blob: cf5d8fd7472e63fd5d9c1d22e7a468da85820770 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Tejun Heoc54fce62010-09-10 16:51:36 +02002 * kernel/workqueue.c - generic async execution with shared worker pool
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
Tejun Heoc54fce62010-09-10 16:51:36 +02004 * Copyright (C) 2002 Ingo Molnar
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 *
Tejun Heoc54fce62010-09-10 16:51:36 +02006 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
Christoph Lameter89ada672005-10-30 15:01:59 -080011 *
Christoph Lametercde53532008-07-04 09:59:22 -070012 * Made to use alloc_percpu by Christoph Lameter.
Tejun Heoc54fce62010-09-10 16:51:36 +020013 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
Libinb11895c2013-08-21 08:50:39 +080019 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
Tejun Heoc54fce62010-09-10 16:51:36 +020023 *
Benjamin Peterson9a261492017-08-06 19:33:22 -070024 * Please read Documentation/core-api/workqueue.rst for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -070025 */
26
Paul Gortmaker9984de12011-05-23 14:51:41 -040027#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
James Bottomley1fa44ec2006-02-23 12:43:43 -060038#include <linux/hardirq.h>
Christoph Lameter46934022006-10-11 01:21:26 -070039#include <linux/mempolicy.h>
Rafael J. Wysocki341a5952006-12-06 20:34:49 -080040#include <linux/freezer.h>
Peter Zijlstrad5abe662006-12-06 20:37:26 -080041#include <linux/kallsyms.h>
42#include <linux/debug_locks.h>
Johannes Berg4e6045f2007-10-18 23:39:55 -070043#include <linux/lockdep.h>
Tejun Heoc34056a2010-06-29 10:07:11 +020044#include <linux/idr.h>
Tejun Heo29c91e92013-03-12 11:30:03 -070045#include <linux/jhash.h>
Sasha Levin42f85702012-12-17 10:01:23 -050046#include <linux/hashtable.h>
Tejun Heo76af4d92013-03-12 11:30:00 -070047#include <linux/rculist.h>
Tejun Heobce90382013-04-01 11:23:32 -070048#include <linux/nodemask.h>
Tejun Heo4c16bd32013-04-01 11:23:36 -070049#include <linux/moduleparam.h>
Tejun Heo3d1cb202013-04-30 15:27:22 -070050#include <linux/uaccess.h>
Sergey Senozhatskya6d59302018-01-11 09:53:35 +090051#include <linux/nmi.h>
Sergey Senozhatskyec32c8a2021-05-20 19:14:22 +090052#include <linux/kvm_para.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
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +080070 * 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 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800127 * A: 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 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700131 * PR: wq_pool_mutex protected for writes. Sched-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
136 * sched-RCU for reads.
137 *
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700138 * WQ: wq->mutex protected.
139 *
Lai Jiangshanb5927602013-03-25 16:57:19 -0700140 * WR: wq->mutex protected for writes. Sched-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 */
157 int nr_workers; /* L: total number of workers */
Lai Jiangshanea1abd62012-09-18 09:59:22 -0700158
159 /* nr_idle includes the ones off idle_list for rebinding */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700160 int nr_idle; /* L: currently idle ones */
161
162 struct list_head idle_list; /* X: list of idle workers */
163 struct timer_list idle_timer; /* L: worker idle timeout */
164 struct timer_list mayday_timer; /* L: SOS timer for workers */
165
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700166 /* a workers is either on busy_hash or idle_list, or the manager */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800167 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
168 /* L: hash of busy workers */
169
Tejun Heobc3a1af2013-03-13 19:47:39 -0700170 /* see manage_workers() for details on the two manager mutexes */
Tejun Heo2607d7a2015-03-09 09:22:28 -0400171 struct worker *manager; /* L: purely informational */
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800172 struct mutex attach_mutex; /* attach/detach exclusion */
173 struct list_head workers; /* A: attached workers */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +0800174 struct completion *detach_completion; /* all workers detached */
Tejun Heoe19e3972013-01-24 11:39:44 -0800175
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +0800176 struct ida worker_ida; /* worker IDs for task name */
Tejun Heoe19e3972013-01-24 11:39:44 -0800177
Tejun Heo7a4e3442013-03-12 11:30:00 -0700178 struct workqueue_attrs *attrs; /* I: worker attributes */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700179 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
180 int refcnt; /* PL: refcnt for unbound pools */
Tejun Heo7a4e3442013-03-12 11:30:00 -0700181
Tejun Heoe19e3972013-01-24 11:39:44 -0800182 /*
183 * The current concurrency level. As it's likely to be accessed
184 * from other CPUs during try_to_wake_up(), put it in a separate
185 * cacheline.
186 */
187 atomic_t nr_running ____cacheline_aligned_in_smp;
Tejun Heo29c91e92013-03-12 11:30:03 -0700188
189 /*
190 * Destruction of pool is sched-RCU protected to allow dereferences
191 * from get_work_pool().
192 */
193 struct rcu_head rcu;
Tejun Heo8b03ae32010-06-29 10:07:12 +0200194} ____cacheline_aligned_in_smp;
195
196/*
Tejun Heo112202d2013-02-13 19:29:12 -0800197 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
198 * of work_struct->data are used for flags and the remaining high bits
199 * point to the pwq; thus, pwqs need to be aligned at two's power of the
200 * number of flag bits.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201 */
Tejun Heo112202d2013-02-13 19:29:12 -0800202struct pool_workqueue {
Tejun Heobd7bdd42012-07-12 14:46:37 -0700203 struct worker_pool *pool; /* I: the associated pool */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200204 struct workqueue_struct *wq; /* I: the owning workqueue */
Tejun Heo73f53c42010-06-29 10:07:11 +0200205 int work_color; /* L: current color */
206 int flush_color; /* L: flushing color */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700207 int refcnt; /* L: reference count */
Tejun Heo73f53c42010-06-29 10:07:11 +0200208 int nr_in_flight[WORK_NR_COLORS];
209 /* L: nr of in_flight works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200210 int nr_active; /* L: nr of active works */
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200211 int max_active; /* L: max active works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200212 struct list_head delayed_works; /* L: delayed works */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700213 struct list_head pwqs_node; /* WR: node on wq->pwqs */
Tejun Heo2e109a22013-03-13 19:47:40 -0700214 struct list_head mayday_node; /* MD: node on wq->maydays */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700215
216 /*
217 * Release of unbound pwq is punted to system_wq. See put_pwq()
218 * and pwq_unbound_release_workfn() for details. pool_workqueue
219 * itself is also sched-RCU protected so that the first pwq can be
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700220 * determined without grabbing wq->mutex.
Tejun Heo8864b4e2013-03-12 11:30:04 -0700221 */
222 struct work_struct unbound_release_work;
223 struct rcu_head rcu;
Tejun Heoe904e6c2013-03-12 11:29:57 -0700224} __aligned(1 << WORK_STRUCT_FLAG_BITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226/*
Tejun Heo73f53c42010-06-29 10:07:11 +0200227 * Structure used to wait for workqueue flush.
228 */
229struct wq_flusher {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700230 struct list_head list; /* WQ: list of flushers */
231 int flush_color; /* WQ: flush color waiting for */
Tejun Heo73f53c42010-06-29 10:07:11 +0200232 struct completion done; /* flush completion */
233};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234
Tejun Heo226223a2013-03-12 11:30:05 -0700235struct wq_device;
236
Tejun Heo73f53c42010-06-29 10:07:11 +0200237/*
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700238 * The externally visible workqueue. It relays the issued work items to
239 * the appropriate worker_pool through its pool_workqueues.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 */
241struct workqueue_struct {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700242 struct list_head pwqs; /* WR: all pwqs of this wq */
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400243 struct list_head list; /* PR: list of all workqueues */
Tejun Heo73f53c42010-06-29 10:07:11 +0200244
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700245 struct mutex mutex; /* protects this wq */
246 int work_color; /* WQ: current work color */
247 int flush_color; /* WQ: current flush color */
Tejun Heo112202d2013-02-13 19:29:12 -0800248 atomic_t nr_pwqs_to_flush; /* flush in progress */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700249 struct wq_flusher *first_flusher; /* WQ: first flusher */
250 struct list_head flusher_queue; /* WQ: flush waiters */
251 struct list_head flusher_overflow; /* WQ: flush overflow list */
Tejun Heo73f53c42010-06-29 10:07:11 +0200252
Tejun Heo2e109a22013-03-13 19:47:40 -0700253 struct list_head maydays; /* MD: pwqs requesting rescue */
Tejun Heoe22bee72010-06-29 10:07:14 +0200254 struct worker *rescuer; /* I: rescue worker */
255
Lai Jiangshan87fc7412013-03-25 16:57:18 -0700256 int nr_drainers; /* WQ: drain in progress */
Lai Jiangshana357fc02013-03-25 16:57:19 -0700257 int saved_max_active; /* WQ: saved pwq max_active */
Tejun Heo226223a2013-03-12 11:30:05 -0700258
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800259 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
260 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
Tejun Heo6029a912013-04-01 11:23:34 -0700261
Tejun Heo226223a2013-03-12 11:30:05 -0700262#ifdef CONFIG_SYSFS
263 struct wq_device *wq_dev; /* I: for sysfs interface */
264#endif
Johannes Berg4e6045f2007-10-18 23:39:55 -0700265#ifdef CONFIG_LOCKDEP
Tejun Heo4690c4a2010-06-29 10:07:10 +0200266 struct lockdep_map lockdep_map;
Johannes Berg4e6045f2007-10-18 23:39:55 -0700267#endif
Tejun Heoecf68812013-04-01 11:23:34 -0700268 char name[WQ_NAME_LEN]; /* I: workqueue name */
Tejun Heo2728fd22013-04-01 11:23:35 -0700269
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400270 /*
271 * Destruction of workqueue_struct is sched-RCU protected to allow
272 * walking the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
274 */
275 struct rcu_head rcu;
276
Tejun Heo2728fd22013-04-01 11:23:35 -0700277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800280 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281};
282
Tejun Heoe904e6c2013-03-12 11:29:57 -0700283static struct kmem_cache *pwq_cache;
284
Tejun Heobce90382013-04-01 11:23:32 -0700285static cpumask_var_t *wq_numa_possible_cpumask;
286 /* possible CPUs of each node */
287
Tejun Heod55262c2013-04-01 11:23:38 -0700288static bool wq_disable_numa;
289module_param_named(disable_numa, wq_disable_numa, bool, 0444);
290
Viresh Kumarcee22a12013-04-08 16:45:40 +0530291/* see the comment above the definition of WQ_POWER_EFFICIENT */
Luis R. Rodriguez552f5302015-05-27 11:09:39 +0930292static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
Viresh Kumarcee22a12013-04-08 16:45:40 +0530293module_param_named(power_efficient, wq_power_efficient, bool, 0444);
294
Tejun Heo863b7102016-09-16 15:49:34 -0400295static bool wq_online; /* can kworkers be created yet? */
Tejun Heo3347fa02016-09-16 15:49:32 -0400296
Tejun Heobce90382013-04-01 11:23:32 -0700297static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
298
Tejun Heo4c16bd32013-04-01 11:23:36 -0700299/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
300static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
301
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700302static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
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 Heof303fcc2016-02-09 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 Heo7d19c5c2013-03-13 19:47:40 -0700358
Tejun Heo97bd2342010-10-05 10:41:14 +0200359#define CREATE_TRACE_POINTS
360#include <trace/events/workqueue.h>
361
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700362#define assert_rcu_or_pool_mutex() \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq_pool_mutex), \
365 "sched RCU or wq_pool_mutex should be held")
Tejun Heo5bcab332013-03-13 19:47:40 -0700366
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700367#define assert_rcu_or_wq_mutex(wq) \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex), \
370 "sched RCU or wq->mutex should be held")
Tejun Heo76af4d92013-03-12 11:30:00 -0700371
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800372#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700373 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
374 !lockdep_is_held(&wq->mutex) && \
375 !lockdep_is_held(&wq_pool_mutex), \
376 "sched RCU, wq->mutex or wq_pool_mutex should be held")
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800377
Tejun Heof02ae732013-03-12 11:30:03 -0700378#define for_each_cpu_worker_pool(pool, cpu) \
379 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
380 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
Tejun Heo7a62c2c2013-03-12 11:30:03 -0700381 (pool)++)
Tejun Heo4ce62e92012-07-13 22:16:44 -0700382
Tejun Heo49e3cf42013-03-12 11:29:58 -0700383/**
Tejun Heo17116962013-03-12 11:29:58 -0700384 * for_each_pool - iterate through all worker_pools in the system
385 * @pool: iteration cursor
Tejun Heo611c92a2013-03-13 16:51:36 -0700386 * @pi: integer used for iteration
Tejun Heofa1b54e2013-03-12 11:30:00 -0700387 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700388 * This must be called either with wq_pool_mutex held or sched RCU read
389 * locked. If the pool needs to be used beyond the locking in effect, the
390 * caller is responsible for guaranteeing that the pool stays online.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700391 *
392 * The if/else clause exists only for the lockdep assertion and can be
393 * ignored.
Tejun Heo17116962013-03-12 11:29:58 -0700394 */
Tejun Heo611c92a2013-03-13 16:51:36 -0700395#define for_each_pool(pool, pi) \
396 idr_for_each_entry(&worker_pool_idr, pool, pi) \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700397 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
Tejun Heofa1b54e2013-03-12 11:30:00 -0700398 else
Tejun Heo17116962013-03-12 11:29:58 -0700399
400/**
Tejun Heo822d8402013-03-19 13:45:21 -0700401 * for_each_pool_worker - iterate through all workers of a worker_pool
402 * @worker: iteration cursor
Tejun Heo822d8402013-03-19 13:45:21 -0700403 * @pool: worker_pool to iterate workers of
404 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800405 * This must be called with @pool->attach_mutex.
Tejun Heo822d8402013-03-19 13:45:21 -0700406 *
407 * The if/else clause exists only for the lockdep assertion and can be
408 * ignored.
409 */
Lai Jiangshanda028462014-05-20 17:46:31 +0800410#define for_each_pool_worker(worker, pool) \
411 list_for_each_entry((worker), &(pool)->workers, node) \
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800412 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
Tejun Heo822d8402013-03-19 13:45:21 -0700413 else
414
415/**
Tejun Heo49e3cf42013-03-12 11:29:58 -0700416 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
417 * @pwq: iteration cursor
418 * @wq: the target workqueue
Tejun Heo76af4d92013-03-12 11:30:00 -0700419 *
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700420 * This must be called either with wq->mutex held or sched RCU read locked.
Tejun Heo794b18b2013-03-13 19:47:40 -0700421 * If the pwq needs to be used beyond the locking in effect, the caller is
422 * responsible for guaranteeing that the pwq stays online.
Tejun Heo76af4d92013-03-12 11:30:00 -0700423 *
424 * The if/else clause exists only for the lockdep assertion and can be
425 * ignored.
Tejun Heo49e3cf42013-03-12 11:29:58 -0700426 */
427#define for_each_pwq(pwq, wq) \
Tejun Heo76af4d92013-03-12 11:30:00 -0700428 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700429 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
Tejun Heo76af4d92013-03-12 11:30:00 -0700430 else
Tejun Heof3421792010-07-02 10:03:51 +0200431
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900432#ifdef CONFIG_DEBUG_OBJECTS_WORK
433
434static struct debug_obj_descr work_debug_descr;
435
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100436static void *work_debug_hint(void *addr)
437{
438 return ((struct work_struct *) addr)->func;
439}
440
Du, Changbinb9fdac72016-05-19 17:09:41 -0700441static bool work_is_static_object(void *addr)
442{
443 struct work_struct *work = addr;
444
445 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
446}
447
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900448/*
449 * fixup_init is called when:
450 * - an active object is initialized
451 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700452static bool work_fixup_init(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900453{
454 struct work_struct *work = addr;
455
456 switch (state) {
457 case ODEBUG_STATE_ACTIVE:
458 cancel_work_sync(work);
459 debug_object_init(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700460 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900461 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700462 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900463 }
464}
465
466/*
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900467 * fixup_free is called when:
468 * - an active object is freed
469 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700470static bool work_fixup_free(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900471{
472 struct work_struct *work = addr;
473
474 switch (state) {
475 case ODEBUG_STATE_ACTIVE:
476 cancel_work_sync(work);
477 debug_object_free(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700478 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900479 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700480 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900481 }
482}
483
484static struct debug_obj_descr work_debug_descr = {
485 .name = "work_struct",
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100486 .debug_hint = work_debug_hint,
Du, Changbinb9fdac72016-05-19 17:09:41 -0700487 .is_static_object = work_is_static_object,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900488 .fixup_init = work_fixup_init,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900489 .fixup_free = work_fixup_free,
490};
491
492static inline void debug_work_activate(struct work_struct *work)
493{
494 debug_object_activate(work, &work_debug_descr);
495}
496
497static inline void debug_work_deactivate(struct work_struct *work)
498{
499 debug_object_deactivate(work, &work_debug_descr);
500}
501
502void __init_work(struct work_struct *work, int onstack)
503{
504 if (onstack)
505 debug_object_init_on_stack(work, &work_debug_descr);
506 else
507 debug_object_init(work, &work_debug_descr);
508}
509EXPORT_SYMBOL_GPL(__init_work);
510
511void destroy_work_on_stack(struct work_struct *work)
512{
513 debug_object_free(work, &work_debug_descr);
514}
515EXPORT_SYMBOL_GPL(destroy_work_on_stack);
516
Thomas Gleixnerea2e64f2014-03-23 14:20:44 +0000517void destroy_delayed_work_on_stack(struct delayed_work *work)
518{
519 destroy_timer_on_stack(&work->timer);
520 debug_object_free(&work->work, &work_debug_descr);
521}
522EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
523
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900524#else
525static inline void debug_work_activate(struct work_struct *work) { }
526static inline void debug_work_deactivate(struct work_struct *work) { }
527#endif
528
Li Bin4e8b22b2013-09-10 09:52:35 +0800529/**
530 * worker_pool_assign_id - allocate ID and assing it to @pool
531 * @pool: the pool pointer of interest
532 *
533 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
534 * successfully, -errno on failure.
535 */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800536static int worker_pool_assign_id(struct worker_pool *pool)
537{
538 int ret;
539
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700540 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo5bcab332013-03-13 19:47:40 -0700541
Li Bin4e8b22b2013-09-10 09:52:35 +0800542 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
543 GFP_KERNEL);
Tejun Heo229641a2013-04-01 17:08:13 -0700544 if (ret >= 0) {
Tejun Heoe68035f2013-03-13 14:59:38 -0700545 pool->id = ret;
Tejun Heo229641a2013-04-01 17:08:13 -0700546 return 0;
547 }
Tejun Heo9daf9e62013-01-24 11:01:33 -0800548 return ret;
549}
550
Tejun Heo76af4d92013-03-12 11:30:00 -0700551/**
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700552 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
553 * @wq: the target workqueue
554 * @node: the node ID
555 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800556 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
557 * read locked.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700558 * If the pwq needs to be used beyond the locking in effect, the caller is
559 * responsible for guaranteeing that the pwq stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700560 *
561 * Return: The unbound pool_workqueue for @node.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700562 */
563static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
564 int node)
565{
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800566 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
Tejun Heod6e022f2016-02-03 13:54:25 -0500567
568 /*
569 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
570 * delayed item is pending. The plan is to keep CPU -> NODE
571 * mapping valid and stable across CPU on/offlines. Once that
572 * happens, this workaround can be removed.
573 */
574 if (unlikely(node == NUMA_NO_NODE))
575 return wq->dfl_pwq;
576
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700577 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
578}
579
Tejun Heo73f53c42010-06-29 10:07:11 +0200580static unsigned int work_color_to_flags(int color)
581{
582 return color << WORK_STRUCT_COLOR_SHIFT;
583}
584
585static int get_work_color(struct work_struct *work)
586{
587 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
588 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
589}
590
591static int work_next_color(int color)
592{
593 return (color + 1) % WORK_NR_COLORS;
Oleg Nesterova848e3b2007-05-09 02:34:17 -0700594}
595
David Howells4594bf12006-12-07 11:33:26 +0000596/*
Tejun Heo112202d2013-02-13 19:29:12 -0800597 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
598 * contain the pointer to the queued pwq. Once execution starts, the flag
Tejun Heo7c3eed52013-01-24 11:01:33 -0800599 * is cleared and the high bits contain OFFQ flags and pool ID.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200600 *
Tejun Heo112202d2013-02-13 19:29:12 -0800601 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
602 * and clear_work_data() can be used to set the pwq, pool or clear
Tejun Heobbb68df2012-08-03 10:30:46 -0700603 * work->data. These functions should only be called while the work is
604 * owned - ie. while the PENDING bit is set.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200605 *
Tejun Heo112202d2013-02-13 19:29:12 -0800606 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
Tejun Heo7c3eed52013-01-24 11:01:33 -0800607 * corresponding to a work. Pool is available once the work has been
Tejun Heo112202d2013-02-13 19:29:12 -0800608 * queued anywhere after initialization until it is sync canceled. pwq is
Tejun Heo7c3eed52013-01-24 11:01:33 -0800609 * available only while the work item is queued.
Tejun Heobbb68df2012-08-03 10:30:46 -0700610 *
611 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
612 * canceled. While being canceled, a work item may have its PENDING set
613 * but stay off timer and worklist for arbitrarily long and nobody should
614 * try to steal the PENDING bit.
David Howells4594bf12006-12-07 11:33:26 +0000615 */
Tejun Heo7a22ad72010-06-29 10:07:13 +0200616static inline void set_work_data(struct work_struct *work, unsigned long data,
617 unsigned long flags)
David Howells365970a2006-11-22 14:54:49 +0000618{
Tejun Heo6183c002013-03-12 11:29:57 -0700619 WARN_ON_ONCE(!work_pending(work));
Tejun Heo7a22ad72010-06-29 10:07:13 +0200620 atomic_long_set(&work->data, data | flags | work_static(work));
David Howells365970a2006-11-22 14:54:49 +0000621}
David Howells365970a2006-11-22 14:54:49 +0000622
Tejun Heo112202d2013-02-13 19:29:12 -0800623static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
Tejun Heo7a22ad72010-06-29 10:07:13 +0200624 unsigned long extra_flags)
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200625{
Tejun Heo112202d2013-02-13 19:29:12 -0800626 set_work_data(work, (unsigned long)pwq,
627 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200628}
629
Lai Jiangshan4468a002013-02-06 18:04:53 -0800630static void set_work_pool_and_keep_pending(struct work_struct *work,
631 int pool_id)
632{
633 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
634 WORK_STRUCT_PENDING);
635}
636
Tejun Heo7c3eed52013-01-24 11:01:33 -0800637static void set_work_pool_and_clear_pending(struct work_struct *work,
638 int pool_id)
David Howells365970a2006-11-22 14:54:49 +0000639{
Tejun Heo23657bb2012-08-13 17:08:19 -0700640 /*
641 * The following wmb is paired with the implied mb in
642 * test_and_set_bit(PENDING) and ensures all updates to @work made
643 * here are visible to and precede any updates by the next PENDING
644 * owner.
645 */
646 smp_wmb();
Tejun Heo7c3eed52013-01-24 11:01:33 -0800647 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
Roman Pen346c09f2016-04-26 13:15:35 +0200648 /*
649 * The following mb guarantees that previous clear of a PENDING bit
650 * will not be reordered with any speculative LOADS or STORES from
651 * work->current_func, which is executed afterwards. This possible
652 * reordering can lead to a missed execution on attempt to qeueue
653 * the same @work. E.g. consider this case:
654 *
655 * CPU#0 CPU#1
656 * ---------------------------- --------------------------------
657 *
658 * 1 STORE event_indicated
659 * 2 queue_work_on() {
660 * 3 test_and_set_bit(PENDING)
661 * 4 } set_..._and_clear_pending() {
662 * 5 set_work_data() # clear bit
663 * 6 smp_mb()
664 * 7 work->current_func() {
665 * 8 LOAD event_indicated
666 * }
667 *
668 * Without an explicit full barrier speculative LOAD on line 8 can
669 * be executed before CPU#0 does STORE on line 1. If that happens,
670 * CPU#0 observes the PENDING bit is still set and new execution of
671 * a @work is not queued in a hope, that CPU#1 will eventually
672 * finish the queued @work. Meanwhile CPU#1 does not see
673 * event_indicated is set, because speculative LOAD was executed
674 * before actual STORE.
675 */
676 smp_mb();
Tejun Heo7a22ad72010-06-29 10:07:13 +0200677}
678
679static void clear_work_data(struct work_struct *work)
680{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800681 smp_wmb(); /* see set_work_pool_and_clear_pending() */
682 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200683}
684
Tejun Heo112202d2013-02-13 19:29:12 -0800685static struct pool_workqueue *get_work_pwq(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200686{
Tejun Heoe1201532010-07-22 14:14:25 +0200687 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200688
Tejun Heo112202d2013-02-13 19:29:12 -0800689 if (data & WORK_STRUCT_PWQ)
Tejun Heoe1201532010-07-22 14:14:25 +0200690 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
691 else
692 return NULL;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200693}
694
Tejun Heo7c3eed52013-01-24 11:01:33 -0800695/**
696 * get_work_pool - return the worker_pool a given work was associated with
697 * @work: the work item of interest
698 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700699 * Pools are created and destroyed under wq_pool_mutex, and allows read
700 * access under sched-RCU read lock. As such, this function should be
701 * called under wq_pool_mutex or with preemption disabled.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700702 *
703 * All fields of the returned pool are accessible as long as the above
704 * mentioned locking is in effect. If the returned pool needs to be used
705 * beyond the critical section, the caller is responsible for ensuring the
706 * returned pool is and stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700707 *
708 * Return: The worker_pool @work was last associated with. %NULL if none.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800709 */
710static struct worker_pool *get_work_pool(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200711{
Tejun Heoe1201532010-07-22 14:14:25 +0200712 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800713 int pool_id;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200714
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700715 assert_rcu_or_pool_mutex();
Tejun Heofa1b54e2013-03-12 11:30:00 -0700716
Tejun Heo112202d2013-02-13 19:29:12 -0800717 if (data & WORK_STRUCT_PWQ)
718 return ((struct pool_workqueue *)
Tejun Heo7c3eed52013-01-24 11:01:33 -0800719 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200720
Tejun Heo7c3eed52013-01-24 11:01:33 -0800721 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
722 if (pool_id == WORK_OFFQ_POOL_NONE)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200723 return NULL;
724
Tejun Heofa1b54e2013-03-12 11:30:00 -0700725 return idr_find(&worker_pool_idr, pool_id);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800726}
727
728/**
729 * get_work_pool_id - return the worker pool ID a given work is associated with
730 * @work: the work item of interest
731 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700732 * Return: The worker_pool ID @work was last associated with.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800733 * %WORK_OFFQ_POOL_NONE if none.
734 */
735static int get_work_pool_id(struct work_struct *work)
736{
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800737 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800738
Tejun Heo112202d2013-02-13 19:29:12 -0800739 if (data & WORK_STRUCT_PWQ)
740 return ((struct pool_workqueue *)
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800741 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
742
743 return data >> WORK_OFFQ_POOL_SHIFT;
Tejun Heo7c3eed52013-01-24 11:01:33 -0800744}
745
Tejun Heobbb68df2012-08-03 10:30:46 -0700746static void mark_work_canceling(struct work_struct *work)
747{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800748 unsigned long pool_id = get_work_pool_id(work);
Tejun Heobbb68df2012-08-03 10:30:46 -0700749
Tejun Heo7c3eed52013-01-24 11:01:33 -0800750 pool_id <<= WORK_OFFQ_POOL_SHIFT;
751 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700752}
753
754static bool work_is_canceling(struct work_struct *work)
755{
756 unsigned long data = atomic_long_read(&work->data);
757
Tejun Heo112202d2013-02-13 19:29:12 -0800758 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700759}
760
David Howells365970a2006-11-22 14:54:49 +0000761/*
Tejun Heo32704762012-07-13 22:16:45 -0700762 * Policy functions. These define the policies on how the global worker
763 * pools are managed. Unless noted otherwise, these functions assume that
Tejun Heod565ed62013-01-24 11:01:33 -0800764 * they're being called with pool->lock held.
David Howells365970a2006-11-22 14:54:49 +0000765 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200766
Tejun Heo63d95a92012-07-12 14:46:37 -0700767static bool __need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000768{
Tejun Heoe19e3972013-01-24 11:39:44 -0800769 return !atomic_read(&pool->nr_running);
David Howells365970a2006-11-22 14:54:49 +0000770}
771
Tejun Heoe22bee72010-06-29 10:07:14 +0200772/*
773 * Need to wake up a worker? Called from anything but currently
774 * running workers.
Tejun Heo974271c2012-07-12 14:46:37 -0700775 *
776 * Note that, because unbound workers never contribute to nr_running, this
Tejun Heo706026c2013-01-24 11:01:34 -0800777 * function will always return %true for unbound pools as long as the
Tejun Heo974271c2012-07-12 14:46:37 -0700778 * worklist isn't empty.
Tejun Heoe22bee72010-06-29 10:07:14 +0200779 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700780static bool need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000781{
Tejun Heo63d95a92012-07-12 14:46:37 -0700782 return !list_empty(&pool->worklist) && __need_more_worker(pool);
David Howells365970a2006-11-22 14:54:49 +0000783}
784
Tejun Heoe22bee72010-06-29 10:07:14 +0200785/* Can I start working? Called from busy but !running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700786static bool may_start_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200787{
Tejun Heo63d95a92012-07-12 14:46:37 -0700788 return pool->nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200789}
790
791/* Do I need to keep working? Called from currently running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700792static bool keep_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200793{
Tejun Heoe19e3972013-01-24 11:39:44 -0800794 return !list_empty(&pool->worklist) &&
795 atomic_read(&pool->nr_running) <= 1;
Tejun Heoe22bee72010-06-29 10:07:14 +0200796}
797
798/* Do we need a new worker? Called from manager. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700799static bool need_to_create_worker(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200800{
Tejun Heo63d95a92012-07-12 14:46:37 -0700801 return need_more_worker(pool) && !may_start_working(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200802}
803
Tejun Heoe22bee72010-06-29 10:07:14 +0200804/* Do we have too many workers and should some go away? */
Tejun Heo63d95a92012-07-12 14:46:37 -0700805static bool too_many_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200806{
Tejun Heo692b4822017-10-09 08:04:13 -0700807 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
Tejun Heo63d95a92012-07-12 14:46:37 -0700808 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
809 int nr_busy = pool->nr_workers - nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200810
811 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
812}
813
814/*
815 * Wake up functions.
