -07001168 /* uncolored work items don't participate in flushing or nr_active */
Tejun Heobf4ede02012-08-03 10:30:46 -07001169 if (color == WORK_NO_COLOR)
Tejun Heo8864b4e2013-03-12 11:30:04 -07001170 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001171
Tejun Heo112202d2013-02-13 19:29:12 -08001172 pwq->nr_in_flight[color]--;
Tejun Heobf4ede02012-08-03 10:30:46 -07001173
Tejun Heo112202d2013-02-13 19:29:12 -08001174 pwq->nr_active--;
1175 if (!list_empty(&pwq->delayed_works)) {
Lai Jiangshanb3f9f402012-09-18 10:40:00 -07001176 /* one down, submit a delayed one */
Tejun Heo112202d2013-02-13 19:29:12 -08001177 if (pwq->nr_active < pwq->max_active)
1178 pwq_activate_first_delayed(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001179 }
1180
1181 /* is flush in progress and are we at the flushing tip? */
Tejun Heo112202d2013-02-13 19:29:12 -08001182 if (likely(pwq->flush_color != color))
Tejun Heo8864b4e2013-03-12 11:30:04 -07001183 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001184
1185 /* are there still in-flight works? */
Tejun Heo112202d2013-02-13 19:29:12 -08001186 if (pwq->nr_in_flight[color])
Tejun Heo8864b4e2013-03-12 11:30:04 -07001187 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001188
Tejun Heo112202d2013-02-13 19:29:12 -08001189 /* this pwq is done, clear flush_color */
1190 pwq->flush_color = -1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001191
1192 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001193 * If this was the last pwq, wake up the first flusher. It
Tejun Heobf4ede02012-08-03 10:30:46 -07001194 * will handle the rest.
1195 */
Tejun Heo112202d2013-02-13 19:29:12 -08001196 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1197 complete(&pwq->wq->first_flusher->done);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001198out_put:
1199 put_pwq(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001200}
1201
Tejun Heo36e227d2012-08-03 10:30:46 -07001202/**
Tejun Heobbb68df2012-08-03 10:30:46 -07001203 * try_to_grab_pending - steal work item from worklist and disable irq
Tejun Heo36e227d2012-08-03 10:30:46 -07001204 * @work: work item to steal
1205 * @is_dwork: @work is a delayed_work
Tejun Heobbb68df2012-08-03 10:30:46 -07001206 * @flags: place to store irq state
Tejun Heo36e227d2012-08-03 10:30:46 -07001207 *
1208 * Try to grab PENDING bit of @work. This function can handle @work in any
Yacine Belkadid185af32013-07-31 14:59:24 -07001209 * stable state - idle, on timer or on worklist.
Tejun Heo36e227d2012-08-03 10:30:46 -07001210 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001211 * Return:
Tejun Heo36e227d2012-08-03 10:30:46 -07001212 * 1 if @work was pending and we successfully stole PENDING
1213 * 0 if @work was idle and we claimed PENDING
1214 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
Tejun Heobbb68df2012-08-03 10:30:46 -07001215 * -ENOENT if someone else is canceling @work, this state may persist
1216 * for arbitrarily long
Tejun Heo36e227d2012-08-03 10:30:46 -07001217 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001218 * Note:
Tejun Heobbb68df2012-08-03 10:30:46 -07001219 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001220 * interrupted while holding PENDING and @work off queue, irq must be
1221 * disabled on entry. This, combined with delayed_work->timer being
1222 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
Tejun Heobbb68df2012-08-03 10:30:46 -07001223 *
1224 * On successful return, >= 0, irq is disabled and the caller is
1225 * responsible for releasing it using local_irq_restore(*@flags).
1226 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001227 * This function is safe to call from any context including IRQ handler.
Tejun Heobf4ede02012-08-03 10:30:46 -07001228 */
Tejun Heobbb68df2012-08-03 10:30:46 -07001229static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1230 unsigned long *flags)
Tejun Heobf4ede02012-08-03 10:30:46 -07001231{
Tejun Heod565ed62013-01-24 11:01:33 -08001232 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001233 struct pool_workqueue *pwq;
Tejun Heobf4ede02012-08-03 10:30:46 -07001234
Tejun Heobbb68df2012-08-03 10:30:46 -07001235 local_irq_save(*flags);
1236
Tejun Heo36e227d2012-08-03 10:30:46 -07001237 /* try to steal the timer if it exists */
1238 if (is_dwork) {
1239 struct delayed_work *dwork = to_delayed_work(work);
1240
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001241 /*
1242 * dwork->timer is irqsafe. If del_timer() fails, it's
1243 * guaranteed that the timer is not queued anywhere and not
1244 * running on the local CPU.
1245 */
Tejun Heo36e227d2012-08-03 10:30:46 -07001246 if (likely(del_timer(&dwork->timer)))
1247 return 1;
1248 }
1249
1250 /* try to claim PENDING the normal way */
Tejun Heobf4ede02012-08-03 10:30:46 -07001251 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1252 return 0;
1253
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001254 rcu_read_lock();
Tejun Heobf4ede02012-08-03 10:30:46 -07001255 /*
1256 * The queueing is in progress, or it is already queued. Try to
1257 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1258 */
Tejun Heod565ed62013-01-24 11:01:33 -08001259 pool = get_work_pool(work);
1260 if (!pool)
Tejun Heobbb68df2012-08-03 10:30:46 -07001261 goto fail;
Tejun Heobf4ede02012-08-03 10:30:46 -07001262
Tejun Heod565ed62013-01-24 11:01:33 -08001263 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001264 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001265 * work->data is guaranteed to point to pwq only while the work
1266 * item is queued on pwq->wq, and both updating work->data to point
1267 * to pwq on queueing and to pool on dequeueing are done under
1268 * pwq->pool->lock. This in turn guarantees that, if work->data
1269 * points to pwq which is associated with a locked pool, the work
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001270 * item is currently queued on that pool.
1271 */
Tejun Heo112202d2013-02-13 19:29:12 -08001272 pwq = get_work_pwq(work);
1273 if (pwq && pwq->pool == pool) {
Tejun Heo16062832013-02-06 18:04:53 -08001274 debug_work_deactivate(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001275
Tejun Heo16062832013-02-06 18:04:53 -08001276 /*
1277 * A delayed work item cannot be grabbed directly because
1278 * it might have linked NO_COLOR work items which, if left
Tejun Heo112202d2013-02-13 19:29:12 -08001279 * on the delayed_list, will confuse pwq->nr_active
Tejun Heo16062832013-02-06 18:04:53 -08001280 * management later on and cause stall. Make sure the work
1281 * item is activated before grabbing.
1282 */
1283 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
Tejun Heo112202d2013-02-13 19:29:12 -08001284 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001285
Tejun Heo16062832013-02-06 18:04:53 -08001286 list_del_init(&work->entry);
Lai Jiangshan9c34a702014-07-11 00:11:13 +08001287 pwq_dec_nr_in_flight(pwq, get_work_color(work));
Tejun Heo36e227d2012-08-03 10:30:46 -07001288
Tejun Heo112202d2013-02-13 19:29:12 -08001289 /* work->data points to pwq iff queued, point to pool */
Tejun Heo16062832013-02-06 18:04:53 -08001290 set_work_pool_and_keep_pending(work, pool->id);
Lai Jiangshan4468a002013-02-06 18:04:53 -08001291
Tejun Heo16062832013-02-06 18:04:53 -08001292 spin_unlock(&pool->lock);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001293 rcu_read_unlock();
Tejun Heo16062832013-02-06 18:04:53 -08001294 return 1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001295 }
Tejun Heod565ed62013-01-24 11:01:33 -08001296 spin_unlock(&pool->lock);
Tejun Heobbb68df2012-08-03 10:30:46 -07001297fail:
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001298 rcu_read_unlock();
Tejun Heobbb68df2012-08-03 10:30:46 -07001299 local_irq_restore(*flags);
1300 if (work_is_canceling(work))
1301 return -ENOENT;
1302 cpu_relax();
Tejun Heo36e227d2012-08-03 10:30:46 -07001303 return -EAGAIN;
Tejun Heobf4ede02012-08-03 10:30:46 -07001304}
1305
1306/**
Tejun Heo706026c2013-01-24 11:01:34 -08001307 * insert_work - insert a work into a pool
Tejun Heo112202d2013-02-13 19:29:12 -08001308 * @pwq: pwq @work belongs to
Tejun Heo4690c4a2010-06-29 10:07:10 +02001309 * @work: work to insert
1310 * @head: insertion point
1311 * @extra_flags: extra WORK_STRUCT_* flags to set
1312 *
Tejun Heo112202d2013-02-13 19:29:12 -08001313 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
Tejun Heo706026c2013-01-24 11:01:34 -08001314 * work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +02001315 *
1316 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001317 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02001318 */
Tejun Heo112202d2013-02-13 19:29:12 -08001319static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1320 struct list_head *head, unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001321{
Tejun Heo112202d2013-02-13 19:29:12 -08001322 struct worker_pool *pool = pwq->pool;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01001323
Tejun Heo4690c4a2010-06-29 10:07:10 +02001324 /* we own @work, set data and link */
Tejun Heo112202d2013-02-13 19:29:12 -08001325 set_work_pwq(work, pwq, extra_flags);
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -07001326 list_add_tail(&work->entry, head);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001327 get_pwq(pwq);
Tejun Heoe22bee72010-06-29 10:07:14 +02001328
1329 /*
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001330 * Ensure either wq_worker_sleeping() sees the above
1331 * list_add_tail() or we see zero nr_running to avoid workers lying
1332 * around lazily while there are works to be processed.
Tejun Heoe22bee72010-06-29 10:07:14 +02001333 */
1334 smp_mb();
1335
Tejun Heo63d95a92012-07-12 14:46:37 -07001336 if (__need_more_worker(pool))
1337 wake_up_worker(pool);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001338}
1339
Tejun Heoc8efcc22010-12-20 19:32:04 +01001340/*
1341 * Test whether @work is being queued from another work executing on the
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001342 * same workqueue.
Tejun Heoc8efcc22010-12-20 19:32:04 +01001343 */
1344static bool is_chained_work(struct workqueue_struct *wq)
1345{
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001346 struct worker *worker;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001347
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001348 worker = current_wq_worker();
1349 /*
Bart Van Asschebf393fd2019-03-01 13:57:25 -08001350 * Return %true iff I'm a worker executing a work item on @wq. If
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001351 * I'm @worker, it's safe to dereference it without locking.
1352 */
Tejun Heo112202d2013-02-13 19:29:12 -08001353 return worker && worker->current_pwq->wq == wq;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001354}
1355
Mike Galbraithef5571802016-02-09 17:59:38 -05001356/*
1357 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1358 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1359 * avoid perturbing sensitive tasks.
1360 */
1361static int wq_select_unbound_cpu(int cpu)
1362{
Tejun Heof303fccb2016-02-09 17:59:38 -05001363 static bool printed_dbg_warning;
Mike Galbraithef5571802016-02-09 17:59:38 -05001364 int new_cpu;
1365
Tejun Heof303fccb2016-02-09 17:59:38 -05001366 if (likely(!wq_debug_force_rr_cpu)) {
1367 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1368 return cpu;
1369 } else if (!printed_dbg_warning) {
1370 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1371 printed_dbg_warning = true;
1372 }
1373
Mike Galbraithef5571802016-02-09 17:59:38 -05001374 if (cpumask_empty(wq_unbound_cpumask))
1375 return cpu;
1376
1377 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1378 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1379 if (unlikely(new_cpu >= nr_cpu_ids)) {
1380 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1381 if (unlikely(new_cpu >= nr_cpu_ids))
1382 return cpu;
1383 }
1384 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1385
1386 return new_cpu;
1387}
1388
Tejun Heod84ff052013-03-12 11:29:59 -07001389static void __queue_work(int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 struct work_struct *work)
1391{
Tejun Heo112202d2013-02-13 19:29:12 -08001392 struct pool_workqueue *pwq;
Tejun Heoc9178082013-03-12 11:30:04 -07001393 struct worker_pool *last_pool;
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001394 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001395 unsigned int work_flags;
Joonsoo Kimb75cac92012-08-15 23:25:37 +09001396 unsigned int req_cpu = cpu;
Tejun Heo8930cab2012-08-03 10:30:45 -07001397
1398 /*
1399 * While a work item is PENDING && off queue, a task trying to
1400 * steal the PENDING will busy-loop waiting for it to either get
1401 * queued or lose PENDING. Grabbing PENDING and queueing should
1402 * happen with IRQ disabled.
1403 */
Frederic Weisbecker8e8eb732017-11-06 16:01:19 +01001404 lockdep_assert_irqs_disabled();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09001406 debug_work_activate(work);
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001407
Li Bin9ef28a72013-09-09 13:13:58 +08001408 /* if draining, only works from the same workqueue are allowed */
Tejun Heo618b01e2013-03-12 11:30:04 -07001409 if (unlikely(wq->flags & __WQ_DRAINING) &&
Tejun Heoc8efcc22010-12-20 19:32:04 +01001410 WARN_ON_ONCE(!is_chained_work(wq)))
Tejun Heoe41e7042010-08-24 14:22:47 +02001411 return;
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001412 rcu_read_lock();
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001413retry:
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001414 if (req_cpu == WORK_CPU_UNBOUND)
Mike Galbraithef5571802016-02-09 17:59:38 -05001415 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001416
Tejun Heoc9178082013-03-12 11:30:04 -07001417 /* pwq which will be used unless @work is executing elsewhere */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001418 if (!(wq->flags & WQ_UNBOUND))
Tejun Heo7fb98ea2013-03-12 11:30:00 -07001419 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001420 else
1421 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodbf25762012-08-20 14:51:23 -07001422
Tejun Heoc9178082013-03-12 11:30:04 -07001423 /*
1424 * If @work was previously on a different pool, it might still be
1425 * running there, in which case the work needs to be queued on that
1426 * pool to guarantee non-reentrancy.
1427 */
1428 last_pool = get_work_pool(work);
1429 if (last_pool && last_pool != pwq->pool) {
1430 struct worker *worker;
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001431
Tejun Heoc9178082013-03-12 11:30:04 -07001432 spin_lock(&last_pool->lock);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001433
Tejun Heoc9178082013-03-12 11:30:04 -07001434 worker = find_worker_executing_work(last_pool, work);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001435
Tejun Heoc9178082013-03-12 11:30:04 -07001436 if (worker && worker->current_pwq->wq == wq) {
1437 pwq = worker->current_pwq;
Tejun Heo8930cab2012-08-03 10:30:45 -07001438 } else {
Tejun Heoc9178082013-03-12 11:30:04 -07001439 /* meh... not running there, queue here */
1440 spin_unlock(&last_pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08001441 spin_lock(&pwq->pool->lock);
Tejun Heo8930cab2012-08-03 10:30:45 -07001442 }
Tejun Heof3421792010-07-02 10:03:51 +02001443 } else {
Tejun Heo112202d2013-02-13 19:29:12 -08001444 spin_lock(&pwq->pool->lock);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001445 }
1446
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001447 /*
1448 * pwq is determined and locked. For unbound pools, we could have
1449 * raced with pwq release and it could already be dead. If its
1450 * refcnt is zero, repeat pwq selection. Note that pwqs never die
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001451 * without another pwq replacing it in the numa_pwq_tbl or while
1452 * work items are executing on it, so the retrying is guaranteed to
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001453 * make forward-progress.
1454 */
1455 if (unlikely(!pwq->refcnt)) {
1456 if (wq->flags & WQ_UNBOUND) {
1457 spin_unlock(&pwq->pool->lock);
1458 cpu_relax();
1459 goto retry;
1460 }
1461 /* oops */
1462 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1463 wq->name, cpu);
1464 }
1465
Tejun Heo112202d2013-02-13 19:29:12 -08001466 /* pwq determined, queue */
1467 trace_workqueue_queue_work(req_cpu, pwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001468
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001469 if (WARN_ON(!list_empty(&work->entry)))
1470 goto out;
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001471
Tejun Heo112202d2013-02-13 19:29:12 -08001472 pwq->nr_in_flight[pwq->work_color]++;
1473 work_flags = work_color_to_flags(pwq->work_color);
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001474
Tejun Heo112202d2013-02-13 19:29:12 -08001475 if (likely(pwq->nr_active < pwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001476 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001477 pwq->nr_active++;
1478 worklist = &pwq->pool->worklist;
Tejun Heo82607adc2015-12-08 11:28:04 -05001479 if (list_empty(worklist))
1480 pwq->pool->watchdog_ts = jiffies;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001481 } else {
1482 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo112202d2013-02-13 19:29:12 -08001483 worklist = &pwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001484 }
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001485
Tejun Heo112202d2013-02-13 19:29:12 -08001486 insert_work(pwq, work, worklist, work_flags);
Tejun Heo1e19ffc62010-06-29 10:07:12 +02001487
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001488out:
Tejun Heo112202d2013-02-13 19:29:12 -08001489 spin_unlock(&pwq->pool->lock);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001490 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491}
1492
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001493/**
Zhang Ruic1a220e2008-07-23 21:28:39 -07001494 * queue_work_on - queue work on specific cpu
1495 * @cpu: CPU number to execute work on
1496 * @wq: workqueue to use
1497 * @work: work to queue
1498 *
Zhang Ruic1a220e2008-07-23 21:28:39 -07001499 * We queue the work to a specific CPU, the caller must ensure it
1500 * can't go away.
Yacine Belkadid185af32013-07-31 14:59:24 -07001501 *
1502 * Return: %false if @work was already on a queue, %true otherwise.
Zhang Ruic1a220e2008-07-23 21:28:39 -07001503 */
Tejun Heod4283e92012-08-03 10:30:44 -07001504bool queue_work_on(int cpu, struct workqueue_struct *wq,
1505 struct work_struct *work)
Zhang Ruic1a220e2008-07-23 21:28:39 -07001506{
Tejun Heod4283e92012-08-03 10:30:44 -07001507 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001508 unsigned long flags;
1509
1510 local_irq_save(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001511
Tejun Heo22df02b2010-06-29 10:07:10 +02001512 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001513 __queue_work(cpu, wq, work);
Tejun Heod4283e92012-08-03 10:30:44 -07001514 ret = true;
Zhang Ruic1a220e2008-07-23 21:28:39 -07001515 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001516
1517 local_irq_restore(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001518 return ret;
1519}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001520EXPORT_SYMBOL(queue_work_on);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001521
Alexander Duyck8204e0c2019-01-22 10:39:26 -08001522/**
1523 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1524 * @node: NUMA node ID that we want to select a CPU from
1525 *
1526 * This function will attempt to find a "random" cpu available on a given
1527 * node. If there are no CPUs available on the given node it will return
1528 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1529 * available CPU if we need to schedule this work.
1530 */
1531static int workqueue_select_cpu_near(int node)
1532{
1533 int cpu;
1534
1535 /* No point in doing this if NUMA isn't enabled for workqueues */
1536 if (!wq_numa_enabled)
1537 return WORK_CPU_UNBOUND;
1538
1539 /* Delay binding to CPU if node is not valid or online */
1540 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1541 return WORK_CPU_UNBOUND;
1542
1543 /* Use local node/cpu if we are already there */
1544 cpu = raw_smp_processor_id();
1545 if (node == cpu_to_node(cpu))
1546 return cpu;
1547
1548 /* Use "random" otherwise know as "first" online CPU of node */
1549 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1550
1551 /* If CPU is valid return that, otherwise just defer */
1552 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1553}
1554
1555/**
1556 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1557 * @node: NUMA node that we are targeting the work for
1558 * @wq: workqueue to use
1559 * @work: work to queue
1560 *
1561 * We queue the work to a "random" CPU within a given NUMA node. The basic
1562 * idea here is to provide a way to somehow associate work with a given
1563 * NUMA node.
1564 *
1565 * This function will only make a best effort attempt at getting this onto
1566 * the right NUMA node. If no node is requested or the requested node is
1567 * offline then we just fall back to standard queue_work behavior.
1568 *
1569 * Currently the "random" CPU ends up being the first available CPU in the
1570 * intersection of cpu_online_mask and the cpumask of the node, unless we
1571 * are running on the node. In that case we just use the current CPU.
1572 *
1573 * Return: %false if @work was already on a queue, %true otherwise.
1574 */
1575bool queue_work_node(int node, struct workqueue_struct *wq,
1576 struct work_struct *work)
1577{
1578 unsigned long flags;
1579 bool ret = false;
1580
1581 /*
1582 * This current implementation is specific to unbound workqueues.
1583 * Specifically we only return the first available CPU for a given
1584 * node instead of cycling through individual CPUs within the node.
1585 *
1586 * If this is used with a per-cpu workqueue then the logic in
1587 * workqueue_select_cpu_near would need to be updated to allow for
1588 * some round robin type logic.
1589 */
1590 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1591
1592 local_irq_save(flags);
1593
1594 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1595 int cpu = workqueue_select_cpu_near(node);
1596
1597 __queue_work(cpu, wq, work);
1598 ret = true;
1599 }
1600
1601 local_irq_restore(flags);
1602 return ret;
1603}
1604EXPORT_SYMBOL_GPL(queue_work_node);
1605
Kees Cook8c20feb2017-10-04 16:27:07 -07001606void delayed_work_timer_fn(struct timer_list *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607{
Kees Cook8c20feb2017-10-04 16:27:07 -07001608 struct delayed_work *dwork = from_timer(dwork, t, timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001610 /* should have been called from irqsafe timer with irq already off */
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001611 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612}
Konstantin Khlebnikov1438ade52013-01-24 16:36:31 +04001613EXPORT_SYMBOL(delayed_work_timer_fn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001615static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1616 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617{
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001618 struct timer_list *timer = &dwork->timer;
1619 struct work_struct *work = &dwork->work;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620
Tejun Heo637fdba2017-03-06 15:33:42 -05001621 WARN_ON_ONCE(!wq);
Kees Cook841b86f2017-10-23 09:40:42 +02001622 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
Tejun Heofc4b5142012-12-04 07:40:39 -08001623 WARN_ON_ONCE(timer_pending(timer));
1624 WARN_ON_ONCE(!list_empty(&work->entry));
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001625
Tejun Heo8852aac2012-12-01 16:23:42 -08001626 /*
1627 * If @delay is 0, queue @dwork->work immediately. This is for
1628 * both optimization and correctness. The earliest @timer can
1629 * expire is on the closest next tick and delayed_work users depend
1630 * on that there's no such delay when @delay is 0.
1631 */
1632 if (!delay) {
1633 __queue_work(cpu, wq, &dwork->work);
1634 return;
1635 }
1636
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001637 dwork->wq = wq;
Tejun Heo12650572012-08-08 09:38:42 -07001638 dwork->cpu = cpu;
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001639 timer->expires = jiffies + delay;
1640
Tejun Heo041bd122016-02-09 16:11:26 -05001641 if (unlikely(cpu != WORK_CPU_UNBOUND))
1642 add_timer_on(timer, cpu);
1643 else
1644 add_timer(timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645}
1646
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001647/**
1648 * queue_delayed_work_on - queue work on specific CPU after delay
1649 * @cpu: CPU number to execute work on
1650 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001651 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001652 * @delay: number of jiffies to wait before queueing
1653 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001654 * Return: %false if @work was already on a queue, %true otherwise. If
Tejun Heo715f1302012-08-03 10:30:46 -07001655 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1656 * execution.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001657 */
Tejun Heod4283e92012-08-03 10:30:44 -07001658bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1659 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001660{
David Howells52bad642006-11-22 14:54:01 +00001661 struct work_struct *work = &dwork->work;
Tejun Heod4283e92012-08-03 10:30:44 -07001662 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001663 unsigned long flags;
1664
1665 /* read the comment in __queue_work() */
1666 local_irq_save(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001667
Tejun Heo22df02b2010-06-29 10:07:10 +02001668 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001669 __queue_delayed_work(cpu, wq, dwork, delay);
Tejun Heod4283e92012-08-03 10:30:44 -07001670 ret = true;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001671 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001672
1673 local_irq_restore(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001674 return ret;
1675}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001676EXPORT_SYMBOL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677
Tejun Heoc8e55f32010-06-29 10:07:12 +02001678/**
Tejun Heo8376fe22012-08-03 10:30:47 -07001679 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1680 * @cpu: CPU number to execute work on
1681 * @wq: workqueue to use
1682 * @dwork: work to queue
1683 * @delay: number of jiffies to wait before queueing
1684 *
1685 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1686 * modify @dwork's timer so that it expires after @delay. If @delay is
1687 * zero, @work is guaranteed to be scheduled immediately regardless of its
1688 * current state.
1689 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001690 * Return: %false if @dwork was idle and queued, %true if @dwork was
Tejun Heo8376fe22012-08-03 10:30:47 -07001691 * pending and its timer was modified.
1692 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001693 * This function is safe to call from any context including IRQ handler.
Tejun Heo8376fe22012-08-03 10:30:47 -07001694 * See try_to_grab_pending() for details.
1695 */
1696bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1697 struct delayed_work *dwork, unsigned long delay)
1698{
1699 unsigned long flags;
1700 int ret;
1701
1702 do {
1703 ret = try_to_grab_pending(&dwork->work, true, &flags);
1704 } while (unlikely(ret == -EAGAIN));
1705
1706 if (likely(ret >= 0)) {
1707 __queue_delayed_work(cpu, wq, dwork, delay);
1708 local_irq_restore(flags);
1709 }
1710
1711 /* -ENOENT from try_to_grab_pending() becomes %true */
1712 return ret;
1713}
1714EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1715
Tejun Heo05f0fe62018-03-14 12:45:13 -07001716static void rcu_work_rcufn(struct rcu_head *rcu)
1717{
1718 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1719
1720 /* read the comment in __queue_work() */
1721 local_irq_disable();
1722 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1723 local_irq_enable();
1724}
1725
1726/**
1727 * queue_rcu_work - queue work after a RCU grace period
1728 * @wq: workqueue to use
1729 * @rwork: work to queue
1730 *
1731 * Return: %false if @rwork was already pending, %true otherwise. Note
1732 * that a full RCU grace period is guaranteed only after a %true return.
