blob: ad8dc2b9efc3786abddd6ba38e23b641e1e0daee [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Tejun Heoc54fce62010-09-10 16:51:36 +02002 * kernel/workqueue.c - generic async execution with shared worker pool
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
Tejun Heoc54fce62010-09-10 16:51:36 +02004 * Copyright (C) 2002 Ingo Molnar
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 *
Tejun Heoc54fce62010-09-10 16:51:36 +02006 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
Christoph Lameter89ada672005-10-30 15:01:59 -080011 *
Christoph Lametercde53532008-07-04 09:59:22 -070012 * Made to use alloc_percpu by Christoph Lameter.
Tejun Heoc54fce62010-09-10 16:51:36 +020013 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
Libinb11895c2013-08-21 08:50:39 +080019 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
Tejun Heoc54fce62010-09-10 16:51:36 +020023 *
24 * Please read Documentation/workqueue.txt for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -070025 */
26
Paul Gortmaker9984de12011-05-23 14:51:41 -040027#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
James Bottomley1fa44ec2006-02-23 12:43:43 -060038#include <linux/hardirq.h>
Christoph 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>
Tejun Heoe22bee72010-06-29 10:07:14 +020051
Tejun Heoea138442013-01-18 14:05:55 -080052#include "workqueue_internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
Tejun Heoc8e55f32010-06-29 10:07:12 +020054enum {
Tejun Heobc2ae0f2012-07-17 12:39:27 -070055 /*
Tejun Heo24647572013-01-24 11:01:33 -080056 * worker_pool flags
Tejun Heobc2ae0f2012-07-17 12:39:27 -070057 *
Tejun Heo24647572013-01-24 11:01:33 -080058 * A bound pool is either associated or disassociated with its CPU.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070059 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
61 * is in effect.
62 *
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
Tejun Heo24647572013-01-24 11:01:33 -080065 * be executing on any CPU. The pool behaves as an unbound one.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070066 *
Tejun Heobc3a1af2013-03-13 19:47:39 -070067 * Note that DISASSOCIATED should be flipped only while holding
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +080068 * attach_mutex to avoid changing binding state while
Lai Jiangshan4736cbf2014-05-20 17:46:35 +080069 * worker_attach_to_pool() is in progress.
Tejun Heobc2ae0f2012-07-17 12:39:27 -070070 */
Tejun Heo24647572013-01-24 11:01:33 -080071 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020072
Tejun Heoc8e55f32010-06-29 10:07:12 +020073 /* worker flags */
Tejun Heoc8e55f32010-06-29 10:07:12 +020074 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
Tejun Heoe22bee72010-06-29 10:07:14 +020076 WORKER_PREP = 1 << 3, /* preparing to run works */
Tejun Heofb0e7be2010-06-29 10:07:15 +020077 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
Tejun Heof3421792010-07-02 10:03:51 +020078 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
Tejun Heoa9ab7752013-03-19 13:45:21 -070079 WORKER_REBOUND = 1 << 8, /* worker was rebound */
Tejun Heoe22bee72010-06-29 10:07:14 +020080
Tejun Heoa9ab7752013-03-19 13:45:21 -070081 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
Tejun Heodb7bccf2010-06-29 10:07:12 +020083
Tejun Heoe34cdddb2013-01-24 11:01:33 -080084 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
Tejun Heo4ce62e92012-07-13 22:16:44 -070085
Tejun Heo29c91e92013-03-12 11:30:03 -070086 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
Tejun Heoc8e55f32010-06-29 10:07:12 +020087 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020088
Tejun Heoe22bee72010-06-29 10:07:14 +020089 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91
Tejun Heo3233cdb2011-02-16 18:10:19 +010092 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
94 (min two ticks) */
Tejun Heoe22bee72010-06-29 10:07:14 +020095 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
Tejun Heoe22bee72010-06-29 10:07:14 +020097
98 /*
99 * Rescue workers are used only on emergencies and shared by
Dongsheng Yang8698a742014-03-11 18:09:12 +0800100 * all cpus. Give MIN_NICE.
Tejun Heoe22bee72010-06-29 10:07:14 +0200101 */
Dongsheng Yang8698a742014-03-11 18:09:12 +0800102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
Tejun Heoecf68812013-04-01 11:23:34 -0700104
105 WQ_NAME_LEN = 24,
Tejun Heoc8e55f32010-06-29 10:07:12 +0200106};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
108/*
Tejun Heo4690c4a2010-06-29 10:07:10 +0200109 * Structure fields follow one of the following exclusion rules.
110 *
Tejun Heoe41e7042010-08-24 14:22:47 +0200111 * I: Modifiable by initialization/destruction paths and read-only for
112 * everyone else.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200113 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
116 *
Tejun Heod565ed62013-01-24 11:01:33 -0800117 * L: pool->lock protected. Access with pool->lock held.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200118 *
Tejun Heod565ed62013-01-24 11:01:33 -0800119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
Tejun Heoe22bee72010-06-29 10:07:14 +0200123 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800124 * A: pool->attach_mutex protected.
Tejun Heo822d8402013-03-19 13:45:21 -0700125 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700126 * PL: wq_pool_mutex protected.
Tejun Heo76af4d92013-03-12 11:30:00 -0700127 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo5bcab332013-03-13 19:47:40 -0700129 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
131 *
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
134 *
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700135 * WQ: wq->mutex protected.
136 *
Lai Jiangshanb5927602013-03-25 16:57:19 -0700137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
Tejun Heo2e109a22013-03-13 19:47:40 -0700138 *
139 * MD: wq_mayday_lock protected.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200140 */
141
Tejun Heo2eaebdb2013-01-18 14:05:55 -0800142/* struct worker is defined in workqueue_internal.h */
Tejun Heoc34056a2010-06-29 10:07:11 +0200143
Tejun Heobd7bdd42012-07-12 14:46:37 -0700144struct worker_pool {
Tejun Heod565ed62013-01-24 11:01:33 -0800145 spinlock_t lock; /* the pool lock */
Tejun Heod84ff052013-03-12 11:29:59 -0700146 int cpu; /* I: the associated cpu */
Tejun Heof3f90ad2013-04-01 11:23:34 -0700147 int node; /* I: the associated node ID */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800148 int id; /* I: pool ID */
Tejun Heo11ebea52012-07-12 14:46:37 -0700149 unsigned int flags; /* X: flags */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700150
151 struct list_head worklist; /* L: list of pending works */
152 int nr_workers; /* L: total number of workers */
Lai Jiangshanea1abd62012-09-18 09:59:22 -0700153
154 /* nr_idle includes the ones off idle_list for rebinding */
Tejun Heobd7bdd42012-07-12 14:46:37 -0700155 int nr_idle; /* L: currently idle ones */
156
157 struct list_head idle_list; /* X: list of idle workers */
158 struct timer_list idle_timer; /* L: worker idle timeout */
159 struct timer_list mayday_timer; /* L: SOS timer for workers */
160
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700161 /* a workers is either on busy_hash or idle_list, or the manager */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800162 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
163 /* L: hash of busy workers */
164
Tejun Heobc3a1af2013-03-13 19:47:39 -0700165 /* see manage_workers() for details on the two manager mutexes */
Tejun Heo34a06bd2013-03-12 11:30:00 -0700166 struct mutex manager_arb; /* manager arbitration */
Tejun Heo2607d7a2015-03-09 09:22:28 -0400167 struct worker *manager; /* L: purely informational */
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800168 struct mutex attach_mutex; /* attach/detach exclusion */
169 struct list_head workers; /* A: attached workers */
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +0800170 struct completion *detach_completion; /* all workers detached */
Tejun Heoe19e3972013-01-24 11:39:44 -0800171
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +0800172 struct ida worker_ida; /* worker IDs for task name */
Tejun Heoe19e3972013-01-24 11:39:44 -0800173
Tejun Heo7a4e3442013-03-12 11:30:00 -0700174 struct workqueue_attrs *attrs; /* I: worker attributes */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700175 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
176 int refcnt; /* PL: refcnt for unbound pools */
Tejun Heo7a4e3442013-03-12 11:30:00 -0700177
Tejun Heoe19e3972013-01-24 11:39:44 -0800178 /*
179 * The current concurrency level. As it's likely to be accessed
180 * from other CPUs during try_to_wake_up(), put it in a separate
181 * cacheline.
182 */
183 atomic_t nr_running ____cacheline_aligned_in_smp;
Tejun Heo29c91e92013-03-12 11:30:03 -0700184
185 /*
186 * Destruction of pool is sched-RCU protected to allow dereferences
187 * from get_work_pool().
188 */
189 struct rcu_head rcu;
Tejun Heo8b03ae32010-06-29 10:07:12 +0200190} ____cacheline_aligned_in_smp;
191
192/*
Tejun Heo112202d2013-02-13 19:29:12 -0800193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
194 * of work_struct->data are used for flags and the remaining high bits
195 * point to the pwq; thus, pwqs need to be aligned at two's power of the
196 * number of flag bits.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 */
Tejun Heo112202d2013-02-13 19:29:12 -0800198struct pool_workqueue {
Tejun Heobd7bdd42012-07-12 14:46:37 -0700199 struct worker_pool *pool; /* I: the associated pool */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200200 struct workqueue_struct *wq; /* I: the owning workqueue */
Tejun Heo73f53c42010-06-29 10:07:11 +0200201 int work_color; /* L: current color */
202 int flush_color; /* L: flushing color */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700203 int refcnt; /* L: reference count */
Tejun Heo73f53c42010-06-29 10:07:11 +0200204 int nr_in_flight[WORK_NR_COLORS];
205 /* L: nr of in_flight works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200206 int nr_active; /* L: nr of active works */
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200207 int max_active; /* L: max active works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200208 struct list_head delayed_works; /* L: delayed works */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700209 struct list_head pwqs_node; /* WR: node on wq->pwqs */
Tejun Heo2e109a22013-03-13 19:47:40 -0700210 struct list_head mayday_node; /* MD: node on wq->maydays */
Tejun Heo8864b4e2013-03-12 11:30:04 -0700211
212 /*
213 * Release of unbound pwq is punted to system_wq. See put_pwq()
214 * and pwq_unbound_release_workfn() for details. pool_workqueue
215 * itself is also sched-RCU protected so that the first pwq can be
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700216 * determined without grabbing wq->mutex.
Tejun Heo8864b4e2013-03-12 11:30:04 -0700217 */
218 struct work_struct unbound_release_work;
219 struct rcu_head rcu;
Tejun Heoe904e6c2013-03-12 11:29:57 -0700220} __aligned(1 << WORK_STRUCT_FLAG_BITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222/*
Tejun Heo73f53c42010-06-29 10:07:11 +0200223 * Structure used to wait for workqueue flush.
224 */
225struct wq_flusher {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700226 struct list_head list; /* WQ: list of flushers */
227 int flush_color; /* WQ: flush color waiting for */
Tejun Heo73f53c42010-06-29 10:07:11 +0200228 struct completion done; /* flush completion */
229};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230
Tejun Heo226223a2013-03-12 11:30:05 -0700231struct wq_device;
232
Tejun Heo73f53c42010-06-29 10:07:11 +0200233/*
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700234 * The externally visible workqueue. It relays the issued work items to
235 * the appropriate worker_pool through its pool_workqueues.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236 */
237struct workqueue_struct {
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700238 struct list_head pwqs; /* WR: all pwqs of this wq */
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400239 struct list_head list; /* PR: list of all workqueues */
Tejun Heo73f53c42010-06-29 10:07:11 +0200240
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700241 struct mutex mutex; /* protects this wq */
242 int work_color; /* WQ: current work color */
243 int flush_color; /* WQ: current flush color */
Tejun Heo112202d2013-02-13 19:29:12 -0800244 atomic_t nr_pwqs_to_flush; /* flush in progress */
Lai Jiangshan3c25a552013-03-25 16:57:17 -0700245 struct wq_flusher *first_flusher; /* WQ: first flusher */
246 struct list_head flusher_queue; /* WQ: flush waiters */
247 struct list_head flusher_overflow; /* WQ: flush overflow list */
Tejun Heo73f53c42010-06-29 10:07:11 +0200248
Tejun Heo2e109a22013-03-13 19:47:40 -0700249 struct list_head maydays; /* MD: pwqs requesting rescue */
Tejun Heoe22bee72010-06-29 10:07:14 +0200250 struct worker *rescuer; /* I: rescue worker */
251
Lai Jiangshan87fc7412013-03-25 16:57:18 -0700252 int nr_drainers; /* WQ: drain in progress */
Lai Jiangshana357fc02013-03-25 16:57:19 -0700253 int saved_max_active; /* WQ: saved pwq max_active */
Tejun Heo226223a2013-03-12 11:30:05 -0700254
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800255 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
256 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
Tejun Heo6029a912013-04-01 11:23:34 -0700257
Tejun Heo226223a2013-03-12 11:30:05 -0700258#ifdef CONFIG_SYSFS
259 struct wq_device *wq_dev; /* I: for sysfs interface */
260#endif
Johannes Berg4e6045f2007-10-18 23:39:55 -0700261#ifdef CONFIG_LOCKDEP
Tejun Heo4690c4a2010-06-29 10:07:10 +0200262 struct lockdep_map lockdep_map;
Johannes Berg4e6045f2007-10-18 23:39:55 -0700263#endif
Tejun Heoecf68812013-04-01 11:23:34 -0700264 char name[WQ_NAME_LEN]; /* I: workqueue name */
Tejun Heo2728fd22013-04-01 11:23:35 -0700265
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400266 /*
267 * Destruction of workqueue_struct is sched-RCU protected to allow
268 * walking the workqueues list without grabbing wq_pool_mutex.
269 * This is used to dump all workqueues from sysrq.
270 */
271 struct rcu_head rcu;
272
Tejun Heo2728fd22013-04-01 11:23:35 -0700273 /* hot fields used during command issue, aligned to cacheline */
274 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
275 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800276 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277};
278
Tejun Heoe904e6c2013-03-12 11:29:57 -0700279static struct kmem_cache *pwq_cache;
280
Tejun Heobce90382013-04-01 11:23:32 -0700281static cpumask_var_t *wq_numa_possible_cpumask;
282 /* possible CPUs of each node */
283
Tejun Heod55262c2013-04-01 11:23:38 -0700284static bool wq_disable_numa;
285module_param_named(disable_numa, wq_disable_numa, bool, 0444);
286
Viresh Kumarcee22a12013-04-08 16:45:40 +0530287/* see the comment above the definition of WQ_POWER_EFFICIENT */
288#ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
289static bool wq_power_efficient = true;
290#else
291static bool wq_power_efficient;
292#endif
293
294module_param_named(power_efficient, wq_power_efficient, bool, 0444);
295
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 Heo5bcab332013-03-13 19:47:40 -0700303
Tejun Heoe2dca7a2015-03-09 09:22:28 -0400304static LIST_HEAD(workqueues); /* PR: list of all workqueues */
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700305static bool workqueue_freezing; /* PL: have wqs started freezing? */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700306
Lai Jiangshan042f7df2015-04-30 17:16:12 +0800307static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +0800308
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700309/* the per-cpu worker pools */
310static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
311 cpu_worker_pools);
312
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700313static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700314
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700315/* PL: hash of all unbound pools keyed by pool->attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -0700316static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
317
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700318/* I: attributes used when instantiating standard unbound pools on demand */
Tejun Heo29c91e92013-03-12 11:30:03 -0700319static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
320
Tejun Heo8a2b7532013-09-05 12:30:04 -0400321/* I: attributes used when instantiating ordered pools on demand */
322static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
323
Tejun Heod320c032010-06-29 10:07:14 +0200324struct workqueue_struct *system_wq __read_mostly;
Marc Dionnead7b1f82013-05-06 17:44:55 -0400325EXPORT_SYMBOL(system_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300326struct workqueue_struct *system_highpri_wq __read_mostly;
Joonsoo Kim1aabe902012-08-15 23:25:39 +0900327EXPORT_SYMBOL_GPL(system_highpri_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300328struct workqueue_struct *system_long_wq __read_mostly;
Tejun Heod320c032010-06-29 10:07:14 +0200329EXPORT_SYMBOL_GPL(system_long_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300330struct workqueue_struct *system_unbound_wq __read_mostly;
Tejun Heof3421792010-07-02 10:03:51 +0200331EXPORT_SYMBOL_GPL(system_unbound_wq);
Valentin Ilie044c7822012-08-19 00:52:42 +0300332struct workqueue_struct *system_freezable_wq __read_mostly;
Tejun Heo24d51ad2011-02-21 09:52:50 +0100333EXPORT_SYMBOL_GPL(system_freezable_wq);
Viresh Kumar06681062013-04-24 17:12:54 +0530334struct workqueue_struct *system_power_efficient_wq __read_mostly;
335EXPORT_SYMBOL_GPL(system_power_efficient_wq);
336struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
337EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
Tejun Heod320c032010-06-29 10:07:14 +0200338
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700339static int worker_thread(void *__worker);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +0800340static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
Tejun Heo7d19c5c2013-03-13 19:47:40 -0700341
Tejun Heo97bd2342010-10-05 10:41:14 +0200342#define CREATE_TRACE_POINTS
343#include <trace/events/workqueue.h>
344
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700345#define assert_rcu_or_pool_mutex() \
Tejun Heo5bcab332013-03-13 19:47:40 -0700346 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700347 lockdep_is_held(&wq_pool_mutex), \
348 "sched RCU or wq_pool_mutex should be held")
Tejun Heo5bcab332013-03-13 19:47:40 -0700349
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700350#define assert_rcu_or_wq_mutex(wq) \
Tejun Heo76af4d92013-03-12 11:30:00 -0700351 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
Lai Jiangshanb5927602013-03-25 16:57:19 -0700352 lockdep_is_held(&wq->mutex), \
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700353 "sched RCU or wq->mutex should be held")
Tejun Heo76af4d92013-03-12 11:30:00 -0700354
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800355#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
356 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
357 lockdep_is_held(&wq->mutex) || \
358 lockdep_is_held(&wq_pool_mutex), \
359 "sched RCU, wq->mutex or wq_pool_mutex should be held")
360
Tejun Heof02ae732013-03-12 11:30:03 -0700361#define for_each_cpu_worker_pool(pool, cpu) \
362 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
363 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
Tejun Heo7a62c2c2013-03-12 11:30:03 -0700364 (pool)++)
Tejun Heo4ce62e92012-07-13 22:16:44 -0700365
Tejun Heo49e3cf42013-03-12 11:29:58 -0700366/**
Tejun Heo17116962013-03-12 11:29:58 -0700367 * for_each_pool - iterate through all worker_pools in the system
368 * @pool: iteration cursor
Tejun Heo611c92a2013-03-13 16:51:36 -0700369 * @pi: integer used for iteration
Tejun Heofa1b54e2013-03-12 11:30:00 -0700370 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700371 * This must be called either with wq_pool_mutex held or sched RCU read
372 * locked. If the pool needs to be used beyond the locking in effect, the
373 * caller is responsible for guaranteeing that the pool stays online.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700374 *
375 * The if/else clause exists only for the lockdep assertion and can be
376 * ignored.
Tejun Heo17116962013-03-12 11:29:58 -0700377 */
Tejun Heo611c92a2013-03-13 16:51:36 -0700378#define for_each_pool(pool, pi) \
379 idr_for_each_entry(&worker_pool_idr, pool, pi) \
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700380 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
Tejun Heofa1b54e2013-03-12 11:30:00 -0700381 else
Tejun Heo17116962013-03-12 11:29:58 -0700382
383/**
Tejun Heo822d8402013-03-19 13:45:21 -0700384 * for_each_pool_worker - iterate through all workers of a worker_pool
385 * @worker: iteration cursor
Tejun Heo822d8402013-03-19 13:45:21 -0700386 * @pool: worker_pool to iterate workers of
387 *
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800388 * This must be called with @pool->attach_mutex.
Tejun Heo822d8402013-03-19 13:45:21 -0700389 *
390 * The if/else clause exists only for the lockdep assertion and can be
391 * ignored.
392 */
Lai Jiangshanda028462014-05-20 17:46:31 +0800393#define for_each_pool_worker(worker, pool) \
394 list_for_each_entry((worker), &(pool)->workers, node) \
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +0800395 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
Tejun Heo822d8402013-03-19 13:45:21 -0700396 else
397
398/**
Tejun Heo49e3cf42013-03-12 11:29:58 -0700399 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
400 * @pwq: iteration cursor
401 * @wq: the target workqueue
Tejun Heo76af4d92013-03-12 11:30:00 -0700402 *
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700403 * This must be called either with wq->mutex held or sched RCU read locked.
Tejun Heo794b18b2013-03-13 19:47:40 -0700404 * If the pwq needs to be used beyond the locking in effect, the caller is
405 * responsible for guaranteeing that the pwq stays online.
Tejun Heo76af4d92013-03-12 11:30:00 -0700406 *
407 * The if/else clause exists only for the lockdep assertion and can be
408 * ignored.
Tejun Heo49e3cf42013-03-12 11:29:58 -0700409 */
410#define for_each_pwq(pwq, wq) \
Tejun Heo76af4d92013-03-12 11:30:00 -0700411 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -0700412 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
Tejun Heo76af4d92013-03-12 11:30:00 -0700413 else
Tejun Heof3421792010-07-02 10:03:51 +0200414
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900415#ifdef CONFIG_DEBUG_OBJECTS_WORK
416
417static struct debug_obj_descr work_debug_descr;
418
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100419static void *work_debug_hint(void *addr)
420{
421 return ((struct work_struct *) addr)->func;
422}
423
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900424/*
425 * fixup_init is called when:
426 * - an active object is initialized
427 */
428static int work_fixup_init(void *addr, enum debug_obj_state state)
429{
430 struct work_struct *work = addr;
431
432 switch (state) {
433 case ODEBUG_STATE_ACTIVE:
434 cancel_work_sync(work);
435 debug_object_init(work, &work_debug_descr);
436 return 1;
437 default:
438 return 0;
439 }
440}
441
442/*
443 * fixup_activate is called when:
444 * - an active object is activated
445 * - an unknown object is activated (might be a statically initialized object)
446 */
447static int work_fixup_activate(void *addr, enum debug_obj_state state)
448{
449 struct work_struct *work = addr;
450
451 switch (state) {
452
453 case ODEBUG_STATE_NOTAVAILABLE:
454 /*
455 * This is not really a fixup. The work struct was
456 * statically initialized. We just make sure that it
457 * is tracked in the object tracker.
458 */
Tejun Heo22df02b2010-06-29 10:07:10 +0200459 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900460 debug_object_init(work, &work_debug_descr);
461 debug_object_activate(work, &work_debug_descr);
462 return 0;
463 }
464 WARN_ON_ONCE(1);
465 return 0;
466
467 case ODEBUG_STATE_ACTIVE:
468 WARN_ON(1);
469
470 default:
471 return 0;
472 }
473}
474
475/*
476 * fixup_free is called when:
477 * - an active object is freed
478 */
479static int work_fixup_free(void *addr, enum debug_obj_state state)
480{
481 struct work_struct *work = addr;
482
483 switch (state) {
484 case ODEBUG_STATE_ACTIVE:
485 cancel_work_sync(work);
486 debug_object_free(work, &work_debug_descr);
487 return 1;
488 default:
489 return 0;
490 }
491}
492
493static struct debug_obj_descr work_debug_descr = {
494 .name = "work_struct",
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100495 .debug_hint = work_debug_hint,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900496 .fixup_init = work_fixup_init,
497 .fixup_activate = work_fixup_activate,
498 .fixup_free = work_fixup_free,
499};
500
501static inline void debug_work_activate(struct work_struct *work)
502{
503 debug_object_activate(work, &work_debug_descr);
504}
505
506static inline void debug_work_deactivate(struct work_struct *work)
507{
508 debug_object_deactivate(work, &work_debug_descr);
509}
510
511void __init_work(struct work_struct *work, int onstack)
512{
513 if (onstack)
514 debug_object_init_on_stack(work, &work_debug_descr);
515 else
516 debug_object_init(work, &work_debug_descr);
517}
518EXPORT_SYMBOL_GPL(__init_work);
519
520void destroy_work_on_stack(struct work_struct *work)
521{
522 debug_object_free(work, &work_debug_descr);
523}
524EXPORT_SYMBOL_GPL(destroy_work_on_stack);
525
Thomas Gleixnerea2e64f2014-03-23 14:20:44 +0000526void destroy_delayed_work_on_stack(struct delayed_work *work)
527{
528 destroy_timer_on_stack(&work->timer);
529 debug_object_free(&work->work, &work_debug_descr);
530}
531EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
532
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900533#else
534static inline void debug_work_activate(struct work_struct *work) { }
535static inline void debug_work_deactivate(struct work_struct *work) { }
536#endif
537
Li Bin4e8b22b2013-09-10 09:52:35 +0800538/**
539 * worker_pool_assign_id - allocate ID and assing it to @pool
540 * @pool: the pool pointer of interest
541 *
542 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
543 * successfully, -errno on failure.
544 */
Tejun Heo9daf9e62013-01-24 11:01:33 -0800545static int worker_pool_assign_id(struct worker_pool *pool)
546{
547 int ret;
548
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700549 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo5bcab332013-03-13 19:47:40 -0700550
Li Bin4e8b22b2013-09-10 09:52:35 +0800551 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
552 GFP_KERNEL);
Tejun Heo229641a2013-04-01 17:08:13 -0700553 if (ret >= 0) {
Tejun Heoe68035f2013-03-13 14:59:38 -0700554 pool->id = ret;
Tejun Heo229641a2013-04-01 17:08:13 -0700555 return 0;
556 }
Tejun Heo9daf9e62013-01-24 11:01:33 -0800557 return ret;
558}
559
Tejun Heo76af4d92013-03-12 11:30:00 -0700560/**
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700561 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
562 * @wq: the target workqueue
563 * @node: the node ID
564 *
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800565 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
566 * read locked.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700567 * If the pwq needs to be used beyond the locking in effect, the caller is
568 * responsible for guaranteeing that the pwq stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700569 *
570 * Return: The unbound pool_workqueue for @node.
Tejun Heodf2d5ae2013-04-01 11:23:35 -0700571 */
572static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
573 int node)
574{
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +0800575 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
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);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200647}
648
649static void clear_work_data(struct work_struct *work)
650{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800651 smp_wmb(); /* see set_work_pool_and_clear_pending() */
652 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200653}
654
Tejun Heo112202d2013-02-13 19:29:12 -0800655static struct pool_workqueue *get_work_pwq(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200656{
Tejun Heoe1201532010-07-22 14:14:25 +0200657 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200658
Tejun Heo112202d2013-02-13 19:29:12 -0800659 if (data & WORK_STRUCT_PWQ)
Tejun Heoe1201532010-07-22 14:14:25 +0200660 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
661 else
662 return NULL;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200663}
664
Tejun Heo7c3eed52013-01-24 11:01:33 -0800665/**
666 * get_work_pool - return the worker_pool a given work was associated with
667 * @work: the work item of interest
668 *
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700669 * Pools are created and destroyed under wq_pool_mutex, and allows read
670 * access under sched-RCU read lock. As such, this function should be
671 * called under wq_pool_mutex or with preemption disabled.
Tejun Heofa1b54e2013-03-12 11:30:00 -0700672 *
673 * All fields of the returned pool are accessible as long as the above
674 * mentioned locking is in effect. If the returned pool needs to be used
675 * beyond the critical section, the caller is responsible for ensuring the
676 * returned pool is and stays online.
Yacine Belkadid185af32013-07-31 14:59:24 -0700677 *
678 * Return: The worker_pool @work was last associated with. %NULL if none.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800679 */
680static struct worker_pool *get_work_pool(struct work_struct *work)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200681{
Tejun Heoe1201532010-07-22 14:14:25 +0200682 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800683 int pool_id;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200684
Lai Jiangshan68e13a62013-03-25 16:57:17 -0700685 assert_rcu_or_pool_mutex();
Tejun Heofa1b54e2013-03-12 11:30:00 -0700686
Tejun Heo112202d2013-02-13 19:29:12 -0800687 if (data & WORK_STRUCT_PWQ)
688 return ((struct pool_workqueue *)
Tejun Heo7c3eed52013-01-24 11:01:33 -0800689 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200690
Tejun Heo7c3eed52013-01-24 11:01:33 -0800691 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
692 if (pool_id == WORK_OFFQ_POOL_NONE)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200693 return NULL;
694
Tejun Heofa1b54e2013-03-12 11:30:00 -0700695 return idr_find(&worker_pool_idr, pool_id);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800696}
697
698/**
699 * get_work_pool_id - return the worker pool ID a given work is associated with
700 * @work: the work item of interest
701 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700702 * Return: The worker_pool ID @work was last associated with.
Tejun Heo7c3eed52013-01-24 11:01:33 -0800703 * %WORK_OFFQ_POOL_NONE if none.
704 */
705static int get_work_pool_id(struct work_struct *work)
706{
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800707 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7c3eed52013-01-24 11:01:33 -0800708
Tejun Heo112202d2013-02-13 19:29:12 -0800709 if (data & WORK_STRUCT_PWQ)
710 return ((struct pool_workqueue *)
Lai Jiangshan54d5b7d2013-02-07 13:14:20 -0800711 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
712
713 return data >> WORK_OFFQ_POOL_SHIFT;
Tejun Heo7c3eed52013-01-24 11:01:33 -0800714}
715
Tejun Heobbb68df2012-08-03 10:30:46 -0700716static void mark_work_canceling(struct work_struct *work)
717{
Tejun Heo7c3eed52013-01-24 11:01:33 -0800718 unsigned long pool_id = get_work_pool_id(work);
Tejun Heobbb68df2012-08-03 10:30:46 -0700719
Tejun Heo7c3eed52013-01-24 11:01:33 -0800720 pool_id <<= WORK_OFFQ_POOL_SHIFT;
721 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700722}
723
724static bool work_is_canceling(struct work_struct *work)
725{
726 unsigned long data = atomic_long_read(&work->data);
727
Tejun Heo112202d2013-02-13 19:29:12 -0800728 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
Tejun Heobbb68df2012-08-03 10:30:46 -0700729}
730
David Howells365970a2006-11-22 14:54:49 +0000731/*
Tejun Heo32704762012-07-13 22:16:45 -0700732 * Policy functions. These define the policies on how the global worker
733 * pools are managed. Unless noted otherwise, these functions assume that
Tejun Heod565ed62013-01-24 11:01:33 -0800734 * they're being called with pool->lock held.
