Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 | 2 | /* |
c54fce6e | 3 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 4 | * |
c54fce6e | 5 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 6 | * |
c54fce6e TH |
7 | * Derived from the taskqueue/keventd code by: |
8 | * David Woodhouse <dwmw2@infradead.org> | |
9 | * Andrew Morton | |
10 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
11 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 12 | * |
c54fce6e | 13 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 14 | * |
c54fce6e TH |
15 | * Copyright (C) 2010 SUSE Linux Products GmbH |
16 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 17 | * |
c54fce6e TH |
18 | * This is the generic async execution mechanism. Work items as are |
19 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
20 | * automatically managed. There are two worker pools for each CPU (one for |
21 | * normal work items and the other for high priority ones) and some extra | |
22 | * pools for workqueues which are not bound to any specific CPU - the | |
23 | * number of these backing pools is dynamic. | |
c54fce6e | 24 | * |
9a261491 | 25 | * Please read Documentation/core-api/workqueue.rst for details. |
1da177e4 LT |
26 | */ |
27 | ||
9984de1a | 28 | #include <linux/export.h> |
1da177e4 LT |
29 | #include <linux/kernel.h> |
30 | #include <linux/sched.h> | |
31 | #include <linux/init.h> | |
32 | #include <linux/signal.h> | |
33 | #include <linux/completion.h> | |
34 | #include <linux/workqueue.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/cpu.h> | |
37 | #include <linux/notifier.h> | |
38 | #include <linux/kthread.h> | |
1fa44eca | 39 | #include <linux/hardirq.h> |
46934023 | 40 | #include <linux/mempolicy.h> |
341a5958 | 41 | #include <linux/freezer.h> |
d5abe669 | 42 | #include <linux/debug_locks.h> |
4e6045f1 | 43 | #include <linux/lockdep.h> |
c34056a3 | 44 | #include <linux/idr.h> |
29c91e99 | 45 | #include <linux/jhash.h> |
42f8570f | 46 | #include <linux/hashtable.h> |
76af4d93 | 47 | #include <linux/rculist.h> |
bce90380 | 48 | #include <linux/nodemask.h> |
4c16bd32 | 49 | #include <linux/moduleparam.h> |
3d1cb205 | 50 | #include <linux/uaccess.h> |
c98a9805 | 51 | #include <linux/sched/isolation.h> |
62635ea8 | 52 | #include <linux/nmi.h> |
e22bee78 | 53 | |
ea138446 | 54 | #include "workqueue_internal.h" |
1da177e4 | 55 | |
c8e55f36 | 56 | enum { |
24647570 TH |
57 | /* |
58 | * worker_pool flags | |
bc2ae0f5 | 59 | * |
24647570 | 60 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
61 | * While associated (!DISASSOCIATED), all workers are bound to the |
62 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
63 | * is in effect. | |
64 | * | |
65 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
66 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 67 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 68 | * |
bc3a1afc | 69 | * Note that DISASSOCIATED should be flipped only while holding |
1258fae7 | 70 | * wq_pool_attach_mutex to avoid changing binding state while |
4736cbf7 | 71 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 72 | */ |
692b4825 | 73 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
24647570 | 74 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
db7bccf4 | 75 | |
c8e55f36 | 76 | /* worker flags */ |
c8e55f36 TH |
77 | WORKER_DIE = 1 << 1, /* die die die */ |
78 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 79 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 80 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 81 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 82 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 83 | |
a9ab775b TH |
84 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
85 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 86 | |
e34cdddb | 87 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 88 | |
29c91e99 | 89 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 90 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 91 | |
e22bee78 TH |
92 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
93 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
94 | ||
3233cdbd TH |
95 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
96 | /* call for help after 10ms | |
97 | (min two ticks) */ | |
e22bee78 TH |
98 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
99 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
100 | |
101 | /* | |
102 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 103 | * all cpus. Give MIN_NICE. |
e22bee78 | 104 | */ |
8698a745 DY |
105 | RESCUER_NICE_LEVEL = MIN_NICE, |
106 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f TH |
107 | |
108 | WQ_NAME_LEN = 24, | |
c8e55f36 | 109 | }; |
1da177e4 LT |
110 | |
111 | /* | |
4690c4ab TH |
112 | * Structure fields follow one of the following exclusion rules. |
113 | * | |
e41e704b TH |
114 | * I: Modifiable by initialization/destruction paths and read-only for |
115 | * everyone else. | |
4690c4ab | 116 | * |
e22bee78 TH |
117 | * P: Preemption protected. Disabling preemption is enough and should |
118 | * only be modified and accessed from the local cpu. | |
119 | * | |
d565ed63 | 120 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 121 | * |
d565ed63 TH |
122 | * X: During normal operation, modification requires pool->lock and should |
123 | * be done only from local cpu. Either disabling preemption on local | |
124 | * cpu or grabbing pool->lock is enough for read access. If | |
125 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 126 | * |
1258fae7 | 127 | * A: wq_pool_attach_mutex protected. |
822d8405 | 128 | * |
68e13a67 | 129 | * PL: wq_pool_mutex protected. |
5bcab335 | 130 | * |
24acfb71 | 131 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
76af4d93 | 132 | * |
5b95e1af LJ |
133 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
134 | * | |
135 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
24acfb71 | 136 | * RCU for reads. |
5b95e1af | 137 | * |
3c25a55d LJ |
138 | * WQ: wq->mutex protected. |
139 | * | |
24acfb71 | 140 | * WR: wq->mutex protected for writes. RCU protected for reads. |
2e109a28 TH |
141 | * |
142 | * MD: wq_mayday_lock protected. | |
1da177e4 | 143 | */ |
1da177e4 | 144 | |
2eaebdb3 | 145 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 146 | |
bd7bdd43 | 147 | struct worker_pool { |
d565ed63 | 148 | spinlock_t lock; /* the pool lock */ |
d84ff051 | 149 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 150 | int node; /* I: the associated node ID */ |
9daf9e67 | 151 | int id; /* I: pool ID */ |
11ebea50 | 152 | unsigned int flags; /* X: flags */ |
bd7bdd43 | 153 | |
82607adc TH |
154 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
155 | ||
bd7bdd43 | 156 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 157 | |
5826cc8f LJ |
158 | int nr_workers; /* L: total number of workers */ |
159 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 TH |
160 | |
161 | struct list_head idle_list; /* X: list of idle workers */ | |
162 | struct timer_list idle_timer; /* L: worker idle timeout */ | |
163 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
164 | ||
c5aa87bb | 165 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
166 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
167 | /* L: hash of busy workers */ | |
168 | ||
2607d7a6 | 169 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 170 | struct list_head workers; /* A: attached workers */ |
60f5a4bc | 171 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 172 | |
7cda9aae | 173 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 174 | |
7a4e344c | 175 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
176 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
177 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 178 | |
e19e397a TH |
179 | /* |
180 | * The current concurrency level. As it's likely to be accessed | |
181 | * from other CPUs during try_to_wake_up(), put it in a separate | |
182 | * cacheline. | |
183 | */ | |
184 | atomic_t nr_running ____cacheline_aligned_in_smp; | |
29c91e99 TH |
185 | |
186 | /* | |
24acfb71 | 187 | * Destruction of pool is RCU protected to allow dereferences |
29c91e99 TH |
188 | * from get_work_pool(). |
189 | */ | |
190 | struct rcu_head rcu; | |
8b03ae3c TH |
191 | } ____cacheline_aligned_in_smp; |
192 | ||
1da177e4 | 193 | /* |
112202d9 TH |
194 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
195 | * of work_struct->data are used for flags and the remaining high bits | |
196 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
197 | * number of flag bits. | |
1da177e4 | 198 | */ |
112202d9 | 199 | struct pool_workqueue { |
bd7bdd43 | 200 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 201 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
202 | int work_color; /* L: current color */ |
203 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 204 | int refcnt; /* L: reference count */ |
73f53c4a TH |
205 | int nr_in_flight[WORK_NR_COLORS]; |
206 | /* L: nr of in_flight works */ | |
1e19ffc6 | 207 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 208 | int max_active; /* L: max active works */ |
1e19ffc6 | 209 | struct list_head delayed_works; /* L: delayed works */ |
3c25a55d | 210 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 211 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 TH |
212 | |
213 | /* | |
214 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
215 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
24acfb71 | 216 | * itself is also RCU protected so that the first pwq can be |
b09f4fd3 | 217 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
218 | */ |
219 | struct work_struct unbound_release_work; | |
220 | struct rcu_head rcu; | |
e904e6c2 | 221 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 222 | |
73f53c4a TH |
223 | /* |
224 | * Structure used to wait for workqueue flush. | |
225 | */ | |
226 | struct wq_flusher { | |
3c25a55d LJ |
227 | struct list_head list; /* WQ: list of flushers */ |
228 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
229 | struct completion done; /* flush completion */ |
230 | }; | |
231 | ||
226223ab TH |
232 | struct wq_device; |
233 | ||
1da177e4 | 234 | /* |
c5aa87bb TH |
235 | * The externally visible workqueue. It relays the issued work items to |
236 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
237 | */ |
238 | struct workqueue_struct { | |
3c25a55d | 239 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 240 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 241 | |
3c25a55d LJ |
242 | struct mutex mutex; /* protects this wq */ |
243 | int work_color; /* WQ: current work color */ | |
244 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 245 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
246 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
247 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
248 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 249 | |
2e109a28 | 250 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
30ae2fc0 | 251 | struct worker *rescuer; /* MD: rescue worker */ |
e22bee78 | 252 | |
87fc741e | 253 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 254 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 255 | |
5b95e1af LJ |
256 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
257 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ | |
6029a918 | 258 | |
226223ab TH |
259 | #ifdef CONFIG_SYSFS |
260 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
261 | #endif | |
4e6045f1 | 262 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
263 | char *lock_name; |
264 | struct lock_class_key key; | |
4690c4ab | 265 | struct lockdep_map lockdep_map; |
4e6045f1 | 266 | #endif |
ecf6881f | 267 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 268 | |
e2dca7ad | 269 | /* |
24acfb71 TG |
270 | * Destruction of workqueue_struct is RCU protected to allow walking |
271 | * the workqueues list without grabbing wq_pool_mutex. | |
e2dca7ad TH |
272 | * This is used to dump all workqueues from sysrq. |
273 | */ | |
274 | struct rcu_head rcu; | |
275 | ||
2728fd2f TH |
276 | /* hot fields used during command issue, aligned to cacheline */ |
277 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
278 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ | |
5b95e1af | 279 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ |
1da177e4 LT |
280 | }; |
281 | ||
e904e6c2 TH |
282 | static struct kmem_cache *pwq_cache; |
283 | ||
bce90380 TH |
284 | static cpumask_var_t *wq_numa_possible_cpumask; |
285 | /* possible CPUs of each node */ | |
286 | ||
d55262c4 TH |
287 | static bool wq_disable_numa; |
288 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); | |
289 | ||
cee22a15 | 290 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 291 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
292 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
293 | ||
863b710b | 294 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 295 | |
bce90380 TH |
296 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
297 | ||
4c16bd32 TH |
298 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
299 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
300 | ||
68e13a67 | 301 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 302 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
2e109a28 | 303 | static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
692b4825 | 304 | static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ |
5bcab335 | 305 | |
e2dca7ad | 306 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 307 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 308 | |
ef557180 MG |
309 | /* PL: allowable cpus for unbound wqs and work items */ |
310 | static cpumask_var_t wq_unbound_cpumask; | |
311 | ||
312 | /* CPU where unbound work was last round robin scheduled from this CPU */ | |
313 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 314 | |
f303fccb TH |
315 | /* |
316 | * Local execution of unbound work items is no longer guaranteed. The | |
317 | * following always forces round-robin CPU selection on unbound work items | |
318 | * to uncover usages which depend on it. | |
319 | */ | |
320 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
321 | static bool wq_debug_force_rr_cpu = true; | |
322 | #else | |
323 | static bool wq_debug_force_rr_cpu = false; | |
324 | #endif | |
325 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
326 | ||
7d19c5ce | 327 | /* the per-cpu worker pools */ |
25528213 | 328 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 329 | |
68e13a67 | 330 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 331 | |
68e13a67 | 332 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
333 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
334 | ||
c5aa87bb | 335 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
336 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
337 | ||
8a2b7538 TH |
338 | /* I: attributes used when instantiating ordered pools on demand */ |
339 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
340 | ||
d320c038 | 341 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 342 | EXPORT_SYMBOL(system_wq); |
044c782c | 343 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 344 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 345 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 346 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 347 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 348 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 349 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 350 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
351 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
352 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
353 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
354 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 355 | |
7d19c5ce | 356 | static int worker_thread(void *__worker); |
6ba94429 | 357 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
c29eb853 | 358 | static void show_pwq(struct pool_workqueue *pwq); |
7d19c5ce | 359 | |
97bd2347 TH |
360 | #define CREATE_TRACE_POINTS |
361 | #include <trace/events/workqueue.h> | |
362 | ||
68e13a67 | 363 | #define assert_rcu_or_pool_mutex() \ |
24acfb71 | 364 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 | 365 | !lockdep_is_held(&wq_pool_mutex), \ |
24acfb71 | 366 | "RCU or wq_pool_mutex should be held") |
5bcab335 | 367 | |
5b95e1af | 368 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
24acfb71 | 369 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 PM |
370 | !lockdep_is_held(&wq->mutex) && \ |
371 | !lockdep_is_held(&wq_pool_mutex), \ | |
24acfb71 | 372 | "RCU, wq->mutex or wq_pool_mutex should be held") |
5b95e1af | 373 | |
f02ae73a TH |
374 | #define for_each_cpu_worker_pool(pool, cpu) \ |
375 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
376 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 377 | (pool)++) |
4ce62e9e | 378 | |
17116969 TH |
379 | /** |
380 | * for_each_pool - iterate through all worker_pools in the system | |
381 | * @pool: iteration cursor | |
611c92a0 | 382 | * @pi: integer used for iteration |
fa1b54e6 | 383 | * |
24acfb71 | 384 | * This must be called either with wq_pool_mutex held or RCU read |
68e13a67 LJ |
385 | * locked. If the pool needs to be used beyond the locking in effect, the |
386 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
387 | * |
388 | * The if/else clause exists only for the lockdep assertion and can be | |
389 | * ignored. | |
17116969 | 390 | */ |
611c92a0 TH |
391 | #define for_each_pool(pool, pi) \ |
392 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 393 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 394 | else |
17116969 | 395 | |
822d8405 TH |
396 | /** |
397 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
398 | * @worker: iteration cursor | |
822d8405 TH |
399 | * @pool: worker_pool to iterate workers of |
400 | * | |
1258fae7 | 401 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
402 | * |
403 | * The if/else clause exists only for the lockdep assertion and can be | |
404 | * ignored. | |
405 | */ | |
da028469 LJ |
406 | #define for_each_pool_worker(worker, pool) \ |
407 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 408 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
409 | else |
410 | ||
49e3cf44 TH |
411 | /** |
412 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
413 | * @pwq: iteration cursor | |
414 | * @wq: the target workqueue | |
76af4d93 | 415 | * |
24acfb71 | 416 | * This must be called either with wq->mutex held or RCU read locked. |
794b18bc TH |
417 | * If the pwq needs to be used beyond the locking in effect, the caller is |
418 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
419 | * |
420 | * The if/else clause exists only for the lockdep assertion and can be | |
421 | * ignored. | |
49e3cf44 TH |
422 | */ |
423 | #define for_each_pwq(pwq, wq) \ | |
49e9d1a9 | 424 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
5a644662 | 425 | lockdep_is_held(&(wq->mutex))) |
f3421797 | 426 | |
dc186ad7 TG |
427 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
428 | ||
429 | static struct debug_obj_descr work_debug_descr; | |
430 | ||
99777288 SG |
431 | static void *work_debug_hint(void *addr) |
432 | { | |
433 | return ((struct work_struct *) addr)->func; | |
434 | } | |
435 | ||
b9fdac7f DC |
436 | static bool work_is_static_object(void *addr) |
437 | { | |
438 | struct work_struct *work = addr; | |
439 | ||
440 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
441 | } | |
442 | ||
dc186ad7 TG |
443 | /* |
444 | * fixup_init is called when: | |
445 | * - an active object is initialized | |
446 | */ | |
02a982a6 | 447 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
448 | { |
449 | struct work_struct *work = addr; | |
450 | ||
451 | switch (state) { | |
452 | case ODEBUG_STATE_ACTIVE: | |
453 | cancel_work_sync(work); | |
454 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 455 | return true; |
dc186ad7 | 456 | default: |
02a982a6 | 457 | return false; |
dc186ad7 TG |
458 | } |
459 | } | |
460 | ||
dc186ad7 TG |
461 | /* |
462 | * fixup_free is called when: | |
463 | * - an active object is freed | |
464 | */ | |
02a982a6 | 465 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
466 | { |
467 | struct work_struct *work = addr; | |
468 | ||
469 | switch (state) { | |
470 | case ODEBUG_STATE_ACTIVE: | |
471 | cancel_work_sync(work); | |
472 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 473 | return true; |
dc186ad7 | 474 | default: |
02a982a6 | 475 | return false; |
dc186ad7 TG |
476 | } |
477 | } | |
478 | ||
479 | static struct debug_obj_descr work_debug_descr = { | |
480 | .name = "work_struct", | |
99777288 | 481 | .debug_hint = work_debug_hint, |
b9fdac7f | 482 | .is_static_object = work_is_static_object, |
dc186ad7 | 483 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
484 | .fixup_free = work_fixup_free, |
485 | }; | |
486 | ||
487 | static inline void debug_work_activate(struct work_struct *work) | |
488 | { | |
489 | debug_object_activate(work, &work_debug_descr); | |
490 | } | |
491 | ||
492 | static inline void debug_work_deactivate(struct work_struct *work) | |
493 | { | |
494 | debug_object_deactivate(work, &work_debug_descr); | |
495 | } | |
496 | ||
497 | void __init_work(struct work_struct *work, int onstack) | |
498 | { | |
499 | if (onstack) | |
500 | debug_object_init_on_stack(work, &work_debug_descr); | |
501 | else | |
502 | debug_object_init(work, &work_debug_descr); | |
503 | } | |
504 | EXPORT_SYMBOL_GPL(__init_work); | |
505 | ||
506 | void destroy_work_on_stack(struct work_struct *work) | |
507 | { | |
508 | debug_object_free(work, &work_debug_descr); | |
509 | } | |
510 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
511 | ||
ea2e64f2 TG |
512 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
513 | { | |
514 | destroy_timer_on_stack(&work->timer); | |
515 | debug_object_free(&work->work, &work_debug_descr); | |
516 | } | |
517 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
518 | ||
dc186ad7 TG |
519 | #else |
520 | static inline void debug_work_activate(struct work_struct *work) { } | |
521 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
522 | #endif | |
523 | ||
4e8b22bd LB |
524 | /** |
525 | * worker_pool_assign_id - allocate ID and assing it to @pool | |
526 | * @pool: the pool pointer of interest | |
527 | * | |
528 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
529 | * successfully, -errno on failure. | |
530 | */ | |
9daf9e67 TH |
531 | static int worker_pool_assign_id(struct worker_pool *pool) |
532 | { | |
533 | int ret; | |
534 | ||
68e13a67 | 535 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 536 | |
4e8b22bd LB |
537 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
538 | GFP_KERNEL); | |
229641a6 | 539 | if (ret >= 0) { |
e68035fb | 540 | pool->id = ret; |
229641a6 TH |
541 | return 0; |
542 | } | |
fa1b54e6 | 543 | return ret; |
7c3eed5c TH |
544 | } |
545 | ||
df2d5ae4 TH |
546 | /** |
547 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
548 | * @wq: the target workqueue | |
549 | * @node: the node ID | |
550 | * | |
24acfb71 | 551 | * This must be called with any of wq_pool_mutex, wq->mutex or RCU |
5b95e1af | 552 | * read locked. |
df2d5ae4 TH |
553 | * If the pwq needs to be used beyond the locking in effect, the caller is |
554 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
555 | * |
556 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
557 | */ |
558 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
559 | int node) | |
560 | { | |
5b95e1af | 561 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
d6e022f1 TH |
562 | |
563 | /* | |
564 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
565 | * delayed item is pending. The plan is to keep CPU -> NODE | |
566 | * mapping valid and stable across CPU on/offlines. Once that | |
567 | * happens, this workaround can be removed. | |
568 | */ | |
569 | if (unlikely(node == NUMA_NO_NODE)) | |
570 | return wq->dfl_pwq; | |
571 | ||
df2d5ae4 TH |
572 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
573 | } | |
574 | ||
73f53c4a TH |
575 | static unsigned int work_color_to_flags(int color) |
576 | { | |
577 | return color << WORK_STRUCT_COLOR_SHIFT; | |
578 | } | |
579 | ||
580 | static int get_work_color(struct work_struct *work) | |
581 | { | |
582 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
583 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
584 | } | |
585 | ||
586 | static int work_next_color(int color) | |
587 | { | |
588 | return (color + 1) % WORK_NR_COLORS; | |
589 | } | |
1da177e4 | 590 | |
14441960 | 591 | /* |
112202d9 TH |
592 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
593 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 594 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 595 | * |
112202d9 TH |
596 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
597 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
598 | * work->data. These functions should only be called while the work is |
599 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 600 | * |
112202d9 | 601 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 602 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 603 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 604 | * available only while the work item is queued. |
7a22ad75 | 605 | * |
bbb68dfa TH |
606 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
607 | * canceled. While being canceled, a work item may have its PENDING set | |
608 | * but stay off timer and worklist for arbitrarily long and nobody should | |
609 | * try to steal the PENDING bit. | |
14441960 | 610 | */ |
7a22ad75 TH |
611 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
612 | unsigned long flags) | |
365970a1 | 613 | { |
6183c009 | 614 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
615 | atomic_long_set(&work->data, data | flags | work_static(work)); |
616 | } | |
365970a1 | 617 | |
112202d9 | 618 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
619 | unsigned long extra_flags) |
620 | { | |
112202d9 TH |
621 | set_work_data(work, (unsigned long)pwq, |
622 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
623 | } |
624 | ||
4468a00f LJ |
625 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
626 | int pool_id) | |
627 | { | |
628 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
629 | WORK_STRUCT_PENDING); | |
630 | } | |
631 | ||
7c3eed5c TH |
632 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
633 | int pool_id) | |
7a22ad75 | 634 | { |
23657bb1 TH |
635 | /* |
636 | * The following wmb is paired with the implied mb in | |
637 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
638 | * here are visible to and precede any updates by the next PENDING | |
639 | * owner. | |
640 | */ | |
641 | smp_wmb(); | |
7c3eed5c | 642 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
643 | /* |
644 | * The following mb guarantees that previous clear of a PENDING bit | |
645 | * will not be reordered with any speculative LOADS or STORES from | |
646 | * work->current_func, which is executed afterwards. This possible | |
8bdc6201 | 647 | * reordering can lead to a missed execution on attempt to queue |
346c09f8 RP |
648 | * the same @work. E.g. consider this case: |
649 | * | |
650 | * CPU#0 CPU#1 | |
651 | * ---------------------------- -------------------------------- | |
652 | * | |
653 | * 1 STORE event_indicated | |
654 | * 2 queue_work_on() { | |
655 | * 3 test_and_set_bit(PENDING) | |
656 | * 4 } set_..._and_clear_pending() { | |
657 | * 5 set_work_data() # clear bit | |
658 | * 6 smp_mb() | |
659 | * 7 work->current_func() { | |
660 | * 8 LOAD event_indicated | |
661 | * } | |
662 | * | |
663 | * Without an explicit full barrier speculative LOAD on line 8 can | |
664 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
665 | * CPU#0 observes the PENDING bit is still set and new execution of | |
666 | * a @work is not queued in a hope, that CPU#1 will eventually | |
667 | * finish the queued @work. Meanwhile CPU#1 does not see | |
668 | * event_indicated is set, because speculative LOAD was executed | |
669 | * before actual STORE. | |
670 | */ | |
671 | smp_mb(); | |
7a22ad75 | 672 | } |
f756d5e2 | 673 | |
7a22ad75 | 674 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 675 | { |
7c3eed5c TH |
676 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
677 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
678 | } |
679 | ||
112202d9 | 680 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 681 | { |
e120153d | 682 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 683 | |
112202d9 | 684 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
685 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
686 | else | |
687 | return NULL; | |
4d707b9f ON |
688 | } |
689 | ||
7c3eed5c TH |
690 | /** |
691 | * get_work_pool - return the worker_pool a given work was associated with | |
692 | * @work: the work item of interest | |
693 | * | |
68e13a67 | 694 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
24acfb71 TG |
695 | * access under RCU read lock. As such, this function should be |
696 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. | |
fa1b54e6 TH |
697 | * |
698 | * All fields of the returned pool are accessible as long as the above | |