Bart Van Asschebf393fd2019-03-01 13:57:25 -08001733 * While @rwork is guaranteed to be executed after a %false return, the
Tejun Heo05f0fe62018-03-14 12:45:13 -07001734 * execution may happen before a full RCU grace period has passed.
1735 */
1736bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1737{
1738 struct work_struct *work = &rwork->work;
1739
1740 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1741 rwork->wq = wq;
1742 call_rcu(&rwork->rcu, rcu_work_rcufn);
1743 return true;
1744 }
1745
1746 return false;
1747}
1748EXPORT_SYMBOL(queue_rcu_work);
1749
Tejun Heo8376fe22012-08-03 10:30:47 -07001750/**
Tejun Heoc8e55f32010-06-29 10:07:12 +02001751 * worker_enter_idle - enter idle state
1752 * @worker: worker which is entering idle state
1753 *
1754 * @worker is entering idle state. Update stats and idle timer if
1755 * necessary.
1756 *
1757 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001758 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001759 */
1760static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001762 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763
Tejun Heo6183c002013-03-12 11:29:57 -07001764 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1765 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1766 (worker->hentry.next || worker->hentry.pprev)))
1767 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768
Lai Jiangshan051e1852014-07-22 13:03:02 +08001769 /* can't use worker_set_flags(), also called from create_worker() */
Tejun Heocb444762010-07-02 10:03:50 +02001770 worker->flags |= WORKER_IDLE;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001771 pool->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001772 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001773
Tejun Heoc8e55f32010-06-29 10:07:12 +02001774 /* idle_list is LIFO */
Tejun Heobd7bdd42012-07-12 14:46:37 -07001775 list_add(&worker->entry, &pool->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001776
Tejun Heo628c78e2012-07-17 12:39:27 -07001777 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1778 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
Tejun Heocb444762010-07-02 10:03:50 +02001779
Tejun Heo544ecf32012-05-14 15:04:50 -07001780 /*
Lai Jiangshane8b3f8d2017-12-01 22:20:36 +08001781 * Sanity check nr_running. Because unbind_workers() releases
Tejun Heod565ed62013-01-24 11:01:33 -08001782 * pool->lock between setting %WORKER_UNBOUND and zapping
Tejun Heo628c78e2012-07-17 12:39:27 -07001783 * nr_running, the warning may trigger spuriously. Check iff
1784 * unbind is not in progress.
Tejun Heo544ecf32012-05-14 15:04:50 -07001785 */
Tejun Heo24647572013-01-24 11:01:33 -08001786 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -07001787 pool->nr_workers == pool->nr_idle &&
Tejun Heoe19e3972013-01-24 11:39:44 -08001788 atomic_read(&pool->nr_running));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789}
1790
Tejun Heoc8e55f32010-06-29 10:07:12 +02001791/**
1792 * worker_leave_idle - leave idle state
1793 * @worker: worker which is leaving idle state
1794 *
1795 * @worker is leaving idle state. Update stats.
1796 *
1797 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001798 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001799 */
1800static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001802 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803
Tejun Heo6183c002013-03-12 11:29:57 -07001804 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1805 return;
Tejun Heod302f012010-06-29 10:07:13 +02001806 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001807 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001808 list_del_init(&worker->entry);
1809}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001811static struct worker *alloc_worker(int node)
Tejun Heoc34056a2010-06-29 10:07:11 +02001812{
1813 struct worker *worker;
1814
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001815 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001816 if (worker) {
1817 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001818 INIT_LIST_HEAD(&worker->scheduled);
Lai Jiangshanda028462014-05-20 17:46:31 +08001819 INIT_LIST_HEAD(&worker->node);
Tejun Heoe22bee72010-06-29 10:07:14 +02001820 /* on creation a worker is in !idle && prep state */
1821 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001822 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001823 return worker;
1824}
1825
1826/**
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001827 * worker_attach_to_pool() - attach a worker to a pool
1828 * @worker: worker to be attached
1829 * @pool: the target pool
1830 *
1831 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1832 * cpu-binding of @worker are kept coordinated with the pool across
1833 * cpu-[un]hotplugs.
1834 */
1835static void worker_attach_to_pool(struct worker *worker,
1836 struct worker_pool *pool)
1837{
Tejun Heo1258fae2018-05-18 08:47:13 -07001838 mutex_lock(&wq_pool_attach_mutex);
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001839
1840 /*
1841 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1842 * online CPUs. It'll be re-applied when any of the CPUs come up.
1843 */
1844 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1845
1846 /*
Tejun Heo1258fae2018-05-18 08:47:13 -07001847 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1848 * stable across this function. See the comments above the flag
1849 * definition for details.
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001850 */
1851 if (pool->flags & POOL_DISASSOCIATED)
1852 worker->flags |= WORKER_UNBOUND;
1853
1854 list_add_tail(&worker->node, &pool->workers);
Tejun Heoa2d812a2018-05-18 08:47:13 -07001855 worker->pool = pool;
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001856
Tejun Heo1258fae2018-05-18 08:47:13 -07001857 mutex_unlock(&wq_pool_attach_mutex);
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001858}
1859
1860/**
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001861 * worker_detach_from_pool() - detach a worker from its pool
1862 * @worker: worker which is attached to its pool
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001863 *
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001864 * Undo the attaching which had been done in worker_attach_to_pool(). The
1865 * caller worker shouldn't access to the pool after detached except it has
1866 * other reference to the pool.
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001867 */
Tejun Heoa2d812a2018-05-18 08:47:13 -07001868static void worker_detach_from_pool(struct worker *worker)
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001869{
Tejun Heoa2d812a2018-05-18 08:47:13 -07001870 struct worker_pool *pool = worker->pool;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001871 struct completion *detach_completion = NULL;
1872
Tejun Heo1258fae2018-05-18 08:47:13 -07001873 mutex_lock(&wq_pool_attach_mutex);
Tejun Heoa2d812a2018-05-18 08:47:13 -07001874
Lai Jiangshanda028462014-05-20 17:46:31 +08001875 list_del(&worker->node);
Tejun Heoa2d812a2018-05-18 08:47:13 -07001876 worker->pool = NULL;
1877
Lai Jiangshanda028462014-05-20 17:46:31 +08001878 if (list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001879 detach_completion = pool->detach_completion;
Tejun Heo1258fae2018-05-18 08:47:13 -07001880 mutex_unlock(&wq_pool_attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001881
Lai Jiangshanb62c0752014-06-03 15:32:52 +08001882 /* clear leftover flags without pool->lock after it is detached */
1883 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1884
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001885 if (detach_completion)
1886 complete(detach_completion);
1887}
1888
1889/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001890 * create_worker - create a new workqueue worker
Tejun Heo63d95a92012-07-12 14:46:37 -07001891 * @pool: pool the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001892 *
Lai Jiangshan051e1852014-07-22 13:03:02 +08001893 * Create and start a new worker which is attached to @pool.
Tejun Heoc34056a2010-06-29 10:07:11 +02001894 *
1895 * CONTEXT:
1896 * Might sleep. Does GFP_KERNEL allocations.
1897 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001898 * Return:
Tejun Heoc34056a2010-06-29 10:07:11 +02001899 * Pointer to the newly created worker.
1900 */
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001901static struct worker *create_worker(struct worker_pool *pool)
Tejun Heoc34056a2010-06-29 10:07:11 +02001902{
Tejun Heoc34056a2010-06-29 10:07:11 +02001903 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001904 int id = -1;
Tejun Heoe3c916a2013-04-01 11:23:32 -07001905 char id_buf[16];
Tejun Heoc34056a2010-06-29 10:07:11 +02001906
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001907 /* ID is needed to determine kthread name */
1908 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
Tejun Heo822d8402013-03-19 13:45:21 -07001909 if (id < 0)
1910 goto fail;
Tejun Heoc34056a2010-06-29 10:07:11 +02001911
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001912 worker = alloc_worker(pool->node);
Tejun Heoc34056a2010-06-29 10:07:11 +02001913 if (!worker)
1914 goto fail;
1915
Tejun Heoc34056a2010-06-29 10:07:11 +02001916 worker->id = id;
1917
Tejun Heo29c91e92013-03-12 11:30:03 -07001918 if (pool->cpu >= 0)
Tejun Heoe3c916a2013-04-01 11:23:32 -07001919 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1920 pool->attrs->nice < 0 ? "H" : "");
Tejun Heof3421792010-07-02 10:03:51 +02001921 else
Tejun Heoe3c916a2013-04-01 11:23:32 -07001922 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1923
Tejun Heof3f90ad2013-04-01 11:23:34 -07001924 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
Tejun Heoe3c916a2013-04-01 11:23:32 -07001925 "kworker/%s", id_buf);
Tejun Heoc34056a2010-06-29 10:07:11 +02001926 if (IS_ERR(worker->task))
1927 goto fail;
1928
Oleg Nesterov91151222013-11-14 12:56:18 +01001929 set_user_nice(worker->task, pool->attrs->nice);
Peter Zijlstra25834c72015-05-15 17:43:34 +02001930 kthread_bind_mask(worker->task, pool->attrs->cpumask);
Oleg Nesterov91151222013-11-14 12:56:18 +01001931
Lai Jiangshanda028462014-05-20 17:46:31 +08001932 /* successful, attach the worker to the pool */
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001933 worker_attach_to_pool(worker, pool);
Tejun Heo822d8402013-03-19 13:45:21 -07001934
Lai Jiangshan051e1852014-07-22 13:03:02 +08001935 /* start the newly created worker */
1936 spin_lock_irq(&pool->lock);
1937 worker->pool->nr_workers++;
1938 worker_enter_idle(worker);
1939 wake_up_process(worker->task);
1940 spin_unlock_irq(&pool->lock);
1941
Tejun Heoc34056a2010-06-29 10:07:11 +02001942 return worker;
Tejun Heo822d8402013-03-19 13:45:21 -07001943
Tejun Heoc34056a2010-06-29 10:07:11 +02001944fail:
Lai Jiangshan9625ab12014-05-20 17:46:27 +08001945 if (id >= 0)
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001946 ida_simple_remove(&pool->worker_ida, id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001947 kfree(worker);
1948 return NULL;
1949}
1950
1951/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001952 * destroy_worker - destroy a workqueue worker
1953 * @worker: worker to be destroyed
1954 *
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001955 * Destroy @worker and adjust @pool stats accordingly. The worker should
1956 * be idle.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001957 *
1958 * CONTEXT:
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001959 * spin_lock_irq(pool->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001960 */
1961static void destroy_worker(struct worker *worker)
1962{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001963 struct worker_pool *pool = worker->pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001964
Tejun Heocd549682013-03-13 19:47:39 -07001965 lockdep_assert_held(&pool->lock);
1966
Tejun Heoc34056a2010-06-29 10:07:11 +02001967 /* sanity check frenzy */
Tejun Heo6183c002013-03-12 11:29:57 -07001968 if (WARN_ON(worker->current_work) ||
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001969 WARN_ON(!list_empty(&worker->scheduled)) ||
1970 WARN_ON(!(worker->flags & WORKER_IDLE)))
Tejun Heo6183c002013-03-12 11:29:57 -07001971 return;
Tejun Heoc34056a2010-06-29 10:07:11 +02001972
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001973 pool->nr_workers--;
1974 pool->nr_idle--;
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001975
Tejun Heoc8e55f32010-06-29 10:07:12 +02001976 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001977 worker->flags |= WORKER_DIE;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001978 wake_up_process(worker->task);
Tejun Heoc34056a2010-06-29 10:07:11 +02001979}
1980
Kees Cook32a6c722017-10-16 15:58:25 -07001981static void idle_worker_timeout(struct timer_list *t)
Tejun Heoe22bee72010-06-29 10:07:14 +02001982{
Kees Cook32a6c722017-10-16 15:58:25 -07001983 struct worker_pool *pool = from_timer(pool, t, idle_timer);
Tejun Heoe22bee72010-06-29 10:07:14 +02001984
Tejun Heod565ed62013-01-24 11:01:33 -08001985 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001986
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001987 while (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001988 struct worker *worker;
1989 unsigned long expires;
1990
1991 /* idle_list is kept in LIFO order, check the last one */
Tejun Heo63d95a92012-07-12 14:46:37 -07001992 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02001993 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1994
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001995 if (time_before(jiffies, expires)) {
Tejun Heo63d95a92012-07-12 14:46:37 -07001996 mod_timer(&pool->idle_timer, expires);
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001997 break;
Tejun Heoe22bee72010-06-29 10:07:14 +02001998 }
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001999
2000 destroy_worker(worker);
Tejun Heoe22bee72010-06-29 10:07:14 +02002001 }
2002
Tejun Heod565ed62013-01-24 11:01:33 -08002003 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002004}
2005
Tejun Heo493a1722013-03-12 11:29:59 -07002006static void send_mayday(struct work_struct *work)
Tejun Heoe22bee72010-06-29 10:07:14 +02002007{
Tejun Heo112202d2013-02-13 19:29:12 -08002008 struct pool_workqueue *pwq = get_work_pwq(work);
2009 struct workqueue_struct *wq = pwq->wq;
Tejun Heo493a1722013-03-12 11:29:59 -07002010
Tejun Heo2e109a22013-03-13 19:47:40 -07002011 lockdep_assert_held(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002012
Tejun Heo493008a2013-03-12 11:30:03 -07002013 if (!wq->rescuer)
Tejun Heo493a1722013-03-12 11:29:59 -07002014 return;
Tejun Heoe22bee72010-06-29 10:07:14 +02002015
2016 /* mayday mayday mayday */
Tejun Heo493a1722013-03-12 11:29:59 -07002017 if (list_empty(&pwq->mayday_node)) {
Lai Jiangshan77668c82014-04-18 11:04:16 -04002018 /*
2019 * If @pwq is for an unbound wq, its base ref may be put at
2020 * any time due to an attribute change. Pin @pwq until the
2021 * rescuer is done with it.
2022 */
2023 get_pwq(pwq);
Tejun Heo493a1722013-03-12 11:29:59 -07002024 list_add_tail(&pwq->mayday_node, &wq->maydays);
Tejun Heoe22bee72010-06-29 10:07:14 +02002025 wake_up_process(wq->rescuer->task);
Tejun Heo493a1722013-03-12 11:29:59 -07002026 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002027}
2028
Kees Cook32a6c722017-10-16 15:58:25 -07002029static void pool_mayday_timeout(struct timer_list *t)
Tejun Heoe22bee72010-06-29 10:07:14 +02002030{
Kees Cook32a6c722017-10-16 15:58:25 -07002031 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
Tejun Heoe22bee72010-06-29 10:07:14 +02002032 struct work_struct *work;
2033
Tejun Heob2d82902014-12-08 12:39:16 -05002034 spin_lock_irq(&pool->lock);
2035 spin_lock(&wq_mayday_lock); /* for wq->maydays */
Tejun Heoe22bee72010-06-29 10:07:14 +02002036
Tejun Heo63d95a92012-07-12 14:46:37 -07002037 if (need_to_create_worker(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02002038 /*
2039 * We've been trying to create a new worker but
2040 * haven't been successful. We might be hitting an
2041 * allocation deadlock. Send distress signals to
2042 * rescuers.
2043 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002044 list_for_each_entry(work, &pool->worklist, entry)
Tejun Heoe22bee72010-06-29 10:07:14 +02002045 send_mayday(work);
2046 }
2047
Tejun Heob2d82902014-12-08 12:39:16 -05002048 spin_unlock(&wq_mayday_lock);
2049 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002050
Tejun Heo63d95a92012-07-12 14:46:37 -07002051 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
Tejun Heoe22bee72010-06-29 10:07:14 +02002052}
2053
2054/**
2055 * maybe_create_worker - create a new worker if necessary
Tejun Heo63d95a92012-07-12 14:46:37 -07002056 * @pool: pool to create a new worker for
Tejun Heoe22bee72010-06-29 10:07:14 +02002057 *
Tejun Heo63d95a92012-07-12 14:46:37 -07002058 * Create a new worker for @pool if necessary. @pool is guaranteed to
Tejun Heoe22bee72010-06-29 10:07:14 +02002059 * have at least one idle worker on return from this function. If
2060 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
Tejun Heo63d95a92012-07-12 14:46:37 -07002061 * sent to all rescuers with works scheduled on @pool to resolve
Tejun Heoe22bee72010-06-29 10:07:14 +02002062 * possible allocation deadlock.
2063 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002064 * On return, need_to_create_worker() is guaranteed to be %false and
2065 * may_start_working() %true.
Tejun Heoe22bee72010-06-29 10:07:14 +02002066 *
2067 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08002068 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02002069 * multiple times. Does GFP_KERNEL allocations. Called only from
2070 * manager.
Tejun Heoe22bee72010-06-29 10:07:14 +02002071 */
Tejun Heo29187a92015-01-16 14:21:16 -05002072static void maybe_create_worker(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08002073__releases(&pool->lock)
2074__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02002075{
Tejun Heoe22bee72010-06-29 10:07:14 +02002076restart:
Tejun Heod565ed62013-01-24 11:01:33 -08002077 spin_unlock_irq(&pool->lock);
Tejun Heo9f9c236442010-07-14 11:31:20 +02002078
Tejun Heoe22bee72010-06-29 10:07:14 +02002079 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
Tejun Heo63d95a92012-07-12 14:46:37 -07002080 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
Tejun Heoe22bee72010-06-29 10:07:14 +02002081
2082 while (true) {
Lai Jiangshan051e1852014-07-22 13:03:02 +08002083 if (create_worker(pool) || !need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002084 break;
2085
Lai Jiangshane212f362014-06-03 15:32:17 +08002086 schedule_timeout_interruptible(CREATE_COOLDOWN);
Tejun Heo9f9c236442010-07-14 11:31:20 +02002087
Tejun Heo63d95a92012-07-12 14:46:37 -07002088 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002089 break;
2090 }
2091
Tejun Heo63d95a92012-07-12 14:46:37 -07002092 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08002093 spin_lock_irq(&pool->lock);
Lai Jiangshan051e1852014-07-22 13:03:02 +08002094 /*
2095 * This is necessary even after a new worker was just successfully
2096 * created as @pool->lock was dropped and the new worker might have
2097 * already become busy.
2098 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002099 if (need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002100 goto restart;
Tejun Heoe22bee72010-06-29 10:07:14 +02002101}
2102
2103/**
Tejun Heoe22bee72010-06-29 10:07:14 +02002104 * manage_workers - manage worker pool
2105 * @worker: self
2106 *
Tejun Heo706026c2013-01-24 11:01:34 -08002107 * Assume the manager role and manage the worker pool @worker belongs
Tejun Heoe22bee72010-06-29 10:07:14 +02002108 * to. At any given time, there can be only zero or one manager per
Tejun Heo706026c2013-01-24 11:01:34 -08002109 * pool. The exclusion is handled automatically by this function.
Tejun Heoe22bee72010-06-29 10:07:14 +02002110 *
2111 * The caller can safely start processing works on false return. On
2112 * true return, it's guaranteed that need_to_create_worker() is false
2113 * and may_start_working() is true.
2114 *
2115 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002116 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02002117 * multiple times. Does GFP_KERNEL allocations.
2118 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002119 * Return:
Tejun Heo29187a92015-01-16 14:21:16 -05002120 * %false if the pool doesn't need management and the caller can safely
2121 * start processing works, %true if management function was performed and
2122 * the conditions that the caller verified before calling the function may
2123 * no longer be true.
Tejun Heoe22bee72010-06-29 10:07:14 +02002124 */
2125static bool manage_workers(struct worker *worker)
2126{
Tejun Heo63d95a92012-07-12 14:46:37 -07002127 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002128
Tejun Heo692b4822017-10-09 08:04:13 -07002129 if (pool->flags & POOL_MANAGER_ACTIVE)
Tejun Heo29187a92015-01-16 14:21:16 -05002130 return false;
Tejun Heo692b4822017-10-09 08:04:13 -07002131
2132 pool->flags |= POOL_MANAGER_ACTIVE;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002133 pool->manager = worker;
Tejun Heoe22bee72010-06-29 10:07:14 +02002134
Tejun Heo29187a92015-01-16 14:21:16 -05002135 maybe_create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02002136
Tejun Heo2607d7a2015-03-09 09:22:28 -04002137 pool->manager = NULL;
Tejun Heo692b4822017-10-09 08:04:13 -07002138 pool->flags &= ~POOL_MANAGER_ACTIVE;
2139 wake_up(&wq_manager_wait);
Tejun Heo29187a92015-01-16 14:21:16 -05002140 return true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002141}
2142
Tejun Heoa62428c2010-06-29 10:07:10 +02002143/**
2144 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02002145 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02002146 * @work: work to process
2147 *
2148 * Process @work. This function contains all the logics necessary to
2149 * process a single work including synchronization against and
2150 * interaction with other workers on the same cpu, queueing and
2151 * flushing. As long as context requirement is met, any worker can
2152 * call this function to process a work.
2153 *
2154 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002155 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02002156 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002157static void process_one_work(struct worker *worker, struct work_struct *work)
Tejun Heod565ed62013-01-24 11:01:33 -08002158__releases(&pool->lock)
2159__acquires(&pool->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02002160{
Tejun Heo112202d2013-02-13 19:29:12 -08002161 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobd7bdd42012-07-12 14:46:37 -07002162 struct worker_pool *pool = worker->pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002163 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heo73f53c42010-06-29 10:07:11 +02002164 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02002165 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02002166#ifdef CONFIG_LOCKDEP
2167 /*
2168 * It is permissible to free the struct work_struct from
2169 * inside the function that is called from it, this we need to
2170 * take into account for lockdep too. To avoid bogus "held
2171 * lock freed" warnings as well as problems when looking into
2172 * work->lockdep_map, make a copy and use that here.
2173 */
Peter Zijlstra4d82a1d2012-05-15 08:06:19 -07002174 struct lockdep_map lockdep_map;
2175
2176 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002177#endif
Lai Jiangshan807407c2014-06-03 15:33:28 +08002178 /* ensure we're on the correct CPU */
Lai Jiangshan85327af2014-06-03 15:33:28 +08002179 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heoec22ca52013-01-24 11:01:33 -08002180 raw_smp_processor_id() != pool->cpu);
Tejun Heo25511a42012-07-17 12:39:27 -07002181
Tejun Heo7e116292010-06-29 10:07:13 +02002182 /*
2183 * A single work shouldn't be executed concurrently by
2184 * multiple workers on a single cpu. Check whether anyone is
2185 * already processing the work. If so, defer the work to the
2186 * currently executing one.
2187 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002188 collision = find_worker_executing_work(pool, work);
Tejun Heo7e116292010-06-29 10:07:13 +02002189 if (unlikely(collision)) {
2190 move_linked_works(work, &collision->scheduled, NULL);
2191 return;
2192 }
2193
Tejun Heo8930cab2012-08-03 10:30:45 -07002194 /* claim and dequeue */
Tejun Heoa62428c2010-06-29 10:07:10 +02002195 debug_work_deactivate(work);
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002196 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
Tejun Heoc34056a2010-06-29 10:07:11 +02002197 worker->current_work = work;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002198 worker->current_func = work->func;
Tejun Heo112202d2013-02-13 19:29:12 -08002199 worker->current_pwq = pwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002200 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002201
Tejun Heo8bf89592018-05-18 08:47:13 -07002202 /*
2203 * Record wq name for cmdline and debug reporting, may get
2204 * overridden through set_worker_desc().
2205 */
2206 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2207
Tejun Heoa62428c2010-06-29 10:07:10 +02002208 list_del_init(&work->entry);
2209
Tejun Heo649027d2010-06-29 10:07:14 +02002210 /*
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002211 * CPU intensive works don't participate in concurrency management.
2212 * They're the scheduler's responsibility. This takes @worker out
2213 * of concurrency management and the next code block will chain
2214 * execution of the pending work items.
Tejun Heofb0e7be2010-06-29 10:07:15 +02002215 */
2216 if (unlikely(cpu_intensive))
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002217 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
Tejun Heofb0e7be2010-06-29 10:07:15 +02002218
Tejun Heo974271c2012-07-12 14:46:37 -07002219 /*
Lai Jiangshana489a032014-07-22 13:01:59 +08002220 * Wake up another worker if necessary. The condition is always
2221 * false for normal per-cpu workers since nr_running would always
2222 * be >= 1 at this point. This is used to chain execution of the
2223 * pending work items for WORKER_NOT_RUNNING workers such as the
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002224 * UNBOUND and CPU_INTENSIVE ones.
Tejun Heo974271c2012-07-12 14:46:37 -07002225 */
Lai Jiangshana489a032014-07-22 13:01:59 +08002226 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002227 wake_up_worker(pool);
Tejun Heo974271c2012-07-12 14:46:37 -07002228
Tejun Heo8930cab2012-08-03 10:30:45 -07002229 /*
Tejun Heo7c3eed52013-01-24 11:01:33 -08002230 * Record the last pool and clear PENDING which should be the last
Tejun Heod565ed62013-01-24 11:01:33 -08002231 * update to @work. Also, do this inside @pool->lock so that
Tejun Heo23657bb2012-08-13 17:08:19 -07002232 * PENDING and queued state changes happen together while IRQ is
2233 * disabled.
Tejun Heo8930cab2012-08-03 10:30:45 -07002234 */
Tejun Heo7c3eed52013-01-24 11:01:33 -08002235 set_work_pool_and_clear_pending(work, pool->id);
Tejun Heoa62428c2010-06-29 10:07:10 +02002236
Tejun Heod565ed62013-01-24 11:01:33 -08002237 spin_unlock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002238
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002239 lock_map_acquire(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002240 lock_map_acquire(&lockdep_map);
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002241 /*
Peter Zijlstraf52be572017-08-29 10:59:39 +02002242 * Strictly speaking we should mark the invariant state without holding
2243 * any locks, that is, before these two lock_map_acquire()'s.
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002244 *
2245 * However, that would result in:
2246 *
2247 * A(W1)
2248 * WFC(C)
2249 * A(W1)
2250 * C(C)
2251 *
2252 * Which would create W1->C->W1 dependencies, even though there is no
2253 * actual deadlock possible. There are two solutions, using a
2254 * read-recursive acquire on the work(queue) 'locks', but this will then
Peter Zijlstraf52be572017-08-29 10:59:39 +02002255 * hit the lockdep limitation on recursive locks, or simply discard
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002256 * these locks.