816 */
817
Lai Jiangshan1037de32014-05-22 16:44:07 +0800818/* Return the first idle worker. Safe with preemption disabled */
819static struct worker *first_idle_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200820{
Tejun Heo63d95a92012-07-12 14:46:37 -0700821 if (unlikely(list_empty(&pool->idle_list)))
Tejun Heo7e116292010-06-29 10:07:13 +0200822 return NULL;
823
Tejun Heo63d95a92012-07-12 14:46:37 -0700824 return list_first_entry(&pool->idle_list, struct worker, entry);
Tejun Heo7e116292010-06-29 10:07:13 +0200825}
826
827/**
828 * wake_up_worker - wake up an idle worker
Tejun Heo63d95a92012-07-12 14:46:37 -0700829 * @pool: worker pool to wake worker from
Tejun Heo7e116292010-06-29 10:07:13 +0200830 *
Tejun Heo63d95a92012-07-12 14:46:37 -0700831 * Wake up the first idle worker of @pool.
Tejun Heo7e116292010-06-29 10:07:13 +0200832 *
833 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800834 * spin_lock_irq(pool->lock).
Tejun Heo7e116292010-06-29 10:07:13 +0200835 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700836static void wake_up_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200837{
Lai Jiangshan1037de32014-05-22 16:44:07 +0800838 struct worker *worker = first_idle_worker(pool);
Tejun Heo7e116292010-06-29 10:07:13 +0200839
840 if (likely(worker))
841 wake_up_process(worker->task);
842}
843
Tejun Heo4690c4a2010-06-29 10:07:10 +0200844/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200845 * wq_worker_waking_up - a worker is waking up
846 * @task: task waking up
847 * @cpu: CPU @task is waking up to
848 *
849 * This function is called during try_to_wake_up() when a worker is
850 * being awoken.
851 *
852 * CONTEXT:
853 * spin_lock_irq(rq->lock)
854 */
Tejun Heod84ff052013-03-12 11:29:59 -0700855void wq_worker_waking_up(struct task_struct *task, int cpu)
Tejun Heoe22bee72010-06-29 10:07:14 +0200856{
857 struct worker *worker = kthread_data(task);
858
Joonsoo Kim36576002012-10-26 23:03:49 +0900859 if (!(worker->flags & WORKER_NOT_RUNNING)) {
Tejun Heoec22ca52013-01-24 11:01:33 -0800860 WARN_ON_ONCE(worker->pool->cpu != cpu);
Tejun Heoe19e3972013-01-24 11:39:44 -0800861 atomic_inc(&worker->pool->nr_running);
Joonsoo Kim36576002012-10-26 23:03:49 +0900862 }
Tejun Heoe22bee72010-06-29 10:07:14 +0200863}
864
865/**
866 * wq_worker_sleeping - a worker is going to sleep
867 * @task: task going to sleep
Tejun Heoe22bee72010-06-29 10:07:14 +0200868 *
869 * This function is called during schedule() when a busy worker is
870 * going to sleep. Worker on the same cpu can be woken up by
871 * returning pointer to its task.
872 *
873 * CONTEXT:
874 * spin_lock_irq(rq->lock)
875 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700876 * Return:
Tejun Heoe22bee72010-06-29 10:07:14 +0200877 * Worker task on @cpu to wake up, %NULL if none.
878 */
Alexander Gordeev9b7f6592016-03-02 12:53:31 +0100879struct task_struct *wq_worker_sleeping(struct task_struct *task)
Tejun Heoe22bee72010-06-29 10:07:14 +0200880{
881 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
Tejun Heo111c2252013-01-17 17:16:24 -0800882 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200883
Tejun Heo111c2252013-01-17 17:16:24 -0800884 /*
885 * Rescuers, which may not have all the fields set up like normal
886 * workers, also reach here, let's not access anything before
887 * checking NOT_RUNNING.
888 */
Steven Rostedt2d646722010-12-03 23:12:33 -0500889 if (worker->flags & WORKER_NOT_RUNNING)
Tejun Heoe22bee72010-06-29 10:07:14 +0200890 return NULL;
891
Tejun Heo111c2252013-01-17 17:16:24 -0800892 pool = worker->pool;
Tejun Heo111c2252013-01-17 17:16:24 -0800893
Tejun Heoe22bee72010-06-29 10:07:14 +0200894 /* this can only happen on the local cpu */
Alexander Gordeev9b7f6592016-03-02 12:53:31 +0100895 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
Tejun Heo6183c002013-03-12 11:29:57 -0700896 return NULL;
Tejun Heoe22bee72010-06-29 10:07:14 +0200897
898 /*
899 * The counterpart of the following dec_and_test, implied mb,
900 * worklist not empty test sequence is in insert_work().
901 * Please read comment there.
902 *
Tejun Heo628c78e2012-07-17 12:39:27 -0700903 * NOT_RUNNING is clear. This means that we're bound to and
904 * running on the local cpu w/ rq lock held and preemption
905 * disabled, which in turn means that none else could be
Tejun Heod565ed62013-01-24 11:01:33 -0800906 * manipulating idle_list, so dereferencing idle_list without pool
Tejun Heo628c78e2012-07-17 12:39:27 -0700907 * lock is safe.
Tejun Heoe22bee72010-06-29 10:07:14 +0200908 */
Tejun Heoe19e3972013-01-24 11:39:44 -0800909 if (atomic_dec_and_test(&pool->nr_running) &&
910 !list_empty(&pool->worklist))
Lai Jiangshan1037de32014-05-22 16:44:07 +0800911 to_wakeup = first_idle_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200912 return to_wakeup ? to_wakeup->task : NULL;
913}
914
915/**
916 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200917 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200918 * @flags: flags to set
Tejun Heod302f012010-06-29 10:07:13 +0200919 *
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800920 * Set @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200921 *
Tejun Heocb444762010-07-02 10:03:50 +0200922 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800923 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200924 */
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800925static inline void worker_set_flags(struct worker *worker, unsigned int flags)
Tejun Heod302f012010-06-29 10:07:13 +0200926{
Tejun Heobd7bdd42012-07-12 14:46:37 -0700927 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200928
Tejun Heocb444762010-07-02 10:03:50 +0200929 WARN_ON_ONCE(worker->task != current);
930
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800931 /* If transitioning into NOT_RUNNING, adjust nr_running. */
Tejun Heoe22bee72010-06-29 10:07:14 +0200932 if ((flags & WORKER_NOT_RUNNING) &&
933 !(worker->flags & WORKER_NOT_RUNNING)) {
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800934 atomic_dec(&pool->nr_running);
Tejun Heoe22bee72010-06-29 10:07:14 +0200935 }
936
Tejun Heod302f012010-06-29 10:07:13 +0200937 worker->flags |= flags;
938}
939
940/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200941 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200942 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200943 * @flags: flags to clear
944 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200945 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200946 *
Tejun Heocb444762010-07-02 10:03:50 +0200947 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800948 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200949 */
950static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
951{
Tejun Heo63d95a92012-07-12 14:46:37 -0700952 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200953 unsigned int oflags = worker->flags;
954
Tejun Heocb444762010-07-02 10:03:50 +0200955 WARN_ON_ONCE(worker->task != current);
956
Tejun Heod302f012010-06-29 10:07:13 +0200957 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200958
Tejun Heo42c025f2011-01-11 15:58:49 +0100959 /*
960 * If transitioning out of NOT_RUNNING, increment nr_running. Note
961 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
962 * of multiple flags, not a single flag.
963 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200964 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
965 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800966 atomic_inc(&pool->nr_running);
Tejun Heod302f012010-06-29 10:07:13 +0200967}
968
969/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200970 * find_worker_executing_work - find worker which is executing a work
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800971 * @pool: pool of interest
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200972 * @work: work to find worker for
973 *
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800974 * Find a worker which is executing @work on @pool by searching
975 * @pool->busy_hash which is keyed by the address of @work. For a worker
Tejun Heoa2c1c572012-12-18 10:35:02 -0800976 * to match, its current execution should match the address of @work and
977 * its work function. This is to avoid unwanted dependency between
978 * unrelated work executions through a work item being recycled while still
979 * being executed.
980 *
981 * This is a bit tricky. A work item may be freed once its execution
982 * starts and nothing prevents the freed area from being recycled for
983 * another work item. If the same work item address ends up being reused
984 * before the original execution finishes, workqueue will identify the
985 * recycled work item as currently executing and make it wait until the
986 * current execution finishes, introducing an unwanted dependency.
987 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700988 * This function checks the work item address and work function to avoid
989 * false positives. Note that this isn't complete as one may construct a
990 * work function which can introduce dependency onto itself through a
991 * recycled work item. Well, if somebody wants to shoot oneself in the
992 * foot that badly, there's only so much we can do, and if such deadlock
993 * actually occurs, it should be easy to locate the culprit work function.
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200994 *
995 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800996 * spin_lock_irq(pool->lock).
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200997 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700998 * Return:
999 * Pointer to worker which is executing @work if found, %NULL
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001000 * otherwise.
1001 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08001002static struct worker *find_worker_executing_work(struct worker_pool *pool,
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001003 struct work_struct *work)
1004{
Sasha Levin42f85702012-12-17 10:01:23 -05001005 struct worker *worker;
Sasha Levin42f85702012-12-17 10:01:23 -05001006
Sasha Levinb67bfe02013-02-27 17:06:00 -08001007 hash_for_each_possible(pool->busy_hash, worker, hentry,
Tejun Heoa2c1c572012-12-18 10:35:02 -08001008 (unsigned long)work)
1009 if (worker->current_work == work &&
1010 worker->current_func == work->func)
Sasha Levin42f85702012-12-17 10:01:23 -05001011 return worker;
1012
1013 return NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001014}
1015
1016/**
Tejun Heobf4ede02012-08-03 10:30:46 -07001017 * move_linked_works - move linked works to a list
1018 * @work: start of series of works to be scheduled
1019 * @head: target list to append @work to
Shailendra Verma402dd892015-05-23 10:38:14 +05301020 * @nextp: out parameter for nested worklist walking
Tejun Heobf4ede02012-08-03 10:30:46 -07001021 *
1022 * Schedule linked works starting from @work to @head. Work series to
1023 * be scheduled starts at @work and includes any consecutive work with
1024 * WORK_STRUCT_LINKED set in its predecessor.
1025 *
1026 * If @nextp is not NULL, it's updated to point to the next work of
1027 * the last scheduled work. This allows move_linked_works() to be
1028 * nested inside outer list_for_each_entry_safe().
1029 *
1030 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001031 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001032 */
1033static void move_linked_works(struct work_struct *work, struct list_head *head,
1034 struct work_struct **nextp)
1035{
1036 struct work_struct *n;
1037
1038 /*
1039 * Linked worklist will always end before the end of the list,
1040 * use NULL for list head.
1041 */
1042 list_for_each_entry_safe_from(work, n, NULL, entry) {
1043 list_move_tail(&work->entry, head);
1044 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1045 break;
1046 }
1047
1048 /*
1049 * If we're already inside safe list traversal and have moved
1050 * multiple works to the scheduled queue, the next position
1051 * needs to be updated.
1052 */
1053 if (nextp)
1054 *nextp = n;
1055}
1056
Tejun Heo8864b4e2013-03-12 11:30:04 -07001057/**
1058 * get_pwq - get an extra reference on the specified pool_workqueue
1059 * @pwq: pool_workqueue to get
1060 *
1061 * Obtain an extra reference on @pwq. The caller should guarantee that
1062 * @pwq has positive refcnt and be holding the matching pool->lock.
1063 */
1064static void get_pwq(struct pool_workqueue *pwq)
1065{
1066 lockdep_assert_held(&pwq->pool->lock);
1067 WARN_ON_ONCE(pwq->refcnt <= 0);
1068 pwq->refcnt++;
1069}
1070
1071/**
1072 * put_pwq - put a pool_workqueue reference
1073 * @pwq: pool_workqueue to put
1074 *
1075 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1076 * destruction. The caller should be holding the matching pool->lock.
1077 */
1078static void put_pwq(struct pool_workqueue *pwq)
1079{
1080 lockdep_assert_held(&pwq->pool->lock);
1081 if (likely(--pwq->refcnt))
1082 return;
1083 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1084 return;
1085 /*
1086 * @pwq can't be released under pool->lock, bounce to
1087 * pwq_unbound_release_workfn(). This never recurses on the same
1088 * pool->lock as this path is taken only for unbound workqueues and
1089 * the release work item is scheduled on a per-cpu workqueue. To
1090 * avoid lockdep warning, unbound pool->locks are given lockdep
1091 * subclass of 1 in get_unbound_pool().
1092 */
1093 schedule_work(&pwq->unbound_release_work);
1094}
1095
Tejun Heodce90d42013-04-01 11:23:35 -07001096/**
1097 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1098 * @pwq: pool_workqueue to put (can be %NULL)
1099 *
1100 * put_pwq() with locking. This function also allows %NULL @pwq.
1101 */
1102static void put_pwq_unlocked(struct pool_workqueue *pwq)
1103{
1104 if (pwq) {
1105 /*
1106 * As both pwqs and pools are sched-RCU protected, the
1107 * following lock operations are safe.
1108 */
1109 spin_lock_irq(&pwq->pool->lock);
1110 put_pwq(pwq);
1111 spin_unlock_irq(&pwq->pool->lock);
1112 }
1113}
1114
Tejun Heo112202d2013-02-13 19:29:12 -08001115static void pwq_activate_delayed_work(struct work_struct *work)
Tejun Heobf4ede02012-08-03 10:30:46 -07001116{
Tejun Heo112202d2013-02-13 19:29:12 -08001117 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobf4ede02012-08-03 10:30:46 -07001118
1119 trace_workqueue_activate_work(work);
Tejun Heo82607adc2015-12-08 11:28:04 -05001120 if (list_empty(&pwq->pool->worklist))
1121 pwq->pool->watchdog_ts = jiffies;
Tejun Heo112202d2013-02-13 19:29:12 -08001122 move_linked_works(work, &pwq->pool->worklist, NULL);
Tejun Heobf4ede02012-08-03 10:30:46 -07001123 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo112202d2013-02-13 19:29:12 -08001124 pwq->nr_active++;
Tejun Heobf4ede02012-08-03 10:30:46 -07001125}
1126
Tejun Heo112202d2013-02-13 19:29:12 -08001127static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001128{
Tejun Heo112202d2013-02-13 19:29:12 -08001129 struct work_struct *work = list_first_entry(&pwq->delayed_works,
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001130 struct work_struct, entry);
1131
Tejun Heo112202d2013-02-13 19:29:12 -08001132 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001133}
1134
Tejun Heobf4ede02012-08-03 10:30:46 -07001135/**
Tejun Heo112202d2013-02-13 19:29:12 -08001136 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1137 * @pwq: pwq of interest
Tejun Heobf4ede02012-08-03 10:30:46 -07001138 * @color: color of work which left the queue
Tejun Heobf4ede02012-08-03 10:30:46 -07001139 *
1140 * A work either has completed or is removed from pending queue,
Tejun Heo112202d2013-02-13 19:29:12 -08001141 * decrement nr_in_flight of its pwq and handle workqueue flushing.
Tejun Heobf4ede02012-08-03 10:30:46 -07001142 *
1143 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001144 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001145 */
Tejun Heo112202d2013-02-13 19:29:12 -08001146static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
Tejun Heobf4ede02012-08-03 10:30:46 -07001147{
Tejun Heo8864b4e2013-03-12 11:30:04 -07001148 /* uncolored work items don't participate in flushing or nr_active */
Tejun Heobf4ede02012-08-03 10:30:46 -07001149 if (color == WORK_NO_COLOR)
Tejun Heo8864b4e2013-03-12 11:30:04 -07001150 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001151
Tejun Heo112202d2013-02-13 19:29:12 -08001152 pwq->nr_in_flight[color]--;
Tejun Heobf4ede02012-08-03 10:30:46 -07001153
Tejun Heo112202d2013-02-13 19:29:12 -08001154 pwq->nr_active--;
1155 if (!list_empty(&pwq->delayed_works)) {
Lai Jiangshanb3f9f402012-09-18 10:40:00 -07001156 /* one down, submit a delayed one */
Tejun Heo112202d2013-02-13 19:29:12 -08001157 if (pwq->nr_active < pwq->max_active)
1158 pwq_activate_first_delayed(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001159 }
1160
1161 /* is flush in progress and are we at the flushing tip? */
Tejun Heo112202d2013-02-13 19:29:12 -08001162 if (likely(pwq->flush_color != color))
Tejun Heo8864b4e2013-03-12 11:30:04 -07001163 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001164
1165 /* are there still in-flight works? */
Tejun Heo112202d2013-02-13 19:29:12 -08001166 if (pwq->nr_in_flight[color])
Tejun Heo8864b4e2013-03-12 11:30:04 -07001167 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001168
Tejun Heo112202d2013-02-13 19:29:12 -08001169 /* this pwq is done, clear flush_color */
1170 pwq->flush_color = -1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001171
1172 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001173 * If this was the last pwq, wake up the first flusher. It
Tejun Heobf4ede02012-08-03 10:30:46 -07001174 * will handle the rest.
1175 */
Tejun Heo112202d2013-02-13 19:29:12 -08001176 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1177 complete(&pwq->wq->first_flusher->done);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001178out_put:
1179 put_pwq(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001180}
1181
Tejun Heo36e227d2012-08-03 10:30:46 -07001182/**
Tejun Heobbb68df2012-08-03 10:30:46 -07001183 * try_to_grab_pending - steal work item from worklist and disable irq
Tejun Heo36e227d2012-08-03 10:30:46 -07001184 * @work: work item to steal
1185 * @is_dwork: @work is a delayed_work
Tejun Heobbb68df2012-08-03 10:30:46 -07001186 * @flags: place to store irq state
Tejun Heo36e227d2012-08-03 10:30:46 -07001187 *
1188 * Try to grab PENDING bit of @work. This function can handle @work in any
Yacine Belkadid185af32013-07-31 14:59:24 -07001189 * stable state - idle, on timer or on worklist.
Tejun Heo36e227d2012-08-03 10:30:46 -07001190 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001191 * Return:
Tejun Heo36e227d2012-08-03 10:30:46 -07001192 * 1 if @work was pending and we successfully stole PENDING
1193 * 0 if @work was idle and we claimed PENDING
1194 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
Tejun Heobbb68df2012-08-03 10:30:46 -07001195 * -ENOENT if someone else is canceling @work, this state may persist
1196 * for arbitrarily long
Tejun Heo36e227d2012-08-03 10:30:46 -07001197 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001198 * Note:
Tejun Heobbb68df2012-08-03 10:30:46 -07001199 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001200 * interrupted while holding PENDING and @work off queue, irq must be
1201 * disabled on entry. This, combined with delayed_work->timer being
1202 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
Tejun Heobbb68df2012-08-03 10:30:46 -07001203 *
1204 * On successful return, >= 0, irq is disabled and the caller is
1205 * responsible for releasing it using local_irq_restore(*@flags).
1206 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001207 * This function is safe to call from any context including IRQ handler.
Tejun Heobf4ede02012-08-03 10:30:46 -07001208 */
Tejun Heobbb68df2012-08-03 10:30:46 -07001209static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1210 unsigned long *flags)
Tejun Heobf4ede02012-08-03 10:30:46 -07001211{
Tejun Heod565ed62013-01-24 11:01:33 -08001212 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001213 struct pool_workqueue *pwq;
Tejun Heobf4ede02012-08-03 10:30:46 -07001214
Tejun Heobbb68df2012-08-03 10:30:46 -07001215 local_irq_save(*flags);
1216
Tejun Heo36e227d2012-08-03 10:30:46 -07001217 /* try to steal the timer if it exists */
1218 if (is_dwork) {
1219 struct delayed_work *dwork = to_delayed_work(work);
1220
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001221 /*
1222 * dwork->timer is irqsafe. If del_timer() fails, it's
1223 * guaranteed that the timer is not queued anywhere and not
1224 * running on the local CPU.
1225 */
Tejun Heo36e227d2012-08-03 10:30:46 -07001226 if (likely(del_timer(&dwork->timer)))
1227 return 1;
1228 }
1229
1230 /* try to claim PENDING the normal way */
Tejun Heobf4ede02012-08-03 10:30:46 -07001231 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1232 return 0;
1233
1234 /*
1235 * The queueing is in progress, or it is already queued. Try to
1236 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1237 */
Tejun Heod565ed62013-01-24 11:01:33 -08001238 pool = get_work_pool(work);
1239 if (!pool)
Tejun Heobbb68df2012-08-03 10:30:46 -07001240 goto fail;
Tejun Heobf4ede02012-08-03 10:30:46 -07001241
Tejun Heod565ed62013-01-24 11:01:33 -08001242 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001243 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001244 * work->data is guaranteed to point to pwq only while the work
1245 * item is queued on pwq->wq, and both updating work->data to point
1246 * to pwq on queueing and to pool on dequeueing are done under
1247 * pwq->pool->lock. This in turn guarantees that, if work->data
1248 * points to pwq which is associated with a locked pool, the work
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001249 * item is currently queued on that pool.
1250 */
Tejun Heo112202d2013-02-13 19:29:12 -08001251 pwq = get_work_pwq(work);
1252 if (pwq && pwq->pool == pool) {
Tejun Heo16062832013-02-06 18:04:53 -08001253 debug_work_deactivate(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001254
Tejun Heo16062832013-02-06 18:04:53 -08001255 /*
1256 * A delayed work item cannot be grabbed directly because
1257 * it might have linked NO_COLOR work items which, if left
Tejun Heo112202d2013-02-13 19:29:12 -08001258 * on the delayed_list, will confuse pwq->nr_active
Tejun Heo16062832013-02-06 18:04:53 -08001259 * management later on and cause stall. Make sure the work
1260 * item is activated before grabbing.
1261 */
1262 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
Tejun Heo112202d2013-02-13 19:29:12 -08001263 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001264
Tejun Heo16062832013-02-06 18:04:53 -08001265 list_del_init(&work->entry);
Lai Jiangshan9c34a702014-07-11 00:11:13 +08001266 pwq_dec_nr_in_flight(pwq, get_work_color(work));
Tejun Heo36e227d2012-08-03 10:30:46 -07001267
Tejun Heo112202d2013-02-13 19:29:12 -08001268 /* work->data points to pwq iff queued, point to pool */
Tejun Heo16062832013-02-06 18:04:53 -08001269 set_work_pool_and_keep_pending(work, pool->id);
Lai Jiangshan4468a002013-02-06 18:04:53 -08001270
Tejun Heo16062832013-02-06 18:04:53 -08001271 spin_unlock(&pool->lock);
1272 return 1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001273 }
Tejun Heod565ed62013-01-24 11:01:33 -08001274 spin_unlock(&pool->lock);
Tejun Heobbb68df2012-08-03 10:30:46 -07001275fail:
1276 local_irq_restore(*flags);
1277 if (work_is_canceling(work))
1278 return -ENOENT;
1279 cpu_relax();
Tejun Heo36e227d2012-08-03 10:30:46 -07001280 return -EAGAIN;
Tejun Heobf4ede02012-08-03 10:30:46 -07001281}
1282
1283/**
Tejun Heo706026c2013-01-24 11:01:34 -08001284 * insert_work - insert a work into a pool
Tejun Heo112202d2013-02-13 19:29:12 -08001285 * @pwq: pwq @work belongs to
Tejun Heo4690c4a2010-06-29 10:07:10 +02001286 * @work: work to insert
1287 * @head: insertion point
1288 * @extra_flags: extra WORK_STRUCT_* flags to set
1289 *
Tejun Heo112202d2013-02-13 19:29:12 -08001290 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
Tejun Heo706026c2013-01-24 11:01:34 -08001291 * work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +02001292 *
1293 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001294 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02001295 */
Tejun Heo112202d2013-02-13 19:29:12 -08001296static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1297 struct list_head *head, unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001298{
Tejun Heo112202d2013-02-13 19:29:12 -08001299 struct worker_pool *pool = pwq->pool;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01001300
Tejun Heo4690c4a2010-06-29 10:07:10 +02001301 /* we own @work, set data and link */
Tejun Heo112202d2013-02-13 19:29:12 -08001302 set_work_pwq(work, pwq, extra_flags);
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -07001303 list_add_tail(&work->entry, head);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001304 get_pwq(pwq);
Tejun Heoe22bee72010-06-29 10:07:14 +02001305
1306 /*
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001307 * Ensure either wq_worker_sleeping() sees the above
1308 * list_add_tail() or we see zero nr_running to avoid workers lying
1309 * around lazily while there are works to be processed.
Tejun Heoe22bee72010-06-29 10:07:14 +02001310 */
1311 smp_mb();
1312
Tejun Heo63d95a92012-07-12 14:46:37 -07001313 if (__need_more_worker(pool))
1314 wake_up_worker(pool);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001315}
1316
Tejun Heoc8efcc22010-12-20 19:32:04 +01001317/*
1318 * Test whether @work is being queued from another work executing on the
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001319 * same workqueue.
Tejun Heoc8efcc22010-12-20 19:32:04 +01001320 */
1321static bool is_chained_work(struct workqueue_struct *wq)
1322{
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001323 struct worker *worker;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001324
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001325 worker = current_wq_worker();
1326 /*
1327 * Return %true iff I'm a worker execuing a work item on @wq. If
1328 * I'm @worker, it's safe to dereference it without locking.
1329 */
Tejun Heo112202d2013-02-13 19:29:12 -08001330 return worker && worker->current_pwq->wq == wq;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001331}
1332
Mike Galbraithef5571802016-02-09 17:59:38 -05001333/*
1334 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1335 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1336 * avoid perturbing sensitive tasks.
1337 */
1338static int wq_select_unbound_cpu(int cpu)
1339{
Tejun Heof303fcc2016-02-09 17:59:38 -05001340 static bool printed_dbg_warning;
Mike Galbraithef5571802016-02-09 17:59:38 -05001341 int new_cpu;
1342
Tejun Heof303fcc2016-02-09 17:59:38 -05001343 if (likely(!wq_debug_force_rr_cpu)) {
1344 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1345 return cpu;
1346 } else if (!printed_dbg_warning) {
1347 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1348 printed_dbg_warning = true;
1349 }
1350
Mike Galbraithef5571802016-02-09 17:59:38 -05001351 if (cpumask_empty(wq_unbound_cpumask))
1352 return cpu;
1353
1354 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1355 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1356 if (unlikely(new_cpu >= nr_cpu_ids)) {
1357 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1358 if (unlikely(new_cpu >= nr_cpu_ids))
1359 return cpu;
1360 }
1361 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1362
1363 return new_cpu;
1364}
1365
Tejun Heod84ff052013-03-12 11:29:59 -07001366static void __queue_work(int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 struct work_struct *work)
1368{
Tejun Heo112202d2013-02-13 19:29:12 -08001369 struct pool_workqueue *pwq;
Tejun Heoc9178082013-03-12 11:30:04 -07001370 struct worker_pool *last_pool;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001371 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001372 unsigned int work_flags;
Joonsoo Kimb75cac92012-08-15 23:25:37 +09001373 unsigned int req_cpu = cpu;
Tejun Heo8930cab2012-08-03 10:30:45 -07001374
1375 /*
1376 * While a work item is PENDING && off queue, a task trying to
1377 * steal the PENDING will busy-loop waiting for it to either get
1378 * queued or lose PENDING. Grabbing PENDING and queueing should
1379 * happen with IRQ disabled.
1380 */
1381 WARN_ON_ONCE(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001383
Li Bin9ef28a72013-09-09 13:13:58 +08001384 /* if draining, only works from the same workqueue are allowed */
Tejun Heo618b01e2013-03-12 11:30:04 -07001385 if (unlikely(wq->flags & __WQ_DRAINING) &&
Tejun Heoc8efcc22010-12-20 19:32:04 +01001386 WARN_ON_ONCE(!is_chained_work(wq)))
Tejun Heoe41e7042010-08-24 14:22:47 +02001387 return;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001388retry:
Tejun Heoc9178082013-03-12 11:30:04 -07001389 /* pwq which will be used unless @work is executing elsewhere */
Hillf Danton48c33622020-01-24 20:14:45 -05001390 if (wq->flags & WQ_UNBOUND) {
1391 if (req_cpu == WORK_CPU_UNBOUND)
1392 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001393 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Hillf Danton48c33622020-01-24 20:14:45 -05001394 } else {
1395 if (req_cpu == WORK_CPU_UNBOUND)
1396 cpu = raw_smp_processor_id();
1397 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1398 }
Tejun Heodbf25762012-08-20 14:51:23 -07001399
Tejun Heoc9178082013-03-12 11:30:04 -07001400 /*
1401 * If @work was previously on a different pool, it might still be
1402 * running there, in which case the work needs to be queued on that
1403 * pool to guarantee non-reentrancy.
