blob: d0c6b50792c876bf814616e30715c52768152c54 [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>
Tejun Heoe22bee72010-06-29 10:07:14 +020052
Tejun Heoea138442013-01-18 14:05:55 -080053#include "workqueue_internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070054
Tejun Heoc8e55f32010-06-29 10:07:12 +020055enum {
Tejun Heobc2ae0f2012-07-17 12:39:27 -070056 /*
Tejun Heo24647572013-01-24 11:01:33 -080057 * worker_pool flags
Tejun Heobc2ae0f2012-07-17 12:39:27 -070058 *
Tejun Heo24647572013-01-24 11:01:33 -080059 * A bound pool is either associated or disassociated with its CPU.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070060 * While associated (!DISASSOCIATED), all workers are bound to the
61 * CPU and none has %WORKER_UNBOUND set and concurrency management
62 * is in effect.
63 *
64 * While DISASSOCIATED, the cpu may be offline and all workers have
65 * %WORKER_UNBOUND set and concurrency management disabled, and may
Tejun Heo24647572013-01-24 11:01:33 -080066 * be executing on any CPU. The pool behaves as an unbound one.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070067 *
Tejun Heobc3a1af2013-03-13 19:47:39 -070068 * Note that DISASSOCIATED should be flipped only while holding
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +080069 * attach_mutex to avoid changing binding state while
Lai Jiangshan4736cbf2014-05-20 17:46:35 +080070 * worker_attach_to_pool() is in progress.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070071 */
Tejun Heo692b4822017-10-09 08:04:13 -070072 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
Tejun Heo24647572013-01-24 11:01:33 -080073 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020074
Tejun Heoc8e55f32010-06-29 10:07:12 +020075 /* worker flags */
Tejun Heoc8e55f32010-06-29 10:07:12 +020076 WORKER_DIE = 1 << 1, /* die die die */
77 WORKER_IDLE = 1 << 2, /* is idle */
Tejun Heoe22bee72010-06-29 10:07:14 +020078 WORKER_PREP = 1 << 3, /* preparing to run works */
Tejun Heofb0e7be2010-06-29 10:07:15 +020079 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
Tejun Heof3421792010-07-02 10:03:51 +020080 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
Tejun Heoa9ab7752013-03-19 13:45:21 -070081 WORKER_REBOUND = 1 << 8, /* worker was rebound */
Tejun Heoe22bee72010-06-29 10:07:14 +020082
Tejun Heoa9ab7752013-03-19 13:45:21 -070083 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84 WORKER_UNBOUND | WORKER_REBOUND,
Tejun Heodb7bccf2010-06-29 10:07:12 +020085
Tejun Heoe34cdddb2013-01-24 11:01:33 -080086 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
Tejun Heo4ce62e92012-07-13 22:16:44 -070087
Tejun Heo29c91e92013-03-12 11:30:03 -070088 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
Tejun Heoc8e55f32010-06-29 10:07:12 +020089 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020090
Tejun Heoe22bee72010-06-29 10:07:14 +020091 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
93
Tejun Heo3233cdb2011-02-16 18:10:19 +010094 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95 /* call for help after 10ms
96 (min two ticks) */
Tejun Heoe22bee72010-06-29 10:07:14 +020097 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
Tejun Heoe22bee72010-06-29 10:07:14 +020099
100 /*
101 * Rescue workers are used only on emergencies and shared by
Dongsheng Yang8698a742014-03-11 18:09:12 +0800102 * all cpus. Give MIN_NICE.
Tejun Heoe22bee72010-06-29 10:07:14 +0200103 */
Dongsheng Yang8698a742014-03-11 18:09:12 +0800104 RESCUER_NICE_LEVEL = MIN_NICE,
105 HIGHPRI_NICE_LEVEL = MIN_NICE,
Tejun Heoecf68812013-04-01 11:23:34 -0700106
107 WQ_NAME_LEN = 24,
Tejun Heoc8e55f32010-06-29 10:07:12 +0200108};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110/*
Tejun Heo4690c4a2010-06-29 10:07:10 +0200111 * Structure fields follow one of the following exclusion rules.
112 *
Tejun Heoe41e7042010-08-24 14:22:47 +0200113 * I: Modifiable by initialization/destruction paths and read-only for
114 * everyone else.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200115 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
118 *
Tejun Heod565ed62013-01-24 11:01:33 -0800119 * L: pool->lock protected. Access with pool->lock held.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200120 *
Tejun Heod565ed62013-01-24 11:01:33 -0800121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
Tejun Heoe22bee72010-06-29 10:07:14 +0200125 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800126 * A: pool->attach_mutex protected.
Tejun Heo822d8402013-03-19 13:45:21 -0700127 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700128 * PL: wq_pool_mutex protected.
Tejun Heo76af4d92013-03-12 11:30:00 -0700129 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700130 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo5bcab332013-03-13 19:47:40 -0700131 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800132 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
133 *
134 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
135 * sched-RCU for reads.
136 *
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700137 * WQ: wq->mutex protected.
138 *
Lai Jiangshanb5927602013-03-25 16:57:19 -0700139 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo2e109a22013-03-13 19:47:40 -0700140 *
141 * MD: wq_mayday_lock protected.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200142 */
143
Tejun Heo2eaebdb2013-01-18 14:05:55 -0800144/* struct worker is defined in workqueue_internal.h */
Tejun Heoc34056a2010-06-29 10:07:11 +0200145
Tejun Heobd7bdd42012-07-12 14:46:37 -0700146struct worker_pool {
Tejun Heod565ed62013-01-24 11:01:33 -0800147 spinlock_t lock; /* the pool lock */
Tejun Heod84ff052013-03-12 11:29:59 -0700148 int cpu; /* I: the associated cpu */
Tejun Heof3f90ad2013-04-01 11:23:34 -0700149 int node; /* I: the associated node ID */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800150 int id; /* I: pool ID */
Tejun Heo11ebea52012-07-12 14:46:37 -0700151 unsigned int flags; /* X: flags */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700152
Tejun Heo82607adc2015-12-08 11:28:04 -0500153 unsigned long watchdog_ts; /* L: watchdog timestamp */
154
Tejun Heobd7bdd42012-07-12 14:46:37 -0700155 struct list_head worklist; /* L: list of pending works */
156 int nr_workers; /* L: total number of workers */
Lai Jiangshanea1abd62012-09-18 09:59:22 -0700157
158 /* nr_idle includes the ones off idle_list for rebinding */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700159 int nr_idle; /* L: currently idle ones */
160
161 struct list_head idle_list; /* X: list of idle workers */
162 struct timer_list idle_timer; /* L: worker idle timeout */
163 struct timer_list mayday_timer; /* L: SOS timer for workers */
164
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700165 /* a workers is either on busy_hash or idle_list, or the manager */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800166 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
167 /* L: hash of busy workers */
168
Tejun Heobc3a1af2013-03-13 19:47:39 -0700169 /* see manage_workers() for details on the two manager mutexes */
Tejun Heo2607d7a2015-03-09 09:22:28 -0400170 struct worker *manager; /* L: purely informational */
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800171 struct mutex attach_mutex; /* attach/detach exclusion */
172 struct list_head workers; /* A: attached workers */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +0800173 struct completion *detach_completion; /* all workers detached */
Tejun Heoe19e3972013-01-24 11:39:44 -0800174
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +0800175 struct ida worker_ida; /* worker IDs for task name */
Tejun Heoe19e3972013-01-24 11:39:44 -0800176
Tejun Heo7a4e3442013-03-12 11:30:00 -0700177 struct workqueue_attrs *attrs; /* I: worker attributes */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700178 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
179 int refcnt; /* PL: refcnt for unbound pools */
Tejun Heo7a4e3442013-03-12 11:30:00 -0700180
Tejun Heoe19e3972013-01-24 11:39:44 -0800181 /*
182 * The current concurrency level. As it's likely to be accessed
183 * from other CPUs during try_to_wake_up(), put it in a separate
184 * cacheline.
185 */
186 atomic_t nr_running ____cacheline_aligned_in_smp;
Tejun Heo29c91e92013-03-12 11:30:03 -0700187
188 /*
189 * Destruction of pool is sched-RCU protected to allow dereferences
190 * from get_work_pool().
191 */
192 struct rcu_head rcu;
Tejun Heo8b03ae32010-06-29 10:07:12 +0200193} ____cacheline_aligned_in_smp;
194
195/*
Tejun Heo112202d2013-02-13 19:29:12 -0800196 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
197 * of work_struct->data are used for flags and the remaining high bits
198 * point to the pwq; thus, pwqs need to be aligned at two's power of the
199 * number of flag bits.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 */
Tejun Heo112202d2013-02-13 19:29:12 -0800201struct pool_workqueue {
Tejun Heobd7bdd42012-07-12 14:46:37 -0700202 struct worker_pool *pool; /* I: the associated pool */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200203 struct workqueue_struct *wq; /* I: the owning workqueue */
Tejun Heo73f53c42010-06-29 10:07:11 +0200204 int work_color; /* L: current color */
205 int flush_color; /* L: flushing color */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700206 int refcnt; /* L: reference count */
Tejun Heo73f53c42010-06-29 10:07:11 +0200207 int nr_in_flight[WORK_NR_COLORS];
208 /* L: nr of in_flight works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200209 int nr_active; /* L: nr of active works */
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200210 int max_active; /* L: max active works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200211 struct list_head delayed_works; /* L: delayed works */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700212 struct list_head pwqs_node; /* WR: node on wq->pwqs */
Tejun Heo2e109a22013-03-13 19:47:40 -0700213 struct list_head mayday_node; /* MD: node on wq->maydays */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700214
215 /*
216 * Release of unbound pwq is punted to system_wq. See put_pwq()
217 * and pwq_unbound_release_workfn() for details. pool_workqueue
218 * itself is also sched-RCU protected so that the first pwq can be
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700219 * determined without grabbing wq->mutex.
Tejun Heo8864b4e2013-03-12 11:30:04 -0700220 */
221 struct work_struct unbound_release_work;
222 struct rcu_head rcu;
Tejun Heoe904e6c2013-03-12 11:29:57 -0700223} __aligned(1 << WORK_STRUCT_FLAG_BITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225/*
Tejun Heo73f53c42010-06-29 10:07:11 +0200226 * Structure used to wait for workqueue flush.
227 */
228struct wq_flusher {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700229 struct list_head list; /* WQ: list of flushers */
230 int flush_color; /* WQ: flush color waiting for */
Tejun Heo73f53c42010-06-29 10:07:11 +0200231 struct completion done; /* flush completion */
232};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233
Tejun Heo226223a2013-03-12 11:30:05 -0700234struct wq_device;
235
Tejun Heo73f53c42010-06-29 10:07:11 +0200236/*
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700237 * The externally visible workqueue. It relays the issued work items to
238 * the appropriate worker_pool through its pool_workqueues.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 */
240struct workqueue_struct {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700241 struct list_head pwqs; /* WR: all pwqs of this wq */
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400242 struct list_head list; /* PR: list of all workqueues */
Tejun Heo73f53c42010-06-29 10:07:11 +0200243
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700244 struct mutex mutex; /* protects this wq */
245 int work_color; /* WQ: current work color */
246 int flush_color; /* WQ: current flush color */
Tejun Heo112202d2013-02-13 19:29:12 -0800247 atomic_t nr_pwqs_to_flush; /* flush in progress */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700248 struct wq_flusher *first_flusher; /* WQ: first flusher */
249 struct list_head flusher_queue; /* WQ: flush waiters */
250 struct list_head flusher_overflow; /* WQ: flush overflow list */
Tejun Heo73f53c42010-06-29 10:07:11 +0200251
Tejun Heo2e109a22013-03-13 19:47:40 -0700252 struct list_head maydays; /* MD: pwqs requesting rescue */
Tejun Heoe22bee72010-06-29 10:07:14 +0200253 struct worker *rescuer; /* I: rescue worker */
254
Lai Jiangshan87fc7412013-03-25 16:57:18 -0700255 int nr_drainers; /* WQ: drain in progress */
Lai Jiangshana357fc02013-03-25 16:57:19 -0700256 int saved_max_active; /* WQ: saved pwq max_active */
Tejun Heo226223a2013-03-12 11:30:05 -0700257
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800258 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
259 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
Tejun Heo6029a912013-04-01 11:23:34 -0700260
Tejun Heo226223a2013-03-12 11:30:05 -0700261#ifdef CONFIG_SYSFS
262 struct wq_device *wq_dev; /* I: for sysfs interface */
263#endif
Johannes Berg4e6045f2007-10-18 23:39:55 -0700264#ifdef CONFIG_LOCKDEP
Tejun Heo4690c4a2010-06-29 10:07:10 +0200265 struct lockdep_map lockdep_map;
Johannes Berg4e6045f2007-10-18 23:39:55 -0700266#endif
Tejun Heoecf68812013-04-01 11:23:34 -0700267 char name[WQ_NAME_LEN]; /* I: workqueue name */
Tejun Heo2728fd22013-04-01 11:23:35 -0700268
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400269 /*
270 * Destruction of workqueue_struct is sched-RCU protected to allow
271 * walking the workqueues list without grabbing wq_pool_mutex.
272 * This is used to dump all workqueues from sysrq.
273 */
274 struct rcu_head rcu;
275
Tejun Heo2728fd22013-04-01 11:23:35 -0700276 /* hot fields used during command issue, aligned to cacheline */
277 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
278 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800279 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280};
281
Tejun Heoe904e6c2013-03-12 11:29:57 -0700282static struct kmem_cache *pwq_cache;
283
Tejun Heobce90382013-04-01 11:23:32 -0700284static cpumask_var_t *wq_numa_possible_cpumask;
285 /* possible CPUs of each node */
286
Tejun Heod55262c2013-04-01 11:23:38 -0700287static bool wq_disable_numa;
288module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289
Viresh Kumarcee22a12013-04-08 16:45:40 +0530290/* see the comment above the definition of WQ_POWER_EFFICIENT */
Luis R. Rodriguez552f5302015-05-27 11:09:39 +0930291static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
Viresh Kumarcee22a12013-04-08 16:45:40 +0530292module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293
Tejun Heo863b7102016-09-16 15:49:34 -0400294static bool wq_online; /* can kworkers be created yet? */
Tejun Heo3347fa02016-09-16 15:49:32 -0400295
Tejun Heobce90382013-04-01 11:23:32 -0700296static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297
Tejun Heo4c16bd32013-04-01 11:23:36 -0700298/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700301static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
Tejun Heo2e109a22013-03-13 19:47:40 -0700302static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
Tejun Heo692b4822017-10-09 08:04:13 -0700303static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
Tejun Heo5bcab332013-03-13 19:47:40 -0700304
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400305static LIST_HEAD(workqueues); /* PR: list of all workqueues */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700306static bool workqueue_freezing; /* PL: have wqs started freezing? */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700307
Mike Galbraithef5571802016-02-09 17:59:38 -0500308/* PL: allowable cpus for unbound wqs and work items */
309static cpumask_var_t wq_unbound_cpumask;
310
311/* CPU where unbound work was last round robin scheduled from this CPU */
312static DEFINE_PER_CPU(int, wq_rr_cpu_last);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +0800313
Tejun Heof303fcc2016-02-09 17:59:38 -0500314/*
315 * Local execution of unbound work items is no longer guaranteed. The
316 * following always forces round-robin CPU selection on unbound work items
317 * to uncover usages which depend on it.
318 */
319#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
320static bool wq_debug_force_rr_cpu = true;
321#else
322static bool wq_debug_force_rr_cpu = false;
323#endif
324module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
325
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700326/* the per-cpu worker pools */
Peter Zijlstra25528212016-03-15 14:52:49 -0700327static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700328
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700329static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700330
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700331/* PL: hash of all unbound pools keyed by pool->attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -0700332static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
333
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700334/* I: attributes used when instantiating standard unbound pools on demand */
Tejun Heo29c91e92013-03-12 11:30:03 -0700335static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
336
Tejun Heo8a2b7532013-09-05 12:30:04 -0400337/* I: attributes used when instantiating ordered pools on demand */
338static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
339
Tejun Heod320c032010-06-29 10:07:14 +0200340struct workqueue_struct *system_wq __read_mostly;
Marc Dionnead7b1f82013-05-06 17:44:55 -0400341EXPORT_SYMBOL(system_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300342struct workqueue_struct *system_highpri_wq __read_mostly;
Joonsoo Kim1aabe902012-08-15 23:25:39 +0900343EXPORT_SYMBOL_GPL(system_highpri_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300344struct workqueue_struct *system_long_wq __read_mostly;
Tejun Heod320c032010-06-29 10:07:14 +0200345EXPORT_SYMBOL_GPL(system_long_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300346struct workqueue_struct *system_unbound_wq __read_mostly;
Tejun Heof3421792010-07-02 10:03:51 +0200347EXPORT_SYMBOL_GPL(system_unbound_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300348struct workqueue_struct *system_freezable_wq __read_mostly;
Tejun Heo24d51ad2011-02-21 09:52:50 +0100349EXPORT_SYMBOL_GPL(system_freezable_wq);
Viresh Kumar06681062013-04-24 17:12:54 +0530350struct workqueue_struct *system_power_efficient_wq __read_mostly;
351EXPORT_SYMBOL_GPL(system_power_efficient_wq);
352struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
353EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
Tejun Heod320c032010-06-29 10:07:14 +0200354
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700355static int worker_thread(void *__worker);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +0800356static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700357
Tejun Heo97bd2342010-10-05 10:41:14 +0200358#define CREATE_TRACE_POINTS
359#include <trace/events/workqueue.h>
360
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700361#define assert_rcu_or_pool_mutex() \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700362 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
363 !lockdep_is_held(&wq_pool_mutex), \
364 "sched RCU or wq_pool_mutex should be held")
Tejun Heo5bcab332013-03-13 19:47:40 -0700365
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700366#define assert_rcu_or_wq_mutex(wq) \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700367 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
368 !lockdep_is_held(&wq->mutex), \
369 "sched RCU or wq->mutex should be held")
Tejun Heo76af4d92013-03-12 11:30:00 -0700370
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800371#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -0700372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
373 !lockdep_is_held(&wq->mutex) && \
374 !lockdep_is_held(&wq_pool_mutex), \
375 "sched RCU, wq->mutex or wq_pool_mutex should be held")
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800376
Tejun Heof02ae732013-03-12 11:30:03 -0700377#define for_each_cpu_worker_pool(pool, cpu) \
378 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
379 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
Tejun Heo7a62c2c2013-03-12 11:30:03 -0700380 (pool)++)
Tejun Heo4ce62e92012-07-13 22:16:44 -0700381
Tejun Heo49e3cf42013-03-12 11:29:58 -0700382/**
Tejun Heo17116962013-03-12 11:29:58 -0700383 * for_each_pool - iterate through all worker_pools in the system
384 * @pool: iteration cursor
Tejun Heo611c92a2013-03-13 16:51:36 -0700385 * @pi: integer used for iteration
Tejun Heofa1b54e2013-03-12 11:30:00 -0700386 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700387 * This must be called either with wq_pool_mutex held or sched RCU read
388 * locked. If the pool needs to be used beyond the locking in effect, the
389 * caller is responsible for guaranteeing that the pool stays online.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700390 *
391 * The if/else clause exists only for the lockdep assertion and can be
392 * ignored.
Tejun Heo17116962013-03-12 11:29:58 -0700393 */
Tejun Heo611c92a2013-03-13 16:51:36 -0700394#define for_each_pool(pool, pi) \
395 idr_for_each_entry(&worker_pool_idr, pool, pi) \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700396 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
Tejun Heofa1b54e2013-03-12 11:30:00 -0700397 else
Tejun Heo17116962013-03-12 11:29:58 -0700398
399/**
Tejun Heo822d8402013-03-19 13:45:21 -0700400 * for_each_pool_worker - iterate through all workers of a worker_pool
401 * @worker: iteration cursor
Tejun Heo822d8402013-03-19 13:45:21 -0700402 * @pool: worker_pool to iterate workers of
403 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800404 * This must be called with @pool->attach_mutex.
Tejun Heo822d8402013-03-19 13:45:21 -0700405 *
406 * The if/else clause exists only for the lockdep assertion and can be
407 * ignored.
408 */
Lai Jiangshanda028462014-05-20 17:46:31 +0800409#define for_each_pool_worker(worker, pool) \
410 list_for_each_entry((worker), &(pool)->workers, node) \
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800411 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
Tejun Heo822d8402013-03-19 13:45:21 -0700412 else
413
414/**
Tejun Heo49e3cf42013-03-12 11:29:58 -0700415 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
416 * @pwq: iteration cursor
417 * @wq: the target workqueue
Tejun Heo76af4d92013-03-12 11:30:00 -0700418 *
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700419 * This must be called either with wq->mutex held or sched RCU read locked.
Tejun Heo794b18b2013-03-13 19:47:40 -0700420 * If the pwq needs to be used beyond the locking in effect, the caller is
421 * responsible for guaranteeing that the pwq stays online.
Tejun Heo76af4d92013-03-12 11:30:00 -0700422 *
423 * The if/else clause exists only for the lockdep assertion and can be
424 * ignored.
Tejun Heo49e3cf42013-03-12 11:29:58 -0700425 */
426#define for_each_pwq(pwq, wq) \
Tejun Heo76af4d92013-03-12 11:30:00 -0700427 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700428 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
Tejun Heo76af4d92013-03-12 11:30:00 -0700429 else
Tejun Heof3421792010-07-02 10:03:51 +0200430
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900431#ifdef CONFIG_DEBUG_OBJECTS_WORK
432
433static struct debug_obj_descr work_debug_descr;
434
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100435static void *work_debug_hint(void *addr)
436{
437 return ((struct work_struct *) addr)->func;
438}
439
Du, Changbinb9fdac72016-05-19 17:09:41 -0700440static bool work_is_static_object(void *addr)
441{
442 struct work_struct *work = addr;
443
444 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
445}
446
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900447/*
448 * fixup_init is called when:
449 * - an active object is initialized
450 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700451static bool work_fixup_init(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900452{
453 struct work_struct *work = addr;
454
455 switch (state) {
456 case ODEBUG_STATE_ACTIVE:
457 cancel_work_sync(work);
458 debug_object_init(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700459 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900460 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700461 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900462 }
463}
464
465/*
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900466 * fixup_free is called when:
467 * - an active object is freed
468 */
Du, Changbin02a982a2016-05-19 17:09:26 -0700469static bool work_fixup_free(void *addr, enum debug_obj_state state)
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900470{
471 struct work_struct *work = addr;
472
473 switch (state) {
474 case ODEBUG_STATE_ACTIVE:
475 cancel_work_sync(work);
476 debug_object_free(work, &work_debug_descr);
Du, Changbin02a982a2016-05-19 17:09:26 -0700477 return true;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900478 default:
Du, Changbin02a982a2016-05-19 17:09:26 -0700479 return false;
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900480 }
481}
482
483static struct debug_obj_descr work_debug_descr = {
484 .name = "work_struct",
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100485 .debug_hint = work_debug_hint,
Du, Changbinb9fdac72016-05-19 17:09:41 -0700486 .is_static_object = work_is_static_object,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900487 .fixup_init = work_fixup_init,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900488 .fixup_free = work_fixup_free,
489};
490
491static inline void debug_work_activate(struct work_struct *work)
492{
493 debug_object_activate(work, &work_debug_descr);
494}
495
496static inline void debug_work_deactivate(struct work_struct *work)
497{
498 debug_object_deactivate(work, &work_debug_descr);
499}
500
501void __init_work(struct work_struct *work, int onstack)
502{
503 if (onstack)
504 debug_object_init_on_stack(work, &work_debug_descr);
505 else
506 debug_object_init(work, &work_debug_descr);
507}
508EXPORT_SYMBOL_GPL(__init_work);
509
510void destroy_work_on_stack(struct work_struct *work)
511{
512 debug_object_free(work, &work_debug_descr);
513}
514EXPORT_SYMBOL_GPL(destroy_work_on_stack);
515
Thomas Gleixnerea2e64f2014-03-23 14:20:44 +0000516void destroy_delayed_work_on_stack(struct delayed_work *work)
517{
518 destroy_timer_on_stack(&work->timer);
519 debug_object_free(&work->work, &work_debug_descr);
520}
521EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
522
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900523#else
524static inline void debug_work_activate(struct work_struct *work) { }
525static inline void debug_work_deactivate(struct work_struct *work) { }
526#endif
527
Li Bin4e8b22b2013-09-10 09:52:35 +0800528/**
529 * worker_pool_assign_id - allocate ID and assing it to @pool
530 * @pool: the pool pointer of interest
531 *
532 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
533 * successfully, -errno on failure.
534 */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800535static int worker_pool_assign_id(struct worker_pool *pool)
536{
537 int ret;
538
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700539 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo5bcab332013-03-13 19:47:40 -0700540
Li Bin4e8b22b2013-09-10 09:52:35 +0800541 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
542 GFP_KERNEL);
Tejun Heo229641a2013-04-01 17:08:13 -0700543 if (ret >= 0) {
Tejun Heoe68035f2013-03-13 14:59:38 -0700544 pool->id = ret;
Tejun Heo229641a2013-04-01 17:08:13 -0700545 return 0;
546 }
Tejun Heo9daf9e62013-01-24 11:01:33 -0800547 return ret;
548}
549
Tejun Heo76af4d92013-03-12 11:30:00 -0700550/**
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700551 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
552 * @wq: the target workqueue
553 * @node: the node ID
554 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800555 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
556 * read locked.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700557 * If the pwq needs to be used beyond the locking in effect, the caller is
558 * responsible for guaranteeing that the pwq stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700559 *
560 * Return: The unbound pool_workqueue for @node.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700561 */
562static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
563 int node)
564{
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800565 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
Tejun Heod6e022f2016-02-03 13:54:25 -0500566
567 /*
568 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
569 * delayed item is pending. The plan is to keep CPU -> NODE
570 * mapping valid and stable across CPU on/offlines. Once that
571 * happens, this workaround can be removed.
572 */
573 if (unlikely(node == NUMA_NO_NODE))
574 return wq->dfl_pwq;
575
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700576 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
577}
578
Tejun Heo73f53c42010-06-29 10:07:11 +0200579static unsigned int work_color_to_flags(int color)
580{
581 return color << WORK_STRUCT_COLOR_SHIFT;
582}
583
584static int get_work_color(struct work_struct *work)
585{
586 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
587 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
588}
589
590static int work_next_color(int color)
591{
592 return (color + 1) % WORK_NR_COLORS;
Oleg Nesterova848e3b2007-05-09 02:34:17 -0700593}
594
David Howells4594bf12006-12-07 11:33:26 +0000595/*
Tejun Heo112202d2013-02-13 19:29:12 -0800596 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
597 * contain the pointer to the queued pwq. Once execution starts, the flag
Tejun Heo7c3eed52013-01-24 11:01:33 -0800598 * is cleared and the high bits contain OFFQ flags and pool ID.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200599 *
Tejun Heo112202d2013-02-13 19:29:12 -0800600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
601 * and clear_work_data() can be used to set the pwq, pool or clear
Tejun Heobbb68df2012-08-03 10:30:46 -0700602 * work->data. These functions should only be called while the work is
603 * owned - ie. while the PENDING bit is set.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200604 *
Tejun Heo112202d2013-02-13 19:29:12 -0800605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
Tejun Heo7c3eed52013-01-24 11:01:33 -0800606 * corresponding to a work. Pool is available once the work has been
Tejun Heo112202d2013-02-13 19:29:12 -0800607 * queued anywhere after initialization until it is sync canceled. pwq is
Tejun Heo7c3eed52013-01-24 11:01:33 -0800608 * available only while the work item is queued.
Tejun Heobbb68df2012-08-03 10:30:46 -0700609 *
610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
611 * canceled. While being canceled, a work item may have its PENDING set
612 * but stay off timer and worklist for arbitrarily long and nobody should
613 * try to steal the PENDING bit.
David Howells4594bf12006-12-07 11:33:26 +0000614 */
Tejun Heo7a22ad72010-06-29 10:07:13 +0200615static inline void set_work_data(struct work_struct *work, unsigned long data,
616 unsigned long flags)
David Howells365970a2006-11-22 14:54:49 +0000617{
Tejun Heo6183c002013-03-12 11:29:57 -0700618 WARN_ON_ONCE(!work_pending(work));
Tejun Heo7a22ad72010-06-29 10:07:13 +0200619 atomic_long_set(&work->data, data | flags | work_static(work));
David Howells365970a2006-11-22 14:54:49 +0000620}
David Howells365970a2006-11-22 14:54:49 +0000621
Tejun Heo112202d2013-02-13 19:29:12 -0800622static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
Tejun Heo7a22ad72010-06-29 10:07:13 +0200623 unsigned long extra_flags)
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200624{
Tejun Heo112202d2013-02-13 19:29:12 -0800625 set_work_data(work, (unsigned long)pwq,
626 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200627}
628
Lai Jiangshan4468a002013-02-06 18:04:53 -0800629static void set_work_pool_and_keep_pending(struct work_struct *work,
630 int pool_id)
631{
632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
633 WORK_STRUCT_PENDING);
634}
635
Tejun Heo7c3eed52013-01-24 11:01:33 -0800636static void set_work_pool_and_clear_pending(struct work_struct *work,
637 int pool_id)
David Howells365970a2006-11-22 14:54:49 +0000638{
Tejun Heo23657bb2012-08-13 17:08:19 -0700639 /*
640 * The following wmb is paired with the implied mb in
641 * test_and_set_bit(PENDING) and ensures all updates to @work made
642 * here are visible to and precede any updates by the next PENDING
643 * owner.
644 */
645 smp_wmb();
Tejun Heo7c3eed52013-01-24 11:01:33 -0800646 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
Roman Pen346c09f2016-04-26 13:15:35 +0200647 /*
648 * The following mb guarantees that previous clear of a PENDING bit
649 * will not be reordered with any speculative LOADS or STORES from
650 * work->current_func, which is executed afterwards. This possible
651 * reordering can lead to a missed execution on attempt to qeueue
652 * the same @work. E.g. consider this case:
653 *
654 * CPU#0 CPU#1
655 * ---------------------------- --------------------------------
656 *
657 * 1 STORE event_indicated
658 * 2 queue_work_on() {
659 * 3 test_and_set_bit(PENDING)
660 * 4 } set_..._and_clear_pending() {
661 * 5 set_work_data() # clear bit
662 * 6 smp_mb()
663 * 7 work->current_func() {
664 * 8 LOAD event_indicated
665 * }
666 *
667 * Without an explicit full barrier speculative LOAD on line 8 can
668 * be executed before CPU#0 does STORE on line 1. If that happens,
669 * CPU#0 observes the PENDING bit is still set and new execution of
670 * a @work is not queued in a hope, that CPU#1 will eventually
671 * finish the queued @work. Meanwhile CPU#1 does not see
672 * event_indicated is set, because speculative LOAD was executed
673 * before actual STORE.