David Howells365970a2006-11-22 14:54:49 +0000735 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200736
Tejun Heo63d95a92012-07-12 14:46:37 -0700737static bool __need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000738{
Tejun Heoe19e3972013-01-24 11:39:44 -0800739 return !atomic_read(&pool->nr_running);
David Howells365970a2006-11-22 14:54:49 +0000740}
741
Tejun Heoe22bee72010-06-29 10:07:14 +0200742/*
743 * Need to wake up a worker? Called from anything but currently
744 * running workers.
Tejun Heo974271c2012-07-12 14:46:37 -0700745 *
746 * Note that, because unbound workers never contribute to nr_running, this
Tejun Heo706026c2013-01-24 11:01:34 -0800747 * function will always return %true for unbound pools as long as the
Tejun Heo974271c2012-07-12 14:46:37 -0700748 * worklist isn't empty.
Tejun Heoe22bee72010-06-29 10:07:14 +0200749 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700750static bool need_more_worker(struct worker_pool *pool)
David Howells365970a2006-11-22 14:54:49 +0000751{
Tejun Heo63d95a92012-07-12 14:46:37 -0700752 return !list_empty(&pool->worklist) && __need_more_worker(pool);
David Howells365970a2006-11-22 14:54:49 +0000753}
754
Tejun Heoe22bee72010-06-29 10:07:14 +0200755/* Can I start working? Called from busy but !running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700756static bool may_start_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200757{
Tejun Heo63d95a92012-07-12 14:46:37 -0700758 return pool->nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200759}
760
761/* Do I need to keep working? Called from currently running workers. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700762static bool keep_working(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200763{
Tejun Heoe19e3972013-01-24 11:39:44 -0800764 return !list_empty(&pool->worklist) &&
765 atomic_read(&pool->nr_running) <= 1;
Tejun Heoe22bee72010-06-29 10:07:14 +0200766}
767
768/* Do we need a new worker? Called from manager. */
Tejun Heo63d95a92012-07-12 14:46:37 -0700769static bool need_to_create_worker(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200770{
Tejun Heo63d95a92012-07-12 14:46:37 -0700771 return need_more_worker(pool) && !may_start_working(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200772}
773
Tejun Heoe22bee72010-06-29 10:07:14 +0200774/* Do we have too many workers and should some go away? */
Tejun Heo63d95a92012-07-12 14:46:37 -0700775static bool too_many_workers(struct worker_pool *pool)
Tejun Heoe22bee72010-06-29 10:07:14 +0200776{
Tejun Heo34a06bd2013-03-12 11:30:00 -0700777 bool managing = mutex_is_locked(&pool->manager_arb);
Tejun Heo63d95a92012-07-12 14:46:37 -0700778 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
779 int nr_busy = pool->nr_workers - nr_idle;
Tejun Heoe22bee72010-06-29 10:07:14 +0200780
781 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
782}
783
784/*
785 * Wake up functions.
786 */
787
Lai Jiangshan1037de32014-05-22 16:44:07 +0800788/* Return the first idle worker. Safe with preemption disabled */
789static struct worker *first_idle_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200790{
Tejun Heo63d95a92012-07-12 14:46:37 -0700791 if (unlikely(list_empty(&pool->idle_list)))
Tejun Heo7e116292010-06-29 10:07:13 +0200792 return NULL;
793
Tejun Heo63d95a92012-07-12 14:46:37 -0700794 return list_first_entry(&pool->idle_list, struct worker, entry);
Tejun Heo7e116292010-06-29 10:07:13 +0200795}
796
797/**
798 * wake_up_worker - wake up an idle worker
Tejun Heo63d95a92012-07-12 14:46:37 -0700799 * @pool: worker pool to wake worker from
Tejun Heo7e116292010-06-29 10:07:13 +0200800 *
Tejun Heo63d95a92012-07-12 14:46:37 -0700801 * Wake up the first idle worker of @pool.
Tejun Heo7e116292010-06-29 10:07:13 +0200802 *
803 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800804 * spin_lock_irq(pool->lock).
Tejun Heo7e116292010-06-29 10:07:13 +0200805 */
Tejun Heo63d95a92012-07-12 14:46:37 -0700806static void wake_up_worker(struct worker_pool *pool)
Tejun Heo7e116292010-06-29 10:07:13 +0200807{
Lai Jiangshan1037de32014-05-22 16:44:07 +0800808 struct worker *worker = first_idle_worker(pool);
Tejun Heo7e116292010-06-29 10:07:13 +0200809
810 if (likely(worker))
811 wake_up_process(worker->task);
812}
813
Tejun Heo4690c4a2010-06-29 10:07:10 +0200814/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200815 * wq_worker_waking_up - a worker is waking up
816 * @task: task waking up
817 * @cpu: CPU @task is waking up to
818 *
819 * This function is called during try_to_wake_up() when a worker is
820 * being awoken.
821 *
822 * CONTEXT:
823 * spin_lock_irq(rq->lock)
824 */
Tejun Heod84ff052013-03-12 11:29:59 -0700825void wq_worker_waking_up(struct task_struct *task, int cpu)
Tejun Heoe22bee72010-06-29 10:07:14 +0200826{
827 struct worker *worker = kthread_data(task);
828
Joonsoo Kim36576002012-10-26 23:03:49 +0900829 if (!(worker->flags & WORKER_NOT_RUNNING)) {
Tejun Heoec22ca52013-01-24 11:01:33 -0800830 WARN_ON_ONCE(worker->pool->cpu != cpu);
Tejun Heoe19e3972013-01-24 11:39:44 -0800831 atomic_inc(&worker->pool->nr_running);
Joonsoo Kim36576002012-10-26 23:03:49 +0900832 }
Tejun Heoe22bee72010-06-29 10:07:14 +0200833}
834
835/**
836 * wq_worker_sleeping - a worker is going to sleep
837 * @task: task going to sleep
838 * @cpu: CPU in question, must be the current CPU number
839 *
840 * This function is called during schedule() when a busy worker is
841 * going to sleep. Worker on the same cpu can be woken up by
842 * returning pointer to its task.
843 *
844 * CONTEXT:
845 * spin_lock_irq(rq->lock)
846 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700847 * Return:
Tejun Heoe22bee72010-06-29 10:07:14 +0200848 * Worker task on @cpu to wake up, %NULL if none.
849 */
Tejun Heod84ff052013-03-12 11:29:59 -0700850struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
Tejun Heoe22bee72010-06-29 10:07:14 +0200851{
852 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
Tejun Heo111c2252013-01-17 17:16:24 -0800853 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200854
Tejun Heo111c2252013-01-17 17:16:24 -0800855 /*
856 * Rescuers, which may not have all the fields set up like normal
857 * workers, also reach here, let's not access anything before
858 * checking NOT_RUNNING.
859 */
Steven Rostedt2d646722010-12-03 23:12:33 -0500860 if (worker->flags & WORKER_NOT_RUNNING)
Tejun Heoe22bee72010-06-29 10:07:14 +0200861 return NULL;
862
Tejun Heo111c2252013-01-17 17:16:24 -0800863 pool = worker->pool;
Tejun Heo111c2252013-01-17 17:16:24 -0800864
Tejun Heoe22bee72010-06-29 10:07:14 +0200865 /* this can only happen on the local cpu */
Lai Jiangshan92b69f52014-06-03 15:33:08 +0800866 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
Tejun Heo6183c002013-03-12 11:29:57 -0700867 return NULL;
Tejun Heoe22bee72010-06-29 10:07:14 +0200868
869 /*
870 * The counterpart of the following dec_and_test, implied mb,
871 * worklist not empty test sequence is in insert_work().
872 * Please read comment there.
873 *
Tejun Heo628c78e2012-07-17 12:39:27 -0700874 * NOT_RUNNING is clear. This means that we're bound to and
875 * running on the local cpu w/ rq lock held and preemption
876 * disabled, which in turn means that none else could be
Tejun Heod565ed62013-01-24 11:01:33 -0800877 * manipulating idle_list, so dereferencing idle_list without pool
Tejun Heo628c78e2012-07-17 12:39:27 -0700878 * lock is safe.
Tejun Heoe22bee72010-06-29 10:07:14 +0200879 */
Tejun Heoe19e3972013-01-24 11:39:44 -0800880 if (atomic_dec_and_test(&pool->nr_running) &&
881 !list_empty(&pool->worklist))
Lai Jiangshan1037de32014-05-22 16:44:07 +0800882 to_wakeup = first_idle_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +0200883 return to_wakeup ? to_wakeup->task : NULL;
884}
885
886/**
887 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200888 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200889 * @flags: flags to set
Tejun Heod302f012010-06-29 10:07:13 +0200890 *
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800891 * Set @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200892 *
Tejun Heocb444762010-07-02 10:03:50 +0200893 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800894 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200895 */
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800896static inline void worker_set_flags(struct worker *worker, unsigned int flags)
Tejun Heod302f012010-06-29 10:07:13 +0200897{
Tejun Heobd7bdd42012-07-12 14:46:37 -0700898 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200899
Tejun Heocb444762010-07-02 10:03:50 +0200900 WARN_ON_ONCE(worker->task != current);
901
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800902 /* If transitioning into NOT_RUNNING, adjust nr_running. */
Tejun Heoe22bee72010-06-29 10:07:14 +0200903 if ((flags & WORKER_NOT_RUNNING) &&
904 !(worker->flags & WORKER_NOT_RUNNING)) {
Lai Jiangshan228f1d02014-07-22 13:02:00 +0800905 atomic_dec(&pool->nr_running);
Tejun Heoe22bee72010-06-29 10:07:14 +0200906 }
907
Tejun Heod302f012010-06-29 10:07:13 +0200908 worker->flags |= flags;
909}
910
911/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200912 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200913 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200914 * @flags: flags to clear
915 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200916 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200917 *
Tejun Heocb444762010-07-02 10:03:50 +0200918 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800919 * spin_lock_irq(pool->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200920 */
921static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
922{
Tejun Heo63d95a92012-07-12 14:46:37 -0700923 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +0200924 unsigned int oflags = worker->flags;
925
Tejun Heocb444762010-07-02 10:03:50 +0200926 WARN_ON_ONCE(worker->task != current);
927
Tejun Heod302f012010-06-29 10:07:13 +0200928 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200929
Tejun Heo42c025f2011-01-11 15:58:49 +0100930 /*
931 * If transitioning out of NOT_RUNNING, increment nr_running. Note
932 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
933 * of multiple flags, not a single flag.
934 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200935 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
936 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe19e3972013-01-24 11:39:44 -0800937 atomic_inc(&pool->nr_running);
Tejun Heod302f012010-06-29 10:07:13 +0200938}
939
940/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200941 * find_worker_executing_work - find worker which is executing a work
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800942 * @pool: pool of interest
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200943 * @work: work to find worker for
944 *
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800945 * Find a worker which is executing @work on @pool by searching
946 * @pool->busy_hash which is keyed by the address of @work. For a worker
Tejun Heoa2c1c572012-12-18 10:35:02 -0800947 * to match, its current execution should match the address of @work and
948 * its work function. This is to avoid unwanted dependency between
949 * unrelated work executions through a work item being recycled while still
950 * being executed.
951 *
952 * This is a bit tricky. A work item may be freed once its execution
953 * starts and nothing prevents the freed area from being recycled for
954 * another work item. If the same work item address ends up being reused
955 * before the original execution finishes, workqueue will identify the
956 * recycled work item as currently executing and make it wait until the
957 * current execution finishes, introducing an unwanted dependency.
958 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -0700959 * This function checks the work item address and work function to avoid
960 * false positives. Note that this isn't complete as one may construct a
961 * work function which can introduce dependency onto itself through a
962 * recycled work item. Well, if somebody wants to shoot oneself in the
963 * foot that badly, there's only so much we can do, and if such deadlock
964 * actually occurs, it should be easy to locate the culprit work function.
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200965 *
966 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -0800967 * spin_lock_irq(pool->lock).
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200968 *
Yacine Belkadid185af32013-07-31 14:59:24 -0700969 * Return:
970 * Pointer to worker which is executing @work if found, %NULL
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200971 * otherwise.
972 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -0800973static struct worker *find_worker_executing_work(struct worker_pool *pool,
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200974 struct work_struct *work)
975{
Sasha Levin42f85702012-12-17 10:01:23 -0500976 struct worker *worker;
Sasha Levin42f85702012-12-17 10:01:23 -0500977
Sasha Levinb67bfe02013-02-27 17:06:00 -0800978 hash_for_each_possible(pool->busy_hash, worker, hentry,
Tejun Heoa2c1c572012-12-18 10:35:02 -0800979 (unsigned long)work)
980 if (worker->current_work == work &&
981 worker->current_func == work->func)
Sasha Levin42f85702012-12-17 10:01:23 -0500982 return worker;
983
984 return NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200985}
986
987/**
Tejun Heobf4ede02012-08-03 10:30:46 -0700988 * move_linked_works - move linked works to a list
989 * @work: start of series of works to be scheduled
990 * @head: target list to append @work to
991 * @nextp: out paramter for nested worklist walking
992 *
993 * Schedule linked works starting from @work to @head. Work series to
994 * be scheduled starts at @work and includes any consecutive work with
995 * WORK_STRUCT_LINKED set in its predecessor.
996 *
997 * If @nextp is not NULL, it's updated to point to the next work of
998 * the last scheduled work. This allows move_linked_works() to be
999 * nested inside outer list_for_each_entry_safe().
1000 *
1001 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001002 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001003 */
1004static void move_linked_works(struct work_struct *work, struct list_head *head,
1005 struct work_struct **nextp)
1006{
1007 struct work_struct *n;
1008
1009 /*
1010 * Linked worklist will always end before the end of the list,
1011 * use NULL for list head.
1012 */
1013 list_for_each_entry_safe_from(work, n, NULL, entry) {
1014 list_move_tail(&work->entry, head);
1015 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1016 break;
1017 }
1018
1019 /*
1020 * If we're already inside safe list traversal and have moved
1021 * multiple works to the scheduled queue, the next position
1022 * needs to be updated.
1023 */
1024 if (nextp)
1025 *nextp = n;
1026}
1027
Tejun Heo8864b4e2013-03-12 11:30:04 -07001028/**
1029 * get_pwq - get an extra reference on the specified pool_workqueue
1030 * @pwq: pool_workqueue to get
1031 *
1032 * Obtain an extra reference on @pwq. The caller should guarantee that
1033 * @pwq has positive refcnt and be holding the matching pool->lock.
1034 */
1035static void get_pwq(struct pool_workqueue *pwq)
1036{
1037 lockdep_assert_held(&pwq->pool->lock);
1038 WARN_ON_ONCE(pwq->refcnt <= 0);
1039 pwq->refcnt++;
1040}
1041
1042/**
1043 * put_pwq - put a pool_workqueue reference
1044 * @pwq: pool_workqueue to put
1045 *
1046 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1047 * destruction. The caller should be holding the matching pool->lock.
1048 */
1049static void put_pwq(struct pool_workqueue *pwq)
1050{
1051 lockdep_assert_held(&pwq->pool->lock);
1052 if (likely(--pwq->refcnt))
1053 return;
1054 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1055 return;
1056 /*
1057 * @pwq can't be released under pool->lock, bounce to
1058 * pwq_unbound_release_workfn(). This never recurses on the same
1059 * pool->lock as this path is taken only for unbound workqueues and
1060 * the release work item is scheduled on a per-cpu workqueue. To
1061 * avoid lockdep warning, unbound pool->locks are given lockdep
1062 * subclass of 1 in get_unbound_pool().
1063 */
1064 schedule_work(&pwq->unbound_release_work);
1065}
1066
Tejun Heodce90d42013-04-01 11:23:35 -07001067/**
1068 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1069 * @pwq: pool_workqueue to put (can be %NULL)
1070 *
1071 * put_pwq() with locking. This function also allows %NULL @pwq.
1072 */
1073static void put_pwq_unlocked(struct pool_workqueue *pwq)
1074{
1075 if (pwq) {
1076 /*
1077 * As both pwqs and pools are sched-RCU protected, the
1078 * following lock operations are safe.
1079 */
1080 spin_lock_irq(&pwq->pool->lock);
1081 put_pwq(pwq);
1082 spin_unlock_irq(&pwq->pool->lock);
1083 }
1084}
1085
Tejun Heo112202d2013-02-13 19:29:12 -08001086static void pwq_activate_delayed_work(struct work_struct *work)
Tejun Heobf4ede02012-08-03 10:30:46 -07001087{
Tejun Heo112202d2013-02-13 19:29:12 -08001088 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobf4ede02012-08-03 10:30:46 -07001089
1090 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001091 move_linked_works(work, &pwq->pool->worklist, NULL);
Tejun Heobf4ede02012-08-03 10:30:46 -07001092 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo112202d2013-02-13 19:29:12 -08001093 pwq->nr_active++;
Tejun Heobf4ede02012-08-03 10:30:46 -07001094}
1095
Tejun Heo112202d2013-02-13 19:29:12 -08001096static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001097{
Tejun Heo112202d2013-02-13 19:29:12 -08001098 struct work_struct *work = list_first_entry(&pwq->delayed_works,
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001099 struct work_struct, entry);
1100
Tejun Heo112202d2013-02-13 19:29:12 -08001101 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001102}
1103
Tejun Heobf4ede02012-08-03 10:30:46 -07001104/**
Tejun Heo112202d2013-02-13 19:29:12 -08001105 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1106 * @pwq: pwq of interest
Tejun Heobf4ede02012-08-03 10:30:46 -07001107 * @color: color of work which left the queue
Tejun Heobf4ede02012-08-03 10:30:46 -07001108 *
1109 * A work either has completed or is removed from pending queue,
Tejun Heo112202d2013-02-13 19:29:12 -08001110 * decrement nr_in_flight of its pwq and handle workqueue flushing.
Tejun Heobf4ede02012-08-03 10:30:46 -07001111 *
1112 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001113 * spin_lock_irq(pool->lock).
Tejun Heobf4ede02012-08-03 10:30:46 -07001114 */
Tejun Heo112202d2013-02-13 19:29:12 -08001115static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
Tejun Heobf4ede02012-08-03 10:30:46 -07001116{
Tejun Heo8864b4e2013-03-12 11:30:04 -07001117 /* uncolored work items don't participate in flushing or nr_active */
Tejun Heobf4ede02012-08-03 10:30:46 -07001118 if (color == WORK_NO_COLOR)
Tejun Heo8864b4e2013-03-12 11:30:04 -07001119 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001120
Tejun Heo112202d2013-02-13 19:29:12 -08001121 pwq->nr_in_flight[color]--;
Tejun Heobf4ede02012-08-03 10:30:46 -07001122
Tejun Heo112202d2013-02-13 19:29:12 -08001123 pwq->nr_active--;
1124 if (!list_empty(&pwq->delayed_works)) {
Lai Jiangshanb3f9f402012-09-18 10:40:00 -07001125 /* one down, submit a delayed one */
Tejun Heo112202d2013-02-13 19:29:12 -08001126 if (pwq->nr_active < pwq->max_active)
1127 pwq_activate_first_delayed(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001128 }
1129
1130 /* is flush in progress and are we at the flushing tip? */
Tejun Heo112202d2013-02-13 19:29:12 -08001131 if (likely(pwq->flush_color != color))
Tejun Heo8864b4e2013-03-12 11:30:04 -07001132 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001133
1134 /* are there still in-flight works? */
Tejun Heo112202d2013-02-13 19:29:12 -08001135 if (pwq->nr_in_flight[color])
Tejun Heo8864b4e2013-03-12 11:30:04 -07001136 goto out_put;
Tejun Heobf4ede02012-08-03 10:30:46 -07001137
Tejun Heo112202d2013-02-13 19:29:12 -08001138 /* this pwq is done, clear flush_color */
1139 pwq->flush_color = -1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001140
1141 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001142 * If this was the last pwq, wake up the first flusher. It
Tejun Heobf4ede02012-08-03 10:30:46 -07001143 * will handle the rest.
1144 */
Tejun Heo112202d2013-02-13 19:29:12 -08001145 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1146 complete(&pwq->wq->first_flusher->done);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001147out_put:
1148 put_pwq(pwq);
Tejun Heobf4ede02012-08-03 10:30:46 -07001149}
1150
Tejun Heo36e227d2012-08-03 10:30:46 -07001151/**
Tejun Heobbb68df2012-08-03 10:30:46 -07001152 * try_to_grab_pending - steal work item from worklist and disable irq
Tejun Heo36e227d2012-08-03 10:30:46 -07001153 * @work: work item to steal
1154 * @is_dwork: @work is a delayed_work
Tejun Heobbb68df2012-08-03 10:30:46 -07001155 * @flags: place to store irq state
Tejun Heo36e227d2012-08-03 10:30:46 -07001156 *
1157 * Try to grab PENDING bit of @work. This function can handle @work in any
Yacine Belkadid185af32013-07-31 14:59:24 -07001158 * stable state - idle, on timer or on worklist.
Tejun Heo36e227d2012-08-03 10:30:46 -07001159 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001160 * Return:
Tejun Heo36e227d2012-08-03 10:30:46 -07001161 * 1 if @work was pending and we successfully stole PENDING
1162 * 0 if @work was idle and we claimed PENDING
1163 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
Tejun Heobbb68df2012-08-03 10:30:46 -07001164 * -ENOENT if someone else is canceling @work, this state may persist
1165 * for arbitrarily long
Tejun Heo36e227d2012-08-03 10:30:46 -07001166 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001167 * Note:
Tejun Heobbb68df2012-08-03 10:30:46 -07001168 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001169 * interrupted while holding PENDING and @work off queue, irq must be
1170 * disabled on entry. This, combined with delayed_work->timer being
1171 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
Tejun Heobbb68df2012-08-03 10:30:46 -07001172 *
1173 * On successful return, >= 0, irq is disabled and the caller is
1174 * responsible for releasing it using local_irq_restore(*@flags).
1175 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001176 * This function is safe to call from any context including IRQ handler.
Tejun Heobf4ede02012-08-03 10:30:46 -07001177 */
Tejun Heobbb68df2012-08-03 10:30:46 -07001178static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1179 unsigned long *flags)
Tejun Heobf4ede02012-08-03 10:30:46 -07001180{
Tejun Heod565ed62013-01-24 11:01:33 -08001181 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001182 struct pool_workqueue *pwq;
Tejun Heobf4ede02012-08-03 10:30:46 -07001183
Tejun Heobbb68df2012-08-03 10:30:46 -07001184 local_irq_save(*flags);
1185
Tejun Heo36e227d2012-08-03 10:30:46 -07001186 /* try to steal the timer if it exists */
1187 if (is_dwork) {
1188 struct delayed_work *dwork = to_delayed_work(work);
1189
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001190 /*
1191 * dwork->timer is irqsafe. If del_timer() fails, it's
1192 * guaranteed that the timer is not queued anywhere and not
1193 * running on the local CPU.
1194 */
Tejun Heo36e227d2012-08-03 10:30:46 -07001195 if (likely(del_timer(&dwork->timer)))
1196 return 1;
1197 }
1198
1199 /* try to claim PENDING the normal way */
Tejun Heobf4ede02012-08-03 10:30:46 -07001200 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1201 return 0;
1202
1203 /*
1204 * The queueing is in progress, or it is already queued. Try to
1205 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1206 */
Tejun Heod565ed62013-01-24 11:01:33 -08001207 pool = get_work_pool(work);
1208 if (!pool)
Tejun Heobbb68df2012-08-03 10:30:46 -07001209 goto fail;
Tejun Heobf4ede02012-08-03 10:30:46 -07001210
Tejun Heod565ed62013-01-24 11:01:33 -08001211 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001212 /*
Tejun Heo112202d2013-02-13 19:29:12 -08001213 * work->data is guaranteed to point to pwq only while the work
1214 * item is queued on pwq->wq, and both updating work->data to point
1215 * to pwq on queueing and to pool on dequeueing are done under
1216 * pwq->pool->lock. This in turn guarantees that, if work->data
1217 * points to pwq which is associated with a locked pool, the work
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08001218 * item is currently queued on that pool.
1219 */
Tejun Heo112202d2013-02-13 19:29:12 -08001220 pwq = get_work_pwq(work);
1221 if (pwq && pwq->pool == pool) {
Tejun Heo16062832013-02-06 18:04:53 -08001222 debug_work_deactivate(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001223
Tejun Heo16062832013-02-06 18:04:53 -08001224 /*
1225 * A delayed work item cannot be grabbed directly because
1226 * it might have linked NO_COLOR work items which, if left
Tejun Heo112202d2013-02-13 19:29:12 -08001227 * on the delayed_list, will confuse pwq->nr_active
Tejun Heo16062832013-02-06 18:04:53 -08001228 * management later on and cause stall. Make sure the work
1229 * item is activated before grabbing.
1230 */
1231 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
Tejun Heo112202d2013-02-13 19:29:12 -08001232 pwq_activate_delayed_work(work);
Lai Jiangshan3aa62492012-09-18 10:40:00 -07001233
Tejun Heo16062832013-02-06 18:04:53 -08001234 list_del_init(&work->entry);
Lai Jiangshan9c34a702014-07-11 00:11:13 +08001235 pwq_dec_nr_in_flight(pwq, get_work_color(work));
Tejun Heo36e227d2012-08-03 10:30:46 -07001236
Tejun Heo112202d2013-02-13 19:29:12 -08001237 /* work->data points to pwq iff queued, point to pool */
Tejun Heo16062832013-02-06 18:04:53 -08001238 set_work_pool_and_keep_pending(work, pool->id);
Lai Jiangshan4468a002013-02-06 18:04:53 -08001239
Tejun Heo16062832013-02-06 18:04:53 -08001240 spin_unlock(&pool->lock);
1241 return 1;
Tejun Heobf4ede02012-08-03 10:30:46 -07001242 }
Tejun Heod565ed62013-01-24 11:01:33 -08001243 spin_unlock(&pool->lock);
Tejun Heobbb68df2012-08-03 10:30:46 -07001244fail:
1245 local_irq_restore(*flags);
1246 if (work_is_canceling(work))
1247 return -ENOENT;
1248 cpu_relax();
Tejun Heo36e227d2012-08-03 10:30:46 -07001249 return -EAGAIN;
Tejun Heobf4ede02012-08-03 10:30:46 -07001250}
1251
1252/**
Tejun Heo706026c2013-01-24 11:01:34 -08001253 * insert_work - insert a work into a pool
Tejun Heo112202d2013-02-13 19:29:12 -08001254 * @pwq: pwq @work belongs to
Tejun Heo4690c4a2010-06-29 10:07:10 +02001255 * @work: work to insert
1256 * @head: insertion point
1257 * @extra_flags: extra WORK_STRUCT_* flags to set
1258 *
Tejun Heo112202d2013-02-13 19:29:12 -08001259 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
Tejun Heo706026c2013-01-24 11:01:34 -08001260 * work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +02001261 *
1262 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001263 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02001264 */
Tejun Heo112202d2013-02-13 19:29:12 -08001265static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1266 struct list_head *head, unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001267{
Tejun Heo112202d2013-02-13 19:29:12 -08001268 struct worker_pool *pool = pwq->pool;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01001269
Tejun Heo4690c4a2010-06-29 10:07:10 +02001270 /* we own @work, set data and link */
Tejun Heo112202d2013-02-13 19:29:12 -08001271 set_work_pwq(work, pwq, extra_flags);
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -07001272 list_add_tail(&work->entry, head);
Tejun Heo8864b4e2013-03-12 11:30:04 -07001273 get_pwq(pwq);
Tejun Heoe22bee72010-06-29 10:07:14 +02001274
1275 /*
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001276 * Ensure either wq_worker_sleeping() sees the above
1277 * list_add_tail() or we see zero nr_running to avoid workers lying
1278 * around lazily while there are works to be processed.
Tejun Heoe22bee72010-06-29 10:07:14 +02001279 */
1280 smp_mb();
1281
Tejun Heo63d95a92012-07-12 14:46:37 -07001282 if (__need_more_worker(pool))
1283 wake_up_worker(pool);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07001284}
1285
Tejun Heoc8efcc22010-12-20 19:32:04 +01001286/*
1287 * Test whether @work is being queued from another work executing on the
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001288 * same workqueue.
Tejun Heoc8efcc22010-12-20 19:32:04 +01001289 */
1290static bool is_chained_work(struct workqueue_struct *wq)
1291{
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001292 struct worker *worker;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001293
Tejun Heo8d03ecf2013-02-13 19:29:10 -08001294 worker = current_wq_worker();
1295 /*
1296 * Return %true iff I'm a worker execuing a work item on @wq. If
1297 * I'm @worker, it's safe to dereference it without locking.
1298 */
Tejun Heo112202d2013-02-13 19:29:12 -08001299 return worker && worker->current_pwq->wq == wq;
Tejun Heoc8efcc22010-12-20 19:32:04 +01001300}
1301
Tejun Heod84ff052013-03-12 11:29:59 -07001302static void __queue_work(int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303 struct work_struct *work)
1304{
Tejun Heo112202d2013-02-13 19:29:12 -08001305 struct pool_workqueue *pwq;
Tejun Heoc9178082013-03-12 11:30:04 -07001306 struct worker_pool *last_pool;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001307 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001308 unsigned int work_flags;
Joonsoo Kimb75cac92012-08-15 23:25:37 +09001309 unsigned int req_cpu = cpu;
Tejun Heo8930cab2012-08-03 10:30:45 -07001310
1311 /*
1312 * While a work item is PENDING && off queue, a task trying to
1313 * steal the PENDING will busy-loop waiting for it to either get
1314 * queued or lose PENDING. Grabbing PENDING and queueing should
1315 * happen with IRQ disabled.