1404 */
1405 last_pool = get_work_pool(work);
1406 if (last_pool && last_pool != pwq->pool) {
1407 struct worker *worker;
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001408
Tejun Heoc9178082013-03-12 11:30:04 -07001409 spin_lock(&last_pool->lock);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001410
Tejun Heoc9178082013-03-12 11:30:04 -07001411 worker = find_worker_executing_work(last_pool, work);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001412
Tejun Heoc9178082013-03-12 11:30:04 -07001413 if (worker && worker->current_pwq->wq == wq) {
1414 pwq = worker->current_pwq;
Tejun Heo8930cab2012-08-03 10:30:45 -07001415 } else {
Tejun Heoc9178082013-03-12 11:30:04 -07001416 /* meh... not running there, queue here */
1417 spin_unlock(&last_pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08001418 spin_lock(&pwq->pool->lock);
Tejun Heo8930cab2012-08-03 10:30:45 -07001419 }
Tejun Heof3421792010-07-02 10:03:51 +02001420 } else {
Tejun Heo112202d2013-02-13 19:29:12 -08001421 spin_lock(&pwq->pool->lock);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001422 }
1423
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001424 /*
1425 * pwq is determined and locked. For unbound pools, we could have
1426 * raced with pwq release and it could already be dead. If its
1427 * refcnt is zero, repeat pwq selection. Note that pwqs never die
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001428 * without another pwq replacing it in the numa_pwq_tbl or while
1429 * work items are executing on it, so the retrying is guaranteed to
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001430 * make forward-progress.
1431 */
1432 if (unlikely(!pwq->refcnt)) {
1433 if (wq->flags & WQ_UNBOUND) {
1434 spin_unlock(&pwq->pool->lock);
1435 cpu_relax();
1436 goto retry;
1437 }
1438 /* oops */
1439 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1440 wq->name, cpu);
1441 }
1442
Tejun Heo112202d2013-02-13 19:29:12 -08001443 /* pwq determined, queue */
1444 trace_workqueue_queue_work(req_cpu, pwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001445
Dan Carpenterf5b25522012-04-13 22:06:58 +03001446 if (WARN_ON(!list_empty(&work->entry))) {
Tejun Heo112202d2013-02-13 19:29:12 -08001447 spin_unlock(&pwq->pool->lock);
Dan Carpenterf5b25522012-04-13 22:06:58 +03001448 return;
1449 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001450
Tejun Heo112202d2013-02-13 19:29:12 -08001451 pwq->nr_in_flight[pwq->work_color]++;
1452 work_flags = work_color_to_flags(pwq->work_color);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001453
Tejun Heo112202d2013-02-13 19:29:12 -08001454 if (likely(pwq->nr_active < pwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001455 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001456 pwq->nr_active++;
1457 worklist = &pwq->pool->worklist;
Tejun Heo82607adc2015-12-08 11:28:04 -05001458 if (list_empty(worklist))
1459 pwq->pool->watchdog_ts = jiffies;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001460 } else {
1461 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo112202d2013-02-13 19:29:12 -08001462 worklist = &pwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001463 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001464
Zqiang9948ff52021-02-18 11:16:49 +08001465 debug_work_activate(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001466 insert_work(pwq, work, worklist, work_flags);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001467
Tejun Heo112202d2013-02-13 19:29:12 -08001468 spin_unlock(&pwq->pool->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469}
1470
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001471/**
Zhang Ruic1a220e2008-07-23 21:28:39 -07001472 * queue_work_on - queue work on specific cpu
1473 * @cpu: CPU number to execute work on
1474 * @wq: workqueue to use
1475 * @work: work to queue
1476 *
Zhang Ruic1a220e2008-07-23 21:28:39 -07001477 * We queue the work to a specific CPU, the caller must ensure it
1478 * can't go away.
Yacine Belkadid185af32013-07-31 14:59:24 -07001479 *
1480 * Return: %false if @work was already on a queue, %true otherwise.
Zhang Ruic1a220e2008-07-23 21:28:39 -07001481 */
Tejun Heod4283e92012-08-03 10:30:44 -07001482bool queue_work_on(int cpu, struct workqueue_struct *wq,
1483 struct work_struct *work)
Zhang Ruic1a220e2008-07-23 21:28:39 -07001484{
Tejun Heod4283e92012-08-03 10:30:44 -07001485 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001486 unsigned long flags;
1487
1488 local_irq_save(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001489
Tejun Heo22df02b2010-06-29 10:07:10 +02001490 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001491 __queue_work(cpu, wq, work);
Tejun Heod4283e92012-08-03 10:30:44 -07001492 ret = true;
Zhang Ruic1a220e2008-07-23 21:28:39 -07001493 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001494
1495 local_irq_restore(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001496 return ret;
1497}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001498EXPORT_SYMBOL(queue_work_on);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001499
Tejun Heod8e794d2012-08-03 10:30:45 -07001500void delayed_work_timer_fn(unsigned long __data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501{
David Howells52bad642006-11-22 14:54:01 +00001502 struct delayed_work *dwork = (struct delayed_work *)__data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001504 /* should have been called from irqsafe timer with irq already off */
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001505 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506}
Konstantin Khlebnikov1438ade52013-01-24 16:36:31 +04001507EXPORT_SYMBOL(delayed_work_timer_fn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001509static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1510 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511{
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001512 struct timer_list *timer = &dwork->timer;
1513 struct work_struct *work = &dwork->work;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514
Tejun Heo637fdba2017-03-06 15:33:42 -05001515 WARN_ON_ONCE(!wq);
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001516 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1517 timer->data != (unsigned long)dwork);
Tejun Heofc4b5142012-12-04 07:40:39 -08001518 WARN_ON_ONCE(timer_pending(timer));
1519 WARN_ON_ONCE(!list_empty(&work->entry));
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001520
Tejun Heo8852aac2012-12-01 16:23:42 -08001521 /*
1522 * If @delay is 0, queue @dwork->work immediately. This is for
1523 * both optimization and correctness. The earliest @timer can
1524 * expire is on the closest next tick and delayed_work users depend
1525 * on that there's no such delay when @delay is 0.
1526 */
1527 if (!delay) {
1528 __queue_work(cpu, wq, &dwork->work);
1529 return;
1530 }
1531
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001532 dwork->wq = wq;
Tejun Heo12650572012-08-08 09:38:42 -07001533 dwork->cpu = cpu;
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001534 timer->expires = jiffies + delay;
1535
Tejun Heo041bd122016-02-09 16:11:26 -05001536 if (unlikely(cpu != WORK_CPU_UNBOUND))
1537 add_timer_on(timer, cpu);
1538 else
1539 add_timer(timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540}
1541
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001542/**
1543 * queue_delayed_work_on - queue work on specific CPU after delay
1544 * @cpu: CPU number to execute work on
1545 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001546 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001547 * @delay: number of jiffies to wait before queueing
1548 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001549 * Return: %false if @work was already on a queue, %true otherwise. If
Tejun Heo715f1302012-08-03 10:30:46 -07001550 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1551 * execution.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001552 */
Tejun Heod4283e92012-08-03 10:30:44 -07001553bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1554 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001555{
David Howells52bad642006-11-22 14:54:01 +00001556 struct work_struct *work = &dwork->work;
Tejun Heod4283e92012-08-03 10:30:44 -07001557 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001558 unsigned long flags;
1559
1560 /* read the comment in __queue_work() */
1561 local_irq_save(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001562
Tejun Heo22df02b2010-06-29 10:07:10 +02001563 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001564 __queue_delayed_work(cpu, wq, dwork, delay);
Tejun Heod4283e92012-08-03 10:30:44 -07001565 ret = true;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001566 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001567
1568 local_irq_restore(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001569 return ret;
1570}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001571EXPORT_SYMBOL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572
Tejun Heoc8e55f32010-06-29 10:07:12 +02001573/**
Tejun Heo8376fe22012-08-03 10:30:47 -07001574 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1575 * @cpu: CPU number to execute work on
1576 * @wq: workqueue to use
1577 * @dwork: work to queue
1578 * @delay: number of jiffies to wait before queueing
1579 *
1580 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1581 * modify @dwork's timer so that it expires after @delay. If @delay is
1582 * zero, @work is guaranteed to be scheduled immediately regardless of its
1583 * current state.
1584 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001585 * Return: %false if @dwork was idle and queued, %true if @dwork was
Tejun Heo8376fe22012-08-03 10:30:47 -07001586 * pending and its timer was modified.
1587 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001588 * This function is safe to call from any context including IRQ handler.
Tejun Heo8376fe22012-08-03 10:30:47 -07001589 * See try_to_grab_pending() for details.
1590 */
1591bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1592 struct delayed_work *dwork, unsigned long delay)
1593{
1594 unsigned long flags;
1595 int ret;
1596
1597 do {
1598 ret = try_to_grab_pending(&dwork->work, true, &flags);
1599 } while (unlikely(ret == -EAGAIN));
1600
1601 if (likely(ret >= 0)) {
1602 __queue_delayed_work(cpu, wq, dwork, delay);
1603 local_irq_restore(flags);
1604 }
1605
1606 /* -ENOENT from try_to_grab_pending() becomes %true */
1607 return ret;
1608}
1609EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1610
1611/**
Tejun Heoc8e55f32010-06-29 10:07:12 +02001612 * worker_enter_idle - enter idle state
1613 * @worker: worker which is entering idle state
1614 *
1615 * @worker is entering idle state. Update stats and idle timer if
1616 * necessary.
1617 *
1618 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001619 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001620 */
1621static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001623 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624
Tejun Heo6183c002013-03-12 11:29:57 -07001625 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1626 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1627 (worker->hentry.next || worker->hentry.pprev)))
1628 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629
Lai Jiangshan051e1852014-07-22 13:03:02 +08001630 /* can't use worker_set_flags(), also called from create_worker() */
Tejun Heocb444762010-07-02 10:03:50 +02001631 worker->flags |= WORKER_IDLE;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001632 pool->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001633 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001634
Tejun Heoc8e55f32010-06-29 10:07:12 +02001635 /* idle_list is LIFO */
Tejun Heobd7bdd42012-07-12 14:46:37 -07001636 list_add(&worker->entry, &pool->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001637
Tejun Heo628c78e2012-07-17 12:39:27 -07001638 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1639 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
Tejun Heocb444762010-07-02 10:03:50 +02001640
Tejun Heo544ecf32012-05-14 15:04:50 -07001641 /*
Tejun Heo706026c2013-01-24 11:01:34 -08001642 * Sanity check nr_running. Because wq_unbind_fn() releases
Tejun Heod565ed62013-01-24 11:01:33 -08001643 * pool->lock between setting %WORKER_UNBOUND and zapping
Tejun Heo628c78e2012-07-17 12:39:27 -07001644 * nr_running, the warning may trigger spuriously. Check iff
1645 * unbind is not in progress.
Tejun Heo544ecf32012-05-14 15:04:50 -07001646 */
Tejun Heo24647572013-01-24 11:01:33 -08001647 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -07001648 pool->nr_workers == pool->nr_idle &&
Tejun Heoe19e3972013-01-24 11:39:44 -08001649 atomic_read(&pool->nr_running));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650}
1651
Tejun Heoc8e55f32010-06-29 10:07:12 +02001652/**
1653 * worker_leave_idle - leave idle state
1654 * @worker: worker which is leaving idle state
1655 *
1656 * @worker is leaving idle state. Update stats.
1657 *
1658 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001659 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001660 */
1661static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001663 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664
Tejun Heo6183c002013-03-12 11:29:57 -07001665 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1666 return;
Tejun Heod302f012010-06-29 10:07:13 +02001667 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001668 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001669 list_del_init(&worker->entry);
1670}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001672static struct worker *alloc_worker(int node)
Tejun Heoc34056a2010-06-29 10:07:11 +02001673{
1674 struct worker *worker;
1675
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001676 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001677 if (worker) {
1678 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001679 INIT_LIST_HEAD(&worker->scheduled);
Lai Jiangshanda028462014-05-20 17:46:31 +08001680 INIT_LIST_HEAD(&worker->node);
Tejun Heoe22bee72010-06-29 10:07:14 +02001681 /* on creation a worker is in !idle && prep state */
1682 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001683 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001684 return worker;
1685}
1686
1687/**
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001688 * worker_attach_to_pool() - attach a worker to a pool
1689 * @worker: worker to be attached
1690 * @pool: the target pool
1691 *
1692 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1693 * cpu-binding of @worker are kept coordinated with the pool across
1694 * cpu-[un]hotplugs.
1695 */
1696static void worker_attach_to_pool(struct worker *worker,
1697 struct worker_pool *pool)
1698{
1699 mutex_lock(&pool->attach_mutex);
1700
1701 /*
1702 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1703 * online CPUs. It'll be re-applied when any of the CPUs come up.
1704 */
1705 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1706
1707 /*
1708 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1709 * stable across this function. See the comments above the
1710 * flag definition for details.
1711 */
1712 if (pool->flags & POOL_DISASSOCIATED)
1713 worker->flags |= WORKER_UNBOUND;
1714
1715 list_add_tail(&worker->node, &pool->workers);
1716
1717 mutex_unlock(&pool->attach_mutex);
1718}
1719
1720/**
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001721 * worker_detach_from_pool() - detach a worker from its pool
1722 * @worker: worker which is attached to its pool
1723 * @pool: the pool @worker is attached to
1724 *
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001725 * Undo the attaching which had been done in worker_attach_to_pool(). The
1726 * caller worker shouldn't access to the pool after detached except it has
1727 * other reference to the pool.
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001728 */
1729static void worker_detach_from_pool(struct worker *worker,
1730 struct worker_pool *pool)
1731{
1732 struct completion *detach_completion = NULL;
1733
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001734 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08001735 list_del(&worker->node);
1736 if (list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001737 detach_completion = pool->detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001738 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001739
Lai Jiangshanb62c0752014-06-03 15:32:52 +08001740 /* clear leftover flags without pool->lock after it is detached */
1741 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1742
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001743 if (detach_completion)
1744 complete(detach_completion);
1745}
1746
1747/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001748 * create_worker - create a new workqueue worker
Tejun Heo63d95a92012-07-12 14:46:37 -07001749 * @pool: pool the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001750 *
Lai Jiangshan051e1852014-07-22 13:03:02 +08001751 * Create and start a new worker which is attached to @pool.
Tejun Heoc34056a2010-06-29 10:07:11 +02001752 *
1753 * CONTEXT:
1754 * Might sleep. Does GFP_KERNEL allocations.
1755 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001756 * Return:
Tejun Heoc34056a2010-06-29 10:07:11 +02001757 * Pointer to the newly created worker.
1758 */
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001759static struct worker *create_worker(struct worker_pool *pool)
Tejun Heoc34056a2010-06-29 10:07:11 +02001760{
Tejun Heoc34056a2010-06-29 10:07:11 +02001761 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001762 int id = -1;
Tejun Heoe3c916a2013-04-01 11:23:32 -07001763 char id_buf[16];
Tejun Heoc34056a2010-06-29 10:07:11 +02001764
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001765 /* ID is needed to determine kthread name */
1766 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
Tejun Heo822d8402013-03-19 13:45:21 -07001767 if (id < 0)
1768 goto fail;
Tejun Heoc34056a2010-06-29 10:07:11 +02001769
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001770 worker = alloc_worker(pool->node);
Tejun Heoc34056a2010-06-29 10:07:11 +02001771 if (!worker)
1772 goto fail;
1773
Tejun Heobd7bdd42012-07-12 14:46:37 -07001774 worker->pool = pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001775 worker->id = id;
1776
Tejun Heo29c91e92013-03-12 11:30:03 -07001777 if (pool->cpu >= 0)
Tejun Heoe3c916a2013-04-01 11:23:32 -07001778 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1779 pool->attrs->nice < 0 ? "H" : "");
Tejun Heof3421792010-07-02 10:03:51 +02001780 else
Tejun Heoe3c916a2013-04-01 11:23:32 -07001781 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1782
Tejun Heof3f90ad2013-04-01 11:23:34 -07001783 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
Tejun Heoe3c916a2013-04-01 11:23:32 -07001784 "kworker/%s", id_buf);
Tejun Heoc34056a2010-06-29 10:07:11 +02001785 if (IS_ERR(worker->task))
1786 goto fail;
1787
Oleg Nesterov91151222013-11-14 12:56:18 +01001788 set_user_nice(worker->task, pool->attrs->nice);
Peter Zijlstra25834c72015-05-15 17:43:34 +02001789 kthread_bind_mask(worker->task, pool->attrs->cpumask);
Oleg Nesterov91151222013-11-14 12:56:18 +01001790
Lai Jiangshanda028462014-05-20 17:46:31 +08001791 /* successful, attach the worker to the pool */
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001792 worker_attach_to_pool(worker, pool);
Tejun Heo822d8402013-03-19 13:45:21 -07001793
Lai Jiangshan051e1852014-07-22 13:03:02 +08001794 /* start the newly created worker */
1795 spin_lock_irq(&pool->lock);
1796 worker->pool->nr_workers++;
1797 worker_enter_idle(worker);
1798 wake_up_process(worker->task);
1799 spin_unlock_irq(&pool->lock);
1800
Tejun Heoc34056a2010-06-29 10:07:11 +02001801 return worker;
Tejun Heo822d8402013-03-19 13:45:21 -07001802
Tejun Heoc34056a2010-06-29 10:07:11 +02001803fail:
Lai Jiangshan9625ab12014-05-20 17:46:27 +08001804 if (id >= 0)
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001805 ida_simple_remove(&pool->worker_ida, id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001806 kfree(worker);
1807 return NULL;
1808}
1809
1810/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001811 * destroy_worker - destroy a workqueue worker
1812 * @worker: worker to be destroyed
1813 *
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001814 * Destroy @worker and adjust @pool stats accordingly. The worker should
1815 * be idle.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001816 *
1817 * CONTEXT:
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001818 * spin_lock_irq(pool->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001819 */
1820static void destroy_worker(struct worker *worker)
1821{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001822 struct worker_pool *pool = worker->pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001823
Tejun Heocd549682013-03-13 19:47:39 -07001824 lockdep_assert_held(&pool->lock);
1825
Tejun Heoc34056a2010-06-29 10:07:11 +02001826 /* sanity check frenzy */
Tejun Heo6183c002013-03-12 11:29:57 -07001827 if (WARN_ON(worker->current_work) ||
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001828 WARN_ON(!list_empty(&worker->scheduled)) ||
1829 WARN_ON(!(worker->flags & WORKER_IDLE)))
Tejun Heo6183c002013-03-12 11:29:57 -07001830 return;
Tejun Heoc34056a2010-06-29 10:07:11 +02001831
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001832 pool->nr_workers--;
1833 pool->nr_idle--;
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001834
Tejun Heoc8e55f32010-06-29 10:07:12 +02001835 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001836 worker->flags |= WORKER_DIE;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001837 wake_up_process(worker->task);
Tejun Heoc34056a2010-06-29 10:07:11 +02001838}
1839
Tejun Heo63d95a92012-07-12 14:46:37 -07001840static void idle_worker_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001841{
Tejun Heo63d95a92012-07-12 14:46:37 -07001842 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001843
Tejun Heod565ed62013-01-24 11:01:33 -08001844 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001845
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001846 while (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001847 struct worker *worker;
1848 unsigned long expires;
1849
1850 /* idle_list is kept in LIFO order, check the last one */
Tejun Heo63d95a92012-07-12 14:46:37 -07001851 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02001852 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1853
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001854 if (time_before(jiffies, expires)) {
Tejun Heo63d95a92012-07-12 14:46:37 -07001855 mod_timer(&pool->idle_timer, expires);
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001856 break;
Tejun Heoe22bee72010-06-29 10:07:14 +02001857 }
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001858
1859 destroy_worker(worker);
Tejun Heoe22bee72010-06-29 10:07:14 +02001860 }
1861
Tejun Heod565ed62013-01-24 11:01:33 -08001862 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001863}
1864
Tejun Heo493a1722013-03-12 11:29:59 -07001865static void send_mayday(struct work_struct *work)
Tejun Heoe22bee72010-06-29 10:07:14 +02001866{
Tejun Heo112202d2013-02-13 19:29:12 -08001867 struct pool_workqueue *pwq = get_work_pwq(work);
1868 struct workqueue_struct *wq = pwq->wq;
Tejun Heo493a1722013-03-12 11:29:59 -07001869
Tejun Heo2e109a22013-03-13 19:47:40 -07001870 lockdep_assert_held(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001871
Tejun Heo493008a2013-03-12 11:30:03 -07001872 if (!wq->rescuer)
Tejun Heo493a1722013-03-12 11:29:59 -07001873 return;
Tejun Heoe22bee72010-06-29 10:07:14 +02001874
1875 /* mayday mayday mayday */
Tejun Heo493a1722013-03-12 11:29:59 -07001876 if (list_empty(&pwq->mayday_node)) {
Lai Jiangshan77668c82014-04-18 11:04:16 -04001877 /*
1878 * If @pwq is for an unbound wq, its base ref may be put at
1879 * any time due to an attribute change. Pin @pwq until the
1880 * rescuer is done with it.
1881 */
1882 get_pwq(pwq);
Tejun Heo493a1722013-03-12 11:29:59 -07001883 list_add_tail(&pwq->mayday_node, &wq->maydays);
Tejun Heoe22bee72010-06-29 10:07:14 +02001884 wake_up_process(wq->rescuer->task);
Tejun Heo493a1722013-03-12 11:29:59 -07001885 }
Tejun Heoe22bee72010-06-29 10:07:14 +02001886}
1887
Tejun Heo706026c2013-01-24 11:01:34 -08001888static void pool_mayday_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001889{
Tejun Heo63d95a92012-07-12 14:46:37 -07001890 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001891 struct work_struct *work;
1892
Tejun Heob2d82902014-12-08 12:39:16 -05001893 spin_lock_irq(&pool->lock);
1894 spin_lock(&wq_mayday_lock); /* for wq->maydays */
Tejun Heoe22bee72010-06-29 10:07:14 +02001895
Tejun Heo63d95a92012-07-12 14:46:37 -07001896 if (need_to_create_worker(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001897 /*
1898 * We've been trying to create a new worker but
1899 * haven't been successful. We might be hitting an
1900 * allocation deadlock. Send distress signals to
1901 * rescuers.
1902 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001903 list_for_each_entry(work, &pool->worklist, entry)
Tejun Heoe22bee72010-06-29 10:07:14 +02001904 send_mayday(work);
1905 }
1906
Tejun Heob2d82902014-12-08 12:39:16 -05001907 spin_unlock(&wq_mayday_lock);
1908 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001909
Tejun Heo63d95a92012-07-12 14:46:37 -07001910 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
Tejun Heoe22bee72010-06-29 10:07:14 +02001911}
1912
1913/**
1914 * maybe_create_worker - create a new worker if necessary
Tejun Heo63d95a92012-07-12 14:46:37 -07001915 * @pool: pool to create a new worker for
Tejun Heoe22bee72010-06-29 10:07:14 +02001916 *
Tejun Heo63d95a92012-07-12 14:46:37 -07001917 * Create a new worker for @pool if necessary. @pool is guaranteed to
Tejun Heoe22bee72010-06-29 10:07:14 +02001918 * have at least one idle worker on return from this function. If
1919 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
Tejun Heo63d95a92012-07-12 14:46:37 -07001920 * sent to all rescuers with works scheduled on @pool to resolve
Tejun Heoe22bee72010-06-29 10:07:14 +02001921 * possible allocation deadlock.
1922 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001923 * On return, need_to_create_worker() is guaranteed to be %false and
1924 * may_start_working() %true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001925 *
1926 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001927 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001928 * multiple times. Does GFP_KERNEL allocations. Called only from
1929 * manager.
Tejun Heoe22bee72010-06-29 10:07:14 +02001930 */
Tejun Heo29187a92015-01-16 14:21:16 -05001931static void maybe_create_worker(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08001932__releases(&pool->lock)
1933__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001934{
Tejun Heoe22bee72010-06-29 10:07:14 +02001935restart:
Tejun Heod565ed62013-01-24 11:01:33 -08001936 spin_unlock_irq(&pool->lock);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001937
Tejun Heoe22bee72010-06-29 10:07:14 +02001938 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
Tejun Heo63d95a92012-07-12 14:46:37 -07001939 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
Tejun Heoe22bee72010-06-29 10:07:14 +02001940
1941 while (true) {
Lai Jiangshan051e1852014-07-22 13:03:02 +08001942 if (create_worker(pool) || !need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001943 break;
1944
Lai Jiangshane212f362014-06-03 15:32:17 +08001945 schedule_timeout_interruptible(CREATE_COOLDOWN);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001946
Tejun Heo63d95a92012-07-12 14:46:37 -07001947 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001948 break;
1949 }
1950
Tejun Heo63d95a92012-07-12 14:46:37 -07001951 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08001952 spin_lock_irq(&pool->lock);
Lai Jiangshan051e1852014-07-22 13:03:02 +08001953 /*
1954 * This is necessary even after a new worker was just successfully
1955 * created as @pool->lock was dropped and the new worker might have
1956 * already become busy.
1957 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001958 if (need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001959 goto restart;
Tejun Heoe22bee72010-06-29 10:07:14 +02001960}
1961
1962/**
Tejun Heoe22bee72010-06-29 10:07:14 +02001963 * manage_workers - manage worker pool
1964 * @worker: self
1965 *
Tejun Heo706026c2013-01-24 11:01:34 -08001966 * Assume the manager role and manage the worker pool @worker belongs
Tejun Heoe22bee72010-06-29 10:07:14 +02001967 * to. At any given time, there can be only zero or one manager per
Tejun Heo706026c2013-01-24 11:01:34 -08001968 * pool. The exclusion is handled automatically by this function.
Tejun Heoe22bee72010-06-29 10:07:14 +02001969 *
1970 * The caller can safely start processing works on false return. On
1971 * true return, it's guaranteed that need_to_create_worker() is false
1972 * and may_start_working() is true.
1973 *
1974 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001975 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001976 * multiple times. Does GFP_KERNEL allocations.
1977 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001978 * Return:
Tejun Heo29187a92015-01-16 14:21:16 -05001979 * %false if the pool doesn't need management and the caller can safely
1980 * start processing works, %true if management function was performed and
1981 * the conditions that the caller verified before calling the function may
1982 * no longer be true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001983 */
1984static bool manage_workers(struct worker *worker)
1985{
Tejun Heo63d95a92012-07-12 14:46:37 -07001986 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001987
Tejun Heo692b4822017-10-09 08:04:13 -07001988 if (pool->flags & POOL_MANAGER_ACTIVE)
Tejun Heo29187a92015-01-16 14:21:16 -05001989 return false;
Tejun Heo692b4822017-10-09 08:04:13 -07001990
1991 pool->flags |= POOL_MANAGER_ACTIVE;
Tejun Heo2607d7a2015-03-09 09:22:28 -04001992 pool->manager = worker;
Tejun Heoe22bee72010-06-29 10:07:14 +02001993
Tejun Heo29187a92015-01-16 14:21:16 -05001994 maybe_create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02001995
Tejun Heo2607d7a2015-03-09 09:22:28 -04001996 pool->manager = NULL;
Tejun Heo692b4822017-10-09 08:04:13 -07001997 pool->flags &= ~POOL_MANAGER_ACTIVE;
1998 wake_up(&wq_manager_wait);
Tejun Heo29187a92015-01-16 14:21:16 -05001999 return true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002000}
2001
Tejun Heoa62428c2010-06-29 10:07:10 +02002002/**
2003 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02002004 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02002005 * @work: work to process
2006 *
2007 * Process @work. This function contains all the logics necessary to
2008 * process a single work including synchronization against and
2009 * interaction with other workers on the same cpu, queueing and
2010 * flushing. As long as context requirement is met, any worker can
2011 * call this function to process a work.
2012 *
2013 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002014 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02002015 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002016static void process_one_work(struct worker *worker, struct work_struct *work)
Tejun Heod565ed62013-01-24 11:01:33 -08002017__releases(&pool->lock)
2018__acquires(&pool->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02002019{
Tejun Heo112202d2013-02-13 19:29:12 -08002020 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobd7bdd42012-07-12 14:46:37 -07002021 struct worker_pool *pool = worker->pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002022 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heo73f53c42010-06-29 10:07:11 +02002023 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02002024 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02002025#ifdef CONFIG_LOCKDEP
2026 /*
2027 * It is permissible to free the struct work_struct from
2028 * inside the function that is called from it, this we need to
2029 * take into account for lockdep too. To avoid bogus "held
2030 * lock freed" warnings as well as problems when looking into
2031 * work->lockdep_map, make a copy and use that here.
2032 */
Peter Zijlstra4d82a1d2012-05-15 08:06:19 -07002033 struct lockdep_map lockdep_map;
2034
2035 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002036#endif
Lai Jiangshan807407c2014-06-03 15:33:28 +08002037 /* ensure we're on the correct CPU */
Lai Jiangshan85327af2014-06-03 15:33:28 +08002038 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heoec22ca52013-01-24 11:01:33 -08002039 raw_smp_processor_id() != pool->cpu);
Tejun Heo25511a42012-07-17 12:39:27 -07002040
Tejun Heo7e116292010-06-29 10:07:13 +02002041 /*
2042 * A single work shouldn't be executed concurrently by
2043 * multiple workers on a single cpu. Check whether anyone is
2044 * already processing the work. If so, defer the work to the
2045 * currently executing one.
2046 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002047 collision = find_worker_executing_work(pool, work);
Tejun Heo7e116292010-06-29 10:07:13 +02002048 if (unlikely(collision)) {
2049 move_linked_works(work, &collision->scheduled, NULL);
2050 return;
2051 }
2052
Tejun Heo8930cab2012-08-03 10:30:45 -07002053 /* claim and dequeue */
Tejun Heoa62428c2010-06-29 10:07:10 +02002054 debug_work_deactivate(work);
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002055 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
Tejun Heoc34056a2010-06-29 10:07:11 +02002056 worker->current_work = work;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002057 worker->current_func = work->func;
Tejun Heo112202d2013-02-13 19:29:12 -08002058 worker->current_pwq = pwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002059 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002060
Tejun Heoa62428c2010-06-29 10:07:10 +02002061 list_del_init(&work->entry);
2062
Tejun Heo649027d2010-06-29 10:07:14 +02002063 /*
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002064 * CPU intensive works don't participate in concurrency management.