674 */
675 smp_mb();
Tejun Heo7a22ad72010-06-29 10:07:13 +0200676}
677
678static void clear_work_data(struct work_struct *work)
679{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800680 smp_wmb(); /* see set_work_pool_and_clear_pending() */
681 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200682}
683
Tejun Heo112202d2013-02-13 19:29:12 -0800684static struct pool_workqueue *get_work_pwq(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200685{
Tejun Heoe1201532010-07-22 14:14:25 +0200686 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200687
Tejun Heo112202d2013-02-13 19:29:12 -0800688 if (data & WORK_STRUCT_PWQ)
Tejun Heoe1201532010-07-22 14:14:25 +0200689 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
690 else
691 return NULL;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200692}
693
Tejun Heo7c3eed52013-01-24 11:01:33 -0800694/**
695 * get_work_pool - return the worker_pool a given work was associated with
696 * @work: the work item of interest
697 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700698 * Pools are created and destroyed under wq_pool_mutex, and allows read
699 * access under sched-RCU read lock. As such, this function should be
700 * called under wq_pool_mutex or with preemption disabled.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700701 *
702 * All fields of the returned pool are accessible as long as the above
703 * mentioned locking is in effect. If the returned pool needs to be used
704 * beyond the critical section, the caller is responsible for ensuring the
705 * returned pool is and stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700706 *
707 * Return: The worker_pool @work was last associated with. %NULL if none.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800708 */
709static struct worker_pool *get_work_pool(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200710{
Tejun Heoe1201532010-07-22 14:14:25 +0200711 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800712 int pool_id;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200713
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700714 assert_rcu_or_pool_mutex();
Tejun Heofa1b54e2013-03-12 11:30:00 -0700715
Tejun Heo112202d2013-02-13 19:29:12 -0800716 if (data & WORK_STRUCT_PWQ)
717 return ((struct pool_workqueue *)
Tejun Heo7c3eed52013-01-24 11:01:33 -0800718 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200719
Tejun Heo7c3eed52013-01-24 11:01:33 -0800720 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
721 if (pool_id == WORK_OFFQ_POOL_NONE)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200722 return NULL;
723
Tejun Heofa1b54e2013-03-12 11:30:00 -0700724 return idr_find(&worker_pool_idr, pool_id);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800725}
726
727/**
728 * get_work_pool_id - return the worker pool ID a given work is associated with
729 * @work: the work item of interest
730 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700731 * Return: The worker_pool ID @work was last associated with.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800732 * %WORK_OFFQ_POOL_NONE if none.
733 */
734static int get_work_pool_id(struct work_struct *work)
735{
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800736 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800737
Tejun Heo112202d2013-02-13 19:29:12 -0800738 if (data & WORK_STRUCT_PWQ)
739 return ((struct pool_workqueue *)
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
741
742 return data >> WORK_OFFQ_POOL_SHIFT;
Tejun Heo7c3eed52013-01-24 11:01:33 -0800743}
744
Tejun Heobbb68df2012-08-03 10:30:46 -0700745static void mark_work_canceling(struct work_struct *work)
746{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800747 unsigned long pool_id = get_work_pool_id(work);
Tejun Heobbb68df2012-08-03 10:30:46 -0700748
Tejun Heo7c3eed52013-01-24 11:01:33 -0800749 pool_id <<= WORK_OFFQ_POOL_SHIFT;
750 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700751}
752
753static bool work_is_canceling(struct work_struct *work)
754{
755 unsigned long data = atomic_long_read(&work->data);
756
Tejun Heo112202d2013-02-13 19:29:12 -0800757 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700758}
759
David Howells365970a2006-11-22 14:54:49 +0000760/*
Tejun Heo32704762012-07-13 22:16:45 -0700761 * Policy functions. These define the policies on how the global worker
762 * pools are managed. Unless noted otherwise, these functions assume that
Tejun Heod565ed62013-01-24 11:01:33 -0800763 * they're being called with pool->lock held.
David Howells365970a2006-11-22 14:54:49 +0000764 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200765
Tejun Heo63d95a92012-07-12 14:46:37 -0700766static bool __need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000767{
Tejun Heoe19e3972013-01-24 11:39:44 -0800768 return !atomic_read(&pool->nr_running);
David Howells365970a2006-11-22 14:54:49 +0000769}
770
Tejun Heoe22bee72010-06-29 10:07:14 +0200771/*
772 * Need to wake up a worker? Called from anything but currently
773 * running workers.
Tejun Heo974271c2012-07-12 14:46:37 -0700774 *
775 * Note that, because unbound workers never contribute to nr_running, this
Tejun Heo706026c2013-01-24 11:01:34 -0800776 * function will always return %true for unbound pools as long as the
Tejun Heo974271c2012-07-12 14:46:37 -0700777 * worklist isn't empty.
Tejun Heoe22bee72010-06-29 10:07:14 +0200778 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700779static bool need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000780{
Tejun Heo63d95a92012-07-12 14:46:37 -0700781 return !list_empty(&pool->worklist) && __need_more_worker(pool);
David Howells365970a2006-11-22 14:54:49 +0000782}
783
Tejun Heoe22bee72010-06-29 10:07:14 +0200784/* Can I start working? Called from busy but !running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700785static bool may_start_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200786{
Tejun Heo63d95a92012-07-12 14:46:37 -0700787 return pool->nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200788}
789
790/* Do I need to keep working? Called from currently running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700791static bool keep_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200792{
Tejun Heoe19e3972013-01-24 11:39:44 -0800793 return !list_empty(&pool->worklist) &&
794 atomic_read(&pool->nr_running) <= 1;
Tejun Heoe22bee72010-06-29 10:07:14 +0200795}
796
797/* Do we need a new worker? Called from manager. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700798static bool need_to_create_worker(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200799{
Tejun Heo63d95a92012-07-12 14:46:37 -0700800 return need_more_worker(pool) && !may_start_working(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200801}
802
Tejun Heoe22bee72010-06-29 10:07:14 +0200803/* Do we have too many workers and should some go away? */
Tejun Heo63d95a92012-07-12 14:46:37 -0700804static bool too_many_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200805{
Tejun Heo692b4822017-10-09 08:04:13 -0700806 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
Tejun Heo63d95a92012-07-12 14:46:37 -0700807 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
808 int nr_busy = pool->nr_workers - nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200809
810 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
811}
812
813/*
814 * Wake up functions.
815 */
816
Lai Jiangshan1037de32014-05-22 16:44:07 +0800817/* Return the first idle worker. Safe with preemption disabled */
818static struct worker *first_idle_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200819{
Tejun Heo63d95a92012-07-12 14:46:37 -0700820 if (unlikely(list_empty(&pool->idle_list)))
Tejun Heo7e116292010-06-29 10:07:13 +0200821 return NULL;
822
Tejun Heo63d95a92012-07-12 14:46:37 -0700823 return list_first_entry(&pool->idle_list, struct worker, entry);
Tejun Heo7e116292010-06-29 10:07:13 +0200824}
825
826/**
827 * wake_up_worker - wake up an idle worker
Tejun Heo63d95a92012-07-12 14:46:37 -0700828 * @pool: worker pool to wake worker from
Tejun Heo7e116292010-06-29 10:07:13 +0200829 *
Tejun Heo63d95a92012-07-12 14:46:37 -0700830 * Wake up the first idle worker of @pool.
Tejun Heo7e116292010-06-29 10:07:13 +0200831 *
832 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800833 * spin_lock_irq(pool->lock).
Tejun Heo7e116292010-06-29 10:07:13 +0200834 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700835static void wake_up_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200836{
Lai Jiangshan1037de32014-05-22 16:44:07 +0800837 struct worker *worker = first_idle_worker(pool);
Tejun Heo7e116292010-06-29 10:07:13 +0200838
839 if (likely(worker))
840 wake_up_process(worker->task);
841}
842
Tejun Heo4690c4a2010-06-29 10:07:10 +0200843/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200844 * wq_worker_waking_up - a worker is waking up
845 * @task: task waking up
846 * @cpu: CPU @task is waking up to
847 *
848 * This function is called during try_to_wake_up() when a worker is
849 * being awoken.
850 *
851 * CONTEXT:
852 * spin_lock_irq(rq->lock)
853 */
Tejun Heod84ff052013-03-12 11:29:59 -0700854void wq_worker_waking_up(struct task_struct *task, int cpu)
Tejun Heoe22bee72010-06-29 10:07:14 +0200855{
856 struct worker *worker = kthread_data(task);
857
Joonsoo Kim36576002012-10-26 23:03:49 +0900858 if (!(worker->flags & WORKER_NOT_RUNNING)) {
Tejun Heoec22ca52013-01-24 11:01:33 -0800859 WARN_ON_ONCE(worker->pool->cpu != cpu);
Tejun Heoe19e3972013-01-24 11:39:44 -0800860 atomic_inc(&worker->pool->nr_running);
Joonsoo Kim36576002012-10-26 23:03:49 +0900861 }
Tejun Heoe22bee72010-06-29 10:07:14 +0200862}
863
864/**
865 * wq_worker_sleeping - a worker is going to sleep
866 * @task: task going to sleep
Tejun Heoe22bee72010-06-29 10:07:14 +0200867 *
868 * This function is called during schedule() when a busy worker is
869 * going to sleep. Worker on the same cpu can be woken up by
870 * returning pointer to its task.
871 *
872 * CONTEXT:
873 * spin_lock_irq(rq->lock)
874 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700875 * Return:
Tejun Heoe22bee72010-06-29 10:07:14 +0200876 * Worker task on @cpu to wake up, %NULL if none.
877 */
Alexander Gordeev9b7f6592016-03-02 12:53:31 +0100878struct task_struct *wq_worker_sleeping(struct task_struct *task)
Tejun Heoe22bee72010-06-29 10:07:14 +0200879{
880 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
Tejun Heo111c2252013-01-17 17:16:24 -0800881 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200882
Tejun Heo111c2252013-01-17 17:16:24 -0800883 /*
884 * Rescuers, which may not have all the fields set up like normal
885 * workers, also reach here, let's not access anything before
886 * checking NOT_RUNNING.
887 */
Steven Rostedt2d646722010-12-03 23:12:33 -0500888 if (worker->flags & WORKER_NOT_RUNNING)
Tejun Heoe22bee72010-06-29 10:07:14 +0200889 return NULL;
890
Tejun Heo111c2252013-01-17 17:16:24 -0800891 pool = worker->pool;
Tejun Heo111c2252013-01-17 17:16:24 -0800892
Tejun Heoe22bee72010-06-29 10:07:14 +0200893 /* this can only happen on the local cpu */
Alexander Gordeev9b7f6592016-03-02 12:53:31 +0100894 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
Tejun Heo6183c002013-03-12 11:29:57 -0700895 return NULL;
Tejun Heoe22bee72010-06-29 10:07:14 +0200896
897 /*
898 * The counterpart of the following dec_and_test, implied mb,
899 * worklist not empty test sequence is in insert_work().
900 * Please read comment there.
901 *
Tejun Heo628c78e2012-07-17 12:39:27 -0700902 * NOT_RUNNING is clear. This means that we're bound to and
903 * running on the local cpu w/ rq lock held and preemption
904 * disabled, which in turn means that none else could be
Tejun Heod565ed62013-01-24 11:01:33 -0800905 * manipulating idle_list, so dereferencing idle_list without pool
Tejun Heo628c78e2012-07-17 12:39:27 -0700906 * lock is safe.
Tejun Heoe22bee72010-06-29 10:07:14 +0200907 */
Tejun Heoe19e3972013-01-24 11:39:44 -0800908 if (atomic_dec_and_test(&pool->nr_running) &&
909 !list_empty(&pool->worklist))
Lai Jiangshan1037de32014-05-22 16:44:07 +0800910 to_wakeup = first_idle_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200911 return to_wakeup ? to_wakeup->task : NULL;
912}
913
914/**
915 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200916 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200917 * @flags: flags to set
Tejun Heod302f012010-06-29 10:07:13 +0200918 *
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800919 * Set @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200920 *
Tejun Heocb444762010-07-02 10:03:50 +0200921 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800922 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200923 */
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800924static inline void worker_set_flags(struct worker *worker, unsigned int flags)
Tejun Heod302f012010-06-29 10:07:13 +0200925{
Tejun Heobd7bdd42012-07-12 14:46:37 -0700926 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200927
Tejun Heocb444762010-07-02 10:03:50 +0200928 WARN_ON_ONCE(worker->task != current);
929
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800930 /* If transitioning into NOT_RUNNING, adjust nr_running. */
Tejun Heoe22bee72010-06-29 10:07:14 +0200931 if ((flags & WORKER_NOT_RUNNING) &&
932 !(worker->flags & WORKER_NOT_RUNNING)) {
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800933 atomic_dec(&pool->nr_running);
Tejun Heoe22bee72010-06-29 10:07:14 +0200934 }
935
Tejun Heod302f012010-06-29 10:07:13 +0200936 worker->flags |= flags;
937}
938
939/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200940 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200941 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200942 * @flags: flags to clear
943 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200944 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200945 *
Tejun Heocb444762010-07-02 10:03:50 +0200946 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800947 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200948 */
949static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
950{
Tejun Heo63d95a92012-07-12 14:46:37 -0700951 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200952 unsigned int oflags = worker->flags;
953
Tejun Heocb444762010-07-02 10:03:50 +0200954 WARN_ON_ONCE(worker->task != current);
955
Tejun Heod302f012010-06-29 10:07:13 +0200956 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200957
Tejun Heo42c025f2011-01-11 15:58:49 +0100958 /*
959 * If transitioning out of NOT_RUNNING, increment nr_running. Note
960 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
961 * of multiple flags, not a single flag.
962 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200963 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
964 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800965 atomic_inc(&pool->nr_running);
Tejun Heod302f012010-06-29 10:07:13 +0200966}
967
968/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200969 * find_worker_executing_work - find worker which is executing a work
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800970 * @pool: pool of interest
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200971 * @work: work to find worker for
972 *
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800973 * Find a worker which is executing @work on @pool by searching
974 * @pool->busy_hash which is keyed by the address of @work. For a worker
Tejun Heoa2c1c572012-12-18 10:35:02 -0800975 * to match, its current execution should match the address of @work and
976 * its work function. This is to avoid unwanted dependency between
977 * unrelated work executions through a work item being recycled while still
978 * being executed.
979 *
980 * This is a bit tricky. A work item may be freed once its execution
981 * starts and nothing prevents the freed area from being recycled for
982 * another work item. If the same work item address ends up being reused
983 * before the original execution finishes, workqueue will identify the
984 * recycled work item as currently executing and make it wait until the
985 * current execution finishes, introducing an unwanted dependency.
986 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700987 * This function checks the work item address and work function to avoid
988 * false positives. Note that this isn't complete as one may construct a
989 * work function which can introduce dependency onto itself through a
990 * recycled work item. Well, if somebody wants to shoot oneself in the
991 * foot that badly, there's only so much we can do, and if such deadlock
992 * actually occurs, it should be easy to locate the culprit work function.
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200993 *
994 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800995 * spin_lock_irq(pool->lock).
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200996 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700997 * Return:
998 * Pointer to worker which is executing @work if found, %NULL
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200999 * otherwise.
1000 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08001001static struct worker *find_worker_executing_work(struct worker_pool *pool,
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001002 struct work_struct *work)
1003{
Sasha Levin42f85702012-12-17 10:01:23 -05001004 struct worker *worker;
Sasha Levin42f85702012-12-17 10:01:23 -05001005
Sasha Levinb67bfe02013-02-27 17:06:00 -08001006 hash_for_each_possible(pool->busy_hash, worker, hentry,
Tejun Heoa2c1c572012-12-18 10:35:02 -08001007 (unsigned long)work)
1008 if (worker->current_work == work &&
1009 worker->current_func == work->func)
Sasha Levin42f85702012-12-17 10:01:23 -05001010 return worker;
1011
1012 return NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001013}
1014
1015/**
Tejun Heobf4ede02012-08-03 10:30:46 -07001016 * move_linked_works - move linked works to a list
1017 * @work: start of series of works to be scheduled
1018 * @head: target list to append @work to
Shailendra Verma402dd892015-05-23 10:38:14 +05301019 * @nextp: out parameter for nested worklist walking
Tejun Heobf4ede02012-08-03 10:30:46 -07001020 *
1021 * Schedule linked works starting from @work to @head. Work series to
1022 * be scheduled starts at @work and includes any consecutive work with
1023 * WORK_STRUCT_LINKED set in its predecessor.
1024 *
1025 * If @nextp is not NULL, it's updated to point to the next work of
1026 * the last scheduled work. This allows move_linked_works() to be
1027 * nested inside outer list_for_each_entry_safe().
1028 *
1029 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001030 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001031 */
1032static void move_linked_works(struct work_struct *work, struct list_head *head,
1033 struct work_struct **nextp)
1034{
1035 struct work_struct *n;
1036
1037 /*
1038 * Linked worklist will always end before the end of the list,
1039 * use NULL for list head.
1040 */
1041 list_for_each_entry_safe_from(work, n, NULL, entry) {
1042 list_move_tail(&work->entry, head);
1043 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1044 break;
1045 }
1046
1047 /*
1048 * If we're already inside safe list traversal and have moved
1049 * multiple works to the scheduled queue, the next position
1050 * needs to be updated.
1051 */
1052 if (nextp)
1053 *nextp = n;
1054}
1055
Tejun Heo8864b4e2013-03-12 11:30:04 -07001056/**
1057 * get_pwq - get an extra reference on the specified pool_workqueue
1058 * @pwq: pool_workqueue to get
1059 *
1060 * Obtain an extra reference on @pwq. The caller should guarantee that
1061 * @pwq has positive refcnt and be holding the matching pool->lock.
1062 */
1063static void get_pwq(struct pool_workqueue *pwq)
1064{
1065 lockdep_assert_held(&pwq->pool->lock);
1066 WARN_ON_ONCE(pwq->refcnt <= 0);
1067 pwq->refcnt++;
1068}
1069
1070/**
1071 * put_pwq - put a pool_workqueue reference
1072 * @pwq: pool_workqueue to put
1073 *
1074 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1075 * destruction. The caller should be holding the matching pool->lock.
1076 */
1077static void put_pwq(struct pool_workqueue *pwq)
1078{
1079 lockdep_assert_held(&pwq->pool->lock);
1080 if (likely(--pwq->refcnt))
1081 return;
1082 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1083 return;
1084 /*
1085 * @pwq can't be released under pool->lock, bounce to
1086 * pwq_unbound_release_workfn(). This never recurses on the same
1087 * pool->lock as this path is taken only for unbound workqueues and
1088 * the release work item is scheduled on a per-cpu workqueue. To
1089 * avoid lockdep warning, unbound pool->locks are given lockdep
1090 * subclass of 1 in get_unbound_pool().
1091 */
1092 schedule_work(&pwq->unbound_release_work);
1093}
1094
Tejun Heodce90d42013-04-01 11:23:35 -07001095/**
1096 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1097 * @pwq: pool_workqueue to put (can be %NULL)
1098 *
1099 * put_pwq() with locking. This function also allows %NULL @pwq.
1100 */
1101static void put_pwq_unlocked(struct pool_workqueue *pwq)
1102{
1103 if (pwq) {
1104 /*
1105 * As both pwqs and pools are sched-RCU protected, the
1106 * following lock operations are safe.
1107 */
1108 spin_lock_irq(&pwq->pool->lock);
1109 put_pwq(pwq);
1110 spin_unlock_irq(&pwq->pool->lock);
1111 }
1112}
1113
Tejun Heo112202d2013-02-13 19:29:12 -08001114static void pwq_activate_delayed_work(struct work_struct *work)
Tejun Heobf4ede02012-08-03 10:30:46 -07001115{
Tejun Heo112202d2013-02-13 19:29:12 -08001116 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobf4ede02012-08-03 10:30:46 -07001117
1118 trace_workqueue_activate_work(work);
Tejun Heo82607adc2015-12-08 11:28:04 -05001119 if (list_empty(&pwq->pool->worklist))
1120 pwq->pool->watchdog_ts = jiffies;
Tejun Heo112202d2013-02-13 19:29:12 -08001121 move_linked_works(work, &pwq->pool->worklist, NULL);
Tejun Heobf4ede02012-08-03 10:30:46 -07001122 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo112202d2013-02-13 19:29:12 -08001123 pwq->nr_active++;
Tejun Heobf4ede02012-08-03 10:30:46 -07001124}
1125
Tejun Heo112202d2013-02-13 19:29:12 -08001126static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001127{
Tejun Heo112202d2013-02-13 19:29:12 -08001128 struct work_struct *work = list_first_entry(&pwq->delayed_works,
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001129 struct work_struct, entry);
1130
Tejun Heo112202d2013-02-13 19:29:12 -08001131 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001132}
1133
Tejun Heobf4ede02012-08-03 10:30:46 -07001134/**
Tejun Heo112202d2013-02-13 19:29:12 -08001135 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1136 * @pwq: pwq of interest
Tejun Heobf4ede02012-08-03 10:30:46 -07001137 * @color: color of work which left the queue
Tejun Heobf4ede02012-08-03 10:30:46 -07001138 *
1139 * A work either has completed or is removed from pending queue,
Tejun Heo112202d2013-02-13 19:29:12 -08001140 * decrement nr_in_flight of its pwq and handle workqueue flushing.
Tejun Heobf4ede02012-08-03 10:30:46 -07001141 *
1142 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001143 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001144 */
Tejun Heo112202d2013-02-13 19:29:12 -08001145static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
Tejun Heobf4ede02012-08-03 10:30:46 -07001146{
Tejun Heo8864b4e2013-03-12 11:30:04 -07001147 /* uncolored work items don't participate in flushing or nr_active */
Tejun Heobf4ede02012-08-03 10:30:46 -07001148 if (color == WORK_NO_COLOR)
Tejun Heo8864b4e2013-03-12 11:30:04 -07001149 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001150
Tejun Heo112202d2013-02-13 19:29:12 -08001151 pwq->nr_in_flight[color]--;
Tejun Heobf4ede02012-08-03 10:30:46 -07001152
Tejun Heo112202d2013-02-13 19:29:12 -08001153 pwq->nr_active--;
1154 if (!list_empty(&pwq->delayed_works)) {
Lai Jiangshanb3f9f402012-09-18 10:40:00 -07001155 /* one down, submit a delayed one */
Tejun Heo112202d2013-02-13 19:29:12 -08001156 if (pwq->nr_active < pwq->max_active)
1157 pwq_activate_first_delayed(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001158 }
1159
1160 /* is flush in progress and are we at the flushing tip? */
Tejun Heo112202d2013-02-13 19:29:12 -08001161 if (likely(pwq->flush_color != color))
Tejun Heo8864b4e2013-03-12 11:30:04 -07001162 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001163
1164 /* are there still in-flight works? */
Tejun Heo112202d2013-02-13 19:29:12 -08001165 if (pwq->nr_in_flight[color])
Tejun Heo8864b4e2013-03-12 11:30:04 -07001166 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001167
Tejun Heo112202d2013-02-13 19:29:12 -08001168 /* this pwq is done, clear flush_color */
1169 pwq->flush_color = -1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001170
1171 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001172 * If this was the last pwq, wake up the first flusher. It
Tejun Heobf4ede02012-08-03 10:30:46 -07001173 * will handle the rest.
1174 */
Tejun Heo112202d2013-02-13 19:29:12 -08001175 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1176 complete(&pwq->wq->first_flusher->done);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001177out_put:
1178 put_pwq(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001179}
1180
Tejun Heo36e227d2012-08-03 10:30:46 -07001181/**
Tejun Heobbb68df2012-08-03 10:30:46 -07001182 * try_to_grab_pending - steal work item from worklist and disable irq
Tejun Heo36e227d2012-08-03 10:30:46 -07001183 * @work: work item to steal
1184 * @is_dwork: @work is a delayed_work
Tejun Heobbb68df2012-08-03 10:30:46 -07001185 * @flags: place to store irq state
Tejun Heo36e227d2012-08-03 10:30:46 -07001186 *
1187 * Try to grab PENDING bit of @work. This function can handle @work in any
Yacine Belkadid185af32013-07-31 14:59:24 -07001188 * stable state - idle, on timer or on worklist.
Tejun Heo36e227d2012-08-03 10:30:46 -07001189 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001190 * Return:
Tejun Heo36e227d2012-08-03 10:30:46 -07001191 * 1 if @work was pending and we successfully stole PENDING
1192 * 0 if @work was idle and we claimed PENDING
1193 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
Tejun Heobbb68df2012-08-03 10:30:46 -07001194 * -ENOENT if someone else is canceling @work, this state may persist
1195 * for arbitrarily long
Tejun Heo36e227d2012-08-03 10:30:46 -07001196 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001197 * Note:
Tejun Heobbb68df2012-08-03 10:30:46 -07001198 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001199 * interrupted while holding PENDING and @work off queue, irq must be
1200 * disabled on entry. This, combined with delayed_work->timer being
1201 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
Tejun Heobbb68df2012-08-03 10:30:46 -07001202 *
1203 * On successful return, >= 0, irq is disabled and the caller is
1204 * responsible for releasing it using local_irq_restore(*@flags).
1205 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001206 * This function is safe to call from any context including IRQ handler.
Tejun Heobf4ede02012-08-03 10:30:46 -07001207 */
Tejun Heobbb68df2012-08-03 10:30:46 -07001208static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1209 unsigned long *flags)
Tejun Heobf4ede02012-08-03 10:30:46 -07001210{
Tejun Heod565ed62013-01-24 11:01:33 -08001211 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001212 struct pool_workqueue *pwq;
Tejun Heobf4ede02012-08-03 10:30:46 -07001213
Tejun Heobbb68df2012-08-03 10:30:46 -07001214 local_irq_save(*flags);
1215
Tejun Heo36e227d2012-08-03 10:30:46 -07001216 /* try to steal the timer if it exists */
1217 if (is_dwork) {
1218 struct delayed_work *dwork = to_delayed_work(work);
1219
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001220 /*
1221 * dwork->timer is irqsafe. If del_timer() fails, it's
1222 * guaranteed that the timer is not queued anywhere and not
1223 * running on the local CPU.
1224 */
Tejun Heo36e227d2012-08-03 10:30:46 -07001225 if (likely(del_timer(&dwork->timer)))
1226 return 1;
1227 }
1228
1229 /* try to claim PENDING the normal way */
Tejun Heobf4ede02012-08-03 10:30:46 -07001230 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1231 return 0;
1232
1233 /*
1234 * The queueing is in progress, or it is already queued. Try to
1235 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1236 */
Tejun Heod565ed62013-01-24 11:01:33 -08001237 pool = get_work_pool(work);
1238 if (!pool)
Tejun Heobbb68df2012-08-03 10:30:46 -07001239 goto fail;
Tejun Heobf4ede02012-08-03 10:30:46 -07001240
Tejun Heod565ed62013-01-24 11:01:33 -08001241 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001242 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001243 * work->data is guaranteed to point to pwq only while the work
1244 * item is queued on pwq->wq, and both updating work->data to point
1245 * to pwq on queueing and to pool on dequeueing are done under
1246 * pwq->pool->lock. This in turn guarantees that, if work->data
1247 * points to pwq which is associated with a locked pool, the work
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001248 * item is currently queued on that pool.
1249 */
Tejun Heo112202d2013-02-13 19:29:12 -08001250 pwq = get_work_pwq(work);
1251 if (pwq && pwq->pool == pool) {
Tejun Heo16062832013-02-06 18:04:53 -08001252 debug_work_deactivate(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001253
Tejun Heo16062832013-02-06 18:04:53 -08001254 /*
1255 * A delayed work item cannot be grabbed directly because
1256 * it might have linked NO_COLOR work items which, if left
Tejun Heo112202d2013-02-13 19:29:12 -08001257 * on the delayed_list, will confuse pwq->nr_active
Tejun Heo16062832013-02-06 18:04:53 -08001258 * management later on and cause stall. Make sure the work
1259 * item is activated before grabbing.
1260 */
1261 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
Tejun Heo112202d2013-02-13 19:29:12 -08001262 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001263
Tejun Heo16062832013-02-06 18:04:53 -08001264 list_del_init(&work->entry);
Lai Jiangshan9c34a702014-07-11 00:11:13 +08001265 pwq_dec_nr_in_flight(pwq, get_work_color(work));
Tejun Heo36e227d2012-08-03 10:30:46 -07001266
Tejun Heo112202d2013-02-13 19:29:12 -08001267 /* work->data points to pwq iff queued, point to pool */
Tejun Heo16062832013-02-06 18:04:53 -08001268 set_work_pool_and_keep_pending(work, pool->id);
Lai Jiangshan4468a002013-02-06 18:04:53 -08001269
Tejun Heo16062832013-02-06 18:04:53 -08001270 spin_unlock(&pool->lock);
1271 return 1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001272 }
Tejun Heod565ed62013-01-24 11:01:33 -08001273 spin_unlock(&pool->lock);
Tejun Heobbb68df2012-08-03 10:30:46 -07001274fail:
1275 local_irq_restore(*flags);
1276 if (work_is_canceling(work))
1277 return -ENOENT;
1278 cpu_relax();
Tejun Heo36e227d2012-08-03 10:30:46 -07001279 return -EAGAIN;
Tejun Heobf4ede02012-08-03 10:30:46 -07001280}
1281
1282/**
Tejun Heo706026c2013-01-24 11:01:34 -08001283 * insert_work - insert a work into a pool
Tejun Heo112202d2013-02-13 19:29:12 -08001284 * @pwq: pwq @work belongs to
Tejun Heo4690c4a2010-06-29 10:07:10 +02001285 * @work: work to insert
1286 * @head: insertion point
1287 * @extra_flags: extra WORK_STRUCT_* flags to set
1288 *
Tejun Heo112202d2013-02-13 19:29:12 -08001289 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
Tejun Heo706026c2013-01-24 11:01:34 -08001290 * work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +02001291 *
1292 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001293 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02001294 */
Tejun Heo112202d2013-02-13 19:29:12 -08001295static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1296 struct list_head *head, unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001297{
Tejun Heo112202d2013-02-13 19:29:12 -08001298 struct worker_pool *pool = pwq->pool;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01001299
Tejun Heo4690c4a2010-06-29 10:07:10 +02001300 /* we own @work, set data and link */
Tejun Heo112202d2013-02-13 19:29:12 -08001301 set_work_pwq(work, pwq, extra_flags);
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -07001302 list_add_tail(&work->entry, head);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001303 get_pwq(pwq);
Tejun Heoe22bee72010-06-29 10:07:14 +02001304
1305 /*
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001306 * Ensure either wq_worker_sleeping() sees the above
1307 * list_add_tail() or we see zero nr_running to avoid workers lying
1308 * around lazily while there are works to be processed.