699 | * mentioned locking is in effect. If the returned pool needs to be used | |
700 | * beyond the critical section, the caller is responsible for ensuring the | |
701 | * returned pool is and stays online. | |
d185af30 YB |
702 | * |
703 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
704 | */ |
705 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 706 | { |
e120153d | 707 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 708 | int pool_id; |
7a22ad75 | 709 | |
68e13a67 | 710 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 711 | |
112202d9 TH |
712 | if (data & WORK_STRUCT_PWQ) |
713 | return ((struct pool_workqueue *) | |
7c3eed5c | 714 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 715 | |
7c3eed5c TH |
716 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
717 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
718 | return NULL; |
719 | ||
fa1b54e6 | 720 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
721 | } |
722 | ||
723 | /** | |
724 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
725 | * @work: the work item of interest | |
726 | * | |
d185af30 | 727 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
728 | * %WORK_OFFQ_POOL_NONE if none. |
729 | */ | |
730 | static int get_work_pool_id(struct work_struct *work) | |
731 | { | |
54d5b7d0 LJ |
732 | unsigned long data = atomic_long_read(&work->data); |
733 | ||
112202d9 TH |
734 | if (data & WORK_STRUCT_PWQ) |
735 | return ((struct pool_workqueue *) | |
54d5b7d0 | 736 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 737 | |
54d5b7d0 | 738 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
739 | } |
740 | ||
bbb68dfa TH |
741 | static void mark_work_canceling(struct work_struct *work) |
742 | { | |
7c3eed5c | 743 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 744 | |
7c3eed5c TH |
745 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
746 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
747 | } |
748 | ||
749 | static bool work_is_canceling(struct work_struct *work) | |
750 | { | |
751 | unsigned long data = atomic_long_read(&work->data); | |
752 | ||
112202d9 | 753 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
754 | } |
755 | ||
e22bee78 | 756 | /* |
3270476a TH |
757 | * Policy functions. These define the policies on how the global worker |
758 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 759 | * they're being called with pool->lock held. |
e22bee78 TH |
760 | */ |
761 | ||
63d95a91 | 762 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 763 | { |
e19e397a | 764 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
765 | } |
766 | ||
4594bf15 | 767 | /* |
e22bee78 TH |
768 | * Need to wake up a worker? Called from anything but currently |
769 | * running workers. | |
974271c4 TH |
770 | * |
771 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 772 | * function will always return %true for unbound pools as long as the |
974271c4 | 773 | * worklist isn't empty. |
4594bf15 | 774 | */ |
63d95a91 | 775 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 776 | { |
63d95a91 | 777 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 778 | } |
4594bf15 | 779 | |
e22bee78 | 780 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 781 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 782 | { |
63d95a91 | 783 | return pool->nr_idle; |
e22bee78 TH |
784 | } |
785 | ||
786 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 787 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 788 | { |
e19e397a TH |
789 | return !list_empty(&pool->worklist) && |
790 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
791 | } |
792 | ||
793 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 794 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 795 | { |
63d95a91 | 796 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 797 | } |
365970a1 | 798 | |
e22bee78 | 799 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 800 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 801 | { |
692b4825 | 802 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
803 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
804 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
805 | |
806 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
807 | } |
808 | ||
4d707b9f | 809 | /* |
e22bee78 TH |
810 | * Wake up functions. |
811 | */ | |
812 | ||
1037de36 LJ |
813 | /* Return the first idle worker. Safe with preemption disabled */ |
814 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
7e11629d | 815 | { |
63d95a91 | 816 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
817 | return NULL; |
818 | ||
63d95a91 | 819 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
820 | } |
821 | ||
822 | /** | |
823 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 824 | * @pool: worker pool to wake worker from |
7e11629d | 825 | * |
63d95a91 | 826 | * Wake up the first idle worker of @pool. |
7e11629d TH |
827 | * |
828 | * CONTEXT: | |
d565ed63 | 829 | * spin_lock_irq(pool->lock). |
7e11629d | 830 | */ |
63d95a91 | 831 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 832 | { |
1037de36 | 833 | struct worker *worker = first_idle_worker(pool); |
7e11629d TH |
834 | |
835 | if (likely(worker)) | |
836 | wake_up_process(worker->task); | |
837 | } | |
838 | ||
d302f017 | 839 | /** |
6d25be57 | 840 | * wq_worker_running - a worker is running again |
e22bee78 | 841 | * @task: task waking up |
e22bee78 | 842 | * |
6d25be57 | 843 | * This function is called when a worker returns from schedule() |
e22bee78 | 844 | */ |
6d25be57 | 845 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
846 | { |
847 | struct worker *worker = kthread_data(task); | |
848 | ||
6d25be57 TG |
849 | if (!worker->sleeping) |
850 | return; | |
851 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 852 | atomic_inc(&worker->pool->nr_running); |
6d25be57 | 853 | worker->sleeping = 0; |
e22bee78 TH |
854 | } |
855 | ||
856 | /** | |
857 | * wq_worker_sleeping - a worker is going to sleep | |
858 | * @task: task going to sleep | |
e22bee78 | 859 | * |
6d25be57 TG |
860 | * This function is called from schedule() when a busy worker is |
861 | * going to sleep. | |
e22bee78 | 862 | */ |
6d25be57 | 863 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 864 | { |
6d25be57 | 865 | struct worker *next, *worker = kthread_data(task); |
111c225a | 866 | struct worker_pool *pool; |
e22bee78 | 867 | |
111c225a TH |
868 | /* |
869 | * Rescuers, which may not have all the fields set up like normal | |
870 | * workers, also reach here, let's not access anything before | |
871 | * checking NOT_RUNNING. | |
872 | */ | |
2d64672e | 873 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 874 | return; |
e22bee78 | 875 | |
111c225a | 876 | pool = worker->pool; |
111c225a | 877 | |
6d25be57 TG |
878 | if (WARN_ON_ONCE(worker->sleeping)) |
879 | return; | |
880 | ||
881 | worker->sleeping = 1; | |
882 | spin_lock_irq(&pool->lock); | |
e22bee78 TH |
883 | |
884 | /* | |
885 | * The counterpart of the following dec_and_test, implied mb, | |
886 | * worklist not empty test sequence is in insert_work(). | |
887 | * Please read comment there. | |
888 | * | |
628c78e7 TH |
889 | * NOT_RUNNING is clear. This means that we're bound to and |
890 | * running on the local cpu w/ rq lock held and preemption | |
891 | * disabled, which in turn means that none else could be | |
d565ed63 | 892 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 893 | * lock is safe. |
e22bee78 | 894 | */ |
e19e397a | 895 | if (atomic_dec_and_test(&pool->nr_running) && |
6d25be57 TG |
896 | !list_empty(&pool->worklist)) { |
897 | next = first_idle_worker(pool); | |
898 | if (next) | |
899 | wake_up_process(next->task); | |
900 | } | |
901 | spin_unlock_irq(&pool->lock); | |
e22bee78 TH |
902 | } |
903 | ||
1b69ac6b JW |
904 | /** |
905 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 906 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
907 | * |
908 | * Determine the last function a worker executed. This is called from | |
909 | * the scheduler to get a worker's last known identity. | |
910 | * | |
911 | * CONTEXT: | |
912 | * spin_lock_irq(rq->lock) | |
913 | * | |
4b047002 JW |
914 | * This function is called during schedule() when a kworker is going |
915 | * to sleep. It's used by psi to identify aggregation workers during | |
916 | * dequeuing, to allow periodic aggregation to shut-off when that | |
917 | * worker is the last task in the system or cgroup to go to sleep. | |
918 | * | |
919 | * As this function doesn't involve any workqueue-related locking, it | |
920 | * only returns stable values when called from inside the scheduler's | |
921 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
922 | * is guaranteed to not be processing any works. | |
923 | * | |
1b69ac6b JW |
924 | * Return: |
925 | * The last work function %current executed as a worker, NULL if it | |
926 | * hasn't executed any work yet. | |
927 | */ | |
928 | work_func_t wq_worker_last_func(struct task_struct *task) | |
929 | { | |
930 | struct worker *worker = kthread_data(task); | |
931 | ||
932 | return worker->last_func; | |
933 | } | |
934 | ||
e22bee78 TH |
935 | /** |
936 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 937 | * @worker: self |
d302f017 | 938 | * @flags: flags to set |
d302f017 | 939 | * |
228f1d00 | 940 | * Set @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 941 | * |
cb444766 | 942 | * CONTEXT: |
d565ed63 | 943 | * spin_lock_irq(pool->lock) |
d302f017 | 944 | */ |
228f1d00 | 945 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) |
d302f017 | 946 | { |
bd7bdd43 | 947 | struct worker_pool *pool = worker->pool; |
e22bee78 | 948 | |
cb444766 TH |
949 | WARN_ON_ONCE(worker->task != current); |
950 | ||
228f1d00 | 951 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ |
e22bee78 TH |
952 | if ((flags & WORKER_NOT_RUNNING) && |
953 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
228f1d00 | 954 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
955 | } |
956 | ||
d302f017 TH |
957 | worker->flags |= flags; |
958 | } | |
959 | ||
960 | /** | |
e22bee78 | 961 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 962 | * @worker: self |
d302f017 TH |
963 | * @flags: flags to clear |
964 | * | |
e22bee78 | 965 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 966 | * |
cb444766 | 967 | * CONTEXT: |
d565ed63 | 968 | * spin_lock_irq(pool->lock) |
d302f017 TH |
969 | */ |
970 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
971 | { | |
63d95a91 | 972 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
973 | unsigned int oflags = worker->flags; |
974 | ||
cb444766 TH |
975 | WARN_ON_ONCE(worker->task != current); |
976 | ||
d302f017 | 977 | worker->flags &= ~flags; |
e22bee78 | 978 | |
42c025f3 TH |
979 | /* |
980 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
981 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
982 | * of multiple flags, not a single flag. | |
983 | */ | |
e22bee78 TH |
984 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
985 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 986 | atomic_inc(&pool->nr_running); |
d302f017 TH |
987 | } |
988 | ||
8cca0eea TH |
989 | /** |
990 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 991 | * @pool: pool of interest |
8cca0eea TH |
992 | * @work: work to find worker for |
993 | * | |
c9e7cf27 TH |
994 | * Find a worker which is executing @work on @pool by searching |
995 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
996 | * to match, its current execution should match the address of @work and |
997 | * its work function. This is to avoid unwanted dependency between | |
998 | * unrelated work executions through a work item being recycled while still | |
999 | * being executed. | |
1000 | * | |
1001 | * This is a bit tricky. A work item may be freed once its execution | |
1002 | * starts and nothing prevents the freed area from being recycled for | |
1003 | * another work item. If the same work item address ends up being reused | |
1004 | * before the original execution finishes, workqueue will identify the | |
1005 | * recycled work item as currently executing and make it wait until the | |
1006 | * current execution finishes, introducing an unwanted dependency. | |
1007 | * | |
c5aa87bb TH |
1008 | * This function checks the work item address and work function to avoid |
1009 | * false positives. Note that this isn't complete as one may construct a | |
1010 | * work function which can introduce dependency onto itself through a | |
1011 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1012 | * foot that badly, there's only so much we can do, and if such deadlock | |
1013 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
1014 | * |
1015 | * CONTEXT: | |
d565ed63 | 1016 | * spin_lock_irq(pool->lock). |
8cca0eea | 1017 | * |
d185af30 YB |
1018 | * Return: |
1019 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 1020 | * otherwise. |
4d707b9f | 1021 | */ |
c9e7cf27 | 1022 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1023 | struct work_struct *work) |
4d707b9f | 1024 | { |
42f8570f | 1025 | struct worker *worker; |
42f8570f | 1026 | |
b67bfe0d | 1027 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1028 | (unsigned long)work) |
1029 | if (worker->current_work == work && | |
1030 | worker->current_func == work->func) | |
42f8570f SL |
1031 | return worker; |
1032 | ||
1033 | return NULL; | |
4d707b9f ON |
1034 | } |
1035 | ||
bf4ede01 TH |
1036 | /** |
1037 | * move_linked_works - move linked works to a list | |
1038 | * @work: start of series of works to be scheduled | |
1039 | * @head: target list to append @work to | |
402dd89d | 1040 | * @nextp: out parameter for nested worklist walking |
bf4ede01 TH |
1041 | * |
1042 | * Schedule linked works starting from @work to @head. Work series to | |
1043 | * be scheduled starts at @work and includes any consecutive work with | |
1044 | * WORK_STRUCT_LINKED set in its predecessor. | |
1045 | * | |
1046 | * If @nextp is not NULL, it's updated to point to the next work of | |
1047 | * the last scheduled work. This allows move_linked_works() to be | |
1048 | * nested inside outer list_for_each_entry_safe(). | |
1049 | * | |
1050 | * CONTEXT: | |
d565ed63 | 1051 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
1052 | */ |
1053 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1054 | struct work_struct **nextp) | |
1055 | { | |
1056 | struct work_struct *n; | |
1057 | ||
1058 | /* | |
1059 | * Linked worklist will always end before the end of the list, | |
1060 | * use NULL for list head. | |
1061 | */ | |
1062 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1063 | list_move_tail(&work->entry, head); | |
1064 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1065 | break; | |
1066 | } | |
1067 | ||
1068 | /* | |
1069 | * If we're already inside safe list traversal and have moved | |
1070 | * multiple works to the scheduled queue, the next position | |
1071 | * needs to be updated. | |
1072 | */ | |
1073 | if (nextp) | |
1074 | *nextp = n; | |
1075 | } | |
1076 | ||
8864b4e5 TH |
1077 | /** |
1078 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1079 | * @pwq: pool_workqueue to get | |
1080 | * | |
1081 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1082 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1083 | */ | |
1084 | static void get_pwq(struct pool_workqueue *pwq) | |
1085 | { | |
1086 | lockdep_assert_held(&pwq->pool->lock); | |
1087 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1088 | pwq->refcnt++; | |
1089 | } | |
1090 | ||
1091 | /** | |
1092 | * put_pwq - put a pool_workqueue reference | |
1093 | * @pwq: pool_workqueue to put | |
1094 | * | |
1095 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1096 | * destruction. The caller should be holding the matching pool->lock. | |
1097 | */ | |
1098 | static void put_pwq(struct pool_workqueue *pwq) | |
1099 | { | |
1100 | lockdep_assert_held(&pwq->pool->lock); | |
1101 | if (likely(--pwq->refcnt)) | |
1102 | return; | |
1103 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1104 | return; | |
1105 | /* | |
1106 | * @pwq can't be released under pool->lock, bounce to | |
1107 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1108 | * pool->lock as this path is taken only for unbound workqueues and | |
1109 | * the release work item is scheduled on a per-cpu workqueue. To | |
1110 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1111 | * subclass of 1 in get_unbound_pool(). | |
1112 | */ | |
1113 | schedule_work(&pwq->unbound_release_work); | |
1114 | } | |
1115 | ||
dce90d47 TH |
1116 | /** |
1117 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1118 | * @pwq: pool_workqueue to put (can be %NULL) | |
1119 | * | |
1120 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1121 | */ | |
1122 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1123 | { | |
1124 | if (pwq) { | |
1125 | /* | |
24acfb71 | 1126 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1127 | * following lock operations are safe. |
1128 | */ | |
1129 | spin_lock_irq(&pwq->pool->lock); | |
1130 | put_pwq(pwq); | |
1131 | spin_unlock_irq(&pwq->pool->lock); | |
1132 | } | |
1133 | } | |
1134 | ||
112202d9 | 1135 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 1136 | { |
112202d9 | 1137 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1138 | |
1139 | trace_workqueue_activate_work(work); | |
82607adc TH |
1140 | if (list_empty(&pwq->pool->worklist)) |
1141 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1142 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1143 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1144 | pwq->nr_active++; |
bf4ede01 TH |
1145 | } |
1146 | ||
112202d9 | 1147 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1148 | { |
112202d9 | 1149 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1150 | struct work_struct, entry); |
1151 | ||
112202d9 | 1152 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1153 | } |
1154 | ||
bf4ede01 | 1155 | /** |
112202d9 TH |
1156 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1157 | * @pwq: pwq of interest | |
bf4ede01 | 1158 | * @color: color of work which left the queue |
bf4ede01 TH |
1159 | * |
1160 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1161 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1162 | * |
1163 | * CONTEXT: | |
d565ed63 | 1164 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1165 | */ |
112202d9 | 1166 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 | 1167 | { |
8864b4e5 | 1168 | /* uncolored work items don't participate in flushing or nr_active */ |
bf4ede01 | 1169 | if (color == WORK_NO_COLOR) |
8864b4e5 | 1170 | goto out_put; |
bf4ede01 | 1171 | |
112202d9 | 1172 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1173 | |
112202d9 TH |
1174 | pwq->nr_active--; |
1175 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1176 | /* one down, submit a delayed one */ |
112202d9 TH |
1177 | if (pwq->nr_active < pwq->max_active) |
1178 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1179 | } |
1180 | ||
1181 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1182 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1183 | goto out_put; |
bf4ede01 TH |
1184 | |
1185 | /* are there still in-flight works? */ | |
112202d9 | 1186 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1187 | goto out_put; |
bf4ede01 | 1188 | |
112202d9 TH |
1189 | /* this pwq is done, clear flush_color */ |
1190 | pwq->flush_color = -1; | |
bf4ede01 TH |
1191 | |
1192 | /* | |
112202d9 | 1193 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1194 | * will handle the rest. |
1195 | */ | |
112202d9 TH |
1196 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1197 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1198 | out_put: |
1199 | put_pwq(pwq); | |
bf4ede01 TH |
1200 | } |
1201 | ||
36e227d2 | 1202 | /** |
bbb68dfa | 1203 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1204 | * @work: work item to steal |
1205 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1206 | * @flags: place to store irq state |
36e227d2 TH |
1207 | * |
1208 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1209 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1210 | * |
d185af30 | 1211 | * Return: |
36e227d2 TH |
1212 | * 1 if @work was pending and we successfully stole PENDING |
1213 | * 0 if @work was idle and we claimed PENDING | |
1214 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1215 | * -ENOENT if someone else is canceling @work, this state may persist |
1216 | * for arbitrarily long | |
36e227d2 | 1217 | * |
d185af30 | 1218 | * Note: |
bbb68dfa | 1219 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1220 | * interrupted while holding PENDING and @work off queue, irq must be |
1221 | * disabled on entry. This, combined with delayed_work->timer being | |
1222 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1223 | * |
1224 | * On successful return, >= 0, irq is disabled and the caller is | |
1225 | * responsible for releasing it using local_irq_restore(*@flags). | |
1226 | * | |
e0aecdd8 | 1227 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1228 | */ |
bbb68dfa TH |
1229 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1230 | unsigned long *flags) | |
bf4ede01 | 1231 | { |
d565ed63 | 1232 | struct worker_pool *pool; |
112202d9 | 1233 | struct pool_workqueue *pwq; |
bf4ede01 | 1234 | |
bbb68dfa TH |
1235 | local_irq_save(*flags); |
1236 | ||
36e227d2 TH |
1237 | /* try to steal the timer if it exists */ |
1238 | if (is_dwork) { | |
1239 | struct delayed_work *dwork = to_delayed_work(work); | |
1240 | ||
e0aecdd8 TH |
1241 | /* |
1242 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1243 | * guaranteed that the timer is not queued anywhere and not | |
1244 | * running on the local CPU. | |
1245 | */ | |
36e227d2 TH |
1246 | if (likely(del_timer(&dwork->timer))) |
1247 | return 1; | |
1248 | } | |
1249 | ||
1250 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1251 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1252 | return 0; | |
1253 | ||
24acfb71 | 1254 | rcu_read_lock(); |
bf4ede01 TH |
1255 | /* |
1256 | * The queueing is in progress, or it is already queued. Try to | |
1257 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1258 | */ | |
d565ed63 TH |
1259 | pool = get_work_pool(work); |
1260 | if (!pool) | |
bbb68dfa | 1261 | goto fail; |
bf4ede01 | 1262 | |
d565ed63 | 1263 | spin_lock(&pool->lock); |
0b3dae68 | 1264 | /* |
112202d9 TH |
1265 | * work->data is guaranteed to point to pwq only while the work |
1266 | * item is queued on pwq->wq, and both updating work->data to point | |
1267 | * to pwq on queueing and to pool on dequeueing are done under | |
1268 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1269 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1270 | * item is currently queued on that pool. |
1271 | */ | |
112202d9 TH |
1272 | pwq = get_work_pwq(work); |
1273 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1274 | debug_work_deactivate(work); |
1275 | ||
1276 | /* | |
1277 | * A delayed work item cannot be grabbed directly because | |
1278 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1279 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1280 | * management later on and cause stall. Make sure the work |
1281 | * item is activated before grabbing. | |
1282 | */ | |
1283 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1284 | pwq_activate_delayed_work(work); |
16062836 TH |
1285 | |
1286 | list_del_init(&work->entry); | |
9c34a704 | 1287 | pwq_dec_nr_in_flight(pwq, get_work_color(work)); |
16062836 | 1288 | |
112202d9 | 1289 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1290 | set_work_pool_and_keep_pending(work, pool->id); |
1291 | ||
1292 | spin_unlock(&pool->lock); | |
24acfb71 | 1293 | rcu_read_unlock(); |
16062836 | 1294 | return 1; |
bf4ede01 | 1295 | } |
d565ed63 | 1296 | spin_unlock(&pool->lock); |
bbb68dfa | 1297 | fail: |
24acfb71 | 1298 | rcu_read_unlock(); |
bbb68dfa TH |
1299 | local_irq_restore(*flags); |
1300 | if (work_is_canceling(work)) | |
1301 | return -ENOENT; | |
1302 | cpu_relax(); | |
36e227d2 | 1303 | return -EAGAIN; |
bf4ede01 TH |
1304 | } |
1305 | ||
4690c4ab | 1306 | /** |
706026c2 | 1307 | * insert_work - insert a work into a pool |
112202d9 | 1308 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1309 | * @work: work to insert |
1310 | * @head: insertion point | |
1311 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1312 | * | |
112202d9 | 1313 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1314 | * work_struct flags. |
4690c4ab TH |
1315 | * |
1316 | * CONTEXT: | |
d565ed63 | 1317 | * spin_lock_irq(pool->lock). |
4690c4ab | 1318 | */ |
112202d9 TH |
1319 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1320 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1321 | { |
112202d9 | 1322 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1323 | |
4690c4ab | 1324 | /* we own @work, set data and link */ |
112202d9 | 1325 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1326 | list_add_tail(&work->entry, head); |
8864b4e5 | 1327 | get_pwq(pwq); |
e22bee78 TH |
1328 | |
1329 | /* | |
c5aa87bb TH |
1330 | * Ensure either wq_worker_sleeping() sees the above |
1331 | * list_add_tail() or we see zero nr_running to avoid workers lying | |
1332 | * around lazily while there are works to be processed. | |
e22bee78 TH |
1333 | */ |
1334 | smp_mb(); | |
1335 | ||
63d95a91 TH |
1336 | if (__need_more_worker(pool)) |
1337 | wake_up_worker(pool); | |
b89deed3 ON |
1338 | } |
1339 | ||
c8efcc25 TH |
1340 | /* |
1341 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1342 | * same workqueue. |
c8efcc25 TH |
1343 | */ |
1344 | static bool is_chained_work(struct workqueue_struct *wq) | |
1345 | { | |
8d03ecfe TH |
1346 | struct worker *worker; |
1347 | ||
1348 | worker = current_wq_worker(); | |
1349 | /* | |
bf393fd4 | 1350 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
1351 | * I'm @worker, it's safe to dereference it without locking. |
1352 | */ | |
112202d9 | 1353 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1354 | } |
1355 | ||
ef557180 MG |
1356 | /* |
1357 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1358 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1359 | * avoid perturbing sensitive tasks. | |
1360 | */ | |
1361 | static int wq_select_unbound_cpu(int cpu) | |
1362 | { | |
f303fccb | 1363 | static bool printed_dbg_warning; |
ef557180 MG |
1364 | int new_cpu; |
1365 | ||
f303fccb TH |
1366 | if (likely(!wq_debug_force_rr_cpu)) { |
1367 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1368 | return cpu; | |
1369 | } else if (!printed_dbg_warning) { | |
1370 | pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
1371 | printed_dbg_warning = true; | |
1372 | } | |
1373 | ||
ef557180 MG |
1374 | if (cpumask_empty(wq_unbound_cpumask)) |
1375 | return cpu; | |
1376 | ||
1377 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1378 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1379 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1380 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1381 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1382 | return cpu; | |
1383 | } | |
1384 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1385 | ||
1386 | return new_cpu; | |
1387 | } | |
1388 | ||
d84ff051 | 1389 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1390 | struct work_struct *work) |
1391 | { | |
112202d9 | 1392 | struct pool_workqueue *pwq; |
c9178087 | 1393 | struct worker_pool *last_pool; |
1e19ffc6 | 1394 | struct list_head *worklist; |
8a2e8e5d | 1395 | unsigned int work_flags; |
b75cac93 | 1396 | unsigned int req_cpu = cpu; |
8930caba TH |
1397 | |
1398 | /* | |
1399 | * While a work item is PENDING && off queue, a task trying to | |
1400 | * steal the PENDING will busy-loop waiting for it to either get | |
1401 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1402 | * happen with IRQ disabled. | |
1403 | */ | |
8e8eb730 | 1404 | lockdep_assert_irqs_disabled(); |
1da177e4 | 1405 | |
dc186ad7 | 1406 | debug_work_activate(work); |
1e19ffc6 | 1407 | |
9ef28a73 | 1408 | /* if draining, only works from the same workqueue are allowed */ |
618b01eb | 1409 | if (unlikely(wq->flags & __WQ_DRAINING) && |
c8efcc25 | 1410 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b | 1411 | return; |
24acfb71 | 1412 | rcu_read_lock(); |
9e8cd2f5 | 1413 | retry: |
df2d5ae4 | 1414 | if (req_cpu == WORK_CPU_UNBOUND) |
ef557180 | 1415 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
df2d5ae4 | 1416 | |
c9178087 | 1417 | /* pwq which will be used unless @work is executing elsewhere */ |
df2d5ae4 | 1418 | if (!(wq->flags & WQ_UNBOUND)) |
7fb98ea7 | 1419 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
df2d5ae4 TH |
1420 | else |
1421 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); | |
dbf2576e | 1422 | |
c9178087 TH |
1423 | /* |
1424 | * If @work was previously on a different pool, it might still be | |
1425 | * running there, in which case the work needs to be queued on that | |
1426 | * pool to guarantee non-reentrancy. | |
1427 | */ | |
1428 | last_pool = get_work_pool(work); | |
1429 | if (last_pool && last_pool != pwq->pool) { | |
1430 | struct worker *worker; | |
18aa9eff | 1431 | |
c9178087 | 1432 | spin_lock(&last_pool->lock); |
18aa9eff | 1433 | |