2257 *
2258 * AFAICT there is no possible deadlock scenario between the
2259 * flush_work() and complete() primitives (except for single-threaded
2260 * workqueues), so hiding them isn't a problem.
2261 */
Peter Zijlstraf52be572017-08-29 10:59:39 +02002262 lockdep_invariant_state(true);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002263 trace_workqueue_execute_start(work);
Tejun Heoa2c1c572012-12-18 10:35:02 -08002264 worker->current_func(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002265 /*
2266 * While we must be careful to not use "work" after this, the trace
2267 * point will only record its address.
2268 */
2269 trace_workqueue_execute_end(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02002270 lock_map_release(&lockdep_map);
Tejun Heo112202d2013-02-13 19:29:12 -08002271 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002272
2273 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
Valentin Ilie044c7822012-08-19 00:52:42 +03002274 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
Sakari Ailusd75f7732019-03-25 21:32:28 +02002275 " last function: %ps\n",
Tejun Heoa2c1c572012-12-18 10:35:02 -08002276 current->comm, preempt_count(), task_pid_nr(current),
2277 worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002278 debug_show_held_locks(current);
2279 dump_stack();
2280 }
2281
Tejun Heob22ce272013-08-28 17:33:37 -04002282 /*
Sebastian Andrzej Siewior025f50f2019-10-15 21:18:21 +02002283 * The following prevents a kworker from hogging CPU on !PREEMPTION
Tejun Heob22ce272013-08-28 17:33:37 -04002284 * kernels, where a requeueing work item waiting for something to
2285 * happen could deadlock with stop_machine as such work item could
2286 * indefinitely requeue itself while all other CPUs are trapped in
Joe Lawrence789cbbe2014-10-05 13:24:21 -04002287 * stop_machine. At the same time, report a quiescent RCU state so
2288 * the same condition doesn't freeze RCU.
Tejun Heob22ce272013-08-28 17:33:37 -04002289 */
Paul E. McKenneya7e64252017-10-24 08:25:02 -07002290 cond_resched();
Tejun Heob22ce272013-08-28 17:33:37 -04002291
Tejun Heod565ed62013-01-24 11:01:33 -08002292 spin_lock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002293
Tejun Heofb0e7be2010-06-29 10:07:15 +02002294 /* clear cpu intensive status */
2295 if (unlikely(cpu_intensive))
2296 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2297
Johannes Weiner1b69ac62019-02-01 14:20:42 -08002298 /* tag the worker for identification in schedule() */
2299 worker->last_func = worker->current_func;
2300
Tejun Heoa62428c2010-06-29 10:07:10 +02002301 /* we're done with it, release */
Sasha Levin42f85702012-12-17 10:01:23 -05002302 hash_del(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002303 worker->current_work = NULL;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002304 worker->current_func = NULL;
Tejun Heo112202d2013-02-13 19:29:12 -08002305 worker->current_pwq = NULL;
2306 pwq_dec_nr_in_flight(pwq, work_color);
Tejun Heoa62428c2010-06-29 10:07:10 +02002307}
2308
Tejun Heoaffee4b2010-06-29 10:07:12 +02002309/**
2310 * process_scheduled_works - process scheduled works
2311 * @worker: self
2312 *
2313 * Process all scheduled works. Please note that the scheduled list
2314 * may change while processing a work, so this function repeatedly
2315 * fetches a work from the top and executes it.
2316 *
2317 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002318 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02002319 * multiple times.
2320 */
2321static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002323 while (!list_empty(&worker->scheduled)) {
2324 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002326 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328}
2329
Tejun Heo197f6ac2018-05-21 08:04:35 -07002330static void set_pf_worker(bool val)
2331{
2332 mutex_lock(&wq_pool_attach_mutex);
2333 if (val)
2334 current->flags |= PF_WQ_WORKER;
2335 else
2336 current->flags &= ~PF_WQ_WORKER;
2337 mutex_unlock(&wq_pool_attach_mutex);
2338}
2339
Tejun Heo4690c4a2010-06-29 10:07:10 +02002340/**
2341 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02002342 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02002343 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002344 * The worker thread function. All workers belong to a worker_pool -
2345 * either a per-cpu one or dynamic unbound one. These workers process all
2346 * work items regardless of their specific target workqueue. The only
2347 * exception is work items which belong to workqueues with a rescuer which
2348 * will be explained in rescuer_thread().
Yacine Belkadid185af32013-07-31 14:59:24 -07002349 *
2350 * Return: 0
Tejun Heo4690c4a2010-06-29 10:07:10 +02002351 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002352static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353{
Tejun Heoc34056a2010-06-29 10:07:11 +02002354 struct worker *worker = __worker;
Tejun Heobd7bdd42012-07-12 14:46:37 -07002355 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356
Tejun Heoe22bee72010-06-29 10:07:14 +02002357 /* tell the scheduler that this is a workqueue worker */
Tejun Heo197f6ac2018-05-21 08:04:35 -07002358 set_pf_worker(true);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002359woke_up:
Tejun Heod565ed62013-01-24 11:01:33 -08002360 spin_lock_irq(&pool->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002361
Tejun Heoa9ab7752013-03-19 13:45:21 -07002362 /* am I supposed to die? */
2363 if (unlikely(worker->flags & WORKER_DIE)) {
Tejun Heod565ed62013-01-24 11:01:33 -08002364 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002365 WARN_ON_ONCE(!list_empty(&worker->entry));
Tejun Heo197f6ac2018-05-21 08:04:35 -07002366 set_pf_worker(false);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002367
2368 set_task_comm(worker->task, "kworker/dying");
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08002369 ida_simple_remove(&pool->worker_ida, worker->id);
Tejun Heoa2d812a2018-05-18 08:47:13 -07002370 worker_detach_from_pool(worker);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002371 kfree(worker);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002372 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07002374
Tejun Heoc8e55f32010-06-29 10:07:12 +02002375 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02002376recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02002377 /* no more worker necessary? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002378 if (!need_more_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002379 goto sleep;
2380
2381 /* do we need to manage? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002382 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002383 goto recheck;
2384
Tejun Heoc8e55f32010-06-29 10:07:12 +02002385 /*
2386 * ->scheduled list can only be filled while a worker is
2387 * preparing to process a work or actually processing it.
2388 * Make sure nobody diddled with it while I was sleeping.
2389 */
Tejun Heo6183c002013-03-12 11:29:57 -07002390 WARN_ON_ONCE(!list_empty(&worker->scheduled));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002391
Tejun Heoe22bee72010-06-29 10:07:14 +02002392 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07002393 * Finish PREP stage. We're guaranteed to have at least one idle
2394 * worker or that someone else has already assumed the manager
2395 * role. This is where @worker starts participating in concurrency
2396 * management if applicable and concurrency management is restored
2397 * after being rebound. See rebind_workers() for details.
Tejun Heoe22bee72010-06-29 10:07:14 +02002398 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07002399 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
Tejun Heoe22bee72010-06-29 10:07:14 +02002400
2401 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02002402 struct work_struct *work =
Tejun Heobd7bdd42012-07-12 14:46:37 -07002403 list_first_entry(&pool->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02002404 struct work_struct, entry);
2405
Tejun Heo82607adc2015-12-08 11:28:04 -05002406 pool->watchdog_ts = jiffies;
2407
Tejun Heoc8e55f32010-06-29 10:07:12 +02002408 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2409 /* optimization path, not strictly necessary */
2410 process_one_work(worker, work);
2411 if (unlikely(!list_empty(&worker->scheduled)))
2412 process_scheduled_works(worker);
2413 } else {
2414 move_linked_works(work, &worker->scheduled, NULL);
2415 process_scheduled_works(worker);
2416 }
Tejun Heo63d95a92012-07-12 14:46:37 -07002417 } while (keep_working(pool));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002418
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002419 worker_set_flags(worker, WORKER_PREP);
Tejun Heod313dd82010-07-02 10:03:51 +02002420sleep:
Tejun Heoc8e55f32010-06-29 10:07:12 +02002421 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002422 * pool->lock is held and there's no work to process and no need to
2423 * manage, sleep. Workers are woken up only while holding
2424 * pool->lock or from local cpu, so setting the current state
2425 * before releasing pool->lock is enough to prevent losing any
2426 * event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02002427 */
2428 worker_enter_idle(worker);
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002429 __set_current_state(TASK_IDLE);
Tejun Heod565ed62013-01-24 11:01:33 -08002430 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002431 schedule();
2432 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433}
2434
Tejun Heoe22bee72010-06-29 10:07:14 +02002435/**
2436 * rescuer_thread - the rescuer thread function
Tejun Heo111c2252013-01-17 17:16:24 -08002437 * @__rescuer: self
Tejun Heoe22bee72010-06-29 10:07:14 +02002438 *
2439 * Workqueue rescuer thread function. There's one rescuer for each
Tejun Heo493008a2013-03-12 11:30:03 -07002440 * workqueue which has WQ_MEM_RECLAIM set.
Tejun Heoe22bee72010-06-29 10:07:14 +02002441 *
Tejun Heo706026c2013-01-24 11:01:34 -08002442 * Regular work processing on a pool may block trying to create a new
Tejun Heoe22bee72010-06-29 10:07:14 +02002443 * worker which uses GFP_KERNEL allocation which has slight chance of
2444 * developing into deadlock if some works currently on the same queue
2445 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2446 * the problem rescuer solves.
2447 *
Tejun Heo706026c2013-01-24 11:01:34 -08002448 * When such condition is possible, the pool summons rescuers of all
2449 * workqueues which have works queued on the pool and let them process
Tejun Heoe22bee72010-06-29 10:07:14 +02002450 * those works so that forward progress can be guaranteed.
2451 *
2452 * This should happen rarely.
Yacine Belkadid185af32013-07-31 14:59:24 -07002453 *
2454 * Return: 0
Tejun Heoe22bee72010-06-29 10:07:14 +02002455 */
Tejun Heo111c2252013-01-17 17:16:24 -08002456static int rescuer_thread(void *__rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02002457{
Tejun Heo111c2252013-01-17 17:16:24 -08002458 struct worker *rescuer = __rescuer;
2459 struct workqueue_struct *wq = rescuer->rescue_wq;
Tejun Heoe22bee72010-06-29 10:07:14 +02002460 struct list_head *scheduled = &rescuer->scheduled;
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002461 bool should_stop;
Tejun Heoe22bee72010-06-29 10:07:14 +02002462
2463 set_user_nice(current, RESCUER_NICE_LEVEL);
Tejun Heo111c2252013-01-17 17:16:24 -08002464
2465 /*
2466 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2467 * doesn't participate in concurrency management.
2468 */
Tejun Heo197f6ac2018-05-21 08:04:35 -07002469 set_pf_worker(true);
Tejun Heoe22bee72010-06-29 10:07:14 +02002470repeat:
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002471 set_current_state(TASK_IDLE);
Tejun Heoe22bee72010-06-29 10:07:14 +02002472
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002473 /*
2474 * By the time the rescuer is requested to stop, the workqueue
2475 * shouldn't have any work pending, but @wq->maydays may still have
2476 * pwq(s) queued. This can happen by non-rescuer workers consuming
2477 * all the work items before the rescuer got to them. Go through
2478 * @wq->maydays processing before acting on should_stop so that the
2479 * list is always empty on exit.
2480 */
2481 should_stop = kthread_should_stop();
Tejun Heoe22bee72010-06-29 10:07:14 +02002482
Tejun Heo493a1722013-03-12 11:29:59 -07002483 /* see whether any pwq is asking for help */
Tejun Heo2e109a22013-03-13 19:47:40 -07002484 spin_lock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002485
2486 while (!list_empty(&wq->maydays)) {
2487 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2488 struct pool_workqueue, mayday_node);
Tejun Heo112202d2013-02-13 19:29:12 -08002489 struct worker_pool *pool = pwq->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002490 struct work_struct *work, *n;
Tejun Heo82607adc2015-12-08 11:28:04 -05002491 bool first = true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002492
2493 __set_current_state(TASK_RUNNING);
Tejun Heo493a1722013-03-12 11:29:59 -07002494 list_del_init(&pwq->mayday_node);
2495
Tejun Heo2e109a22013-03-13 19:47:40 -07002496 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002497
Lai Jiangshan51697d392014-05-20 17:46:36 +08002498 worker_attach_to_pool(rescuer, pool);
2499
2500 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002501
2502 /*
2503 * Slurp in all works issued via this workqueue and
2504 * process'em.
2505 */
Tejun Heo0479c8c2014-12-04 10:14:13 -05002506 WARN_ON_ONCE(!list_empty(scheduled));
Tejun Heo82607adc2015-12-08 11:28:04 -05002507 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2508 if (get_work_pwq(work) == pwq) {
2509 if (first)
2510 pool->watchdog_ts = jiffies;
Tejun Heoe22bee72010-06-29 10:07:14 +02002511 move_linked_works(work, scheduled, &n);
Tejun Heo82607adc2015-12-08 11:28:04 -05002512 }
2513 first = false;
2514 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002515
NeilBrown008847f2014-12-08 12:39:16 -05002516 if (!list_empty(scheduled)) {
2517 process_scheduled_works(rescuer);
2518
2519 /*
2520 * The above execution of rescued work items could
2521 * have created more to rescue through
2522 * pwq_activate_first_delayed() or chained
2523 * queueing. Let's put @pwq back on mayday list so
2524 * that such back-to-back work items, which may be
2525 * being used to relieve memory pressure, don't
2526 * incur MAYDAY_INTERVAL delay inbetween.
2527 */
2528 if (need_to_create_worker(pool)) {
2529 spin_lock(&wq_mayday_lock);
Tejun Heoe66b39a2019-09-25 06:59:15 -07002530 /*
2531 * Queue iff we aren't racing destruction
2532 * and somebody else hasn't queued it already.
2533 */
2534 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2535 get_pwq(pwq);
2536 list_add_tail(&pwq->mayday_node, &wq->maydays);
2537 }
NeilBrown008847f2014-12-08 12:39:16 -05002538 spin_unlock(&wq_mayday_lock);
2539 }
2540 }
Tejun Heo75769582011-02-14 14:04:46 +01002541
2542 /*
Lai Jiangshan77668c82014-04-18 11:04:16 -04002543 * Put the reference grabbed by send_mayday(). @pool won't
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002544 * go away while we're still attached to it.
Lai Jiangshan77668c82014-04-18 11:04:16 -04002545 */
2546 put_pwq(pwq);
2547
2548 /*
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002549 * Leave this pool. If need_more_worker() is %true, notify a
Tejun Heo75769582011-02-14 14:04:46 +01002550 * regular worker; otherwise, we end up with 0 concurrency
2551 * and stalling the execution.
2552 */
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002553 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002554 wake_up_worker(pool);
Tejun Heo75769582011-02-14 14:04:46 +01002555
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002556 spin_unlock_irq(&pool->lock);
2557
Tejun Heoa2d812a2018-05-18 08:47:13 -07002558 worker_detach_from_pool(rescuer);
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002559
2560 spin_lock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002561 }
2562
Tejun Heo2e109a22013-03-13 19:47:40 -07002563 spin_unlock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002564
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002565 if (should_stop) {
2566 __set_current_state(TASK_RUNNING);
Tejun Heo197f6ac2018-05-21 08:04:35 -07002567 set_pf_worker(false);
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002568 return 0;
2569 }
2570
Tejun Heo111c2252013-01-17 17:16:24 -08002571 /* rescuers should never participate in concurrency management */
2572 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
Tejun Heoe22bee72010-06-29 10:07:14 +02002573 schedule();
2574 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575}
2576
Tejun Heofca839c2015-12-07 10:58:57 -05002577/**
2578 * check_flush_dependency - check for flush dependency sanity
2579 * @target_wq: workqueue being flushed
2580 * @target_work: work item being flushed (NULL for workqueue flushes)
2581 *
2582 * %current is trying to flush the whole @target_wq or @target_work on it.
2583 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2584 * reclaiming memory or running on a workqueue which doesn't have
2585 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2586 * a deadlock.
2587 */
2588static void check_flush_dependency(struct workqueue_struct *target_wq,
2589 struct work_struct *target_work)
2590{
2591 work_func_t target_func = target_work ? target_work->func : NULL;
2592 struct worker *worker;
2593
2594 if (target_wq->flags & WQ_MEM_RECLAIM)
2595 return;
2596
2597 worker = current_wq_worker();
2598
2599 WARN_ONCE(current->flags & PF_MEMALLOC,
Sakari Ailusd75f7732019-03-25 21:32:28 +02002600 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
Tejun Heofca839c2015-12-07 10:58:57 -05002601 current->pid, current->comm, target_wq->name, target_func);
Tejun Heo23d11a52016-01-29 05:59:46 -05002602 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2603 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
Sakari Ailusd75f7732019-03-25 21:32:28 +02002604 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
Tejun Heofca839c2015-12-07 10:58:57 -05002605 worker->current_pwq->wq->name, worker->current_func,
2606 target_wq->name, target_func);
2607}
2608
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002609struct wq_barrier {
2610 struct work_struct work;
2611 struct completion done;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002612 struct task_struct *task; /* purely informational */
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002613};
2614
2615static void wq_barrier_func(struct work_struct *work)
2616{
2617 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2618 complete(&barr->done);
2619}
2620
Tejun Heo4690c4a2010-06-29 10:07:10 +02002621/**
2622 * insert_wq_barrier - insert a barrier work
Tejun Heo112202d2013-02-13 19:29:12 -08002623 * @pwq: pwq to insert barrier into
Tejun Heo4690c4a2010-06-29 10:07:10 +02002624 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002625 * @target: target work to attach @barr to
2626 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002627 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002628 * @barr is linked to @target such that @barr is completed only after
2629 * @target finishes execution. Please note that the ordering
2630 * guarantee is observed only with respect to @target and on the local
2631 * cpu.
2632 *
2633 * Currently, a queued barrier can't be canceled. This is because
2634 * try_to_grab_pending() can't determine whether the work to be
2635 * grabbed is at the head of the queue and thus can't clear LINKED
2636 * flag of the previous work while there must be a valid next work
2637 * after a work with LINKED flag set.
2638 *
2639 * Note that when @worker is non-NULL, @target may be modified
Tejun Heo112202d2013-02-13 19:29:12 -08002640 * underneath us, so we can't reliably determine pwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002641 *
2642 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002643 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002644 */
Tejun Heo112202d2013-02-13 19:29:12 -08002645static void insert_wq_barrier(struct pool_workqueue *pwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002646 struct wq_barrier *barr,
2647 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002648{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002649 struct list_head *head;
2650 unsigned int linked = 0;
2651
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002652 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002653 * debugobject calls are safe here even with pool->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002654 * as we know for sure that this will not trigger any of the
2655 * checks and call back into the fixup functions where we
2656 * might deadlock.
2657 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002658 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002659 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Boqun Feng52fa5bc2017-08-17 17:46:12 +08002660
Byungchul Parkfd1a5b02017-10-25 17:56:04 +09002661 init_completion_map(&barr->done, &target->lockdep_map);
2662
Tejun Heo2607d7a2015-03-09 09:22:28 -04002663 barr->task = current;
Oleg Nesterov83c22522007-05-09 02:33:54 -07002664
Tejun Heoaffee4b2010-06-29 10:07:12 +02002665 /*
2666 * If @target is currently being executed, schedule the
2667 * barrier to the worker; otherwise, put it after @target.
2668 */
2669 if (worker)
2670 head = worker->scheduled.next;
2671 else {
2672 unsigned long *bits = work_data_bits(target);
2673
2674 head = target->entry.next;
2675 /* there can already be other linked works, inherit and set */
2676 linked = *bits & WORK_STRUCT_LINKED;
2677 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2678 }
2679
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002680 debug_work_activate(&barr->work);
Tejun Heo112202d2013-02-13 19:29:12 -08002681 insert_work(pwq, &barr->work, head,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002682 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002683}
2684
Tejun Heo73f53c42010-06-29 10:07:11 +02002685/**
Tejun Heo112202d2013-02-13 19:29:12 -08002686 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
Tejun Heo73f53c42010-06-29 10:07:11 +02002687 * @wq: workqueue being flushed
2688 * @flush_color: new flush color, < 0 for no-op
2689 * @work_color: new work color, < 0 for no-op
2690 *
Tejun Heo112202d2013-02-13 19:29:12 -08002691 * Prepare pwqs for workqueue flushing.
Tejun Heo73f53c42010-06-29 10:07:11 +02002692 *
Tejun Heo112202d2013-02-13 19:29:12 -08002693 * If @flush_color is non-negative, flush_color on all pwqs should be
2694 * -1. If no pwq has in-flight commands at the specified color, all
2695 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2696 * has in flight commands, its pwq->flush_color is set to
2697 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
Tejun Heo73f53c42010-06-29 10:07:11 +02002698 * wakeup logic is armed and %true is returned.
2699 *
2700 * The caller should have initialized @wq->first_flusher prior to
2701 * calling this function with non-negative @flush_color. If
2702 * @flush_color is negative, no flush color update is done and %false
2703 * is returned.
2704 *
Tejun Heo112202d2013-02-13 19:29:12 -08002705 * If @work_color is non-negative, all pwqs should have the same
Tejun Heo73f53c42010-06-29 10:07:11 +02002706 * work_color which is previous to @work_color and all will be
2707 * advanced to @work_color.
2708 *
2709 * CONTEXT:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002710 * mutex_lock(wq->mutex).
Tejun Heo73f53c42010-06-29 10:07:11 +02002711 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002712 * Return:
Tejun Heo73f53c42010-06-29 10:07:11 +02002713 * %true if @flush_color >= 0 and there's something to flush. %false
2714 * otherwise.
2715 */
Tejun Heo112202d2013-02-13 19:29:12 -08002716static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
Tejun Heo73f53c42010-06-29 10:07:11 +02002717 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718{
Tejun Heo73f53c42010-06-29 10:07:11 +02002719 bool wait = false;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002720 struct pool_workqueue *pwq;
Oleg Nesterov14441962007-05-23 13:57:57 -07002721
Tejun Heo73f53c42010-06-29 10:07:11 +02002722 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002723 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
Tejun Heo112202d2013-02-13 19:29:12 -08002724 atomic_set(&wq->nr_pwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002725 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002726
Tejun Heo49e3cf42013-03-12 11:29:58 -07002727 for_each_pwq(pwq, wq) {
Tejun Heo112202d2013-02-13 19:29:12 -08002728 struct worker_pool *pool = pwq->pool;
Tejun Heo73f53c42010-06-29 10:07:11 +02002729
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002730 spin_lock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002731
2732 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002733 WARN_ON_ONCE(pwq->flush_color != -1);
Tejun Heo73f53c42010-06-29 10:07:11 +02002734
Tejun Heo112202d2013-02-13 19:29:12 -08002735 if (pwq->nr_in_flight[flush_color]) {
2736 pwq->flush_color = flush_color;
2737 atomic_inc(&wq->nr_pwqs_to_flush);
Tejun Heo73f53c42010-06-29 10:07:11 +02002738 wait = true;
2739 }
2740 }
2741
2742 if (work_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002743 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
Tejun Heo112202d2013-02-13 19:29:12 -08002744 pwq->work_color = work_color;
Tejun Heo73f53c42010-06-29 10:07:11 +02002745 }
2746
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002747 spin_unlock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002748 }
2749
Tejun Heo112202d2013-02-13 19:29:12 -08002750 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
Tejun Heo73f53c42010-06-29 10:07:11 +02002751 complete(&wq->first_flusher->done);
2752
2753 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754}
2755
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002756/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002758 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002760 * This function sleeps until all work items which were queued on entry
2761 * have finished execution, but it is not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002763void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764{
Tejun Heo73f53c42010-06-29 10:07:11 +02002765 struct wq_flusher this_flusher = {
2766 .list = LIST_HEAD_INIT(this_flusher.list),
2767 .flush_color = -1,
Byungchul Parkfd1a5b02017-10-25 17:56:04 +09002768 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
Tejun Heo73f53c42010-06-29 10:07:11 +02002769 };
2770 int next_color;
Oleg Nesterovb1f4ec172007-05-09 02:34:12 -07002771
Tejun Heo3347fa02016-09-16 15:49:32 -04002772 if (WARN_ON(!wq_online))
2773 return;
2774
Johannes Berg87915ad2018-08-22 11:49:04 +02002775 lock_map_acquire(&wq->lockdep_map);
2776 lock_map_release(&wq->lockdep_map);
2777
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002778 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002779
2780 /*
2781 * Start-to-wait phase
2782 */
2783 next_color = work_next_color(wq->work_color);
2784
2785 if (next_color != wq->flush_color) {
2786 /*
2787 * Color space is not full. The current work_color
2788 * becomes our flush_color and work_color is advanced
2789 * by one.
2790 */
Tejun Heo6183c002013-03-12 11:29:57 -07002791 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
Tejun Heo73f53c42010-06-29 10:07:11 +02002792 this_flusher.flush_color = wq->work_color;
2793 wq->work_color = next_color;
2794
2795 if (!wq->first_flusher) {
2796 /* no flush in progress, become the first flusher */
Tejun Heo6183c002013-03-12 11:29:57 -07002797 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002798
2799 wq->first_flusher = &this_flusher;
2800
Tejun Heo112202d2013-02-13 19:29:12 -08002801 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
Tejun Heo73f53c42010-06-29 10:07:11 +02002802 wq->work_color)) {
2803 /* nothing to flush, done */
2804 wq->flush_color = next_color;
2805 wq->first_flusher = NULL;
2806 goto out_unlock;
2807 }
2808 } else {
2809 /* wait in queue */
Tejun Heo6183c002013-03-12 11:29:57 -07002810 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002811 list_add_tail(&this_flusher.list, &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002812 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002813 }
2814 } else {
2815 /*
2816 * Oops, color space is full, wait on overflow queue.
2817 * The next flush completion will assign us
2818 * flush_color and transfer to flusher_queue.