Tejun Heoe22bee72010-06-29 10:07:14 +02001309 */
1310 smp_mb();
1311
Tejun Heo63d95a92012-07-12 14:46:37 -07001312 if (__need_more_worker(pool))
1313 wake_up_worker(pool);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001314}
1315
Tejun Heoc8efcc22010-12-20 19:32:04 +01001316/*
1317 * Test whether @work is being queued from another work executing on the
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001318 * same workqueue.
Tejun Heoc8efcc22010-12-20 19:32:04 +01001319 */
1320static bool is_chained_work(struct workqueue_struct *wq)
1321{
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001322 struct worker *worker;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001323
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001324 worker = current_wq_worker();
1325 /*
1326 * Return %true iff I'm a worker execuing a work item on @wq. If
1327 * I'm @worker, it's safe to dereference it without locking.
1328 */
Tejun Heo112202d2013-02-13 19:29:12 -08001329 return worker && worker->current_pwq->wq == wq;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001330}
1331
Mike Galbraithef5571802016-02-09 17:59:38 -05001332/*
1333 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1334 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1335 * avoid perturbing sensitive tasks.
1336 */
1337static int wq_select_unbound_cpu(int cpu)
1338{
Tejun Heof303fcc2016-02-09 17:59:38 -05001339 static bool printed_dbg_warning;
Mike Galbraithef5571802016-02-09 17:59:38 -05001340 int new_cpu;
1341
Tejun Heof303fcc2016-02-09 17:59:38 -05001342 if (likely(!wq_debug_force_rr_cpu)) {
1343 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1344 return cpu;
1345 } else if (!printed_dbg_warning) {
1346 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1347 printed_dbg_warning = true;
1348 }
1349
Mike Galbraithef5571802016-02-09 17:59:38 -05001350 if (cpumask_empty(wq_unbound_cpumask))
1351 return cpu;
1352
1353 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1354 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1355 if (unlikely(new_cpu >= nr_cpu_ids)) {
1356 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1357 if (unlikely(new_cpu >= nr_cpu_ids))
1358 return cpu;
1359 }
1360 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1361
1362 return new_cpu;
1363}
1364
Tejun Heod84ff052013-03-12 11:29:59 -07001365static void __queue_work(int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 struct work_struct *work)
1367{
Tejun Heo112202d2013-02-13 19:29:12 -08001368 struct pool_workqueue *pwq;
Tejun Heoc9178082013-03-12 11:30:04 -07001369 struct worker_pool *last_pool;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001370 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001371 unsigned int work_flags;
Joonsoo Kimb75cac92012-08-15 23:25:37 +09001372 unsigned int req_cpu = cpu;
Tejun Heo8930cab2012-08-03 10:30:45 -07001373
1374 /*
1375 * While a work item is PENDING && off queue, a task trying to
1376 * steal the PENDING will busy-loop waiting for it to either get
1377 * queued or lose PENDING. Grabbing PENDING and queueing should
1378 * happen with IRQ disabled.
1379 */
1380 WARN_ON_ONCE(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09001382 debug_work_activate(work);
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 Heodf2d5ae2013-04-01 11:23:35 -07001389 if (req_cpu == WORK_CPU_UNBOUND)
Mike Galbraithef5571802016-02-09 17:59:38 -05001390 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001391
Tejun Heoc9178082013-03-12 11:30:04 -07001392 /* pwq which will be used unless @work is executing elsewhere */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001393 if (!(wq->flags & WQ_UNBOUND))
Tejun Heo7fb98ea2013-03-12 11:30:00 -07001394 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001395 else
1396 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodbf25762012-08-20 14:51:23 -07001397
Tejun Heoc9178082013-03-12 11:30:04 -07001398 /*
1399 * If @work was previously on a different pool, it might still be
1400 * running there, in which case the work needs to be queued on that
1401 * pool to guarantee non-reentrancy.
1402 */
1403 last_pool = get_work_pool(work);
1404 if (last_pool && last_pool != pwq->pool) {
1405 struct worker *worker;
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001406
Tejun Heoc9178082013-03-12 11:30:04 -07001407 spin_lock(&last_pool->lock);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001408
Tejun Heoc9178082013-03-12 11:30:04 -07001409 worker = find_worker_executing_work(last_pool, work);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001410
Tejun Heoc9178082013-03-12 11:30:04 -07001411 if (worker && worker->current_pwq->wq == wq) {
1412 pwq = worker->current_pwq;
Tejun Heo8930cab2012-08-03 10:30:45 -07001413 } else {
Tejun Heoc9178082013-03-12 11:30:04 -07001414 /* meh... not running there, queue here */
1415 spin_unlock(&last_pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08001416 spin_lock(&pwq->pool->lock);
Tejun Heo8930cab2012-08-03 10:30:45 -07001417 }
Tejun Heof3421792010-07-02 10:03:51 +02001418 } else {
Tejun Heo112202d2013-02-13 19:29:12 -08001419 spin_lock(&pwq->pool->lock);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001420 }
1421
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001422 /*
1423 * pwq is determined and locked. For unbound pools, we could have
1424 * raced with pwq release and it could already be dead. If its
1425 * refcnt is zero, repeat pwq selection. Note that pwqs never die
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001426 * without another pwq replacing it in the numa_pwq_tbl or while
1427 * work items are executing on it, so the retrying is guaranteed to
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001428 * make forward-progress.
1429 */
1430 if (unlikely(!pwq->refcnt)) {
1431 if (wq->flags & WQ_UNBOUND) {
1432 spin_unlock(&pwq->pool->lock);
1433 cpu_relax();
1434 goto retry;
1435 }
1436 /* oops */
1437 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1438 wq->name, cpu);
1439 }
1440
Tejun Heo112202d2013-02-13 19:29:12 -08001441 /* pwq determined, queue */
1442 trace_workqueue_queue_work(req_cpu, pwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001443
Dan Carpenterf5b25522012-04-13 22:06:58 +03001444 if (WARN_ON(!list_empty(&work->entry))) {
Tejun Heo112202d2013-02-13 19:29:12 -08001445 spin_unlock(&pwq->pool->lock);
Dan Carpenterf5b25522012-04-13 22:06:58 +03001446 return;
1447 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001448
Tejun Heo112202d2013-02-13 19:29:12 -08001449 pwq->nr_in_flight[pwq->work_color]++;
1450 work_flags = work_color_to_flags(pwq->work_color);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001451
Tejun Heo112202d2013-02-13 19:29:12 -08001452 if (likely(pwq->nr_active < pwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001453 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001454 pwq->nr_active++;
1455 worklist = &pwq->pool->worklist;
Tejun Heo82607adc2015-12-08 11:28:04 -05001456 if (list_empty(worklist))
1457 pwq->pool->watchdog_ts = jiffies;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001458 } else {
1459 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo112202d2013-02-13 19:29:12 -08001460 worklist = &pwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001461 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001462
Tejun Heo112202d2013-02-13 19:29:12 -08001463 insert_work(pwq, work, worklist, work_flags);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001464
Tejun Heo112202d2013-02-13 19:29:12 -08001465 spin_unlock(&pwq->pool->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466}
1467
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001468/**
Zhang Ruic1a220e2008-07-23 21:28:39 -07001469 * queue_work_on - queue work on specific cpu
1470 * @cpu: CPU number to execute work on
1471 * @wq: workqueue to use
1472 * @work: work to queue
1473 *
Zhang Ruic1a220e2008-07-23 21:28:39 -07001474 * We queue the work to a specific CPU, the caller must ensure it
1475 * can't go away.
Yacine Belkadid185af32013-07-31 14:59:24 -07001476 *
1477 * Return: %false if @work was already on a queue, %true otherwise.
Zhang Ruic1a220e2008-07-23 21:28:39 -07001478 */
Tejun Heod4283e92012-08-03 10:30:44 -07001479bool queue_work_on(int cpu, struct workqueue_struct *wq,
1480 struct work_struct *work)
Zhang Ruic1a220e2008-07-23 21:28:39 -07001481{
Tejun Heod4283e92012-08-03 10:30:44 -07001482 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001483 unsigned long flags;
1484
1485 local_irq_save(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001486
Tejun Heo22df02b2010-06-29 10:07:10 +02001487 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001488 __queue_work(cpu, wq, work);
Tejun Heod4283e92012-08-03 10:30:44 -07001489 ret = true;
Zhang Ruic1a220e2008-07-23 21:28:39 -07001490 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001491
1492 local_irq_restore(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001493 return ret;
1494}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001495EXPORT_SYMBOL(queue_work_on);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001496
Tejun Heod8e794d2012-08-03 10:30:45 -07001497void delayed_work_timer_fn(unsigned long __data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498{
David Howells52bad642006-11-22 14:54:01 +00001499 struct delayed_work *dwork = (struct delayed_work *)__data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001501 /* should have been called from irqsafe timer with irq already off */
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001502 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503}
Konstantin Khlebnikov1438ade52013-01-24 16:36:31 +04001504EXPORT_SYMBOL(delayed_work_timer_fn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001506static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1507 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508{
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001509 struct timer_list *timer = &dwork->timer;
1510 struct work_struct *work = &dwork->work;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
Tejun Heo637fdba2017-03-06 15:33:42 -05001512 WARN_ON_ONCE(!wq);
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001513 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1514 timer->data != (unsigned long)dwork);
Tejun Heofc4b5142012-12-04 07:40:39 -08001515 WARN_ON_ONCE(timer_pending(timer));
1516 WARN_ON_ONCE(!list_empty(&work->entry));
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001517
Tejun Heo8852aac2012-12-01 16:23:42 -08001518 /*
1519 * If @delay is 0, queue @dwork->work immediately. This is for
1520 * both optimization and correctness. The earliest @timer can
1521 * expire is on the closest next tick and delayed_work users depend
1522 * on that there's no such delay when @delay is 0.
1523 */
1524 if (!delay) {
1525 __queue_work(cpu, wq, &dwork->work);
1526 return;
1527 }
1528
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001529 dwork->wq = wq;
Tejun Heo12650572012-08-08 09:38:42 -07001530 dwork->cpu = cpu;
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001531 timer->expires = jiffies + delay;
1532
Tejun Heo041bd122016-02-09 16:11:26 -05001533 if (unlikely(cpu != WORK_CPU_UNBOUND))
1534 add_timer_on(timer, cpu);
1535 else
1536 add_timer(timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537}
1538
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001539/**
1540 * queue_delayed_work_on - queue work on specific CPU after delay
1541 * @cpu: CPU number to execute work on
1542 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001543 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001544 * @delay: number of jiffies to wait before queueing
1545 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001546 * Return: %false if @work was already on a queue, %true otherwise. If
Tejun Heo715f1302012-08-03 10:30:46 -07001547 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1548 * execution.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001549 */
Tejun Heod4283e92012-08-03 10:30:44 -07001550bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1551 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001552{
David Howells52bad642006-11-22 14:54:01 +00001553 struct work_struct *work = &dwork->work;
Tejun Heod4283e92012-08-03 10:30:44 -07001554 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001555 unsigned long flags;
1556
1557 /* read the comment in __queue_work() */
1558 local_irq_save(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001559
Tejun Heo22df02b2010-06-29 10:07:10 +02001560 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001561 __queue_delayed_work(cpu, wq, dwork, delay);
Tejun Heod4283e92012-08-03 10:30:44 -07001562 ret = true;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001563 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001564
1565 local_irq_restore(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001566 return ret;
1567}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001568EXPORT_SYMBOL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569
Tejun Heoc8e55f32010-06-29 10:07:12 +02001570/**
Tejun Heo8376fe22012-08-03 10:30:47 -07001571 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1572 * @cpu: CPU number to execute work on
1573 * @wq: workqueue to use
1574 * @dwork: work to queue
1575 * @delay: number of jiffies to wait before queueing
1576 *
1577 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1578 * modify @dwork's timer so that it expires after @delay. If @delay is
1579 * zero, @work is guaranteed to be scheduled immediately regardless of its
1580 * current state.
1581 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001582 * Return: %false if @dwork was idle and queued, %true if @dwork was
Tejun Heo8376fe22012-08-03 10:30:47 -07001583 * pending and its timer was modified.
1584 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001585 * This function is safe to call from any context including IRQ handler.
Tejun Heo8376fe22012-08-03 10:30:47 -07001586 * See try_to_grab_pending() for details.
1587 */
1588bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1589 struct delayed_work *dwork, unsigned long delay)
1590{
1591 unsigned long flags;
1592 int ret;
1593
1594 do {
1595 ret = try_to_grab_pending(&dwork->work, true, &flags);
1596 } while (unlikely(ret == -EAGAIN));
1597
1598 if (likely(ret >= 0)) {
1599 __queue_delayed_work(cpu, wq, dwork, delay);
1600 local_irq_restore(flags);
1601 }
1602
1603 /* -ENOENT from try_to_grab_pending() becomes %true */
1604 return ret;
1605}
1606EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1607
1608/**
Tejun Heoc8e55f32010-06-29 10:07:12 +02001609 * worker_enter_idle - enter idle state
1610 * @worker: worker which is entering idle state
1611 *
1612 * @worker is entering idle state. Update stats and idle timer if
1613 * necessary.
1614 *
1615 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001616 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001617 */
1618static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001620 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
Tejun Heo6183c002013-03-12 11:29:57 -07001622 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1623 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1624 (worker->hentry.next || worker->hentry.pprev)))
1625 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626
Lai Jiangshan051e1852014-07-22 13:03:02 +08001627 /* can't use worker_set_flags(), also called from create_worker() */
Tejun Heocb444762010-07-02 10:03:50 +02001628 worker->flags |= WORKER_IDLE;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001629 pool->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001630 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001631
Tejun Heoc8e55f32010-06-29 10:07:12 +02001632 /* idle_list is LIFO */
Tejun Heobd7bdd42012-07-12 14:46:37 -07001633 list_add(&worker->entry, &pool->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001634
Tejun Heo628c78e2012-07-17 12:39:27 -07001635 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1636 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
Tejun Heocb444762010-07-02 10:03:50 +02001637
Tejun Heo544ecf32012-05-14 15:04:50 -07001638 /*
Tejun Heo706026c2013-01-24 11:01:34 -08001639 * Sanity check nr_running. Because wq_unbind_fn() releases
Tejun Heod565ed62013-01-24 11:01:33 -08001640 * pool->lock between setting %WORKER_UNBOUND and zapping
Tejun Heo628c78e2012-07-17 12:39:27 -07001641 * nr_running, the warning may trigger spuriously. Check iff
1642 * unbind is not in progress.
Tejun Heo544ecf32012-05-14 15:04:50 -07001643 */
Tejun Heo24647572013-01-24 11:01:33 -08001644 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -07001645 pool->nr_workers == pool->nr_idle &&
Tejun Heoe19e3972013-01-24 11:39:44 -08001646 atomic_read(&pool->nr_running));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647}
1648
Tejun Heoc8e55f32010-06-29 10:07:12 +02001649/**
1650 * worker_leave_idle - leave idle state
1651 * @worker: worker which is leaving idle state
1652 *
1653 * @worker is leaving idle state. Update stats.
1654 *
1655 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001656 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001657 */
1658static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001660 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661
Tejun Heo6183c002013-03-12 11:29:57 -07001662 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1663 return;
Tejun Heod302f012010-06-29 10:07:13 +02001664 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001665 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001666 list_del_init(&worker->entry);
1667}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001669static struct worker *alloc_worker(int node)
Tejun Heoc34056a2010-06-29 10:07:11 +02001670{
1671 struct worker *worker;
1672
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001673 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001674 if (worker) {
1675 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001676 INIT_LIST_HEAD(&worker->scheduled);
Lai Jiangshanda028462014-05-20 17:46:31 +08001677 INIT_LIST_HEAD(&worker->node);
Tejun Heoe22bee72010-06-29 10:07:14 +02001678 /* on creation a worker is in !idle && prep state */
1679 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001680 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001681 return worker;
1682}
1683
1684/**
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001685 * worker_attach_to_pool() - attach a worker to a pool
1686 * @worker: worker to be attached
1687 * @pool: the target pool
1688 *
1689 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1690 * cpu-binding of @worker are kept coordinated with the pool across
1691 * cpu-[un]hotplugs.
1692 */
1693static void worker_attach_to_pool(struct worker *worker,
1694 struct worker_pool *pool)
1695{
1696 mutex_lock(&pool->attach_mutex);
1697
1698 /*
1699 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1700 * online CPUs. It'll be re-applied when any of the CPUs come up.
1701 */
1702 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1703
1704 /*
1705 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1706 * stable across this function. See the comments above the
1707 * flag definition for details.
1708 */
1709 if (pool->flags & POOL_DISASSOCIATED)
1710 worker->flags |= WORKER_UNBOUND;
1711
1712 list_add_tail(&worker->node, &pool->workers);
1713
1714 mutex_unlock(&pool->attach_mutex);
1715}
1716
1717/**
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001718 * worker_detach_from_pool() - detach a worker from its pool
1719 * @worker: worker which is attached to its pool
1720 * @pool: the pool @worker is attached to
1721 *
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001722 * Undo the attaching which had been done in worker_attach_to_pool(). The
1723 * caller worker shouldn't access to the pool after detached except it has
1724 * other reference to the pool.
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001725 */
1726static void worker_detach_from_pool(struct worker *worker,
1727 struct worker_pool *pool)
1728{
1729 struct completion *detach_completion = NULL;
1730
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001731 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08001732 list_del(&worker->node);
1733 if (list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001734 detach_completion = pool->detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001735 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001736
Lai Jiangshanb62c0752014-06-03 15:32:52 +08001737 /* clear leftover flags without pool->lock after it is detached */
1738 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1739
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001740 if (detach_completion)
1741 complete(detach_completion);
1742}
1743
1744/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001745 * create_worker - create a new workqueue worker
Tejun Heo63d95a92012-07-12 14:46:37 -07001746 * @pool: pool the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001747 *
Lai Jiangshan051e1852014-07-22 13:03:02 +08001748 * Create and start a new worker which is attached to @pool.
Tejun Heoc34056a2010-06-29 10:07:11 +02001749 *
1750 * CONTEXT:
1751 * Might sleep. Does GFP_KERNEL allocations.
1752 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001753 * Return:
Tejun Heoc34056a2010-06-29 10:07:11 +02001754 * Pointer to the newly created worker.
1755 */
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001756static struct worker *create_worker(struct worker_pool *pool)
Tejun Heoc34056a2010-06-29 10:07:11 +02001757{
Tejun Heoc34056a2010-06-29 10:07:11 +02001758 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001759 int id = -1;
Tejun Heoe3c916a2013-04-01 11:23:32 -07001760 char id_buf[16];
Tejun Heoc34056a2010-06-29 10:07:11 +02001761
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001762 /* ID is needed to determine kthread name */
1763 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
Tejun Heo822d8402013-03-19 13:45:21 -07001764 if (id < 0)
1765 goto fail;
Tejun Heoc34056a2010-06-29 10:07:11 +02001766
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001767 worker = alloc_worker(pool->node);
Tejun Heoc34056a2010-06-29 10:07:11 +02001768 if (!worker)
1769 goto fail;
1770
Tejun Heobd7bdd42012-07-12 14:46:37 -07001771 worker->pool = pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001772 worker->id = id;
1773
Tejun Heo29c91e92013-03-12 11:30:03 -07001774 if (pool->cpu >= 0)
Tejun Heoe3c916a2013-04-01 11:23:32 -07001775 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1776 pool->attrs->nice < 0 ? "H" : "");
Tejun Heof3421792010-07-02 10:03:51 +02001777 else
Tejun Heoe3c916a2013-04-01 11:23:32 -07001778 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1779
Tejun Heof3f90ad2013-04-01 11:23:34 -07001780 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
Tejun Heoe3c916a2013-04-01 11:23:32 -07001781 "kworker/%s", id_buf);
Tejun Heoc34056a2010-06-29 10:07:11 +02001782 if (IS_ERR(worker->task))
1783 goto fail;
1784
Oleg Nesterov91151222013-11-14 12:56:18 +01001785 set_user_nice(worker->task, pool->attrs->nice);
Peter Zijlstra25834c72015-05-15 17:43:34 +02001786 kthread_bind_mask(worker->task, pool->attrs->cpumask);
Oleg Nesterov91151222013-11-14 12:56:18 +01001787
Lai Jiangshanda028462014-05-20 17:46:31 +08001788 /* successful, attach the worker to the pool */
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001789 worker_attach_to_pool(worker, pool);
Tejun Heo822d8402013-03-19 13:45:21 -07001790
Lai Jiangshan051e1852014-07-22 13:03:02 +08001791 /* start the newly created worker */
1792 spin_lock_irq(&pool->lock);
1793 worker->pool->nr_workers++;
1794 worker_enter_idle(worker);
1795 wake_up_process(worker->task);
1796 spin_unlock_irq(&pool->lock);
1797
Tejun Heoc34056a2010-06-29 10:07:11 +02001798 return worker;
Tejun Heo822d8402013-03-19 13:45:21 -07001799
Tejun Heoc34056a2010-06-29 10:07:11 +02001800fail:
Lai Jiangshan9625ab12014-05-20 17:46:27 +08001801 if (id >= 0)
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001802 ida_simple_remove(&pool->worker_ida, id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001803 kfree(worker);
1804 return NULL;
1805}
1806
1807/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001808 * destroy_worker - destroy a workqueue worker
1809 * @worker: worker to be destroyed
1810 *
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001811 * Destroy @worker and adjust @pool stats accordingly. The worker should
1812 * be idle.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001813 *
1814 * CONTEXT:
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001815 * spin_lock_irq(pool->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001816 */
1817static void destroy_worker(struct worker *worker)
1818{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001819 struct worker_pool *pool = worker->pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001820
Tejun Heocd549682013-03-13 19:47:39 -07001821 lockdep_assert_held(&pool->lock);
1822
Tejun Heoc34056a2010-06-29 10:07:11 +02001823 /* sanity check frenzy */
Tejun Heo6183c002013-03-12 11:29:57 -07001824 if (WARN_ON(worker->current_work) ||
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001825 WARN_ON(!list_empty(&worker->scheduled)) ||
1826 WARN_ON(!(worker->flags & WORKER_IDLE)))
Tejun Heo6183c002013-03-12 11:29:57 -07001827 return;
Tejun Heoc34056a2010-06-29 10:07:11 +02001828
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001829 pool->nr_workers--;
1830 pool->nr_idle--;
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001831
Tejun Heoc8e55f32010-06-29 10:07:12 +02001832 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001833 worker->flags |= WORKER_DIE;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001834 wake_up_process(worker->task);
Tejun Heoc34056a2010-06-29 10:07:11 +02001835}
1836
Tejun Heo63d95a92012-07-12 14:46:37 -07001837static void idle_worker_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001838{
Tejun Heo63d95a92012-07-12 14:46:37 -07001839 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001840
Tejun Heod565ed62013-01-24 11:01:33 -08001841 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001842
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001843 while (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001844 struct worker *worker;
1845 unsigned long expires;
1846
1847 /* idle_list is kept in LIFO order, check the last one */
Tejun Heo63d95a92012-07-12 14:46:37 -07001848 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02001849 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1850
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001851 if (time_before(jiffies, expires)) {
Tejun Heo63d95a92012-07-12 14:46:37 -07001852 mod_timer(&pool->idle_timer, expires);
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001853 break;
Tejun Heoe22bee72010-06-29 10:07:14 +02001854 }
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001855
1856 destroy_worker(worker);
Tejun Heoe22bee72010-06-29 10:07:14 +02001857 }
1858
Tejun Heod565ed62013-01-24 11:01:33 -08001859 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001860}
1861
Tejun Heo493a1722013-03-12 11:29:59 -07001862static void send_mayday(struct work_struct *work)
Tejun Heoe22bee72010-06-29 10:07:14 +02001863{
Tejun Heo112202d2013-02-13 19:29:12 -08001864 struct pool_workqueue *pwq = get_work_pwq(work);
1865 struct workqueue_struct *wq = pwq->wq;
Tejun Heo493a1722013-03-12 11:29:59 -07001866
Tejun Heo2e109a22013-03-13 19:47:40 -07001867 lockdep_assert_held(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001868
Tejun Heo493008a2013-03-12 11:30:03 -07001869 if (!wq->rescuer)
Tejun Heo493a1722013-03-12 11:29:59 -07001870 return;
Tejun Heoe22bee72010-06-29 10:07:14 +02001871
1872 /* mayday mayday mayday */
Tejun Heo493a1722013-03-12 11:29:59 -07001873 if (list_empty(&pwq->mayday_node)) {
Lai Jiangshan77668c82014-04-18 11:04:16 -04001874 /*
1875 * If @pwq is for an unbound wq, its base ref may be put at
1876 * any time due to an attribute change. Pin @pwq until the
1877 * rescuer is done with it.
1878 */
1879 get_pwq(pwq);
Tejun Heo493a1722013-03-12 11:29:59 -07001880 list_add_tail(&pwq->mayday_node, &wq->maydays);
Tejun Heoe22bee72010-06-29 10:07:14 +02001881 wake_up_process(wq->rescuer->task);
Tejun Heo493a1722013-03-12 11:29:59 -07001882 }
Tejun Heoe22bee72010-06-29 10:07:14 +02001883}
1884
Tejun Heo706026c2013-01-24 11:01:34 -08001885static void pool_mayday_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001886{
Tejun Heo63d95a92012-07-12 14:46:37 -07001887 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001888 struct work_struct *work;
1889
Tejun Heob2d82902014-12-08 12:39:16 -05001890 spin_lock_irq(&pool->lock);
1891 spin_lock(&wq_mayday_lock); /* for wq->maydays */
Tejun Heoe22bee72010-06-29 10:07:14 +02001892
Tejun Heo63d95a92012-07-12 14:46:37 -07001893 if (need_to_create_worker(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001894 /*
1895 * We've been trying to create a new worker but
1896 * haven't been successful. We might be hitting an
1897 * allocation deadlock. Send distress signals to
1898 * rescuers.
1899 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001900 list_for_each_entry(work, &pool->worklist, entry)
Tejun Heoe22bee72010-06-29 10:07:14 +02001901 send_mayday(work);
1902 }
1903
Tejun Heob2d82902014-12-08 12:39:16 -05001904 spin_unlock(&wq_mayday_lock);
1905 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001906
Tejun Heo63d95a92012-07-12 14:46:37 -07001907 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
Tejun Heoe22bee72010-06-29 10:07:14 +02001908}
1909
1910/**
1911 * maybe_create_worker - create a new worker if necessary
Tejun Heo63d95a92012-07-12 14:46:37 -07001912 * @pool: pool to create a new worker for
Tejun Heoe22bee72010-06-29 10:07:14 +02001913 *
Tejun Heo63d95a92012-07-12 14:46:37 -07001914 * Create a new worker for @pool if necessary. @pool is guaranteed to
Tejun Heoe22bee72010-06-29 10:07:14 +02001915 * have at least one idle worker on return from this function. If
1916 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
Tejun Heo63d95a92012-07-12 14:46:37 -07001917 * sent to all rescuers with works scheduled on @pool to resolve
Tejun Heoe22bee72010-06-29 10:07:14 +02001918 * possible allocation deadlock.
1919 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001920 * On return, need_to_create_worker() is guaranteed to be %false and
1921 * may_start_working() %true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001922 *
1923 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001924 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001925 * multiple times. Does GFP_KERNEL allocations. Called only from
1926 * manager.
Tejun Heoe22bee72010-06-29 10:07:14 +02001927 */
Tejun Heo29187a92015-01-16 14:21:16 -05001928static void maybe_create_worker(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08001929__releases(&pool->lock)
1930__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001931{
Tejun Heoe22bee72010-06-29 10:07:14 +02001932restart:
Tejun Heod565ed62013-01-24 11:01:33 -08001933 spin_unlock_irq(&pool->lock);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001934
Tejun Heoe22bee72010-06-29 10:07:14 +02001935 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
Tejun Heo63d95a92012-07-12 14:46:37 -07001936 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
Tejun Heoe22bee72010-06-29 10:07:14 +02001937
1938 while (true) {
Lai Jiangshan051e1852014-07-22 13:03:02 +08001939 if (create_worker(pool) || !need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001940 break;
1941
Lai Jiangshane212f362014-06-03 15:32:17 +08001942 schedule_timeout_interruptible(CREATE_COOLDOWN);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001943
Tejun Heo63d95a92012-07-12 14:46:37 -07001944 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001945 break;
1946 }
1947
Tejun Heo63d95a92012-07-12 14:46:37 -07001948 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08001949 spin_lock_irq(&pool->lock);
Lai Jiangshan051e1852014-07-22 13:03:02 +08001950 /*
1951 * This is necessary even after a new worker was just successfully
1952 * created as @pool->lock was dropped and the new worker might have
1953 * already become busy.
1954 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001955 if (need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001956 goto restart;
Tejun Heoe22bee72010-06-29 10:07:14 +02001957}
1958
1959/**
Tejun Heoe22bee72010-06-29 10:07:14 +02001960 * manage_workers - manage worker pool
1961 * @worker: self
1962 *
Tejun Heo706026c2013-01-24 11:01:34 -08001963 * Assume the manager role and manage the worker pool @worker belongs
Tejun Heoe22bee72010-06-29 10:07:14 +02001964 * to. At any given time, there can be only zero or one manager per
Tejun Heo706026c2013-01-24 11:01:34 -08001965 * pool. The exclusion is handled automatically by this function.
Tejun Heoe22bee72010-06-29 10:07:14 +02001966 *
1967 * The caller can safely start processing works on false return. On
1968 * true return, it's guaranteed that need_to_create_worker() is false
1969 * and may_start_working() is true.
