blob: ddf0537dce140dd88e042a1395d5ca90657d9c06 [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 Heo2eaebdb2013-01-18 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 Heo1e19ffc2010-06-29 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 Heo1e19ffc2010-06-29 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 Heo794b18bc2013-03-13 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 Heo974271c42012-07-12 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 Heo974271c42012-07-12 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
Sebastian Andrzej Siewior62849a92020-03-28 00:29:59 +0100861 * going to sleep. Preemption needs to be disabled to protect ->sleeping
862 * assignment.
Tejun Heoe22bee72010-06-29 10:07:14 +0200863 */
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100864void wq_worker_sleeping(struct task_struct *task)
Tejun Heoe22bee72010-06-29 10:07:14 +0200865{
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100866 struct worker *next, *worker = kthread_data(task);
Tejun Heo111c2252013-01-17 17:16:24 -0800867 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200868
Tejun Heo111c2252013-01-17 17:16:24 -0800869 /*
870 * Rescuers, which may not have all the fields set up like normal
871 * workers, also reach here, let's not access anything before
872 * checking NOT_RUNNING.
873 */
Steven Rostedt2d646722010-12-03 23:12:33 -0500874 if (worker->flags & WORKER_NOT_RUNNING)
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100875 return;
Tejun Heoe22bee72010-06-29 10:07:14 +0200876
Tejun Heo111c2252013-01-17 17:16:24 -0800877 pool = worker->pool;
Tejun Heo111c2252013-01-17 17:16:24 -0800878
Sebastian Andrzej Siewior62849a92020-03-28 00:29:59 +0100879 /* Return if preempted before wq_worker_running() was reached */
880 if (worker->sleeping)
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100881 return;
882
883 worker->sleeping = 1;
884 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +0200885
886 /*
887 * The counterpart of the following dec_and_test, implied mb,
888 * worklist not empty test sequence is in insert_work().
889 * Please read comment there.
890 *
Tejun Heo628c78e2012-07-17 12:39:27 -0700891 * NOT_RUNNING is clear. This means that we're bound to and
892 * running on the local cpu w/ rq lock held and preemption
893 * disabled, which in turn means that none else could be
Tejun Heod565ed62013-01-24 11:01:33 -0800894 * manipulating idle_list, so dereferencing idle_list without pool
Tejun Heo628c78e2012-07-17 12:39:27 -0700895 * lock is safe.
Tejun Heoe22bee72010-06-29 10:07:14 +0200896 */
Tejun Heoe19e3972013-01-24 11:39:44 -0800897 if (atomic_dec_and_test(&pool->nr_running) &&
Thomas Gleixner6d25be52019-03-13 17:55:48 +0100898 !list_empty(&pool->worklist)) {
899 next = first_idle_worker(pool);
900 if (next)
901 wake_up_process(next->task);
902 }
903 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +0200904}
905
906/**
Johannes Weiner1b69ac62019-02-01 14:20:42 -0800907 * wq_worker_last_func - retrieve worker's last work function
Bart Van Assche8194fe92019-03-19 10:45:09 -0700908 * @task: Task to retrieve last work function of.
Johannes Weiner1b69ac62019-02-01 14:20:42 -0800909 *
910 * Determine the last function a worker executed. This is called from
911 * the scheduler to get a worker's last known identity.
912 *
913 * CONTEXT:
914 * spin_lock_irq(rq->lock)
915 *
Johannes Weiner4b047002019-03-07 16:29:30 -0800916 * This function is called during schedule() when a kworker is going
917 * to sleep. It's used by psi to identify aggregation workers during
918 * dequeuing, to allow periodic aggregation to shut-off when that
919 * worker is the last task in the system or cgroup to go to sleep.
920 *
921 * As this function doesn't involve any workqueue-related locking, it
922 * only returns stable values when called from inside the scheduler's
923 * queuing and dequeuing paths, when @task, which must be a kworker,
924 * is guaranteed to not be processing any works.
925 *
Johannes Weiner1b69ac62019-02-01 14:20:42 -0800926 * Return:
927 * The last work function %current executed as a worker, NULL if it
928 * hasn't executed any work yet.
929 */
930work_func_t wq_worker_last_func(struct task_struct *task)
931{
932 struct worker *worker = kthread_data(task);
933
934 return worker->last_func;
935}
936
937/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200938 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200939 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200940 * @flags: flags to set
Tejun Heod302f012010-06-29 10:07:13 +0200941 *
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800942 * Set @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200943 *
Tejun Heocb444762010-07-02 10:03:50 +0200944 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800945 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200946 */
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800947static inline void worker_set_flags(struct worker *worker, unsigned int flags)
Tejun Heod302f012010-06-29 10:07:13 +0200948{
Tejun Heobd7bdd42012-07-12 14:46:37 -0700949 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200950
Tejun Heocb444762010-07-02 10:03:50 +0200951 WARN_ON_ONCE(worker->task != current);
952
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800953 /* If transitioning into NOT_RUNNING, adjust nr_running. */
Tejun Heoe22bee72010-06-29 10:07:14 +0200954 if ((flags & WORKER_NOT_RUNNING) &&
955 !(worker->flags & WORKER_NOT_RUNNING)) {
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800956 atomic_dec(&pool->nr_running);
Tejun Heoe22bee72010-06-29 10:07:14 +0200957 }
958
Tejun Heod302f012010-06-29 10:07:13 +0200959 worker->flags |= flags;
960}
961
962/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200963 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200964 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200965 * @flags: flags to clear
966 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200967 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200968 *
Tejun Heocb444762010-07-02 10:03:50 +0200969 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800970 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200971 */
972static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
973{
Tejun Heo63d95a92012-07-12 14:46:37 -0700974 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200975 unsigned int oflags = worker->flags;
976
Tejun Heocb444762010-07-02 10:03:50 +0200977 WARN_ON_ONCE(worker->task != current);
978
Tejun Heod302f012010-06-29 10:07:13 +0200979 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200980
Tejun Heo42c025f2011-01-11 15:58:49 +0100981 /*
982 * If transitioning out of NOT_RUNNING, increment nr_running. Note
983 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
984 * of multiple flags, not a single flag.
985 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200986 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
987 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800988 atomic_inc(&pool->nr_running);
Tejun Heod302f012010-06-29 10:07:13 +0200989}
990
991/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200992 * find_worker_executing_work - find worker which is executing a work
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800993 * @pool: pool of interest
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200994 * @work: work to find worker for
995 *
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800996 * Find a worker which is executing @work on @pool by searching
997 * @pool->busy_hash which is keyed by the address of @work. For a worker
Tejun Heoa2c1c572012-12-18 10:35:02 -0800998 * to match, its current execution should match the address of @work and
999 * its work function. This is to avoid unwanted dependency between
1000 * unrelated work executions through a work item being recycled while still
1001 * being executed.
1002 *
1003 * This is a bit tricky. A work item may be freed once its execution
1004 * starts and nothing prevents the freed area from being recycled for
1005 * another work item. If the same work item address ends up being reused
1006 * before the original execution finishes, workqueue will identify the
1007 * recycled work item as currently executing and make it wait until the
1008 * current execution finishes, introducing an unwanted dependency.
1009 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001010 * This function checks the work item address and work function to avoid
1011 * false positives. Note that this isn't complete as one may construct a
1012 * work function which can introduce dependency onto itself through a
1013 * recycled work item. Well, if somebody wants to shoot oneself in the
1014 * foot that badly, there's only so much we can do, and if such deadlock
1015 * actually occurs, it should be easy to locate the culprit work function.
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001016 *
1017 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001018 * spin_lock_irq(pool->lock).
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001019 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001020 * Return:
1021 * Pointer to worker which is executing @work if found, %NULL
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001022 * otherwise.
1023 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08001024static struct worker *find_worker_executing_work(struct worker_pool *pool,
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001025 struct work_struct *work)
1026{
Sasha Levin42f85702012-12-17 10:01:23 -05001027 struct worker *worker;
Sasha Levin42f85702012-12-17 10:01:23 -05001028
Sasha Levinb67bfe02013-02-27 17:06:00 -08001029 hash_for_each_possible(pool->busy_hash, worker, hentry,
Tejun Heoa2c1c572012-12-18 10:35:02 -08001030 (unsigned long)work)
1031 if (worker->current_work == work &&
1032 worker->current_func == work->func)
Sasha Levin42f85702012-12-17 10:01:23 -05001033 return worker;
1034
1035 return NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001036}
1037
1038/**
Tejun Heobf4ede02012-08-03 10:30:46 -07001039 * move_linked_works - move linked works to a list
1040 * @work: start of series of works to be scheduled
1041 * @head: target list to append @work to
Shailendra Verma402dd892015-05-23 10:38:14 +05301042 * @nextp: out parameter for nested worklist walking
Tejun Heobf4ede02012-08-03 10:30:46 -07001043 *
1044 * Schedule linked works starting from @work to @head. Work series to
1045 * be scheduled starts at @work and includes any consecutive work with
1046 * WORK_STRUCT_LINKED set in its predecessor.
1047 *
1048 * If @nextp is not NULL, it's updated to point to the next work of
1049 * the last scheduled work. This allows move_linked_works() to be
1050 * nested inside outer list_for_each_entry_safe().
1051 *
1052 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001053 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001054 */
1055static void move_linked_works(struct work_struct *work, struct list_head *head,
1056 struct work_struct **nextp)
1057{
1058 struct work_struct *n;
1059
1060 /*
1061 * Linked worklist will always end before the end of the list,
1062 * use NULL for list head.
1063 */
1064 list_for_each_entry_safe_from(work, n, NULL, entry) {
1065 list_move_tail(&work->entry, head);
1066 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1067 break;
1068 }
1069
1070 /*
1071 * If we're already inside safe list traversal and have moved
1072 * multiple works to the scheduled queue, the next position
1073 * needs to be updated.
1074 */
1075 if (nextp)
1076 *nextp = n;
1077}
1078
Tejun Heo8864b4e2013-03-12 11:30:04 -07001079/**
1080 * get_pwq - get an extra reference on the specified pool_workqueue
1081 * @pwq: pool_workqueue to get
1082 *
1083 * Obtain an extra reference on @pwq. The caller should guarantee that
1084 * @pwq has positive refcnt and be holding the matching pool->lock.
1085 */
1086static void get_pwq(struct pool_workqueue *pwq)
1087{
1088 lockdep_assert_held(&pwq->pool->lock);
1089 WARN_ON_ONCE(pwq->refcnt <= 0);
1090 pwq->refcnt++;
1091}
1092
1093/**
1094 * put_pwq - put a pool_workqueue reference
1095 * @pwq: pool_workqueue to put
1096 *
1097 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1098 * destruction. The caller should be holding the matching pool->lock.
1099 */
1100static void put_pwq(struct pool_workqueue *pwq)
1101{
1102 lockdep_assert_held(&pwq->pool->lock);
1103 if (likely(--pwq->refcnt))
1104 return;
1105 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1106 return;
1107 /*
1108 * @pwq can't be released under pool->lock, bounce to
1109 * pwq_unbound_release_workfn(). This never recurses on the same
1110 * pool->lock as this path is taken only for unbound workqueues and
1111 * the release work item is scheduled on a per-cpu workqueue. To
1112 * avoid lockdep warning, unbound pool->locks are given lockdep
1113 * subclass of 1 in get_unbound_pool().
1114 */
1115 schedule_work(&pwq->unbound_release_work);
1116}
1117
Tejun Heodce90d42013-04-01 11:23:35 -07001118/**
1119 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1120 * @pwq: pool_workqueue to put (can be %NULL)
1121 *
1122 * put_pwq() with locking. This function also allows %NULL @pwq.
1123 */
1124static void put_pwq_unlocked(struct pool_workqueue *pwq)
1125{
1126 if (pwq) {
1127 /*
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001128 * As both pwqs and pools are RCU protected, the
Tejun Heodce90d42013-04-01 11:23:35 -07001129 * following lock operations are safe.
1130 */
1131 spin_lock_irq(&pwq->pool->lock);
1132 put_pwq(pwq);
1133 spin_unlock_irq(&pwq->pool->lock);
1134 }
1135}
1136
Tejun Heo112202d2013-02-13 19:29:12 -08001137static void pwq_activate_delayed_work(struct work_struct *work)
Tejun Heobf4ede02012-08-03 10:30:46 -07001138{
Tejun Heo112202d2013-02-13 19:29:12 -08001139 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobf4ede02012-08-03 10:30:46 -07001140
1141 trace_workqueue_activate_work(work);
Tejun Heo82607adc2015-12-08 11:28:04 -05001142 if (list_empty(&pwq->pool->worklist))
1143 pwq->pool->watchdog_ts = jiffies;
Tejun Heo112202d2013-02-13 19:29:12 -08001144 move_linked_works(work, &pwq->pool->worklist, NULL);
Tejun Heobf4ede02012-08-03 10:30:46 -07001145 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo112202d2013-02-13 19:29:12 -08001146 pwq->nr_active++;
Tejun Heobf4ede02012-08-03 10:30:46 -07001147}
1148
Tejun Heo112202d2013-02-13 19:29:12 -08001149static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001150{
Tejun Heo112202d2013-02-13 19:29:12 -08001151 struct work_struct *work = list_first_entry(&pwq->delayed_works,
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001152 struct work_struct, entry);
1153
Tejun Heo112202d2013-02-13 19:29:12 -08001154 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001155}
1156
Tejun Heobf4ede02012-08-03 10:30:46 -07001157/**
Tejun Heo112202d2013-02-13 19:29:12 -08001158 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1159 * @pwq: pwq of interest
Tejun Heobf4ede02012-08-03 10:30:46 -07001160 * @color: color of work which left the queue
Tejun Heobf4ede02012-08-03 10:30:46 -07001161 *
1162 * A work either has completed or is removed from pending queue,
Tejun Heo112202d2013-02-13 19:29:12 -08001163 * decrement nr_in_flight of its pwq and handle workqueue flushing.
Tejun Heobf4ede02012-08-03 10:30:46 -07001164 *
1165 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001166 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001167 */
Tejun Heo112202d2013-02-13 19:29:12 -08001168static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
Tejun Heobf4ede02012-08-03 10:30:46 -07001169{
Tejun Heo8864b4e2013-03-12 11:30:04 -07001170 /* uncolored work items don't participate in flushing or nr_active */
Tejun Heobf4ede02012-08-03 10:30:46 -07001171 if (color == WORK_NO_COLOR)
Tejun Heo8864b4e2013-03-12 11:30:04 -07001172 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001173
Tejun Heo112202d2013-02-13 19:29:12 -08001174 pwq->nr_in_flight[color]--;
Tejun Heobf4ede02012-08-03 10:30:46 -07001175
Tejun Heo112202d2013-02-13 19:29:12 -08001176 pwq->nr_active--;
1177 if (!list_empty(&pwq->delayed_works)) {
Lai Jiangshanb3f9f402012-09-18 10:40:00 -07001178 /* one down, submit a delayed one */
Tejun Heo112202d2013-02-13 19:29:12 -08001179 if (pwq->nr_active < pwq->max_active)
1180 pwq_activate_first_delayed(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001181 }
1182
1183 /* is flush in progress and are we at the flushing tip? */
Tejun Heo112202d2013-02-13 19:29:12 -08001184 if (likely(pwq->flush_color != color))
Tejun Heo8864b4e2013-03-12 11:30:04 -07001185 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001186
1187 /* are there still in-flight works? */
Tejun Heo112202d2013-02-13 19:29:12 -08001188 if (pwq->nr_in_flight[color])
Tejun Heo8864b4e2013-03-12 11:30:04 -07001189 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001190
Tejun Heo112202d2013-02-13 19:29:12 -08001191 /* this pwq is done, clear flush_color */
1192 pwq->flush_color = -1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001193
1194 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001195 * If this was the last pwq, wake up the first flusher. It
Tejun Heobf4ede02012-08-03 10:30:46 -07001196 * will handle the rest.
1197 */
Tejun Heo112202d2013-02-13 19:29:12 -08001198 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1199 complete(&pwq->wq->first_flusher->done);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001200out_put:
1201 put_pwq(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001202}
1203
Tejun Heo36e227d2012-08-03 10:30:46 -07001204/**
Tejun Heobbb68df2012-08-03 10:30:46 -07001205 * try_to_grab_pending - steal work item from worklist and disable irq
Tejun Heo36e227d2012-08-03 10:30:46 -07001206 * @work: work item to steal
1207 * @is_dwork: @work is a delayed_work
Tejun Heobbb68df2012-08-03 10:30:46 -07001208 * @flags: place to store irq state
Tejun Heo36e227d2012-08-03 10:30:46 -07001209 *
1210 * Try to grab PENDING bit of @work. This function can handle @work in any
Yacine Belkadid185af32013-07-31 14:59:24 -07001211 * stable state - idle, on timer or on worklist.
Tejun Heo36e227d2012-08-03 10:30:46 -07001212 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001213 * Return:
Tejun Heo36e227d2012-08-03 10:30:46 -07001214 * 1 if @work was pending and we successfully stole PENDING
1215 * 0 if @work was idle and we claimed PENDING
1216 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
Tejun Heobbb68df2012-08-03 10:30:46 -07001217 * -ENOENT if someone else is canceling @work, this state may persist
1218 * for arbitrarily long
Tejun Heo36e227d2012-08-03 10:30:46 -07001219 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001220 * Note:
Tejun Heobbb68df2012-08-03 10:30:46 -07001221 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001222 * interrupted while holding PENDING and @work off queue, irq must be
1223 * disabled on entry. This, combined with delayed_work->timer being
1224 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
Tejun Heobbb68df2012-08-03 10:30:46 -07001225 *
1226 * On successful return, >= 0, irq is disabled and the caller is
1227 * responsible for releasing it using local_irq_restore(*@flags).
1228 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001229 * This function is safe to call from any context including IRQ handler.
Tejun Heobf4ede02012-08-03 10:30:46 -07001230 */
Tejun Heobbb68df2012-08-03 10:30:46 -07001231static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1232 unsigned long *flags)
Tejun Heobf4ede02012-08-03 10:30:46 -07001233{
Tejun Heod565ed62013-01-24 11:01:33 -08001234 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001235 struct pool_workqueue *pwq;
Tejun Heobf4ede02012-08-03 10:30:46 -07001236
Tejun Heobbb68df2012-08-03 10:30:46 -07001237 local_irq_save(*flags);
1238
Tejun Heo36e227d2012-08-03 10:30:46 -07001239 /* try to steal the timer if it exists */
1240 if (is_dwork) {
1241 struct delayed_work *dwork = to_delayed_work(work);
1242
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001243 /*
1244 * dwork->timer is irqsafe. If del_timer() fails, it's
1245 * guaranteed that the timer is not queued anywhere and not
1246 * running on the local CPU.
1247 */
Tejun Heo36e227d2012-08-03 10:30:46 -07001248 if (likely(del_timer(&dwork->timer)))
1249 return 1;
1250 }
1251
1252 /* try to claim PENDING the normal way */
Tejun Heobf4ede02012-08-03 10:30:46 -07001253 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1254 return 0;
1255
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001256 rcu_read_lock();
Tejun Heobf4ede02012-08-03 10:30:46 -07001257 /*
1258 * The queueing is in progress, or it is already queued. Try to
1259 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1260 */
Tejun Heod565ed62013-01-24 11:01:33 -08001261 pool = get_work_pool(work);
1262 if (!pool)
Tejun Heobbb68df2012-08-03 10:30:46 -07001263 goto fail;
Tejun Heobf4ede02012-08-03 10:30:46 -07001264
Tejun Heod565ed62013-01-24 11:01:33 -08001265 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001266 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001267 * work->data is guaranteed to point to pwq only while the work
1268 * item is queued on pwq->wq, and both updating work->data to point
1269 * to pwq on queueing and to pool on dequeueing are done under
1270 * pwq->pool->lock. This in turn guarantees that, if work->data
1271 * points to pwq which is associated with a locked pool, the work
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001272 * item is currently queued on that pool.
1273 */
Tejun Heo112202d2013-02-13 19:29:12 -08001274 pwq = get_work_pwq(work);
1275 if (pwq && pwq->pool == pool) {
Tejun Heo16062832013-02-06 18:04:53 -08001276 debug_work_deactivate(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001277
Tejun Heo16062832013-02-06 18:04:53 -08001278 /*
1279 * A delayed work item cannot be grabbed directly because
1280 * it might have linked NO_COLOR work items which, if left
Tejun Heo112202d2013-02-13 19:29:12 -08001281 * on the delayed_list, will confuse pwq->nr_active
Tejun Heo16062832013-02-06 18:04:53 -08001282 * management later on and cause stall. Make sure the work
1283 * item is activated before grabbing.
1284 */
1285 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
Tejun Heo112202d2013-02-13 19:29:12 -08001286 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001287
Tejun Heo16062832013-02-06 18:04:53 -08001288 list_del_init(&work->entry);
Lai Jiangshan9c34a702014-07-11 00:11:13 +08001289 pwq_dec_nr_in_flight(pwq, get_work_color(work));
Tejun Heo36e227d2012-08-03 10:30:46 -07001290
Tejun Heo112202d2013-02-13 19:29:12 -08001291 /* work->data points to pwq iff queued, point to pool */
Tejun Heo16062832013-02-06 18:04:53 -08001292 set_work_pool_and_keep_pending(work, pool->id);
Lai Jiangshan4468a002013-02-06 18:04:53 -08001293
Tejun Heo16062832013-02-06 18:04:53 -08001294 spin_unlock(&pool->lock);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001295 rcu_read_unlock();
Tejun Heo16062832013-02-06 18:04:53 -08001296 return 1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001297 }
Tejun Heod565ed62013-01-24 11:01:33 -08001298 spin_unlock(&pool->lock);
Tejun Heobbb68df2012-08-03 10:30:46 -07001299fail:
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001300 rcu_read_unlock();
Tejun Heobbb68df2012-08-03 10:30:46 -07001301 local_irq_restore(*flags);
1302 if (work_is_canceling(work))
1303 return -ENOENT;
1304 cpu_relax();
Tejun Heo36e227d2012-08-03 10:30:46 -07001305 return -EAGAIN;
Tejun Heobf4ede02012-08-03 10:30:46 -07001306}
1307
1308/**
Tejun Heo706026c2013-01-24 11:01:34 -08001309 * insert_work - insert a work into a pool
Tejun Heo112202d2013-02-13 19:29:12 -08001310 * @pwq: pwq @work belongs to
Tejun Heo4690c4a2010-06-29 10:07:10 +02001311 * @work: work to insert
1312 * @head: insertion point
1313 * @extra_flags: extra WORK_STRUCT_* flags to set
1314 *
Tejun Heo112202d2013-02-13 19:29:12 -08001315 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
Tejun Heo706026c2013-01-24 11:01:34 -08001316 * work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +02001317 *
1318 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001319 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02001320 */
Tejun Heo112202d2013-02-13 19:29:12 -08001321static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1322 struct list_head *head, unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001323{
Tejun Heo112202d2013-02-13 19:29:12 -08001324 struct worker_pool *pool = pwq->pool;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01001325
Tejun Heo4690c4a2010-06-29 10:07:10 +02001326 /* we own @work, set data and link */
Tejun Heo112202d2013-02-13 19:29:12 -08001327 set_work_pwq(work, pwq, extra_flags);
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -07001328 list_add_tail(&work->entry, head);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001329 get_pwq(pwq);
Tejun Heoe22bee72010-06-29 10:07:14 +02001330
1331 /*
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001332 * Ensure either wq_worker_sleeping() sees the above
1333 * list_add_tail() or we see zero nr_running to avoid workers lying
1334 * around lazily while there are works to be processed.
Tejun Heoe22bee72010-06-29 10:07:14 +02001335 */
1336 smp_mb();
1337
Tejun Heo63d95a92012-07-12 14:46:37 -07001338 if (__need_more_worker(pool))
1339 wake_up_worker(pool);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001340}
1341
Tejun Heoc8efcc22010-12-20 19:32:04 +01001342/*
1343 * Test whether @work is being queued from another work executing on the
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001344 * same workqueue.
Tejun Heoc8efcc22010-12-20 19:32:04 +01001345 */
1346static bool is_chained_work(struct workqueue_struct *wq)
1347{
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001348 struct worker *worker;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001349
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001350 worker = current_wq_worker();
1351 /*
Bart Van Asschebf393fd2019-03-01 13:57:25 -08001352 * Return %true iff I'm a worker executing a work item on @wq. If
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001353 * I'm @worker, it's safe to dereference it without locking.
1354 */
Tejun Heo112202d2013-02-13 19:29:12 -08001355 return worker && worker->current_pwq->wq == wq;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001356}
1357
Mike Galbraithef5571802016-02-09 17:59:38 -05001358/*
1359 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1360 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1361 * avoid perturbing sensitive tasks.
1362 */
1363static int wq_select_unbound_cpu(int cpu)
1364{
Tejun Heof303fccb2016-02-09 17:59:38 -05001365 static bool printed_dbg_warning;
Mike Galbraithef5571802016-02-09 17:59:38 -05001366 int new_cpu;
1367
Tejun Heof303fccb2016-02-09 17:59:38 -05001368 if (likely(!wq_debug_force_rr_cpu)) {
1369 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1370 return cpu;
1371 } else if (!printed_dbg_warning) {
1372 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1373 printed_dbg_warning = true;
1374 }
1375
Mike Galbraithef5571802016-02-09 17:59:38 -05001376 if (cpumask_empty(wq_unbound_cpumask))
1377 return cpu;
1378
1379 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1380 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1381 if (unlikely(new_cpu >= nr_cpu_ids)) {
1382 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1383 if (unlikely(new_cpu >= nr_cpu_ids))
1384 return cpu;
1385 }
1386 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1387
1388 return new_cpu;
1389}
1390
Tejun Heod84ff052013-03-12 11:29:59 -07001391static void __queue_work(int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 struct work_struct *work)
1393{
Tejun Heo112202d2013-02-13 19:29:12 -08001394 struct pool_workqueue *pwq;
Tejun Heoc9178082013-03-12 11:30:04 -07001395 struct worker_pool *last_pool;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001396 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001397 unsigned int work_flags;
Joonsoo Kimb75cac92012-08-15 23:25:37 +09001398 unsigned int req_cpu = cpu;
Tejun Heo8930cab2012-08-03 10:30:45 -07001399
1400 /*
1401 * While a work item is PENDING && off queue, a task trying to
1402 * steal the PENDING will busy-loop waiting for it to either get
1403 * queued or lose PENDING. Grabbing PENDING and queueing should
1404 * happen with IRQ disabled.
1405 */
Frederic Weisbecker8e8eb732017-11-06 16:01:19 +01001406 lockdep_assert_irqs_disabled();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09001408 debug_work_activate(work);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001409
Li Bin9ef28a72013-09-09 13:13:58 +08001410 /* if draining, only works from the same workqueue are allowed */
Tejun Heo618b01e2013-03-12 11:30:04 -07001411 if (unlikely(wq->flags & __WQ_DRAINING) &&
Tejun Heoc8efcc22010-12-20 19:32:04 +01001412 WARN_ON_ONCE(!is_chained_work(wq)))
Tejun Heoe41e7042010-08-24 14:22:47 +02001413 return;
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001414 rcu_read_lock();
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001415retry:
Tejun Heoc9178082013-03-12 11:30:04 -07001416 /* pwq which will be used unless @work is executing elsewhere */
Hillf Dantonaa202f12020-01-24 20:14:45 -05001417 if (wq->flags & WQ_UNBOUND) {
1418 if (req_cpu == WORK_CPU_UNBOUND)
1419 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001420 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Hillf Dantonaa202f12020-01-24 20:14:45 -05001421 } else {
1422 if (req_cpu == WORK_CPU_UNBOUND)
1423 cpu = raw_smp_processor_id();
1424 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1425 }
Tejun Heodbf25762012-08-20 14:51:23 -07001426
Tejun Heoc9178082013-03-12 11:30:04 -07001427 /*
1428 * If @work was previously on a different pool, it might still be
1429 * running there, in which case the work needs to be queued on that
1430 * pool to guarantee non-reentrancy.