2065 * They're the scheduler's responsibility. This takes @worker out
2066 * of concurrency management and the next code block will chain
2067 * execution of the pending work items.
Tejun Heofb0e7be2010-06-29 10:07:15 +02002068 */
2069 if (unlikely(cpu_intensive))
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002070 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
Tejun Heofb0e7be2010-06-29 10:07:15 +02002071
Tejun Heo974271c2012-07-12 14:46:37 -07002072 /*
Lai Jiangshana489a032014-07-22 13:01:59 +08002073 * Wake up another worker if necessary. The condition is always
2074 * false for normal per-cpu workers since nr_running would always
2075 * be >= 1 at this point. This is used to chain execution of the
2076 * pending work items for WORKER_NOT_RUNNING workers such as the
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002077 * UNBOUND and CPU_INTENSIVE ones.
Tejun Heo974271c2012-07-12 14:46:37 -07002078 */
Lai Jiangshana489a032014-07-22 13:01:59 +08002079 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002080 wake_up_worker(pool);
Tejun Heo974271c2012-07-12 14:46:37 -07002081
Tejun Heo8930cab2012-08-03 10:30:45 -07002082 /*
Tejun Heo7c3eed52013-01-24 11:01:33 -08002083 * Record the last pool and clear PENDING which should be the last
Tejun Heod565ed62013-01-24 11:01:33 -08002084 * update to @work. Also, do this inside @pool->lock so that
Tejun Heo23657bb2012-08-13 17:08:19 -07002085 * PENDING and queued state changes happen together while IRQ is
2086 * disabled.
Tejun Heo8930cab2012-08-03 10:30:45 -07002087 */
Tejun Heo7c3eed52013-01-24 11:01:33 -08002088 set_work_pool_and_clear_pending(work, pool->id);
Tejun Heoa62428c2010-06-29 10:07:10 +02002089
Tejun Heod565ed62013-01-24 11:01:33 -08002090 spin_unlock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002091
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002092 lock_map_acquire(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002093 lock_map_acquire(&lockdep_map);
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002094 /*
Peter Zijlstraf52be572017-08-29 10:59:39 +02002095 * Strictly speaking we should mark the invariant state without holding
2096 * any locks, that is, before these two lock_map_acquire()'s.
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002097 *
2098 * However, that would result in:
2099 *
2100 * A(W1)
2101 * WFC(C)
2102 * A(W1)
2103 * C(C)
2104 *
2105 * Which would create W1->C->W1 dependencies, even though there is no
2106 * actual deadlock possible. There are two solutions, using a
2107 * read-recursive acquire on the work(queue) 'locks', but this will then
Peter Zijlstraf52be572017-08-29 10:59:39 +02002108 * hit the lockdep limitation on recursive locks, or simply discard
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002109 * these locks.
2110 *
2111 * AFAICT there is no possible deadlock scenario between the
2112 * flush_work() and complete() primitives (except for single-threaded
2113 * workqueues), so hiding them isn't a problem.
2114 */
Peter Zijlstraf52be572017-08-29 10:59:39 +02002115 lockdep_invariant_state(true);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002116 trace_workqueue_execute_start(work);
Tejun Heoa2c1c572012-12-18 10:35:02 -08002117 worker->current_func(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002118 /*
2119 * While we must be careful to not use "work" after this, the trace
2120 * point will only record its address.
2121 */
2122 trace_workqueue_execute_end(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02002123 lock_map_release(&lockdep_map);
Tejun Heo112202d2013-02-13 19:29:12 -08002124 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002125
2126 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
Valentin Ilie044c7822012-08-19 00:52:42 +03002127 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2128 " last function: %pf\n",
Tejun Heoa2c1c572012-12-18 10:35:02 -08002129 current->comm, preempt_count(), task_pid_nr(current),
2130 worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002131 debug_show_held_locks(current);
2132 dump_stack();
2133 }
2134
Tejun Heob22ce272013-08-28 17:33:37 -04002135 /*
2136 * The following prevents a kworker from hogging CPU on !PREEMPT
2137 * kernels, where a requeueing work item waiting for something to
2138 * happen could deadlock with stop_machine as such work item could
2139 * indefinitely requeue itself while all other CPUs are trapped in
Joe Lawrence789cbbe2014-10-05 13:24:21 -04002140 * stop_machine. At the same time, report a quiescent RCU state so
2141 * the same condition doesn't freeze RCU.
Tejun Heob22ce272013-08-28 17:33:37 -04002142 */
Joe Lawrence3e28e372014-10-05 13:24:22 -04002143 cond_resched_rcu_qs();
Tejun Heob22ce272013-08-28 17:33:37 -04002144
Tejun Heod565ed62013-01-24 11:01:33 -08002145 spin_lock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002146
Tejun Heofb0e7be2010-06-29 10:07:15 +02002147 /* clear cpu intensive status */
2148 if (unlikely(cpu_intensive))
2149 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2150
Tejun Heoa62428c2010-06-29 10:07:10 +02002151 /* we're done with it, release */
Sasha Levin42f85702012-12-17 10:01:23 -05002152 hash_del(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002153 worker->current_work = NULL;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002154 worker->current_func = NULL;
Tejun Heo112202d2013-02-13 19:29:12 -08002155 worker->current_pwq = NULL;
Tejun Heo3d1cb202013-04-30 15:27:22 -07002156 worker->desc_valid = false;
Tejun Heo112202d2013-02-13 19:29:12 -08002157 pwq_dec_nr_in_flight(pwq, work_color);
Tejun Heoa62428c2010-06-29 10:07:10 +02002158}
2159
Tejun Heoaffee4b2010-06-29 10:07:12 +02002160/**
2161 * process_scheduled_works - process scheduled works
2162 * @worker: self
2163 *
2164 * Process all scheduled works. Please note that the scheduled list
2165 * may change while processing a work, so this function repeatedly
2166 * fetches a work from the top and executes it.
2167 *
2168 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002169 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02002170 * multiple times.
2171 */
2172static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002174 while (!list_empty(&worker->scheduled)) {
2175 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002177 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179}
2180
Tejun Heo4690c4a2010-06-29 10:07:10 +02002181/**
2182 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02002183 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02002184 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002185 * The worker thread function. All workers belong to a worker_pool -
2186 * either a per-cpu one or dynamic unbound one. These workers process all
2187 * work items regardless of their specific target workqueue. The only
2188 * exception is work items which belong to workqueues with a rescuer which
2189 * will be explained in rescuer_thread().
Yacine Belkadid185af32013-07-31 14:59:24 -07002190 *
2191 * Return: 0
Tejun Heo4690c4a2010-06-29 10:07:10 +02002192 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002193static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194{
Tejun Heoc34056a2010-06-29 10:07:11 +02002195 struct worker *worker = __worker;
Tejun Heobd7bdd42012-07-12 14:46:37 -07002196 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197
Tejun Heoe22bee72010-06-29 10:07:14 +02002198 /* tell the scheduler that this is a workqueue worker */
2199 worker->task->flags |= PF_WQ_WORKER;
Tejun Heoc8e55f32010-06-29 10:07:12 +02002200woke_up:
Tejun Heod565ed62013-01-24 11:01:33 -08002201 spin_lock_irq(&pool->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002202
Tejun Heoa9ab7752013-03-19 13:45:21 -07002203 /* am I supposed to die? */
2204 if (unlikely(worker->flags & WORKER_DIE)) {
Tejun Heod565ed62013-01-24 11:01:33 -08002205 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002206 WARN_ON_ONCE(!list_empty(&worker->entry));
2207 worker->task->flags &= ~PF_WQ_WORKER;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002208
2209 set_task_comm(worker->task, "kworker/dying");
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08002210 ida_simple_remove(&pool->worker_ida, worker->id);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002211 worker_detach_from_pool(worker, pool);
2212 kfree(worker);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002213 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07002215
Tejun Heoc8e55f32010-06-29 10:07:12 +02002216 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02002217recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02002218 /* no more worker necessary? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002219 if (!need_more_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002220 goto sleep;
2221
2222 /* do we need to manage? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002223 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002224 goto recheck;
2225
Tejun Heoc8e55f32010-06-29 10:07:12 +02002226 /*
2227 * ->scheduled list can only be filled while a worker is
2228 * preparing to process a work or actually processing it.
2229 * Make sure nobody diddled with it while I was sleeping.
2230 */
Tejun Heo6183c002013-03-12 11:29:57 -07002231 WARN_ON_ONCE(!list_empty(&worker->scheduled));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002232
Tejun Heoe22bee72010-06-29 10:07:14 +02002233 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07002234 * Finish PREP stage. We're guaranteed to have at least one idle
2235 * worker or that someone else has already assumed the manager
2236 * role. This is where @worker starts participating in concurrency
2237 * management if applicable and concurrency management is restored
2238 * after being rebound. See rebind_workers() for details.
Tejun Heoe22bee72010-06-29 10:07:14 +02002239 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07002240 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
Tejun Heoe22bee72010-06-29 10:07:14 +02002241
2242 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02002243 struct work_struct *work =
Tejun Heobd7bdd42012-07-12 14:46:37 -07002244 list_first_entry(&pool->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02002245 struct work_struct, entry);
2246
Tejun Heo82607adc2015-12-08 11:28:04 -05002247 pool->watchdog_ts = jiffies;
2248
Tejun Heoc8e55f32010-06-29 10:07:12 +02002249 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2250 /* optimization path, not strictly necessary */
2251 process_one_work(worker, work);
2252 if (unlikely(!list_empty(&worker->scheduled)))
2253 process_scheduled_works(worker);
2254 } else {
2255 move_linked_works(work, &worker->scheduled, NULL);
2256 process_scheduled_works(worker);
2257 }
Tejun Heo63d95a92012-07-12 14:46:37 -07002258 } while (keep_working(pool));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002259
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002260 worker_set_flags(worker, WORKER_PREP);
Tejun Heod313dd82010-07-02 10:03:51 +02002261sleep:
Tejun Heoc8e55f32010-06-29 10:07:12 +02002262 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002263 * pool->lock is held and there's no work to process and no need to
2264 * manage, sleep. Workers are woken up only while holding
2265 * pool->lock or from local cpu, so setting the current state
2266 * before releasing pool->lock is enough to prevent losing any
2267 * event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02002268 */
2269 worker_enter_idle(worker);
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002270 __set_current_state(TASK_IDLE);
Tejun Heod565ed62013-01-24 11:01:33 -08002271 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002272 schedule();
2273 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274}
2275
Tejun Heoe22bee72010-06-29 10:07:14 +02002276/**
2277 * rescuer_thread - the rescuer thread function
Tejun Heo111c2252013-01-17 17:16:24 -08002278 * @__rescuer: self
Tejun Heoe22bee72010-06-29 10:07:14 +02002279 *
2280 * Workqueue rescuer thread function. There's one rescuer for each
Tejun Heo493008a2013-03-12 11:30:03 -07002281 * workqueue which has WQ_MEM_RECLAIM set.
Tejun Heoe22bee72010-06-29 10:07:14 +02002282 *
Tejun Heo706026c2013-01-24 11:01:34 -08002283 * Regular work processing on a pool may block trying to create a new
Tejun Heoe22bee72010-06-29 10:07:14 +02002284 * worker which uses GFP_KERNEL allocation which has slight chance of
2285 * developing into deadlock if some works currently on the same queue
2286 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2287 * the problem rescuer solves.
2288 *
Tejun Heo706026c2013-01-24 11:01:34 -08002289 * When such condition is possible, the pool summons rescuers of all
2290 * workqueues which have works queued on the pool and let them process
Tejun Heoe22bee72010-06-29 10:07:14 +02002291 * those works so that forward progress can be guaranteed.
2292 *
2293 * This should happen rarely.
Yacine Belkadid185af32013-07-31 14:59:24 -07002294 *
2295 * Return: 0
Tejun Heoe22bee72010-06-29 10:07:14 +02002296 */
Tejun Heo111c2252013-01-17 17:16:24 -08002297static int rescuer_thread(void *__rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02002298{
Tejun Heo111c2252013-01-17 17:16:24 -08002299 struct worker *rescuer = __rescuer;
2300 struct workqueue_struct *wq = rescuer->rescue_wq;
Tejun Heoe22bee72010-06-29 10:07:14 +02002301 struct list_head *scheduled = &rescuer->scheduled;
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002302 bool should_stop;
Tejun Heoe22bee72010-06-29 10:07:14 +02002303
2304 set_user_nice(current, RESCUER_NICE_LEVEL);
Tejun Heo111c2252013-01-17 17:16:24 -08002305
2306 /*
2307 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2308 * doesn't participate in concurrency management.
2309 */
2310 rescuer->task->flags |= PF_WQ_WORKER;
Tejun Heoe22bee72010-06-29 10:07:14 +02002311repeat:
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002312 set_current_state(TASK_IDLE);
Tejun Heoe22bee72010-06-29 10:07:14 +02002313
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002314 /*
2315 * By the time the rescuer is requested to stop, the workqueue
2316 * shouldn't have any work pending, but @wq->maydays may still have
2317 * pwq(s) queued. This can happen by non-rescuer workers consuming
2318 * all the work items before the rescuer got to them. Go through
2319 * @wq->maydays processing before acting on should_stop so that the
2320 * list is always empty on exit.
2321 */
2322 should_stop = kthread_should_stop();
Tejun Heoe22bee72010-06-29 10:07:14 +02002323
Tejun Heo493a1722013-03-12 11:29:59 -07002324 /* see whether any pwq is asking for help */
Tejun Heo2e109a22013-03-13 19:47:40 -07002325 spin_lock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002326
2327 while (!list_empty(&wq->maydays)) {
2328 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2329 struct pool_workqueue, mayday_node);
Tejun Heo112202d2013-02-13 19:29:12 -08002330 struct worker_pool *pool = pwq->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002331 struct work_struct *work, *n;
Tejun Heo82607adc2015-12-08 11:28:04 -05002332 bool first = true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002333
2334 __set_current_state(TASK_RUNNING);
Tejun Heo493a1722013-03-12 11:29:59 -07002335 list_del_init(&pwq->mayday_node);
2336
Tejun Heo2e109a22013-03-13 19:47:40 -07002337 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002338
Lai Jiangshan51697d392014-05-20 17:46:36 +08002339 worker_attach_to_pool(rescuer, pool);
2340
2341 spin_lock_irq(&pool->lock);
Lai Jiangshanb3104102013-02-19 12:17:02 -08002342 rescuer->pool = pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002343
2344 /*
2345 * Slurp in all works issued via this workqueue and
2346 * process'em.
2347 */
Tejun Heo0479c8c2014-12-04 10:14:13 -05002348 WARN_ON_ONCE(!list_empty(scheduled));
Tejun Heo82607adc2015-12-08 11:28:04 -05002349 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2350 if (get_work_pwq(work) == pwq) {
2351 if (first)
2352 pool->watchdog_ts = jiffies;
Tejun Heoe22bee72010-06-29 10:07:14 +02002353 move_linked_works(work, scheduled, &n);
Tejun Heo82607adc2015-12-08 11:28:04 -05002354 }
2355 first = false;
2356 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002357
NeilBrown008847f2014-12-08 12:39:16 -05002358 if (!list_empty(scheduled)) {
2359 process_scheduled_works(rescuer);
2360
2361 /*
2362 * The above execution of rescued work items could
2363 * have created more to rescue through
2364 * pwq_activate_first_delayed() or chained
2365 * queueing. Let's put @pwq back on mayday list so
2366 * that such back-to-back work items, which may be
2367 * being used to relieve memory pressure, don't
2368 * incur MAYDAY_INTERVAL delay inbetween.
2369 */
2370 if (need_to_create_worker(pool)) {
2371 spin_lock(&wq_mayday_lock);
Tejun Heo9e0f33d2019-09-25 06:59:15 -07002372 /*
2373 * Queue iff we aren't racing destruction
2374 * and somebody else hasn't queued it already.
2375 */
2376 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2377 get_pwq(pwq);
2378 list_add_tail(&pwq->mayday_node, &wq->maydays);
2379 }
NeilBrown008847f2014-12-08 12:39:16 -05002380 spin_unlock(&wq_mayday_lock);
2381 }
2382 }
Tejun Heo75769582011-02-14 14:04:46 +01002383
2384 /*
Lai Jiangshan77668c82014-04-18 11:04:16 -04002385 * Put the reference grabbed by send_mayday(). @pool won't
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002386 * go away while we're still attached to it.
Lai Jiangshan77668c82014-04-18 11:04:16 -04002387 */
2388 put_pwq(pwq);
2389
2390 /*
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002391 * Leave this pool. If need_more_worker() is %true, notify a
Tejun Heo75769582011-02-14 14:04:46 +01002392 * regular worker; otherwise, we end up with 0 concurrency
2393 * and stalling the execution.
2394 */
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002395 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002396 wake_up_worker(pool);
Tejun Heo75769582011-02-14 14:04:46 +01002397
Lai Jiangshanb3104102013-02-19 12:17:02 -08002398 rescuer->pool = NULL;
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002399 spin_unlock_irq(&pool->lock);
2400
2401 worker_detach_from_pool(rescuer, pool);
2402
2403 spin_lock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002404 }
2405
Tejun Heo2e109a22013-03-13 19:47:40 -07002406 spin_unlock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002407
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002408 if (should_stop) {
2409 __set_current_state(TASK_RUNNING);
2410 rescuer->task->flags &= ~PF_WQ_WORKER;
2411 return 0;
2412 }
2413
Tejun Heo111c2252013-01-17 17:16:24 -08002414 /* rescuers should never participate in concurrency management */
2415 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
Tejun Heoe22bee72010-06-29 10:07:14 +02002416 schedule();
2417 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418}
2419
Tejun Heofca839c2015-12-07 10:58:57 -05002420/**
2421 * check_flush_dependency - check for flush dependency sanity
2422 * @target_wq: workqueue being flushed
2423 * @target_work: work item being flushed (NULL for workqueue flushes)
2424 *
2425 * %current is trying to flush the whole @target_wq or @target_work on it.
2426 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2427 * reclaiming memory or running on a workqueue which doesn't have
2428 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2429 * a deadlock.
2430 */
2431static void check_flush_dependency(struct workqueue_struct *target_wq,
2432 struct work_struct *target_work)
2433{
2434 work_func_t target_func = target_work ? target_work->func : NULL;
2435 struct worker *worker;
2436
2437 if (target_wq->flags & WQ_MEM_RECLAIM)
2438 return;
2439
2440 worker = current_wq_worker();
2441
2442 WARN_ONCE(current->flags & PF_MEMALLOC,
2443 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2444 current->pid, current->comm, target_wq->name, target_func);
Tejun Heo23d11a52016-01-29 05:59:46 -05002445 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2446 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
Tejun Heofca839c2015-12-07 10:58:57 -05002447 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2448 worker->current_pwq->wq->name, worker->current_func,
2449 target_wq->name, target_func);
2450}
2451
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002452struct wq_barrier {
2453 struct work_struct work;
2454 struct completion done;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002455 struct task_struct *task; /* purely informational */
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002456};
2457
2458static void wq_barrier_func(struct work_struct *work)
2459{
2460 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2461 complete(&barr->done);
2462}
2463
Tejun Heo4690c4a2010-06-29 10:07:10 +02002464/**
2465 * insert_wq_barrier - insert a barrier work
Tejun Heo112202d2013-02-13 19:29:12 -08002466 * @pwq: pwq to insert barrier into
Tejun Heo4690c4a2010-06-29 10:07:10 +02002467 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002468 * @target: target work to attach @barr to
2469 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002470 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002471 * @barr is linked to @target such that @barr is completed only after
2472 * @target finishes execution. Please note that the ordering
2473 * guarantee is observed only with respect to @target and on the local
2474 * cpu.
2475 *
2476 * Currently, a queued barrier can't be canceled. This is because
2477 * try_to_grab_pending() can't determine whether the work to be
2478 * grabbed is at the head of the queue and thus can't clear LINKED
2479 * flag of the previous work while there must be a valid next work
2480 * after a work with LINKED flag set.
2481 *
2482 * Note that when @worker is non-NULL, @target may be modified
Tejun Heo112202d2013-02-13 19:29:12 -08002483 * underneath us, so we can't reliably determine pwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002484 *
2485 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002486 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002487 */
Tejun Heo112202d2013-02-13 19:29:12 -08002488static void insert_wq_barrier(struct pool_workqueue *pwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002489 struct wq_barrier *barr,
2490 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002491{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002492 struct list_head *head;
2493 unsigned int linked = 0;
2494
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002495 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002496 * debugobject calls are safe here even with pool->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002497 * as we know for sure that this will not trigger any of the
2498 * checks and call back into the fixup functions where we
2499 * might deadlock.
2500 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002501 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002502 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Boqun Feng52fa5bc2017-08-17 17:46:12 +08002503
2504 /*
2505 * Explicitly init the crosslock for wq_barrier::done, make its lock
2506 * key a subkey of the corresponding work. As a result we won't
2507 * build a dependency between wq_barrier::done and unrelated work.
2508 */
2509 lockdep_init_map_crosslock((struct lockdep_map *)&barr->done.map,
2510 "(complete)wq_barr::done",
2511 target->lockdep_map.key, 1);
2512 __init_completion(&barr->done);
Tejun Heo2607d7a2015-03-09 09:22:28 -04002513 barr->task = current;
Oleg Nesterov83c22522007-05-09 02:33:54 -07002514
Tejun Heoaffee4b2010-06-29 10:07:12 +02002515 /*
2516 * If @target is currently being executed, schedule the
2517 * barrier to the worker; otherwise, put it after @target.
2518 */
2519 if (worker)
2520 head = worker->scheduled.next;
2521 else {
2522 unsigned long *bits = work_data_bits(target);
2523
2524 head = target->entry.next;
2525 /* there can already be other linked works, inherit and set */
2526 linked = *bits & WORK_STRUCT_LINKED;
2527 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2528 }
2529
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002530 debug_work_activate(&barr->work);
Tejun Heo112202d2013-02-13 19:29:12 -08002531 insert_work(pwq, &barr->work, head,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002532 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002533}
2534
Tejun Heo73f53c42010-06-29 10:07:11 +02002535/**
Tejun Heo112202d2013-02-13 19:29:12 -08002536 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
Tejun Heo73f53c42010-06-29 10:07:11 +02002537 * @wq: workqueue being flushed
2538 * @flush_color: new flush color, < 0 for no-op
2539 * @work_color: new work color, < 0 for no-op
2540 *
Tejun Heo112202d2013-02-13 19:29:12 -08002541 * Prepare pwqs for workqueue flushing.
Tejun Heo73f53c42010-06-29 10:07:11 +02002542 *
Tejun Heo112202d2013-02-13 19:29:12 -08002543 * If @flush_color is non-negative, flush_color on all pwqs should be
2544 * -1. If no pwq has in-flight commands at the specified color, all
2545 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2546 * has in flight commands, its pwq->flush_color is set to
2547 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
Tejun Heo73f53c42010-06-29 10:07:11 +02002548 * wakeup logic is armed and %true is returned.
2549 *
2550 * The caller should have initialized @wq->first_flusher prior to
2551 * calling this function with non-negative @flush_color. If
2552 * @flush_color is negative, no flush color update is done and %false
2553 * is returned.
2554 *
Tejun Heo112202d2013-02-13 19:29:12 -08002555 * If @work_color is non-negative, all pwqs should have the same
Tejun Heo73f53c42010-06-29 10:07:11 +02002556 * work_color which is previous to @work_color and all will be
2557 * advanced to @work_color.
2558 *
2559 * CONTEXT:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002560 * mutex_lock(wq->mutex).
Tejun Heo73f53c42010-06-29 10:07:11 +02002561 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002562 * Return:
Tejun Heo73f53c42010-06-29 10:07:11 +02002563 * %true if @flush_color >= 0 and there's something to flush. %false
2564 * otherwise.
2565 */
Tejun Heo112202d2013-02-13 19:29:12 -08002566static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
Tejun Heo73f53c42010-06-29 10:07:11 +02002567 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568{
Tejun Heo73f53c42010-06-29 10:07:11 +02002569 bool wait = false;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002570 struct pool_workqueue *pwq;
Oleg Nesterov14441962007-05-23 13:57:57 -07002571
Tejun Heo73f53c42010-06-29 10:07:11 +02002572 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002573 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
Tejun Heo112202d2013-02-13 19:29:12 -08002574 atomic_set(&wq->nr_pwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002575 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002576
Tejun Heo49e3cf42013-03-12 11:29:58 -07002577 for_each_pwq(pwq, wq) {
Tejun Heo112202d2013-02-13 19:29:12 -08002578 struct worker_pool *pool = pwq->pool;
Tejun Heo73f53c42010-06-29 10:07:11 +02002579
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002580 spin_lock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002581
2582 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002583 WARN_ON_ONCE(pwq->flush_color != -1);
Tejun Heo73f53c42010-06-29 10:07:11 +02002584
Tejun Heo112202d2013-02-13 19:29:12 -08002585 if (pwq->nr_in_flight[flush_color]) {
2586 pwq->flush_color = flush_color;
2587 atomic_inc(&wq->nr_pwqs_to_flush);
Tejun Heo73f53c42010-06-29 10:07:11 +02002588 wait = true;
2589 }
2590 }
2591
2592 if (work_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002593 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
Tejun Heo112202d2013-02-13 19:29:12 -08002594 pwq->work_color = work_color;
Tejun Heo73f53c42010-06-29 10:07:11 +02002595 }
2596
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002597 spin_unlock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002598 }
2599
Tejun Heo112202d2013-02-13 19:29:12 -08002600 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
Tejun Heo73f53c42010-06-29 10:07:11 +02002601 complete(&wq->first_flusher->done);
2602
2603 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604}
2605
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002606/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002608 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002610 * This function sleeps until all work items which were queued on entry
2611 * have finished execution, but it is not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002613void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614{
Tejun Heo73f53c42010-06-29 10:07:11 +02002615 struct wq_flusher this_flusher = {
2616 .list = LIST_HEAD_INIT(this_flusher.list),
2617 .flush_color = -1,
2618 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2619 };
2620 int next_color;
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -07002621
Tejun Heo3347fa02016-09-16 15:49:32 -04002622 if (WARN_ON(!wq_online))
2623 return;
2624
Ingo Molnar3295f0e2008-08-11 10:30:30 +02002625 lock_map_acquire(&wq->lockdep_map);
2626 lock_map_release(&wq->lockdep_map);
Tejun Heo73f53c42010-06-29 10:07:11 +02002627
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002628 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002629
2630 /*
2631 * Start-to-wait phase
2632 */
2633 next_color = work_next_color(wq->work_color);
2634
2635 if (next_color != wq->flush_color) {
2636 /*
2637 * Color space is not full. The current work_color
2638 * becomes our flush_color and work_color is advanced
2639 * by one.
2640 */
Tejun Heo6183c002013-03-12 11:29:57 -07002641 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
Tejun Heo73f53c42010-06-29 10:07:11 +02002642 this_flusher.flush_color = wq->work_color;
2643 wq->work_color = next_color;
2644
2645 if (!wq->first_flusher) {
2646 /* no flush in progress, become the first flusher */
Tejun Heo6183c002013-03-12 11:29:57 -07002647 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002648
2649 wq->first_flusher = &this_flusher;
2650
Tejun Heo112202d2013-02-13 19:29:12 -08002651 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
Tejun Heo73f53c42010-06-29 10:07:11 +02002652 wq->work_color)) {
2653 /* nothing to flush, done */
2654 wq->flush_color = next_color;
2655 wq->first_flusher = NULL;
2656 goto out_unlock;
2657 }
2658 } else {
2659 /* wait in queue */
Tejun Heo6183c002013-03-12 11:29:57 -07002660 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002661 list_add_tail(&this_flusher.list, &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002662 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002663 }
2664 } else {
2665 /*
2666 * Oops, color space is full, wait on overflow queue.
2667 * The next flush completion will assign us
2668 * flush_color and transfer to flusher_queue.
2669 */
2670 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2671 }
2672
Tejun Heofca839c2015-12-07 10:58:57 -05002673 check_flush_dependency(wq, NULL);
2674
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002675 mutex_unlock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002676
2677 wait_for_completion(&this_flusher.done);
2678
2679 /*
2680 * Wake-up-and-cascade phase
2681 *
2682 * First flushers are responsible for cascading flushes and
2683 * handling overflow. Non-first flushers can simply return.
2684 */
2685 if (wq->first_flusher != &this_flusher)
2686 return;
2687
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002688 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002689
Tejun Heo4ce48b32010-07-02 10:03:51 +02002690 /* we might have raced, check again with mutex held */
2691 if (wq->first_flusher != &this_flusher)
2692 goto out_unlock;
2693
Tejun Heo73f53c42010-06-29 10:07:11 +02002694 wq->first_flusher = NULL;
2695
Tejun Heo6183c002013-03-12 11:29:57 -07002696 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2697 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002698
2699 while (true) {
2700 struct wq_flusher *next, *tmp;
2701
2702 /* complete all the flushers sharing the current flush color */
2703 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2704 if (next->flush_color != wq->flush_color)
2705 break;
2706 list_del_init(&next->list);
2707 complete(&next->done);
2708 }
2709
Tejun Heo6183c002013-03-12 11:29:57 -07002710 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2711 wq->flush_color != work_next_color(wq->work_color));
Tejun Heo73f53c42010-06-29 10:07:11 +02002712
2713 /* this flush_color is finished, advance by one */
2714 wq->flush_color = work_next_color(wq->flush_color);
2715
2716 /* one color has been freed, handle overflow queue */
2717 if (!list_empty(&wq->flusher_overflow)) {
2718 /*
2719 * Assign the same color to all overflowed
2720 * flushers, advance work_color and append to
2721 * flusher_queue. This is the start-to-wait
2722 * phase for these overflowed flushers.