1970 *
1971 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001972 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001973 * multiple times. Does GFP_KERNEL allocations.
1974 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001975 * Return:
Tejun Heo29187a92015-01-16 14:21:16 -05001976 * %false if the pool doesn't need management and the caller can safely
1977 * start processing works, %true if management function was performed and
1978 * the conditions that the caller verified before calling the function may
1979 * no longer be true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001980 */
1981static bool manage_workers(struct worker *worker)
1982{
Tejun Heo63d95a92012-07-12 14:46:37 -07001983 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001984
Tejun Heo692b4822017-10-09 08:04:13 -07001985 if (pool->flags & POOL_MANAGER_ACTIVE)
Tejun Heo29187a92015-01-16 14:21:16 -05001986 return false;
Tejun Heo692b4822017-10-09 08:04:13 -07001987
1988 pool->flags |= POOL_MANAGER_ACTIVE;
Tejun Heo2607d7a2015-03-09 09:22:28 -04001989 pool->manager = worker;
Tejun Heoe22bee72010-06-29 10:07:14 +02001990
Tejun Heo29187a92015-01-16 14:21:16 -05001991 maybe_create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02001992
Tejun Heo2607d7a2015-03-09 09:22:28 -04001993 pool->manager = NULL;
Tejun Heo692b4822017-10-09 08:04:13 -07001994 pool->flags &= ~POOL_MANAGER_ACTIVE;
1995 wake_up(&wq_manager_wait);
Tejun Heo29187a92015-01-16 14:21:16 -05001996 return true;
Tejun Heoe22bee72010-06-29 10:07:14 +02001997}
1998
Tejun Heoa62428c2010-06-29 10:07:10 +02001999/**
2000 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02002001 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02002002 * @work: work to process
2003 *
2004 * Process @work. This function contains all the logics necessary to
2005 * process a single work including synchronization against and
2006 * interaction with other workers on the same cpu, queueing and
2007 * flushing. As long as context requirement is met, any worker can
2008 * call this function to process a work.
2009 *
2010 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002011 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02002012 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002013static void process_one_work(struct worker *worker, struct work_struct *work)
Tejun Heod565ed62013-01-24 11:01:33 -08002014__releases(&pool->lock)
2015__acquires(&pool->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02002016{
Tejun Heo112202d2013-02-13 19:29:12 -08002017 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobd7bdd42012-07-12 14:46:37 -07002018 struct worker_pool *pool = worker->pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002019 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heo73f53c42010-06-29 10:07:11 +02002020 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02002021 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02002022#ifdef CONFIG_LOCKDEP
2023 /*
2024 * It is permissible to free the struct work_struct from
2025 * inside the function that is called from it, this we need to
2026 * take into account for lockdep too. To avoid bogus "held
2027 * lock freed" warnings as well as problems when looking into
2028 * work->lockdep_map, make a copy and use that here.
2029 */
Peter Zijlstra4d82a1d2012-05-15 08:06:19 -07002030 struct lockdep_map lockdep_map;
2031
2032 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002033#endif
Lai Jiangshan807407c2014-06-03 15:33:28 +08002034 /* ensure we're on the correct CPU */
Lai Jiangshan85327af2014-06-03 15:33:28 +08002035 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heoec22ca52013-01-24 11:01:33 -08002036 raw_smp_processor_id() != pool->cpu);
Tejun Heo25511a42012-07-17 12:39:27 -07002037
Tejun Heo7e116292010-06-29 10:07:13 +02002038 /*
2039 * A single work shouldn't be executed concurrently by
2040 * multiple workers on a single cpu. Check whether anyone is
2041 * already processing the work. If so, defer the work to the
2042 * currently executing one.
2043 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002044 collision = find_worker_executing_work(pool, work);
Tejun Heo7e116292010-06-29 10:07:13 +02002045 if (unlikely(collision)) {
2046 move_linked_works(work, &collision->scheduled, NULL);
2047 return;
2048 }
2049
Tejun Heo8930cab2012-08-03 10:30:45 -07002050 /* claim and dequeue */
Tejun Heoa62428c2010-06-29 10:07:10 +02002051 debug_work_deactivate(work);
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002052 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
Tejun Heoc34056a2010-06-29 10:07:11 +02002053 worker->current_work = work;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002054 worker->current_func = work->func;
Tejun Heo112202d2013-02-13 19:29:12 -08002055 worker->current_pwq = pwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002056 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002057
Tejun Heoa62428c2010-06-29 10:07:10 +02002058 list_del_init(&work->entry);
2059
Tejun Heo649027d2010-06-29 10:07:14 +02002060 /*
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002061 * CPU intensive works don't participate in concurrency management.
2062 * They're the scheduler's responsibility. This takes @worker out
2063 * of concurrency management and the next code block will chain
2064 * execution of the pending work items.
Tejun Heofb0e7be2010-06-29 10:07:15 +02002065 */
2066 if (unlikely(cpu_intensive))
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002067 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
Tejun Heofb0e7be2010-06-29 10:07:15 +02002068
Tejun Heo974271c2012-07-12 14:46:37 -07002069 /*
Lai Jiangshana489a032014-07-22 13:01:59 +08002070 * Wake up another worker if necessary. The condition is always
2071 * false for normal per-cpu workers since nr_running would always
2072 * be >= 1 at this point. This is used to chain execution of the
2073 * pending work items for WORKER_NOT_RUNNING workers such as the
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002074 * UNBOUND and CPU_INTENSIVE ones.
Tejun Heo974271c2012-07-12 14:46:37 -07002075 */
Lai Jiangshana489a032014-07-22 13:01:59 +08002076 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002077 wake_up_worker(pool);
Tejun Heo974271c2012-07-12 14:46:37 -07002078
Tejun Heo8930cab2012-08-03 10:30:45 -07002079 /*
Tejun Heo7c3eed52013-01-24 11:01:33 -08002080 * Record the last pool and clear PENDING which should be the last
Tejun Heod565ed62013-01-24 11:01:33 -08002081 * update to @work. Also, do this inside @pool->lock so that
Tejun Heo23657bb2012-08-13 17:08:19 -07002082 * PENDING and queued state changes happen together while IRQ is
2083 * disabled.
Tejun Heo8930cab2012-08-03 10:30:45 -07002084 */
Tejun Heo7c3eed52013-01-24 11:01:33 -08002085 set_work_pool_and_clear_pending(work, pool->id);
Tejun Heoa62428c2010-06-29 10:07:10 +02002086
Tejun Heod565ed62013-01-24 11:01:33 -08002087 spin_unlock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002088
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002089 lock_map_acquire(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002090 lock_map_acquire(&lockdep_map);
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002091 /*
Peter Zijlstraf52be572017-08-29 10:59:39 +02002092 * Strictly speaking we should mark the invariant state without holding
2093 * any locks, that is, before these two lock_map_acquire()'s.
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002094 *
2095 * However, that would result in:
2096 *
2097 * A(W1)
2098 * WFC(C)
2099 * A(W1)
2100 * C(C)
2101 *
2102 * Which would create W1->C->W1 dependencies, even though there is no
2103 * actual deadlock possible. There are two solutions, using a
2104 * read-recursive acquire on the work(queue) 'locks', but this will then
Peter Zijlstraf52be572017-08-29 10:59:39 +02002105 * hit the lockdep limitation on recursive locks, or simply discard
Peter Zijlstrae6f3faa2017-08-23 13:23:30 +02002106 * these locks.
2107 *
2108 * AFAICT there is no possible deadlock scenario between the
2109 * flush_work() and complete() primitives (except for single-threaded
2110 * workqueues), so hiding them isn't a problem.
2111 */
Peter Zijlstraf52be572017-08-29 10:59:39 +02002112 lockdep_invariant_state(true);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002113 trace_workqueue_execute_start(work);
Tejun Heoa2c1c572012-12-18 10:35:02 -08002114 worker->current_func(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002115 /*
2116 * While we must be careful to not use "work" after this, the trace
2117 * point will only record its address.
2118 */
2119 trace_workqueue_execute_end(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02002120 lock_map_release(&lockdep_map);
Tejun Heo112202d2013-02-13 19:29:12 -08002121 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002122
2123 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
Valentin Ilie044c7822012-08-19 00:52:42 +03002124 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2125 " last function: %pf\n",
Tejun Heoa2c1c572012-12-18 10:35:02 -08002126 current->comm, preempt_count(), task_pid_nr(current),
2127 worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002128 debug_show_held_locks(current);
2129 dump_stack();
2130 }
2131
Tejun Heob22ce272013-08-28 17:33:37 -04002132 /*
2133 * The following prevents a kworker from hogging CPU on !PREEMPT
2134 * kernels, where a requeueing work item waiting for something to
2135 * happen could deadlock with stop_machine as such work item could
2136 * indefinitely requeue itself while all other CPUs are trapped in
Joe Lawrence789cbbe2014-10-05 13:24:21 -04002137 * stop_machine. At the same time, report a quiescent RCU state so
2138 * the same condition doesn't freeze RCU.
Tejun Heob22ce272013-08-28 17:33:37 -04002139 */
Joe Lawrence3e28e372014-10-05 13:24:22 -04002140 cond_resched_rcu_qs();
Tejun Heob22ce272013-08-28 17:33:37 -04002141
Tejun Heod565ed62013-01-24 11:01:33 -08002142 spin_lock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002143
Tejun Heofb0e7be2010-06-29 10:07:15 +02002144 /* clear cpu intensive status */
2145 if (unlikely(cpu_intensive))
2146 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2147
Tejun Heoa62428c2010-06-29 10:07:10 +02002148 /* we're done with it, release */
Sasha Levin42f85702012-12-17 10:01:23 -05002149 hash_del(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002150 worker->current_work = NULL;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002151 worker->current_func = NULL;
Tejun Heo112202d2013-02-13 19:29:12 -08002152 worker->current_pwq = NULL;
Tejun Heo3d1cb202013-04-30 15:27:22 -07002153 worker->desc_valid = false;
Tejun Heo112202d2013-02-13 19:29:12 -08002154 pwq_dec_nr_in_flight(pwq, work_color);
Tejun Heoa62428c2010-06-29 10:07:10 +02002155}
2156
Tejun Heoaffee4b2010-06-29 10:07:12 +02002157/**
2158 * process_scheduled_works - process scheduled works
2159 * @worker: self
2160 *
2161 * Process all scheduled works. Please note that the scheduled list
2162 * may change while processing a work, so this function repeatedly
2163 * fetches a work from the top and executes it.
2164 *
2165 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002166 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02002167 * multiple times.
2168 */
2169static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002171 while (!list_empty(&worker->scheduled)) {
2172 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002174 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176}
2177
Tejun Heo4690c4a2010-06-29 10:07:10 +02002178/**
2179 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02002180 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02002181 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002182 * The worker thread function. All workers belong to a worker_pool -
2183 * either a per-cpu one or dynamic unbound one. These workers process all
2184 * work items regardless of their specific target workqueue. The only
2185 * exception is work items which belong to workqueues with a rescuer which
2186 * will be explained in rescuer_thread().
Yacine Belkadid185af32013-07-31 14:59:24 -07002187 *
2188 * Return: 0
Tejun Heo4690c4a2010-06-29 10:07:10 +02002189 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002190static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191{
Tejun Heoc34056a2010-06-29 10:07:11 +02002192 struct worker *worker = __worker;
Tejun Heobd7bdd42012-07-12 14:46:37 -07002193 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194
Tejun Heoe22bee72010-06-29 10:07:14 +02002195 /* tell the scheduler that this is a workqueue worker */
2196 worker->task->flags |= PF_WQ_WORKER;
Tejun Heoc8e55f32010-06-29 10:07:12 +02002197woke_up:
Tejun Heod565ed62013-01-24 11:01:33 -08002198 spin_lock_irq(&pool->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002199
Tejun Heoa9ab7752013-03-19 13:45:21 -07002200 /* am I supposed to die? */
2201 if (unlikely(worker->flags & WORKER_DIE)) {
Tejun Heod565ed62013-01-24 11:01:33 -08002202 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002203 WARN_ON_ONCE(!list_empty(&worker->entry));
2204 worker->task->flags &= ~PF_WQ_WORKER;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002205
2206 set_task_comm(worker->task, "kworker/dying");
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08002207 ida_simple_remove(&pool->worker_ida, worker->id);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002208 worker_detach_from_pool(worker, pool);
2209 kfree(worker);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002210 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07002212
Tejun Heoc8e55f32010-06-29 10:07:12 +02002213 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02002214recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02002215 /* no more worker necessary? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002216 if (!need_more_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002217 goto sleep;
2218
2219 /* do we need to manage? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002220 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002221 goto recheck;
2222
Tejun Heoc8e55f32010-06-29 10:07:12 +02002223 /*
2224 * ->scheduled list can only be filled while a worker is
2225 * preparing to process a work or actually processing it.
2226 * Make sure nobody diddled with it while I was sleeping.
2227 */
Tejun Heo6183c002013-03-12 11:29:57 -07002228 WARN_ON_ONCE(!list_empty(&worker->scheduled));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002229
Tejun Heoe22bee72010-06-29 10:07:14 +02002230 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07002231 * Finish PREP stage. We're guaranteed to have at least one idle
2232 * worker or that someone else has already assumed the manager
2233 * role. This is where @worker starts participating in concurrency
2234 * management if applicable and concurrency management is restored
2235 * after being rebound. See rebind_workers() for details.
Tejun Heoe22bee72010-06-29 10:07:14 +02002236 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07002237 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
Tejun Heoe22bee72010-06-29 10:07:14 +02002238
2239 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02002240 struct work_struct *work =
Tejun Heobd7bdd42012-07-12 14:46:37 -07002241 list_first_entry(&pool->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02002242 struct work_struct, entry);
2243
Tejun Heo82607adc2015-12-08 11:28:04 -05002244 pool->watchdog_ts = jiffies;
2245
Tejun Heoc8e55f32010-06-29 10:07:12 +02002246 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2247 /* optimization path, not strictly necessary */
2248 process_one_work(worker, work);
2249 if (unlikely(!list_empty(&worker->scheduled)))
2250 process_scheduled_works(worker);
2251 } else {
2252 move_linked_works(work, &worker->scheduled, NULL);
2253 process_scheduled_works(worker);
2254 }
Tejun Heo63d95a92012-07-12 14:46:37 -07002255 } while (keep_working(pool));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002256
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002257 worker_set_flags(worker, WORKER_PREP);
Tejun Heod313dd82010-07-02 10:03:51 +02002258sleep:
Tejun Heoc8e55f32010-06-29 10:07:12 +02002259 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002260 * pool->lock is held and there's no work to process and no need to
2261 * manage, sleep. Workers are woken up only while holding
2262 * pool->lock or from local cpu, so setting the current state
2263 * before releasing pool->lock is enough to prevent losing any
2264 * event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02002265 */
2266 worker_enter_idle(worker);
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002267 __set_current_state(TASK_IDLE);
Tejun Heod565ed62013-01-24 11:01:33 -08002268 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002269 schedule();
2270 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271}
2272
Tejun Heoe22bee72010-06-29 10:07:14 +02002273/**
2274 * rescuer_thread - the rescuer thread function
Tejun Heo111c2252013-01-17 17:16:24 -08002275 * @__rescuer: self
Tejun Heoe22bee72010-06-29 10:07:14 +02002276 *
2277 * Workqueue rescuer thread function. There's one rescuer for each
Tejun Heo493008a2013-03-12 11:30:03 -07002278 * workqueue which has WQ_MEM_RECLAIM set.
Tejun Heoe22bee72010-06-29 10:07:14 +02002279 *
Tejun Heo706026c2013-01-24 11:01:34 -08002280 * Regular work processing on a pool may block trying to create a new
Tejun Heoe22bee72010-06-29 10:07:14 +02002281 * worker which uses GFP_KERNEL allocation which has slight chance of
2282 * developing into deadlock if some works currently on the same queue
2283 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2284 * the problem rescuer solves.
2285 *
Tejun Heo706026c2013-01-24 11:01:34 -08002286 * When such condition is possible, the pool summons rescuers of all
2287 * workqueues which have works queued on the pool and let them process
Tejun Heoe22bee72010-06-29 10:07:14 +02002288 * those works so that forward progress can be guaranteed.
2289 *
2290 * This should happen rarely.
Yacine Belkadid185af32013-07-31 14:59:24 -07002291 *
2292 * Return: 0
Tejun Heoe22bee72010-06-29 10:07:14 +02002293 */
Tejun Heo111c2252013-01-17 17:16:24 -08002294static int rescuer_thread(void *__rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02002295{
Tejun Heo111c2252013-01-17 17:16:24 -08002296 struct worker *rescuer = __rescuer;
2297 struct workqueue_struct *wq = rescuer->rescue_wq;
Tejun Heoe22bee72010-06-29 10:07:14 +02002298 struct list_head *scheduled = &rescuer->scheduled;
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002299 bool should_stop;
Tejun Heoe22bee72010-06-29 10:07:14 +02002300
2301 set_user_nice(current, RESCUER_NICE_LEVEL);
Tejun Heo111c2252013-01-17 17:16:24 -08002302
2303 /*
2304 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2305 * doesn't participate in concurrency management.
2306 */
2307 rescuer->task->flags |= PF_WQ_WORKER;
Tejun Heoe22bee72010-06-29 10:07:14 +02002308repeat:
Peter Zijlstrac5a94a62017-08-23 13:58:44 +02002309 set_current_state(TASK_IDLE);
Tejun Heoe22bee72010-06-29 10:07:14 +02002310
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002311 /*
2312 * By the time the rescuer is requested to stop, the workqueue
2313 * shouldn't have any work pending, but @wq->maydays may still have
2314 * pwq(s) queued. This can happen by non-rescuer workers consuming
2315 * all the work items before the rescuer got to them. Go through
2316 * @wq->maydays processing before acting on should_stop so that the
2317 * list is always empty on exit.
2318 */
2319 should_stop = kthread_should_stop();
Tejun Heoe22bee72010-06-29 10:07:14 +02002320
Tejun Heo493a1722013-03-12 11:29:59 -07002321 /* see whether any pwq is asking for help */
Tejun Heo2e109a22013-03-13 19:47:40 -07002322 spin_lock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002323
2324 while (!list_empty(&wq->maydays)) {
2325 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2326 struct pool_workqueue, mayday_node);
Tejun Heo112202d2013-02-13 19:29:12 -08002327 struct worker_pool *pool = pwq->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002328 struct work_struct *work, *n;
Tejun Heo82607adc2015-12-08 11:28:04 -05002329 bool first = true;
Tejun Heoe22bee72010-06-29 10:07:14 +02002330
2331 __set_current_state(TASK_RUNNING);
Tejun Heo493a1722013-03-12 11:29:59 -07002332 list_del_init(&pwq->mayday_node);
2333
Tejun Heo2e109a22013-03-13 19:47:40 -07002334 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002335
Lai Jiangshan51697d392014-05-20 17:46:36 +08002336 worker_attach_to_pool(rescuer, pool);
2337
2338 spin_lock_irq(&pool->lock);
Lai Jiangshanb3104102013-02-19 12:17:02 -08002339 rescuer->pool = pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002340
2341 /*
2342 * Slurp in all works issued via this workqueue and
2343 * process'em.
2344 */
Tejun Heo0479c8c2014-12-04 10:14:13 -05002345 WARN_ON_ONCE(!list_empty(scheduled));
Tejun Heo82607adc2015-12-08 11:28:04 -05002346 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2347 if (get_work_pwq(work) == pwq) {
2348 if (first)
2349 pool->watchdog_ts = jiffies;
Tejun Heoe22bee72010-06-29 10:07:14 +02002350 move_linked_works(work, scheduled, &n);
Tejun Heo82607adc2015-12-08 11:28:04 -05002351 }
2352 first = false;
2353 }
Tejun Heoe22bee72010-06-29 10:07:14 +02002354
NeilBrown008847f2014-12-08 12:39:16 -05002355 if (!list_empty(scheduled)) {
2356 process_scheduled_works(rescuer);
2357
2358 /*
2359 * The above execution of rescued work items could
2360 * have created more to rescue through
2361 * pwq_activate_first_delayed() or chained
2362 * queueing. Let's put @pwq back on mayday list so
2363 * that such back-to-back work items, which may be
2364 * being used to relieve memory pressure, don't
2365 * incur MAYDAY_INTERVAL delay inbetween.
2366 */
2367 if (need_to_create_worker(pool)) {
2368 spin_lock(&wq_mayday_lock);
2369 get_pwq(pwq);
2370 list_move_tail(&pwq->mayday_node, &wq->maydays);
2371 spin_unlock(&wq_mayday_lock);
2372 }
2373 }
Tejun Heo75769582011-02-14 14:04:46 +01002374
2375 /*
Lai Jiangshan77668c82014-04-18 11:04:16 -04002376 * Put the reference grabbed by send_mayday(). @pool won't
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002377 * go away while we're still attached to it.
Lai Jiangshan77668c82014-04-18 11:04:16 -04002378 */
2379 put_pwq(pwq);
2380
2381 /*
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002382 * Leave this pool. If need_more_worker() is %true, notify a
Tejun Heo75769582011-02-14 14:04:46 +01002383 * regular worker; otherwise, we end up with 0 concurrency
2384 * and stalling the execution.
2385 */
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002386 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002387 wake_up_worker(pool);
Tejun Heo75769582011-02-14 14:04:46 +01002388
Lai Jiangshanb3104102013-02-19 12:17:02 -08002389 rescuer->pool = NULL;
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002390 spin_unlock_irq(&pool->lock);
2391
2392 worker_detach_from_pool(rescuer, pool);
2393
2394 spin_lock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002395 }
2396
Tejun Heo2e109a22013-03-13 19:47:40 -07002397 spin_unlock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002398
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002399 if (should_stop) {
2400 __set_current_state(TASK_RUNNING);
2401 rescuer->task->flags &= ~PF_WQ_WORKER;
2402 return 0;
2403 }
2404
Tejun Heo111c2252013-01-17 17:16:24 -08002405 /* rescuers should never participate in concurrency management */
2406 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
Tejun Heoe22bee72010-06-29 10:07:14 +02002407 schedule();
2408 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409}
2410
Tejun Heofca839c2015-12-07 10:58:57 -05002411/**
2412 * check_flush_dependency - check for flush dependency sanity
2413 * @target_wq: workqueue being flushed
2414 * @target_work: work item being flushed (NULL for workqueue flushes)
2415 *
2416 * %current is trying to flush the whole @target_wq or @target_work on it.
2417 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2418 * reclaiming memory or running on a workqueue which doesn't have
2419 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2420 * a deadlock.
2421 */
2422static void check_flush_dependency(struct workqueue_struct *target_wq,
2423 struct work_struct *target_work)
2424{
2425 work_func_t target_func = target_work ? target_work->func : NULL;
2426 struct worker *worker;
2427
2428 if (target_wq->flags & WQ_MEM_RECLAIM)
2429 return;
2430
2431 worker = current_wq_worker();
2432
2433 WARN_ONCE(current->flags & PF_MEMALLOC,
2434 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2435 current->pid, current->comm, target_wq->name, target_func);
Tejun Heo23d11a52016-01-29 05:59:46 -05002436 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2437 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
Tejun Heofca839c2015-12-07 10:58:57 -05002438 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2439 worker->current_pwq->wq->name, worker->current_func,
2440 target_wq->name, target_func);
2441}
2442
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002443struct wq_barrier {
2444 struct work_struct work;
2445 struct completion done;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002446 struct task_struct *task; /* purely informational */
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002447};
2448
2449static void wq_barrier_func(struct work_struct *work)
2450{
2451 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2452 complete(&barr->done);
2453}
2454
Tejun Heo4690c4a2010-06-29 10:07:10 +02002455/**
2456 * insert_wq_barrier - insert a barrier work
Tejun Heo112202d2013-02-13 19:29:12 -08002457 * @pwq: pwq to insert barrier into
Tejun Heo4690c4a2010-06-29 10:07:10 +02002458 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002459 * @target: target work to attach @barr to
2460 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002461 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002462 * @barr is linked to @target such that @barr is completed only after
2463 * @target finishes execution. Please note that the ordering
2464 * guarantee is observed only with respect to @target and on the local
2465 * cpu.
2466 *
2467 * Currently, a queued barrier can't be canceled. This is because
2468 * try_to_grab_pending() can't determine whether the work to be
2469 * grabbed is at the head of the queue and thus can't clear LINKED
2470 * flag of the previous work while there must be a valid next work
2471 * after a work with LINKED flag set.
2472 *
2473 * Note that when @worker is non-NULL, @target may be modified
Tejun Heo112202d2013-02-13 19:29:12 -08002474 * underneath us, so we can't reliably determine pwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002475 *
2476 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002477 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002478 */
Tejun Heo112202d2013-02-13 19:29:12 -08002479static void insert_wq_barrier(struct pool_workqueue *pwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002480 struct wq_barrier *barr,
2481 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002482{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002483 struct list_head *head;
2484 unsigned int linked = 0;
2485
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002486 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002487 * debugobject calls are safe here even with pool->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002488 * as we know for sure that this will not trigger any of the
2489 * checks and call back into the fixup functions where we
2490 * might deadlock.
2491 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002492 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002493 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Boqun Feng52fa5bc2017-08-17 17:46:12 +08002494
2495 /*
2496 * Explicitly init the crosslock for wq_barrier::done, make its lock
2497 * key a subkey of the corresponding work. As a result we won't
2498 * build a dependency between wq_barrier::done and unrelated work.
2499 */
2500 lockdep_init_map_crosslock((struct lockdep_map *)&barr->done.map,
2501 "(complete)wq_barr::done",
2502 target->lockdep_map.key, 1);
2503 __init_completion(&barr->done);
Tejun Heo2607d7a2015-03-09 09:22:28 -04002504 barr->task = current;
Oleg Nesterov83c22522007-05-09 02:33:54 -07002505
Tejun Heoaffee4b2010-06-29 10:07:12 +02002506 /*
2507 * If @target is currently being executed, schedule the
2508 * barrier to the worker; otherwise, put it after @target.
2509 */
2510 if (worker)
2511 head = worker->scheduled.next;
2512 else {
2513 unsigned long *bits = work_data_bits(target);
2514
2515 head = target->entry.next;
2516 /* there can already be other linked works, inherit and set */
2517 linked = *bits & WORK_STRUCT_LINKED;
2518 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2519 }
2520
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002521 debug_work_activate(&barr->work);
Tejun Heo112202d2013-02-13 19:29:12 -08002522 insert_work(pwq, &barr->work, head,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002523 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002524}
2525
Tejun Heo73f53c42010-06-29 10:07:11 +02002526/**
Tejun Heo112202d2013-02-13 19:29:12 -08002527 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
Tejun Heo73f53c42010-06-29 10:07:11 +02002528 * @wq: workqueue being flushed
2529 * @flush_color: new flush color, < 0 for no-op
2530 * @work_color: new work color, < 0 for no-op
2531 *
Tejun Heo112202d2013-02-13 19:29:12 -08002532 * Prepare pwqs for workqueue flushing.
Tejun Heo73f53c42010-06-29 10:07:11 +02002533 *
Tejun Heo112202d2013-02-13 19:29:12 -08002534 * If @flush_color is non-negative, flush_color on all pwqs should be
2535 * -1. If no pwq has in-flight commands at the specified color, all
2536 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2537 * has in flight commands, its pwq->flush_color is set to
2538 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
Tejun Heo73f53c42010-06-29 10:07:11 +02002539 * wakeup logic is armed and %true is returned.
2540 *
2541 * The caller should have initialized @wq->first_flusher prior to
2542 * calling this function with non-negative @flush_color. If
2543 * @flush_color is negative, no flush color update is done and %false
2544 * is returned.
2545 *
Tejun Heo112202d2013-02-13 19:29:12 -08002546 * If @work_color is non-negative, all pwqs should have the same
Tejun Heo73f53c42010-06-29 10:07:11 +02002547 * work_color which is previous to @work_color and all will be
2548 * advanced to @work_color.
2549 *
2550 * CONTEXT:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002551 * mutex_lock(wq->mutex).
Tejun Heo73f53c42010-06-29 10:07:11 +02002552 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002553 * Return:
Tejun Heo73f53c42010-06-29 10:07:11 +02002554 * %true if @flush_color >= 0 and there's something to flush. %false
2555 * otherwise.
2556 */
Tejun Heo112202d2013-02-13 19:29:12 -08002557static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
Tejun Heo73f53c42010-06-29 10:07:11 +02002558 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559{
Tejun Heo73f53c42010-06-29 10:07:11 +02002560 bool wait = false;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002561 struct pool_workqueue *pwq;
Oleg Nesterov14441962007-05-23 13:57:57 -07002562
Tejun Heo73f53c42010-06-29 10:07:11 +02002563 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002564 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
Tejun Heo112202d2013-02-13 19:29:12 -08002565 atomic_set(&wq->nr_pwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002566 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002567
Tejun Heo49e3cf42013-03-12 11:29:58 -07002568 for_each_pwq(pwq, wq) {
Tejun Heo112202d2013-02-13 19:29:12 -08002569 struct worker_pool *pool = pwq->pool;
Tejun Heo73f53c42010-06-29 10:07:11 +02002570
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002571 spin_lock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002572
2573 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002574 WARN_ON_ONCE(pwq->flush_color != -1);
Tejun Heo73f53c42010-06-29 10:07:11 +02002575
Tejun Heo112202d2013-02-13 19:29:12 -08002576 if (pwq->nr_in_flight[flush_color]) {
2577 pwq->flush_color = flush_color;
2578 atomic_inc(&wq->nr_pwqs_to_flush);
Tejun Heo73f53c42010-06-29 10:07:11 +02002579 wait = true;
2580 }
2581 }
2582
2583 if (work_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002584 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
Tejun Heo112202d2013-02-13 19:29:12 -08002585 pwq->work_color = work_color;
Tejun Heo73f53c42010-06-29 10:07:11 +02002586 }
2587
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002588 spin_unlock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002589 }
2590
Tejun Heo112202d2013-02-13 19:29:12 -08002591 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
Tejun Heo73f53c42010-06-29 10:07:11 +02002592 complete(&wq->first_flusher->done);
2593
2594 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595}
2596
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002597/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002599 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002601 * This function sleeps until all work items which were queued on entry
2602 * have finished execution, but it is not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002604void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605{
Tejun Heo73f53c42010-06-29 10:07:11 +02002606 struct wq_flusher this_flusher = {
2607 .list = LIST_HEAD_INIT(this_flusher.list),
2608 .flush_color = -1,
2609 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2610 };
2611 int next_color;
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -07002612
Tejun Heo3347fa02016-09-16 15:49:32 -04002613 if (WARN_ON(!wq_online))
2614 return;
2615
Ingo Molnar3295f0e2008-08-11 10:30:30 +02002616 lock_map_acquire(&wq->lockdep_map);
2617 lock_map_release(&wq->lockdep_map);
Tejun Heo73f53c42010-06-29 10:07:11 +02002618
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002619 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002620
2621 /*
2622 * Start-to-wait phase
2623 */
2624 next_color = work_next_color(wq->work_color);
2625
2626 if (next_color != wq->flush_color) {
2627 /*
2628 * Color space is not full. The current work_color
2629 * becomes our flush_color and work_color is advanced
2630 * by one.