1431 */
1432 last_pool = get_work_pool(work);
1433 if (last_pool && last_pool != pwq->pool) {
1434 struct worker *worker;
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001435
Tejun Heoc9178082013-03-12 11:30:04 -07001436 spin_lock(&last_pool->lock);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001437
Tejun Heoc9178082013-03-12 11:30:04 -07001438 worker = find_worker_executing_work(last_pool, work);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001439
Tejun Heoc9178082013-03-12 11:30:04 -07001440 if (worker && worker->current_pwq->wq == wq) {
1441 pwq = worker->current_pwq;
Tejun Heo8930cab2012-08-03 10:30:45 -07001442 } else {
Tejun Heoc9178082013-03-12 11:30:04 -07001443 /* meh... not running there, queue here */
1444 spin_unlock(&last_pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08001445 spin_lock(&pwq->pool->lock);
Tejun Heo8930cab2012-08-03 10:30:45 -07001446 }
Tejun Heof3421792010-07-02 10:03:51 +02001447 } else {
Tejun Heo112202d2013-02-13 19:29:12 -08001448 spin_lock(&pwq->pool->lock);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001449 }
1450
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001451 /*
1452 * pwq is determined and locked. For unbound pools, we could have
1453 * raced with pwq release and it could already be dead. If its
1454 * refcnt is zero, repeat pwq selection. Note that pwqs never die
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001455 * without another pwq replacing it in the numa_pwq_tbl or while
1456 * work items are executing on it, so the retrying is guaranteed to
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001457 * make forward-progress.
1458 */
1459 if (unlikely(!pwq->refcnt)) {
1460 if (wq->flags & WQ_UNBOUND) {
1461 spin_unlock(&pwq->pool->lock);
1462 cpu_relax();
1463 goto retry;
1464 }
1465 /* oops */
1466 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1467 wq->name, cpu);
1468 }
1469
Tejun Heo112202d2013-02-13 19:29:12 -08001470 /* pwq determined, queue */
1471 trace_workqueue_queue_work(req_cpu, pwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001472
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001473 if (WARN_ON(!list_empty(&work->entry)))
1474 goto out;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001475
Tejun Heo112202d2013-02-13 19:29:12 -08001476 pwq->nr_in_flight[pwq->work_color]++;
1477 work_flags = work_color_to_flags(pwq->work_color);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001478
Tejun Heo112202d2013-02-13 19:29:12 -08001479 if (likely(pwq->nr_active < pwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001480 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001481 pwq->nr_active++;
1482 worklist = &pwq->pool->worklist;
Tejun Heo82607adc2015-12-08 11:28:04 -05001483 if (list_empty(worklist))
1484 pwq->pool->watchdog_ts = jiffies;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001485 } else {
1486 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo112202d2013-02-13 19:29:12 -08001487 worklist = &pwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001488 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001489
Tejun Heo112202d2013-02-13 19:29:12 -08001490 insert_work(pwq, work, worklist, work_flags);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001491
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001492out:
Tejun Heo112202d2013-02-13 19:29:12 -08001493 spin_unlock(&pwq->pool->lock);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01001494 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495}
1496
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001497/**
Zhang Ruic1a220e2008-07-23 21:28:39 -07001498 * queue_work_on - queue work on specific cpu
1499 * @cpu: CPU number to execute work on
1500 * @wq: workqueue to use
1501 * @work: work to queue
1502 *
Zhang Ruic1a220e2008-07-23 21:28:39 -07001503 * We queue the work to a specific CPU, the caller must ensure it
1504 * can't go away.
Yacine Belkadid185af32013-07-31 14:59:24 -07001505 *
1506 * Return: %false if @work was already on a queue, %true otherwise.
Zhang Ruic1a220e2008-07-23 21:28:39 -07001507 */
Tejun Heod4283e92012-08-03 10:30:44 -07001508bool queue_work_on(int cpu, struct workqueue_struct *wq,
1509 struct work_struct *work)
Zhang Ruic1a220e2008-07-23 21:28:39 -07001510{
Tejun Heod4283e92012-08-03 10:30:44 -07001511 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001512 unsigned long flags;
1513
1514 local_irq_save(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001515
Tejun Heo22df02b2010-06-29 10:07:10 +02001516 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001517 __queue_work(cpu, wq, work);
Tejun Heod4283e92012-08-03 10:30:44 -07001518 ret = true;
Zhang Ruic1a220e2008-07-23 21:28:39 -07001519 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001520
1521 local_irq_restore(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001522 return ret;
1523}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001524EXPORT_SYMBOL(queue_work_on);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001525
Alexander Duyck8204e0c2019-01-22 10:39:26 -08001526/**
1527 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1528 * @node: NUMA node ID that we want to select a CPU from
1529 *
1530 * This function will attempt to find a "random" cpu available on a given
1531 * node. If there are no CPUs available on the given node it will return
1532 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1533 * available CPU if we need to schedule this work.
1534 */
1535static int workqueue_select_cpu_near(int node)
1536{
1537 int cpu;
1538
1539 /* No point in doing this if NUMA isn't enabled for workqueues */
1540 if (!wq_numa_enabled)
1541 return WORK_CPU_UNBOUND;
1542
1543 /* Delay binding to CPU if node is not valid or online */
1544 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1545 return WORK_CPU_UNBOUND;
1546
1547 /* Use local node/cpu if we are already there */
1548 cpu = raw_smp_processor_id();
1549 if (node == cpu_to_node(cpu))
1550 return cpu;
1551
1552 /* Use "random" otherwise know as "first" online CPU of node */
1553 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1554
1555 /* If CPU is valid return that, otherwise just defer */
1556 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1557}
1558
1559/**
1560 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1561 * @node: NUMA node that we are targeting the work for
1562 * @wq: workqueue to use
1563 * @work: work to queue
1564 *
1565 * We queue the work to a "random" CPU within a given NUMA node. The basic
1566 * idea here is to provide a way to somehow associate work with a given
1567 * NUMA node.
1568 *
1569 * This function will only make a best effort attempt at getting this onto
1570 * the right NUMA node. If no node is requested or the requested node is
1571 * offline then we just fall back to standard queue_work behavior.
1572 *
1573 * Currently the "random" CPU ends up being the first available CPU in the
1574 * intersection of cpu_online_mask and the cpumask of the node, unless we
1575 * are running on the node. In that case we just use the current CPU.
1576 *
1577 * Return: %false if @work was already on a queue, %true otherwise.
1578 */
1579bool queue_work_node(int node, struct workqueue_struct *wq,
1580 struct work_struct *work)
1581{
1582 unsigned long flags;
1583 bool ret = false;
1584
1585 /*
1586 * This current implementation is specific to unbound workqueues.
1587 * Specifically we only return the first available CPU for a given
1588 * node instead of cycling through individual CPUs within the node.
1589 *
1590 * If this is used with a per-cpu workqueue then the logic in
1591 * workqueue_select_cpu_near would need to be updated to allow for
1592 * some round robin type logic.
1593 */
1594 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1595
1596 local_irq_save(flags);
1597
1598 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1599 int cpu = workqueue_select_cpu_near(node);
1600
1601 __queue_work(cpu, wq, work);
1602 ret = true;
1603 }
1604
1605 local_irq_restore(flags);
1606 return ret;
1607}
1608EXPORT_SYMBOL_GPL(queue_work_node);
1609
Kees Cook8c20feb2017-10-04 16:27:07 -07001610void delayed_work_timer_fn(struct timer_list *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611{
Kees Cook8c20feb2017-10-04 16:27:07 -07001612 struct delayed_work *dwork = from_timer(dwork, t, timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001614 /* should have been called from irqsafe timer with irq already off */
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001615 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616}
Konstantin Khlebnikov1438ade52013-01-24 16:36:31 +04001617EXPORT_SYMBOL(delayed_work_timer_fn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001619static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1620 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621{
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001622 struct timer_list *timer = &dwork->timer;
1623 struct work_struct *work = &dwork->work;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624
Tejun Heo637fdba2017-03-06 15:33:42 -05001625 WARN_ON_ONCE(!wq);
Kees Cook841b86f2017-10-23 09:40:42 +02001626 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
Tejun Heofc4b5142012-12-04 07:40:39 -08001627 WARN_ON_ONCE(timer_pending(timer));
1628 WARN_ON_ONCE(!list_empty(&work->entry));
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001629
Tejun Heo8852aac2012-12-01 16:23:42 -08001630 /*
1631 * If @delay is 0, queue @dwork->work immediately. This is for
1632 * both optimization and correctness. The earliest @timer can
1633 * expire is on the closest next tick and delayed_work users depend
1634 * on that there's no such delay when @delay is 0.
1635 */
1636 if (!delay) {
1637 __queue_work(cpu, wq, &dwork->work);
1638 return;
1639 }
1640
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001641 dwork->wq = wq;
Tejun Heo12650572012-08-08 09:38:42 -07001642 dwork->cpu = cpu;
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001643 timer->expires = jiffies + delay;
1644
Tejun Heo041bd122016-02-09 16:11:26 -05001645 if (unlikely(cpu != WORK_CPU_UNBOUND))
1646 add_timer_on(timer, cpu);
1647 else
1648 add_timer(timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649}
1650
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001651/**
1652 * queue_delayed_work_on - queue work on specific CPU after delay
1653 * @cpu: CPU number to execute work on
1654 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001655 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001656 * @delay: number of jiffies to wait before queueing
1657 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001658 * Return: %false if @work was already on a queue, %true otherwise. If
Tejun Heo715f1302012-08-03 10:30:46 -07001659 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1660 * execution.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001661 */
Tejun Heod4283e92012-08-03 10:30:44 -07001662bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1663 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001664{
David Howells52bad642006-11-22 14:54:01 +00001665 struct work_struct *work = &dwork->work;
Tejun Heod4283e92012-08-03 10:30:44 -07001666 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001667 unsigned long flags;
1668
1669 /* read the comment in __queue_work() */
1670 local_irq_save(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001671
Tejun Heo22df02b2010-06-29 10:07:10 +02001672 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001673 __queue_delayed_work(cpu, wq, dwork, delay);
Tejun Heod4283e92012-08-03 10:30:44 -07001674 ret = true;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001675 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001676
1677 local_irq_restore(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001678 return ret;
1679}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001680EXPORT_SYMBOL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
Tejun Heoc8e55f32010-06-29 10:07:12 +02001682/**
Tejun Heo8376fe22012-08-03 10:30:47 -07001683 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1684 * @cpu: CPU number to execute work on
1685 * @wq: workqueue to use
1686 * @dwork: work to queue
1687 * @delay: number of jiffies to wait before queueing
1688 *
1689 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1690 * modify @dwork's timer so that it expires after @delay. If @delay is
1691 * zero, @work is guaranteed to be scheduled immediately regardless of its
1692 * current state.
1693 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001694 * Return: %false if @dwork was idle and queued, %true if @dwork was
Tejun Heo8376fe22012-08-03 10:30:47 -07001695 * pending and its timer was modified.
1696 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001697 * This function is safe to call from any context including IRQ handler.
Tejun Heo8376fe22012-08-03 10:30:47 -07001698 * See try_to_grab_pending() for details.
1699 */
1700bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1701 struct delayed_work *dwork, unsigned long delay)
1702{
1703 unsigned long flags;
1704 int ret;
1705
1706 do {
1707 ret = try_to_grab_pending(&dwork->work, true, &flags);
1708 } while (unlikely(ret == -EAGAIN));
1709
1710 if (likely(ret >= 0)) {
1711 __queue_delayed_work(cpu, wq, dwork, delay);
1712 local_irq_restore(flags);
1713 }
1714
1715 /* -ENOENT from try_to_grab_pending() becomes %true */
1716 return ret;
1717}
1718EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1719
Tejun Heo05f0fe62018-03-14 12:45:13 -07001720static void rcu_work_rcufn(struct rcu_head *rcu)
1721{
1722 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1723
1724 /* read the comment in __queue_work() */
1725 local_irq_disable();
1726 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1727 local_irq_enable();
1728}
1729
1730/**
1731 * queue_rcu_work - queue work after a RCU grace period
1732 * @wq: workqueue to use
1733 * @rwork: work to queue
1734 *
1735 * Return: %false if @rwork was already pending, %true otherwise. Note
1736 * that a full RCU grace period is guaranteed only after a %true return.
Bart Van Asschebf393fd2019-03-01 13:57:25 -08001737 * While @rwork is guaranteed to be executed after a %false return, the
Tejun Heo05f0fe62018-03-14 12:45:13 -07001738 * execution may happen before a full RCU grace period has passed.
1739 */
1740bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1741{
1742 struct work_struct *work = &rwork->work;
1743
1744 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1745 rwork->wq = wq;
1746 call_rcu(&rwork->rcu, rcu_work_rcufn);
1747 return true;
1748 }
1749
1750 return false;
1751}
1752EXPORT_SYMBOL(queue_rcu_work);
1753
Tejun Heo8376fe22012-08-03 10:30:47 -07001754/**
Tejun Heoc8e55f32010-06-29 10:07:12 +02001755 * worker_enter_idle - enter idle state
1756 * @worker: worker which is entering idle state
1757 *
1758 * @worker is entering idle state. Update stats and idle timer if
1759 * necessary.
1760 *
1761 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001762 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001763 */
1764static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001766 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767
Tejun Heo6183c002013-03-12 11:29:57 -07001768 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1769 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1770 (worker->hentry.next || worker->hentry.pprev)))
1771 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
Lai Jiangshan051e1852014-07-22 13:03:02 +08001773 /* can't use worker_set_flags(), also called from create_worker() */
Tejun Heocb444762010-07-02 10:03:50 +02001774 worker->flags |= WORKER_IDLE;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001775 pool->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001776 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001777
Tejun Heoc8e55f32010-06-29 10:07:12 +02001778 /* idle_list is LIFO */
Tejun Heobd7bdd42012-07-12 14:46:37 -07001779 list_add(&worker->entry, &pool->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001780
Tejun Heo628c78e2012-07-17 12:39:27 -07001781 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1782 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
Tejun Heocb444762010-07-02 10:03:50 +02001783
Tejun Heo544ecf32012-05-14 15:04:50 -07001784 /*
Lai Jiangshane8b3f8d2017-12-01 22:20:36 +08001785 * Sanity check nr_running. Because unbind_workers() releases
Tejun Heod565ed62013-01-24 11:01:33 -08001786 * pool->lock between setting %WORKER_UNBOUND and zapping
Tejun Heo628c78e2012-07-17 12:39:27 -07001787 * nr_running, the warning may trigger spuriously. Check iff
1788 * unbind is not in progress.
Tejun Heo544ecf32012-05-14 15:04:50 -07001789 */
Tejun Heo24647572013-01-24 11:01:33 -08001790 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -07001791 pool->nr_workers == pool->nr_idle &&
Tejun Heoe19e3972013-01-24 11:39:44 -08001792 atomic_read(&pool->nr_running));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793}
1794
Tejun Heoc8e55f32010-06-29 10:07:12 +02001795/**
1796 * worker_leave_idle - leave idle state
1797 * @worker: worker which is leaving idle state
1798 *
1799 * @worker is leaving idle state. Update stats.
1800 *
1801 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001802 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001803 */
1804static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001806 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807
Tejun Heo6183c002013-03-12 11:29:57 -07001808 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1809 return;
Tejun Heod302f012010-06-29 10:07:13 +02001810 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001811 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001812 list_del_init(&worker->entry);
1813}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001815static struct worker *alloc_worker(int node)
Tejun Heoc34056a2010-06-29 10:07:11 +02001816{
1817 struct worker *worker;
1818
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001819 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001820 if (worker) {
1821 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001822 INIT_LIST_HEAD(&worker->scheduled);
Lai Jiangshanda028462014-05-20 17:46:31 +08001823 INIT_LIST_HEAD(&worker->node);
Tejun Heoe22bee72010-06-29 10:07:14 +02001824 /* on creation a worker is in !idle && prep state */
1825 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001826 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001827 return worker;
1828}
1829
1830/**
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001831 * worker_attach_to_pool() - attach a worker to a pool
1832 * @worker: worker to be attached
1833 * @pool: the target pool
1834 *
1835 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1836 * cpu-binding of @worker are kept coordinated with the pool across
1837 * cpu-[un]hotplugs.
1838 */
1839static void worker_attach_to_pool(struct worker *worker,
1840 struct worker_pool *pool)
1841{
Tejun Heo1258fae2018-05-18 08:47:13 -07001842 mutex_lock(&wq_pool_attach_mutex);
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001843
1844 /*
1845 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1846 * online CPUs. It'll be re-applied when any of the CPUs come up.
1847 */
1848 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1849
1850 /*
Tejun Heo1258fae2018-05-18 08:47:13 -07001851 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1852 * stable across this function. See the comments above the flag
1853 * definition for details.
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001854 */
1855 if (pool->flags & POOL_DISASSOCIATED)
1856 worker->flags |= WORKER_UNBOUND;
1857
1858 list_add_tail(&worker->node, &pool->workers);
Tejun Heoa2d812a2018-05-18 08:47:13 -07001859 worker->pool = pool;
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001860
Tejun Heo1258fae2018-05-18 08:47:13 -07001861 mutex_unlock(&wq_pool_attach_mutex);
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001862}
1863
1864/**
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001865 * worker_detach_from_pool() - detach a worker from its pool
1866 * @worker: worker which is attached to its pool
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001867 *
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001868 * Undo the attaching which had been done in worker_attach_to_pool(). The
1869 * caller worker shouldn't access to the pool after detached except it has
1870 * other reference to the pool.
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001871 */
Tejun Heoa2d812a2018-05-18 08:47:13 -07001872static void worker_detach_from_pool(struct worker *worker)
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001873{
Tejun Heoa2d812a2018-05-18 08:47:13 -07001874 struct worker_pool *pool = worker->pool;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001875 struct completion *detach_completion = NULL;
1876
Tejun Heo1258fae2018-05-18 08:47:13 -07001877 mutex_lock(&wq_pool_attach_mutex);
Tejun Heoa2d812a2018-05-18 08:47:13 -07001878
Lai Jiangshanda028462014-05-20 17:46:31 +08001879 list_del(&worker->node);
Tejun Heoa2d812a2018-05-18 08:47:13 -07001880 worker->pool = NULL;
1881
Lai Jiangshanda028462014-05-20 17:46:31 +08001882 if (list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001883 detach_completion = pool->detach_completion;
Tejun Heo1258fae2018-05-18 08:47:13 -07001884 mutex_unlock(&wq_pool_attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001885
Lai Jiangshanb62c0752014-06-03 15:32:52 +08001886 /* clear leftover flags without pool->lock after it is detached */
1887 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1888
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001889 if (detach_completion)
1890 complete(detach_completion);
1891}
1892
1893/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001894 * create_worker - create a new workqueue worker
Tejun Heo63d95a92012-07-12 14:46:37 -07001895 * @pool: pool the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001896 *
Lai Jiangshan051e1852014-07-22 13:03:02 +08001897 * Create and start a new worker which is attached to @pool.
Tejun Heoc34056a2010-06-29 10:07:11 +02001898 *
1899 * CONTEXT:
1900 * Might sleep. Does GFP_KERNEL allocations.
1901 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001902 * Return:
Tejun Heoc34056a2010-06-29 10:07:11 +02001903 * Pointer to the newly created worker.
1904 */
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001905static struct worker *create_worker(struct worker_pool *pool)
Tejun Heoc34056a2010-06-29 10:07:11 +02001906{
Tejun Heoc34056a2010-06-29 10:07:11 +02001907 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001908 int id = -1;
Tejun Heoe3c916a2013-04-01 11:23:32 -07001909 char id_buf[16];
Tejun Heoc34056a2010-06-29 10:07:11 +02001910
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001911 /* ID is needed to determine kthread name */
1912 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
Tejun Heo822d8402013-03-19 13:45:21 -07001913 if (id < 0)
1914 goto fail;
Tejun Heoc34056a2010-06-29 10:07:11 +02001915
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001916 worker = alloc_worker(pool->node);
Tejun Heoc34056a2010-06-29 10:07:11 +02001917 if (!worker)
1918 goto fail;
1919
Tejun Heoc34056a2010-06-29 10:07:11 +02001920 worker->id = id;
1921
Tejun Heo29c91e92013-03-12 11:30:03 -07001922 if (pool->cpu >= 0)
Tejun Heoe3c916a2013-04-01 11:23:32 -07001923 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1924 pool->attrs->nice < 0 ? "H" : "");
Tejun Heof3421792010-07-02 10:03:51 +02001925 else
Tejun Heoe3c916a2013-04-01 11:23:32 -07001926 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1927
Tejun Heof3f90ad2013-04-01 11:23:34 -07001928 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
Tejun Heoe3c916a2013-04-01 11:23:32 -07001929 "kworker/%s", id_buf);
Tejun Heoc34056a2010-06-29 10:07:11 +02001930 if (IS_ERR(worker->task))
1931 goto fail;
1932
Oleg Nesterov91151222013-11-14 12:56:18 +01001933 set_user_nice(worker->task, pool->attrs->nice);
Peter Zijlstra25834c72015-05-15 17:43:34 +02001934 kthread_bind_mask(worker->task, pool->attrs->cpumask);
Oleg Nesterov91151222013-11-14 12:56:18 +01001935
Lai Jiangshanda028462014-05-20 17:46:31 +08001936 /* successful, attach the worker to the pool */
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001937 worker_attach_to_pool(worker, pool);
Tejun Heo822d8402013-03-19 13:45:21 -07001938
Lai Jiangshan051e1852014-07-22 13:03:02 +08001939 /* start the newly created worker */
1940 spin_lock_irq(&pool->lock);
1941 worker->pool->nr_workers++;
1942 worker_enter_idle(worker);
1943 wake_up_process(worker->task);
1944 spin_unlock_irq(&pool->lock);
1945
Tejun Heoc34056a2010-06-29 10:07:11 +02001946 return worker;
Tejun Heo822d8402013-03-19 13:45:21 -07001947
Tejun Heoc34056a2010-06-29 10:07:11 +02001948fail:
Lai Jiangshan9625ab12014-05-20 17:46:27 +08001949 if (id >= 0)
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001950 ida_simple_remove(&pool->worker_ida, id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001951 kfree(worker);
1952 return NULL;
1953}
1954
1955/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001956 * destroy_worker - destroy a workqueue worker
1957 * @worker: worker to be destroyed
1958 *
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001959 * Destroy @worker and adjust @pool stats accordingly. The worker should
1960 * be idle.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001961 *
1962 * CONTEXT:
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001963 * spin_lock_irq(pool->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001964 */
1965static void destroy_worker(struct worker *worker)
1966{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001967 struct worker_pool *pool = worker->pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001968
Tejun Heocd549682013-03-13 19:47:39 -07001969 lockdep_assert_held(&pool->lock);
1970
Tejun Heoc34056a2010-06-29 10:07:11 +02001971 /* sanity check frenzy */
Tejun Heo6183c002013-03-12 11:29:57 -07001972 if (WARN_ON(worker->current_work) ||
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001973 WARN_ON(!list_empty(&worker->scheduled)) ||
1974 WARN_ON(!(worker->flags & WORKER_IDLE)))
Tejun Heo6183c002013-03-12 11:29:57 -07001975 return;
Tejun Heoc34056a2010-06-29 10:07:11 +02001976
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001977 pool->nr_workers--;
1978 pool->nr_idle--;
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001979
Tejun Heoc8e55f32010-06-29 10:07:12 +02001980 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001981 worker->flags |= WORKER_DIE;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001982 wake_up_process(worker->task);
Tejun Heoc34056a2010-06-29 10:07:11 +02001983}
1984
Kees Cook32a6c722017-10-16 15:58:25 -07001985static void idle_worker_timeout(struct timer_list *t)
Tejun Heoe22bee72010-06-29 10:07:14 +02001986{
Kees Cook32a6c722017-10-16 15:58:25 -07001987 struct worker_pool *pool = from_timer(pool, t, idle_timer);
Tejun Heoe22bee72010-06-29 10:07:14 +02001988
Tejun Heod565ed62013-01-24 11:01:33 -08001989 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001990
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001991 while (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001992 struct worker *worker;
1993 unsigned long expires;
1994
1995 /* idle_list is kept in LIFO order, check the last one */
Tejun Heo63d95a92012-07-12 14:46:37 -07001996 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02001997 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1998
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001999 if (time_before(jiffies, expires)) {
Tejun Heo63d95a92012-07-12 14:46:37 -07002000 mod_timer(&pool->idle_timer, expires);
Lai Jiangshan3347fc92014-05-20 17:46:30 +08002001 break;
Tejun Heoe22bee72010-06-29 10:07:14 +02002002 }
Lai Jiangshan3347fc92014-05-20 17:46:30 +08002003
2004 destroy_worker(worker);
Tejun Heoe22bee72010-06-29 10:07:14 +02002005 }
2006
Tejun Heod565ed62013-01-24 11:01:33 -08002007 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002008}
2009
Tejun Heo493a1722013-03-12 11:29:59 -07002010static void send_mayday(struct work_struct *work)
Tejun Heoe22bee72010-06-29 10:07:14 +02002011{
Tejun Heo112202d2013-02-13 19:29:12 -08002012 struct pool_workqueue *pwq = get_work_pwq(work);
2013 struct workqueue_struct *wq = pwq->wq;
Tejun Heo493a1722013-03-12 11:29:59 -07002014
Tejun Heo2e109a22013-03-13 19:47:40 -07002015 lockdep_assert_held(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002016
Tejun Heo493008a2013-03-12 11:30:03 -07002017 if (!wq->rescuer)
Tejun Heo493a1722013-03-12 11:29:59 -07002018 return;
Tejun Heoe22bee72010-06-29 10:07:14 +02002019
2020 /* mayday mayday mayday */
Tejun Heo493a1722013-03-12 11:29:59 -07002021 if (list_empty(&pwq->mayday_node)) {
Lai Jiangshan77668c82014-04-18 11:04:16 -04002022 /*
2023 * If @pwq is for an unbound wq, its base ref may be put at
2024 * any time due to an attribute change. Pin @pwq until the
2025 * rescuer is done with it.
2026 */
2027 get_pwq(pwq);
Tejun Heo493a1722013-03-12 11:29:59 -07002028 list_add_tail(&pwq->mayday_node, &wq->maydays);
Tejun Heoe22bee72010-06-29 10:07:14 +02002029 wake_up_process(wq->rescuer->task);
Tejun Heo493a1722013-03-12 11:29:59 -07002030 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002031}
2032
Kees Cook32a6c722017-10-16 15:58:25 -07002033static void pool_mayday_timeout(struct timer_list *t)
Tejun Heoe22bee72010-06-29 10:07:14 +02002034{
Kees Cook32a6c722017-10-16 15:58:25 -07002035 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
Tejun Heoe22bee72010-06-29 10:07:14 +02002036 struct work_struct *work;
2037
Tejun Heob2d82902014-12-08 12:39:16 -05002038 spin_lock_irq(&pool->lock);
2039 spin_lock(&wq_mayday_lock); /* for wq->maydays */
Tejun Heoe22bee72010-06-29 10:07:14 +02002040
Tejun Heo63d95a92012-07-12 14:46:37 -07002041 if (need_to_create_worker(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02002042 /*
2043 * We've been trying to create a new worker but
2044 * haven't been successful. We might be hitting an
2045 * allocation deadlock. Send distress signals to
2046 * rescuers.