2723 */
2724 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2725 tmp->flush_color = wq->work_color;
2726
2727 wq->work_color = work_next_color(wq->work_color);
2728
2729 list_splice_tail_init(&wq->flusher_overflow,
2730 &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002731 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002732 }
2733
2734 if (list_empty(&wq->flusher_queue)) {
Tejun Heo6183c002013-03-12 11:29:57 -07002735 WARN_ON_ONCE(wq->flush_color != wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002736 break;
2737 }
2738
2739 /*
2740 * Need to flush more colors. Make the next flusher
Tejun Heo112202d2013-02-13 19:29:12 -08002741 * the new first flusher and arm pwqs.
Tejun Heo73f53c42010-06-29 10:07:11 +02002742 */
Tejun Heo6183c002013-03-12 11:29:57 -07002743 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2744 WARN_ON_ONCE(wq->flush_color != next->flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002745
2746 list_del_init(&next->list);
2747 wq->first_flusher = next;
2748
Tejun Heo112202d2013-02-13 19:29:12 -08002749 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
Tejun Heo73f53c42010-06-29 10:07:11 +02002750 break;
2751
2752 /*
2753 * Meh... this color is already done, clear first
2754 * flusher and repeat cascading.
2755 */
2756 wq->first_flusher = NULL;
2757 }
2758
2759out_unlock:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002760 mutex_unlock(&wq->mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761}
Tim Gardner1dadafa2015-08-04 11:26:04 -06002762EXPORT_SYMBOL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002764/**
2765 * drain_workqueue - drain a workqueue
2766 * @wq: workqueue to drain
2767 *
2768 * Wait until the workqueue becomes empty. While draining is in progress,
2769 * only chain queueing is allowed. IOW, only currently pending or running
2770 * work items on @wq can queue further work items on it. @wq is flushed
Chen Hanxiaob749b1b2015-05-13 06:10:05 -04002771 * repeatedly until it becomes empty. The number of flushing is determined
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002772 * by the depth of chaining and should be relatively short. Whine if it
2773 * takes too long.
2774 */
2775void drain_workqueue(struct workqueue_struct *wq)
2776{
2777 unsigned int flush_cnt = 0;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002778 struct pool_workqueue *pwq;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002779
2780 /*
2781 * __queue_work() needs to test whether there are drainers, is much
2782 * hotter than drain_workqueue() and already looks at @wq->flags.
Tejun Heo618b01e2013-03-12 11:30:04 -07002783 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002784 */
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002785 mutex_lock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002786 if (!wq->nr_drainers++)
Tejun Heo618b01e2013-03-12 11:30:04 -07002787 wq->flags |= __WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002788 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002789reflush:
2790 flush_workqueue(wq);
2791
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002792 mutex_lock(&wq->mutex);
Tejun Heo76af4d92013-03-12 11:30:00 -07002793
Tejun Heo49e3cf42013-03-12 11:29:58 -07002794 for_each_pwq(pwq, wq) {
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002795 bool drained;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002796
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002797 spin_lock_irq(&pwq->pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08002798 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002799 spin_unlock_irq(&pwq->pool->lock);
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002800
2801 if (drained)
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002802 continue;
2803
2804 if (++flush_cnt == 10 ||
2805 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002806 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
Valentin Ilie044c7822012-08-19 00:52:42 +03002807 wq->name, flush_cnt);
Tejun Heo76af4d92013-03-12 11:30:00 -07002808
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002809 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002810 goto reflush;
2811 }
2812
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002813 if (!--wq->nr_drainers)
Tejun Heo618b01e2013-03-12 11:30:04 -07002814 wq->flags &= ~__WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002815 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002816}
2817EXPORT_SYMBOL_GPL(drain_workqueue);
2818
Tejun Heo606a5022012-08-20 14:51:23 -07002819static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
Tejun Heobaf59022010-09-16 10:42:16 +02002820{
2821 struct worker *worker = NULL;
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002822 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002823 struct pool_workqueue *pwq;
Tejun Heobaf59022010-09-16 10:42:16 +02002824
2825 might_sleep();
Tejun Heobaf59022010-09-16 10:42:16 +02002826
Tejun Heofa1b54e2013-03-12 11:30:00 -07002827 local_irq_disable();
2828 pool = get_work_pool(work);
2829 if (!pool) {
2830 local_irq_enable();
2831 return false;
2832 }
2833
2834 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08002835 /* see the comment in try_to_grab_pending() with the same code */
Tejun Heo112202d2013-02-13 19:29:12 -08002836 pwq = get_work_pwq(work);
2837 if (pwq) {
2838 if (unlikely(pwq->pool != pool))
Tejun Heobaf59022010-09-16 10:42:16 +02002839 goto already_gone;
Tejun Heo606a5022012-08-20 14:51:23 -07002840 } else {
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002841 worker = find_worker_executing_work(pool, work);
Tejun Heobaf59022010-09-16 10:42:16 +02002842 if (!worker)
2843 goto already_gone;
Tejun Heo112202d2013-02-13 19:29:12 -08002844 pwq = worker->current_pwq;
Tejun Heo606a5022012-08-20 14:51:23 -07002845 }
Tejun Heobaf59022010-09-16 10:42:16 +02002846
Tejun Heofca839c2015-12-07 10:58:57 -05002847 check_flush_dependency(pwq->wq, work);
2848
Tejun Heo112202d2013-02-13 19:29:12 -08002849 insert_wq_barrier(pwq, barr, work, worker);
Tejun Heod565ed62013-01-24 11:01:33 -08002850 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002851
Tejun Heoe1594892011-01-09 23:32:15 +01002852 /*
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002853 * Force a lock recursion deadlock when using flush_work() inside a
2854 * single-threaded or rescuer equipped workqueue.
2855 *
2856 * For single threaded workqueues the deadlock happens when the work
2857 * is after the work issuing the flush_work(). For rescuer equipped
2858 * workqueues the deadlock happens when the rescuer stalls, blocking
2859 * forward progress.
Tejun Heoe1594892011-01-09 23:32:15 +01002860 */
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002861 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) {
Tejun Heo112202d2013-02-13 19:29:12 -08002862 lock_map_acquire(&pwq->wq->lockdep_map);
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002863 lock_map_release(&pwq->wq->lockdep_map);
2864 }
Tejun Heoe1594892011-01-09 23:32:15 +01002865
Tejun Heobaf59022010-09-16 10:42:16 +02002866 return true;
2867already_gone:
Tejun Heod565ed62013-01-24 11:01:33 -08002868 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002869 return false;
2870}
2871
Oleg Nesterovdb700892008-07-25 01:47:49 -07002872/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002873 * flush_work - wait for a work to finish executing the last queueing instance
2874 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07002875 *
Tejun Heo606a5022012-08-20 14:51:23 -07002876 * Wait until @work has finished execution. @work is guaranteed to be idle
2877 * on return if it hasn't been requeued since flush started.
Tejun Heo401a8d02010-09-16 10:36:00 +02002878 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002879 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002880 * %true if flush_work() waited for the work to finish execution,
2881 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07002882 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002883bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07002884{
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002885 struct wq_barrier barr;
2886
Tejun Heo3347fa02016-09-16 15:49:32 -04002887 if (WARN_ON(!wq_online))
2888 return false;
2889
Stephen Boyd0976dfc2012-04-20 17:28:50 -07002890 lock_map_acquire(&work->lockdep_map);
2891 lock_map_release(&work->lockdep_map);
2892
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002893 if (start_flush_work(work, &barr)) {
2894 wait_for_completion(&barr.done);
2895 destroy_work_on_stack(&barr.work);
2896 return true;
2897 } else {
2898 return false;
2899 }
Oleg Nesterovdb700892008-07-25 01:47:49 -07002900}
2901EXPORT_SYMBOL_GPL(flush_work);
2902
Tejun Heo8603e1b32015-03-05 08:04:13 -05002903struct cwt_wait {
Ingo Molnarac6424b2017-06-20 12:06:13 +02002904 wait_queue_entry_t wait;
Tejun Heo8603e1b32015-03-05 08:04:13 -05002905 struct work_struct *work;
2906};
2907
Ingo Molnarac6424b2017-06-20 12:06:13 +02002908static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
Tejun Heo8603e1b32015-03-05 08:04:13 -05002909{
2910 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2911
2912 if (cwait->work != key)
2913 return 0;
2914 return autoremove_wake_function(wait, mode, sync, key);
2915}
2916
Tejun Heo36e227d2012-08-03 10:30:46 -07002917static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
Tejun Heo401a8d02010-09-16 10:36:00 +02002918{
Tejun Heo8603e1b32015-03-05 08:04:13 -05002919 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
Tejun Heobbb68df2012-08-03 10:30:46 -07002920 unsigned long flags;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002921 int ret;
2922
2923 do {
Tejun Heobbb68df2012-08-03 10:30:46 -07002924 ret = try_to_grab_pending(work, is_dwork, &flags);
2925 /*
Tejun Heo8603e1b32015-03-05 08:04:13 -05002926 * If someone else is already canceling, wait for it to
2927 * finish. flush_work() doesn't work for PREEMPT_NONE
2928 * because we may get scheduled between @work's completion
2929 * and the other canceling task resuming and clearing
2930 * CANCELING - flush_work() will return false immediately
2931 * as @work is no longer busy, try_to_grab_pending() will
2932 * return -ENOENT as @work is still being canceled and the
2933 * other canceling task won't be able to clear CANCELING as
2934 * we're hogging the CPU.
2935 *
2936 * Let's wait for completion using a waitqueue. As this
2937 * may lead to the thundering herd problem, use a custom
2938 * wake function which matches @work along with exclusive
2939 * wait and wakeup.
Tejun Heobbb68df2012-08-03 10:30:46 -07002940 */
Tejun Heo8603e1b32015-03-05 08:04:13 -05002941 if (unlikely(ret == -ENOENT)) {
2942 struct cwt_wait cwait;
2943
2944 init_wait(&cwait.wait);
2945 cwait.wait.func = cwt_wakefn;
2946 cwait.work = work;
2947
2948 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2949 TASK_UNINTERRUPTIBLE);
2950 if (work_is_canceling(work))
2951 schedule();
2952 finish_wait(&cancel_waitq, &cwait.wait);
2953 }
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002954 } while (unlikely(ret < 0));
2955
Tejun Heobbb68df2012-08-03 10:30:46 -07002956 /* tell other tasks trying to grab @work to back off */
2957 mark_work_canceling(work);
2958 local_irq_restore(flags);
2959
Tejun Heo3347fa02016-09-16 15:49:32 -04002960 /*
2961 * This allows canceling during early boot. We know that @work
2962 * isn't executing.
2963 */
2964 if (wq_online)
2965 flush_work(work);
2966
Tejun Heo7a22ad72010-06-29 10:07:13 +02002967 clear_work_data(work);
Tejun Heo8603e1b32015-03-05 08:04:13 -05002968
2969 /*
2970 * Paired with prepare_to_wait() above so that either
2971 * waitqueue_active() is visible here or !work_is_canceling() is
2972 * visible there.
2973 */
2974 smp_mb();
2975 if (waitqueue_active(&cancel_waitq))
2976 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2977
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002978 return ret;
2979}
2980
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002981/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002982 * cancel_work_sync - cancel a work and wait for it to finish
2983 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002984 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002985 * Cancel @work and wait for its execution to finish. This function
2986 * can be used even if the work re-queues itself or migrates to
2987 * another workqueue. On return from this function, @work is
2988 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002989 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002990 * cancel_work_sync(&delayed_work->work) must not be used for
2991 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002992 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002993 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002994 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02002995 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002996 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002997 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002998 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002999bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003000{
Tejun Heo36e227d2012-08-03 10:30:46 -07003001 return __cancel_work_timer(work, false);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07003002}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07003003EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07003004
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003005/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003006 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3007 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003008 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003009 * Delayed timer is cancelled and the pending work is queued for
3010 * immediate execution. Like flush_work(), this function only
3011 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003012 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003013 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003014 * %true if flush_work() waited for the work to finish execution,
3015 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003016 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003017bool flush_delayed_work(struct delayed_work *dwork)
3018{
Tejun Heo8930cab2012-08-03 10:30:45 -07003019 local_irq_disable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003020 if (del_timer_sync(&dwork->timer))
Lai Jiangshan60c057b2013-02-06 18:04:53 -08003021 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Tejun Heo8930cab2012-08-03 10:30:45 -07003022 local_irq_enable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003023 return flush_work(&dwork->work);
3024}
3025EXPORT_SYMBOL(flush_delayed_work);
3026
Jens Axboef72b8792016-08-24 15:51:50 -06003027static bool __cancel_work(struct work_struct *work, bool is_dwork)
3028{
3029 unsigned long flags;
3030 int ret;
3031
3032 do {
3033 ret = try_to_grab_pending(work, is_dwork, &flags);
3034 } while (unlikely(ret == -EAGAIN));
3035
3036 if (unlikely(ret < 0))
3037 return false;
3038
3039 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3040 local_irq_restore(flags);
3041 return ret;
3042}
3043
3044/*
3045 * See cancel_delayed_work()
3046 */
3047bool cancel_work(struct work_struct *work)
3048{
3049 return __cancel_work(work, false);
3050}
3051
Tejun Heo401a8d02010-09-16 10:36:00 +02003052/**
Tejun Heo57b30ae2012-08-21 13:18:24 -07003053 * cancel_delayed_work - cancel a delayed work
3054 * @dwork: delayed_work to cancel
Tejun Heo09383492010-09-16 10:48:29 +02003055 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003056 * Kill off a pending delayed_work.
3057 *
3058 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3059 * pending.
3060 *
3061 * Note:
3062 * The work callback function may still be running on return, unless
3063 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3064 * use cancel_delayed_work_sync() to wait on it.
Tejun Heo09383492010-09-16 10:48:29 +02003065 *
Tejun Heo57b30ae2012-08-21 13:18:24 -07003066 * This function is safe to call from any context including IRQ handler.
Tejun Heo09383492010-09-16 10:48:29 +02003067 */
Tejun Heo57b30ae2012-08-21 13:18:24 -07003068bool cancel_delayed_work(struct delayed_work *dwork)
Tejun Heo09383492010-09-16 10:48:29 +02003069{
Jens Axboef72b8792016-08-24 15:51:50 -06003070 return __cancel_work(&dwork->work, true);
Tejun Heo09383492010-09-16 10:48:29 +02003071}
Tejun Heo57b30ae2012-08-21 13:18:24 -07003072EXPORT_SYMBOL(cancel_delayed_work);
Tejun Heo09383492010-09-16 10:48:29 +02003073
3074/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003075 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3076 * @dwork: the delayed work cancel
3077 *
3078 * This is cancel_work_sync() for delayed works.
3079 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003080 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003081 * %true if @dwork was pending, %false otherwise.
3082 */
3083bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003084{
Tejun Heo36e227d2012-08-03 10:30:46 -07003085 return __cancel_work_timer(&dwork->work, true);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003086}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07003087EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07003089/**
Tejun Heo31ddd872010-10-19 11:14:49 +02003090 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07003091 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07003092 *
Tejun Heo31ddd872010-10-19 11:14:49 +02003093 * schedule_on_each_cpu() executes @func on each online CPU using the
3094 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07003095 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02003096 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003097 * Return:
Tejun Heo31ddd872010-10-19 11:14:49 +02003098 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07003099 */
David Howells65f27f32006-11-22 14:55:48 +00003100int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003101{
3102 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02003103 struct work_struct __percpu *works;
Christoph Lameter15316ba2006-01-08 01:00:43 -08003104
Andrew Mortonb6136772006-06-25 05:47:49 -07003105 works = alloc_percpu(struct work_struct);
3106 if (!works)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003107 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07003108
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003109 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08003110
Christoph Lameter15316ba2006-01-08 01:00:43 -08003111 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01003112 struct work_struct *work = per_cpu_ptr(works, cpu);
3113
3114 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003115 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02003116 }
Tejun Heo93981802009-11-17 14:06:20 -08003117
3118 for_each_online_cpu(cpu)
3119 flush_work(per_cpu_ptr(works, cpu));
3120
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003121 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07003122 free_percpu(works);
Christoph Lameter15316ba2006-01-08 01:00:43 -08003123 return 0;
3124}
3125
Alan Sterneef6a7d2010-02-12 17:39:21 +09003126/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06003127 * execute_in_process_context - reliably execute the routine with user context
3128 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06003129 * @ew: guaranteed storage for the execute work structure (must
3130 * be available when the work executes)
3131 *
3132 * Executes the function immediately if process context is available,
3133 * otherwise schedules the function for delayed execution.
3134 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003135 * Return: 0 - function was executed
James Bottomley1fa44ec2006-02-23 12:43:43 -06003136 * 1 - function was scheduled for execution
3137 */
David Howells65f27f32006-11-22 14:55:48 +00003138int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06003139{
3140 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00003141 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003142 return 0;
3143 }
3144
David Howells65f27f32006-11-22 14:55:48 +00003145 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003146 schedule_work(&ew->work);
3147
3148 return 1;
3149}
3150EXPORT_SYMBOL_GPL(execute_in_process_context);
3151
Tejun Heo7a4e3442013-03-12 11:30:00 -07003152/**
3153 * free_workqueue_attrs - free a workqueue_attrs
3154 * @attrs: workqueue_attrs to free
3155 *
3156 * Undo alloc_workqueue_attrs().
3157 */
3158void free_workqueue_attrs(struct workqueue_attrs *attrs)
3159{
3160 if (attrs) {
3161 free_cpumask_var(attrs->cpumask);
3162 kfree(attrs);
3163 }
3164}
3165
3166/**
3167 * alloc_workqueue_attrs - allocate a workqueue_attrs
3168 * @gfp_mask: allocation mask to use
3169 *
3170 * Allocate a new workqueue_attrs, initialize with default settings and
Yacine Belkadid185af32013-07-31 14:59:24 -07003171 * return it.
3172 *
3173 * Return: The allocated new workqueue_attr on success. %NULL on failure.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003174 */
3175struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3176{
3177 struct workqueue_attrs *attrs;
3178
3179 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3180 if (!attrs)
3181 goto fail;
3182 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3183 goto fail;
3184
Tejun Heo13e2e552013-04-01 11:23:31 -07003185 cpumask_copy(attrs->cpumask, cpu_possible_mask);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003186 return attrs;
3187fail:
3188 free_workqueue_attrs(attrs);
3189 return NULL;
3190}
3191
Tejun Heo29c91e92013-03-12 11:30:03 -07003192static void copy_workqueue_attrs(struct workqueue_attrs *to,
3193 const struct workqueue_attrs *from)
3194{
3195 to->nice = from->nice;
3196 cpumask_copy(to->cpumask, from->cpumask);
Shaohua Li2865a8f2013-08-01 09:56:36 +08003197 /*
3198 * Unlike hash and equality test, this function doesn't ignore
3199 * ->no_numa as it is used for both pool and wq attrs. Instead,
3200 * get_unbound_pool() explicitly clears ->no_numa after copying.
3201 */
3202 to->no_numa = from->no_numa;
Tejun Heo29c91e92013-03-12 11:30:03 -07003203}
3204
Tejun Heo29c91e92013-03-12 11:30:03 -07003205/* hash value of the content of @attr */
3206static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3207{
3208 u32 hash = 0;
3209
3210 hash = jhash_1word(attrs->nice, hash);
Tejun Heo13e2e552013-04-01 11:23:31 -07003211 hash = jhash(cpumask_bits(attrs->cpumask),
3212 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
Tejun Heo29c91e92013-03-12 11:30:03 -07003213 return hash;
3214}
3215
3216/* content equality test */
3217static bool wqattrs_equal(const struct workqueue_attrs *a,
3218 const struct workqueue_attrs *b)
3219{
3220 if (a->nice != b->nice)
3221 return false;
3222 if (!cpumask_equal(a->cpumask, b->cpumask))
3223 return false;
3224 return true;
3225}
3226
Tejun Heo7a4e3442013-03-12 11:30:00 -07003227/**
3228 * init_worker_pool - initialize a newly zalloc'd worker_pool
3229 * @pool: worker_pool to initialize
3230 *
Shailendra Verma402dd892015-05-23 10:38:14 +05303231 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
Yacine Belkadid185af32013-07-31 14:59:24 -07003232 *
3233 * Return: 0 on success, -errno on failure. Even on failure, all fields
Tejun Heo29c91e92013-03-12 11:30:03 -07003234 * inside @pool proper are initialized and put_unbound_pool() can be called
3235 * on @pool safely to release it.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003236 */
3237static int init_worker_pool(struct worker_pool *pool)
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003238{
3239 spin_lock_init(&pool->lock);
Tejun Heo29c91e92013-03-12 11:30:03 -07003240 pool->id = -1;
3241 pool->cpu = -1;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003242 pool->node = NUMA_NO_NODE;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003243 pool->flags |= POOL_DISASSOCIATED;
Tejun Heo82607adc2015-12-08 11:28:04 -05003244 pool->watchdog_ts = jiffies;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003245 INIT_LIST_HEAD(&pool->worklist);
3246 INIT_LIST_HEAD(&pool->idle_list);
3247 hash_init(pool->busy_hash);
3248
Geliang Tangc30fb262017-02-26 08:58:21 +08003249 setup_deferrable_timer(&pool->idle_timer, idle_worker_timeout,
3250 (unsigned long)pool);
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003251
3252 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3253 (unsigned long)pool);
3254
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003255 mutex_init(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003256 INIT_LIST_HEAD(&pool->workers);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003257
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003258 ida_init(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003259 INIT_HLIST_NODE(&pool->hash_node);
3260 pool->refcnt = 1;
3261
3262 /* shouldn't fail above this point */
Tejun Heo7a4e3442013-03-12 11:30:00 -07003263 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3264 if (!pool->attrs)
3265 return -ENOMEM;
3266 return 0;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003267}
3268
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003269static void rcu_free_wq(struct rcu_head *rcu)
3270{
3271 struct workqueue_struct *wq =
3272 container_of(rcu, struct workqueue_struct, rcu);
3273
3274 if (!(wq->flags & WQ_UNBOUND))
3275 free_percpu(wq->cpu_pwqs);
3276 else
3277 free_workqueue_attrs(wq->unbound_attrs);
3278
3279 kfree(wq->rescuer);
3280 kfree(wq);
3281}
3282
Tejun Heo29c91e92013-03-12 11:30:03 -07003283static void rcu_free_pool(struct rcu_head *rcu)
3284{
3285 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3286
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003287 ida_destroy(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003288 free_workqueue_attrs(pool->attrs);
3289 kfree(pool);
3290}
3291
3292/**
3293 * put_unbound_pool - put a worker_pool
3294 * @pool: worker_pool to put
3295 *
3296 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003297 * safe manner. get_unbound_pool() calls this function on its failure path
3298 * and this function should be able to release pools which went through,
3299 * successfully or not, init_worker_pool().
Tejun Heoa892cac2013-04-01 11:23:32 -07003300 *
3301 * Should be called with wq_pool_mutex held.
Tejun Heo29c91e92013-03-12 11:30:03 -07003302 */
3303static void put_unbound_pool(struct worker_pool *pool)
3304{
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003305 DECLARE_COMPLETION_ONSTACK(detach_completion);
Tejun Heo29c91e92013-03-12 11:30:03 -07003306 struct worker *worker;
3307
Tejun Heoa892cac2013-04-01 11:23:32 -07003308 lockdep_assert_held(&wq_pool_mutex);
3309
3310 if (--pool->refcnt)
Tejun Heo29c91e92013-03-12 11:30:03 -07003311 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003312
3313 /* sanity checks */
Lai Jiangshan61d0fbb2014-06-03 15:31:45 +08003314 if (WARN_ON(!(pool->cpu < 0)) ||
Tejun Heoa892cac2013-04-01 11:23:32 -07003315 WARN_ON(!list_empty(&pool->worklist)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003316 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003317
3318 /* release id and unhash */
3319 if (pool->id >= 0)
3320 idr_remove(&worker_pool_idr, pool->id);
3321 hash_del(&pool->hash_node);
3322
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003323 /*
Tejun Heo692b4822017-10-09 08:04:13 -07003324 * Become the manager and destroy all workers. This prevents
3325 * @pool's workers from blocking on attach_mutex. We're the last
3326 * manager and @pool gets freed with the flag set.
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003327 */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003328 spin_lock_irq(&pool->lock);
Tejun Heo692b4822017-10-09 08:04:13 -07003329 wait_event_lock_irq(wq_manager_wait,
3330 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3331 pool->flags |= POOL_MANAGER_ACTIVE;
3332
Lai Jiangshan1037de32014-05-22 16:44:07 +08003333 while ((worker = first_idle_worker(pool)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003334 destroy_worker(worker);
3335 WARN_ON(pool->nr_workers || pool->nr_idle);
Tejun Heo29c91e92013-03-12 11:30:03 -07003336 spin_unlock_irq(&pool->lock);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003337
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003338 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003339 if (!list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003340 pool->detach_completion = &detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003341 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003342
3343 if (pool->detach_completion)
3344 wait_for_completion(pool->detach_completion);
3345
Tejun Heo29c91e92013-03-12 11:30:03 -07003346 /* shut down the timers */
3347 del_timer_sync(&pool->idle_timer);
3348 del_timer_sync(&pool->mayday_timer);
3349
3350 /* sched-RCU protected to allow dereferences from get_work_pool() */
3351 call_rcu_sched(&pool->rcu, rcu_free_pool);
3352}
3353
3354/**
3355 * get_unbound_pool - get a worker_pool with the specified attributes
3356 * @attrs: the attributes of the worker_pool to get
3357 *
3358 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3359 * reference count and return it. If there already is a matching
3360 * worker_pool, it will be used; otherwise, this function attempts to
Yacine Belkadid185af32013-07-31 14:59:24 -07003361 * create a new one.
Tejun Heoa892cac2013-04-01 11:23:32 -07003362 *
3363 * Should be called with wq_pool_mutex held.
Yacine Belkadid185af32013-07-31 14:59:24 -07003364 *
3365 * Return: On success, a worker_pool with the same attributes as @attrs.
3366 * On failure, %NULL.
Tejun Heo29c91e92013-03-12 11:30:03 -07003367 */
3368static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3369{
Tejun Heo29c91e92013-03-12 11:30:03 -07003370 u32 hash = wqattrs_hash(attrs);
3371 struct worker_pool *pool;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003372 int node;
Xunlei Pange2273582015-10-09 11:53:12 +08003373 int target_node = NUMA_NO_NODE;
Tejun Heo29c91e92013-03-12 11:30:03 -07003374
Tejun Heoa892cac2013-04-01 11:23:32 -07003375 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003376
3377 /* do we already have a matching pool? */
Tejun Heo29c91e92013-03-12 11:30:03 -07003378 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3379 if (wqattrs_equal(pool->attrs, attrs)) {
3380 pool->refcnt++;
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003381 return pool;
Tejun Heo29c91e92013-03-12 11:30:03 -07003382 }
3383 }
Tejun Heo29c91e92013-03-12 11:30:03 -07003384
Xunlei Pange2273582015-10-09 11:53:12 +08003385 /* if cpumask is contained inside a NUMA node, we belong to that node */
3386 if (wq_numa_enabled) {
3387 for_each_node(node) {
3388 if (cpumask_subset(attrs->cpumask,
3389 wq_numa_possible_cpumask[node])) {
3390 target_node = node;
3391 break;
3392 }
3393 }
3394 }
3395
Tejun Heo29c91e92013-03-12 11:30:03 -07003396 /* nope, create a new one */
Xunlei Pange2273582015-10-09 11:53:12 +08003397 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
Tejun Heo29c91e92013-03-12 11:30:03 -07003398 if (!pool || init_worker_pool(pool) < 0)
3399 goto fail;
3400
Tejun Heo8864b4e2013-03-12 11:30:04 -07003401 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
Tejun Heo29c91e92013-03-12 11:30:03 -07003402 copy_workqueue_attrs(pool->attrs, attrs);
Xunlei Pange2273582015-10-09 11:53:12 +08003403 pool->node = target_node;
Tejun Heo29c91e92013-03-12 11:30:03 -07003404
Shaohua Li2865a8f2013-08-01 09:56:36 +08003405 /*
3406 * no_numa isn't a worker_pool attribute, always clear it. See
3407 * 'struct workqueue_attrs' comments for detail.
3408 */
3409 pool->attrs->no_numa = false;
3410
Tejun Heo29c91e92013-03-12 11:30:03 -07003411 if (worker_pool_assign_id(pool) < 0)
3412 goto fail;
3413
3414 /* create and start the initial worker */
Tejun Heo3347fa02016-09-16 15:49:32 -04003415 if (wq_online && !create_worker(pool))
Tejun Heo29c91e92013-03-12 11:30:03 -07003416 goto fail;
3417
Tejun Heo29c91e92013-03-12 11:30:03 -07003418 /* install */
Tejun Heo29c91e92013-03-12 11:30:03 -07003419 hash_add(unbound_pool_hash, &pool->hash_node, hash);
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003420
Tejun Heo29c91e92013-03-12 11:30:03 -07003421 return pool;
3422fail:
Tejun Heo29c91e92013-03-12 11:30:03 -07003423 if (pool)
3424 put_unbound_pool(pool);
3425 return NULL;
3426}
3427
Tejun Heo8864b4e2013-03-12 11:30:04 -07003428static void rcu_free_pwq(struct rcu_head *rcu)
3429{
3430 kmem_cache_free(pwq_cache,
3431 container_of(rcu, struct pool_workqueue, rcu));
3432}
3433
3434/*
3435 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3436 * and needs to be destroyed.
3437 */
3438static void pwq_unbound_release_workfn(struct work_struct *work)
3439{
3440 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3441 unbound_release_work);
3442 struct workqueue_struct *wq = pwq->wq;
3443 struct worker_pool *pool = pwq->pool;
Yang Yingliangad7ae9e2021-07-14 17:19:33 +08003444 bool is_last = false;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003445
Yang Yingliangad7ae9e2021-07-14 17:19:33 +08003446 /*
3447 * when @pwq is not linked, it doesn't hold any reference to the
3448 * @wq, and @wq is invalid to access.
