Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
c54fce6e | 2 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 3 | * |
c54fce6e | 4 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 5 | * |
c54fce6e TH |
6 | * Derived from the taskqueue/keventd code by: |
7 | * David Woodhouse <dwmw2@infradead.org> | |
8 | * Andrew Morton | |
9 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
10 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 11 | * |
c54fce6e | 12 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 13 | * |
c54fce6e TH |
14 | * Copyright (C) 2010 SUSE Linux Products GmbH |
15 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 16 | * |
c54fce6e TH |
17 | * This is the generic async execution mechanism. Work items as are |
18 | * executed in process context. The worker pool is shared and | |
19 | * automatically managed. There is one worker pool for each CPU and | |
20 | * one extra for works which are better served by workers which are | |
21 | * not bound to any specific CPU. | |
22 | * | |
23 | * Please read Documentation/workqueue.txt for details. | |
1da177e4 LT |
24 | */ |
25 | ||
9984de1a | 26 | #include <linux/export.h> |
1da177e4 LT |
27 | #include <linux/kernel.h> |
28 | #include <linux/sched.h> | |
29 | #include <linux/init.h> | |
30 | #include <linux/signal.h> | |
31 | #include <linux/completion.h> | |
32 | #include <linux/workqueue.h> | |
33 | #include <linux/slab.h> | |
34 | #include <linux/cpu.h> | |
35 | #include <linux/notifier.h> | |
36 | #include <linux/kthread.h> | |
1fa44eca | 37 | #include <linux/hardirq.h> |
46934023 | 38 | #include <linux/mempolicy.h> |
341a5958 | 39 | #include <linux/freezer.h> |
d5abe669 PZ |
40 | #include <linux/kallsyms.h> |
41 | #include <linux/debug_locks.h> | |
4e6045f1 | 42 | #include <linux/lockdep.h> |
c34056a3 | 43 | #include <linux/idr.h> |
29c91e99 | 44 | #include <linux/jhash.h> |
42f8570f | 45 | #include <linux/hashtable.h> |
76af4d93 | 46 | #include <linux/rculist.h> |
bce90380 | 47 | #include <linux/nodemask.h> |
4c16bd32 | 48 | #include <linux/moduleparam.h> |
3d1cb205 | 49 | #include <linux/uaccess.h> |
3c2a0909 | 50 | #include <linux/exynos-ss.h> |
e22bee78 | 51 | |
ea138446 | 52 | #include "workqueue_internal.h" |
1da177e4 | 53 | |
c8e55f36 | 54 | enum { |
24647570 TH |
55 | /* |
56 | * worker_pool flags | |
bc2ae0f5 | 57 | * |
24647570 | 58 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
59 | * While associated (!DISASSOCIATED), all workers are bound to the |
60 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
61 | * is in effect. | |
62 | * | |
63 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
64 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 65 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 66 | * |
bc3a1afc TH |
67 | * Note that DISASSOCIATED should be flipped only while holding |
68 | * manager_mutex to avoid changing binding state while | |
24647570 | 69 | * create_worker() is in progress. |
bc2ae0f5 | 70 | */ |
11ebea50 | 71 | POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ |
24647570 | 72 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
35b6bb63 | 73 | POOL_FREEZING = 1 << 3, /* freeze in progress */ |
db7bccf4 | 74 | |
c8e55f36 TH |
75 | /* worker flags */ |
76 | WORKER_STARTED = 1 << 0, /* started */ | |
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 | |
103 | * all cpus. Give -20. | |
104 | */ | |
105 | RESCUER_NICE_LEVEL = -20, | |
3270476a | 106 | HIGHPRI_NICE_LEVEL = -20, |
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 | * |
822d8405 TH |
127 | * MG: pool->manager_mutex and pool->lock protected. Writes require both |
128 | * locks. Reads can happen under either lock. | |
129 | * | |
68e13a67 | 130 | * PL: wq_pool_mutex protected. |
5bcab335 | 131 | * |
68e13a67 | 132 | * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. |
76af4d93 | 133 | * |
3c25a55d LJ |
134 | * WQ: wq->mutex protected. |
135 | * | |
b5927605 | 136 | * WR: wq->mutex protected for writes. Sched-RCU protected for reads. |
2e109a28 TH |
137 | * |
138 | * MD: wq_mayday_lock protected. | |
1da177e4 | 139 | */ |
1da177e4 | 140 | |
2eaebdb3 | 141 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 142 | |
bd7bdd43 | 143 | struct worker_pool { |
d565ed63 | 144 | spinlock_t lock; /* the pool lock */ |
d84ff051 | 145 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 146 | int node; /* I: the associated node ID */ |
9daf9e67 | 147 | int id; /* I: pool ID */ |
11ebea50 | 148 | unsigned int flags; /* X: flags */ |
bd7bdd43 TH |
149 | |
150 | struct list_head worklist; /* L: list of pending works */ | |
151 | int nr_workers; /* L: total number of workers */ | |
ea1abd61 LJ |
152 | |
153 | /* nr_idle includes the ones off idle_list for rebinding */ | |
bd7bdd43 TH |
154 | int nr_idle; /* L: currently idle ones */ |
155 | ||
156 | struct list_head idle_list; /* X: list of idle workers */ | |
157 | struct timer_list idle_timer; /* L: worker idle timeout */ | |
158 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
159 | ||
c5aa87bb | 160 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
161 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
162 | /* L: hash of busy workers */ | |
163 | ||
bc3a1afc | 164 | /* see manage_workers() for details on the two manager mutexes */ |
34a06bd6 | 165 | struct mutex manager_arb; /* manager arbitration */ |
bc3a1afc | 166 | struct mutex manager_mutex; /* manager exclusion */ |
822d8405 | 167 | struct idr worker_idr; /* MG: worker IDs and iteration */ |
e19e397a | 168 | |
7a4e344c | 169 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
170 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
171 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 172 | |
e19e397a TH |
173 | /* |
174 | * The current concurrency level. As it's likely to be accessed | |
175 | * from other CPUs during try_to_wake_up(), put it in a separate | |
176 | * cacheline. | |
177 | */ | |
178 | atomic_t nr_running ____cacheline_aligned_in_smp; | |
29c91e99 TH |
179 | |
180 | /* | |
181 | * Destruction of pool is sched-RCU protected to allow dereferences | |
182 | * from get_work_pool(). | |
183 | */ | |
184 | struct rcu_head rcu; | |
8b03ae3c TH |
185 | } ____cacheline_aligned_in_smp; |
186 | ||
1da177e4 | 187 | /* |
112202d9 TH |
188 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
189 | * of work_struct->data are used for flags and the remaining high bits | |
190 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
191 | * number of flag bits. | |
1da177e4 | 192 | */ |
112202d9 | 193 | struct pool_workqueue { |
bd7bdd43 | 194 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 195 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
196 | int work_color; /* L: current color */ |
197 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 198 | int refcnt; /* L: reference count */ |
73f53c4a TH |
199 | int nr_in_flight[WORK_NR_COLORS]; |
200 | /* L: nr of in_flight works */ | |
1e19ffc6 | 201 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 202 | int max_active; /* L: max active works */ |
1e19ffc6 | 203 | struct list_head delayed_works; /* L: delayed works */ |
3c25a55d | 204 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 205 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 TH |
206 | |
207 | /* | |
208 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
209 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
210 | * itself is also sched-RCU protected so that the first pwq can be | |
b09f4fd3 | 211 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
212 | */ |
213 | struct work_struct unbound_release_work; | |
214 | struct rcu_head rcu; | |
e904e6c2 | 215 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 216 | |
73f53c4a TH |
217 | /* |
218 | * Structure used to wait for workqueue flush. | |
219 | */ | |
220 | struct wq_flusher { | |
3c25a55d LJ |
221 | struct list_head list; /* WQ: list of flushers */ |
222 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
223 | struct completion done; /* flush completion */ |
224 | }; | |
225 | ||
226223ab TH |
226 | struct wq_device; |
227 | ||
1da177e4 | 228 | /* |
c5aa87bb TH |
229 | * The externally visible workqueue. It relays the issued work items to |
230 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
231 | */ |
232 | struct workqueue_struct { | |
3c25a55d | 233 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
68e13a67 | 234 | struct list_head list; /* PL: list of all workqueues */ |
73f53c4a | 235 | |
3c25a55d LJ |
236 | struct mutex mutex; /* protects this wq */ |
237 | int work_color; /* WQ: current work color */ | |
238 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 239 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
240 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
241 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
242 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 243 | |
2e109a28 | 244 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
e22bee78 TH |
245 | struct worker *rescuer; /* I: rescue worker */ |
246 | ||
87fc741e | 247 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 248 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 249 | |
6029a918 | 250 | struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ |
4c16bd32 | 251 | struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ |
6029a918 | 252 | |
226223ab TH |
253 | #ifdef CONFIG_SYSFS |
254 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
255 | #endif | |
4e6045f1 | 256 | #ifdef CONFIG_LOCKDEP |
4690c4ab | 257 | struct lockdep_map lockdep_map; |
4e6045f1 | 258 | #endif |
ecf6881f | 259 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f TH |
260 | |
261 | /* hot fields used during command issue, aligned to cacheline */ | |
262 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
263 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ | |
df2d5ae4 | 264 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ |
1da177e4 LT |
265 | }; |
266 | ||
e904e6c2 TH |
267 | static struct kmem_cache *pwq_cache; |
268 | ||
bce90380 TH |
269 | static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ |
270 | static cpumask_var_t *wq_numa_possible_cpumask; | |
271 | /* possible CPUs of each node */ | |
272 | ||
d55262c4 TH |
273 | static bool wq_disable_numa; |
274 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); | |
275 | ||
3c2a0909 S |
276 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
277 | #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT | |
278 | static bool wq_power_efficient = true; | |
279 | #else | |
280 | static bool wq_power_efficient; | |
281 | #endif | |
282 | ||
283 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); | |
284 | ||
bce90380 TH |
285 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
286 | ||
4c16bd32 TH |
287 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
288 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
289 | ||
68e13a67 | 290 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
2e109a28 | 291 | static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
5bcab335 | 292 | |
68e13a67 LJ |
293 | static LIST_HEAD(workqueues); /* PL: list of all workqueues */ |
294 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ | |
7d19c5ce TH |
295 | |
296 | /* the per-cpu worker pools */ | |
297 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], | |
298 | cpu_worker_pools); | |
299 | ||
68e13a67 | 300 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 301 | |
68e13a67 | 302 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
303 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
304 | ||
c5aa87bb | 305 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
306 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
307 | ||
ced4ac92 TH |
308 | /* I: attributes used when instantiating ordered pools on demand */ |
309 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
310 | ||
d320c038 | 311 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 312 | EXPORT_SYMBOL(system_wq); |
044c782c | 313 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 314 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 315 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 316 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 317 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 318 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 319 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 320 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
3c2a0909 S |
321 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
322 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
323 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
324 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 325 | |
7d19c5ce TH |
326 | static int worker_thread(void *__worker); |
327 | static void copy_workqueue_attrs(struct workqueue_attrs *to, | |
328 | const struct workqueue_attrs *from); | |
329 | ||
97bd2347 TH |
330 | #define CREATE_TRACE_POINTS |
331 | #include <trace/events/workqueue.h> | |
332 | ||
68e13a67 | 333 | #define assert_rcu_or_pool_mutex() \ |
5bcab335 | 334 | rcu_lockdep_assert(rcu_read_lock_sched_held() || \ |
68e13a67 LJ |
335 | lockdep_is_held(&wq_pool_mutex), \ |
336 | "sched RCU or wq_pool_mutex should be held") | |
5bcab335 | 337 | |
b09f4fd3 | 338 | #define assert_rcu_or_wq_mutex(wq) \ |
76af4d93 | 339 | rcu_lockdep_assert(rcu_read_lock_sched_held() || \ |
b5927605 | 340 | lockdep_is_held(&wq->mutex), \ |
b09f4fd3 | 341 | "sched RCU or wq->mutex should be held") |
76af4d93 | 342 | |
822d8405 TH |
343 | #ifdef CONFIG_LOCKDEP |
344 | #define assert_manager_or_pool_lock(pool) \ | |
519e3c11 LJ |
345 | WARN_ONCE(debug_locks && \ |
346 | !lockdep_is_held(&(pool)->manager_mutex) && \ | |
822d8405 TH |
347 | !lockdep_is_held(&(pool)->lock), \ |
348 | "pool->manager_mutex or ->lock should be held") | |
349 | #else | |
350 | #define assert_manager_or_pool_lock(pool) do { } while (0) | |
351 | #endif | |
352 | ||
f02ae73a TH |
353 | #define for_each_cpu_worker_pool(pool, cpu) \ |
354 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
355 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 356 | (pool)++) |
4ce62e9e | 357 | |
17116969 TH |
358 | /** |
359 | * for_each_pool - iterate through all worker_pools in the system | |
360 | * @pool: iteration cursor | |
611c92a0 | 361 | * @pi: integer used for iteration |
fa1b54e6 | 362 | * |
68e13a67 LJ |
363 | * This must be called either with wq_pool_mutex held or sched RCU read |