2047 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002048 list_for_each_entry(work, &pool->worklist, entry)
Tejun Heoe22bee72010-06-29 10:07:14 +02002049 send_mayday(work);
2050 }
2051
Tejun Heob2d82902014-12-08 12:39:16 -05002052 spin_unlock(&wq_mayday_lock);
2053 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002054
Tejun Heo63d95a92012-07-12 14:46:37 -07002055 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
Tejun Heoe22bee72010-06-29 10:07:14 +02002056}
2057
2058/**
2059 * maybe_create_worker - create a new worker if necessary
Tejun Heo63d95a92012-07-12 14:46:37 -07002060 * @pool: pool to create a new worker for
Tejun Heoe22bee72010-06-29 10:07:14 +02002061 *
Tejun Heo63d95a92012-07-12 14:46:37 -07002062 * Create a new worker for @pool if necessary. @pool is guaranteed to
Tejun Heoe22bee72010-06-29 10:07:14 +02002063 * have at least one idle worker on return from this function. If
2064 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
Tejun Heo63d95a92012-07-12 14:46:37 -07002065 * sent to all rescuers with works scheduled on @pool to resolve
Tejun Heoe22bee72010-06-29 10:07:14 +02002066 * possible allocation deadlock.
2067 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002068 * On return, need_to_create_worker() is guaranteed to be %false and
2069 * may_start_working() %true.
Tejun Heoe22bee72010-06-29 10:07:14 +02002070 *
2071 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08002072 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02002073 * multiple times. Does GFP_KERNEL allocations. Called only from
2074 * manager.
Tejun Heoe22bee72010-06-29 10:07:14 +02002075 */
Tejun Heo29187a92015-01-16 14:21:16 -05002076static void maybe_create_worker(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08002077__releases(&pool->lock)
2078__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02002079{
Tejun Heoe22bee72010-06-29 10:07:14 +02002080restart:
Tejun Heod565ed62013-01-24 11:01:33 -08002081 spin_unlock_irq(&pool->lock);
Tejun Heo9f9c236442010-07-14 11:31:20 +02002082
Tejun Heoe22bee72010-06-29 10:07:14 +02002083 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
Tejun Heo63d95a92012-07-12 14:46:37 -07002084 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
Tejun Heoe22bee72010-06-29 10:07:14 +02002085
2086 while (true) {
Lai Jiangshan051e1852014-07-22 13:03:02 +08002087 if (create_worker(pool) || !need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002088 break;
2089
Lai Jiangshane212f362014-06-03 15:32:17 +08002090 schedule_timeout_interruptible(CREATE_COOLDOWN);
Tejun Heo9f9c236442010-07-14 11:31:20 +02002091
Tejun Heo63d95a92012-07-12 14:46:37 -07002092 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002093 break;
2094 }
2095
Tejun Heo63d95a92012-07-12 14:46:37 -07002096 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08002097 spin_lock_irq(&pool->lock);
Lai Jiangshan051e1852014-07-22 13:03:02 +08002098 /*
2099 * This is necessary even after a new worker was just successfully
2100 * created as @pool->lock was dropped and the new worker might have
2101 * already become busy.
2102 */
Tejun Heo63d95a92012-07-12 14:46:37 -07002103 if (need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002104 goto restart;
Tejun Heoe22bee72010-06-29 10:07:14 +02002105}
2106
2107/**
Tejun Heoe22bee72010-06-29 10:07:14 +02002108 * manage_workers - manage worker pool
2109 * @worker: self
2110 *
Tejun Heo706026c2013-01-24 11:01:34 -08002111 * Assume the manager role and manage the worker pool @worker belongs
Tejun Heoe22bee72010-06-29 10:07:14 +02002112 * to. At any given time, there can be only zero or one manager per
Tejun Heo706026c2013-01-24 11:01:34 -08002113 * pool. The exclusion is handled automatically by this function.
Tejun Heoe22bee72010-06-29 10:07:14 +02002114 *
2115 * The caller can safely start processing works on false return. On
2116 * true return, it's guaranteed that need_to_create_worker() is false
2117 * and may_start_working() is true.
2118 *
2119 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002120 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02002121 * multiple times. Does GFP_KERNEL allocations.
2122 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002123 * Return:
Tejun Heo29187a92015-01-16 14:21:16 -05002124 * %false if the pool doesn't need management and the caller can safely
2125 * start processing works, %true if management function was performed and
2126 * the conditions that the caller verified before calling the function may
2127 * no longer be true.
Tejun Heoe22bee72010-06-29 10:07:14 +02002128 */
2129static bool manage_workers(struct worker *worker)
2130{
Tejun Heo63d95a92012-07-12 14:46:37 -07002131 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002132
Tejun Heo692b4822017-10-09 08:04:13 -07002133 if (pool->flags & POOL_MANAGER_ACTIVE)
Tejun Heo29187a92015-01-16 14:21:16 -05002134 return false;
Tejun Heo692b4822017-10-09 08:04:13 -07002135
2136 pool->flags |= POOL_MANAGER_ACTIVE;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002137 pool->manager = worker;
Tejun Heoe22bee72010-06-29 10:07:14 +02002138
Tejun Heo29187a92015-01-16 14:21:16 -05002139 maybe_create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02002140
Tejun Heo2607d7a2015-03-09 09:22:28 -04002141 pool->manager = NULL;
Tejun Heo692b4822017-10-09 08:04:13 -07002142 pool->flags &= ~POOL_MANAGER_ACTIVE;
2143 wake_up(&wq_manager_wait);
Tejun Heo29187a92015-01-16 14:21:16 -05002144 return true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002145}
2146
Tejun Heoa62428c2010-06-29 10:07:10 +02002147/**
2148 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02002149 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02002150 * @work: work to process
2151 *
2152 * Process @work. This function contains all the logics necessary to
2153 * process a single work including synchronization against and
2154 * interaction with other workers on the same cpu, queueing and
2155 * flushing. As long as context requirement is met, any worker can
2156 * call this function to process a work.
2157 *
2158 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002159 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02002160 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002161static void process_one_work(struct worker *worker, struct work_struct *work)
Tejun Heod565ed62013-01-24 11:01:33 -08002162__releases(&pool->lock)
2163__acquires(&pool->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02002164{
Tejun Heo112202d2013-02-13 19:29:12 -08002165 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobd7bdd42012-07-12 14:46:37 -07002166 struct worker_pool *pool = worker->pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002167 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heo73f53c42010-06-29 10:07:11 +02002168 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02002169 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02002170#ifdef CONFIG_LOCKDEP
2171 /*
2172 * It is permissible to free the struct work_struct from
2173 * inside the function that is called from it, this we need to
2174 * take into account for lockdep too. To avoid bogus "held
2175 * lock freed" warnings as well as problems when looking into
2176 * work->lockdep_map, make a copy and use that here.
2177 */
Peter Zijlstra4d82a1d2012-05-15 08:06:19 -07002178 struct lockdep_map lockdep_map;
2179
2180 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002181#endif
Lai Jiangshan807407c2014-06-03 15:33:28 +08002182 /* ensure we're on the correct CPU */
Lai Jiangshan85327af2014-06-03 15:33:28 +08002183 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heoec22ca52013-01-24 11:01:33 -08002184 raw_smp_processor_id() != pool->cpu);
Tejun Heo25511a42012-07-17 12:39:27 -07002185
Tejun Heo7e116292010-06-29 10:07:13 +02002186 /*
2187 * A single work shouldn't be executed concurrently by
2188 * multiple workers on a single cpu. Check whether anyone is
2189 * already processing the work. If so, defer the work to the
2190 * currently executing one.
2191 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002192 collision = find_worker_executing_work(pool, work);
Tejun Heo7e116292010-06-29 10:07:13 +02002193 if (unlikely(collision)) {
2194 move_linked_works(work, &collision->scheduled, NULL);
2195 return;
2196 }
2197
Tejun Heo8930cab2012-08-03 10:30:45 -07002198 /* claim and dequeue */
Tejun Heoa62428c2010-06-29 10:07:10 +02002199 debug_work_deactivate(work);
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002200 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
Tejun Heoc34056a2010-06-29 10:07:11 +02002201 worker->current_work = work;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002202 worker->current_func = work->func;
Tejun Heo112202d2013-02-13 19:29:12 -08002203 worker->current_pwq = pwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002204 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002205
Tejun Heo8bf89592018-05-18 08:47:13 -07002206 /*
2207 * Record wq name for cmdline and debug reporting, may get
2208 * overridden through set_worker_desc().
2209 */
2210 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2211
Tejun Heoa62428c2010-06-29 10:07:10 +02002212 list_del_init(&work->entry);
2213
Tejun Heo649027d2010-06-29 10:07:14 +02002214 /*
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002215 * CPU intensive works don't participate in concurrency management.
2216 * They're the scheduler's responsibility. This takes @worker out
2217 * of concurrency management and the next code block will chain
2218 * execution of the pending work items.
Tejun Heofb0e7be2010-06-29 10:07:15 +02002219 */
2220 if (unlikely(cpu_intensive))
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002221 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
Tejun Heofb0e7be2010-06-29 10:07:15 +02002222
Tejun Heo974271c42012-07-12 14:46:37 -07002223 /*
Lai Jiangshana489a032014-07-22 13:01:59 +08002224 * Wake up another worker if necessary. The condition is always
2225 * false for normal per-cpu workers since nr_running would always
2226 * be >= 1 at this point. This is used to chain execution of the
2227 * pending work items for WORKER_NOT_RUNNING workers such as the
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002228 * UNBOUND and CPU_INTENSIVE ones.
Tejun Heo974271c42012-07-12 14:46:37 -07002229 */
Lai Jiangshana489a032014-07-22 13:01:59 +08002230 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002231 wake_up_worker(pool);
Tejun Heo974271c42012-07-12 14:46:37 -07002232
Tejun Heo8930cab2012-08-03 10:30:45 -07002233 /*
Tejun Heo7c3eed52013-01-24 11:01:33 -08002234 * Record the last pool and clear PENDING which should be the last
Tejun Heod565ed62013-01-24 11:01:33 -08002235 * update to @work. Also, do this inside @pool->lock so that
Tejun Heo23657bb2012-08-13 17:08:19 -07002236 * PENDING and queued state changes happen together while IRQ is
2237 * disabled.
Tejun Heo8930cab2012-08-03 10:30:45 -07002238 */
Tejun Heo7c3eed52013-01-24 11:01:33 -08002239 set_work_pool_and_clear_pending(work, pool->id);
Tejun Heoa62428c2010-06-29 10:07:10 +02002240
Tejun Heod565ed62013-01-24 11:01:33 -08002241 spin_unlock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002242
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002243 lock_map_acquire(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002244 lock_map_acquire(&lockdep_map);
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002245 /*
Peter Zijlstraf52be572017-08-29 10:59:39 +02002246 * Strictly speaking we should mark the invariant state without holding
2247 * any locks, that is, before these two lock_map_acquire()'s.
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002248 *
2249 * However, that would result in:
2250 *
2251 * A(W1)
2252 * WFC(C)
2253 * A(W1)
2254 * C(C)
2255 *
2256 * Which would create W1->C->W1 dependencies, even though there is no
2257 * actual deadlock possible. There are two solutions, using a
2258 * read-recursive acquire on the work(queue) 'locks', but this will then
Peter Zijlstraf52be572017-08-29 10:59:39 +02002259 * hit the lockdep limitation on recursive locks, or simply discard
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002260 * these locks.
2261 *
2262 * AFAICT there is no possible deadlock scenario between the
2263 * flush_work() and complete() primitives (except for single-threaded
2264 * workqueues), so hiding them isn't a problem.
2265 */
Peter Zijlstraf52be572017-08-29 10:59:39 +02002266 lockdep_invariant_state(true);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002267 trace_workqueue_execute_start(work);
Tejun Heoa2c1c572012-12-18 10:35:02 -08002268 worker->current_func(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002269 /*
2270 * While we must be careful to not use "work" after this, the trace
2271 * point will only record its address.
2272 */
Daniel Jordan1c5da0e2020-01-13 17:52:39 -05002273 trace_workqueue_execute_end(work, worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002274 lock_map_release(&lockdep_map);
Tejun Heo112202d2013-02-13 19:29:12 -08002275 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002276
2277 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
Valentin Ilie044c7822012-08-19 00:52:42 +03002278 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
Sakari Ailusd75f7732019-03-25 21:32:28 +02002279 " last function: %ps\n",
Tejun Heoa2c1c572012-12-18 10:35:02 -08002280 current->comm, preempt_count(), task_pid_nr(current),
2281 worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002282 debug_show_held_locks(current);
2283 dump_stack();
2284 }
2285
Tejun Heob22ce272013-08-28 17:33:37 -04002286 /*
Sebastian Andrzej Siewior025f50f2019-10-15 21:18:21 +02002287 * The following prevents a kworker from hogging CPU on !PREEMPTION
Tejun Heob22ce272013-08-28 17:33:37 -04002288 * kernels, where a requeueing work item waiting for something to
2289 * happen could deadlock with stop_machine as such work item could
2290 * indefinitely requeue itself while all other CPUs are trapped in
Joe Lawrence789cbbe2014-10-05 13:24:21 -04002291 * stop_machine. At the same time, report a quiescent RCU state so
2292 * the same condition doesn't freeze RCU.
Tejun Heob22ce272013-08-28 17:33:37 -04002293 */
Paul E. McKenneya7e64252017-10-24 08:25:02 -07002294 cond_resched();
Tejun Heob22ce272013-08-28 17:33:37 -04002295
Tejun Heod565ed62013-01-24 11:01:33 -08002296 spin_lock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002297
Tejun Heofb0e7be2010-06-29 10:07:15 +02002298 /* clear cpu intensive status */
2299 if (unlikely(cpu_intensive))
2300 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2301
Johannes Weiner1b69ac62019-02-01 14:20:42 -08002302 /* tag the worker for identification in schedule() */
2303 worker->last_func = worker->current_func;
2304
Tejun Heoa62428c2010-06-29 10:07:10 +02002305 /* we're done with it, release */
Sasha Levin42f85702012-12-17 10:01:23 -05002306 hash_del(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002307 worker->current_work = NULL;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002308 worker->current_func = NULL;
Tejun Heo112202d2013-02-13 19:29:12 -08002309 worker->current_pwq = NULL;
2310 pwq_dec_nr_in_flight(pwq, work_color);
Tejun Heoa62428c2010-06-29 10:07:10 +02002311}
2312
Tejun Heoaffee4b2010-06-29 10:07:12 +02002313/**
2314 * process_scheduled_works - process scheduled works
2315 * @worker: self
2316 *
2317 * Process all scheduled works. Please note that the scheduled list
2318 * may change while processing a work, so this function repeatedly
2319 * fetches a work from the top and executes it.
2320 *
2321 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002322 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02002323 * multiple times.
2324 */
2325static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002327 while (!list_empty(&worker->scheduled)) {
2328 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002330 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332}
2333
Tejun Heo197f6ac2018-05-21 08:04:35 -07002334static void set_pf_worker(bool val)
2335{
2336 mutex_lock(&wq_pool_attach_mutex);
2337 if (val)
2338 current->flags |= PF_WQ_WORKER;
2339 else
2340 current->flags &= ~PF_WQ_WORKER;
2341 mutex_unlock(&wq_pool_attach_mutex);
2342}
2343
Tejun Heo4690c4a2010-06-29 10:07:10 +02002344/**
2345 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02002346 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02002347 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002348 * The worker thread function. All workers belong to a worker_pool -
2349 * either a per-cpu one or dynamic unbound one. These workers process all
2350 * work items regardless of their specific target workqueue. The only
2351 * exception is work items which belong to workqueues with a rescuer which
2352 * will be explained in rescuer_thread().
Yacine Belkadid185af32013-07-31 14:59:24 -07002353 *
2354 * Return: 0
Tejun Heo4690c4a2010-06-29 10:07:10 +02002355 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002356static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357{
Tejun Heoc34056a2010-06-29 10:07:11 +02002358 struct worker *worker = __worker;
Tejun Heobd7bdd42012-07-12 14:46:37 -07002359 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360
Tejun Heoe22bee72010-06-29 10:07:14 +02002361 /* tell the scheduler that this is a workqueue worker */
Tejun Heo197f6ac2018-05-21 08:04:35 -07002362 set_pf_worker(true);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002363woke_up:
Tejun Heod565ed62013-01-24 11:01:33 -08002364 spin_lock_irq(&pool->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002365
Tejun Heoa9ab7752013-03-19 13:45:21 -07002366 /* am I supposed to die? */
2367 if (unlikely(worker->flags & WORKER_DIE)) {
Tejun Heod565ed62013-01-24 11:01:33 -08002368 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002369 WARN_ON_ONCE(!list_empty(&worker->entry));
Tejun Heo197f6ac2018-05-21 08:04:35 -07002370 set_pf_worker(false);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002371
2372 set_task_comm(worker->task, "kworker/dying");
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08002373 ida_simple_remove(&pool->worker_ida, worker->id);
Tejun Heoa2d812a2018-05-18 08:47:13 -07002374 worker_detach_from_pool(worker);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002375 kfree(worker);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002376 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07002378
Tejun Heoc8e55f32010-06-29 10:07:12 +02002379 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02002380recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02002381 /* no more worker necessary? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002382 if (!need_more_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002383 goto sleep;
2384
2385 /* do we need to manage? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002386 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002387 goto recheck;
2388
Tejun Heoc8e55f32010-06-29 10:07:12 +02002389 /*
2390 * ->scheduled list can only be filled while a worker is
2391 * preparing to process a work or actually processing it.
2392 * Make sure nobody diddled with it while I was sleeping.
2393 */
Tejun Heo6183c002013-03-12 11:29:57 -07002394 WARN_ON_ONCE(!list_empty(&worker->scheduled));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002395
Tejun Heoe22bee72010-06-29 10:07:14 +02002396 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07002397 * Finish PREP stage. We're guaranteed to have at least one idle
2398 * worker or that someone else has already assumed the manager
2399 * role. This is where @worker starts participating in concurrency
2400 * management if applicable and concurrency management is restored
2401 * after being rebound. See rebind_workers() for details.
Tejun Heoe22bee72010-06-29 10:07:14 +02002402 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07002403 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
Tejun Heoe22bee72010-06-29 10:07:14 +02002404
2405 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02002406 struct work_struct *work =
Tejun Heobd7bdd42012-07-12 14:46:37 -07002407 list_first_entry(&pool->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02002408 struct work_struct, entry);
2409
Tejun Heo82607adc2015-12-08 11:28:04 -05002410 pool->watchdog_ts = jiffies;
2411
Tejun Heoc8e55f32010-06-29 10:07:12 +02002412 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2413 /* optimization path, not strictly necessary */
2414 process_one_work(worker, work);
2415 if (unlikely(!list_empty(&worker->scheduled)))
2416 process_scheduled_works(worker);
2417 } else {
2418 move_linked_works(work, &worker->scheduled, NULL);
2419 process_scheduled_works(worker);
2420 }
Tejun Heo63d95a92012-07-12 14:46:37 -07002421 } while (keep_working(pool));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002422
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002423 worker_set_flags(worker, WORKER_PREP);
Tejun Heod313dd82010-07-02 10:03:51 +02002424sleep:
Tejun Heoc8e55f32010-06-29 10:07:12 +02002425 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002426 * pool->lock is held and there's no work to process and no need to
2427 * manage, sleep. Workers are woken up only while holding
2428 * pool->lock or from local cpu, so setting the current state
2429 * before releasing pool->lock is enough to prevent losing any
2430 * event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02002431 */
2432 worker_enter_idle(worker);
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002433 __set_current_state(TASK_IDLE);
Tejun Heod565ed62013-01-24 11:01:33 -08002434 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002435 schedule();
2436 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437}
2438
Tejun Heoe22bee72010-06-29 10:07:14 +02002439/**
2440 * rescuer_thread - the rescuer thread function
Tejun Heo111c2252013-01-17 17:16:24 -08002441 * @__rescuer: self
Tejun Heoe22bee72010-06-29 10:07:14 +02002442 *
2443 * Workqueue rescuer thread function. There's one rescuer for each
Tejun Heo493008a2013-03-12 11:30:03 -07002444 * workqueue which has WQ_MEM_RECLAIM set.
Tejun Heoe22bee72010-06-29 10:07:14 +02002445 *
Tejun Heo706026c2013-01-24 11:01:34 -08002446 * Regular work processing on a pool may block trying to create a new
Tejun Heoe22bee72010-06-29 10:07:14 +02002447 * worker which uses GFP_KERNEL allocation which has slight chance of
2448 * developing into deadlock if some works currently on the same queue
2449 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2450 * the problem rescuer solves.
2451 *
Tejun Heo706026c2013-01-24 11:01:34 -08002452 * When such condition is possible, the pool summons rescuers of all
2453 * workqueues which have works queued on the pool and let them process
Tejun Heoe22bee72010-06-29 10:07:14 +02002454 * those works so that forward progress can be guaranteed.
2455 *
2456 * This should happen rarely.
Yacine Belkadid185af32013-07-31 14:59:24 -07002457 *
2458 * Return: 0
Tejun Heoe22bee72010-06-29 10:07:14 +02002459 */
Tejun Heo111c2252013-01-17 17:16:24 -08002460static int rescuer_thread(void *__rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02002461{
Tejun Heo111c2252013-01-17 17:16:24 -08002462 struct worker *rescuer = __rescuer;
2463 struct workqueue_struct *wq = rescuer->rescue_wq;
Tejun Heoe22bee72010-06-29 10:07:14 +02002464 struct list_head *scheduled = &rescuer->scheduled;
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002465 bool should_stop;
Tejun Heoe22bee72010-06-29 10:07:14 +02002466
2467 set_user_nice(current, RESCUER_NICE_LEVEL);
Tejun Heo111c2252013-01-17 17:16:24 -08002468
2469 /*
2470 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2471 * doesn't participate in concurrency management.
2472 */
Tejun Heo197f6ac2018-05-21 08:04:35 -07002473 set_pf_worker(true);
Tejun Heoe22bee72010-06-29 10:07:14 +02002474repeat:
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002475 set_current_state(TASK_IDLE);
Tejun Heoe22bee72010-06-29 10:07:14 +02002476
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002477 /*
2478 * By the time the rescuer is requested to stop, the workqueue
2479 * shouldn't have any work pending, but @wq->maydays may still have
2480 * pwq(s) queued. This can happen by non-rescuer workers consuming
2481 * all the work items before the rescuer got to them. Go through
2482 * @wq->maydays processing before acting on should_stop so that the
2483 * list is always empty on exit.
2484 */
2485 should_stop = kthread_should_stop();
Tejun Heoe22bee72010-06-29 10:07:14 +02002486
Tejun Heo493a1722013-03-12 11:29:59 -07002487 /* see whether any pwq is asking for help */
Tejun Heo2e109a22013-03-13 19:47:40 -07002488 spin_lock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002489
2490 while (!list_empty(&wq->maydays)) {
2491 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2492 struct pool_workqueue, mayday_node);
Tejun Heo112202d2013-02-13 19:29:12 -08002493 struct worker_pool *pool = pwq->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002494 struct work_struct *work, *n;
Tejun Heo82607adc2015-12-08 11:28:04 -05002495 bool first = true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002496
2497 __set_current_state(TASK_RUNNING);
Tejun Heo493a1722013-03-12 11:29:59 -07002498 list_del_init(&pwq->mayday_node);
2499
Tejun Heo2e109a22013-03-13 19:47:40 -07002500 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002501
Lai Jiangshan51697d392014-05-20 17:46:36 +08002502 worker_attach_to_pool(rescuer, pool);
2503
2504 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002505
2506 /*
2507 * Slurp in all works issued via this workqueue and
2508 * process'em.
2509 */
Tejun Heo0479c8c2014-12-04 10:14:13 -05002510 WARN_ON_ONCE(!list_empty(scheduled));
Tejun Heo82607adc2015-12-08 11:28:04 -05002511 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2512 if (get_work_pwq(work) == pwq) {
2513 if (first)
2514 pool->watchdog_ts = jiffies;
Tejun Heoe22bee72010-06-29 10:07:14 +02002515 move_linked_works(work, scheduled, &n);
Tejun Heo82607adc2015-12-08 11:28:04 -05002516 }
2517 first = false;
2518 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002519
NeilBrown008847f2014-12-08 12:39:16 -05002520 if (!list_empty(scheduled)) {
2521 process_scheduled_works(rescuer);
2522
2523 /*
2524 * The above execution of rescued work items could
2525 * have created more to rescue through
2526 * pwq_activate_first_delayed() or chained
2527 * queueing. Let's put @pwq back on mayday list so
2528 * that such back-to-back work items, which may be
2529 * being used to relieve memory pressure, don't
2530 * incur MAYDAY_INTERVAL delay inbetween.
2531 */
2532 if (need_to_create_worker(pool)) {
2533 spin_lock(&wq_mayday_lock);
Tejun Heoe66b39a2019-09-25 06:59:15 -07002534 /*
2535 * Queue iff we aren't racing destruction
2536 * and somebody else hasn't queued it already.
2537 */
2538 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2539 get_pwq(pwq);
2540 list_add_tail(&pwq->mayday_node, &wq->maydays);
2541 }
NeilBrown008847f2014-12-08 12:39:16 -05002542 spin_unlock(&wq_mayday_lock);
2543 }
2544 }
Tejun Heo75769582011-02-14 14:04:46 +01002545
2546 /*
Lai Jiangshan77668c82014-04-18 11:04:16 -04002547 * Put the reference grabbed by send_mayday(). @pool won't
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002548 * go away while we're still attached to it.
Lai Jiangshan77668c82014-04-18 11:04:16 -04002549 */
2550 put_pwq(pwq);
2551
2552 /*
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002553 * Leave this pool. If need_more_worker() is %true, notify a
Tejun Heo75769582011-02-14 14:04:46 +01002554 * regular worker; otherwise, we end up with 0 concurrency
2555 * and stalling the execution.
2556 */
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002557 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002558 wake_up_worker(pool);
Tejun Heo75769582011-02-14 14:04:46 +01002559
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002560 spin_unlock_irq(&pool->lock);
2561
Tejun Heoa2d812a2018-05-18 08:47:13 -07002562 worker_detach_from_pool(rescuer);
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002563
2564 spin_lock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002565 }
2566
Tejun Heo2e109a22013-03-13 19:47:40 -07002567 spin_unlock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002568
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002569 if (should_stop) {
2570 __set_current_state(TASK_RUNNING);
Tejun Heo197f6ac2018-05-21 08:04:35 -07002571 set_pf_worker(false);
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002572 return 0;
2573 }
2574
Tejun Heo111c2252013-01-17 17:16:24 -08002575 /* rescuers should never participate in concurrency management */
2576 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
Tejun Heoe22bee72010-06-29 10:07:14 +02002577 schedule();
2578 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579}
2580
Tejun Heofca839c2015-12-07 10:58:57 -05002581/**
2582 * check_flush_dependency - check for flush dependency sanity
2583 * @target_wq: workqueue being flushed
2584 * @target_work: work item being flushed (NULL for workqueue flushes)
2585 *
2586 * %current is trying to flush the whole @target_wq or @target_work on it.
2587 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2588 * reclaiming memory or running on a workqueue which doesn't have
2589 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2590 * a deadlock.