c9178087 | 1434 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1435 | |
c9178087 TH |
1436 | if (worker && worker->current_pwq->wq == wq) { |
1437 | pwq = worker->current_pwq; | |
8930caba | 1438 | } else { |
c9178087 TH |
1439 | /* meh... not running there, queue here */ |
1440 | spin_unlock(&last_pool->lock); | |
112202d9 | 1441 | spin_lock(&pwq->pool->lock); |
8930caba | 1442 | } |
f3421797 | 1443 | } else { |
112202d9 | 1444 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1445 | } |
1446 | ||
9e8cd2f5 TH |
1447 | /* |
1448 | * pwq is determined and locked. For unbound pools, we could have | |
1449 | * raced with pwq release and it could already be dead. If its | |
1450 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1451 | * without another pwq replacing it in the numa_pwq_tbl or while |
1452 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1453 | * make forward-progress. |
1454 | */ | |
1455 | if (unlikely(!pwq->refcnt)) { | |
1456 | if (wq->flags & WQ_UNBOUND) { | |
1457 | spin_unlock(&pwq->pool->lock); | |
1458 | cpu_relax(); | |
1459 | goto retry; | |
1460 | } | |
1461 | /* oops */ | |
1462 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1463 | wq->name, cpu); | |
1464 | } | |
1465 | ||
112202d9 TH |
1466 | /* pwq determined, queue */ |
1467 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1468 | |
24acfb71 TG |
1469 | if (WARN_ON(!list_empty(&work->entry))) |
1470 | goto out; | |
1e19ffc6 | 1471 | |
112202d9 TH |
1472 | pwq->nr_in_flight[pwq->work_color]++; |
1473 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1474 | |
112202d9 | 1475 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1476 | trace_workqueue_activate_work(work); |
112202d9 TH |
1477 | pwq->nr_active++; |
1478 | worklist = &pwq->pool->worklist; | |
82607adc TH |
1479 | if (list_empty(worklist)) |
1480 | pwq->pool->watchdog_ts = jiffies; | |
8a2e8e5d TH |
1481 | } else { |
1482 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1483 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1484 | } |
1e19ffc6 | 1485 | |
112202d9 | 1486 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1487 | |
24acfb71 | 1488 | out: |
112202d9 | 1489 | spin_unlock(&pwq->pool->lock); |
24acfb71 | 1490 | rcu_read_unlock(); |
1da177e4 LT |
1491 | } |
1492 | ||
0fcb78c2 | 1493 | /** |
c1a220e7 ZR |
1494 | * queue_work_on - queue work on specific cpu |
1495 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1496 | * @wq: workqueue to use |
1497 | * @work: work to queue | |
1498 | * | |
c1a220e7 ZR |
1499 | * We queue the work to a specific CPU, the caller must ensure it |
1500 | * can't go away. | |
d185af30 YB |
1501 | * |
1502 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1503 | */ |
d4283e93 TH |
1504 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1505 | struct work_struct *work) | |
1da177e4 | 1506 | { |
d4283e93 | 1507 | bool ret = false; |
8930caba | 1508 | unsigned long flags; |
ef1ca236 | 1509 | |
8930caba | 1510 | local_irq_save(flags); |
c1a220e7 | 1511 | |
22df02bb | 1512 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1513 | __queue_work(cpu, wq, work); |
d4283e93 | 1514 | ret = true; |
c1a220e7 | 1515 | } |
ef1ca236 | 1516 | |
8930caba | 1517 | local_irq_restore(flags); |
1da177e4 LT |
1518 | return ret; |
1519 | } | |
ad7b1f84 | 1520 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1521 | |
8204e0c1 AD |
1522 | /** |
1523 | * workqueue_select_cpu_near - Select a CPU based on NUMA node | |
1524 | * @node: NUMA node ID that we want to select a CPU from | |
1525 | * | |
1526 | * This function will attempt to find a "random" cpu available on a given | |
1527 | * node. If there are no CPUs available on the given node it will return | |
1528 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
1529 | * available CPU if we need to schedule this work. | |
1530 | */ | |
1531 | static int workqueue_select_cpu_near(int node) | |
1532 | { | |
1533 | int cpu; | |
1534 | ||
1535 | /* No point in doing this if NUMA isn't enabled for workqueues */ | |
1536 | if (!wq_numa_enabled) | |
1537 | return WORK_CPU_UNBOUND; | |
1538 | ||
1539 | /* Delay binding to CPU if node is not valid or online */ | |
1540 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
1541 | return WORK_CPU_UNBOUND; | |
1542 | ||
1543 | /* Use local node/cpu if we are already there */ | |
1544 | cpu = raw_smp_processor_id(); | |
1545 | if (node == cpu_to_node(cpu)) | |
1546 | return cpu; | |
1547 | ||
1548 | /* Use "random" otherwise know as "first" online CPU of node */ | |
1549 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
1550 | ||
1551 | /* If CPU is valid return that, otherwise just defer */ | |
1552 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
1553 | } | |
1554 | ||
1555 | /** | |
1556 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
1557 | * @node: NUMA node that we are targeting the work for | |
1558 | * @wq: workqueue to use | |
1559 | * @work: work to queue | |
1560 | * | |
1561 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
1562 | * idea here is to provide a way to somehow associate work with a given | |
1563 | * NUMA node. | |
1564 | * | |
1565 | * This function will only make a best effort attempt at getting this onto | |
1566 | * the right NUMA node. If no node is requested or the requested node is | |
1567 | * offline then we just fall back to standard queue_work behavior. | |
1568 | * | |
1569 | * Currently the "random" CPU ends up being the first available CPU in the | |
1570 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
1571 | * are running on the node. In that case we just use the current CPU. | |
1572 | * | |
1573 | * Return: %false if @work was already on a queue, %true otherwise. | |
1574 | */ | |
1575 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
1576 | struct work_struct *work) | |
1577 | { | |
1578 | unsigned long flags; | |
1579 | bool ret = false; | |
1580 | ||
1581 | /* | |
1582 | * This current implementation is specific to unbound workqueues. | |
1583 | * Specifically we only return the first available CPU for a given | |
1584 | * node instead of cycling through individual CPUs within the node. | |
1585 | * | |
1586 | * If this is used with a per-cpu workqueue then the logic in | |
1587 | * workqueue_select_cpu_near would need to be updated to allow for | |
1588 | * some round robin type logic. | |
1589 | */ | |
1590 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
1591 | ||
1592 | local_irq_save(flags); | |
1593 | ||
1594 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
1595 | int cpu = workqueue_select_cpu_near(node); | |
1596 | ||
1597 | __queue_work(cpu, wq, work); | |
1598 | ret = true; | |
1599 | } | |
1600 | ||
1601 | local_irq_restore(flags); | |
1602 | return ret; | |
1603 | } | |
1604 | EXPORT_SYMBOL_GPL(queue_work_node); | |
1605 | ||
8c20feb6 | 1606 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 1607 | { |
8c20feb6 | 1608 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 1609 | |
e0aecdd8 | 1610 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1611 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1612 | } |
1438ade5 | 1613 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1614 | |
7beb2edf TH |
1615 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1616 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1617 | { |
7beb2edf TH |
1618 | struct timer_list *timer = &dwork->timer; |
1619 | struct work_struct *work = &dwork->work; | |
7beb2edf | 1620 | |
637fdbae | 1621 | WARN_ON_ONCE(!wq); |
841b86f3 | 1622 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
1623 | WARN_ON_ONCE(timer_pending(timer)); |
1624 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1625 | |
8852aac2 TH |
1626 | /* |
1627 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1628 | * both optimization and correctness. The earliest @timer can | |
1629 | * expire is on the closest next tick and delayed_work users depend | |
1630 | * on that there's no such delay when @delay is 0. | |
1631 | */ | |
1632 | if (!delay) { | |
1633 | __queue_work(cpu, wq, &dwork->work); | |
1634 | return; | |
1635 | } | |
1636 | ||
60c057bc | 1637 | dwork->wq = wq; |
1265057f | 1638 | dwork->cpu = cpu; |
7beb2edf TH |
1639 | timer->expires = jiffies + delay; |
1640 | ||
041bd12e TH |
1641 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1642 | add_timer_on(timer, cpu); | |
1643 | else | |
1644 | add_timer(timer); | |
1da177e4 LT |
1645 | } |
1646 | ||
0fcb78c2 REB |
1647 | /** |
1648 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1649 | * @cpu: CPU number to execute work on | |
1650 | * @wq: workqueue to use | |
af9997e4 | 1651 | * @dwork: work to queue |
0fcb78c2 REB |
1652 | * @delay: number of jiffies to wait before queueing |
1653 | * | |
d185af30 | 1654 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1655 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1656 | * execution. | |
0fcb78c2 | 1657 | */ |
d4283e93 TH |
1658 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1659 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1660 | { |
52bad64d | 1661 | struct work_struct *work = &dwork->work; |
d4283e93 | 1662 | bool ret = false; |
8930caba | 1663 | unsigned long flags; |
7a6bc1cd | 1664 | |
8930caba TH |
1665 | /* read the comment in __queue_work() */ |
1666 | local_irq_save(flags); | |
7a6bc1cd | 1667 | |
22df02bb | 1668 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1669 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1670 | ret = true; |
7a6bc1cd | 1671 | } |
8a3e77cc | 1672 | |
8930caba | 1673 | local_irq_restore(flags); |
7a6bc1cd VP |
1674 | return ret; |
1675 | } | |
ad7b1f84 | 1676 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1677 | |
8376fe22 TH |
1678 | /** |
1679 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1680 | * @cpu: CPU number to execute work on | |
1681 | * @wq: workqueue to use | |
1682 | * @dwork: work to queue | |
1683 | * @delay: number of jiffies to wait before queueing | |
1684 | * | |
1685 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1686 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1687 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1688 | * current state. | |
1689 | * | |
d185af30 | 1690 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1691 | * pending and its timer was modified. |
1692 | * | |
e0aecdd8 | 1693 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1694 | * See try_to_grab_pending() for details. |
1695 | */ | |
1696 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1697 | struct delayed_work *dwork, unsigned long delay) | |
1698 | { | |
1699 | unsigned long flags; | |
1700 | int ret; | |
c7fc77f7 | 1701 | |
8376fe22 TH |
1702 | do { |
1703 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1704 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1705 | |
8376fe22 TH |
1706 | if (likely(ret >= 0)) { |
1707 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1708 | local_irq_restore(flags); | |
7a6bc1cd | 1709 | } |
8376fe22 TH |
1710 | |
1711 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1712 | return ret; |
1713 | } | |
8376fe22 TH |
1714 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1715 | ||
05f0fe6b TH |
1716 | static void rcu_work_rcufn(struct rcu_head *rcu) |
1717 | { | |
1718 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
1719 | ||
1720 | /* read the comment in __queue_work() */ | |
1721 | local_irq_disable(); | |
1722 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
1723 | local_irq_enable(); | |
1724 | } | |
1725 | ||
1726 | /** | |
1727 | * queue_rcu_work - queue work after a RCU grace period | |
1728 | * @wq: workqueue to use | |
1729 | * @rwork: work to queue | |
1730 | * | |
1731 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
1732 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 1733 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
1734 | * execution may happen before a full RCU grace period has passed. |
1735 | */ | |
1736 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
1737 | { | |
1738 | struct work_struct *work = &rwork->work; | |
1739 | ||
1740 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
1741 | rwork->wq = wq; | |
1742 | call_rcu(&rwork->rcu, rcu_work_rcufn); | |
1743 | return true; | |
1744 | } | |
1745 | ||
1746 | return false; | |
1747 | } | |
1748 | EXPORT_SYMBOL(queue_rcu_work); | |
1749 | ||
c8e55f36 TH |
1750 | /** |
1751 | * worker_enter_idle - enter idle state | |
1752 | * @worker: worker which is entering idle state | |
1753 | * | |
1754 | * @worker is entering idle state. Update stats and idle timer if | |
1755 | * necessary. | |
1756 | * | |
1757 | * LOCKING: | |
d565ed63 | 1758 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1759 | */ |
1760 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1761 | { |
bd7bdd43 | 1762 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1763 | |
6183c009 TH |
1764 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1765 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1766 | (worker->hentry.next || worker->hentry.pprev))) | |
1767 | return; | |
c8e55f36 | 1768 | |
051e1850 | 1769 | /* can't use worker_set_flags(), also called from create_worker() */ |
cb444766 | 1770 | worker->flags |= WORKER_IDLE; |
bd7bdd43 | 1771 | pool->nr_idle++; |
e22bee78 | 1772 | worker->last_active = jiffies; |
c8e55f36 TH |
1773 | |
1774 | /* idle_list is LIFO */ | |
bd7bdd43 | 1775 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1776 | |
628c78e7 TH |
1777 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1778 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1779 | |
544ecf31 | 1780 | /* |
e8b3f8db | 1781 | * Sanity check nr_running. Because unbind_workers() releases |
d565ed63 | 1782 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1783 | * nr_running, the warning may trigger spuriously. Check iff |
1784 | * unbind is not in progress. | |
544ecf31 | 1785 | */ |
24647570 | 1786 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1787 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1788 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1789 | } |
1790 | ||
1791 | /** | |
1792 | * worker_leave_idle - leave idle state | |
1793 | * @worker: worker which is leaving idle state | |
1794 | * | |
1795 | * @worker is leaving idle state. Update stats. | |
1796 | * | |
1797 | * LOCKING: | |
d565ed63 | 1798 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1799 | */ |
1800 | static void worker_leave_idle(struct worker *worker) | |
1801 | { | |
bd7bdd43 | 1802 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1803 | |
6183c009 TH |
1804 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1805 | return; | |
d302f017 | 1806 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1807 | pool->nr_idle--; |
c8e55f36 TH |
1808 | list_del_init(&worker->entry); |
1809 | } | |
1810 | ||
f7537df5 | 1811 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
1812 | { |
1813 | struct worker *worker; | |
1814 | ||
f7537df5 | 1815 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
1816 | if (worker) { |
1817 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1818 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 1819 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
1820 | /* on creation a worker is in !idle && prep state */ |
1821 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1822 | } |
c34056a3 TH |
1823 | return worker; |
1824 | } | |
1825 | ||
4736cbf7 LJ |
1826 | /** |
1827 | * worker_attach_to_pool() - attach a worker to a pool | |
1828 | * @worker: worker to be attached | |
1829 | * @pool: the target pool | |
1830 | * | |
1831 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
1832 | * cpu-binding of @worker are kept coordinated with the pool across | |
1833 | * cpu-[un]hotplugs. | |
1834 | */ | |
1835 | static void worker_attach_to_pool(struct worker *worker, | |
1836 | struct worker_pool *pool) | |
1837 | { | |
1258fae7 | 1838 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
1839 | |
1840 | /* | |
1841 | * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any | |
1842 | * online CPUs. It'll be re-applied when any of the CPUs come up. | |
1843 | */ | |
1844 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
1845 | ||
1846 | /* | |
1258fae7 TH |
1847 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
1848 | * stable across this function. See the comments above the flag | |
1849 | * definition for details. | |
4736cbf7 LJ |
1850 | */ |
1851 | if (pool->flags & POOL_DISASSOCIATED) | |
1852 | worker->flags |= WORKER_UNBOUND; | |
1853 | ||
1854 | list_add_tail(&worker->node, &pool->workers); | |
a2d812a2 | 1855 | worker->pool = pool; |
4736cbf7 | 1856 | |
1258fae7 | 1857 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
1858 | } |
1859 | ||
60f5a4bc LJ |
1860 | /** |
1861 | * worker_detach_from_pool() - detach a worker from its pool | |
1862 | * @worker: worker which is attached to its pool | |
60f5a4bc | 1863 | * |
4736cbf7 LJ |
1864 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
1865 | * caller worker shouldn't access to the pool after detached except it has | |
1866 | * other reference to the pool. | |
60f5a4bc | 1867 | */ |
a2d812a2 | 1868 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 1869 | { |
a2d812a2 | 1870 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
1871 | struct completion *detach_completion = NULL; |
1872 | ||
1258fae7 | 1873 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 1874 | |
da028469 | 1875 | list_del(&worker->node); |
a2d812a2 TH |
1876 | worker->pool = NULL; |
1877 | ||
da028469 | 1878 | if (list_empty(&pool->workers)) |
60f5a4bc | 1879 | detach_completion = pool->detach_completion; |
1258fae7 | 1880 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 1881 | |
b62c0751 LJ |
1882 | /* clear leftover flags without pool->lock after it is detached */ |
1883 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
1884 | ||
60f5a4bc LJ |
1885 | if (detach_completion) |
1886 | complete(detach_completion); | |
1887 | } | |
1888 | ||
c34056a3 TH |
1889 | /** |
1890 | * create_worker - create a new workqueue worker | |
63d95a91 | 1891 | * @pool: pool the new worker will belong to |
c34056a3 | 1892 | * |
051e1850 | 1893 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
1894 | * |
1895 | * CONTEXT: | |
1896 | * Might sleep. Does GFP_KERNEL allocations. | |
1897 | * | |
d185af30 | 1898 | * Return: |
c34056a3 TH |
1899 | * Pointer to the newly created worker. |
1900 | */ | |
bc2ae0f5 | 1901 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1902 | { |
c34056a3 | 1903 | struct worker *worker = NULL; |
f3421797 | 1904 | int id = -1; |
e3c916a4 | 1905 | char id_buf[16]; |
c34056a3 | 1906 | |
7cda9aae LJ |
1907 | /* ID is needed to determine kthread name */ |
1908 | id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); | |
822d8405 TH |
1909 | if (id < 0) |
1910 | goto fail; | |
c34056a3 | 1911 | |
f7537df5 | 1912 | worker = alloc_worker(pool->node); |
c34056a3 TH |
1913 | if (!worker) |
1914 | goto fail; | |
1915 | ||
c34056a3 TH |
1916 | worker->id = id; |
1917 | ||
29c91e99 | 1918 | if (pool->cpu >= 0) |
e3c916a4 TH |
1919 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1920 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1921 | else |
e3c916a4 TH |
1922 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1923 | ||
f3f90ad4 | 1924 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1925 | "kworker/%s", id_buf); |
c34056a3 TH |
1926 | if (IS_ERR(worker->task)) |
1927 | goto fail; | |
1928 | ||
91151228 | 1929 | set_user_nice(worker->task, pool->attrs->nice); |
25834c73 | 1930 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
91151228 | 1931 | |
da028469 | 1932 | /* successful, attach the worker to the pool */ |
4736cbf7 | 1933 | worker_attach_to_pool(worker, pool); |
822d8405 | 1934 | |
051e1850 LJ |
1935 | /* start the newly created worker */ |
1936 | spin_lock_irq(&pool->lock); | |
1937 | worker->pool->nr_workers++; | |
1938 | worker_enter_idle(worker); | |
1939 | wake_up_process(worker->task); | |
1940 | spin_unlock_irq(&pool->lock); | |
1941 | ||
c34056a3 | 1942 | return worker; |
822d8405 | 1943 | |
c34056a3 | 1944 | fail: |
9625ab17 | 1945 | if (id >= 0) |
7cda9aae | 1946 | ida_simple_remove(&pool->worker_ida, id); |
c34056a3 TH |
1947 | kfree(worker); |
1948 | return NULL; | |
1949 | } | |
1950 | ||
c34056a3 TH |
1951 | /** |
1952 | * destroy_worker - destroy a workqueue worker | |
1953 | * @worker: worker to be destroyed | |
1954 | * | |
73eb7fe7 LJ |
1955 | * Destroy @worker and adjust @pool stats accordingly. The worker should |
1956 | * be idle. | |
c8e55f36 TH |
1957 | * |
1958 | * CONTEXT: | |
60f5a4bc | 1959 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1960 | */ |
1961 | static void destroy_worker(struct worker *worker) | |
1962 | { | |
bd7bdd43 | 1963 | struct worker_pool *pool = worker->pool; |
c34056a3 | 1964 | |
cd549687 TH |
1965 | lockdep_assert_held(&pool->lock); |
1966 | ||
c34056a3 | 1967 | /* sanity check frenzy */ |
6183c009 | 1968 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
1969 | WARN_ON(!list_empty(&worker->scheduled)) || |
1970 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 1971 | return; |
c34056a3 | 1972 | |
73eb7fe7 LJ |
1973 | pool->nr_workers--; |
1974 | pool->nr_idle--; | |
5bdfff96 | 1975 | |
c8e55f36 | 1976 | list_del_init(&worker->entry); |
cb444766 | 1977 | worker->flags |= WORKER_DIE; |
60f5a4bc | 1978 | wake_up_process(worker->task); |
c34056a3 TH |
1979 | } |
1980 | ||
32a6c723 | 1981 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 1982 | { |
32a6c723 | 1983 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
e22bee78 | 1984 | |
d565ed63 | 1985 | spin_lock_irq(&pool->lock); |
e22bee78 | 1986 | |
3347fc9f | 1987 | while (too_many_workers(pool)) { |
e22bee78 TH |
1988 | struct worker *worker; |
1989 | unsigned long expires; | |
1990 | ||
1991 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1992 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1993 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1994 | ||
3347fc9f | 1995 | if (time_before(jiffies, expires)) { |
63d95a91 | 1996 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 1997 | break; |
d5abe669 | 1998 | } |
3347fc9f LJ |
1999 | |
2000 | destroy_worker(worker); | |
e22bee78 TH |
2001 | } |
2002 | ||
d565ed63 | 2003 | spin_unlock_irq(&pool->lock); |
e22bee78 | 2004 | } |
d5abe669 | 2005 | |
493a1724 | 2006 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2007 | { |
112202d9 TH |
2008 | struct pool_workqueue *pwq = get_work_pwq(work); |
2009 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2010 | |
2e109a28 | 2011 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2012 | |
493008a8 | 2013 | if (!wq->rescuer) |
493a1724 | 2014 | return; |
e22bee78 TH |
2015 | |
2016 | /* mayday mayday mayday */ | |
493a1724 | 2017 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2018 | /* |
2019 | * If @pwq is for an unbound wq, its base ref may be put at | |
2020 | * any time due to an attribute change. Pin @pwq until the | |
2021 | * rescuer is done with it. | |
2022 | */ | |
2023 | get_pwq(pwq); | |
493a1724 | 2024 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2025 | wake_up_process(wq->rescuer->task); |
493a1724 | 2026 | } |
e22bee78 TH |
2027 | } |
2028 | ||
32a6c723 | 2029 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 2030 | { |
32a6c723 | 2031 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
2032 | struct work_struct *work; |
2033 | ||
b2d82909 TH |
2034 | spin_lock_irq(&pool->lock); |
2035 | spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 2036 | |
63d95a91 | 2037 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
2038 | /* |
2039 | * We've been trying to create a new worker but | |
2040 | * haven't been successful. We might be hitting an | |
2041 | * allocation deadlock. Send distress signals to | |
2042 | * rescuers. | |
2043 | */ | |
63d95a91 | 2044 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 2045 | send_mayday(work); |
1da177e4 | 2046 | } |
e22bee78 | 2047 | |
b2d82909 TH |
2048 | spin_unlock(&wq_mayday_lock); |
2049 | spin_unlock_irq(&pool->lock); | |
e22bee78 | 2050 | |
63d95a91 | 2051 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
2052 | } |
2053 | ||
e22bee78 TH |
2054 | /** |
2055 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 2056 | * @pool: pool to create a new worker for |
e22bee78 | 2057 | * |
63d95a91 | 2058 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
2059 | * have at least one idle worker on return from this function. If |
2060 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 2061 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
2062 | * possible allocation deadlock. |
2063 | * | |
c5aa87bb TH |
2064 | * On return, need_to_create_worker() is guaranteed to be %false and |
2065 | * may_start_working() %true. | |
e22bee78 TH |
2066 | * |
2067 | * LOCKING: | |
d565ed63 | 2068 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2069 | * multiple times. Does GFP_KERNEL allocations. Called only from |
2070 | * manager. | |
e22bee78 | 2071 | */ |
29187a9e | 2072 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
2073 | __releases(&pool->lock) |
2074 | __acquires(&pool->lock) | |
1da177e4 | 2075 | { |
e22bee78 | 2076 | restart: |
d565ed63 | 2077 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 2078 | |
e22bee78 | 2079 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 2080 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
2081 | |
2082 | while (true) { | |
051e1850 | 2083 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 2084 | break; |
1da177e4 | 2085 | |
e212f361 | 2086 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 2087 | |
63d95a91 | 2088 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2089 | break; |
2090 | } | |
2091 | ||
63d95a91 | 2092 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 2093 | spin_lock_irq(&pool->lock); |
051e1850 LJ |
2094 | /* |
2095 | * This is necessary even after a new worker was just successfully | |
2096 | * created as @pool->lock was dropped and the new worker might have | |
2097 | * already become busy. | |
2098 | */ | |
63d95a91 | 2099 | if (need_to_create_worker(pool)) |
e22bee78 | 2100 | goto restart; |
e22bee78 TH |
2101 | } |
2102 | ||
73f53c4a | 2103 | /** |
e22bee78 TH |
2104 | * manage_workers - manage worker pool |
2105 | * @worker: self | |
73f53c4a | 2106 | * |
706026c2 | 2107 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2108 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2109 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2110 | * |
2111 | * The caller can safely start processing works on false return. On | |
2112 | * true return, it's guaranteed that need_to_create_worker() is false | |
2113 | * and may_start_working() is true. | |
73f53c4a TH |
2114 | * |
2115 | * CONTEXT: | |
d565ed63 | 2116 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2117 | * multiple times. Does GFP_KERNEL allocations. |
2118 | * | |
d185af30 | 2119 | * Return: |
29187a9e TH |
2120 | * %false if the pool doesn't need management and the caller can safely |
2121 | * start processing works, %true if management function was performed and | |
2122 | * the conditions that the caller verified before calling the function may | |
2123 | * no longer be true. | |
73f53c4a | 2124 | */ |
e22bee78 | 2125 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2126 | { |
63d95a91 | 2127 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2128 | |
692b4825 | 2129 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2130 | return false; |
692b4825 TH |
2131 | |
2132 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2133 | pool->manager = worker; |
1e19ffc6 | 2134 | |
29187a9e | 2135 | maybe_create_worker(pool); |
e22bee78 | 2136 | |
2607d7a6 | 2137 | pool->manager = NULL; |
692b4825 TH |
2138 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
2139 | wake_up(&wq_manager_wait); | |
29187a9e | 2140 | return true; |
73f53c4a TH |
2141 | } |
2142 | ||
a62428c0 TH |
2143 | /** |
2144 | * process_one_work - process single work | |
c34056a3 | 2145 | * @worker: self |
a62428c0 TH |
2146 | * @work: work to process |
2147 | * | |
2148 | * Process @work. This function contains all the logics necessary to | |
2149 | * process a single work including synchronization against and | |
2150 | * interaction with other workers on the same cpu, queueing and | |
2151 | * flushing. As long as context requirement is met, any worker can | |
2152 | * call this function to process a work. | |
2153 | * | |
2154 | * CONTEXT: | |
d565ed63 | 2155 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2156 | */ |
c34056a3 | 2157 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2158 | __releases(&pool->lock) |
2159 | __acquires(&pool->lock) | |
a62428c0 | 2160 | { |
112202d9 | 2161 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2162 | struct worker_pool *pool = worker->pool; |
112202d9 | 2163 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2164 | int work_color; |
7e11629d | 2165 | struct worker *collision; |
a62428c0 TH |
2166 | #ifdef CONFIG_LOCKDEP |
2167 | /* | |
2168 | * It is permissible to free the struct work_struct from | |
2169 | * inside the function that is called from it, this we need to | |
2170 | * take into account for lockdep too. To avoid bogus "held | |
2171 | * lock freed" warnings as well as problems when looking into | |
2172 | * work->lockdep_map, make a copy and use that here. | |