3449 */
3450 if (!list_empty(&pwq->pwqs_node)) {
3451 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3452 return;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003453
Yang Yingliangad7ae9e2021-07-14 17:19:33 +08003454 mutex_lock(&wq->mutex);
3455 list_del_rcu(&pwq->pwqs_node);
3456 is_last = list_empty(&wq->pwqs);
3457 mutex_unlock(&wq->mutex);
3458 }
Tejun Heo8864b4e2013-03-12 11:30:04 -07003459
Tejun Heoa892cac2013-04-01 11:23:32 -07003460 mutex_lock(&wq_pool_mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003461 put_unbound_pool(pool);
Tejun Heoa892cac2013-04-01 11:23:32 -07003462 mutex_unlock(&wq_pool_mutex);
3463
Tejun Heo8864b4e2013-03-12 11:30:04 -07003464 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3465
3466 /*
3467 * If we're the last pwq going away, @wq is already dead and no one
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003468 * is gonna access it anymore. Schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003469 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003470 if (is_last)
3471 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003472}
3473
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003474/**
Tejun Heo699ce092013-03-13 16:51:35 -07003475 * pwq_adjust_max_active - update a pwq's max_active to the current setting
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003476 * @pwq: target pool_workqueue
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003477 *
Tejun Heo699ce092013-03-13 16:51:35 -07003478 * If @pwq isn't freezing, set @pwq->max_active to the associated
3479 * workqueue's saved_max_active and activate delayed work items
3480 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003481 */
Tejun Heo699ce092013-03-13 16:51:35 -07003482static void pwq_adjust_max_active(struct pool_workqueue *pwq)
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003483{
Tejun Heo699ce092013-03-13 16:51:35 -07003484 struct workqueue_struct *wq = pwq->wq;
3485 bool freezable = wq->flags & WQ_FREEZABLE;
Tejun Heo3347fa02016-09-16 15:49:32 -04003486 unsigned long flags;
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003487
Tejun Heo699ce092013-03-13 16:51:35 -07003488 /* for @wq->saved_max_active */
Lai Jiangshana357fc02013-03-25 16:57:19 -07003489 lockdep_assert_held(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003490
3491 /* fast exit for non-freezable wqs */
3492 if (!freezable && pwq->max_active == wq->saved_max_active)
3493 return;
3494
Tejun Heo3347fa02016-09-16 15:49:32 -04003495 /* this function can be called during early boot w/ irq disabled */
3496 spin_lock_irqsave(&pwq->pool->lock, flags);
Tejun Heo699ce092013-03-13 16:51:35 -07003497
Lai Jiangshan74b414e2014-05-22 19:01:16 +08003498 /*
3499 * During [un]freezing, the caller is responsible for ensuring that
3500 * this function is called at least once after @workqueue_freezing
3501 * is updated and visible.
3502 */
3503 if (!freezable || !workqueue_freezing) {
Yunfeng Yef6d739c2020-11-19 14:21:25 +08003504 bool kick = false;
3505
Tejun Heo699ce092013-03-13 16:51:35 -07003506 pwq->max_active = wq->saved_max_active;
3507
3508 while (!list_empty(&pwq->delayed_works) &&
Yunfeng Yef6d739c2020-11-19 14:21:25 +08003509 pwq->nr_active < pwq->max_active) {
Tejun Heo699ce092013-03-13 16:51:35 -07003510 pwq_activate_first_delayed(pwq);
Yunfeng Yef6d739c2020-11-19 14:21:25 +08003511 kick = true;
3512 }
Lai Jiangshan951a0782013-03-20 10:52:30 -07003513
3514 /*
3515 * Need to kick a worker after thawed or an unbound wq's
Yunfeng Yef6d739c2020-11-19 14:21:25 +08003516 * max_active is bumped. In realtime scenarios, always kicking a
3517 * worker will cause interference on the isolated cpu cores, so
3518 * let's kick iff work items were activated.
Lai Jiangshan951a0782013-03-20 10:52:30 -07003519 */
Yunfeng Yef6d739c2020-11-19 14:21:25 +08003520 if (kick)
3521 wake_up_worker(pwq->pool);
Tejun Heo699ce092013-03-13 16:51:35 -07003522 } else {
3523 pwq->max_active = 0;
3524 }
3525
Tejun Heo3347fa02016-09-16 15:49:32 -04003526 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003527}
3528
Tejun Heoe50aba92013-04-01 11:23:35 -07003529/* initialize newly alloced @pwq which is associated with @wq and @pool */
Tejun Heof147f292013-04-01 11:23:35 -07003530static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3531 struct worker_pool *pool)
Tejun Heod2c1d402013-03-12 11:30:04 -07003532{
3533 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3534
Tejun Heoe50aba92013-04-01 11:23:35 -07003535 memset(pwq, 0, sizeof(*pwq));
3536
Tejun Heod2c1d402013-03-12 11:30:04 -07003537 pwq->pool = pool;
3538 pwq->wq = wq;
3539 pwq->flush_color = -1;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003540 pwq->refcnt = 1;
Tejun Heod2c1d402013-03-12 11:30:04 -07003541 INIT_LIST_HEAD(&pwq->delayed_works);
Tejun Heo1befcf32013-04-01 11:23:35 -07003542 INIT_LIST_HEAD(&pwq->pwqs_node);
Tejun Heod2c1d402013-03-12 11:30:04 -07003543 INIT_LIST_HEAD(&pwq->mayday_node);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003544 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
Tejun Heof147f292013-04-01 11:23:35 -07003545}
Tejun Heod2c1d402013-03-12 11:30:04 -07003546
Tejun Heof147f292013-04-01 11:23:35 -07003547/* sync @pwq with the current state of its associated wq and link it */
Tejun Heo1befcf32013-04-01 11:23:35 -07003548static void link_pwq(struct pool_workqueue *pwq)
Tejun Heof147f292013-04-01 11:23:35 -07003549{
3550 struct workqueue_struct *wq = pwq->wq;
3551
3552 lockdep_assert_held(&wq->mutex);
Tejun Heo75ccf592013-03-12 11:30:04 -07003553
Tejun Heo1befcf32013-04-01 11:23:35 -07003554 /* may be called multiple times, ignore if already linked */
3555 if (!list_empty(&pwq->pwqs_node))
3556 return;
3557
Lai Jiangshan29b1cb42014-07-22 13:04:27 +08003558 /* set the matching work_color */
Tejun Heo75ccf592013-03-12 11:30:04 -07003559 pwq->work_color = wq->work_color;
Tejun Heo983ca252013-03-13 16:51:35 -07003560
3561 /* sync max_active to the current setting */
3562 pwq_adjust_max_active(pwq);
3563
3564 /* link in @pwq */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003565 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
Tejun Heof147f292013-04-01 11:23:35 -07003566}
Lai Jiangshana357fc02013-03-25 16:57:19 -07003567
Tejun Heof147f292013-04-01 11:23:35 -07003568/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3569static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3570 const struct workqueue_attrs *attrs)
3571{
3572 struct worker_pool *pool;
3573 struct pool_workqueue *pwq;
3574
3575 lockdep_assert_held(&wq_pool_mutex);
3576
3577 pool = get_unbound_pool(attrs);
3578 if (!pool)
3579 return NULL;
3580
Tejun Heoe50aba92013-04-01 11:23:35 -07003581 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
Tejun Heof147f292013-04-01 11:23:35 -07003582 if (!pwq) {
3583 put_unbound_pool(pool);
3584 return NULL;
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003585 }
Tejun Heo6029a912013-04-01 11:23:34 -07003586
Tejun Heof147f292013-04-01 11:23:35 -07003587 init_pwq(pwq, wq, pool);
3588 return pwq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003589}
3590
Tejun Heo4c16bd32013-04-01 11:23:36 -07003591/**
Gong Zhaogang30186c62015-05-11 11:02:47 -04003592 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003593 * @attrs: the wq_attrs of the default pwq of the target workqueue
Tejun Heo4c16bd32013-04-01 11:23:36 -07003594 * @node: the target NUMA node
3595 * @cpu_going_down: if >= 0, the CPU to consider as offline
3596 * @cpumask: outarg, the resulting cpumask
3597 *
3598 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3599 * @cpu_going_down is >= 0, that cpu is considered offline during
Yacine Belkadid185af32013-07-31 14:59:24 -07003600 * calculation. The result is stored in @cpumask.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003601 *
3602 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3603 * enabled and @node has online CPUs requested by @attrs, the returned
3604 * cpumask is the intersection of the possible CPUs of @node and
3605 * @attrs->cpumask.
3606 *
3607 * The caller is responsible for ensuring that the cpumask of @node stays
3608 * stable.
Yacine Belkadid185af32013-07-31 14:59:24 -07003609 *
3610 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3611 * %false if equal.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003612 */
3613static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3614 int cpu_going_down, cpumask_t *cpumask)
3615{
Tejun Heod55262c2013-04-01 11:23:38 -07003616 if (!wq_numa_enabled || attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003617 goto use_dfl;
3618
3619 /* does @node have any online CPUs @attrs wants? */
3620 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3621 if (cpu_going_down >= 0)
3622 cpumask_clear_cpu(cpu_going_down, cpumask);
3623
3624 if (cpumask_empty(cpumask))
3625 goto use_dfl;
3626
3627 /* yeap, return possible CPUs in @node that @attrs wants */
3628 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
Michael Bringmann1ad0f0a2017-07-27 16:27:14 -05003629
3630 if (cpumask_empty(cpumask)) {
3631 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3632 "possible intersect\n");
3633 return false;
3634 }
3635
Tejun Heo4c16bd32013-04-01 11:23:36 -07003636 return !cpumask_equal(cpumask, attrs->cpumask);
3637
3638use_dfl:
3639 cpumask_copy(cpumask, attrs->cpumask);
3640 return false;
3641}
3642
Tejun Heo1befcf32013-04-01 11:23:35 -07003643/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3644static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3645 int node,
3646 struct pool_workqueue *pwq)
3647{
3648 struct pool_workqueue *old_pwq;
3649
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +08003650 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003651 lockdep_assert_held(&wq->mutex);
3652
3653 /* link_pwq() can handle duplicate calls */
3654 link_pwq(pwq);
3655
3656 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3657 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3658 return old_pwq;
3659}
3660
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003661/* context to store the prepared attrs & pwqs before applying */
3662struct apply_wqattrs_ctx {
3663 struct workqueue_struct *wq; /* target workqueue */
3664 struct workqueue_attrs *attrs; /* attrs to apply */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003665 struct list_head list; /* queued for batching commit */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003666 struct pool_workqueue *dfl_pwq;
3667 struct pool_workqueue *pwq_tbl[];
3668};
3669
3670/* free the resources after success or abort */
3671static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3672{
3673 if (ctx) {
3674 int node;
3675
3676 for_each_node(node)
3677 put_pwq_unlocked(ctx->pwq_tbl[node]);
3678 put_pwq_unlocked(ctx->dfl_pwq);
3679
3680 free_workqueue_attrs(ctx->attrs);
3681
3682 kfree(ctx);
3683 }
3684}
3685
3686/* allocate the attrs and pwqs for later installation */
3687static struct apply_wqattrs_ctx *
3688apply_wqattrs_prepare(struct workqueue_struct *wq,
3689 const struct workqueue_attrs *attrs)
3690{
3691 struct apply_wqattrs_ctx *ctx;
3692 struct workqueue_attrs *new_attrs, *tmp_attrs;
3693 int node;
3694
3695 lockdep_assert_held(&wq_pool_mutex);
3696
3697 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3698 GFP_KERNEL);
3699
3700 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3701 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3702 if (!ctx || !new_attrs || !tmp_attrs)
3703 goto out_free;
3704
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003705 /*
3706 * Calculate the attrs of the default pwq.
3707 * If the user configured cpumask doesn't overlap with the
3708 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3709 */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003710 copy_workqueue_attrs(new_attrs, attrs);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08003711 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003712 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3713 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003714
3715 /*
3716 * We may create multiple pwqs with differing cpumasks. Make a
3717 * copy of @new_attrs which will be modified and used to obtain
3718 * pools.
3719 */
3720 copy_workqueue_attrs(tmp_attrs, new_attrs);
3721
3722 /*
3723 * If something goes wrong during CPU up/down, we'll fall back to
3724 * the default pwq covering whole @attrs->cpumask. Always create
3725 * it even if we don't use it immediately.
3726 */
3727 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3728 if (!ctx->dfl_pwq)
3729 goto out_free;
3730
3731 for_each_node(node) {
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003732 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003733 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3734 if (!ctx->pwq_tbl[node])
3735 goto out_free;
3736 } else {
3737 ctx->dfl_pwq->refcnt++;
3738 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3739 }
3740 }
3741
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003742 /* save the user configured attrs and sanitize it. */
3743 copy_workqueue_attrs(new_attrs, attrs);
3744 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003745 ctx->attrs = new_attrs;
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003746
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003747 ctx->wq = wq;
3748 free_workqueue_attrs(tmp_attrs);
3749 return ctx;
3750
3751out_free:
3752 free_workqueue_attrs(tmp_attrs);
3753 free_workqueue_attrs(new_attrs);
3754 apply_wqattrs_cleanup(ctx);
3755 return NULL;
3756}
3757
3758/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3759static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3760{
3761 int node;
3762
3763 /* all pwqs have been created successfully, let's install'em */
3764 mutex_lock(&ctx->wq->mutex);
3765
3766 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3767
3768 /* save the previous pwq and install the new one */
3769 for_each_node(node)
3770 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3771 ctx->pwq_tbl[node]);
3772
3773 /* @dfl_pwq might not have been used, ensure it's linked */
3774 link_pwq(ctx->dfl_pwq);
3775 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3776
3777 mutex_unlock(&ctx->wq->mutex);
3778}
3779
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003780static void apply_wqattrs_lock(void)
3781{
3782 /* CPUs should stay stable across pwq creations and installations */
3783 get_online_cpus();
3784 mutex_lock(&wq_pool_mutex);
3785}
3786
3787static void apply_wqattrs_unlock(void)
3788{
3789 mutex_unlock(&wq_pool_mutex);
3790 put_online_cpus();
3791}
3792
3793static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3794 const struct workqueue_attrs *attrs)
3795{
3796 struct apply_wqattrs_ctx *ctx;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003797
3798 /* only unbound workqueues can change attributes */
3799 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3800 return -EINVAL;
3801
3802 /* creating multiple pwqs breaks ordering guarantee */
Tejun Heo0a94efb2017-07-23 08:36:15 -04003803 if (!list_empty(&wq->pwqs)) {
3804 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3805 return -EINVAL;
3806
3807 wq->flags &= ~__WQ_ORDERED;
3808 }
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003809
3810 ctx = apply_wqattrs_prepare(wq, attrs);
wanghaibin62011712016-01-07 20:38:59 +08003811 if (!ctx)
3812 return -ENOMEM;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003813
3814 /* the ctx has been prepared successfully, let's commit it */
wanghaibin62011712016-01-07 20:38:59 +08003815 apply_wqattrs_commit(ctx);
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003816 apply_wqattrs_cleanup(ctx);
3817
wanghaibin62011712016-01-07 20:38:59 +08003818 return 0;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003819}
3820
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003821/**
3822 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3823 * @wq: the target workqueue
3824 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3825 *
Tejun Heo4c16bd32013-04-01 11:23:36 -07003826 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3827 * machines, this function maps a separate pwq to each NUMA node with
3828 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3829 * NUMA node it was issued on. Older pwqs are released as in-flight work
3830 * items finish. Note that a work item which repeatedly requeues itself
3831 * back-to-back will stay on its current pwq.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003832 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003833 * Performs GFP_KERNEL allocations.
3834 *
3835 * Return: 0 on success and -errno on failure.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003836 */
3837int apply_workqueue_attrs(struct workqueue_struct *wq,
3838 const struct workqueue_attrs *attrs)
3839{
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003840 int ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003841
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003842 apply_wqattrs_lock();
3843 ret = apply_workqueue_attrs_locked(wq, attrs);
3844 apply_wqattrs_unlock();
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003845
Tejun Heo48621252013-04-01 11:23:31 -07003846 return ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003847}
3848
Tejun Heo4c16bd32013-04-01 11:23:36 -07003849/**
3850 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3851 * @wq: the target workqueue
3852 * @cpu: the CPU coming up or going down
3853 * @online: whether @cpu is coming up or going down
3854 *
3855 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3856 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3857 * @wq accordingly.
3858 *
3859 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3860 * falls back to @wq->dfl_pwq which may not be optimal but is always
3861 * correct.
3862 *
3863 * Note that when the last allowed CPU of a NUMA node goes offline for a
3864 * workqueue with a cpumask spanning multiple nodes, the workers which were
3865 * already executing the work items for the workqueue will lose their CPU
3866 * affinity and may execute on any CPU. This is similar to how per-cpu
3867 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3868 * affinity, it's the user's responsibility to flush the work item from
3869 * CPU_DOWN_PREPARE.
3870 */
3871static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3872 bool online)
3873{
3874 int node = cpu_to_node(cpu);
3875 int cpu_off = online ? -1 : cpu;
3876 struct pool_workqueue *old_pwq = NULL, *pwq;
3877 struct workqueue_attrs *target_attrs;
3878 cpumask_t *cpumask;
3879
3880 lockdep_assert_held(&wq_pool_mutex);
3881
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003882 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3883 wq->unbound_attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003884 return;
3885
3886 /*
3887 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3888 * Let's use a preallocated one. The following buf is protected by
3889 * CPU hotplug exclusion.
3890 */
3891 target_attrs = wq_update_unbound_numa_attrs_buf;
3892 cpumask = target_attrs->cpumask;
3893
Tejun Heo4c16bd32013-04-01 11:23:36 -07003894 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3895 pwq = unbound_pwq_by_node(wq, node);
3896
3897 /*
3898 * Let's determine what needs to be done. If the target cpumask is
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003899 * different from the default pwq's, we need to compare it to @pwq's
3900 * and create a new one if they don't match. If the target cpumask
3901 * equals the default pwq's, the default pwq should be used.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003902 */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003903 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
Tejun Heo4c16bd32013-04-01 11:23:36 -07003904 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003905 return;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003906 } else {
Daeseok Youn534a3fb2014-04-18 09:08:14 +09003907 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003908 }
3909
Tejun Heo4c16bd32013-04-01 11:23:36 -07003910 /* create a new pwq */
3911 pwq = alloc_unbound_pwq(wq, target_attrs);
3912 if (!pwq) {
Fabian Frederick2d916032014-05-12 13:59:35 -04003913 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3914 wq->name);
Daeseok Youn77f300b2014-04-16 14:32:29 +09003915 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003916 }
3917
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003918 /* Install the new pwq. */
Tejun Heo4c16bd32013-04-01 11:23:36 -07003919 mutex_lock(&wq->mutex);
3920 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3921 goto out_unlock;
3922
3923use_dfl_pwq:
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003924 mutex_lock(&wq->mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07003925 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3926 get_pwq(wq->dfl_pwq);
3927 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3928 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3929out_unlock:
3930 mutex_unlock(&wq->mutex);
3931 put_pwq_unlocked(old_pwq);
3932}
3933
Tejun Heo30cdf242013-03-12 11:29:57 -07003934static int alloc_and_link_pwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935{
Tejun Heo49e3cf42013-03-12 11:29:58 -07003936 bool highpri = wq->flags & WQ_HIGHPRI;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003937 int cpu, ret;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01003938
Tejun Heo30cdf242013-03-12 11:29:57 -07003939 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo420c0dd2013-03-12 11:29:59 -07003940 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3941 if (!wq->cpu_pwqs)
Tejun Heo30cdf242013-03-12 11:29:57 -07003942 return -ENOMEM;
3943
3944 for_each_possible_cpu(cpu) {
Tejun Heo7fb98ea2013-03-12 11:30:00 -07003945 struct pool_workqueue *pwq =
3946 per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heo7a62c2c2013-03-12 11:30:03 -07003947 struct worker_pool *cpu_pools =
Tejun Heof02ae732013-03-12 11:30:03 -07003948 per_cpu(cpu_worker_pools, cpu);
Tejun Heo30cdf242013-03-12 11:29:57 -07003949
Tejun Heof147f292013-04-01 11:23:35 -07003950 init_pwq(pwq, wq, &cpu_pools[highpri]);
3951
3952 mutex_lock(&wq->mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003953 link_pwq(pwq);
Tejun Heof147f292013-04-01 11:23:35 -07003954 mutex_unlock(&wq->mutex);
Tejun Heo30cdf242013-03-12 11:29:57 -07003955 }
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003956 return 0;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003957 } else if (wq->flags & __WQ_ORDERED) {
3958 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3959 /* there should only be single pwq for ordering guarantee */
3960 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3961 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3962 "ordering guarantee broken for workqueue %s\n", wq->name);
3963 return ret;
Tejun Heo30cdf242013-03-12 11:29:57 -07003964 } else {
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003965 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
Tejun Heo30cdf242013-03-12 11:29:57 -07003966 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07003967}
3968
Tejun Heof3421792010-07-02 10:03:51 +02003969static int wq_clamp_max_active(int max_active, unsigned int flags,
3970 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02003971{
Tejun Heof3421792010-07-02 10:03:51 +02003972 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3973
3974 if (max_active < 1 || max_active > lim)
Valentin Ilie044c7822012-08-19 00:52:42 +03003975 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3976 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003977
Tejun Heof3421792010-07-02 10:03:51 +02003978 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003979}
3980
Tejun Heob196be82012-01-10 15:11:35 -08003981struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
Tejun Heod320c032010-06-29 10:07:14 +02003982 unsigned int flags,
3983 int max_active,
3984 struct lock_class_key *key,
Tejun Heob196be82012-01-10 15:11:35 -08003985 const char *lock_name, ...)
Oleg Nesterov3af244332007-05-09 02:34:09 -07003986{
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003987 size_t tbl_size = 0;
Tejun Heoecf68812013-04-01 11:23:34 -07003988 va_list args;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003989 struct workqueue_struct *wq;
Tejun Heo49e3cf42013-03-12 11:29:58 -07003990 struct pool_workqueue *pwq;
Tejun Heob196be82012-01-10 15:11:35 -08003991
Tejun Heo5c0338c2017-07-18 18:41:52 -04003992 /*
3993 * Unbound && max_active == 1 used to imply ordered, which is no
3994 * longer the case on NUMA machines due to per-node pools. While
3995 * alloc_ordered_workqueue() is the right way to create an ordered
3996 * workqueue, keep the previous behavior to avoid subtle breakages
3997 * on NUMA.
3998 */
3999 if ((flags & WQ_UNBOUND) && max_active == 1)
4000 flags |= __WQ_ORDERED;
4001
Viresh Kumarcee22a12013-04-08 16:45:40 +05304002 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4003 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4004 flags |= WQ_UNBOUND;
4005
Tejun Heoecf68812013-04-01 11:23:34 -07004006 /* allocate wq and format name */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004007 if (flags & WQ_UNBOUND)
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08004008 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004009
4010 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
Tejun Heob196be82012-01-10 15:11:35 -08004011 if (!wq)
Tejun Heod2c1d402013-03-12 11:30:04 -07004012 return NULL;
Tejun Heob196be82012-01-10 15:11:35 -08004013
Tejun Heo6029a912013-04-01 11:23:34 -07004014 if (flags & WQ_UNBOUND) {
4015 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4016 if (!wq->unbound_attrs)
4017 goto err_free_wq;
4018 }
4019
Tejun Heoecf68812013-04-01 11:23:34 -07004020 va_start(args, lock_name);
4021 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
Tejun Heob196be82012-01-10 15:11:35 -08004022 va_end(args);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004023
Tejun Heod320c032010-06-29 10:07:14 +02004024 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heob196be82012-01-10 15:11:35 -08004025 max_active = wq_clamp_max_active(max_active, flags, wq->name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004026
Tejun Heob196be82012-01-10 15:11:35 -08004027 /* init wq */
Tejun Heo97e37d72010-06-29 10:07:10 +02004028 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004029 wq->saved_max_active = max_active;
Lai Jiangshan3c25a552013-03-25 16:57:17 -07004030 mutex_init(&wq->mutex);
Tejun Heo112202d2013-02-13 19:29:12 -08004031 atomic_set(&wq->nr_pwqs_to_flush, 0);
Tejun Heo30cdf242013-03-12 11:29:57 -07004032 INIT_LIST_HEAD(&wq->pwqs);
Tejun Heo73f53c42010-06-29 10:07:11 +02004033 INIT_LIST_HEAD(&wq->flusher_queue);
4034 INIT_LIST_HEAD(&wq->flusher_overflow);
Tejun Heo493a1722013-03-12 11:29:59 -07004035 INIT_LIST_HEAD(&wq->maydays);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004036
Johannes Bergeb13ba82008-01-16 09:51:58 +01004037 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07004038 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004039
Tejun Heo30cdf242013-03-12 11:29:57 -07004040 if (alloc_and_link_pwqs(wq) < 0)
Tejun Heod2c1d402013-03-12 11:30:04 -07004041 goto err_free_wq;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004042
Tejun Heo493008a2013-03-12 11:30:03 -07004043 /*
4044 * Workqueues which may be used during memory reclaim should
4045 * have a rescuer to guarantee forward progress.
4046 */
4047 if (flags & WQ_MEM_RECLAIM) {
Tejun Heoe22bee72010-06-29 10:07:14 +02004048 struct worker *rescuer;
4049
Lai Jiangshanf7537df2014-07-15 17:24:15 +08004050 rescuer = alloc_worker(NUMA_NO_NODE);
Tejun Heoe22bee72010-06-29 10:07:14 +02004051 if (!rescuer)
Tejun Heod2c1d402013-03-12 11:30:04 -07004052 goto err_destroy;
Tejun Heoe22bee72010-06-29 10:07:14 +02004053
Tejun Heo111c2252013-01-17 17:16:24 -08004054 rescuer->rescue_wq = wq;
4055 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
Tejun Heob196be82012-01-10 15:11:35 -08004056 wq->name);
Tejun Heod2c1d402013-03-12 11:30:04 -07004057 if (IS_ERR(rescuer->task)) {
4058 kfree(rescuer);
4059 goto err_destroy;
4060 }
Tejun Heoe22bee72010-06-29 10:07:14 +02004061
Tejun Heod2c1d402013-03-12 11:30:04 -07004062 wq->rescuer = rescuer;
Peter Zijlstra25834c72015-05-15 17:43:34 +02004063 kthread_bind_mask(rescuer->task, cpu_possible_mask);
Tejun Heoe22bee72010-06-29 10:07:14 +02004064 wake_up_process(rescuer->task);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004065 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004066
Tejun Heo226223a2013-03-12 11:30:05 -07004067 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4068 goto err_destroy;
4069
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004070 /*
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004071 * wq_pool_mutex protects global freeze state and workqueues list.
4072 * Grab it, adjust max_active and add the new @wq to workqueues
4073 * list.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004074 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004075 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004076
Lai Jiangshana357fc02013-03-25 16:57:19 -07004077 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004078 for_each_pwq(pwq, wq)
4079 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004080 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004081
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004082 list_add_tail_rcu(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004083
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004084 mutex_unlock(&wq_pool_mutex);
Tejun Heo15376632010-06-29 10:07:11 +02004085
Oleg Nesterov3af244332007-05-09 02:34:09 -07004086 return wq;
Tejun Heod2c1d402013-03-12 11:30:04 -07004087
4088err_free_wq:
Tejun Heo6029a912013-04-01 11:23:34 -07004089 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heod2c1d402013-03-12 11:30:04 -07004090 kfree(wq);
4091 return NULL;
4092err_destroy:
4093 destroy_workqueue(wq);
Tejun Heo4690c4a2010-06-29 10:07:10 +02004094 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004095}
Tejun Heod320c032010-06-29 10:07:14 +02004096EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004097
4098/**
4099 * destroy_workqueue - safely terminate a workqueue
4100 * @wq: target workqueue
4101 *
4102 * Safely destroy a workqueue. All work currently pending will be done first.
4103 */
4104void destroy_workqueue(struct workqueue_struct *wq)
4105{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004106 struct pool_workqueue *pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004107 int node;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004108
Tejun Heo05905c22019-09-18 18:43:40 -07004109 /*
4110 * Remove it from sysfs first so that sanity check failure doesn't
4111 * lead to sysfs name conflicts.
4112 */
4113 workqueue_sysfs_unregister(wq);
4114
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02004115 /* drain it before proceeding with destruction */
4116 drain_workqueue(wq);
Tejun Heoc8efcc22010-12-20 19:32:04 +01004117
Tejun Heo05905c22019-09-18 18:43:40 -07004118 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4119 if (wq->rescuer) {
4120 struct worker *rescuer = wq->rescuer;
4121
4122 /* this prevents new queueing */
4123 spin_lock_irq(&wq_mayday_lock);
4124 wq->rescuer = NULL;
4125 spin_unlock_irq(&wq_mayday_lock);
4126
4127 /* rescuer will empty maydays list before exiting */
4128 kthread_stop(rescuer->task);
Tejun Heo80797bd2019-09-20 13:39:57 -07004129 kfree(rescuer);
Tejun Heo05905c22019-09-18 18:43:40 -07004130 }
4131
Tejun Heo6183c002013-03-12 11:29:57 -07004132 /* sanity checks */
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004133 mutex_lock(&wq->mutex);
Tejun Heo49e3cf42013-03-12 11:29:58 -07004134 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004135 int i;
4136
Tejun Heo76af4d92013-03-12 11:30:00 -07004137 for (i = 0; i < WORK_NR_COLORS; i++) {
4138 if (WARN_ON(pwq->nr_in_flight[i])) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004139 mutex_unlock(&wq->mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004140 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004141 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004142 }
4143 }
4144
Lai Jiangshan5c529592013-04-04 10:05:38 +08004145 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
Tejun Heo8864b4e2013-03-12 11:30:04 -07004146 WARN_ON(pwq->nr_active) ||
Tejun Heo76af4d92013-03-12 11:30:00 -07004147 WARN_ON(!list_empty(&pwq->delayed_works))) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004148 mutex_unlock(&wq->mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004149 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004150 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004151 }
Tejun Heo6183c002013-03-12 11:29:57 -07004152 }
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004153 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004154
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004155 /*
4156 * wq list is used to freeze wq, remove from list after
4157 * flushing is complete in case freeze races us.