2819 */
2820 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2821 }
2822
Tejun Heofca839c2015-12-07 10:58:57 -05002823 check_flush_dependency(wq, NULL);
2824
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002825 mutex_unlock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002826
2827 wait_for_completion(&this_flusher.done);
2828
2829 /*
2830 * Wake-up-and-cascade phase
2831 *
2832 * First flushers are responsible for cascading flushes and
2833 * handling overflow. Non-first flushers can simply return.
2834 */
2835 if (wq->first_flusher != &this_flusher)
2836 return;
2837
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002838 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002839
Tejun Heo4ce48b32010-07-02 10:03:51 +02002840 /* we might have raced, check again with mutex held */
2841 if (wq->first_flusher != &this_flusher)
2842 goto out_unlock;
2843
Tejun Heo73f53c42010-06-29 10:07:11 +02002844 wq->first_flusher = NULL;
2845
Tejun Heo6183c002013-03-12 11:29:57 -07002846 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2847 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002848
2849 while (true) {
2850 struct wq_flusher *next, *tmp;
2851
2852 /* complete all the flushers sharing the current flush color */
2853 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2854 if (next->flush_color != wq->flush_color)
2855 break;
2856 list_del_init(&next->list);
2857 complete(&next->done);
2858 }
2859
Tejun Heo6183c002013-03-12 11:29:57 -07002860 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2861 wq->flush_color != work_next_color(wq->work_color));
Tejun Heo73f53c42010-06-29 10:07:11 +02002862
2863 /* this flush_color is finished, advance by one */
2864 wq->flush_color = work_next_color(wq->flush_color);
2865
2866 /* one color has been freed, handle overflow queue */
2867 if (!list_empty(&wq->flusher_overflow)) {
2868 /*
2869 * Assign the same color to all overflowed
2870 * flushers, advance work_color and append to
2871 * flusher_queue. This is the start-to-wait
2872 * phase for these overflowed flushers.
2873 */
2874 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2875 tmp->flush_color = wq->work_color;
2876
2877 wq->work_color = work_next_color(wq->work_color);
2878
2879 list_splice_tail_init(&wq->flusher_overflow,
2880 &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002881 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002882 }
2883
2884 if (list_empty(&wq->flusher_queue)) {
Tejun Heo6183c002013-03-12 11:29:57 -07002885 WARN_ON_ONCE(wq->flush_color != wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002886 break;
2887 }
2888
2889 /*
2890 * Need to flush more colors. Make the next flusher
Tejun Heo112202d2013-02-13 19:29:12 -08002891 * the new first flusher and arm pwqs.
Tejun Heo73f53c42010-06-29 10:07:11 +02002892 */
Tejun Heo6183c002013-03-12 11:29:57 -07002893 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2894 WARN_ON_ONCE(wq->flush_color != next->flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002895
2896 list_del_init(&next->list);
2897 wq->first_flusher = next;
2898
Tejun Heo112202d2013-02-13 19:29:12 -08002899 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
Tejun Heo73f53c42010-06-29 10:07:11 +02002900 break;
2901
2902 /*
2903 * Meh... this color is already done, clear first
2904 * flusher and repeat cascading.
2905 */
2906 wq->first_flusher = NULL;
2907 }
2908
2909out_unlock:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002910 mutex_unlock(&wq->mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911}
Tim Gardner1dadafa2015-08-04 11:26:04 -06002912EXPORT_SYMBOL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002914/**
2915 * drain_workqueue - drain a workqueue
2916 * @wq: workqueue to drain
2917 *
2918 * Wait until the workqueue becomes empty. While draining is in progress,
2919 * only chain queueing is allowed. IOW, only currently pending or running
2920 * work items on @wq can queue further work items on it. @wq is flushed
Chen Hanxiaob749b1b2015-05-13 06:10:05 -04002921 * repeatedly until it becomes empty. The number of flushing is determined
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002922 * by the depth of chaining and should be relatively short. Whine if it
2923 * takes too long.
2924 */
2925void drain_workqueue(struct workqueue_struct *wq)
2926{
2927 unsigned int flush_cnt = 0;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002928 struct pool_workqueue *pwq;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002929
2930 /*
2931 * __queue_work() needs to test whether there are drainers, is much
2932 * hotter than drain_workqueue() and already looks at @wq->flags.
Tejun Heo618b01e2013-03-12 11:30:04 -07002933 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002934 */
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002935 mutex_lock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002936 if (!wq->nr_drainers++)
Tejun Heo618b01e2013-03-12 11:30:04 -07002937 wq->flags |= __WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002938 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002939reflush:
2940 flush_workqueue(wq);
2941
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002942 mutex_lock(&wq->mutex);
Tejun Heo76af4d92013-03-12 11:30:00 -07002943
Tejun Heo49e3cf42013-03-12 11:29:58 -07002944 for_each_pwq(pwq, wq) {
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002945 bool drained;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002946
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002947 spin_lock_irq(&pwq->pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08002948 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002949 spin_unlock_irq(&pwq->pool->lock);
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002950
2951 if (drained)
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002952 continue;
2953
2954 if (++flush_cnt == 10 ||
2955 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002956 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
Valentin Ilie044c7822012-08-19 00:52:42 +03002957 wq->name, flush_cnt);
Tejun Heo76af4d92013-03-12 11:30:00 -07002958
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002959 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002960 goto reflush;
2961 }
2962
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002963 if (!--wq->nr_drainers)
Tejun Heo618b01e2013-03-12 11:30:04 -07002964 wq->flags &= ~__WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002965 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002966}
2967EXPORT_SYMBOL_GPL(drain_workqueue);
2968
Johannes Bergd6e89782018-08-22 11:49:03 +02002969static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2970 bool from_cancel)
Tejun Heobaf59022010-09-16 10:42:16 +02002971{
2972 struct worker *worker = NULL;
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002973 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002974 struct pool_workqueue *pwq;
Tejun Heobaf59022010-09-16 10:42:16 +02002975
2976 might_sleep();
Tejun Heobaf59022010-09-16 10:42:16 +02002977
Thomas Gleixner24acfb72019-03-13 17:55:47 +01002978 rcu_read_lock();
Tejun Heofa1b54e2013-03-12 11:30:00 -07002979 pool = get_work_pool(work);
2980 if (!pool) {
Thomas Gleixner24acfb72019-03-13 17:55:47 +01002981 rcu_read_unlock();
Tejun Heofa1b54e2013-03-12 11:30:00 -07002982 return false;
2983 }
2984
Thomas Gleixner24acfb72019-03-13 17:55:47 +01002985 spin_lock_irq(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08002986 /* see the comment in try_to_grab_pending() with the same code */
Tejun Heo112202d2013-02-13 19:29:12 -08002987 pwq = get_work_pwq(work);
2988 if (pwq) {
2989 if (unlikely(pwq->pool != pool))
Tejun Heobaf59022010-09-16 10:42:16 +02002990 goto already_gone;
Tejun Heo606a5022012-08-20 14:51:23 -07002991 } else {
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002992 worker = find_worker_executing_work(pool, work);
Tejun Heobaf59022010-09-16 10:42:16 +02002993 if (!worker)
2994 goto already_gone;
Tejun Heo112202d2013-02-13 19:29:12 -08002995 pwq = worker->current_pwq;
Tejun Heo606a5022012-08-20 14:51:23 -07002996 }
Tejun Heobaf59022010-09-16 10:42:16 +02002997
Tejun Heofca839c2015-12-07 10:58:57 -05002998 check_flush_dependency(pwq->wq, work);
2999
Tejun Heo112202d2013-02-13 19:29:12 -08003000 insert_wq_barrier(pwq, barr, work, worker);
Tejun Heod565ed62013-01-24 11:01:33 -08003001 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02003002
Tejun Heoe159489b2011-01-09 23:32:15 +01003003 /*
Peter Zijlstraa1d14932017-08-23 12:52:32 +02003004 * Force a lock recursion deadlock when using flush_work() inside a
3005 * single-threaded or rescuer equipped workqueue.
3006 *
3007 * For single threaded workqueues the deadlock happens when the work
3008 * is after the work issuing the flush_work(). For rescuer equipped
3009 * workqueues the deadlock happens when the rescuer stalls, blocking
3010 * forward progress.
Tejun Heoe159489b2011-01-09 23:32:15 +01003011 */
Johannes Bergd6e89782018-08-22 11:49:03 +02003012 if (!from_cancel &&
3013 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
Tejun Heo112202d2013-02-13 19:29:12 -08003014 lock_map_acquire(&pwq->wq->lockdep_map);
Peter Zijlstraa1d14932017-08-23 12:52:32 +02003015 lock_map_release(&pwq->wq->lockdep_map);
3016 }
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003017 rcu_read_unlock();
Tejun Heobaf59022010-09-16 10:42:16 +02003018 return true;
3019already_gone:
Tejun Heod565ed62013-01-24 11:01:33 -08003020 spin_unlock_irq(&pool->lock);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003021 rcu_read_unlock();
Tejun Heobaf59022010-09-16 10:42:16 +02003022 return false;
3023}
3024
Johannes Bergd6e89782018-08-22 11:49:03 +02003025static bool __flush_work(struct work_struct *work, bool from_cancel)
3026{
3027 struct wq_barrier barr;
3028
3029 if (WARN_ON(!wq_online))
3030 return false;
3031
Tetsuo Handa4d43d392019-01-23 09:44:12 +09003032 if (WARN_ON(!work->func))
3033 return false;
3034
Johannes Berg87915ad2018-08-22 11:49:04 +02003035 if (!from_cancel) {
3036 lock_map_acquire(&work->lockdep_map);
3037 lock_map_release(&work->lockdep_map);
3038 }
3039
Johannes Bergd6e89782018-08-22 11:49:03 +02003040 if (start_flush_work(work, &barr, from_cancel)) {
3041 wait_for_completion(&barr.done);
3042 destroy_work_on_stack(&barr.work);
3043 return true;
3044 } else {
3045 return false;
3046 }
3047}
3048
Oleg Nesterovdb700892008-07-25 01:47:49 -07003049/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003050 * flush_work - wait for a work to finish executing the last queueing instance
3051 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07003052 *
Tejun Heo606a5022012-08-20 14:51:23 -07003053 * Wait until @work has finished execution. @work is guaranteed to be idle
3054 * on return if it hasn't been requeued since flush started.
Tejun Heo401a8d02010-09-16 10:36:00 +02003055 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003056 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003057 * %true if flush_work() waited for the work to finish execution,
3058 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07003059 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003060bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07003061{
Johannes Bergd6e89782018-08-22 11:49:03 +02003062 return __flush_work(work, false);
Oleg Nesterovdb700892008-07-25 01:47:49 -07003063}
3064EXPORT_SYMBOL_GPL(flush_work);
3065
Tejun Heo8603e1b32015-03-05 08:04:13 -05003066struct cwt_wait {
Ingo Molnarac6424b2017-06-20 12:06:13 +02003067 wait_queue_entry_t wait;
Tejun Heo8603e1b32015-03-05 08:04:13 -05003068 struct work_struct *work;
3069};
3070
Ingo Molnarac6424b2017-06-20 12:06:13 +02003071static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
Tejun Heo8603e1b32015-03-05 08:04:13 -05003072{
3073 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3074
3075 if (cwait->work != key)
3076 return 0;
3077 return autoremove_wake_function(wait, mode, sync, key);
3078}
3079
Tejun Heo36e227d2012-08-03 10:30:46 -07003080static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
Tejun Heo401a8d02010-09-16 10:36:00 +02003081{
Tejun Heo8603e1b32015-03-05 08:04:13 -05003082 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
Tejun Heobbb68df2012-08-03 10:30:46 -07003083 unsigned long flags;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003084 int ret;
3085
3086 do {
Tejun Heobbb68df2012-08-03 10:30:46 -07003087 ret = try_to_grab_pending(work, is_dwork, &flags);
3088 /*
Tejun Heo8603e1b32015-03-05 08:04:13 -05003089 * If someone else is already canceling, wait for it to
3090 * finish. flush_work() doesn't work for PREEMPT_NONE
3091 * because we may get scheduled between @work's completion
3092 * and the other canceling task resuming and clearing
3093 * CANCELING - flush_work() will return false immediately
3094 * as @work is no longer busy, try_to_grab_pending() will
3095 * return -ENOENT as @work is still being canceled and the
3096 * other canceling task won't be able to clear CANCELING as
3097 * we're hogging the CPU.
3098 *
3099 * Let's wait for completion using a waitqueue. As this
3100 * may lead to the thundering herd problem, use a custom
3101 * wake function which matches @work along with exclusive
3102 * wait and wakeup.
Tejun Heobbb68df2012-08-03 10:30:46 -07003103 */
Tejun Heo8603e1b32015-03-05 08:04:13 -05003104 if (unlikely(ret == -ENOENT)) {
3105 struct cwt_wait cwait;
3106
3107 init_wait(&cwait.wait);
3108 cwait.wait.func = cwt_wakefn;
3109 cwait.work = work;
3110
3111 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3112 TASK_UNINTERRUPTIBLE);
3113 if (work_is_canceling(work))
3114 schedule();
3115 finish_wait(&cancel_waitq, &cwait.wait);
3116 }
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003117 } while (unlikely(ret < 0));
3118
Tejun Heobbb68df2012-08-03 10:30:46 -07003119 /* tell other tasks trying to grab @work to back off */
3120 mark_work_canceling(work);
3121 local_irq_restore(flags);
3122
Tejun Heo3347fa02016-09-16 15:49:32 -04003123 /*
3124 * This allows canceling during early boot. We know that @work
3125 * isn't executing.
3126 */
3127 if (wq_online)
Johannes Bergd6e89782018-08-22 11:49:03 +02003128 __flush_work(work, true);
Tejun Heo3347fa02016-09-16 15:49:32 -04003129
Tejun Heo7a22ad72010-06-29 10:07:13 +02003130 clear_work_data(work);
Tejun Heo8603e1b32015-03-05 08:04:13 -05003131
3132 /*
3133 * Paired with prepare_to_wait() above so that either
3134 * waitqueue_active() is visible here or !work_is_canceling() is
3135 * visible there.
3136 */
3137 smp_mb();
3138 if (waitqueue_active(&cancel_waitq))
3139 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3140
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003141 return ret;
3142}
3143
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003144/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003145 * cancel_work_sync - cancel a work and wait for it to finish
3146 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003147 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003148 * Cancel @work and wait for its execution to finish. This function
3149 * can be used even if the work re-queues itself or migrates to
3150 * another workqueue. On return from this function, @work is
3151 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003152 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003153 * cancel_work_sync(&delayed_work->work) must not be used for
3154 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003155 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003156 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003157 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02003158 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003159 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003160 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003161 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003162bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003163{
Tejun Heo36e227d2012-08-03 10:30:46 -07003164 return __cancel_work_timer(work, false);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07003165}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07003166EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07003167
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003168/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003169 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3170 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003171 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003172 * Delayed timer is cancelled and the pending work is queued for
3173 * immediate execution. Like flush_work(), this function only
3174 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003175 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003176 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003177 * %true if flush_work() waited for the work to finish execution,
3178 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003179 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003180bool flush_delayed_work(struct delayed_work *dwork)
3181{
Tejun Heo8930cab2012-08-03 10:30:45 -07003182 local_irq_disable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003183 if (del_timer_sync(&dwork->timer))
Lai Jiangshan60c057b2013-02-06 18:04:53 -08003184 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Tejun Heo8930cab2012-08-03 10:30:45 -07003185 local_irq_enable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003186 return flush_work(&dwork->work);
3187}
3188EXPORT_SYMBOL(flush_delayed_work);
3189
Tejun Heo05f0fe62018-03-14 12:45:13 -07003190/**
3191 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3192 * @rwork: the rcu work to flush
3193 *
3194 * Return:
3195 * %true if flush_rcu_work() waited for the work to finish execution,
3196 * %false if it was already idle.
3197 */
3198bool flush_rcu_work(struct rcu_work *rwork)
3199{
3200 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3201 rcu_barrier();
3202 flush_work(&rwork->work);
3203 return true;
3204 } else {
3205 return flush_work(&rwork->work);
3206 }
3207}
3208EXPORT_SYMBOL(flush_rcu_work);
3209
Jens Axboef72b8792016-08-24 15:51:50 -06003210static bool __cancel_work(struct work_struct *work, bool is_dwork)
3211{
3212 unsigned long flags;
3213 int ret;
3214
3215 do {
3216 ret = try_to_grab_pending(work, is_dwork, &flags);
3217 } while (unlikely(ret == -EAGAIN));
3218
3219 if (unlikely(ret < 0))
3220 return false;
3221
3222 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3223 local_irq_restore(flags);
3224 return ret;
3225}
3226
Tejun Heo401a8d02010-09-16 10:36:00 +02003227/**
Tejun Heo57b30ae2012-08-21 13:18:24 -07003228 * cancel_delayed_work - cancel a delayed work
3229 * @dwork: delayed_work to cancel
Tejun Heo09383492010-09-16 10:48:29 +02003230 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003231 * Kill off a pending delayed_work.
3232 *
3233 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3234 * pending.
3235 *
3236 * Note:
3237 * The work callback function may still be running on return, unless
3238 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3239 * use cancel_delayed_work_sync() to wait on it.
Tejun Heo09383492010-09-16 10:48:29 +02003240 *
Tejun Heo57b30ae2012-08-21 13:18:24 -07003241 * This function is safe to call from any context including IRQ handler.
Tejun Heo09383492010-09-16 10:48:29 +02003242 */
Tejun Heo57b30ae2012-08-21 13:18:24 -07003243bool cancel_delayed_work(struct delayed_work *dwork)
Tejun Heo09383492010-09-16 10:48:29 +02003244{
Jens Axboef72b8792016-08-24 15:51:50 -06003245 return __cancel_work(&dwork->work, true);
Tejun Heo09383492010-09-16 10:48:29 +02003246}
Tejun Heo57b30ae2012-08-21 13:18:24 -07003247EXPORT_SYMBOL(cancel_delayed_work);
Tejun Heo09383492010-09-16 10:48:29 +02003248
3249/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003250 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3251 * @dwork: the delayed work cancel
3252 *
3253 * This is cancel_work_sync() for delayed works.
3254 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003255 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003256 * %true if @dwork was pending, %false otherwise.
3257 */
3258bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003259{
Tejun Heo36e227d2012-08-03 10:30:46 -07003260 return __cancel_work_timer(&dwork->work, true);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003261}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07003262EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07003264/**
Tejun Heo31ddd872010-10-19 11:14:49 +02003265 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07003266 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07003267 *
Tejun Heo31ddd872010-10-19 11:14:49 +02003268 * schedule_on_each_cpu() executes @func on each online CPU using the
3269 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07003270 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02003271 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003272 * Return:
Tejun Heo31ddd872010-10-19 11:14:49 +02003273 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07003274 */
David Howells65f27f32006-11-22 14:55:48 +00003275int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003276{
3277 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02003278 struct work_struct __percpu *works;
Christoph Lameter15316ba2006-01-08 01:00:43 -08003279
Andrew Mortonb6136772006-06-25 05:47:49 -07003280 works = alloc_percpu(struct work_struct);
3281 if (!works)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003282 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07003283
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003284 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08003285
Christoph Lameter15316ba2006-01-08 01:00:43 -08003286 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01003287 struct work_struct *work = per_cpu_ptr(works, cpu);
3288
3289 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003290 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02003291 }
Tejun Heo93981802009-11-17 14:06:20 -08003292
3293 for_each_online_cpu(cpu)
3294 flush_work(per_cpu_ptr(works, cpu));
3295
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003296 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07003297 free_percpu(works);
Christoph Lameter15316ba2006-01-08 01:00:43 -08003298 return 0;
3299}
3300
Alan Sterneef6a7d2010-02-12 17:39:21 +09003301/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06003302 * execute_in_process_context - reliably execute the routine with user context
3303 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06003304 * @ew: guaranteed storage for the execute work structure (must
3305 * be available when the work executes)
3306 *
3307 * Executes the function immediately if process context is available,
3308 * otherwise schedules the function for delayed execution.
3309 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003310 * Return: 0 - function was executed
James Bottomley1fa44ec2006-02-23 12:43:43 -06003311 * 1 - function was scheduled for execution
3312 */
David Howells65f27f32006-11-22 14:55:48 +00003313int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06003314{
3315 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00003316 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003317 return 0;
3318 }
3319
David Howells65f27f32006-11-22 14:55:48 +00003320 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003321 schedule_work(&ew->work);
3322
3323 return 1;
3324}
3325EXPORT_SYMBOL_GPL(execute_in_process_context);
3326
Tejun Heo7a4e3442013-03-12 11:30:00 -07003327/**
3328 * free_workqueue_attrs - free a workqueue_attrs
3329 * @attrs: workqueue_attrs to free
3330 *
3331 * Undo alloc_workqueue_attrs().
3332 */
Daniel Jordan513c98d2019-09-05 21:40:22 -04003333void free_workqueue_attrs(struct workqueue_attrs *attrs)
Tejun Heo7a4e3442013-03-12 11:30:00 -07003334{
3335 if (attrs) {
3336 free_cpumask_var(attrs->cpumask);
3337 kfree(attrs);
3338 }
3339}
3340
3341/**
3342 * alloc_workqueue_attrs - allocate a workqueue_attrs
Tejun Heo7a4e3442013-03-12 11:30:00 -07003343 *
3344 * Allocate a new workqueue_attrs, initialize with default settings and
Yacine Belkadid185af32013-07-31 14:59:24 -07003345 * return it.
3346 *
3347 * Return: The allocated new workqueue_attr on success. %NULL on failure.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003348 */
Daniel Jordan513c98d2019-09-05 21:40:22 -04003349struct workqueue_attrs *alloc_workqueue_attrs(void)
Tejun Heo7a4e3442013-03-12 11:30:00 -07003350{
3351 struct workqueue_attrs *attrs;
3352
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003353 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003354 if (!attrs)
3355 goto fail;
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003356 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
Tejun Heo7a4e3442013-03-12 11:30:00 -07003357 goto fail;
3358
Tejun Heo13e2e552013-04-01 11:23:31 -07003359 cpumask_copy(attrs->cpumask, cpu_possible_mask);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003360 return attrs;
3361fail:
3362 free_workqueue_attrs(attrs);
3363 return NULL;
3364}
3365
Tejun Heo29c91e92013-03-12 11:30:03 -07003366static void copy_workqueue_attrs(struct workqueue_attrs *to,
3367 const struct workqueue_attrs *from)
3368{
3369 to->nice = from->nice;
3370 cpumask_copy(to->cpumask, from->cpumask);
Shaohua Li2865a8f2013-08-01 09:56:36 +08003371 /*
3372 * Unlike hash and equality test, this function doesn't ignore
3373 * ->no_numa as it is used for both pool and wq attrs. Instead,
3374 * get_unbound_pool() explicitly clears ->no_numa after copying.
3375 */
3376 to->no_numa = from->no_numa;
Tejun Heo29c91e92013-03-12 11:30:03 -07003377}
3378
Tejun Heo29c91e92013-03-12 11:30:03 -07003379/* hash value of the content of @attr */
3380static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3381{
3382 u32 hash = 0;
3383
3384 hash = jhash_1word(attrs->nice, hash);
Tejun Heo13e2e552013-04-01 11:23:31 -07003385 hash = jhash(cpumask_bits(attrs->cpumask),
3386 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
Tejun Heo29c91e92013-03-12 11:30:03 -07003387 return hash;
3388}
3389
3390/* content equality test */
3391static bool wqattrs_equal(const struct workqueue_attrs *a,
3392 const struct workqueue_attrs *b)
3393{
3394 if (a->nice != b->nice)
3395 return false;
3396 if (!cpumask_equal(a->cpumask, b->cpumask))
3397 return false;
3398 return true;
3399}
3400
Tejun Heo7a4e3442013-03-12 11:30:00 -07003401/**
3402 * init_worker_pool - initialize a newly zalloc'd worker_pool
3403 * @pool: worker_pool to initialize
3404 *
Shailendra Verma402dd892015-05-23 10:38:14 +05303405 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
Yacine Belkadid185af32013-07-31 14:59:24 -07003406 *
3407 * Return: 0 on success, -errno on failure. Even on failure, all fields
Tejun Heo29c91e92013-03-12 11:30:03 -07003408 * inside @pool proper are initialized and put_unbound_pool() can be called
3409 * on @pool safely to release it.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003410 */
3411static int init_worker_pool(struct worker_pool *pool)
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003412{
3413 spin_lock_init(&pool->lock);
Tejun Heo29c91e92013-03-12 11:30:03 -07003414 pool->id = -1;
3415 pool->cpu = -1;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003416 pool->node = NUMA_NO_NODE;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003417 pool->flags |= POOL_DISASSOCIATED;
Tejun Heo82607adc2015-12-08 11:28:04 -05003418 pool->watchdog_ts = jiffies;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003419 INIT_LIST_HEAD(&pool->worklist);
3420 INIT_LIST_HEAD(&pool->idle_list);
3421 hash_init(pool->busy_hash);
3422
Kees Cook32a6c722017-10-16 15:58:25 -07003423 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003424
Kees Cook32a6c722017-10-16 15:58:25 -07003425 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003426
Lai Jiangshanda028462014-05-20 17:46:31 +08003427 INIT_LIST_HEAD(&pool->workers);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003428
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003429 ida_init(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003430 INIT_HLIST_NODE(&pool->hash_node);
3431 pool->refcnt = 1;
3432
3433 /* shouldn't fail above this point */
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003434 pool->attrs = alloc_workqueue_attrs();
Tejun Heo7a4e3442013-03-12 11:30:00 -07003435 if (!pool->attrs)
3436 return -ENOMEM;
3437 return 0;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003438}
3439
Bart Van Assche669de8b2019-02-14 15:00:54 -08003440#ifdef CONFIG_LOCKDEP
3441static void wq_init_lockdep(struct workqueue_struct *wq)
3442{
3443 char *lock_name;
3444
3445 lockdep_register_key(&wq->key);
3446 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3447 if (!lock_name)
3448 lock_name = wq->name;
Qian Cai69a106c2019-03-06 19:27:31 -05003449
3450 wq->lock_name = lock_name;
Bart Van Assche669de8b2019-02-14 15:00:54 -08003451 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3452}
3453
3454static void wq_unregister_lockdep(struct workqueue_struct *wq)
3455{
3456 lockdep_unregister_key(&wq->key);
3457}
3458
3459static void wq_free_lockdep(struct workqueue_struct *wq)
3460{
3461 if (wq->lock_name != wq->name)
3462 kfree(wq->lock_name);
3463}
3464#else
3465static void wq_init_lockdep(struct workqueue_struct *wq)
3466{
3467}
3468
3469static void wq_unregister_lockdep(struct workqueue_struct *wq)
3470{
3471}
3472
3473static void wq_free_lockdep(struct workqueue_struct *wq)
3474{
3475}
3476#endif
3477
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003478static void rcu_free_wq(struct rcu_head *rcu)
3479{
3480 struct workqueue_struct *wq =
3481 container_of(rcu, struct workqueue_struct, rcu);
3482
Bart Van Assche669de8b2019-02-14 15:00:54 -08003483 wq_free_lockdep(wq);
3484
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003485 if (!(wq->flags & WQ_UNBOUND))
3486 free_percpu(wq->cpu_pwqs);
3487 else
3488 free_workqueue_attrs(wq->unbound_attrs);
3489
3490 kfree(wq->rescuer);
3491 kfree(wq);
3492}
3493
Tejun Heo29c91e92013-03-12 11:30:03 -07003494static void rcu_free_pool(struct rcu_head *rcu)
3495{
3496 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3497
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003498 ida_destroy(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003499 free_workqueue_attrs(pool->attrs);
3500 kfree(pool);
3501}
3502
3503/**
3504 * put_unbound_pool - put a worker_pool
3505 * @pool: worker_pool to put
3506 *
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003507 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003508 * safe manner. get_unbound_pool() calls this function on its failure path
3509 * and this function should be able to release pools which went through,
3510 * successfully or not, init_worker_pool().