364 | * locked. If the pool needs to be used beyond the locking in effect, the | |
365 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
366 | * |
367 | * The if/else clause exists only for the lockdep assertion and can be | |
368 | * ignored. | |
17116969 | 369 | */ |
611c92a0 TH |
370 | #define for_each_pool(pool, pi) \ |
371 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 372 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 373 | else |
17116969 | 374 | |
822d8405 TH |
375 | /** |
376 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
377 | * @worker: iteration cursor | |
378 | * @wi: integer used for iteration | |
379 | * @pool: worker_pool to iterate workers of | |
380 | * | |
381 | * This must be called with either @pool->manager_mutex or ->lock held. | |
382 | * | |
383 | * The if/else clause exists only for the lockdep assertion and can be | |
384 | * ignored. | |
385 | */ | |
386 | #define for_each_pool_worker(worker, wi, pool) \ | |
387 | idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ | |
388 | if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ | |
389 | else | |
390 | ||
49e3cf44 TH |
391 | /** |
392 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
393 | * @pwq: iteration cursor | |
394 | * @wq: the target workqueue | |
76af4d93 | 395 | * |
b09f4fd3 | 396 | * This must be called either with wq->mutex held or sched RCU read locked. |
794b18bc TH |
397 | * If the pwq needs to be used beyond the locking in effect, the caller is |
398 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
399 | * |
400 | * The if/else clause exists only for the lockdep assertion and can be | |
401 | * ignored. | |
49e3cf44 TH |
402 | */ |
403 | #define for_each_pwq(pwq, wq) \ | |
76af4d93 | 404 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ |
b09f4fd3 | 405 | if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ |
76af4d93 | 406 | else |
f3421797 | 407 | |
dc186ad7 TG |
408 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
409 | ||
410 | static struct debug_obj_descr work_debug_descr; | |
411 | ||
99777288 SG |
412 | static void *work_debug_hint(void *addr) |
413 | { | |
414 | return ((struct work_struct *) addr)->func; | |
415 | } | |
416 | ||
dc186ad7 TG |
417 | /* |
418 | * fixup_init is called when: | |
419 | * - an active object is initialized | |
420 | */ | |
421 | static int work_fixup_init(void *addr, enum debug_obj_state state) | |
422 | { | |
423 | struct work_struct *work = addr; | |
424 | ||
425 | switch (state) { | |
426 | case ODEBUG_STATE_ACTIVE: | |
427 | cancel_work_sync(work); | |
428 | debug_object_init(work, &work_debug_descr); | |
429 | return 1; | |
430 | default: | |
431 | return 0; | |
432 | } | |
433 | } | |
434 | ||
435 | /* | |
436 | * fixup_activate is called when: | |
437 | * - an active object is activated | |
438 | * - an unknown object is activated (might be a statically initialized object) | |
439 | */ | |
440 | static int work_fixup_activate(void *addr, enum debug_obj_state state) | |
441 | { | |
442 | struct work_struct *work = addr; | |
443 | ||
444 | switch (state) { | |
445 | ||
446 | case ODEBUG_STATE_NOTAVAILABLE: | |
447 | /* | |
448 | * This is not really a fixup. The work struct was | |
449 | * statically initialized. We just make sure that it | |
450 | * is tracked in the object tracker. | |
451 | */ | |
22df02bb | 452 | if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { |
dc186ad7 TG |
453 | debug_object_init(work, &work_debug_descr); |
454 | debug_object_activate(work, &work_debug_descr); | |
455 | return 0; | |
456 | } | |
457 | WARN_ON_ONCE(1); | |
458 | return 0; | |
459 | ||
460 | case ODEBUG_STATE_ACTIVE: | |
461 | WARN_ON(1); | |
462 | ||
463 | default: | |
464 | return 0; | |
465 | } | |
466 | } | |
467 | ||
468 | /* | |
469 | * fixup_free is called when: | |
470 | * - an active object is freed | |
471 | */ | |
472 | static int work_fixup_free(void *addr, enum debug_obj_state state) | |
473 | { | |
474 | struct work_struct *work = addr; | |
475 | ||
476 | switch (state) { | |
477 | case ODEBUG_STATE_ACTIVE: | |
478 | cancel_work_sync(work); | |
479 | debug_object_free(work, &work_debug_descr); | |
480 | return 1; | |
481 | default: | |
482 | return 0; | |
483 | } | |
484 | } | |
485 | ||
486 | static struct debug_obj_descr work_debug_descr = { | |
487 | .name = "work_struct", | |
99777288 | 488 | .debug_hint = work_debug_hint, |
dc186ad7 TG |
489 | .fixup_init = work_fixup_init, |
490 | .fixup_activate = work_fixup_activate, | |
491 | .fixup_free = work_fixup_free, | |
492 | }; | |
493 | ||
494 | static inline void debug_work_activate(struct work_struct *work) | |
495 | { | |
496 | debug_object_activate(work, &work_debug_descr); | |
497 | } | |
498 | ||
499 | static inline void debug_work_deactivate(struct work_struct *work) | |
500 | { | |
501 | debug_object_deactivate(work, &work_debug_descr); | |
502 | } | |
503 | ||
504 | void __init_work(struct work_struct *work, int onstack) | |
505 | { | |
506 | if (onstack) | |
507 | debug_object_init_on_stack(work, &work_debug_descr); | |
508 | else | |
509 | debug_object_init(work, &work_debug_descr); | |
510 | } | |
511 | EXPORT_SYMBOL_GPL(__init_work); | |
512 | ||
513 | void destroy_work_on_stack(struct work_struct *work) | |
514 | { | |
515 | debug_object_free(work, &work_debug_descr); | |
516 | } | |
517 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
518 | ||
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 | ||
9daf9e67 TH |
524 | /* allocate ID and assign it to @pool */ |
525 | static int worker_pool_assign_id(struct worker_pool *pool) | |
526 | { | |
527 | int ret; | |
528 | ||
68e13a67 | 529 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 530 | |
e68035fb | 531 | ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL); |
229641a6 | 532 | if (ret >= 0) { |
e68035fb | 533 | pool->id = ret; |
229641a6 TH |
534 | return 0; |
535 | } | |
fa1b54e6 | 536 | return ret; |
7c3eed5c TH |
537 | } |
538 | ||
df2d5ae4 TH |
539 | /** |
540 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
541 | * @wq: the target workqueue | |
542 | * @node: the node ID | |
543 | * | |
544 | * This must be called either with pwq_lock held or sched RCU read locked. | |
545 | * If the pwq needs to be used beyond the locking in effect, the caller is | |
546 | * responsible for guaranteeing that the pwq stays online. | |
547 | */ | |
548 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
549 | int node) | |
550 | { | |
551 | assert_rcu_or_wq_mutex(wq); | |
552 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); | |
553 | } | |
554 | ||
73f53c4a TH |
555 | static unsigned int work_color_to_flags(int color) |
556 | { | |
557 | return color << WORK_STRUCT_COLOR_SHIFT; | |
558 | } | |
559 | ||
560 | static int get_work_color(struct work_struct *work) | |
561 | { | |
562 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
563 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
564 | } | |
565 | ||
566 | static int work_next_color(int color) | |
567 | { | |
568 | return (color + 1) % WORK_NR_COLORS; | |
569 | } | |
1da177e4 | 570 | |
14441960 | 571 | /* |
112202d9 TH |
572 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
573 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 574 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 575 | * |
112202d9 TH |
576 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
577 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
578 | * work->data. These functions should only be called while the work is |
579 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 580 | * |
112202d9 | 581 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 582 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 583 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 584 | * available only while the work item is queued. |
7a22ad75 | 585 | * |
bbb68dfa TH |
586 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
587 | * canceled. While being canceled, a work item may have its PENDING set | |
588 | * but stay off timer and worklist for arbitrarily long and nobody should | |
589 | * try to steal the PENDING bit. | |
14441960 | 590 | */ |
7a22ad75 TH |
591 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
592 | unsigned long flags) | |
365970a1 | 593 | { |
6183c009 | 594 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
595 | atomic_long_set(&work->data, data | flags | work_static(work)); |
596 | } | |
365970a1 | 597 | |
112202d9 | 598 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
599 | unsigned long extra_flags) |
600 | { | |
112202d9 TH |
601 | set_work_data(work, (unsigned long)pwq, |
602 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
603 | } |
604 | ||
4468a00f LJ |
605 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
606 | int pool_id) | |
607 | { | |
608 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
609 | WORK_STRUCT_PENDING); | |
610 | } | |
611 | ||
7c3eed5c TH |
612 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
613 | int pool_id) | |
7a22ad75 | 614 | { |
23657bb1 TH |
615 | /* |
616 | * The following wmb is paired with the implied mb in | |
617 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
618 | * here are visible to and precede any updates by the next PENDING | |
619 | * owner. | |
620 | */ | |
621 | smp_wmb(); | |
7c3eed5c | 622 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
9a9850c7 RP |
623 | /* |
624 | * The following mb guarantees that previous clear of a PENDING bit | |
625 | * will not be reordered with any speculative LOADS or STORES from | |
626 | * work->current_func, which is executed afterwards. This possible | |
627 | * reordering can lead to a missed execution on attempt to qeueue | |
628 | * the same @work. E.g. consider this case: | |
629 | * | |
630 | * CPU#0 CPU#1 | |
631 | * ---------------------------- -------------------------------- | |
632 | * | |
633 | * 1 STORE event_indicated | |
634 | * 2 queue_work_on() { | |
635 | * 3 test_and_set_bit(PENDING) | |
636 | * 4 } set_..._and_clear_pending() { | |
637 | * 5 set_work_data() # clear bit | |
638 | * 6 smp_mb() | |
639 | * 7 work->current_func() { | |
640 | * 8 LOAD event_indicated | |
641 | * } | |
642 | * | |
643 | * Without an explicit full barrier speculative LOAD on line 8 can | |
644 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
645 | * CPU#0 observes the PENDING bit is still set and new execution of | |
646 | * a @work is not queued in a hope, that CPU#1 will eventually | |
647 | * finish the queued @work. Meanwhile CPU#1 does not see | |
648 | * event_indicated is set, because speculative LOAD was executed | |
649 | * before actual STORE. | |
650 | */ | |
651 | smp_mb(); | |
7a22ad75 | 652 | } |
f756d5e2 | 653 | |
7a22ad75 | 654 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 655 | { |
7c3eed5c TH |
656 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
657 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
658 | } |
659 | ||
112202d9 | 660 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 661 | { |
e120153d | 662 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 663 | |
112202d9 | 664 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
665 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
666 | else | |
667 | return NULL; | |
4d707b9f ON |
668 | } |
669 | ||
7c3eed5c TH |
670 | /** |
671 | * get_work_pool - return the worker_pool a given work was associated with | |
672 | * @work: the work item of interest | |
673 | * | |
674 | * Return the worker_pool @work was last associated with. %NULL if none. | |
fa1b54e6 | 675 | * |
68e13a67 LJ |
676 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
677 | * access under sched-RCU read lock. As such, this function should be | |
678 | * called under wq_pool_mutex or with preemption disabled. | |
fa1b54e6 TH |
679 | * |
680 | * All fields of the returned pool are accessible as long as the above | |
681 | * mentioned locking is in effect. If the returned pool needs to be used | |
682 | * beyond the critical section, the caller is responsible for ensuring the | |
683 | * returned pool is and stays online. | |
7c3eed5c TH |
684 | */ |
685 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 686 | { |
e120153d | 687 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 688 | int pool_id; |
7a22ad75 | 689 | |
68e13a67 | 690 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 691 | |
112202d9 TH |
692 | if (data & WORK_STRUCT_PWQ) |
693 | return ((struct pool_workqueue *) | |
7c3eed5c | 694 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 695 | |
7c3eed5c TH |
696 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
697 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
698 | return NULL; |
699 | ||
fa1b54e6 | 700 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
701 | } |
702 | ||
703 | /** | |
704 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
705 | * @work: the work item of interest | |
706 | * | |
707 | * Return the worker_pool ID @work was last associated with. | |
708 | * %WORK_OFFQ_POOL_NONE if none. | |
709 | */ | |
710 | static int get_work_pool_id(struct work_struct *work) | |
711 | { | |
54d5b7d0 LJ |
712 | unsigned long data = atomic_long_read(&work->data); |
713 | ||
112202d9 TH |
714 | if (data & WORK_STRUCT_PWQ) |
715 | return ((struct pool_workqueue *) | |
54d5b7d0 | 716 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 717 | |
54d5b7d0 | 718 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
719 | } |
720 | ||
bbb68dfa TH |
721 | static void mark_work_canceling(struct work_struct *work) |
722 | { | |
7c3eed5c | 723 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 724 | |
7c3eed5c TH |
725 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
726 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
727 | } |
728 | ||
729 | static bool work_is_canceling(struct work_struct *work) | |
730 | { | |
731 | unsigned long data = atomic_long_read(&work->data); | |
732 | ||
112202d9 | 733 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
734 | } |
735 | ||
e22bee78 | 736 | /* |
3270476a TH |
737 | * Policy functions. These define the policies on how the global worker |
738 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 739 | * they're being called with pool->lock held. |
e22bee78 TH |
740 | */ |
741 | ||
63d95a91 | 742 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 743 | { |
e19e397a | 744 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
745 | } |
746 | ||
4594bf15 | 747 | /* |
e22bee78 TH |
748 | * Need to wake up a worker? Called from anything but currently |