1316 */
1317 WARN_ON_ONCE(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09001319 debug_work_activate(work);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001320
Li Bin9ef28a72013-09-09 13:13:58 +08001321 /* if draining, only works from the same workqueue are allowed */
Tejun Heo618b01e2013-03-12 11:30:04 -07001322 if (unlikely(wq->flags & __WQ_DRAINING) &&
Tejun Heoc8efcc22010-12-20 19:32:04 +01001323 WARN_ON_ONCE(!is_chained_work(wq)))
Tejun Heoe41e7042010-08-24 14:22:47 +02001324 return;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001325retry:
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001326 if (req_cpu == WORK_CPU_UNBOUND)
1327 cpu = raw_smp_processor_id();
1328
Tejun Heoc9178082013-03-12 11:30:04 -07001329 /* pwq which will be used unless @work is executing elsewhere */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001330 if (!(wq->flags & WQ_UNBOUND))
Tejun Heo7fb98ea2013-03-12 11:30:00 -07001331 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001332 else
1333 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodbf25762012-08-20 14:51:23 -07001334
Tejun Heoc9178082013-03-12 11:30:04 -07001335 /*
1336 * If @work was previously on a different pool, it might still be
1337 * running there, in which case the work needs to be queued on that
1338 * pool to guarantee non-reentrancy.
1339 */
1340 last_pool = get_work_pool(work);
1341 if (last_pool && last_pool != pwq->pool) {
1342 struct worker *worker;
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001343
Tejun Heoc9178082013-03-12 11:30:04 -07001344 spin_lock(&last_pool->lock);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001345
Tejun Heoc9178082013-03-12 11:30:04 -07001346 worker = find_worker_executing_work(last_pool, work);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001347
Tejun Heoc9178082013-03-12 11:30:04 -07001348 if (worker && worker->current_pwq->wq == wq) {
1349 pwq = worker->current_pwq;
Tejun Heo8930cab2012-08-03 10:30:45 -07001350 } else {
Tejun Heoc9178082013-03-12 11:30:04 -07001351 /* meh... not running there, queue here */
1352 spin_unlock(&last_pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08001353 spin_lock(&pwq->pool->lock);
Tejun Heo8930cab2012-08-03 10:30:45 -07001354 }
Tejun Heof3421792010-07-02 10:03:51 +02001355 } else {
Tejun Heo112202d2013-02-13 19:29:12 -08001356 spin_lock(&pwq->pool->lock);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001357 }
1358
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001359 /*
1360 * pwq is determined and locked. For unbound pools, we could have
1361 * raced with pwq release and it could already be dead. If its
1362 * refcnt is zero, repeat pwq selection. Note that pwqs never die
Tejun Heodf2d5ae2013-04-01 11:23:35 -07001363 * without another pwq replacing it in the numa_pwq_tbl or while
1364 * work items are executing on it, so the retrying is guaranteed to
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07001365 * make forward-progress.
1366 */
1367 if (unlikely(!pwq->refcnt)) {
1368 if (wq->flags & WQ_UNBOUND) {
1369 spin_unlock(&pwq->pool->lock);
1370 cpu_relax();
1371 goto retry;
1372 }
1373 /* oops */
1374 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1375 wq->name, cpu);
1376 }
1377
Tejun Heo112202d2013-02-13 19:29:12 -08001378 /* pwq determined, queue */
1379 trace_workqueue_queue_work(req_cpu, pwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001380
Dan Carpenterf5b25522012-04-13 22:06:58 +03001381 if (WARN_ON(!list_empty(&work->entry))) {
Tejun Heo112202d2013-02-13 19:29:12 -08001382 spin_unlock(&pwq->pool->lock);
Dan Carpenterf5b25522012-04-13 22:06:58 +03001383 return;
1384 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001385
Tejun Heo112202d2013-02-13 19:29:12 -08001386 pwq->nr_in_flight[pwq->work_color]++;
1387 work_flags = work_color_to_flags(pwq->work_color);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001388
Tejun Heo112202d2013-02-13 19:29:12 -08001389 if (likely(pwq->nr_active < pwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001390 trace_workqueue_activate_work(work);
Tejun Heo112202d2013-02-13 19:29:12 -08001391 pwq->nr_active++;
1392 worklist = &pwq->pool->worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001393 } else {
1394 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo112202d2013-02-13 19:29:12 -08001395 worklist = &pwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001396 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001397
Tejun Heo112202d2013-02-13 19:29:12 -08001398 insert_work(pwq, work, worklist, work_flags);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001399
Tejun Heo112202d2013-02-13 19:29:12 -08001400 spin_unlock(&pwq->pool->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401}
1402
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001403/**
Zhang Ruic1a220e2008-07-23 21:28:39 -07001404 * queue_work_on - queue work on specific cpu
1405 * @cpu: CPU number to execute work on
1406 * @wq: workqueue to use
1407 * @work: work to queue
1408 *
Zhang Ruic1a220e2008-07-23 21:28:39 -07001409 * We queue the work to a specific CPU, the caller must ensure it
1410 * can't go away.
Yacine Belkadid185af32013-07-31 14:59:24 -07001411 *
1412 * Return: %false if @work was already on a queue, %true otherwise.
Zhang Ruic1a220e2008-07-23 21:28:39 -07001413 */
Tejun Heod4283e92012-08-03 10:30:44 -07001414bool queue_work_on(int cpu, struct workqueue_struct *wq,
1415 struct work_struct *work)
Zhang Ruic1a220e2008-07-23 21:28:39 -07001416{
Tejun Heod4283e92012-08-03 10:30:44 -07001417 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001418 unsigned long flags;
1419
1420 local_irq_save(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001421
Tejun Heo22df02b2010-06-29 10:07:10 +02001422 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001423 __queue_work(cpu, wq, work);
Tejun Heod4283e92012-08-03 10:30:44 -07001424 ret = true;
Zhang Ruic1a220e2008-07-23 21:28:39 -07001425 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001426
1427 local_irq_restore(flags);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001428 return ret;
1429}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001430EXPORT_SYMBOL(queue_work_on);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001431
Tejun Heod8e794d2012-08-03 10:30:45 -07001432void delayed_work_timer_fn(unsigned long __data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433{
David Howells52bad642006-11-22 14:54:01 +00001434 struct delayed_work *dwork = (struct delayed_work *)__data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001436 /* should have been called from irqsafe timer with irq already off */
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001437 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438}
Konstantin Khlebnikov1438ade52013-01-24 16:36:31 +04001439EXPORT_SYMBOL(delayed_work_timer_fn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001441static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1442 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443{
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001444 struct timer_list *timer = &dwork->timer;
1445 struct work_struct *work = &dwork->work;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001447 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1448 timer->data != (unsigned long)dwork);
Tejun Heofc4b5142012-12-04 07:40:39 -08001449 WARN_ON_ONCE(timer_pending(timer));
1450 WARN_ON_ONCE(!list_empty(&work->entry));
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001451
Tejun Heo8852aac2012-12-01 16:23:42 -08001452 /*
1453 * If @delay is 0, queue @dwork->work immediately. This is for
1454 * both optimization and correctness. The earliest @timer can
1455 * expire is on the closest next tick and delayed_work users depend
1456 * on that there's no such delay when @delay is 0.
1457 */
1458 if (!delay) {
1459 __queue_work(cpu, wq, &dwork->work);
1460 return;
1461 }
1462
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001463 timer_stats_timer_set_start_info(&dwork->timer);
1464
Lai Jiangshan60c057b2013-02-06 18:04:53 -08001465 dwork->wq = wq;
Tejun Heo12650572012-08-08 09:38:42 -07001466 dwork->cpu = cpu;
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001467 timer->expires = jiffies + delay;
1468
1469 if (unlikely(cpu != WORK_CPU_UNBOUND))
1470 add_timer_on(timer, cpu);
1471 else
1472 add_timer(timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473}
1474
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001475/**
1476 * queue_delayed_work_on - queue work on specific CPU after delay
1477 * @cpu: CPU number to execute work on
1478 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001479 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001480 * @delay: number of jiffies to wait before queueing
1481 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001482 * Return: %false if @work was already on a queue, %true otherwise. If
Tejun Heo715f1302012-08-03 10:30:46 -07001483 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1484 * execution.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001485 */
Tejun Heod4283e92012-08-03 10:30:44 -07001486bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1487 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001488{
David Howells52bad642006-11-22 14:54:01 +00001489 struct work_struct *work = &dwork->work;
Tejun Heod4283e92012-08-03 10:30:44 -07001490 bool ret = false;
Tejun Heo8930cab2012-08-03 10:30:45 -07001491 unsigned long flags;
1492
1493 /* read the comment in __queue_work() */
1494 local_irq_save(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001495
Tejun Heo22df02b2010-06-29 10:07:10 +02001496 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo7beb2ed2012-08-03 10:30:46 -07001497 __queue_delayed_work(cpu, wq, dwork, delay);
Tejun Heod4283e92012-08-03 10:30:44 -07001498 ret = true;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001499 }
Tejun Heo8930cab2012-08-03 10:30:45 -07001500
1501 local_irq_restore(flags);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001502 return ret;
1503}
Marc Dionnead7b1f82013-05-06 17:44:55 -04001504EXPORT_SYMBOL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
Tejun Heoc8e55f32010-06-29 10:07:12 +02001506/**
Tejun Heo8376fe22012-08-03 10:30:47 -07001507 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1508 * @cpu: CPU number to execute work on
1509 * @wq: workqueue to use
1510 * @dwork: work to queue
1511 * @delay: number of jiffies to wait before queueing
1512 *
1513 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1514 * modify @dwork's timer so that it expires after @delay. If @delay is
1515 * zero, @work is guaranteed to be scheduled immediately regardless of its
1516 * current state.
1517 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001518 * Return: %false if @dwork was idle and queued, %true if @dwork was
Tejun Heo8376fe22012-08-03 10:30:47 -07001519 * pending and its timer was modified.
1520 *
Tejun Heoe0aecdd2012-08-21 13:18:24 -07001521 * This function is safe to call from any context including IRQ handler.
Tejun Heo8376fe22012-08-03 10:30:47 -07001522 * See try_to_grab_pending() for details.
1523 */
1524bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1525 struct delayed_work *dwork, unsigned long delay)
1526{
1527 unsigned long flags;
1528 int ret;
1529
1530 do {
1531 ret = try_to_grab_pending(&dwork->work, true, &flags);
1532 } while (unlikely(ret == -EAGAIN));
1533
1534 if (likely(ret >= 0)) {
1535 __queue_delayed_work(cpu, wq, dwork, delay);
1536 local_irq_restore(flags);
1537 }
1538
1539 /* -ENOENT from try_to_grab_pending() becomes %true */
1540 return ret;
1541}
1542EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1543
1544/**
Tejun Heoc8e55f32010-06-29 10:07:12 +02001545 * worker_enter_idle - enter idle state
1546 * @worker: worker which is entering idle state
1547 *
1548 * @worker is entering idle state. Update stats and idle timer if
1549 * necessary.
1550 *
1551 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001552 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001553 */
1554static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001556 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557
Tejun Heo6183c002013-03-12 11:29:57 -07001558 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1559 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1560 (worker->hentry.next || worker->hentry.pprev)))
1561 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562
Lai Jiangshan051e1852014-07-22 13:03:02 +08001563 /* can't use worker_set_flags(), also called from create_worker() */
Tejun Heocb444762010-07-02 10:03:50 +02001564 worker->flags |= WORKER_IDLE;
Tejun Heobd7bdd42012-07-12 14:46:37 -07001565 pool->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001566 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001567
Tejun Heoc8e55f32010-06-29 10:07:12 +02001568 /* idle_list is LIFO */
Tejun Heobd7bdd42012-07-12 14:46:37 -07001569 list_add(&worker->entry, &pool->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001570
Tejun Heo628c78e2012-07-17 12:39:27 -07001571 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1572 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
Tejun Heocb444762010-07-02 10:03:50 +02001573
Tejun Heo544ecf32012-05-14 15:04:50 -07001574 /*
Tejun Heo706026c2013-01-24 11:01:34 -08001575 * Sanity check nr_running. Because wq_unbind_fn() releases
Tejun Heod565ed62013-01-24 11:01:33 -08001576 * pool->lock between setting %WORKER_UNBOUND and zapping
Tejun Heo628c78e2012-07-17 12:39:27 -07001577 * nr_running, the warning may trigger spuriously. Check iff
1578 * unbind is not in progress.
Tejun Heo544ecf32012-05-14 15:04:50 -07001579 */
Tejun Heo24647572013-01-24 11:01:33 -08001580 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heobd7bdd42012-07-12 14:46:37 -07001581 pool->nr_workers == pool->nr_idle &&
Tejun Heoe19e3972013-01-24 11:39:44 -08001582 atomic_read(&pool->nr_running));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583}
1584
Tejun Heoc8e55f32010-06-29 10:07:12 +02001585/**
1586 * worker_leave_idle - leave idle state
1587 * @worker: worker which is leaving idle state
1588 *
1589 * @worker is leaving idle state. Update stats.
1590 *
1591 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001592 * spin_lock_irq(pool->lock).
Tejun Heoc8e55f32010-06-29 10:07:12 +02001593 */
1594static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001596 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597
Tejun Heo6183c002013-03-12 11:29:57 -07001598 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1599 return;
Tejun Heod302f012010-06-29 10:07:13 +02001600 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001601 pool->nr_idle--;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001602 list_del_init(&worker->entry);
1603}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001605static struct worker *alloc_worker(int node)
Tejun Heoc34056a2010-06-29 10:07:11 +02001606{
1607 struct worker *worker;
1608
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001609 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001610 if (worker) {
1611 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001612 INIT_LIST_HEAD(&worker->scheduled);
Lai Jiangshanda028462014-05-20 17:46:31 +08001613 INIT_LIST_HEAD(&worker->node);
Tejun Heoe22bee72010-06-29 10:07:14 +02001614 /* on creation a worker is in !idle && prep state */
1615 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001616 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001617 return worker;
1618}
1619
1620/**
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001621 * worker_attach_to_pool() - attach a worker to a pool
1622 * @worker: worker to be attached
1623 * @pool: the target pool
1624 *
1625 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1626 * cpu-binding of @worker are kept coordinated with the pool across
1627 * cpu-[un]hotplugs.
1628 */
1629static void worker_attach_to_pool(struct worker *worker,
1630 struct worker_pool *pool)
1631{
1632 mutex_lock(&pool->attach_mutex);
1633
1634 /*
1635 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1636 * online CPUs. It'll be re-applied when any of the CPUs come up.
1637 */
1638 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1639
1640 /*
1641 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1642 * stable across this function. See the comments above the
1643 * flag definition for details.
1644 */
1645 if (pool->flags & POOL_DISASSOCIATED)
1646 worker->flags |= WORKER_UNBOUND;
1647
1648 list_add_tail(&worker->node, &pool->workers);
1649
1650 mutex_unlock(&pool->attach_mutex);
1651}
1652
1653/**
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001654 * worker_detach_from_pool() - detach a worker from its pool
1655 * @worker: worker which is attached to its pool
1656 * @pool: the pool @worker is attached to
1657 *
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001658 * Undo the attaching which had been done in worker_attach_to_pool(). The
1659 * caller worker shouldn't access to the pool after detached except it has
1660 * other reference to the pool.
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001661 */
1662static void worker_detach_from_pool(struct worker *worker,
1663 struct worker_pool *pool)
1664{
1665 struct completion *detach_completion = NULL;
1666
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001667 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08001668 list_del(&worker->node);
1669 if (list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001670 detach_completion = pool->detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08001671 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001672
Lai Jiangshanb62c0752014-06-03 15:32:52 +08001673 /* clear leftover flags without pool->lock after it is detached */
1674 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1675
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001676 if (detach_completion)
1677 complete(detach_completion);
1678}
1679
1680/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001681 * create_worker - create a new workqueue worker
Tejun Heo63d95a92012-07-12 14:46:37 -07001682 * @pool: pool the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001683 *
Lai Jiangshan051e1852014-07-22 13:03:02 +08001684 * Create and start a new worker which is attached to @pool.
Tejun Heoc34056a2010-06-29 10:07:11 +02001685 *
1686 * CONTEXT:
1687 * Might sleep. Does GFP_KERNEL allocations.
1688 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001689 * Return:
Tejun Heoc34056a2010-06-29 10:07:11 +02001690 * Pointer to the newly created worker.
1691 */
Tejun Heobc2ae0f2012-07-17 12:39:27 -07001692static struct worker *create_worker(struct worker_pool *pool)
Tejun Heoc34056a2010-06-29 10:07:11 +02001693{
Tejun Heoc34056a2010-06-29 10:07:11 +02001694 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001695 int id = -1;
Tejun Heoe3c916a2013-04-01 11:23:32 -07001696 char id_buf[16];
Tejun Heoc34056a2010-06-29 10:07:11 +02001697
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001698 /* ID is needed to determine kthread name */
1699 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
Tejun Heo822d8402013-03-19 13:45:21 -07001700 if (id < 0)
1701 goto fail;
Tejun Heoc34056a2010-06-29 10:07:11 +02001702
Lai Jiangshanf7537df2014-07-15 17:24:15 +08001703 worker = alloc_worker(pool->node);
Tejun Heoc34056a2010-06-29 10:07:11 +02001704 if (!worker)
1705 goto fail;
1706
Tejun Heobd7bdd42012-07-12 14:46:37 -07001707 worker->pool = pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001708 worker->id = id;
1709
Tejun Heo29c91e92013-03-12 11:30:03 -07001710 if (pool->cpu >= 0)
Tejun Heoe3c916a2013-04-01 11:23:32 -07001711 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1712 pool->attrs->nice < 0 ? "H" : "");
Tejun Heof3421792010-07-02 10:03:51 +02001713 else
Tejun Heoe3c916a2013-04-01 11:23:32 -07001714 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1715
Tejun Heof3f90ad2013-04-01 11:23:34 -07001716 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
Tejun Heoe3c916a2013-04-01 11:23:32 -07001717 "kworker/%s", id_buf);
Tejun Heoc34056a2010-06-29 10:07:11 +02001718 if (IS_ERR(worker->task))
1719 goto fail;
1720
Oleg Nesterov91151222013-11-14 12:56:18 +01001721 set_user_nice(worker->task, pool->attrs->nice);
1722
1723 /* prevent userland from meddling with cpumask of workqueue workers */
1724 worker->task->flags |= PF_NO_SETAFFINITY;
1725
Lai Jiangshanda028462014-05-20 17:46:31 +08001726 /* successful, attach the worker to the pool */
Lai Jiangshan4736cbf2014-05-20 17:46:35 +08001727 worker_attach_to_pool(worker, pool);
Tejun Heo822d8402013-03-19 13:45:21 -07001728
Lai Jiangshan051e1852014-07-22 13:03:02 +08001729 /* start the newly created worker */
1730 spin_lock_irq(&pool->lock);
1731 worker->pool->nr_workers++;
1732 worker_enter_idle(worker);
1733 wake_up_process(worker->task);
1734 spin_unlock_irq(&pool->lock);
1735
Tejun Heoc34056a2010-06-29 10:07:11 +02001736 return worker;
Tejun Heo822d8402013-03-19 13:45:21 -07001737
Tejun Heoc34056a2010-06-29 10:07:11 +02001738fail:
Lai Jiangshan9625ab12014-05-20 17:46:27 +08001739 if (id >= 0)
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08001740 ida_simple_remove(&pool->worker_ida, id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001741 kfree(worker);
1742 return NULL;
1743}
1744
1745/**
Tejun Heoc34056a2010-06-29 10:07:11 +02001746 * destroy_worker - destroy a workqueue worker
1747 * @worker: worker to be destroyed
1748 *
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001749 * Destroy @worker and adjust @pool stats accordingly. The worker should
1750 * be idle.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001751 *
1752 * CONTEXT:
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001753 * spin_lock_irq(pool->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001754 */
1755static void destroy_worker(struct worker *worker)
1756{
Tejun Heobd7bdd42012-07-12 14:46:37 -07001757 struct worker_pool *pool = worker->pool;
Tejun Heoc34056a2010-06-29 10:07:11 +02001758
Tejun Heocd549682013-03-13 19:47:39 -07001759 lockdep_assert_held(&pool->lock);
1760
Tejun Heoc34056a2010-06-29 10:07:11 +02001761 /* sanity check frenzy */
Tejun Heo6183c002013-03-12 11:29:57 -07001762 if (WARN_ON(worker->current_work) ||
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001763 WARN_ON(!list_empty(&worker->scheduled)) ||
1764 WARN_ON(!(worker->flags & WORKER_IDLE)))
Tejun Heo6183c002013-03-12 11:29:57 -07001765 return;
Tejun Heoc34056a2010-06-29 10:07:11 +02001766
Lai Jiangshan73eb7fe2014-05-20 17:46:28 +08001767 pool->nr_workers--;
1768 pool->nr_idle--;
Lai Jiangshan5bdfff92014-02-15 22:02:28 +08001769
Tejun Heoc8e55f32010-06-29 10:07:12 +02001770 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001771 worker->flags |= WORKER_DIE;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08001772 wake_up_process(worker->task);
Tejun Heoc34056a2010-06-29 10:07:11 +02001773}
1774
Tejun Heo63d95a92012-07-12 14:46:37 -07001775static void idle_worker_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001776{
Tejun Heo63d95a92012-07-12 14:46:37 -07001777 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001778
Tejun Heod565ed62013-01-24 11:01:33 -08001779 spin_lock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001780
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001781 while (too_many_workers(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001782 struct worker *worker;
1783 unsigned long expires;
1784
1785 /* idle_list is kept in LIFO order, check the last one */
Tejun Heo63d95a92012-07-12 14:46:37 -07001786 worker = list_entry(pool->idle_list.prev, struct worker, entry);
Tejun Heoe22bee72010-06-29 10:07:14 +02001787 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1788
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001789 if (time_before(jiffies, expires)) {
Tejun Heo63d95a92012-07-12 14:46:37 -07001790 mod_timer(&pool->idle_timer, expires);
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001791 break;
Tejun Heoe22bee72010-06-29 10:07:14 +02001792 }
Lai Jiangshan3347fc92014-05-20 17:46:30 +08001793
1794 destroy_worker(worker);
Tejun Heoe22bee72010-06-29 10:07:14 +02001795 }
1796
Tejun Heod565ed62013-01-24 11:01:33 -08001797 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001798}
1799
Tejun Heo493a1722013-03-12 11:29:59 -07001800static void send_mayday(struct work_struct *work)
Tejun Heoe22bee72010-06-29 10:07:14 +02001801{
Tejun Heo112202d2013-02-13 19:29:12 -08001802 struct pool_workqueue *pwq = get_work_pwq(work);
1803 struct workqueue_struct *wq = pwq->wq;
Tejun Heo493a1722013-03-12 11:29:59 -07001804
Tejun Heo2e109a22013-03-13 19:47:40 -07001805 lockdep_assert_held(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001806
Tejun Heo493008a2013-03-12 11:30:03 -07001807 if (!wq->rescuer)
Tejun Heo493a1722013-03-12 11:29:59 -07001808 return;
Tejun Heoe22bee72010-06-29 10:07:14 +02001809
1810 /* mayday mayday mayday */
Tejun Heo493a1722013-03-12 11:29:59 -07001811 if (list_empty(&pwq->mayday_node)) {
Lai Jiangshan77668c82014-04-18 11:04:16 -04001812 /*
1813 * If @pwq is for an unbound wq, its base ref may be put at
1814 * any time due to an attribute change. Pin @pwq until the
1815 * rescuer is done with it.
1816 */
1817 get_pwq(pwq);
Tejun Heo493a1722013-03-12 11:29:59 -07001818 list_add_tail(&pwq->mayday_node, &wq->maydays);
Tejun Heoe22bee72010-06-29 10:07:14 +02001819 wake_up_process(wq->rescuer->task);
Tejun Heo493a1722013-03-12 11:29:59 -07001820 }
Tejun Heoe22bee72010-06-29 10:07:14 +02001821}
1822
Tejun Heo706026c2013-01-24 11:01:34 -08001823static void pool_mayday_timeout(unsigned long __pool)
Tejun Heoe22bee72010-06-29 10:07:14 +02001824{
Tejun Heo63d95a92012-07-12 14:46:37 -07001825 struct worker_pool *pool = (void *)__pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001826 struct work_struct *work;
1827
Tejun Heob2d82902014-12-08 12:39:16 -05001828 spin_lock_irq(&pool->lock);
1829 spin_lock(&wq_mayday_lock); /* for wq->maydays */
Tejun Heoe22bee72010-06-29 10:07:14 +02001830
Tejun Heo63d95a92012-07-12 14:46:37 -07001831 if (need_to_create_worker(pool)) {
Tejun Heoe22bee72010-06-29 10:07:14 +02001832 /*
1833 * We've been trying to create a new worker but
1834 * haven't been successful. We might be hitting an
1835 * allocation deadlock. Send distress signals to
1836 * rescuers.
1837 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001838 list_for_each_entry(work, &pool->worklist, entry)
Tejun Heoe22bee72010-06-29 10:07:14 +02001839 send_mayday(work);
1840 }
1841
Tejun Heob2d82902014-12-08 12:39:16 -05001842 spin_unlock(&wq_mayday_lock);
1843 spin_unlock_irq(&pool->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001844
Tejun Heo63d95a92012-07-12 14:46:37 -07001845 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
Tejun Heoe22bee72010-06-29 10:07:14 +02001846}
1847
1848/**
1849 * maybe_create_worker - create a new worker if necessary
Tejun Heo63d95a92012-07-12 14:46:37 -07001850 * @pool: pool to create a new worker for
Tejun Heoe22bee72010-06-29 10:07:14 +02001851 *
Tejun Heo63d95a92012-07-12 14:46:37 -07001852 * Create a new worker for @pool if necessary. @pool is guaranteed to
Tejun Heoe22bee72010-06-29 10:07:14 +02001853 * have at least one idle worker on return from this function. If
1854 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
Tejun Heo63d95a92012-07-12 14:46:37 -07001855 * sent to all rescuers with works scheduled on @pool to resolve
Tejun Heoe22bee72010-06-29 10:07:14 +02001856 * possible allocation deadlock.
1857 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07001858 * On return, need_to_create_worker() is guaranteed to be %false and
1859 * may_start_working() %true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001860 *
1861 * LOCKING:
Tejun Heod565ed62013-01-24 11:01:33 -08001862 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001863 * multiple times. Does GFP_KERNEL allocations. Called only from
1864 * manager.
Tejun Heoe22bee72010-06-29 10:07:14 +02001865 */
Tejun Heo29187a92015-01-16 14:21:16 -05001866static void maybe_create_worker(struct worker_pool *pool)
Tejun Heod565ed62013-01-24 11:01:33 -08001867__releases(&pool->lock)
1868__acquires(&pool->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001869{
Tejun Heoe22bee72010-06-29 10:07:14 +02001870restart:
Tejun Heod565ed62013-01-24 11:01:33 -08001871 spin_unlock_irq(&pool->lock);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001872
Tejun Heoe22bee72010-06-29 10:07:14 +02001873 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
Tejun Heo63d95a92012-07-12 14:46:37 -07001874 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
Tejun Heoe22bee72010-06-29 10:07:14 +02001875
1876 while (true) {
Lai Jiangshan051e1852014-07-22 13:03:02 +08001877 if (create_worker(pool) || !need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001878 break;
1879
Lai Jiangshane212f362014-06-03 15:32:17 +08001880 schedule_timeout_interruptible(CREATE_COOLDOWN);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001881
Tejun Heo63d95a92012-07-12 14:46:37 -07001882 if (!need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001883 break;
1884 }
1885
Tejun Heo63d95a92012-07-12 14:46:37 -07001886 del_timer_sync(&pool->mayday_timer);
Tejun Heod565ed62013-01-24 11:01:33 -08001887 spin_lock_irq(&pool->lock);
Lai Jiangshan051e1852014-07-22 13:03:02 +08001888 /*
1889 * This is necessary even after a new worker was just successfully
1890 * created as @pool->lock was dropped and the new worker might have
1891 * already become busy.
1892 */
Tejun Heo63d95a92012-07-12 14:46:37 -07001893 if (need_to_create_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02001894 goto restart;
Tejun Heoe22bee72010-06-29 10:07:14 +02001895}
1896
1897/**
Tejun Heoe22bee72010-06-29 10:07:14 +02001898 * manage_workers - manage worker pool
1899 * @worker: self
1900 *
Tejun Heo706026c2013-01-24 11:01:34 -08001901 * Assume the manager role and manage the worker pool @worker belongs
Tejun Heoe22bee72010-06-29 10:07:14 +02001902 * to. At any given time, there can be only zero or one manager per
Tejun Heo706026c2013-01-24 11:01:34 -08001903 * pool. The exclusion is handled automatically by this function.
Tejun Heoe22bee72010-06-29 10:07:14 +02001904 *
1905 * The caller can safely start processing works on false return. On
1906 * true return, it's guaranteed that need_to_create_worker() is false
1907 * and may_start_working() is true.
1908 *
1909 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001910 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoe22bee72010-06-29 10:07:14 +02001911 * multiple times. Does GFP_KERNEL allocations.
1912 *
Yacine Belkadid185af32013-07-31 14:59:24 -07001913 * Return:
Tejun Heo29187a92015-01-16 14:21:16 -05001914 * %false if the pool doesn't need management and the caller can safely
1915 * start processing works, %true if management function was performed and
1916 * the conditions that the caller verified before calling the function may
1917 * no longer be true.