2591 */
2592static void check_flush_dependency(struct workqueue_struct *target_wq,
2593 struct work_struct *target_work)
2594{
2595 work_func_t target_func = target_work ? target_work->func : NULL;
2596 struct worker *worker;
2597
2598 if (target_wq->flags & WQ_MEM_RECLAIM)
2599 return;
2600
2601 worker = current_wq_worker();
2602
2603 WARN_ONCE(current->flags & PF_MEMALLOC,
Sakari Ailusd75f7732019-03-25 21:32:28 +02002604 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
Tejun Heofca839c2015-12-07 10:58:57 -05002605 current->pid, current->comm, target_wq->name, target_func);
Tejun Heo23d11a52016-01-29 05:59:46 -05002606 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2607 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
Sakari Ailusd75f7732019-03-25 21:32:28 +02002608 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
Tejun Heofca839c2015-12-07 10:58:57 -05002609 worker->current_pwq->wq->name, worker->current_func,
2610 target_wq->name, target_func);
2611}
2612
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002613struct wq_barrier {
2614 struct work_struct work;
2615 struct completion done;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002616 struct task_struct *task; /* purely informational */
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002617};
2618
2619static void wq_barrier_func(struct work_struct *work)
2620{
2621 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2622 complete(&barr->done);
2623}
2624
Tejun Heo4690c4a2010-06-29 10:07:10 +02002625/**
2626 * insert_wq_barrier - insert a barrier work
Tejun Heo112202d2013-02-13 19:29:12 -08002627 * @pwq: pwq to insert barrier into
Tejun Heo4690c4a2010-06-29 10:07:10 +02002628 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002629 * @target: target work to attach @barr to
2630 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002631 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002632 * @barr is linked to @target such that @barr is completed only after
2633 * @target finishes execution. Please note that the ordering
2634 * guarantee is observed only with respect to @target and on the local
2635 * cpu.
2636 *
2637 * Currently, a queued barrier can't be canceled. This is because
2638 * try_to_grab_pending() can't determine whether the work to be
2639 * grabbed is at the head of the queue and thus can't clear LINKED
2640 * flag of the previous work while there must be a valid next work
2641 * after a work with LINKED flag set.
2642 *
2643 * Note that when @worker is non-NULL, @target may be modified
Tejun Heo112202d2013-02-13 19:29:12 -08002644 * underneath us, so we can't reliably determine pwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002645 *
2646 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002647 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002648 */
Tejun Heo112202d2013-02-13 19:29:12 -08002649static void insert_wq_barrier(struct pool_workqueue *pwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002650 struct wq_barrier *barr,
2651 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002652{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002653 struct list_head *head;
2654 unsigned int linked = 0;
2655
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002656 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002657 * debugobject calls are safe here even with pool->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002658 * as we know for sure that this will not trigger any of the
2659 * checks and call back into the fixup functions where we
2660 * might deadlock.
2661 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002662 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002663 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Boqun Feng52fa5bc2017-08-17 17:46:12 +08002664
Byungchul Parkfd1a5b02017-10-25 17:56:04 +09002665 init_completion_map(&barr->done, &target->lockdep_map);
2666
Tejun Heo2607d7a2015-03-09 09:22:28 -04002667 barr->task = current;
Oleg Nesterov83c22522007-05-09 02:33:54 -07002668
Tejun Heoaffee4b2010-06-29 10:07:12 +02002669 /*
2670 * If @target is currently being executed, schedule the
2671 * barrier to the worker; otherwise, put it after @target.
2672 */
2673 if (worker)
2674 head = worker->scheduled.next;
2675 else {
2676 unsigned long *bits = work_data_bits(target);
2677
2678 head = target->entry.next;
2679 /* there can already be other linked works, inherit and set */
2680 linked = *bits & WORK_STRUCT_LINKED;
2681 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2682 }
2683
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002684 debug_work_activate(&barr->work);
Tejun Heo112202d2013-02-13 19:29:12 -08002685 insert_work(pwq, &barr->work, head,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002686 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002687}
2688
Tejun Heo73f53c42010-06-29 10:07:11 +02002689/**
Tejun Heo112202d2013-02-13 19:29:12 -08002690 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
Tejun Heo73f53c42010-06-29 10:07:11 +02002691 * @wq: workqueue being flushed
2692 * @flush_color: new flush color, < 0 for no-op
2693 * @work_color: new work color, < 0 for no-op
2694 *
Tejun Heo112202d2013-02-13 19:29:12 -08002695 * Prepare pwqs for workqueue flushing.
Tejun Heo73f53c42010-06-29 10:07:11 +02002696 *
Tejun Heo112202d2013-02-13 19:29:12 -08002697 * If @flush_color is non-negative, flush_color on all pwqs should be
2698 * -1. If no pwq has in-flight commands at the specified color, all
2699 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2700 * has in flight commands, its pwq->flush_color is set to
2701 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
Tejun Heo73f53c42010-06-29 10:07:11 +02002702 * wakeup logic is armed and %true is returned.
2703 *
2704 * The caller should have initialized @wq->first_flusher prior to
2705 * calling this function with non-negative @flush_color. If
2706 * @flush_color is negative, no flush color update is done and %false
2707 * is returned.
2708 *
Tejun Heo112202d2013-02-13 19:29:12 -08002709 * If @work_color is non-negative, all pwqs should have the same
Tejun Heo73f53c42010-06-29 10:07:11 +02002710 * work_color which is previous to @work_color and all will be
2711 * advanced to @work_color.
2712 *
2713 * CONTEXT:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002714 * mutex_lock(wq->mutex).
Tejun Heo73f53c42010-06-29 10:07:11 +02002715 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002716 * Return:
Tejun Heo73f53c42010-06-29 10:07:11 +02002717 * %true if @flush_color >= 0 and there's something to flush. %false
2718 * otherwise.
2719 */
Tejun Heo112202d2013-02-13 19:29:12 -08002720static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
Tejun Heo73f53c42010-06-29 10:07:11 +02002721 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722{
Tejun Heo73f53c42010-06-29 10:07:11 +02002723 bool wait = false;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002724 struct pool_workqueue *pwq;
Oleg Nesterov14441962007-05-23 13:57:57 -07002725
Tejun Heo73f53c42010-06-29 10:07:11 +02002726 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002727 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
Tejun Heo112202d2013-02-13 19:29:12 -08002728 atomic_set(&wq->nr_pwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002729 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002730
Tejun Heo49e3cf42013-03-12 11:29:58 -07002731 for_each_pwq(pwq, wq) {
Tejun Heo112202d2013-02-13 19:29:12 -08002732 struct worker_pool *pool = pwq->pool;
Tejun Heo73f53c42010-06-29 10:07:11 +02002733
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002734 spin_lock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002735
2736 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002737 WARN_ON_ONCE(pwq->flush_color != -1);
Tejun Heo73f53c42010-06-29 10:07:11 +02002738
Tejun Heo112202d2013-02-13 19:29:12 -08002739 if (pwq->nr_in_flight[flush_color]) {
2740 pwq->flush_color = flush_color;
2741 atomic_inc(&wq->nr_pwqs_to_flush);
Tejun Heo73f53c42010-06-29 10:07:11 +02002742 wait = true;
2743 }
2744 }
2745
2746 if (work_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002747 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
Tejun Heo112202d2013-02-13 19:29:12 -08002748 pwq->work_color = work_color;
Tejun Heo73f53c42010-06-29 10:07:11 +02002749 }
2750
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002751 spin_unlock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002752 }
2753
Tejun Heo112202d2013-02-13 19:29:12 -08002754 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
Tejun Heo73f53c42010-06-29 10:07:11 +02002755 complete(&wq->first_flusher->done);
2756
2757 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758}
2759
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002760/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002762 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002764 * This function sleeps until all work items which were queued on entry
2765 * have finished execution, but it is not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002767void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768{
Tejun Heo73f53c42010-06-29 10:07:11 +02002769 struct wq_flusher this_flusher = {
2770 .list = LIST_HEAD_INIT(this_flusher.list),
2771 .flush_color = -1,
Byungchul Parkfd1a5b02017-10-25 17:56:04 +09002772 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
Tejun Heo73f53c42010-06-29 10:07:11 +02002773 };
2774 int next_color;
Oleg Nesterovb1f4ec172007-05-09 02:34:12 -07002775
Tejun Heo3347fa02016-09-16 15:49:32 -04002776 if (WARN_ON(!wq_online))
2777 return;
2778
Johannes Berg87915ad2018-08-22 11:49:04 +02002779 lock_map_acquire(&wq->lockdep_map);
2780 lock_map_release(&wq->lockdep_map);
2781
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002782 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002783
2784 /*
2785 * Start-to-wait phase
2786 */
2787 next_color = work_next_color(wq->work_color);
2788
2789 if (next_color != wq->flush_color) {
2790 /*
2791 * Color space is not full. The current work_color
2792 * becomes our flush_color and work_color is advanced
2793 * by one.
2794 */
Tejun Heo6183c002013-03-12 11:29:57 -07002795 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
Tejun Heo73f53c42010-06-29 10:07:11 +02002796 this_flusher.flush_color = wq->work_color;
2797 wq->work_color = next_color;
2798
2799 if (!wq->first_flusher) {
2800 /* no flush in progress, become the first flusher */
Tejun Heo6183c002013-03-12 11:29:57 -07002801 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002802
2803 wq->first_flusher = &this_flusher;
2804
Tejun Heo112202d2013-02-13 19:29:12 -08002805 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
Tejun Heo73f53c42010-06-29 10:07:11 +02002806 wq->work_color)) {
2807 /* nothing to flush, done */
2808 wq->flush_color = next_color;
2809 wq->first_flusher = NULL;
2810 goto out_unlock;
2811 }
2812 } else {
2813 /* wait in queue */
Tejun Heo6183c002013-03-12 11:29:57 -07002814 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002815 list_add_tail(&this_flusher.list, &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002816 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002817 }
2818 } else {
2819 /*
2820 * Oops, color space is full, wait on overflow queue.
2821 * The next flush completion will assign us
2822 * flush_color and transfer to flusher_queue.
2823 */
2824 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2825 }
2826
Tejun Heofca839c2015-12-07 10:58:57 -05002827 check_flush_dependency(wq, NULL);
2828
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002829 mutex_unlock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002830
2831 wait_for_completion(&this_flusher.done);
2832
2833 /*
2834 * Wake-up-and-cascade phase
2835 *
2836 * First flushers are responsible for cascading flushes and
2837 * handling overflow. Non-first flushers can simply return.
2838 */
Chris Wilson00d5d152020-03-10 16:23:19 +00002839 if (READ_ONCE(wq->first_flusher) != &this_flusher)
Tejun Heo73f53c42010-06-29 10:07:11 +02002840 return;
2841
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002842 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002843
Tejun Heo4ce48b32010-07-02 10:03:51 +02002844 /* we might have raced, check again with mutex held */
2845 if (wq->first_flusher != &this_flusher)
2846 goto out_unlock;
2847
Chris Wilson00d5d152020-03-10 16:23:19 +00002848 WRITE_ONCE(wq->first_flusher, NULL);
Tejun Heo73f53c42010-06-29 10:07:11 +02002849
Tejun Heo6183c002013-03-12 11:29:57 -07002850 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2851 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002852
2853 while (true) {
2854 struct wq_flusher *next, *tmp;
2855
2856 /* complete all the flushers sharing the current flush color */
2857 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2858 if (next->flush_color != wq->flush_color)
2859 break;
2860 list_del_init(&next->list);
2861 complete(&next->done);
2862 }
2863
Tejun Heo6183c002013-03-12 11:29:57 -07002864 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2865 wq->flush_color != work_next_color(wq->work_color));
Tejun Heo73f53c42010-06-29 10:07:11 +02002866
2867 /* this flush_color is finished, advance by one */
2868 wq->flush_color = work_next_color(wq->flush_color);
2869
2870 /* one color has been freed, handle overflow queue */
2871 if (!list_empty(&wq->flusher_overflow)) {
2872 /*
2873 * Assign the same color to all overflowed
2874 * flushers, advance work_color and append to
2875 * flusher_queue. This is the start-to-wait
2876 * phase for these overflowed flushers.
2877 */
2878 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2879 tmp->flush_color = wq->work_color;
2880
2881 wq->work_color = work_next_color(wq->work_color);
2882
2883 list_splice_tail_init(&wq->flusher_overflow,
2884 &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002885 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002886 }
2887
2888 if (list_empty(&wq->flusher_queue)) {
Tejun Heo6183c002013-03-12 11:29:57 -07002889 WARN_ON_ONCE(wq->flush_color != wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002890 break;
2891 }
2892
2893 /*
2894 * Need to flush more colors. Make the next flusher
Tejun Heo112202d2013-02-13 19:29:12 -08002895 * the new first flusher and arm pwqs.
Tejun Heo73f53c42010-06-29 10:07:11 +02002896 */
Tejun Heo6183c002013-03-12 11:29:57 -07002897 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2898 WARN_ON_ONCE(wq->flush_color != next->flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002899
2900 list_del_init(&next->list);
2901 wq->first_flusher = next;
2902
Tejun Heo112202d2013-02-13 19:29:12 -08002903 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
Tejun Heo73f53c42010-06-29 10:07:11 +02002904 break;
2905
2906 /*
2907 * Meh... this color is already done, clear first
2908 * flusher and repeat cascading.
2909 */
2910 wq->first_flusher = NULL;
2911 }
2912
2913out_unlock:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002914 mutex_unlock(&wq->mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915}
Tim Gardner1dadafa2015-08-04 11:26:04 -06002916EXPORT_SYMBOL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002918/**
2919 * drain_workqueue - drain a workqueue
2920 * @wq: workqueue to drain
2921 *
2922 * Wait until the workqueue becomes empty. While draining is in progress,
2923 * only chain queueing is allowed. IOW, only currently pending or running
2924 * work items on @wq can queue further work items on it. @wq is flushed
Chen Hanxiaob749b1b2015-05-13 06:10:05 -04002925 * repeatedly until it becomes empty. The number of flushing is determined
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002926 * by the depth of chaining and should be relatively short. Whine if it
2927 * takes too long.
2928 */
2929void drain_workqueue(struct workqueue_struct *wq)
2930{
2931 unsigned int flush_cnt = 0;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002932 struct pool_workqueue *pwq;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002933
2934 /*
2935 * __queue_work() needs to test whether there are drainers, is much
2936 * hotter than drain_workqueue() and already looks at @wq->flags.
Tejun Heo618b01e2013-03-12 11:30:04 -07002937 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002938 */
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002939 mutex_lock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002940 if (!wq->nr_drainers++)
Tejun Heo618b01e2013-03-12 11:30:04 -07002941 wq->flags |= __WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002942 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002943reflush:
2944 flush_workqueue(wq);
2945
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002946 mutex_lock(&wq->mutex);
Tejun Heo76af4d92013-03-12 11:30:00 -07002947
Tejun Heo49e3cf42013-03-12 11:29:58 -07002948 for_each_pwq(pwq, wq) {
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002949 bool drained;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002950
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002951 spin_lock_irq(&pwq->pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08002952 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002953 spin_unlock_irq(&pwq->pool->lock);
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002954
2955 if (drained)
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002956 continue;
2957
2958 if (++flush_cnt == 10 ||
2959 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002960 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
Valentin Ilie044c7822012-08-19 00:52:42 +03002961 wq->name, flush_cnt);
Tejun Heo76af4d92013-03-12 11:30:00 -07002962
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002963 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002964 goto reflush;
2965 }
2966
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002967 if (!--wq->nr_drainers)
Tejun Heo618b01e2013-03-12 11:30:04 -07002968 wq->flags &= ~__WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002969 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002970}
2971EXPORT_SYMBOL_GPL(drain_workqueue);
2972
Johannes Bergd6e89782018-08-22 11:49:03 +02002973static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2974 bool from_cancel)
Tejun Heobaf59022010-09-16 10:42:16 +02002975{
2976 struct worker *worker = NULL;
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002977 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002978 struct pool_workqueue *pwq;
Tejun Heobaf59022010-09-16 10:42:16 +02002979
2980 might_sleep();
Tejun Heobaf59022010-09-16 10:42:16 +02002981
Thomas Gleixner24acfb72019-03-13 17:55:47 +01002982 rcu_read_lock();
Tejun Heofa1b54e2013-03-12 11:30:00 -07002983 pool = get_work_pool(work);
2984 if (!pool) {
Thomas Gleixner24acfb72019-03-13 17:55:47 +01002985 rcu_read_unlock();
Tejun Heofa1b54e2013-03-12 11:30:00 -07002986 return false;
2987 }
2988
Thomas Gleixner24acfb72019-03-13 17:55:47 +01002989 spin_lock_irq(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08002990 /* see the comment in try_to_grab_pending() with the same code */
Tejun Heo112202d2013-02-13 19:29:12 -08002991 pwq = get_work_pwq(work);
2992 if (pwq) {
2993 if (unlikely(pwq->pool != pool))
Tejun Heobaf59022010-09-16 10:42:16 +02002994 goto already_gone;
Tejun Heo606a5022012-08-20 14:51:23 -07002995 } else {
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002996 worker = find_worker_executing_work(pool, work);
Tejun Heobaf59022010-09-16 10:42:16 +02002997 if (!worker)
2998 goto already_gone;
Tejun Heo112202d2013-02-13 19:29:12 -08002999 pwq = worker->current_pwq;
Tejun Heo606a5022012-08-20 14:51:23 -07003000 }
Tejun Heobaf59022010-09-16 10:42:16 +02003001
Tejun Heofca839c2015-12-07 10:58:57 -05003002 check_flush_dependency(pwq->wq, work);
3003
Tejun Heo112202d2013-02-13 19:29:12 -08003004 insert_wq_barrier(pwq, barr, work, worker);
Tejun Heod565ed62013-01-24 11:01:33 -08003005 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02003006
Tejun Heoe159489b2011-01-09 23:32:15 +01003007 /*
Peter Zijlstraa1d14932017-08-23 12:52:32 +02003008 * Force a lock recursion deadlock when using flush_work() inside a
3009 * single-threaded or rescuer equipped workqueue.
3010 *
3011 * For single threaded workqueues the deadlock happens when the work
3012 * is after the work issuing the flush_work(). For rescuer equipped
3013 * workqueues the deadlock happens when the rescuer stalls, blocking
3014 * forward progress.
Tejun Heoe159489b2011-01-09 23:32:15 +01003015 */
Johannes Bergd6e89782018-08-22 11:49:03 +02003016 if (!from_cancel &&
3017 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
Tejun Heo112202d2013-02-13 19:29:12 -08003018 lock_map_acquire(&pwq->wq->lockdep_map);
Peter Zijlstraa1d14932017-08-23 12:52:32 +02003019 lock_map_release(&pwq->wq->lockdep_map);
3020 }
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003021 rcu_read_unlock();
Tejun Heobaf59022010-09-16 10:42:16 +02003022 return true;
3023already_gone:
Tejun Heod565ed62013-01-24 11:01:33 -08003024 spin_unlock_irq(&pool->lock);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003025 rcu_read_unlock();
Tejun Heobaf59022010-09-16 10:42:16 +02003026 return false;
3027}
3028
Johannes Bergd6e89782018-08-22 11:49:03 +02003029static bool __flush_work(struct work_struct *work, bool from_cancel)
3030{
3031 struct wq_barrier barr;
3032
3033 if (WARN_ON(!wq_online))
3034 return false;
3035
Tetsuo Handa4d43d392019-01-23 09:44:12 +09003036 if (WARN_ON(!work->func))
3037 return false;
3038
Johannes Berg87915ad2018-08-22 11:49:04 +02003039 if (!from_cancel) {
3040 lock_map_acquire(&work->lockdep_map);
3041 lock_map_release(&work->lockdep_map);
3042 }
3043
Johannes Bergd6e89782018-08-22 11:49:03 +02003044 if (start_flush_work(work, &barr, from_cancel)) {
3045 wait_for_completion(&barr.done);
3046 destroy_work_on_stack(&barr.work);
3047 return true;
3048 } else {
3049 return false;
3050 }
3051}
3052
Oleg Nesterovdb700892008-07-25 01:47:49 -07003053/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003054 * flush_work - wait for a work to finish executing the last queueing instance
3055 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07003056 *
Tejun Heo606a5022012-08-20 14:51:23 -07003057 * Wait until @work has finished execution. @work is guaranteed to be idle
3058 * on return if it hasn't been requeued since flush started.
Tejun Heo401a8d02010-09-16 10:36:00 +02003059 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003060 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003061 * %true if flush_work() waited for the work to finish execution,
3062 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07003063 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003064bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07003065{
Johannes Bergd6e89782018-08-22 11:49:03 +02003066 return __flush_work(work, false);
Oleg Nesterovdb700892008-07-25 01:47:49 -07003067}
3068EXPORT_SYMBOL_GPL(flush_work);
3069
Tejun Heo8603e1b32015-03-05 08:04:13 -05003070struct cwt_wait {
Ingo Molnarac6424b2017-06-20 12:06:13 +02003071 wait_queue_entry_t wait;
Tejun Heo8603e1b32015-03-05 08:04:13 -05003072 struct work_struct *work;
3073};
3074
Ingo Molnarac6424b2017-06-20 12:06:13 +02003075static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
Tejun Heo8603e1b32015-03-05 08:04:13 -05003076{
3077 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3078
3079 if (cwait->work != key)
3080 return 0;
3081 return autoremove_wake_function(wait, mode, sync, key);
3082}
3083
Tejun Heo36e227d2012-08-03 10:30:46 -07003084static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
Tejun Heo401a8d02010-09-16 10:36:00 +02003085{
Tejun Heo8603e1b32015-03-05 08:04:13 -05003086 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
Tejun Heobbb68df2012-08-03 10:30:46 -07003087 unsigned long flags;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003088 int ret;
3089
3090 do {
Tejun Heobbb68df2012-08-03 10:30:46 -07003091 ret = try_to_grab_pending(work, is_dwork, &flags);
3092 /*
Tejun Heo8603e1b32015-03-05 08:04:13 -05003093 * If someone else is already canceling, wait for it to
3094 * finish. flush_work() doesn't work for PREEMPT_NONE
3095 * because we may get scheduled between @work's completion
3096 * and the other canceling task resuming and clearing
3097 * CANCELING - flush_work() will return false immediately
3098 * as @work is no longer busy, try_to_grab_pending() will
3099 * return -ENOENT as @work is still being canceled and the
3100 * other canceling task won't be able to clear CANCELING as
3101 * we're hogging the CPU.
3102 *
3103 * Let's wait for completion using a waitqueue. As this
3104 * may lead to the thundering herd problem, use a custom
3105 * wake function which matches @work along with exclusive
3106 * wait and wakeup.
Tejun Heobbb68df2012-08-03 10:30:46 -07003107 */
Tejun Heo8603e1b32015-03-05 08:04:13 -05003108 if (unlikely(ret == -ENOENT)) {
3109 struct cwt_wait cwait;
3110
3111 init_wait(&cwait.wait);
3112 cwait.wait.func = cwt_wakefn;
3113 cwait.work = work;
3114
3115 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3116 TASK_UNINTERRUPTIBLE);
3117 if (work_is_canceling(work))
3118 schedule();
3119 finish_wait(&cancel_waitq, &cwait.wait);
3120 }
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003121 } while (unlikely(ret < 0));
3122
Tejun Heobbb68df2012-08-03 10:30:46 -07003123 /* tell other tasks trying to grab @work to back off */
3124 mark_work_canceling(work);
3125 local_irq_restore(flags);
3126
Tejun Heo3347fa02016-09-16 15:49:32 -04003127 /*
3128 * This allows canceling during early boot. We know that @work
3129 * isn't executing.
3130 */
3131 if (wq_online)
Johannes Bergd6e89782018-08-22 11:49:03 +02003132 __flush_work(work, true);
Tejun Heo3347fa02016-09-16 15:49:32 -04003133
Tejun Heo7a22ad72010-06-29 10:07:13 +02003134 clear_work_data(work);
Tejun Heo8603e1b32015-03-05 08:04:13 -05003135
3136 /*
3137 * Paired with prepare_to_wait() above so that either
3138 * waitqueue_active() is visible here or !work_is_canceling() is
3139 * visible there.
3140 */
3141 smp_mb();
3142 if (waitqueue_active(&cancel_waitq))
3143 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3144
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003145 return ret;
3146}
3147
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003148/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003149 * cancel_work_sync - cancel a work and wait for it to finish
3150 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003151 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003152 * Cancel @work and wait for its execution to finish. This function
3153 * can be used even if the work re-queues itself or migrates to
3154 * another workqueue. On return from this function, @work is
3155 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003156 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003157 * cancel_work_sync(&delayed_work->work) must not be used for
3158 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003159 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003160 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003161 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02003162 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003163 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003164 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003165 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003166bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003167{
Tejun Heo36e227d2012-08-03 10:30:46 -07003168 return __cancel_work_timer(work, false);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07003169}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07003170EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07003171
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003172/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003173 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3174 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003175 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003176 * Delayed timer is cancelled and the pending work is queued for
3177 * immediate execution. Like flush_work(), this function only
3178 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003179 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003180 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003181 * %true if flush_work() waited for the work to finish execution,
3182 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003183 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003184bool flush_delayed_work(struct delayed_work *dwork)
3185{
Tejun Heo8930cab2012-08-03 10:30:45 -07003186 local_irq_disable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003187 if (del_timer_sync(&dwork->timer))
Lai Jiangshan60c057b2013-02-06 18:04:53 -08003188 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Tejun Heo8930cab2012-08-03 10:30:45 -07003189 local_irq_enable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003190 return flush_work(&dwork->work);
3191}
3192EXPORT_SYMBOL(flush_delayed_work);
3193
Tejun Heo05f0fe62018-03-14 12:45:13 -07003194/**
3195 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3196 * @rwork: the rcu work to flush
3197 *
3198 * Return:
3199 * %true if flush_rcu_work() waited for the work to finish execution,
3200 * %false if it was already idle.
3201 */
3202bool flush_rcu_work(struct rcu_work *rwork)
3203{
3204 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3205 rcu_barrier();
3206 flush_work(&rwork->work);
3207 return true;
3208 } else {
3209 return flush_work(&rwork->work);
3210 }
3211}
3212EXPORT_SYMBOL(flush_rcu_work);
3213
Jens Axboef72b8792016-08-24 15:51:50 -06003214static bool __cancel_work(struct work_struct *work, bool is_dwork)
3215{
3216 unsigned long flags;
3217 int ret;
3218
3219 do {
3220 ret = try_to_grab_pending(work, is_dwork, &flags);
3221 } while (unlikely(ret == -EAGAIN));
3222
3223 if (unlikely(ret < 0))
3224 return false;
3225
3226 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3227 local_irq_restore(flags);
3228 return ret;
3229}
3230
Tejun Heo401a8d02010-09-16 10:36:00 +02003231/**
Tejun Heo57b30ae2012-08-21 13:18:24 -07003232 * cancel_delayed_work - cancel a delayed work
3233 * @dwork: delayed_work to cancel
Tejun Heo09383492010-09-16 10:48:29 +02003234 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003235 * Kill off a pending delayed_work.
3236 *
3237 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3238 * pending.
3239 *
3240 * Note:
3241 * The work callback function may still be running on return, unless
3242 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3243 * use cancel_delayed_work_sync() to wait on it.
Tejun Heo09383492010-09-16 10:48:29 +02003244 *
Tejun Heo57b30ae2012-08-21 13:18:24 -07003245 * This function is safe to call from any context including IRQ handler.
Tejun Heo09383492010-09-16 10:48:29 +02003246 */
Tejun Heo57b30ae2012-08-21 13:18:24 -07003247bool cancel_delayed_work(struct delayed_work *dwork)
Tejun Heo09383492010-09-16 10:48:29 +02003248{
Jens Axboef72b8792016-08-24 15:51:50 -06003249 return __cancel_work(&dwork->work, true);
Tejun Heo09383492010-09-16 10:48:29 +02003250}
Tejun Heo57b30ae2012-08-21 13:18:24 -07003251EXPORT_SYMBOL(cancel_delayed_work);
Tejun Heo09383492010-09-16 10:48:29 +02003252
3253/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003254 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3255 * @dwork: the delayed work cancel
3256 *
3257 * This is cancel_work_sync() for delayed works.
3258 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003259 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003260 * %true if @dwork was pending, %false otherwise.