2631 */
Tejun Heo6183c002013-03-12 11:29:57 -07002632 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
Tejun Heo73f53c42010-06-29 10:07:11 +02002633 this_flusher.flush_color = wq->work_color;
2634 wq->work_color = next_color;
2635
2636 if (!wq->first_flusher) {
2637 /* no flush in progress, become the first flusher */
Tejun Heo6183c002013-03-12 11:29:57 -07002638 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002639
2640 wq->first_flusher = &this_flusher;
2641
Tejun Heo112202d2013-02-13 19:29:12 -08002642 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
Tejun Heo73f53c42010-06-29 10:07:11 +02002643 wq->work_color)) {
2644 /* nothing to flush, done */
2645 wq->flush_color = next_color;
2646 wq->first_flusher = NULL;
2647 goto out_unlock;
2648 }
2649 } else {
2650 /* wait in queue */
Tejun Heo6183c002013-03-12 11:29:57 -07002651 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002652 list_add_tail(&this_flusher.list, &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002653 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002654 }
2655 } else {
2656 /*
2657 * Oops, color space is full, wait on overflow queue.
2658 * The next flush completion will assign us
2659 * flush_color and transfer to flusher_queue.
2660 */
2661 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2662 }
2663
Tejun Heofca839c2015-12-07 10:58:57 -05002664 check_flush_dependency(wq, NULL);
2665
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002666 mutex_unlock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002667
2668 wait_for_completion(&this_flusher.done);
2669
2670 /*
2671 * Wake-up-and-cascade phase
2672 *
2673 * First flushers are responsible for cascading flushes and
2674 * handling overflow. Non-first flushers can simply return.
2675 */
2676 if (wq->first_flusher != &this_flusher)
2677 return;
2678
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002679 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002680
Tejun Heo4ce48b32010-07-02 10:03:51 +02002681 /* we might have raced, check again with mutex held */
2682 if (wq->first_flusher != &this_flusher)
2683 goto out_unlock;
2684
Tejun Heo73f53c42010-06-29 10:07:11 +02002685 wq->first_flusher = NULL;
2686
Tejun Heo6183c002013-03-12 11:29:57 -07002687 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2688 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002689
2690 while (true) {
2691 struct wq_flusher *next, *tmp;
2692
2693 /* complete all the flushers sharing the current flush color */
2694 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2695 if (next->flush_color != wq->flush_color)
2696 break;
2697 list_del_init(&next->list);
2698 complete(&next->done);
2699 }
2700
Tejun Heo6183c002013-03-12 11:29:57 -07002701 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2702 wq->flush_color != work_next_color(wq->work_color));
Tejun Heo73f53c42010-06-29 10:07:11 +02002703
2704 /* this flush_color is finished, advance by one */
2705 wq->flush_color = work_next_color(wq->flush_color);
2706
2707 /* one color has been freed, handle overflow queue */
2708 if (!list_empty(&wq->flusher_overflow)) {
2709 /*
2710 * Assign the same color to all overflowed
2711 * flushers, advance work_color and append to
2712 * flusher_queue. This is the start-to-wait
2713 * phase for these overflowed flushers.
2714 */
2715 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2716 tmp->flush_color = wq->work_color;
2717
2718 wq->work_color = work_next_color(wq->work_color);
2719
2720 list_splice_tail_init(&wq->flusher_overflow,
2721 &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002722 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002723 }
2724
2725 if (list_empty(&wq->flusher_queue)) {
Tejun Heo6183c002013-03-12 11:29:57 -07002726 WARN_ON_ONCE(wq->flush_color != wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002727 break;
2728 }
2729
2730 /*
2731 * Need to flush more colors. Make the next flusher
Tejun Heo112202d2013-02-13 19:29:12 -08002732 * the new first flusher and arm pwqs.
Tejun Heo73f53c42010-06-29 10:07:11 +02002733 */
Tejun Heo6183c002013-03-12 11:29:57 -07002734 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2735 WARN_ON_ONCE(wq->flush_color != next->flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002736
2737 list_del_init(&next->list);
2738 wq->first_flusher = next;
2739
Tejun Heo112202d2013-02-13 19:29:12 -08002740 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
Tejun Heo73f53c42010-06-29 10:07:11 +02002741 break;
2742
2743 /*
2744 * Meh... this color is already done, clear first
2745 * flusher and repeat cascading.
2746 */
2747 wq->first_flusher = NULL;
2748 }
2749
2750out_unlock:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002751 mutex_unlock(&wq->mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752}
Tim Gardner1dadafa2015-08-04 11:26:04 -06002753EXPORT_SYMBOL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002755/**
2756 * drain_workqueue - drain a workqueue
2757 * @wq: workqueue to drain
2758 *
2759 * Wait until the workqueue becomes empty. While draining is in progress,
2760 * only chain queueing is allowed. IOW, only currently pending or running
2761 * work items on @wq can queue further work items on it. @wq is flushed
Chen Hanxiaob749b1b2015-05-13 06:10:05 -04002762 * repeatedly until it becomes empty. The number of flushing is determined
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002763 * by the depth of chaining and should be relatively short. Whine if it
2764 * takes too long.
2765 */
2766void drain_workqueue(struct workqueue_struct *wq)
2767{
2768 unsigned int flush_cnt = 0;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002769 struct pool_workqueue *pwq;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002770
2771 /*
2772 * __queue_work() needs to test whether there are drainers, is much
2773 * hotter than drain_workqueue() and already looks at @wq->flags.
Tejun Heo618b01e2013-03-12 11:30:04 -07002774 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002775 */
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002776 mutex_lock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002777 if (!wq->nr_drainers++)
Tejun Heo618b01e2013-03-12 11:30:04 -07002778 wq->flags |= __WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002779 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002780reflush:
2781 flush_workqueue(wq);
2782
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002783 mutex_lock(&wq->mutex);
Tejun Heo76af4d92013-03-12 11:30:00 -07002784
Tejun Heo49e3cf42013-03-12 11:29:58 -07002785 for_each_pwq(pwq, wq) {
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002786 bool drained;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002787
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002788 spin_lock_irq(&pwq->pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08002789 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002790 spin_unlock_irq(&pwq->pool->lock);
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002791
2792 if (drained)
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002793 continue;
2794
2795 if (++flush_cnt == 10 ||
2796 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002797 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
Valentin Ilie044c7822012-08-19 00:52:42 +03002798 wq->name, flush_cnt);
Tejun Heo76af4d92013-03-12 11:30:00 -07002799
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002800 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002801 goto reflush;
2802 }
2803
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002804 if (!--wq->nr_drainers)
Tejun Heo618b01e2013-03-12 11:30:04 -07002805 wq->flags &= ~__WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002806 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002807}
2808EXPORT_SYMBOL_GPL(drain_workqueue);
2809
Tejun Heo606a5022012-08-20 14:51:23 -07002810static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
Tejun Heobaf59022010-09-16 10:42:16 +02002811{
2812 struct worker *worker = NULL;
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002813 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002814 struct pool_workqueue *pwq;
Tejun Heobaf59022010-09-16 10:42:16 +02002815
2816 might_sleep();
Tejun Heobaf59022010-09-16 10:42:16 +02002817
Tejun Heofa1b54e2013-03-12 11:30:00 -07002818 local_irq_disable();
2819 pool = get_work_pool(work);
2820 if (!pool) {
2821 local_irq_enable();
2822 return false;
2823 }
2824
2825 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08002826 /* see the comment in try_to_grab_pending() with the same code */
Tejun Heo112202d2013-02-13 19:29:12 -08002827 pwq = get_work_pwq(work);
2828 if (pwq) {
2829 if (unlikely(pwq->pool != pool))
Tejun Heobaf59022010-09-16 10:42:16 +02002830 goto already_gone;
Tejun Heo606a5022012-08-20 14:51:23 -07002831 } else {
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002832 worker = find_worker_executing_work(pool, work);
Tejun Heobaf59022010-09-16 10:42:16 +02002833 if (!worker)
2834 goto already_gone;
Tejun Heo112202d2013-02-13 19:29:12 -08002835 pwq = worker->current_pwq;
Tejun Heo606a5022012-08-20 14:51:23 -07002836 }
Tejun Heobaf59022010-09-16 10:42:16 +02002837
Tejun Heofca839c2015-12-07 10:58:57 -05002838 check_flush_dependency(pwq->wq, work);
2839
Tejun Heo112202d2013-02-13 19:29:12 -08002840 insert_wq_barrier(pwq, barr, work, worker);
Tejun Heod565ed62013-01-24 11:01:33 -08002841 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002842
Tejun Heoe1594892011-01-09 23:32:15 +01002843 /*
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002844 * Force a lock recursion deadlock when using flush_work() inside a
2845 * single-threaded or rescuer equipped workqueue.
2846 *
2847 * For single threaded workqueues the deadlock happens when the work
2848 * is after the work issuing the flush_work(). For rescuer equipped
2849 * workqueues the deadlock happens when the rescuer stalls, blocking
2850 * forward progress.
Tejun Heoe1594892011-01-09 23:32:15 +01002851 */
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002852 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) {
Tejun Heo112202d2013-02-13 19:29:12 -08002853 lock_map_acquire(&pwq->wq->lockdep_map);
Peter Zijlstraa1d14932017-08-23 12:52:32 +02002854 lock_map_release(&pwq->wq->lockdep_map);
2855 }
Tejun Heoe1594892011-01-09 23:32:15 +01002856
Tejun Heobaf59022010-09-16 10:42:16 +02002857 return true;
2858already_gone:
Tejun Heod565ed62013-01-24 11:01:33 -08002859 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002860 return false;
2861}
2862
Oleg Nesterovdb700892008-07-25 01:47:49 -07002863/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002864 * flush_work - wait for a work to finish executing the last queueing instance
2865 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07002866 *
Tejun Heo606a5022012-08-20 14:51:23 -07002867 * Wait until @work has finished execution. @work is guaranteed to be idle
2868 * on return if it hasn't been requeued since flush started.
Tejun Heo401a8d02010-09-16 10:36:00 +02002869 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002870 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002871 * %true if flush_work() waited for the work to finish execution,
2872 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07002873 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002874bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07002875{
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002876 struct wq_barrier barr;
2877
Tejun Heo3347fa02016-09-16 15:49:32 -04002878 if (WARN_ON(!wq_online))
2879 return false;
2880
Stephen Boyd0976dfc2012-04-20 17:28:50 -07002881 lock_map_acquire(&work->lockdep_map);
2882 lock_map_release(&work->lockdep_map);
2883
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002884 if (start_flush_work(work, &barr)) {
2885 wait_for_completion(&barr.done);
2886 destroy_work_on_stack(&barr.work);
2887 return true;
2888 } else {
2889 return false;
2890 }
Oleg Nesterovdb700892008-07-25 01:47:49 -07002891}
2892EXPORT_SYMBOL_GPL(flush_work);
2893
Tejun Heo8603e1b32015-03-05 08:04:13 -05002894struct cwt_wait {
Ingo Molnarac6424b2017-06-20 12:06:13 +02002895 wait_queue_entry_t wait;
Tejun Heo8603e1b32015-03-05 08:04:13 -05002896 struct work_struct *work;
2897};
2898
Ingo Molnarac6424b2017-06-20 12:06:13 +02002899static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
Tejun Heo8603e1b32015-03-05 08:04:13 -05002900{
2901 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2902
2903 if (cwait->work != key)
2904 return 0;
2905 return autoremove_wake_function(wait, mode, sync, key);
2906}
2907
Tejun Heo36e227d2012-08-03 10:30:46 -07002908static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
Tejun Heo401a8d02010-09-16 10:36:00 +02002909{
Tejun Heo8603e1b32015-03-05 08:04:13 -05002910 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
Tejun Heobbb68df2012-08-03 10:30:46 -07002911 unsigned long flags;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002912 int ret;
2913
2914 do {
Tejun Heobbb68df2012-08-03 10:30:46 -07002915 ret = try_to_grab_pending(work, is_dwork, &flags);
2916 /*
Tejun Heo8603e1b32015-03-05 08:04:13 -05002917 * If someone else is already canceling, wait for it to
2918 * finish. flush_work() doesn't work for PREEMPT_NONE
2919 * because we may get scheduled between @work's completion
2920 * and the other canceling task resuming and clearing
2921 * CANCELING - flush_work() will return false immediately
2922 * as @work is no longer busy, try_to_grab_pending() will
2923 * return -ENOENT as @work is still being canceled and the
2924 * other canceling task won't be able to clear CANCELING as
2925 * we're hogging the CPU.
2926 *
2927 * Let's wait for completion using a waitqueue. As this
2928 * may lead to the thundering herd problem, use a custom
2929 * wake function which matches @work along with exclusive
2930 * wait and wakeup.
Tejun Heobbb68df2012-08-03 10:30:46 -07002931 */
Tejun Heo8603e1b32015-03-05 08:04:13 -05002932 if (unlikely(ret == -ENOENT)) {
2933 struct cwt_wait cwait;
2934
2935 init_wait(&cwait.wait);
2936 cwait.wait.func = cwt_wakefn;
2937 cwait.work = work;
2938
2939 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2940 TASK_UNINTERRUPTIBLE);
2941 if (work_is_canceling(work))
2942 schedule();
2943 finish_wait(&cancel_waitq, &cwait.wait);
2944 }
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002945 } while (unlikely(ret < 0));
2946
Tejun Heobbb68df2012-08-03 10:30:46 -07002947 /* tell other tasks trying to grab @work to back off */
2948 mark_work_canceling(work);
2949 local_irq_restore(flags);
2950
Tejun Heo3347fa02016-09-16 15:49:32 -04002951 /*
2952 * This allows canceling during early boot. We know that @work
2953 * isn't executing.
2954 */
2955 if (wq_online)
2956 flush_work(work);
2957
Tejun Heo7a22ad72010-06-29 10:07:13 +02002958 clear_work_data(work);
Tejun Heo8603e1b32015-03-05 08:04:13 -05002959
2960 /*
2961 * Paired with prepare_to_wait() above so that either
2962 * waitqueue_active() is visible here or !work_is_canceling() is
2963 * visible there.
2964 */
2965 smp_mb();
2966 if (waitqueue_active(&cancel_waitq))
2967 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2968
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002969 return ret;
2970}
2971
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002972/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002973 * cancel_work_sync - cancel a work and wait for it to finish
2974 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002975 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002976 * Cancel @work and wait for its execution to finish. This function
2977 * can be used even if the work re-queues itself or migrates to
2978 * another workqueue. On return from this function, @work is
2979 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002980 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002981 * cancel_work_sync(&delayed_work->work) must not be used for
2982 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002983 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002984 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002985 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02002986 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002987 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002988 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002989 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002990bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002991{
Tejun Heo36e227d2012-08-03 10:30:46 -07002992 return __cancel_work_timer(work, false);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002993}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07002994EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002995
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002996/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002997 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2998 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002999 *
Tejun Heo401a8d02010-09-16 10:36:00 +02003000 * Delayed timer is cancelled and the pending work is queued for
3001 * immediate execution. Like flush_work(), this function only
3002 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07003003 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003004 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003005 * %true if flush_work() waited for the work to finish execution,
3006 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003007 */
Tejun Heo401a8d02010-09-16 10:36:00 +02003008bool flush_delayed_work(struct delayed_work *dwork)
3009{
Tejun Heo8930cab2012-08-03 10:30:45 -07003010 local_irq_disable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003011 if (del_timer_sync(&dwork->timer))
Lai Jiangshan60c057b2013-02-06 18:04:53 -08003012 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Tejun Heo8930cab2012-08-03 10:30:45 -07003013 local_irq_enable();
Tejun Heo401a8d02010-09-16 10:36:00 +02003014 return flush_work(&dwork->work);
3015}
3016EXPORT_SYMBOL(flush_delayed_work);
3017
Jens Axboef72b8792016-08-24 15:51:50 -06003018static bool __cancel_work(struct work_struct *work, bool is_dwork)
3019{
3020 unsigned long flags;
3021 int ret;
3022
3023 do {
3024 ret = try_to_grab_pending(work, is_dwork, &flags);
3025 } while (unlikely(ret == -EAGAIN));
3026
3027 if (unlikely(ret < 0))
3028 return false;
3029
3030 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3031 local_irq_restore(flags);
3032 return ret;
3033}
3034
3035/*
3036 * See cancel_delayed_work()
3037 */
3038bool cancel_work(struct work_struct *work)
3039{
3040 return __cancel_work(work, false);
3041}
3042
Tejun Heo401a8d02010-09-16 10:36:00 +02003043/**
Tejun Heo57b30ae2012-08-21 13:18:24 -07003044 * cancel_delayed_work - cancel a delayed work
3045 * @dwork: delayed_work to cancel
Tejun Heo09383492010-09-16 10:48:29 +02003046 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003047 * Kill off a pending delayed_work.
3048 *
3049 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3050 * pending.
3051 *
3052 * Note:
3053 * The work callback function may still be running on return, unless
3054 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3055 * use cancel_delayed_work_sync() to wait on it.
Tejun Heo09383492010-09-16 10:48:29 +02003056 *
Tejun Heo57b30ae2012-08-21 13:18:24 -07003057 * This function is safe to call from any context including IRQ handler.
Tejun Heo09383492010-09-16 10:48:29 +02003058 */
Tejun Heo57b30ae2012-08-21 13:18:24 -07003059bool cancel_delayed_work(struct delayed_work *dwork)
Tejun Heo09383492010-09-16 10:48:29 +02003060{
Jens Axboef72b8792016-08-24 15:51:50 -06003061 return __cancel_work(&dwork->work, true);
Tejun Heo09383492010-09-16 10:48:29 +02003062}
Tejun Heo57b30ae2012-08-21 13:18:24 -07003063EXPORT_SYMBOL(cancel_delayed_work);
Tejun Heo09383492010-09-16 10:48:29 +02003064
3065/**
Tejun Heo401a8d02010-09-16 10:36:00 +02003066 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3067 * @dwork: the delayed work cancel
3068 *
3069 * This is cancel_work_sync() for delayed works.
3070 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003071 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02003072 * %true if @dwork was pending, %false otherwise.
3073 */
3074bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003075{
Tejun Heo36e227d2012-08-03 10:30:46 -07003076 return __cancel_work_timer(&dwork->work, true);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07003077}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07003078EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07003080/**
Tejun Heo31ddd872010-10-19 11:14:49 +02003081 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07003082 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07003083 *
Tejun Heo31ddd872010-10-19 11:14:49 +02003084 * schedule_on_each_cpu() executes @func on each online CPU using the
3085 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07003086 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02003087 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003088 * Return:
Tejun Heo31ddd872010-10-19 11:14:49 +02003089 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07003090 */
David Howells65f27f32006-11-22 14:55:48 +00003091int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003092{
3093 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02003094 struct work_struct __percpu *works;
Christoph Lameter15316ba2006-01-08 01:00:43 -08003095
Andrew Mortonb6136772006-06-25 05:47:49 -07003096 works = alloc_percpu(struct work_struct);
3097 if (!works)
Christoph Lameter15316ba2006-01-08 01:00:43 -08003098 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07003099
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003100 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08003101
Christoph Lameter15316ba2006-01-08 01:00:43 -08003102 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01003103 struct work_struct *work = per_cpu_ptr(works, cpu);
3104
3105 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003106 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02003107 }
Tejun Heo93981802009-11-17 14:06:20 -08003108
3109 for_each_online_cpu(cpu)
3110 flush_work(per_cpu_ptr(works, cpu));
3111
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003112 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07003113 free_percpu(works);
Christoph Lameter15316ba2006-01-08 01:00:43 -08003114 return 0;
3115}
3116
Alan Sterneef6a7d2010-02-12 17:39:21 +09003117/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06003118 * execute_in_process_context - reliably execute the routine with user context
3119 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06003120 * @ew: guaranteed storage for the execute work structure (must
3121 * be available when the work executes)
3122 *
3123 * Executes the function immediately if process context is available,
3124 * otherwise schedules the function for delayed execution.
3125 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003126 * Return: 0 - function was executed
James Bottomley1fa44ec2006-02-23 12:43:43 -06003127 * 1 - function was scheduled for execution
3128 */
David Howells65f27f32006-11-22 14:55:48 +00003129int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06003130{
3131 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00003132 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003133 return 0;
3134 }
3135
David Howells65f27f32006-11-22 14:55:48 +00003136 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003137 schedule_work(&ew->work);
3138
3139 return 1;
3140}
3141EXPORT_SYMBOL_GPL(execute_in_process_context);
3142
Tejun Heo7a4e3442013-03-12 11:30:00 -07003143/**
3144 * free_workqueue_attrs - free a workqueue_attrs
3145 * @attrs: workqueue_attrs to free
3146 *
3147 * Undo alloc_workqueue_attrs().
3148 */
3149void free_workqueue_attrs(struct workqueue_attrs *attrs)
3150{
3151 if (attrs) {
3152 free_cpumask_var(attrs->cpumask);
3153 kfree(attrs);
3154 }
3155}
3156
3157/**
3158 * alloc_workqueue_attrs - allocate a workqueue_attrs
3159 * @gfp_mask: allocation mask to use
3160 *
3161 * Allocate a new workqueue_attrs, initialize with default settings and
Yacine Belkadid185af32013-07-31 14:59:24 -07003162 * return it.
3163 *
3164 * Return: The allocated new workqueue_attr on success. %NULL on failure.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003165 */
3166struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3167{
3168 struct workqueue_attrs *attrs;
3169
3170 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3171 if (!attrs)
3172 goto fail;
3173 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3174 goto fail;
3175
Tejun Heo13e2e552013-04-01 11:23:31 -07003176 cpumask_copy(attrs->cpumask, cpu_possible_mask);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003177 return attrs;
3178fail:
3179 free_workqueue_attrs(attrs);
3180 return NULL;
3181}
3182
Tejun Heo29c91e92013-03-12 11:30:03 -07003183static void copy_workqueue_attrs(struct workqueue_attrs *to,
3184 const struct workqueue_attrs *from)
3185{
3186 to->nice = from->nice;
3187 cpumask_copy(to->cpumask, from->cpumask);
Shaohua Li2865a8f2013-08-01 09:56:36 +08003188 /*
3189 * Unlike hash and equality test, this function doesn't ignore
3190 * ->no_numa as it is used for both pool and wq attrs. Instead,
3191 * get_unbound_pool() explicitly clears ->no_numa after copying.
3192 */
3193 to->no_numa = from->no_numa;
Tejun Heo29c91e92013-03-12 11:30:03 -07003194}
3195
Tejun Heo29c91e92013-03-12 11:30:03 -07003196/* hash value of the content of @attr */
3197static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3198{
3199 u32 hash = 0;
3200
3201 hash = jhash_1word(attrs->nice, hash);
Tejun Heo13e2e552013-04-01 11:23:31 -07003202 hash = jhash(cpumask_bits(attrs->cpumask),
3203 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
Tejun Heo29c91e92013-03-12 11:30:03 -07003204 return hash;
3205}
3206
3207/* content equality test */
3208static bool wqattrs_equal(const struct workqueue_attrs *a,
3209 const struct workqueue_attrs *b)
3210{
3211 if (a->nice != b->nice)
3212 return false;
3213 if (!cpumask_equal(a->cpumask, b->cpumask))
3214 return false;
3215 return true;
3216}
3217
Tejun Heo7a4e3442013-03-12 11:30:00 -07003218/**
3219 * init_worker_pool - initialize a newly zalloc'd worker_pool
3220 * @pool: worker_pool to initialize
3221 *
Shailendra Verma402dd892015-05-23 10:38:14 +05303222 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
Yacine Belkadid185af32013-07-31 14:59:24 -07003223 *
3224 * Return: 0 on success, -errno on failure. Even on failure, all fields
Tejun Heo29c91e92013-03-12 11:30:03 -07003225 * inside @pool proper are initialized and put_unbound_pool() can be called
3226 * on @pool safely to release it.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003227 */
3228static int init_worker_pool(struct worker_pool *pool)
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003229{
3230 spin_lock_init(&pool->lock);
Tejun Heo29c91e92013-03-12 11:30:03 -07003231 pool->id = -1;
3232 pool->cpu = -1;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003233 pool->node = NUMA_NO_NODE;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003234 pool->flags |= POOL_DISASSOCIATED;
Tejun Heo82607adc2015-12-08 11:28:04 -05003235 pool->watchdog_ts = jiffies;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003236 INIT_LIST_HEAD(&pool->worklist);
3237 INIT_LIST_HEAD(&pool->idle_list);
3238 hash_init(pool->busy_hash);
3239
Geliang Tangc30fb262017-02-26 08:58:21 +08003240 setup_deferrable_timer(&pool->idle_timer, idle_worker_timeout,
3241 (unsigned long)pool);
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003242
3243 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3244 (unsigned long)pool);
3245
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003246 mutex_init(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003247 INIT_LIST_HEAD(&pool->workers);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003248
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003249 ida_init(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003250 INIT_HLIST_NODE(&pool->hash_node);
3251 pool->refcnt = 1;
3252
3253 /* shouldn't fail above this point */
Tejun Heo7a4e3442013-03-12 11:30:00 -07003254 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3255 if (!pool->attrs)
3256 return -ENOMEM;
3257 return 0;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003258}
3259
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003260static void rcu_free_wq(struct rcu_head *rcu)
3261{
3262 struct workqueue_struct *wq =
3263 container_of(rcu, struct workqueue_struct, rcu);
3264
3265 if (!(wq->flags & WQ_UNBOUND))
3266 free_percpu(wq->cpu_pwqs);
3267 else
3268 free_workqueue_attrs(wq->unbound_attrs);
3269
3270 kfree(wq->rescuer);
3271 kfree(wq);
3272}
3273
Tejun Heo29c91e92013-03-12 11:30:03 -07003274static void rcu_free_pool(struct rcu_head *rcu)
3275{
3276 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3277
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003278 ida_destroy(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003279 free_workqueue_attrs(pool->attrs);
3280 kfree(pool);
3281}
3282
3283/**
3284 * put_unbound_pool - put a worker_pool
3285 * @pool: worker_pool to put
3286 *
3287 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003288 * safe manner. get_unbound_pool() calls this function on its failure path
3289 * and this function should be able to release pools which went through,
3290 * successfully or not, init_worker_pool().
Tejun Heoa892cac2013-04-01 11:23:32 -07003291 *
3292 * Should be called with wq_pool_mutex held.
Tejun Heo29c91e92013-03-12 11:30:03 -07003293 */
3294static void put_unbound_pool(struct worker_pool *pool)
3295{
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003296 DECLARE_COMPLETION_ONSTACK(detach_completion);
Tejun Heo29c91e92013-03-12 11:30:03 -07003297 struct worker *worker;
3298
Tejun Heoa892cac2013-04-01 11:23:32 -07003299 lockdep_assert_held(&wq_pool_mutex);
3300
3301 if (--pool->refcnt)
Tejun Heo29c91e92013-03-12 11:30:03 -07003302 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003303
3304 /* sanity checks */
Lai Jiangshan61d0fbb2014-06-03 15:31:45 +08003305 if (WARN_ON(!(pool->cpu < 0)) ||
Tejun Heoa892cac2013-04-01 11:23:32 -07003306 WARN_ON(!list_empty(&pool->worklist)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003307 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003308
3309 /* release id and unhash */
3310 if (pool->id >= 0)
3311 idr_remove(&worker_pool_idr, pool->id);
3312 hash_del(&pool->hash_node);
3313
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003314 /*
Tejun Heo692b4822017-10-09 08:04:13 -07003315 * Become the manager and destroy all workers. This prevents
3316 * @pool's workers from blocking on attach_mutex. We're the last
3317 * manager and @pool gets freed with the flag set.
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003318 */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003319 spin_lock_irq(&pool->lock);
Tejun Heo692b4822017-10-09 08:04:13 -07003320 wait_event_lock_irq(wq_manager_wait,
3321 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3322 pool->flags |= POOL_MANAGER_ACTIVE;
3323
Lai Jiangshan1037de32014-05-22 16:44:07 +08003324 while ((worker = first_idle_worker(pool)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003325 destroy_worker(worker);
3326 WARN_ON(pool->nr_workers || pool->nr_idle);
Tejun Heo29c91e92013-03-12 11:30:03 -07003327 spin_unlock_irq(&pool->lock);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003328
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003329 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003330 if (!list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003331 pool->detach_completion = &detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003332 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003333
3334 if (pool->detach_completion)
3335 wait_for_completion(pool->detach_completion);
3336
Tejun Heo29c91e92013-03-12 11:30:03 -07003337 /* shut down the timers */
3338 del_timer_sync(&pool->idle_timer);
3339 del_timer_sync(&pool->mayday_timer);
3340
3341 /* sched-RCU protected to allow dereferences from get_work_pool() */
3342 call_rcu_sched(&pool->rcu, rcu_free_pool);
3343}
3344
3345/**
3346 * get_unbound_pool - get a worker_pool with the specified attributes
3347 * @attrs: the attributes of the worker_pool to get
3348 *
3349 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3350 * reference count and return it. If there already is a matching
3351 * worker_pool, it will be used; otherwise, this function attempts to
Yacine Belkadid185af32013-07-31 14:59:24 -07003352 * create a new one.
Tejun Heoa892cac2013-04-01 11:23:32 -07003353 *
3354 * Should be called with wq_pool_mutex held.
Yacine Belkadid185af32013-07-31 14:59:24 -07003355 *
3356 * Return: On success, a worker_pool with the same attributes as @attrs.
3357 * On failure, %NULL.