Tejun Heoe22bee72010-06-29 10:07:14 +02001918 */
1919static bool manage_workers(struct worker *worker)
1920{
Tejun Heo63d95a92012-07-12 14:46:37 -07001921 struct worker_pool *pool = worker->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02001922
Tejun Heobc3a1af2013-03-13 19:47:39 -07001923 /*
Tejun Heobc3a1af2013-03-13 19:47:39 -07001924 * Anyone who successfully grabs manager_arb wins the arbitration
1925 * and becomes the manager. mutex_trylock() on pool->manager_arb
1926 * failure while holding pool->lock reliably indicates that someone
1927 * else is managing the pool and the worker which failed trylock
1928 * can proceed to executing work items. This means that anyone
1929 * grabbing manager_arb is responsible for actually performing
1930 * manager duties. If manager_arb is grabbed and released without
1931 * actual management, the pool may stall indefinitely.
Tejun Heobc3a1af2013-03-13 19:47:39 -07001932 */
Tejun Heo34a06bd2013-03-12 11:30:00 -07001933 if (!mutex_trylock(&pool->manager_arb))
Tejun Heo29187a92015-01-16 14:21:16 -05001934 return false;
Tejun Heo2607d7a2015-03-09 09:22:28 -04001935 pool->manager = worker;
Tejun Heoe22bee72010-06-29 10:07:14 +02001936
Tejun Heo29187a92015-01-16 14:21:16 -05001937 maybe_create_worker(pool);
Tejun Heoe22bee72010-06-29 10:07:14 +02001938
Tejun Heo2607d7a2015-03-09 09:22:28 -04001939 pool->manager = NULL;
Tejun Heo34a06bd2013-03-12 11:30:00 -07001940 mutex_unlock(&pool->manager_arb);
Tejun Heo29187a92015-01-16 14:21:16 -05001941 return true;
Tejun Heoe22bee72010-06-29 10:07:14 +02001942}
1943
Tejun Heoa62428c2010-06-29 10:07:10 +02001944/**
1945 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02001946 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02001947 * @work: work to process
1948 *
1949 * Process @work. This function contains all the logics necessary to
1950 * process a single work including synchronization against and
1951 * interaction with other workers on the same cpu, queueing and
1952 * flushing. As long as context requirement is met, any worker can
1953 * call this function to process a work.
1954 *
1955 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08001956 * spin_lock_irq(pool->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02001957 */
Tejun Heoc34056a2010-06-29 10:07:11 +02001958static void process_one_work(struct worker *worker, struct work_struct *work)
Tejun Heod565ed62013-01-24 11:01:33 -08001959__releases(&pool->lock)
1960__acquires(&pool->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02001961{
Tejun Heo112202d2013-02-13 19:29:12 -08001962 struct pool_workqueue *pwq = get_work_pwq(work);
Tejun Heobd7bdd42012-07-12 14:46:37 -07001963 struct worker_pool *pool = worker->pool;
Tejun Heo112202d2013-02-13 19:29:12 -08001964 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heo73f53c42010-06-29 10:07:11 +02001965 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02001966 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02001967#ifdef CONFIG_LOCKDEP
1968 /*
1969 * It is permissible to free the struct work_struct from
1970 * inside the function that is called from it, this we need to
1971 * take into account for lockdep too. To avoid bogus "held
1972 * lock freed" warnings as well as problems when looking into
1973 * work->lockdep_map, make a copy and use that here.
1974 */
Peter Zijlstra4d82a1d2012-05-15 08:06:19 -07001975 struct lockdep_map lockdep_map;
1976
1977 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02001978#endif
Lai Jiangshan807407c2014-06-03 15:33:28 +08001979 /* ensure we're on the correct CPU */
Lai Jiangshan85327af2014-06-03 15:33:28 +08001980 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
Tejun Heoec22ca52013-01-24 11:01:33 -08001981 raw_smp_processor_id() != pool->cpu);
Tejun Heo25511a42012-07-17 12:39:27 -07001982
Tejun Heo7e116292010-06-29 10:07:13 +02001983 /*
1984 * A single work shouldn't be executed concurrently by
1985 * multiple workers on a single cpu. Check whether anyone is
1986 * already processing the work. If so, defer the work to the
1987 * currently executing one.
1988 */
Tejun Heoc9e7cf22013-01-24 11:01:33 -08001989 collision = find_worker_executing_work(pool, work);
Tejun Heo7e116292010-06-29 10:07:13 +02001990 if (unlikely(collision)) {
1991 move_linked_works(work, &collision->scheduled, NULL);
1992 return;
1993 }
1994
Tejun Heo8930cab2012-08-03 10:30:45 -07001995 /* claim and dequeue */
Tejun Heoa62428c2010-06-29 10:07:10 +02001996 debug_work_deactivate(work);
Tejun Heoc9e7cf22013-01-24 11:01:33 -08001997 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
Tejun Heoc34056a2010-06-29 10:07:11 +02001998 worker->current_work = work;
Tejun Heoa2c1c572012-12-18 10:35:02 -08001999 worker->current_func = work->func;
Tejun Heo112202d2013-02-13 19:29:12 -08002000 worker->current_pwq = pwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002001 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002002
Tejun Heoa62428c2010-06-29 10:07:10 +02002003 list_del_init(&work->entry);
2004
Tejun Heo649027d2010-06-29 10:07:14 +02002005 /*
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002006 * CPU intensive works don't participate in concurrency management.
2007 * They're the scheduler's responsibility. This takes @worker out
2008 * of concurrency management and the next code block will chain
2009 * execution of the pending work items.
Tejun Heofb0e7be2010-06-29 10:07:15 +02002010 */
2011 if (unlikely(cpu_intensive))
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002012 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
Tejun Heofb0e7be2010-06-29 10:07:15 +02002013
Tejun Heo974271c2012-07-12 14:46:37 -07002014 /*
Lai Jiangshana489a032014-07-22 13:01:59 +08002015 * Wake up another worker if necessary. The condition is always
2016 * false for normal per-cpu workers since nr_running would always
2017 * be >= 1 at this point. This is used to chain execution of the
2018 * pending work items for WORKER_NOT_RUNNING workers such as the
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002019 * UNBOUND and CPU_INTENSIVE ones.
Tejun Heo974271c2012-07-12 14:46:37 -07002020 */
Lai Jiangshana489a032014-07-22 13:01:59 +08002021 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002022 wake_up_worker(pool);
Tejun Heo974271c2012-07-12 14:46:37 -07002023
Tejun Heo8930cab2012-08-03 10:30:45 -07002024 /*
Tejun Heo7c3eed52013-01-24 11:01:33 -08002025 * Record the last pool and clear PENDING which should be the last
Tejun Heod565ed62013-01-24 11:01:33 -08002026 * update to @work. Also, do this inside @pool->lock so that
Tejun Heo23657bb2012-08-13 17:08:19 -07002027 * PENDING and queued state changes happen together while IRQ is
2028 * disabled.
Tejun Heo8930cab2012-08-03 10:30:45 -07002029 */
Tejun Heo7c3eed52013-01-24 11:01:33 -08002030 set_work_pool_and_clear_pending(work, pool->id);
Tejun Heoa62428c2010-06-29 10:07:10 +02002031
Tejun Heod565ed62013-01-24 11:01:33 -08002032 spin_unlock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002033
Tejun Heo112202d2013-02-13 19:29:12 -08002034 lock_map_acquire_read(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002035 lock_map_acquire(&lockdep_map);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002036 trace_workqueue_execute_start(work);
Tejun Heoa2c1c572012-12-18 10:35:02 -08002037 worker->current_func(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07002038 /*
2039 * While we must be careful to not use "work" after this, the trace
2040 * point will only record its address.
2041 */
2042 trace_workqueue_execute_end(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02002043 lock_map_release(&lockdep_map);
Tejun Heo112202d2013-02-13 19:29:12 -08002044 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02002045
2046 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
Valentin Ilie044c7822012-08-19 00:52:42 +03002047 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2048 " last function: %pf\n",
Tejun Heoa2c1c572012-12-18 10:35:02 -08002049 current->comm, preempt_count(), task_pid_nr(current),
2050 worker->current_func);
Tejun Heoa62428c2010-06-29 10:07:10 +02002051 debug_show_held_locks(current);
2052 dump_stack();
2053 }
2054
Tejun Heob22ce272013-08-28 17:33:37 -04002055 /*
2056 * The following prevents a kworker from hogging CPU on !PREEMPT
2057 * kernels, where a requeueing work item waiting for something to
2058 * happen could deadlock with stop_machine as such work item could
2059 * indefinitely requeue itself while all other CPUs are trapped in
Joe Lawrence789cbbe2014-10-05 13:24:21 -04002060 * stop_machine. At the same time, report a quiescent RCU state so
2061 * the same condition doesn't freeze RCU.
Tejun Heob22ce272013-08-28 17:33:37 -04002062 */
Joe Lawrence3e28e372014-10-05 13:24:22 -04002063 cond_resched_rcu_qs();
Tejun Heob22ce272013-08-28 17:33:37 -04002064
Tejun Heod565ed62013-01-24 11:01:33 -08002065 spin_lock_irq(&pool->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02002066
Tejun Heofb0e7be2010-06-29 10:07:15 +02002067 /* clear cpu intensive status */
2068 if (unlikely(cpu_intensive))
2069 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2070
Tejun Heoa62428c2010-06-29 10:07:10 +02002071 /* we're done with it, release */
Sasha Levin42f85702012-12-17 10:01:23 -05002072 hash_del(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002073 worker->current_work = NULL;
Tejun Heoa2c1c572012-12-18 10:35:02 -08002074 worker->current_func = NULL;
Tejun Heo112202d2013-02-13 19:29:12 -08002075 worker->current_pwq = NULL;
Tejun Heo3d1cb202013-04-30 15:27:22 -07002076 worker->desc_valid = false;
Tejun Heo112202d2013-02-13 19:29:12 -08002077 pwq_dec_nr_in_flight(pwq, work_color);
Tejun Heoa62428c2010-06-29 10:07:10 +02002078}
2079
Tejun Heoaffee4b2010-06-29 10:07:12 +02002080/**
2081 * process_scheduled_works - process scheduled works
2082 * @worker: self
2083 *
2084 * Process all scheduled works. Please note that the scheduled list
2085 * may change while processing a work, so this function repeatedly
2086 * fetches a work from the top and executes it.
2087 *
2088 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002089 * spin_lock_irq(pool->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02002090 * multiple times.
2091 */
2092static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002094 while (!list_empty(&worker->scheduled)) {
2095 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02002097 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099}
2100
Tejun Heo4690c4a2010-06-29 10:07:10 +02002101/**
2102 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02002103 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02002104 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002105 * The worker thread function. All workers belong to a worker_pool -
2106 * either a per-cpu one or dynamic unbound one. These workers process all
2107 * work items regardless of their specific target workqueue. The only
2108 * exception is work items which belong to workqueues with a rescuer which
2109 * will be explained in rescuer_thread().
Yacine Belkadid185af32013-07-31 14:59:24 -07002110 *
2111 * Return: 0
Tejun Heo4690c4a2010-06-29 10:07:10 +02002112 */
Tejun Heoc34056a2010-06-29 10:07:11 +02002113static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114{
Tejun Heoc34056a2010-06-29 10:07:11 +02002115 struct worker *worker = __worker;
Tejun Heobd7bdd42012-07-12 14:46:37 -07002116 struct worker_pool *pool = worker->pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Tejun Heoe22bee72010-06-29 10:07:14 +02002118 /* tell the scheduler that this is a workqueue worker */
2119 worker->task->flags |= PF_WQ_WORKER;
Tejun Heoc8e55f32010-06-29 10:07:12 +02002120woke_up:
Tejun Heod565ed62013-01-24 11:01:33 -08002121 spin_lock_irq(&pool->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002122
Tejun Heoa9ab7752013-03-19 13:45:21 -07002123 /* am I supposed to die? */
2124 if (unlikely(worker->flags & WORKER_DIE)) {
Tejun Heod565ed62013-01-24 11:01:33 -08002125 spin_unlock_irq(&pool->lock);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002126 WARN_ON_ONCE(!list_empty(&worker->entry));
2127 worker->task->flags &= ~PF_WQ_WORKER;
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002128
2129 set_task_comm(worker->task, "kworker/dying");
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08002130 ida_simple_remove(&pool->worker_ida, worker->id);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08002131 worker_detach_from_pool(worker, pool);
2132 kfree(worker);
Tejun Heoa9ab7752013-03-19 13:45:21 -07002133 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07002135
Tejun Heoc8e55f32010-06-29 10:07:12 +02002136 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02002137recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02002138 /* no more worker necessary? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002139 if (!need_more_worker(pool))
Tejun Heoe22bee72010-06-29 10:07:14 +02002140 goto sleep;
2141
2142 /* do we need to manage? */
Tejun Heo63d95a92012-07-12 14:46:37 -07002143 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
Tejun Heoe22bee72010-06-29 10:07:14 +02002144 goto recheck;
2145
Tejun Heoc8e55f32010-06-29 10:07:12 +02002146 /*
2147 * ->scheduled list can only be filled while a worker is
2148 * preparing to process a work or actually processing it.
2149 * Make sure nobody diddled with it while I was sleeping.
2150 */
Tejun Heo6183c002013-03-12 11:29:57 -07002151 WARN_ON_ONCE(!list_empty(&worker->scheduled));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002152
Tejun Heoe22bee72010-06-29 10:07:14 +02002153 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07002154 * Finish PREP stage. We're guaranteed to have at least one idle
2155 * worker or that someone else has already assumed the manager
2156 * role. This is where @worker starts participating in concurrency
2157 * management if applicable and concurrency management is restored
2158 * after being rebound. See rebind_workers() for details.
Tejun Heoe22bee72010-06-29 10:07:14 +02002159 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07002160 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
Tejun Heoe22bee72010-06-29 10:07:14 +02002161
2162 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02002163 struct work_struct *work =
Tejun Heobd7bdd42012-07-12 14:46:37 -07002164 list_first_entry(&pool->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02002165 struct work_struct, entry);
2166
2167 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2168 /* optimization path, not strictly necessary */
2169 process_one_work(worker, work);
2170 if (unlikely(!list_empty(&worker->scheduled)))
2171 process_scheduled_works(worker);
2172 } else {
2173 move_linked_works(work, &worker->scheduled, NULL);
2174 process_scheduled_works(worker);
2175 }
Tejun Heo63d95a92012-07-12 14:46:37 -07002176 } while (keep_working(pool));
Tejun Heoc8e55f32010-06-29 10:07:12 +02002177
Lai Jiangshan228f1d02014-07-22 13:02:00 +08002178 worker_set_flags(worker, WORKER_PREP);
Tejun Heod313dd82010-07-02 10:03:51 +02002179sleep:
Tejun Heoc8e55f32010-06-29 10:07:12 +02002180 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002181 * pool->lock is held and there's no work to process and no need to
2182 * manage, sleep. Workers are woken up only while holding
2183 * pool->lock or from local cpu, so setting the current state
2184 * before releasing pool->lock is enough to prevent losing any
2185 * event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02002186 */
2187 worker_enter_idle(worker);
2188 __set_current_state(TASK_INTERRUPTIBLE);
Tejun Heod565ed62013-01-24 11:01:33 -08002189 spin_unlock_irq(&pool->lock);
Tejun Heoc8e55f32010-06-29 10:07:12 +02002190 schedule();
2191 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192}
2193
Tejun Heoe22bee72010-06-29 10:07:14 +02002194/**
2195 * rescuer_thread - the rescuer thread function
Tejun Heo111c2252013-01-17 17:16:24 -08002196 * @__rescuer: self
Tejun Heoe22bee72010-06-29 10:07:14 +02002197 *
2198 * Workqueue rescuer thread function. There's one rescuer for each
Tejun Heo493008a2013-03-12 11:30:03 -07002199 * workqueue which has WQ_MEM_RECLAIM set.
Tejun Heoe22bee72010-06-29 10:07:14 +02002200 *
Tejun Heo706026c2013-01-24 11:01:34 -08002201 * Regular work processing on a pool may block trying to create a new
Tejun Heoe22bee72010-06-29 10:07:14 +02002202 * worker which uses GFP_KERNEL allocation which has slight chance of
2203 * developing into deadlock if some works currently on the same queue
2204 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2205 * the problem rescuer solves.
2206 *
Tejun Heo706026c2013-01-24 11:01:34 -08002207 * When such condition is possible, the pool summons rescuers of all
2208 * workqueues which have works queued on the pool and let them process
Tejun Heoe22bee72010-06-29 10:07:14 +02002209 * those works so that forward progress can be guaranteed.
2210 *
2211 * This should happen rarely.
Yacine Belkadid185af32013-07-31 14:59:24 -07002212 *
2213 * Return: 0
Tejun Heoe22bee72010-06-29 10:07:14 +02002214 */
Tejun Heo111c2252013-01-17 17:16:24 -08002215static int rescuer_thread(void *__rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02002216{
Tejun Heo111c2252013-01-17 17:16:24 -08002217 struct worker *rescuer = __rescuer;
2218 struct workqueue_struct *wq = rescuer->rescue_wq;
Tejun Heoe22bee72010-06-29 10:07:14 +02002219 struct list_head *scheduled = &rescuer->scheduled;
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002220 bool should_stop;
Tejun Heoe22bee72010-06-29 10:07:14 +02002221
2222 set_user_nice(current, RESCUER_NICE_LEVEL);
Tejun Heo111c2252013-01-17 17:16:24 -08002223
2224 /*
2225 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2226 * doesn't participate in concurrency management.
2227 */
2228 rescuer->task->flags |= PF_WQ_WORKER;
Tejun Heoe22bee72010-06-29 10:07:14 +02002229repeat:
2230 set_current_state(TASK_INTERRUPTIBLE);
2231
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002232 /*
2233 * By the time the rescuer is requested to stop, the workqueue
2234 * shouldn't have any work pending, but @wq->maydays may still have
2235 * pwq(s) queued. This can happen by non-rescuer workers consuming
2236 * all the work items before the rescuer got to them. Go through
2237 * @wq->maydays processing before acting on should_stop so that the
2238 * list is always empty on exit.
2239 */
2240 should_stop = kthread_should_stop();
Tejun Heoe22bee72010-06-29 10:07:14 +02002241
Tejun Heo493a1722013-03-12 11:29:59 -07002242 /* see whether any pwq is asking for help */
Tejun Heo2e109a22013-03-13 19:47:40 -07002243 spin_lock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002244
2245 while (!list_empty(&wq->maydays)) {
2246 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2247 struct pool_workqueue, mayday_node);
Tejun Heo112202d2013-02-13 19:29:12 -08002248 struct worker_pool *pool = pwq->pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002249 struct work_struct *work, *n;
2250
2251 __set_current_state(TASK_RUNNING);
Tejun Heo493a1722013-03-12 11:29:59 -07002252 list_del_init(&pwq->mayday_node);
2253
Tejun Heo2e109a22013-03-13 19:47:40 -07002254 spin_unlock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002255
Lai Jiangshan51697d392014-05-20 17:46:36 +08002256 worker_attach_to_pool(rescuer, pool);
2257
2258 spin_lock_irq(&pool->lock);
Lai Jiangshanb3104102013-02-19 12:17:02 -08002259 rescuer->pool = pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02002260
2261 /*
2262 * Slurp in all works issued via this workqueue and
2263 * process'em.
2264 */
Tejun Heo0479c8c2014-12-04 10:14:13 -05002265 WARN_ON_ONCE(!list_empty(scheduled));
Tejun Heobd7bdd42012-07-12 14:46:37 -07002266 list_for_each_entry_safe(work, n, &pool->worklist, entry)
Tejun Heo112202d2013-02-13 19:29:12 -08002267 if (get_work_pwq(work) == pwq)
Tejun Heoe22bee72010-06-29 10:07:14 +02002268 move_linked_works(work, scheduled, &n);
2269
NeilBrown008847f2014-12-08 12:39:16 -05002270 if (!list_empty(scheduled)) {
2271 process_scheduled_works(rescuer);
2272
2273 /*
2274 * The above execution of rescued work items could
2275 * have created more to rescue through
2276 * pwq_activate_first_delayed() or chained
2277 * queueing. Let's put @pwq back on mayday list so
2278 * that such back-to-back work items, which may be
2279 * being used to relieve memory pressure, don't
2280 * incur MAYDAY_INTERVAL delay inbetween.
2281 */
2282 if (need_to_create_worker(pool)) {
2283 spin_lock(&wq_mayday_lock);
2284 get_pwq(pwq);
2285 list_move_tail(&pwq->mayday_node, &wq->maydays);
2286 spin_unlock(&wq_mayday_lock);
2287 }
2288 }
Tejun Heo75769582011-02-14 14:04:46 +01002289
2290 /*
Lai Jiangshan77668c82014-04-18 11:04:16 -04002291 * Put the reference grabbed by send_mayday(). @pool won't
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002292 * go away while we're still attached to it.
Lai Jiangshan77668c82014-04-18 11:04:16 -04002293 */
2294 put_pwq(pwq);
2295
2296 /*
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002297 * Leave this pool. If need_more_worker() is %true, notify a
Tejun Heo75769582011-02-14 14:04:46 +01002298 * regular worker; otherwise, we end up with 0 concurrency
2299 * and stalling the execution.
2300 */
Lai Jiangshand8ca83e2014-07-16 14:56:36 +08002301 if (need_more_worker(pool))
Tejun Heo63d95a92012-07-12 14:46:37 -07002302 wake_up_worker(pool);
Tejun Heo75769582011-02-14 14:04:46 +01002303
Lai Jiangshanb3104102013-02-19 12:17:02 -08002304 rescuer->pool = NULL;
Lai Jiangshan13b1d622014-07-22 13:03:47 +08002305 spin_unlock_irq(&pool->lock);
2306
2307 worker_detach_from_pool(rescuer, pool);
2308
2309 spin_lock_irq(&wq_mayday_lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02002310 }
2311
Tejun Heo2e109a22013-03-13 19:47:40 -07002312 spin_unlock_irq(&wq_mayday_lock);
Tejun Heo493a1722013-03-12 11:29:59 -07002313
Lai Jiangshan4d595b82014-04-18 11:04:16 -04002314 if (should_stop) {
2315 __set_current_state(TASK_RUNNING);
2316 rescuer->task->flags &= ~PF_WQ_WORKER;
2317 return 0;
2318 }
2319
Tejun Heo111c2252013-01-17 17:16:24 -08002320 /* rescuers should never participate in concurrency management */
2321 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
Tejun Heoe22bee72010-06-29 10:07:14 +02002322 schedule();
2323 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324}
2325
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002326struct wq_barrier {
2327 struct work_struct work;
2328 struct completion done;
Tejun Heo2607d7a2015-03-09 09:22:28 -04002329 struct task_struct *task; /* purely informational */
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002330};
2331
2332static void wq_barrier_func(struct work_struct *work)
2333{
2334 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2335 complete(&barr->done);
2336}
2337
Tejun Heo4690c4a2010-06-29 10:07:10 +02002338/**
2339 * insert_wq_barrier - insert a barrier work
Tejun Heo112202d2013-02-13 19:29:12 -08002340 * @pwq: pwq to insert barrier into
Tejun Heo4690c4a2010-06-29 10:07:10 +02002341 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002342 * @target: target work to attach @barr to
2343 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002344 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002345 * @barr is linked to @target such that @barr is completed only after
2346 * @target finishes execution. Please note that the ordering
2347 * guarantee is observed only with respect to @target and on the local
2348 * cpu.
2349 *
2350 * Currently, a queued barrier can't be canceled. This is because
2351 * try_to_grab_pending() can't determine whether the work to be
2352 * grabbed is at the head of the queue and thus can't clear LINKED
2353 * flag of the previous work while there must be a valid next work
2354 * after a work with LINKED flag set.
2355 *
2356 * Note that when @worker is non-NULL, @target may be modified
Tejun Heo112202d2013-02-13 19:29:12 -08002357 * underneath us, so we can't reliably determine pwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002358 *
2359 * CONTEXT:
Tejun Heod565ed62013-01-24 11:01:33 -08002360 * spin_lock_irq(pool->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002361 */
Tejun Heo112202d2013-02-13 19:29:12 -08002362static void insert_wq_barrier(struct pool_workqueue *pwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002363 struct wq_barrier *barr,
2364 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002365{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002366 struct list_head *head;
2367 unsigned int linked = 0;
2368
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002369 /*
Tejun Heod565ed62013-01-24 11:01:33 -08002370 * debugobject calls are safe here even with pool->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002371 * as we know for sure that this will not trigger any of the
2372 * checks and call back into the fixup functions where we
2373 * might deadlock.
2374 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002375 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002376 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002377 init_completion(&barr->done);
Tejun Heo2607d7a2015-03-09 09:22:28 -04002378 barr->task = current;
Oleg Nesterov83c22522007-05-09 02:33:54 -07002379
Tejun Heoaffee4b2010-06-29 10:07:12 +02002380 /*
2381 * If @target is currently being executed, schedule the
2382 * barrier to the worker; otherwise, put it after @target.
2383 */
2384 if (worker)
2385 head = worker->scheduled.next;
2386 else {
2387 unsigned long *bits = work_data_bits(target);
2388
2389 head = target->entry.next;
2390 /* there can already be other linked works, inherit and set */
2391 linked = *bits & WORK_STRUCT_LINKED;
2392 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2393 }
2394
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002395 debug_work_activate(&barr->work);
Tejun Heo112202d2013-02-13 19:29:12 -08002396 insert_work(pwq, &barr->work, head,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002397 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002398}
2399
Tejun Heo73f53c42010-06-29 10:07:11 +02002400/**
Tejun Heo112202d2013-02-13 19:29:12 -08002401 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
Tejun Heo73f53c42010-06-29 10:07:11 +02002402 * @wq: workqueue being flushed
2403 * @flush_color: new flush color, < 0 for no-op
2404 * @work_color: new work color, < 0 for no-op
2405 *
Tejun Heo112202d2013-02-13 19:29:12 -08002406 * Prepare pwqs for workqueue flushing.
Tejun Heo73f53c42010-06-29 10:07:11 +02002407 *
Tejun Heo112202d2013-02-13 19:29:12 -08002408 * If @flush_color is non-negative, flush_color on all pwqs should be
2409 * -1. If no pwq has in-flight commands at the specified color, all
2410 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2411 * has in flight commands, its pwq->flush_color is set to
2412 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
Tejun Heo73f53c42010-06-29 10:07:11 +02002413 * wakeup logic is armed and %true is returned.
2414 *
2415 * The caller should have initialized @wq->first_flusher prior to
2416 * calling this function with non-negative @flush_color. If
2417 * @flush_color is negative, no flush color update is done and %false
2418 * is returned.
2419 *
Tejun Heo112202d2013-02-13 19:29:12 -08002420 * If @work_color is non-negative, all pwqs should have the same
Tejun Heo73f53c42010-06-29 10:07:11 +02002421 * work_color which is previous to @work_color and all will be
2422 * advanced to @work_color.
2423 *
2424 * CONTEXT:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002425 * mutex_lock(wq->mutex).
Tejun Heo73f53c42010-06-29 10:07:11 +02002426 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002427 * Return:
Tejun Heo73f53c42010-06-29 10:07:11 +02002428 * %true if @flush_color >= 0 and there's something to flush. %false
2429 * otherwise.
2430 */
Tejun Heo112202d2013-02-13 19:29:12 -08002431static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
Tejun Heo73f53c42010-06-29 10:07:11 +02002432 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433{
Tejun Heo73f53c42010-06-29 10:07:11 +02002434 bool wait = false;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002435 struct pool_workqueue *pwq;
Oleg Nesterov14441962007-05-23 13:57:57 -07002436
Tejun Heo73f53c42010-06-29 10:07:11 +02002437 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002438 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
Tejun Heo112202d2013-02-13 19:29:12 -08002439 atomic_set(&wq->nr_pwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002440 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002441
Tejun Heo49e3cf42013-03-12 11:29:58 -07002442 for_each_pwq(pwq, wq) {
Tejun Heo112202d2013-02-13 19:29:12 -08002443 struct worker_pool *pool = pwq->pool;
Tejun Heo73f53c42010-06-29 10:07:11 +02002444
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002445 spin_lock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002446
2447 if (flush_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002448 WARN_ON_ONCE(pwq->flush_color != -1);
Tejun Heo73f53c42010-06-29 10:07:11 +02002449
Tejun Heo112202d2013-02-13 19:29:12 -08002450 if (pwq->nr_in_flight[flush_color]) {
2451 pwq->flush_color = flush_color;
2452 atomic_inc(&wq->nr_pwqs_to_flush);
Tejun Heo73f53c42010-06-29 10:07:11 +02002453 wait = true;
2454 }
2455 }
2456
2457 if (work_color >= 0) {
Tejun Heo6183c002013-03-12 11:29:57 -07002458 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
Tejun Heo112202d2013-02-13 19:29:12 -08002459 pwq->work_color = work_color;
Tejun Heo73f53c42010-06-29 10:07:11 +02002460 }
2461
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002462 spin_unlock_irq(&pool->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002463 }
2464
Tejun Heo112202d2013-02-13 19:29:12 -08002465 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
Tejun Heo73f53c42010-06-29 10:07:11 +02002466 complete(&wq->first_flusher->done);
2467
2468 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469}
2470
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002471/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002473 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 *
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002475 * This function sleeps until all work items which were queued on entry
2476 * have finished execution, but it is not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002478void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479{
Tejun Heo73f53c42010-06-29 10:07:11 +02002480 struct wq_flusher this_flusher = {
2481 .list = LIST_HEAD_INIT(this_flusher.list),
2482 .flush_color = -1,
2483 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2484 };
2485 int next_color;
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -07002486
Ingo Molnar3295f0e2008-08-11 10:30:30 +02002487 lock_map_acquire(&wq->lockdep_map);
2488 lock_map_release(&wq->lockdep_map);
Tejun Heo73f53c42010-06-29 10:07:11 +02002489
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002490 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002491
2492 /*
2493 * Start-to-wait phase
2494 */
2495 next_color = work_next_color(wq->work_color);
2496
2497 if (next_color != wq->flush_color) {
2498 /*
2499 * Color space is not full. The current work_color
2500 * becomes our flush_color and work_color is advanced
2501 * by one.