749 | * running workers. | |
974271c4 TH |
750 | * |
751 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 752 | * function will always return %true for unbound pools as long as the |
974271c4 | 753 | * worklist isn't empty. |
4594bf15 | 754 | */ |
63d95a91 | 755 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 756 | { |
63d95a91 | 757 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 758 | } |
4594bf15 | 759 | |
e22bee78 | 760 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 761 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 762 | { |
63d95a91 | 763 | return pool->nr_idle; |
e22bee78 TH |
764 | } |
765 | ||
766 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 767 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 768 | { |
e19e397a TH |
769 | return !list_empty(&pool->worklist) && |
770 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
771 | } |
772 | ||
773 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 774 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 775 | { |
63d95a91 | 776 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 777 | } |
365970a1 | 778 | |
e22bee78 | 779 | /* Do I need to be the manager? */ |
63d95a91 | 780 | static bool need_to_manage_workers(struct worker_pool *pool) |
e22bee78 | 781 | { |
63d95a91 | 782 | return need_to_create_worker(pool) || |
11ebea50 | 783 | (pool->flags & POOL_MANAGE_WORKERS); |
e22bee78 TH |
784 | } |
785 | ||
786 | /* Do we have too many workers and should some go away? */ | |
63d95a91 | 787 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 788 | { |
34a06bd6 | 789 | bool managing = mutex_is_locked(&pool->manager_arb); |
63d95a91 TH |
790 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
791 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 | 792 | |
ea1abd61 LJ |
793 | /* |
794 | * nr_idle and idle_list may disagree if idle rebinding is in | |
795 | * progress. Never return %true if idle_list is empty. | |
796 | */ | |
797 | if (list_empty(&pool->idle_list)) | |
798 | return false; | |
799 | ||
e22bee78 | 800 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; |
365970a1 DH |
801 | } |
802 | ||
4d707b9f | 803 | /* |
e22bee78 TH |
804 | * Wake up functions. |
805 | */ | |
806 | ||
7e11629d | 807 | /* Return the first worker. Safe with preemption disabled */ |
63d95a91 | 808 | static struct worker *first_worker(struct worker_pool *pool) |
7e11629d | 809 | { |
63d95a91 | 810 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
811 | return NULL; |
812 | ||
63d95a91 | 813 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
814 | } |
815 | ||
816 | /** | |
817 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 818 | * @pool: worker pool to wake worker from |
7e11629d | 819 | * |
63d95a91 | 820 | * Wake up the first idle worker of @pool. |
7e11629d TH |
821 | * |
822 | * CONTEXT: | |
d565ed63 | 823 | * spin_lock_irq(pool->lock). |
7e11629d | 824 | */ |
63d95a91 | 825 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 826 | { |
63d95a91 | 827 | struct worker *worker = first_worker(pool); |
7e11629d TH |
828 | |
829 | if (likely(worker)) | |
830 | wake_up_process(worker->task); | |
831 | } | |
832 | ||
d302f017 | 833 | /** |
e22bee78 TH |
834 | * wq_worker_waking_up - a worker is waking up |
835 | * @task: task waking up | |
836 | * @cpu: CPU @task is waking up to | |
837 | * | |
838 | * This function is called during try_to_wake_up() when a worker is | |
839 | * being awoken. | |
840 | * | |
841 | * CONTEXT: | |
842 | * spin_lock_irq(rq->lock) | |
843 | */ | |
d84ff051 | 844 | void wq_worker_waking_up(struct task_struct *task, int cpu) |
e22bee78 TH |
845 | { |
846 | struct worker *worker = kthread_data(task); | |
847 | ||
36576000 | 848 | if (!(worker->flags & WORKER_NOT_RUNNING)) { |
ec22ca5e | 849 | WARN_ON_ONCE(worker->pool->cpu != cpu); |
e19e397a | 850 | atomic_inc(&worker->pool->nr_running); |
36576000 | 851 | } |
e22bee78 TH |
852 | } |
853 | ||
854 | /** | |
855 | * wq_worker_sleeping - a worker is going to sleep | |
856 | * @task: task going to sleep | |
857 | * @cpu: CPU in question, must be the current CPU number | |
858 | * | |
859 | * This function is called during schedule() when a busy worker is | |
860 | * going to sleep. Worker on the same cpu can be woken up by | |
861 | * returning pointer to its task. | |
862 | * | |
863 | * CONTEXT: | |
864 | * spin_lock_irq(rq->lock) | |
865 | * | |
866 | * RETURNS: | |
867 | * Worker task on @cpu to wake up, %NULL if none. | |
868 | */ | |
d84ff051 | 869 | struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) |
e22bee78 TH |
870 | { |
871 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
111c225a | 872 | struct worker_pool *pool; |
e22bee78 | 873 | |
111c225a TH |
874 | /* |
875 | * Rescuers, which may not have all the fields set up like normal | |
876 | * workers, also reach here, let's not access anything before | |
877 | * checking NOT_RUNNING. | |
878 | */ | |
2d64672e | 879 | if (worker->flags & WORKER_NOT_RUNNING) |
e22bee78 TH |
880 | return NULL; |
881 | ||
111c225a | 882 | pool = worker->pool; |
111c225a | 883 | |
e22bee78 | 884 | /* this can only happen on the local cpu */ |
6183c009 TH |
885 | if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) |
886 | return NULL; | |
e22bee78 TH |
887 | |
888 | /* | |
889 | * The counterpart of the following dec_and_test, implied mb, | |
890 | * worklist not empty test sequence is in insert_work(). | |
891 | * Please read comment there. | |
892 | * | |
628c78e7 TH |
893 | * NOT_RUNNING is clear. This means that we're bound to and |
894 | * running on the local cpu w/ rq lock held and preemption | |
895 | * disabled, which in turn means that none else could be | |
d565ed63 | 896 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 897 | * lock is safe. |
e22bee78 | 898 | */ |
e19e397a TH |
899 | if (atomic_dec_and_test(&pool->nr_running) && |
900 | !list_empty(&pool->worklist)) | |
63d95a91 | 901 | to_wakeup = first_worker(pool); |
e22bee78 TH |
902 | return to_wakeup ? to_wakeup->task : NULL; |
903 | } | |
904 | ||
905 | /** | |
906 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 907 | * @worker: self |
d302f017 TH |
908 | * @flags: flags to set |
909 | * @wakeup: wakeup an idle worker if necessary | |
910 | * | |
e22bee78 TH |
911 | * Set @flags in @worker->flags and adjust nr_running accordingly. If |
912 | * nr_running becomes zero and @wakeup is %true, an idle worker is | |
913 | * woken up. | |
d302f017 | 914 | * |
cb444766 | 915 | * CONTEXT: |
d565ed63 | 916 | * spin_lock_irq(pool->lock) |
d302f017 TH |
917 | */ |
918 | static inline void worker_set_flags(struct worker *worker, unsigned int flags, | |
919 | bool wakeup) | |
920 | { | |
bd7bdd43 | 921 | struct worker_pool *pool = worker->pool; |
e22bee78 | 922 | |
cb444766 TH |
923 | WARN_ON_ONCE(worker->task != current); |
924 | ||
e22bee78 TH |
925 | /* |
926 | * If transitioning into NOT_RUNNING, adjust nr_running and | |
927 | * wake up an idle worker as necessary if requested by | |
928 | * @wakeup. | |
929 | */ | |
930 | if ((flags & WORKER_NOT_RUNNING) && | |
931 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
e22bee78 | 932 | if (wakeup) { |
e19e397a | 933 | if (atomic_dec_and_test(&pool->nr_running) && |
bd7bdd43 | 934 | !list_empty(&pool->worklist)) |
63d95a91 | 935 | wake_up_worker(pool); |
e22bee78 | 936 | } else |
e19e397a | 937 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
938 | } |
939 | ||
d302f017 TH |
940 | worker->flags |= flags; |
941 | } | |
942 | ||
943 | /** | |
e22bee78 | 944 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 945 | * @worker: self |
d302f017 TH |
946 | * @flags: flags to clear |
947 | * | |
e22bee78 | 948 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 949 | * |
cb444766 | 950 | * CONTEXT: |
d565ed63 | 951 | * spin_lock_irq(pool->lock) |
d302f017 TH |
952 | */ |
953 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
954 | { | |
63d95a91 | 955 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
956 | unsigned int oflags = worker->flags; |
957 | ||
cb444766 TH |
958 | WARN_ON_ONCE(worker->task != current); |
959 | ||
d302f017 | 960 | worker->flags &= ~flags; |
e22bee78 | 961 | |
42c025f3 TH |
962 | /* |
963 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
964 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
965 | * of multiple flags, not a single flag. | |
966 | */ | |
e22bee78 TH |
967 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
968 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 969 | atomic_inc(&pool->nr_running); |
d302f017 TH |
970 | } |
971 | ||
8cca0eea TH |
972 | /** |
973 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 974 | * @pool: pool of interest |
8cca0eea TH |
975 | * @work: work to find worker for |
976 | * | |
c9e7cf27 TH |
977 | * Find a worker which is executing @work on @pool by searching |
978 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
979 | * to match, its current execution should match the address of @work and |
980 | * its work function. This is to avoid unwanted dependency between | |
981 | * unrelated work executions through a work item being recycled while still | |
982 | * being executed. | |
983 | * | |
984 | * This is a bit tricky. A work item may be freed once its execution | |
985 | * starts and nothing prevents the freed area from being recycled for | |
986 | * another work item. If the same work item address ends up being reused | |
987 | * before the original execution finishes, workqueue will identify the | |
988 | * recycled work item as currently executing and make it wait until the | |
989 | * current execution finishes, introducing an unwanted dependency. | |
990 | * | |
c5aa87bb TH |
991 | * This function checks the work item address and work function to avoid |
992 | * false positives. Note that this isn't complete as one may construct a | |
993 | * work function which can introduce dependency onto itself through a | |
994 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
995 | * foot that badly, there's only so much we can do, and if such deadlock | |
996 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
997 | * |
998 | * CONTEXT: | |
d565ed63 | 999 | * spin_lock_irq(pool->lock). |
8cca0eea TH |
1000 | * |
1001 | * RETURNS: | |
1002 | * Pointer to worker which is executing @work if found, NULL | |
1003 | * otherwise. | |
4d707b9f | 1004 | */ |
c9e7cf27 | 1005 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1006 | struct work_struct *work) |
4d707b9f | 1007 | { |
42f8570f | 1008 | struct worker *worker; |
42f8570f | 1009 | |
b67bfe0d | 1010 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1011 | (unsigned long)work) |
1012 | if (worker->current_work == work && | |
1013 | worker->current_func == work->func) | |
42f8570f SL |
1014 | return worker; |
1015 | ||
1016 | return NULL; | |
4d707b9f ON |
1017 | } |
1018 | ||
bf4ede01 TH |
1019 | /** |
1020 | * move_linked_works - move linked works to a list | |
1021 | * @work: start of series of works to be scheduled | |
1022 | * @head: target list to append @work to | |
1023 | * @nextp: out paramter for nested worklist walking | |
1024 | * | |
1025 | * Schedule linked works starting from @work to @head. Work series to | |
1026 | * be scheduled starts at @work and includes any consecutive work with | |
1027 | * WORK_STRUCT_LINKED set in its predecessor. | |
1028 | * | |
1029 | * If @nextp is not NULL, it's updated to point to the next work of | |
1030 | * the last scheduled work. This allows move_linked_works() to be | |
1031 | * nested inside outer list_for_each_entry_safe(). | |
1032 | * | |
1033 | * CONTEXT: | |
d565ed63 | 1034 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
1035 | */ |
1036 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1037 | struct work_struct **nextp) | |
1038 | { | |
1039 | struct work_struct *n; | |
1040 | ||
1041 | /* | |
1042 | * Linked worklist will always end before the end of the list, | |
1043 | * use NULL for list head. | |
1044 | */ | |
1045 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1046 | list_move_tail(&work->entry, head); | |
1047 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1048 | break; | |
1049 | } | |
1050 | ||
1051 | /* | |
1052 | * If we're already inside safe list traversal and have moved | |
1053 | * multiple works to the scheduled queue, the next position | |
1054 | * needs to be updated. | |
1055 | */ | |
1056 | if (nextp) | |
1057 | *nextp = n; | |
1058 | } | |
1059 | ||
8864b4e5 TH |
1060 | /** |
1061 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1062 | * @pwq: pool_workqueue to get | |
1063 | * | |
1064 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1065 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1066 | */ | |
1067 | static void get_pwq(struct pool_workqueue *pwq) | |
1068 | { | |
1069 | lockdep_assert_held(&pwq->pool->lock); | |
1070 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1071 | pwq->refcnt++; | |
1072 | } | |
1073 | ||
1074 | /** | |
1075 | * put_pwq - put a pool_workqueue reference | |
1076 | * @pwq: pool_workqueue to put | |
1077 | * | |
1078 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1079 | * destruction. The caller should be holding the matching pool->lock. | |
1080 | */ | |
1081 | static void put_pwq(struct pool_workqueue *pwq) | |
1082 | { | |
1083 | lockdep_assert_held(&pwq->pool->lock); | |
1084 | if (likely(--pwq->refcnt)) | |
1085 | return; | |
1086 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1087 | return; | |
1088 | /* | |
1089 | * @pwq can't be released under pool->lock, bounce to | |
1090 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1091 | * pool->lock as this path is taken only for unbound workqueues and | |
1092 | * the release work item is scheduled on a per-cpu workqueue. To | |
1093 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1094 | * subclass of 1 in get_unbound_pool(). | |
1095 | */ | |
1096 | schedule_work(&pwq->unbound_release_work); | |
1097 | } | |
1098 | ||
dce90d47 TH |
1099 | /** |
1100 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1101 | * @pwq: pool_workqueue to put (can be %NULL) | |
1102 | * | |
1103 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1104 | */ | |
1105 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1106 | { | |
1107 | if (pwq) { | |
1108 | /* | |
1109 | * As both pwqs and pools are sched-RCU protected, the | |
1110 | * following lock operations are safe. | |
1111 | */ | |
1112 | spin_lock_irq(&pwq->pool->lock); | |