3261 */
3262bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003263{
Tejun Heo36e227d2012-08-03 10:30:46 -07003264 return __cancel_work_timer(&dwork->work, true);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003265}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07003266EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07003268/**
Tejun Heo31ddd872010-10-19 11:14:49 +02003269 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07003270 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07003271 *
Tejun Heo31ddd872010-10-19 11:14:49 +02003272 * schedule_on_each_cpu() executes @func on each online CPU using the
3273 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07003274 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02003275 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003276 * Return:
Tejun Heo31ddd872010-10-19 11:14:49 +02003277 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07003278 */
David Howells65f27f32006-11-22 14:55:48 +00003279int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba82006-01-08 01:00:43 -08003280{
3281 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02003282 struct work_struct __percpu *works;
Christoph Lameter15316ba82006-01-08 01:00:43 -08003283
Andrew Mortonb6136772006-06-25 05:47:49 -07003284 works = alloc_percpu(struct work_struct);
3285 if (!works)
Christoph Lameter15316ba82006-01-08 01:00:43 -08003286 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07003287
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003288 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08003289
Christoph Lameter15316ba82006-01-08 01:00:43 -08003290 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01003291 struct work_struct *work = per_cpu_ptr(works, cpu);
3292
3293 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003294 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02003295 }
Tejun Heo93981802009-11-17 14:06:20 -08003296
3297 for_each_online_cpu(cpu)
3298 flush_work(per_cpu_ptr(works, cpu));
3299
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003300 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07003301 free_percpu(works);
Christoph Lameter15316ba82006-01-08 01:00:43 -08003302 return 0;
3303}
3304
Alan Sterneef6a7d2010-02-12 17:39:21 +09003305/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06003306 * execute_in_process_context - reliably execute the routine with user context
3307 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06003308 * @ew: guaranteed storage for the execute work structure (must
3309 * be available when the work executes)
3310 *
3311 * Executes the function immediately if process context is available,
3312 * otherwise schedules the function for delayed execution.
3313 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003314 * Return: 0 - function was executed
James Bottomley1fa44ec2006-02-23 12:43:43 -06003315 * 1 - function was scheduled for execution
3316 */
David Howells65f27f32006-11-22 14:55:48 +00003317int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06003318{
3319 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00003320 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003321 return 0;
3322 }
3323
David Howells65f27f32006-11-22 14:55:48 +00003324 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003325 schedule_work(&ew->work);
3326
3327 return 1;
3328}
3329EXPORT_SYMBOL_GPL(execute_in_process_context);
3330
Tejun Heo7a4e3442013-03-12 11:30:00 -07003331/**
3332 * free_workqueue_attrs - free a workqueue_attrs
3333 * @attrs: workqueue_attrs to free
3334 *
3335 * Undo alloc_workqueue_attrs().
3336 */
Daniel Jordan513c98d2019-09-05 21:40:22 -04003337void free_workqueue_attrs(struct workqueue_attrs *attrs)
Tejun Heo7a4e3442013-03-12 11:30:00 -07003338{
3339 if (attrs) {
3340 free_cpumask_var(attrs->cpumask);
3341 kfree(attrs);
3342 }
3343}
3344
3345/**
3346 * alloc_workqueue_attrs - allocate a workqueue_attrs
Tejun Heo7a4e3442013-03-12 11:30:00 -07003347 *
3348 * Allocate a new workqueue_attrs, initialize with default settings and
Yacine Belkadid185af32013-07-31 14:59:24 -07003349 * return it.
3350 *
3351 * Return: The allocated new workqueue_attr on success. %NULL on failure.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003352 */
Daniel Jordan513c98d2019-09-05 21:40:22 -04003353struct workqueue_attrs *alloc_workqueue_attrs(void)
Tejun Heo7a4e3442013-03-12 11:30:00 -07003354{
3355 struct workqueue_attrs *attrs;
3356
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003357 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003358 if (!attrs)
3359 goto fail;
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003360 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
Tejun Heo7a4e3442013-03-12 11:30:00 -07003361 goto fail;
3362
Tejun Heo13e2e552013-04-01 11:23:31 -07003363 cpumask_copy(attrs->cpumask, cpu_possible_mask);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003364 return attrs;
3365fail:
3366 free_workqueue_attrs(attrs);
3367 return NULL;
3368}
3369
Tejun Heo29c91e92013-03-12 11:30:03 -07003370static void copy_workqueue_attrs(struct workqueue_attrs *to,
3371 const struct workqueue_attrs *from)
3372{
3373 to->nice = from->nice;
3374 cpumask_copy(to->cpumask, from->cpumask);
Shaohua Li2865a8f2013-08-01 09:56:36 +08003375 /*
3376 * Unlike hash and equality test, this function doesn't ignore
3377 * ->no_numa as it is used for both pool and wq attrs. Instead,
3378 * get_unbound_pool() explicitly clears ->no_numa after copying.
3379 */
3380 to->no_numa = from->no_numa;
Tejun Heo29c91e92013-03-12 11:30:03 -07003381}
3382
Tejun Heo29c91e92013-03-12 11:30:03 -07003383/* hash value of the content of @attr */
3384static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3385{
3386 u32 hash = 0;
3387
3388 hash = jhash_1word(attrs->nice, hash);
Tejun Heo13e2e552013-04-01 11:23:31 -07003389 hash = jhash(cpumask_bits(attrs->cpumask),
3390 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
Tejun Heo29c91e92013-03-12 11:30:03 -07003391 return hash;
3392}
3393
3394/* content equality test */
3395static bool wqattrs_equal(const struct workqueue_attrs *a,
3396 const struct workqueue_attrs *b)
3397{
3398 if (a->nice != b->nice)
3399 return false;
3400 if (!cpumask_equal(a->cpumask, b->cpumask))
3401 return false;
3402 return true;
3403}
3404
Tejun Heo7a4e3442013-03-12 11:30:00 -07003405/**
3406 * init_worker_pool - initialize a newly zalloc'd worker_pool
3407 * @pool: worker_pool to initialize
3408 *
Shailendra Verma402dd892015-05-23 10:38:14 +05303409 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
Yacine Belkadid185af32013-07-31 14:59:24 -07003410 *
3411 * Return: 0 on success, -errno on failure. Even on failure, all fields
Tejun Heo29c91e92013-03-12 11:30:03 -07003412 * inside @pool proper are initialized and put_unbound_pool() can be called
3413 * on @pool safely to release it.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003414 */
3415static int init_worker_pool(struct worker_pool *pool)
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003416{
3417 spin_lock_init(&pool->lock);
Tejun Heo29c91e92013-03-12 11:30:03 -07003418 pool->id = -1;
3419 pool->cpu = -1;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003420 pool->node = NUMA_NO_NODE;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003421 pool->flags |= POOL_DISASSOCIATED;
Tejun Heo82607adc2015-12-08 11:28:04 -05003422 pool->watchdog_ts = jiffies;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003423 INIT_LIST_HEAD(&pool->worklist);
3424 INIT_LIST_HEAD(&pool->idle_list);
3425 hash_init(pool->busy_hash);
3426
Kees Cook32a6c722017-10-16 15:58:25 -07003427 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003428
Kees Cook32a6c722017-10-16 15:58:25 -07003429 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003430
Lai Jiangshanda028462014-05-20 17:46:31 +08003431 INIT_LIST_HEAD(&pool->workers);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003432
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003433 ida_init(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003434 INIT_HLIST_NODE(&pool->hash_node);
3435 pool->refcnt = 1;
3436
3437 /* shouldn't fail above this point */
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003438 pool->attrs = alloc_workqueue_attrs();
Tejun Heo7a4e3442013-03-12 11:30:00 -07003439 if (!pool->attrs)
3440 return -ENOMEM;
3441 return 0;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003442}
3443
Bart Van Assche669de8b2019-02-14 15:00:54 -08003444#ifdef CONFIG_LOCKDEP
3445static void wq_init_lockdep(struct workqueue_struct *wq)
3446{
3447 char *lock_name;
3448
3449 lockdep_register_key(&wq->key);
3450 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3451 if (!lock_name)
3452 lock_name = wq->name;
Qian Cai69a106c2019-03-06 19:27:31 -05003453
3454 wq->lock_name = lock_name;
Bart Van Assche669de8b2019-02-14 15:00:54 -08003455 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3456}
3457
3458static void wq_unregister_lockdep(struct workqueue_struct *wq)
3459{
3460 lockdep_unregister_key(&wq->key);
3461}
3462
3463static void wq_free_lockdep(struct workqueue_struct *wq)
3464{
3465 if (wq->lock_name != wq->name)
3466 kfree(wq->lock_name);
3467}
3468#else
3469static void wq_init_lockdep(struct workqueue_struct *wq)
3470{
3471}
3472
3473static void wq_unregister_lockdep(struct workqueue_struct *wq)
3474{
3475}
3476
3477static void wq_free_lockdep(struct workqueue_struct *wq)
3478{
3479}
3480#endif
3481
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003482static void rcu_free_wq(struct rcu_head *rcu)
3483{
3484 struct workqueue_struct *wq =
3485 container_of(rcu, struct workqueue_struct, rcu);
3486
Bart Van Assche669de8b2019-02-14 15:00:54 -08003487 wq_free_lockdep(wq);
3488
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003489 if (!(wq->flags & WQ_UNBOUND))
3490 free_percpu(wq->cpu_pwqs);
3491 else
3492 free_workqueue_attrs(wq->unbound_attrs);
3493
3494 kfree(wq->rescuer);
3495 kfree(wq);
3496}
3497
Tejun Heo29c91e92013-03-12 11:30:03 -07003498static void rcu_free_pool(struct rcu_head *rcu)
3499{
3500 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3501
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003502 ida_destroy(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003503 free_workqueue_attrs(pool->attrs);
3504 kfree(pool);
3505}
3506
3507/**
3508 * put_unbound_pool - put a worker_pool
3509 * @pool: worker_pool to put
3510 *
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003511 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003512 * safe manner. get_unbound_pool() calls this function on its failure path
3513 * and this function should be able to release pools which went through,
3514 * successfully or not, init_worker_pool().
Tejun Heoa892cac2013-04-01 11:23:32 -07003515 *
3516 * Should be called with wq_pool_mutex held.
Tejun Heo29c91e92013-03-12 11:30:03 -07003517 */
3518static void put_unbound_pool(struct worker_pool *pool)
3519{
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003520 DECLARE_COMPLETION_ONSTACK(detach_completion);
Tejun Heo29c91e92013-03-12 11:30:03 -07003521 struct worker *worker;
3522
Tejun Heoa892cac2013-04-01 11:23:32 -07003523 lockdep_assert_held(&wq_pool_mutex);
3524
3525 if (--pool->refcnt)
Tejun Heo29c91e92013-03-12 11:30:03 -07003526 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003527
3528 /* sanity checks */
Lai Jiangshan61d0fbb2014-06-03 15:31:45 +08003529 if (WARN_ON(!(pool->cpu < 0)) ||
Tejun Heoa892cac2013-04-01 11:23:32 -07003530 WARN_ON(!list_empty(&pool->worklist)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003531 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003532
3533 /* release id and unhash */
3534 if (pool->id >= 0)
3535 idr_remove(&worker_pool_idr, pool->id);
3536 hash_del(&pool->hash_node);
3537
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003538 /*
Tejun Heo692b4822017-10-09 08:04:13 -07003539 * Become the manager and destroy all workers. This prevents
3540 * @pool's workers from blocking on attach_mutex. We're the last
3541 * manager and @pool gets freed with the flag set.
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003542 */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003543 spin_lock_irq(&pool->lock);
Tejun Heo692b4822017-10-09 08:04:13 -07003544 wait_event_lock_irq(wq_manager_wait,
3545 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3546 pool->flags |= POOL_MANAGER_ACTIVE;
3547
Lai Jiangshan1037de32014-05-22 16:44:07 +08003548 while ((worker = first_idle_worker(pool)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003549 destroy_worker(worker);
3550 WARN_ON(pool->nr_workers || pool->nr_idle);
Tejun Heo29c91e92013-03-12 11:30:03 -07003551 spin_unlock_irq(&pool->lock);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003552
Tejun Heo1258fae2018-05-18 08:47:13 -07003553 mutex_lock(&wq_pool_attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003554 if (!list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003555 pool->detach_completion = &detach_completion;
Tejun Heo1258fae2018-05-18 08:47:13 -07003556 mutex_unlock(&wq_pool_attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003557
3558 if (pool->detach_completion)
3559 wait_for_completion(pool->detach_completion);
3560
Tejun Heo29c91e92013-03-12 11:30:03 -07003561 /* shut down the timers */
3562 del_timer_sync(&pool->idle_timer);
3563 del_timer_sync(&pool->mayday_timer);
3564
Thomas Gleixner24acfb72019-03-13 17:55:47 +01003565 /* RCU protected to allow dereferences from get_work_pool() */
Paul E. McKenney25b00772018-11-06 19:18:45 -08003566 call_rcu(&pool->rcu, rcu_free_pool);
Tejun Heo29c91e92013-03-12 11:30:03 -07003567}
3568
3569/**
3570 * get_unbound_pool - get a worker_pool with the specified attributes
3571 * @attrs: the attributes of the worker_pool to get
3572 *
3573 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3574 * reference count and return it. If there already is a matching
3575 * worker_pool, it will be used; otherwise, this function attempts to
Yacine Belkadid185af32013-07-31 14:59:24 -07003576 * create a new one.
Tejun Heoa892cac2013-04-01 11:23:32 -07003577 *
3578 * Should be called with wq_pool_mutex held.
Yacine Belkadid185af32013-07-31 14:59:24 -07003579 *
3580 * Return: On success, a worker_pool with the same attributes as @attrs.
3581 * On failure, %NULL.
Tejun Heo29c91e92013-03-12 11:30:03 -07003582 */
3583static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3584{
Tejun Heo29c91e92013-03-12 11:30:03 -07003585 u32 hash = wqattrs_hash(attrs);
3586 struct worker_pool *pool;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003587 int node;
Xunlei Pange22735842015-10-09 11:53:12 +08003588 int target_node = NUMA_NO_NODE;
Tejun Heo29c91e92013-03-12 11:30:03 -07003589
Tejun Heoa892cac2013-04-01 11:23:32 -07003590 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003591
3592 /* do we already have a matching pool? */
Tejun Heo29c91e92013-03-12 11:30:03 -07003593 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3594 if (wqattrs_equal(pool->attrs, attrs)) {
3595 pool->refcnt++;
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003596 return pool;
Tejun Heo29c91e92013-03-12 11:30:03 -07003597 }
3598 }
Tejun Heo29c91e92013-03-12 11:30:03 -07003599
Xunlei Pange22735842015-10-09 11:53:12 +08003600 /* if cpumask is contained inside a NUMA node, we belong to that node */
3601 if (wq_numa_enabled) {
3602 for_each_node(node) {
3603 if (cpumask_subset(attrs->cpumask,
3604 wq_numa_possible_cpumask[node])) {
3605 target_node = node;
3606 break;
3607 }
3608 }
3609 }
3610
Tejun Heo29c91e92013-03-12 11:30:03 -07003611 /* nope, create a new one */
Xunlei Pange22735842015-10-09 11:53:12 +08003612 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
Tejun Heo29c91e92013-03-12 11:30:03 -07003613 if (!pool || init_worker_pool(pool) < 0)
3614 goto fail;
3615
Tejun Heo8864b4e2013-03-12 11:30:04 -07003616 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
Tejun Heo29c91e92013-03-12 11:30:03 -07003617 copy_workqueue_attrs(pool->attrs, attrs);
Xunlei Pange22735842015-10-09 11:53:12 +08003618 pool->node = target_node;
Tejun Heo29c91e92013-03-12 11:30:03 -07003619
Shaohua Li2865a8f2013-08-01 09:56:36 +08003620 /*
3621 * no_numa isn't a worker_pool attribute, always clear it. See
3622 * 'struct workqueue_attrs' comments for detail.
3623 */
3624 pool->attrs->no_numa = false;
3625
Tejun Heo29c91e92013-03-12 11:30:03 -07003626 if (worker_pool_assign_id(pool) < 0)
3627 goto fail;
3628
3629 /* create and start the initial worker */
Tejun Heo3347fa02016-09-16 15:49:32 -04003630 if (wq_online && !create_worker(pool))
Tejun Heo29c91e92013-03-12 11:30:03 -07003631 goto fail;
3632
Tejun Heo29c91e92013-03-12 11:30:03 -07003633 /* install */
Tejun Heo29c91e92013-03-12 11:30:03 -07003634 hash_add(unbound_pool_hash, &pool->hash_node, hash);
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003635
Tejun Heo29c91e92013-03-12 11:30:03 -07003636 return pool;
3637fail:
Tejun Heo29c91e92013-03-12 11:30:03 -07003638 if (pool)
3639 put_unbound_pool(pool);
3640 return NULL;
3641}
3642
Tejun Heo8864b4e2013-03-12 11:30:04 -07003643static void rcu_free_pwq(struct rcu_head *rcu)
3644{
3645 kmem_cache_free(pwq_cache,
3646 container_of(rcu, struct pool_workqueue, rcu));
3647}
3648
3649/*
3650 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3651 * and needs to be destroyed.
3652 */
3653static void pwq_unbound_release_workfn(struct work_struct *work)
3654{
3655 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3656 unbound_release_work);
3657 struct workqueue_struct *wq = pwq->wq;
3658 struct worker_pool *pool = pwq->pool;
Tejun Heobc0caf02013-04-01 11:23:31 -07003659 bool is_last;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003660
3661 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3662 return;
3663
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003664 mutex_lock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003665 list_del_rcu(&pwq->pwqs_node);
Tejun Heobc0caf02013-04-01 11:23:31 -07003666 is_last = list_empty(&wq->pwqs);
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003667 mutex_unlock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003668
Tejun Heoa892cac2013-04-01 11:23:32 -07003669 mutex_lock(&wq_pool_mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003670 put_unbound_pool(pool);
Tejun Heoa892cac2013-04-01 11:23:32 -07003671 mutex_unlock(&wq_pool_mutex);
3672
Paul E. McKenney25b00772018-11-06 19:18:45 -08003673 call_rcu(&pwq->rcu, rcu_free_pwq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003674
3675 /*
3676 * If we're the last pwq going away, @wq is already dead and no one
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003677 * is gonna access it anymore. Schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003678 */
Bart Van Assche669de8b2019-02-14 15:00:54 -08003679 if (is_last) {
3680 wq_unregister_lockdep(wq);
Paul E. McKenney25b00772018-11-06 19:18:45 -08003681 call_rcu(&wq->rcu, rcu_free_wq);
Bart Van Assche669de8b2019-02-14 15:00:54 -08003682 }
Tejun Heo8864b4e2013-03-12 11:30:04 -07003683}
3684
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003685/**
Tejun Heo699ce092013-03-13 16:51:35 -07003686 * pwq_adjust_max_active - update a pwq's max_active to the current setting
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003687 * @pwq: target pool_workqueue
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003688 *
Tejun Heo699ce092013-03-13 16:51:35 -07003689 * If @pwq isn't freezing, set @pwq->max_active to the associated
3690 * workqueue's saved_max_active and activate delayed work items
3691 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003692 */
Tejun Heo699ce092013-03-13 16:51:35 -07003693static void pwq_adjust_max_active(struct pool_workqueue *pwq)
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003694{
Tejun Heo699ce092013-03-13 16:51:35 -07003695 struct workqueue_struct *wq = pwq->wq;
3696 bool freezable = wq->flags & WQ_FREEZABLE;
Tejun Heo3347fa02016-09-16 15:49:32 -04003697 unsigned long flags;
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003698
Tejun Heo699ce092013-03-13 16:51:35 -07003699 /* for @wq->saved_max_active */
Lai Jiangshana357fc02013-03-25 16:57:19 -07003700 lockdep_assert_held(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003701
3702 /* fast exit for non-freezable wqs */
3703 if (!freezable && pwq->max_active == wq->saved_max_active)
3704 return;
3705
Tejun Heo3347fa02016-09-16 15:49:32 -04003706 /* this function can be called during early boot w/ irq disabled */
3707 spin_lock_irqsave(&pwq->pool->lock, flags);
Tejun Heo699ce092013-03-13 16:51:35 -07003708
Lai Jiangshan74b414e2014-05-22 19:01:16 +08003709 /*
3710 * During [un]freezing, the caller is responsible for ensuring that
3711 * this function is called at least once after @workqueue_freezing
3712 * is updated and visible.
3713 */
3714 if (!freezable || !workqueue_freezing) {
Tejun Heo699ce092013-03-13 16:51:35 -07003715 pwq->max_active = wq->saved_max_active;
3716
3717 while (!list_empty(&pwq->delayed_works) &&
3718 pwq->nr_active < pwq->max_active)
3719 pwq_activate_first_delayed(pwq);
Lai Jiangshan951a0782013-03-20 10:52:30 -07003720
3721 /*
3722 * Need to kick a worker after thawed or an unbound wq's
3723 * max_active is bumped. It's a slow path. Do it always.
3724 */
3725 wake_up_worker(pwq->pool);
Tejun Heo699ce092013-03-13 16:51:35 -07003726 } else {
3727 pwq->max_active = 0;
3728 }
3729
Tejun Heo3347fa02016-09-16 15:49:32 -04003730 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003731}
3732
Tejun Heoe50aba92013-04-01 11:23:35 -07003733/* initialize newly alloced @pwq which is associated with @wq and @pool */
Tejun Heof147f292013-04-01 11:23:35 -07003734static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3735 struct worker_pool *pool)
Tejun Heod2c1d402013-03-12 11:30:04 -07003736{
3737 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3738
Tejun Heoe50aba92013-04-01 11:23:35 -07003739 memset(pwq, 0, sizeof(*pwq));
3740
Tejun Heod2c1d402013-03-12 11:30:04 -07003741 pwq->pool = pool;
3742 pwq->wq = wq;
3743 pwq->flush_color = -1;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003744 pwq->refcnt = 1;
Tejun Heod2c1d402013-03-12 11:30:04 -07003745 INIT_LIST_HEAD(&pwq->delayed_works);
Tejun Heo1befcf32013-04-01 11:23:35 -07003746 INIT_LIST_HEAD(&pwq->pwqs_node);
Tejun Heod2c1d402013-03-12 11:30:04 -07003747 INIT_LIST_HEAD(&pwq->mayday_node);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003748 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
Tejun Heof147f292013-04-01 11:23:35 -07003749}
Tejun Heod2c1d402013-03-12 11:30:04 -07003750
Tejun Heof147f292013-04-01 11:23:35 -07003751/* sync @pwq with the current state of its associated wq and link it */
Tejun Heo1befcf32013-04-01 11:23:35 -07003752static void link_pwq(struct pool_workqueue *pwq)
Tejun Heof147f292013-04-01 11:23:35 -07003753{
3754 struct workqueue_struct *wq = pwq->wq;
3755
3756 lockdep_assert_held(&wq->mutex);
Tejun Heo75ccf592013-03-12 11:30:04 -07003757
Tejun Heo1befcf32013-04-01 11:23:35 -07003758 /* may be called multiple times, ignore if already linked */
3759 if (!list_empty(&pwq->pwqs_node))
3760 return;
3761
Lai Jiangshan29b1cb42014-07-22 13:04:27 +08003762 /* set the matching work_color */
Tejun Heo75ccf592013-03-12 11:30:04 -07003763 pwq->work_color = wq->work_color;
Tejun Heo983ca252013-03-13 16:51:35 -07003764
3765 /* sync max_active to the current setting */
3766 pwq_adjust_max_active(pwq);
3767
3768 /* link in @pwq */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003769 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
Tejun Heof147f292013-04-01 11:23:35 -07003770}
Lai Jiangshana357fc02013-03-25 16:57:19 -07003771
Tejun Heof147f292013-04-01 11:23:35 -07003772/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3773static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3774 const struct workqueue_attrs *attrs)
3775{
3776 struct worker_pool *pool;
3777 struct pool_workqueue *pwq;
3778
3779 lockdep_assert_held(&wq_pool_mutex);
3780
3781 pool = get_unbound_pool(attrs);
3782 if (!pool)
3783 return NULL;
3784
Tejun Heoe50aba92013-04-01 11:23:35 -07003785 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
Tejun Heof147f292013-04-01 11:23:35 -07003786 if (!pwq) {
3787 put_unbound_pool(pool);
3788 return NULL;
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003789 }
Tejun Heo6029a912013-04-01 11:23:34 -07003790
Tejun Heof147f292013-04-01 11:23:35 -07003791 init_pwq(pwq, wq, pool);
3792 return pwq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003793}
3794
Tejun Heo4c16bd32013-04-01 11:23:36 -07003795/**
Gong Zhaogang30186c62015-05-11 11:02:47 -04003796 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
Lai Jiangshan042f7df12015-04-30 17:16:12 +08003797 * @attrs: the wq_attrs of the default pwq of the target workqueue
Tejun Heo4c16bd32013-04-01 11:23:36 -07003798 * @node: the target NUMA node
3799 * @cpu_going_down: if >= 0, the CPU to consider as offline
3800 * @cpumask: outarg, the resulting cpumask
3801 *
3802 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3803 * @cpu_going_down is >= 0, that cpu is considered offline during
Yacine Belkadid185af32013-07-31 14:59:24 -07003804 * calculation. The result is stored in @cpumask.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003805 *
3806 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3807 * enabled and @node has online CPUs requested by @attrs, the returned
3808 * cpumask is the intersection of the possible CPUs of @node and
3809 * @attrs->cpumask.
3810 *
3811 * The caller is responsible for ensuring that the cpumask of @node stays
3812 * stable.
Yacine Belkadid185af32013-07-31 14:59:24 -07003813 *
3814 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3815 * %false if equal.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003816 */
3817static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3818 int cpu_going_down, cpumask_t *cpumask)
3819{
Tejun Heod55262c2013-04-01 11:23:38 -07003820 if (!wq_numa_enabled || attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003821 goto use_dfl;
3822
3823 /* does @node have any online CPUs @attrs wants? */
3824 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3825 if (cpu_going_down >= 0)
3826 cpumask_clear_cpu(cpu_going_down, cpumask);
3827
3828 if (cpumask_empty(cpumask))
3829 goto use_dfl;
3830
3831 /* yeap, return possible CPUs in @node that @attrs wants */
3832 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
Michael Bringmann1ad0f0a2017-07-27 16:27:14 -05003833
3834 if (cpumask_empty(cpumask)) {
3835 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3836 "possible intersect\n");
3837 return false;
3838 }
3839
Tejun Heo4c16bd32013-04-01 11:23:36 -07003840 return !cpumask_equal(cpumask, attrs->cpumask);
3841
3842use_dfl:
3843 cpumask_copy(cpumask, attrs->cpumask);
3844 return false;
3845}
3846
Tejun Heo1befcf32013-04-01 11:23:35 -07003847/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3848static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3849 int node,
3850 struct pool_workqueue *pwq)
3851{
3852 struct pool_workqueue *old_pwq;
3853
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +08003854 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003855 lockdep_assert_held(&wq->mutex);
3856
3857 /* link_pwq() can handle duplicate calls */
3858 link_pwq(pwq);
3859
3860 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3861 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3862 return old_pwq;
3863}
3864
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003865/* context to store the prepared attrs & pwqs before applying */
3866struct apply_wqattrs_ctx {
3867 struct workqueue_struct *wq; /* target workqueue */
3868 struct workqueue_attrs *attrs; /* attrs to apply */
Lai Jiangshan042f7df12015-04-30 17:16:12 +08003869 struct list_head list; /* queued for batching commit */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003870 struct pool_workqueue *dfl_pwq;
3871 struct pool_workqueue *pwq_tbl[];
3872};
3873
3874/* free the resources after success or abort */
3875static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3876{
3877 if (ctx) {
3878 int node;
3879
3880 for_each_node(node)
3881 put_pwq_unlocked(ctx->pwq_tbl[node]);
3882 put_pwq_unlocked(ctx->dfl_pwq);
3883
3884 free_workqueue_attrs(ctx->attrs);
3885
3886 kfree(ctx);
3887 }
3888}
3889
3890/* allocate the attrs and pwqs for later installation */
3891static struct apply_wqattrs_ctx *
3892apply_wqattrs_prepare(struct workqueue_struct *wq,
3893 const struct workqueue_attrs *attrs)
3894{
3895 struct apply_wqattrs_ctx *ctx;
3896 struct workqueue_attrs *new_attrs, *tmp_attrs;
3897 int node;
3898
3899 lockdep_assert_held(&wq_pool_mutex);
3900
Kees Cookacafe7e2018-05-08 13:45:50 -07003901 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003902
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02003903 new_attrs = alloc_workqueue_attrs();
3904 tmp_attrs = alloc_workqueue_attrs();
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003905 if (!ctx || !new_attrs || !tmp_attrs)
3906 goto out_free;
3907
Lai Jiangshan042f7df12015-04-30 17:16:12 +08003908 /*
3909 * Calculate the attrs of the default pwq.