Tejun Heoa892cac2013-04-01 11:23:32 -07003511 *
3512 * Should be called with wq_pool_mutex held.
Tejun Heo29c91e92013-03-12 11:30:03 -07003513 */
3514static void put_unbound_pool(struct worker_pool *pool)
3515{
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003516 DECLARE_COMPLETION_ONSTACK(detach_completion);
Tejun Heo29c91e92013-03-12 11:30:03 -07003517 struct worker *worker;
3518
Tejun Heoa892cac2013-04-01 11:23:32 -07003519 lockdep_assert_held(&wq_pool_mutex);
3520
3521 if (--pool->refcnt)
Tejun Heo29c91e92013-03-12 11:30:03 -07003522 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003523
3524 /* sanity checks */
Lai Jiangshan61d0fbb2014-06-03 15:31:45 +08003525 if (WARN_ON(!(pool->cpu < 0)) ||
Tejun Heoa892cac2013-04-01 11:23:32 -07003526 WARN_ON(!list_empty(&pool->worklist)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003527 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003528
3529 /* release id and unhash */
3530 if (pool->id >= 0)
3531 idr_remove(&worker_pool_idr, pool->id);
3532 hash_del(&pool->hash_node);
3533
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003534 /*
Tejun Heo692b4822017-10-09 08:04:13 -07003535 * Become the manager and destroy all workers. This prevents
3536 * @pool's workers from blocking on attach_mutex. We're the last
3537 * manager and @pool gets freed with the flag set.
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003538 */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003539 spin_lock_irq(&pool->lock);
Tejun Heo692b4822017-10-09 08:04:13 -07003540 wait_event_lock_irq(wq_manager_wait,
3541 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3542 pool->flags |= POOL_MANAGER_ACTIVE;
3543
Lai Jiangshan1037de32014-05-22 16:44:07 +08003544 while ((worker = first_idle_worker(pool)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003545 destroy_worker(worker);
3546 WARN_ON(pool->nr_workers || pool->nr_idle);
Tejun Heo29c91e92013-03-12 11:30:03 -07003547 spin_unlock_irq(&pool->lock);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003548
Tejun Heo1258fae2018-05-18 08:47:13 -07003549 mutex_lock(&wq_pool_attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003550 if (!list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003551 pool->detach_completion = &detach_completion;
Tejun Heo1258fae2018-05-18 08:47:13 -07003552 mutex_unlock(&wq_pool_attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003553
3554 if (pool->detach_completion)
3555 wait_for_completion(pool->detach_completion);
3556
Tejun Heo29c91e92013-03-12 11:30:03 -07003557 /* shut down the timers */
3558 del_timer_sync(&pool->idle_timer);
3559 del_timer_sync(&pool->mayday_timer);
3560
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003561 /* RCU protected to allow dereferences from get_work_pool() */
Paul E. McKenney25b00772018-11-06 19:18:45 -08003562 call_rcu(&pool->rcu, rcu_free_pool);
Tejun Heo29c91e92013-03-12 11:30:03 -07003563}
3564
3565/**
3566 * get_unbound_pool - get a worker_pool with the specified attributes
3567 * @attrs: the attributes of the worker_pool to get
3568 *
3569 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3570 * reference count and return it. If there already is a matching
3571 * worker_pool, it will be used; otherwise, this function attempts to
Yacine Belkadid185af32013-07-31 14:59:24 -07003572 * create a new one.
Tejun Heoa892cac2013-04-01 11:23:32 -07003573 *
3574 * Should be called with wq_pool_mutex held.
Yacine Belkadid185af32013-07-31 14:59:24 -07003575 *
3576 * Return: On success, a worker_pool with the same attributes as @attrs.
3577 * On failure, %NULL.
Tejun Heo29c91e92013-03-12 11:30:03 -07003578 */
3579static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3580{
Tejun Heo29c91e92013-03-12 11:30:03 -07003581 u32 hash = wqattrs_hash(attrs);
3582 struct worker_pool *pool;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003583 int node;
Xunlei Pange22735842015-10-09 11:53:12 +08003584 int target_node = NUMA_NO_NODE;
Tejun Heo29c91e92013-03-12 11:30:03 -07003585
Tejun Heoa892cac2013-04-01 11:23:32 -07003586 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003587
3588 /* do we already have a matching pool? */
Tejun Heo29c91e92013-03-12 11:30:03 -07003589 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3590 if (wqattrs_equal(pool->attrs, attrs)) {
3591 pool->refcnt++;
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003592 return pool;
Tejun Heo29c91e92013-03-12 11:30:03 -07003593 }
3594 }
Tejun Heo29c91e92013-03-12 11:30:03 -07003595
Xunlei Pange22735842015-10-09 11:53:12 +08003596 /* if cpumask is contained inside a NUMA node, we belong to that node */
3597 if (wq_numa_enabled) {
3598 for_each_node(node) {
3599 if (cpumask_subset(attrs->cpumask,
3600 wq_numa_possible_cpumask[node])) {
3601 target_node = node;
3602 break;
3603 }
3604 }
3605 }
3606
Tejun Heo29c91e92013-03-12 11:30:03 -07003607 /* nope, create a new one */
Xunlei Pange22735842015-10-09 11:53:12 +08003608 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
Tejun Heo29c91e92013-03-12 11:30:03 -07003609 if (!pool || init_worker_pool(pool) < 0)
3610 goto fail;
3611
Tejun Heo8864b4e2013-03-12 11:30:04 -07003612 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
Tejun Heo29c91e92013-03-12 11:30:03 -07003613 copy_workqueue_attrs(pool->attrs, attrs);
Xunlei Pange22735842015-10-09 11:53:12 +08003614 pool->node = target_node;
Tejun Heo29c91e92013-03-12 11:30:03 -07003615
Shaohua Li2865a8f2013-08-01 09:56:36 +08003616 /*
3617 * no_numa isn't a worker_pool attribute, always clear it. See
3618 * 'struct workqueue_attrs' comments for detail.
3619 */
3620 pool->attrs->no_numa = false;
3621
Tejun Heo29c91e92013-03-12 11:30:03 -07003622 if (worker_pool_assign_id(pool) < 0)
3623 goto fail;
3624
3625 /* create and start the initial worker */
Tejun Heo3347fa02016-09-16 15:49:32 -04003626 if (wq_online && !create_worker(pool))
Tejun Heo29c91e92013-03-12 11:30:03 -07003627 goto fail;
3628
Tejun Heo29c91e92013-03-12 11:30:03 -07003629 /* install */
Tejun Heo29c91e92013-03-12 11:30:03 -07003630 hash_add(unbound_pool_hash, &pool->hash_node, hash);
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003631
Tejun Heo29c91e92013-03-12 11:30:03 -07003632 return pool;
3633fail:
Tejun Heo29c91e92013-03-12 11:30:03 -07003634 if (pool)
3635 put_unbound_pool(pool);
3636 return NULL;
3637}
3638
Tejun Heo8864b4e2013-03-12 11:30:04 -07003639static void rcu_free_pwq(struct rcu_head *rcu)
3640{
3641 kmem_cache_free(pwq_cache,
3642 container_of(rcu, struct pool_workqueue, rcu));
3643}
3644
3645/*
3646 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3647 * and needs to be destroyed.
3648 */
3649static void pwq_unbound_release_workfn(struct work_struct *work)
3650{
3651 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3652 unbound_release_work);
3653 struct workqueue_struct *wq = pwq->wq;
3654 struct worker_pool *pool = pwq->pool;
Tejun Heobc0caf02013-04-01 11:23:31 -07003655 bool is_last;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003656
3657 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3658 return;
3659
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003660 mutex_lock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003661 list_del_rcu(&pwq->pwqs_node);
Tejun Heobc0caf02013-04-01 11:23:31 -07003662 is_last = list_empty(&wq->pwqs);
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003663 mutex_unlock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003664
Tejun Heoa892cac2013-04-01 11:23:32 -07003665 mutex_lock(&wq_pool_mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003666 put_unbound_pool(pool);
Tejun Heoa892cac2013-04-01 11:23:32 -07003667 mutex_unlock(&wq_pool_mutex);
3668
Paul E. McKenney25b00772018-11-06 19:18:45 -08003669 call_rcu(&pwq->rcu, rcu_free_pwq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003670
3671 /*
3672 * If we're the last pwq going away, @wq is already dead and no one
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003673 * is gonna access it anymore. Schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003674 */
Bart Van Assche669de8b2019-02-14 15:00:54 -08003675 if (is_last) {
3676 wq_unregister_lockdep(wq);
Paul E. McKenney25b00772018-11-06 19:18:45 -08003677 call_rcu(&wq->rcu, rcu_free_wq);
Bart Van Assche669de8b2019-02-14 15:00:54 -08003678 }
Tejun Heo8864b4e2013-03-12 11:30:04 -07003679}
3680
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003681/**
Tejun Heo699ce092013-03-13 16:51:35 -07003682 * pwq_adjust_max_active - update a pwq's max_active to the current setting
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003683 * @pwq: target pool_workqueue
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003684 *
Tejun Heo699ce092013-03-13 16:51:35 -07003685 * If @pwq isn't freezing, set @pwq->max_active to the associated
3686 * workqueue's saved_max_active and activate delayed work items
3687 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003688 */
Tejun Heo699ce092013-03-13 16:51:35 -07003689static void pwq_adjust_max_active(struct pool_workqueue *pwq)
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003690{
Tejun Heo699ce092013-03-13 16:51:35 -07003691 struct workqueue_struct *wq = pwq->wq;
3692 bool freezable = wq->flags & WQ_FREEZABLE;
Tejun Heo3347fa02016-09-16 15:49:32 -04003693 unsigned long flags;
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003694
Tejun Heo699ce092013-03-13 16:51:35 -07003695 /* for @wq->saved_max_active */
Lai Jiangshana357fc02013-03-25 16:57:19 -07003696 lockdep_assert_held(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003697
3698 /* fast exit for non-freezable wqs */
3699 if (!freezable && pwq->max_active == wq->saved_max_active)
3700 return;
3701
Tejun Heo3347fa02016-09-16 15:49:32 -04003702 /* this function can be called during early boot w/ irq disabled */
3703 spin_lock_irqsave(&pwq->pool->lock, flags);
Tejun Heo699ce092013-03-13 16:51:35 -07003704
Lai Jiangshan74b414e2014-05-22 19:01:16 +08003705 /*
3706 * During [un]freezing, the caller is responsible for ensuring that
3707 * this function is called at least once after @workqueue_freezing
3708 * is updated and visible.
3709 */
3710 if (!freezable || !workqueue_freezing) {
Tejun Heo699ce092013-03-13 16:51:35 -07003711 pwq->max_active = wq->saved_max_active;
3712
3713 while (!list_empty(&pwq->delayed_works) &&
3714 pwq->nr_active < pwq->max_active)
3715 pwq_activate_first_delayed(pwq);
Lai Jiangshan951a0782013-03-20 10:52:30 -07003716
3717 /*
3718 * Need to kick a worker after thawed or an unbound wq's
3719 * max_active is bumped. It's a slow path. Do it always.
3720 */
3721 wake_up_worker(pwq->pool);
Tejun Heo699ce092013-03-13 16:51:35 -07003722 } else {
3723 pwq->max_active = 0;
3724 }
3725
Tejun Heo3347fa02016-09-16 15:49:32 -04003726 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003727}
3728
Tejun Heoe50aba92013-04-01 11:23:35 -07003729/* initialize newly alloced @pwq which is associated with @wq and @pool */
Tejun Heof147f292013-04-01 11:23:35 -07003730static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3731 struct worker_pool *pool)
Tejun Heod2c1d402013-03-12 11:30:04 -07003732{
3733 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3734
Tejun Heoe50aba92013-04-01 11:23:35 -07003735 memset(pwq, 0, sizeof(*pwq));
3736
Tejun Heod2c1d402013-03-12 11:30:04 -07003737 pwq->pool = pool;
3738 pwq->wq = wq;
3739 pwq->flush_color = -1;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003740 pwq->refcnt = 1;
Tejun Heod2c1d402013-03-12 11:30:04 -07003741 INIT_LIST_HEAD(&pwq->delayed_works);
Tejun Heo1befcf32013-04-01 11:23:35 -07003742 INIT_LIST_HEAD(&pwq->pwqs_node);
Tejun Heod2c1d402013-03-12 11:30:04 -07003743 INIT_LIST_HEAD(&pwq->mayday_node);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003744 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
Tejun Heof147f292013-04-01 11:23:35 -07003745}
Tejun Heod2c1d402013-03-12 11:30:04 -07003746
Tejun Heof147f292013-04-01 11:23:35 -07003747/* sync @pwq with the current state of its associated wq and link it */
Tejun Heo1befcf32013-04-01 11:23:35 -07003748static void link_pwq(struct pool_workqueue *pwq)
Tejun Heof147f292013-04-01 11:23:35 -07003749{
3750 struct workqueue_struct *wq = pwq->wq;
3751
3752 lockdep_assert_held(&wq->mutex);
Tejun Heo75ccf592013-03-12 11:30:04 -07003753
Tejun Heo1befcf32013-04-01 11:23:35 -07003754 /* may be called multiple times, ignore if already linked */
3755 if (!list_empty(&pwq->pwqs_node))
3756 return;
3757
Lai Jiangshan29b1cb42014-07-22 13:04:27 +08003758 /* set the matching work_color */
Tejun Heo75ccf592013-03-12 11:30:04 -07003759 pwq->work_color = wq->work_color;
Tejun Heo983ca252013-03-13 16:51:35 -07003760
3761 /* sync max_active to the current setting */
3762 pwq_adjust_max_active(pwq);
3763
3764 /* link in @pwq */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003765 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
Tejun Heof147f292013-04-01 11:23:35 -07003766}
Lai Jiangshana357fc02013-03-25 16:57:19 -07003767
Tejun Heof147f292013-04-01 11:23:35 -07003768/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3769static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3770 const struct workqueue_attrs *attrs)
3771{
3772 struct worker_pool *pool;
3773 struct pool_workqueue *pwq;
3774
3775 lockdep_assert_held(&wq_pool_mutex);
3776
3777 pool = get_unbound_pool(attrs);
3778 if (!pool)
3779 return NULL;
3780
Tejun Heoe50aba92013-04-01 11:23:35 -07003781 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
Tejun Heof147f292013-04-01 11:23:35 -07003782 if (!pwq) {
3783 put_unbound_pool(pool);
3784 return NULL;
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003785 }
Tejun Heo6029a912013-04-01 11:23:34 -07003786
Tejun Heof147f292013-04-01 11:23:35 -07003787 init_pwq(pwq, wq, pool);
3788 return pwq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003789}
3790
Tejun Heo4c16bd32013-04-01 11:23:36 -07003791/**
Gong Zhaogang30186c62015-05-11 11:02:47 -04003792 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003793 * @attrs: the wq_attrs of the default pwq of the target workqueue
Tejun Heo4c16bd32013-04-01 11:23:36 -07003794 * @node: the target NUMA node
3795 * @cpu_going_down: if >= 0, the CPU to consider as offline
3796 * @cpumask: outarg, the resulting cpumask
3797 *
3798 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3799 * @cpu_going_down is >= 0, that cpu is considered offline during
Yacine Belkadid185af32013-07-31 14:59:24 -07003800 * calculation. The result is stored in @cpumask.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003801 *
3802 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3803 * enabled and @node has online CPUs requested by @attrs, the returned
3804 * cpumask is the intersection of the possible CPUs of @node and
3805 * @attrs->cpumask.
3806 *
3807 * The caller is responsible for ensuring that the cpumask of @node stays
3808 * stable.
Yacine Belkadid185af32013-07-31 14:59:24 -07003809 *
3810 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3811 * %false if equal.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003812 */
3813static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3814 int cpu_going_down, cpumask_t *cpumask)
3815{
Tejun Heod55262c2013-04-01 11:23:38 -07003816 if (!wq_numa_enabled || attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003817 goto use_dfl;
3818
3819 /* does @node have any online CPUs @attrs wants? */
3820 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3821 if (cpu_going_down >= 0)
3822 cpumask_clear_cpu(cpu_going_down, cpumask);
3823
3824 if (cpumask_empty(cpumask))
3825 goto use_dfl;
3826
3827 /* yeap, return possible CPUs in @node that @attrs wants */
3828 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
Michael Bringmann1ad0f0a2017-07-27 16:27:14 -05003829
3830 if (cpumask_empty(cpumask)) {
3831 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3832 "possible intersect\n");
3833 return false;
3834 }
3835
Tejun Heo4c16bd32013-04-01 11:23:36 -07003836 return !cpumask_equal(cpumask, attrs->cpumask);
3837
3838use_dfl:
3839 cpumask_copy(cpumask, attrs->cpumask);
3840 return false;
3841}
3842
Tejun Heo1befcf32013-04-01 11:23:35 -07003843/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3844static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3845 int node,
3846 struct pool_workqueue *pwq)
3847{
3848 struct pool_workqueue *old_pwq;
3849
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +08003850 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003851 lockdep_assert_held(&wq->mutex);
3852
3853 /* link_pwq() can handle duplicate calls */
3854 link_pwq(pwq);
3855
3856 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3857 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3858 return old_pwq;
3859}
3860
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003861/* context to store the prepared attrs & pwqs before applying */
3862struct apply_wqattrs_ctx {
3863 struct workqueue_struct *wq; /* target workqueue */
3864 struct workqueue_attrs *attrs; /* attrs to apply */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003865 struct list_head list; /* queued for batching commit */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003866 struct pool_workqueue *dfl_pwq;
3867 struct pool_workqueue *pwq_tbl[];
3868};
3869
3870/* free the resources after success or abort */
3871static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3872{
3873 if (ctx) {
3874 int node;
3875
3876 for_each_node(node)
3877 put_pwq_unlocked(ctx->pwq_tbl[node]);
3878 put_pwq_unlocked(ctx->dfl_pwq);
3879
3880 free_workqueue_attrs(ctx->attrs);
3881
3882 kfree(ctx);
3883 }
3884}
3885
3886/* allocate the attrs and pwqs for later installation */
3887static struct apply_wqattrs_ctx *
3888apply_wqattrs_prepare(struct workqueue_struct *wq,
3889 const struct workqueue_attrs *attrs)
3890{
3891 struct apply_wqattrs_ctx *ctx;
3892 struct workqueue_attrs *new_attrs, *tmp_attrs;
3893 int node;
3894
3895 lockdep_assert_held(&wq_pool_mutex);
3896
Kees Cookacafe7e2018-05-08 13:45:50 -07003897 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003898
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003899 new_attrs = alloc_workqueue_attrs();
3900 tmp_attrs = alloc_workqueue_attrs();
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003901 if (!ctx || !new_attrs || !tmp_attrs)
3902 goto out_free;
3903
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003904 /*
3905 * Calculate the attrs of the default pwq.
3906 * If the user configured cpumask doesn't overlap with the
3907 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3908 */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003909 copy_workqueue_attrs(new_attrs, attrs);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08003910 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003911 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3912 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003913
3914 /*
3915 * We may create multiple pwqs with differing cpumasks. Make a
3916 * copy of @new_attrs which will be modified and used to obtain
3917 * pools.
3918 */
3919 copy_workqueue_attrs(tmp_attrs, new_attrs);
3920
3921 /*
3922 * If something goes wrong during CPU up/down, we'll fall back to
3923 * the default pwq covering whole @attrs->cpumask. Always create
3924 * it even if we don't use it immediately.
3925 */
3926 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3927 if (!ctx->dfl_pwq)
3928 goto out_free;
3929
3930 for_each_node(node) {
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003931 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003932 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3933 if (!ctx->pwq_tbl[node])
3934 goto out_free;
3935 } else {
3936 ctx->dfl_pwq->refcnt++;
3937 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3938 }
3939 }
3940
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003941 /* save the user configured attrs and sanitize it. */
3942 copy_workqueue_attrs(new_attrs, attrs);
3943 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003944 ctx->attrs = new_attrs;
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003945
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003946 ctx->wq = wq;
3947 free_workqueue_attrs(tmp_attrs);
3948 return ctx;
3949
3950out_free:
3951 free_workqueue_attrs(tmp_attrs);
3952 free_workqueue_attrs(new_attrs);
3953 apply_wqattrs_cleanup(ctx);
3954 return NULL;
3955}
3956
3957/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3958static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3959{
3960 int node;
3961
3962 /* all pwqs have been created successfully, let's install'em */
3963 mutex_lock(&ctx->wq->mutex);
3964
3965 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3966
3967 /* save the previous pwq and install the new one */
3968 for_each_node(node)
3969 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3970 ctx->pwq_tbl[node]);
3971
3972 /* @dfl_pwq might not have been used, ensure it's linked */
3973 link_pwq(ctx->dfl_pwq);
3974 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3975
3976 mutex_unlock(&ctx->wq->mutex);
3977}
3978
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003979static void apply_wqattrs_lock(void)
3980{
3981 /* CPUs should stay stable across pwq creations and installations */
3982 get_online_cpus();
3983 mutex_lock(&wq_pool_mutex);
3984}
3985
3986static void apply_wqattrs_unlock(void)
3987{
3988 mutex_unlock(&wq_pool_mutex);
3989 put_online_cpus();
3990}
3991
3992static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3993 const struct workqueue_attrs *attrs)
3994{
3995 struct apply_wqattrs_ctx *ctx;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003996
3997 /* only unbound workqueues can change attributes */
3998 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3999 return -EINVAL;
4000
4001 /* creating multiple pwqs breaks ordering guarantee */
Tejun Heo0a94efb2017-07-23 08:36:15 -04004002 if (!list_empty(&wq->pwqs)) {
4003 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4004 return -EINVAL;
4005
4006 wq->flags &= ~__WQ_ORDERED;
4007 }
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004008
4009 ctx = apply_wqattrs_prepare(wq, attrs);
wanghaibin62011712016-01-07 20:38:59 +08004010 if (!ctx)
4011 return -ENOMEM;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004012
4013 /* the ctx has been prepared successfully, let's commit it */
wanghaibin62011712016-01-07 20:38:59 +08004014 apply_wqattrs_commit(ctx);
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004015 apply_wqattrs_cleanup(ctx);
4016
wanghaibin62011712016-01-07 20:38:59 +08004017 return 0;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004018}
4019
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004020/**
4021 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4022 * @wq: the target workqueue
4023 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4024 *
Tejun Heo4c16bd32013-04-01 11:23:36 -07004025 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4026 * machines, this function maps a separate pwq to each NUMA node with
4027 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4028 * NUMA node it was issued on. Older pwqs are released as in-flight work
4029 * items finish. Note that a work item which repeatedly requeues itself
4030 * back-to-back will stay on its current pwq.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004031 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004032 * Performs GFP_KERNEL allocations.
4033 *
Daniel Jordan509b3202019-09-05 21:40:23 -04004034 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4035 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004036 * Return: 0 on success and -errno on failure.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004037 */
Daniel Jordan513c98d2019-09-05 21:40:22 -04004038int apply_workqueue_attrs(struct workqueue_struct *wq,
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004039 const struct workqueue_attrs *attrs)
4040{
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004041 int ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004042
Daniel Jordan509b3202019-09-05 21:40:23 -04004043 lockdep_assert_cpus_held();
4044
4045 mutex_lock(&wq_pool_mutex);
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004046 ret = apply_workqueue_attrs_locked(wq, attrs);
Daniel Jordan509b3202019-09-05 21:40:23 -04004047 mutex_unlock(&wq_pool_mutex);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08004048
Tejun Heo48621252013-04-01 11:23:31 -07004049 return ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004050}
4051
Tejun Heo4c16bd32013-04-01 11:23:36 -07004052/**
4053 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4054 * @wq: the target workqueue
4055 * @cpu: the CPU coming up or going down
4056 * @online: whether @cpu is coming up or going down
4057 *
4058 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4059 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4060 * @wq accordingly.
4061 *
4062 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4063 * falls back to @wq->dfl_pwq which may not be optimal but is always
4064 * correct.
4065 *
4066 * Note that when the last allowed CPU of a NUMA node goes offline for a
4067 * workqueue with a cpumask spanning multiple nodes, the workers which were
4068 * already executing the work items for the workqueue will lose their CPU
4069 * affinity and may execute on any CPU. This is similar to how per-cpu
4070 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4071 * affinity, it's the user's responsibility to flush the work item from
4072 * CPU_DOWN_PREPARE.