Tejun Heo29c91e92013-03-12 11:30:03 -07003358 */
3359static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3360{
Tejun Heo29c91e92013-03-12 11:30:03 -07003361 u32 hash = wqattrs_hash(attrs);
3362 struct worker_pool *pool;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003363 int node;
Xunlei Pange2273582015-10-09 11:53:12 +08003364 int target_node = NUMA_NO_NODE;
Tejun Heo29c91e92013-03-12 11:30:03 -07003365
Tejun Heoa892cac2013-04-01 11:23:32 -07003366 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003367
3368 /* do we already have a matching pool? */
Tejun Heo29c91e92013-03-12 11:30:03 -07003369 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3370 if (wqattrs_equal(pool->attrs, attrs)) {
3371 pool->refcnt++;
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003372 return pool;
Tejun Heo29c91e92013-03-12 11:30:03 -07003373 }
3374 }
Tejun Heo29c91e92013-03-12 11:30:03 -07003375
Xunlei Pange2273582015-10-09 11:53:12 +08003376 /* if cpumask is contained inside a NUMA node, we belong to that node */
3377 if (wq_numa_enabled) {
3378 for_each_node(node) {
3379 if (cpumask_subset(attrs->cpumask,
3380 wq_numa_possible_cpumask[node])) {
3381 target_node = node;
3382 break;
3383 }
3384 }
3385 }
3386
Tejun Heo29c91e92013-03-12 11:30:03 -07003387 /* nope, create a new one */
Xunlei Pange2273582015-10-09 11:53:12 +08003388 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
Tejun Heo29c91e92013-03-12 11:30:03 -07003389 if (!pool || init_worker_pool(pool) < 0)
3390 goto fail;
3391
Tejun Heo8864b4e2013-03-12 11:30:04 -07003392 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
Tejun Heo29c91e92013-03-12 11:30:03 -07003393 copy_workqueue_attrs(pool->attrs, attrs);
Xunlei Pange2273582015-10-09 11:53:12 +08003394 pool->node = target_node;
Tejun Heo29c91e92013-03-12 11:30:03 -07003395
Shaohua Li2865a8f2013-08-01 09:56:36 +08003396 /*
3397 * no_numa isn't a worker_pool attribute, always clear it. See
3398 * 'struct workqueue_attrs' comments for detail.
3399 */
3400 pool->attrs->no_numa = false;
3401
Tejun Heo29c91e92013-03-12 11:30:03 -07003402 if (worker_pool_assign_id(pool) < 0)
3403 goto fail;
3404
3405 /* create and start the initial worker */
Tejun Heo3347fa02016-09-16 15:49:32 -04003406 if (wq_online && !create_worker(pool))
Tejun Heo29c91e92013-03-12 11:30:03 -07003407 goto fail;
3408
Tejun Heo29c91e92013-03-12 11:30:03 -07003409 /* install */
Tejun Heo29c91e92013-03-12 11:30:03 -07003410 hash_add(unbound_pool_hash, &pool->hash_node, hash);
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003411
Tejun Heo29c91e92013-03-12 11:30:03 -07003412 return pool;
3413fail:
Tejun Heo29c91e92013-03-12 11:30:03 -07003414 if (pool)
3415 put_unbound_pool(pool);
3416 return NULL;
3417}
3418
Tejun Heo8864b4e2013-03-12 11:30:04 -07003419static void rcu_free_pwq(struct rcu_head *rcu)
3420{
3421 kmem_cache_free(pwq_cache,
3422 container_of(rcu, struct pool_workqueue, rcu));
3423}
3424
3425/*
3426 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3427 * and needs to be destroyed.
3428 */
3429static void pwq_unbound_release_workfn(struct work_struct *work)
3430{
3431 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3432 unbound_release_work);
3433 struct workqueue_struct *wq = pwq->wq;
3434 struct worker_pool *pool = pwq->pool;
Tejun Heobc0caf02013-04-01 11:23:31 -07003435 bool is_last;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003436
3437 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3438 return;
3439
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003440 mutex_lock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003441 list_del_rcu(&pwq->pwqs_node);
Tejun Heobc0caf02013-04-01 11:23:31 -07003442 is_last = list_empty(&wq->pwqs);
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003443 mutex_unlock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003444
Tejun Heoa892cac2013-04-01 11:23:32 -07003445 mutex_lock(&wq_pool_mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003446 put_unbound_pool(pool);
Tejun Heoa892cac2013-04-01 11:23:32 -07003447 mutex_unlock(&wq_pool_mutex);
3448
Tejun Heo8864b4e2013-03-12 11:30:04 -07003449 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3450
3451 /*
3452 * If we're the last pwq going away, @wq is already dead and no one
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003453 * is gonna access it anymore. Schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003454 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003455 if (is_last)
3456 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003457}
3458
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003459/**
Tejun Heo699ce092013-03-13 16:51:35 -07003460 * pwq_adjust_max_active - update a pwq's max_active to the current setting
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003461 * @pwq: target pool_workqueue
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003462 *
Tejun Heo699ce092013-03-13 16:51:35 -07003463 * If @pwq isn't freezing, set @pwq->max_active to the associated
3464 * workqueue's saved_max_active and activate delayed work items
3465 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003466 */
Tejun Heo699ce092013-03-13 16:51:35 -07003467static void pwq_adjust_max_active(struct pool_workqueue *pwq)
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003468{
Tejun Heo699ce092013-03-13 16:51:35 -07003469 struct workqueue_struct *wq = pwq->wq;
3470 bool freezable = wq->flags & WQ_FREEZABLE;
Tejun Heo3347fa02016-09-16 15:49:32 -04003471 unsigned long flags;
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003472
Tejun Heo699ce092013-03-13 16:51:35 -07003473 /* for @wq->saved_max_active */
Lai Jiangshana357fc02013-03-25 16:57:19 -07003474 lockdep_assert_held(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003475
3476 /* fast exit for non-freezable wqs */
3477 if (!freezable && pwq->max_active == wq->saved_max_active)
3478 return;
3479
Tejun Heo3347fa02016-09-16 15:49:32 -04003480 /* this function can be called during early boot w/ irq disabled */
3481 spin_lock_irqsave(&pwq->pool->lock, flags);
Tejun Heo699ce092013-03-13 16:51:35 -07003482
Lai Jiangshan74b414e2014-05-22 19:01:16 +08003483 /*
3484 * During [un]freezing, the caller is responsible for ensuring that
3485 * this function is called at least once after @workqueue_freezing
3486 * is updated and visible.
3487 */
3488 if (!freezable || !workqueue_freezing) {
Tejun Heo699ce092013-03-13 16:51:35 -07003489 pwq->max_active = wq->saved_max_active;
3490
3491 while (!list_empty(&pwq->delayed_works) &&
3492 pwq->nr_active < pwq->max_active)
3493 pwq_activate_first_delayed(pwq);
Lai Jiangshan951a0782013-03-20 10:52:30 -07003494
3495 /*
3496 * Need to kick a worker after thawed or an unbound wq's
3497 * max_active is bumped. It's a slow path. Do it always.
3498 */
3499 wake_up_worker(pwq->pool);
Tejun Heo699ce092013-03-13 16:51:35 -07003500 } else {
3501 pwq->max_active = 0;
3502 }
3503
Tejun Heo3347fa02016-09-16 15:49:32 -04003504 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003505}
3506
Tejun Heoe50aba92013-04-01 11:23:35 -07003507/* initialize newly alloced @pwq which is associated with @wq and @pool */
Tejun Heof147f292013-04-01 11:23:35 -07003508static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3509 struct worker_pool *pool)
Tejun Heod2c1d402013-03-12 11:30:04 -07003510{
3511 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3512
Tejun Heoe50aba92013-04-01 11:23:35 -07003513 memset(pwq, 0, sizeof(*pwq));
3514
Tejun Heod2c1d402013-03-12 11:30:04 -07003515 pwq->pool = pool;
3516 pwq->wq = wq;
3517 pwq->flush_color = -1;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003518 pwq->refcnt = 1;
Tejun Heod2c1d402013-03-12 11:30:04 -07003519 INIT_LIST_HEAD(&pwq->delayed_works);
Tejun Heo1befcf32013-04-01 11:23:35 -07003520 INIT_LIST_HEAD(&pwq->pwqs_node);
Tejun Heod2c1d402013-03-12 11:30:04 -07003521 INIT_LIST_HEAD(&pwq->mayday_node);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003522 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
Tejun Heof147f292013-04-01 11:23:35 -07003523}
Tejun Heod2c1d402013-03-12 11:30:04 -07003524
Tejun Heof147f292013-04-01 11:23:35 -07003525/* sync @pwq with the current state of its associated wq and link it */
Tejun Heo1befcf32013-04-01 11:23:35 -07003526static void link_pwq(struct pool_workqueue *pwq)
Tejun Heof147f292013-04-01 11:23:35 -07003527{
3528 struct workqueue_struct *wq = pwq->wq;
3529
3530 lockdep_assert_held(&wq->mutex);
Tejun Heo75ccf592013-03-12 11:30:04 -07003531
Tejun Heo1befcf32013-04-01 11:23:35 -07003532 /* may be called multiple times, ignore if already linked */
3533 if (!list_empty(&pwq->pwqs_node))
3534 return;
3535
Lai Jiangshan29b1cb42014-07-22 13:04:27 +08003536 /* set the matching work_color */
Tejun Heo75ccf592013-03-12 11:30:04 -07003537 pwq->work_color = wq->work_color;
Tejun Heo983ca252013-03-13 16:51:35 -07003538
3539 /* sync max_active to the current setting */
3540 pwq_adjust_max_active(pwq);
3541
3542 /* link in @pwq */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003543 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
Tejun Heof147f292013-04-01 11:23:35 -07003544}
Lai Jiangshana357fc02013-03-25 16:57:19 -07003545
Tejun Heof147f292013-04-01 11:23:35 -07003546/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3547static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3548 const struct workqueue_attrs *attrs)
3549{
3550 struct worker_pool *pool;
3551 struct pool_workqueue *pwq;
3552
3553 lockdep_assert_held(&wq_pool_mutex);
3554
3555 pool = get_unbound_pool(attrs);
3556 if (!pool)
3557 return NULL;
3558
Tejun Heoe50aba92013-04-01 11:23:35 -07003559 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
Tejun Heof147f292013-04-01 11:23:35 -07003560 if (!pwq) {
3561 put_unbound_pool(pool);
3562 return NULL;
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003563 }
Tejun Heo6029a912013-04-01 11:23:34 -07003564
Tejun Heof147f292013-04-01 11:23:35 -07003565 init_pwq(pwq, wq, pool);
3566 return pwq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003567}
3568
Tejun Heo4c16bd32013-04-01 11:23:36 -07003569/**
Gong Zhaogang30186c62015-05-11 11:02:47 -04003570 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003571 * @attrs: the wq_attrs of the default pwq of the target workqueue
Tejun Heo4c16bd32013-04-01 11:23:36 -07003572 * @node: the target NUMA node
3573 * @cpu_going_down: if >= 0, the CPU to consider as offline
3574 * @cpumask: outarg, the resulting cpumask
3575 *
3576 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3577 * @cpu_going_down is >= 0, that cpu is considered offline during
Yacine Belkadid185af32013-07-31 14:59:24 -07003578 * calculation. The result is stored in @cpumask.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003579 *
3580 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3581 * enabled and @node has online CPUs requested by @attrs, the returned
3582 * cpumask is the intersection of the possible CPUs of @node and
3583 * @attrs->cpumask.
3584 *
3585 * The caller is responsible for ensuring that the cpumask of @node stays
3586 * stable.
Yacine Belkadid185af32013-07-31 14:59:24 -07003587 *
3588 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3589 * %false if equal.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003590 */
3591static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3592 int cpu_going_down, cpumask_t *cpumask)
3593{
Tejun Heod55262c2013-04-01 11:23:38 -07003594 if (!wq_numa_enabled || attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003595 goto use_dfl;
3596
3597 /* does @node have any online CPUs @attrs wants? */
3598 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3599 if (cpu_going_down >= 0)
3600 cpumask_clear_cpu(cpu_going_down, cpumask);
3601
3602 if (cpumask_empty(cpumask))
3603 goto use_dfl;
3604
3605 /* yeap, return possible CPUs in @node that @attrs wants */
3606 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
Michael Bringmann1ad0f0a2017-07-27 16:27:14 -05003607
3608 if (cpumask_empty(cpumask)) {
3609 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3610 "possible intersect\n");
3611 return false;
3612 }
3613
Tejun Heo4c16bd32013-04-01 11:23:36 -07003614 return !cpumask_equal(cpumask, attrs->cpumask);
3615
3616use_dfl:
3617 cpumask_copy(cpumask, attrs->cpumask);
3618 return false;
3619}
3620
Tejun Heo1befcf32013-04-01 11:23:35 -07003621/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3622static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3623 int node,
3624 struct pool_workqueue *pwq)
3625{
3626 struct pool_workqueue *old_pwq;
3627
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +08003628 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003629 lockdep_assert_held(&wq->mutex);
3630
3631 /* link_pwq() can handle duplicate calls */
3632 link_pwq(pwq);
3633
3634 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3635 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3636 return old_pwq;
3637}
3638
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003639/* context to store the prepared attrs & pwqs before applying */
3640struct apply_wqattrs_ctx {
3641 struct workqueue_struct *wq; /* target workqueue */
3642 struct workqueue_attrs *attrs; /* attrs to apply */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003643 struct list_head list; /* queued for batching commit */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003644 struct pool_workqueue *dfl_pwq;
3645 struct pool_workqueue *pwq_tbl[];
3646};
3647
3648/* free the resources after success or abort */
3649static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3650{
3651 if (ctx) {
3652 int node;
3653
3654 for_each_node(node)
3655 put_pwq_unlocked(ctx->pwq_tbl[node]);
3656 put_pwq_unlocked(ctx->dfl_pwq);
3657
3658 free_workqueue_attrs(ctx->attrs);
3659
3660 kfree(ctx);
3661 }
3662}
3663
3664/* allocate the attrs and pwqs for later installation */
3665static struct apply_wqattrs_ctx *
3666apply_wqattrs_prepare(struct workqueue_struct *wq,
3667 const struct workqueue_attrs *attrs)
3668{
3669 struct apply_wqattrs_ctx *ctx;
3670 struct workqueue_attrs *new_attrs, *tmp_attrs;
3671 int node;
3672
3673 lockdep_assert_held(&wq_pool_mutex);
3674
3675 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3676 GFP_KERNEL);
3677
3678 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3679 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3680 if (!ctx || !new_attrs || !tmp_attrs)
3681 goto out_free;
3682
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003683 /*
3684 * Calculate the attrs of the default pwq.
3685 * If the user configured cpumask doesn't overlap with the
3686 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3687 */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003688 copy_workqueue_attrs(new_attrs, attrs);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08003689 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003690 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3691 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003692
3693 /*
3694 * We may create multiple pwqs with differing cpumasks. Make a
3695 * copy of @new_attrs which will be modified and used to obtain
3696 * pools.
3697 */
3698 copy_workqueue_attrs(tmp_attrs, new_attrs);
3699
3700 /*
3701 * If something goes wrong during CPU up/down, we'll fall back to
3702 * the default pwq covering whole @attrs->cpumask. Always create
3703 * it even if we don't use it immediately.
3704 */
3705 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3706 if (!ctx->dfl_pwq)
3707 goto out_free;
3708
3709 for_each_node(node) {
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003710 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003711 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3712 if (!ctx->pwq_tbl[node])
3713 goto out_free;
3714 } else {
3715 ctx->dfl_pwq->refcnt++;
3716 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3717 }
3718 }
3719
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003720 /* save the user configured attrs and sanitize it. */
3721 copy_workqueue_attrs(new_attrs, attrs);
3722 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003723 ctx->attrs = new_attrs;
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003724
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003725 ctx->wq = wq;
3726 free_workqueue_attrs(tmp_attrs);
3727 return ctx;
3728
3729out_free:
3730 free_workqueue_attrs(tmp_attrs);
3731 free_workqueue_attrs(new_attrs);
3732 apply_wqattrs_cleanup(ctx);
3733 return NULL;
3734}
3735
3736/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3737static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3738{
3739 int node;
3740
3741 /* all pwqs have been created successfully, let's install'em */
3742 mutex_lock(&ctx->wq->mutex);
3743
3744 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3745
3746 /* save the previous pwq and install the new one */
3747 for_each_node(node)
3748 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3749 ctx->pwq_tbl[node]);
3750
3751 /* @dfl_pwq might not have been used, ensure it's linked */
3752 link_pwq(ctx->dfl_pwq);
3753 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3754
3755 mutex_unlock(&ctx->wq->mutex);
3756}
3757
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003758static void apply_wqattrs_lock(void)
3759{
3760 /* CPUs should stay stable across pwq creations and installations */
3761 get_online_cpus();
3762 mutex_lock(&wq_pool_mutex);
3763}
3764
3765static void apply_wqattrs_unlock(void)
3766{
3767 mutex_unlock(&wq_pool_mutex);
3768 put_online_cpus();
3769}
3770
3771static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3772 const struct workqueue_attrs *attrs)
3773{
3774 struct apply_wqattrs_ctx *ctx;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003775
3776 /* only unbound workqueues can change attributes */
3777 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3778 return -EINVAL;
3779
3780 /* creating multiple pwqs breaks ordering guarantee */
Tejun Heo0a94efb2017-07-23 08:36:15 -04003781 if (!list_empty(&wq->pwqs)) {
3782 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3783 return -EINVAL;
3784
3785 wq->flags &= ~__WQ_ORDERED;
3786 }
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003787
3788 ctx = apply_wqattrs_prepare(wq, attrs);
wanghaibin62011712016-01-07 20:38:59 +08003789 if (!ctx)
3790 return -ENOMEM;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003791
3792 /* the ctx has been prepared successfully, let's commit it */
wanghaibin62011712016-01-07 20:38:59 +08003793 apply_wqattrs_commit(ctx);
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003794 apply_wqattrs_cleanup(ctx);
3795
wanghaibin62011712016-01-07 20:38:59 +08003796 return 0;
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003797}
3798
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003799/**
3800 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3801 * @wq: the target workqueue
3802 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3803 *
Tejun Heo4c16bd32013-04-01 11:23:36 -07003804 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3805 * machines, this function maps a separate pwq to each NUMA node with
3806 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3807 * NUMA node it was issued on. Older pwqs are released as in-flight work
3808 * items finish. Note that a work item which repeatedly requeues itself
3809 * back-to-back will stay on its current pwq.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003810 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003811 * Performs GFP_KERNEL allocations.
3812 *
3813 * Return: 0 on success and -errno on failure.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003814 */
3815int apply_workqueue_attrs(struct workqueue_struct *wq,
3816 const struct workqueue_attrs *attrs)
3817{
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003818 int ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003819
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003820 apply_wqattrs_lock();
3821 ret = apply_workqueue_attrs_locked(wq, attrs);
3822 apply_wqattrs_unlock();
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003823
Tejun Heo48621252013-04-01 11:23:31 -07003824 return ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003825}
3826
Tejun Heo4c16bd32013-04-01 11:23:36 -07003827/**
3828 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3829 * @wq: the target workqueue
3830 * @cpu: the CPU coming up or going down
3831 * @online: whether @cpu is coming up or going down
3832 *
3833 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3834 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3835 * @wq accordingly.
3836 *
3837 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3838 * falls back to @wq->dfl_pwq which may not be optimal but is always
3839 * correct.
3840 *
3841 * Note that when the last allowed CPU of a NUMA node goes offline for a
3842 * workqueue with a cpumask spanning multiple nodes, the workers which were
3843 * already executing the work items for the workqueue will lose their CPU
3844 * affinity and may execute on any CPU. This is similar to how per-cpu
3845 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3846 * affinity, it's the user's responsibility to flush the work item from
3847 * CPU_DOWN_PREPARE.
3848 */
3849static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3850 bool online)
3851{
3852 int node = cpu_to_node(cpu);
3853 int cpu_off = online ? -1 : cpu;
3854 struct pool_workqueue *old_pwq = NULL, *pwq;
3855 struct workqueue_attrs *target_attrs;
3856 cpumask_t *cpumask;
3857
3858 lockdep_assert_held(&wq_pool_mutex);
3859
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003860 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3861 wq->unbound_attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003862 return;
3863
3864 /*
3865 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3866 * Let's use a preallocated one. The following buf is protected by
3867 * CPU hotplug exclusion.
3868 */
3869 target_attrs = wq_update_unbound_numa_attrs_buf;
3870 cpumask = target_attrs->cpumask;
3871
Tejun Heo4c16bd32013-04-01 11:23:36 -07003872 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3873 pwq = unbound_pwq_by_node(wq, node);
3874
3875 /*
3876 * Let's determine what needs to be done. If the target cpumask is
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003877 * different from the default pwq's, we need to compare it to @pwq's
3878 * and create a new one if they don't match. If the target cpumask
3879 * equals the default pwq's, the default pwq should be used.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003880 */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003881 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
Tejun Heo4c16bd32013-04-01 11:23:36 -07003882 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003883 return;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003884 } else {
Daeseok Youn534a3fb2014-04-18 09:08:14 +09003885 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003886 }
3887
Tejun Heo4c16bd32013-04-01 11:23:36 -07003888 /* create a new pwq */
3889 pwq = alloc_unbound_pwq(wq, target_attrs);
3890 if (!pwq) {
Fabian Frederick2d916032014-05-12 13:59:35 -04003891 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3892 wq->name);
Daeseok Youn77f300b2014-04-16 14:32:29 +09003893 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003894 }
3895
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003896 /* Install the new pwq. */
Tejun Heo4c16bd32013-04-01 11:23:36 -07003897 mutex_lock(&wq->mutex);
3898 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3899 goto out_unlock;
3900
3901use_dfl_pwq:
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003902 mutex_lock(&wq->mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07003903 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3904 get_pwq(wq->dfl_pwq);
3905 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3906 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3907out_unlock:
3908 mutex_unlock(&wq->mutex);
3909 put_pwq_unlocked(old_pwq);
3910}
3911
Tejun Heo30cdf242013-03-12 11:29:57 -07003912static int alloc_and_link_pwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913{
Tejun Heo49e3cf42013-03-12 11:29:58 -07003914 bool highpri = wq->flags & WQ_HIGHPRI;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003915 int cpu, ret;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01003916
Tejun Heo30cdf242013-03-12 11:29:57 -07003917 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo420c0dd2013-03-12 11:29:59 -07003918 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3919 if (!wq->cpu_pwqs)
Tejun Heo30cdf242013-03-12 11:29:57 -07003920 return -ENOMEM;
3921
3922 for_each_possible_cpu(cpu) {
Tejun Heo7fb98ea2013-03-12 11:30:00 -07003923 struct pool_workqueue *pwq =
3924 per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heo7a62c2c2013-03-12 11:30:03 -07003925 struct worker_pool *cpu_pools =
Tejun Heof02ae732013-03-12 11:30:03 -07003926 per_cpu(cpu_worker_pools, cpu);
Tejun Heo30cdf242013-03-12 11:29:57 -07003927
Tejun Heof147f292013-04-01 11:23:35 -07003928 init_pwq(pwq, wq, &cpu_pools[highpri]);
3929
3930 mutex_lock(&wq->mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003931 link_pwq(pwq);
Tejun Heof147f292013-04-01 11:23:35 -07003932 mutex_unlock(&wq->mutex);
Tejun Heo30cdf242013-03-12 11:29:57 -07003933 }
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003934 return 0;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003935 } else if (wq->flags & __WQ_ORDERED) {
3936 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3937 /* there should only be single pwq for ordering guarantee */
3938 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3939 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3940 "ordering guarantee broken for workqueue %s\n", wq->name);
3941 return ret;
Tejun Heo30cdf242013-03-12 11:29:57 -07003942 } else {
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003943 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
Tejun Heo30cdf242013-03-12 11:29:57 -07003944 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07003945}
3946
Tejun Heof3421792010-07-02 10:03:51 +02003947static int wq_clamp_max_active(int max_active, unsigned int flags,
3948 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02003949{
Tejun Heof3421792010-07-02 10:03:51 +02003950 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3951
3952 if (max_active < 1 || max_active > lim)
Valentin Ilie044c7822012-08-19 00:52:42 +03003953 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3954 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003955
Tejun Heof3421792010-07-02 10:03:51 +02003956 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003957}
3958
Tejun Heob196be82012-01-10 15:11:35 -08003959struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
Tejun Heod320c032010-06-29 10:07:14 +02003960 unsigned int flags,
3961 int max_active,
3962 struct lock_class_key *key,
Tejun Heob196be82012-01-10 15:11:35 -08003963 const char *lock_name, ...)
Oleg Nesterov3af244332007-05-09 02:34:09 -07003964{
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003965 size_t tbl_size = 0;
Tejun Heoecf68812013-04-01 11:23:34 -07003966 va_list args;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003967 struct workqueue_struct *wq;
Tejun Heo49e3cf42013-03-12 11:29:58 -07003968 struct pool_workqueue *pwq;
Tejun Heob196be82012-01-10 15:11:35 -08003969
Tejun Heo5c0338c2017-07-18 18:41:52 -04003970 /*
3971 * Unbound && max_active == 1 used to imply ordered, which is no
3972 * longer the case on NUMA machines due to per-node pools. While
3973 * alloc_ordered_workqueue() is the right way to create an ordered
3974 * workqueue, keep the previous behavior to avoid subtle breakages
3975 * on NUMA.
3976 */
3977 if ((flags & WQ_UNBOUND) && max_active == 1)
3978 flags |= __WQ_ORDERED;
3979
Viresh Kumarcee22a12013-04-08 16:45:40 +05303980 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3981 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3982 flags |= WQ_UNBOUND;
3983
Tejun Heoecf68812013-04-01 11:23:34 -07003984 /* allocate wq and format name */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003985 if (flags & WQ_UNBOUND)
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08003986 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003987
3988 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
Tejun Heob196be82012-01-10 15:11:35 -08003989 if (!wq)
Tejun Heod2c1d402013-03-12 11:30:04 -07003990 return NULL;
Tejun Heob196be82012-01-10 15:11:35 -08003991
Tejun Heo6029a912013-04-01 11:23:34 -07003992 if (flags & WQ_UNBOUND) {
3993 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3994 if (!wq->unbound_attrs)
3995 goto err_free_wq;
3996 }
3997
Tejun Heoecf68812013-04-01 11:23:34 -07003998 va_start(args, lock_name);
3999 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
Tejun Heob196be82012-01-10 15:11:35 -08004000 va_end(args);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004001
Tejun Heod320c032010-06-29 10:07:14 +02004002 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heob196be82012-01-10 15:11:35 -08004003 max_active = wq_clamp_max_active(max_active, flags, wq->name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004004
Tejun Heob196be82012-01-10 15:11:35 -08004005 /* init wq */
Tejun Heo97e37d72010-06-29 10:07:10 +02004006 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004007 wq->saved_max_active = max_active;
Lai Jiangshan3c25a552013-03-25 16:57:17 -07004008 mutex_init(&wq->mutex);
Tejun Heo112202d2013-02-13 19:29:12 -08004009 atomic_set(&wq->nr_pwqs_to_flush, 0);
Tejun Heo30cdf242013-03-12 11:29:57 -07004010 INIT_LIST_HEAD(&wq->pwqs);
Tejun Heo73f53c42010-06-29 10:07:11 +02004011 INIT_LIST_HEAD(&wq->flusher_queue);
4012 INIT_LIST_HEAD(&wq->flusher_overflow);
Tejun Heo493a1722013-03-12 11:29:59 -07004013 INIT_LIST_HEAD(&wq->maydays);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004014
Johannes Bergeb13ba82008-01-16 09:51:58 +01004015 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07004016 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004017
Tejun Heo30cdf242013-03-12 11:29:57 -07004018 if (alloc_and_link_pwqs(wq) < 0)
Tejun Heod2c1d402013-03-12 11:30:04 -07004019 goto err_free_wq;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004020
Tejun Heo493008a2013-03-12 11:30:03 -07004021 /*
4022 * Workqueues which may be used during memory reclaim should
4023 * have a rescuer to guarantee forward progress.
4024 */
4025 if (flags & WQ_MEM_RECLAIM) {
Tejun Heoe22bee72010-06-29 10:07:14 +02004026 struct worker *rescuer;
4027
Lai Jiangshanf7537df2014-07-15 17:24:15 +08004028 rescuer = alloc_worker(NUMA_NO_NODE);
Tejun Heoe22bee72010-06-29 10:07:14 +02004029 if (!rescuer)
Tejun Heod2c1d402013-03-12 11:30:04 -07004030 goto err_destroy;
Tejun Heoe22bee72010-06-29 10:07:14 +02004031
Tejun Heo111c2252013-01-17 17:16:24 -08004032 rescuer->rescue_wq = wq;
4033 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
Tejun Heob196be82012-01-10 15:11:35 -08004034 wq->name);
Tejun Heod2c1d402013-03-12 11:30:04 -07004035 if (IS_ERR(rescuer->task)) {
4036 kfree(rescuer);
4037 goto err_destroy;
4038 }
Tejun Heoe22bee72010-06-29 10:07:14 +02004039
Tejun Heod2c1d402013-03-12 11:30:04 -07004040 wq->rescuer = rescuer;
Peter Zijlstra25834c72015-05-15 17:43:34 +02004041 kthread_bind_mask(rescuer->task, cpu_possible_mask);
Tejun Heoe22bee72010-06-29 10:07:14 +02004042 wake_up_process(rescuer->task);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004043 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004044
Tejun Heo226223a2013-03-12 11:30:05 -07004045 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4046 goto err_destroy;
4047
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004048 /*
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004049 * wq_pool_mutex protects global freeze state and workqueues list.
4050 * Grab it, adjust max_active and add the new @wq to workqueues
4051 * list.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004052 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004053 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004054
Lai Jiangshana357fc02013-03-25 16:57:19 -07004055 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004056 for_each_pwq(pwq, wq)
4057 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004058 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004059
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004060 list_add_tail_rcu(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004061
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004062 mutex_unlock(&wq_pool_mutex);
Tejun Heo15376632010-06-29 10:07:11 +02004063
Oleg Nesterov3af244332007-05-09 02:34:09 -07004064 return wq;
Tejun Heod2c1d402013-03-12 11:30:04 -07004065
4066err_free_wq:
Tejun Heo6029a912013-04-01 11:23:34 -07004067 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heod2c1d402013-03-12 11:30:04 -07004068 kfree(wq);
4069 return NULL;
4070err_destroy:
4071 destroy_workqueue(wq);
Tejun Heo4690c4a2010-06-29 10:07:10 +02004072 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004073}
Tejun Heod320c032010-06-29 10:07:14 +02004074EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004075
4076/**
4077 * destroy_workqueue - safely terminate a workqueue
4078 * @wq: target workqueue
4079 *
4080 * Safely destroy a workqueue. All work currently pending will be done first.