3910 * If the user configured cpumask doesn't overlap with the
3911 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3912 */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003913 copy_workqueue_attrs(new_attrs, attrs);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08003914 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan042f7df12015-04-30 17:16:12 +08003915 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3916 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003917
3918 /*
3919 * We may create multiple pwqs with differing cpumasks. Make a
3920 * copy of @new_attrs which will be modified and used to obtain
3921 * pools.
3922 */
3923 copy_workqueue_attrs(tmp_attrs, new_attrs);
3924
3925 /*
3926 * If something goes wrong during CPU up/down, we'll fall back to
3927 * the default pwq covering whole @attrs->cpumask. Always create
3928 * it even if we don't use it immediately.
3929 */
3930 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3931 if (!ctx->dfl_pwq)
3932 goto out_free;
3933
3934 for_each_node(node) {
Lai Jiangshan042f7df12015-04-30 17:16:12 +08003935 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003936 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3937 if (!ctx->pwq_tbl[node])
3938 goto out_free;
3939 } else {
3940 ctx->dfl_pwq->refcnt++;
3941 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3942 }
3943 }
3944
Lai Jiangshan042f7df12015-04-30 17:16:12 +08003945 /* save the user configured attrs and sanitize it. */
3946 copy_workqueue_attrs(new_attrs, attrs);
3947 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003948 ctx->attrs = new_attrs;
Lai Jiangshan042f7df12015-04-30 17:16:12 +08003949
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003950 ctx->wq = wq;
3951 free_workqueue_attrs(tmp_attrs);
3952 return ctx;
3953
3954out_free:
3955 free_workqueue_attrs(tmp_attrs);
3956 free_workqueue_attrs(new_attrs);
3957 apply_wqattrs_cleanup(ctx);
3958 return NULL;
3959}
3960
3961/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3962static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3963{
3964 int node;
3965
3966 /* all pwqs have been created successfully, let's install'em */
3967 mutex_lock(&ctx->wq->mutex);
3968
3969 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3970
3971 /* save the previous pwq and install the new one */
3972 for_each_node(node)
3973 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3974 ctx->pwq_tbl[node]);
3975
3976 /* @dfl_pwq might not have been used, ensure it's linked */
3977 link_pwq(ctx->dfl_pwq);
3978 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3979
3980 mutex_unlock(&ctx->wq->mutex);
3981}
3982
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003983static void apply_wqattrs_lock(void)
3984{
3985 /* CPUs should stay stable across pwq creations and installations */
3986 get_online_cpus();
3987 mutex_lock(&wq_pool_mutex);
3988}
3989
3990static void apply_wqattrs_unlock(void)
3991{
3992 mutex_unlock(&wq_pool_mutex);
3993 put_online_cpus();
3994}
3995
3996static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3997 const struct workqueue_attrs *attrs)
3998{
3999 struct apply_wqattrs_ctx *ctx;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004000
4001 /* only unbound workqueues can change attributes */
4002 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4003 return -EINVAL;
4004
4005 /* creating multiple pwqs breaks ordering guarantee */
Tejun Heo0a94efb2017-07-23 08:36:15 -04004006 if (!list_empty(&wq->pwqs)) {
4007 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4008 return -EINVAL;
4009
4010 wq->flags &= ~__WQ_ORDERED;
4011 }
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004012
4013 ctx = apply_wqattrs_prepare(wq, attrs);
wanghaibin62011712016-01-07 20:38:59 +08004014 if (!ctx)
4015 return -ENOMEM;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004016
4017 /* the ctx has been prepared successfully, let's commit it */
wanghaibin62011712016-01-07 20:38:59 +08004018 apply_wqattrs_commit(ctx);
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004019 apply_wqattrs_cleanup(ctx);
4020
wanghaibin62011712016-01-07 20:38:59 +08004021 return 0;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004022}
4023
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004024/**
4025 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4026 * @wq: the target workqueue
4027 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4028 *
Tejun Heo4c16bd32013-04-01 11:23:36 -07004029 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4030 * machines, this function maps a separate pwq to each NUMA node with
4031 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4032 * NUMA node it was issued on. Older pwqs are released as in-flight work
4033 * items finish. Note that a work item which repeatedly requeues itself
4034 * back-to-back will stay on its current pwq.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004035 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004036 * Performs GFP_KERNEL allocations.
4037 *
Daniel Jordan509b3202019-09-05 21:40:23 -04004038 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4039 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004040 * Return: 0 on success and -errno on failure.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004041 */
Daniel Jordan513c98d2019-09-05 21:40:22 -04004042int apply_workqueue_attrs(struct workqueue_struct *wq,
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004043 const struct workqueue_attrs *attrs)
4044{
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004045 int ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004046
Daniel Jordan509b3202019-09-05 21:40:23 -04004047 lockdep_assert_cpus_held();
4048
4049 mutex_lock(&wq_pool_mutex);
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004050 ret = apply_workqueue_attrs_locked(wq, attrs);
Daniel Jordan509b3202019-09-05 21:40:23 -04004051 mutex_unlock(&wq_pool_mutex);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08004052
Tejun Heo48621252013-04-01 11:23:31 -07004053 return ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004054}
4055
Tejun Heo4c16bd32013-04-01 11:23:36 -07004056/**
4057 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4058 * @wq: the target workqueue
4059 * @cpu: the CPU coming up or going down
4060 * @online: whether @cpu is coming up or going down
4061 *
4062 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4063 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4064 * @wq accordingly.
4065 *
4066 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4067 * falls back to @wq->dfl_pwq which may not be optimal but is always
4068 * correct.
4069 *
4070 * Note that when the last allowed CPU of a NUMA node goes offline for a
4071 * workqueue with a cpumask spanning multiple nodes, the workers which were
4072 * already executing the work items for the workqueue will lose their CPU
4073 * affinity and may execute on any CPU. This is similar to how per-cpu
4074 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4075 * affinity, it's the user's responsibility to flush the work item from
4076 * CPU_DOWN_PREPARE.
4077 */
4078static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4079 bool online)
4080{
4081 int node = cpu_to_node(cpu);
4082 int cpu_off = online ? -1 : cpu;
4083 struct pool_workqueue *old_pwq = NULL, *pwq;
4084 struct workqueue_attrs *target_attrs;
4085 cpumask_t *cpumask;
4086
4087 lockdep_assert_held(&wq_pool_mutex);
4088
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004089 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4090 wq->unbound_attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07004091 return;
4092
4093 /*
4094 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4095 * Let's use a preallocated one. The following buf is protected by
4096 * CPU hotplug exclusion.
4097 */
4098 target_attrs = wq_update_unbound_numa_attrs_buf;
4099 cpumask = target_attrs->cpumask;
4100
Tejun Heo4c16bd32013-04-01 11:23:36 -07004101 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4102 pwq = unbound_pwq_by_node(wq, node);
4103
4104 /*
4105 * Let's determine what needs to be done. If the target cpumask is
Lai Jiangshan042f7df12015-04-30 17:16:12 +08004106 * different from the default pwq's, we need to compare it to @pwq's
4107 * and create a new one if they don't match. If the target cpumask
4108 * equals the default pwq's, the default pwq should be used.
Tejun Heo4c16bd32013-04-01 11:23:36 -07004109 */
Lai Jiangshan042f7df12015-04-30 17:16:12 +08004110 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
Tejun Heo4c16bd32013-04-01 11:23:36 -07004111 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004112 return;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004113 } else {
Daeseok Youn534a3fb2014-04-18 09:08:14 +09004114 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004115 }
4116
Tejun Heo4c16bd32013-04-01 11:23:36 -07004117 /* create a new pwq */
4118 pwq = alloc_unbound_pwq(wq, target_attrs);
4119 if (!pwq) {
Fabian Frederick2d916032014-05-12 13:59:35 -04004120 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4121 wq->name);
Daeseok Youn77f300b2014-04-16 14:32:29 +09004122 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004123 }
4124
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004125 /* Install the new pwq. */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004126 mutex_lock(&wq->mutex);
4127 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4128 goto out_unlock;
4129
4130use_dfl_pwq:
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08004131 mutex_lock(&wq->mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004132 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4133 get_pwq(wq->dfl_pwq);
4134 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4135 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4136out_unlock:
4137 mutex_unlock(&wq->mutex);
4138 put_pwq_unlocked(old_pwq);
4139}
4140
Tejun Heo30cdf2492013-03-12 11:29:57 -07004141static int alloc_and_link_pwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004143 bool highpri = wq->flags & WQ_HIGHPRI;
Tejun Heo8a2b7532013-09-05 12:30:04 -04004144 int cpu, ret;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01004145
Tejun Heo30cdf2492013-03-12 11:29:57 -07004146 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo420c0dd2013-03-12 11:29:59 -07004147 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4148 if (!wq->cpu_pwqs)
Tejun Heo30cdf2492013-03-12 11:29:57 -07004149 return -ENOMEM;
4150
4151 for_each_possible_cpu(cpu) {
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004152 struct pool_workqueue *pwq =
4153 per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heo7a62c2c2013-03-12 11:30:03 -07004154 struct worker_pool *cpu_pools =
Tejun Heof02ae732013-03-12 11:30:03 -07004155 per_cpu(cpu_worker_pools, cpu);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004156
Tejun Heof147f292013-04-01 11:23:35 -07004157 init_pwq(pwq, wq, &cpu_pools[highpri]);
4158
4159 mutex_lock(&wq->mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07004160 link_pwq(pwq);
Tejun Heof147f292013-04-01 11:23:35 -07004161 mutex_unlock(&wq->mutex);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004162 }
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07004163 return 0;
Daniel Jordan509b3202019-09-05 21:40:23 -04004164 }
4165
4166 get_online_cpus();
4167 if (wq->flags & __WQ_ORDERED) {
Tejun Heo8a2b7532013-09-05 12:30:04 -04004168 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4169 /* there should only be single pwq for ordering guarantee */
4170 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4171 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4172 "ordering guarantee broken for workqueue %s\n", wq->name);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004173 } else {
Daniel Jordan509b3202019-09-05 21:40:23 -04004174 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004175 }
Daniel Jordan509b3202019-09-05 21:40:23 -04004176 put_online_cpus();
4177
4178 return ret;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004179}
4180
Tejun Heof3421792010-07-02 10:03:51 +02004181static int wq_clamp_max_active(int max_active, unsigned int flags,
4182 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02004183{
Tejun Heof3421792010-07-02 10:03:51 +02004184 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4185
4186 if (max_active < 1 || max_active > lim)
Valentin Ilie044c7822012-08-19 00:52:42 +03004187 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4188 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02004189
Tejun Heof3421792010-07-02 10:03:51 +02004190 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02004191}
4192
Tejun Heo983c7512018-01-08 05:38:32 -08004193/*
4194 * Workqueues which may be used during memory reclaim should have a rescuer
4195 * to guarantee forward progress.
4196 */
4197static int init_rescuer(struct workqueue_struct *wq)
4198{
4199 struct worker *rescuer;
Tejun Heo983c7512018-01-08 05:38:32 -08004200
4201 if (!(wq->flags & WQ_MEM_RECLAIM))
4202 return 0;
4203
4204 rescuer = alloc_worker(NUMA_NO_NODE);
4205 if (!rescuer)
4206 return -ENOMEM;
4207
4208 rescuer->rescue_wq = wq;
4209 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
Sean Fuf187b692020-04-29 12:04:13 +08004210 if (IS_ERR(rescuer->task)) {
Tejun Heo983c7512018-01-08 05:38:32 -08004211 kfree(rescuer);
Sean Fuf187b692020-04-29 12:04:13 +08004212 return PTR_ERR(rescuer->task);
Tejun Heo983c7512018-01-08 05:38:32 -08004213 }
4214
4215 wq->rescuer = rescuer;
4216 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4217 wake_up_process(rescuer->task);
4218
4219 return 0;
4220}
4221
Mathieu Malaterrea2775bb2019-03-12 21:21:26 +01004222__printf(1, 4)
Bart Van Assche669de8b2019-02-14 15:00:54 -08004223struct workqueue_struct *alloc_workqueue(const char *fmt,
4224 unsigned int flags,
4225 int max_active, ...)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004226{
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004227 size_t tbl_size = 0;
Tejun Heoecf68812013-04-01 11:23:34 -07004228 va_list args;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004229 struct workqueue_struct *wq;
Tejun Heo49e3cf42013-03-12 11:29:58 -07004230 struct pool_workqueue *pwq;
Tejun Heob196be82012-01-10 15:11:35 -08004231
Tejun Heo5c0338c2017-07-18 18:41:52 -04004232 /*
4233 * Unbound && max_active == 1 used to imply ordered, which is no
4234 * longer the case on NUMA machines due to per-node pools. While
4235 * alloc_ordered_workqueue() is the right way to create an ordered
4236 * workqueue, keep the previous behavior to avoid subtle breakages
4237 * on NUMA.
4238 */
4239 if ((flags & WQ_UNBOUND) && max_active == 1)
4240 flags |= __WQ_ORDERED;
4241
Viresh Kumarcee22a12013-04-08 16:45:40 +05304242 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4243 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4244 flags |= WQ_UNBOUND;
4245
Tejun Heoecf68812013-04-01 11:23:34 -07004246 /* allocate wq and format name */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004247 if (flags & WQ_UNBOUND)
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08004248 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004249
4250 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
Tejun Heob196be82012-01-10 15:11:35 -08004251 if (!wq)
Tejun Heod2c1d402013-03-12 11:30:04 -07004252 return NULL;
Tejun Heob196be82012-01-10 15:11:35 -08004253
Tejun Heo6029a912013-04-01 11:23:34 -07004254 if (flags & WQ_UNBOUND) {
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02004255 wq->unbound_attrs = alloc_workqueue_attrs();
Tejun Heo6029a912013-04-01 11:23:34 -07004256 if (!wq->unbound_attrs)
4257 goto err_free_wq;
4258 }
4259
Bart Van Assche669de8b2019-02-14 15:00:54 -08004260 va_start(args, max_active);
Tejun Heoecf68812013-04-01 11:23:34 -07004261 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
Tejun Heob196be82012-01-10 15:11:35 -08004262 va_end(args);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004263
Tejun Heod320c032010-06-29 10:07:14 +02004264 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heob196be82012-01-10 15:11:35 -08004265 max_active = wq_clamp_max_active(max_active, flags, wq->name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004266
Tejun Heob196be82012-01-10 15:11:35 -08004267 /* init wq */
Tejun Heo97e37d72010-06-29 10:07:10 +02004268 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004269 wq->saved_max_active = max_active;
Lai Jiangshan3c25a552013-03-25 16:57:17 -07004270 mutex_init(&wq->mutex);
Tejun Heo112202d2013-02-13 19:29:12 -08004271 atomic_set(&wq->nr_pwqs_to_flush, 0);
Tejun Heo30cdf2492013-03-12 11:29:57 -07004272 INIT_LIST_HEAD(&wq->pwqs);
Tejun Heo73f53c42010-06-29 10:07:11 +02004273 INIT_LIST_HEAD(&wq->flusher_queue);
4274 INIT_LIST_HEAD(&wq->flusher_overflow);
Tejun Heo493a1722013-03-12 11:29:59 -07004275 INIT_LIST_HEAD(&wq->maydays);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004276
Bart Van Assche669de8b2019-02-14 15:00:54 -08004277 wq_init_lockdep(wq);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07004278 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004279
Tejun Heo30cdf2492013-03-12 11:29:57 -07004280 if (alloc_and_link_pwqs(wq) < 0)
Bart Van Assche82efcab2019-03-11 16:02:55 -07004281 goto err_unreg_lockdep;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004282
Tejun Heo40c17f72018-01-08 05:38:37 -08004283 if (wq_online && init_rescuer(wq) < 0)
Tejun Heo983c7512018-01-08 05:38:32 -08004284 goto err_destroy;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004285
Tejun Heo226223a2013-03-12 11:30:05 -07004286 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4287 goto err_destroy;
4288
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004289 /*
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004290 * wq_pool_mutex protects global freeze state and workqueues list.
4291 * Grab it, adjust max_active and add the new @wq to workqueues
4292 * list.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004293 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004294 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004295
Lai Jiangshana357fc02013-03-25 16:57:19 -07004296 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004297 for_each_pwq(pwq, wq)
4298 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004299 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004300
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004301 list_add_tail_rcu(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004302
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004303 mutex_unlock(&wq_pool_mutex);
Tejun Heo15376632010-06-29 10:07:11 +02004304
Oleg Nesterov3af244332007-05-09 02:34:09 -07004305 return wq;
Tejun Heod2c1d402013-03-12 11:30:04 -07004306
Bart Van Assche82efcab2019-03-11 16:02:55 -07004307err_unreg_lockdep:
Bart Van Assche009bb422019-03-03 14:00:46 -08004308 wq_unregister_lockdep(wq);
4309 wq_free_lockdep(wq);
Bart Van Assche82efcab2019-03-11 16:02:55 -07004310err_free_wq:
Tejun Heo6029a912013-04-01 11:23:34 -07004311 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heod2c1d402013-03-12 11:30:04 -07004312 kfree(wq);
4313 return NULL;
4314err_destroy:
4315 destroy_workqueue(wq);
Tejun Heo4690c4a2010-06-29 10:07:10 +02004316 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004317}
Bart Van Assche669de8b2019-02-14 15:00:54 -08004318EXPORT_SYMBOL_GPL(alloc_workqueue);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004319
Tejun Heoc29eb852019-09-23 11:08:58 -07004320static bool pwq_busy(struct pool_workqueue *pwq)
4321{
4322 int i;
4323
4324 for (i = 0; i < WORK_NR_COLORS; i++)
4325 if (pwq->nr_in_flight[i])
4326 return true;
4327
4328 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4329 return true;
4330 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4331 return true;
4332
4333 return false;
4334}
4335
Oleg Nesterov3af244332007-05-09 02:34:09 -07004336/**
4337 * destroy_workqueue - safely terminate a workqueue
4338 * @wq: target workqueue
4339 *
4340 * Safely destroy a workqueue. All work currently pending will be done first.
4341 */
4342void destroy_workqueue(struct workqueue_struct *wq)
4343{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004344 struct pool_workqueue *pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004345 int node;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004346
Tejun Heodef98c82019-09-18 18:43:40 -07004347 /*
4348 * Remove it from sysfs first so that sanity check failure doesn't
4349 * lead to sysfs name conflicts.
4350 */
4351 workqueue_sysfs_unregister(wq);
4352
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02004353 /* drain it before proceeding with destruction */
4354 drain_workqueue(wq);
Tejun Heoc8efcc22010-12-20 19:32:04 +01004355
Tejun Heodef98c82019-09-18 18:43:40 -07004356 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4357 if (wq->rescuer) {
4358 struct worker *rescuer = wq->rescuer;
4359
4360 /* this prevents new queueing */
4361 spin_lock_irq(&wq_mayday_lock);
4362 wq->rescuer = NULL;
4363 spin_unlock_irq(&wq_mayday_lock);
4364
4365 /* rescuer will empty maydays list before exiting */
4366 kthread_stop(rescuer->task);
Tejun Heo8efe1222019-09-20 13:39:57 -07004367 kfree(rescuer);
Tejun Heodef98c82019-09-18 18:43:40 -07004368 }
4369
Tejun Heoc29eb852019-09-23 11:08:58 -07004370 /*
4371 * Sanity checks - grab all the locks so that we wait for all
4372 * in-flight operations which may do put_pwq().
4373 */
4374 mutex_lock(&wq_pool_mutex);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004375 mutex_lock(&wq->mutex);
Tejun Heo49e3cf42013-03-12 11:29:58 -07004376 for_each_pwq(pwq, wq) {
Tejun Heoc29eb852019-09-23 11:08:58 -07004377 spin_lock_irq(&pwq->pool->lock);
4378 if (WARN_ON(pwq_busy(pwq))) {
Kefeng Wang1d9a6152019-11-28 08:47:49 +08004379 pr_warn("%s: %s has the following busy pwq\n",
4380 __func__, wq->name);
Tejun Heoc29eb852019-09-23 11:08:58 -07004381 show_pwq(pwq);
4382 spin_unlock_irq(&pwq->pool->lock);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004383 mutex_unlock(&wq->mutex);
Tejun Heoc29eb852019-09-23 11:08:58 -07004384 mutex_unlock(&wq_pool_mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004385 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004386 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004387 }
Tejun Heoc29eb852019-09-23 11:08:58 -07004388 spin_unlock_irq(&pwq->pool->lock);
Tejun Heo6183c002013-03-12 11:29:57 -07004389 }
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004390 mutex_unlock(&wq->mutex);
Tejun Heoc29eb852019-09-23 11:08:58 -07004391 mutex_unlock(&wq_pool_mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004392
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004393 /*
4394 * wq list is used to freeze wq, remove from list after
4395 * flushing is complete in case freeze races us.
4396 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004397 mutex_lock(&wq_pool_mutex);
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004398 list_del_rcu(&wq->list);
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004399 mutex_unlock(&wq_pool_mutex);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004400
Tejun Heo8864b4e2013-03-12 11:30:04 -07004401 if (!(wq->flags & WQ_UNBOUND)) {
Bart Van Assche669de8b2019-02-14 15:00:54 -08004402 wq_unregister_lockdep(wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07004403 /*
4404 * The base ref is never dropped on per-cpu pwqs. Directly
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004405 * schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004406 */
Paul E. McKenney25b00772018-11-06 19:18:45 -08004407 call_rcu(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07004408 } else {
4409 /*
4410 * We're the sole accessor of @wq at this point. Directly
Tejun Heo4c16bd32013-04-01 11:23:36 -07004411 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4412 * @wq will be freed when the last pwq is released.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004413 */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004414 for_each_node(node) {
4415 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4416 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4417 put_pwq_unlocked(pwq);
4418 }
4419
4420 /*
4421 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4422 * put. Don't access it afterwards.
4423 */
4424 pwq = wq->dfl_pwq;
4425 wq->dfl_pwq = NULL;
Tejun Heodce90d42013-04-01 11:23:35 -07004426 put_pwq_unlocked(pwq);
Tejun Heo29c91e92013-03-12 11:30:03 -07004427 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004428}
4429EXPORT_SYMBOL_GPL(destroy_workqueue);
4430
Tejun Heodcd989c2010-06-29 10:07:14 +02004431/**
4432 * workqueue_set_max_active - adjust max_active of a workqueue
4433 * @wq: target workqueue
4434 * @max_active: new max_active value.
4435 *
4436 * Set max_active of @wq to @max_active.
4437 *
4438 * CONTEXT:
4439 * Don't call from IRQ context.
4440 */
4441void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4442{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004443 struct pool_workqueue *pwq;
Tejun Heodcd989c2010-06-29 10:07:14 +02004444
Tejun Heo8719dce2013-03-12 11:30:04 -07004445 /* disallow meddling with max_active for ordered workqueues */
Tejun Heo0a94efb2017-07-23 08:36:15 -04004446 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Tejun Heo8719dce2013-03-12 11:30:04 -07004447 return;
4448
Tejun Heof3421792010-07-02 10:03:51 +02004449 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02004450
Lai Jiangshana357fc02013-03-25 16:57:19 -07004451 mutex_lock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004452
Tejun Heo0a94efb2017-07-23 08:36:15 -04004453 wq->flags &= ~__WQ_ORDERED;
Tejun Heodcd989c2010-06-29 10:07:14 +02004454 wq->saved_max_active = max_active;
4455
Tejun Heo699ce092013-03-13 16:51:35 -07004456 for_each_pwq(pwq, wq)
4457 pwq_adjust_max_active(pwq);
Tejun Heodcd989c2010-06-29 10:07:14 +02004458
Lai Jiangshana357fc02013-03-25 16:57:19 -07004459 mutex_unlock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004460}
4461EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4462
4463/**
Lukas Wunner27d4ee02018-02-11 10:38:28 +01004464 * current_work - retrieve %current task's work struct
4465 *
4466 * Determine if %current task is a workqueue worker and what it's working on.
4467 * Useful to find out the context that the %current task is running in.
4468 *
4469 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4470 */
4471struct work_struct *current_work(void)
4472{
4473 struct worker *worker = current_wq_worker();
4474
4475 return worker ? worker->current_work : NULL;
4476}
4477EXPORT_SYMBOL(current_work);
4478
4479/**
Tejun Heoe62676162013-03-12 17:41:37 -07004480 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4481 *
4482 * Determine whether %current is a workqueue rescuer. Can be used from
4483 * work functions to determine whether it's being run off the rescuer task.
Yacine Belkadid185af32013-07-31 14:59:24 -07004484 *
4485 * Return: %true if %current is a workqueue rescuer. %false otherwise.
Tejun Heoe62676162013-03-12 17:41:37 -07004486 */
4487bool current_is_workqueue_rescuer(void)
4488{
4489 struct worker *worker = current_wq_worker();
4490
Lai Jiangshan6a092df2013-03-20 03:28:03 +08004491 return worker && worker->rescue_wq;
Tejun Heoe62676162013-03-12 17:41:37 -07004492}
4493
4494/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004495 * workqueue_congested - test whether a workqueue is congested
4496 * @cpu: CPU in question
4497 * @wq: target workqueue
4498 *
4499 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4500 * no synchronization around this function and the test result is
4501 * unreliable and only useful as advisory hints or for debugging.
4502 *
Tejun Heod3251852013-05-10 11:10:17 -07004503 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4504 * Note that both per-cpu and unbound workqueues may be associated with
4505 * multiple pool_workqueues which have separate congested states. A
4506 * workqueue being congested on one CPU doesn't mean the workqueue is also
4507 * contested on other CPUs / NUMA nodes.
4508 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004509 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004510 * %true if congested, %false otherwise.
4511 */
Tejun Heod84ff052013-03-12 11:29:59 -07004512bool workqueue_congested(int cpu, struct workqueue_struct *wq)
Tejun Heodcd989c2010-06-29 10:07:14 +02004513{
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004514 struct pool_workqueue *pwq;
Tejun Heo76af4d92013-03-12 11:30:00 -07004515 bool ret;
4516
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004517 rcu_read_lock();
4518 preempt_disable();
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004519
Tejun Heod3251852013-05-10 11:10:17 -07004520 if (cpu == WORK_CPU_UNBOUND)
4521 cpu = smp_processor_id();
4522
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004523 if (!(wq->flags & WQ_UNBOUND))
4524 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4525 else
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004526 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodcd989c2010-06-29 10:07:14 +02004527
Tejun Heo76af4d92013-03-12 11:30:00 -07004528 ret = !list_empty(&pwq->delayed_works);
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004529 preempt_enable();
4530 rcu_read_unlock();
Tejun Heo76af4d92013-03-12 11:30:00 -07004531
4532 return ret;
Tejun Heodcd989c2010-06-29 10:07:14 +02004533}
4534EXPORT_SYMBOL_GPL(workqueue_congested);
4535
4536/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004537 * work_busy - test whether a work is currently pending or running
4538 * @work: the work to be tested
4539 *
4540 * Test whether @work is currently pending or running. There is no
4541 * synchronization around this function and the test result is
4542 * unreliable and only useful as advisory hints or for debugging.