4073 */
4074static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4075 bool online)
4076{
4077 int node = cpu_to_node(cpu);
4078 int cpu_off = online ? -1 : cpu;
4079 struct pool_workqueue *old_pwq = NULL, *pwq;
4080 struct workqueue_attrs *target_attrs;
4081 cpumask_t *cpumask;
4082
4083 lockdep_assert_held(&wq_pool_mutex);
4084
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004085 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4086 wq->unbound_attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07004087 return;
4088
4089 /*
4090 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4091 * Let's use a preallocated one. The following buf is protected by
4092 * CPU hotplug exclusion.
4093 */
4094 target_attrs = wq_update_unbound_numa_attrs_buf;
4095 cpumask = target_attrs->cpumask;
4096
Tejun Heo4c16bd32013-04-01 11:23:36 -07004097 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4098 pwq = unbound_pwq_by_node(wq, node);
4099
4100 /*
4101 * Let's determine what needs to be done. If the target cpumask is
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004102 * different from the default pwq's, we need to compare it to @pwq's
4103 * and create a new one if they don't match. If the target cpumask
4104 * equals the default pwq's, the default pwq should be used.
Tejun Heo4c16bd32013-04-01 11:23:36 -07004105 */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004106 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
Tejun Heo4c16bd32013-04-01 11:23:36 -07004107 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004108 return;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004109 } else {
Daeseok Youn534a3fb2014-04-18 09:08:14 +09004110 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004111 }
4112
Tejun Heo4c16bd32013-04-01 11:23:36 -07004113 /* create a new pwq */
4114 pwq = alloc_unbound_pwq(wq, target_attrs);
4115 if (!pwq) {
Fabian Frederick2d916032014-05-12 13:59:35 -04004116 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4117 wq->name);
Daeseok Youn77f300b2014-04-16 14:32:29 +09004118 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004119 }
4120
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004121 /* Install the new pwq. */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004122 mutex_lock(&wq->mutex);
4123 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4124 goto out_unlock;
4125
4126use_dfl_pwq:
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004127 mutex_lock(&wq->mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004128 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4129 get_pwq(wq->dfl_pwq);
4130 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4131 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4132out_unlock:
4133 mutex_unlock(&wq->mutex);
4134 put_pwq_unlocked(old_pwq);
4135}
4136
Tejun Heo30cdf2492013-03-12 11:29:57 -07004137static int alloc_and_link_pwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004139 bool highpri = wq->flags & WQ_HIGHPRI;
Tejun Heo8a2b7532013-09-05 12:30:04 -04004140 int cpu, ret;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01004141
Tejun Heo30cdf2492013-03-12 11:29:57 -07004142 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo420c0dd2013-03-12 11:29:59 -07004143 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4144 if (!wq->cpu_pwqs)
Tejun Heo30cdf2492013-03-12 11:29:57 -07004145 return -ENOMEM;
4146
4147 for_each_possible_cpu(cpu) {
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004148 struct pool_workqueue *pwq =
4149 per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heo7a62c2c2013-03-12 11:30:03 -07004150 struct worker_pool *cpu_pools =
Tejun Heof02ae732013-03-12 11:30:03 -07004151 per_cpu(cpu_worker_pools, cpu);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004152
Tejun Heof147f292013-04-01 11:23:35 -07004153 init_pwq(pwq, wq, &cpu_pools[highpri]);
4154
4155 mutex_lock(&wq->mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07004156 link_pwq(pwq);
Tejun Heof147f292013-04-01 11:23:35 -07004157 mutex_unlock(&wq->mutex);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004158 }
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004159 return 0;
Daniel Jordan509b3202019-09-05 21:40:23 -04004160 }
4161
4162 get_online_cpus();
4163 if (wq->flags & __WQ_ORDERED) {
Tejun Heo8a2b7532013-09-05 12:30:04 -04004164 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4165 /* there should only be single pwq for ordering guarantee */
4166 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4167 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4168 "ordering guarantee broken for workqueue %s\n", wq->name);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004169 } else {
Daniel Jordan509b3202019-09-05 21:40:23 -04004170 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004171 }
Daniel Jordan509b3202019-09-05 21:40:23 -04004172 put_online_cpus();
4173
4174 return ret;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004175}
4176
Tejun Heof3421792010-07-02 10:03:51 +02004177static int wq_clamp_max_active(int max_active, unsigned int flags,
4178 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02004179{
Tejun Heof3421792010-07-02 10:03:51 +02004180 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4181
4182 if (max_active < 1 || max_active > lim)
Valentin Ilie044c7822012-08-19 00:52:42 +03004183 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4184 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02004185
Tejun Heof3421792010-07-02 10:03:51 +02004186 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02004187}
4188
Tejun Heo983c7512018-01-08 05:38:32 -08004189/*
4190 * Workqueues which may be used during memory reclaim should have a rescuer
4191 * to guarantee forward progress.
4192 */
4193static int init_rescuer(struct workqueue_struct *wq)
4194{
4195 struct worker *rescuer;
4196 int ret;
4197
4198 if (!(wq->flags & WQ_MEM_RECLAIM))
4199 return 0;
4200
4201 rescuer = alloc_worker(NUMA_NO_NODE);
4202 if (!rescuer)
4203 return -ENOMEM;
4204
4205 rescuer->rescue_wq = wq;
4206 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4207 ret = PTR_ERR_OR_ZERO(rescuer->task);
4208 if (ret) {
4209 kfree(rescuer);
4210 return ret;
4211 }
4212
4213 wq->rescuer = rescuer;
4214 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4215 wake_up_process(rescuer->task);
4216
4217 return 0;
4218}
4219
Mathieu Malaterrea2775bb2019-03-12 21:21:26 +01004220__printf(1, 4)
Bart Van Assche669de8b2019-02-14 15:00:54 -08004221struct workqueue_struct *alloc_workqueue(const char *fmt,
4222 unsigned int flags,
4223 int max_active, ...)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004224{
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004225 size_t tbl_size = 0;
Tejun Heoecf68812013-04-01 11:23:34 -07004226 va_list args;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004227 struct workqueue_struct *wq;
Tejun Heo49e3cf42013-03-12 11:29:58 -07004228 struct pool_workqueue *pwq;
Tejun Heob196be82012-01-10 15:11:35 -08004229
Tejun Heo5c0338c2017-07-18 18:41:52 -04004230 /*
4231 * Unbound && max_active == 1 used to imply ordered, which is no
4232 * longer the case on NUMA machines due to per-node pools. While
4233 * alloc_ordered_workqueue() is the right way to create an ordered
4234 * workqueue, keep the previous behavior to avoid subtle breakages
4235 * on NUMA.
4236 */
4237 if ((flags & WQ_UNBOUND) && max_active == 1)
4238 flags |= __WQ_ORDERED;
4239
Viresh Kumarcee22a12013-04-08 16:45:40 +05304240 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4241 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4242 flags |= WQ_UNBOUND;
4243
Tejun Heoecf68812013-04-01 11:23:34 -07004244 /* allocate wq and format name */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004245 if (flags & WQ_UNBOUND)
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08004246 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004247
4248 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
Tejun Heob196be82012-01-10 15:11:35 -08004249 if (!wq)
Tejun Heod2c1d402013-03-12 11:30:04 -07004250 return NULL;
Tejun Heob196be82012-01-10 15:11:35 -08004251
Tejun Heo6029a912013-04-01 11:23:34 -07004252 if (flags & WQ_UNBOUND) {
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02004253 wq->unbound_attrs = alloc_workqueue_attrs();
Tejun Heo6029a912013-04-01 11:23:34 -07004254 if (!wq->unbound_attrs)
4255 goto err_free_wq;
4256 }
4257
Bart Van Assche669de8b2019-02-14 15:00:54 -08004258 va_start(args, max_active);
Tejun Heoecf68812013-04-01 11:23:34 -07004259 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
Tejun Heob196be82012-01-10 15:11:35 -08004260 va_end(args);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004261
Tejun Heod320c032010-06-29 10:07:14 +02004262 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heob196be82012-01-10 15:11:35 -08004263 max_active = wq_clamp_max_active(max_active, flags, wq->name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004264
Tejun Heob196be82012-01-10 15:11:35 -08004265 /* init wq */
Tejun Heo97e37d72010-06-29 10:07:10 +02004266 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004267 wq->saved_max_active = max_active;
Lai Jiangshan3c25a552013-03-25 16:57:17 -07004268 mutex_init(&wq->mutex);
Tejun Heo112202d2013-02-13 19:29:12 -08004269 atomic_set(&wq->nr_pwqs_to_flush, 0);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004270 INIT_LIST_HEAD(&wq->pwqs);
Tejun Heo73f53c42010-06-29 10:07:11 +02004271 INIT_LIST_HEAD(&wq->flusher_queue);
4272 INIT_LIST_HEAD(&wq->flusher_overflow);
Tejun Heo493a1722013-03-12 11:29:59 -07004273 INIT_LIST_HEAD(&wq->maydays);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004274
Bart Van Assche669de8b2019-02-14 15:00:54 -08004275 wq_init_lockdep(wq);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07004276 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004277
Tejun Heo30cdf2492013-03-12 11:29:57 -07004278 if (alloc_and_link_pwqs(wq) < 0)
Bart Van Assche82efcab2019-03-11 16:02:55 -07004279 goto err_unreg_lockdep;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004280
Tejun Heo40c17f72018-01-08 05:38:37 -08004281 if (wq_online && init_rescuer(wq) < 0)
Tejun Heo983c7512018-01-08 05:38:32 -08004282 goto err_destroy;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004283
Tejun Heo226223a2013-03-12 11:30:05 -07004284 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4285 goto err_destroy;
4286
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004287 /*
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004288 * wq_pool_mutex protects global freeze state and workqueues list.
4289 * Grab it, adjust max_active and add the new @wq to workqueues
4290 * list.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004291 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004292 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004293
Lai Jiangshana357fc02013-03-25 16:57:19 -07004294 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004295 for_each_pwq(pwq, wq)
4296 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004297 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004298
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004299 list_add_tail_rcu(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004300
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004301 mutex_unlock(&wq_pool_mutex);
Tejun Heo15376632010-06-29 10:07:11 +02004302
Oleg Nesterov3af244332007-05-09 02:34:09 -07004303 return wq;
Tejun Heod2c1d402013-03-12 11:30:04 -07004304
Bart Van Assche82efcab2019-03-11 16:02:55 -07004305err_unreg_lockdep:
Bart Van Assche009bb422019-03-03 14:00:46 -08004306 wq_unregister_lockdep(wq);
4307 wq_free_lockdep(wq);
Bart Van Assche82efcab2019-03-11 16:02:55 -07004308err_free_wq:
Tejun Heo6029a912013-04-01 11:23:34 -07004309 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heod2c1d402013-03-12 11:30:04 -07004310 kfree(wq);
4311 return NULL;
4312err_destroy:
4313 destroy_workqueue(wq);
Tejun Heo4690c4a2010-06-29 10:07:10 +02004314 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004315}
Bart Van Assche669de8b2019-02-14 15:00:54 -08004316EXPORT_SYMBOL_GPL(alloc_workqueue);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004317
Tejun Heoc29eb852019-09-23 11:08:58 -07004318static bool pwq_busy(struct pool_workqueue *pwq)
4319{
4320 int i;
4321
4322 for (i = 0; i < WORK_NR_COLORS; i++)
4323 if (pwq->nr_in_flight[i])
4324 return true;
4325
4326 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4327 return true;
4328 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4329 return true;
4330
4331 return false;
4332}
4333
Oleg Nesterov3af244332007-05-09 02:34:09 -07004334/**
4335 * destroy_workqueue - safely terminate a workqueue
4336 * @wq: target workqueue
4337 *
4338 * Safely destroy a workqueue. All work currently pending will be done first.
4339 */
4340void destroy_workqueue(struct workqueue_struct *wq)
4341{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004342 struct pool_workqueue *pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004343 int node;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004344
Tejun Heodef98c82019-09-18 18:43:40 -07004345 /*
4346 * Remove it from sysfs first so that sanity check failure doesn't
4347 * lead to sysfs name conflicts.
4348 */
4349 workqueue_sysfs_unregister(wq);
4350
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02004351 /* drain it before proceeding with destruction */
4352 drain_workqueue(wq);
Tejun Heoc8efcc22010-12-20 19:32:04 +01004353
Tejun Heodef98c82019-09-18 18:43:40 -07004354 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4355 if (wq->rescuer) {
4356 struct worker *rescuer = wq->rescuer;
4357
4358 /* this prevents new queueing */
4359 spin_lock_irq(&wq_mayday_lock);
4360 wq->rescuer = NULL;
4361 spin_unlock_irq(&wq_mayday_lock);
4362
4363 /* rescuer will empty maydays list before exiting */
4364 kthread_stop(rescuer->task);
Tejun Heo8efe1222019-09-20 13:39:57 -07004365 kfree(rescuer);
Tejun Heodef98c82019-09-18 18:43:40 -07004366 }
4367
Tejun Heoc29eb852019-09-23 11:08:58 -07004368 /*
4369 * Sanity checks - grab all the locks so that we wait for all
4370 * in-flight operations which may do put_pwq().
4371 */
4372 mutex_lock(&wq_pool_mutex);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004373 mutex_lock(&wq->mutex);
Tejun Heo49e3cf42013-03-12 11:29:58 -07004374 for_each_pwq(pwq, wq) {
Tejun Heoc29eb852019-09-23 11:08:58 -07004375 spin_lock_irq(&pwq->pool->lock);
4376 if (WARN_ON(pwq_busy(pwq))) {
Kefeng Wang1d9a6152019-11-28 08:47:49 +08004377 pr_warn("%s: %s has the following busy pwq\n",
4378 __func__, wq->name);
Tejun Heoc29eb852019-09-23 11:08:58 -07004379 show_pwq(pwq);
4380 spin_unlock_irq(&pwq->pool->lock);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004381 mutex_unlock(&wq->mutex);
Tejun Heoc29eb852019-09-23 11:08:58 -07004382 mutex_unlock(&wq_pool_mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004383 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004384 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004385 }
Tejun Heoc29eb852019-09-23 11:08:58 -07004386 spin_unlock_irq(&pwq->pool->lock);
Tejun Heo6183c002013-03-12 11:29:57 -07004387 }
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004388 mutex_unlock(&wq->mutex);
Tejun Heoc29eb852019-09-23 11:08:58 -07004389 mutex_unlock(&wq_pool_mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004390
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004391 /*
4392 * wq list is used to freeze wq, remove from list after
4393 * flushing is complete in case freeze races us.
4394 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004395 mutex_lock(&wq_pool_mutex);
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004396 list_del_rcu(&wq->list);
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004397 mutex_unlock(&wq_pool_mutex);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004398
Tejun Heo8864b4e2013-03-12 11:30:04 -07004399 if (!(wq->flags & WQ_UNBOUND)) {
Bart Van Assche669de8b2019-02-14 15:00:54 -08004400 wq_unregister_lockdep(wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07004401 /*
4402 * The base ref is never dropped on per-cpu pwqs. Directly
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004403 * schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004404 */
Paul E. McKenney25b00772018-11-06 19:18:45 -08004405 call_rcu(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07004406 } else {
4407 /*
4408 * We're the sole accessor of @wq at this point. Directly
Tejun Heo4c16bd32013-04-01 11:23:36 -07004409 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4410 * @wq will be freed when the last pwq is released.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004411 */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004412 for_each_node(node) {
4413 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4414 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4415 put_pwq_unlocked(pwq);
4416 }
4417
4418 /*
4419 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4420 * put. Don't access it afterwards.
4421 */
4422 pwq = wq->dfl_pwq;
4423 wq->dfl_pwq = NULL;
Tejun Heodce90d42013-04-01 11:23:35 -07004424 put_pwq_unlocked(pwq);
Tejun Heo29c91e92013-03-12 11:30:03 -07004425 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004426}
4427EXPORT_SYMBOL_GPL(destroy_workqueue);
4428
Tejun Heodcd989c2010-06-29 10:07:14 +02004429/**
4430 * workqueue_set_max_active - adjust max_active of a workqueue
4431 * @wq: target workqueue
4432 * @max_active: new max_active value.
4433 *
4434 * Set max_active of @wq to @max_active.
4435 *
4436 * CONTEXT:
4437 * Don't call from IRQ context.
4438 */
4439void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4440{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004441 struct pool_workqueue *pwq;
Tejun Heodcd989c2010-06-29 10:07:14 +02004442
Tejun Heo8719dce2013-03-12 11:30:04 -07004443 /* disallow meddling with max_active for ordered workqueues */
Tejun Heo0a94efb2017-07-23 08:36:15 -04004444 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Tejun Heo8719dce2013-03-12 11:30:04 -07004445 return;
4446
Tejun Heof3421792010-07-02 10:03:51 +02004447 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02004448
Lai Jiangshana357fc02013-03-25 16:57:19 -07004449 mutex_lock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004450
Tejun Heo0a94efb2017-07-23 08:36:15 -04004451 wq->flags &= ~__WQ_ORDERED;
Tejun Heodcd989c2010-06-29 10:07:14 +02004452 wq->saved_max_active = max_active;
4453
Tejun Heo699ce092013-03-13 16:51:35 -07004454 for_each_pwq(pwq, wq)
4455 pwq_adjust_max_active(pwq);
Tejun Heodcd989c2010-06-29 10:07:14 +02004456
Lai Jiangshana357fc02013-03-25 16:57:19 -07004457 mutex_unlock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004458}
4459EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4460
4461/**
Lukas Wunner27d4ee02018-02-11 10:38:28 +01004462 * current_work - retrieve %current task's work struct
4463 *
4464 * Determine if %current task is a workqueue worker and what it's working on.
4465 * Useful to find out the context that the %current task is running in.
4466 *
4467 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4468 */
4469struct work_struct *current_work(void)
4470{
4471 struct worker *worker = current_wq_worker();
4472
4473 return worker ? worker->current_work : NULL;
4474}
4475EXPORT_SYMBOL(current_work);
4476
4477/**
Tejun Heoe62676162013-03-12 17:41:37 -07004478 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4479 *
4480 * Determine whether %current is a workqueue rescuer. Can be used from
4481 * work functions to determine whether it's being run off the rescuer task.
Yacine Belkadid185af32013-07-31 14:59:24 -07004482 *
4483 * Return: %true if %current is a workqueue rescuer. %false otherwise.
Tejun Heoe62676162013-03-12 17:41:37 -07004484 */
4485bool current_is_workqueue_rescuer(void)
4486{
4487 struct worker *worker = current_wq_worker();
4488
Lai Jiangshan6a092df2013-03-20 03:28:03 +08004489 return worker && worker->rescue_wq;
Tejun Heoe62676162013-03-12 17:41:37 -07004490}
4491
4492/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004493 * workqueue_congested - test whether a workqueue is congested
4494 * @cpu: CPU in question
4495 * @wq: target workqueue
4496 *
4497 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4498 * no synchronization around this function and the test result is
4499 * unreliable and only useful as advisory hints or for debugging.
4500 *
Tejun Heod3251852013-05-10 11:10:17 -07004501 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4502 * Note that both per-cpu and unbound workqueues may be associated with
4503 * multiple pool_workqueues which have separate congested states. A
4504 * workqueue being congested on one CPU doesn't mean the workqueue is also
4505 * contested on other CPUs / NUMA nodes.
4506 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004507 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004508 * %true if congested, %false otherwise.
4509 */
Tejun Heod84ff052013-03-12 11:29:59 -07004510bool workqueue_congested(int cpu, struct workqueue_struct *wq)
Tejun Heodcd989c2010-06-29 10:07:14 +02004511{
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004512 struct pool_workqueue *pwq;
Tejun Heo76af4d92013-03-12 11:30:00 -07004513 bool ret;
4514
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004515 rcu_read_lock();
4516 preempt_disable();
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004517
Tejun Heod3251852013-05-10 11:10:17 -07004518 if (cpu == WORK_CPU_UNBOUND)
4519 cpu = smp_processor_id();
4520
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004521 if (!(wq->flags & WQ_UNBOUND))
4522 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4523 else
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004524 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodcd989c2010-06-29 10:07:14 +02004525
Tejun Heo76af4d92013-03-12 11:30:00 -07004526 ret = !list_empty(&pwq->delayed_works);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004527 preempt_enable();
4528 rcu_read_unlock();
Tejun Heo76af4d92013-03-12 11:30:00 -07004529
4530 return ret;
Tejun Heodcd989c2010-06-29 10:07:14 +02004531}
4532EXPORT_SYMBOL_GPL(workqueue_congested);
4533
4534/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004535 * work_busy - test whether a work is currently pending or running
4536 * @work: the work to be tested
4537 *
4538 * Test whether @work is currently pending or running. There is no
4539 * synchronization around this function and the test result is
4540 * unreliable and only useful as advisory hints or for debugging.
Tejun Heodcd989c2010-06-29 10:07:14 +02004541 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004542 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004543 * OR'd bitmask of WORK_BUSY_* bits.
4544 */
4545unsigned int work_busy(struct work_struct *work)
4546{
Tejun Heofa1b54e2013-03-12 11:30:00 -07004547 struct worker_pool *pool;
Tejun Heodcd989c2010-06-29 10:07:14 +02004548 unsigned long flags;
4549 unsigned int ret = 0;
4550
Tejun Heodcd989c2010-06-29 10:07:14 +02004551 if (work_pending(work))
4552 ret |= WORK_BUSY_PENDING;
Tejun Heodcd989c2010-06-29 10:07:14 +02004553
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004554 rcu_read_lock();
Tejun Heofa1b54e2013-03-12 11:30:00 -07004555 pool = get_work_pool(work);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004556 if (pool) {
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004557 spin_lock_irqsave(&pool->lock, flags);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004558 if (find_worker_executing_work(pool, work))
4559 ret |= WORK_BUSY_RUNNING;
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004560 spin_unlock_irqrestore(&pool->lock, flags);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004561 }
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004562 rcu_read_unlock();
Tejun Heodcd989c2010-06-29 10:07:14 +02004563
4564 return ret;
4565}
4566EXPORT_SYMBOL_GPL(work_busy);
4567
Tejun Heo3d1cb202013-04-30 15:27:22 -07004568/**
4569 * set_worker_desc - set description for the current work item
4570 * @fmt: printf-style format string
4571 * @...: arguments for the format string
4572 *
4573 * This function can be called by a running work function to describe what
4574 * the work item is about. If the worker task gets dumped, this
4575 * information will be printed out together to help debugging. The
4576 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4577 */
4578void set_worker_desc(const char *fmt, ...)
4579{
4580 struct worker *worker = current_wq_worker();
4581 va_list args;
4582
4583 if (worker) {
4584 va_start(args, fmt);
4585 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4586 va_end(args);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004587 }
4588}
Steffen Maier5c750d52018-05-17 19:14:57 +02004589EXPORT_SYMBOL_GPL(set_worker_desc);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004590
4591/**
4592 * print_worker_info - print out worker information and description
4593 * @log_lvl: the log level to use when printing
4594 * @task: target task
4595 *
4596 * If @task is a worker and currently executing a work item, print out the
4597 * name of the workqueue being serviced and worker description set with
4598 * set_worker_desc() by the currently executing work item.
4599 *
4600 * This function can be safely called on any task as long as the
4601 * task_struct itself is accessible. While safe, this function isn't
4602 * synchronized and may print out mixups or garbages of limited length.
4603 */
4604void print_worker_info(const char *log_lvl, struct task_struct *task)
4605{
4606 work_func_t *fn = NULL;
4607 char name[WQ_NAME_LEN] = { };
4608 char desc[WORKER_DESC_LEN] = { };
4609 struct pool_workqueue *pwq = NULL;
4610 struct workqueue_struct *wq = NULL;
Tejun Heo3d1cb202013-04-30 15:27:22 -07004611 struct worker *worker;
4612
4613 if (!(task->flags & PF_WQ_WORKER))
4614 return;
4615
4616 /*
4617 * This function is called without any synchronization and @task
4618 * could be in any state. Be careful with dereferences.
4619 */
Petr Mladeke7005912016-10-11 13:55:17 -07004620 worker = kthread_probe_data(task);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004621
4622 /*
Tejun Heo8bf89592018-05-18 08:47:13 -07004623 * Carefully copy the associated workqueue's workfn, name and desc.
4624 * Keep the original last '\0' in case the original is garbage.