2502 */
Tejun Heo6183c002013-03-12 11:29:57 -07002503 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
Tejun Heo73f53c42010-06-29 10:07:11 +02002504 this_flusher.flush_color = wq->work_color;
2505 wq->work_color = next_color;
2506
2507 if (!wq->first_flusher) {
2508 /* no flush in progress, become the first flusher */
Tejun Heo6183c002013-03-12 11:29:57 -07002509 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002510
2511 wq->first_flusher = &this_flusher;
2512
Tejun Heo112202d2013-02-13 19:29:12 -08002513 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
Tejun Heo73f53c42010-06-29 10:07:11 +02002514 wq->work_color)) {
2515 /* nothing to flush, done */
2516 wq->flush_color = next_color;
2517 wq->first_flusher = NULL;
2518 goto out_unlock;
2519 }
2520 } else {
2521 /* wait in queue */
Tejun Heo6183c002013-03-12 11:29:57 -07002522 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002523 list_add_tail(&this_flusher.list, &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002524 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002525 }
2526 } else {
2527 /*
2528 * Oops, color space is full, wait on overflow queue.
2529 * The next flush completion will assign us
2530 * flush_color and transfer to flusher_queue.
2531 */
2532 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2533 }
2534
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002535 mutex_unlock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002536
2537 wait_for_completion(&this_flusher.done);
2538
2539 /*
2540 * Wake-up-and-cascade phase
2541 *
2542 * First flushers are responsible for cascading flushes and
2543 * handling overflow. Non-first flushers can simply return.
2544 */
2545 if (wq->first_flusher != &this_flusher)
2546 return;
2547
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002548 mutex_lock(&wq->mutex);
Tejun Heo73f53c42010-06-29 10:07:11 +02002549
Tejun Heo4ce48b32010-07-02 10:03:51 +02002550 /* we might have raced, check again with mutex held */
2551 if (wq->first_flusher != &this_flusher)
2552 goto out_unlock;
2553
Tejun Heo73f53c42010-06-29 10:07:11 +02002554 wq->first_flusher = NULL;
2555
Tejun Heo6183c002013-03-12 11:29:57 -07002556 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2557 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002558
2559 while (true) {
2560 struct wq_flusher *next, *tmp;
2561
2562 /* complete all the flushers sharing the current flush color */
2563 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2564 if (next->flush_color != wq->flush_color)
2565 break;
2566 list_del_init(&next->list);
2567 complete(&next->done);
2568 }
2569
Tejun Heo6183c002013-03-12 11:29:57 -07002570 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2571 wq->flush_color != work_next_color(wq->work_color));
Tejun Heo73f53c42010-06-29 10:07:11 +02002572
2573 /* this flush_color is finished, advance by one */
2574 wq->flush_color = work_next_color(wq->flush_color);
2575
2576 /* one color has been freed, handle overflow queue */
2577 if (!list_empty(&wq->flusher_overflow)) {
2578 /*
2579 * Assign the same color to all overflowed
2580 * flushers, advance work_color and append to
2581 * flusher_queue. This is the start-to-wait
2582 * phase for these overflowed flushers.
2583 */
2584 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2585 tmp->flush_color = wq->work_color;
2586
2587 wq->work_color = work_next_color(wq->work_color);
2588
2589 list_splice_tail_init(&wq->flusher_overflow,
2590 &wq->flusher_queue);
Tejun Heo112202d2013-02-13 19:29:12 -08002591 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002592 }
2593
2594 if (list_empty(&wq->flusher_queue)) {
Tejun Heo6183c002013-03-12 11:29:57 -07002595 WARN_ON_ONCE(wq->flush_color != wq->work_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002596 break;
2597 }
2598
2599 /*
2600 * Need to flush more colors. Make the next flusher
Tejun Heo112202d2013-02-13 19:29:12 -08002601 * the new first flusher and arm pwqs.
Tejun Heo73f53c42010-06-29 10:07:11 +02002602 */
Tejun Heo6183c002013-03-12 11:29:57 -07002603 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2604 WARN_ON_ONCE(wq->flush_color != next->flush_color);
Tejun Heo73f53c42010-06-29 10:07:11 +02002605
2606 list_del_init(&next->list);
2607 wq->first_flusher = next;
2608
Tejun Heo112202d2013-02-13 19:29:12 -08002609 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
Tejun Heo73f53c42010-06-29 10:07:11 +02002610 break;
2611
2612 /*
2613 * Meh... this color is already done, clear first
2614 * flusher and repeat cascading.
2615 */
2616 wq->first_flusher = NULL;
2617 }
2618
2619out_unlock:
Lai Jiangshan3c25a552013-03-25 16:57:17 -07002620 mutex_unlock(&wq->mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621}
Dave Jonesae90dd52006-06-30 01:40:45 -04002622EXPORT_SYMBOL_GPL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002624/**
2625 * drain_workqueue - drain a workqueue
2626 * @wq: workqueue to drain
2627 *
2628 * Wait until the workqueue becomes empty. While draining is in progress,
2629 * only chain queueing is allowed. IOW, only currently pending or running
2630 * work items on @wq can queue further work items on it. @wq is flushed
Chen Hanxiaob749b1b2015-05-13 06:10:05 -04002631 * repeatedly until it becomes empty. The number of flushing is determined
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002632 * by the depth of chaining and should be relatively short. Whine if it
2633 * takes too long.
2634 */
2635void drain_workqueue(struct workqueue_struct *wq)
2636{
2637 unsigned int flush_cnt = 0;
Tejun Heo49e3cf42013-03-12 11:29:58 -07002638 struct pool_workqueue *pwq;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002639
2640 /*
2641 * __queue_work() needs to test whether there are drainers, is much
2642 * hotter than drain_workqueue() and already looks at @wq->flags.
Tejun Heo618b01e2013-03-12 11:30:04 -07002643 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002644 */
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002645 mutex_lock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002646 if (!wq->nr_drainers++)
Tejun Heo618b01e2013-03-12 11:30:04 -07002647 wq->flags |= __WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002648 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002649reflush:
2650 flush_workqueue(wq);
2651
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002652 mutex_lock(&wq->mutex);
Tejun Heo76af4d92013-03-12 11:30:00 -07002653
Tejun Heo49e3cf42013-03-12 11:29:58 -07002654 for_each_pwq(pwq, wq) {
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002655 bool drained;
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002656
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002657 spin_lock_irq(&pwq->pool->lock);
Tejun Heo112202d2013-02-13 19:29:12 -08002658 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002659 spin_unlock_irq(&pwq->pool->lock);
Thomas Tuttlefa2563e2011-09-14 16:22:28 -07002660
2661 if (drained)
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002662 continue;
2663
2664 if (++flush_cnt == 10 ||
2665 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
Tejun Heoc5aa87b2013-03-13 16:51:36 -07002666 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
Valentin Ilie044c7822012-08-19 00:52:42 +03002667 wq->name, flush_cnt);
Tejun Heo76af4d92013-03-12 11:30:00 -07002668
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07002669 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002670 goto reflush;
2671 }
2672
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002673 if (!--wq->nr_drainers)
Tejun Heo618b01e2013-03-12 11:30:04 -07002674 wq->flags &= ~__WQ_DRAINING;
Lai Jiangshan87fc7412013-03-25 16:57:18 -07002675 mutex_unlock(&wq->mutex);
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02002676}
2677EXPORT_SYMBOL_GPL(drain_workqueue);
2678
Tejun Heo606a5022012-08-20 14:51:23 -07002679static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
Tejun Heobaf59022010-09-16 10:42:16 +02002680{
2681 struct worker *worker = NULL;
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002682 struct worker_pool *pool;
Tejun Heo112202d2013-02-13 19:29:12 -08002683 struct pool_workqueue *pwq;
Tejun Heobaf59022010-09-16 10:42:16 +02002684
2685 might_sleep();
Tejun Heobaf59022010-09-16 10:42:16 +02002686
Tejun Heofa1b54e2013-03-12 11:30:00 -07002687 local_irq_disable();
2688 pool = get_work_pool(work);
2689 if (!pool) {
2690 local_irq_enable();
2691 return false;
2692 }
2693
2694 spin_lock(&pool->lock);
Lai Jiangshan0b3dae62013-02-06 18:04:53 -08002695 /* see the comment in try_to_grab_pending() with the same code */
Tejun Heo112202d2013-02-13 19:29:12 -08002696 pwq = get_work_pwq(work);
2697 if (pwq) {
2698 if (unlikely(pwq->pool != pool))
Tejun Heobaf59022010-09-16 10:42:16 +02002699 goto already_gone;
Tejun Heo606a5022012-08-20 14:51:23 -07002700 } else {
Tejun Heoc9e7cf22013-01-24 11:01:33 -08002701 worker = find_worker_executing_work(pool, work);
Tejun Heobaf59022010-09-16 10:42:16 +02002702 if (!worker)
2703 goto already_gone;
Tejun Heo112202d2013-02-13 19:29:12 -08002704 pwq = worker->current_pwq;
Tejun Heo606a5022012-08-20 14:51:23 -07002705 }
Tejun Heobaf59022010-09-16 10:42:16 +02002706
Tejun Heo112202d2013-02-13 19:29:12 -08002707 insert_wq_barrier(pwq, barr, work, worker);
Tejun Heod565ed62013-01-24 11:01:33 -08002708 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002709
Tejun Heoe1594892011-01-09 23:32:15 +01002710 /*
2711 * If @max_active is 1 or rescuer is in use, flushing another work
2712 * item on the same workqueue may lead to deadlock. Make sure the
2713 * flusher is not running on the same workqueue by verifying write
2714 * access.
2715 */
Tejun Heo493008a2013-03-12 11:30:03 -07002716 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
Tejun Heo112202d2013-02-13 19:29:12 -08002717 lock_map_acquire(&pwq->wq->lockdep_map);
Tejun Heoe1594892011-01-09 23:32:15 +01002718 else
Tejun Heo112202d2013-02-13 19:29:12 -08002719 lock_map_acquire_read(&pwq->wq->lockdep_map);
2720 lock_map_release(&pwq->wq->lockdep_map);
Tejun Heoe1594892011-01-09 23:32:15 +01002721
Tejun Heobaf59022010-09-16 10:42:16 +02002722 return true;
2723already_gone:
Tejun Heod565ed62013-01-24 11:01:33 -08002724 spin_unlock_irq(&pool->lock);
Tejun Heobaf59022010-09-16 10:42:16 +02002725 return false;
2726}
2727
Oleg Nesterovdb700892008-07-25 01:47:49 -07002728/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002729 * flush_work - wait for a work to finish executing the last queueing instance
2730 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07002731 *
Tejun Heo606a5022012-08-20 14:51:23 -07002732 * Wait until @work has finished execution. @work is guaranteed to be idle
2733 * on return if it hasn't been requeued since flush started.
Tejun Heo401a8d02010-09-16 10:36:00 +02002734 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002735 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002736 * %true if flush_work() waited for the work to finish execution,
2737 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07002738 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002739bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07002740{
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002741 struct wq_barrier barr;
2742
Stephen Boyd0976dfc2012-04-20 17:28:50 -07002743 lock_map_acquire(&work->lockdep_map);
2744 lock_map_release(&work->lockdep_map);
2745
Bjorn Helgaas12997d12013-11-18 11:00:29 -07002746 if (start_flush_work(work, &barr)) {
2747 wait_for_completion(&barr.done);
2748 destroy_work_on_stack(&barr.work);
2749 return true;
2750 } else {
2751 return false;
2752 }
Oleg Nesterovdb700892008-07-25 01:47:49 -07002753}
2754EXPORT_SYMBOL_GPL(flush_work);
2755
Tejun Heo8603e1b32015-03-05 08:04:13 -05002756struct cwt_wait {
2757 wait_queue_t wait;
2758 struct work_struct *work;
2759};
2760
2761static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2762{
2763 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2764
2765 if (cwait->work != key)
2766 return 0;
2767 return autoremove_wake_function(wait, mode, sync, key);
2768}
2769
Tejun Heo36e227d2012-08-03 10:30:46 -07002770static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
Tejun Heo401a8d02010-09-16 10:36:00 +02002771{
Tejun Heo8603e1b32015-03-05 08:04:13 -05002772 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
Tejun Heobbb68df2012-08-03 10:30:46 -07002773 unsigned long flags;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002774 int ret;
2775
2776 do {
Tejun Heobbb68df2012-08-03 10:30:46 -07002777 ret = try_to_grab_pending(work, is_dwork, &flags);
2778 /*
Tejun Heo8603e1b32015-03-05 08:04:13 -05002779 * If someone else is already canceling, wait for it to
2780 * finish. flush_work() doesn't work for PREEMPT_NONE
2781 * because we may get scheduled between @work's completion
2782 * and the other canceling task resuming and clearing
2783 * CANCELING - flush_work() will return false immediately
2784 * as @work is no longer busy, try_to_grab_pending() will
2785 * return -ENOENT as @work is still being canceled and the
2786 * other canceling task won't be able to clear CANCELING as
2787 * we're hogging the CPU.
2788 *
2789 * Let's wait for completion using a waitqueue. As this
2790 * may lead to the thundering herd problem, use a custom
2791 * wake function which matches @work along with exclusive
2792 * wait and wakeup.
Tejun Heobbb68df2012-08-03 10:30:46 -07002793 */
Tejun Heo8603e1b32015-03-05 08:04:13 -05002794 if (unlikely(ret == -ENOENT)) {
2795 struct cwt_wait cwait;
2796
2797 init_wait(&cwait.wait);
2798 cwait.wait.func = cwt_wakefn;
2799 cwait.work = work;
2800
2801 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2802 TASK_UNINTERRUPTIBLE);
2803 if (work_is_canceling(work))
2804 schedule();
2805 finish_wait(&cancel_waitq, &cwait.wait);
2806 }
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002807 } while (unlikely(ret < 0));
2808
Tejun Heobbb68df2012-08-03 10:30:46 -07002809 /* tell other tasks trying to grab @work to back off */
2810 mark_work_canceling(work);
2811 local_irq_restore(flags);
2812
Tejun Heo606a5022012-08-20 14:51:23 -07002813 flush_work(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002814 clear_work_data(work);
Tejun Heo8603e1b32015-03-05 08:04:13 -05002815
2816 /*
2817 * Paired with prepare_to_wait() above so that either
2818 * waitqueue_active() is visible here or !work_is_canceling() is
2819 * visible there.
2820 */
2821 smp_mb();
2822 if (waitqueue_active(&cancel_waitq))
2823 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2824
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002825 return ret;
2826}
2827
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002828/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002829 * cancel_work_sync - cancel a work and wait for it to finish
2830 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002831 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002832 * Cancel @work and wait for its execution to finish. This function
2833 * can be used even if the work re-queues itself or migrates to
2834 * another workqueue. On return from this function, @work is
2835 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002836 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002837 * cancel_work_sync(&delayed_work->work) must not be used for
2838 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002839 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002840 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002841 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02002842 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002843 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002844 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002845 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002846bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002847{
Tejun Heo36e227d2012-08-03 10:30:46 -07002848 return __cancel_work_timer(work, false);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002849}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07002850EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002851
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002852/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002853 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2854 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002855 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002856 * Delayed timer is cancelled and the pending work is queued for
2857 * immediate execution. Like flush_work(), this function only
2858 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002859 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002860 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002861 * %true if flush_work() waited for the work to finish execution,
2862 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002863 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002864bool flush_delayed_work(struct delayed_work *dwork)
2865{
Tejun Heo8930cab2012-08-03 10:30:45 -07002866 local_irq_disable();
Tejun Heo401a8d02010-09-16 10:36:00 +02002867 if (del_timer_sync(&dwork->timer))
Lai Jiangshan60c057b2013-02-06 18:04:53 -08002868 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
Tejun Heo8930cab2012-08-03 10:30:45 -07002869 local_irq_enable();
Tejun Heo401a8d02010-09-16 10:36:00 +02002870 return flush_work(&dwork->work);
2871}
2872EXPORT_SYMBOL(flush_delayed_work);
2873
2874/**
Tejun Heo57b30ae2012-08-21 13:18:24 -07002875 * cancel_delayed_work - cancel a delayed work
2876 * @dwork: delayed_work to cancel
Tejun Heo09383492010-09-16 10:48:29 +02002877 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002878 * Kill off a pending delayed_work.
2879 *
2880 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2881 * pending.
2882 *
2883 * Note:
2884 * The work callback function may still be running on return, unless
2885 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2886 * use cancel_delayed_work_sync() to wait on it.
Tejun Heo09383492010-09-16 10:48:29 +02002887 *
Tejun Heo57b30ae2012-08-21 13:18:24 -07002888 * This function is safe to call from any context including IRQ handler.
Tejun Heo09383492010-09-16 10:48:29 +02002889 */
Tejun Heo57b30ae2012-08-21 13:18:24 -07002890bool cancel_delayed_work(struct delayed_work *dwork)
Tejun Heo09383492010-09-16 10:48:29 +02002891{
Tejun Heo57b30ae2012-08-21 13:18:24 -07002892 unsigned long flags;
2893 int ret;
2894
2895 do {
2896 ret = try_to_grab_pending(&dwork->work, true, &flags);
2897 } while (unlikely(ret == -EAGAIN));
2898
2899 if (unlikely(ret < 0))
2900 return false;
2901
Tejun Heo7c3eed52013-01-24 11:01:33 -08002902 set_work_pool_and_clear_pending(&dwork->work,
2903 get_work_pool_id(&dwork->work));
Tejun Heo57b30ae2012-08-21 13:18:24 -07002904 local_irq_restore(flags);
Dan Magenheimerc0158ca2012-10-18 16:31:37 -07002905 return ret;
Tejun Heo09383492010-09-16 10:48:29 +02002906}
Tejun Heo57b30ae2012-08-21 13:18:24 -07002907EXPORT_SYMBOL(cancel_delayed_work);
Tejun Heo09383492010-09-16 10:48:29 +02002908
2909/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002910 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2911 * @dwork: the delayed work cancel
2912 *
2913 * This is cancel_work_sync() for delayed works.
2914 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002915 * Return:
Tejun Heo401a8d02010-09-16 10:36:00 +02002916 * %true if @dwork was pending, %false otherwise.
2917 */
2918bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002919{
Tejun Heo36e227d2012-08-03 10:30:46 -07002920 return __cancel_work_timer(&dwork->work, true);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002921}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07002922EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002924/**
Tejun Heo31ddd872010-10-19 11:14:49 +02002925 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07002926 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07002927 *
Tejun Heo31ddd872010-10-19 11:14:49 +02002928 * schedule_on_each_cpu() executes @func on each online CPU using the
2929 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07002930 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02002931 *
Yacine Belkadid185af32013-07-31 14:59:24 -07002932 * Return:
Tejun Heo31ddd872010-10-19 11:14:49 +02002933 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07002934 */
David Howells65f27f32006-11-22 14:55:48 +00002935int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba2006-01-08 01:00:43 -08002936{
2937 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02002938 struct work_struct __percpu *works;
Christoph Lameter15316ba2006-01-08 01:00:43 -08002939
Andrew Mortonb6136772006-06-25 05:47:49 -07002940 works = alloc_percpu(struct work_struct);
2941 if (!works)
Christoph Lameter15316ba2006-01-08 01:00:43 -08002942 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07002943
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002944 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08002945
Christoph Lameter15316ba2006-01-08 01:00:43 -08002946 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01002947 struct work_struct *work = per_cpu_ptr(works, cpu);
2948
2949 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02002950 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02002951 }
Tejun Heo93981802009-11-17 14:06:20 -08002952
2953 for_each_online_cpu(cpu)
2954 flush_work(per_cpu_ptr(works, cpu));
2955
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002956 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07002957 free_percpu(works);
Christoph Lameter15316ba2006-01-08 01:00:43 -08002958 return 0;
2959}
2960
Alan Sterneef6a7d2010-02-12 17:39:21 +09002961/**
2962 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2963 *
2964 * Forces execution of the kernel-global workqueue and blocks until its
2965 * completion.
2966 *
2967 * Think twice before calling this function! It's very easy to get into
2968 * trouble if you don't take great care. Either of the following situations
2969 * will lead to deadlock:
2970 *
2971 * One of the work items currently on the workqueue needs to acquire
2972 * a lock held by your code or its caller.
2973 *
2974 * Your code is running in the context of a work routine.
2975 *
2976 * They will be detected by lockdep when they occur, but the first might not
2977 * occur very often. It depends on what work items are on the workqueue and
2978 * what locks they need, which you have no control over.
2979 *
2980 * In most situations flushing the entire workqueue is overkill; you merely
2981 * need to know that a particular work item isn't queued and isn't running.
2982 * In such cases you should use cancel_delayed_work_sync() or
2983 * cancel_work_sync() instead.
2984 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985void flush_scheduled_work(void)
2986{
Tejun Heod320c032010-06-29 10:07:14 +02002987 flush_workqueue(system_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988}
Dave Jonesae90dd52006-06-30 01:40:45 -04002989EXPORT_SYMBOL(flush_scheduled_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990
2991/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06002992 * execute_in_process_context - reliably execute the routine with user context
2993 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06002994 * @ew: guaranteed storage for the execute work structure (must
2995 * be available when the work executes)
2996 *
2997 * Executes the function immediately if process context is available,
2998 * otherwise schedules the function for delayed execution.
2999 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003000 * Return: 0 - function was executed
James Bottomley1fa44ec2006-02-23 12:43:43 -06003001 * 1 - function was scheduled for execution
3002 */
David Howells65f27f32006-11-22 14:55:48 +00003003int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06003004{
3005 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00003006 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003007 return 0;
3008 }
3009
David Howells65f27f32006-11-22 14:55:48 +00003010 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06003011 schedule_work(&ew->work);
3012
3013 return 1;
3014}
3015EXPORT_SYMBOL_GPL(execute_in_process_context);
3016
Tejun Heo7a4e3442013-03-12 11:30:00 -07003017/**
3018 * free_workqueue_attrs - free a workqueue_attrs
3019 * @attrs: workqueue_attrs to free
3020 *
3021 * Undo alloc_workqueue_attrs().
3022 */
3023void free_workqueue_attrs(struct workqueue_attrs *attrs)
3024{
3025 if (attrs) {
3026 free_cpumask_var(attrs->cpumask);
3027 kfree(attrs);
3028 }
3029}
3030
3031/**
3032 * alloc_workqueue_attrs - allocate a workqueue_attrs
3033 * @gfp_mask: allocation mask to use
3034 *
3035 * Allocate a new workqueue_attrs, initialize with default settings and
Yacine Belkadid185af32013-07-31 14:59:24 -07003036 * return it.
3037 *
3038 * Return: The allocated new workqueue_attr on success. %NULL on failure.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003039 */
3040struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3041{
3042 struct workqueue_attrs *attrs;
3043
3044 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3045 if (!attrs)
3046 goto fail;
3047 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3048 goto fail;
3049
Tejun Heo13e2e552013-04-01 11:23:31 -07003050 cpumask_copy(attrs->cpumask, cpu_possible_mask);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003051 return attrs;
3052fail:
3053 free_workqueue_attrs(attrs);
3054 return NULL;
3055}
3056
Tejun Heo29c91e92013-03-12 11:30:03 -07003057static void copy_workqueue_attrs(struct workqueue_attrs *to,
3058 const struct workqueue_attrs *from)
3059{
3060 to->nice = from->nice;
3061 cpumask_copy(to->cpumask, from->cpumask);
Shaohua Li2865a8f2013-08-01 09:56:36 +08003062 /*
3063 * Unlike hash and equality test, this function doesn't ignore
3064 * ->no_numa as it is used for both pool and wq attrs. Instead,
3065 * get_unbound_pool() explicitly clears ->no_numa after copying.
3066 */
3067 to->no_numa = from->no_numa;
Tejun Heo29c91e92013-03-12 11:30:03 -07003068}
3069
Tejun Heo29c91e92013-03-12 11:30:03 -07003070/* hash value of the content of @attr */
3071static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3072{
3073 u32 hash = 0;
3074
3075 hash = jhash_1word(attrs->nice, hash);
Tejun Heo13e2e552013-04-01 11:23:31 -07003076 hash = jhash(cpumask_bits(attrs->cpumask),
3077 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
Tejun Heo29c91e92013-03-12 11:30:03 -07003078 return hash;
3079}
3080
3081/* content equality test */
3082static bool wqattrs_equal(const struct workqueue_attrs *a,
3083 const struct workqueue_attrs *b)
3084{
3085 if (a->nice != b->nice)
3086 return false;
3087 if (!cpumask_equal(a->cpumask, b->cpumask))
3088 return false;
3089 return true;
3090}
3091
Tejun Heo7a4e3442013-03-12 11:30:00 -07003092/**
3093 * init_worker_pool - initialize a newly zalloc'd worker_pool
3094 * @pool: worker_pool to initialize
3095 *
3096 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
Yacine Belkadid185af32013-07-31 14:59:24 -07003097 *
3098 * Return: 0 on success, -errno on failure. Even on failure, all fields
Tejun Heo29c91e92013-03-12 11:30:03 -07003099 * inside @pool proper are initialized and put_unbound_pool() can be called
3100 * on @pool safely to release it.
Tejun Heo7a4e3442013-03-12 11:30:00 -07003101 */
3102static int init_worker_pool(struct worker_pool *pool)
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003103{
3104 spin_lock_init(&pool->lock);
Tejun Heo29c91e92013-03-12 11:30:03 -07003105 pool->id = -1;
3106 pool->cpu = -1;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003107 pool->node = NUMA_NO_NODE;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003108 pool->flags |= POOL_DISASSOCIATED;
3109 INIT_LIST_HEAD(&pool->worklist);
3110 INIT_LIST_HEAD(&pool->idle_list);
3111 hash_init(pool->busy_hash);
3112
3113 init_timer_deferrable(&pool->idle_timer);
3114 pool->idle_timer.function = idle_worker_timeout;
3115 pool->idle_timer.data = (unsigned long)pool;
3116
3117 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3118 (unsigned long)pool);
3119
3120 mutex_init(&pool->manager_arb);
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003121 mutex_init(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003122 INIT_LIST_HEAD(&pool->workers);
Tejun Heo7a4e3442013-03-12 11:30:00 -07003123
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003124 ida_init(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003125 INIT_HLIST_NODE(&pool->hash_node);
3126 pool->refcnt = 1;
3127
3128 /* shouldn't fail above this point */
Tejun Heo7a4e3442013-03-12 11:30:00 -07003129 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3130 if (!pool->attrs)
3131 return -ENOMEM;
3132 return 0;
Tejun Heo4e1a1f92013-03-12 11:30:00 -07003133}
3134
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003135static void rcu_free_wq(struct rcu_head *rcu)
3136{
3137 struct workqueue_struct *wq =
3138 container_of(rcu, struct workqueue_struct, rcu);
3139
3140 if (!(wq->flags & WQ_UNBOUND))
3141 free_percpu(wq->cpu_pwqs);
3142 else
3143 free_workqueue_attrs(wq->unbound_attrs);
3144
3145 kfree(wq->rescuer);
3146 kfree(wq);
3147}
3148
Tejun Heo29c91e92013-03-12 11:30:03 -07003149static void rcu_free_pool(struct rcu_head *rcu)
3150{
3151 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3152
Lai Jiangshan7cda9aa2014-05-20 17:46:32 +08003153 ida_destroy(&pool->worker_ida);
Tejun Heo29c91e92013-03-12 11:30:03 -07003154 free_workqueue_attrs(pool->attrs);
3155 kfree(pool);
3156}
3157
3158/**
3159 * put_unbound_pool - put a worker_pool
3160 * @pool: worker_pool to put
3161 *
3162 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003163 * safe manner. get_unbound_pool() calls this function on its failure path
3164 * and this function should be able to release pools which went through,
3165 * successfully or not, init_worker_pool().
Tejun Heoa892cac2013-04-01 11:23:32 -07003166 *
3167 * Should be called with wq_pool_mutex held.
Tejun Heo29c91e92013-03-12 11:30:03 -07003168 */
3169static void put_unbound_pool(struct worker_pool *pool)
3170{
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003171 DECLARE_COMPLETION_ONSTACK(detach_completion);
Tejun Heo29c91e92013-03-12 11:30:03 -07003172 struct worker *worker;
3173
Tejun Heoa892cac2013-04-01 11:23:32 -07003174 lockdep_assert_held(&wq_pool_mutex);
3175
3176 if (--pool->refcnt)
Tejun Heo29c91e92013-03-12 11:30:03 -07003177 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003178
3179 /* sanity checks */
Lai Jiangshan61d0fbb2014-06-03 15:31:45 +08003180 if (WARN_ON(!(pool->cpu < 0)) ||
Tejun Heoa892cac2013-04-01 11:23:32 -07003181 WARN_ON(!list_empty(&pool->worklist)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003182 return;
Tejun Heo29c91e92013-03-12 11:30:03 -07003183
3184 /* release id and unhash */
3185 if (pool->id >= 0)
3186 idr_remove(&worker_pool_idr, pool->id);
3187 hash_del(&pool->hash_node);
3188
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003189 /*
3190 * Become the manager and destroy all workers. Grabbing
3191 * manager_arb prevents @pool's workers from blocking on
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003192 * attach_mutex.