Tejun Heodcd989c2010-06-29 10:07:14 +02004543 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004544 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004545 * OR'd bitmask of WORK_BUSY_* bits.
4546 */
4547unsigned int work_busy(struct work_struct *work)
4548{
Tejun Heofa1b54e2013-03-12 11:30:00 -07004549 struct worker_pool *pool;
Tejun Heodcd989c2010-06-29 10:07:14 +02004550 unsigned long flags;
4551 unsigned int ret = 0;
4552
Tejun Heodcd989c2010-06-29 10:07:14 +02004553 if (work_pending(work))
4554 ret |= WORK_BUSY_PENDING;
Tejun Heodcd989c2010-06-29 10:07:14 +02004555
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004556 rcu_read_lock();
Tejun Heofa1b54e2013-03-12 11:30:00 -07004557 pool = get_work_pool(work);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004558 if (pool) {
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004559 spin_lock_irqsave(&pool->lock, flags);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004560 if (find_worker_executing_work(pool, work))
4561 ret |= WORK_BUSY_RUNNING;
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004562 spin_unlock_irqrestore(&pool->lock, flags);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004563 }
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004564 rcu_read_unlock();
Tejun Heodcd989c2010-06-29 10:07:14 +02004565
4566 return ret;
4567}
4568EXPORT_SYMBOL_GPL(work_busy);
4569
Tejun Heo3d1cb202013-04-30 15:27:22 -07004570/**
4571 * set_worker_desc - set description for the current work item
4572 * @fmt: printf-style format string
4573 * @...: arguments for the format string
4574 *
4575 * This function can be called by a running work function to describe what
4576 * the work item is about. If the worker task gets dumped, this
4577 * information will be printed out together to help debugging. The
4578 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4579 */
4580void set_worker_desc(const char *fmt, ...)
4581{
4582 struct worker *worker = current_wq_worker();
4583 va_list args;
4584
4585 if (worker) {
4586 va_start(args, fmt);
4587 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4588 va_end(args);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004589 }
4590}
Steffen Maier5c750d52018-05-17 19:14:57 +02004591EXPORT_SYMBOL_GPL(set_worker_desc);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004592
4593/**
4594 * print_worker_info - print out worker information and description
4595 * @log_lvl: the log level to use when printing
4596 * @task: target task
4597 *
4598 * If @task is a worker and currently executing a work item, print out the
4599 * name of the workqueue being serviced and worker description set with
4600 * set_worker_desc() by the currently executing work item.
4601 *
4602 * This function can be safely called on any task as long as the
4603 * task_struct itself is accessible. While safe, this function isn't
4604 * synchronized and may print out mixups or garbages of limited length.
4605 */
4606void print_worker_info(const char *log_lvl, struct task_struct *task)
4607{
4608 work_func_t *fn = NULL;
4609 char name[WQ_NAME_LEN] = { };
4610 char desc[WORKER_DESC_LEN] = { };
4611 struct pool_workqueue *pwq = NULL;
4612 struct workqueue_struct *wq = NULL;
Tejun Heo3d1cb202013-04-30 15:27:22 -07004613 struct worker *worker;
4614
4615 if (!(task->flags & PF_WQ_WORKER))
4616 return;
4617
4618 /*
4619 * This function is called without any synchronization and @task
4620 * could be in any state. Be careful with dereferences.
4621 */
Petr Mladeke7005912016-10-11 13:55:17 -07004622 worker = kthread_probe_data(task);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004623
4624 /*
Tejun Heo8bf89592018-05-18 08:47:13 -07004625 * Carefully copy the associated workqueue's workfn, name and desc.
4626 * Keep the original last '\0' in case the original is garbage.
Tejun Heo3d1cb202013-04-30 15:27:22 -07004627 */
4628 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4629 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4630 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4631 probe_kernel_read(name, wq->name, sizeof(name) - 1);
Tejun Heo8bf89592018-05-18 08:47:13 -07004632 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004633
4634 if (fn || name[0] || desc[0]) {
Sakari Ailusd75f7732019-03-25 21:32:28 +02004635 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
Tejun Heo8bf89592018-05-18 08:47:13 -07004636 if (strcmp(name, desc))
Tejun Heo3d1cb202013-04-30 15:27:22 -07004637 pr_cont(" (%s)", desc);
4638 pr_cont("\n");
4639 }
4640}
4641
Tejun Heo3494fc32015-03-09 09:22:28 -04004642static void pr_cont_pool_info(struct worker_pool *pool)
4643{
4644 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4645 if (pool->node != NUMA_NO_NODE)
4646 pr_cont(" node=%d", pool->node);
4647 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4648}
4649
4650static void pr_cont_work(bool comma, struct work_struct *work)
4651{
4652 if (work->func == wq_barrier_func) {
4653 struct wq_barrier *barr;
4654
4655 barr = container_of(work, struct wq_barrier, work);
4656
4657 pr_cont("%s BAR(%d)", comma ? "," : "",
4658 task_pid_nr(barr->task));
4659 } else {
Sakari Ailusd75f7732019-03-25 21:32:28 +02004660 pr_cont("%s %ps", comma ? "," : "", work->func);
Tejun Heo3494fc32015-03-09 09:22:28 -04004661 }
4662}
4663
4664static void show_pwq(struct pool_workqueue *pwq)
4665{
4666 struct worker_pool *pool = pwq->pool;
4667 struct work_struct *work;
4668 struct worker *worker;
4669 bool has_in_flight = false, has_pending = false;
4670 int bkt;
4671
4672 pr_info(" pwq %d:", pool->id);
4673 pr_cont_pool_info(pool);
4674
Tejun Heoe66b39a2019-09-25 06:59:15 -07004675 pr_cont(" active=%d/%d refcnt=%d%s\n",
4676 pwq->nr_active, pwq->max_active, pwq->refcnt,
Tejun Heo3494fc32015-03-09 09:22:28 -04004677 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4678
4679 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4680 if (worker->current_pwq == pwq) {
4681 has_in_flight = true;
4682 break;
4683 }
4684 }
4685 if (has_in_flight) {
4686 bool comma = false;
4687
4688 pr_info(" in-flight:");
4689 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4690 if (worker->current_pwq != pwq)
4691 continue;
4692
Sakari Ailusd75f7732019-03-25 21:32:28 +02004693 pr_cont("%s %d%s:%ps", comma ? "," : "",
Tejun Heo3494fc32015-03-09 09:22:28 -04004694 task_pid_nr(worker->task),
Tejun Heo30ae2fc2019-09-20 14:09:14 -07004695 worker->rescue_wq ? "(RESCUER)" : "",
Tejun Heo3494fc32015-03-09 09:22:28 -04004696 worker->current_func);
4697 list_for_each_entry(work, &worker->scheduled, entry)
4698 pr_cont_work(false, work);
4699 comma = true;
4700 }
4701 pr_cont("\n");
4702 }
4703
4704 list_for_each_entry(work, &pool->worklist, entry) {
4705 if (get_work_pwq(work) == pwq) {
4706 has_pending = true;
4707 break;
4708 }
4709 }
4710 if (has_pending) {
4711 bool comma = false;
4712
4713 pr_info(" pending:");
4714 list_for_each_entry(work, &pool->worklist, entry) {
4715 if (get_work_pwq(work) != pwq)
4716 continue;
4717
4718 pr_cont_work(comma, work);
4719 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4720 }
4721 pr_cont("\n");
4722 }
4723
4724 if (!list_empty(&pwq->delayed_works)) {
4725 bool comma = false;
4726
4727 pr_info(" delayed:");
4728 list_for_each_entry(work, &pwq->delayed_works, entry) {
4729 pr_cont_work(comma, work);
4730 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4731 }
4732 pr_cont("\n");
4733 }
4734}
4735
4736/**
4737 * show_workqueue_state - dump workqueue state
4738 *
Roger Lu7b776af2016-07-01 11:05:02 +08004739 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4740 * all busy workqueues and pools.
Tejun Heo3494fc32015-03-09 09:22:28 -04004741 */
4742void show_workqueue_state(void)
4743{
4744 struct workqueue_struct *wq;
4745 struct worker_pool *pool;
4746 unsigned long flags;
4747 int pi;
4748
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004749 rcu_read_lock();
Tejun Heo3494fc32015-03-09 09:22:28 -04004750
4751 pr_info("Showing busy workqueues and worker pools:\n");
4752
4753 list_for_each_entry_rcu(wq, &workqueues, list) {
4754 struct pool_workqueue *pwq;
4755 bool idle = true;
4756
4757 for_each_pwq(pwq, wq) {
4758 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4759 idle = false;
4760 break;
4761 }
4762 }
4763 if (idle)
4764 continue;
4765
4766 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4767
4768 for_each_pwq(pwq, wq) {
4769 spin_lock_irqsave(&pwq->pool->lock, flags);
4770 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4771 show_pwq(pwq);
4772 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Sergey Senozhatsky62635ea2018-01-11 09:53:35 +09004773 /*
4774 * We could be printing a lot from atomic context, e.g.
4775 * sysrq-t -> show_workqueue_state(). Avoid triggering
4776 * hard lockup.
4777 */
4778 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004779 }
4780 }
4781
4782 for_each_pool(pool, pi) {
4783 struct worker *worker;
4784 bool first = true;
4785
4786 spin_lock_irqsave(&pool->lock, flags);
4787 if (pool->nr_workers == pool->nr_idle)
4788 goto next_pool;
4789
4790 pr_info("pool %d:", pool->id);
4791 pr_cont_pool_info(pool);
Tejun Heo82607adc2015-12-08 11:28:04 -05004792 pr_cont(" hung=%us workers=%d",
4793 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4794 pool->nr_workers);
Tejun Heo3494fc32015-03-09 09:22:28 -04004795 if (pool->manager)
4796 pr_cont(" manager: %d",
4797 task_pid_nr(pool->manager->task));
4798 list_for_each_entry(worker, &pool->idle_list, entry) {
4799 pr_cont(" %s%d", first ? "idle: " : "",
4800 task_pid_nr(worker->task));
4801 first = false;
4802 }
4803 pr_cont("\n");
4804 next_pool:
4805 spin_unlock_irqrestore(&pool->lock, flags);
Sergey Senozhatsky62635ea2018-01-11 09:53:35 +09004806 /*
4807 * We could be printing a lot from atomic context, e.g.
4808 * sysrq-t -> show_workqueue_state(). Avoid triggering
4809 * hard lockup.
4810 */
4811 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004812 }
4813
Thomas Gleixner24acfb72019-03-13 17:55:47 +01004814 rcu_read_unlock();
Tejun Heo3494fc32015-03-09 09:22:28 -04004815}
4816
Tejun Heo6b598082018-05-18 08:47:13 -07004817/* used to show worker information through /proc/PID/{comm,stat,status} */
4818void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4819{
Tejun Heo6b598082018-05-18 08:47:13 -07004820 int off;
4821
4822 /* always show the actual comm */
4823 off = strscpy(buf, task->comm, size);
4824 if (off < 0)
4825 return;
4826
Tejun Heo197f6ac2018-05-21 08:04:35 -07004827 /* stabilize PF_WQ_WORKER and worker pool association */
Tejun Heo6b598082018-05-18 08:47:13 -07004828 mutex_lock(&wq_pool_attach_mutex);
4829
Tejun Heo197f6ac2018-05-21 08:04:35 -07004830 if (task->flags & PF_WQ_WORKER) {
4831 struct worker *worker = kthread_data(task);
4832 struct worker_pool *pool = worker->pool;
Tejun Heo6b598082018-05-18 08:47:13 -07004833
Tejun Heo197f6ac2018-05-21 08:04:35 -07004834 if (pool) {
4835 spin_lock_irq(&pool->lock);
4836 /*
4837 * ->desc tracks information (wq name or
4838 * set_worker_desc()) for the latest execution. If
4839 * current, prepend '+', otherwise '-'.
4840 */
4841 if (worker->desc[0] != '\0') {
4842 if (worker->current_work)
4843 scnprintf(buf + off, size - off, "+%s",
4844 worker->desc);
4845 else
4846 scnprintf(buf + off, size - off, "-%s",
4847 worker->desc);
4848 }
4849 spin_unlock_irq(&pool->lock);
Tejun Heo6b598082018-05-18 08:47:13 -07004850 }
Tejun Heo6b598082018-05-18 08:47:13 -07004851 }
4852
4853 mutex_unlock(&wq_pool_attach_mutex);
4854}
4855
Mathieu Malaterre66448bc2018-05-22 21:47:32 +02004856#ifdef CONFIG_SMP
4857
Tejun Heodb7bccf2010-06-29 10:07:12 +02004858/*
4859 * CPU hotplug.
4860 *
Tejun Heoe22bee72010-06-29 10:07:14 +02004861 * There are two challenges in supporting CPU hotplug. Firstly, there
Tejun Heo112202d2013-02-13 19:29:12 -08004862 * are a lot of assumptions on strong associations among work, pwq and
Tejun Heo706026c2013-01-24 11:01:34 -08004863 * pool which make migrating pending and scheduled works very
Tejun Heoe22bee72010-06-29 10:07:14 +02004864 * difficult to implement without impacting hot paths. Secondly,
Tejun Heo94cf58b2013-01-24 11:01:33 -08004865 * worker pools serve mix of short, long and very long running works making
Tejun Heoe22bee72010-06-29 10:07:14 +02004866 * blocked draining impractical.
4867 *
Tejun Heo24647572013-01-24 11:01:33 -08004868 * This is solved by allowing the pools to be disassociated from the CPU
Tejun Heo628c78e2012-07-17 12:39:27 -07004869 * running as an unbound one and allowing it to be reattached later if the
4870 * cpu comes back online.
Tejun Heodb7bccf2010-06-29 10:07:12 +02004871 */
4872
Lai Jiangshane8b3f8d2017-12-01 22:20:36 +08004873static void unbind_workers(int cpu)
Tejun Heodb7bccf2010-06-29 10:07:12 +02004874{
Tejun Heo4ce62e92012-07-13 22:16:44 -07004875 struct worker_pool *pool;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004876 struct worker *worker;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004877
Tejun Heof02ae732013-03-12 11:30:03 -07004878 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo1258fae2018-05-18 08:47:13 -07004879 mutex_lock(&wq_pool_attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004880 spin_lock_irq(&pool->lock);
4881
4882 /*
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004883 * We've blocked all attach/detach operations. Make all workers
Tejun Heo94cf58b2013-01-24 11:01:33 -08004884 * unbound and set DISASSOCIATED. Before this, all workers
4885 * except for the ones which are still executing works from
4886 * before the last CPU down must be on the cpu. After
4887 * this, they may become diasporas.
4888 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004889 for_each_pool_worker(worker, pool)
Tejun Heoc9e7cf22013-01-24 11:01:33 -08004890 worker->flags |= WORKER_UNBOUND;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004891
Tejun Heo24647572013-01-24 11:01:33 -08004892 pool->flags |= POOL_DISASSOCIATED;
Tejun Heof2d5a0e2012-07-17 12:39:26 -07004893
Tejun Heo94cf58b2013-01-24 11:01:33 -08004894 spin_unlock_irq(&pool->lock);
Tejun Heo1258fae2018-05-18 08:47:13 -07004895 mutex_unlock(&wq_pool_attach_mutex);
Tejun Heoe22bee72010-06-29 10:07:14 +02004896
Lai Jiangshaneb283422013-03-08 15:18:28 -08004897 /*
4898 * Call schedule() so that we cross rq->lock and thus can
4899 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4900 * This is necessary as scheduler callbacks may be invoked
4901 * from other cpus.
4902 */
4903 schedule();
Tejun Heo628c78e2012-07-17 12:39:27 -07004904
Lai Jiangshaneb283422013-03-08 15:18:28 -08004905 /*
4906 * Sched callbacks are disabled now. Zap nr_running.
4907 * After this, nr_running stays zero and need_more_worker()
4908 * and keep_working() are always true as long as the
4909 * worklist is not empty. This pool now behaves as an
4910 * unbound (in terms of concurrency management) pool which
4911 * are served by workers tied to the pool.
4912 */
Tejun Heoe19e3972013-01-24 11:39:44 -08004913 atomic_set(&pool->nr_running, 0);
Lai Jiangshaneb283422013-03-08 15:18:28 -08004914
4915 /*
4916 * With concurrency management just turned off, a busy
4917 * worker blocking could lead to lengthy stalls. Kick off
4918 * unbound chain execution of currently pending work items.
4919 */
4920 spin_lock_irq(&pool->lock);
4921 wake_up_worker(pool);
4922 spin_unlock_irq(&pool->lock);
4923 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004924}
4925
Tejun Heobd7c0892013-03-19 13:45:21 -07004926/**
4927 * rebind_workers - rebind all workers of a pool to the associated CPU
4928 * @pool: pool of interest
4929 *
Tejun Heoa9ab7752013-03-19 13:45:21 -07004930 * @pool->cpu is coming online. Rebind all workers to the CPU.
Tejun Heobd7c0892013-03-19 13:45:21 -07004931 */
4932static void rebind_workers(struct worker_pool *pool)
4933{
Tejun Heoa9ab7752013-03-19 13:45:21 -07004934 struct worker *worker;
Tejun Heobd7c0892013-03-19 13:45:21 -07004935
Tejun Heo1258fae2018-05-18 08:47:13 -07004936 lockdep_assert_held(&wq_pool_attach_mutex);
Tejun Heobd7c0892013-03-19 13:45:21 -07004937
Tejun Heoa9ab7752013-03-19 13:45:21 -07004938 /*
4939 * Restore CPU affinity of all workers. As all idle workers should
4940 * be on the run-queue of the associated CPU before any local
Shailendra Verma402dd892015-05-23 10:38:14 +05304941 * wake-ups for concurrency management happen, restore CPU affinity
Tejun Heoa9ab7752013-03-19 13:45:21 -07004942 * of all workers first and then clear UNBOUND. As we're called
4943 * from CPU_ONLINE, the following shouldn't fail.
4944 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004945 for_each_pool_worker(worker, pool)
Tejun Heoa9ab7752013-03-19 13:45:21 -07004946 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4947 pool->attrs->cpumask) < 0);
4948
4949 spin_lock_irq(&pool->lock);
Wanpeng Lif7c17d22016-05-11 17:55:18 +08004950
Lai Jiangshan3de5e882014-06-03 15:33:27 +08004951 pool->flags &= ~POOL_DISASSOCIATED;
Tejun Heoa9ab7752013-03-19 13:45:21 -07004952
Lai Jiangshanda028462014-05-20 17:46:31 +08004953 for_each_pool_worker(worker, pool) {
Tejun Heoa9ab7752013-03-19 13:45:21 -07004954 unsigned int worker_flags = worker->flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004955
4956 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07004957 * A bound idle worker should actually be on the runqueue
4958 * of the associated CPU for local wake-ups targeting it to
4959 * work. Kick all idle workers so that they migrate to the
4960 * associated CPU. Doing this in the same loop as
4961 * replacing UNBOUND with REBOUND is safe as no worker will
4962 * be bound before @pool->lock is released.
Tejun Heobd7c0892013-03-19 13:45:21 -07004963 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004964 if (worker_flags & WORKER_IDLE)
4965 wake_up_process(worker->task);
4966
4967 /*
4968 * We want to clear UNBOUND but can't directly call
4969 * worker_clr_flags() or adjust nr_running. Atomically
4970 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4971 * @worker will clear REBOUND using worker_clr_flags() when
4972 * it initiates the next execution cycle thus restoring
4973 * concurrency management. Note that when or whether
4974 * @worker clears REBOUND doesn't affect correctness.
4975 *
Mark Rutlandc95491e2017-10-23 14:07:22 -07004976 * WRITE_ONCE() is necessary because @worker->flags may be
Tejun Heoa9ab7752013-03-19 13:45:21 -07004977 * tested without holding any lock in
Thomas Gleixner6d25be52019-03-13 17:55:48 +01004978 * wq_worker_running(). Without it, NOT_RUNNING test may
Tejun Heoa9ab7752013-03-19 13:45:21 -07004979 * fail incorrectly leading to premature concurrency
4980 * management operations.
4981 */
4982 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4983 worker_flags |= WORKER_REBOUND;
4984 worker_flags &= ~WORKER_UNBOUND;
Mark Rutlandc95491e2017-10-23 14:07:22 -07004985 WRITE_ONCE(worker->flags, worker_flags);
Tejun Heobd7c0892013-03-19 13:45:21 -07004986 }
4987
Tejun Heoa9ab7752013-03-19 13:45:21 -07004988 spin_unlock_irq(&pool->lock);
Tejun Heobd7c0892013-03-19 13:45:21 -07004989}
4990
Tejun Heo7dbc7252013-03-19 13:45:21 -07004991/**
4992 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4993 * @pool: unbound pool of interest
4994 * @cpu: the CPU which is coming up
4995 *
4996 * An unbound pool may end up with a cpumask which doesn't have any online
4997 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4998 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4999 * online CPU before, cpus_allowed of all its workers should be restored.
5000 */
5001static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5002{
5003 static cpumask_t cpumask;
5004 struct worker *worker;
Tejun Heo7dbc7252013-03-19 13:45:21 -07005005
Tejun Heo1258fae2018-05-18 08:47:13 -07005006 lockdep_assert_held(&wq_pool_attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005007
5008 /* is @cpu allowed for @pool? */
5009 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5010 return;
5011
Tejun Heo7dbc7252013-03-19 13:45:21 -07005012 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005013
5014 /* as we're called from CPU_ONLINE, the following shouldn't fail */
Lai Jiangshanda028462014-05-20 17:46:31 +08005015 for_each_pool_worker(worker, pool)
Peter Zijlstrad945b5e2016-06-16 14:38:42 +02005016 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005017}
5018
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005019int workqueue_prepare_cpu(unsigned int cpu)
Oleg Nesterov3af244332007-05-09 02:34:09 -07005020{
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005021 struct worker_pool *pool;
5022
5023 for_each_cpu_worker_pool(pool, cpu) {
5024 if (pool->nr_workers)
5025 continue;
5026 if (!create_worker(pool))
5027 return -ENOMEM;
5028 }
5029 return 0;
5030}
5031
5032int workqueue_online_cpu(unsigned int cpu)
5033{
Tejun Heo4ce62e92012-07-13 22:16:44 -07005034 struct worker_pool *pool;
Tejun Heo4c16bd32013-04-01 11:23:36 -07005035 struct workqueue_struct *wq;
Tejun Heo7dbc7252013-03-19 13:45:21 -07005036 int pi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005037
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005038 mutex_lock(&wq_pool_mutex);
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07005039
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005040 for_each_pool(pool, pi) {
Tejun Heo1258fae2018-05-18 08:47:13 -07005041 mutex_lock(&wq_pool_attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07005042
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005043 if (pool->cpu == cpu)
5044 rebind_workers(pool);
5045 else if (pool->cpu < 0)
5046 restore_unbound_workers_cpumask(pool, cpu);
Tejun Heo94cf58b2013-01-24 11:01:33 -08005047
Tejun Heo1258fae2018-05-18 08:47:13 -07005048 mutex_unlock(&wq_pool_attach_mutex);
Tejun Heo65758202012-07-17 12:39:26 -07005049 }
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005050
5051 /* update NUMA affinity of unbound workqueues */
5052 list_for_each_entry(wq, &workqueues, list)
5053 wq_update_unbound_numa(wq, cpu, true);
5054
5055 mutex_unlock(&wq_pool_mutex);
5056 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07005057}
5058
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005059int workqueue_offline_cpu(unsigned int cpu)
Tejun Heo65758202012-07-17 12:39:26 -07005060{
Tejun Heo4c16bd32013-04-01 11:23:36 -07005061 struct workqueue_struct *wq;
Tejun Heo8db25e72012-07-17 12:39:28 -07005062
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005063 /* unbinding per-cpu workers should happen on the local CPU */
Lai Jiangshane8b3f8d2017-12-01 22:20:36 +08005064 if (WARN_ON(cpu != smp_processor_id()))
5065 return -1;
5066
5067 unbind_workers(cpu);
Tejun Heo4c16bd32013-04-01 11:23:36 -07005068
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005069 /* update NUMA affinity of unbound workqueues */
5070 mutex_lock(&wq_pool_mutex);
5071 list_for_each_entry(wq, &workqueues, list)
5072 wq_update_unbound_numa(wq, cpu, false);
5073 mutex_unlock(&wq_pool_mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07005074
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00005075 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07005076}
5077
Rusty Russell2d3854a2008-11-05 13:39:10 +11005078struct work_for_cpu {
Tejun Heoed48ece2012-09-18 12:48:43 -07005079 struct work_struct work;
Rusty Russell2d3854a2008-11-05 13:39:10 +11005080 long (*fn)(void *);
5081 void *arg;
5082 long ret;
5083};
5084
Tejun Heoed48ece2012-09-18 12:48:43 -07005085static void work_for_cpu_fn(struct work_struct *work)
Rusty Russell2d3854a2008-11-05 13:39:10 +11005086{
Tejun Heoed48ece2012-09-18 12:48:43 -07005087 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5088
Rusty Russell2d3854a2008-11-05 13:39:10 +11005089 wfc->ret = wfc->fn(wfc->arg);
5090}
5091
5092/**
Anna-Maria Gleixner22aceb32016-03-10 12:07:38 +01005093 * work_on_cpu - run a function in thread context on a particular cpu
Rusty Russell2d3854a2008-11-05 13:39:10 +11005094 * @cpu: the cpu to run on
5095 * @fn: the function to run
5096 * @arg: the function arg
5097 *
Rusty Russell31ad9082009-01-16 15:31:15 -08005098 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06005099 * The caller must not hold any locks which would prevent @fn from completing.
Yacine Belkadid185af32013-07-31 14:59:24 -07005100 *
5101 * Return: The value @fn returns.
Rusty Russell2d3854a2008-11-05 13:39:10 +11005102 */
Tejun Heod84ff052013-03-12 11:29:59 -07005103long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
Rusty Russell2d3854a2008-11-05 13:39:10 +11005104{
Tejun Heoed48ece2012-09-18 12:48:43 -07005105 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
Rusty Russell2d3854a2008-11-05 13:39:10 +11005106
Tejun Heoed48ece2012-09-18 12:48:43 -07005107 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5108 schedule_work_on(cpu, &wfc.work);
Bjorn Helgaas12997d12013-11-18 11:00:29 -07005109 flush_work(&wfc.work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05005110 destroy_work_on_stack(&wfc.work);
Rusty Russell2d3854a2008-11-05 13:39:10 +11005111 return wfc.ret;
5112}
5113EXPORT_SYMBOL_GPL(work_on_cpu);
Thomas Gleixner0e8d6a92017-04-12 22:07:28 +02005114
5115/**
5116 * work_on_cpu_safe - run a function in thread context on a particular cpu
5117 * @cpu: the cpu to run on
5118 * @fn: the function to run
5119 * @arg: the function argument
5120 *
5121 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5122 * any locks which would prevent @fn from completing.
5123 *
5124 * Return: The value @fn returns.