Tejun Heo3d1cb202013-04-30 15:27:22 -07004625 */
4626 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4627 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4628 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4629 probe_kernel_read(name, wq->name, sizeof(name) - 1);
Tejun Heo8bf89592018-05-18 08:47:13 -07004630 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004631
4632 if (fn || name[0] || desc[0]) {
Sakari Ailusd75f7732019-03-25 21:32:28 +02004633 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
Tejun Heo8bf89592018-05-18 08:47:13 -07004634 if (strcmp(name, desc))
Tejun Heo3d1cb202013-04-30 15:27:22 -07004635 pr_cont(" (%s)", desc);
4636 pr_cont("\n");
4637 }
4638}
4639
Tejun Heo3494fc32015-03-09 09:22:28 -04004640static void pr_cont_pool_info(struct worker_pool *pool)
4641{
4642 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4643 if (pool->node != NUMA_NO_NODE)
4644 pr_cont(" node=%d", pool->node);
4645 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4646}
4647
4648static void pr_cont_work(bool comma, struct work_struct *work)
4649{
4650 if (work->func == wq_barrier_func) {
4651 struct wq_barrier *barr;
4652
4653 barr = container_of(work, struct wq_barrier, work);
4654
4655 pr_cont("%s BAR(%d)", comma ? "," : "",
4656 task_pid_nr(barr->task));
4657 } else {
Sakari Ailusd75f7732019-03-25 21:32:28 +02004658 pr_cont("%s %ps", comma ? "," : "", work->func);
Tejun Heo3494fc32015-03-09 09:22:28 -04004659 }
4660}
4661
4662static void show_pwq(struct pool_workqueue *pwq)
4663{
4664 struct worker_pool *pool = pwq->pool;
4665 struct work_struct *work;
4666 struct worker *worker;
4667 bool has_in_flight = false, has_pending = false;
4668 int bkt;
4669
4670 pr_info(" pwq %d:", pool->id);
4671 pr_cont_pool_info(pool);
4672
Tejun Heoe66b39a2019-09-25 06:59:15 -07004673 pr_cont(" active=%d/%d refcnt=%d%s\n",
4674 pwq->nr_active, pwq->max_active, pwq->refcnt,
Tejun Heo3494fc32015-03-09 09:22:28 -04004675 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4676
4677 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4678 if (worker->current_pwq == pwq) {
4679 has_in_flight = true;
4680 break;
4681 }
4682 }
4683 if (has_in_flight) {
4684 bool comma = false;
4685
4686 pr_info(" in-flight:");
4687 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4688 if (worker->current_pwq != pwq)
4689 continue;
4690
Sakari Ailusd75f7732019-03-25 21:32:28 +02004691 pr_cont("%s %d%s:%ps", comma ? "," : "",
Tejun Heo3494fc32015-03-09 09:22:28 -04004692 task_pid_nr(worker->task),
Tejun Heo30ae2fc2019-09-20 14:09:14 -07004693 worker->rescue_wq ? "(RESCUER)" : "",
Tejun Heo3494fc32015-03-09 09:22:28 -04004694 worker->current_func);
4695 list_for_each_entry(work, &worker->scheduled, entry)
4696 pr_cont_work(false, work);
4697 comma = true;
4698 }
4699 pr_cont("\n");
4700 }
4701
4702 list_for_each_entry(work, &pool->worklist, entry) {
4703 if (get_work_pwq(work) == pwq) {
4704 has_pending = true;
4705 break;
4706 }
4707 }
4708 if (has_pending) {
4709 bool comma = false;
4710
4711 pr_info(" pending:");
4712 list_for_each_entry(work, &pool->worklist, entry) {
4713 if (get_work_pwq(work) != pwq)
4714 continue;
4715
4716 pr_cont_work(comma, work);
4717 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4718 }
4719 pr_cont("\n");
4720 }
4721
4722 if (!list_empty(&pwq->delayed_works)) {
4723 bool comma = false;
4724
4725 pr_info(" delayed:");
4726 list_for_each_entry(work, &pwq->delayed_works, entry) {
4727 pr_cont_work(comma, work);
4728 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4729 }
4730 pr_cont("\n");
4731 }
4732}
4733
4734/**
4735 * show_workqueue_state - dump workqueue state
4736 *
Roger Lu7b776af2016-07-01 11:05:02 +08004737 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4738 * all busy workqueues and pools.
Tejun Heo3494fc32015-03-09 09:22:28 -04004739 */
4740void show_workqueue_state(void)
4741{
4742 struct workqueue_struct *wq;
4743 struct worker_pool *pool;
4744 unsigned long flags;
4745 int pi;
4746
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004747 rcu_read_lock();
Tejun Heo3494fc32015-03-09 09:22:28 -04004748
4749 pr_info("Showing busy workqueues and worker pools:\n");
4750
4751 list_for_each_entry_rcu(wq, &workqueues, list) {
4752 struct pool_workqueue *pwq;
4753 bool idle = true;
4754
4755 for_each_pwq(pwq, wq) {
4756 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4757 idle = false;
4758 break;
4759 }
4760 }
4761 if (idle)
4762 continue;
4763
4764 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4765
4766 for_each_pwq(pwq, wq) {
4767 spin_lock_irqsave(&pwq->pool->lock, flags);
4768 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4769 show_pwq(pwq);
4770 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Sergey Senozhatsky62635ea2018-01-11 09:53:35 +09004771 /*
4772 * We could be printing a lot from atomic context, e.g.
4773 * sysrq-t -> show_workqueue_state(). Avoid triggering
4774 * hard lockup.
4775 */
4776 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004777 }
4778 }
4779
4780 for_each_pool(pool, pi) {
4781 struct worker *worker;
4782 bool first = true;
4783
4784 spin_lock_irqsave(&pool->lock, flags);
4785 if (pool->nr_workers == pool->nr_idle)
4786 goto next_pool;
4787
4788 pr_info("pool %d:", pool->id);
4789 pr_cont_pool_info(pool);
Tejun Heo82607adc2015-12-08 11:28:04 -05004790 pr_cont(" hung=%us workers=%d",
4791 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4792 pool->nr_workers);
Tejun Heo3494fc32015-03-09 09:22:28 -04004793 if (pool->manager)
4794 pr_cont(" manager: %d",
4795 task_pid_nr(pool->manager->task));
4796 list_for_each_entry(worker, &pool->idle_list, entry) {
4797 pr_cont(" %s%d", first ? "idle: " : "",
4798 task_pid_nr(worker->task));
4799 first = false;
4800 }
4801 pr_cont("\n");
4802 next_pool:
4803 spin_unlock_irqrestore(&pool->lock, flags);
Sergey Senozhatsky62635ea2018-01-11 09:53:35 +09004804 /*
4805 * We could be printing a lot from atomic context, e.g.
4806 * sysrq-t -> show_workqueue_state(). Avoid triggering
4807 * hard lockup.
4808 */
4809 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004810 }
4811
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004812 rcu_read_unlock();
Tejun Heo3494fc32015-03-09 09:22:28 -04004813}
4814
Tejun Heo6b598082018-05-18 08:47:13 -07004815/* used to show worker information through /proc/PID/{comm,stat,status} */
4816void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4817{
Tejun Heo6b598082018-05-18 08:47:13 -07004818 int off;
4819
4820 /* always show the actual comm */
4821 off = strscpy(buf, task->comm, size);
4822 if (off < 0)
4823 return;
4824
Tejun Heo197f6ac2018-05-21 08:04:35 -07004825 /* stabilize PF_WQ_WORKER and worker pool association */
Tejun Heo6b598082018-05-18 08:47:13 -07004826 mutex_lock(&wq_pool_attach_mutex);
4827
Tejun Heo197f6ac2018-05-21 08:04:35 -07004828 if (task->flags & PF_WQ_WORKER) {
4829 struct worker *worker = kthread_data(task);
4830 struct worker_pool *pool = worker->pool;
Tejun Heo6b598082018-05-18 08:47:13 -07004831
Tejun Heo197f6ac2018-05-21 08:04:35 -07004832 if (pool) {
4833 spin_lock_irq(&pool->lock);
4834 /*
4835 * ->desc tracks information (wq name or
4836 * set_worker_desc()) for the latest execution. If
4837 * current, prepend '+', otherwise '-'.
4838 */
4839 if (worker->desc[0] != '\0') {
4840 if (worker->current_work)
4841 scnprintf(buf + off, size - off, "+%s",
4842 worker->desc);
4843 else
4844 scnprintf(buf + off, size - off, "-%s",
4845 worker->desc);
4846 }
4847 spin_unlock_irq(&pool->lock);
Tejun Heo6b598082018-05-18 08:47:13 -07004848 }
Tejun Heo6b598082018-05-18 08:47:13 -07004849 }
4850
4851 mutex_unlock(&wq_pool_attach_mutex);
4852}
4853
Mathieu Malaterre66448bc2018-05-22 21:47:32 +02004854#ifdef CONFIG_SMP
4855
Tejun Heodb7bccf2010-06-29 10:07:12 +02004856/*
4857 * CPU hotplug.
4858 *
Tejun Heoe22bee72010-06-29 10:07:14 +02004859 * There are two challenges in supporting CPU hotplug. Firstly, there
Tejun Heo112202d2013-02-13 19:29:12 -08004860 * are a lot of assumptions on strong associations among work, pwq and
Tejun Heo706026c2013-01-24 11:01:34 -08004861 * pool which make migrating pending and scheduled works very
Tejun Heoe22bee72010-06-29 10:07:14 +02004862 * difficult to implement without impacting hot paths. Secondly,
Tejun Heo94cf58b2013-01-24 11:01:33 -08004863 * worker pools serve mix of short, long and very long running works making
Tejun Heoe22bee72010-06-29 10:07:14 +02004864 * blocked draining impractical.
4865 *
Tejun Heo24647572013-01-24 11:01:33 -08004866 * This is solved by allowing the pools to be disassociated from the CPU
Tejun Heo628c78e2012-07-17 12:39:27 -07004867 * running as an unbound one and allowing it to be reattached later if the
4868 * cpu comes back online.
Tejun Heodb7bccf2010-06-29 10:07:12 +02004869 */
4870
Lai Jiangshane8b3f8d2017-12-01 22:20:36 +08004871static void unbind_workers(int cpu)
Tejun Heodb7bccf2010-06-29 10:07:12 +02004872{
Tejun Heo4ce62e92012-07-13 22:16:44 -07004873 struct worker_pool *pool;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004874 struct worker *worker;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004875
Tejun Heof02ae732013-03-12 11:30:03 -07004876 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo1258fae2018-05-18 08:47:13 -07004877 mutex_lock(&wq_pool_attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004878 spin_lock_irq(&pool->lock);
4879
4880 /*
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004881 * We've blocked all attach/detach operations. Make all workers
Tejun Heo94cf58b2013-01-24 11:01:33 -08004882 * unbound and set DISASSOCIATED. Before this, all workers
4883 * except for the ones which are still executing works from
4884 * before the last CPU down must be on the cpu. After
4885 * this, they may become diasporas.
4886 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004887 for_each_pool_worker(worker, pool)
Tejun Heoc9e7cf22013-01-24 11:01:33 -08004888 worker->flags |= WORKER_UNBOUND;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004889
Tejun Heo24647572013-01-24 11:01:33 -08004890 pool->flags |= POOL_DISASSOCIATED;
Tejun Heof2d5a0e2012-07-17 12:39:26 -07004891
Tejun Heo94cf58b2013-01-24 11:01:33 -08004892 spin_unlock_irq(&pool->lock);
Tejun Heo1258fae2018-05-18 08:47:13 -07004893 mutex_unlock(&wq_pool_attach_mutex);
Tejun Heoe22bee72010-06-29 10:07:14 +02004894
Lai Jiangshaneb283422013-03-08 15:18:28 -08004895 /*
4896 * Call schedule() so that we cross rq->lock and thus can
4897 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4898 * This is necessary as scheduler callbacks may be invoked
4899 * from other cpus.
4900 */
4901 schedule();
Tejun Heo628c78e2012-07-17 12:39:27 -07004902
Lai Jiangshaneb283422013-03-08 15:18:28 -08004903 /*
4904 * Sched callbacks are disabled now. Zap nr_running.
4905 * After this, nr_running stays zero and need_more_worker()
4906 * and keep_working() are always true as long as the
4907 * worklist is not empty. This pool now behaves as an
4908 * unbound (in terms of concurrency management) pool which
4909 * are served by workers tied to the pool.
4910 */
Tejun Heoe19e3972013-01-24 11:39:44 -08004911 atomic_set(&pool->nr_running, 0);
Lai Jiangshaneb283422013-03-08 15:18:28 -08004912
4913 /*
4914 * With concurrency management just turned off, a busy
4915 * worker blocking could lead to lengthy stalls. Kick off
4916 * unbound chain execution of currently pending work items.
4917 */
4918 spin_lock_irq(&pool->lock);
4919 wake_up_worker(pool);
4920 spin_unlock_irq(&pool->lock);
4921 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004922}
4923
Tejun Heobd7c0892013-03-19 13:45:21 -07004924/**
4925 * rebind_workers - rebind all workers of a pool to the associated CPU
4926 * @pool: pool of interest
4927 *
Tejun Heoa9ab7752013-03-19 13:45:21 -07004928 * @pool->cpu is coming online. Rebind all workers to the CPU.
Tejun Heobd7c0892013-03-19 13:45:21 -07004929 */
4930static void rebind_workers(struct worker_pool *pool)
4931{
Tejun Heoa9ab7752013-03-19 13:45:21 -07004932 struct worker *worker;
Tejun Heobd7c0892013-03-19 13:45:21 -07004933
Tejun Heo1258fae2018-05-18 08:47:13 -07004934 lockdep_assert_held(&wq_pool_attach_mutex);
Tejun Heobd7c0892013-03-19 13:45:21 -07004935
Tejun Heoa9ab7752013-03-19 13:45:21 -07004936 /*
4937 * Restore CPU affinity of all workers. As all idle workers should
4938 * be on the run-queue of the associated CPU before any local
Shailendra Verma402dd892015-05-23 10:38:14 +05304939 * wake-ups for concurrency management happen, restore CPU affinity
Tejun Heoa9ab7752013-03-19 13:45:21 -07004940 * of all workers first and then clear UNBOUND. As we're called
4941 * from CPU_ONLINE, the following shouldn't fail.
4942 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004943 for_each_pool_worker(worker, pool)
Tejun Heoa9ab7752013-03-19 13:45:21 -07004944 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4945 pool->attrs->cpumask) < 0);
4946
4947 spin_lock_irq(&pool->lock);
Wanpeng Lif7c17d22016-05-11 17:55:18 +08004948
Lai Jiangshan3de5e882014-06-03 15:33:27 +08004949 pool->flags &= ~POOL_DISASSOCIATED;
Tejun Heoa9ab7752013-03-19 13:45:21 -07004950
Lai Jiangshanda028462014-05-20 17:46:31 +08004951 for_each_pool_worker(worker, pool) {
Tejun Heoa9ab7752013-03-19 13:45:21 -07004952 unsigned int worker_flags = worker->flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004953
4954 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07004955 * A bound idle worker should actually be on the runqueue
4956 * of the associated CPU for local wake-ups targeting it to
4957 * work. Kick all idle workers so that they migrate to the
4958 * associated CPU. Doing this in the same loop as
4959 * replacing UNBOUND with REBOUND is safe as no worker will
4960 * be bound before @pool->lock is released.
Tejun Heobd7c0892013-03-19 13:45:21 -07004961 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004962 if (worker_flags & WORKER_IDLE)
4963 wake_up_process(worker->task);
4964
4965 /*
4966 * We want to clear UNBOUND but can't directly call
4967 * worker_clr_flags() or adjust nr_running. Atomically
4968 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4969 * @worker will clear REBOUND using worker_clr_flags() when
4970 * it initiates the next execution cycle thus restoring
4971 * concurrency management. Note that when or whether
4972 * @worker clears REBOUND doesn't affect correctness.
4973 *
Mark Rutlandc95491e2017-10-23 14:07:22 -07004974 * WRITE_ONCE() is necessary because @worker->flags may be
Tejun Heoa9ab7752013-03-19 13:45:21 -07004975 * tested without holding any lock in
Thomas Gleixner6d25be52019-03-13 17:55:48 +01004976 * wq_worker_running(). Without it, NOT_RUNNING test may
Tejun Heoa9ab7752013-03-19 13:45:21 -07004977 * fail incorrectly leading to premature concurrency
4978 * management operations.
4979 */
4980 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4981 worker_flags |= WORKER_REBOUND;
4982 worker_flags &= ~WORKER_UNBOUND;
Mark Rutlandc95491e2017-10-23 14:07:22 -07004983 WRITE_ONCE(worker->flags, worker_flags);
Tejun Heobd7c0892013-03-19 13:45:21 -07004984 }
4985
Tejun Heoa9ab7752013-03-19 13:45:21 -07004986 spin_unlock_irq(&pool->lock);
Tejun Heobd7c0892013-03-19 13:45:21 -07004987}
4988
Tejun Heo7dbc7252013-03-19 13:45:21 -07004989/**
4990 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4991 * @pool: unbound pool of interest
4992 * @cpu: the CPU which is coming up
4993 *
4994 * An unbound pool may end up with a cpumask which doesn't have any online
4995 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4996 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4997 * online CPU before, cpus_allowed of all its workers should be restored.
4998 */
4999static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5000{
5001 static cpumask_t cpumask;
5002 struct worker *worker;
Tejun Heo7dbc7252013-03-19 13:45:21 -07005003
Tejun Heo1258fae2018-05-18 08:47:13 -07005004 lockdep_assert_held(&wq_pool_attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005005
5006 /* is @cpu allowed for @pool? */
5007 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5008 return;
5009
Tejun Heo7dbc7252013-03-19 13:45:21 -07005010 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005011
5012 /* as we're called from CPU_ONLINE, the following shouldn't fail */
Lai Jiangshanda028462014-05-20 17:46:31 +08005013 for_each_pool_worker(worker, pool)
Peter Zijlstrad945b5e2016-06-16 14:38:42 +02005014 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005015}
5016
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005017int workqueue_prepare_cpu(unsigned int cpu)
Oleg Nesterov3af244332007-05-09 02:34:09 -07005018{
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005019 struct worker_pool *pool;
5020
5021 for_each_cpu_worker_pool(pool, cpu) {
5022 if (pool->nr_workers)
5023 continue;
5024 if (!create_worker(pool))
5025 return -ENOMEM;
5026 }
5027 return 0;
5028}
5029
5030int workqueue_online_cpu(unsigned int cpu)
5031{
Tejun Heo4ce62e92012-07-13 22:16:44 -07005032 struct worker_pool *pool;
Tejun Heo4c16bd32013-04-01 11:23:36 -07005033 struct workqueue_struct *wq;
Tejun Heo7dbc7252013-03-19 13:45:21 -07005034 int pi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005035
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005036 mutex_lock(&wq_pool_mutex);
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07005037
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005038 for_each_pool(pool, pi) {
Tejun Heo1258fae2018-05-18 08:47:13 -07005039 mutex_lock(&wq_pool_attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005040
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005041 if (pool->cpu == cpu)
5042 rebind_workers(pool);
5043 else if (pool->cpu < 0)
5044 restore_unbound_workers_cpumask(pool, cpu);
Tejun Heo94cf58b2013-01-24 11:01:33 -08005045
Tejun Heo1258fae2018-05-18 08:47:13 -07005046 mutex_unlock(&wq_pool_attach_mutex);
Tejun Heo65758202012-07-17 12:39:26 -07005047 }
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005048
5049 /* update NUMA affinity of unbound workqueues */
5050 list_for_each_entry(wq, &workqueues, list)
5051 wq_update_unbound_numa(wq, cpu, true);
5052
5053 mutex_unlock(&wq_pool_mutex);
5054 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07005055}
5056
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005057int workqueue_offline_cpu(unsigned int cpu)
Tejun Heo65758202012-07-17 12:39:26 -07005058{
Tejun Heo4c16bd32013-04-01 11:23:36 -07005059 struct workqueue_struct *wq;
Tejun Heo8db25e72012-07-17 12:39:28 -07005060
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005061 /* unbinding per-cpu workers should happen on the local CPU */
Lai Jiangshane8b3f8d2017-12-01 22:20:36 +08005062 if (WARN_ON(cpu != smp_processor_id()))
5063 return -1;
5064
5065 unbind_workers(cpu);
Tejun Heo4c16bd32013-04-01 11:23:36 -07005066
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005067 /* update NUMA affinity of unbound workqueues */
5068 mutex_lock(&wq_pool_mutex);
5069 list_for_each_entry(wq, &workqueues, list)
5070 wq_update_unbound_numa(wq, cpu, false);
5071 mutex_unlock(&wq_pool_mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07005072
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005073 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07005074}
5075
Rusty Russell2d3854a2008-11-05 13:39:10 +11005076struct work_for_cpu {
Tejun Heoed48ece2012-09-18 12:48:43 -07005077 struct work_struct work;
Rusty Russell2d3854a2008-11-05 13:39:10 +11005078 long (*fn)(void *);
5079 void *arg;
5080 long ret;
5081};
5082
Tejun Heoed48ece2012-09-18 12:48:43 -07005083static void work_for_cpu_fn(struct work_struct *work)
Rusty Russell2d3854a2008-11-05 13:39:10 +11005084{
Tejun Heoed48ece2012-09-18 12:48:43 -07005085 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5086
Rusty Russell2d3854a2008-11-05 13:39:10 +11005087 wfc->ret = wfc->fn(wfc->arg);
5088}
5089
5090/**
Anna-Maria Gleixner22aceb32016-03-10 12:07:38 +01005091 * work_on_cpu - run a function in thread context on a particular cpu
Rusty Russell2d3854a2008-11-05 13:39:10 +11005092 * @cpu: the cpu to run on
5093 * @fn: the function to run
5094 * @arg: the function arg
5095 *
Rusty Russell31ad9082009-01-16 15:31:15 -08005096 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06005097 * The caller must not hold any locks which would prevent @fn from completing.
Yacine Belkadid185af32013-07-31 14:59:24 -07005098 *
5099 * Return: The value @fn returns.
Rusty Russell2d3854a2008-11-05 13:39:10 +11005100 */
Tejun Heod84ff052013-03-12 11:29:59 -07005101long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
Rusty Russell2d3854a2008-11-05 13:39:10 +11005102{
Tejun Heoed48ece2012-09-18 12:48:43 -07005103 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
Rusty Russell2d3854a2008-11-05 13:39:10 +11005104
Tejun Heoed48ece2012-09-18 12:48:43 -07005105 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5106 schedule_work_on(cpu, &wfc.work);
Bjorn Helgaas12997d12013-11-18 11:00:29 -07005107 flush_work(&wfc.work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05005108 destroy_work_on_stack(&wfc.work);
Rusty Russell2d3854a2008-11-05 13:39:10 +11005109 return wfc.ret;
5110}
5111EXPORT_SYMBOL_GPL(work_on_cpu);
Thomas Gleixner0e8d6a92017-04-12 22:07:28 +02005112
5113/**
5114 * work_on_cpu_safe - run a function in thread context on a particular cpu
5115 * @cpu: the cpu to run on
5116 * @fn: the function to run
5117 * @arg: the function argument
5118 *
5119 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5120 * any locks which would prevent @fn from completing.
5121 *
5122 * Return: The value @fn returns.
5123 */
5124long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5125{
5126 long ret = -ENODEV;
5127
5128 get_online_cpus();
5129 if (cpu_online(cpu))
5130 ret = work_on_cpu(cpu, fn, arg);
5131 put_online_cpus();
5132 return ret;
5133}
5134EXPORT_SYMBOL_GPL(work_on_cpu_safe);
Rusty Russell2d3854a2008-11-05 13:39:10 +11005135#endif /* CONFIG_SMP */
5136
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005137#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10305138
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005139/**
5140 * freeze_workqueues_begin - begin freezing workqueues
5141 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01005142 * Start freezing workqueues. After this function returns, all freezable
Tejun Heoc5aa87b2013-03-13 16:51:36 -07005143 * workqueues will queue new works to their delayed_works list instead of
Tejun Heo706026c2013-01-24 11:01:34 -08005144 * pool->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005145 *
5146 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07005147 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005148 */
5149void freeze_workqueues_begin(void)
5150{
Tejun Heo24b8a842013-03-12 11:29:58 -07005151 struct workqueue_struct *wq;
5152 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005153
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005154 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005155
Tejun Heo6183c002013-03-12 11:29:57 -07005156 WARN_ON_ONCE(workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005157 workqueue_freezing = true;
5158
Tejun Heo24b8a842013-03-12 11:29:58 -07005159 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07005160 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07005161 for_each_pwq(pwq, wq)
5162 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07005163 mutex_unlock(&wq->mutex);
Tejun Heo24b8a842013-03-12 11:29:58 -07005164 }
Tejun Heo5bcab332013-03-13 19:47:40 -07005165
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005166 mutex_unlock(&wq_pool_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005167}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005168
5169/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01005170 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005171 *
5172 * Check whether freezing is complete. This function must be called
5173 * between freeze_workqueues_begin() and thaw_workqueues().
5174 *
5175 * CONTEXT:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005176 * Grabs and releases wq_pool_mutex.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005177 *
Yacine Belkadid185af32013-07-31 14:59:24 -07005178 * Return:
Tejun Heo58a69cb2011-02-16 09:25:31 +01005179 * %true if some freezable workqueues are still busy. %false if freezing
5180 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005181 */
5182bool freeze_workqueues_busy(void)
5183{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005184 bool busy = false;
Tejun Heo24b8a842013-03-12 11:29:58 -07005185 struct workqueue_struct *wq;
5186 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005187
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005188 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005189
Tejun Heo6183c002013-03-12 11:29:57 -07005190 WARN_ON_ONCE(!workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005191
Tejun Heo24b8a842013-03-12 11:29:58 -07005192 list_for_each_entry(wq, &workqueues, list) {
5193 if (!(wq->flags & WQ_FREEZABLE))
5194 continue;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005195 /*
5196 * nr_active is monotonically decreasing. It's safe
5197 * to peek without lock.
5198 */
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005199 rcu_read_lock();
Tejun Heo24b8a842013-03-12 11:29:58 -07005200 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07005201 WARN_ON_ONCE(pwq->nr_active < 0);
Tejun Heo112202d2013-02-13 19:29:12 -08005202 if (pwq->nr_active) {
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005203 busy = true;
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005204 rcu_read_unlock();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005205 goto out_unlock;
5206 }
5207 }
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005208 rcu_read_unlock();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005209 }
5210out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005211 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005212 return busy;
5213}
5214
5215/**
5216 * thaw_workqueues - thaw workqueues
5217 *
5218 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo706026c2013-01-24 11:01:34 -08005219 * frozen works are transferred to their respective pool worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005220 *
5221 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07005222 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005223 */
5224void thaw_workqueues(void)
5225{
Tejun Heo24b8a842013-03-12 11:29:58 -07005226 struct workqueue_struct *wq;
5227 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005228
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005229 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005230
5231 if (!workqueue_freezing)
5232 goto out_unlock;
5233
Lai Jiangshan74b414e2014-05-22 19:01:16 +08005234 workqueue_freezing = false;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005235
Tejun Heo24b8a842013-03-12 11:29:58 -07005236 /* restore max_active and repopulate worklist */
5237 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07005238 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07005239 for_each_pwq(pwq, wq)
5240 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07005241 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005242 }
5243
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005244out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005245 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005246}
5247#endif /* CONFIG_FREEZER */
5248
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005249static int workqueue_apply_unbound_cpumask(void)
5250{
5251 LIST_HEAD(ctxs);
5252 int ret = 0;
5253 struct workqueue_struct *wq;
5254 struct apply_wqattrs_ctx *ctx, *n;
5255
5256 lockdep_assert_held(&wq_pool_mutex);
5257
5258 list_for_each_entry(wq, &workqueues, list) {
5259 if (!(wq->flags & WQ_UNBOUND))
5260 continue;
5261 /* creating multiple pwqs breaks ordering guarantee */
5262 if (wq->flags & __WQ_ORDERED)
5263 continue;
5264
5265 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5266 if (!ctx) {
5267 ret = -ENOMEM;
5268 break;
5269 }
5270
5271 list_add_tail(&ctx->list, &ctxs);
5272 }
5273
5274 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5275 if (!ret)
5276 apply_wqattrs_commit(ctx);
5277 apply_wqattrs_cleanup(ctx);
5278 }
5279
5280 return ret;
5281}
5282
5283/**
5284 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5285 * @cpumask: the cpumask to set
5286 *
5287 * The low-level workqueues cpumask is a global cpumask that limits
5288 * the affinity of all unbound workqueues. This function check the @cpumask
5289 * and apply it to all unbound workqueues and updates all pwqs of them.