4158 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004159 mutex_lock(&wq_pool_mutex);
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004160 list_del_rcu(&wq->list);
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004161 mutex_unlock(&wq_pool_mutex);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004162
Tejun Heo8864b4e2013-03-12 11:30:04 -07004163 if (!(wq->flags & WQ_UNBOUND)) {
4164 /*
4165 * The base ref is never dropped on per-cpu pwqs. Directly
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004166 * schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004167 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004168 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07004169 } else {
4170 /*
4171 * We're the sole accessor of @wq at this point. Directly
Tejun Heo4c16bd32013-04-01 11:23:36 -07004172 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4173 * @wq will be freed when the last pwq is released.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004174 */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004175 for_each_node(node) {
4176 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4177 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4178 put_pwq_unlocked(pwq);
4179 }
4180
4181 /*
4182 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4183 * put. Don't access it afterwards.
4184 */
4185 pwq = wq->dfl_pwq;
4186 wq->dfl_pwq = NULL;
Tejun Heodce90d42013-04-01 11:23:35 -07004187 put_pwq_unlocked(pwq);
Tejun Heo29c91e92013-03-12 11:30:03 -07004188 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004189}
4190EXPORT_SYMBOL_GPL(destroy_workqueue);
4191
Tejun Heodcd989c2010-06-29 10:07:14 +02004192/**
4193 * workqueue_set_max_active - adjust max_active of a workqueue
4194 * @wq: target workqueue
4195 * @max_active: new max_active value.
4196 *
4197 * Set max_active of @wq to @max_active.
4198 *
4199 * CONTEXT:
4200 * Don't call from IRQ context.
4201 */
4202void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4203{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004204 struct pool_workqueue *pwq;
Tejun Heodcd989c2010-06-29 10:07:14 +02004205
Tejun Heo8719dce2013-03-12 11:30:04 -07004206 /* disallow meddling with max_active for ordered workqueues */
Tejun Heo0a94efb2017-07-23 08:36:15 -04004207 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Tejun Heo8719dce2013-03-12 11:30:04 -07004208 return;
4209
Tejun Heof3421792010-07-02 10:03:51 +02004210 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02004211
Lai Jiangshana357fc02013-03-25 16:57:19 -07004212 mutex_lock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004213
Tejun Heo0a94efb2017-07-23 08:36:15 -04004214 wq->flags &= ~__WQ_ORDERED;
Tejun Heodcd989c2010-06-29 10:07:14 +02004215 wq->saved_max_active = max_active;
4216
Tejun Heo699ce092013-03-13 16:51:35 -07004217 for_each_pwq(pwq, wq)
4218 pwq_adjust_max_active(pwq);
Tejun Heodcd989c2010-06-29 10:07:14 +02004219
Lai Jiangshana357fc02013-03-25 16:57:19 -07004220 mutex_unlock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004221}
4222EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4223
4224/**
Lukas Wunner363e3fd2018-02-11 10:38:28 +01004225 * current_work - retrieve %current task's work struct
4226 *
4227 * Determine if %current task is a workqueue worker and what it's working on.
4228 * Useful to find out the context that the %current task is running in.
4229 *
4230 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4231 */
4232struct work_struct *current_work(void)
4233{
4234 struct worker *worker = current_wq_worker();
4235
4236 return worker ? worker->current_work : NULL;
4237}
4238EXPORT_SYMBOL(current_work);
4239
4240/**
Tejun Heoe62676162013-03-12 17:41:37 -07004241 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4242 *
4243 * Determine whether %current is a workqueue rescuer. Can be used from
4244 * work functions to determine whether it's being run off the rescuer task.
Yacine Belkadid185af32013-07-31 14:59:24 -07004245 *
4246 * Return: %true if %current is a workqueue rescuer. %false otherwise.
Tejun Heoe62676162013-03-12 17:41:37 -07004247 */
4248bool current_is_workqueue_rescuer(void)
4249{
4250 struct worker *worker = current_wq_worker();
4251
Lai Jiangshan6a092df2013-03-20 03:28:03 +08004252 return worker && worker->rescue_wq;
Tejun Heoe62676162013-03-12 17:41:37 -07004253}
4254
4255/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004256 * workqueue_congested - test whether a workqueue is congested
4257 * @cpu: CPU in question
4258 * @wq: target workqueue
4259 *
4260 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4261 * no synchronization around this function and the test result is
4262 * unreliable and only useful as advisory hints or for debugging.
4263 *
Tejun Heod3251852013-05-10 11:10:17 -07004264 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4265 * Note that both per-cpu and unbound workqueues may be associated with
4266 * multiple pool_workqueues which have separate congested states. A
4267 * workqueue being congested on one CPU doesn't mean the workqueue is also
4268 * contested on other CPUs / NUMA nodes.
4269 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004270 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004271 * %true if congested, %false otherwise.
4272 */
Tejun Heod84ff052013-03-12 11:29:59 -07004273bool workqueue_congested(int cpu, struct workqueue_struct *wq)
Tejun Heodcd989c2010-06-29 10:07:14 +02004274{
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004275 struct pool_workqueue *pwq;
Tejun Heo76af4d92013-03-12 11:30:00 -07004276 bool ret;
4277
Lai Jiangshan88109452013-03-20 03:28:10 +08004278 rcu_read_lock_sched();
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004279
Tejun Heod3251852013-05-10 11:10:17 -07004280 if (cpu == WORK_CPU_UNBOUND)
4281 cpu = smp_processor_id();
4282
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004283 if (!(wq->flags & WQ_UNBOUND))
4284 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4285 else
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004286 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodcd989c2010-06-29 10:07:14 +02004287
Tejun Heo76af4d92013-03-12 11:30:00 -07004288 ret = !list_empty(&pwq->delayed_works);
Lai Jiangshan88109452013-03-20 03:28:10 +08004289 rcu_read_unlock_sched();
Tejun Heo76af4d92013-03-12 11:30:00 -07004290
4291 return ret;
Tejun Heodcd989c2010-06-29 10:07:14 +02004292}
4293EXPORT_SYMBOL_GPL(workqueue_congested);
4294
4295/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004296 * work_busy - test whether a work is currently pending or running
4297 * @work: the work to be tested
4298 *
4299 * Test whether @work is currently pending or running. There is no
4300 * synchronization around this function and the test result is
4301 * unreliable and only useful as advisory hints or for debugging.
Tejun Heodcd989c2010-06-29 10:07:14 +02004302 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004303 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004304 * OR'd bitmask of WORK_BUSY_* bits.
4305 */
4306unsigned int work_busy(struct work_struct *work)
4307{
Tejun Heofa1b54e2013-03-12 11:30:00 -07004308 struct worker_pool *pool;
Tejun Heodcd989c2010-06-29 10:07:14 +02004309 unsigned long flags;
4310 unsigned int ret = 0;
4311
Tejun Heodcd989c2010-06-29 10:07:14 +02004312 if (work_pending(work))
4313 ret |= WORK_BUSY_PENDING;
Tejun Heodcd989c2010-06-29 10:07:14 +02004314
Tejun Heofa1b54e2013-03-12 11:30:00 -07004315 local_irq_save(flags);
4316 pool = get_work_pool(work);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004317 if (pool) {
Tejun Heofa1b54e2013-03-12 11:30:00 -07004318 spin_lock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004319 if (find_worker_executing_work(pool, work))
4320 ret |= WORK_BUSY_RUNNING;
Tejun Heofa1b54e2013-03-12 11:30:00 -07004321 spin_unlock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004322 }
Tejun Heofa1b54e2013-03-12 11:30:00 -07004323 local_irq_restore(flags);
Tejun Heodcd989c2010-06-29 10:07:14 +02004324
4325 return ret;
4326}
4327EXPORT_SYMBOL_GPL(work_busy);
4328
Tejun Heo3d1cb202013-04-30 15:27:22 -07004329/**
4330 * set_worker_desc - set description for the current work item
4331 * @fmt: printf-style format string
4332 * @...: arguments for the format string
4333 *
4334 * This function can be called by a running work function to describe what
4335 * the work item is about. If the worker task gets dumped, this
4336 * information will be printed out together to help debugging. The
4337 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4338 */
4339void set_worker_desc(const char *fmt, ...)
4340{
4341 struct worker *worker = current_wq_worker();
4342 va_list args;
4343
4344 if (worker) {
4345 va_start(args, fmt);
4346 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4347 va_end(args);
4348 worker->desc_valid = true;
4349 }
4350}
4351
4352/**
4353 * print_worker_info - print out worker information and description
4354 * @log_lvl: the log level to use when printing
4355 * @task: target task
4356 *
4357 * If @task is a worker and currently executing a work item, print out the
4358 * name of the workqueue being serviced and worker description set with
4359 * set_worker_desc() by the currently executing work item.
4360 *
4361 * This function can be safely called on any task as long as the
4362 * task_struct itself is accessible. While safe, this function isn't
4363 * synchronized and may print out mixups or garbages of limited length.
4364 */
4365void print_worker_info(const char *log_lvl, struct task_struct *task)
4366{
4367 work_func_t *fn = NULL;
4368 char name[WQ_NAME_LEN] = { };
4369 char desc[WORKER_DESC_LEN] = { };
4370 struct pool_workqueue *pwq = NULL;
4371 struct workqueue_struct *wq = NULL;
4372 bool desc_valid = false;
4373 struct worker *worker;
4374
4375 if (!(task->flags & PF_WQ_WORKER))
4376 return;
4377
4378 /*
4379 * This function is called without any synchronization and @task
4380 * could be in any state. Be careful with dereferences.
4381 */
Petr Mladeke7005912016-10-11 13:55:17 -07004382 worker = kthread_probe_data(task);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004383
4384 /*
4385 * Carefully copy the associated workqueue's workfn and name. Keep
4386 * the original last '\0' in case the original contains garbage.
4387 */
4388 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4389 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4390 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4391 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4392
4393 /* copy worker description */
4394 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4395 if (desc_valid)
4396 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4397
4398 if (fn || name[0] || desc[0]) {
4399 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4400 if (desc[0])
4401 pr_cont(" (%s)", desc);
4402 pr_cont("\n");
4403 }
4404}
4405
Tejun Heo3494fc32015-03-09 09:22:28 -04004406static void pr_cont_pool_info(struct worker_pool *pool)
4407{
4408 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4409 if (pool->node != NUMA_NO_NODE)
4410 pr_cont(" node=%d", pool->node);
4411 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4412}
4413
4414static void pr_cont_work(bool comma, struct work_struct *work)
4415{
4416 if (work->func == wq_barrier_func) {
4417 struct wq_barrier *barr;
4418
4419 barr = container_of(work, struct wq_barrier, work);
4420
4421 pr_cont("%s BAR(%d)", comma ? "," : "",
4422 task_pid_nr(barr->task));
4423 } else {
4424 pr_cont("%s %pf", comma ? "," : "", work->func);
4425 }
4426}
4427
4428static void show_pwq(struct pool_workqueue *pwq)
4429{
4430 struct worker_pool *pool = pwq->pool;
4431 struct work_struct *work;
4432 struct worker *worker;
4433 bool has_in_flight = false, has_pending = false;
4434 int bkt;
4435
4436 pr_info(" pwq %d:", pool->id);
4437 pr_cont_pool_info(pool);
4438
Tejun Heo9e0f33d2019-09-25 06:59:15 -07004439 pr_cont(" active=%d/%d refcnt=%d%s\n",
4440 pwq->nr_active, pwq->max_active, pwq->refcnt,
Tejun Heo3494fc32015-03-09 09:22:28 -04004441 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4442
4443 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4444 if (worker->current_pwq == pwq) {
4445 has_in_flight = true;
4446 break;
4447 }
4448 }
4449 if (has_in_flight) {
4450 bool comma = false;
4451
4452 pr_info(" in-flight:");
4453 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4454 if (worker->current_pwq != pwq)
4455 continue;
4456
4457 pr_cont("%s %d%s:%pf", comma ? "," : "",
4458 task_pid_nr(worker->task),
4459 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4460 worker->current_func);
4461 list_for_each_entry(work, &worker->scheduled, entry)
4462 pr_cont_work(false, work);
4463 comma = true;
4464 }
4465 pr_cont("\n");
4466 }
4467
4468 list_for_each_entry(work, &pool->worklist, entry) {
4469 if (get_work_pwq(work) == pwq) {
4470 has_pending = true;
4471 break;
4472 }
4473 }
4474 if (has_pending) {
4475 bool comma = false;
4476
4477 pr_info(" pending:");
4478 list_for_each_entry(work, &pool->worklist, entry) {
4479 if (get_work_pwq(work) != pwq)
4480 continue;
4481
4482 pr_cont_work(comma, work);
4483 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4484 }
4485 pr_cont("\n");
4486 }
4487
4488 if (!list_empty(&pwq->delayed_works)) {
4489 bool comma = false;
4490
4491 pr_info(" delayed:");
4492 list_for_each_entry(work, &pwq->delayed_works, entry) {
4493 pr_cont_work(comma, work);
4494 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4495 }
4496 pr_cont("\n");
4497 }
4498}
4499
4500/**
4501 * show_workqueue_state - dump workqueue state
4502 *
Roger Lu7b776af2016-07-01 11:05:02 +08004503 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4504 * all busy workqueues and pools.
Tejun Heo3494fc32015-03-09 09:22:28 -04004505 */
4506void show_workqueue_state(void)
4507{
4508 struct workqueue_struct *wq;
4509 struct worker_pool *pool;
4510 unsigned long flags;
4511 int pi;
4512
4513 rcu_read_lock_sched();
4514
4515 pr_info("Showing busy workqueues and worker pools:\n");
4516
4517 list_for_each_entry_rcu(wq, &workqueues, list) {
4518 struct pool_workqueue *pwq;
4519 bool idle = true;
4520
4521 for_each_pwq(pwq, wq) {
4522 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4523 idle = false;
4524 break;
4525 }
4526 }
4527 if (idle)
4528 continue;
4529
4530 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4531
4532 for_each_pwq(pwq, wq) {
4533 spin_lock_irqsave(&pwq->pool->lock, flags);
4534 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4535 show_pwq(pwq);
4536 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Sergey Senozhatskya6d59302018-01-11 09:53:35 +09004537 /*
4538 * We could be printing a lot from atomic context, e.g.
4539 * sysrq-t -> show_workqueue_state(). Avoid triggering
4540 * hard lockup.
4541 */
4542 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004543 }
4544 }
4545
4546 for_each_pool(pool, pi) {
4547 struct worker *worker;
4548 bool first = true;
4549
4550 spin_lock_irqsave(&pool->lock, flags);
4551 if (pool->nr_workers == pool->nr_idle)
4552 goto next_pool;
4553
4554 pr_info("pool %d:", pool->id);
4555 pr_cont_pool_info(pool);
Tejun Heo82607adc2015-12-08 11:28:04 -05004556 pr_cont(" hung=%us workers=%d",
4557 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4558 pool->nr_workers);
Tejun Heo3494fc32015-03-09 09:22:28 -04004559 if (pool->manager)
4560 pr_cont(" manager: %d",
4561 task_pid_nr(pool->manager->task));
4562 list_for_each_entry(worker, &pool->idle_list, entry) {
4563 pr_cont(" %s%d", first ? "idle: " : "",
4564 task_pid_nr(worker->task));
4565 first = false;
4566 }
4567 pr_cont("\n");
4568 next_pool:
4569 spin_unlock_irqrestore(&pool->lock, flags);
Sergey Senozhatskya6d59302018-01-11 09:53:35 +09004570 /*
4571 * We could be printing a lot from atomic context, e.g.
4572 * sysrq-t -> show_workqueue_state(). Avoid triggering
4573 * hard lockup.
4574 */
4575 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004576 }
4577
4578 rcu_read_unlock_sched();
4579}
4580
Tejun Heodb7bccf2010-06-29 10:07:12 +02004581/*
4582 * CPU hotplug.
4583 *
Tejun Heoe22bee72010-06-29 10:07:14 +02004584 * There are two challenges in supporting CPU hotplug. Firstly, there
Tejun Heo112202d2013-02-13 19:29:12 -08004585 * are a lot of assumptions on strong associations among work, pwq and
Tejun Heo706026c2013-01-24 11:01:34 -08004586 * pool which make migrating pending and scheduled works very
Tejun Heoe22bee72010-06-29 10:07:14 +02004587 * difficult to implement without impacting hot paths. Secondly,
Tejun Heo94cf58b2013-01-24 11:01:33 -08004588 * worker pools serve mix of short, long and very long running works making
Tejun Heoe22bee72010-06-29 10:07:14 +02004589 * blocked draining impractical.
4590 *
Tejun Heo24647572013-01-24 11:01:33 -08004591 * This is solved by allowing the pools to be disassociated from the CPU
Tejun Heo628c78e2012-07-17 12:39:27 -07004592 * running as an unbound one and allowing it to be reattached later if the
4593 * cpu comes back online.
Tejun Heodb7bccf2010-06-29 10:07:12 +02004594 */
4595
Tejun Heo706026c2013-01-24 11:01:34 -08004596static void wq_unbind_fn(struct work_struct *work)
Tejun Heodb7bccf2010-06-29 10:07:12 +02004597{
Tejun Heo38db41d2013-01-24 11:01:34 -08004598 int cpu = smp_processor_id();
Tejun Heo4ce62e92012-07-13 22:16:44 -07004599 struct worker_pool *pool;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004600 struct worker *worker;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004601
Tejun Heof02ae732013-03-12 11:30:03 -07004602 for_each_cpu_worker_pool(pool, cpu) {
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004603 mutex_lock(&pool->attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004604 spin_lock_irq(&pool->lock);
4605
4606 /*
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004607 * We've blocked all attach/detach operations. Make all workers
Tejun Heo94cf58b2013-01-24 11:01:33 -08004608 * unbound and set DISASSOCIATED. Before this, all workers
4609 * except for the ones which are still executing works from
4610 * before the last CPU down must be on the cpu. After
4611 * this, they may become diasporas.
4612 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004613 for_each_pool_worker(worker, pool)
Tejun Heoc9e7cf22013-01-24 11:01:33 -08004614 worker->flags |= WORKER_UNBOUND;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004615
Tejun Heo24647572013-01-24 11:01:33 -08004616 pool->flags |= POOL_DISASSOCIATED;
Tejun Heof2d5a0e2012-07-17 12:39:26 -07004617
Tejun Heo94cf58b2013-01-24 11:01:33 -08004618 spin_unlock_irq(&pool->lock);
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004619 mutex_unlock(&pool->attach_mutex);
Tejun Heoe22bee72010-06-29 10:07:14 +02004620
Lai Jiangshaneb283422013-03-08 15:18:28 -08004621 /*
4622 * Call schedule() so that we cross rq->lock and thus can
4623 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4624 * This is necessary as scheduler callbacks may be invoked
4625 * from other cpus.
4626 */
4627 schedule();
Tejun Heo628c78e2012-07-17 12:39:27 -07004628
Lai Jiangshaneb283422013-03-08 15:18:28 -08004629 /*
4630 * Sched callbacks are disabled now. Zap nr_running.
4631 * After this, nr_running stays zero and need_more_worker()
4632 * and keep_working() are always true as long as the
4633 * worklist is not empty. This pool now behaves as an
4634 * unbound (in terms of concurrency management) pool which
4635 * are served by workers tied to the pool.
4636 */
Tejun Heoe19e3972013-01-24 11:39:44 -08004637 atomic_set(&pool->nr_running, 0);
Lai Jiangshaneb283422013-03-08 15:18:28 -08004638
4639 /*
4640 * With concurrency management just turned off, a busy
4641 * worker blocking could lead to lengthy stalls. Kick off
4642 * unbound chain execution of currently pending work items.
4643 */
4644 spin_lock_irq(&pool->lock);
4645 wake_up_worker(pool);
4646 spin_unlock_irq(&pool->lock);
4647 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004648}
4649
Tejun Heobd7c0892013-03-19 13:45:21 -07004650/**
4651 * rebind_workers - rebind all workers of a pool to the associated CPU
4652 * @pool: pool of interest
4653 *
Tejun Heoa9ab7752013-03-19 13:45:21 -07004654 * @pool->cpu is coming online. Rebind all workers to the CPU.
Tejun Heobd7c0892013-03-19 13:45:21 -07004655 */
4656static void rebind_workers(struct worker_pool *pool)
4657{
Tejun Heoa9ab7752013-03-19 13:45:21 -07004658 struct worker *worker;
Tejun Heobd7c0892013-03-19 13:45:21 -07004659
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004660 lockdep_assert_held(&pool->attach_mutex);
Tejun Heobd7c0892013-03-19 13:45:21 -07004661
Tejun Heoa9ab7752013-03-19 13:45:21 -07004662 /*
4663 * Restore CPU affinity of all workers. As all idle workers should
4664 * be on the run-queue of the associated CPU before any local
Shailendra Verma402dd892015-05-23 10:38:14 +05304665 * wake-ups for concurrency management happen, restore CPU affinity
Tejun Heoa9ab7752013-03-19 13:45:21 -07004666 * of all workers first and then clear UNBOUND. As we're called
4667 * from CPU_ONLINE, the following shouldn't fail.
4668 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004669 for_each_pool_worker(worker, pool)
Tejun Heoa9ab7752013-03-19 13:45:21 -07004670 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4671 pool->attrs->cpumask) < 0);
4672
4673 spin_lock_irq(&pool->lock);
Wanpeng Lif7c17d22016-05-11 17:55:18 +08004674
4675 /*
4676 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4677 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4678 * being reworked and this can go away in time.
4679 */
4680 if (!(pool->flags & POOL_DISASSOCIATED)) {
4681 spin_unlock_irq(&pool->lock);
4682 return;
4683 }
4684
Lai Jiangshan3de5e882014-06-03 15:33:27 +08004685 pool->flags &= ~POOL_DISASSOCIATED;
Tejun Heoa9ab7752013-03-19 13:45:21 -07004686
Lai Jiangshanda028462014-05-20 17:46:31 +08004687 for_each_pool_worker(worker, pool) {
Tejun Heoa9ab7752013-03-19 13:45:21 -07004688 unsigned int worker_flags = worker->flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004689
4690 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07004691 * A bound idle worker should actually be on the runqueue
4692 * of the associated CPU for local wake-ups targeting it to
4693 * work. Kick all idle workers so that they migrate to the
4694 * associated CPU. Doing this in the same loop as
4695 * replacing UNBOUND with REBOUND is safe as no worker will
4696 * be bound before @pool->lock is released.
Tejun Heobd7c0892013-03-19 13:45:21 -07004697 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004698 if (worker_flags & WORKER_IDLE)
4699 wake_up_process(worker->task);
4700
4701 /*
4702 * We want to clear UNBOUND but can't directly call
4703 * worker_clr_flags() or adjust nr_running. Atomically
4704 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4705 * @worker will clear REBOUND using worker_clr_flags() when
4706 * it initiates the next execution cycle thus restoring
4707 * concurrency management. Note that when or whether
4708 * @worker clears REBOUND doesn't affect correctness.
4709 *
4710 * ACCESS_ONCE() is necessary because @worker->flags may be
4711 * tested without holding any lock in
4712 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4713 * fail incorrectly leading to premature concurrency
4714 * management operations.
4715 */
4716 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4717 worker_flags |= WORKER_REBOUND;
4718 worker_flags &= ~WORKER_UNBOUND;
4719 ACCESS_ONCE(worker->flags) = worker_flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004720 }
4721
Tejun Heoa9ab7752013-03-19 13:45:21 -07004722 spin_unlock_irq(&pool->lock);
Tejun Heobd7c0892013-03-19 13:45:21 -07004723}
4724
Tejun Heo7dbc7252013-03-19 13:45:21 -07004725/**
4726 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4727 * @pool: unbound pool of interest
4728 * @cpu: the CPU which is coming up
4729 *
4730 * An unbound pool may end up with a cpumask which doesn't have any online
4731 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4732 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4733 * online CPU before, cpus_allowed of all its workers should be restored.
4734 */
4735static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4736{
4737 static cpumask_t cpumask;
4738 struct worker *worker;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004739
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004740 lockdep_assert_held(&pool->attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004741
4742 /* is @cpu allowed for @pool? */
4743 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4744 return;
4745
Tejun Heo7dbc7252013-03-19 13:45:21 -07004746 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004747
4748 /* as we're called from CPU_ONLINE, the following shouldn't fail */
Lai Jiangshanda028462014-05-20 17:46:31 +08004749 for_each_pool_worker(worker, pool)
Peter Zijlstrad945b5e2016-06-16 14:38:42 +02004750 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004751}
4752
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004753int workqueue_prepare_cpu(unsigned int cpu)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004754{
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004755 struct worker_pool *pool;
4756
4757 for_each_cpu_worker_pool(pool, cpu) {
4758 if (pool->nr_workers)
4759 continue;
4760 if (!create_worker(pool))
4761 return -ENOMEM;
4762 }
4763 return 0;
4764}
4765
4766int workqueue_online_cpu(unsigned int cpu)
4767{
Tejun Heo4ce62e92012-07-13 22:16:44 -07004768 struct worker_pool *pool;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004769 struct workqueue_struct *wq;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004770 int pi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004771
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004772 mutex_lock(&wq_pool_mutex);
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07004773
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004774 for_each_pool(pool, pi) {
4775 mutex_lock(&pool->attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004776
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004777 if (pool->cpu == cpu)
4778 rebind_workers(pool);
4779 else if (pool->cpu < 0)
4780 restore_unbound_workers_cpumask(pool, cpu);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004781
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004782 mutex_unlock(&pool->attach_mutex);
Tejun Heo65758202012-07-17 12:39:26 -07004783 }
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004784
4785 /* update NUMA affinity of unbound workqueues */
4786 list_for_each_entry(wq, &workqueues, list)
4787 wq_update_unbound_numa(wq, cpu, true);
4788
4789 mutex_unlock(&wq_pool_mutex);
4790 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07004791}
4792
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004793int workqueue_offline_cpu(unsigned int cpu)
Tejun Heo65758202012-07-17 12:39:26 -07004794{
Tejun Heo8db25e72012-07-17 12:39:28 -07004795 struct work_struct unbind_work;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004796 struct workqueue_struct *wq;
Tejun Heo8db25e72012-07-17 12:39:28 -07004797
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004798 /* unbinding per-cpu workers should happen on the local CPU */
4799 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4800 queue_work_on(cpu, system_highpri_wq, &unbind_work);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004801
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004802 /* update NUMA affinity of unbound workqueues */
4803 mutex_lock(&wq_pool_mutex);
4804 list_for_each_entry(wq, &workqueues, list)
4805 wq_update_unbound_numa(wq, cpu, false);
4806 mutex_unlock(&wq_pool_mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004807
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004808 /* wait for per-cpu unbinding to finish */
4809 flush_work(&unbind_work);
4810 destroy_work_on_stack(&unbind_work);
4811 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07004812}
4813
Rusty Russell2d3854a2008-11-05 13:39:10 +11004814#ifdef CONFIG_SMP
Rusty Russell8ccad402009-01-16 15:31:15 -08004815
Rusty Russell2d3854a2008-11-05 13:39:10 +11004816struct work_for_cpu {
Tejun Heoed48ece2012-09-18 12:48:43 -07004817 struct work_struct work;
Rusty Russell2d3854a2008-11-05 13:39:10 +11004818 long (*fn)(void *);
4819 void *arg;
4820 long ret;
4821};
4822
Tejun Heoed48ece2012-09-18 12:48:43 -07004823static void work_for_cpu_fn(struct work_struct *work)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004824{
Tejun Heoed48ece2012-09-18 12:48:43 -07004825 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4826
Rusty Russell2d3854a2008-11-05 13:39:10 +11004827 wfc->ret = wfc->fn(wfc->arg);
4828}
4829
4830/**
Anna-Maria Gleixner22aceb32016-03-10 12:07:38 +01004831 * work_on_cpu - run a function in thread context on a particular cpu
Rusty Russell2d3854a2008-11-05 13:39:10 +11004832 * @cpu: the cpu to run on
4833 * @fn: the function to run
4834 * @arg: the function arg
4835 *
Rusty Russell31ad9082009-01-16 15:31:15 -08004836 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06004837 * The caller must not hold any locks which would prevent @fn from completing.
Yacine Belkadid185af32013-07-31 14:59:24 -07004838 *
4839 * Return: The value @fn returns.
Rusty Russell2d3854a2008-11-05 13:39:10 +11004840 */
Tejun Heod84ff052013-03-12 11:29:59 -07004841long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004842{
Tejun Heoed48ece2012-09-18 12:48:43 -07004843 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
Rusty Russell2d3854a2008-11-05 13:39:10 +11004844
Tejun Heoed48ece2012-09-18 12:48:43 -07004845 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4846 schedule_work_on(cpu, &wfc.work);
Bjorn Helgaas12997d12013-11-18 11:00:29 -07004847 flush_work(&wfc.work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05004848 destroy_work_on_stack(&wfc.work);
Rusty Russell2d3854a2008-11-05 13:39:10 +11004849 return wfc.ret;
4850}
4851EXPORT_SYMBOL_GPL(work_on_cpu);
Thomas Gleixner0e8d6a92017-04-12 22:07:28 +02004852
4853/**
4854 * work_on_cpu_safe - run a function in thread context on a particular cpu
4855 * @cpu: the cpu to run on
4856 * @fn: the function to run
4857 * @arg: the function argument
4858 *
4859 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4860 * any locks which would prevent @fn from completing.