Tejun Heoc5aa87b2013-03-13 16:51:36 -07003193 */
Tejun Heo29c91e92013-03-12 11:30:03 -07003194 mutex_lock(&pool->manager_arb);
Tejun Heo29c91e92013-03-12 11:30:03 -07003195
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003196 spin_lock_irq(&pool->lock);
Lai Jiangshan1037de32014-05-22 16:44:07 +08003197 while ((worker = first_idle_worker(pool)))
Tejun Heo29c91e92013-03-12 11:30:03 -07003198 destroy_worker(worker);
3199 WARN_ON(pool->nr_workers || pool->nr_idle);
Tejun Heo29c91e92013-03-12 11:30:03 -07003200 spin_unlock_irq(&pool->lock);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003201
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003202 mutex_lock(&pool->attach_mutex);
Lai Jiangshanda028462014-05-20 17:46:31 +08003203 if (!list_empty(&pool->workers))
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003204 pool->detach_completion = &detach_completion;
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08003205 mutex_unlock(&pool->attach_mutex);
Lai Jiangshan60f5a4b2014-05-20 17:46:29 +08003206
3207 if (pool->detach_completion)
3208 wait_for_completion(pool->detach_completion);
3209
Tejun Heo29c91e92013-03-12 11:30:03 -07003210 mutex_unlock(&pool->manager_arb);
3211
3212 /* shut down the timers */
3213 del_timer_sync(&pool->idle_timer);
3214 del_timer_sync(&pool->mayday_timer);
3215
3216 /* sched-RCU protected to allow dereferences from get_work_pool() */
3217 call_rcu_sched(&pool->rcu, rcu_free_pool);
3218}
3219
3220/**
3221 * get_unbound_pool - get a worker_pool with the specified attributes
3222 * @attrs: the attributes of the worker_pool to get
3223 *
3224 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3225 * reference count and return it. If there already is a matching
3226 * worker_pool, it will be used; otherwise, this function attempts to
Yacine Belkadid185af32013-07-31 14:59:24 -07003227 * create a new one.
Tejun Heoa892cac2013-04-01 11:23:32 -07003228 *
3229 * Should be called with wq_pool_mutex held.
Yacine Belkadid185af32013-07-31 14:59:24 -07003230 *
3231 * Return: On success, a worker_pool with the same attributes as @attrs.
3232 * On failure, %NULL.
Tejun Heo29c91e92013-03-12 11:30:03 -07003233 */
3234static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3235{
Tejun Heo29c91e92013-03-12 11:30:03 -07003236 u32 hash = wqattrs_hash(attrs);
3237 struct worker_pool *pool;
Tejun Heof3f90ad2013-04-01 11:23:34 -07003238 int node;
Tejun Heo29c91e92013-03-12 11:30:03 -07003239
Tejun Heoa892cac2013-04-01 11:23:32 -07003240 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo29c91e92013-03-12 11:30:03 -07003241
3242 /* do we already have a matching pool? */
Tejun Heo29c91e92013-03-12 11:30:03 -07003243 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3244 if (wqattrs_equal(pool->attrs, attrs)) {
3245 pool->refcnt++;
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003246 return pool;
Tejun Heo29c91e92013-03-12 11:30:03 -07003247 }
3248 }
Tejun Heo29c91e92013-03-12 11:30:03 -07003249
3250 /* nope, create a new one */
3251 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3252 if (!pool || init_worker_pool(pool) < 0)
3253 goto fail;
3254
Tejun Heo8864b4e2013-03-12 11:30:04 -07003255 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
Tejun Heo29c91e92013-03-12 11:30:03 -07003256 copy_workqueue_attrs(pool->attrs, attrs);
3257
Shaohua Li2865a8f2013-08-01 09:56:36 +08003258 /*
3259 * no_numa isn't a worker_pool attribute, always clear it. See
3260 * 'struct workqueue_attrs' comments for detail.
3261 */
3262 pool->attrs->no_numa = false;
3263
Tejun Heof3f90ad2013-04-01 11:23:34 -07003264 /* if cpumask is contained inside a NUMA node, we belong to that node */
3265 if (wq_numa_enabled) {
3266 for_each_node(node) {
3267 if (cpumask_subset(pool->attrs->cpumask,
3268 wq_numa_possible_cpumask[node])) {
3269 pool->node = node;
3270 break;
3271 }
3272 }
3273 }
3274
Tejun Heo29c91e92013-03-12 11:30:03 -07003275 if (worker_pool_assign_id(pool) < 0)
3276 goto fail;
3277
3278 /* create and start the initial worker */
Lai Jiangshan051e1852014-07-22 13:03:02 +08003279 if (!create_worker(pool))
Tejun Heo29c91e92013-03-12 11:30:03 -07003280 goto fail;
3281
Tejun Heo29c91e92013-03-12 11:30:03 -07003282 /* install */
Tejun Heo29c91e92013-03-12 11:30:03 -07003283 hash_add(unbound_pool_hash, &pool->hash_node, hash);
Lai Jiangshan3fb18232014-07-22 13:04:49 +08003284
Tejun Heo29c91e92013-03-12 11:30:03 -07003285 return pool;
3286fail:
Tejun Heo29c91e92013-03-12 11:30:03 -07003287 if (pool)
3288 put_unbound_pool(pool);
3289 return NULL;
3290}
3291
Tejun Heo8864b4e2013-03-12 11:30:04 -07003292static void rcu_free_pwq(struct rcu_head *rcu)
3293{
3294 kmem_cache_free(pwq_cache,
3295 container_of(rcu, struct pool_workqueue, rcu));
3296}
3297
3298/*
3299 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3300 * and needs to be destroyed.
3301 */
3302static void pwq_unbound_release_workfn(struct work_struct *work)
3303{
3304 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3305 unbound_release_work);
3306 struct workqueue_struct *wq = pwq->wq;
3307 struct worker_pool *pool = pwq->pool;
Tejun Heobc0caf02013-04-01 11:23:31 -07003308 bool is_last;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003309
3310 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3311 return;
3312
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003313 mutex_lock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003314 list_del_rcu(&pwq->pwqs_node);
Tejun Heobc0caf02013-04-01 11:23:31 -07003315 is_last = list_empty(&wq->pwqs);
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003316 mutex_unlock(&wq->mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003317
Tejun Heoa892cac2013-04-01 11:23:32 -07003318 mutex_lock(&wq_pool_mutex);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003319 put_unbound_pool(pool);
Tejun Heoa892cac2013-04-01 11:23:32 -07003320 mutex_unlock(&wq_pool_mutex);
3321
Tejun Heo8864b4e2013-03-12 11:30:04 -07003322 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3323
3324 /*
3325 * If we're the last pwq going away, @wq is already dead and no one
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003326 * is gonna access it anymore. Schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003327 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003328 if (is_last)
3329 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003330}
3331
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003332/**
Tejun Heo699ce092013-03-13 16:51:35 -07003333 * pwq_adjust_max_active - update a pwq's max_active to the current setting
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003334 * @pwq: target pool_workqueue
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003335 *
Tejun Heo699ce092013-03-13 16:51:35 -07003336 * If @pwq isn't freezing, set @pwq->max_active to the associated
3337 * workqueue's saved_max_active and activate delayed work items
3338 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003339 */
Tejun Heo699ce092013-03-13 16:51:35 -07003340static void pwq_adjust_max_active(struct pool_workqueue *pwq)
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003341{
Tejun Heo699ce092013-03-13 16:51:35 -07003342 struct workqueue_struct *wq = pwq->wq;
3343 bool freezable = wq->flags & WQ_FREEZABLE;
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003344
Tejun Heo699ce092013-03-13 16:51:35 -07003345 /* for @wq->saved_max_active */
Lai Jiangshana357fc02013-03-25 16:57:19 -07003346 lockdep_assert_held(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003347
3348 /* fast exit for non-freezable wqs */
3349 if (!freezable && pwq->max_active == wq->saved_max_active)
3350 return;
3351
Lai Jiangshana357fc02013-03-25 16:57:19 -07003352 spin_lock_irq(&pwq->pool->lock);
Tejun Heo699ce092013-03-13 16:51:35 -07003353
Lai Jiangshan74b414e2014-05-22 19:01:16 +08003354 /*
3355 * During [un]freezing, the caller is responsible for ensuring that
3356 * this function is called at least once after @workqueue_freezing
3357 * is updated and visible.
3358 */
3359 if (!freezable || !workqueue_freezing) {
Tejun Heo699ce092013-03-13 16:51:35 -07003360 pwq->max_active = wq->saved_max_active;
3361
3362 while (!list_empty(&pwq->delayed_works) &&
3363 pwq->nr_active < pwq->max_active)
3364 pwq_activate_first_delayed(pwq);
Lai Jiangshan951a0782013-03-20 10:52:30 -07003365
3366 /*
3367 * Need to kick a worker after thawed or an unbound wq's
3368 * max_active is bumped. It's a slow path. Do it always.
3369 */
3370 wake_up_worker(pwq->pool);
Tejun Heo699ce092013-03-13 16:51:35 -07003371 } else {
3372 pwq->max_active = 0;
3373 }
3374
Lai Jiangshana357fc02013-03-25 16:57:19 -07003375 spin_unlock_irq(&pwq->pool->lock);
Tejun Heo0fbd95a2013-03-13 16:51:35 -07003376}
3377
Tejun Heoe50aba92013-04-01 11:23:35 -07003378/* initialize newly alloced @pwq which is associated with @wq and @pool */
Tejun Heof147f292013-04-01 11:23:35 -07003379static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3380 struct worker_pool *pool)
Tejun Heod2c1d402013-03-12 11:30:04 -07003381{
3382 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3383
Tejun Heoe50aba92013-04-01 11:23:35 -07003384 memset(pwq, 0, sizeof(*pwq));
3385
Tejun Heod2c1d402013-03-12 11:30:04 -07003386 pwq->pool = pool;
3387 pwq->wq = wq;
3388 pwq->flush_color = -1;
Tejun Heo8864b4e2013-03-12 11:30:04 -07003389 pwq->refcnt = 1;
Tejun Heod2c1d402013-03-12 11:30:04 -07003390 INIT_LIST_HEAD(&pwq->delayed_works);
Tejun Heo1befcf32013-04-01 11:23:35 -07003391 INIT_LIST_HEAD(&pwq->pwqs_node);
Tejun Heod2c1d402013-03-12 11:30:04 -07003392 INIT_LIST_HEAD(&pwq->mayday_node);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003393 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
Tejun Heof147f292013-04-01 11:23:35 -07003394}
Tejun Heod2c1d402013-03-12 11:30:04 -07003395
Tejun Heof147f292013-04-01 11:23:35 -07003396/* sync @pwq with the current state of its associated wq and link it */
Tejun Heo1befcf32013-04-01 11:23:35 -07003397static void link_pwq(struct pool_workqueue *pwq)
Tejun Heof147f292013-04-01 11:23:35 -07003398{
3399 struct workqueue_struct *wq = pwq->wq;
3400
3401 lockdep_assert_held(&wq->mutex);
Tejun Heo75ccf592013-03-12 11:30:04 -07003402
Tejun Heo1befcf32013-04-01 11:23:35 -07003403 /* may be called multiple times, ignore if already linked */
3404 if (!list_empty(&pwq->pwqs_node))
3405 return;
3406
Lai Jiangshan29b1cb42014-07-22 13:04:27 +08003407 /* set the matching work_color */
Tejun Heo75ccf592013-03-12 11:30:04 -07003408 pwq->work_color = wq->work_color;
Tejun Heo983ca252013-03-13 16:51:35 -07003409
3410 /* sync max_active to the current setting */
3411 pwq_adjust_max_active(pwq);
3412
3413 /* link in @pwq */
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003414 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
Tejun Heof147f292013-04-01 11:23:35 -07003415}
Lai Jiangshana357fc02013-03-25 16:57:19 -07003416
Tejun Heof147f292013-04-01 11:23:35 -07003417/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3418static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3419 const struct workqueue_attrs *attrs)
3420{
3421 struct worker_pool *pool;
3422 struct pool_workqueue *pwq;
3423
3424 lockdep_assert_held(&wq_pool_mutex);
3425
3426 pool = get_unbound_pool(attrs);
3427 if (!pool)
3428 return NULL;
3429
Tejun Heoe50aba92013-04-01 11:23:35 -07003430 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
Tejun Heof147f292013-04-01 11:23:35 -07003431 if (!pwq) {
3432 put_unbound_pool(pool);
3433 return NULL;
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003434 }
Tejun Heo6029a912013-04-01 11:23:34 -07003435
Tejun Heof147f292013-04-01 11:23:35 -07003436 init_pwq(pwq, wq, pool);
3437 return pwq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003438}
3439
Tejun Heo4c16bd32013-04-01 11:23:36 -07003440/**
Gong Zhaogang30186c62015-05-11 11:02:47 -04003441 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003442 * @attrs: the wq_attrs of the default pwq of the target workqueue
Tejun Heo4c16bd32013-04-01 11:23:36 -07003443 * @node: the target NUMA node
3444 * @cpu_going_down: if >= 0, the CPU to consider as offline
3445 * @cpumask: outarg, the resulting cpumask
3446 *
3447 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3448 * @cpu_going_down is >= 0, that cpu is considered offline during
Yacine Belkadid185af32013-07-31 14:59:24 -07003449 * calculation. The result is stored in @cpumask.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003450 *
3451 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3452 * enabled and @node has online CPUs requested by @attrs, the returned
3453 * cpumask is the intersection of the possible CPUs of @node and
3454 * @attrs->cpumask.
3455 *
3456 * The caller is responsible for ensuring that the cpumask of @node stays
3457 * stable.
Yacine Belkadid185af32013-07-31 14:59:24 -07003458 *
3459 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3460 * %false if equal.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003461 */
3462static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3463 int cpu_going_down, cpumask_t *cpumask)
3464{
Tejun Heod55262c2013-04-01 11:23:38 -07003465 if (!wq_numa_enabled || attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003466 goto use_dfl;
3467
3468 /* does @node have any online CPUs @attrs wants? */
3469 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3470 if (cpu_going_down >= 0)
3471 cpumask_clear_cpu(cpu_going_down, cpumask);
3472
3473 if (cpumask_empty(cpumask))
3474 goto use_dfl;
3475
3476 /* yeap, return possible CPUs in @node that @attrs wants */
3477 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3478 return !cpumask_equal(cpumask, attrs->cpumask);
3479
3480use_dfl:
3481 cpumask_copy(cpumask, attrs->cpumask);
3482 return false;
3483}
3484
Tejun Heo1befcf32013-04-01 11:23:35 -07003485/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3486static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3487 int node,
3488 struct pool_workqueue *pwq)
3489{
3490 struct pool_workqueue *old_pwq;
3491
Lai Jiangshan5b95e1a2015-05-12 20:32:29 +08003492 lockdep_assert_held(&wq_pool_mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003493 lockdep_assert_held(&wq->mutex);
3494
3495 /* link_pwq() can handle duplicate calls */
3496 link_pwq(pwq);
3497
3498 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3499 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3500 return old_pwq;
3501}
3502
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003503/* context to store the prepared attrs & pwqs before applying */
3504struct apply_wqattrs_ctx {
3505 struct workqueue_struct *wq; /* target workqueue */
3506 struct workqueue_attrs *attrs; /* attrs to apply */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003507 struct list_head list; /* queued for batching commit */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003508 struct pool_workqueue *dfl_pwq;
3509 struct pool_workqueue *pwq_tbl[];
3510};
3511
3512/* free the resources after success or abort */
3513static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3514{
3515 if (ctx) {
3516 int node;
3517
3518 for_each_node(node)
3519 put_pwq_unlocked(ctx->pwq_tbl[node]);
3520 put_pwq_unlocked(ctx->dfl_pwq);
3521
3522 free_workqueue_attrs(ctx->attrs);
3523
3524 kfree(ctx);
3525 }
3526}
3527
3528/* allocate the attrs and pwqs for later installation */
3529static struct apply_wqattrs_ctx *
3530apply_wqattrs_prepare(struct workqueue_struct *wq,
3531 const struct workqueue_attrs *attrs)
3532{
3533 struct apply_wqattrs_ctx *ctx;
3534 struct workqueue_attrs *new_attrs, *tmp_attrs;
3535 int node;
3536
3537 lockdep_assert_held(&wq_pool_mutex);
3538
3539 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3540 GFP_KERNEL);
3541
3542 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3543 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3544 if (!ctx || !new_attrs || !tmp_attrs)
3545 goto out_free;
3546
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003547 /*
3548 * Calculate the attrs of the default pwq.
3549 * If the user configured cpumask doesn't overlap with the
3550 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3551 */
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003552 copy_workqueue_attrs(new_attrs, attrs);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08003553 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003554 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3555 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003556
3557 /*
3558 * We may create multiple pwqs with differing cpumasks. Make a
3559 * copy of @new_attrs which will be modified and used to obtain
3560 * pools.
3561 */
3562 copy_workqueue_attrs(tmp_attrs, new_attrs);
3563
3564 /*
3565 * If something goes wrong during CPU up/down, we'll fall back to
3566 * the default pwq covering whole @attrs->cpumask. Always create
3567 * it even if we don't use it immediately.
3568 */
3569 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3570 if (!ctx->dfl_pwq)
3571 goto out_free;
3572
3573 for_each_node(node) {
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003574 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003575 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3576 if (!ctx->pwq_tbl[node])
3577 goto out_free;
3578 } else {
3579 ctx->dfl_pwq->refcnt++;
3580 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3581 }
3582 }
3583
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003584 /* save the user configured attrs and sanitize it. */
3585 copy_workqueue_attrs(new_attrs, attrs);
3586 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003587 ctx->attrs = new_attrs;
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003588
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003589 ctx->wq = wq;
3590 free_workqueue_attrs(tmp_attrs);
3591 return ctx;
3592
3593out_free:
3594 free_workqueue_attrs(tmp_attrs);
3595 free_workqueue_attrs(new_attrs);
3596 apply_wqattrs_cleanup(ctx);
3597 return NULL;
3598}
3599
3600/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3601static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3602{
3603 int node;
3604
3605 /* all pwqs have been created successfully, let's install'em */
3606 mutex_lock(&ctx->wq->mutex);
3607
3608 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3609
3610 /* save the previous pwq and install the new one */
3611 for_each_node(node)
3612 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3613 ctx->pwq_tbl[node]);
3614
3615 /* @dfl_pwq might not have been used, ensure it's linked */
3616 link_pwq(ctx->dfl_pwq);
3617 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3618
3619 mutex_unlock(&ctx->wq->mutex);
3620}
3621
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003622static void apply_wqattrs_lock(void)
3623{
3624 /* CPUs should stay stable across pwq creations and installations */
3625 get_online_cpus();
3626 mutex_lock(&wq_pool_mutex);
3627}
3628
3629static void apply_wqattrs_unlock(void)
3630{
3631 mutex_unlock(&wq_pool_mutex);
3632 put_online_cpus();
3633}
3634
3635static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3636 const struct workqueue_attrs *attrs)
3637{
3638 struct apply_wqattrs_ctx *ctx;
3639 int ret = -ENOMEM;
3640
3641 /* only unbound workqueues can change attributes */
3642 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3643 return -EINVAL;
3644
3645 /* creating multiple pwqs breaks ordering guarantee */
3646 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3647 return -EINVAL;
3648
3649 ctx = apply_wqattrs_prepare(wq, attrs);
3650
3651 /* the ctx has been prepared successfully, let's commit it */
3652 if (ctx) {
3653 apply_wqattrs_commit(ctx);
3654 ret = 0;
3655 }
3656
3657 apply_wqattrs_cleanup(ctx);
3658
3659 return ret;
3660}
3661
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003662/**
3663 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3664 * @wq: the target workqueue
3665 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3666 *
Tejun Heo4c16bd32013-04-01 11:23:36 -07003667 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3668 * machines, this function maps a separate pwq to each NUMA node with
3669 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3670 * NUMA node it was issued on. Older pwqs are released as in-flight work
3671 * items finish. Note that a work item which repeatedly requeues itself
3672 * back-to-back will stay on its current pwq.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003673 *
Yacine Belkadid185af32013-07-31 14:59:24 -07003674 * Performs GFP_KERNEL allocations.
3675 *
3676 * Return: 0 on success and -errno on failure.
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003677 */
3678int apply_workqueue_attrs(struct workqueue_struct *wq,
3679 const struct workqueue_attrs *attrs)
3680{
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003681 int ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003682
Lai Jiangshana0111cf2015-05-19 18:03:47 +08003683 apply_wqattrs_lock();
3684 ret = apply_workqueue_attrs_locked(wq, attrs);
3685 apply_wqattrs_unlock();
Lai Jiangshan2d5f0762015-04-27 17:58:38 +08003686
Tejun Heo48621252013-04-01 11:23:31 -07003687 return ret;
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003688}
3689
Tejun Heo4c16bd32013-04-01 11:23:36 -07003690/**
3691 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3692 * @wq: the target workqueue
3693 * @cpu: the CPU coming up or going down
3694 * @online: whether @cpu is coming up or going down
3695 *
3696 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3697 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3698 * @wq accordingly.
3699 *
3700 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3701 * falls back to @wq->dfl_pwq which may not be optimal but is always
3702 * correct.
3703 *
3704 * Note that when the last allowed CPU of a NUMA node goes offline for a
3705 * workqueue with a cpumask spanning multiple nodes, the workers which were
3706 * already executing the work items for the workqueue will lose their CPU
3707 * affinity and may execute on any CPU. This is similar to how per-cpu
3708 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3709 * affinity, it's the user's responsibility to flush the work item from
3710 * CPU_DOWN_PREPARE.
3711 */
3712static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3713 bool online)
3714{
3715 int node = cpu_to_node(cpu);
3716 int cpu_off = online ? -1 : cpu;
3717 struct pool_workqueue *old_pwq = NULL, *pwq;
3718 struct workqueue_attrs *target_attrs;
3719 cpumask_t *cpumask;
3720
3721 lockdep_assert_held(&wq_pool_mutex);
3722
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003723 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3724 wq->unbound_attrs->no_numa)
Tejun Heo4c16bd32013-04-01 11:23:36 -07003725 return;
3726
3727 /*
3728 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3729 * Let's use a preallocated one. The following buf is protected by
3730 * CPU hotplug exclusion.
3731 */
3732 target_attrs = wq_update_unbound_numa_attrs_buf;
3733 cpumask = target_attrs->cpumask;
3734
Tejun Heo4c16bd32013-04-01 11:23:36 -07003735 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3736 pwq = unbound_pwq_by_node(wq, node);
3737
3738 /*
3739 * Let's determine what needs to be done. If the target cpumask is
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003740 * different from the default pwq's, we need to compare it to @pwq's
3741 * and create a new one if they don't match. If the target cpumask
3742 * equals the default pwq's, the default pwq should be used.
Tejun Heo4c16bd32013-04-01 11:23:36 -07003743 */
Lai Jiangshan042f7df2015-04-30 17:16:12 +08003744 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
Tejun Heo4c16bd32013-04-01 11:23:36 -07003745 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003746 return;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003747 } else {
Daeseok Youn534a3fb2014-04-18 09:08:14 +09003748 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003749 }
3750
Tejun Heo4c16bd32013-04-01 11:23:36 -07003751 /* create a new pwq */
3752 pwq = alloc_unbound_pwq(wq, target_attrs);
3753 if (!pwq) {
Fabian Frederick2d916032014-05-12 13:59:35 -04003754 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3755 wq->name);
Daeseok Youn77f300b2014-04-16 14:32:29 +09003756 goto use_dfl_pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003757 }
3758
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003759 /* Install the new pwq. */
Tejun Heo4c16bd32013-04-01 11:23:36 -07003760 mutex_lock(&wq->mutex);
3761 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3762 goto out_unlock;
3763
3764use_dfl_pwq:
Lai Jiangshanf7142ed2015-05-12 20:32:30 +08003765 mutex_lock(&wq->mutex);
Tejun Heo4c16bd32013-04-01 11:23:36 -07003766 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3767 get_pwq(wq->dfl_pwq);
3768 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3769 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3770out_unlock:
3771 mutex_unlock(&wq->mutex);
3772 put_pwq_unlocked(old_pwq);
3773}
3774
Tejun Heo30cdf242013-03-12 11:29:57 -07003775static int alloc_and_link_pwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776{
Tejun Heo49e3cf42013-03-12 11:29:58 -07003777 bool highpri = wq->flags & WQ_HIGHPRI;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003778 int cpu, ret;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01003779
Tejun Heo30cdf242013-03-12 11:29:57 -07003780 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo420c0dd2013-03-12 11:29:59 -07003781 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3782 if (!wq->cpu_pwqs)
Tejun Heo30cdf242013-03-12 11:29:57 -07003783 return -ENOMEM;
3784
3785 for_each_possible_cpu(cpu) {
Tejun Heo7fb98ea2013-03-12 11:30:00 -07003786 struct pool_workqueue *pwq =
3787 per_cpu_ptr(wq->cpu_pwqs, cpu);
Tejun Heo7a62c2c2013-03-12 11:30:03 -07003788 struct worker_pool *cpu_pools =
Tejun Heof02ae732013-03-12 11:30:03 -07003789 per_cpu(cpu_worker_pools, cpu);
Tejun Heo30cdf242013-03-12 11:29:57 -07003790
Tejun Heof147f292013-04-01 11:23:35 -07003791 init_pwq(pwq, wq, &cpu_pools[highpri]);
3792
3793 mutex_lock(&wq->mutex);
Tejun Heo1befcf32013-04-01 11:23:35 -07003794 link_pwq(pwq);
Tejun Heof147f292013-04-01 11:23:35 -07003795 mutex_unlock(&wq->mutex);
Tejun Heo30cdf242013-03-12 11:29:57 -07003796 }
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003797 return 0;
Tejun Heo8a2b7532013-09-05 12:30:04 -04003798 } else if (wq->flags & __WQ_ORDERED) {
3799 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3800 /* there should only be single pwq for ordering guarantee */
3801 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3802 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3803 "ordering guarantee broken for workqueue %s\n", wq->name);
3804 return ret;
Tejun Heo30cdf242013-03-12 11:29:57 -07003805 } else {
Tejun Heo9e8cd2f2013-03-12 11:30:04 -07003806 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
Tejun Heo30cdf242013-03-12 11:29:57 -07003807 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07003808}
3809
Tejun Heof3421792010-07-02 10:03:51 +02003810static int wq_clamp_max_active(int max_active, unsigned int flags,
3811 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02003812{
Tejun Heof3421792010-07-02 10:03:51 +02003813 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3814
3815 if (max_active < 1 || max_active > lim)
Valentin Ilie044c7822012-08-19 00:52:42 +03003816 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3817 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003818
Tejun Heof3421792010-07-02 10:03:51 +02003819 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02003820}
3821
Tejun Heob196be82012-01-10 15:11:35 -08003822struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
Tejun Heod320c032010-06-29 10:07:14 +02003823 unsigned int flags,
3824 int max_active,
3825 struct lock_class_key *key,
Tejun Heob196be82012-01-10 15:11:35 -08003826 const char *lock_name, ...)
Oleg Nesterov3af244332007-05-09 02:34:09 -07003827{
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003828 size_t tbl_size = 0;
Tejun Heoecf68812013-04-01 11:23:34 -07003829 va_list args;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003830 struct workqueue_struct *wq;
Tejun Heo49e3cf42013-03-12 11:29:58 -07003831 struct pool_workqueue *pwq;
Tejun Heob196be82012-01-10 15:11:35 -08003832
Viresh Kumarcee22a12013-04-08 16:45:40 +05303833 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3834 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3835 flags |= WQ_UNBOUND;
3836
Tejun Heoecf68812013-04-01 11:23:34 -07003837 /* allocate wq and format name */
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003838 if (flags & WQ_UNBOUND)
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08003839 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
Tejun Heodf2d5ae2013-04-01 11:23:35 -07003840
3841 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
Tejun Heob196be82012-01-10 15:11:35 -08003842 if (!wq)
Tejun Heod2c1d402013-03-12 11:30:04 -07003843 return NULL;
Tejun Heob196be82012-01-10 15:11:35 -08003844
Tejun Heo6029a912013-04-01 11:23:34 -07003845 if (flags & WQ_UNBOUND) {
3846 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3847 if (!wq->unbound_attrs)
3848 goto err_free_wq;
3849 }
3850
Tejun Heoecf68812013-04-01 11:23:34 -07003851 va_start(args, lock_name);
3852 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
Tejun Heob196be82012-01-10 15:11:35 -08003853 va_end(args);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003854
Tejun Heod320c032010-06-29 10:07:14 +02003855 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heob196be82012-01-10 15:11:35 -08003856 max_active = wq_clamp_max_active(max_active, flags, wq->name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003857
Tejun Heob196be82012-01-10 15:11:35 -08003858 /* init wq */
Tejun Heo97e37d72010-06-29 10:07:10 +02003859 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003860 wq->saved_max_active = max_active;
Lai Jiangshan3c25a552013-03-25 16:57:17 -07003861 mutex_init(&wq->mutex);
Tejun Heo112202d2013-02-13 19:29:12 -08003862 atomic_set(&wq->nr_pwqs_to_flush, 0);
Tejun Heo30cdf242013-03-12 11:29:57 -07003863 INIT_LIST_HEAD(&wq->pwqs);
Tejun Heo73f53c42010-06-29 10:07:11 +02003864 INIT_LIST_HEAD(&wq->flusher_queue);
3865 INIT_LIST_HEAD(&wq->flusher_overflow);
Tejun Heo493a1722013-03-12 11:29:59 -07003866 INIT_LIST_HEAD(&wq->maydays);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003867
Johannes Bergeb13ba82008-01-16 09:51:58 +01003868 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07003869 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003870
Tejun Heo30cdf242013-03-12 11:29:57 -07003871 if (alloc_and_link_pwqs(wq) < 0)
Tejun Heod2c1d402013-03-12 11:30:04 -07003872 goto err_free_wq;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003873
Tejun Heo493008a2013-03-12 11:30:03 -07003874 /*
3875 * Workqueues which may be used during memory reclaim should
3876 * have a rescuer to guarantee forward progress.