5290 *
5291 * Retun: 0 - Success
5292 * -EINVAL - Invalid @cpumask
5293 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5294 */
5295int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5296{
5297 int ret = -EINVAL;
5298 cpumask_var_t saved_cpumask;
5299
5300 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5301 return -ENOMEM;
5302
Tal Shorerc98a9802017-11-03 17:27:50 +02005303 /*
5304 * Not excluding isolated cpus on purpose.
5305 * If the user wishes to include them, we allow that.
5306 */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005307 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5308 if (!cpumask_empty(cpumask)) {
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005309 apply_wqattrs_lock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005310
5311 /* save the old wq_unbound_cpumask. */
5312 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5313
5314 /* update wq_unbound_cpumask at first and apply it to wqs. */
5315 cpumask_copy(wq_unbound_cpumask, cpumask);
5316 ret = workqueue_apply_unbound_cpumask();
5317
5318 /* restore the wq_unbound_cpumask when failed. */
5319 if (ret < 0)
5320 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5321
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005322 apply_wqattrs_unlock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005323 }
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005324
5325 free_cpumask_var(saved_cpumask);
5326 return ret;
5327}
5328
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005329#ifdef CONFIG_SYSFS
5330/*
5331 * Workqueues with WQ_SYSFS flag set is visible to userland via
5332 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5333 * following attributes.
5334 *
5335 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5336 * max_active RW int : maximum number of in-flight work items
5337 *
5338 * Unbound workqueues have the following extra attributes.
5339 *
Wang Long9a19b462017-11-02 23:05:12 -04005340 * pool_ids RO int : the associated pool IDs for each node
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005341 * nice RW int : nice value of the workers
5342 * cpumask RW mask : bitmask of allowed CPUs for the workers
Wang Long9a19b462017-11-02 23:05:12 -04005343 * numa RW bool : whether enable NUMA affinity
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005344 */
5345struct wq_device {
5346 struct workqueue_struct *wq;
5347 struct device dev;
5348};
5349
5350static struct workqueue_struct *dev_to_wq(struct device *dev)
5351{
5352 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5353
5354 return wq_dev->wq;
5355}
5356
5357static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5358 char *buf)
5359{
5360 struct workqueue_struct *wq = dev_to_wq(dev);
5361
5362 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5363}
5364static DEVICE_ATTR_RO(per_cpu);
5365
5366static ssize_t max_active_show(struct device *dev,
5367 struct device_attribute *attr, char *buf)
5368{
5369 struct workqueue_struct *wq = dev_to_wq(dev);
5370
5371 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5372}
5373
5374static ssize_t max_active_store(struct device *dev,
5375 struct device_attribute *attr, const char *buf,
5376 size_t count)
5377{
5378 struct workqueue_struct *wq = dev_to_wq(dev);
5379 int val;
5380
5381 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5382 return -EINVAL;
5383
5384 workqueue_set_max_active(wq, val);
5385 return count;
5386}
5387static DEVICE_ATTR_RW(max_active);
5388
5389static struct attribute *wq_sysfs_attrs[] = {
5390 &dev_attr_per_cpu.attr,
5391 &dev_attr_max_active.attr,
5392 NULL,
5393};
5394ATTRIBUTE_GROUPS(wq_sysfs);
5395
5396static ssize_t wq_pool_ids_show(struct device *dev,
5397 struct device_attribute *attr, char *buf)
5398{
5399 struct workqueue_struct *wq = dev_to_wq(dev);
5400 const char *delim = "";
5401 int node, written = 0;
5402
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005403 get_online_cpus();
5404 rcu_read_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005405 for_each_node(node) {
5406 written += scnprintf(buf + written, PAGE_SIZE - written,
5407 "%s%d:%d", delim, node,
5408 unbound_pwq_by_node(wq, node)->pool->id);
5409 delim = " ";
5410 }
5411 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005412 rcu_read_unlock();
5413 put_online_cpus();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005414
5415 return written;
5416}
5417
5418static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5419 char *buf)
5420{
5421 struct workqueue_struct *wq = dev_to_wq(dev);
5422 int written;
5423
5424 mutex_lock(&wq->mutex);
5425 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5426 mutex_unlock(&wq->mutex);
5427
5428 return written;
5429}
5430
5431/* prepare workqueue_attrs for sysfs store operations */
5432static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5433{
5434 struct workqueue_attrs *attrs;
5435
Lai Jiangshan899a94f2015-05-20 14:41:18 +08005436 lockdep_assert_held(&wq_pool_mutex);
5437
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005438 attrs = alloc_workqueue_attrs();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005439 if (!attrs)
5440 return NULL;
5441
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005442 copy_workqueue_attrs(attrs, wq->unbound_attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005443 return attrs;
5444}
5445
5446static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5447 const char *buf, size_t count)
5448{
5449 struct workqueue_struct *wq = dev_to_wq(dev);
5450 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005451 int ret = -ENOMEM;
5452
5453 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005454
5455 attrs = wq_sysfs_prep_attrs(wq);
5456 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005457 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005458
5459 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5460 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005461 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005462 else
5463 ret = -EINVAL;
5464
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005465out_unlock:
5466 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005467 free_workqueue_attrs(attrs);
5468 return ret ?: count;
5469}
5470
5471static ssize_t wq_cpumask_show(struct device *dev,
5472 struct device_attribute *attr, char *buf)
5473{
5474 struct workqueue_struct *wq = dev_to_wq(dev);
5475 int written;
5476
5477 mutex_lock(&wq->mutex);
5478 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5479 cpumask_pr_args(wq->unbound_attrs->cpumask));
5480 mutex_unlock(&wq->mutex);
5481 return written;
5482}
5483
5484static ssize_t wq_cpumask_store(struct device *dev,
5485 struct device_attribute *attr,
5486 const char *buf, size_t count)
5487{
5488 struct workqueue_struct *wq = dev_to_wq(dev);
5489 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005490 int ret = -ENOMEM;
5491
5492 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005493
5494 attrs = wq_sysfs_prep_attrs(wq);
5495 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005496 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005497
5498 ret = cpumask_parse(buf, attrs->cpumask);
5499 if (!ret)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005500 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005501
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005502out_unlock:
5503 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005504 free_workqueue_attrs(attrs);
5505 return ret ?: count;
5506}
5507
5508static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5509 char *buf)
5510{
5511 struct workqueue_struct *wq = dev_to_wq(dev);
5512 int written;
5513
5514 mutex_lock(&wq->mutex);
5515 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5516 !wq->unbound_attrs->no_numa);
5517 mutex_unlock(&wq->mutex);
5518
5519 return written;
5520}
5521
5522static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5523 const char *buf, size_t count)
5524{
5525 struct workqueue_struct *wq = dev_to_wq(dev);
5526 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005527 int v, ret = -ENOMEM;
5528
5529 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005530
5531 attrs = wq_sysfs_prep_attrs(wq);
5532 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005533 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005534
5535 ret = -EINVAL;
5536 if (sscanf(buf, "%d", &v) == 1) {
5537 attrs->no_numa = !v;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005538 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005539 }
5540
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005541out_unlock:
5542 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005543 free_workqueue_attrs(attrs);
5544 return ret ?: count;
5545}
5546
5547static struct device_attribute wq_sysfs_unbound_attrs[] = {
5548 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5549 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5550 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5551 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5552 __ATTR_NULL,
5553};
5554
5555static struct bus_type wq_subsys = {
5556 .name = "workqueue",
5557 .dev_groups = wq_sysfs_groups,
5558};
5559
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005560static ssize_t wq_unbound_cpumask_show(struct device *dev,
5561 struct device_attribute *attr, char *buf)
5562{
5563 int written;
5564
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005565 mutex_lock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005566 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5567 cpumask_pr_args(wq_unbound_cpumask));
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005568 mutex_unlock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005569
5570 return written;
5571}
5572
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005573static ssize_t wq_unbound_cpumask_store(struct device *dev,
5574 struct device_attribute *attr, const char *buf, size_t count)
5575{
5576 cpumask_var_t cpumask;
5577 int ret;
5578
5579 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5580 return -ENOMEM;
5581
5582 ret = cpumask_parse(buf, cpumask);
5583 if (!ret)
5584 ret = workqueue_set_unbound_cpumask(cpumask);
5585
5586 free_cpumask_var(cpumask);
5587 return ret ? ret : count;
5588}
5589
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005590static struct device_attribute wq_sysfs_cpumask_attr =
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005591 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5592 wq_unbound_cpumask_store);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005593
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005594static int __init wq_sysfs_init(void)
5595{
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005596 int err;
5597
5598 err = subsys_virtual_register(&wq_subsys, NULL);
5599 if (err)
5600 return err;
5601
5602 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005603}
5604core_initcall(wq_sysfs_init);
5605
5606static void wq_device_release(struct device *dev)
5607{
5608 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5609
5610 kfree(wq_dev);
5611}
5612
5613/**
5614 * workqueue_sysfs_register - make a workqueue visible in sysfs
5615 * @wq: the workqueue to register
5616 *
5617 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5618 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5619 * which is the preferred method.
5620 *
5621 * Workqueue user should use this function directly iff it wants to apply
5622 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5623 * apply_workqueue_attrs() may race against userland updating the
5624 * attributes.
5625 *
5626 * Return: 0 on success, -errno on failure.
5627 */
5628int workqueue_sysfs_register(struct workqueue_struct *wq)
5629{
5630 struct wq_device *wq_dev;
5631 int ret;
5632
5633 /*
Shailendra Verma402dd892015-05-23 10:38:14 +05305634 * Adjusting max_active or creating new pwqs by applying
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005635 * attributes breaks ordering guarantee. Disallow exposing ordered
5636 * workqueues.
5637 */
Tejun Heo0a94efb2017-07-23 08:36:15 -04005638 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005639 return -EINVAL;
5640
5641 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5642 if (!wq_dev)
5643 return -ENOMEM;
5644
5645 wq_dev->wq = wq;
5646 wq_dev->dev.bus = &wq_subsys;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005647 wq_dev->dev.release = wq_device_release;
Lars-Peter Clausen23217b42016-02-17 21:04:41 +01005648 dev_set_name(&wq_dev->dev, "%s", wq->name);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005649
5650 /*
5651 * unbound_attrs are created separately. Suppress uevent until
5652 * everything is ready.
5653 */
5654 dev_set_uevent_suppress(&wq_dev->dev, true);
5655
5656 ret = device_register(&wq_dev->dev);
5657 if (ret) {
Arvind Yadav537f4142018-03-06 15:35:43 +05305658 put_device(&wq_dev->dev);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005659 wq->wq_dev = NULL;
5660 return ret;
5661 }
5662
5663 if (wq->flags & WQ_UNBOUND) {
5664 struct device_attribute *attr;
5665
5666 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5667 ret = device_create_file(&wq_dev->dev, attr);
5668 if (ret) {
5669 device_unregister(&wq_dev->dev);
5670 wq->wq_dev = NULL;
5671 return ret;
5672 }
5673 }
5674 }
5675
5676 dev_set_uevent_suppress(&wq_dev->dev, false);
5677 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5678 return 0;
5679}
5680
5681/**
5682 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5683 * @wq: the workqueue to unregister
5684 *
5685 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5686 */
5687static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5688{
5689 struct wq_device *wq_dev = wq->wq_dev;
5690
5691 if (!wq->wq_dev)
5692 return;
5693
5694 wq->wq_dev = NULL;
5695 device_unregister(&wq_dev->dev);
5696}
5697#else /* CONFIG_SYSFS */
5698static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5699#endif /* CONFIG_SYSFS */
5700
Tejun Heo82607adc2015-12-08 11:28:04 -05005701/*
5702 * Workqueue watchdog.
5703 *
5704 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5705 * flush dependency, a concurrency managed work item which stays RUNNING
5706 * indefinitely. Workqueue stalls can be very difficult to debug as the
5707 * usual warning mechanisms don't trigger and internal workqueue state is
5708 * largely opaque.
5709 *
5710 * Workqueue watchdog monitors all worker pools periodically and dumps
5711 * state if some pools failed to make forward progress for a while where
5712 * forward progress is defined as the first item on ->worklist changing.
5713 *
5714 * This mechanism is controlled through the kernel parameter
5715 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5716 * corresponding sysfs parameter file.
5717 */
5718#ifdef CONFIG_WQ_WATCHDOG
5719
Tejun Heo82607adc2015-12-08 11:28:04 -05005720static unsigned long wq_watchdog_thresh = 30;
Kees Cook5cd79d62017-10-04 16:27:00 -07005721static struct timer_list wq_watchdog_timer;
Tejun Heo82607adc2015-12-08 11:28:04 -05005722
5723static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5724static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5725
5726static void wq_watchdog_reset_touched(void)
5727{
5728 int cpu;
5729
5730 wq_watchdog_touched = jiffies;
5731 for_each_possible_cpu(cpu)
5732 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5733}
5734
Kees Cook5cd79d62017-10-04 16:27:00 -07005735static void wq_watchdog_timer_fn(struct timer_list *unused)
Tejun Heo82607adc2015-12-08 11:28:04 -05005736{
5737 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5738 bool lockup_detected = false;
5739 struct worker_pool *pool;
5740 int pi;
5741
5742 if (!thresh)
5743 return;
5744
5745 rcu_read_lock();
5746
5747 for_each_pool(pool, pi) {
5748 unsigned long pool_ts, touched, ts;
5749
5750 if (list_empty(&pool->worklist))
5751 continue;
5752
5753 /* get the latest of pool and touched timestamps */
5754 pool_ts = READ_ONCE(pool->watchdog_ts);
5755 touched = READ_ONCE(wq_watchdog_touched);
5756
5757 if (time_after(pool_ts, touched))
5758 ts = pool_ts;
5759 else
5760 ts = touched;
5761
5762 if (pool->cpu >= 0) {
5763 unsigned long cpu_touched =
5764 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5765 pool->cpu));
5766 if (time_after(cpu_touched, ts))
5767 ts = cpu_touched;
5768 }
5769
5770 /* did we stall? */
5771 if (time_after(jiffies, ts + thresh)) {
5772 lockup_detected = true;
5773 pr_emerg("BUG: workqueue lockup - pool");
5774 pr_cont_pool_info(pool);
5775 pr_cont(" stuck for %us!\n",
5776 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5777 }
5778 }
5779
5780 rcu_read_unlock();
5781
5782 if (lockup_detected)
5783 show_workqueue_state();
5784
5785 wq_watchdog_reset_touched();
5786 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5787}
5788
Vincent Whitchurchcb9d7fd2018-08-21 17:25:07 +02005789notrace void wq_watchdog_touch(int cpu)
Tejun Heo82607adc2015-12-08 11:28:04 -05005790{
5791 if (cpu >= 0)
5792 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5793 else
5794 wq_watchdog_touched = jiffies;
5795}
5796
5797static void wq_watchdog_set_thresh(unsigned long thresh)
5798{
5799 wq_watchdog_thresh = 0;
5800 del_timer_sync(&wq_watchdog_timer);
5801
5802 if (thresh) {
5803 wq_watchdog_thresh = thresh;
5804 wq_watchdog_reset_touched();
5805 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5806 }
5807}
5808
5809static int wq_watchdog_param_set_thresh(const char *val,
5810 const struct kernel_param *kp)
5811{
5812 unsigned long thresh;
5813 int ret;
5814
5815 ret = kstrtoul(val, 0, &thresh);
5816 if (ret)
5817 return ret;
5818
5819 if (system_wq)
5820 wq_watchdog_set_thresh(thresh);
5821 else
5822 wq_watchdog_thresh = thresh;
5823
5824 return 0;
5825}
5826
5827static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5828 .set = wq_watchdog_param_set_thresh,
5829 .get = param_get_ulong,
5830};
5831
5832module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5833 0644);
5834
5835static void wq_watchdog_init(void)
5836{
Kees Cook5cd79d62017-10-04 16:27:00 -07005837 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
Tejun Heo82607adc2015-12-08 11:28:04 -05005838 wq_watchdog_set_thresh(wq_watchdog_thresh);
5839}
5840
5841#else /* CONFIG_WQ_WATCHDOG */
5842
5843static inline void wq_watchdog_init(void) { }
5844
5845#endif /* CONFIG_WQ_WATCHDOG */
5846
Tejun Heobce90382013-04-01 11:23:32 -07005847static void __init wq_numa_init(void)
5848{
5849 cpumask_var_t *tbl;
5850 int node, cpu;
5851
Tejun Heobce90382013-04-01 11:23:32 -07005852 if (num_possible_nodes() <= 1)
5853 return;
5854
Tejun Heod55262c2013-04-01 11:23:38 -07005855 if (wq_disable_numa) {
5856 pr_info("workqueue: NUMA affinity support disabled\n");
5857 return;
5858 }
5859
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005860 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
Tejun Heo4c16bd32013-04-01 11:23:36 -07005861 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5862
Tejun Heobce90382013-04-01 11:23:32 -07005863 /*
5864 * We want masks of possible CPUs of each node which isn't readily
5865 * available. Build one from cpu_to_node() which should have been
5866 * fully initialized by now.
5867 */
Kees Cook6396bb22018-06-12 14:03:40 -07005868 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
Tejun Heobce90382013-04-01 11:23:32 -07005869 BUG_ON(!tbl);
5870
5871 for_each_node(node)
Yasuaki Ishimatsu5a6024f2014-07-07 09:56:48 -04005872 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
Tejun Heo1be0c252013-05-15 14:24:24 -07005873 node_online(node) ? node : NUMA_NO_NODE));
Tejun Heobce90382013-04-01 11:23:32 -07005874
5875 for_each_possible_cpu(cpu) {
5876 node = cpu_to_node(cpu);
5877 if (WARN_ON(node == NUMA_NO_NODE)) {
5878 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5879 /* happens iff arch is bonkers, let's just proceed */
5880 return;
5881 }
5882 cpumask_set_cpu(cpu, tbl[node]);
5883 }
5884
5885 wq_numa_possible_cpumask = tbl;
5886 wq_numa_enabled = true;
5887}
5888
Tejun Heo3347fa02016-09-16 15:49:32 -04005889/**
5890 * workqueue_init_early - early init for workqueue subsystem
5891 *
5892 * This is the first half of two-staged workqueue subsystem initialization
5893 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5894 * idr are up. It sets up all the data structures and system workqueues
5895 * and allows early boot code to create workqueues and queue/cancel work
5896 * items. Actual work item execution starts only after kthreads can be
5897 * created and scheduled right before early initcalls.
5898 */
5899int __init workqueue_init_early(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005900{
Tejun Heo7a4e3442013-03-12 11:30:00 -07005901 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
Frederic Weisbecker1bda3f82018-02-21 05:17:26 +01005902 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
Tejun Heo7a4e3442013-03-12 11:30:00 -07005903 int i, cpu;
Tejun Heoc34056a2010-06-29 10:07:11 +02005904
Tejun Heoe904e6c2013-03-12 11:29:57 -07005905 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5906
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005907 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
Frederic Weisbecker1bda3f82018-02-21 05:17:26 +01005908 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005909
Tejun Heoe904e6c2013-03-12 11:29:57 -07005910 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5911
Tejun Heo706026c2013-01-24 11:01:34 -08005912 /* initialize CPU pools */
Tejun Heo29c91e92013-03-12 11:30:03 -07005913 for_each_possible_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005914 struct worker_pool *pool;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005915
Tejun Heo7a4e3442013-03-12 11:30:00 -07005916 i = 0;
Tejun Heof02ae732013-03-12 11:30:03 -07005917 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo7a4e3442013-03-12 11:30:00 -07005918 BUG_ON(init_worker_pool(pool));
Tejun Heoec22ca52013-01-24 11:01:33 -08005919 pool->cpu = cpu;
Tejun Heo29c91e92013-03-12 11:30:03 -07005920 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
Tejun Heo7a4e3442013-03-12 11:30:00 -07005921 pool->attrs->nice = std_nice[i++];
Tejun Heof3f90ad2013-04-01 11:23:34 -07005922 pool->node = cpu_to_node(cpu);
Tejun Heo7a4e3442013-03-12 11:30:00 -07005923
Tejun Heo9daf9e62013-01-24 11:01:33 -08005924 /* alloc pool ID */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005925 mutex_lock(&wq_pool_mutex);
Tejun Heo9daf9e62013-01-24 11:01:33 -08005926 BUG_ON(worker_pool_assign_id(pool));
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005927 mutex_unlock(&wq_pool_mutex);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005928 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02005929 }
5930
Tejun Heo8a2b7532013-09-05 12:30:04 -04005931 /* create default unbound and ordered wq attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -07005932 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5933 struct workqueue_attrs *attrs;
5934
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005935 BUG_ON(!(attrs = alloc_workqueue_attrs()));
Tejun Heo29c91e92013-03-12 11:30:03 -07005936 attrs->nice = std_nice[i];
Tejun Heo29c91e92013-03-12 11:30:03 -07005937 unbound_std_wq_attrs[i] = attrs;
Tejun Heo8a2b7532013-09-05 12:30:04 -04005938
5939 /*
5940 * An ordered wq should have only one pwq as ordering is
5941 * guaranteed by max_active which is enforced by pwqs.
5942 * Turn off NUMA so that dfl_pwq is used for all nodes.
5943 */
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005944 BUG_ON(!(attrs = alloc_workqueue_attrs()));
Tejun Heo8a2b7532013-09-05 12:30:04 -04005945 attrs->nice = std_nice[i];
5946 attrs->no_numa = true;
5947 ordered_wq_attrs[i] = attrs;
Tejun Heo29c91e92013-03-12 11:30:03 -07005948 }
5949
Tejun Heod320c032010-06-29 10:07:14 +02005950 system_wq = alloc_workqueue("events", 0, 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005951 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
Tejun Heod320c032010-06-29 10:07:14 +02005952 system_long_wq = alloc_workqueue("events_long", 0, 0);
Tejun Heof3421792010-07-02 10:03:51 +02005953 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5954 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01005955 system_freezable_wq = alloc_workqueue("events_freezable",
5956 WQ_FREEZABLE, 0);
Viresh Kumar06681062013-04-24 17:12:54 +05305957 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5958 WQ_POWER_EFFICIENT, 0);
5959 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5960 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5961 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005962 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
Viresh Kumar06681062013-04-24 17:12:54 +05305963 !system_unbound_wq || !system_freezable_wq ||
5964 !system_power_efficient_wq ||
5965 !system_freezable_power_efficient_wq);
Tejun Heo82607adc2015-12-08 11:28:04 -05005966
Tejun Heo3347fa02016-09-16 15:49:32 -04005967 return 0;
5968}
5969
5970/**
5971 * workqueue_init - bring workqueue subsystem fully online
5972 *
5973 * This is the latter half of two-staged workqueue subsystem initialization
5974 * and invoked as soon as kthreads can be created and scheduled.
5975 * Workqueues have been created and work items queued on them, but there
5976 * are no kworkers executing the work items yet. Populate the worker pools
5977 * with the initial workers and enable future kworker creations.
5978 */
5979int __init workqueue_init(void)
5980{
Tejun Heo2186d9f2016-10-19 12:01:27 -04005981 struct workqueue_struct *wq;
Tejun Heo3347fa02016-09-16 15:49:32 -04005982 struct worker_pool *pool;
5983 int cpu, bkt;
5984
Tejun Heo2186d9f2016-10-19 12:01:27 -04005985 /*
5986 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5987 * CPU to node mapping may not be available that early on some
5988 * archs such as power and arm64. As per-cpu pools created
5989 * previously could be missing node hint and unbound pools NUMA
5990 * affinity, fix them up.
Tejun Heo40c17f72018-01-08 05:38:37 -08005991 *
5992 * Also, while iterating workqueues, create rescuers if requested.
Tejun Heo2186d9f2016-10-19 12:01:27 -04005993 */
5994 wq_numa_init();
5995
5996 mutex_lock(&wq_pool_mutex);
5997
5998 for_each_possible_cpu(cpu) {
5999 for_each_cpu_worker_pool(pool, cpu) {
6000 pool->node = cpu_to_node(cpu);
6001 }
6002 }
6003
Tejun Heo40c17f72018-01-08 05:38:37 -08006004 list_for_each_entry(wq, &workqueues, list) {
Tejun Heo2186d9f2016-10-19 12:01:27 -04006005 wq_update_unbound_numa(wq, smp_processor_id(), true);
Tejun Heo40c17f72018-01-08 05:38:37 -08006006 WARN(init_rescuer(wq),
6007 "workqueue: failed to create early rescuer for %s",
6008 wq->name);
6009 }
Tejun Heo2186d9f2016-10-19 12:01:27 -04006010
6011 mutex_unlock(&wq_pool_mutex);
6012
Tejun Heo3347fa02016-09-16 15:49:32 -04006013 /* create the initial workers */
6014 for_each_online_cpu(cpu) {
6015 for_each_cpu_worker_pool(pool, cpu) {
6016 pool->flags &= ~POOL_DISASSOCIATED;
6017 BUG_ON(!create_worker(pool));
6018 }
6019 }
6020
6021 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6022 BUG_ON(!create_worker(pool));
6023
6024 wq_online = true;
Tejun Heo82607adc2015-12-08 11:28:04 -05006025 wq_watchdog_init();
6026
Suresh Siddha6ee05782010-07-30 14:57:37 -07006027 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006028}