2173 | */ | |
4d82a1de PZ |
2174 | struct lockdep_map lockdep_map; |
2175 | ||
2176 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2177 | #endif |
807407c0 | 2178 | /* ensure we're on the correct CPU */ |
85327af6 | 2179 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2180 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2181 | |
7e11629d TH |
2182 | /* |
2183 | * A single work shouldn't be executed concurrently by | |
2184 | * multiple workers on a single cpu. Check whether anyone is | |
2185 | * already processing the work. If so, defer the work to the | |
2186 | * currently executing one. | |
2187 | */ | |
c9e7cf27 | 2188 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2189 | if (unlikely(collision)) { |
2190 | move_linked_works(work, &collision->scheduled, NULL); | |
2191 | return; | |
2192 | } | |
2193 | ||
8930caba | 2194 | /* claim and dequeue */ |
a62428c0 | 2195 | debug_work_deactivate(work); |
c9e7cf27 | 2196 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2197 | worker->current_work = work; |
a2c1c57b | 2198 | worker->current_func = work->func; |
112202d9 | 2199 | worker->current_pwq = pwq; |
73f53c4a | 2200 | work_color = get_work_color(work); |
7a22ad75 | 2201 | |
8bf89593 TH |
2202 | /* |
2203 | * Record wq name for cmdline and debug reporting, may get | |
2204 | * overridden through set_worker_desc(). | |
2205 | */ | |
2206 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
2207 | ||
a62428c0 TH |
2208 | list_del_init(&work->entry); |
2209 | ||
fb0e7beb | 2210 | /* |
228f1d00 LJ |
2211 | * CPU intensive works don't participate in concurrency management. |
2212 | * They're the scheduler's responsibility. This takes @worker out | |
2213 | * of concurrency management and the next code block will chain | |
2214 | * execution of the pending work items. | |
fb0e7beb TH |
2215 | */ |
2216 | if (unlikely(cpu_intensive)) | |
228f1d00 | 2217 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2218 | |
974271c4 | 2219 | /* |
a489a03e LJ |
2220 | * Wake up another worker if necessary. The condition is always |
2221 | * false for normal per-cpu workers since nr_running would always | |
2222 | * be >= 1 at this point. This is used to chain execution of the | |
2223 | * pending work items for WORKER_NOT_RUNNING workers such as the | |
228f1d00 | 2224 | * UNBOUND and CPU_INTENSIVE ones. |
974271c4 | 2225 | */ |
a489a03e | 2226 | if (need_more_worker(pool)) |
63d95a91 | 2227 | wake_up_worker(pool); |
974271c4 | 2228 | |
8930caba | 2229 | /* |
7c3eed5c | 2230 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2231 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2232 | * PENDING and queued state changes happen together while IRQ is |
2233 | * disabled. | |
8930caba | 2234 | */ |
7c3eed5c | 2235 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2236 | |
d565ed63 | 2237 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2238 | |
a1d14934 | 2239 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 2240 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 2241 | /* |
f52be570 PZ |
2242 | * Strictly speaking we should mark the invariant state without holding |
2243 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
2244 | * |
2245 | * However, that would result in: | |
2246 | * | |
2247 | * A(W1) | |
2248 | * WFC(C) | |
2249 | * A(W1) | |
2250 | * C(C) | |
2251 | * | |
2252 | * Which would create W1->C->W1 dependencies, even though there is no | |
2253 | * actual deadlock possible. There are two solutions, using a | |
2254 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 2255 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
2256 | * these locks. |
2257 | * | |
2258 | * AFAICT there is no possible deadlock scenario between the | |
2259 | * flush_work() and complete() primitives (except for single-threaded | |
2260 | * workqueues), so hiding them isn't a problem. | |
2261 | */ | |
f52be570 | 2262 | lockdep_invariant_state(true); |
e36c886a | 2263 | trace_workqueue_execute_start(work); |
a2c1c57b | 2264 | worker->current_func(work); |
e36c886a AV |
2265 | /* |
2266 | * While we must be careful to not use "work" after this, the trace | |
2267 | * point will only record its address. | |
2268 | */ | |
1c5da0ec | 2269 | trace_workqueue_execute_end(work, worker->current_func); |
a62428c0 | 2270 | lock_map_release(&lockdep_map); |
112202d9 | 2271 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2272 | |
2273 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c | 2274 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
d75f773c | 2275 | " last function: %ps\n", |
a2c1c57b TH |
2276 | current->comm, preempt_count(), task_pid_nr(current), |
2277 | worker->current_func); | |
a62428c0 TH |
2278 | debug_show_held_locks(current); |
2279 | dump_stack(); | |
2280 | } | |
2281 | ||
b22ce278 | 2282 | /* |
025f50f3 | 2283 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
2284 | * kernels, where a requeueing work item waiting for something to |
2285 | * happen could deadlock with stop_machine as such work item could | |
2286 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2287 | * stop_machine. At the same time, report a quiescent RCU state so |
2288 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2289 | */ |
a7e6425e | 2290 | cond_resched(); |
b22ce278 | 2291 | |
d565ed63 | 2292 | spin_lock_irq(&pool->lock); |
a62428c0 | 2293 | |
fb0e7beb TH |
2294 | /* clear cpu intensive status */ |
2295 | if (unlikely(cpu_intensive)) | |
2296 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2297 | ||
1b69ac6b JW |
2298 | /* tag the worker for identification in schedule() */ |
2299 | worker->last_func = worker->current_func; | |
2300 | ||
a62428c0 | 2301 | /* we're done with it, release */ |
42f8570f | 2302 | hash_del(&worker->hentry); |
c34056a3 | 2303 | worker->current_work = NULL; |
a2c1c57b | 2304 | worker->current_func = NULL; |
112202d9 TH |
2305 | worker->current_pwq = NULL; |
2306 | pwq_dec_nr_in_flight(pwq, work_color); | |
a62428c0 TH |
2307 | } |
2308 | ||
affee4b2 TH |
2309 | /** |
2310 | * process_scheduled_works - process scheduled works | |
2311 | * @worker: self | |
2312 | * | |
2313 | * Process all scheduled works. Please note that the scheduled list | |
2314 | * may change while processing a work, so this function repeatedly | |
2315 | * fetches a work from the top and executes it. | |
2316 | * | |
2317 | * CONTEXT: | |
d565ed63 | 2318 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2319 | * multiple times. |
2320 | */ | |
2321 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2322 | { |
affee4b2 TH |
2323 | while (!list_empty(&worker->scheduled)) { |
2324 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2325 | struct work_struct, entry); |
c34056a3 | 2326 | process_one_work(worker, work); |
1da177e4 | 2327 | } |
1da177e4 LT |
2328 | } |
2329 | ||
197f6acc TH |
2330 | static void set_pf_worker(bool val) |
2331 | { | |
2332 | mutex_lock(&wq_pool_attach_mutex); | |
2333 | if (val) | |
2334 | current->flags |= PF_WQ_WORKER; | |
2335 | else | |
2336 | current->flags &= ~PF_WQ_WORKER; | |
2337 | mutex_unlock(&wq_pool_attach_mutex); | |
2338 | } | |
2339 | ||
4690c4ab TH |
2340 | /** |
2341 | * worker_thread - the worker thread function | |
c34056a3 | 2342 | * @__worker: self |
4690c4ab | 2343 | * |
c5aa87bb TH |
2344 | * The worker thread function. All workers belong to a worker_pool - |
2345 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2346 | * work items regardless of their specific target workqueue. The only | |
2347 | * exception is work items which belong to workqueues with a rescuer which | |
2348 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2349 | * |
2350 | * Return: 0 | |
4690c4ab | 2351 | */ |
c34056a3 | 2352 | static int worker_thread(void *__worker) |
1da177e4 | 2353 | { |
c34056a3 | 2354 | struct worker *worker = __worker; |
bd7bdd43 | 2355 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2356 | |
e22bee78 | 2357 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 2358 | set_pf_worker(true); |
c8e55f36 | 2359 | woke_up: |
d565ed63 | 2360 | spin_lock_irq(&pool->lock); |
1da177e4 | 2361 | |
a9ab775b TH |
2362 | /* am I supposed to die? */ |
2363 | if (unlikely(worker->flags & WORKER_DIE)) { | |
d565ed63 | 2364 | spin_unlock_irq(&pool->lock); |
a9ab775b | 2365 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
197f6acc | 2366 | set_pf_worker(false); |
60f5a4bc LJ |
2367 | |
2368 | set_task_comm(worker->task, "kworker/dying"); | |
7cda9aae | 2369 | ida_simple_remove(&pool->worker_ida, worker->id); |
a2d812a2 | 2370 | worker_detach_from_pool(worker); |
60f5a4bc | 2371 | kfree(worker); |
a9ab775b | 2372 | return 0; |
c8e55f36 | 2373 | } |
affee4b2 | 2374 | |
c8e55f36 | 2375 | worker_leave_idle(worker); |
db7bccf4 | 2376 | recheck: |
e22bee78 | 2377 | /* no more worker necessary? */ |
63d95a91 | 2378 | if (!need_more_worker(pool)) |
e22bee78 TH |
2379 | goto sleep; |
2380 | ||
2381 | /* do we need to manage? */ | |
63d95a91 | 2382 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2383 | goto recheck; |
2384 | ||
c8e55f36 TH |
2385 | /* |
2386 | * ->scheduled list can only be filled while a worker is | |
2387 | * preparing to process a work or actually processing it. | |
2388 | * Make sure nobody diddled with it while I was sleeping. | |
2389 | */ | |
6183c009 | 2390 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2391 | |
e22bee78 | 2392 | /* |
a9ab775b TH |
2393 | * Finish PREP stage. We're guaranteed to have at least one idle |
2394 | * worker or that someone else has already assumed the manager | |
2395 | * role. This is where @worker starts participating in concurrency | |
2396 | * management if applicable and concurrency management is restored | |
2397 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2398 | */ |
a9ab775b | 2399 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2400 | |
2401 | do { | |
c8e55f36 | 2402 | struct work_struct *work = |
bd7bdd43 | 2403 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2404 | struct work_struct, entry); |
2405 | ||
82607adc TH |
2406 | pool->watchdog_ts = jiffies; |
2407 | ||
c8e55f36 TH |
2408 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
2409 | /* optimization path, not strictly necessary */ | |
2410 | process_one_work(worker, work); | |
2411 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2412 | process_scheduled_works(worker); |
c8e55f36 TH |
2413 | } else { |
2414 | move_linked_works(work, &worker->scheduled, NULL); | |
2415 | process_scheduled_works(worker); | |
affee4b2 | 2416 | } |
63d95a91 | 2417 | } while (keep_working(pool)); |
e22bee78 | 2418 | |
228f1d00 | 2419 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2420 | sleep: |
c8e55f36 | 2421 | /* |
d565ed63 TH |
2422 | * pool->lock is held and there's no work to process and no need to |
2423 | * manage, sleep. Workers are woken up only while holding | |
2424 | * pool->lock or from local cpu, so setting the current state | |
2425 | * before releasing pool->lock is enough to prevent losing any | |
2426 | * event. | |
c8e55f36 TH |
2427 | */ |
2428 | worker_enter_idle(worker); | |
c5a94a61 | 2429 | __set_current_state(TASK_IDLE); |
d565ed63 | 2430 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2431 | schedule(); |
2432 | goto woke_up; | |
1da177e4 LT |
2433 | } |
2434 | ||
e22bee78 TH |
2435 | /** |
2436 | * rescuer_thread - the rescuer thread function | |
111c225a | 2437 | * @__rescuer: self |
e22bee78 TH |
2438 | * |
2439 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2440 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2441 | * |
706026c2 | 2442 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2443 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2444 | * developing into deadlock if some works currently on the same queue | |
2445 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2446 | * the problem rescuer solves. | |
2447 | * | |
706026c2 TH |
2448 | * When such condition is possible, the pool summons rescuers of all |
2449 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2450 | * those works so that forward progress can be guaranteed. |
2451 | * | |
2452 | * This should happen rarely. | |
d185af30 YB |
2453 | * |
2454 | * Return: 0 | |
e22bee78 | 2455 | */ |
111c225a | 2456 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2457 | { |
111c225a TH |
2458 | struct worker *rescuer = __rescuer; |
2459 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2460 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2461 | bool should_stop; |
e22bee78 TH |
2462 | |
2463 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2464 | |
2465 | /* | |
2466 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2467 | * doesn't participate in concurrency management. | |
2468 | */ | |
197f6acc | 2469 | set_pf_worker(true); |
e22bee78 | 2470 | repeat: |
c5a94a61 | 2471 | set_current_state(TASK_IDLE); |
e22bee78 | 2472 | |
4d595b86 LJ |
2473 | /* |
2474 | * By the time the rescuer is requested to stop, the workqueue | |
2475 | * shouldn't have any work pending, but @wq->maydays may still have | |
2476 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2477 | * all the work items before the rescuer got to them. Go through | |
2478 | * @wq->maydays processing before acting on should_stop so that the | |
2479 | * list is always empty on exit. | |
2480 | */ | |
2481 | should_stop = kthread_should_stop(); | |
e22bee78 | 2482 | |
493a1724 | 2483 | /* see whether any pwq is asking for help */ |
2e109a28 | 2484 | spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2485 | |
2486 | while (!list_empty(&wq->maydays)) { | |
2487 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2488 | struct pool_workqueue, mayday_node); | |
112202d9 | 2489 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 2490 | struct work_struct *work, *n; |
82607adc | 2491 | bool first = true; |
e22bee78 TH |
2492 | |
2493 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2494 | list_del_init(&pwq->mayday_node); |
2495 | ||
2e109a28 | 2496 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2497 | |
51697d39 LJ |
2498 | worker_attach_to_pool(rescuer, pool); |
2499 | ||
2500 | spin_lock_irq(&pool->lock); | |
e22bee78 TH |
2501 | |
2502 | /* | |
2503 | * Slurp in all works issued via this workqueue and | |
2504 | * process'em. | |
2505 | */ | |
0479c8c5 | 2506 | WARN_ON_ONCE(!list_empty(scheduled)); |
82607adc TH |
2507 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
2508 | if (get_work_pwq(work) == pwq) { | |
2509 | if (first) | |
2510 | pool->watchdog_ts = jiffies; | |
e22bee78 | 2511 | move_linked_works(work, scheduled, &n); |
82607adc TH |
2512 | } |
2513 | first = false; | |
2514 | } | |
e22bee78 | 2515 | |
008847f6 N |
2516 | if (!list_empty(scheduled)) { |
2517 | process_scheduled_works(rescuer); | |
2518 | ||
2519 | /* | |
2520 | * The above execution of rescued work items could | |
2521 | * have created more to rescue through | |
2522 | * pwq_activate_first_delayed() or chained | |
2523 | * queueing. Let's put @pwq back on mayday list so | |
2524 | * that such back-to-back work items, which may be | |
2525 | * being used to relieve memory pressure, don't | |
2526 | * incur MAYDAY_INTERVAL delay inbetween. | |
2527 | */ | |
2528 | if (need_to_create_worker(pool)) { | |
2529 | spin_lock(&wq_mayday_lock); | |
e66b39af TH |
2530 | /* |
2531 | * Queue iff we aren't racing destruction | |
2532 | * and somebody else hasn't queued it already. | |
2533 | */ | |
2534 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
2535 | get_pwq(pwq); | |
2536 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
2537 | } | |
008847f6 N |
2538 | spin_unlock(&wq_mayday_lock); |
2539 | } | |
2540 | } | |
7576958a | 2541 | |
77668c8b LJ |
2542 | /* |
2543 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2544 | * go away while we're still attached to it. |
77668c8b LJ |
2545 | */ |
2546 | put_pwq(pwq); | |
2547 | ||
7576958a | 2548 | /* |
d8ca83e6 | 2549 | * Leave this pool. If need_more_worker() is %true, notify a |
7576958a TH |
2550 | * regular worker; otherwise, we end up with 0 concurrency |
2551 | * and stalling the execution. | |
2552 | */ | |
d8ca83e6 | 2553 | if (need_more_worker(pool)) |
63d95a91 | 2554 | wake_up_worker(pool); |
7576958a | 2555 | |
13b1d625 LJ |
2556 | spin_unlock_irq(&pool->lock); |
2557 | ||
a2d812a2 | 2558 | worker_detach_from_pool(rescuer); |
13b1d625 LJ |
2559 | |
2560 | spin_lock_irq(&wq_mayday_lock); | |
e22bee78 TH |
2561 | } |
2562 | ||
2e109a28 | 2563 | spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2564 | |
4d595b86 LJ |
2565 | if (should_stop) { |
2566 | __set_current_state(TASK_RUNNING); | |
197f6acc | 2567 | set_pf_worker(false); |
4d595b86 LJ |
2568 | return 0; |
2569 | } | |
2570 | ||
111c225a TH |
2571 | /* rescuers should never participate in concurrency management */ |
2572 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2573 | schedule(); |
2574 | goto repeat; | |
1da177e4 LT |
2575 | } |
2576 | ||
fca839c0 TH |
2577 | /** |
2578 | * check_flush_dependency - check for flush dependency sanity | |
2579 | * @target_wq: workqueue being flushed | |
2580 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2581 | * | |
2582 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2583 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2584 | * reclaiming memory or running on a workqueue which doesn't have | |
2585 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2586 | * a deadlock. | |
2587 | */ | |
2588 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2589 | struct work_struct *target_work) | |
2590 | { | |
2591 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2592 | struct worker *worker; | |
2593 | ||
2594 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2595 | return; | |
2596 | ||
2597 | worker = current_wq_worker(); | |
2598 | ||
2599 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 2600 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 2601 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
2602 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2603 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 2604 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
2605 | worker->current_pwq->wq->name, worker->current_func, |
2606 | target_wq->name, target_func); | |
2607 | } | |
2608 | ||
fc2e4d70 ON |
2609 | struct wq_barrier { |
2610 | struct work_struct work; | |
2611 | struct completion done; | |
2607d7a6 | 2612 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2613 | }; |
2614 | ||
2615 | static void wq_barrier_func(struct work_struct *work) | |
2616 | { | |
2617 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2618 | complete(&barr->done); | |
2619 | } | |
2620 | ||
4690c4ab TH |
2621 | /** |
2622 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2623 | * @pwq: pwq to insert barrier into |
4690c4ab | 2624 | * @barr: wq_barrier to insert |
affee4b2 TH |
2625 | * @target: target work to attach @barr to |
2626 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2627 | * |
affee4b2 TH |
2628 | * @barr is linked to @target such that @barr is completed only after |
2629 | * @target finishes execution. Please note that the ordering | |
2630 | * guarantee is observed only with respect to @target and on the local | |
2631 | * cpu. | |
2632 | * | |
2633 | * Currently, a queued barrier can't be canceled. This is because | |
2634 | * try_to_grab_pending() can't determine whether the work to be | |
2635 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2636 | * flag of the previous work while there must be a valid next work | |
2637 | * after a work with LINKED flag set. | |
2638 | * | |
2639 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2640 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2641 | * |
2642 | * CONTEXT: | |
d565ed63 | 2643 | * spin_lock_irq(pool->lock). |
4690c4ab | 2644 | */ |
112202d9 | 2645 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2646 | struct wq_barrier *barr, |
2647 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2648 | { |
affee4b2 TH |
2649 | struct list_head *head; |
2650 | unsigned int linked = 0; | |
2651 | ||
dc186ad7 | 2652 | /* |
d565ed63 | 2653 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2654 | * as we know for sure that this will not trigger any of the |
2655 | * checks and call back into the fixup functions where we | |
2656 | * might deadlock. | |
2657 | */ | |
ca1cab37 | 2658 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2659 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 2660 | |
fd1a5b04 BP |
2661 | init_completion_map(&barr->done, &target->lockdep_map); |
2662 | ||
2607d7a6 | 2663 | barr->task = current; |
83c22520 | 2664 | |
affee4b2 TH |
2665 | /* |
2666 | * If @target is currently being executed, schedule the | |
2667 | * barrier to the worker; otherwise, put it after @target. | |
2668 | */ | |
2669 | if (worker) | |
2670 | head = worker->scheduled.next; | |
2671 | else { | |
2672 | unsigned long *bits = work_data_bits(target); | |
2673 | ||
2674 | head = target->entry.next; | |
2675 | /* there can already be other linked works, inherit and set */ | |
2676 | linked = *bits & WORK_STRUCT_LINKED; | |
2677 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2678 | } | |
2679 | ||
dc186ad7 | 2680 | debug_work_activate(&barr->work); |
112202d9 | 2681 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2682 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2683 | } |
2684 | ||
73f53c4a | 2685 | /** |
112202d9 | 2686 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2687 | * @wq: workqueue being flushed |
2688 | * @flush_color: new flush color, < 0 for no-op | |
2689 | * @work_color: new work color, < 0 for no-op | |
2690 | * | |
112202d9 | 2691 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2692 | * |
112202d9 TH |
2693 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2694 | * -1. If no pwq has in-flight commands at the specified color, all | |
2695 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2696 | * has in flight commands, its pwq->flush_color is set to | |
2697 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2698 | * wakeup logic is armed and %true is returned. |
2699 | * | |
2700 | * The caller should have initialized @wq->first_flusher prior to | |
2701 | * calling this function with non-negative @flush_color. If | |
2702 | * @flush_color is negative, no flush color update is done and %false | |
2703 | * is returned. | |
2704 | * | |
112202d9 | 2705 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2706 | * work_color which is previous to @work_color and all will be |
2707 | * advanced to @work_color. | |
2708 | * | |
2709 | * CONTEXT: | |
3c25a55d | 2710 | * mutex_lock(wq->mutex). |
73f53c4a | 2711 | * |
d185af30 | 2712 | * Return: |
73f53c4a TH |
2713 | * %true if @flush_color >= 0 and there's something to flush. %false |
2714 | * otherwise. | |
2715 | */ | |
112202d9 | 2716 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2717 | int flush_color, int work_color) |
1da177e4 | 2718 | { |
73f53c4a | 2719 | bool wait = false; |
49e3cf44 | 2720 | struct pool_workqueue *pwq; |
1da177e4 | 2721 | |
73f53c4a | 2722 | if (flush_color >= 0) { |
6183c009 | 2723 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2724 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2725 | } |
2355b70f | 2726 | |
49e3cf44 | 2727 | for_each_pwq(pwq, wq) { |
112202d9 | 2728 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2729 | |
b09f4fd3 | 2730 | spin_lock_irq(&pool->lock); |
83c22520 | 2731 | |
73f53c4a | 2732 | if (flush_color >= 0) { |
6183c009 | 2733 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2734 | |
112202d9 TH |
2735 | if (pwq->nr_in_flight[flush_color]) { |
2736 | pwq->flush_color = flush_color; | |
2737 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2738 | wait = true; |
2739 | } | |
2740 | } | |
1da177e4 | 2741 | |
73f53c4a | 2742 | if (work_color >= 0) { |
6183c009 | 2743 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2744 | pwq->work_color = work_color; |
73f53c4a | 2745 | } |
1da177e4 | 2746 | |
b09f4fd3 | 2747 | spin_unlock_irq(&pool->lock); |
1da177e4 | 2748 | } |
2355b70f | 2749 | |
112202d9 | 2750 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2751 | complete(&wq->first_flusher->done); |
14441960 | 2752 | |
73f53c4a | 2753 | return wait; |
1da177e4 LT |
2754 | } |
2755 | ||
0fcb78c2 | 2756 | /** |
1da177e4 | 2757 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2758 | * @wq: workqueue to flush |
1da177e4 | 2759 | * |
c5aa87bb TH |
2760 | * This function sleeps until all work items which were queued on entry |
2761 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2762 | */ |
7ad5b3a5 | 2763 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2764 | { |
73f53c4a TH |
2765 | struct wq_flusher this_flusher = { |
2766 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2767 | .flush_color = -1, | |
fd1a5b04 | 2768 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
2769 | }; |
2770 | int next_color; | |
1da177e4 | 2771 | |
3347fa09 TH |
2772 | if (WARN_ON(!wq_online)) |
2773 | return; | |
2774 | ||
87915adc JB |
2775 | lock_map_acquire(&wq->lockdep_map); |
2776 | lock_map_release(&wq->lockdep_map); | |
2777 | ||
3c25a55d | 2778 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2779 | |
2780 | /* | |
2781 | * Start-to-wait phase | |
2782 | */ | |
2783 | next_color = work_next_color(wq->work_color); | |
2784 | ||
2785 | if (next_color != wq->flush_color) { | |
2786 | /* | |
2787 | * Color space is not full. The current work_color | |
2788 | * becomes our flush_color and work_color is advanced | |
2789 | * by one. | |
2790 | */ | |
6183c009 | 2791 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2792 | this_flusher.flush_color = wq->work_color; |
2793 | wq->work_color = next_color; | |
2794 | ||
2795 | if (!wq->first_flusher) { | |
2796 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2797 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2798 | |
2799 | wq->first_flusher = &this_flusher; | |
2800 | ||
112202d9 | 2801 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2802 | wq->work_color)) { |
2803 | /* nothing to flush, done */ | |
2804 | wq->flush_color = next_color; | |
2805 | wq->first_flusher = NULL; | |
2806 | goto out_unlock; | |
2807 | } | |
2808 | } else { | |
2809 | /* wait in queue */ | |
6183c009 | 2810 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2811 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2812 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2813 | } |
2814 | } else { | |
2815 | /* | |
2816 | * Oops, color space is full, wait on overflow queue. | |
2817 | * The next flush completion will assign us | |
2818 | * flush_color and transfer to flusher_queue. | |
2819 | */ | |
2820 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2821 | } | |
2822 | ||
fca839c0 TH |
2823 | check_flush_dependency(wq, NULL); |
2824 | ||
3c25a55d | 2825 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2826 | |
2827 | wait_for_completion(&this_flusher.done); | |
2828 | ||
2829 | /* | |
2830 | * Wake-up-and-cascade phase | |
2831 | * | |
2832 | * First flushers are responsible for cascading flushes and | |
2833 | * handling overflow. Non-first flushers can simply return. | |
2834 | */ | |
2835 | if (wq->first_flusher != &this_flusher) | |
2836 | return; | |
2837 | ||
3c25a55d | 2838 | mutex_lock(&wq->mutex); |
73f53c4a | 2839 | |
4ce48b37 TH |
2840 | /* we might have raced, check again with mutex held */ |
2841 | if (wq->first_flusher != &this_flusher) | |
2842 | goto out_unlock; | |
2843 | ||
73f53c4a TH |
2844 | wq->first_flusher = NULL; |
2845 | ||
6183c009 TH |
2846 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2847 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2848 | |
2849 | while (true) { | |
2850 | struct wq_flusher *next, *tmp; | |
2851 | ||
2852 | /* complete all the flushers sharing the current flush color */ | |
2853 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2854 | if (next->flush_color != wq->flush_color) | |
2855 | break; | |
2856 | list_del_init(&next->list); | |
2857 | complete(&next->done); | |
2858 | } | |
2859 | ||
6183c009 TH |
2860 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2861 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2862 | |
2863 | /* this flush_color is finished, advance by one */ | |
2864 | wq->flush_color = work_next_color(wq->flush_color); | |
2865 | ||
2866 | /* one color has been freed, handle overflow queue */ | |
2867 | if (!list_empty(&wq->flusher_overflow)) { | |
2868 | /* | |
2869 | * Assign the same color to all overflowed | |
2870 | * flushers, advance work_color and append to | |
2871 | * flusher_queue. This is the start-to-wait | |
2872 | * phase for these overflowed flushers. | |
2873 | */ | |
2874 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2875 | tmp->flush_color = wq->work_color; | |
2876 | ||
2877 | wq->work_color = work_next_color(wq->work_color); | |
2878 | ||
2879 | list_splice_tail_init(&wq->flusher_overflow, | |
2880 | &wq->flusher_queue); | |
112202d9 | 2881 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2882 | } |
2883 | ||
2884 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2885 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2886 | break; |
2887 | } | |
2888 | ||
2889 | /* | |
2890 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2891 | * the new first flusher and arm pwqs. |