3877 */
3878 if (flags & WQ_MEM_RECLAIM) {
Tejun Heoe22bee72010-06-29 10:07:14 +02003879 struct worker *rescuer;
3880
Lai Jiangshanf7537df2014-07-15 17:24:15 +08003881 rescuer = alloc_worker(NUMA_NO_NODE);
Tejun Heoe22bee72010-06-29 10:07:14 +02003882 if (!rescuer)
Tejun Heod2c1d402013-03-12 11:30:04 -07003883 goto err_destroy;
Tejun Heoe22bee72010-06-29 10:07:14 +02003884
Tejun Heo111c2252013-01-17 17:16:24 -08003885 rescuer->rescue_wq = wq;
3886 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
Tejun Heob196be82012-01-10 15:11:35 -08003887 wq->name);
Tejun Heod2c1d402013-03-12 11:30:04 -07003888 if (IS_ERR(rescuer->task)) {
3889 kfree(rescuer);
3890 goto err_destroy;
3891 }
Tejun Heoe22bee72010-06-29 10:07:14 +02003892
Tejun Heod2c1d402013-03-12 11:30:04 -07003893 wq->rescuer = rescuer;
Tejun Heo14a40ff2013-03-19 13:45:20 -07003894 rescuer->task->flags |= PF_NO_SETAFFINITY;
Tejun Heoe22bee72010-06-29 10:07:14 +02003895 wake_up_process(rescuer->task);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003896 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07003897
Tejun Heo226223a2013-03-12 11:30:05 -07003898 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3899 goto err_destroy;
3900
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003901 /*
Lai Jiangshan68e13a62013-03-25 16:57:17 -07003902 * wq_pool_mutex protects global freeze state and workqueues list.
3903 * Grab it, adjust max_active and add the new @wq to workqueues
3904 * list.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003905 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07003906 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003907
Lai Jiangshana357fc02013-03-25 16:57:19 -07003908 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07003909 for_each_pwq(pwq, wq)
3910 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07003911 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003912
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003913 list_add_tail_rcu(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003914
Lai Jiangshan68e13a62013-03-25 16:57:17 -07003915 mutex_unlock(&wq_pool_mutex);
Tejun Heo15376632010-06-29 10:07:11 +02003916
Oleg Nesterov3af244332007-05-09 02:34:09 -07003917 return wq;
Tejun Heod2c1d402013-03-12 11:30:04 -07003918
3919err_free_wq:
Tejun Heo6029a912013-04-01 11:23:34 -07003920 free_workqueue_attrs(wq->unbound_attrs);
Tejun Heod2c1d402013-03-12 11:30:04 -07003921 kfree(wq);
3922 return NULL;
3923err_destroy:
3924 destroy_workqueue(wq);
Tejun Heo4690c4a2010-06-29 10:07:10 +02003925 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003926}
Tejun Heod320c032010-06-29 10:07:14 +02003927EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003928
3929/**
3930 * destroy_workqueue - safely terminate a workqueue
3931 * @wq: target workqueue
3932 *
3933 * Safely destroy a workqueue. All work currently pending will be done first.
3934 */
3935void destroy_workqueue(struct workqueue_struct *wq)
3936{
Tejun Heo49e3cf42013-03-12 11:29:58 -07003937 struct pool_workqueue *pwq;
Tejun Heo4c16bd32013-04-01 11:23:36 -07003938 int node;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003939
Tejun Heo9c5a2ba2011-04-05 18:01:44 +02003940 /* drain it before proceeding with destruction */
3941 drain_workqueue(wq);
Tejun Heoc8efcc22010-12-20 19:32:04 +01003942
Tejun Heo6183c002013-03-12 11:29:57 -07003943 /* sanity checks */
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07003944 mutex_lock(&wq->mutex);
Tejun Heo49e3cf42013-03-12 11:29:58 -07003945 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07003946 int i;
3947
Tejun Heo76af4d92013-03-12 11:30:00 -07003948 for (i = 0; i < WORK_NR_COLORS; i++) {
3949 if (WARN_ON(pwq->nr_in_flight[i])) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07003950 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07003951 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07003952 }
3953 }
3954
Lai Jiangshan5c529592013-04-04 10:05:38 +08003955 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
Tejun Heo8864b4e2013-03-12 11:30:04 -07003956 WARN_ON(pwq->nr_active) ||
Tejun Heo76af4d92013-03-12 11:30:00 -07003957 WARN_ON(!list_empty(&pwq->delayed_works))) {
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07003958 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07003959 return;
Tejun Heo76af4d92013-03-12 11:30:00 -07003960 }
Tejun Heo6183c002013-03-12 11:29:57 -07003961 }
Lai Jiangshanb09f4fd2013-03-25 16:57:18 -07003962 mutex_unlock(&wq->mutex);
Tejun Heo6183c002013-03-12 11:29:57 -07003963
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003964 /*
3965 * wq list is used to freeze wq, remove from list after
3966 * flushing is complete in case freeze races us.
3967 */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07003968 mutex_lock(&wq_pool_mutex);
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003969 list_del_rcu(&wq->list);
Lai Jiangshan68e13a62013-03-25 16:57:17 -07003970 mutex_unlock(&wq_pool_mutex);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003971
Tejun Heo226223a2013-03-12 11:30:05 -07003972 workqueue_sysfs_unregister(wq);
3973
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003974 if (wq->rescuer)
Tejun Heoe22bee72010-06-29 10:07:14 +02003975 kthread_stop(wq->rescuer->task);
Tejun Heoe22bee72010-06-29 10:07:14 +02003976
Tejun Heo8864b4e2013-03-12 11:30:04 -07003977 if (!(wq->flags & WQ_UNBOUND)) {
3978 /*
3979 * The base ref is never dropped on per-cpu pwqs. Directly
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003980 * schedule RCU free.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003981 */
Tejun Heoe2dca7a2015-03-09 09:22:28 -04003982 call_rcu_sched(&wq->rcu, rcu_free_wq);
Tejun Heo8864b4e2013-03-12 11:30:04 -07003983 } else {
3984 /*
3985 * We're the sole accessor of @wq at this point. Directly
Tejun Heo4c16bd32013-04-01 11:23:36 -07003986 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3987 * @wq will be freed when the last pwq is released.
Tejun Heo8864b4e2013-03-12 11:30:04 -07003988 */
Tejun Heo4c16bd32013-04-01 11:23:36 -07003989 for_each_node(node) {
3990 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3991 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
3992 put_pwq_unlocked(pwq);
3993 }
3994
3995 /*
3996 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
3997 * put. Don't access it afterwards.
3998 */
3999 pwq = wq->dfl_pwq;
4000 wq->dfl_pwq = NULL;
Tejun Heodce90d42013-04-01 11:23:35 -07004001 put_pwq_unlocked(pwq);
Tejun Heo29c91e92013-03-12 11:30:03 -07004002 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07004003}
4004EXPORT_SYMBOL_GPL(destroy_workqueue);
4005
Tejun Heodcd989c2010-06-29 10:07:14 +02004006/**
4007 * workqueue_set_max_active - adjust max_active of a workqueue
4008 * @wq: target workqueue
4009 * @max_active: new max_active value.
4010 *
4011 * Set max_active of @wq to @max_active.
4012 *
4013 * CONTEXT:
4014 * Don't call from IRQ context.
4015 */
4016void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4017{
Tejun Heo49e3cf42013-03-12 11:29:58 -07004018 struct pool_workqueue *pwq;
Tejun Heodcd989c2010-06-29 10:07:14 +02004019
Tejun Heo8719dce2013-03-12 11:30:04 -07004020 /* disallow meddling with max_active for ordered workqueues */
4021 if (WARN_ON(wq->flags & __WQ_ORDERED))
4022 return;
4023
Tejun Heof3421792010-07-02 10:03:51 +02004024 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02004025
Lai Jiangshana357fc02013-03-25 16:57:19 -07004026 mutex_lock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004027
4028 wq->saved_max_active = max_active;
4029
Tejun Heo699ce092013-03-13 16:51:35 -07004030 for_each_pwq(pwq, wq)
4031 pwq_adjust_max_active(pwq);
Tejun Heodcd989c2010-06-29 10:07:14 +02004032
Lai Jiangshana357fc02013-03-25 16:57:19 -07004033 mutex_unlock(&wq->mutex);
Tejun Heodcd989c2010-06-29 10:07:14 +02004034}
4035EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4036
4037/**
Tejun Heoe62676162013-03-12 17:41:37 -07004038 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4039 *
4040 * Determine whether %current is a workqueue rescuer. Can be used from
4041 * work functions to determine whether it's being run off the rescuer task.
Yacine Belkadid185af32013-07-31 14:59:24 -07004042 *
4043 * Return: %true if %current is a workqueue rescuer. %false otherwise.
Tejun Heoe62676162013-03-12 17:41:37 -07004044 */
4045bool current_is_workqueue_rescuer(void)
4046{
4047 struct worker *worker = current_wq_worker();
4048
Lai Jiangshan6a092df2013-03-20 03:28:03 +08004049 return worker && worker->rescue_wq;
Tejun Heoe62676162013-03-12 17:41:37 -07004050}
4051
4052/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004053 * workqueue_congested - test whether a workqueue is congested
4054 * @cpu: CPU in question
4055 * @wq: target workqueue
4056 *
4057 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4058 * no synchronization around this function and the test result is
4059 * unreliable and only useful as advisory hints or for debugging.
4060 *
Tejun Heod3251852013-05-10 11:10:17 -07004061 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4062 * Note that both per-cpu and unbound workqueues may be associated with
4063 * multiple pool_workqueues which have separate congested states. A
4064 * workqueue being congested on one CPU doesn't mean the workqueue is also
4065 * contested on other CPUs / NUMA nodes.
4066 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004067 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004068 * %true if congested, %false otherwise.
4069 */
Tejun Heod84ff052013-03-12 11:29:59 -07004070bool workqueue_congested(int cpu, struct workqueue_struct *wq)
Tejun Heodcd989c2010-06-29 10:07:14 +02004071{
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004072 struct pool_workqueue *pwq;
Tejun Heo76af4d92013-03-12 11:30:00 -07004073 bool ret;
4074
Lai Jiangshan88109452013-03-20 03:28:10 +08004075 rcu_read_lock_sched();
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004076
Tejun Heod3251852013-05-10 11:10:17 -07004077 if (cpu == WORK_CPU_UNBOUND)
4078 cpu = smp_processor_id();
4079
Tejun Heo7fb98ea2013-03-12 11:30:00 -07004080 if (!(wq->flags & WQ_UNBOUND))
4081 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4082 else
Tejun Heodf2d5ae2013-04-01 11:23:35 -07004083 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
Tejun Heodcd989c2010-06-29 10:07:14 +02004084
Tejun Heo76af4d92013-03-12 11:30:00 -07004085 ret = !list_empty(&pwq->delayed_works);
Lai Jiangshan88109452013-03-20 03:28:10 +08004086 rcu_read_unlock_sched();
Tejun Heo76af4d92013-03-12 11:30:00 -07004087
4088 return ret;
Tejun Heodcd989c2010-06-29 10:07:14 +02004089}
4090EXPORT_SYMBOL_GPL(workqueue_congested);
4091
4092/**
Tejun Heodcd989c2010-06-29 10:07:14 +02004093 * work_busy - test whether a work is currently pending or running
4094 * @work: the work to be tested
4095 *
4096 * Test whether @work is currently pending or running. There is no
4097 * synchronization around this function and the test result is
4098 * unreliable and only useful as advisory hints or for debugging.
Tejun Heodcd989c2010-06-29 10:07:14 +02004099 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004100 * Return:
Tejun Heodcd989c2010-06-29 10:07:14 +02004101 * OR'd bitmask of WORK_BUSY_* bits.
4102 */
4103unsigned int work_busy(struct work_struct *work)
4104{
Tejun Heofa1b54e2013-03-12 11:30:00 -07004105 struct worker_pool *pool;
Tejun Heodcd989c2010-06-29 10:07:14 +02004106 unsigned long flags;
4107 unsigned int ret = 0;
4108
Tejun Heodcd989c2010-06-29 10:07:14 +02004109 if (work_pending(work))
4110 ret |= WORK_BUSY_PENDING;
Tejun Heodcd989c2010-06-29 10:07:14 +02004111
Tejun Heofa1b54e2013-03-12 11:30:00 -07004112 local_irq_save(flags);
4113 pool = get_work_pool(work);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004114 if (pool) {
Tejun Heofa1b54e2013-03-12 11:30:00 -07004115 spin_lock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004116 if (find_worker_executing_work(pool, work))
4117 ret |= WORK_BUSY_RUNNING;
Tejun Heofa1b54e2013-03-12 11:30:00 -07004118 spin_unlock(&pool->lock);
Lai Jiangshan038366c2013-02-06 18:04:53 -08004119 }
Tejun Heofa1b54e2013-03-12 11:30:00 -07004120 local_irq_restore(flags);
Tejun Heodcd989c2010-06-29 10:07:14 +02004121
4122 return ret;
4123}
4124EXPORT_SYMBOL_GPL(work_busy);
4125
Tejun Heo3d1cb202013-04-30 15:27:22 -07004126/**
4127 * set_worker_desc - set description for the current work item
4128 * @fmt: printf-style format string
4129 * @...: arguments for the format string
4130 *
4131 * This function can be called by a running work function to describe what
4132 * the work item is about. If the worker task gets dumped, this
4133 * information will be printed out together to help debugging. The
4134 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4135 */
4136void set_worker_desc(const char *fmt, ...)
4137{
4138 struct worker *worker = current_wq_worker();
4139 va_list args;
4140
4141 if (worker) {
4142 va_start(args, fmt);
4143 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4144 va_end(args);
4145 worker->desc_valid = true;
4146 }
4147}
4148
4149/**
4150 * print_worker_info - print out worker information and description
4151 * @log_lvl: the log level to use when printing
4152 * @task: target task
4153 *
4154 * If @task is a worker and currently executing a work item, print out the
4155 * name of the workqueue being serviced and worker description set with
4156 * set_worker_desc() by the currently executing work item.
4157 *
4158 * This function can be safely called on any task as long as the
4159 * task_struct itself is accessible. While safe, this function isn't
4160 * synchronized and may print out mixups or garbages of limited length.
4161 */
4162void print_worker_info(const char *log_lvl, struct task_struct *task)
4163{
4164 work_func_t *fn = NULL;
4165 char name[WQ_NAME_LEN] = { };
4166 char desc[WORKER_DESC_LEN] = { };
4167 struct pool_workqueue *pwq = NULL;
4168 struct workqueue_struct *wq = NULL;
4169 bool desc_valid = false;
4170 struct worker *worker;
4171
4172 if (!(task->flags & PF_WQ_WORKER))
4173 return;
4174
4175 /*
4176 * This function is called without any synchronization and @task
4177 * could be in any state. Be careful with dereferences.
4178 */
4179 worker = probe_kthread_data(task);
4180
4181 /*
4182 * Carefully copy the associated workqueue's workfn and name. Keep
4183 * the original last '\0' in case the original contains garbage.
4184 */
4185 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4186 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4187 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4188 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4189
4190 /* copy worker description */
4191 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4192 if (desc_valid)
4193 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4194
4195 if (fn || name[0] || desc[0]) {
4196 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4197 if (desc[0])
4198 pr_cont(" (%s)", desc);
4199 pr_cont("\n");
4200 }
4201}
4202
Tejun Heo3494fc32015-03-09 09:22:28 -04004203static void pr_cont_pool_info(struct worker_pool *pool)
4204{
4205 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4206 if (pool->node != NUMA_NO_NODE)
4207 pr_cont(" node=%d", pool->node);
4208 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4209}
4210
4211static void pr_cont_work(bool comma, struct work_struct *work)
4212{
4213 if (work->func == wq_barrier_func) {
4214 struct wq_barrier *barr;
4215
4216 barr = container_of(work, struct wq_barrier, work);
4217
4218 pr_cont("%s BAR(%d)", comma ? "," : "",
4219 task_pid_nr(barr->task));
4220 } else {
4221 pr_cont("%s %pf", comma ? "," : "", work->func);
4222 }
4223}
4224
4225static void show_pwq(struct pool_workqueue *pwq)
4226{
4227 struct worker_pool *pool = pwq->pool;
4228 struct work_struct *work;
4229 struct worker *worker;
4230 bool has_in_flight = false, has_pending = false;
4231 int bkt;
4232
4233 pr_info(" pwq %d:", pool->id);
4234 pr_cont_pool_info(pool);
4235
4236 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4237 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4238
4239 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4240 if (worker->current_pwq == pwq) {
4241 has_in_flight = true;
4242 break;
4243 }
4244 }
4245 if (has_in_flight) {
4246 bool comma = false;
4247
4248 pr_info(" in-flight:");
4249 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4250 if (worker->current_pwq != pwq)
4251 continue;
4252
4253 pr_cont("%s %d%s:%pf", comma ? "," : "",
4254 task_pid_nr(worker->task),
4255 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4256 worker->current_func);
4257 list_for_each_entry(work, &worker->scheduled, entry)
4258 pr_cont_work(false, work);
4259 comma = true;
4260 }
4261 pr_cont("\n");
4262 }
4263
4264 list_for_each_entry(work, &pool->worklist, entry) {
4265 if (get_work_pwq(work) == pwq) {
4266 has_pending = true;
4267 break;
4268 }
4269 }
4270 if (has_pending) {
4271 bool comma = false;
4272
4273 pr_info(" pending:");
4274 list_for_each_entry(work, &pool->worklist, entry) {
4275 if (get_work_pwq(work) != pwq)
4276 continue;
4277
4278 pr_cont_work(comma, work);
4279 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4280 }
4281 pr_cont("\n");
4282 }
4283
4284 if (!list_empty(&pwq->delayed_works)) {
4285 bool comma = false;
4286
4287 pr_info(" delayed:");
4288 list_for_each_entry(work, &pwq->delayed_works, entry) {
4289 pr_cont_work(comma, work);
4290 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4291 }
4292 pr_cont("\n");
4293 }
4294}
4295
4296/**
4297 * show_workqueue_state - dump workqueue state
4298 *
4299 * Called from a sysrq handler and prints out all busy workqueues and
4300 * pools.
4301 */
4302void show_workqueue_state(void)
4303{
4304 struct workqueue_struct *wq;
4305 struct worker_pool *pool;
4306 unsigned long flags;
4307 int pi;
4308
4309 rcu_read_lock_sched();
4310
4311 pr_info("Showing busy workqueues and worker pools:\n");
4312
4313 list_for_each_entry_rcu(wq, &workqueues, list) {
4314 struct pool_workqueue *pwq;
4315 bool idle = true;
4316
4317 for_each_pwq(pwq, wq) {
4318 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4319 idle = false;
4320 break;
4321 }
4322 }
4323 if (idle)
4324 continue;
4325
4326 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4327
4328 for_each_pwq(pwq, wq) {
4329 spin_lock_irqsave(&pwq->pool->lock, flags);
4330 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4331 show_pwq(pwq);
4332 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4333 }
4334 }
4335
4336 for_each_pool(pool, pi) {
4337 struct worker *worker;
4338 bool first = true;
4339
4340 spin_lock_irqsave(&pool->lock, flags);
4341 if (pool->nr_workers == pool->nr_idle)
4342 goto next_pool;
4343
4344 pr_info("pool %d:", pool->id);
4345 pr_cont_pool_info(pool);
4346 pr_cont(" workers=%d", pool->nr_workers);
4347 if (pool->manager)
4348 pr_cont(" manager: %d",
4349 task_pid_nr(pool->manager->task));
4350 list_for_each_entry(worker, &pool->idle_list, entry) {
4351 pr_cont(" %s%d", first ? "idle: " : "",
4352 task_pid_nr(worker->task));
4353 first = false;
4354 }
4355 pr_cont("\n");
4356 next_pool:
4357 spin_unlock_irqrestore(&pool->lock, flags);
4358 }
4359
4360 rcu_read_unlock_sched();
4361}
4362
Tejun Heodb7bccf2010-06-29 10:07:12 +02004363/*
4364 * CPU hotplug.
4365 *
Tejun Heoe22bee72010-06-29 10:07:14 +02004366 * There are two challenges in supporting CPU hotplug. Firstly, there
Tejun Heo112202d2013-02-13 19:29:12 -08004367 * are a lot of assumptions on strong associations among work, pwq and
Tejun Heo706026c2013-01-24 11:01:34 -08004368 * pool which make migrating pending and scheduled works very
Tejun Heoe22bee72010-06-29 10:07:14 +02004369 * difficult to implement without impacting hot paths. Secondly,
Tejun Heo94cf58b2013-01-24 11:01:33 -08004370 * worker pools serve mix of short, long and very long running works making
Tejun Heoe22bee72010-06-29 10:07:14 +02004371 * blocked draining impractical.
4372 *
Tejun Heo24647572013-01-24 11:01:33 -08004373 * This is solved by allowing the pools to be disassociated from the CPU
Tejun Heo628c78e2012-07-17 12:39:27 -07004374 * running as an unbound one and allowing it to be reattached later if the
4375 * cpu comes back online.
Tejun Heodb7bccf2010-06-29 10:07:12 +02004376 */
4377
Tejun Heo706026c2013-01-24 11:01:34 -08004378static void wq_unbind_fn(struct work_struct *work)
Tejun Heodb7bccf2010-06-29 10:07:12 +02004379{
Tejun Heo38db41d2013-01-24 11:01:34 -08004380 int cpu = smp_processor_id();
Tejun Heo4ce62e92012-07-13 22:16:44 -07004381 struct worker_pool *pool;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004382 struct worker *worker;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004383
Tejun Heof02ae732013-03-12 11:30:03 -07004384 for_each_cpu_worker_pool(pool, cpu) {
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004385 mutex_lock(&pool->attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004386 spin_lock_irq(&pool->lock);
4387
4388 /*
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004389 * We've blocked all attach/detach operations. Make all workers
Tejun Heo94cf58b2013-01-24 11:01:33 -08004390 * unbound and set DISASSOCIATED. Before this, all workers
4391 * except for the ones which are still executing works from
4392 * before the last CPU down must be on the cpu. After
4393 * this, they may become diasporas.
4394 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004395 for_each_pool_worker(worker, pool)
Tejun Heoc9e7cf22013-01-24 11:01:33 -08004396 worker->flags |= WORKER_UNBOUND;
Tejun Heodb7bccf2010-06-29 10:07:12 +02004397
Tejun Heo24647572013-01-24 11:01:33 -08004398 pool->flags |= POOL_DISASSOCIATED;
Tejun Heof2d5a0e2012-07-17 12:39:26 -07004399
Tejun Heo94cf58b2013-01-24 11:01:33 -08004400 spin_unlock_irq(&pool->lock);
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004401 mutex_unlock(&pool->attach_mutex);
Tejun Heoe22bee72010-06-29 10:07:14 +02004402
Lai Jiangshaneb283422013-03-08 15:18:28 -08004403 /*
4404 * Call schedule() so that we cross rq->lock and thus can
4405 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4406 * This is necessary as scheduler callbacks may be invoked
4407 * from other cpus.
4408 */
4409 schedule();
Tejun Heo628c78e2012-07-17 12:39:27 -07004410
Lai Jiangshaneb283422013-03-08 15:18:28 -08004411 /*
4412 * Sched callbacks are disabled now. Zap nr_running.
4413 * After this, nr_running stays zero and need_more_worker()
4414 * and keep_working() are always true as long as the
4415 * worklist is not empty. This pool now behaves as an
4416 * unbound (in terms of concurrency management) pool which
4417 * are served by workers tied to the pool.
4418 */
Tejun Heoe19e3972013-01-24 11:39:44 -08004419 atomic_set(&pool->nr_running, 0);
Lai Jiangshaneb283422013-03-08 15:18:28 -08004420
4421 /*
4422 * With concurrency management just turned off, a busy
4423 * worker blocking could lead to lengthy stalls. Kick off
4424 * unbound chain execution of currently pending work items.
4425 */
4426 spin_lock_irq(&pool->lock);
4427 wake_up_worker(pool);
4428 spin_unlock_irq(&pool->lock);
4429 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004430}
4431
Tejun Heobd7c0892013-03-19 13:45:21 -07004432/**
4433 * rebind_workers - rebind all workers of a pool to the associated CPU
4434 * @pool: pool of interest
4435 *
Tejun Heoa9ab7752013-03-19 13:45:21 -07004436 * @pool->cpu is coming online. Rebind all workers to the CPU.
Tejun Heobd7c0892013-03-19 13:45:21 -07004437 */
4438static void rebind_workers(struct worker_pool *pool)
4439{
Tejun Heoa9ab7752013-03-19 13:45:21 -07004440 struct worker *worker;
Tejun Heobd7c0892013-03-19 13:45:21 -07004441
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004442 lockdep_assert_held(&pool->attach_mutex);
Tejun Heobd7c0892013-03-19 13:45:21 -07004443
Tejun Heoa9ab7752013-03-19 13:45:21 -07004444 /*
4445 * Restore CPU affinity of all workers. As all idle workers should
4446 * be on the run-queue of the associated CPU before any local
4447 * wake-ups for concurrency management happen, restore CPU affinty
4448 * of all workers first and then clear UNBOUND. As we're called
4449 * from CPU_ONLINE, the following shouldn't fail.
4450 */
Lai Jiangshanda028462014-05-20 17:46:31 +08004451 for_each_pool_worker(worker, pool)
Tejun Heoa9ab7752013-03-19 13:45:21 -07004452 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4453 pool->attrs->cpumask) < 0);
4454
4455 spin_lock_irq(&pool->lock);
Lai Jiangshan3de5e882014-06-03 15:33:27 +08004456 pool->flags &= ~POOL_DISASSOCIATED;
Tejun Heoa9ab7752013-03-19 13:45:21 -07004457
Lai Jiangshanda028462014-05-20 17:46:31 +08004458 for_each_pool_worker(worker, pool) {
Tejun Heoa9ab7752013-03-19 13:45:21 -07004459 unsigned int worker_flags = worker->flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004460
4461 /*
Tejun Heoa9ab7752013-03-19 13:45:21 -07004462 * A bound idle worker should actually be on the runqueue
4463 * of the associated CPU for local wake-ups targeting it to
4464 * work. Kick all idle workers so that they migrate to the
4465 * associated CPU. Doing this in the same loop as
4466 * replacing UNBOUND with REBOUND is safe as no worker will
4467 * be bound before @pool->lock is released.
Tejun Heobd7c0892013-03-19 13:45:21 -07004468 */
Tejun Heoa9ab7752013-03-19 13:45:21 -07004469 if (worker_flags & WORKER_IDLE)
4470 wake_up_process(worker->task);
4471
4472 /*
4473 * We want to clear UNBOUND but can't directly call
4474 * worker_clr_flags() or adjust nr_running. Atomically
4475 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4476 * @worker will clear REBOUND using worker_clr_flags() when
4477 * it initiates the next execution cycle thus restoring
4478 * concurrency management. Note that when or whether
4479 * @worker clears REBOUND doesn't affect correctness.
4480 *
4481 * ACCESS_ONCE() is necessary because @worker->flags may be
4482 * tested without holding any lock in
4483 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4484 * fail incorrectly leading to premature concurrency
4485 * management operations.
4486 */
4487 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4488 worker_flags |= WORKER_REBOUND;
4489 worker_flags &= ~WORKER_UNBOUND;
4490 ACCESS_ONCE(worker->flags) = worker_flags;
Tejun Heobd7c0892013-03-19 13:45:21 -07004491 }
4492
Tejun Heoa9ab7752013-03-19 13:45:21 -07004493 spin_unlock_irq(&pool->lock);
Tejun Heobd7c0892013-03-19 13:45:21 -07004494}
4495
Tejun Heo7dbc7252013-03-19 13:45:21 -07004496/**
4497 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4498 * @pool: unbound pool of interest
4499 * @cpu: the CPU which is coming up
4500 *
4501 * An unbound pool may end up with a cpumask which doesn't have any online
4502 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4503 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4504 * online CPU before, cpus_allowed of all its workers should be restored.
4505 */
4506static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4507{
4508 static cpumask_t cpumask;
4509 struct worker *worker;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004510
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004511 lockdep_assert_held(&pool->attach_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004512
4513 /* is @cpu allowed for @pool? */
4514 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4515 return;
4516
4517 /* is @cpu the only online CPU? */
4518 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4519 if (cpumask_weight(&cpumask) != 1)
4520 return;
4521
4522 /* as we're called from CPU_ONLINE, the following shouldn't fail */
Lai Jiangshanda028462014-05-20 17:46:31 +08004523 for_each_pool_worker(worker, pool)
Tejun Heo7dbc7252013-03-19 13:45:21 -07004524 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4525 pool->attrs->cpumask) < 0);
4526}
4527
Tejun Heo8db25e72012-07-17 12:39:28 -07004528/*
4529 * Workqueues should be brought up before normal priority CPU notifiers.
4530 * This will be registered high priority CPU notifier.
4531 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04004532static int workqueue_cpu_up_callback(struct notifier_block *nfb,
Tejun Heo8db25e72012-07-17 12:39:28 -07004533 unsigned long action,
4534 void *hcpu)
Oleg Nesterov3af244332007-05-09 02:34:09 -07004535{
Tejun Heod84ff052013-03-12 11:29:59 -07004536 int cpu = (unsigned long)hcpu;
Tejun Heo4ce62e92012-07-13 22:16:44 -07004537 struct worker_pool *pool;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004538 struct workqueue_struct *wq;
Tejun Heo7dbc7252013-03-19 13:45:21 -07004539 int pi;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004540
Tejun Heo8db25e72012-07-17 12:39:28 -07004541 switch (action & ~CPU_TASKS_FROZEN) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004542 case CPU_UP_PREPARE:
Tejun Heof02ae732013-03-12 11:30:03 -07004543 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo3ce63372012-07-17 12:39:27 -07004544 if (pool->nr_workers)
4545 continue;
Lai Jiangshan051e1852014-07-22 13:03:02 +08004546 if (!create_worker(pool))
Tejun Heo3ce63372012-07-17 12:39:27 -07004547 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004548 }
Tejun Heodb7bccf2010-06-29 10:07:12 +02004549 break;
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07004550
Tejun Heo65758202012-07-17 12:39:26 -07004551 case CPU_DOWN_FAILED:
4552 case CPU_ONLINE:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004553 mutex_lock(&wq_pool_mutex);
Tejun Heo7dbc7252013-03-19 13:45:21 -07004554
4555 for_each_pool(pool, pi) {
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004556 mutex_lock(&pool->attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004557
Lai Jiangshanf05b5582014-06-03 15:33:27 +08004558 if (pool->cpu == cpu)
Tejun Heo7dbc7252013-03-19 13:45:21 -07004559 rebind_workers(pool);
Lai Jiangshanf05b5582014-06-03 15:33:27 +08004560 else if (pool->cpu < 0)
Tejun Heo7dbc7252013-03-19 13:45:21 -07004561 restore_unbound_workers_cpumask(pool, cpu);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004562
Lai Jiangshan92f9c5c2014-05-20 17:46:34 +08004563 mutex_unlock(&pool->attach_mutex);
Tejun Heo94cf58b2013-01-24 11:01:33 -08004564 }
Tejun Heo7dbc7252013-03-19 13:45:21 -07004565
Tejun Heo4c16bd32013-04-01 11:23:36 -07004566 /* update NUMA affinity of unbound workqueues */
4567 list_for_each_entry(wq, &workqueues, list)
4568 wq_update_unbound_numa(wq, cpu, true);
4569
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004570 mutex_unlock(&wq_pool_mutex);
Tejun Heo8db25e72012-07-17 12:39:28 -07004571 break;
Tejun Heo65758202012-07-17 12:39:26 -07004572 }
4573 return NOTIFY_OK;
4574}
4575
4576/*
4577 * Workqueues should be brought down after normal priority CPU notifiers.