1113 | put_pwq(pwq); | |
1114 | spin_unlock_irq(&pwq->pool->lock); | |
1115 | } | |
1116 | } | |
1117 | ||
112202d9 | 1118 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 1119 | { |
112202d9 | 1120 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1121 | |
1122 | trace_workqueue_activate_work(work); | |
112202d9 | 1123 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1124 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1125 | pwq->nr_active++; |
bf4ede01 TH |
1126 | } |
1127 | ||
112202d9 | 1128 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1129 | { |
112202d9 | 1130 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1131 | struct work_struct, entry); |
1132 | ||
112202d9 | 1133 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1134 | } |
1135 | ||
bf4ede01 | 1136 | /** |
112202d9 TH |
1137 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1138 | * @pwq: pwq of interest | |
bf4ede01 | 1139 | * @color: color of work which left the queue |
bf4ede01 TH |
1140 | * |
1141 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1142 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1143 | * |
1144 | * CONTEXT: | |
d565ed63 | 1145 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1146 | */ |
112202d9 | 1147 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 | 1148 | { |
8864b4e5 | 1149 | /* uncolored work items don't participate in flushing or nr_active */ |
bf4ede01 | 1150 | if (color == WORK_NO_COLOR) |
8864b4e5 | 1151 | goto out_put; |
bf4ede01 | 1152 | |
112202d9 | 1153 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1154 | |
112202d9 TH |
1155 | pwq->nr_active--; |
1156 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1157 | /* one down, submit a delayed one */ |
112202d9 TH |
1158 | if (pwq->nr_active < pwq->max_active) |
1159 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1160 | } |
1161 | ||
1162 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1163 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1164 | goto out_put; |
bf4ede01 TH |
1165 | |
1166 | /* are there still in-flight works? */ | |
112202d9 | 1167 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1168 | goto out_put; |
bf4ede01 | 1169 | |
112202d9 TH |
1170 | /* this pwq is done, clear flush_color */ |
1171 | pwq->flush_color = -1; | |
bf4ede01 TH |
1172 | |
1173 | /* | |
112202d9 | 1174 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1175 | * will handle the rest. |
1176 | */ | |
112202d9 TH |
1177 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1178 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1179 | out_put: |
1180 | put_pwq(pwq); | |
bf4ede01 TH |
1181 | } |
1182 | ||
36e227d2 | 1183 | /** |
bbb68dfa | 1184 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1185 | * @work: work item to steal |
1186 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1187 | * @flags: place to store irq state |
36e227d2 TH |
1188 | * |
1189 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
1190 | * stable state - idle, on timer or on worklist. Return values are | |
1191 | * | |
1192 | * 1 if @work was pending and we successfully stole PENDING | |
1193 | * 0 if @work was idle and we claimed PENDING | |
1194 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1195 | * -ENOENT if someone else is canceling @work, this state may persist |
1196 | * for arbitrarily long | |
36e227d2 | 1197 | * |
bbb68dfa | 1198 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1199 | * interrupted while holding PENDING and @work off queue, irq must be |
1200 | * disabled on entry. This, combined with delayed_work->timer being | |
1201 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1202 | * |
1203 | * On successful return, >= 0, irq is disabled and the caller is | |
1204 | * responsible for releasing it using local_irq_restore(*@flags). | |
1205 | * | |
e0aecdd8 | 1206 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1207 | */ |
bbb68dfa TH |
1208 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1209 | unsigned long *flags) | |
bf4ede01 | 1210 | { |
d565ed63 | 1211 | struct worker_pool *pool; |
112202d9 | 1212 | struct pool_workqueue *pwq; |
bf4ede01 | 1213 | |
bbb68dfa TH |
1214 | local_irq_save(*flags); |
1215 | ||
36e227d2 TH |
1216 | /* try to steal the timer if it exists */ |
1217 | if (is_dwork) { | |
1218 | struct delayed_work *dwork = to_delayed_work(work); | |
1219 | ||
e0aecdd8 TH |
1220 | /* |
1221 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1222 | * guaranteed that the timer is not queued anywhere and not | |
1223 | * running on the local CPU. | |
1224 | */ | |
36e227d2 TH |
1225 | if (likely(del_timer(&dwork->timer))) |
1226 | return 1; | |
1227 | } | |
1228 | ||
1229 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1230 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1231 | return 0; | |
1232 | ||
1233 | /* | |
1234 | * The queueing is in progress, or it is already queued. Try to | |
1235 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1236 | */ | |
d565ed63 TH |
1237 | pool = get_work_pool(work); |
1238 | if (!pool) | |
bbb68dfa | 1239 | goto fail; |
bf4ede01 | 1240 | |
d565ed63 | 1241 | spin_lock(&pool->lock); |
0b3dae68 | 1242 | /* |
112202d9 TH |
1243 | * work->data is guaranteed to point to pwq only while the work |
1244 | * item is queued on pwq->wq, and both updating work->data to point | |
1245 | * to pwq on queueing and to pool on dequeueing are done under | |
1246 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1247 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1248 | * item is currently queued on that pool. |
1249 | */ | |
112202d9 TH |
1250 | pwq = get_work_pwq(work); |
1251 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1252 | debug_work_deactivate(work); |
1253 | ||
1254 | /* | |
1255 | * A delayed work item cannot be grabbed directly because | |
1256 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1257 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1258 | * management later on and cause stall. Make sure the work |
1259 | * item is activated before grabbing. | |
1260 | */ | |
1261 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1262 | pwq_activate_delayed_work(work); |
16062836 TH |
1263 | |
1264 | list_del_init(&work->entry); | |
112202d9 | 1265 | pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); |
16062836 | 1266 | |
112202d9 | 1267 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1268 | set_work_pool_and_keep_pending(work, pool->id); |
1269 | ||
1270 | spin_unlock(&pool->lock); | |
1271 | return 1; | |
bf4ede01 | 1272 | } |
d565ed63 | 1273 | spin_unlock(&pool->lock); |
bbb68dfa TH |
1274 | fail: |
1275 | local_irq_restore(*flags); | |
1276 | if (work_is_canceling(work)) | |
1277 | return -ENOENT; | |
1278 | cpu_relax(); | |
36e227d2 | 1279 | return -EAGAIN; |
bf4ede01 TH |
1280 | } |
1281 | ||
4690c4ab | 1282 | /** |
706026c2 | 1283 | * insert_work - insert a work into a pool |
112202d9 | 1284 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1285 | * @work: work to insert |
1286 | * @head: insertion point | |
1287 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1288 | * | |
112202d9 | 1289 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1290 | * work_struct flags. |
4690c4ab TH |
1291 | * |
1292 | * CONTEXT: | |
d565ed63 | 1293 | * spin_lock_irq(pool->lock). |
4690c4ab | 1294 | */ |
112202d9 TH |
1295 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1296 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1297 | { |
112202d9 | 1298 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1299 | |
4690c4ab | 1300 | /* we own @work, set data and link */ |
112202d9 | 1301 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1302 | list_add_tail(&work->entry, head); |
8864b4e5 | 1303 | get_pwq(pwq); |
e22bee78 TH |
1304 | |
1305 | /* | |
c5aa87bb TH |
1306 | * Ensure either wq_worker_sleeping() sees the above |
1307 | * list_add_tail() or we see zero nr_running to avoid workers lying | |
1308 | * around lazily while there are works to be processed. | |
e22bee78 TH |
1309 | */ |
1310 | smp_mb(); | |
1311 | ||
63d95a91 TH |
1312 | if (__need_more_worker(pool)) |
1313 | wake_up_worker(pool); | |
b89deed3 ON |
1314 | } |
1315 | ||
c8efcc25 TH |
1316 | /* |
1317 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1318 | * same workqueue. |
c8efcc25 TH |
1319 | */ |
1320 | static bool is_chained_work(struct workqueue_struct *wq) | |
1321 | { | |
8d03ecfe TH |
1322 | struct worker *worker; |
1323 | ||
1324 | worker = current_wq_worker(); | |
1325 | /* | |
1326 | * Return %true iff I'm a worker execuing a work item on @wq. If | |
1327 | * I'm @worker, it's safe to dereference it without locking. | |
1328 | */ | |
112202d9 | 1329 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1330 | } |
1331 | ||
d84ff051 | 1332 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1333 | struct work_struct *work) |
1334 | { | |
112202d9 | 1335 | struct pool_workqueue *pwq; |
c9178087 | 1336 | struct worker_pool *last_pool; |
1e19ffc6 | 1337 | struct list_head *worklist; |
8a2e8e5d | 1338 | unsigned int work_flags; |
b75cac93 | 1339 | unsigned int req_cpu = cpu; |
8930caba TH |
1340 | |
1341 | /* | |
1342 | * While a work item is PENDING && off queue, a task trying to | |
1343 | * steal the PENDING will busy-loop waiting for it to either get | |
1344 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1345 | * happen with IRQ disabled. | |
1346 | */ | |
1347 | WARN_ON_ONCE(!irqs_disabled()); | |
1da177e4 | 1348 | |
dc186ad7 | 1349 | debug_work_activate(work); |
1e19ffc6 | 1350 | |
c8efcc25 | 1351 | /* if dying, only works from the same workqueue are allowed */ |
618b01eb | 1352 | if (unlikely(wq->flags & __WQ_DRAINING) && |
c8efcc25 | 1353 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b | 1354 | return; |
9e8cd2f5 | 1355 | retry: |
df2d5ae4 TH |
1356 | if (req_cpu == WORK_CPU_UNBOUND) |
1357 | cpu = raw_smp_processor_id(); | |
1358 | ||
c9178087 | 1359 | /* pwq which will be used unless @work is executing elsewhere */ |
df2d5ae4 | 1360 | if (!(wq->flags & WQ_UNBOUND)) |
7fb98ea7 | 1361 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
df2d5ae4 TH |
1362 | else |
1363 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); | |
dbf2576e | 1364 | |
c9178087 TH |
1365 | /* |
1366 | * If @work was previously on a different pool, it might still be | |
1367 | * running there, in which case the work needs to be queued on that | |
1368 | * pool to guarantee non-reentrancy. | |
1369 | */ | |
1370 | last_pool = get_work_pool(work); | |
1371 | if (last_pool && last_pool != pwq->pool) { | |
1372 | struct worker *worker; | |
18aa9eff | 1373 | |
c9178087 | 1374 | spin_lock(&last_pool->lock); |
18aa9eff | 1375 | |
c9178087 | 1376 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1377 | |
c9178087 TH |
1378 | if (worker && worker->current_pwq->wq == wq) { |
1379 | pwq = worker->current_pwq; | |
8930caba | 1380 | } else { |
c9178087 TH |
1381 | /* meh... not running there, queue here */ |
1382 | spin_unlock(&last_pool->lock); | |
112202d9 | 1383 | spin_lock(&pwq->pool->lock); |
8930caba | 1384 | } |
f3421797 | 1385 | } else { |
112202d9 | 1386 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1387 | } |
1388 | ||
9e8cd2f5 TH |
1389 | /* |
1390 | * pwq is determined and locked. For unbound pools, we could have | |
1391 | * raced with pwq release and it could already be dead. If its | |
1392 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1393 | * without another pwq replacing it in the numa_pwq_tbl or while |
1394 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1395 | * make forward-progress. |
1396 | */ | |
1397 | if (unlikely(!pwq->refcnt)) { | |
1398 | if (wq->flags & WQ_UNBOUND) { | |
1399 | spin_unlock(&pwq->pool->lock); | |
1400 | cpu_relax(); | |
1401 | goto retry; | |
1402 | } | |
1403 | /* oops */ | |
1404 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1405 | wq->name, cpu); | |
1406 | } | |
1407 | ||
112202d9 TH |
1408 | /* pwq determined, queue */ |
1409 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1410 | |
f5b2552b | 1411 | if (WARN_ON(!list_empty(&work->entry))) { |
112202d9 | 1412 | spin_unlock(&pwq->pool->lock); |
f5b2552b DC |
1413 | return; |
1414 | } | |
1e19ffc6 | 1415 | |
112202d9 TH |
1416 | pwq->nr_in_flight[pwq->work_color]++; |
1417 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1418 | |
112202d9 | 1419 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1420 | trace_workqueue_activate_work(work); |
112202d9 TH |
1421 | pwq->nr_active++; |
1422 | worklist = &pwq->pool->worklist; | |
8a2e8e5d TH |
1423 | } else { |
1424 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1425 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1426 | } |
1e19ffc6 | 1427 | |
112202d9 | 1428 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1429 | |
112202d9 | 1430 | spin_unlock(&pwq->pool->lock); |
1da177e4 LT |
1431 | } |
1432 | ||
0fcb78c2 | 1433 | /** |
c1a220e7 ZR |
1434 | * queue_work_on - queue work on specific cpu |
1435 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1436 | * @wq: workqueue to use |
1437 | * @work: work to queue | |
1438 | * | |
d4283e93 | 1439 | * Returns %false if @work was already on a queue, %true otherwise. |
1da177e4 | 1440 | * |
c1a220e7 ZR |
1441 | * We queue the work to a specific CPU, the caller must ensure it |
1442 | * can't go away. | |
1da177e4 | 1443 | */ |
d4283e93 TH |
1444 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1445 | struct work_struct *work) | |
1da177e4 | 1446 | { |
d4283e93 | 1447 | bool ret = false; |
8930caba | 1448 | unsigned long flags; |
ef1ca236 | 1449 | |
8930caba | 1450 | local_irq_save(flags); |
c1a220e7 | 1451 | |
22df02bb | 1452 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1453 | __queue_work(cpu, wq, work); |
d4283e93 | 1454 | ret = true; |
c1a220e7 | 1455 | } |
ef1ca236 | 1456 | |
8930caba | 1457 | local_irq_restore(flags); |
1da177e4 LT |
1458 | return ret; |
1459 | } | |
ad7b1f84 | 1460 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1461 | |
d8e794df | 1462 | void delayed_work_timer_fn(unsigned long __data) |
1da177e4 | 1463 | { |
52bad64d | 1464 | struct delayed_work *dwork = (struct delayed_work *)__data; |
1da177e4 | 1465 | |
e0aecdd8 | 1466 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1467 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1468 | } |
1438ade5 | 1469 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1470 | |
7beb2edf TH |