4081 */
4082void destroy_workqueue(struct workqueue_struct *wq)
4083{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004084 struct pool_workqueue *pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004085 int node;
Oleg Nesterov3af244332007-05-09 02:34:09 -07004086
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02004087 /* drain it before proceeding with destruction */
4088 drain_workqueue(wq);
Tejun Heoc8efcc22010-12-20 19:32:04 +01004089
Tejun Heo6183c002013-03-12 11:29:57 -07004090 /* sanity checks */
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004091 mutex_lock(&wq->mutex);
Tejun Heo49e3cf42013-03-12 11:29:58 -07004092 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004093 int i;
4094
Tejun Heo76af4d92013-03-12 11:30:00 -07004095 for (i = 0; i < WORK_NR_COLORS; i++) {
4096 if (WARN_ON(pwq->nr_in_flight[i])) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004097 mutex_unlock(&wq->mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004098 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004099 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004100 }
4101 }
4102
Lai Jiangshan5c529592013-04-04 10:05:38 +08004103 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
Tejun Heo8864b4e2013-03-12 11:30:04 -07004104 WARN_ON(pwq->nr_active) ||
Tejun Heo76af4d92013-03-12 11:30:00 -07004105 WARN_ON(!list_empty(&pwq->delayed_works))) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004106 mutex_unlock(&wq->mutex);
Tejun Heofa07fb62016-09-05 08:54:06 -04004107 show_workqueue_state();
Tejun Heo6183c002013-03-12 11:29:57 -07004108 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07004109 }
Tejun Heo6183c002013-03-12 11:29:57 -07004110 }
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07004111 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07004112
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004113 /*
4114 * wq list is used to freeze wq, remove from list after
4115 * flushing is complete in case freeze races us.
4116 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004117 mutex_lock(&wq_pool_mutex);
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004118 list_del_rcu(&wq->list);
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004119 mutex_unlock(&wq_pool_mutex);
Oleg Nesterov3af244332007-05-09 02:34:09 -07004120
Tejun Heo226223a2013-03-12 11:30:05 -07004121 workqueue_sysfs_unregister(wq);
4122
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004123 if (wq->rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02004124 kthread_stop(wq->rescuer->task);
Tejun Heoe22bee72010-06-29 10:07:14 +02004125
Tejun Heo8864b4e2013-03-12 11:30:04 -07004126 if (!(wq->flags & WQ_UNBOUND)) {
4127 /*
4128 * The base ref is never dropped on per-cpu pwqs. Directly
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004129 * schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004130 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04004131 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07004132 } else {
4133 /*
4134 * We're the sole accessor of @wq at this point. Directly
Tejun Heo4c16bd32013-04-01 11:23:36 -07004135 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4136 * @wq will be freed when the last pwq is released.
Tejun Heo8864b4e2013-03-12 11:30:04 -07004137 */
Tejun Heo4c16bd32013-04-01 11:23:36 -07004138 for_each_node(node) {
4139 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4140 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4141 put_pwq_unlocked(pwq);
4142 }
4143
4144 /*
4145 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4146 * put. Don't access it afterwards.
4147 */
4148 pwq = wq->dfl_pwq;
4149 wq->dfl_pwq = NULL;
Tejun Heodce90d42013-04-01 11:23:35 -07004150 put_pwq_unlocked(pwq);
Tejun Heo29c91e92013-03-12 11:30:03 -07004151 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004152}
4153EXPORT_SYMBOL_GPL(destroy_workqueue);
4154
Tejun Heodcd989c2010-06-29 10:07:14 +02004155/**
4156 * workqueue_set_max_active - adjust max_active of a workqueue
4157 * @wq: target workqueue
4158 * @max_active: new max_active value.
4159 *
4160 * Set max_active of @wq to @max_active.
4161 *
4162 * CONTEXT:
4163 * Don't call from IRQ context.
4164 */
4165void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4166{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004167 struct pool_workqueue *pwq;
Tejun Heodcd989c2010-06-29 10:07:14 +02004168
Tejun Heo8719dce2013-03-12 11:30:04 -07004169 /* disallow meddling with max_active for ordered workqueues */
Tejun Heo0a94efb2017-07-23 08:36:15 -04004170 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Tejun Heo8719dce2013-03-12 11:30:04 -07004171 return;
4172
Tejun Heof3421792010-07-02 10:03:51 +02004173 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02004174
Lai Jiangshana357fc02013-03-25 16:57:19 -07004175 mutex_lock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004176
Tejun Heo0a94efb2017-07-23 08:36:15 -04004177 wq->flags &= ~__WQ_ORDERED;
Tejun Heodcd989c2010-06-29 10:07:14 +02004178 wq->saved_max_active = max_active;
4179
Tejun Heo699ce092013-03-13 16:51:35 -07004180 for_each_pwq(pwq, wq)
4181 pwq_adjust_max_active(pwq);
Tejun Heodcd989c2010-06-29 10:07:14 +02004182
Lai Jiangshana357fc02013-03-25 16:57:19 -07004183 mutex_unlock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004184}
4185EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4186
4187/**
Lukas Wunner363e3fd2018-02-11 10:38:28 +01004188 * current_work - retrieve %current task's work struct
4189 *
4190 * Determine if %current task is a workqueue worker and what it's working on.
4191 * Useful to find out the context that the %current task is running in.
4192 *
4193 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4194 */
4195struct work_struct *current_work(void)
4196{
4197 struct worker *worker = current_wq_worker();
4198
4199 return worker ? worker->current_work : NULL;
4200}
4201EXPORT_SYMBOL(current_work);
4202
4203/**
Tejun Heoe62676162013-03-12 17:41:37 -07004204 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4205 *
4206 * Determine whether %current is a workqueue rescuer. Can be used from
4207 * work functions to determine whether it's being run off the rescuer task.
Yacine Belkadid185af32013-07-31 14:59:24 -07004208 *
4209 * Return: %true if %current is a workqueue rescuer. %false otherwise.
Tejun Heoe62676162013-03-12 17:41:37 -07004210 */
4211bool current_is_workqueue_rescuer(void)
4212{
4213 struct worker *worker = current_wq_worker();
4214
Lai Jiangshan6a092df2013-03-20 03:28:03 +08004215 return worker && worker->rescue_wq;
Tejun Heoe62676162013-03-12 17:41:37 -07004216}
4217
4218/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004219 * workqueue_congested - test whether a workqueue is congested
4220 * @cpu: CPU in question
4221 * @wq: target workqueue
4222 *
4223 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4224 * no synchronization around this function and the test result is
4225 * unreliable and only useful as advisory hints or for debugging.
4226 *
Tejun Heod3251852013-05-10 11:10:17 -07004227 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4228 * Note that both per-cpu and unbound workqueues may be associated with
4229 * multiple pool_workqueues which have separate congested states. A
4230 * workqueue being congested on one CPU doesn't mean the workqueue is also
4231 * contested on other CPUs / NUMA nodes.
4232 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004233 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004234 * %true if congested, %false otherwise.
4235 */
Tejun Heod84ff052013-03-12 11:29:59 -07004236bool workqueue_congested(int cpu, struct workqueue_struct *wq)
Tejun Heodcd989c2010-06-29 10:07:14 +02004237{
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004238 struct pool_workqueue *pwq;
Tejun Heo76af4d92013-03-12 11:30:00 -07004239 bool ret;
4240
Lai Jiangshan88109452013-03-20 03:28:10 +08004241 rcu_read_lock_sched();
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004242
Tejun Heod3251852013-05-10 11:10:17 -07004243 if (cpu == WORK_CPU_UNBOUND)
4244 cpu = smp_processor_id();
4245
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004246 if (!(wq->flags & WQ_UNBOUND))
4247 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4248 else
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004249 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodcd989c2010-06-29 10:07:14 +02004250
Tejun Heo76af4d92013-03-12 11:30:00 -07004251 ret = !list_empty(&pwq->delayed_works);
Lai Jiangshan88109452013-03-20 03:28:10 +08004252 rcu_read_unlock_sched();
Tejun Heo76af4d92013-03-12 11:30:00 -07004253
4254 return ret;
Tejun Heodcd989c2010-06-29 10:07:14 +02004255}
4256EXPORT_SYMBOL_GPL(workqueue_congested);
4257
4258/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004259 * work_busy - test whether a work is currently pending or running
4260 * @work: the work to be tested
4261 *
4262 * Test whether @work is currently pending or running. There is no
4263 * synchronization around this function and the test result is
4264 * unreliable and only useful as advisory hints or for debugging.
Tejun Heodcd989c2010-06-29 10:07:14 +02004265 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004266 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004267 * OR'd bitmask of WORK_BUSY_* bits.
4268 */
4269unsigned int work_busy(struct work_struct *work)
4270{
Tejun Heofa1b54e2013-03-12 11:30:00 -07004271 struct worker_pool *pool;
Tejun Heodcd989c2010-06-29 10:07:14 +02004272 unsigned long flags;
4273 unsigned int ret = 0;
4274
Tejun Heodcd989c2010-06-29 10:07:14 +02004275 if (work_pending(work))
4276 ret |= WORK_BUSY_PENDING;
Tejun Heodcd989c2010-06-29 10:07:14 +02004277
Tejun Heofa1b54e2013-03-12 11:30:00 -07004278 local_irq_save(flags);
4279 pool = get_work_pool(work);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004280 if (pool) {
Tejun Heofa1b54e2013-03-12 11:30:00 -07004281 spin_lock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004282 if (find_worker_executing_work(pool, work))
4283 ret |= WORK_BUSY_RUNNING;
Tejun Heofa1b54e2013-03-12 11:30:00 -07004284 spin_unlock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004285 }
Tejun Heofa1b54e2013-03-12 11:30:00 -07004286 local_irq_restore(flags);
Tejun Heodcd989c2010-06-29 10:07:14 +02004287
4288 return ret;
4289}
4290EXPORT_SYMBOL_GPL(work_busy);
4291
Tejun Heo3d1cb202013-04-30 15:27:22 -07004292/**
4293 * set_worker_desc - set description for the current work item
4294 * @fmt: printf-style format string
4295 * @...: arguments for the format string
4296 *
4297 * This function can be called by a running work function to describe what
4298 * the work item is about. If the worker task gets dumped, this
4299 * information will be printed out together to help debugging. The
4300 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4301 */
4302void set_worker_desc(const char *fmt, ...)
4303{
4304 struct worker *worker = current_wq_worker();
4305 va_list args;
4306
4307 if (worker) {
4308 va_start(args, fmt);
4309 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4310 va_end(args);
4311 worker->desc_valid = true;
4312 }
4313}
4314
4315/**
4316 * print_worker_info - print out worker information and description
4317 * @log_lvl: the log level to use when printing
4318 * @task: target task
4319 *
4320 * If @task is a worker and currently executing a work item, print out the
4321 * name of the workqueue being serviced and worker description set with
4322 * set_worker_desc() by the currently executing work item.
4323 *
4324 * This function can be safely called on any task as long as the
4325 * task_struct itself is accessible. While safe, this function isn't
4326 * synchronized and may print out mixups or garbages of limited length.
4327 */
4328void print_worker_info(const char *log_lvl, struct task_struct *task)
4329{
4330 work_func_t *fn = NULL;
4331 char name[WQ_NAME_LEN] = { };
4332 char desc[WORKER_DESC_LEN] = { };
4333 struct pool_workqueue *pwq = NULL;
4334 struct workqueue_struct *wq = NULL;
4335 bool desc_valid = false;
4336 struct worker *worker;
4337
4338 if (!(task->flags & PF_WQ_WORKER))
4339 return;
4340
4341 /*
4342 * This function is called without any synchronization and @task
4343 * could be in any state. Be careful with dereferences.
4344 */
Petr Mladeke7005912016-10-11 13:55:17 -07004345 worker = kthread_probe_data(task);
Tejun Heo3d1cb202013-04-30 15:27:22 -07004346
4347 /*
4348 * Carefully copy the associated workqueue's workfn and name. Keep
4349 * the original last '\0' in case the original contains garbage.
4350 */
4351 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4352 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4353 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4354 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4355
4356 /* copy worker description */
4357 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4358 if (desc_valid)
4359 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4360
4361 if (fn || name[0] || desc[0]) {
4362 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4363 if (desc[0])
4364 pr_cont(" (%s)", desc);
4365 pr_cont("\n");
4366 }
4367}
4368
Tejun Heo3494fc32015-03-09 09:22:28 -04004369static void pr_cont_pool_info(struct worker_pool *pool)
4370{
4371 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4372 if (pool->node != NUMA_NO_NODE)
4373 pr_cont(" node=%d", pool->node);
4374 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4375}
4376
4377static void pr_cont_work(bool comma, struct work_struct *work)
4378{
4379 if (work->func == wq_barrier_func) {
4380 struct wq_barrier *barr;
4381
4382 barr = container_of(work, struct wq_barrier, work);
4383
4384 pr_cont("%s BAR(%d)", comma ? "," : "",
4385 task_pid_nr(barr->task));
4386 } else {
4387 pr_cont("%s %pf", comma ? "," : "", work->func);
4388 }
4389}
4390
4391static void show_pwq(struct pool_workqueue *pwq)
4392{
4393 struct worker_pool *pool = pwq->pool;
4394 struct work_struct *work;
4395 struct worker *worker;
4396 bool has_in_flight = false, has_pending = false;
4397 int bkt;
4398
4399 pr_info(" pwq %d:", pool->id);
4400 pr_cont_pool_info(pool);
4401
4402 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4403 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4404
4405 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4406 if (worker->current_pwq == pwq) {
4407 has_in_flight = true;
4408 break;
4409 }
4410 }
4411 if (has_in_flight) {
4412 bool comma = false;
4413
4414 pr_info(" in-flight:");
4415 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4416 if (worker->current_pwq != pwq)
4417 continue;
4418
4419 pr_cont("%s %d%s:%pf", comma ? "," : "",
4420 task_pid_nr(worker->task),
4421 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4422 worker->current_func);
4423 list_for_each_entry(work, &worker->scheduled, entry)
4424 pr_cont_work(false, work);
4425 comma = true;
4426 }
4427 pr_cont("\n");
4428 }
4429
4430 list_for_each_entry(work, &pool->worklist, entry) {
4431 if (get_work_pwq(work) == pwq) {
4432 has_pending = true;
4433 break;
4434 }
4435 }
4436 if (has_pending) {
4437 bool comma = false;
4438
4439 pr_info(" pending:");
4440 list_for_each_entry(work, &pool->worklist, entry) {
4441 if (get_work_pwq(work) != pwq)
4442 continue;
4443
4444 pr_cont_work(comma, work);
4445 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4446 }
4447 pr_cont("\n");
4448 }
4449
4450 if (!list_empty(&pwq->delayed_works)) {
4451 bool comma = false;
4452
4453 pr_info(" delayed:");
4454 list_for_each_entry(work, &pwq->delayed_works, entry) {
4455 pr_cont_work(comma, work);
4456 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4457 }
4458 pr_cont("\n");
4459 }
4460}
4461
4462/**
4463 * show_workqueue_state - dump workqueue state
4464 *
Roger Lu7b776af2016-07-01 11:05:02 +08004465 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4466 * all busy workqueues and pools.
Tejun Heo3494fc32015-03-09 09:22:28 -04004467 */
4468void show_workqueue_state(void)
4469{
4470 struct workqueue_struct *wq;
4471 struct worker_pool *pool;
4472 unsigned long flags;
4473 int pi;
4474
4475 rcu_read_lock_sched();
4476
4477 pr_info("Showing busy workqueues and worker pools:\n");
4478
4479 list_for_each_entry_rcu(wq, &workqueues, list) {
4480 struct pool_workqueue *pwq;
4481 bool idle = true;
4482
4483 for_each_pwq(pwq, wq) {
4484 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4485 idle = false;
4486 break;
4487 }
4488 }
4489 if (idle)
4490 continue;
4491
4492 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4493
4494 for_each_pwq(pwq, wq) {
4495 spin_lock_irqsave(&pwq->pool->lock, flags);
4496 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4497 show_pwq(pwq);
4498 spin_unlock_irqrestore(&pwq->pool->lock, flags);
Sergey Senozhatskya6d59302018-01-11 09:53:35 +09004499 /*
4500 * We could be printing a lot from atomic context, e.g.
4501 * sysrq-t -> show_workqueue_state(). Avoid triggering
4502 * hard lockup.
4503 */
4504 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004505 }
4506 }
4507
4508 for_each_pool(pool, pi) {
4509 struct worker *worker;
4510 bool first = true;
4511
4512 spin_lock_irqsave(&pool->lock, flags);
4513 if (pool->nr_workers == pool->nr_idle)
4514 goto next_pool;
4515
4516 pr_info("pool %d:", pool->id);
4517 pr_cont_pool_info(pool);
Tejun Heo82607adc2015-12-08 11:28:04 -05004518 pr_cont(" hung=%us workers=%d",
4519 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4520 pool->nr_workers);
Tejun Heo3494fc32015-03-09 09:22:28 -04004521 if (pool->manager)
4522 pr_cont(" manager: %d",
4523 task_pid_nr(pool->manager->task));
4524 list_for_each_entry(worker, &pool->idle_list, entry) {
4525 pr_cont(" %s%d", first ? "idle: " : "",
4526 task_pid_nr(worker->task));
4527 first = false;
4528 }
4529 pr_cont("\n");
4530 next_pool:
4531 spin_unlock_irqrestore(&pool->lock, flags);
Sergey Senozhatskya6d59302018-01-11 09:53:35 +09004532 /*
4533 * We could be printing a lot from atomic context, e.g.
4534 * sysrq-t -> show_workqueue_state(). Avoid triggering
4535 * hard lockup.
4536 */
4537 touch_nmi_watchdog();
Tejun Heo3494fc32015-03-09 09:22:28 -04004538 }
4539
4540 rcu_read_unlock_sched();
4541}
4542
Tejun Heodb7bccf2010-06-29 10:07:12 +02004543/*
4544 * CPU hotplug.
4545 *
Tejun Heoe22bee72010-06-29 10:07:14 +02004546 * There are two challenges in supporting CPU hotplug. Firstly, there
Tejun Heo112202d2013-02-13 19:29:12 -08004547 * are a lot of assumptions on strong associations among work, pwq and
Tejun Heo706026c2013-01-24 11:01:34 -08004548 * pool which make migrating pending and scheduled works very
Tejun Heoe22bee72010-06-29 10:07:14 +02004549 * difficult to implement without impacting hot paths. Secondly,
Tejun Heo94cf58b2013-01-24 11:01:33 -08004550 * worker pools serve mix of short, long and very long running works making
Tejun Heoe22bee72010-06-29 10:07:14 +02004551 * blocked draining impractical.
4552 *
Tejun Heo24647572013-01-24 11:01:33 -08004553 * This is solved by allowing the pools to be disassociated from the CPU
Tejun Heo628c78e2012-07-17 12:39:27 -07004554 * running as an unbound one and allowing it to be reattached later if the
4555 * cpu comes back online.
Tejun Heodb7bccf2010-06-29 10:07:12 +02004556 */
4557
Tejun Heo706026c2013-01-24 11:01:34 -08004558static void wq_unbind_fn(struct work_struct *work)
Tejun Heodb7bccf2010-06-29 10:07:12 +02004559{
Tejun Heo38db41d2013-01-24 11:01:34 -08004560 int cpu = smp_processor_id();
Tejun Heo4ce62e92012-07-13 22:16:44 -07004561 struct worker_pool *pool;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004562 struct worker *worker;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004563
Tejun Heof02ae732013-03-12 11:30:03 -07004564 for_each_cpu_worker_pool(pool, cpu) {
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004565 mutex_lock(&pool->attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004566 spin_lock_irq(&pool->lock);
4567
4568 /*
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004569 * We've blocked all attach/detach operations. Make all workers
Tejun Heo94cf58b2013-01-24 11:01:33 -08004570 * unbound and set DISASSOCIATED. Before this, all workers
4571 * except for the ones which are still executing works from
4572 * before the last CPU down must be on the cpu. After
4573 * this, they may become diasporas.
4574 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004575 for_each_pool_worker(worker, pool)
Tejun Heoc9e7cf22013-01-24 11:01:33 -08004576 worker->flags |= WORKER_UNBOUND;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004577
Tejun Heo24647572013-01-24 11:01:33 -08004578 pool->flags |= POOL_DISASSOCIATED;
Tejun Heof2d5a0e2012-07-17 12:39:26 -07004579
Tejun Heo94cf58b2013-01-24 11:01:33 -08004580 spin_unlock_irq(&pool->lock);
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004581 mutex_unlock(&pool->attach_mutex);
Tejun Heoe22bee72010-06-29 10:07:14 +02004582
Lai Jiangshaneb283422013-03-08 15:18:28 -08004583 /*
4584 * Call schedule() so that we cross rq->lock and thus can
4585 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4586 * This is necessary as scheduler callbacks may be invoked
4587 * from other cpus.
4588 */
4589 schedule();
Tejun Heo628c78e2012-07-17 12:39:27 -07004590
Lai Jiangshaneb283422013-03-08 15:18:28 -08004591 /*
4592 * Sched callbacks are disabled now. Zap nr_running.
4593 * After this, nr_running stays zero and need_more_worker()
4594 * and keep_working() are always true as long as the
4595 * worklist is not empty. This pool now behaves as an
4596 * unbound (in terms of concurrency management) pool which
4597 * are served by workers tied to the pool.
4598 */
Tejun Heoe19e3972013-01-24 11:39:44 -08004599 atomic_set(&pool->nr_running, 0);
Lai Jiangshaneb283422013-03-08 15:18:28 -08004600
4601 /*
4602 * With concurrency management just turned off, a busy
4603 * worker blocking could lead to lengthy stalls. Kick off
4604 * unbound chain execution of currently pending work items.
4605 */
4606 spin_lock_irq(&pool->lock);
4607 wake_up_worker(pool);
4608 spin_unlock_irq(&pool->lock);
4609 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004610}
4611
Tejun Heobd7c0892013-03-19 13:45:21 -07004612/**
4613 * rebind_workers - rebind all workers of a pool to the associated CPU
4614 * @pool: pool of interest
4615 *
Tejun Heoa9ab7752013-03-19 13:45:21 -07004616 * @pool->cpu is coming online. Rebind all workers to the CPU.
Tejun Heobd7c0892013-03-19 13:45:21 -07004617 */
4618static void rebind_workers(struct worker_pool *pool)
4619{
Tejun Heoa9ab7752013-03-19 13:45:21 -07004620 struct worker *worker;
Tejun Heobd7c0892013-03-19 13:45:21 -07004621
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004622 lockdep_assert_held(&pool->attach_mutex);
Tejun Heobd7c0892013-03-19 13:45:21 -07004623
Tejun Heoa9ab7752013-03-19 13:45:21 -07004624 /*
4625 * Restore CPU affinity of all workers. As all idle workers should
4626 * be on the run-queue of the associated CPU before any local
Shailendra Verma402dd892015-05-23 10:38:14 +05304627 * wake-ups for concurrency management happen, restore CPU affinity
Tejun Heoa9ab7752013-03-19 13:45:21 -07004628 * of all workers first and then clear UNBOUND. As we're called
4629 * from CPU_ONLINE, the following shouldn't fail.
4630 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004631 for_each_pool_worker(worker, pool)
Tejun Heoa9ab7752013-03-19 13:45:21 -07004632 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4633 pool->attrs->cpumask) < 0);
4634
4635 spin_lock_irq(&pool->lock);
Wanpeng Lif7c17d22016-05-11 17:55:18 +08004636
4637 /*
4638 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4639 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4640 * being reworked and this can go away in time.
4641 */
4642 if (!(pool->flags & POOL_DISASSOCIATED)) {
4643 spin_unlock_irq(&pool->lock);
4644 return;
4645 }
4646
Lai Jiangshan3de5e882014-06-03 15:33:27 +08004647 pool->flags &= ~POOL_DISASSOCIATED;
Tejun Heoa9ab7752013-03-19 13:45:21 -07004648
Lai Jiangshanda028462014-05-20 17:46:31 +08004649 for_each_pool_worker(worker, pool) {
Tejun Heoa9ab7752013-03-19 13:45:21 -07004650 unsigned int worker_flags = worker->flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004651
4652 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07004653 * A bound idle worker should actually be on the runqueue
4654 * of the associated CPU for local wake-ups targeting it to
4655 * work. Kick all idle workers so that they migrate to the
4656 * associated CPU. Doing this in the same loop as
4657 * replacing UNBOUND with REBOUND is safe as no worker will
4658 * be bound before @pool->lock is released.
Tejun Heobd7c0892013-03-19 13:45:21 -07004659 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004660 if (worker_flags & WORKER_IDLE)
4661 wake_up_process(worker->task);
4662
4663 /*
4664 * We want to clear UNBOUND but can't directly call
4665 * worker_clr_flags() or adjust nr_running. Atomically
4666 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4667 * @worker will clear REBOUND using worker_clr_flags() when
4668 * it initiates the next execution cycle thus restoring
4669 * concurrency management. Note that when or whether
4670 * @worker clears REBOUND doesn't affect correctness.
4671 *
4672 * ACCESS_ONCE() is necessary because @worker->flags may be
4673 * tested without holding any lock in
4674 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4675 * fail incorrectly leading to premature concurrency
4676 * management operations.
4677 */
4678 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4679 worker_flags |= WORKER_REBOUND;
4680 worker_flags &= ~WORKER_UNBOUND;
4681 ACCESS_ONCE(worker->flags) = worker_flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004682 }
4683
Tejun Heoa9ab7752013-03-19 13:45:21 -07004684 spin_unlock_irq(&pool->lock);
Tejun Heobd7c0892013-03-19 13:45:21 -07004685}
4686
Tejun Heo7dbc7252013-03-19 13:45:21 -07004687/**
4688 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4689 * @pool: unbound pool of interest
4690 * @cpu: the CPU which is coming up
4691 *
4692 * An unbound pool may end up with a cpumask which doesn't have any online
4693 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4694 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4695 * online CPU before, cpus_allowed of all its workers should be restored.
4696 */
4697static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4698{
4699 static cpumask_t cpumask;
4700 struct worker *worker;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004701
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004702 lockdep_assert_held(&pool->attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004703
4704 /* is @cpu allowed for @pool? */
4705 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4706 return;
4707
Tejun Heo7dbc7252013-03-19 13:45:21 -07004708 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004709
4710 /* as we're called from CPU_ONLINE, the following shouldn't fail */
Lai Jiangshanda028462014-05-20 17:46:31 +08004711 for_each_pool_worker(worker, pool)
Peter Zijlstrad945b5e2016-06-16 14:38:42 +02004712 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004713}
4714
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004715int workqueue_prepare_cpu(unsigned int cpu)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004716{
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004717 struct worker_pool *pool;
4718
4719 for_each_cpu_worker_pool(pool, cpu) {
4720 if (pool->nr_workers)
4721 continue;
4722 if (!create_worker(pool))
4723 return -ENOMEM;
4724 }
4725 return 0;
4726}
4727
4728int workqueue_online_cpu(unsigned int cpu)
4729{
Tejun Heo4ce62e92012-07-13 22:16:44 -07004730 struct worker_pool *pool;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004731 struct workqueue_struct *wq;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004732 int pi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004733
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004734 mutex_lock(&wq_pool_mutex);
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07004735
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004736 for_each_pool(pool, pi) {
4737 mutex_lock(&pool->attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004738
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004739 if (pool->cpu == cpu)
4740 rebind_workers(pool);
4741 else if (pool->cpu < 0)
4742 restore_unbound_workers_cpumask(pool, cpu);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004743
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004744 mutex_unlock(&pool->attach_mutex);
Tejun Heo65758202012-07-17 12:39:26 -07004745 }
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004746
4747 /* update NUMA affinity of unbound workqueues */
4748 list_for_each_entry(wq, &workqueues, list)
4749 wq_update_unbound_numa(wq, cpu, true);
4750
4751 mutex_unlock(&wq_pool_mutex);
4752 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07004753}
4754
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004755int workqueue_offline_cpu(unsigned int cpu)
Tejun Heo65758202012-07-17 12:39:26 -07004756{
Tejun Heo8db25e72012-07-17 12:39:28 -07004757 struct work_struct unbind_work;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004758 struct workqueue_struct *wq;
Tejun Heo8db25e72012-07-17 12:39:28 -07004759
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004760 /* unbinding per-cpu workers should happen on the local CPU */
4761 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4762 queue_work_on(cpu, system_highpri_wq, &unbind_work);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004763
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004764 /* update NUMA affinity of unbound workqueues */
4765 mutex_lock(&wq_pool_mutex);
4766 list_for_each_entry(wq, &workqueues, list)
4767 wq_update_unbound_numa(wq, cpu, false);
4768 mutex_unlock(&wq_pool_mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004769
Thomas Gleixner7ee681b2016-07-13 17:16:29 +00004770 /* wait for per-cpu unbinding to finish */
4771 flush_work(&unbind_work);
4772 destroy_work_on_stack(&unbind_work);
4773 return 0;
Tejun Heo65758202012-07-17 12:39:26 -07004774}
4775
Rusty Russell2d3854a2008-11-05 13:39:10 +11004776#ifdef CONFIG_SMP
Rusty Russell8ccad402009-01-16 15:31:15 -08004777
Rusty Russell2d3854a2008-11-05 13:39:10 +11004778struct work_for_cpu {
Tejun Heoed48ece2012-09-18 12:48:43 -07004779 struct work_struct work;
Rusty Russell2d3854a2008-11-05 13:39:10 +11004780 long (*fn)(void *);
4781 void *arg;
4782 long ret;
4783};
4784
Tejun Heoed48ece2012-09-18 12:48:43 -07004785static void work_for_cpu_fn(struct work_struct *work)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004786{
Tejun Heoed48ece2012-09-18 12:48:43 -07004787 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4788
Rusty Russell2d3854a2008-11-05 13:39:10 +11004789 wfc->ret = wfc->fn(wfc->arg);
4790}
4791
4792/**
Anna-Maria Gleixner22aceb32016-03-10 12:07:38 +01004793 * work_on_cpu - run a function in thread context on a particular cpu
Rusty Russell2d3854a2008-11-05 13:39:10 +11004794 * @cpu: the cpu to run on
4795 * @fn: the function to run
4796 * @arg: the function arg
4797 *
Rusty Russell31ad9082009-01-16 15:31:15 -08004798 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06004799 * The caller must not hold any locks which would prevent @fn from completing.