5125 */
5126long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5127{
5128 long ret = -ENODEV;
5129
5130 get_online_cpus();
5131 if (cpu_online(cpu))
5132 ret = work_on_cpu(cpu, fn, arg);
5133 put_online_cpus();
5134 return ret;
5135}
5136EXPORT_SYMBOL_GPL(work_on_cpu_safe);
Rusty Russell2d3854a2008-11-05 13:39:10 +11005137#endif /* CONFIG_SMP */
5138
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005139#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10305140
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005141/**
5142 * freeze_workqueues_begin - begin freezing workqueues
5143 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01005144 * Start freezing workqueues. After this function returns, all freezable
Tejun Heoc5aa87b2013-03-13 16:51:36 -07005145 * workqueues will queue new works to their delayed_works list instead of
Tejun Heo706026c2013-01-24 11:01:34 -08005146 * pool->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005147 *
5148 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07005149 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005150 */
5151void freeze_workqueues_begin(void)
5152{
Tejun Heo24b8a842013-03-12 11:29:58 -07005153 struct workqueue_struct *wq;
5154 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005155
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005156 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005157
Tejun Heo6183c002013-03-12 11:29:57 -07005158 WARN_ON_ONCE(workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005159 workqueue_freezing = true;
5160
Tejun Heo24b8a842013-03-12 11:29:58 -07005161 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07005162 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07005163 for_each_pwq(pwq, wq)
5164 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07005165 mutex_unlock(&wq->mutex);
Tejun Heo24b8a842013-03-12 11:29:58 -07005166 }
Tejun Heo5bcab332013-03-13 19:47:40 -07005167
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005168 mutex_unlock(&wq_pool_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005169}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005170
5171/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01005172 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005173 *
5174 * Check whether freezing is complete. This function must be called
5175 * between freeze_workqueues_begin() and thaw_workqueues().
5176 *
5177 * CONTEXT:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005178 * Grabs and releases wq_pool_mutex.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005179 *
Yacine Belkadid185af32013-07-31 14:59:24 -07005180 * Return:
Tejun Heo58a69cb2011-02-16 09:25:31 +01005181 * %true if some freezable workqueues are still busy. %false if freezing
5182 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005183 */
5184bool freeze_workqueues_busy(void)
5185{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005186 bool busy = false;
Tejun Heo24b8a842013-03-12 11:29:58 -07005187 struct workqueue_struct *wq;
5188 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005189
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005190 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005191
Tejun Heo6183c002013-03-12 11:29:57 -07005192 WARN_ON_ONCE(!workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005193
Tejun Heo24b8a842013-03-12 11:29:58 -07005194 list_for_each_entry(wq, &workqueues, list) {
5195 if (!(wq->flags & WQ_FREEZABLE))
5196 continue;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005197 /*
5198 * nr_active is monotonically decreasing. It's safe
5199 * to peek without lock.
5200 */
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005201 rcu_read_lock();
Tejun Heo24b8a842013-03-12 11:29:58 -07005202 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07005203 WARN_ON_ONCE(pwq->nr_active < 0);
Tejun Heo112202d2013-02-13 19:29:12 -08005204 if (pwq->nr_active) {
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005205 busy = true;
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005206 rcu_read_unlock();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005207 goto out_unlock;
5208 }
5209 }
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005210 rcu_read_unlock();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005211 }
5212out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005213 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005214 return busy;
5215}
5216
5217/**
5218 * thaw_workqueues - thaw workqueues
5219 *
5220 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo706026c2013-01-24 11:01:34 -08005221 * frozen works are transferred to their respective pool worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005222 *
5223 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07005224 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005225 */
5226void thaw_workqueues(void)
5227{
Tejun Heo24b8a842013-03-12 11:29:58 -07005228 struct workqueue_struct *wq;
5229 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005230
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005231 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005232
5233 if (!workqueue_freezing)
5234 goto out_unlock;
5235
Lai Jiangshan74b414e2014-05-22 19:01:16 +08005236 workqueue_freezing = false;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005237
Tejun Heo24b8a842013-03-12 11:29:58 -07005238 /* restore max_active and repopulate worklist */
5239 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07005240 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07005241 for_each_pwq(pwq, wq)
5242 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07005243 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005244 }
5245
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005246out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005247 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02005248}
5249#endif /* CONFIG_FREEZER */
5250
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005251static int workqueue_apply_unbound_cpumask(void)
5252{
5253 LIST_HEAD(ctxs);
5254 int ret = 0;
5255 struct workqueue_struct *wq;
5256 struct apply_wqattrs_ctx *ctx, *n;
5257
5258 lockdep_assert_held(&wq_pool_mutex);
5259
5260 list_for_each_entry(wq, &workqueues, list) {
5261 if (!(wq->flags & WQ_UNBOUND))
5262 continue;
5263 /* creating multiple pwqs breaks ordering guarantee */
5264 if (wq->flags & __WQ_ORDERED)
5265 continue;
5266
5267 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5268 if (!ctx) {
5269 ret = -ENOMEM;
5270 break;
5271 }
5272
5273 list_add_tail(&ctx->list, &ctxs);
5274 }
5275
5276 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5277 if (!ret)
5278 apply_wqattrs_commit(ctx);
5279 apply_wqattrs_cleanup(ctx);
5280 }
5281
5282 return ret;
5283}
5284
5285/**
5286 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5287 * @cpumask: the cpumask to set
5288 *
5289 * The low-level workqueues cpumask is a global cpumask that limits
5290 * the affinity of all unbound workqueues. This function check the @cpumask
5291 * and apply it to all unbound workqueues and updates all pwqs of them.
5292 *
5293 * Retun: 0 - Success
5294 * -EINVAL - Invalid @cpumask
5295 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5296 */
5297int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5298{
5299 int ret = -EINVAL;
5300 cpumask_var_t saved_cpumask;
5301
5302 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5303 return -ENOMEM;
5304
Tal Shorerc98a9802017-11-03 17:27:50 +02005305 /*
5306 * Not excluding isolated cpus on purpose.
5307 * If the user wishes to include them, we allow that.
5308 */
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005309 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5310 if (!cpumask_empty(cpumask)) {
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005311 apply_wqattrs_lock();
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005312
5313 /* save the old wq_unbound_cpumask. */
5314 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5315
5316 /* update wq_unbound_cpumask at first and apply it to wqs. */
5317 cpumask_copy(wq_unbound_cpumask, cpumask);
5318 ret = workqueue_apply_unbound_cpumask();
5319
5320 /* restore the wq_unbound_cpumask when failed. */
5321 if (ret < 0)
5322 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5323
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005324 apply_wqattrs_unlock();
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005325 }
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005326
5327 free_cpumask_var(saved_cpumask);
5328 return ret;
5329}
5330
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005331#ifdef CONFIG_SYSFS
5332/*
5333 * Workqueues with WQ_SYSFS flag set is visible to userland via
5334 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5335 * following attributes.
5336 *
5337 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5338 * max_active RW int : maximum number of in-flight work items
5339 *
5340 * Unbound workqueues have the following extra attributes.
5341 *
Wang Long9a19b462017-11-02 23:05:12 -04005342 * pool_ids RO int : the associated pool IDs for each node
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005343 * nice RW int : nice value of the workers
5344 * cpumask RW mask : bitmask of allowed CPUs for the workers
Wang Long9a19b462017-11-02 23:05:12 -04005345 * numa RW bool : whether enable NUMA affinity
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005346 */
5347struct wq_device {
5348 struct workqueue_struct *wq;
5349 struct device dev;
5350};
5351
5352static struct workqueue_struct *dev_to_wq(struct device *dev)
5353{
5354 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5355
5356 return wq_dev->wq;
5357}
5358
5359static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5360 char *buf)
5361{
5362 struct workqueue_struct *wq = dev_to_wq(dev);
5363
5364 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5365}
5366static DEVICE_ATTR_RO(per_cpu);
5367
5368static ssize_t max_active_show(struct device *dev,
5369 struct device_attribute *attr, char *buf)
5370{
5371 struct workqueue_struct *wq = dev_to_wq(dev);
5372
5373 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5374}
5375
5376static ssize_t max_active_store(struct device *dev,
5377 struct device_attribute *attr, const char *buf,
5378 size_t count)
5379{
5380 struct workqueue_struct *wq = dev_to_wq(dev);
5381 int val;
5382
5383 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5384 return -EINVAL;
5385
5386 workqueue_set_max_active(wq, val);
5387 return count;
5388}
5389static DEVICE_ATTR_RW(max_active);
5390
5391static struct attribute *wq_sysfs_attrs[] = {
5392 &dev_attr_per_cpu.attr,
5393 &dev_attr_max_active.attr,
5394 NULL,
5395};
5396ATTRIBUTE_GROUPS(wq_sysfs);
5397
5398static ssize_t wq_pool_ids_show(struct device *dev,
5399 struct device_attribute *attr, char *buf)
5400{
5401 struct workqueue_struct *wq = dev_to_wq(dev);
5402 const char *delim = "";
5403 int node, written = 0;
5404
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005405 get_online_cpus();
5406 rcu_read_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005407 for_each_node(node) {
5408 written += scnprintf(buf + written, PAGE_SIZE - written,
5409 "%s%d:%d", delim, node,
5410 unbound_pwq_by_node(wq, node)->pool->id);
5411 delim = " ";
5412 }
5413 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
Thomas Gleixner24acfb72019-03-13 17:55:47 +01005414 rcu_read_unlock();
5415 put_online_cpus();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005416
5417 return written;
5418}
5419
5420static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5421 char *buf)
5422{
5423 struct workqueue_struct *wq = dev_to_wq(dev);
5424 int written;
5425
5426 mutex_lock(&wq->mutex);
5427 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5428 mutex_unlock(&wq->mutex);
5429
5430 return written;
5431}
5432
5433/* prepare workqueue_attrs for sysfs store operations */
5434static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5435{
5436 struct workqueue_attrs *attrs;
5437
Lai Jiangshan899a94f2015-05-20 14:41:18 +08005438 lockdep_assert_held(&wq_pool_mutex);
5439
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005440 attrs = alloc_workqueue_attrs();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005441 if (!attrs)
5442 return NULL;
5443
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005444 copy_workqueue_attrs(attrs, wq->unbound_attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005445 return attrs;
5446}
5447
5448static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5449 const char *buf, size_t count)
5450{
5451 struct workqueue_struct *wq = dev_to_wq(dev);
5452 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005453 int ret = -ENOMEM;
5454
5455 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005456
5457 attrs = wq_sysfs_prep_attrs(wq);
5458 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005459 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005460
5461 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5462 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005463 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005464 else
5465 ret = -EINVAL;
5466
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005467out_unlock:
5468 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005469 free_workqueue_attrs(attrs);
5470 return ret ?: count;
5471}
5472
5473static ssize_t wq_cpumask_show(struct device *dev,
5474 struct device_attribute *attr, char *buf)
5475{
5476 struct workqueue_struct *wq = dev_to_wq(dev);
5477 int written;
5478
5479 mutex_lock(&wq->mutex);
5480 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5481 cpumask_pr_args(wq->unbound_attrs->cpumask));
5482 mutex_unlock(&wq->mutex);
5483 return written;
5484}
5485
5486static ssize_t wq_cpumask_store(struct device *dev,
5487 struct device_attribute *attr,
5488 const char *buf, size_t count)
5489{
5490 struct workqueue_struct *wq = dev_to_wq(dev);
5491 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005492 int ret = -ENOMEM;
5493
5494 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005495
5496 attrs = wq_sysfs_prep_attrs(wq);
5497 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005498 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005499
5500 ret = cpumask_parse(buf, attrs->cpumask);
5501 if (!ret)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005502 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005503
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005504out_unlock:
5505 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005506 free_workqueue_attrs(attrs);
5507 return ret ?: count;
5508}
5509
5510static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5511 char *buf)
5512{
5513 struct workqueue_struct *wq = dev_to_wq(dev);
5514 int written;
5515
5516 mutex_lock(&wq->mutex);
5517 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5518 !wq->unbound_attrs->no_numa);
5519 mutex_unlock(&wq->mutex);
5520
5521 return written;
5522}
5523
5524static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5525 const char *buf, size_t count)
5526{
5527 struct workqueue_struct *wq = dev_to_wq(dev);
5528 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005529 int v, ret = -ENOMEM;
5530
5531 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005532
5533 attrs = wq_sysfs_prep_attrs(wq);
5534 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005535 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005536
5537 ret = -EINVAL;
5538 if (sscanf(buf, "%d", &v) == 1) {
5539 attrs->no_numa = !v;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005540 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005541 }
5542
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005543out_unlock:
5544 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005545 free_workqueue_attrs(attrs);
5546 return ret ?: count;
5547}
5548
5549static struct device_attribute wq_sysfs_unbound_attrs[] = {
5550 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5551 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5552 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5553 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5554 __ATTR_NULL,
5555};
5556
5557static struct bus_type wq_subsys = {
5558 .name = "workqueue",
5559 .dev_groups = wq_sysfs_groups,
5560};
5561
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005562static ssize_t wq_unbound_cpumask_show(struct device *dev,
5563 struct device_attribute *attr, char *buf)
5564{
5565 int written;
5566
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005567 mutex_lock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005568 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5569 cpumask_pr_args(wq_unbound_cpumask));
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005570 mutex_unlock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005571
5572 return written;
5573}
5574
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005575static ssize_t wq_unbound_cpumask_store(struct device *dev,
5576 struct device_attribute *attr, const char *buf, size_t count)
5577{
5578 cpumask_var_t cpumask;
5579 int ret;
5580
5581 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5582 return -ENOMEM;
5583
5584 ret = cpumask_parse(buf, cpumask);
5585 if (!ret)
5586 ret = workqueue_set_unbound_cpumask(cpumask);
5587
5588 free_cpumask_var(cpumask);
5589 return ret ? ret : count;
5590}
5591
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005592static struct device_attribute wq_sysfs_cpumask_attr =
Lai Jiangshan042f7df12015-04-30 17:16:12 +08005593 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5594 wq_unbound_cpumask_store);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005595
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005596static int __init wq_sysfs_init(void)
5597{
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005598 int err;
5599
5600 err = subsys_virtual_register(&wq_subsys, NULL);
5601 if (err)
5602 return err;
5603
5604 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005605}
5606core_initcall(wq_sysfs_init);
5607
5608static void wq_device_release(struct device *dev)
5609{
5610 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5611
5612 kfree(wq_dev);
5613}
5614
5615/**
5616 * workqueue_sysfs_register - make a workqueue visible in sysfs
5617 * @wq: the workqueue to register
5618 *
5619 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5620 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5621 * which is the preferred method.
5622 *
5623 * Workqueue user should use this function directly iff it wants to apply
5624 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5625 * apply_workqueue_attrs() may race against userland updating the
5626 * attributes.
5627 *
5628 * Return: 0 on success, -errno on failure.
5629 */
5630int workqueue_sysfs_register(struct workqueue_struct *wq)
5631{
5632 struct wq_device *wq_dev;
5633 int ret;
5634
5635 /*
Shailendra Verma402dd892015-05-23 10:38:14 +05305636 * Adjusting max_active or creating new pwqs by applying
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005637 * attributes breaks ordering guarantee. Disallow exposing ordered
5638 * workqueues.
5639 */
Tejun Heo0a94efb2017-07-23 08:36:15 -04005640 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005641 return -EINVAL;
5642
5643 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5644 if (!wq_dev)
5645 return -ENOMEM;
5646
5647 wq_dev->wq = wq;
5648 wq_dev->dev.bus = &wq_subsys;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005649 wq_dev->dev.release = wq_device_release;
Lars-Peter Clausen23217b42016-02-17 21:04:41 +01005650 dev_set_name(&wq_dev->dev, "%s", wq->name);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005651
5652 /*
5653 * unbound_attrs are created separately. Suppress uevent until
5654 * everything is ready.
5655 */
5656 dev_set_uevent_suppress(&wq_dev->dev, true);
5657
5658 ret = device_register(&wq_dev->dev);
5659 if (ret) {
Arvind Yadav537f4142018-03-06 15:35:43 +05305660 put_device(&wq_dev->dev);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005661 wq->wq_dev = NULL;
5662 return ret;
5663 }
5664
5665 if (wq->flags & WQ_UNBOUND) {
5666 struct device_attribute *attr;
5667
5668 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5669 ret = device_create_file(&wq_dev->dev, attr);
5670 if (ret) {
5671 device_unregister(&wq_dev->dev);
5672 wq->wq_dev = NULL;
5673 return ret;
5674 }
5675 }
5676 }
5677
5678 dev_set_uevent_suppress(&wq_dev->dev, false);
5679 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5680 return 0;
5681}
5682
5683/**
5684 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5685 * @wq: the workqueue to unregister
5686 *
5687 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5688 */
5689static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5690{
5691 struct wq_device *wq_dev = wq->wq_dev;
5692
5693 if (!wq->wq_dev)
5694 return;
5695
5696 wq->wq_dev = NULL;
5697 device_unregister(&wq_dev->dev);
5698}
5699#else /* CONFIG_SYSFS */
5700static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5701#endif /* CONFIG_SYSFS */
5702
Tejun Heo82607adc2015-12-08 11:28:04 -05005703/*
5704 * Workqueue watchdog.
5705 *
5706 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5707 * flush dependency, a concurrency managed work item which stays RUNNING
5708 * indefinitely. Workqueue stalls can be very difficult to debug as the
5709 * usual warning mechanisms don't trigger and internal workqueue state is
5710 * largely opaque.
5711 *
5712 * Workqueue watchdog monitors all worker pools periodically and dumps
5713 * state if some pools failed to make forward progress for a while where
5714 * forward progress is defined as the first item on ->worklist changing.
5715 *
5716 * This mechanism is controlled through the kernel parameter
5717 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5718 * corresponding sysfs parameter file.
5719 */
5720#ifdef CONFIG_WQ_WATCHDOG
5721
Tejun Heo82607adc2015-12-08 11:28:04 -05005722static unsigned long wq_watchdog_thresh = 30;
Kees Cook5cd79d62017-10-04 16:27:00 -07005723static struct timer_list wq_watchdog_timer;
Tejun Heo82607adc2015-12-08 11:28:04 -05005724
5725static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5726static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5727
5728static void wq_watchdog_reset_touched(void)
5729{
5730 int cpu;
5731
5732 wq_watchdog_touched = jiffies;
5733 for_each_possible_cpu(cpu)
5734 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5735}
5736
Kees Cook5cd79d62017-10-04 16:27:00 -07005737static void wq_watchdog_timer_fn(struct timer_list *unused)
Tejun Heo82607adc2015-12-08 11:28:04 -05005738{
5739 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5740 bool lockup_detected = false;
5741 struct worker_pool *pool;
5742 int pi;
5743
5744 if (!thresh)
5745 return;
5746
5747 rcu_read_lock();
5748
5749 for_each_pool(pool, pi) {
5750 unsigned long pool_ts, touched, ts;
5751
5752 if (list_empty(&pool->worklist))
5753 continue;
5754
5755 /* get the latest of pool and touched timestamps */
5756 pool_ts = READ_ONCE(pool->watchdog_ts);
5757 touched = READ_ONCE(wq_watchdog_touched);
5758
5759 if (time_after(pool_ts, touched))
5760 ts = pool_ts;
5761 else
5762 ts = touched;
5763
5764 if (pool->cpu >= 0) {
5765 unsigned long cpu_touched =
5766 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5767 pool->cpu));
5768 if (time_after(cpu_touched, ts))
5769 ts = cpu_touched;
5770 }
5771
5772 /* did we stall? */
5773 if (time_after(jiffies, ts + thresh)) {
5774 lockup_detected = true;
5775 pr_emerg("BUG: workqueue lockup - pool");
5776 pr_cont_pool_info(pool);
5777 pr_cont(" stuck for %us!\n",
5778 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5779 }
5780 }
5781
5782 rcu_read_unlock();
5783
5784 if (lockup_detected)
5785 show_workqueue_state();
5786
5787 wq_watchdog_reset_touched();
5788 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5789}
5790
Vincent Whitchurchcb9d7fd2018-08-21 17:25:07 +02005791notrace void wq_watchdog_touch(int cpu)
Tejun Heo82607adc2015-12-08 11:28:04 -05005792{
5793 if (cpu >= 0)
5794 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5795 else
5796 wq_watchdog_touched = jiffies;
5797}
5798
5799static void wq_watchdog_set_thresh(unsigned long thresh)
5800{
5801 wq_watchdog_thresh = 0;
5802 del_timer_sync(&wq_watchdog_timer);
5803
5804 if (thresh) {
5805 wq_watchdog_thresh = thresh;
5806 wq_watchdog_reset_touched();
5807 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5808 }
5809}
5810
5811static int wq_watchdog_param_set_thresh(const char *val,
5812 const struct kernel_param *kp)
5813{
5814 unsigned long thresh;
5815 int ret;
5816
5817 ret = kstrtoul(val, 0, &thresh);
5818 if (ret)
5819 return ret;
5820
5821 if (system_wq)
5822 wq_watchdog_set_thresh(thresh);
5823 else
5824 wq_watchdog_thresh = thresh;
5825
5826 return 0;
5827}
5828
5829static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5830 .set = wq_watchdog_param_set_thresh,
5831 .get = param_get_ulong,
5832};
5833
5834module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5835 0644);
5836
5837static void wq_watchdog_init(void)
5838{
Kees Cook5cd79d62017-10-04 16:27:00 -07005839 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
Tejun Heo82607adc2015-12-08 11:28:04 -05005840 wq_watchdog_set_thresh(wq_watchdog_thresh);
5841}
5842
5843#else /* CONFIG_WQ_WATCHDOG */
5844
5845static inline void wq_watchdog_init(void) { }
5846
5847#endif /* CONFIG_WQ_WATCHDOG */
5848
Tejun Heobce90382013-04-01 11:23:32 -07005849static void __init wq_numa_init(void)
5850{
5851 cpumask_var_t *tbl;
5852 int node, cpu;
5853
Tejun Heobce90382013-04-01 11:23:32 -07005854 if (num_possible_nodes() <= 1)
5855 return;
5856
Tejun Heod55262c2013-04-01 11:23:38 -07005857 if (wq_disable_numa) {
5858 pr_info("workqueue: NUMA affinity support disabled\n");
5859 return;
5860 }
5861
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005862 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
Tejun Heo4c16bd32013-04-01 11:23:36 -07005863 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5864
Tejun Heobce90382013-04-01 11:23:32 -07005865 /*
5866 * We want masks of possible CPUs of each node which isn't readily
5867 * available. Build one from cpu_to_node() which should have been
5868 * fully initialized by now.
5869 */
Kees Cook6396bb22018-06-12 14:03:40 -07005870 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
Tejun Heobce90382013-04-01 11:23:32 -07005871 BUG_ON(!tbl);
5872
5873 for_each_node(node)
Yasuaki Ishimatsu5a6024f2014-07-07 09:56:48 -04005874 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
Tejun Heo1be0c252013-05-15 14:24:24 -07005875 node_online(node) ? node : NUMA_NO_NODE));
Tejun Heobce90382013-04-01 11:23:32 -07005876
5877 for_each_possible_cpu(cpu) {
5878 node = cpu_to_node(cpu);
5879 if (WARN_ON(node == NUMA_NO_NODE)) {
5880 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5881 /* happens iff arch is bonkers, let's just proceed */
5882 return;
5883 }
5884 cpumask_set_cpu(cpu, tbl[node]);
5885 }
5886
5887 wq_numa_possible_cpumask = tbl;
5888 wq_numa_enabled = true;
5889}
5890
Tejun Heo3347fa02016-09-16 15:49:32 -04005891/**
5892 * workqueue_init_early - early init for workqueue subsystem
5893 *
5894 * This is the first half of two-staged workqueue subsystem initialization
5895 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5896 * idr are up. It sets up all the data structures and system workqueues
5897 * and allows early boot code to create workqueues and queue/cancel work
5898 * items. Actual work item execution starts only after kthreads can be
5899 * created and scheduled right before early initcalls.
5900 */
Yu Chen2333e822020-02-23 15:28:52 +08005901void __init workqueue_init_early(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005902{
Tejun Heo7a4e3442013-03-12 11:30:00 -07005903 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
Frederic Weisbecker1bda3f82018-02-21 05:17:26 +01005904 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
Tejun Heo7a4e3442013-03-12 11:30:00 -07005905 int i, cpu;
Tejun Heoc34056a2010-06-29 10:07:11 +02005906
Tejun Heoe904e6c2013-03-12 11:29:57 -07005907 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5908
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005909 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
Frederic Weisbecker1bda3f82018-02-21 05:17:26 +01005910 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005911
Tejun Heoe904e6c2013-03-12 11:29:57 -07005912 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5913
Tejun Heo706026c2013-01-24 11:01:34 -08005914 /* initialize CPU pools */
Tejun Heo29c91e92013-03-12 11:30:03 -07005915 for_each_possible_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005916 struct worker_pool *pool;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005917
Tejun Heo7a4e3442013-03-12 11:30:00 -07005918 i = 0;
Tejun Heof02ae732013-03-12 11:30:03 -07005919 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo7a4e3442013-03-12 11:30:00 -07005920 BUG_ON(init_worker_pool(pool));
Tejun Heoec22ca52013-01-24 11:01:33 -08005921 pool->cpu = cpu;
Tejun Heo29c91e92013-03-12 11:30:03 -07005922 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
Tejun Heo7a4e3442013-03-12 11:30:00 -07005923 pool->attrs->nice = std_nice[i++];
Tejun Heof3f90ad2013-04-01 11:23:34 -07005924 pool->node = cpu_to_node(cpu);
Tejun Heo7a4e3442013-03-12 11:30:00 -07005925
Tejun Heo9daf9e62013-01-24 11:01:33 -08005926 /* alloc pool ID */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005927 mutex_lock(&wq_pool_mutex);
Tejun Heo9daf9e62013-01-24 11:01:33 -08005928 BUG_ON(worker_pool_assign_id(pool));
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005929 mutex_unlock(&wq_pool_mutex);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005930 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02005931 }
5932
Tejun Heo8a2b7532013-09-05 12:30:04 -04005933 /* create default unbound and ordered wq attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -07005934 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5935 struct workqueue_attrs *attrs;
5936
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005937 BUG_ON(!(attrs = alloc_workqueue_attrs()));
Tejun Heo29c91e92013-03-12 11:30:03 -07005938 attrs->nice = std_nice[i];
Tejun Heo29c91e92013-03-12 11:30:03 -07005939 unbound_std_wq_attrs[i] = attrs;
Tejun Heo8a2b7532013-09-05 12:30:04 -04005940
5941 /*
5942 * An ordered wq should have only one pwq as ordering is
5943 * guaranteed by max_active which is enforced by pwqs.
5944 * Turn off NUMA so that dfl_pwq is used for all nodes.
5945 */
Thomas Gleixnerbe69d002019-06-26 16:52:38 +02005946 BUG_ON(!(attrs = alloc_workqueue_attrs()));
Tejun Heo8a2b7532013-09-05 12:30:04 -04005947 attrs->nice = std_nice[i];
5948 attrs->no_numa = true;
5949 ordered_wq_attrs[i] = attrs;
Tejun Heo29c91e92013-03-12 11:30:03 -07005950 }
5951
Tejun Heod320c032010-06-29 10:07:14 +02005952 system_wq = alloc_workqueue("events", 0, 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005953 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
Tejun Heod320c032010-06-29 10:07:14 +02005954 system_long_wq = alloc_workqueue("events_long", 0, 0);
Tejun Heof3421792010-07-02 10:03:51 +02005955 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5956 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01005957 system_freezable_wq = alloc_workqueue("events_freezable",
5958 WQ_FREEZABLE, 0);
Viresh Kumar06681062013-04-24 17:12:54 +05305959 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5960 WQ_POWER_EFFICIENT, 0);
5961 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5962 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5963 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005964 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
Viresh Kumar06681062013-04-24 17:12:54 +05305965 !system_unbound_wq || !system_freezable_wq ||
5966 !system_power_efficient_wq ||
5967 !system_freezable_power_efficient_wq);
Tejun Heo3347fa02016-09-16 15:49:32 -04005968}
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 */
Yu Chen2333e822020-02-23 15:28:52 +08005979void __init workqueue_init(void)
Tejun Heo3347fa02016-09-16 15:49:32 -04005980{
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();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006026}