1471 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1472 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1473 | { |
7beb2edf TH |
1474 | struct timer_list *timer = &dwork->timer; |
1475 | struct work_struct *work = &dwork->work; | |
7beb2edf TH |
1476 | |
1477 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn || | |
1478 | timer->data != (unsigned long)dwork); | |
fc4b514f TH |
1479 | WARN_ON_ONCE(timer_pending(timer)); |
1480 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1481 | |
8852aac2 TH |
1482 | /* |
1483 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1484 | * both optimization and correctness. The earliest @timer can | |
1485 | * expire is on the closest next tick and delayed_work users depend | |
1486 | * on that there's no such delay when @delay is 0. | |
1487 | */ | |
1488 | if (!delay) { | |
1489 | __queue_work(cpu, wq, &dwork->work); | |
1490 | return; | |
1491 | } | |
1492 | ||
7beb2edf | 1493 | timer_stats_timer_set_start_info(&dwork->timer); |
1da177e4 | 1494 | |
60c057bc | 1495 | dwork->wq = wq; |
1265057f | 1496 | dwork->cpu = cpu; |
7beb2edf TH |
1497 | timer->expires = jiffies + delay; |
1498 | ||
a7b053c7 TH |
1499 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1500 | add_timer_on(timer, cpu); | |
1501 | else | |
1502 | add_timer(timer); | |
1da177e4 LT |
1503 | } |
1504 | ||
0fcb78c2 REB |
1505 | /** |
1506 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1507 | * @cpu: CPU number to execute work on | |
1508 | * @wq: workqueue to use | |
af9997e4 | 1509 | * @dwork: work to queue |
0fcb78c2 REB |
1510 | * @delay: number of jiffies to wait before queueing |
1511 | * | |
715f1300 TH |
1512 | * Returns %false if @work was already on a queue, %true otherwise. If |
1513 | * @delay is zero and @dwork is idle, it will be scheduled for immediate | |
1514 | * execution. | |
0fcb78c2 | 1515 | */ |
d4283e93 TH |
1516 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1517 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1518 | { |
52bad64d | 1519 | struct work_struct *work = &dwork->work; |
d4283e93 | 1520 | bool ret = false; |
8930caba | 1521 | unsigned long flags; |
7a6bc1cd | 1522 | |
8930caba TH |
1523 | /* read the comment in __queue_work() */ |
1524 | local_irq_save(flags); | |
7a6bc1cd | 1525 | |
22df02bb | 1526 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1527 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1528 | ret = true; |
7a6bc1cd | 1529 | } |
8a3e77cc | 1530 | |
8930caba | 1531 | local_irq_restore(flags); |
7a6bc1cd VP |
1532 | return ret; |
1533 | } | |
ad7b1f84 | 1534 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1535 | |
8376fe22 TH |
1536 | /** |
1537 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1538 | * @cpu: CPU number to execute work on | |
1539 | * @wq: workqueue to use | |
1540 | * @dwork: work to queue | |
1541 | * @delay: number of jiffies to wait before queueing | |
1542 | * | |
1543 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1544 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1545 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1546 | * current state. | |
1547 | * | |
1548 | * Returns %false if @dwork was idle and queued, %true if @dwork was | |
1549 | * pending and its timer was modified. | |
1550 | * | |
e0aecdd8 | 1551 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1552 | * See try_to_grab_pending() for details. |
1553 | */ | |
1554 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1555 | struct delayed_work *dwork, unsigned long delay) | |
1556 | { | |
1557 | unsigned long flags; | |
1558 | int ret; | |
c7fc77f7 | 1559 | |
8376fe22 TH |
1560 | do { |
1561 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1562 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1563 | |
8376fe22 TH |
1564 | if (likely(ret >= 0)) { |
1565 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1566 | local_irq_restore(flags); | |
7a6bc1cd | 1567 | } |
8376fe22 TH |
1568 | |
1569 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1570 | return ret; |
1571 | } | |
8376fe22 TH |
1572 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1573 | ||
c8e55f36 TH |
1574 | /** |
1575 | * worker_enter_idle - enter idle state | |
1576 | * @worker: worker which is entering idle state | |
1577 | * | |
1578 | * @worker is entering idle state. Update stats and idle timer if | |
1579 | * necessary. | |
1580 | * | |
1581 | * LOCKING: | |
d565ed63 | 1582 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1583 | */ |
1584 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1585 | { |
bd7bdd43 | 1586 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1587 | |
6183c009 TH |
1588 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1589 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1590 | (worker->hentry.next || worker->hentry.pprev))) | |
1591 | return; | |
c8e55f36 | 1592 | |
cb444766 TH |
1593 | /* can't use worker_set_flags(), also called from start_worker() */ |
1594 | worker->flags |= WORKER_IDLE; | |
bd7bdd43 | 1595 | pool->nr_idle++; |
e22bee78 | 1596 | worker->last_active = jiffies; |
c8e55f36 TH |
1597 | |
1598 | /* idle_list is LIFO */ | |
bd7bdd43 | 1599 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1600 | |
628c78e7 TH |
1601 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1602 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1603 | |
544ecf31 | 1604 | /* |
706026c2 | 1605 | * Sanity check nr_running. Because wq_unbind_fn() releases |
d565ed63 | 1606 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1607 | * nr_running, the warning may trigger spuriously. Check iff |
1608 | * unbind is not in progress. | |
544ecf31 | 1609 | */ |
24647570 | 1610 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1611 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1612 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1613 | } |
1614 | ||
1615 | /** | |
1616 | * worker_leave_idle - leave idle state | |
1617 | * @worker: worker which is leaving idle state | |
1618 | * | |
1619 | * @worker is leaving idle state. Update stats. | |
1620 | * | |
1621 | * LOCKING: | |
d565ed63 | 1622 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1623 | */ |
1624 | static void worker_leave_idle(struct worker *worker) | |
1625 | { | |
bd7bdd43 | 1626 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1627 | |
6183c009 TH |
1628 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1629 | return; | |
d302f017 | 1630 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1631 | pool->nr_idle--; |
c8e55f36 TH |
1632 | list_del_init(&worker->entry); |
1633 | } | |
1634 | ||
e22bee78 | 1635 | /** |
f36dc67b LJ |
1636 | * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it |
1637 | * @pool: target worker_pool | |
1638 | * | |
1639 | * Bind %current to the cpu of @pool if it is associated and lock @pool. | |
e22bee78 TH |
1640 | * |
1641 | * Works which are scheduled while the cpu is online must at least be | |
1642 | * scheduled to a worker which is bound to the cpu so that if they are | |
1643 | * flushed from cpu callbacks while cpu is going down, they are | |
1644 | * guaranteed to execute on the cpu. | |
1645 | * | |
f5faa077 | 1646 | * This function is to be used by unbound workers and rescuers to bind |
e22bee78 TH |
1647 | * themselves to the target cpu and may race with cpu going down or |
1648 | * coming online. kthread_bind() can't be used because it may put the | |
1649 | * worker to already dead cpu and set_cpus_allowed_ptr() can't be used | |
706026c2 | 1650 | * verbatim as it's best effort and blocking and pool may be |
e22bee78 TH |
1651 | * [dis]associated in the meantime. |
1652 | * | |
706026c2 | 1653 | * This function tries set_cpus_allowed() and locks pool and verifies the |
24647570 | 1654 | * binding against %POOL_DISASSOCIATED which is set during |
f2d5a0ee TH |
1655 | * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker |
1656 | * enters idle state or fetches works without dropping lock, it can | |
1657 | * guarantee the scheduling requirement described in the first paragraph. | |
e22bee78 TH |
1658 | * |
1659 | * CONTEXT: | |
d565ed63 | 1660 | * Might sleep. Called without any lock but returns with pool->lock |
e22bee78 TH |
1661 | * held. |
1662 | * | |
1663 | * RETURNS: | |
706026c2 | 1664 | * %true if the associated pool is online (@worker is successfully |
e22bee78 TH |
1665 | * bound), %false if offline. |
1666 | */ | |
f36dc67b | 1667 | static bool worker_maybe_bind_and_lock(struct worker_pool *pool) |
d565ed63 | 1668 | __acquires(&pool->lock) |
e22bee78 | 1669 | { |
e22bee78 | 1670 | while (true) { |
4e6045f1 | 1671 | /* |
e22bee78 TH |
1672 | * The following call may fail, succeed or succeed |
1673 | * without actually migrating the task to the cpu if | |
1674 | * it races with cpu hotunplug operation. Verify | |
24647570 | 1675 | * against POOL_DISASSOCIATED. |
4e6045f1 | 1676 | */ |
24647570 | 1677 | if (!(pool->flags & POOL_DISASSOCIATED)) |
7a4e344c | 1678 | set_cpus_allowed_ptr(current, pool->attrs->cpumask); |
e22bee78 | 1679 | |
d565ed63 | 1680 | spin_lock_irq(&pool->lock); |
24647570 | 1681 | if (pool->flags & POOL_DISASSOCIATED) |
e22bee78 | 1682 | return false; |
f5faa077 | 1683 | if (task_cpu(current) == pool->cpu && |
7a4e344c | 1684 | cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) |
e22bee78 | 1685 | return true; |
d565ed63 | 1686 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1687 | |
5035b20f TH |
1688 | /* |
1689 | * We've raced with CPU hot[un]plug. Give it a breather | |
1690 | * and retry migration. cond_resched() is required here; | |
1691 | * otherwise, we might deadlock against cpu_stop trying to | |
1692 | * bring down the CPU on non-preemptive kernel. | |
1693 | */ | |
e22bee78 | 1694 | cpu_relax(); |
5035b20f | 1695 | cond_resched(); |
e22bee78 TH |
1696 | } |
1697 | } | |
1698 | ||
c34056a3 TH |
1699 | static struct worker *alloc_worker(void) |
1700 | { | |
1701 | struct worker *worker; | |
1702 | ||
1703 | worker = kzalloc(sizeof(*worker), GFP_KERNEL); | |
c8e55f36 TH |
1704 | if (worker) { |
1705 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1706 | INIT_LIST_HEAD(&worker->scheduled); |
e22bee78 TH |
1707 | /* on creation a worker is in !idle && prep state */ |
1708 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1709 | } |
c34056a3 TH |
1710 | return worker; |
1711 | } | |
1712 | ||
1713 | /** | |
1714 | * create_worker - create a new workqueue worker | |
63d95a91 | 1715 | * @pool: pool the new worker will belong to |
c34056a3 | 1716 | * |
63d95a91 | 1717 | * Create a new worker which is bound to @pool. The returned worker |
c34056a3 TH |
1718 | * can be started by calling start_worker() or destroyed using |
1719 | * destroy_worker(). | |
1720 | * | |
1721 | * CONTEXT: | |
1722 | * Might sleep. Does GFP_KERNEL allocations. | |
1723 | * | |
1724 | * RETURNS: | |
1725 | * Pointer to the newly created worker. | |
1726 | */ | |
bc2ae0f5 | 1727 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1728 | { |
c34056a3 | 1729 | struct worker *worker = NULL; |
f3421797 | 1730 | int id = -1; |
e3c916a4 | 1731 | char id_buf[16]; |
c34056a3 | 1732 | |
cd549687 TH |
1733 | lockdep_assert_held(&pool->manager_mutex); |
1734 | ||
822d8405 TH |
1735 | /* |
1736 | * ID is needed to determine kthread name. Allocate ID first | |
1737 | * without installing the pointer. | |
1738 | */ | |
1739 | idr_preload(GFP_KERNEL); | |
d565ed63 | 1740 | spin_lock_irq(&pool->lock); |
822d8405 TH |
1741 | |
1742 | id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); | |
1743 | ||
d565ed63 | 1744 | spin_unlock_irq(&pool->lock); |
822d8405 TH |
1745 | idr_preload_end(); |
1746 | if (id < 0) | |
1747 | goto fail; | |
c34056a3 TH |
1748 | |
1749 | worker = alloc_worker(); | |
1750 | if (!worker) | |
1751 | goto fail; | |
1752 | ||
bd7bdd43 | 1753 | worker->pool = pool; |
c34056a3 TH |
1754 | worker->id = id; |
1755 | ||
29c91e99 | 1756 | if (pool->cpu >= 0) |
e3c916a4 TH |
1757 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1758 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1759 | else |
e3c916a4 TH |
1760 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1761 | ||
f3f90ad4 | 1762 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1763 | "kworker/%s", id_buf); |
c34056a3 TH |
1764 | if (IS_ERR(worker->task)) |
1765 | goto fail; | |
1766 | ||
c5aa87bb TH |
1767 | /* |
1768 | * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any | |
1769 | * online CPUs. It'll be re-applied when any of the CPUs come up. | |
1770 | */ | |
7a4e344c TH |
1771 | set_user_nice(worker->task, pool->attrs->nice); |
1772 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
3270476a | 1773 | |
14a40ffc TH |
1774 | /* prevent userland from meddling with cpumask of workqueue workers */ |
1775 | worker->task->flags |= PF_NO_SETAFFINITY; | |
7a4e344c TH |
1776 | |
1777 | /* | |
1778 | * The caller is responsible for ensuring %POOL_DISASSOCIATED | |
1779 | * remains stable across this function. See the comments above the | |
1780 | * flag definition for details. | |
1781 | */ | |
1782 | if (pool->flags & POOL_DISASSOCIATED) | |
bc2ae0f5 | 1783 | worker->flags |= WORKER_UNBOUND; |
c34056a3 | 1784 | |
822d8405 TH |
1785 | /* successful, commit the pointer to idr */ |
1786 | spin_lock_irq(&pool->lock); | |
1787 | idr_replace(&pool->worker_idr, worker, worker->id); | |
1788 | spin_unlock_irq(&pool->lock); | |
1789 | ||
c34056a3 | 1790 | return worker; |
822d8405 | 1791 | |
c34056a3 TH |
1792 | fail: |
1793 | if (id >= 0) { | |
d565ed63 | 1794 | spin_lock_irq(&pool->lock); |
822d8405 | 1795 | idr_remove(&pool->worker_idr, id); |
d565ed63 | 1796 | spin_unlock_irq(&pool->lock); |
c34056a3 TH |
1797 | } |
1798 | kfree(worker); | |
1799 | return NULL; | |
1800 | } | |
1801 | ||
1802 | /** | |
1803 | * start_worker - start a newly created worker | |
1804 | * @worker: worker to start | |
1805 | * | |
706026c2 | 1806 | * Make the pool aware of @worker and start it. |
c34056a3 TH |
1807 | * |
1808 | * CONTEXT: | |
d565ed63 | 1809 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1810 | */ |
1811 | static void start_worker(struct worker *worker) | |
1812 | { | |
cb444766 | 1813 | worker->flags |= WORKER_STARTED; |
bd7bdd43 | 1814 | worker->pool->nr_workers++; |
c8e55f36 | 1815 | worker_enter_idle(worker); |
c34056a3 TH |
1816 | wake_up_process(worker->task); |
1817 | } | |
1818 | ||
ebf44d16 TH |
1819 | /** |
1820 | * create_and_start_worker - create and start a worker for a pool | |
1821 | * @pool: the target pool | |
1822 | * | |