73f53c4a | 2892 | */ |
6183c009 TH |
2893 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2894 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2895 | |
2896 | list_del_init(&next->list); | |
2897 | wq->first_flusher = next; | |
2898 | ||
112202d9 | 2899 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2900 | break; |
2901 | ||
2902 | /* | |
2903 | * Meh... this color is already done, clear first | |
2904 | * flusher and repeat cascading. | |
2905 | */ | |
2906 | wq->first_flusher = NULL; | |
2907 | } | |
2908 | ||
2909 | out_unlock: | |
3c25a55d | 2910 | mutex_unlock(&wq->mutex); |
1da177e4 | 2911 | } |
1dadafa8 | 2912 | EXPORT_SYMBOL(flush_workqueue); |
1da177e4 | 2913 | |
9c5a2ba7 TH |
2914 | /** |
2915 | * drain_workqueue - drain a workqueue | |
2916 | * @wq: workqueue to drain | |
2917 | * | |
2918 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2919 | * only chain queueing is allowed. IOW, only currently pending or running | |
2920 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 2921 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
2922 | * by the depth of chaining and should be relatively short. Whine if it |
2923 | * takes too long. | |
2924 | */ | |
2925 | void drain_workqueue(struct workqueue_struct *wq) | |
2926 | { | |
2927 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2928 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2929 | |
2930 | /* | |
2931 | * __queue_work() needs to test whether there are drainers, is much | |
2932 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 2933 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 2934 | */ |
87fc741e | 2935 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 2936 | if (!wq->nr_drainers++) |
618b01eb | 2937 | wq->flags |= __WQ_DRAINING; |
87fc741e | 2938 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2939 | reflush: |
2940 | flush_workqueue(wq); | |
2941 | ||
b09f4fd3 | 2942 | mutex_lock(&wq->mutex); |
76af4d93 | 2943 | |
49e3cf44 | 2944 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2945 | bool drained; |
9c5a2ba7 | 2946 | |
b09f4fd3 | 2947 | spin_lock_irq(&pwq->pool->lock); |
112202d9 | 2948 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
b09f4fd3 | 2949 | spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
2950 | |
2951 | if (drained) | |
9c5a2ba7 TH |
2952 | continue; |
2953 | ||
2954 | if (++flush_cnt == 10 || | |
2955 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
c5aa87bb | 2956 | pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", |
044c782c | 2957 | wq->name, flush_cnt); |
76af4d93 | 2958 | |
b09f4fd3 | 2959 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2960 | goto reflush; |
2961 | } | |
2962 | ||
9c5a2ba7 | 2963 | if (!--wq->nr_drainers) |
618b01eb | 2964 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 2965 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2966 | } |
2967 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2968 | ||
d6e89786 JB |
2969 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
2970 | bool from_cancel) | |
db700897 | 2971 | { |
affee4b2 | 2972 | struct worker *worker = NULL; |
c9e7cf27 | 2973 | struct worker_pool *pool; |
112202d9 | 2974 | struct pool_workqueue *pwq; |
db700897 ON |
2975 | |
2976 | might_sleep(); | |
fa1b54e6 | 2977 | |
24acfb71 | 2978 | rcu_read_lock(); |
c9e7cf27 | 2979 | pool = get_work_pool(work); |
fa1b54e6 | 2980 | if (!pool) { |
24acfb71 | 2981 | rcu_read_unlock(); |
baf59022 | 2982 | return false; |
fa1b54e6 | 2983 | } |
db700897 | 2984 | |
24acfb71 | 2985 | spin_lock_irq(&pool->lock); |
0b3dae68 | 2986 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2987 | pwq = get_work_pwq(work); |
2988 | if (pwq) { | |
2989 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2990 | goto already_gone; |
606a5020 | 2991 | } else { |
c9e7cf27 | 2992 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2993 | if (!worker) |
4690c4ab | 2994 | goto already_gone; |
112202d9 | 2995 | pwq = worker->current_pwq; |
606a5020 | 2996 | } |
db700897 | 2997 | |
fca839c0 TH |
2998 | check_flush_dependency(pwq->wq, work); |
2999 | ||
112202d9 | 3000 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 3001 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 3002 | |
e159489b | 3003 | /* |
a1d14934 PZ |
3004 | * Force a lock recursion deadlock when using flush_work() inside a |
3005 | * single-threaded or rescuer equipped workqueue. | |
3006 | * | |
3007 | * For single threaded workqueues the deadlock happens when the work | |
3008 | * is after the work issuing the flush_work(). For rescuer equipped | |
3009 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
3010 | * forward progress. | |
e159489b | 3011 | */ |
d6e89786 JB |
3012 | if (!from_cancel && |
3013 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { | |
112202d9 | 3014 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
3015 | lock_map_release(&pwq->wq->lockdep_map); |
3016 | } | |
24acfb71 | 3017 | rcu_read_unlock(); |
401a8d04 | 3018 | return true; |
4690c4ab | 3019 | already_gone: |
d565ed63 | 3020 | spin_unlock_irq(&pool->lock); |
24acfb71 | 3021 | rcu_read_unlock(); |
401a8d04 | 3022 | return false; |
db700897 | 3023 | } |
baf59022 | 3024 | |
d6e89786 JB |
3025 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
3026 | { | |
3027 | struct wq_barrier barr; | |
3028 | ||
3029 | if (WARN_ON(!wq_online)) | |
3030 | return false; | |
3031 | ||
4d43d395 TH |
3032 | if (WARN_ON(!work->func)) |
3033 | return false; | |
3034 | ||
87915adc JB |
3035 | if (!from_cancel) { |
3036 | lock_map_acquire(&work->lockdep_map); | |
3037 | lock_map_release(&work->lockdep_map); | |
3038 | } | |
3039 | ||
d6e89786 JB |
3040 | if (start_flush_work(work, &barr, from_cancel)) { |
3041 | wait_for_completion(&barr.done); | |
3042 | destroy_work_on_stack(&barr.work); | |
3043 | return true; | |
3044 | } else { | |
3045 | return false; | |
3046 | } | |
3047 | } | |
3048 | ||
baf59022 TH |
3049 | /** |
3050 | * flush_work - wait for a work to finish executing the last queueing instance | |
3051 | * @work: the work to flush | |
3052 | * | |
606a5020 TH |
3053 | * Wait until @work has finished execution. @work is guaranteed to be idle |
3054 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 3055 | * |
d185af30 | 3056 | * Return: |
baf59022 TH |
3057 | * %true if flush_work() waited for the work to finish execution, |
3058 | * %false if it was already idle. | |
3059 | */ | |
3060 | bool flush_work(struct work_struct *work) | |
3061 | { | |
d6e89786 | 3062 | return __flush_work(work, false); |
6e84d644 | 3063 | } |
606a5020 | 3064 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 3065 | |
8603e1b3 | 3066 | struct cwt_wait { |
ac6424b9 | 3067 | wait_queue_entry_t wait; |
8603e1b3 TH |
3068 | struct work_struct *work; |
3069 | }; | |
3070 | ||
ac6424b9 | 3071 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
3072 | { |
3073 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
3074 | ||
3075 | if (cwait->work != key) | |
3076 | return 0; | |
3077 | return autoremove_wake_function(wait, mode, sync, key); | |
3078 | } | |
3079 | ||
36e227d2 | 3080 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 3081 | { |
8603e1b3 | 3082 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 3083 | unsigned long flags; |
1f1f642e ON |
3084 | int ret; |
3085 | ||
3086 | do { | |
bbb68dfa TH |
3087 | ret = try_to_grab_pending(work, is_dwork, &flags); |
3088 | /* | |
8603e1b3 TH |
3089 | * If someone else is already canceling, wait for it to |
3090 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
3091 | * because we may get scheduled between @work's completion | |
3092 | * and the other canceling task resuming and clearing | |
3093 | * CANCELING - flush_work() will return false immediately | |
3094 | * as @work is no longer busy, try_to_grab_pending() will | |
3095 | * return -ENOENT as @work is still being canceled and the | |
3096 | * other canceling task won't be able to clear CANCELING as | |
3097 | * we're hogging the CPU. | |
3098 | * | |
3099 | * Let's wait for completion using a waitqueue. As this | |
3100 | * may lead to the thundering herd problem, use a custom | |
3101 | * wake function which matches @work along with exclusive | |
3102 | * wait and wakeup. | |
bbb68dfa | 3103 | */ |
8603e1b3 TH |
3104 | if (unlikely(ret == -ENOENT)) { |
3105 | struct cwt_wait cwait; | |
3106 | ||
3107 | init_wait(&cwait.wait); | |
3108 | cwait.wait.func = cwt_wakefn; | |
3109 | cwait.work = work; | |
3110 | ||
3111 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
3112 | TASK_UNINTERRUPTIBLE); | |
3113 | if (work_is_canceling(work)) | |
3114 | schedule(); | |
3115 | finish_wait(&cancel_waitq, &cwait.wait); | |
3116 | } | |
1f1f642e ON |
3117 | } while (unlikely(ret < 0)); |
3118 | ||
bbb68dfa TH |
3119 | /* tell other tasks trying to grab @work to back off */ |
3120 | mark_work_canceling(work); | |
3121 | local_irq_restore(flags); | |
3122 | ||
3347fa09 TH |
3123 | /* |
3124 | * This allows canceling during early boot. We know that @work | |
3125 | * isn't executing. | |
3126 | */ | |
3127 | if (wq_online) | |
d6e89786 | 3128 | __flush_work(work, true); |
3347fa09 | 3129 | |
7a22ad75 | 3130 | clear_work_data(work); |
8603e1b3 TH |
3131 | |
3132 | /* | |
3133 | * Paired with prepare_to_wait() above so that either | |
3134 | * waitqueue_active() is visible here or !work_is_canceling() is | |
3135 | * visible there. | |
3136 | */ | |
3137 | smp_mb(); | |
3138 | if (waitqueue_active(&cancel_waitq)) | |
3139 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
3140 | ||
1f1f642e ON |
3141 | return ret; |
3142 | } | |
3143 | ||
6e84d644 | 3144 | /** |
401a8d04 TH |
3145 | * cancel_work_sync - cancel a work and wait for it to finish |
3146 | * @work: the work to cancel | |
6e84d644 | 3147 | * |
401a8d04 TH |
3148 | * Cancel @work and wait for its execution to finish. This function |
3149 | * can be used even if the work re-queues itself or migrates to | |
3150 | * another workqueue. On return from this function, @work is | |
3151 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3152 | * |
401a8d04 TH |
3153 | * cancel_work_sync(&delayed_work->work) must not be used for |
3154 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3155 | * |
401a8d04 | 3156 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3157 | * queued can't be destroyed before this function returns. |
401a8d04 | 3158 | * |
d185af30 | 3159 | * Return: |
401a8d04 | 3160 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3161 | */ |
401a8d04 | 3162 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3163 | { |
36e227d2 | 3164 | return __cancel_work_timer(work, false); |
b89deed3 | 3165 | } |
28e53bdd | 3166 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3167 | |
6e84d644 | 3168 | /** |
401a8d04 TH |
3169 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3170 | * @dwork: the delayed work to flush | |
6e84d644 | 3171 | * |
401a8d04 TH |
3172 | * Delayed timer is cancelled and the pending work is queued for |
3173 | * immediate execution. Like flush_work(), this function only | |
3174 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3175 | * |
d185af30 | 3176 | * Return: |
401a8d04 TH |
3177 | * %true if flush_work() waited for the work to finish execution, |
3178 | * %false if it was already idle. | |
6e84d644 | 3179 | */ |
401a8d04 TH |
3180 | bool flush_delayed_work(struct delayed_work *dwork) |
3181 | { | |
8930caba | 3182 | local_irq_disable(); |
401a8d04 | 3183 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3184 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3185 | local_irq_enable(); |
401a8d04 TH |
3186 | return flush_work(&dwork->work); |
3187 | } | |
3188 | EXPORT_SYMBOL(flush_delayed_work); | |
3189 | ||
05f0fe6b TH |
3190 | /** |
3191 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
3192 | * @rwork: the rcu work to flush | |
3193 | * | |
3194 | * Return: | |
3195 | * %true if flush_rcu_work() waited for the work to finish execution, | |
3196 | * %false if it was already idle. | |
3197 | */ | |
3198 | bool flush_rcu_work(struct rcu_work *rwork) | |
3199 | { | |
3200 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
3201 | rcu_barrier(); | |
3202 | flush_work(&rwork->work); | |
3203 | return true; | |
3204 | } else { | |
3205 | return flush_work(&rwork->work); | |
3206 | } | |
3207 | } | |
3208 | EXPORT_SYMBOL(flush_rcu_work); | |
3209 | ||
f72b8792 JA |
3210 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
3211 | { | |
3212 | unsigned long flags; | |
3213 | int ret; | |
3214 | ||
3215 | do { | |
3216 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
3217 | } while (unlikely(ret == -EAGAIN)); | |
3218 | ||
3219 | if (unlikely(ret < 0)) | |
3220 | return false; | |
3221 | ||
3222 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3223 | local_irq_restore(flags); | |
3224 | return ret; | |
3225 | } | |
3226 | ||
09383498 | 3227 | /** |
57b30ae7 TH |
3228 | * cancel_delayed_work - cancel a delayed work |
3229 | * @dwork: delayed_work to cancel | |
09383498 | 3230 | * |
d185af30 YB |
3231 | * Kill off a pending delayed_work. |
3232 | * | |
3233 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3234 | * pending. | |
3235 | * | |
3236 | * Note: | |
3237 | * The work callback function may still be running on return, unless | |
3238 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3239 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3240 | * |
57b30ae7 | 3241 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3242 | */ |
57b30ae7 | 3243 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3244 | { |
f72b8792 | 3245 | return __cancel_work(&dwork->work, true); |
09383498 | 3246 | } |
57b30ae7 | 3247 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3248 | |
401a8d04 TH |
3249 | /** |
3250 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3251 | * @dwork: the delayed work cancel | |
3252 | * | |
3253 | * This is cancel_work_sync() for delayed works. | |
3254 | * | |
d185af30 | 3255 | * Return: |
401a8d04 TH |
3256 | * %true if @dwork was pending, %false otherwise. |
3257 | */ | |
3258 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3259 | { |
36e227d2 | 3260 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3261 | } |
f5a421a4 | 3262 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3263 | |
b6136773 | 3264 | /** |
31ddd871 | 3265 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3266 | * @func: the function to call |
b6136773 | 3267 | * |
31ddd871 TH |
3268 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3269 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3270 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3271 | * |
d185af30 | 3272 | * Return: |
31ddd871 | 3273 | * 0 on success, -errno on failure. |
b6136773 | 3274 | */ |
65f27f38 | 3275 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3276 | { |
3277 | int cpu; | |
38f51568 | 3278 | struct work_struct __percpu *works; |
15316ba8 | 3279 | |
b6136773 AM |
3280 | works = alloc_percpu(struct work_struct); |
3281 | if (!works) | |
15316ba8 | 3282 | return -ENOMEM; |
b6136773 | 3283 | |
93981800 TH |
3284 | get_online_cpus(); |
3285 | ||
15316ba8 | 3286 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3287 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3288 | ||
3289 | INIT_WORK(work, func); | |
b71ab8c2 | 3290 | schedule_work_on(cpu, work); |
65a64464 | 3291 | } |
93981800 TH |
3292 | |
3293 | for_each_online_cpu(cpu) | |
3294 | flush_work(per_cpu_ptr(works, cpu)); | |
3295 | ||
95402b38 | 3296 | put_online_cpus(); |
b6136773 | 3297 | free_percpu(works); |
15316ba8 CL |
3298 | return 0; |
3299 | } | |
3300 | ||
1fa44eca JB |
3301 | /** |
3302 | * execute_in_process_context - reliably execute the routine with user context | |
3303 | * @fn: the function to execute | |
1fa44eca JB |
3304 | * @ew: guaranteed storage for the execute work structure (must |
3305 | * be available when the work executes) | |
3306 | * | |
3307 | * Executes the function immediately if process context is available, | |
3308 | * otherwise schedules the function for delayed execution. | |
3309 | * | |
d185af30 | 3310 | * Return: 0 - function was executed |
1fa44eca JB |
3311 | * 1 - function was scheduled for execution |
3312 | */ | |
65f27f38 | 3313 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3314 | { |
3315 | if (!in_interrupt()) { | |
65f27f38 | 3316 | fn(&ew->work); |
1fa44eca JB |
3317 | return 0; |
3318 | } | |
3319 | ||
65f27f38 | 3320 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3321 | schedule_work(&ew->work); |
3322 | ||
3323 | return 1; | |
3324 | } | |
3325 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3326 | ||
6ba94429 FW |
3327 | /** |
3328 | * free_workqueue_attrs - free a workqueue_attrs | |
3329 | * @attrs: workqueue_attrs to free | |
226223ab | 3330 | * |
6ba94429 | 3331 | * Undo alloc_workqueue_attrs(). |
226223ab | 3332 | */ |
513c98d0 | 3333 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3334 | { |
6ba94429 FW |
3335 | if (attrs) { |
3336 | free_cpumask_var(attrs->cpumask); | |
3337 | kfree(attrs); | |
3338 | } | |
226223ab TH |
3339 | } |
3340 | ||
6ba94429 FW |
3341 | /** |
3342 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
3343 | * |
3344 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3345 | * return it. | |
3346 | * | |
3347 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3348 | */ | |
513c98d0 | 3349 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 3350 | { |
6ba94429 | 3351 | struct workqueue_attrs *attrs; |
226223ab | 3352 | |
be69d00d | 3353 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
3354 | if (!attrs) |
3355 | goto fail; | |
be69d00d | 3356 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 FW |
3357 | goto fail; |
3358 | ||
3359 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
3360 | return attrs; | |
3361 | fail: | |
3362 | free_workqueue_attrs(attrs); | |
3363 | return NULL; | |
226223ab TH |
3364 | } |
3365 | ||
6ba94429 FW |
3366 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3367 | const struct workqueue_attrs *from) | |
226223ab | 3368 | { |
6ba94429 FW |
3369 | to->nice = from->nice; |
3370 | cpumask_copy(to->cpumask, from->cpumask); | |
3371 | /* | |
3372 | * Unlike hash and equality test, this function doesn't ignore | |
3373 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3374 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3375 | */ | |
3376 | to->no_numa = from->no_numa; | |
226223ab TH |
3377 | } |
3378 | ||
6ba94429 FW |
3379 | /* hash value of the content of @attr */ |
3380 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3381 | { |
6ba94429 | 3382 | u32 hash = 0; |
226223ab | 3383 | |
6ba94429 FW |
3384 | hash = jhash_1word(attrs->nice, hash); |
3385 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3386 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
3387 | return hash; | |
226223ab | 3388 | } |
226223ab | 3389 | |
6ba94429 FW |
3390 | /* content equality test */ |
3391 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3392 | const struct workqueue_attrs *b) | |
226223ab | 3393 | { |
6ba94429 FW |
3394 | if (a->nice != b->nice) |
3395 | return false; | |
3396 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3397 | return false; | |
3398 | return true; | |
226223ab TH |
3399 | } |
3400 | ||
6ba94429 FW |
3401 | /** |
3402 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3403 | * @pool: worker_pool to initialize | |
3404 | * | |
402dd89d | 3405 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3406 | * |
3407 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3408 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3409 | * on @pool safely to release it. | |
3410 | */ | |
3411 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3412 | { |
6ba94429 FW |
3413 | spin_lock_init(&pool->lock); |
3414 | pool->id = -1; | |
3415 | pool->cpu = -1; | |
3416 | pool->node = NUMA_NO_NODE; | |
3417 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3418 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3419 | INIT_LIST_HEAD(&pool->worklist); |
3420 | INIT_LIST_HEAD(&pool->idle_list); | |
3421 | hash_init(pool->busy_hash); | |
226223ab | 3422 | |
32a6c723 | 3423 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
226223ab | 3424 | |
32a6c723 | 3425 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 3426 | |
6ba94429 | 3427 | INIT_LIST_HEAD(&pool->workers); |
226223ab | 3428 | |
6ba94429 FW |
3429 | ida_init(&pool->worker_ida); |
3430 | INIT_HLIST_NODE(&pool->hash_node); | |
3431 | pool->refcnt = 1; | |
226223ab | 3432 | |
6ba94429 | 3433 | /* shouldn't fail above this point */ |
be69d00d | 3434 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
3435 | if (!pool->attrs) |
3436 | return -ENOMEM; | |
3437 | return 0; | |
226223ab TH |
3438 | } |
3439 | ||
669de8bd BVA |
3440 | #ifdef CONFIG_LOCKDEP |
3441 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3442 | { | |
3443 | char *lock_name; | |
3444 | ||
3445 | lockdep_register_key(&wq->key); | |
3446 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
3447 | if (!lock_name) | |
3448 | lock_name = wq->name; | |
69a106c0 QC |
3449 | |
3450 | wq->lock_name = lock_name; | |
669de8bd BVA |
3451 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
3452 | } | |
3453 | ||
3454 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3455 | { | |
3456 | lockdep_unregister_key(&wq->key); | |
3457 | } | |
3458 | ||
3459 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3460 | { | |
3461 | if (wq->lock_name != wq->name) | |
3462 | kfree(wq->lock_name); | |
3463 | } | |
3464 | #else | |
3465 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3466 | { | |
3467 | } | |
3468 | ||
3469 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3470 | { | |
3471 | } | |
3472 | ||
3473 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3474 | { | |
3475 | } | |
3476 | #endif | |
3477 | ||
6ba94429 | 3478 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3479 | { |
6ba94429 FW |
3480 | struct workqueue_struct *wq = |
3481 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3482 | |
669de8bd BVA |
3483 | wq_free_lockdep(wq); |
3484 | ||
6ba94429 FW |
3485 | if (!(wq->flags & WQ_UNBOUND)) |
3486 | free_percpu(wq->cpu_pwqs); | |
226223ab | 3487 | else |
6ba94429 | 3488 | free_workqueue_attrs(wq->unbound_attrs); |
226223ab | 3489 | |
6ba94429 FW |
3490 | kfree(wq->rescuer); |
3491 | kfree(wq); | |
226223ab TH |
3492 | } |
3493 | ||
6ba94429 | 3494 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3495 | { |
6ba94429 | 3496 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3497 | |
6ba94429 FW |
3498 | ida_destroy(&pool->worker_ida); |
3499 | free_workqueue_attrs(pool->attrs); | |
3500 | kfree(pool); | |
226223ab TH |
3501 | } |
3502 | ||
6ba94429 FW |
3503 | /** |
3504 | * put_unbound_pool - put a worker_pool | |
3505 | * @pool: worker_pool to put | |
3506 | * | |
24acfb71 | 3507 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
3508 | * safe manner. get_unbound_pool() calls this function on its failure path |
3509 | * and this function should be able to release pools which went through, | |
3510 | * successfully or not, init_worker_pool(). | |
3511 | * | |
3512 | * Should be called with wq_pool_mutex held. | |
3513 | */ | |
3514 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3515 | { |
6ba94429 FW |
3516 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
3517 | struct worker *worker; | |
226223ab | 3518 | |
6ba94429 | 3519 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3520 | |
6ba94429 FW |
3521 | if (--pool->refcnt) |
3522 | return; | |
226223ab | 3523 | |
6ba94429 FW |
3524 | /* sanity checks */ |
3525 | if (WARN_ON(!(pool->cpu < 0)) || | |
3526 | WARN_ON(!list_empty(&pool->worklist))) | |
3527 | return; | |
226223ab | 3528 | |
6ba94429 FW |
3529 | /* release id and unhash */ |
3530 | if (pool->id >= 0) | |
3531 | idr_remove(&worker_pool_idr, pool->id); | |
3532 | hash_del(&pool->hash_node); | |
d55262c4 | 3533 | |
6ba94429 | 3534 | /* |
692b4825 TH |
3535 | * Become the manager and destroy all workers. This prevents |
3536 | * @pool's workers from blocking on attach_mutex. We're the last | |
3537 | * manager and @pool gets freed with the flag set. | |
6ba94429 | 3538 | */ |
6ba94429 | 3539 | spin_lock_irq(&pool->lock); |
692b4825 TH |
3540 | wait_event_lock_irq(wq_manager_wait, |
3541 | !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); | |
3542 | pool->flags |= POOL_MANAGER_ACTIVE; | |
3543 | ||
6ba94429 FW |
3544 | while ((worker = first_idle_worker(pool))) |
3545 | destroy_worker(worker); | |
3546 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
3547 | spin_unlock_irq(&pool->lock); | |
d55262c4 | 3548 | |
1258fae7 | 3549 | mutex_lock(&wq_pool_attach_mutex); |
6ba94429 FW |
3550 | if (!list_empty(&pool->workers)) |
3551 | pool->detach_completion = &detach_completion; | |
1258fae7 | 3552 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 3553 | |
6ba94429 FW |
3554 | if (pool->detach_completion) |
3555 | wait_for_completion(pool->detach_completion); | |
226223ab | 3556 | |
6ba94429 FW |
3557 | /* shut down the timers */ |
3558 | del_timer_sync(&pool->idle_timer); | |
3559 | del_timer_sync(&pool->mayday_timer); | |
226223ab | 3560 | |
24acfb71 | 3561 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 3562 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
3563 | } |
3564 | ||
3565 | /** | |
6ba94429 FW |
3566 | * get_unbound_pool - get a worker_pool with the specified attributes |
3567 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 3568 | * |
6ba94429 FW |
3569 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
3570 | * reference count and return it. If there already is a matching | |
3571 | * worker_pool, it will be used; otherwise, this function attempts to | |
3572 | * create a new one. | |
226223ab | 3573 | * |
6ba94429 | 3574 | * Should be called with wq_pool_mutex held. |
226223ab | 3575 | * |
6ba94429 FW |
3576 | * Return: On success, a worker_pool with the same attributes as @attrs. |
3577 | * On failure, %NULL. | |
226223ab | 3578 | */ |
6ba94429 | 3579 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 3580 | { |
6ba94429 FW |
3581 | u32 hash = wqattrs_hash(attrs); |
3582 | struct worker_pool *pool; | |
3583 | int node; | |
e2273584 | 3584 | int target_node = NUMA_NO_NODE; |
226223ab | 3585 | |
6ba94429 | 3586 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3587 | |
6ba94429 FW |
3588 | /* do we already have a matching pool? */ |
3589 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3590 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3591 | pool->refcnt++; | |
3592 | return pool; | |
3593 | } | |
3594 | } | |
226223ab | 3595 | |
e2273584 XP |
3596 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3597 | if (wq_numa_enabled) { | |
3598 | for_each_node(node) { | |
3599 | if (cpumask_subset(attrs->cpumask, | |
3600 | wq_numa_possible_cpumask[node])) { | |
3601 | target_node = node; | |
3602 | break; | |
3603 | } | |
3604 | } | |
3605 | } | |
3606 | ||
6ba94429 | 3607 | /* nope, create a new one */ |
e2273584 | 3608 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
6ba94429 FW |
3609 | if (!pool || init_worker_pool(pool) < 0) |
3610 | goto fail; | |
3611 | ||
3612 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ | |
3613 | copy_workqueue_attrs(pool->attrs, attrs); | |
e2273584 | 3614 | pool->node = target_node; |
226223ab TH |
3615 | |
3616 | /* | |
6ba94429 FW |
3617 | * no_numa isn't a worker_pool attribute, always clear it. See |
3618 | * 'struct workqueue_attrs' comments for detail. | |
226223ab | 3619 | */ |
6ba94429 | 3620 | pool->attrs->no_numa = false; |
226223ab | 3621 | |
6ba94429 FW |
3622 | if (worker_pool_assign_id(pool) < 0) |
3623 | goto fail; | |
226223ab | 3624 | |
6ba94429 | 3625 | /* create and start the initial worker */ |
3347fa09 | 3626 | if (wq_online && !create_worker(pool)) |
6ba94429 | 3627 | goto fail; |
226223ab | 3628 | |
6ba94429 FW |
3629 | /* install */ |
3630 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 3631 | |
6ba94429 FW |
3632 | return pool; |
3633 | fail: | |
3634 | if (pool) | |
3635 | put_unbound_pool(pool); | |
3636 | return NULL; | |
226223ab | 3637 | } |
226223ab | 3638 | |
6ba94429 | 3639 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 3640 | { |
6ba94429 FW |
3641 | kmem_cache_free(pwq_cache, |
3642 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
3643 | } |
3644 | ||
6ba94429 FW |
3645 | /* |
3646 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3647 | * and needs to be destroyed. | |
7a4e344c | 3648 | */ |
6ba94429 | 3649 | static void pwq_unbound_release_workfn(struct work_struct *work) |
7a4e344c | 3650 | { |
6ba94429 FW |
3651 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
3652 | unbound_release_work); | |
3653 | struct workqueue_struct *wq = pwq->wq; | |
3654 | struct worker_pool *pool = pwq->pool; | |
3655 | bool is_last; | |
7a4e344c | 3656 | |
6ba94429 FW |
3657 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) |
3658 | return; | |
7a4e344c | 3659 | |
6ba94429 FW |
3660 | mutex_lock(&wq->mutex); |
3661 | list_del_rcu(&pwq->pwqs_node); | |
3662 | is_last = list_empty(&wq->pwqs); | |
3663 | mutex_unlock(&wq->mutex); | |
3664 | ||
3665 | mutex_lock(&wq_pool_mutex); | |
3666 | put_unbound_pool(pool); | |
3667 | mutex_unlock(&wq_pool_mutex); | |
3668 | ||
25b00775 | 3669 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 3670 | |