4861 *
4862 * Return: The value @fn returns.
4863 */
4864long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4865{
4866 long ret = -ENODEV;
4867
4868 get_online_cpus();
4869 if (cpu_online(cpu))
4870 ret = work_on_cpu(cpu, fn, arg);
4871 put_online_cpus();
4872 return ret;
4873}
4874EXPORT_SYMBOL_GPL(work_on_cpu_safe);
Rusty Russell2d3854a2008-11-05 13:39:10 +11004875#endif /* CONFIG_SMP */
4876
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004877#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10304878
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004879/**
4880 * freeze_workqueues_begin - begin freezing workqueues
4881 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01004882 * Start freezing workqueues. After this function returns, all freezable
Tejun Heoc5aa87b2013-03-13 16:51:36 -07004883 * workqueues will queue new works to their delayed_works list instead of
Tejun Heo706026c2013-01-24 11:01:34 -08004884 * pool->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004885 *
4886 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004887 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004888 */
4889void freeze_workqueues_begin(void)
4890{
Tejun Heo24b8a842013-03-12 11:29:58 -07004891 struct workqueue_struct *wq;
4892 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004893
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004894 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004895
Tejun Heo6183c002013-03-12 11:29:57 -07004896 WARN_ON_ONCE(workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004897 workqueue_freezing = true;
4898
Tejun Heo24b8a842013-03-12 11:29:58 -07004899 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004900 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004901 for_each_pwq(pwq, wq)
4902 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004903 mutex_unlock(&wq->mutex);
Tejun Heo24b8a842013-03-12 11:29:58 -07004904 }
Tejun Heo5bcab332013-03-13 19:47:40 -07004905
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004906 mutex_unlock(&wq_pool_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004907}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004908
4909/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01004910 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004911 *
4912 * Check whether freezing is complete. This function must be called
4913 * between freeze_workqueues_begin() and thaw_workqueues().
4914 *
4915 * CONTEXT:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004916 * Grabs and releases wq_pool_mutex.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004917 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004918 * Return:
Tejun Heo58a69cb2011-02-16 09:25:31 +01004919 * %true if some freezable workqueues are still busy. %false if freezing
4920 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004921 */
4922bool freeze_workqueues_busy(void)
4923{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004924 bool busy = false;
Tejun Heo24b8a842013-03-12 11:29:58 -07004925 struct workqueue_struct *wq;
4926 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004927
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004928 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004929
Tejun Heo6183c002013-03-12 11:29:57 -07004930 WARN_ON_ONCE(!workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004931
Tejun Heo24b8a842013-03-12 11:29:58 -07004932 list_for_each_entry(wq, &workqueues, list) {
4933 if (!(wq->flags & WQ_FREEZABLE))
4934 continue;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004935 /*
4936 * nr_active is monotonically decreasing. It's safe
4937 * to peek without lock.
4938 */
Lai Jiangshan88109452013-03-20 03:28:10 +08004939 rcu_read_lock_sched();
Tejun Heo24b8a842013-03-12 11:29:58 -07004940 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004941 WARN_ON_ONCE(pwq->nr_active < 0);
Tejun Heo112202d2013-02-13 19:29:12 -08004942 if (pwq->nr_active) {
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004943 busy = true;
Lai Jiangshan88109452013-03-20 03:28:10 +08004944 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004945 goto out_unlock;
4946 }
4947 }
Lai Jiangshan88109452013-03-20 03:28:10 +08004948 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004949 }
4950out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004951 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004952 return busy;
4953}
4954
4955/**
4956 * thaw_workqueues - thaw workqueues
4957 *
4958 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo706026c2013-01-24 11:01:34 -08004959 * frozen works are transferred to their respective pool worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004960 *
4961 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004962 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004963 */
4964void thaw_workqueues(void)
4965{
Tejun Heo24b8a842013-03-12 11:29:58 -07004966 struct workqueue_struct *wq;
4967 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004968
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004969 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004970
4971 if (!workqueue_freezing)
4972 goto out_unlock;
4973
Lai Jiangshan74b414e2014-05-22 19:01:16 +08004974 workqueue_freezing = false;
Tejun Heo8b03ae32010-06-29 10:07:12 +02004975
Tejun Heo24b8a842013-03-12 11:29:58 -07004976 /* restore max_active and repopulate worklist */
4977 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004978 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004979 for_each_pwq(pwq, wq)
4980 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004981 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004982 }
4983
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004984out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004985 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004986}
4987#endif /* CONFIG_FREEZER */
4988
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004989static int workqueue_apply_unbound_cpumask(void)
4990{
4991 LIST_HEAD(ctxs);
4992 int ret = 0;
4993 struct workqueue_struct *wq;
4994 struct apply_wqattrs_ctx *ctx, *n;
4995
4996 lockdep_assert_held(&wq_pool_mutex);
4997
4998 list_for_each_entry(wq, &workqueues, list) {
4999 if (!(wq->flags & WQ_UNBOUND))
5000 continue;
5001 /* creating multiple pwqs breaks ordering guarantee */
5002 if (wq->flags & __WQ_ORDERED)
5003 continue;
5004
5005 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5006 if (!ctx) {
5007 ret = -ENOMEM;
5008 break;
5009 }
5010
5011 list_add_tail(&ctx->list, &ctxs);
5012 }
5013
5014 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5015 if (!ret)
5016 apply_wqattrs_commit(ctx);
5017 apply_wqattrs_cleanup(ctx);
5018 }
5019
5020 return ret;
5021}
5022
5023/**
5024 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5025 * @cpumask: the cpumask to set
5026 *
5027 * The low-level workqueues cpumask is a global cpumask that limits
5028 * the affinity of all unbound workqueues. This function check the @cpumask
5029 * and apply it to all unbound workqueues and updates all pwqs of them.
5030 *
5031 * Retun: 0 - Success
5032 * -EINVAL - Invalid @cpumask
5033 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5034 */
5035int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5036{
5037 int ret = -EINVAL;
5038 cpumask_var_t saved_cpumask;
5039
5040 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5041 return -ENOMEM;
5042
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005043 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5044 if (!cpumask_empty(cpumask)) {
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005045 apply_wqattrs_lock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005046
5047 /* save the old wq_unbound_cpumask. */
5048 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5049
5050 /* update wq_unbound_cpumask at first and apply it to wqs. */
5051 cpumask_copy(wq_unbound_cpumask, cpumask);
5052 ret = workqueue_apply_unbound_cpumask();
5053
5054 /* restore the wq_unbound_cpumask when failed. */
5055 if (ret < 0)
5056 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5057
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005058 apply_wqattrs_unlock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005059 }
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005060
5061 free_cpumask_var(saved_cpumask);
5062 return ret;
5063}
5064
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005065#ifdef CONFIG_SYSFS
5066/*
5067 * Workqueues with WQ_SYSFS flag set is visible to userland via
5068 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5069 * following attributes.
5070 *
5071 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5072 * max_active RW int : maximum number of in-flight work items
5073 *
5074 * Unbound workqueues have the following extra attributes.
5075 *
5076 * id RO int : the associated pool ID
5077 * nice RW int : nice value of the workers
5078 * cpumask RW mask : bitmask of allowed CPUs for the workers
5079 */
5080struct wq_device {
5081 struct workqueue_struct *wq;
5082 struct device dev;
5083};
5084
5085static struct workqueue_struct *dev_to_wq(struct device *dev)
5086{
5087 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5088
5089 return wq_dev->wq;
5090}
5091
5092static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5093 char *buf)
5094{
5095 struct workqueue_struct *wq = dev_to_wq(dev);
5096
5097 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5098}
5099static DEVICE_ATTR_RO(per_cpu);
5100
5101static ssize_t max_active_show(struct device *dev,
5102 struct device_attribute *attr, char *buf)
5103{
5104 struct workqueue_struct *wq = dev_to_wq(dev);
5105
5106 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5107}
5108
5109static ssize_t max_active_store(struct device *dev,
5110 struct device_attribute *attr, const char *buf,
5111 size_t count)
5112{
5113 struct workqueue_struct *wq = dev_to_wq(dev);
5114 int val;
5115
5116 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5117 return -EINVAL;
5118
5119 workqueue_set_max_active(wq, val);
5120 return count;
5121}
5122static DEVICE_ATTR_RW(max_active);
5123
5124static struct attribute *wq_sysfs_attrs[] = {
5125 &dev_attr_per_cpu.attr,
5126 &dev_attr_max_active.attr,
5127 NULL,
5128};
5129ATTRIBUTE_GROUPS(wq_sysfs);
5130
5131static ssize_t wq_pool_ids_show(struct device *dev,
5132 struct device_attribute *attr, char *buf)
5133{
5134 struct workqueue_struct *wq = dev_to_wq(dev);
5135 const char *delim = "";
5136 int node, written = 0;
5137
5138 rcu_read_lock_sched();
5139 for_each_node(node) {
5140 written += scnprintf(buf + written, PAGE_SIZE - written,
5141 "%s%d:%d", delim, node,
5142 unbound_pwq_by_node(wq, node)->pool->id);
5143 delim = " ";
5144 }
5145 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5146 rcu_read_unlock_sched();
5147
5148 return written;
5149}
5150
5151static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5152 char *buf)
5153{
5154 struct workqueue_struct *wq = dev_to_wq(dev);
5155 int written;
5156
5157 mutex_lock(&wq->mutex);
5158 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5159 mutex_unlock(&wq->mutex);
5160
5161 return written;
5162}
5163
5164/* prepare workqueue_attrs for sysfs store operations */
5165static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5166{
5167 struct workqueue_attrs *attrs;
5168
Lai Jiangshan899a94f2015-05-20 14:41:18 +08005169 lockdep_assert_held(&wq_pool_mutex);
5170
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005171 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5172 if (!attrs)
5173 return NULL;
5174
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005175 copy_workqueue_attrs(attrs, wq->unbound_attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005176 return attrs;
5177}
5178
5179static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5180 const char *buf, size_t count)
5181{
5182 struct workqueue_struct *wq = dev_to_wq(dev);
5183 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005184 int ret = -ENOMEM;
5185
5186 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005187
5188 attrs = wq_sysfs_prep_attrs(wq);
5189 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005190 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005191
5192 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5193 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005194 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005195 else
5196 ret = -EINVAL;
5197
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005198out_unlock:
5199 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005200 free_workqueue_attrs(attrs);
5201 return ret ?: count;
5202}
5203
5204static ssize_t wq_cpumask_show(struct device *dev,
5205 struct device_attribute *attr, char *buf)
5206{
5207 struct workqueue_struct *wq = dev_to_wq(dev);
5208 int written;
5209
5210 mutex_lock(&wq->mutex);
5211 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5212 cpumask_pr_args(wq->unbound_attrs->cpumask));
5213 mutex_unlock(&wq->mutex);
5214 return written;
5215}
5216
5217static ssize_t wq_cpumask_store(struct device *dev,
5218 struct device_attribute *attr,
5219 const char *buf, size_t count)
5220{
5221 struct workqueue_struct *wq = dev_to_wq(dev);
5222 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005223 int ret = -ENOMEM;
5224
5225 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005226
5227 attrs = wq_sysfs_prep_attrs(wq);
5228 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005229 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005230
5231 ret = cpumask_parse(buf, attrs->cpumask);
5232 if (!ret)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005233 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005234
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005235out_unlock:
5236 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005237 free_workqueue_attrs(attrs);
5238 return ret ?: count;
5239}
5240
5241static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5242 char *buf)
5243{
5244 struct workqueue_struct *wq = dev_to_wq(dev);
5245 int written;
5246
5247 mutex_lock(&wq->mutex);
5248 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5249 !wq->unbound_attrs->no_numa);
5250 mutex_unlock(&wq->mutex);
5251
5252 return written;
5253}
5254
5255static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5256 const char *buf, size_t count)
5257{
5258 struct workqueue_struct *wq = dev_to_wq(dev);
5259 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005260 int v, ret = -ENOMEM;
5261
5262 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005263
5264 attrs = wq_sysfs_prep_attrs(wq);
5265 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005266 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005267
5268 ret = -EINVAL;
5269 if (sscanf(buf, "%d", &v) == 1) {
5270 attrs->no_numa = !v;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005271 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005272 }
5273
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005274out_unlock:
5275 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005276 free_workqueue_attrs(attrs);
5277 return ret ?: count;
5278}
5279
5280static struct device_attribute wq_sysfs_unbound_attrs[] = {
5281 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5282 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5283 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5284 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5285 __ATTR_NULL,
5286};
5287
5288static struct bus_type wq_subsys = {
5289 .name = "workqueue",
5290 .dev_groups = wq_sysfs_groups,
5291};
5292
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005293static ssize_t wq_unbound_cpumask_show(struct device *dev,
5294 struct device_attribute *attr, char *buf)
5295{
5296 int written;
5297
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005298 mutex_lock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005299 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5300 cpumask_pr_args(wq_unbound_cpumask));
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005301 mutex_unlock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005302
5303 return written;
5304}
5305
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005306static ssize_t wq_unbound_cpumask_store(struct device *dev,
5307 struct device_attribute *attr, const char *buf, size_t count)
5308{
5309 cpumask_var_t cpumask;
5310 int ret;
5311
5312 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5313 return -ENOMEM;
5314
5315 ret = cpumask_parse(buf, cpumask);
5316 if (!ret)
5317 ret = workqueue_set_unbound_cpumask(cpumask);
5318
5319 free_cpumask_var(cpumask);
5320 return ret ? ret : count;
5321}
5322
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005323static struct device_attribute wq_sysfs_cpumask_attr =
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005324 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5325 wq_unbound_cpumask_store);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005326
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005327static int __init wq_sysfs_init(void)
5328{
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005329 int err;
5330
5331 err = subsys_virtual_register(&wq_subsys, NULL);
5332 if (err)
5333 return err;
5334
5335 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005336}
5337core_initcall(wq_sysfs_init);
5338
5339static void wq_device_release(struct device *dev)
5340{
5341 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5342
5343 kfree(wq_dev);
5344}
5345
5346/**
5347 * workqueue_sysfs_register - make a workqueue visible in sysfs
5348 * @wq: the workqueue to register
5349 *
5350 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5351 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5352 * which is the preferred method.
5353 *
5354 * Workqueue user should use this function directly iff it wants to apply
5355 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5356 * apply_workqueue_attrs() may race against userland updating the
5357 * attributes.
5358 *
5359 * Return: 0 on success, -errno on failure.
5360 */
5361int workqueue_sysfs_register(struct workqueue_struct *wq)
5362{
5363 struct wq_device *wq_dev;
5364 int ret;
5365
5366 /*
Shailendra Verma402dd892015-05-23 10:38:14 +05305367 * Adjusting max_active or creating new pwqs by applying
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005368 * attributes breaks ordering guarantee. Disallow exposing ordered
5369 * workqueues.
5370 */
Tejun Heo0a94efb2017-07-23 08:36:15 -04005371 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005372 return -EINVAL;
5373
5374 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5375 if (!wq_dev)
5376 return -ENOMEM;
5377
5378 wq_dev->wq = wq;
5379 wq_dev->dev.bus = &wq_subsys;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005380 wq_dev->dev.release = wq_device_release;
Lars-Peter Clausen23217b42016-02-17 21:04:41 +01005381 dev_set_name(&wq_dev->dev, "%s", wq->name);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005382
5383 /*
5384 * unbound_attrs are created separately. Suppress uevent until
5385 * everything is ready.
5386 */
5387 dev_set_uevent_suppress(&wq_dev->dev, true);
5388
5389 ret = device_register(&wq_dev->dev);
5390 if (ret) {
Arvind Yadav7c84e5e2018-03-06 15:35:43 +05305391 put_device(&wq_dev->dev);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005392 wq->wq_dev = NULL;
5393 return ret;
5394 }
5395
5396 if (wq->flags & WQ_UNBOUND) {
5397 struct device_attribute *attr;
5398
5399 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5400 ret = device_create_file(&wq_dev->dev, attr);
5401 if (ret) {
5402 device_unregister(&wq_dev->dev);
5403 wq->wq_dev = NULL;
5404 return ret;
5405 }
5406 }
5407 }
5408
5409 dev_set_uevent_suppress(&wq_dev->dev, false);
5410 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5411 return 0;
5412}
5413
5414/**
5415 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5416 * @wq: the workqueue to unregister
5417 *
5418 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5419 */
5420static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5421{
5422 struct wq_device *wq_dev = wq->wq_dev;
5423
5424 if (!wq->wq_dev)
5425 return;
5426
5427 wq->wq_dev = NULL;
5428 device_unregister(&wq_dev->dev);
5429}
5430#else /* CONFIG_SYSFS */
5431static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5432#endif /* CONFIG_SYSFS */
5433
Tejun Heo82607adc2015-12-08 11:28:04 -05005434/*
5435 * Workqueue watchdog.
5436 *
5437 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5438 * flush dependency, a concurrency managed work item which stays RUNNING
5439 * indefinitely. Workqueue stalls can be very difficult to debug as the
5440 * usual warning mechanisms don't trigger and internal workqueue state is
5441 * largely opaque.
5442 *
5443 * Workqueue watchdog monitors all worker pools periodically and dumps
5444 * state if some pools failed to make forward progress for a while where
5445 * forward progress is defined as the first item on ->worklist changing.
5446 *
5447 * This mechanism is controlled through the kernel parameter
5448 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5449 * corresponding sysfs parameter file.
5450 */
5451#ifdef CONFIG_WQ_WATCHDOG
5452
5453static void wq_watchdog_timer_fn(unsigned long data);
5454
5455static unsigned long wq_watchdog_thresh = 30;
5456static struct timer_list wq_watchdog_timer =
5457 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5458
5459static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5460static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5461
5462static void wq_watchdog_reset_touched(void)
5463{
5464 int cpu;
5465
5466 wq_watchdog_touched = jiffies;
5467 for_each_possible_cpu(cpu)
5468 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5469}
5470
5471static void wq_watchdog_timer_fn(unsigned long data)
5472{
5473 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5474 bool lockup_detected = false;
Sergey Senozhatskyec32c8a2021-05-20 19:14:22 +09005475 unsigned long now = jiffies;
Tejun Heo82607adc2015-12-08 11:28:04 -05005476 struct worker_pool *pool;
5477 int pi;
5478
5479 if (!thresh)
5480 return;
5481
5482 rcu_read_lock();
5483
5484 for_each_pool(pool, pi) {
5485 unsigned long pool_ts, touched, ts;
5486
5487 if (list_empty(&pool->worklist))
5488 continue;
5489
Sergey Senozhatskyec32c8a2021-05-20 19:14:22 +09005490 /*
5491 * If a virtual machine is stopped by the host it can look to
5492 * the watchdog like a stall.
5493 */
5494 kvm_check_and_clear_guest_paused();
5495
Tejun Heo82607adc2015-12-08 11:28:04 -05005496 /* get the latest of pool and touched timestamps */
5497 pool_ts = READ_ONCE(pool->watchdog_ts);
5498 touched = READ_ONCE(wq_watchdog_touched);
5499
5500 if (time_after(pool_ts, touched))
5501 ts = pool_ts;
5502 else
5503 ts = touched;
5504
5505 if (pool->cpu >= 0) {
5506 unsigned long cpu_touched =
5507 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5508 pool->cpu));
5509 if (time_after(cpu_touched, ts))
5510 ts = cpu_touched;
5511 }
5512
5513 /* did we stall? */
Sergey Senozhatskyec32c8a2021-05-20 19:14:22 +09005514 if (time_after(now, ts + thresh)) {
Tejun Heo82607adc2015-12-08 11:28:04 -05005515 lockup_detected = true;
5516 pr_emerg("BUG: workqueue lockup - pool");
5517 pr_cont_pool_info(pool);
5518 pr_cont(" stuck for %us!\n",
Sergey Senozhatskyec32c8a2021-05-20 19:14:22 +09005519 jiffies_to_msecs(now - pool_ts) / 1000);
Tejun Heo82607adc2015-12-08 11:28:04 -05005520 }
5521 }
5522
5523 rcu_read_unlock();
5524
5525 if (lockup_detected)
5526 show_workqueue_state();
5527
5528 wq_watchdog_reset_touched();
5529 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5530}
5531
Vincent Whitchurch63a0f9d2018-08-21 17:25:07 +02005532notrace void wq_watchdog_touch(int cpu)
Tejun Heo82607adc2015-12-08 11:28:04 -05005533{
5534 if (cpu >= 0)
5535 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5536 else
5537 wq_watchdog_touched = jiffies;
5538}
5539
5540static void wq_watchdog_set_thresh(unsigned long thresh)
5541{
5542 wq_watchdog_thresh = 0;
5543 del_timer_sync(&wq_watchdog_timer);
5544
5545 if (thresh) {
5546 wq_watchdog_thresh = thresh;
5547 wq_watchdog_reset_touched();
5548 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5549 }
5550}
5551
5552static int wq_watchdog_param_set_thresh(const char *val,
5553 const struct kernel_param *kp)
5554{
5555 unsigned long thresh;
5556 int ret;
5557
5558 ret = kstrtoul(val, 0, &thresh);
5559 if (ret)
5560 return ret;
5561
5562 if (system_wq)
5563 wq_watchdog_set_thresh(thresh);
5564 else
5565 wq_watchdog_thresh = thresh;
5566
5567 return 0;
5568}
5569
5570static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5571 .set = wq_watchdog_param_set_thresh,
5572 .get = param_get_ulong,
5573};
5574
5575module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5576 0644);
5577
5578static void wq_watchdog_init(void)
5579{
5580 wq_watchdog_set_thresh(wq_watchdog_thresh);
5581}
5582
5583#else /* CONFIG_WQ_WATCHDOG */
5584
5585static inline void wq_watchdog_init(void) { }
5586
5587#endif /* CONFIG_WQ_WATCHDOG */
5588
Tejun Heobce90382013-04-01 11:23:32 -07005589static void __init wq_numa_init(void)
5590{
5591 cpumask_var_t *tbl;
5592 int node, cpu;
5593
Tejun Heobce90382013-04-01 11:23:32 -07005594 if (num_possible_nodes() <= 1)
5595 return;
5596
Tejun Heod55262c2013-04-01 11:23:38 -07005597 if (wq_disable_numa) {
5598 pr_info("workqueue: NUMA affinity support disabled\n");
5599 return;
5600 }
5601
Tejun Heo4c16bd32013-04-01 11:23:36 -07005602 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5603 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5604
Tejun Heobce90382013-04-01 11:23:32 -07005605 /*
5606 * We want masks of possible CPUs of each node which isn't readily
5607 * available. Build one from cpu_to_node() which should have been
5608 * fully initialized by now.
5609 */
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08005610 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
Tejun Heobce90382013-04-01 11:23:32 -07005611 BUG_ON(!tbl);
5612
5613 for_each_node(node)
Yasuaki Ishimatsu5a6024f2014-07-07 09:56:48 -04005614 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
Tejun Heo1be0c252013-05-15 14:24:24 -07005615 node_online(node) ? node : NUMA_NO_NODE));
Tejun Heobce90382013-04-01 11:23:32 -07005616
5617 for_each_possible_cpu(cpu) {
5618 node = cpu_to_node(cpu);
5619 if (WARN_ON(node == NUMA_NO_NODE)) {
5620 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5621 /* happens iff arch is bonkers, let's just proceed */
5622 return;
5623 }
5624 cpumask_set_cpu(cpu, tbl[node]);
5625 }
5626
5627 wq_numa_possible_cpumask = tbl;
5628 wq_numa_enabled = true;
5629}
5630
Tejun Heo3347fa02016-09-16 15:49:32 -04005631/**
5632 * workqueue_init_early - early init for workqueue subsystem
5633 *
5634 * This is the first half of two-staged workqueue subsystem initialization
5635 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5636 * idr are up. It sets up all the data structures and system workqueues
5637 * and allows early boot code to create workqueues and queue/cancel work
5638 * items. Actual work item execution starts only after kthreads can be
5639 * created and scheduled right before early initcalls.
5640 */
5641int __init workqueue_init_early(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005642{
Tejun Heo7a4e3442013-03-12 11:30:00 -07005643 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5644 int i, cpu;
Tejun Heoc34056a2010-06-29 10:07:11 +02005645
Tejun Heoe904e6c2013-03-12 11:29:57 -07005646 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5647
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005648 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5649 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5650
Tejun Heoe904e6c2013-03-12 11:29:57 -07005651 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5652
Tejun Heo706026c2013-01-24 11:01:34 -08005653 /* initialize CPU pools */
Tejun Heo29c91e92013-03-12 11:30:03 -07005654 for_each_possible_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005655 struct worker_pool *pool;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005656
Tejun Heo7a4e3442013-03-12 11:30:00 -07005657 i = 0;
Tejun Heof02ae732013-03-12 11:30:03 -07005658 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo7a4e3442013-03-12 11:30:00 -07005659 BUG_ON(init_worker_pool(pool));
Tejun Heoec22ca52013-01-24 11:01:33 -08005660 pool->cpu = cpu;
Tejun Heo29c91e92013-03-12 11:30:03 -07005661 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
Tejun Heo7a4e3442013-03-12 11:30:00 -07005662 pool->attrs->nice = std_nice[i++];
Tejun Heof3f90ad2013-04-01 11:23:34 -07005663 pool->node = cpu_to_node(cpu);
Tejun Heo7a4e3442013-03-12 11:30:00 -07005664
Tejun Heo9daf9e62013-01-24 11:01:33 -08005665 /* alloc pool ID */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005666 mutex_lock(&wq_pool_mutex);
Tejun Heo9daf9e62013-01-24 11:01:33 -08005667 BUG_ON(worker_pool_assign_id(pool));
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005668 mutex_unlock(&wq_pool_mutex);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005669 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02005670 }
5671
Tejun Heo8a2b7532013-09-05 12:30:04 -04005672 /* create default unbound and ordered wq attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -07005673 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5674 struct workqueue_attrs *attrs;
5675
5676 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
Tejun Heo29c91e92013-03-12 11:30:03 -07005677 attrs->nice = std_nice[i];
Tejun Heo29c91e92013-03-12 11:30:03 -07005678 unbound_std_wq_attrs[i] = attrs;
Tejun Heo8a2b7532013-09-05 12:30:04 -04005679
5680 /*
5681 * An ordered wq should have only one pwq as ordering is
5682 * guaranteed by max_active which is enforced by pwqs.
5683 * Turn off NUMA so that dfl_pwq is used for all nodes.
5684 */
5685 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5686 attrs->nice = std_nice[i];
5687 attrs->no_numa = true;
5688 ordered_wq_attrs[i] = attrs;
Tejun Heo29c91e92013-03-12 11:30:03 -07005689 }
5690
Tejun Heod320c032010-06-29 10:07:14 +02005691 system_wq = alloc_workqueue("events", 0, 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005692 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
Tejun Heod320c032010-06-29 10:07:14 +02005693 system_long_wq = alloc_workqueue("events_long", 0, 0);
Tejun Heof3421792010-07-02 10:03:51 +02005694 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5695 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01005696 system_freezable_wq = alloc_workqueue("events_freezable",
5697 WQ_FREEZABLE, 0);
Viresh Kumar06681062013-04-24 17:12:54 +05305698 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5699 WQ_POWER_EFFICIENT, 0);
5700 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5701 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5702 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005703 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
Viresh Kumar06681062013-04-24 17:12:54 +05305704 !system_unbound_wq || !system_freezable_wq ||
5705 !system_power_efficient_wq ||
5706 !system_freezable_power_efficient_wq);
Tejun Heo82607adc2015-12-08 11:28:04 -05005707
Tejun Heo3347fa02016-09-16 15:49:32 -04005708 return 0;
5709}
5710
5711/**
5712 * workqueue_init - bring workqueue subsystem fully online
5713 *
5714 * This is the latter half of two-staged workqueue subsystem initialization
5715 * and invoked as soon as kthreads can be created and scheduled.
5716 * Workqueues have been created and work items queued on them, but there
5717 * are no kworkers executing the work items yet. Populate the worker pools
5718 * with the initial workers and enable future kworker creations.
5719 */
5720int __init workqueue_init(void)
5721{
Tejun Heo2186d9f2016-10-19 12:01:27 -04005722 struct workqueue_struct *wq;
Tejun Heo3347fa02016-09-16 15:49:32 -04005723 struct worker_pool *pool;
5724 int cpu, bkt;
5725
Tejun Heo2186d9f2016-10-19 12:01:27 -04005726 /*
5727 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5728 * CPU to node mapping may not be available that early on some
5729 * archs such as power and arm64. As per-cpu pools created
5730 * previously could be missing node hint and unbound pools NUMA
5731 * affinity, fix them up.
5732 */
5733 wq_numa_init();
5734
5735 mutex_lock(&wq_pool_mutex);
5736
5737 for_each_possible_cpu(cpu) {
5738 for_each_cpu_worker_pool(pool, cpu) {
5739 pool->node = cpu_to_node(cpu);
5740 }
5741 }
5742
5743 list_for_each_entry(wq, &workqueues, list)
5744 wq_update_unbound_numa(wq, smp_processor_id(), true);
5745
5746 mutex_unlock(&wq_pool_mutex);
5747
Tejun Heo3347fa02016-09-16 15:49:32 -04005748 /* create the initial workers */
5749 for_each_online_cpu(cpu) {
5750 for_each_cpu_worker_pool(pool, cpu) {
5751 pool->flags &= ~POOL_DISASSOCIATED;
5752 BUG_ON(!create_worker(pool));
5753 }
5754 }
5755
5756 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5757 BUG_ON(!create_worker(pool));
5758
5759 wq_online = true;
Tejun Heo82607adc2015-12-08 11:28:04 -05005760 wq_watchdog_init();
5761
Suresh Siddha6ee05782010-07-30 14:57:37 -07005762 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005763}