cd549687 | 1823 | * Grab the managership of @pool and create and start a new worker for it. |
ebf44d16 TH |
1824 | */ |
1825 | static int create_and_start_worker(struct worker_pool *pool) | |
1826 | { | |
1827 | struct worker *worker; | |
1828 | ||
cd549687 TH |
1829 | mutex_lock(&pool->manager_mutex); |
1830 | ||
ebf44d16 TH |
1831 | worker = create_worker(pool); |
1832 | if (worker) { | |
1833 | spin_lock_irq(&pool->lock); | |
1834 | start_worker(worker); | |
1835 | spin_unlock_irq(&pool->lock); | |
1836 | } | |
1837 | ||
cd549687 TH |
1838 | mutex_unlock(&pool->manager_mutex); |
1839 | ||
ebf44d16 TH |
1840 | return worker ? 0 : -ENOMEM; |
1841 | } | |
1842 | ||
c34056a3 TH |
1843 | /** |
1844 | * destroy_worker - destroy a workqueue worker | |
1845 | * @worker: worker to be destroyed | |
1846 | * | |
706026c2 | 1847 | * Destroy @worker and adjust @pool stats accordingly. |
c8e55f36 TH |
1848 | * |
1849 | * CONTEXT: | |
d565ed63 | 1850 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
c34056a3 TH |
1851 | */ |
1852 | static void destroy_worker(struct worker *worker) | |
1853 | { | |
bd7bdd43 | 1854 | struct worker_pool *pool = worker->pool; |
c34056a3 | 1855 | |
cd549687 TH |
1856 | lockdep_assert_held(&pool->manager_mutex); |
1857 | lockdep_assert_held(&pool->lock); | |
1858 | ||
c34056a3 | 1859 | /* sanity check frenzy */ |
6183c009 TH |
1860 | if (WARN_ON(worker->current_work) || |
1861 | WARN_ON(!list_empty(&worker->scheduled))) | |
1862 | return; | |
c34056a3 | 1863 | |
c8e55f36 | 1864 | if (worker->flags & WORKER_STARTED) |
bd7bdd43 | 1865 | pool->nr_workers--; |
c8e55f36 | 1866 | if (worker->flags & WORKER_IDLE) |
bd7bdd43 | 1867 | pool->nr_idle--; |
c8e55f36 | 1868 | |
4403be9e LJ |
1869 | /* |
1870 | * Once WORKER_DIE is set, the kworker may destroy itself at any | |
1871 | * point. Pin to ensure the task stays until we're done with it. | |
1872 | */ | |
1873 | get_task_struct(worker->task); | |
1874 | ||
c8e55f36 | 1875 | list_del_init(&worker->entry); |
cb444766 | 1876 | worker->flags |= WORKER_DIE; |
c8e55f36 | 1877 | |
822d8405 TH |
1878 | idr_remove(&pool->worker_idr, worker->id); |
1879 | ||
d565ed63 | 1880 | spin_unlock_irq(&pool->lock); |
c8e55f36 | 1881 | |
c34056a3 | 1882 | kthread_stop(worker->task); |
4403be9e | 1883 | put_task_struct(worker->task); |
c34056a3 TH |
1884 | kfree(worker); |
1885 | ||
d565ed63 | 1886 | spin_lock_irq(&pool->lock); |
c34056a3 TH |
1887 | } |
1888 | ||
63d95a91 | 1889 | static void idle_worker_timeout(unsigned long __pool) |
e22bee78 | 1890 | { |
63d95a91 | 1891 | struct worker_pool *pool = (void *)__pool; |
e22bee78 | 1892 | |
d565ed63 | 1893 | spin_lock_irq(&pool->lock); |
e22bee78 | 1894 | |
63d95a91 | 1895 | if (too_many_workers(pool)) { |
e22bee78 TH |
1896 | struct worker *worker; |
1897 | unsigned long expires; | |
1898 | ||
1899 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1900 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1901 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1902 | ||
1903 | if (time_before(jiffies, expires)) | |
63d95a91 | 1904 | mod_timer(&pool->idle_timer, expires); |
e22bee78 TH |
1905 | else { |
1906 | /* it's been idle for too long, wake up manager */ | |
11ebea50 | 1907 | pool->flags |= POOL_MANAGE_WORKERS; |
63d95a91 | 1908 | wake_up_worker(pool); |
d5abe669 | 1909 | } |
e22bee78 TH |
1910 | } |
1911 | ||
d565ed63 | 1912 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1913 | } |
d5abe669 | 1914 | |
493a1724 | 1915 | static void send_mayday(struct work_struct *work) |
e22bee78 | 1916 | { |
112202d9 TH |
1917 | struct pool_workqueue *pwq = get_work_pwq(work); |
1918 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 1919 | |
2e109a28 | 1920 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 1921 | |
493008a8 | 1922 | if (!wq->rescuer) |
493a1724 | 1923 | return; |
e22bee78 TH |
1924 | |
1925 | /* mayday mayday mayday */ | |
493a1724 | 1926 | if (list_empty(&pwq->mayday_node)) { |
aac8b37f LJ |
1927 | /* |
1928 | * If @pwq is for an unbound wq, its base ref may be put at | |
1929 | * any time due to an attribute change. Pin @pwq until the | |
1930 | * rescuer is done with it. | |
1931 | */ | |
1932 | get_pwq(pwq); | |
493a1724 | 1933 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 1934 | wake_up_process(wq->rescuer->task); |
493a1724 | 1935 | } |
e22bee78 TH |
1936 | } |
1937 | ||
706026c2 | 1938 | static void pool_mayday_timeout(unsigned long __pool) |
e22bee78 | 1939 | { |
63d95a91 | 1940 | struct worker_pool *pool = (void *)__pool; |
e22bee78 TH |
1941 | struct work_struct *work; |
1942 | ||
2e109a28 | 1943 | spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ |
493a1724 | 1944 | spin_lock(&pool->lock); |
e22bee78 | 1945 | |
63d95a91 | 1946 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
1947 | /* |
1948 | * We've been trying to create a new worker but | |
1949 | * haven't been successful. We might be hitting an | |
1950 | * allocation deadlock. Send distress signals to | |
1951 | * rescuers. | |
1952 | */ | |
63d95a91 | 1953 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 1954 | send_mayday(work); |
1da177e4 | 1955 | } |
e22bee78 | 1956 | |
493a1724 | 1957 | spin_unlock(&pool->lock); |
2e109a28 | 1958 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 1959 | |
63d95a91 | 1960 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
1961 | } |
1962 | ||
e22bee78 TH |
1963 | /** |
1964 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 1965 | * @pool: pool to create a new worker for |
e22bee78 | 1966 | * |
63d95a91 | 1967 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
1968 | * have at least one idle worker on return from this function. If |
1969 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 1970 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
1971 | * possible allocation deadlock. |
1972 | * | |
c5aa87bb TH |
1973 | * On return, need_to_create_worker() is guaranteed to be %false and |
1974 | * may_start_working() %true. | |
e22bee78 TH |
1975 | * |
1976 | * LOCKING: | |
d565ed63 | 1977 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1978 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1979 | * manager. | |
e22bee78 | 1980 | */ |
12f45445 | 1981 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
1982 | __releases(&pool->lock) |
1983 | __acquires(&pool->lock) | |
1da177e4 | 1984 | { |
63d95a91 | 1985 | if (!need_to_create_worker(pool)) |
12f45445 | 1986 | return; |
e22bee78 | 1987 | restart: |
d565ed63 | 1988 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 1989 | |
e22bee78 | 1990 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 1991 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
1992 | |
1993 | while (true) { | |
1994 | struct worker *worker; | |
1995 | ||
bc2ae0f5 | 1996 | worker = create_worker(pool); |
e22bee78 | 1997 | if (worker) { |
63d95a91 | 1998 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1999 | spin_lock_irq(&pool->lock); |
e22bee78 | 2000 | start_worker(worker); |
6183c009 TH |
2001 | if (WARN_ON_ONCE(need_to_create_worker(pool))) |
2002 | goto restart; | |
12f45445 | 2003 | return; |
e22bee78 TH |
2004 | } |
2005 | ||
63d95a91 | 2006 | if (!need_to_create_worker(pool)) |
e22bee78 | 2007 | break; |
1da177e4 | 2008 | |
e22bee78 TH |
2009 | __set_current_state(TASK_INTERRUPTIBLE); |
2010 | schedule_timeout(CREATE_COOLDOWN); | |
9f9c2364 | 2011 | |
63d95a91 | 2012 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2013 | break; |
2014 | } | |
2015 | ||
63d95a91 | 2016 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 2017 | spin_lock_irq(&pool->lock); |
63d95a91 | 2018 | if (need_to_create_worker(pool)) |
e22bee78 | 2019 | goto restart; |
12f45445 | 2020 | return; |
e22bee78 TH |
2021 | } |
2022 | ||
2023 | /** | |
2024 | * maybe_destroy_worker - destroy workers which have been idle for a while | |
63d95a91 | 2025 | * @pool: pool to destroy workers for |
e22bee78 | 2026 | * |
63d95a91 | 2027 | * Destroy @pool workers which have been idle for longer than |
e22bee78 TH |
2028 | * IDLE_WORKER_TIMEOUT. |
2029 | * | |
2030 | * LOCKING: | |
d565ed63 | 2031 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 | 2032 | * multiple times. Called only from manager. |
e22bee78 | 2033 | */ |
12f45445 | 2034 | static void maybe_destroy_workers(struct worker_pool *pool) |
e22bee78 | 2035 | { |
63d95a91 | 2036 | while (too_many_workers(pool)) { |
e22bee78 TH |
2037 | struct worker *worker; |
2038 | unsigned long expires; | |
3af24433 | 2039 | |
63d95a91 | 2040 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 | 2041 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
85f4186a | 2042 | |
e22bee78 | 2043 | if (time_before(jiffies, expires)) { |
63d95a91 | 2044 | mod_timer(&pool->idle_timer, expires); |
3af24433 | 2045 | break; |
e22bee78 | 2046 | } |
1da177e4 | 2047 | |
e22bee78 | 2048 | destroy_worker(worker); |
1da177e4 | 2049 | } |
1e19ffc6 TH |
2050 | } |
2051 | ||
73f53c4a | 2052 | /** |
e22bee78 TH |
2053 | * manage_workers - manage worker pool |
2054 | * @worker: self | |
73f53c4a | 2055 | * |
706026c2 | 2056 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2057 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2058 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2059 | * |
2060 | * The caller can safely start processing works on false return. On | |
2061 | * true return, it's guaranteed that need_to_create_worker() is false | |
2062 | * and may_start_working() is true. | |
73f53c4a TH |
2063 | * |
2064 | * CONTEXT: | |
d565ed63 | 2065 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2066 | * multiple times. Does GFP_KERNEL allocations. |
2067 | * | |
2068 | * RETURNS: | |
12f45445 TH |
2069 | * %false if the pool doesn't need management and the caller can safely |
2070 | * start processing works, %true if management function was performed and | |
2071 | * the conditions that the caller verified before calling the function may | |
2072 | * no longer be true. | |
73f53c4a | 2073 | */ |
e22bee78 | 2074 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2075 | { |
63d95a91 | 2076 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2077 | |
bc3a1afc TH |
2078 | /* |
2079 | * Managership is governed by two mutexes - manager_arb and | |
2080 | * manager_mutex. manager_arb handles arbitration of manager role. | |
2081 | * Anyone who successfully grabs manager_arb wins the arbitration | |
2082 | * and becomes the manager. mutex_trylock() on pool->manager_arb | |
2083 | * failure while holding pool->lock reliably indicates that someone | |
2084 | * else is managing the pool and the worker which failed trylock | |
2085 | * can proceed to executing work items. This means that anyone | |
2086 | * grabbing manager_arb is responsible for actually performing | |
2087 | * manager duties. If manager_arb is grabbed and released without | |
2088 | * actual management, the pool may stall indefinitely. | |
2089 | * | |
2090 | * manager_mutex is used for exclusion of actual management | |
2091 | * operations. The holder of manager_mutex can be sure that none | |
2092 | * of management operations, including creation and destruction of | |
2093 | * workers, won't take place until the mutex is released. Because | |
2094 | * manager_mutex doesn't interfere with manager role arbitration, | |
2095 | * it is guaranteed that the pool's management, while may be | |
2096 | * delayed, won't be disturbed by someone else grabbing | |
2097 | * manager_mutex. | |
2098 | */ | |
34a06bd6 | 2099 | if (!mutex_trylock(&pool->manager_arb)) |
12f45445 | 2100 | return false; |
1e19ffc6 | 2101 | |
ee378aa4 | 2102 | /* |
bc3a1afc TH |
2103 | * With manager arbitration won, manager_mutex would be free in |
2104 | * most cases. trylock first without dropping @pool->lock. | |
ee378aa4 | 2105 | */ |
bc3a1afc | 2106 | if (unlikely(!mutex_trylock(&pool->manager_mutex))) { |
d565ed63 | 2107 | spin_unlock_irq(&pool->lock); |
bc3a1afc | 2108 | mutex_lock(&pool->manager_mutex); |
8f174b11 | 2109 | spin_lock_irq(&pool->lock); |
ee378aa4 | 2110 | } |
73f53c4a | 2111 | |
11ebea50 | 2112 | pool->flags &= ~POOL_MANAGE_WORKERS; |
73f53c4a TH |
2113 | |
2114 | /* | |
e22bee78 TH |
2115 | * Destroy and then create so that may_start_working() is true |
2116 | * on return. | |
73f53c4a | 2117 | */ |
12f45445 TH |
2118 | maybe_destroy_workers(pool); |
2119 | maybe_create_worker(pool); | |
e22bee78 | 2120 | |
bc3a1afc | 2121 | mutex_unlock(&pool->manager_mutex); |
34a06bd6 | 2122 | mutex_unlock(&pool->manager_arb); |
12f45445 | 2123 | return true; |
73f53c4a TH |
2124 | } |
2125 | ||
a62428c0 TH |
2126 | /** |
2127 | * process_one_work - process single work | |
c34056a3 | 2128 | * @worker: self |
a62428c0 TH |
2129 | * @work: work to process |
2130 | * | |
2131 | * Process @work. This function contains all the logics necessary to | |
2132 | * process a single work including synchronization against and | |
2133 | * interaction with other workers on the same cpu, queueing and | |
2134 | * flushing. As long as context requirement is met, any worker can | |
2135 | * call this function to process a work. | |
2136 | * | |
2137 | * CONTEXT: | |
d565ed63 | 2138 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2139 | */ |
c34056a3 | 2140 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2141 | __releases(&pool->lock) |
2142 | __acquires(&pool->lock) | |
a62428c0 | 2143 | { |
112202d9 | 2144 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2145 | struct worker_pool *pool = worker->pool; |
112202d9 | 2146 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2147 | int work_color; |
7e11629d | 2148 | struct worker *collision; |
a62428c0 TH |
2149 | #ifdef CONFIG_LOCKDEP |
2150 | /* | |
2151 | * It is permissible to free the struct work_struct from | |
2152 | * inside the function that is called from it, this we need to | |
2153 | * take into account for lockdep too. To avoid bogus "held | |
2154 | * lock freed" warnings as well as problems when looking into | |
2155 | * work->lockdep_map, make a copy and use that here. | |
2156 | */ | |
4d82a1de PZ |
2157 | struct lockdep_map lockdep_map; |
2158 | ||
2159 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2160 | #endif |
6fec10a1 TH |
2161 | /* |
2162 | * Ensure we're on the correct CPU. DISASSOCIATED test is | |
2163 | * necessary to avoid spurious warnings from rescuers servicing the | |
24647570 | 2164 | * unbound or a disassociated pool. |
6fec10a1 | 2165 | */ |
5f7dabfd | 2166 | WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && |
24647570 | 2167 | !(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2168 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2169 | |
7e11629d TH |
2170 | /* |
2171 | * A single work shouldn't be executed concurrently by | |
2172 | * multiple workers on a single cpu. Check whether anyone is | |
2173 | * already processing the work. If so, defer the work to the | |
2174 | * currently executing one. | |
2175 | */ | |
c9e7cf27 | 2176 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2177 | if (unlikely(collision)) { |
2178 | move_linked_works(work, &collision->scheduled, NULL); | |
2179 | return; | |
2180 | } | |
2181 | ||
8930caba | 2182 | /* claim and dequeue */ |
a62428c0 | 2183 | debug_work_deactivate(work); |
c9e7cf27 | 2184 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2185 | worker->current_work = work; |
a2c1c57b | 2186 | worker->current_func = work->func; |
112202d9 | 2187 | worker->current_pwq = pwq; |
73f53c4a | 2188 | work_color = get_work_color(work); |
7a22ad75 | 2189 | |
a62428c0 TH |
2190 | list_del_init(&work->entry); |
2191 | ||
fb0e7beb TH |
2192 | /* |
2193 | * CPU intensive works don't participate in concurrency | |
2194 | * management. They're the scheduler's responsibility. | |
2195 | */ | |
2196 | if (unlikely(cpu_intensive)) | |
2197 | worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); | |
2198 | ||
974271c4 | 2199 | /* |
d565ed63 | 2200 | * Unbound pool isn't concurrency managed and work items should be |
974271c4 TH |
2201 | * executed ASAP. Wake up another worker if necessary. |
2202 | */ | |
63d95a91 TH |
2203 | if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) |
2204 | wake_up_worker(pool); | |
974271c4 | 2205 | |
8930caba | 2206 | /* |
7c3eed5c | 2207 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2208 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2209 | * PENDING and queued state changes happen together while IRQ is |
2210 | * disabled. | |
8930caba | 2211 | */ |
7c3eed5c | 2212 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2213 | |
d565ed63 | 2214 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2215 | |
112202d9 | 2216 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
a62428c0 | 2217 | lock_map_acquire(&lockdep_map); |
e36c886a | 2218 | trace_workqueue_execute_start(work); |
3c2a0909 | 2219 | exynos_ss_work(worker, work, worker->current_func, ESS_FLAG_IN); |
a2c1c57b | 2220 | worker->current_func(work); |
3c2a0909 | 2221 | exynos_ss_work(worker, work, worker->current_func, ESS_FLAG_OUT); |
e36c886a AV |
2222 | /* |
2223 | * While we must be careful to not use "work" after this, the trace | |
2224 | * point will only record its address. | |
2225 | */ | |
2226 | trace_workqueue_execute_end(work); | |
a62428c0 | 2227 | lock_map_release(&lockdep_map); |
112202d9 | 2228 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2229 | |
2230 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c VI |
2231 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
2232 | " last function: %pf\n", | |
a2c1c57b TH |
2233 | current->comm, preempt_count(), task_pid_nr(current), |
2234 | worker->current_func); | |
a62428c0 TH |
2235 | debug_show_held_locks(current); |
2236 | dump_stack(); | |
2237 | } | |
2238 | ||
6ff96f73 TH |
2239 | /* |
2240 | * The following prevents a kworker from hogging CPU on !PREEMPT | |
2241 | * kernels, where a requeueing work item waiting for something to | |
2242 | * happen could deadlock with stop_machine as such work item could | |
2243 | * indefinitely requeue itself while all other CPUs are trapped in | |
2244 | * stop_machine. | |
2245 | */ | |
2246 | cond_resched(); | |
2247 | ||
d565ed63 | 2248 | spin_lock_irq(&pool->lock); |
a62428c0 | 2249 | |
fb0e7beb TH |
2250 | /* clear cpu intensive status */ |
2251 | if (unlikely(cpu_intensive)) | |
2252 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2253 | ||
a62428c0 | 2254 | /* we're done with it, release */ |
42f8570f | 2255 | hash_del(&worker->hentry); |
c34056a3 | 2256 | worker->current_work = NULL; |
a2c1c57b | 2257 | worker->current_func = NULL; |
112202d9 | 2258 | worker->current_pwq = NULL; |
3d1cb205 | 2259 | worker->desc_valid = false; |
112202d9 | 2260 | pwq_dec_nr_in_flight(pwq, work_color); |
a62428c0 TH |
2261 | } |
2262 | ||
affee4b2 TH |
2263 | /** |
2264 | * process_scheduled_works - process scheduled works | |
2265 | * @worker: self | |
2266 | * | |
2267 | * Process all scheduled works. Please note that the scheduled list | |
2268 | * may change while processing a work, so this function repeatedly | |
2269 | * fetches a work from the top and executes it. | |
2270 | * | |
2271 | * CONTEXT: | |
d565ed63 | 2272 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2273 | * multiple times. |
2274 | */ | |
2275 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2276 | { |
affee4b2 TH |
2277 | while (!list_empty(&worker->scheduled)) { |
2278 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2279 | struct work_struct, entry); |
c34056a3 | 2280 | process_one_work(worker, work); |
1da177e4 | 2281 | } |
1da177e4 LT |
2282 | } |
2283 | ||
4690c4ab TH |
2284 | /** |
2285 | * worker_thread - the worker thread function | |
c34056a3 | 2286 | * @__worker: self |
4690c4ab | 2287 | * |
c5aa87bb TH |
2288 | * The worker thread function. All workers belong to a worker_pool - |
2289 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2290 | * work items regardless of their specific target workqueue. The only | |
2291 | * exception is work items which belong to workqueues with a rescuer which | |
2292 | * will be explained in rescuer_thread(). | |
4690c4ab | 2293 | */ |
c34056a3 | 2294 | static int worker_thread(void *__worker) |
1da177e4 | 2295 | { |
c34056a3 | 2296 | struct worker *worker = __worker; |
bd7bdd43 | 2297 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2298 | |
e22bee78 TH |
2299 | /* tell the scheduler that this is a workqueue worker */ |
2300 | worker->task->flags |= PF_WQ_WORKER; | |
c8e55f36 | 2301 | woke_up: |
d565ed63 | 2302 | spin_lock_irq(&pool->lock); |
1da177e4 | 2303 | |
a9ab775b TH |
2304 | /* am I supposed to die? */ |
2305 | if (unlikely(worker->flags & WORKER_DIE)) { | |
d565ed63 | 2306 | spin_unlock_irq(&pool->lock); |
a9ab775b TH |
2307 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
2308 | worker->task->flags &= ~PF_WQ_WORKER; | |
2309 | return 0; | |
c8e55f36 | 2310 | } |
affee4b2 | 2311 | |
c8e55f36 | 2312 | worker_leave_idle(worker); |
db7bccf4 | 2313 | recheck: |
e22bee78 | 2314 | /* no more worker necessary? */ |
63d95a91 | 2315 | if (!need_more_worker(pool)) |
e22bee78 TH |
2316 | goto sleep; |
2317 | ||
2318 | /* do we need to manage? */ | |
63d95a91 | 2319 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2320 | goto recheck; |
2321 | ||
c8e55f36 TH |
2322 | /* |
2323 | * ->scheduled list can only be filled while a worker is | |
2324 | * preparing to process a work or actually processing it. | |
2325 | * Make sure nobody diddled with it while I was sleeping. | |
2326 | */ | |
6183c009 | 2327 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2328 | |
e22bee78 | 2329 | /* |
a9ab775b TH |
2330 | * Finish PREP stage. We're guaranteed to have at least one idle |
2331 | * worker or that someone else has already assumed the manager | |
2332 | * role. This is where @worker starts participating in concurrency | |
2333 | * management if applicable and concurrency management is restored | |
2334 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2335 | */ |
a9ab775b | 2336 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2337 | |
2338 | do { | |
c8e55f36 | 2339 | struct work_struct *work = |
bd7bdd43 | 2340 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2341 | struct work_struct, entry); |
2342 | ||
2343 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { | |
2344 | /* optimization path, not strictly necessary */ | |
2345 | process_one_work(worker, work); | |
2346 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2347 | process_scheduled_works(worker); |
c8e55f36 TH |
2348 | } else { |
2349 | move_linked_works(work, &worker->scheduled, NULL); | |
2350 | process_scheduled_works(worker); | |
affee4b2 | 2351 | } |
63d95a91 | 2352 | } while (keep_working(pool)); |
e22bee78 TH |
2353 | |
2354 | worker_set_flags(worker, WORKER_PREP, false); | |
d313dd85 | 2355 | sleep: |
63d95a91 | 2356 | if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) |
e22bee78 | 2357 | goto recheck; |
d313dd85 | 2358 | |
c8e55f36 | 2359 | /* |
d565ed63 TH |
2360 | * pool->lock is held and there's no work to process and no need to |
2361 | * manage, sleep. Workers are woken up only while holding | |
2362 | * pool->lock or from local cpu, so setting the current state | |
2363 | * before releasing pool->lock is enough to prevent losing any | |
2364 | * event. | |
c8e55f36 TH |
2365 | */ |
2366 | worker_enter_idle(worker); | |
2367 | __set_current_state(TASK_INTERRUPTIBLE); | |
d565ed63 | 2368 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2369 | schedule(); |
2370 | goto woke_up; | |
1da177e4 LT |
2371 | } |
2372 | ||
e22bee78 TH |
2373 | /** |
2374 | * rescuer_thread - the rescuer thread function | |
111c225a | 2375 | * @__rescuer: self |
e22bee78 TH |
2376 | * |
2377 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2378 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2379 | * |
706026c2 | 2380 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2381 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2382 | * developing into deadlock if some works currently on the same queue | |
2383 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2384 | * the problem rescuer solves. | |
2385 | * | |
706026c2 TH |
2386 | * When such condition is possible, the pool summons rescuers of all |
2387 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2388 | * those works so that forward progress can be guaranteed. |
2389 | * | |
2390 | * This should happen rarely. | |
2391 | */ | |
111c225a | 2392 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2393 | { |
111c225a TH |
2394 | struct worker *rescuer = __rescuer; |
2395 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2396 | struct list_head *scheduled = &rescuer->scheduled; |
f56fb0d4 | 2397 | bool should_stop; |
e22bee78 TH |
2398 | |
2399 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2400 | |
2401 | /* | |
2402 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2403 | * doesn't participate in concurrency management. | |
2404 | */ | |
2405 | rescuer->task->flags |= PF_WQ_WORKER; | |
e22bee78 TH |
2406 | repeat: |
2407 | set_current_state(TASK_INTERRUPTIBLE); | |
2408 | ||
f56fb0d4 LJ |
2409 | /* |
2410 | * By the time the rescuer is requested to stop, the workqueue | |
2411 | * shouldn't have any work pending, but @wq->maydays may still have | |
2412 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2413 | * all the work items before the rescuer got to them. Go through | |
2414 | * @wq->maydays processing before acting on should_stop so that the | |
2415 | * list is always empty on exit. | |
2416 | */ | |
2417 | should_stop = kthread_should_stop(); | |
e22bee78 | 2418 | |
493a1724 | 2419 | /* see whether any pwq is asking for help */ |
2e109a28 | 2420 | spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2421 | |
2422 | while (!list_empty(&wq->maydays)) { | |
2423 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2424 | struct pool_workqueue, mayday_node); | |
112202d9 | 2425 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
2426 | struct work_struct *work, *n; |
2427 | ||
2428 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2429 | list_del_init(&pwq->mayday_node); |
2430 | ||
2e109a28 | 2431 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 TH |
2432 | |
2433 | /* migrate to the target cpu if possible */ | |
f36dc67b | 2434 | worker_maybe_bind_and_lock(pool); |
b3104104 | 2435 | rescuer->pool = pool; |
e22bee78 TH |
2436 | |
2437 | /* | |
2438 | * Slurp in all works issued via this workqueue and | |
2439 | * process'em. | |
2440 | */ | |
6183c009 | 2441 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
bd7bdd43 | 2442 | list_for_each_entry_safe(work, n, &pool->worklist, entry) |
112202d9 | 2443 | if (get_work_pwq(work) == pwq) |
e22bee78 TH |
2444 | move_linked_works(work, scheduled, &n); |
2445 | ||
2446 | process_scheduled_works(rescuer); | |
7576958a | 2447 | |
aac8b37f LJ |
2448 | /* |
2449 | * Put the reference grabbed by send_mayday(). @pool won't | |
2450 | * go away while we're holding its lock. | |
2451 | */ | |
2452 | put_pwq(pwq); | |
2453 | ||
7576958a | 2454 | /* |
d565ed63 | 2455 | * Leave this pool. If keep_working() is %true, notify a |
7576958a TH |
2456 | * regular worker; otherwise, we end up with 0 concurrency |
2457 | * and stalling the execution. | |
2458 | */ | |
63d95a91 TH |
2459 | if (keep_working(pool)) |
2460 | wake_up_worker(pool); | |
7576958a | 2461 | |
b3104104 | 2462 | rescuer->pool = NULL; |
493a1724 | 2463 | spin_unlock(&pool->lock); |
2e109a28 | 2464 | spin_lock(&wq_mayday_lock); |
e22bee78 TH |
2465 | } |
2466 | ||
2e109a28 | 2467 | spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2468 | |
f56fb0d4 LJ |
2469 | if (should_stop) { |
2470 | __set_current_state(TASK_RUNNING); | |
2471 | rescuer->task->flags &= ~PF_WQ_WORKER; | |
2472 | return 0; | |
2473 | } | |
2474 | ||
111c225a TH |
2475 | /* rescuers should never participate in concurrency management */ |
2476 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2477 | schedule(); |
2478 | goto repeat; | |
1da177e4 LT |
2479 | } |
2480 | ||
fc2e4d70 ON |
2481 | struct wq_barrier { |
2482 | struct work_struct work; | |
2483 | struct completion done; | |
2484 | }; | |
2485 | ||
2486 | static void wq_barrier_func(struct work_struct *work) | |
2487 | { | |
2488 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2489 | complete(&barr->done); | |
2490 | } | |
2491 | ||
4690c4ab TH |
2492 | /** |
2493 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2494 | * @pwq: pwq to insert barrier into |
4690c4ab | 2495 | * @barr: wq_barrier to insert |
affee4b2 TH |
2496 | * @target: target work to attach @barr to |
2497 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2498 | * |
affee4b2 TH |
2499 | * @barr is linked to @target such that @barr is completed only after |
2500 | * @target finishes execution. Please note that the ordering | |
2501 | * guarantee is observed only with respect to @target and on the local | |
2502 | * cpu. | |
2503 | * | |
2504 | * Currently, a queued barrier can't be canceled. This is because | |