2865a8fb | 3671 | /* |
6ba94429 FW |
3672 | * If we're the last pwq going away, @wq is already dead and no one |
3673 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 3674 | */ |
669de8bd BVA |
3675 | if (is_last) { |
3676 | wq_unregister_lockdep(wq); | |
25b00775 | 3677 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 3678 | } |
29c91e99 TH |
3679 | } |
3680 | ||
7a4e344c | 3681 | /** |
6ba94429 FW |
3682 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
3683 | * @pwq: target pool_workqueue | |
d185af30 | 3684 | * |
6ba94429 FW |
3685 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
3686 | * workqueue's saved_max_active and activate delayed work items | |
3687 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. | |
7a4e344c | 3688 | */ |
6ba94429 | 3689 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 3690 | { |
6ba94429 FW |
3691 | struct workqueue_struct *wq = pwq->wq; |
3692 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3347fa09 | 3693 | unsigned long flags; |
4e1a1f9a | 3694 | |
6ba94429 FW |
3695 | /* for @wq->saved_max_active */ |
3696 | lockdep_assert_held(&wq->mutex); | |
4e1a1f9a | 3697 | |
6ba94429 FW |
3698 | /* fast exit for non-freezable wqs */ |
3699 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
3700 | return; | |
7a4e344c | 3701 | |
3347fa09 TH |
3702 | /* this function can be called during early boot w/ irq disabled */ |
3703 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
29c91e99 | 3704 | |
6ba94429 FW |
3705 | /* |
3706 | * During [un]freezing, the caller is responsible for ensuring that | |
3707 | * this function is called at least once after @workqueue_freezing | |
3708 | * is updated and visible. | |
3709 | */ | |
3710 | if (!freezable || !workqueue_freezing) { | |
3711 | pwq->max_active = wq->saved_max_active; | |
4e1a1f9a | 3712 | |
6ba94429 FW |
3713 | while (!list_empty(&pwq->delayed_works) && |
3714 | pwq->nr_active < pwq->max_active) | |
3715 | pwq_activate_first_delayed(pwq); | |
e2dca7ad | 3716 | |
6ba94429 FW |
3717 | /* |
3718 | * Need to kick a worker after thawed or an unbound wq's | |
3719 | * max_active is bumped. It's a slow path. Do it always. | |
3720 | */ | |
3721 | wake_up_worker(pwq->pool); | |
3722 | } else { | |
3723 | pwq->max_active = 0; | |
3724 | } | |
e2dca7ad | 3725 | |
3347fa09 | 3726 | spin_unlock_irqrestore(&pwq->pool->lock, flags); |
e2dca7ad TH |
3727 | } |
3728 | ||
6ba94429 FW |
3729 | /* initialize newly alloced @pwq which is associated with @wq and @pool */ |
3730 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, | |
3731 | struct worker_pool *pool) | |
29c91e99 | 3732 | { |
6ba94429 | 3733 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 3734 | |
6ba94429 FW |
3735 | memset(pwq, 0, sizeof(*pwq)); |
3736 | ||
3737 | pwq->pool = pool; | |
3738 | pwq->wq = wq; | |
3739 | pwq->flush_color = -1; | |
3740 | pwq->refcnt = 1; | |
3741 | INIT_LIST_HEAD(&pwq->delayed_works); | |
3742 | INIT_LIST_HEAD(&pwq->pwqs_node); | |
3743 | INIT_LIST_HEAD(&pwq->mayday_node); | |
3744 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); | |
29c91e99 TH |
3745 | } |
3746 | ||
6ba94429 FW |
3747 | /* sync @pwq with the current state of its associated wq and link it */ |
3748 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 3749 | { |
6ba94429 | 3750 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 3751 | |
6ba94429 | 3752 | lockdep_assert_held(&wq->mutex); |
a892cacc | 3753 | |
6ba94429 FW |
3754 | /* may be called multiple times, ignore if already linked */ |
3755 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 3756 | return; |
29c91e99 | 3757 | |
6ba94429 FW |
3758 | /* set the matching work_color */ |
3759 | pwq->work_color = wq->work_color; | |
29c91e99 | 3760 | |
6ba94429 FW |
3761 | /* sync max_active to the current setting */ |
3762 | pwq_adjust_max_active(pwq); | |
29c91e99 | 3763 | |
6ba94429 FW |
3764 | /* link in @pwq */ |
3765 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
3766 | } | |
29c91e99 | 3767 | |
6ba94429 FW |
3768 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3769 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3770 | const struct workqueue_attrs *attrs) | |
3771 | { | |
3772 | struct worker_pool *pool; | |
3773 | struct pool_workqueue *pwq; | |
60f5a4bc | 3774 | |
6ba94429 | 3775 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 3776 | |
6ba94429 FW |
3777 | pool = get_unbound_pool(attrs); |
3778 | if (!pool) | |
3779 | return NULL; | |
60f5a4bc | 3780 | |
6ba94429 FW |
3781 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
3782 | if (!pwq) { | |
3783 | put_unbound_pool(pool); | |
3784 | return NULL; | |
3785 | } | |
29c91e99 | 3786 | |
6ba94429 FW |
3787 | init_pwq(pwq, wq, pool); |
3788 | return pwq; | |
3789 | } | |
29c91e99 | 3790 | |
29c91e99 | 3791 | /** |
30186c6f | 3792 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
042f7df1 | 3793 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
6ba94429 FW |
3794 | * @node: the target NUMA node |
3795 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3796 | * @cpumask: outarg, the resulting cpumask | |
29c91e99 | 3797 | * |
6ba94429 FW |
3798 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
3799 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
3800 | * calculation. The result is stored in @cpumask. | |
a892cacc | 3801 | * |
6ba94429 FW |
3802 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
3803 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3804 | * cpumask is the intersection of the possible CPUs of @node and | |
3805 | * @attrs->cpumask. | |
d185af30 | 3806 | * |
6ba94429 FW |
3807 | * The caller is responsible for ensuring that the cpumask of @node stays |
3808 | * stable. | |
3809 | * | |
3810 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
3811 | * %false if equal. | |
29c91e99 | 3812 | */ |
6ba94429 FW |
3813 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
3814 | int cpu_going_down, cpumask_t *cpumask) | |
29c91e99 | 3815 | { |
6ba94429 FW |
3816 | if (!wq_numa_enabled || attrs->no_numa) |
3817 | goto use_dfl; | |
29c91e99 | 3818 | |
6ba94429 FW |
3819 | /* does @node have any online CPUs @attrs wants? */ |
3820 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
3821 | if (cpu_going_down >= 0) | |
3822 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
29c91e99 | 3823 | |
6ba94429 FW |
3824 | if (cpumask_empty(cpumask)) |
3825 | goto use_dfl; | |
4c16bd32 TH |
3826 | |
3827 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
3828 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
1ad0f0a7 MB |
3829 | |
3830 | if (cpumask_empty(cpumask)) { | |
3831 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " | |
3832 | "possible intersect\n"); | |
3833 | return false; | |
3834 | } | |
3835 | ||
4c16bd32 TH |
3836 | return !cpumask_equal(cpumask, attrs->cpumask); |
3837 | ||
3838 | use_dfl: | |
3839 | cpumask_copy(cpumask, attrs->cpumask); | |
3840 | return false; | |
3841 | } | |
3842 | ||
1befcf30 TH |
3843 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
3844 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
3845 | int node, | |
3846 | struct pool_workqueue *pwq) | |
3847 | { | |
3848 | struct pool_workqueue *old_pwq; | |
3849 | ||
5b95e1af | 3850 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
3851 | lockdep_assert_held(&wq->mutex); |
3852 | ||
3853 | /* link_pwq() can handle duplicate calls */ | |
3854 | link_pwq(pwq); | |
3855 | ||
3856 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
3857 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
3858 | return old_pwq; | |
3859 | } | |
3860 | ||
2d5f0764 LJ |
3861 | /* context to store the prepared attrs & pwqs before applying */ |
3862 | struct apply_wqattrs_ctx { | |
3863 | struct workqueue_struct *wq; /* target workqueue */ | |
3864 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 3865 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
3866 | struct pool_workqueue *dfl_pwq; |
3867 | struct pool_workqueue *pwq_tbl[]; | |
3868 | }; | |
3869 | ||
3870 | /* free the resources after success or abort */ | |
3871 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
3872 | { | |
3873 | if (ctx) { | |
3874 | int node; | |
3875 | ||
3876 | for_each_node(node) | |
3877 | put_pwq_unlocked(ctx->pwq_tbl[node]); | |
3878 | put_pwq_unlocked(ctx->dfl_pwq); | |
3879 | ||
3880 | free_workqueue_attrs(ctx->attrs); | |
3881 | ||
3882 | kfree(ctx); | |
3883 | } | |
3884 | } | |
3885 | ||
3886 | /* allocate the attrs and pwqs for later installation */ | |
3887 | static struct apply_wqattrs_ctx * | |
3888 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
3889 | const struct workqueue_attrs *attrs) | |
9e8cd2f5 | 3890 | { |
2d5f0764 | 3891 | struct apply_wqattrs_ctx *ctx; |
4c16bd32 | 3892 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
2d5f0764 | 3893 | int node; |
9e8cd2f5 | 3894 | |
2d5f0764 | 3895 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 3896 | |
acafe7e3 | 3897 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); |
8719dcea | 3898 | |
be69d00d TG |
3899 | new_attrs = alloc_workqueue_attrs(); |
3900 | tmp_attrs = alloc_workqueue_attrs(); | |
2d5f0764 LJ |
3901 | if (!ctx || !new_attrs || !tmp_attrs) |
3902 | goto out_free; | |
13e2e556 | 3903 | |
042f7df1 LJ |
3904 | /* |
3905 | * Calculate the attrs of the default pwq. | |
3906 | * If the user configured cpumask doesn't overlap with the | |
3907 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. | |
3908 | */ | |
13e2e556 | 3909 | copy_workqueue_attrs(new_attrs, attrs); |
b05a7928 | 3910 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); |
042f7df1 LJ |
3911 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
3912 | cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); | |
13e2e556 | 3913 | |
4c16bd32 TH |
3914 | /* |
3915 | * We may create multiple pwqs with differing cpumasks. Make a | |
3916 | * copy of @new_attrs which will be modified and used to obtain | |
3917 | * pools. | |
3918 | */ | |
3919 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
3920 | ||
4c16bd32 TH |
3921 | /* |
3922 | * If something goes wrong during CPU up/down, we'll fall back to | |
3923 | * the default pwq covering whole @attrs->cpumask. Always create | |
3924 | * it even if we don't use it immediately. | |
3925 | */ | |
2d5f0764 LJ |
3926 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
3927 | if (!ctx->dfl_pwq) | |
3928 | goto out_free; | |
4c16bd32 TH |
3929 | |
3930 | for_each_node(node) { | |
042f7df1 | 3931 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
2d5f0764 LJ |
3932 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
3933 | if (!ctx->pwq_tbl[node]) | |
3934 | goto out_free; | |
4c16bd32 | 3935 | } else { |
2d5f0764 LJ |
3936 | ctx->dfl_pwq->refcnt++; |
3937 | ctx->pwq_tbl[node] = ctx->dfl_pwq; | |
4c16bd32 TH |
3938 | } |
3939 | } | |
3940 | ||
042f7df1 LJ |
3941 | /* save the user configured attrs and sanitize it. */ |
3942 | copy_workqueue_attrs(new_attrs, attrs); | |
3943 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
2d5f0764 | 3944 | ctx->attrs = new_attrs; |
042f7df1 | 3945 | |
2d5f0764 LJ |
3946 | ctx->wq = wq; |
3947 | free_workqueue_attrs(tmp_attrs); | |
3948 | return ctx; | |
3949 | ||
3950 | out_free: | |
3951 | free_workqueue_attrs(tmp_attrs); | |
3952 | free_workqueue_attrs(new_attrs); | |
3953 | apply_wqattrs_cleanup(ctx); | |
3954 | return NULL; | |
3955 | } | |
3956 | ||
3957 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
3958 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
3959 | { | |
3960 | int node; | |
9e8cd2f5 | 3961 | |
4c16bd32 | 3962 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 3963 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 3964 | |
2d5f0764 | 3965 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
3966 | |
3967 | /* save the previous pwq and install the new one */ | |
f147f29e | 3968 | for_each_node(node) |
2d5f0764 LJ |
3969 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
3970 | ctx->pwq_tbl[node]); | |
4c16bd32 TH |
3971 | |
3972 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
3973 | link_pwq(ctx->dfl_pwq); |
3974 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 3975 | |
2d5f0764 LJ |
3976 | mutex_unlock(&ctx->wq->mutex); |
3977 | } | |
9e8cd2f5 | 3978 | |
a0111cf6 LJ |
3979 | static void apply_wqattrs_lock(void) |
3980 | { | |
3981 | /* CPUs should stay stable across pwq creations and installations */ | |
3982 | get_online_cpus(); | |
3983 | mutex_lock(&wq_pool_mutex); | |
3984 | } | |
3985 | ||
3986 | static void apply_wqattrs_unlock(void) | |
3987 | { | |
3988 | mutex_unlock(&wq_pool_mutex); | |
3989 | put_online_cpus(); | |
3990 | } | |
3991 | ||
3992 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
3993 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
3994 | { |
3995 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 3996 | |
2d5f0764 LJ |
3997 | /* only unbound workqueues can change attributes */ |
3998 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
3999 | return -EINVAL; | |
13e2e556 | 4000 | |
2d5f0764 | 4001 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
4002 | if (!list_empty(&wq->pwqs)) { |
4003 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
4004 | return -EINVAL; | |
4005 | ||
4006 | wq->flags &= ~__WQ_ORDERED; | |
4007 | } | |
2d5f0764 | 4008 | |
2d5f0764 | 4009 | ctx = apply_wqattrs_prepare(wq, attrs); |
6201171e | 4010 | if (!ctx) |
4011 | return -ENOMEM; | |
2d5f0764 LJ |
4012 | |
4013 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 4014 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
4015 | apply_wqattrs_cleanup(ctx); |
4016 | ||
6201171e | 4017 | return 0; |
9e8cd2f5 TH |
4018 | } |
4019 | ||
a0111cf6 LJ |
4020 | /** |
4021 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
4022 | * @wq: the target workqueue | |
4023 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
4024 | * | |
4025 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA | |
4026 | * machines, this function maps a separate pwq to each NUMA node with | |
4027 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
4028 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
4029 | * items finish. Note that a work item which repeatedly requeues itself | |
4030 | * back-to-back will stay on its current pwq. | |
4031 | * | |
4032 | * Performs GFP_KERNEL allocations. | |
4033 | * | |
509b3204 DJ |
4034 | * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus(). |
4035 | * | |
a0111cf6 LJ |
4036 | * Return: 0 on success and -errno on failure. |
4037 | */ | |
513c98d0 | 4038 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
4039 | const struct workqueue_attrs *attrs) |
4040 | { | |
4041 | int ret; | |
4042 | ||
509b3204 DJ |
4043 | lockdep_assert_cpus_held(); |
4044 | ||
4045 | mutex_lock(&wq_pool_mutex); | |
a0111cf6 | 4046 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 4047 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
4048 | |
4049 | return ret; | |
4050 | } | |
4051 | ||
4c16bd32 TH |
4052 | /** |
4053 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
4054 | * @wq: the target workqueue | |
4055 | * @cpu: the CPU coming up or going down | |
4056 | * @online: whether @cpu is coming up or going down | |
4057 | * | |
4058 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
4059 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
4060 | * @wq accordingly. | |
4061 | * | |
4062 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
4063 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
4064 | * correct. | |
4065 | * | |
4066 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
4067 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
4068 | * already executing the work items for the workqueue will lose their CPU | |
4069 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
4070 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
4071 | * affinity, it's the user's responsibility to flush the work item from | |
4072 | * CPU_DOWN_PREPARE. | |
4073 | */ | |
4074 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
4075 | bool online) | |
4076 | { | |
4077 | int node = cpu_to_node(cpu); | |
4078 | int cpu_off = online ? -1 : cpu; | |
4079 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
4080 | struct workqueue_attrs *target_attrs; | |
4081 | cpumask_t *cpumask; | |
4082 | ||
4083 | lockdep_assert_held(&wq_pool_mutex); | |
4084 | ||
f7142ed4 LJ |
4085 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
4086 | wq->unbound_attrs->no_numa) | |
4c16bd32 TH |
4087 | return; |
4088 | ||
4089 | /* | |
4090 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4091 | * Let's use a preallocated one. The following buf is protected by | |
4092 | * CPU hotplug exclusion. | |
4093 | */ | |
4094 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
4095 | cpumask = target_attrs->cpumask; | |
4096 | ||
4c16bd32 TH |
4097 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
4098 | pwq = unbound_pwq_by_node(wq, node); | |
4099 | ||
4100 | /* | |
4101 | * Let's determine what needs to be done. If the target cpumask is | |
042f7df1 LJ |
4102 | * different from the default pwq's, we need to compare it to @pwq's |
4103 | * and create a new one if they don't match. If the target cpumask | |
4104 | * equals the default pwq's, the default pwq should be used. | |
4c16bd32 | 4105 | */ |
042f7df1 | 4106 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
4c16bd32 | 4107 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
f7142ed4 | 4108 | return; |
4c16bd32 | 4109 | } else { |
534a3fbb | 4110 | goto use_dfl_pwq; |
4c16bd32 TH |
4111 | } |
4112 | ||
4c16bd32 TH |
4113 | /* create a new pwq */ |
4114 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4115 | if (!pwq) { | |
2d916033 FF |
4116 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
4117 | wq->name); | |
77f300b1 | 4118 | goto use_dfl_pwq; |
4c16bd32 TH |
4119 | } |
4120 | ||
f7142ed4 | 4121 | /* Install the new pwq. */ |
4c16bd32 TH |
4122 | mutex_lock(&wq->mutex); |
4123 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
4124 | goto out_unlock; | |
4125 | ||
4126 | use_dfl_pwq: | |
f7142ed4 | 4127 | mutex_lock(&wq->mutex); |
4c16bd32 TH |
4128 | spin_lock_irq(&wq->dfl_pwq->pool->lock); |
4129 | get_pwq(wq->dfl_pwq); | |
4130 | spin_unlock_irq(&wq->dfl_pwq->pool->lock); | |
4131 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); | |
4132 | out_unlock: | |
4133 | mutex_unlock(&wq->mutex); | |
4134 | put_pwq_unlocked(old_pwq); | |
4135 | } | |
4136 | ||
30cdf249 | 4137 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4138 | { |
49e3cf44 | 4139 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4140 | int cpu, ret; |
30cdf249 TH |
4141 | |
4142 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
4143 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
4144 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
4145 | return -ENOMEM; |
4146 | ||
4147 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
4148 | struct pool_workqueue *pwq = |
4149 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 4150 | struct worker_pool *cpu_pools = |
f02ae73a | 4151 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 4152 | |
f147f29e TH |
4153 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
4154 | ||
4155 | mutex_lock(&wq->mutex); | |
1befcf30 | 4156 | link_pwq(pwq); |
f147f29e | 4157 | mutex_unlock(&wq->mutex); |
30cdf249 | 4158 | } |
9e8cd2f5 | 4159 | return 0; |
509b3204 DJ |
4160 | } |
4161 | ||
4162 | get_online_cpus(); | |
4163 | if (wq->flags & __WQ_ORDERED) { | |
8a2b7538 TH |
4164 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
4165 | /* there should only be single pwq for ordering guarantee */ | |
4166 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
4167 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
4168 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
30cdf249 | 4169 | } else { |
509b3204 | 4170 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 4171 | } |
509b3204 DJ |
4172 | put_online_cpus(); |
4173 | ||
4174 | return ret; | |
0f900049 TH |
4175 | } |
4176 | ||
f3421797 TH |
4177 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
4178 | const char *name) | |
b71ab8c2 | 4179 | { |
f3421797 TH |
4180 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
4181 | ||
4182 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
4183 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
4184 | max_active, name, 1, lim); | |
b71ab8c2 | 4185 | |
f3421797 | 4186 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
4187 | } |
4188 | ||
983c7515 TH |
4189 | /* |
4190 | * Workqueues which may be used during memory reclaim should have a rescuer | |
4191 | * to guarantee forward progress. | |
4192 | */ | |
4193 | static int init_rescuer(struct workqueue_struct *wq) | |
4194 | { | |
4195 | struct worker *rescuer; | |
4196 | int ret; | |
4197 | ||
4198 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
4199 | return 0; | |
4200 | ||
4201 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4202 | if (!rescuer) | |
4203 | return -ENOMEM; | |
4204 | ||
4205 | rescuer->rescue_wq = wq; | |
4206 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); | |
4207 | ret = PTR_ERR_OR_ZERO(rescuer->task); | |
4208 | if (ret) { | |
4209 | kfree(rescuer); | |
4210 | return ret; | |
4211 | } | |
4212 | ||
4213 | wq->rescuer = rescuer; | |
4214 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
4215 | wake_up_process(rescuer->task); | |
4216 | ||
4217 | return 0; | |
4218 | } | |
4219 | ||
a2775bbc | 4220 | __printf(1, 4) |
669de8bd BVA |
4221 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
4222 | unsigned int flags, | |
4223 | int max_active, ...) | |
1da177e4 | 4224 | { |
df2d5ae4 | 4225 | size_t tbl_size = 0; |
ecf6881f | 4226 | va_list args; |
1da177e4 | 4227 | struct workqueue_struct *wq; |
49e3cf44 | 4228 | struct pool_workqueue *pwq; |
b196be89 | 4229 | |
5c0338c6 TH |
4230 | /* |
4231 | * Unbound && max_active == 1 used to imply ordered, which is no | |
4232 | * longer the case on NUMA machines due to per-node pools. While | |
4233 | * alloc_ordered_workqueue() is the right way to create an ordered | |
4234 | * workqueue, keep the previous behavior to avoid subtle breakages | |
4235 | * on NUMA. | |
4236 | */ | |
4237 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
4238 | flags |= __WQ_ORDERED; | |
4239 | ||
cee22a15 VK |
4240 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4241 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4242 | flags |= WQ_UNBOUND; | |
4243 | ||
ecf6881f | 4244 | /* allocate wq and format name */ |
df2d5ae4 | 4245 | if (flags & WQ_UNBOUND) |
ddcb57e2 | 4246 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
df2d5ae4 TH |
4247 | |
4248 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 4249 | if (!wq) |
d2c1d404 | 4250 | return NULL; |
b196be89 | 4251 | |
6029a918 | 4252 | if (flags & WQ_UNBOUND) { |
be69d00d | 4253 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
4254 | if (!wq->unbound_attrs) |
4255 | goto err_free_wq; | |
4256 | } | |
4257 | ||
669de8bd | 4258 | va_start(args, max_active); |
ecf6881f | 4259 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 4260 | va_end(args); |
1da177e4 | 4261 | |
d320c038 | 4262 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4263 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4264 | |
b196be89 | 4265 | /* init wq */ |
97e37d7b | 4266 | wq->flags = flags; |
a0a1a5fd | 4267 | wq->saved_max_active = max_active; |
3c25a55d | 4268 | mutex_init(&wq->mutex); |
112202d9 | 4269 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4270 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4271 | INIT_LIST_HEAD(&wq->flusher_queue); |
4272 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4273 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4274 | |
669de8bd | 4275 | wq_init_lockdep(wq); |
cce1a165 | 4276 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4277 | |
30cdf249 | 4278 | if (alloc_and_link_pwqs(wq) < 0) |
82efcab3 | 4279 | goto err_unreg_lockdep; |
1537663f | 4280 | |
40c17f75 | 4281 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 4282 | goto err_destroy; |
3af24433 | 4283 | |
226223ab TH |
4284 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4285 | goto err_destroy; | |
4286 | ||
a0a1a5fd | 4287 | /* |
68e13a67 LJ |
4288 | * wq_pool_mutex protects global freeze state and workqueues list. |
4289 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4290 | * list. | |
a0a1a5fd | 4291 | */ |
68e13a67 | 4292 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4293 | |
a357fc03 | 4294 | mutex_lock(&wq->mutex); |
699ce097 TH |
4295 | for_each_pwq(pwq, wq) |
4296 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4297 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4298 | |
e2dca7ad | 4299 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4300 | |
68e13a67 | 4301 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4302 | |
3af24433 | 4303 | return wq; |
d2c1d404 | 4304 | |
82efcab3 | 4305 | err_unreg_lockdep: |
009bb421 BVA |
4306 | wq_unregister_lockdep(wq); |
4307 | wq_free_lockdep(wq); | |
82efcab3 | 4308 | err_free_wq: |
6029a918 | 4309 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4310 | kfree(wq); |
4311 | return NULL; | |
4312 | err_destroy: | |
4313 | destroy_workqueue(wq); | |
4690c4ab | 4314 | return NULL; |
3af24433 | 4315 | } |
669de8bd | 4316 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 4317 | |
c29eb853 TH |
4318 | static bool pwq_busy(struct pool_workqueue *pwq) |
4319 | { | |
4320 | int i; | |
4321 | ||
4322 | for (i = 0; i < WORK_NR_COLORS; i++) | |
4323 | if (pwq->nr_in_flight[i]) | |
4324 | return true; | |
4325 | ||
4326 | if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) | |
4327 | return true; | |
4328 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) | |
4329 | return true; | |
4330 | ||
4331 | return false; | |
4332 | } | |
4333 | ||
3af24433 ON |
4334 | /** |
4335 | * destroy_workqueue - safely terminate a workqueue | |
4336 | * @wq: target workqueue | |
4337 | * | |
4338 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4339 | */ | |
4340 | void destroy_workqueue(struct workqueue_struct *wq) | |
4341 | { | |
49e3cf44 | 4342 | struct pool_workqueue *pwq; |
4c16bd32 | 4343 | int node; |
3af24433 | 4344 | |
def98c84 TH |
4345 | /* |
4346 | * Remove it from sysfs first so that sanity check failure doesn't | |
4347 | * lead to sysfs name conflicts. | |
4348 | */ | |
4349 | workqueue_sysfs_unregister(wq); | |
4350 | ||
9c5a2ba7 TH |
4351 | /* drain it before proceeding with destruction */ |
4352 | drain_workqueue(wq); | |
c8efcc25 | 4353 | |
def98c84 TH |
4354 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
4355 | if (wq->rescuer) { | |
4356 | struct worker *rescuer = wq->rescuer; | |
4357 | ||
4358 | /* this prevents new queueing */ | |
4359 | spin_lock_irq(&wq_mayday_lock); | |
4360 | wq->rescuer = NULL; | |
4361 | spin_unlock_irq(&wq_mayday_lock); | |
4362 | ||
4363 | /* rescuer will empty maydays list before exiting */ | |
4364 | kthread_stop(rescuer->task); | |
8efe1223 | 4365 | kfree(rescuer); |
def98c84 TH |
4366 | } |
4367 | ||
c29eb853 TH |
4368 | /* |
4369 | * Sanity checks - grab all the locks so that we wait for all | |
4370 | * in-flight operations which may do put_pwq(). | |
4371 | */ | |
4372 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 4373 | mutex_lock(&wq->mutex); |
49e3cf44 | 4374 | for_each_pwq(pwq, wq) { |
c29eb853 TH |
4375 | spin_lock_irq(&pwq->pool->lock); |
4376 | if (WARN_ON(pwq_busy(pwq))) { | |
1d9a6159 KW |
4377 | pr_warn("%s: %s has the following busy pwq\n", |
4378 | __func__, wq->name); | |
c29eb853 TH |
4379 | show_pwq(pwq); |
4380 | spin_unlock_irq(&pwq->pool->lock); | |
b09f4fd3 | 4381 | mutex_unlock(&wq->mutex); |
c29eb853 | 4382 | mutex_unlock(&wq_pool_mutex); |
fa07fb6a | 4383 | show_workqueue_state(); |
6183c009 | 4384 | return; |
76af4d93 | 4385 | } |
c29eb853 | 4386 | spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 4387 | } |
b09f4fd3 | 4388 | mutex_unlock(&wq->mutex); |
c29eb853 | 4389 | mutex_unlock(&wq_pool_mutex); |
6183c009 | 4390 | |
a0a1a5fd TH |
4391 | /* |
4392 | * wq list is used to freeze wq, remove from list after | |
4393 | * flushing is complete in case freeze races us. | |
4394 | */ | |
68e13a67 | 4395 | mutex_lock(&wq_pool_mutex); |
e2dca7ad | 4396 | list_del_rcu(&wq->list); |
68e13a67 | 4397 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4398 | |
8864b4e5 | 4399 | if (!(wq->flags & WQ_UNBOUND)) { |
669de8bd | 4400 | wq_unregister_lockdep(wq); |
8864b4e5 TH |
4401 | /* |
4402 | * The base ref is never dropped on per-cpu pwqs. Directly | |
e2dca7ad | 4403 | * schedule RCU free. |
8864b4e5 | 4404 | */ |
25b00775 | 4405 | call_rcu(&wq->rcu, rcu_free_wq); |
8864b4e5 TH |
4406 | } else { |
4407 | /* | |
4408 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4409 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4410 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4411 | */ |
4c16bd32 TH |
4412 | for_each_node(node) { |
4413 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4414 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4415 | put_pwq_unlocked(pwq); | |
4416 | } | |
4417 | ||
4418 | /* | |
4419 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4420 | * put. Don't access it afterwards. | |
4421 | */ | |
4422 | pwq = wq->dfl_pwq; | |
4423 | wq->dfl_pwq = NULL; | |
dce90d47 | 4424 | put_pwq_unlocked(pwq); |