Yacine Belkadid185af32013-07-31 14:59:24 -07004800 *
4801 * Return: The value @fn returns.
Rusty Russell2d3854a2008-11-05 13:39:10 +11004802 */
Tejun Heod84ff052013-03-12 11:29:59 -07004803long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004804{
Tejun Heoed48ece2012-09-18 12:48:43 -07004805 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
Rusty Russell2d3854a2008-11-05 13:39:10 +11004806
Tejun Heoed48ece2012-09-18 12:48:43 -07004807 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4808 schedule_work_on(cpu, &wfc.work);
Bjorn Helgaas12997d12013-11-18 11:00:29 -07004809 flush_work(&wfc.work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05004810 destroy_work_on_stack(&wfc.work);
Rusty Russell2d3854a2008-11-05 13:39:10 +11004811 return wfc.ret;
4812}
4813EXPORT_SYMBOL_GPL(work_on_cpu);
Thomas Gleixner0e8d6a92017-04-12 22:07:28 +02004814
4815/**
4816 * work_on_cpu_safe - run a function in thread context on a particular cpu
4817 * @cpu: the cpu to run on
4818 * @fn: the function to run
4819 * @arg: the function argument
4820 *
4821 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4822 * any locks which would prevent @fn from completing.
4823 *
4824 * Return: The value @fn returns.
4825 */
4826long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4827{
4828 long ret = -ENODEV;
4829
4830 get_online_cpus();
4831 if (cpu_online(cpu))
4832 ret = work_on_cpu(cpu, fn, arg);
4833 put_online_cpus();
4834 return ret;
4835}
4836EXPORT_SYMBOL_GPL(work_on_cpu_safe);
Rusty Russell2d3854a2008-11-05 13:39:10 +11004837#endif /* CONFIG_SMP */
4838
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004839#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10304840
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004841/**
4842 * freeze_workqueues_begin - begin freezing workqueues
4843 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01004844 * Start freezing workqueues. After this function returns, all freezable
Tejun Heoc5aa87b2013-03-13 16:51:36 -07004845 * workqueues will queue new works to their delayed_works list instead of
Tejun Heo706026c2013-01-24 11:01:34 -08004846 * pool->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004847 *
4848 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004849 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004850 */
4851void freeze_workqueues_begin(void)
4852{
Tejun Heo24b8a842013-03-12 11:29:58 -07004853 struct workqueue_struct *wq;
4854 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004855
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004856 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004857
Tejun Heo6183c002013-03-12 11:29:57 -07004858 WARN_ON_ONCE(workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004859 workqueue_freezing = true;
4860
Tejun Heo24b8a842013-03-12 11:29:58 -07004861 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004862 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004863 for_each_pwq(pwq, wq)
4864 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004865 mutex_unlock(&wq->mutex);
Tejun Heo24b8a842013-03-12 11:29:58 -07004866 }
Tejun Heo5bcab332013-03-13 19:47:40 -07004867
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004868 mutex_unlock(&wq_pool_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004869}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004870
4871/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01004872 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004873 *
4874 * Check whether freezing is complete. This function must be called
4875 * between freeze_workqueues_begin() and thaw_workqueues().
4876 *
4877 * CONTEXT:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004878 * Grabs and releases wq_pool_mutex.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004879 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004880 * Return:
Tejun Heo58a69cb2011-02-16 09:25:31 +01004881 * %true if some freezable workqueues are still busy. %false if freezing
4882 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004883 */
4884bool freeze_workqueues_busy(void)
4885{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004886 bool busy = false;
Tejun Heo24b8a842013-03-12 11:29:58 -07004887 struct workqueue_struct *wq;
4888 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004889
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004890 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004891
Tejun Heo6183c002013-03-12 11:29:57 -07004892 WARN_ON_ONCE(!workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004893
Tejun Heo24b8a842013-03-12 11:29:58 -07004894 list_for_each_entry(wq, &workqueues, list) {
4895 if (!(wq->flags & WQ_FREEZABLE))
4896 continue;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004897 /*
4898 * nr_active is monotonically decreasing. It's safe
4899 * to peek without lock.
4900 */
Lai Jiangshan88109452013-03-20 03:28:10 +08004901 rcu_read_lock_sched();
Tejun Heo24b8a842013-03-12 11:29:58 -07004902 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004903 WARN_ON_ONCE(pwq->nr_active < 0);
Tejun Heo112202d2013-02-13 19:29:12 -08004904 if (pwq->nr_active) {
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004905 busy = true;
Lai Jiangshan88109452013-03-20 03:28:10 +08004906 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004907 goto out_unlock;
4908 }
4909 }
Lai Jiangshan88109452013-03-20 03:28:10 +08004910 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004911 }
4912out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004913 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004914 return busy;
4915}
4916
4917/**
4918 * thaw_workqueues - thaw workqueues
4919 *
4920 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo706026c2013-01-24 11:01:34 -08004921 * frozen works are transferred to their respective pool worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004922 *
4923 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004924 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004925 */
4926void thaw_workqueues(void)
4927{
Tejun Heo24b8a842013-03-12 11:29:58 -07004928 struct workqueue_struct *wq;
4929 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004930
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004931 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004932
4933 if (!workqueue_freezing)
4934 goto out_unlock;
4935
Lai Jiangshan74b414e2014-05-22 19:01:16 +08004936 workqueue_freezing = false;
Tejun Heo8b03ae32010-06-29 10:07:12 +02004937
Tejun Heo24b8a842013-03-12 11:29:58 -07004938 /* restore max_active and repopulate worklist */
4939 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004940 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004941 for_each_pwq(pwq, wq)
4942 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004943 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004944 }
4945
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004946out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004947 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004948}
4949#endif /* CONFIG_FREEZER */
4950
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004951static int workqueue_apply_unbound_cpumask(void)
4952{
4953 LIST_HEAD(ctxs);
4954 int ret = 0;
4955 struct workqueue_struct *wq;
4956 struct apply_wqattrs_ctx *ctx, *n;
4957
4958 lockdep_assert_held(&wq_pool_mutex);
4959
4960 list_for_each_entry(wq, &workqueues, list) {
4961 if (!(wq->flags & WQ_UNBOUND))
4962 continue;
4963 /* creating multiple pwqs breaks ordering guarantee */
4964 if (wq->flags & __WQ_ORDERED)
4965 continue;
4966
4967 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4968 if (!ctx) {
4969 ret = -ENOMEM;
4970 break;
4971 }
4972
4973 list_add_tail(&ctx->list, &ctxs);
4974 }
4975
4976 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4977 if (!ret)
4978 apply_wqattrs_commit(ctx);
4979 apply_wqattrs_cleanup(ctx);
4980 }
4981
4982 return ret;
4983}
4984
4985/**
4986 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4987 * @cpumask: the cpumask to set
4988 *
4989 * The low-level workqueues cpumask is a global cpumask that limits
4990 * the affinity of all unbound workqueues. This function check the @cpumask
4991 * and apply it to all unbound workqueues and updates all pwqs of them.
4992 *
4993 * Retun: 0 - Success
4994 * -EINVAL - Invalid @cpumask
4995 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4996 */
4997int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4998{
4999 int ret = -EINVAL;
5000 cpumask_var_t saved_cpumask;
5001
5002 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5003 return -ENOMEM;
5004
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005005 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5006 if (!cpumask_empty(cpumask)) {
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005007 apply_wqattrs_lock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005008
5009 /* save the old wq_unbound_cpumask. */
5010 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5011
5012 /* update wq_unbound_cpumask at first and apply it to wqs. */
5013 cpumask_copy(wq_unbound_cpumask, cpumask);
5014 ret = workqueue_apply_unbound_cpumask();
5015
5016 /* restore the wq_unbound_cpumask when failed. */
5017 if (ret < 0)
5018 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5019
Lai Jiangshana0111cf2015-05-19 18:03:47 +08005020 apply_wqattrs_unlock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005021 }
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005022
5023 free_cpumask_var(saved_cpumask);
5024 return ret;
5025}
5026
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005027#ifdef CONFIG_SYSFS
5028/*
5029 * Workqueues with WQ_SYSFS flag set is visible to userland via
5030 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5031 * following attributes.
5032 *
5033 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5034 * max_active RW int : maximum number of in-flight work items
5035 *
5036 * Unbound workqueues have the following extra attributes.
5037 *
5038 * id RO int : the associated pool ID
5039 * nice RW int : nice value of the workers
5040 * cpumask RW mask : bitmask of allowed CPUs for the workers
5041 */
5042struct wq_device {
5043 struct workqueue_struct *wq;
5044 struct device dev;
5045};
5046
5047static struct workqueue_struct *dev_to_wq(struct device *dev)
5048{
5049 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5050
5051 return wq_dev->wq;
5052}
5053
5054static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5055 char *buf)
5056{
5057 struct workqueue_struct *wq = dev_to_wq(dev);
5058
5059 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5060}
5061static DEVICE_ATTR_RO(per_cpu);
5062
5063static ssize_t max_active_show(struct device *dev,
5064 struct device_attribute *attr, char *buf)
5065{
5066 struct workqueue_struct *wq = dev_to_wq(dev);
5067
5068 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5069}
5070
5071static ssize_t max_active_store(struct device *dev,
5072 struct device_attribute *attr, const char *buf,
5073 size_t count)
5074{
5075 struct workqueue_struct *wq = dev_to_wq(dev);
5076 int val;
5077
5078 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5079 return -EINVAL;
5080
5081 workqueue_set_max_active(wq, val);
5082 return count;
5083}
5084static DEVICE_ATTR_RW(max_active);
5085
5086static struct attribute *wq_sysfs_attrs[] = {
5087 &dev_attr_per_cpu.attr,
5088 &dev_attr_max_active.attr,
5089 NULL,
5090};
5091ATTRIBUTE_GROUPS(wq_sysfs);
5092
5093static ssize_t wq_pool_ids_show(struct device *dev,
5094 struct device_attribute *attr, char *buf)
5095{
5096 struct workqueue_struct *wq = dev_to_wq(dev);
5097 const char *delim = "";
5098 int node, written = 0;
5099
5100 rcu_read_lock_sched();
5101 for_each_node(node) {
5102 written += scnprintf(buf + written, PAGE_SIZE - written,
5103 "%s%d:%d", delim, node,
5104 unbound_pwq_by_node(wq, node)->pool->id);
5105 delim = " ";
5106 }
5107 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5108 rcu_read_unlock_sched();
5109
5110 return written;
5111}
5112
5113static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5114 char *buf)
5115{
5116 struct workqueue_struct *wq = dev_to_wq(dev);
5117 int written;
5118
5119 mutex_lock(&wq->mutex);
5120 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5121 mutex_unlock(&wq->mutex);
5122
5123 return written;
5124}
5125
5126/* prepare workqueue_attrs for sysfs store operations */
5127static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5128{
5129 struct workqueue_attrs *attrs;
5130
Lai Jiangshan899a94f2015-05-20 14:41:18 +08005131 lockdep_assert_held(&wq_pool_mutex);
5132
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005133 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5134 if (!attrs)
5135 return NULL;
5136
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005137 copy_workqueue_attrs(attrs, wq->unbound_attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005138 return attrs;
5139}
5140
5141static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5142 const char *buf, size_t count)
5143{
5144 struct workqueue_struct *wq = dev_to_wq(dev);
5145 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005146 int ret = -ENOMEM;
5147
5148 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005149
5150 attrs = wq_sysfs_prep_attrs(wq);
5151 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005152 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005153
5154 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5155 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005156 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005157 else
5158 ret = -EINVAL;
5159
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005160out_unlock:
5161 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005162 free_workqueue_attrs(attrs);
5163 return ret ?: count;
5164}
5165
5166static ssize_t wq_cpumask_show(struct device *dev,
5167 struct device_attribute *attr, char *buf)
5168{
5169 struct workqueue_struct *wq = dev_to_wq(dev);
5170 int written;
5171
5172 mutex_lock(&wq->mutex);
5173 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5174 cpumask_pr_args(wq->unbound_attrs->cpumask));
5175 mutex_unlock(&wq->mutex);
5176 return written;
5177}
5178
5179static ssize_t wq_cpumask_store(struct device *dev,
5180 struct device_attribute *attr,
5181 const char *buf, size_t count)
5182{
5183 struct workqueue_struct *wq = dev_to_wq(dev);
5184 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005185 int ret = -ENOMEM;
5186
5187 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005188
5189 attrs = wq_sysfs_prep_attrs(wq);
5190 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005191 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005192
5193 ret = cpumask_parse(buf, attrs->cpumask);
5194 if (!ret)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005195 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005196
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005197out_unlock:
5198 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005199 free_workqueue_attrs(attrs);
5200 return ret ?: count;
5201}
5202
5203static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5204 char *buf)
5205{
5206 struct workqueue_struct *wq = dev_to_wq(dev);
5207 int written;
5208
5209 mutex_lock(&wq->mutex);
5210 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5211 !wq->unbound_attrs->no_numa);
5212 mutex_unlock(&wq->mutex);
5213
5214 return written;
5215}
5216
5217static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5218 const char *buf, size_t count)
5219{
5220 struct workqueue_struct *wq = dev_to_wq(dev);
5221 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005222 int v, ret = -ENOMEM;
5223
5224 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005225
5226 attrs = wq_sysfs_prep_attrs(wq);
5227 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005228 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005229
5230 ret = -EINVAL;
5231 if (sscanf(buf, "%d", &v) == 1) {
5232 attrs->no_numa = !v;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005233 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005234 }
5235
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005236out_unlock:
5237 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005238 free_workqueue_attrs(attrs);
5239 return ret ?: count;
5240}
5241
5242static struct device_attribute wq_sysfs_unbound_attrs[] = {
5243 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5244 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5245 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5246 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5247 __ATTR_NULL,
5248};
5249
5250static struct bus_type wq_subsys = {
5251 .name = "workqueue",
5252 .dev_groups = wq_sysfs_groups,
5253};
5254
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005255static ssize_t wq_unbound_cpumask_show(struct device *dev,
5256 struct device_attribute *attr, char *buf)
5257{
5258 int written;
5259
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005260 mutex_lock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005261 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5262 cpumask_pr_args(wq_unbound_cpumask));
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005263 mutex_unlock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005264
5265 return written;
5266}
5267
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005268static ssize_t wq_unbound_cpumask_store(struct device *dev,
5269 struct device_attribute *attr, const char *buf, size_t count)
5270{
5271 cpumask_var_t cpumask;
5272 int ret;
5273
5274 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5275 return -ENOMEM;
5276
5277 ret = cpumask_parse(buf, cpumask);
5278 if (!ret)
5279 ret = workqueue_set_unbound_cpumask(cpumask);
5280
5281 free_cpumask_var(cpumask);
5282 return ret ? ret : count;
5283}
5284
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005285static struct device_attribute wq_sysfs_cpumask_attr =
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005286 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5287 wq_unbound_cpumask_store);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005288
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005289static int __init wq_sysfs_init(void)
5290{
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005291 int err;
5292
5293 err = subsys_virtual_register(&wq_subsys, NULL);
5294 if (err)
5295 return err;
5296
5297 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005298}
5299core_initcall(wq_sysfs_init);
5300
5301static void wq_device_release(struct device *dev)
5302{
5303 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5304
5305 kfree(wq_dev);
5306}
5307
5308/**
5309 * workqueue_sysfs_register - make a workqueue visible in sysfs
5310 * @wq: the workqueue to register
5311 *
5312 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5313 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5314 * which is the preferred method.
5315 *
5316 * Workqueue user should use this function directly iff it wants to apply
5317 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5318 * apply_workqueue_attrs() may race against userland updating the
5319 * attributes.
5320 *
5321 * Return: 0 on success, -errno on failure.
5322 */
5323int workqueue_sysfs_register(struct workqueue_struct *wq)
5324{
5325 struct wq_device *wq_dev;
5326 int ret;
5327
5328 /*
Shailendra Verma402dd892015-05-23 10:38:14 +05305329 * Adjusting max_active or creating new pwqs by applying
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005330 * attributes breaks ordering guarantee. Disallow exposing ordered
5331 * workqueues.
5332 */
Tejun Heo0a94efb2017-07-23 08:36:15 -04005333 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005334 return -EINVAL;
5335
5336 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5337 if (!wq_dev)
5338 return -ENOMEM;
5339
5340 wq_dev->wq = wq;
5341 wq_dev->dev.bus = &wq_subsys;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005342 wq_dev->dev.release = wq_device_release;
Lars-Peter Clausen23217b42016-02-17 21:04:41 +01005343 dev_set_name(&wq_dev->dev, "%s", wq->name);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005344
5345 /*
5346 * unbound_attrs are created separately. Suppress uevent until
5347 * everything is ready.
5348 */
5349 dev_set_uevent_suppress(&wq_dev->dev, true);
5350
5351 ret = device_register(&wq_dev->dev);
5352 if (ret) {
5353 kfree(wq_dev);
5354 wq->wq_dev = NULL;
5355 return ret;
5356 }
5357
5358 if (wq->flags & WQ_UNBOUND) {
5359 struct device_attribute *attr;
5360
5361 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5362 ret = device_create_file(&wq_dev->dev, attr);
5363 if (ret) {
5364 device_unregister(&wq_dev->dev);
5365 wq->wq_dev = NULL;
5366 return ret;
5367 }
5368 }
5369 }
5370
5371 dev_set_uevent_suppress(&wq_dev->dev, false);
5372 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5373 return 0;
5374}
5375
5376/**
5377 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5378 * @wq: the workqueue to unregister
5379 *
5380 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5381 */
5382static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5383{
5384 struct wq_device *wq_dev = wq->wq_dev;
5385
5386 if (!wq->wq_dev)
5387 return;
5388
5389 wq->wq_dev = NULL;
5390 device_unregister(&wq_dev->dev);
5391}
5392#else /* CONFIG_SYSFS */
5393static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5394#endif /* CONFIG_SYSFS */
5395
Tejun Heo82607adc2015-12-08 11:28:04 -05005396/*
5397 * Workqueue watchdog.
5398 *
5399 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5400 * flush dependency, a concurrency managed work item which stays RUNNING
5401 * indefinitely. Workqueue stalls can be very difficult to debug as the
5402 * usual warning mechanisms don't trigger and internal workqueue state is
5403 * largely opaque.
5404 *
5405 * Workqueue watchdog monitors all worker pools periodically and dumps
5406 * state if some pools failed to make forward progress for a while where
5407 * forward progress is defined as the first item on ->worklist changing.
5408 *
5409 * This mechanism is controlled through the kernel parameter
5410 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5411 * corresponding sysfs parameter file.
5412 */
5413#ifdef CONFIG_WQ_WATCHDOG
5414
5415static void wq_watchdog_timer_fn(unsigned long data);
5416
5417static unsigned long wq_watchdog_thresh = 30;
5418static struct timer_list wq_watchdog_timer =
5419 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5420
5421static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5422static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5423
5424static void wq_watchdog_reset_touched(void)
5425{
5426 int cpu;
5427
5428 wq_watchdog_touched = jiffies;
5429 for_each_possible_cpu(cpu)
5430 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5431}
5432
5433static void wq_watchdog_timer_fn(unsigned long data)
5434{
5435 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5436 bool lockup_detected = false;
5437 struct worker_pool *pool;
5438 int pi;
5439
5440 if (!thresh)
5441 return;
5442
5443 rcu_read_lock();
5444
5445 for_each_pool(pool, pi) {
5446 unsigned long pool_ts, touched, ts;
5447
5448 if (list_empty(&pool->worklist))
5449 continue;
5450
5451 /* get the latest of pool and touched timestamps */
5452 pool_ts = READ_ONCE(pool->watchdog_ts);
5453 touched = READ_ONCE(wq_watchdog_touched);
5454
5455 if (time_after(pool_ts, touched))
5456 ts = pool_ts;
5457 else
5458 ts = touched;
5459
5460 if (pool->cpu >= 0) {
5461 unsigned long cpu_touched =
5462 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5463 pool->cpu));
5464 if (time_after(cpu_touched, ts))
5465 ts = cpu_touched;
5466 }
5467
5468 /* did we stall? */
5469 if (time_after(jiffies, ts + thresh)) {
5470 lockup_detected = true;
5471 pr_emerg("BUG: workqueue lockup - pool");
5472 pr_cont_pool_info(pool);
5473 pr_cont(" stuck for %us!\n",
5474 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5475 }
5476 }
5477
5478 rcu_read_unlock();
5479
5480 if (lockup_detected)
5481 show_workqueue_state();
5482
5483 wq_watchdog_reset_touched();
5484 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5485}
5486
5487void wq_watchdog_touch(int cpu)
5488{
5489 if (cpu >= 0)
5490 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5491 else
5492 wq_watchdog_touched = jiffies;
5493}
5494
5495static void wq_watchdog_set_thresh(unsigned long thresh)
5496{
5497 wq_watchdog_thresh = 0;
5498 del_timer_sync(&wq_watchdog_timer);
5499
5500 if (thresh) {
5501 wq_watchdog_thresh = thresh;
5502 wq_watchdog_reset_touched();
5503 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5504 }
5505}
5506
5507static int wq_watchdog_param_set_thresh(const char *val,
5508 const struct kernel_param *kp)
5509{
5510 unsigned long thresh;
5511 int ret;
5512
5513 ret = kstrtoul(val, 0, &thresh);
5514 if (ret)
5515 return ret;
5516
5517 if (system_wq)
5518 wq_watchdog_set_thresh(thresh);
5519 else
5520 wq_watchdog_thresh = thresh;
5521
5522 return 0;
5523}
5524
5525static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5526 .set = wq_watchdog_param_set_thresh,
5527 .get = param_get_ulong,
5528};
5529
5530module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5531 0644);
5532
5533static void wq_watchdog_init(void)
5534{
5535 wq_watchdog_set_thresh(wq_watchdog_thresh);
5536}
5537
5538#else /* CONFIG_WQ_WATCHDOG */
5539
5540static inline void wq_watchdog_init(void) { }
5541
5542#endif /* CONFIG_WQ_WATCHDOG */
5543
Tejun Heobce90382013-04-01 11:23:32 -07005544static void __init wq_numa_init(void)
5545{
5546 cpumask_var_t *tbl;
5547 int node, cpu;
5548
Tejun Heobce90382013-04-01 11:23:32 -07005549 if (num_possible_nodes() <= 1)
5550 return;
5551
Tejun Heod55262c2013-04-01 11:23:38 -07005552 if (wq_disable_numa) {
5553 pr_info("workqueue: NUMA affinity support disabled\n");
5554 return;
5555 }
5556
Tejun Heo4c16bd32013-04-01 11:23:36 -07005557 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5558 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5559
Tejun Heobce90382013-04-01 11:23:32 -07005560 /*
5561 * We want masks of possible CPUs of each node which isn't readily
5562 * available. Build one from cpu_to_node() which should have been
5563 * fully initialized by now.
5564 */
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08005565 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
Tejun Heobce90382013-04-01 11:23:32 -07005566 BUG_ON(!tbl);
5567
5568 for_each_node(node)
Yasuaki Ishimatsu5a6024f2014-07-07 09:56:48 -04005569 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
Tejun Heo1be0c252013-05-15 14:24:24 -07005570 node_online(node) ? node : NUMA_NO_NODE));
Tejun Heobce90382013-04-01 11:23:32 -07005571
5572 for_each_possible_cpu(cpu) {
5573 node = cpu_to_node(cpu);
5574 if (WARN_ON(node == NUMA_NO_NODE)) {
5575 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5576 /* happens iff arch is bonkers, let's just proceed */
5577 return;
5578 }
5579 cpumask_set_cpu(cpu, tbl[node]);
5580 }
5581
5582 wq_numa_possible_cpumask = tbl;
5583 wq_numa_enabled = true;
5584}
5585
Tejun Heo3347fa02016-09-16 15:49:32 -04005586/**
5587 * workqueue_init_early - early init for workqueue subsystem
5588 *
5589 * This is the first half of two-staged workqueue subsystem initialization
5590 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5591 * idr are up. It sets up all the data structures and system workqueues
5592 * and allows early boot code to create workqueues and queue/cancel work
5593 * items. Actual work item execution starts only after kthreads can be
5594 * created and scheduled right before early initcalls.
5595 */
5596int __init workqueue_init_early(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005597{
Tejun Heo7a4e3442013-03-12 11:30:00 -07005598 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5599 int i, cpu;
Tejun Heoc34056a2010-06-29 10:07:11 +02005600
Tejun Heoe904e6c2013-03-12 11:29:57 -07005601 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5602
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005603 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5604 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5605
Tejun Heoe904e6c2013-03-12 11:29:57 -07005606 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5607
Tejun Heo706026c2013-01-24 11:01:34 -08005608 /* initialize CPU pools */
Tejun Heo29c91e92013-03-12 11:30:03 -07005609 for_each_possible_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005610 struct worker_pool *pool;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005611
Tejun Heo7a4e3442013-03-12 11:30:00 -07005612 i = 0;
Tejun Heof02ae732013-03-12 11:30:03 -07005613 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo7a4e3442013-03-12 11:30:00 -07005614 BUG_ON(init_worker_pool(pool));
Tejun Heoec22ca52013-01-24 11:01:33 -08005615 pool->cpu = cpu;
Tejun Heo29c91e92013-03-12 11:30:03 -07005616 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
Tejun Heo7a4e3442013-03-12 11:30:00 -07005617 pool->attrs->nice = std_nice[i++];
Tejun Heof3f90ad2013-04-01 11:23:34 -07005618 pool->node = cpu_to_node(cpu);
Tejun Heo7a4e3442013-03-12 11:30:00 -07005619
Tejun Heo9daf9e62013-01-24 11:01:33 -08005620 /* alloc pool ID */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005621 mutex_lock(&wq_pool_mutex);
Tejun Heo9daf9e62013-01-24 11:01:33 -08005622 BUG_ON(worker_pool_assign_id(pool));
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005623 mutex_unlock(&wq_pool_mutex);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005624 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02005625 }
5626
Tejun Heo8a2b7532013-09-05 12:30:04 -04005627 /* create default unbound and ordered wq attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -07005628 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5629 struct workqueue_attrs *attrs;
5630
5631 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
Tejun Heo29c91e92013-03-12 11:30:03 -07005632 attrs->nice = std_nice[i];
Tejun Heo29c91e92013-03-12 11:30:03 -07005633 unbound_std_wq_attrs[i] = attrs;
Tejun Heo8a2b7532013-09-05 12:30:04 -04005634
5635 /*
5636 * An ordered wq should have only one pwq as ordering is
5637 * guaranteed by max_active which is enforced by pwqs.
5638 * Turn off NUMA so that dfl_pwq is used for all nodes.
5639 */
5640 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5641 attrs->nice = std_nice[i];
5642 attrs->no_numa = true;
5643 ordered_wq_attrs[i] = attrs;
Tejun Heo29c91e92013-03-12 11:30:03 -07005644 }
5645
Tejun Heod320c032010-06-29 10:07:14 +02005646 system_wq = alloc_workqueue("events", 0, 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005647 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
Tejun Heod320c032010-06-29 10:07:14 +02005648 system_long_wq = alloc_workqueue("events_long", 0, 0);
Tejun Heof3421792010-07-02 10:03:51 +02005649 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5650 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01005651 system_freezable_wq = alloc_workqueue("events_freezable",
5652 WQ_FREEZABLE, 0);
Viresh Kumar06681062013-04-24 17:12:54 +05305653 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5654 WQ_POWER_EFFICIENT, 0);
5655 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5656 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5657 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005658 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
Viresh Kumar06681062013-04-24 17:12:54 +05305659 !system_unbound_wq || !system_freezable_wq ||
5660 !system_power_efficient_wq ||
5661 !system_freezable_power_efficient_wq);
Tejun Heo82607adc2015-12-08 11:28:04 -05005662
Tejun Heo3347fa02016-09-16 15:49:32 -04005663 return 0;
5664}
5665
5666/**
5667 * workqueue_init - bring workqueue subsystem fully online
5668 *
5669 * This is the latter half of two-staged workqueue subsystem initialization
5670 * and invoked as soon as kthreads can be created and scheduled.
5671 * Workqueues have been created and work items queued on them, but there
5672 * are no kworkers executing the work items yet. Populate the worker pools
5673 * with the initial workers and enable future kworker creations.
5674 */
5675int __init workqueue_init(void)
5676{
Tejun Heo2186d9f2016-10-19 12:01:27 -04005677 struct workqueue_struct *wq;
Tejun Heo3347fa02016-09-16 15:49:32 -04005678 struct worker_pool *pool;
5679 int cpu, bkt;
5680
Tejun Heo2186d9f2016-10-19 12:01:27 -04005681 /*
5682 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5683 * CPU to node mapping may not be available that early on some
5684 * archs such as power and arm64. As per-cpu pools created
5685 * previously could be missing node hint and unbound pools NUMA
5686 * affinity, fix them up.
5687 */
5688 wq_numa_init();
5689
5690 mutex_lock(&wq_pool_mutex);
5691
5692 for_each_possible_cpu(cpu) {
5693 for_each_cpu_worker_pool(pool, cpu) {
5694 pool->node = cpu_to_node(cpu);
5695 }
5696 }
5697
5698 list_for_each_entry(wq, &workqueues, list)
5699 wq_update_unbound_numa(wq, smp_processor_id(), true);
5700
5701 mutex_unlock(&wq_pool_mutex);
5702
Tejun Heo3347fa02016-09-16 15:49:32 -04005703 /* create the initial workers */
5704 for_each_online_cpu(cpu) {
5705 for_each_cpu_worker_pool(pool, cpu) {
5706 pool->flags &= ~POOL_DISASSOCIATED;
5707 BUG_ON(!create_worker(pool));
5708 }
5709 }
5710
5711 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5712 BUG_ON(!create_worker(pool));
5713
5714 wq_online = true;
Tejun Heo82607adc2015-12-08 11:28:04 -05005715 wq_watchdog_init();
5716
Suresh Siddha6ee05782010-07-30 14:57:37 -07005717 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005718}