4578 * This will be registered as low priority CPU notifier.
4579 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04004580static int workqueue_cpu_down_callback(struct notifier_block *nfb,
Tejun Heo65758202012-07-17 12:39:26 -07004581 unsigned long action,
4582 void *hcpu)
4583{
Tejun Heod84ff052013-03-12 11:29:59 -07004584 int cpu = (unsigned long)hcpu;
Tejun Heo8db25e72012-07-17 12:39:28 -07004585 struct work_struct unbind_work;
Tejun Heo4c16bd32013-04-01 11:23:36 -07004586 struct workqueue_struct *wq;
Tejun Heo8db25e72012-07-17 12:39:28 -07004587
Tejun Heo65758202012-07-17 12:39:26 -07004588 switch (action & ~CPU_TASKS_FROZEN) {
4589 case CPU_DOWN_PREPARE:
Tejun Heo4c16bd32013-04-01 11:23:36 -07004590 /* unbinding per-cpu workers should happen on the local CPU */
Tejun Heo706026c2013-01-24 11:01:34 -08004591 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
Joonsoo Kim7635d2f2012-08-15 23:25:41 +09004592 queue_work_on(cpu, system_highpri_wq, &unbind_work);
Tejun Heo4c16bd32013-04-01 11:23:36 -07004593
4594 /* update NUMA affinity of unbound workqueues */
4595 mutex_lock(&wq_pool_mutex);
4596 list_for_each_entry(wq, &workqueues, list)
4597 wq_update_unbound_numa(wq, cpu, false);
4598 mutex_unlock(&wq_pool_mutex);
4599
4600 /* wait for per-cpu unbinding to finish */
Tejun Heo8db25e72012-07-17 12:39:28 -07004601 flush_work(&unbind_work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05004602 destroy_work_on_stack(&unbind_work);
Tejun Heo8db25e72012-07-17 12:39:28 -07004603 break;
Tejun Heo65758202012-07-17 12:39:26 -07004604 }
4605 return NOTIFY_OK;
4606}
4607
Rusty Russell2d3854a2008-11-05 13:39:10 +11004608#ifdef CONFIG_SMP
Rusty Russell8ccad402009-01-16 15:31:15 -08004609
Rusty Russell2d3854a2008-11-05 13:39:10 +11004610struct work_for_cpu {
Tejun Heoed48ece2012-09-18 12:48:43 -07004611 struct work_struct work;
Rusty Russell2d3854a2008-11-05 13:39:10 +11004612 long (*fn)(void *);
4613 void *arg;
4614 long ret;
4615};
4616
Tejun Heoed48ece2012-09-18 12:48:43 -07004617static void work_for_cpu_fn(struct work_struct *work)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004618{
Tejun Heoed48ece2012-09-18 12:48:43 -07004619 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4620
Rusty Russell2d3854a2008-11-05 13:39:10 +11004621 wfc->ret = wfc->fn(wfc->arg);
4622}
4623
4624/**
4625 * work_on_cpu - run a function in user context on a particular cpu
4626 * @cpu: the cpu to run on
4627 * @fn: the function to run
4628 * @arg: the function arg
4629 *
Rusty Russell31ad9082009-01-16 15:31:15 -08004630 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06004631 * The caller must not hold any locks which would prevent @fn from completing.
Yacine Belkadid185af32013-07-31 14:59:24 -07004632 *
4633 * Return: The value @fn returns.
Rusty Russell2d3854a2008-11-05 13:39:10 +11004634 */
Tejun Heod84ff052013-03-12 11:29:59 -07004635long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
Rusty Russell2d3854a2008-11-05 13:39:10 +11004636{
Tejun Heoed48ece2012-09-18 12:48:43 -07004637 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
Rusty Russell2d3854a2008-11-05 13:39:10 +11004638
Tejun Heoed48ece2012-09-18 12:48:43 -07004639 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4640 schedule_work_on(cpu, &wfc.work);
Bjorn Helgaas12997d12013-11-18 11:00:29 -07004641 flush_work(&wfc.work);
Chuansheng Liu440a1132014-01-11 22:26:33 -05004642 destroy_work_on_stack(&wfc.work);
Rusty Russell2d3854a2008-11-05 13:39:10 +11004643 return wfc.ret;
4644}
4645EXPORT_SYMBOL_GPL(work_on_cpu);
4646#endif /* CONFIG_SMP */
4647
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004648#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10304649
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004650/**
4651 * freeze_workqueues_begin - begin freezing workqueues
4652 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01004653 * Start freezing workqueues. After this function returns, all freezable
Tejun Heoc5aa87b2013-03-13 16:51:36 -07004654 * workqueues will queue new works to their delayed_works list instead of
Tejun Heo706026c2013-01-24 11:01:34 -08004655 * pool->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004656 *
4657 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004658 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004659 */
4660void freeze_workqueues_begin(void)
4661{
Tejun Heo24b8a842013-03-12 11:29:58 -07004662 struct workqueue_struct *wq;
4663 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004664
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004665 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004666
Tejun Heo6183c002013-03-12 11:29:57 -07004667 WARN_ON_ONCE(workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004668 workqueue_freezing = true;
4669
Tejun Heo24b8a842013-03-12 11:29:58 -07004670 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004671 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004672 for_each_pwq(pwq, wq)
4673 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004674 mutex_unlock(&wq->mutex);
Tejun Heo24b8a842013-03-12 11:29:58 -07004675 }
Tejun Heo5bcab332013-03-13 19:47:40 -07004676
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004677 mutex_unlock(&wq_pool_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004678}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004679
4680/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01004681 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004682 *
4683 * Check whether freezing is complete. This function must be called
4684 * between freeze_workqueues_begin() and thaw_workqueues().
4685 *
4686 * CONTEXT:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004687 * Grabs and releases wq_pool_mutex.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004688 *
Yacine Belkadid185af32013-07-31 14:59:24 -07004689 * Return:
Tejun Heo58a69cb2011-02-16 09:25:31 +01004690 * %true if some freezable workqueues are still busy. %false if freezing
4691 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004692 */
4693bool freeze_workqueues_busy(void)
4694{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004695 bool busy = false;
Tejun Heo24b8a842013-03-12 11:29:58 -07004696 struct workqueue_struct *wq;
4697 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004698
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004699 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004700
Tejun Heo6183c002013-03-12 11:29:57 -07004701 WARN_ON_ONCE(!workqueue_freezing);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004702
Tejun Heo24b8a842013-03-12 11:29:58 -07004703 list_for_each_entry(wq, &workqueues, list) {
4704 if (!(wq->flags & WQ_FREEZABLE))
4705 continue;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004706 /*
4707 * nr_active is monotonically decreasing. It's safe
4708 * to peek without lock.
4709 */
Lai Jiangshan88109452013-03-20 03:28:10 +08004710 rcu_read_lock_sched();
Tejun Heo24b8a842013-03-12 11:29:58 -07004711 for_each_pwq(pwq, wq) {
Tejun Heo6183c002013-03-12 11:29:57 -07004712 WARN_ON_ONCE(pwq->nr_active < 0);
Tejun Heo112202d2013-02-13 19:29:12 -08004713 if (pwq->nr_active) {
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004714 busy = true;
Lai Jiangshan88109452013-03-20 03:28:10 +08004715 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004716 goto out_unlock;
4717 }
4718 }
Lai Jiangshan88109452013-03-20 03:28:10 +08004719 rcu_read_unlock_sched();
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004720 }
4721out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004722 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004723 return busy;
4724}
4725
4726/**
4727 * thaw_workqueues - thaw workqueues
4728 *
4729 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo706026c2013-01-24 11:01:34 -08004730 * frozen works are transferred to their respective pool worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004731 *
4732 * CONTEXT:
Lai Jiangshana357fc02013-03-25 16:57:19 -07004733 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004734 */
4735void thaw_workqueues(void)
4736{
Tejun Heo24b8a842013-03-12 11:29:58 -07004737 struct workqueue_struct *wq;
4738 struct pool_workqueue *pwq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004739
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004740 mutex_lock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004741
4742 if (!workqueue_freezing)
4743 goto out_unlock;
4744
Lai Jiangshan74b414e2014-05-22 19:01:16 +08004745 workqueue_freezing = false;
Tejun Heo8b03ae32010-06-29 10:07:12 +02004746
Tejun Heo24b8a842013-03-12 11:29:58 -07004747 /* restore max_active and repopulate worklist */
4748 list_for_each_entry(wq, &workqueues, list) {
Lai Jiangshana357fc02013-03-25 16:57:19 -07004749 mutex_lock(&wq->mutex);
Tejun Heo699ce092013-03-13 16:51:35 -07004750 for_each_pwq(pwq, wq)
4751 pwq_adjust_max_active(pwq);
Lai Jiangshana357fc02013-03-25 16:57:19 -07004752 mutex_unlock(&wq->mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004753 }
4754
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004755out_unlock:
Lai Jiangshan68e13a62013-03-25 16:57:17 -07004756 mutex_unlock(&wq_pool_mutex);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02004757}
4758#endif /* CONFIG_FREEZER */
4759
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004760static int workqueue_apply_unbound_cpumask(void)
4761{
4762 LIST_HEAD(ctxs);
4763 int ret = 0;
4764 struct workqueue_struct *wq;
4765 struct apply_wqattrs_ctx *ctx, *n;
4766
4767 lockdep_assert_held(&wq_pool_mutex);
4768
4769 list_for_each_entry(wq, &workqueues, list) {
4770 if (!(wq->flags & WQ_UNBOUND))
4771 continue;
4772 /* creating multiple pwqs breaks ordering guarantee */
4773 if (wq->flags & __WQ_ORDERED)
4774 continue;
4775
4776 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4777 if (!ctx) {
4778 ret = -ENOMEM;
4779 break;
4780 }
4781
4782 list_add_tail(&ctx->list, &ctxs);
4783 }
4784
4785 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4786 if (!ret)
4787 apply_wqattrs_commit(ctx);
4788 apply_wqattrs_cleanup(ctx);
4789 }
4790
4791 return ret;
4792}
4793
4794/**
4795 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4796 * @cpumask: the cpumask to set
4797 *
4798 * The low-level workqueues cpumask is a global cpumask that limits
4799 * the affinity of all unbound workqueues. This function check the @cpumask
4800 * and apply it to all unbound workqueues and updates all pwqs of them.
4801 *
4802 * Retun: 0 - Success
4803 * -EINVAL - Invalid @cpumask
4804 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4805 */
4806int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4807{
4808 int ret = -EINVAL;
4809 cpumask_var_t saved_cpumask;
4810
4811 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4812 return -ENOMEM;
4813
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004814 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4815 if (!cpumask_empty(cpumask)) {
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004816 apply_wqattrs_lock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004817
4818 /* save the old wq_unbound_cpumask. */
4819 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4820
4821 /* update wq_unbound_cpumask at first and apply it to wqs. */
4822 cpumask_copy(wq_unbound_cpumask, cpumask);
4823 ret = workqueue_apply_unbound_cpumask();
4824
4825 /* restore the wq_unbound_cpumask when failed. */
4826 if (ret < 0)
4827 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4828
Lai Jiangshana0111cf2015-05-19 18:03:47 +08004829 apply_wqattrs_unlock();
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004830 }
Lai Jiangshan042f7df2015-04-30 17:16:12 +08004831
4832 free_cpumask_var(saved_cpumask);
4833 return ret;
4834}
4835
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004836#ifdef CONFIG_SYSFS
4837/*
4838 * Workqueues with WQ_SYSFS flag set is visible to userland via
4839 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4840 * following attributes.
4841 *
4842 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4843 * max_active RW int : maximum number of in-flight work items
4844 *
4845 * Unbound workqueues have the following extra attributes.
4846 *
4847 * id RO int : the associated pool ID
4848 * nice RW int : nice value of the workers
4849 * cpumask RW mask : bitmask of allowed CPUs for the workers
4850 */
4851struct wq_device {
4852 struct workqueue_struct *wq;
4853 struct device dev;
4854};
4855
4856static struct workqueue_struct *dev_to_wq(struct device *dev)
4857{
4858 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4859
4860 return wq_dev->wq;
4861}
4862
4863static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4864 char *buf)
4865{
4866 struct workqueue_struct *wq = dev_to_wq(dev);
4867
4868 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4869}
4870static DEVICE_ATTR_RO(per_cpu);
4871
4872static ssize_t max_active_show(struct device *dev,
4873 struct device_attribute *attr, char *buf)
4874{
4875 struct workqueue_struct *wq = dev_to_wq(dev);
4876
4877 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4878}
4879
4880static ssize_t max_active_store(struct device *dev,
4881 struct device_attribute *attr, const char *buf,
4882 size_t count)
4883{
4884 struct workqueue_struct *wq = dev_to_wq(dev);
4885 int val;
4886
4887 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4888 return -EINVAL;
4889
4890 workqueue_set_max_active(wq, val);
4891 return count;
4892}
4893static DEVICE_ATTR_RW(max_active);
4894
4895static struct attribute *wq_sysfs_attrs[] = {
4896 &dev_attr_per_cpu.attr,
4897 &dev_attr_max_active.attr,
4898 NULL,
4899};
4900ATTRIBUTE_GROUPS(wq_sysfs);
4901
4902static ssize_t wq_pool_ids_show(struct device *dev,
4903 struct device_attribute *attr, char *buf)
4904{
4905 struct workqueue_struct *wq = dev_to_wq(dev);
4906 const char *delim = "";
4907 int node, written = 0;
4908
4909 rcu_read_lock_sched();
4910 for_each_node(node) {
4911 written += scnprintf(buf + written, PAGE_SIZE - written,
4912 "%s%d:%d", delim, node,
4913 unbound_pwq_by_node(wq, node)->pool->id);
4914 delim = " ";
4915 }
4916 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4917 rcu_read_unlock_sched();
4918
4919 return written;
4920}
4921
4922static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4923 char *buf)
4924{
4925 struct workqueue_struct *wq = dev_to_wq(dev);
4926 int written;
4927
4928 mutex_lock(&wq->mutex);
4929 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4930 mutex_unlock(&wq->mutex);
4931
4932 return written;
4933}
4934
4935/* prepare workqueue_attrs for sysfs store operations */
4936static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4937{
4938 struct workqueue_attrs *attrs;
4939
Lai Jiangshan899a94f2015-05-20 14:41:18 +08004940 lockdep_assert_held(&wq_pool_mutex);
4941
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004942 attrs = alloc_workqueue_attrs(GFP_KERNEL);
4943 if (!attrs)
4944 return NULL;
4945
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004946 copy_workqueue_attrs(attrs, wq->unbound_attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004947 return attrs;
4948}
4949
4950static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4951 const char *buf, size_t count)
4952{
4953 struct workqueue_struct *wq = dev_to_wq(dev);
4954 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08004955 int ret = -ENOMEM;
4956
4957 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004958
4959 attrs = wq_sysfs_prep_attrs(wq);
4960 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08004961 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004962
4963 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4964 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08004965 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004966 else
4967 ret = -EINVAL;
4968
Lai Jiangshand4d3e252015-05-19 18:03:48 +08004969out_unlock:
4970 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004971 free_workqueue_attrs(attrs);
4972 return ret ?: count;
4973}
4974
4975static ssize_t wq_cpumask_show(struct device *dev,
4976 struct device_attribute *attr, char *buf)
4977{
4978 struct workqueue_struct *wq = dev_to_wq(dev);
4979 int written;
4980
4981 mutex_lock(&wq->mutex);
4982 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4983 cpumask_pr_args(wq->unbound_attrs->cpumask));
4984 mutex_unlock(&wq->mutex);
4985 return written;
4986}
4987
4988static ssize_t wq_cpumask_store(struct device *dev,
4989 struct device_attribute *attr,
4990 const char *buf, size_t count)
4991{
4992 struct workqueue_struct *wq = dev_to_wq(dev);
4993 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08004994 int ret = -ENOMEM;
4995
4996 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08004997
4998 attrs = wq_sysfs_prep_attrs(wq);
4999 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005000 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005001
5002 ret = cpumask_parse(buf, attrs->cpumask);
5003 if (!ret)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005004 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005005
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005006out_unlock:
5007 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005008 free_workqueue_attrs(attrs);
5009 return ret ?: count;
5010}
5011
5012static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5013 char *buf)
5014{
5015 struct workqueue_struct *wq = dev_to_wq(dev);
5016 int written;
5017
5018 mutex_lock(&wq->mutex);
5019 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5020 !wq->unbound_attrs->no_numa);
5021 mutex_unlock(&wq->mutex);
5022
5023 return written;
5024}
5025
5026static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5027 const char *buf, size_t count)
5028{
5029 struct workqueue_struct *wq = dev_to_wq(dev);
5030 struct workqueue_attrs *attrs;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005031 int v, ret = -ENOMEM;
5032
5033 apply_wqattrs_lock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005034
5035 attrs = wq_sysfs_prep_attrs(wq);
5036 if (!attrs)
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005037 goto out_unlock;
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005038
5039 ret = -EINVAL;
5040 if (sscanf(buf, "%d", &v) == 1) {
5041 attrs->no_numa = !v;
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005042 ret = apply_workqueue_attrs_locked(wq, attrs);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005043 }
5044
Lai Jiangshand4d3e252015-05-19 18:03:48 +08005045out_unlock:
5046 apply_wqattrs_unlock();
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005047 free_workqueue_attrs(attrs);
5048 return ret ?: count;
5049}
5050
5051static struct device_attribute wq_sysfs_unbound_attrs[] = {
5052 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5053 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5054 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5055 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5056 __ATTR_NULL,
5057};
5058
5059static struct bus_type wq_subsys = {
5060 .name = "workqueue",
5061 .dev_groups = wq_sysfs_groups,
5062};
5063
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005064static ssize_t wq_unbound_cpumask_show(struct device *dev,
5065 struct device_attribute *attr, char *buf)
5066{
5067 int written;
5068
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005069 mutex_lock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005070 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5071 cpumask_pr_args(wq_unbound_cpumask));
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005072 mutex_unlock(&wq_pool_mutex);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005073
5074 return written;
5075}
5076
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005077static ssize_t wq_unbound_cpumask_store(struct device *dev,
5078 struct device_attribute *attr, const char *buf, size_t count)
5079{
5080 cpumask_var_t cpumask;
5081 int ret;
5082
5083 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5084 return -ENOMEM;
5085
5086 ret = cpumask_parse(buf, cpumask);
5087 if (!ret)
5088 ret = workqueue_set_unbound_cpumask(cpumask);
5089
5090 free_cpumask_var(cpumask);
5091 return ret ? ret : count;
5092}
5093
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005094static struct device_attribute wq_sysfs_cpumask_attr =
Lai Jiangshan042f7df2015-04-30 17:16:12 +08005095 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5096 wq_unbound_cpumask_store);
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005097
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005098static int __init wq_sysfs_init(void)
5099{
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005100 int err;
5101
5102 err = subsys_virtual_register(&wq_subsys, NULL);
5103 if (err)
5104 return err;
5105
5106 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
Frederic Weisbecker6ba94422015-04-02 19:14:39 +08005107}
5108core_initcall(wq_sysfs_init);
5109
5110static void wq_device_release(struct device *dev)
5111{
5112 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5113
5114 kfree(wq_dev);
5115}
5116
5117/**
5118 * workqueue_sysfs_register - make a workqueue visible in sysfs
5119 * @wq: the workqueue to register
5120 *
5121 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5122 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5123 * which is the preferred method.
5124 *
5125 * Workqueue user should use this function directly iff it wants to apply
5126 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5127 * apply_workqueue_attrs() may race against userland updating the
5128 * attributes.
5129 *
5130 * Return: 0 on success, -errno on failure.
5131 */
5132int workqueue_sysfs_register(struct workqueue_struct *wq)
5133{
5134 struct wq_device *wq_dev;
5135 int ret;
5136
5137 /*
5138 * Adjusting max_active or creating new pwqs by applyting
5139 * attributes breaks ordering guarantee. Disallow exposing ordered
5140 * workqueues.
5141 */
5142 if (WARN_ON(wq->flags & __WQ_ORDERED))
5143 return -EINVAL;
5144
5145 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5146 if (!wq_dev)
5147 return -ENOMEM;
5148
5149 wq_dev->wq = wq;
5150 wq_dev->dev.bus = &wq_subsys;
5151 wq_dev->dev.init_name = wq->name;
5152 wq_dev->dev.release = wq_device_release;
5153
5154 /*
5155 * unbound_attrs are created separately. Suppress uevent until
5156 * everything is ready.
5157 */
5158 dev_set_uevent_suppress(&wq_dev->dev, true);
5159
5160 ret = device_register(&wq_dev->dev);
5161 if (ret) {
5162 kfree(wq_dev);
5163 wq->wq_dev = NULL;
5164 return ret;
5165 }
5166
5167 if (wq->flags & WQ_UNBOUND) {
5168 struct device_attribute *attr;
5169
5170 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5171 ret = device_create_file(&wq_dev->dev, attr);
5172 if (ret) {
5173 device_unregister(&wq_dev->dev);
5174 wq->wq_dev = NULL;
5175 return ret;
5176 }
5177 }
5178 }
5179
5180 dev_set_uevent_suppress(&wq_dev->dev, false);
5181 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5182 return 0;
5183}
5184
5185/**
5186 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5187 * @wq: the workqueue to unregister
5188 *
5189 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5190 */
5191static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5192{
5193 struct wq_device *wq_dev = wq->wq_dev;
5194
5195 if (!wq->wq_dev)
5196 return;
5197
5198 wq->wq_dev = NULL;
5199 device_unregister(&wq_dev->dev);
5200}
5201#else /* CONFIG_SYSFS */
5202static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5203#endif /* CONFIG_SYSFS */
5204
Tejun Heobce90382013-04-01 11:23:32 -07005205static void __init wq_numa_init(void)
5206{
5207 cpumask_var_t *tbl;
5208 int node, cpu;
5209
Tejun Heobce90382013-04-01 11:23:32 -07005210 if (num_possible_nodes() <= 1)
5211 return;
5212
Tejun Heod55262c2013-04-01 11:23:38 -07005213 if (wq_disable_numa) {
5214 pr_info("workqueue: NUMA affinity support disabled\n");
5215 return;
5216 }
5217
Tejun Heo4c16bd32013-04-01 11:23:36 -07005218 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5219 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5220
Tejun Heobce90382013-04-01 11:23:32 -07005221 /*
5222 * We want masks of possible CPUs of each node which isn't readily
5223 * available. Build one from cpu_to_node() which should have been
5224 * fully initialized by now.
5225 */
Lai Jiangshanddcb57e2014-07-22 13:05:40 +08005226 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
Tejun Heobce90382013-04-01 11:23:32 -07005227 BUG_ON(!tbl);
5228
5229 for_each_node(node)
Yasuaki Ishimatsu5a6024f2014-07-07 09:56:48 -04005230 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
Tejun Heo1be0c252013-05-15 14:24:24 -07005231 node_online(node) ? node : NUMA_NO_NODE));
Tejun Heobce90382013-04-01 11:23:32 -07005232
5233 for_each_possible_cpu(cpu) {
5234 node = cpu_to_node(cpu);
5235 if (WARN_ON(node == NUMA_NO_NODE)) {
5236 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5237 /* happens iff arch is bonkers, let's just proceed */
5238 return;
5239 }
5240 cpumask_set_cpu(cpu, tbl[node]);
5241 }
5242
5243 wq_numa_possible_cpumask = tbl;
5244 wq_numa_enabled = true;
5245}
5246
Suresh Siddha6ee05782010-07-30 14:57:37 -07005247static int __init init_workqueues(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005248{
Tejun Heo7a4e3442013-03-12 11:30:00 -07005249 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5250 int i, cpu;
Tejun Heoc34056a2010-06-29 10:07:11 +02005251
Tejun Heoe904e6c2013-03-12 11:29:57 -07005252 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5253
Frederic Weisbeckerb05a7922015-04-27 17:58:39 +08005254 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5255 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5256
Tejun Heoe904e6c2013-03-12 11:29:57 -07005257 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5258
Tejun Heo65758202012-07-17 12:39:26 -07005259 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
Lai Jiangshana5b4e572012-09-18 09:59:23 -07005260 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
Tejun Heo8b03ae32010-06-29 10:07:12 +02005261
Tejun Heobce90382013-04-01 11:23:32 -07005262 wq_numa_init();
5263
Tejun Heo706026c2013-01-24 11:01:34 -08005264 /* initialize CPU pools */
Tejun Heo29c91e92013-03-12 11:30:03 -07005265 for_each_possible_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005266 struct worker_pool *pool;
Tejun Heo8b03ae32010-06-29 10:07:12 +02005267
Tejun Heo7a4e3442013-03-12 11:30:00 -07005268 i = 0;
Tejun Heof02ae732013-03-12 11:30:03 -07005269 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo7a4e3442013-03-12 11:30:00 -07005270 BUG_ON(init_worker_pool(pool));
Tejun Heoec22ca52013-01-24 11:01:33 -08005271 pool->cpu = cpu;
Tejun Heo29c91e92013-03-12 11:30:03 -07005272 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
Tejun Heo7a4e3442013-03-12 11:30:00 -07005273 pool->attrs->nice = std_nice[i++];
Tejun Heof3f90ad2013-04-01 11:23:34 -07005274 pool->node = cpu_to_node(cpu);
Tejun Heo7a4e3442013-03-12 11:30:00 -07005275
Tejun Heo9daf9e62013-01-24 11:01:33 -08005276 /* alloc pool ID */
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005277 mutex_lock(&wq_pool_mutex);
Tejun Heo9daf9e62013-01-24 11:01:33 -08005278 BUG_ON(worker_pool_assign_id(pool));
Lai Jiangshan68e13a62013-03-25 16:57:17 -07005279 mutex_unlock(&wq_pool_mutex);
Tejun Heo4ce62e92012-07-13 22:16:44 -07005280 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02005281 }
5282
Tejun Heoe22bee72010-06-29 10:07:14 +02005283 /* create the initial worker */
Tejun Heo29c91e92013-03-12 11:30:03 -07005284 for_each_online_cpu(cpu) {
Tejun Heo4ce62e92012-07-13 22:16:44 -07005285 struct worker_pool *pool;
Tejun Heoe22bee72010-06-29 10:07:14 +02005286
Tejun Heof02ae732013-03-12 11:30:03 -07005287 for_each_cpu_worker_pool(pool, cpu) {
Tejun Heo29c91e92013-03-12 11:30:03 -07005288 pool->flags &= ~POOL_DISASSOCIATED;
Lai Jiangshan051e1852014-07-22 13:03:02 +08005289 BUG_ON(!create_worker(pool));
Tejun Heo4ce62e92012-07-13 22:16:44 -07005290 }
Tejun Heoe22bee72010-06-29 10:07:14 +02005291 }
5292
Tejun Heo8a2b7532013-09-05 12:30:04 -04005293 /* create default unbound and ordered wq attrs */
Tejun Heo29c91e92013-03-12 11:30:03 -07005294 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5295 struct workqueue_attrs *attrs;
5296
5297 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
Tejun Heo29c91e92013-03-12 11:30:03 -07005298 attrs->nice = std_nice[i];
Tejun Heo29c91e92013-03-12 11:30:03 -07005299 unbound_std_wq_attrs[i] = attrs;
Tejun Heo8a2b7532013-09-05 12:30:04 -04005300
5301 /*
5302 * An ordered wq should have only one pwq as ordering is
5303 * guaranteed by max_active which is enforced by pwqs.
5304 * Turn off NUMA so that dfl_pwq is used for all nodes.
5305 */
5306 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5307 attrs->nice = std_nice[i];
5308 attrs->no_numa = true;
5309 ordered_wq_attrs[i] = attrs;
Tejun Heo29c91e92013-03-12 11:30:03 -07005310 }
5311
Tejun Heod320c032010-06-29 10:07:14 +02005312 system_wq = alloc_workqueue("events", 0, 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005313 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
Tejun Heod320c032010-06-29 10:07:14 +02005314 system_long_wq = alloc_workqueue("events_long", 0, 0);
Tejun Heof3421792010-07-02 10:03:51 +02005315 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5316 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01005317 system_freezable_wq = alloc_workqueue("events_freezable",
5318 WQ_FREEZABLE, 0);
Viresh Kumar06681062013-04-24 17:12:54 +05305319 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5320 WQ_POWER_EFFICIENT, 0);
5321 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5322 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5323 0);
Joonsoo Kim1aabe902012-08-15 23:25:39 +09005324 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
Viresh Kumar06681062013-04-24 17:12:54 +05305325 !system_unbound_wq || !system_freezable_wq ||
5326 !system_power_efficient_wq ||
5327 !system_freezable_power_efficient_wq);
Suresh Siddha6ee05782010-07-30 14:57:37 -07005328 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005329}
Suresh Siddha6ee05782010-07-30 14:57:37 -07005330early_initcall(init_workqueues);