29c91e99 | 4425 | } |
3af24433 ON |
4426 | } |
4427 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4428 | ||
dcd989cb TH |
4429 | /** |
4430 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4431 | * @wq: target workqueue | |
4432 | * @max_active: new max_active value. | |
4433 | * | |
4434 | * Set max_active of @wq to @max_active. | |
4435 | * | |
4436 | * CONTEXT: | |
4437 | * Don't call from IRQ context. | |
4438 | */ | |
4439 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4440 | { | |
49e3cf44 | 4441 | struct pool_workqueue *pwq; |
dcd989cb | 4442 | |
8719dcea | 4443 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 4444 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
4445 | return; |
4446 | ||
f3421797 | 4447 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4448 | |
a357fc03 | 4449 | mutex_lock(&wq->mutex); |
dcd989cb | 4450 | |
0a94efb5 | 4451 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb TH |
4452 | wq->saved_max_active = max_active; |
4453 | ||
699ce097 TH |
4454 | for_each_pwq(pwq, wq) |
4455 | pwq_adjust_max_active(pwq); | |
93981800 | 4456 | |
a357fc03 | 4457 | mutex_unlock(&wq->mutex); |
15316ba8 | 4458 | } |
dcd989cb | 4459 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4460 | |
27d4ee03 LW |
4461 | /** |
4462 | * current_work - retrieve %current task's work struct | |
4463 | * | |
4464 | * Determine if %current task is a workqueue worker and what it's working on. | |
4465 | * Useful to find out the context that the %current task is running in. | |
4466 | * | |
4467 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
4468 | */ | |
4469 | struct work_struct *current_work(void) | |
4470 | { | |
4471 | struct worker *worker = current_wq_worker(); | |
4472 | ||
4473 | return worker ? worker->current_work : NULL; | |
4474 | } | |
4475 | EXPORT_SYMBOL(current_work); | |
4476 | ||
e6267616 TH |
4477 | /** |
4478 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4479 | * | |
4480 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4481 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4482 | * |
4483 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4484 | */ |
4485 | bool current_is_workqueue_rescuer(void) | |
4486 | { | |
4487 | struct worker *worker = current_wq_worker(); | |
4488 | ||
6a092dfd | 4489 | return worker && worker->rescue_wq; |
e6267616 TH |
4490 | } |
4491 | ||
eef6a7d5 | 4492 | /** |
dcd989cb TH |
4493 | * workqueue_congested - test whether a workqueue is congested |
4494 | * @cpu: CPU in question | |
4495 | * @wq: target workqueue | |
eef6a7d5 | 4496 | * |
dcd989cb TH |
4497 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4498 | * no synchronization around this function and the test result is | |
4499 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4500 | * |
d3251859 TH |
4501 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4502 | * Note that both per-cpu and unbound workqueues may be associated with | |
4503 | * multiple pool_workqueues which have separate congested states. A | |
4504 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4505 | * contested on other CPUs / NUMA nodes. | |
4506 | * | |
d185af30 | 4507 | * Return: |
dcd989cb | 4508 | * %true if congested, %false otherwise. |
eef6a7d5 | 4509 | */ |
d84ff051 | 4510 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4511 | { |
7fb98ea7 | 4512 | struct pool_workqueue *pwq; |
76af4d93 TH |
4513 | bool ret; |
4514 | ||
24acfb71 TG |
4515 | rcu_read_lock(); |
4516 | preempt_disable(); | |
7fb98ea7 | 4517 | |
d3251859 TH |
4518 | if (cpu == WORK_CPU_UNBOUND) |
4519 | cpu = smp_processor_id(); | |
4520 | ||
7fb98ea7 TH |
4521 | if (!(wq->flags & WQ_UNBOUND)) |
4522 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4523 | else | |
df2d5ae4 | 4524 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4525 | |
76af4d93 | 4526 | ret = !list_empty(&pwq->delayed_works); |
24acfb71 TG |
4527 | preempt_enable(); |
4528 | rcu_read_unlock(); | |
76af4d93 TH |
4529 | |
4530 | return ret; | |
1da177e4 | 4531 | } |
dcd989cb | 4532 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4533 | |
dcd989cb TH |
4534 | /** |
4535 | * work_busy - test whether a work is currently pending or running | |
4536 | * @work: the work to be tested | |
4537 | * | |
4538 | * Test whether @work is currently pending or running. There is no | |
4539 | * synchronization around this function and the test result is | |
4540 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4541 | * |
d185af30 | 4542 | * Return: |
dcd989cb TH |
4543 | * OR'd bitmask of WORK_BUSY_* bits. |
4544 | */ | |
4545 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4546 | { |
fa1b54e6 | 4547 | struct worker_pool *pool; |
dcd989cb TH |
4548 | unsigned long flags; |
4549 | unsigned int ret = 0; | |
1da177e4 | 4550 | |
dcd989cb TH |
4551 | if (work_pending(work)) |
4552 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4553 | |
24acfb71 | 4554 | rcu_read_lock(); |
fa1b54e6 | 4555 | pool = get_work_pool(work); |
038366c5 | 4556 | if (pool) { |
24acfb71 | 4557 | spin_lock_irqsave(&pool->lock, flags); |
038366c5 LJ |
4558 | if (find_worker_executing_work(pool, work)) |
4559 | ret |= WORK_BUSY_RUNNING; | |
24acfb71 | 4560 | spin_unlock_irqrestore(&pool->lock, flags); |
038366c5 | 4561 | } |
24acfb71 | 4562 | rcu_read_unlock(); |
1da177e4 | 4563 | |
dcd989cb | 4564 | return ret; |
1da177e4 | 4565 | } |
dcd989cb | 4566 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4567 | |
3d1cb205 TH |
4568 | /** |
4569 | * set_worker_desc - set description for the current work item | |
4570 | * @fmt: printf-style format string | |
4571 | * @...: arguments for the format string | |
4572 | * | |
4573 | * This function can be called by a running work function to describe what | |
4574 | * the work item is about. If the worker task gets dumped, this | |
4575 | * information will be printed out together to help debugging. The | |
4576 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4577 | */ | |
4578 | void set_worker_desc(const char *fmt, ...) | |
4579 | { | |
4580 | struct worker *worker = current_wq_worker(); | |
4581 | va_list args; | |
4582 | ||
4583 | if (worker) { | |
4584 | va_start(args, fmt); | |
4585 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4586 | va_end(args); | |
3d1cb205 TH |
4587 | } |
4588 | } | |
5c750d58 | 4589 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
4590 | |
4591 | /** | |
4592 | * print_worker_info - print out worker information and description | |
4593 | * @log_lvl: the log level to use when printing | |
4594 | * @task: target task | |
4595 | * | |
4596 | * If @task is a worker and currently executing a work item, print out the | |
4597 | * name of the workqueue being serviced and worker description set with | |
4598 | * set_worker_desc() by the currently executing work item. | |
4599 | * | |
4600 | * This function can be safely called on any task as long as the | |
4601 | * task_struct itself is accessible. While safe, this function isn't | |
4602 | * synchronized and may print out mixups or garbages of limited length. | |
4603 | */ | |
4604 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4605 | { | |
4606 | work_func_t *fn = NULL; | |
4607 | char name[WQ_NAME_LEN] = { }; | |
4608 | char desc[WORKER_DESC_LEN] = { }; | |
4609 | struct pool_workqueue *pwq = NULL; | |
4610 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
4611 | struct worker *worker; |
4612 | ||
4613 | if (!(task->flags & PF_WQ_WORKER)) | |
4614 | return; | |
4615 | ||
4616 | /* | |
4617 | * This function is called without any synchronization and @task | |
4618 | * could be in any state. Be careful with dereferences. | |
4619 | */ | |
e700591a | 4620 | worker = kthread_probe_data(task); |
3d1cb205 TH |
4621 | |
4622 | /* | |
8bf89593 TH |
4623 | * Carefully copy the associated workqueue's workfn, name and desc. |
4624 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 TH |
4625 | */ |
4626 | probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); | |
4627 | probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4628 | probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); | |
4629 | probe_kernel_read(name, wq->name, sizeof(name) - 1); | |
8bf89593 | 4630 | probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); |
3d1cb205 TH |
4631 | |
4632 | if (fn || name[0] || desc[0]) { | |
d75f773c | 4633 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 4634 | if (strcmp(name, desc)) |
3d1cb205 TH |
4635 | pr_cont(" (%s)", desc); |
4636 | pr_cont("\n"); | |
4637 | } | |
4638 | } | |
4639 | ||
3494fc30 TH |
4640 | static void pr_cont_pool_info(struct worker_pool *pool) |
4641 | { | |
4642 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
4643 | if (pool->node != NUMA_NO_NODE) | |
4644 | pr_cont(" node=%d", pool->node); | |
4645 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
4646 | } | |
4647 | ||
4648 | static void pr_cont_work(bool comma, struct work_struct *work) | |
4649 | { | |
4650 | if (work->func == wq_barrier_func) { | |
4651 | struct wq_barrier *barr; | |
4652 | ||
4653 | barr = container_of(work, struct wq_barrier, work); | |
4654 | ||
4655 | pr_cont("%s BAR(%d)", comma ? "," : "", | |
4656 | task_pid_nr(barr->task)); | |
4657 | } else { | |
d75f773c | 4658 | pr_cont("%s %ps", comma ? "," : "", work->func); |
3494fc30 TH |
4659 | } |
4660 | } | |
4661 | ||
4662 | static void show_pwq(struct pool_workqueue *pwq) | |
4663 | { | |
4664 | struct worker_pool *pool = pwq->pool; | |
4665 | struct work_struct *work; | |
4666 | struct worker *worker; | |
4667 | bool has_in_flight = false, has_pending = false; | |
4668 | int bkt; | |
4669 | ||
4670 | pr_info(" pwq %d:", pool->id); | |
4671 | pr_cont_pool_info(pool); | |
4672 | ||
e66b39af TH |
4673 | pr_cont(" active=%d/%d refcnt=%d%s\n", |
4674 | pwq->nr_active, pwq->max_active, pwq->refcnt, | |
3494fc30 TH |
4675 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
4676 | ||
4677 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4678 | if (worker->current_pwq == pwq) { | |
4679 | has_in_flight = true; | |
4680 | break; | |
4681 | } | |
4682 | } | |
4683 | if (has_in_flight) { | |
4684 | bool comma = false; | |
4685 | ||
4686 | pr_info(" in-flight:"); | |
4687 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4688 | if (worker->current_pwq != pwq) | |
4689 | continue; | |
4690 | ||
d75f773c | 4691 | pr_cont("%s %d%s:%ps", comma ? "," : "", |
3494fc30 | 4692 | task_pid_nr(worker->task), |
30ae2fc0 | 4693 | worker->rescue_wq ? "(RESCUER)" : "", |
3494fc30 TH |
4694 | worker->current_func); |
4695 | list_for_each_entry(work, &worker->scheduled, entry) | |
4696 | pr_cont_work(false, work); | |
4697 | comma = true; | |
4698 | } | |
4699 | pr_cont("\n"); | |
4700 | } | |
4701 | ||
4702 | list_for_each_entry(work, &pool->worklist, entry) { | |
4703 | if (get_work_pwq(work) == pwq) { | |
4704 | has_pending = true; | |
4705 | break; | |
4706 | } | |
4707 | } | |
4708 | if (has_pending) { | |
4709 | bool comma = false; | |
4710 | ||
4711 | pr_info(" pending:"); | |
4712 | list_for_each_entry(work, &pool->worklist, entry) { | |
4713 | if (get_work_pwq(work) != pwq) | |
4714 | continue; | |
4715 | ||
4716 | pr_cont_work(comma, work); | |
4717 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4718 | } | |
4719 | pr_cont("\n"); | |
4720 | } | |
4721 | ||
4722 | if (!list_empty(&pwq->delayed_works)) { | |
4723 | bool comma = false; | |
4724 | ||
4725 | pr_info(" delayed:"); | |
4726 | list_for_each_entry(work, &pwq->delayed_works, entry) { | |
4727 | pr_cont_work(comma, work); | |
4728 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4729 | } | |
4730 | pr_cont("\n"); | |
4731 | } | |
4732 | } | |
4733 | ||
4734 | /** | |
4735 | * show_workqueue_state - dump workqueue state | |
4736 | * | |
7b776af6 RL |
4737 | * Called from a sysrq handler or try_to_freeze_tasks() and prints out |
4738 | * all busy workqueues and pools. | |
3494fc30 TH |
4739 | */ |
4740 | void show_workqueue_state(void) | |
4741 | { | |
4742 | struct workqueue_struct *wq; | |
4743 | struct worker_pool *pool; | |
4744 | unsigned long flags; | |
4745 | int pi; | |
4746 | ||
24acfb71 | 4747 | rcu_read_lock(); |
3494fc30 TH |
4748 | |
4749 | pr_info("Showing busy workqueues and worker pools:\n"); | |
4750 | ||
4751 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
4752 | struct pool_workqueue *pwq; | |
4753 | bool idle = true; | |
4754 | ||
4755 | for_each_pwq(pwq, wq) { | |
4756 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { | |
4757 | idle = false; | |
4758 | break; | |
4759 | } | |
4760 | } | |
4761 | if (idle) | |
4762 | continue; | |
4763 | ||
4764 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); | |
4765 | ||
4766 | for_each_pwq(pwq, wq) { | |
4767 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
4768 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) | |
4769 | show_pwq(pwq); | |
4770 | spin_unlock_irqrestore(&pwq->pool->lock, flags); | |
62635ea8 SS |
4771 | /* |
4772 | * We could be printing a lot from atomic context, e.g. | |
4773 | * sysrq-t -> show_workqueue_state(). Avoid triggering | |
4774 | * hard lockup. | |
4775 | */ | |
4776 | touch_nmi_watchdog(); | |
3494fc30 TH |
4777 | } |
4778 | } | |
4779 | ||
4780 | for_each_pool(pool, pi) { | |
4781 | struct worker *worker; | |
4782 | bool first = true; | |
4783 | ||
4784 | spin_lock_irqsave(&pool->lock, flags); | |
4785 | if (pool->nr_workers == pool->nr_idle) | |
4786 | goto next_pool; | |
4787 | ||
4788 | pr_info("pool %d:", pool->id); | |
4789 | pr_cont_pool_info(pool); | |
82607adc TH |
4790 | pr_cont(" hung=%us workers=%d", |
4791 | jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, | |
4792 | pool->nr_workers); | |
3494fc30 TH |
4793 | if (pool->manager) |
4794 | pr_cont(" manager: %d", | |
4795 | task_pid_nr(pool->manager->task)); | |
4796 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
4797 | pr_cont(" %s%d", first ? "idle: " : "", | |
4798 | task_pid_nr(worker->task)); | |
4799 | first = false; | |
4800 | } | |
4801 | pr_cont("\n"); | |
4802 | next_pool: | |
4803 | spin_unlock_irqrestore(&pool->lock, flags); | |
62635ea8 SS |
4804 | /* |
4805 | * We could be printing a lot from atomic context, e.g. | |
4806 | * sysrq-t -> show_workqueue_state(). Avoid triggering | |
4807 | * hard lockup. | |
4808 | */ | |
4809 | touch_nmi_watchdog(); | |
3494fc30 TH |
4810 | } |
4811 | ||
24acfb71 | 4812 | rcu_read_unlock(); |
3494fc30 TH |
4813 | } |
4814 | ||
6b59808b TH |
4815 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
4816 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
4817 | { | |
6b59808b TH |
4818 | int off; |
4819 | ||
4820 | /* always show the actual comm */ | |
4821 | off = strscpy(buf, task->comm, size); | |
4822 | if (off < 0) | |
4823 | return; | |
4824 | ||
197f6acc | 4825 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
4826 | mutex_lock(&wq_pool_attach_mutex); |
4827 | ||
197f6acc TH |
4828 | if (task->flags & PF_WQ_WORKER) { |
4829 | struct worker *worker = kthread_data(task); | |
4830 | struct worker_pool *pool = worker->pool; | |
6b59808b | 4831 | |
197f6acc TH |
4832 | if (pool) { |
4833 | spin_lock_irq(&pool->lock); | |
4834 | /* | |
4835 | * ->desc tracks information (wq name or | |
4836 | * set_worker_desc()) for the latest execution. If | |
4837 | * current, prepend '+', otherwise '-'. | |
4838 | */ | |
4839 | if (worker->desc[0] != '\0') { | |
4840 | if (worker->current_work) | |
4841 | scnprintf(buf + off, size - off, "+%s", | |
4842 | worker->desc); | |
4843 | else | |
4844 | scnprintf(buf + off, size - off, "-%s", | |
4845 | worker->desc); | |
4846 | } | |
4847 | spin_unlock_irq(&pool->lock); | |
6b59808b | 4848 | } |
6b59808b TH |
4849 | } |
4850 | ||
4851 | mutex_unlock(&wq_pool_attach_mutex); | |
4852 | } | |
4853 | ||
66448bc2 MM |
4854 | #ifdef CONFIG_SMP |
4855 | ||
db7bccf4 TH |
4856 | /* |
4857 | * CPU hotplug. | |
4858 | * | |
e22bee78 | 4859 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 4860 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 4861 | * pool which make migrating pending and scheduled works very |
e22bee78 | 4862 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 4863 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
4864 | * blocked draining impractical. |
4865 | * | |
24647570 | 4866 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
4867 | * running as an unbound one and allowing it to be reattached later if the |
4868 | * cpu comes back online. | |
db7bccf4 | 4869 | */ |
1da177e4 | 4870 | |
e8b3f8db | 4871 | static void unbind_workers(int cpu) |
3af24433 | 4872 | { |
4ce62e9e | 4873 | struct worker_pool *pool; |
db7bccf4 | 4874 | struct worker *worker; |
3af24433 | 4875 | |
f02ae73a | 4876 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 4877 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 4878 | spin_lock_irq(&pool->lock); |
3af24433 | 4879 | |
94cf58bb | 4880 | /* |
92f9c5c4 | 4881 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb TH |
4882 | * unbound and set DISASSOCIATED. Before this, all workers |
4883 | * except for the ones which are still executing works from | |
4884 | * before the last CPU down must be on the cpu. After | |
4885 | * this, they may become diasporas. | |
4886 | */ | |
da028469 | 4887 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 4888 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 4889 | |
24647570 | 4890 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 4891 | |
94cf58bb | 4892 | spin_unlock_irq(&pool->lock); |
1258fae7 | 4893 | mutex_unlock(&wq_pool_attach_mutex); |
628c78e7 | 4894 | |
eb283428 LJ |
4895 | /* |
4896 | * Call schedule() so that we cross rq->lock and thus can | |
4897 | * guarantee sched callbacks see the %WORKER_UNBOUND flag. | |
4898 | * This is necessary as scheduler callbacks may be invoked | |
4899 | * from other cpus. | |
4900 | */ | |
4901 | schedule(); | |
06ba38a9 | 4902 | |
eb283428 LJ |
4903 | /* |
4904 | * Sched callbacks are disabled now. Zap nr_running. | |
4905 | * After this, nr_running stays zero and need_more_worker() | |
4906 | * and keep_working() are always true as long as the | |
4907 | * worklist is not empty. This pool now behaves as an | |
4908 | * unbound (in terms of concurrency management) pool which | |
4909 | * are served by workers tied to the pool. | |
4910 | */ | |
e19e397a | 4911 | atomic_set(&pool->nr_running, 0); |
eb283428 LJ |
4912 | |
4913 | /* | |
4914 | * With concurrency management just turned off, a busy | |
4915 | * worker blocking could lead to lengthy stalls. Kick off | |
4916 | * unbound chain execution of currently pending work items. | |
4917 | */ | |
4918 | spin_lock_irq(&pool->lock); | |
4919 | wake_up_worker(pool); | |
4920 | spin_unlock_irq(&pool->lock); | |
4921 | } | |
3af24433 | 4922 | } |
3af24433 | 4923 | |
bd7c089e TH |
4924 | /** |
4925 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
4926 | * @pool: pool of interest | |
4927 | * | |
a9ab775b | 4928 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
4929 | */ |
4930 | static void rebind_workers(struct worker_pool *pool) | |
4931 | { | |
a9ab775b | 4932 | struct worker *worker; |
bd7c089e | 4933 | |
1258fae7 | 4934 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 4935 | |
a9ab775b TH |
4936 | /* |
4937 | * Restore CPU affinity of all workers. As all idle workers should | |
4938 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 4939 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
4940 | * of all workers first and then clear UNBOUND. As we're called |
4941 | * from CPU_ONLINE, the following shouldn't fail. | |
4942 | */ | |
da028469 | 4943 | for_each_pool_worker(worker, pool) |
a9ab775b TH |
4944 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, |
4945 | pool->attrs->cpumask) < 0); | |
bd7c089e | 4946 | |
a9ab775b | 4947 | spin_lock_irq(&pool->lock); |
f7c17d26 | 4948 | |
3de5e884 | 4949 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 4950 | |
da028469 | 4951 | for_each_pool_worker(worker, pool) { |
a9ab775b | 4952 | unsigned int worker_flags = worker->flags; |
bd7c089e TH |
4953 | |
4954 | /* | |
a9ab775b TH |
4955 | * A bound idle worker should actually be on the runqueue |
4956 | * of the associated CPU for local wake-ups targeting it to | |
4957 | * work. Kick all idle workers so that they migrate to the | |
4958 | * associated CPU. Doing this in the same loop as | |
4959 | * replacing UNBOUND with REBOUND is safe as no worker will | |
4960 | * be bound before @pool->lock is released. | |
bd7c089e | 4961 | */ |
a9ab775b TH |
4962 | if (worker_flags & WORKER_IDLE) |
4963 | wake_up_process(worker->task); | |
bd7c089e | 4964 | |
a9ab775b TH |
4965 | /* |
4966 | * We want to clear UNBOUND but can't directly call | |
4967 | * worker_clr_flags() or adjust nr_running. Atomically | |
4968 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
4969 | * @worker will clear REBOUND using worker_clr_flags() when | |
4970 | * it initiates the next execution cycle thus restoring | |
4971 | * concurrency management. Note that when or whether | |
4972 | * @worker clears REBOUND doesn't affect correctness. | |
4973 | * | |
c95491ed | 4974 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 4975 | * tested without holding any lock in |
6d25be57 | 4976 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
4977 | * fail incorrectly leading to premature concurrency |
4978 | * management operations. | |
4979 | */ | |
4980 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
4981 | worker_flags |= WORKER_REBOUND; | |
4982 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 4983 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 4984 | } |
a9ab775b TH |
4985 | |
4986 | spin_unlock_irq(&pool->lock); | |
bd7c089e TH |
4987 | } |
4988 | ||
7dbc725e TH |
4989 | /** |
4990 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
4991 | * @pool: unbound pool of interest | |
4992 | * @cpu: the CPU which is coming up | |
4993 | * | |
4994 | * An unbound pool may end up with a cpumask which doesn't have any online | |
4995 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
4996 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
4997 | * online CPU before, cpus_allowed of all its workers should be restored. | |
4998 | */ | |
4999 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
5000 | { | |
5001 | static cpumask_t cpumask; | |
5002 | struct worker *worker; | |
7dbc725e | 5003 | |
1258fae7 | 5004 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
5005 | |
5006 | /* is @cpu allowed for @pool? */ | |
5007 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
5008 | return; | |
5009 | ||
7dbc725e | 5010 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
5011 | |
5012 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 5013 | for_each_pool_worker(worker, pool) |
d945b5e9 | 5014 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
5015 | } |
5016 | ||
7ee681b2 TG |
5017 | int workqueue_prepare_cpu(unsigned int cpu) |
5018 | { | |
5019 | struct worker_pool *pool; | |
5020 | ||
5021 | for_each_cpu_worker_pool(pool, cpu) { | |
5022 | if (pool->nr_workers) | |
5023 | continue; | |
5024 | if (!create_worker(pool)) | |
5025 | return -ENOMEM; | |
5026 | } | |
5027 | return 0; | |
5028 | } | |
5029 | ||
5030 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 5031 | { |
4ce62e9e | 5032 | struct worker_pool *pool; |
4c16bd32 | 5033 | struct workqueue_struct *wq; |
7dbc725e | 5034 | int pi; |
3ce63377 | 5035 | |
7ee681b2 | 5036 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 5037 | |
7ee681b2 | 5038 | for_each_pool(pool, pi) { |
1258fae7 | 5039 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 5040 | |
7ee681b2 TG |
5041 | if (pool->cpu == cpu) |
5042 | rebind_workers(pool); | |
5043 | else if (pool->cpu < 0) | |
5044 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 5045 | |
1258fae7 | 5046 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 5047 | } |
6ba94429 | 5048 | |
7ee681b2 TG |
5049 | /* update NUMA affinity of unbound workqueues */ |
5050 | list_for_each_entry(wq, &workqueues, list) | |
5051 | wq_update_unbound_numa(wq, cpu, true); | |
6ba94429 | 5052 | |
7ee681b2 TG |
5053 | mutex_unlock(&wq_pool_mutex); |
5054 | return 0; | |
6ba94429 FW |
5055 | } |
5056 | ||
7ee681b2 | 5057 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 5058 | { |
6ba94429 FW |
5059 | struct workqueue_struct *wq; |
5060 | ||
7ee681b2 | 5061 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
5062 | if (WARN_ON(cpu != smp_processor_id())) |
5063 | return -1; | |
5064 | ||
5065 | unbind_workers(cpu); | |
7ee681b2 TG |
5066 | |
5067 | /* update NUMA affinity of unbound workqueues */ | |
5068 | mutex_lock(&wq_pool_mutex); | |
5069 | list_for_each_entry(wq, &workqueues, list) | |
5070 | wq_update_unbound_numa(wq, cpu, false); | |
5071 | mutex_unlock(&wq_pool_mutex); | |
5072 | ||
7ee681b2 | 5073 | return 0; |
6ba94429 FW |
5074 | } |
5075 | ||
6ba94429 FW |
5076 | struct work_for_cpu { |
5077 | struct work_struct work; | |
5078 | long (*fn)(void *); | |
5079 | void *arg; | |
5080 | long ret; | |
5081 | }; | |
5082 | ||
5083 | static void work_for_cpu_fn(struct work_struct *work) | |
5084 | { | |
5085 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
5086 | ||
5087 | wfc->ret = wfc->fn(wfc->arg); | |
5088 | } | |
5089 | ||
5090 | /** | |
22aceb31 | 5091 | * work_on_cpu - run a function in thread context on a particular cpu |
6ba94429 FW |
5092 | * @cpu: the cpu to run on |
5093 | * @fn: the function to run | |
5094 | * @arg: the function arg | |
5095 | * | |
5096 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
5097 | * The caller must not hold any locks which would prevent @fn from completing. | |
5098 | * | |
5099 | * Return: The value @fn returns. | |
5100 | */ | |
5101 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
5102 | { | |
5103 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
5104 | ||
5105 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
5106 | schedule_work_on(cpu, &wfc.work); | |
5107 | flush_work(&wfc.work); | |
5108 | destroy_work_on_stack(&wfc.work); | |
5109 | return wfc.ret; | |
5110 | } | |
5111 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
0e8d6a93 TG |
5112 | |
5113 | /** | |
5114 | * work_on_cpu_safe - run a function in thread context on a particular cpu | |
5115 | * @cpu: the cpu to run on | |
5116 | * @fn: the function to run | |
5117 | * @arg: the function argument | |
5118 | * | |
5119 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
5120 | * any locks which would prevent @fn from completing. | |
5121 | * | |
5122 | * Return: The value @fn returns. | |
5123 | */ | |
5124 | long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) | |
5125 | { | |
5126 | long ret = -ENODEV; | |
5127 | ||
5128 | get_online_cpus(); | |
5129 | if (cpu_online(cpu)) | |
5130 | ret = work_on_cpu(cpu, fn, arg); | |
5131 | put_online_cpus(); | |
5132 | return ret; | |
5133 | } | |
5134 | EXPORT_SYMBOL_GPL(work_on_cpu_safe); | |
6ba94429 FW |
5135 | #endif /* CONFIG_SMP */ |
5136 | ||
5137 | #ifdef CONFIG_FREEZER | |
5138 | ||
5139 | /** | |
5140 | * freeze_workqueues_begin - begin freezing workqueues | |
5141 | * | |
5142 | * Start freezing workqueues. After this function returns, all freezable | |
5143 | * workqueues will queue new works to their delayed_works list instead of | |
5144 | * pool->worklist. | |
5145 | * | |
5146 | * CONTEXT: | |
5147 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5148 | */ | |
5149 | void freeze_workqueues_begin(void) | |
5150 | { | |
5151 | struct workqueue_struct *wq; | |
5152 | struct pool_workqueue *pwq; | |
5153 | ||
5154 | mutex_lock(&wq_pool_mutex); | |
5155 | ||
5156 | WARN_ON_ONCE(workqueue_freezing); | |
5157 | workqueue_freezing = true; | |
5158 | ||
5159 | list_for_each_entry(wq, &workqueues, list) { | |
5160 | mutex_lock(&wq->mutex); | |
5161 | for_each_pwq(pwq, wq) | |
5162 | pwq_adjust_max_active(pwq); | |
5163 | mutex_unlock(&wq->mutex); | |
5164 | } | |
5165 | ||
5166 | mutex_unlock(&wq_pool_mutex); | |
5167 | } | |
5168 | ||
5169 | /** | |
5170 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
5171 | * | |
5172 | * Check whether freezing is complete. This function must be called | |
5173 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
5174 | * | |
5175 | * CONTEXT: | |
5176 | * Grabs and releases wq_pool_mutex. | |
5177 | * | |
5178 | * Return: | |
5179 | * %true if some freezable workqueues are still busy. %false if freezing | |
5180 | * is complete. | |
5181 | */ | |
5182 | bool freeze_workqueues_busy(void) | |
5183 | { | |
5184 | bool busy = false; | |
5185 | struct workqueue_struct *wq; | |
5186 | struct pool_workqueue *pwq; | |
5187 | ||
5188 | mutex_lock(&wq_pool_mutex); | |
5189 | ||
5190 | WARN_ON_ONCE(!workqueue_freezing); | |
5191 | ||
5192 | list_for_each_entry(wq, &workqueues, list) { | |
5193 | if (!(wq->flags & WQ_FREEZABLE)) | |
5194 | continue; | |
5195 | /* | |
5196 | * nr_active is monotonically decreasing. It's safe | |
5197 | * to peek without lock. | |
5198 | */ | |
24acfb71 | 5199 | rcu_read_lock(); |
6ba94429 FW |
5200 | for_each_pwq(pwq, wq) { |
5201 | WARN_ON_ONCE(pwq->nr_active < 0); | |
5202 | if (pwq->nr_active) { | |
5203 | busy = true; | |
24acfb71 | 5204 | rcu_read_unlock(); |
6ba94429 FW |
5205 | goto out_unlock; |
5206 | } | |
5207 | } | |
24acfb71 | 5208 | rcu_read_unlock(); |
6ba94429 FW |
5209 | } |
5210 | out_unlock: | |
5211 | mutex_unlock(&wq_pool_mutex); | |
5212 | return busy; | |
5213 | } | |
5214 | ||
5215 | /** | |
5216 | * thaw_workqueues - thaw workqueues | |
5217 | * | |
5218 | * Thaw workqueues. Normal queueing is restored and all collected | |
5219 | * frozen works are transferred to their respective pool worklists. | |
5220 | * | |
5221 | * CONTEXT: | |
5222 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5223 | */ | |
5224 | void thaw_workqueues(void) | |
5225 | { | |
5226 | struct workqueue_struct *wq; | |
5227 | struct pool_workqueue *pwq; | |
5228 | ||
5229 | mutex_lock(&wq_pool_mutex); | |
5230 | ||
5231 | if (!workqueue_freezing) | |
5232 | goto out_unlock; | |
5233 | ||
5234 | workqueue_freezing = false; | |
5235 | ||
5236 | /* restore max_active and repopulate worklist */ | |
5237 | list_for_each_entry(wq, &workqueues, list) { | |
5238 | mutex_lock(&wq->mutex); | |
5239 | for_each_pwq(pwq, wq) | |
5240 | pwq_adjust_max_active(pwq); | |
5241 | mutex_unlock(&wq->mutex); | |
5242 | } | |
5243 | ||
5244 | out_unlock: | |
5245 | mutex_unlock(&wq_pool_mutex); | |
5246 | } | |
5247 | #endif /* CONFIG_FREEZER */ | |
5248 | ||
042f7df1 LJ |
5249 | static int workqueue_apply_unbound_cpumask(void) |
5250 | { | |
5251 | LIST_HEAD(ctxs); | |
5252 | int ret = 0; | |
5253 | struct workqueue_struct *wq; | |
5254 | struct apply_wqattrs_ctx *ctx, *n; | |
5255 | ||
5256 | lockdep_assert_held(&wq_pool_mutex); | |
5257 | ||
5258 | list_for_each_entry(wq, &workqueues, list) { | |
5259 | if (!(wq->flags & WQ_UNBOUND)) | |
5260 | continue; | |
5261 | /* creating multiple pwqs breaks ordering guarantee */ | |
5262 | if (wq->flags & __WQ_ORDERED) | |
5263 | continue; | |
5264 | ||
5265 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); | |
5266 | if (!ctx) { | |
5267 | ret = -ENOMEM; | |
5268 | break; | |
5269 | } | |
5270 | ||
5271 | list_add_tail(&ctx->list, &ctxs); | |
5272 | } | |
5273 | ||
5274 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
5275 | if (!ret) | |
5276 | apply_wqattrs_commit(ctx); | |
5277 | apply_wqattrs_cleanup(ctx); | |
5278 | } | |
5279 | ||
5280 | return ret; | |
5281 | } | |
5282 | ||
5283 | /** | |
5284 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
5285 | * @cpumask: the cpumask to set | |
5286 | * | |
5287 | * The low-level workqueues cpumask is a global cpumask that limits | |
5288 | * the affinity of all unbound workqueues. This function check the @cpumask | |
5289 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
5290 | * | |
5291 | * Retun: 0 - Success | |
5292 | * -EINVAL - Invalid @cpumask | |
5293 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
5294 | */ | |
5295 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
5296 | { | |
5297 | int ret = -EINVAL; | |
5298 | cpumask_var_t saved_cpumask; | |
5299 | ||
5300 | if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) | |
5301 | return -ENOMEM; | |
5302 | ||
c98a9805 TS |
5303 | /* |
5304 | * Not excluding isolated cpus on purpose. | |
5305 | * If the user wishes to include them, we allow that. | |
5306 | */ | |
042f7df1 LJ |
5307 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
5308 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 5309 | apply_wqattrs_lock(); |
042f7df1 LJ |
5310 | |
5311 | /* save the old wq_unbound_cpumask. */ | |
5312 | cpumask_copy(saved_cpumask, wq_unbound_cpumask); | |
5313 | ||
5314 | /* update wq_unbound_cpumask at first and apply it to wqs. */ | |
5315 | cpumask_copy(wq_unbound_cpumask, cpumask); | |
5316 | ret = workqueue_apply_unbound_cpumask(); | |
5317 | ||
5318 | /* restore the wq_unbound_cpumask when failed. */ | |
5319 | if (ret < 0) | |
5320 | cpumask_copy(wq_unbound_cpumask, saved_cpumask); | |
5321 | ||
a0111cf6 | 5322 | apply_wqattrs_unlock(); |
042f7df1 | 5323 | } |
042f7df1 LJ |
5324 | |
5325 | free_cpumask_var(saved_cpumask); | |
5326 | return ret; | |
5327 | } | |
5328 | ||
6ba94429 FW |
5329 | #ifdef CONFIG_SYSFS |
5330 | /* | |
5331 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
5332 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
5333 | * following attributes. | |
5334 | * | |
5335 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
5336 | * max_active RW int : maximum number of in-flight work items | |
5337 | * | |
5338 | * Unbound workqueues have the following extra attributes. | |
5339 | * | |
9a19b463 | 5340 | * pool_ids RO int : the associated pool IDs for each node |
6ba94429 FW |
5341 | * nice RW int : nice value of the workers |
5342 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
9a19b463 | 5343 | * numa RW bool : whether enable NUMA affinity |
6ba94429 FW |
5344 | */ |
5345 | struct wq_device { | |
5346 | struct workqueue_struct *wq; | |
5347 | struct device dev; | |
5348 | }; | |
5349 | ||
5350 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
5351 | { | |
5352 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
5353 | ||
5354 | return wq_dev->wq; | |
5355 | } | |
5356 | ||
5357 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
5358 | char *buf) | |
5359 | { | |
5360 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5361 | ||
5362 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
5363 | } | |
5364 | static DEVICE_ATTR_RO(per_cpu); | |
5365 | ||
5366 | static ssize_t max_active_show(struct device *dev, | |
5367 | struct device_attribute *attr, char *buf) | |
5368 | { | |
5369 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5370 | ||
5371 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
5372 | } | |
5373 | ||
5374 | static ssize_t max_active_store(struct device *dev, | |
5375 | struct device_attribute *attr, const char *buf, | |
5376 | size_t count) | |
5377 | { | |
5378 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5379 | int val; | |
5380 | ||
5381 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
5382 | return -EINVAL; | |
5383 | ||
5384 | workqueue_set_max_active(wq, val); | |
5385 | return count; | |
5386 | } | |
5387 | static DEVICE_ATTR_RW(max_active); | |
5388 | ||
5389 | static struct attribute *wq_sysfs_attrs[] = { | |
5390 | &dev_attr_per_cpu.attr, | |
5391 | &dev_attr_max_active.attr, | |
5392 | NULL, | |
5393 | }; | |
5394 | ATTRIBUTE_GROUPS(wq_sysfs); | |
5395 | ||
5396 | static ssize_t wq_pool_ids_show(struct device *dev, | |
5397 | struct device_attribute *attr, char *buf) | |
5398 | { | |
5399 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5400 | const char *delim = ""; | |
5401 | int node, written = 0; | |
5402 | ||
24acfb71 TG |
5403 | get_online_cpus(); |
5404 | rcu_read_lock(); | |
6ba94429 FW |
5405 | for_each_node(node) { |
5406 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
5407 | "%s%d:%d", delim, node, | |
5408 | unbound_pwq_by_node(wq, node)->pool->id); | |
5409 | delim = " "; | |
5410 | } | |
5411 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
24acfb71 TG |
5412 | rcu_read_unlock(); |
5413 | put_online_cpus(); | |
6ba94429 FW |
5414 | |
5415 | return written; | |
5416 | } | |
5417 | ||
5418 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
5419 | char *buf) | |
5420 | { | |
5421 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5422 | int written; | |
5423 | ||
5424 | mutex_lock(&wq->mutex); | |
5425 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5426 | mutex_unlock(&wq->mutex); | |
5427 | ||
5428 | return written; | |
5429 | } | |
5430 | ||
5431 | /* prepare workqueue_attrs for sysfs store operations */ | |
5432 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5433 | { | |
5434 | struct workqueue_attrs *attrs; | |
5435 | ||
899a94fe LJ |
5436 | lockdep_assert_held(&wq_pool_mutex); |
5437 | ||
be69d00d | 5438 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
5439 | if (!attrs) |
5440 | return NULL; | |
5441 | ||
6ba94429 | 5442 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5443 | return attrs; |
5444 | } | |
5445 | ||
5446 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5447 | const char *buf, size_t count) | |
5448 | { | |
5449 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5450 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5451 | int ret = -ENOMEM; |
5452 | ||
5453 | apply_wqattrs_lock(); | |
6ba94429 FW |
5454 | |
5455 | attrs = wq_sysfs_prep_attrs(wq); | |
5456 | if (!attrs) | |
d4d3e257 | 5457 | goto out_unlock; |
6ba94429 FW |
5458 | |
5459 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
5460 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 5461 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
5462 | else |
5463 | ret = -EINVAL; | |
5464 | ||
d4d3e257 LJ |
5465 | out_unlock: |
5466 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5467 | free_workqueue_attrs(attrs); |
5468 | return ret ?: count; | |
5469 | } | |
5470 | ||
5471 | static ssize_t wq_cpumask_show(struct device *dev, | |
5472 | struct device_attribute *attr, char *buf) | |
5473 | { | |
5474 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5475 | int written; | |
5476 | ||
5477 | mutex_lock(&wq->mutex); | |
5478 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
5479 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
5480 | mutex_unlock(&wq->mutex); | |
5481 | return written; | |
5482 | } | |
5483 | ||
5484 | static ssize_t wq_cpumask_store(struct device *dev, | |
5485 | struct device_attribute *attr, | |
5486 | const char *buf, size_t count) | |
5487 | { | |
5488 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5489 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5490 | int ret = -ENOMEM; |
5491 | ||
5492 | apply_wqattrs_lock(); | |
6ba94429 FW |
5493 | |
5494 | attrs = wq_sysfs_prep_attrs(wq); | |
5495 | if (!attrs) | |
d4d3e257 | 5496 | goto out_unlock; |
6ba94429 FW |
5497 | |
5498 | ret = cpumask_parse(buf, attrs->cpumask); | |
5499 | if (!ret) | |
d4d3e257 | 5500 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 5501 | |
d4d3e257 LJ |
5502 | out_unlock: |
5503 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5504 | free_workqueue_attrs(attrs); |
5505 | return ret ?: count; | |
5506 | } | |
5507 | ||
5508 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, | |
5509 | char *buf) | |
5510 | { | |
5511 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5512 | int written; | |
7dbc725e | 5513 | |
6ba94429 FW |
5514 | mutex_lock(&wq->mutex); |
5515 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
5516 | !wq->unbound_attrs->no_numa); | |
5517 | mutex_unlock(&wq->mutex); | |
4c16bd32 | 5518 | |
6ba94429 | 5519 | return written; |
65758202 TH |
5520 | } |
5521 | ||
6ba94429 FW |
5522 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
5523 | const char *buf, size_t count) | |
65758202 | 5524 | { |
6ba94429 FW |
5525 | struct workqueue_struct *wq = dev_to_wq(dev); |
5526 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5527 | int v, ret = -ENOMEM; |
5528 | ||
5529 | apply_wqattrs_lock(); | |
4c16bd32 | 5530 | |
6ba94429 FW |
5531 | attrs = wq_sysfs_prep_attrs(wq); |
5532 | if (!attrs) | |
d4d3e257 | 5533 | goto out_unlock; |
4c16bd32 | 5534 | |
6ba94429 FW |
5535 | ret = -EINVAL; |
5536 | if (sscanf(buf, "%d", &v) == 1) { | |
5537 | attrs->no_numa = !v; | |
d4d3e257 | 5538 | ret = apply_workqueue_attrs_locked(wq, attrs); |
65758202 | 5539 | } |
6ba94429 | 5540 | |
d4d3e257 LJ |
5541 | out_unlock: |
5542 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5543 | free_workqueue_attrs(attrs); |
5544 | return ret ?: count; | |
65758202 TH |
5545 | } |
5546 | ||
6ba94429 FW |
5547 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
5548 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), | |
5549 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), | |
5550 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
5551 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), | |
5552 | __ATTR_NULL, | |
5553 | }; | |
8ccad40d | 5554 | |
6ba94429 FW |
5555 | static struct bus_type wq_subsys = { |
5556 | .name = "workqueue", | |
5557 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
5558 | }; |
5559 | ||
b05a7928 FW |
5560 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
5561 | struct device_attribute *attr, char *buf) | |
5562 | { | |
5563 | int written; | |
5564 | ||
042f7df1 | 5565 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
5566 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
5567 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 5568 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
5569 | |
5570 | return written; | |
5571 | } | |
5572 | ||
042f7df1 LJ |
5573 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
5574 | struct device_attribute *attr, const char *buf, size_t count) | |
5575 | { | |
5576 | cpumask_var_t cpumask; | |
5577 | int ret; | |
5578 | ||
5579 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
5580 | return -ENOMEM; | |
5581 | ||
5582 | ret = cpumask_parse(buf, cpumask); | |
5583 | if (!ret) | |
5584 | ret = workqueue_set_unbound_cpumask(cpumask); | |
5585 | ||
5586 | free_cpumask_var(cpumask); | |
5587 | return ret ? ret : count; | |
5588 | } | |
5589 | ||
b05a7928 | 5590 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
5591 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
5592 | wq_unbound_cpumask_store); | |
b05a7928 | 5593 | |
6ba94429 | 5594 | static int __init wq_sysfs_init(void) |
2d3854a3 | 5595 | { |
b05a7928 FW |
5596 | int err; |
5597 | ||
5598 | err = subsys_virtual_register(&wq_subsys, NULL); | |
5599 | if (err) | |
5600 | return err; | |
5601 | ||
5602 | return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); | |
2d3854a3 | 5603 | } |
6ba94429 | 5604 | core_initcall(wq_sysfs_init); |
2d3854a3 | 5605 | |
6ba94429 | 5606 | static void wq_device_release(struct device *dev) |
2d3854a3 | 5607 | { |
6ba94429 | 5608 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 5609 | |
6ba94429 | 5610 | kfree(wq_dev); |
2d3854a3 | 5611 | } |
a0a1a5fd TH |
5612 | |
5613 | /** | |
6ba94429 FW |
5614 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
5615 | * @wq: the workqueue to register | |
a0a1a5fd | 5616 | * |
6ba94429 FW |
5617 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
5618 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
5619 | * which is the preferred method. | |
a0a1a5fd | 5620 | * |
6ba94429 FW |
5621 | * Workqueue user should use this function directly iff it wants to apply |
5622 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
5623 | * apply_workqueue_attrs() may race against userland updating the | |
5624 | * attributes. | |
5625 | * | |
5626 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 5627 | */ |
6ba94429 | 5628 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 5629 | { |
6ba94429 FW |
5630 | struct wq_device *wq_dev; |
5631 | int ret; | |
a0a1a5fd | 5632 | |
6ba94429 | 5633 | /* |
402dd89d | 5634 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
5635 | * attributes breaks ordering guarantee. Disallow exposing ordered |
5636 | * workqueues. | |
5637 | */ | |
0a94efb5 | 5638 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 5639 | return -EINVAL; |
a0a1a5fd | 5640 | |
6ba94429 FW |
5641 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
5642 | if (!wq_dev) | |
5643 | return -ENOMEM; | |
5bcab335 | 5644 | |
6ba94429 FW |
5645 | wq_dev->wq = wq; |
5646 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 5647 | wq_dev->dev.release = wq_device_release; |
23217b44 | 5648 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 5649 | |
6ba94429 FW |
5650 | /* |
5651 | * unbound_attrs are created separately. Suppress uevent until | |
5652 | * everything is ready. | |
5653 | */ | |
5654 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 5655 | |
6ba94429 FW |
5656 | ret = device_register(&wq_dev->dev); |
5657 | if (ret) { | |
537f4146 | 5658 | put_device(&wq_dev->dev); |
6ba94429 FW |
5659 | wq->wq_dev = NULL; |
5660 | return ret; | |
5661 | } | |
a0a1a5fd | 5662 | |
6ba94429 FW |
5663 | if (wq->flags & WQ_UNBOUND) { |
5664 | struct device_attribute *attr; | |
a0a1a5fd | 5665 | |
6ba94429 FW |
5666 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
5667 | ret = device_create_file(&wq_dev->dev, attr); | |
5668 | if (ret) { | |
5669 | device_unregister(&wq_dev->dev); | |
5670 | wq->wq_dev = NULL; | |
5671 | return ret; | |
a0a1a5fd TH |
5672 | } |
5673 | } | |
5674 | } | |
6ba94429 FW |
5675 | |
5676 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
5677 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
5678 | return 0; | |
a0a1a5fd TH |
5679 | } |
5680 | ||
5681 | /** | |
6ba94429 FW |
5682 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
5683 | * @wq: the workqueue to unregister | |
a0a1a5fd | 5684 | * |
6ba94429 | 5685 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 5686 | */ |
6ba94429 | 5687 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 5688 | { |
6ba94429 | 5689 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 5690 | |
6ba94429 FW |
5691 | if (!wq->wq_dev) |
5692 | return; | |
a0a1a5fd | 5693 | |
6ba94429 FW |
5694 | wq->wq_dev = NULL; |
5695 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 5696 | } |
6ba94429 FW |
5697 | #else /* CONFIG_SYSFS */ |
5698 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
5699 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 5700 | |
82607adc TH |
5701 | /* |
5702 | * Workqueue watchdog. | |
5703 | * | |
5704 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
5705 | * flush dependency, a concurrency managed work item which stays RUNNING | |
5706 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
5707 | * usual warning mechanisms don't trigger and internal workqueue state is | |
5708 | * largely opaque. | |
5709 | * | |
5710 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
5711 | * state if some pools failed to make forward progress for a while where | |
5712 | * forward progress is defined as the first item on ->worklist changing. | |
5713 | * | |
5714 | * This mechanism is controlled through the kernel parameter | |
5715 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
5716 | * corresponding sysfs parameter file. | |
5717 | */ | |
5718 | #ifdef CONFIG_WQ_WATCHDOG | |
5719 | ||
82607adc | 5720 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 5721 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
5722 | |
5723 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
5724 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
5725 | ||
5726 | static void wq_watchdog_reset_touched(void) | |
5727 | { | |
5728 | int cpu; | |
5729 | ||
5730 | wq_watchdog_touched = jiffies; | |
5731 | for_each_possible_cpu(cpu) | |
5732 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5733 | } | |
5734 | ||
5cd79d6a | 5735 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
5736 | { |
5737 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
5738 | bool lockup_detected = false; | |
5739 | struct worker_pool *pool; | |
5740 | int pi; | |
5741 | ||
5742 | if (!thresh) | |
5743 | return; | |
5744 | ||
5745 | rcu_read_lock(); | |
5746 | ||
5747 | for_each_pool(pool, pi) { | |
5748 | unsigned long pool_ts, touched, ts; | |
5749 | ||
5750 | if (list_empty(&pool->worklist)) | |
5751 | continue; | |
5752 | ||
5753 | /* get the latest of pool and touched timestamps */ | |
5754 | pool_ts = READ_ONCE(pool->watchdog_ts); | |
5755 | touched = READ_ONCE(wq_watchdog_touched); | |
5756 | ||
5757 | if (time_after(pool_ts, touched)) | |
5758 | ts = pool_ts; | |
5759 | else | |
5760 | ts = touched; | |
5761 | ||
5762 | if (pool->cpu >= 0) { | |
5763 | unsigned long cpu_touched = | |
5764 | READ_ONCE(per_cpu(wq_watchdog_touched_cpu, | |
5765 | pool->cpu)); | |
5766 | if (time_after(cpu_touched, ts)) | |
5767 | ts = cpu_touched; | |
5768 | } | |
5769 | ||
5770 | /* did we stall? */ | |
5771 | if (time_after(jiffies, ts + thresh)) { | |
5772 | lockup_detected = true; | |
5773 | pr_emerg("BUG: workqueue lockup - pool"); | |
5774 | pr_cont_pool_info(pool); | |
5775 | pr_cont(" stuck for %us!\n", | |
5776 | jiffies_to_msecs(jiffies - pool_ts) / 1000); | |
5777 | } | |
5778 | } | |
5779 | ||
5780 | rcu_read_unlock(); | |
5781 | ||
5782 | if (lockup_detected) | |
5783 | show_workqueue_state(); | |
5784 | ||
5785 | wq_watchdog_reset_touched(); | |
5786 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
5787 | } | |
5788 | ||
cb9d7fd5 | 5789 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
5790 | { |
5791 | if (cpu >= 0) | |
5792 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5793 | else | |
5794 | wq_watchdog_touched = jiffies; | |
5795 | } | |
5796 | ||
5797 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
5798 | { | |
5799 | wq_watchdog_thresh = 0; | |
5800 | del_timer_sync(&wq_watchdog_timer); | |
5801 | ||
5802 | if (thresh) { | |
5803 | wq_watchdog_thresh = thresh; | |
5804 | wq_watchdog_reset_touched(); | |
5805 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
5806 | } | |
5807 | } | |
5808 | ||
5809 | static int wq_watchdog_param_set_thresh(const char *val, | |
5810 | const struct kernel_param *kp) | |
5811 | { | |
5812 | unsigned long thresh; | |
5813 | int ret; | |
5814 | ||
5815 | ret = kstrtoul(val, 0, &thresh); | |
5816 | if (ret) | |
5817 | return ret; | |
5818 | ||
5819 | if (system_wq) | |
5820 | wq_watchdog_set_thresh(thresh); | |
5821 | else | |
5822 | wq_watchdog_thresh = thresh; | |
5823 | ||
5824 | return 0; | |
5825 | } | |
5826 | ||
5827 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
5828 | .set = wq_watchdog_param_set_thresh, | |
5829 | .get = param_get_ulong, | |
5830 | }; | |
5831 | ||
5832 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
5833 | 0644); | |
5834 | ||
5835 | static void wq_watchdog_init(void) | |
5836 | { | |
5cd79d6a | 5837 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
5838 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
5839 | } | |
5840 | ||
5841 | #else /* CONFIG_WQ_WATCHDOG */ | |
5842 | ||
5843 | static inline void wq_watchdog_init(void) { } | |
5844 | ||
5845 | #endif /* CONFIG_WQ_WATCHDOG */ | |
5846 | ||
bce90380 TH |
5847 | static void __init wq_numa_init(void) |
5848 | { | |
5849 | cpumask_var_t *tbl; | |
5850 | int node, cpu; | |
5851 | ||
bce90380 TH |
5852 | if (num_possible_nodes() <= 1) |
5853 | return; | |
5854 | ||
d55262c4 TH |
5855 | if (wq_disable_numa) { |
5856 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
5857 | return; | |
5858 | } | |
5859 | ||
be69d00d | 5860 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); |
4c16bd32 TH |
5861 | BUG_ON(!wq_update_unbound_numa_attrs_buf); |
5862 | ||
bce90380 TH |
5863 | /* |
5864 | * We want masks of possible CPUs of each node which isn't readily | |
5865 | * available. Build one from cpu_to_node() which should have been | |
5866 | * fully initialized by now. | |
5867 | */ | |
6396bb22 | 5868 | tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); |
bce90380 TH |
5869 | BUG_ON(!tbl); |
5870 | ||
5871 | for_each_node(node) | |
5a6024f1 | 5872 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 5873 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
5874 | |
5875 | for_each_possible_cpu(cpu) { | |
5876 | node = cpu_to_node(cpu); | |
5877 | if (WARN_ON(node == NUMA_NO_NODE)) { | |
5878 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
5879 | /* happens iff arch is bonkers, let's just proceed */ | |
5880 | return; | |
5881 | } | |
5882 | cpumask_set_cpu(cpu, tbl[node]); | |
5883 | } | |
5884 | ||
5885 | wq_numa_possible_cpumask = tbl; | |
5886 | wq_numa_enabled = true; | |
5887 | } | |
5888 | ||
3347fa09 TH |
5889 | /** |
5890 | * workqueue_init_early - early init for workqueue subsystem | |
5891 | * | |
5892 | * This is the first half of two-staged workqueue subsystem initialization | |
5893 | * and invoked as soon as the bare basics - memory allocation, cpumasks and | |
5894 | * idr are up. It sets up all the data structures and system workqueues | |
5895 | * and allows early boot code to create workqueues and queue/cancel work | |
5896 | * items. Actual work item execution starts only after kthreads can be | |
5897 | * created and scheduled right before early initcalls. | |
5898 | */ | |
5899 | int __init workqueue_init_early(void) | |
1da177e4 | 5900 | { |
7a4e344c | 5901 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
1bda3f80 | 5902 | int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; |
7a4e344c | 5903 | int i, cpu; |
c34056a3 | 5904 | |
e904e6c2 TH |
5905 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
5906 | ||
b05a7928 | 5907 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
1bda3f80 | 5908 | cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); |
b05a7928 | 5909 | |
e904e6c2 TH |
5910 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
5911 | ||
706026c2 | 5912 | /* initialize CPU pools */ |
29c91e99 | 5913 | for_each_possible_cpu(cpu) { |
4ce62e9e | 5914 | struct worker_pool *pool; |
8b03ae3c | 5915 | |
7a4e344c | 5916 | i = 0; |
f02ae73a | 5917 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 5918 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 5919 | pool->cpu = cpu; |
29c91e99 | 5920 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 5921 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 5922 | pool->node = cpu_to_node(cpu); |
7a4e344c | 5923 | |
9daf9e67 | 5924 | /* alloc pool ID */ |
68e13a67 | 5925 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 5926 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 5927 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 5928 | } |
8b03ae3c TH |
5929 | } |
5930 | ||
8a2b7538 | 5931 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
5932 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
5933 | struct workqueue_attrs *attrs; | |
5934 | ||
be69d00d | 5935 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 5936 | attrs->nice = std_nice[i]; |
29c91e99 | 5937 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
5938 | |
5939 | /* | |
5940 | * An ordered wq should have only one pwq as ordering is | |
5941 | * guaranteed by max_active which is enforced by pwqs. | |
5942 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
5943 | */ | |
be69d00d | 5944 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 TH |
5945 | attrs->nice = std_nice[i]; |
5946 | attrs->no_numa = true; | |
5947 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
5948 | } |
5949 | ||
d320c038 | 5950 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 5951 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 5952 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
5953 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
5954 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
5955 | system_freezable_wq = alloc_workqueue("events_freezable", |
5956 | WQ_FREEZABLE, 0); | |
0668106c VK |
5957 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
5958 | WQ_POWER_EFFICIENT, 0); | |
5959 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
5960 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
5961 | 0); | |
1aabe902 | 5962 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
5963 | !system_unbound_wq || !system_freezable_wq || |
5964 | !system_power_efficient_wq || | |
5965 | !system_freezable_power_efficient_wq); | |
82607adc | 5966 | |
3347fa09 TH |
5967 | return 0; |
5968 | } | |
5969 | ||
5970 | /** | |
5971 | * workqueue_init - bring workqueue subsystem fully online | |
5972 | * | |
5973 | * This is the latter half of two-staged workqueue subsystem initialization | |
5974 | * and invoked as soon as kthreads can be created and scheduled. | |
5975 | * Workqueues have been created and work items queued on them, but there | |
5976 | * are no kworkers executing the work items yet. Populate the worker pools | |
5977 | * with the initial workers and enable future kworker creations. | |
5978 | */ | |
5979 | int __init workqueue_init(void) | |
5980 | { | |
2186d9f9 | 5981 | struct workqueue_struct *wq; |
3347fa09 TH |
5982 | struct worker_pool *pool; |
5983 | int cpu, bkt; | |
5984 | ||
2186d9f9 TH |
5985 | /* |
5986 | * It'd be simpler to initialize NUMA in workqueue_init_early() but | |
5987 | * CPU to node mapping may not be available that early on some | |
5988 | * archs such as power and arm64. As per-cpu pools created | |
5989 | * previously could be missing node hint and unbound pools NUMA | |
5990 | * affinity, fix them up. | |
40c17f75 TH |
5991 | * |
5992 | * Also, while iterating workqueues, create rescuers if requested. | |
2186d9f9 TH |
5993 | */ |
5994 | wq_numa_init(); | |
5995 | ||
5996 | mutex_lock(&wq_pool_mutex); | |
5997 | ||
5998 | for_each_possible_cpu(cpu) { | |
5999 | for_each_cpu_worker_pool(pool, cpu) { | |
6000 | pool->node = cpu_to_node(cpu); | |
6001 | } | |
6002 | } | |
6003 | ||
40c17f75 | 6004 | list_for_each_entry(wq, &workqueues, list) { |
2186d9f9 | 6005 | wq_update_unbound_numa(wq, smp_processor_id(), true); |
40c17f75 TH |
6006 | WARN(init_rescuer(wq), |
6007 | "workqueue: failed to create early rescuer for %s", | |
6008 | wq->name); | |
6009 | } | |
2186d9f9 TH |
6010 | |
6011 | mutex_unlock(&wq_pool_mutex); | |
6012 | ||
3347fa09 TH |
6013 | /* create the initial workers */ |
6014 | for_each_online_cpu(cpu) { | |
6015 | for_each_cpu_worker_pool(pool, cpu) { | |
6016 | pool->flags &= ~POOL_DISASSOCIATED; | |
6017 | BUG_ON(!create_worker(pool)); | |
6018 | } | |
6019 | } | |
6020 | ||
6021 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
6022 | BUG_ON(!create_worker(pool)); | |
6023 | ||
6024 | wq_online = true; | |
82607adc TH |
6025 | wq_watchdog_init(); |
6026 | ||
6ee0578b | 6027 | return 0; |
1da177e4 | 6028 | } |