2505 | * try_to_grab_pending() can't determine whether the work to be | |
2506 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2507 | * flag of the previous work while there must be a valid next work | |
2508 | * after a work with LINKED flag set. | |
2509 | * | |
2510 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2511 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2512 | * |
2513 | * CONTEXT: | |
d565ed63 | 2514 | * spin_lock_irq(pool->lock). |
4690c4ab | 2515 | */ |
112202d9 | 2516 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2517 | struct wq_barrier *barr, |
2518 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2519 | { |
affee4b2 TH |
2520 | struct list_head *head; |
2521 | unsigned int linked = 0; | |
2522 | ||
dc186ad7 | 2523 | /* |
d565ed63 | 2524 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2525 | * as we know for sure that this will not trigger any of the |
2526 | * checks and call back into the fixup functions where we | |
2527 | * might deadlock. | |
2528 | */ | |
ca1cab37 | 2529 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2530 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
fc2e4d70 | 2531 | init_completion(&barr->done); |
83c22520 | 2532 | |
affee4b2 TH |
2533 | /* |
2534 | * If @target is currently being executed, schedule the | |
2535 | * barrier to the worker; otherwise, put it after @target. | |
2536 | */ | |
2537 | if (worker) | |
2538 | head = worker->scheduled.next; | |
2539 | else { | |
2540 | unsigned long *bits = work_data_bits(target); | |
2541 | ||
2542 | head = target->entry.next; | |
2543 | /* there can already be other linked works, inherit and set */ | |
2544 | linked = *bits & WORK_STRUCT_LINKED; | |
2545 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2546 | } | |
2547 | ||
dc186ad7 | 2548 | debug_work_activate(&barr->work); |
112202d9 | 2549 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2550 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2551 | } |
2552 | ||
73f53c4a | 2553 | /** |
112202d9 | 2554 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2555 | * @wq: workqueue being flushed |
2556 | * @flush_color: new flush color, < 0 for no-op | |
2557 | * @work_color: new work color, < 0 for no-op | |
2558 | * | |
112202d9 | 2559 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2560 | * |
112202d9 TH |
2561 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2562 | * -1. If no pwq has in-flight commands at the specified color, all | |
2563 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2564 | * has in flight commands, its pwq->flush_color is set to | |
2565 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2566 | * wakeup logic is armed and %true is returned. |
2567 | * | |
2568 | * The caller should have initialized @wq->first_flusher prior to | |
2569 | * calling this function with non-negative @flush_color. If | |
2570 | * @flush_color is negative, no flush color update is done and %false | |
2571 | * is returned. | |
2572 | * | |
112202d9 | 2573 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2574 | * work_color which is previous to @work_color and all will be |
2575 | * advanced to @work_color. | |
2576 | * | |
2577 | * CONTEXT: | |
3c25a55d | 2578 | * mutex_lock(wq->mutex). |
73f53c4a TH |
2579 | * |
2580 | * RETURNS: | |
2581 | * %true if @flush_color >= 0 and there's something to flush. %false | |
2582 | * otherwise. | |
2583 | */ | |
112202d9 | 2584 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2585 | int flush_color, int work_color) |
1da177e4 | 2586 | { |
73f53c4a | 2587 | bool wait = false; |
49e3cf44 | 2588 | struct pool_workqueue *pwq; |
1da177e4 | 2589 | |
73f53c4a | 2590 | if (flush_color >= 0) { |
6183c009 | 2591 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2592 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2593 | } |
2355b70f | 2594 | |
49e3cf44 | 2595 | for_each_pwq(pwq, wq) { |
112202d9 | 2596 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2597 | |
b09f4fd3 | 2598 | spin_lock_irq(&pool->lock); |
83c22520 | 2599 | |
73f53c4a | 2600 | if (flush_color >= 0) { |
6183c009 | 2601 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2602 | |
112202d9 TH |
2603 | if (pwq->nr_in_flight[flush_color]) { |
2604 | pwq->flush_color = flush_color; | |
2605 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2606 | wait = true; |
2607 | } | |
2608 | } | |
1da177e4 | 2609 | |
73f53c4a | 2610 | if (work_color >= 0) { |
6183c009 | 2611 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2612 | pwq->work_color = work_color; |
73f53c4a | 2613 | } |
1da177e4 | 2614 | |
b09f4fd3 | 2615 | spin_unlock_irq(&pool->lock); |
1da177e4 | 2616 | } |
2355b70f | 2617 | |
112202d9 | 2618 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2619 | complete(&wq->first_flusher->done); |
14441960 | 2620 | |
73f53c4a | 2621 | return wait; |
1da177e4 LT |
2622 | } |
2623 | ||
0fcb78c2 | 2624 | /** |
1da177e4 | 2625 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2626 | * @wq: workqueue to flush |
1da177e4 | 2627 | * |
c5aa87bb TH |
2628 | * This function sleeps until all work items which were queued on entry |
2629 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2630 | */ |
7ad5b3a5 | 2631 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2632 | { |
73f53c4a TH |
2633 | struct wq_flusher this_flusher = { |
2634 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2635 | .flush_color = -1, | |
2636 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), | |
2637 | }; | |
2638 | int next_color; | |
1da177e4 | 2639 | |
3295f0ef IM |
2640 | lock_map_acquire(&wq->lockdep_map); |
2641 | lock_map_release(&wq->lockdep_map); | |
73f53c4a | 2642 | |
3c25a55d | 2643 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2644 | |
2645 | /* | |
2646 | * Start-to-wait phase | |
2647 | */ | |
2648 | next_color = work_next_color(wq->work_color); | |
2649 | ||
2650 | if (next_color != wq->flush_color) { | |
2651 | /* | |
2652 | * Color space is not full. The current work_color | |
2653 | * becomes our flush_color and work_color is advanced | |
2654 | * by one. | |
2655 | */ | |
6183c009 | 2656 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2657 | this_flusher.flush_color = wq->work_color; |
2658 | wq->work_color = next_color; | |
2659 | ||
2660 | if (!wq->first_flusher) { | |
2661 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2662 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2663 | |
2664 | wq->first_flusher = &this_flusher; | |
2665 | ||
112202d9 | 2666 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2667 | wq->work_color)) { |
2668 | /* nothing to flush, done */ | |
2669 | wq->flush_color = next_color; | |
2670 | wq->first_flusher = NULL; | |
2671 | goto out_unlock; | |
2672 | } | |
2673 | } else { | |
2674 | /* wait in queue */ | |
6183c009 | 2675 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2676 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2677 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2678 | } |
2679 | } else { | |
2680 | /* | |
2681 | * Oops, color space is full, wait on overflow queue. | |
2682 | * The next flush completion will assign us | |
2683 | * flush_color and transfer to flusher_queue. | |
2684 | */ | |
2685 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2686 | } | |
2687 | ||
3c25a55d | 2688 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2689 | |
2690 | wait_for_completion(&this_flusher.done); | |
2691 | ||
2692 | /* | |
2693 | * Wake-up-and-cascade phase | |
2694 | * | |
2695 | * First flushers are responsible for cascading flushes and | |
2696 | * handling overflow. Non-first flushers can simply return. | |
2697 | */ | |
2698 | if (wq->first_flusher != &this_flusher) | |
2699 | return; | |
2700 | ||
3c25a55d | 2701 | mutex_lock(&wq->mutex); |
73f53c4a | 2702 | |
4ce48b37 TH |
2703 | /* we might have raced, check again with mutex held */ |
2704 | if (wq->first_flusher != &this_flusher) | |
2705 | goto out_unlock; | |
2706 | ||
73f53c4a TH |
2707 | wq->first_flusher = NULL; |
2708 | ||
6183c009 TH |
2709 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2710 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2711 | |
2712 | while (true) { | |
2713 | struct wq_flusher *next, *tmp; | |
2714 | ||
2715 | /* complete all the flushers sharing the current flush color */ | |
2716 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2717 | if (next->flush_color != wq->flush_color) | |
2718 | break; | |
2719 | list_del_init(&next->list); | |
2720 | complete(&next->done); | |
2721 | } | |
2722 | ||
6183c009 TH |
2723 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2724 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2725 | |
2726 | /* this flush_color is finished, advance by one */ | |
2727 | wq->flush_color = work_next_color(wq->flush_color); | |
2728 | ||
2729 | /* one color has been freed, handle overflow queue */ | |
2730 | if (!list_empty(&wq->flusher_overflow)) { | |
2731 | /* | |
2732 | * Assign the same color to all overflowed | |
2733 | * flushers, advance work_color and append to | |
2734 | * flusher_queue. This is the start-to-wait | |
2735 | * phase for these overflowed flushers. | |
2736 | */ | |
2737 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2738 | tmp->flush_color = wq->work_color; | |
2739 | ||
2740 | wq->work_color = work_next_color(wq->work_color); | |
2741 | ||
2742 | list_splice_tail_init(&wq->flusher_overflow, | |
2743 | &wq->flusher_queue); | |
112202d9 | 2744 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2745 | } |
2746 | ||
2747 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2748 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2749 | break; |
2750 | } | |
2751 | ||
2752 | /* | |
2753 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2754 | * the new first flusher and arm pwqs. |
73f53c4a | 2755 | */ |
6183c009 TH |
2756 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2757 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2758 | |
2759 | list_del_init(&next->list); | |
2760 | wq->first_flusher = next; | |
2761 | ||
112202d9 | 2762 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2763 | break; |
2764 | ||
2765 | /* | |
2766 | * Meh... this color is already done, clear first | |
2767 | * flusher and repeat cascading. | |
2768 | */ | |
2769 | wq->first_flusher = NULL; | |
2770 | } | |
2771 | ||
2772 | out_unlock: | |
3c25a55d | 2773 | mutex_unlock(&wq->mutex); |
1da177e4 | 2774 | } |
ae90dd5d | 2775 | EXPORT_SYMBOL_GPL(flush_workqueue); |
1da177e4 | 2776 | |
9c5a2ba7 TH |
2777 | /** |
2778 | * drain_workqueue - drain a workqueue | |
2779 | * @wq: workqueue to drain | |
2780 | * | |
2781 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2782 | * only chain queueing is allowed. IOW, only currently pending or running | |
2783 | * work items on @wq can queue further work items on it. @wq is flushed | |
2784 | * repeatedly until it becomes empty. The number of flushing is detemined | |
2785 | * by the depth of chaining and should be relatively short. Whine if it | |
2786 | * takes too long. | |
2787 | */ | |
2788 | void drain_workqueue(struct workqueue_struct *wq) | |
2789 | { | |
2790 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2791 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2792 | |
2793 | /* | |
2794 | * __queue_work() needs to test whether there are drainers, is much | |
2795 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 2796 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 2797 | */ |
87fc741e | 2798 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 2799 | if (!wq->nr_drainers++) |
618b01eb | 2800 | wq->flags |= __WQ_DRAINING; |
87fc741e | 2801 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2802 | reflush: |
2803 | flush_workqueue(wq); | |
2804 | ||
b09f4fd3 | 2805 | mutex_lock(&wq->mutex); |
76af4d93 | 2806 | |
49e3cf44 | 2807 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2808 | bool drained; |
9c5a2ba7 | 2809 | |
b09f4fd3 | 2810 | spin_lock_irq(&pwq->pool->lock); |
112202d9 | 2811 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
b09f4fd3 | 2812 | spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
2813 | |
2814 | if (drained) | |
9c5a2ba7 TH |
2815 | continue; |
2816 | ||
2817 | if (++flush_cnt == 10 || | |
2818 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
c5aa87bb | 2819 | pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", |
044c782c | 2820 | wq->name, flush_cnt); |
76af4d93 | 2821 | |
b09f4fd3 | 2822 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2823 | goto reflush; |
2824 | } | |
2825 | ||
9c5a2ba7 | 2826 | if (!--wq->nr_drainers) |
618b01eb | 2827 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 2828 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2829 | } |
2830 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2831 | ||
606a5020 | 2832 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) |
db700897 | 2833 | { |
affee4b2 | 2834 | struct worker *worker = NULL; |
c9e7cf27 | 2835 | struct worker_pool *pool; |
112202d9 | 2836 | struct pool_workqueue *pwq; |
db700897 ON |
2837 | |
2838 | might_sleep(); | |
fa1b54e6 TH |
2839 | |
2840 | local_irq_disable(); | |
c9e7cf27 | 2841 | pool = get_work_pool(work); |
fa1b54e6 TH |
2842 | if (!pool) { |
2843 | local_irq_enable(); | |
baf59022 | 2844 | return false; |
fa1b54e6 | 2845 | } |
db700897 | 2846 | |
fa1b54e6 | 2847 | spin_lock(&pool->lock); |
0b3dae68 | 2848 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2849 | pwq = get_work_pwq(work); |
2850 | if (pwq) { | |
2851 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2852 | goto already_gone; |
606a5020 | 2853 | } else { |
c9e7cf27 | 2854 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2855 | if (!worker) |
4690c4ab | 2856 | goto already_gone; |
112202d9 | 2857 | pwq = worker->current_pwq; |
606a5020 | 2858 | } |
db700897 | 2859 | |
112202d9 | 2860 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 2861 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 2862 | |
e159489b TH |
2863 | /* |
2864 | * If @max_active is 1 or rescuer is in use, flushing another work | |
2865 | * item on the same workqueue may lead to deadlock. Make sure the | |
2866 | * flusher is not running on the same workqueue by verifying write | |
2867 | * access. | |
2868 | */ | |
493008a8 | 2869 | if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) |
112202d9 | 2870 | lock_map_acquire(&pwq->wq->lockdep_map); |
e159489b | 2871 | else |
112202d9 TH |
2872 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
2873 | lock_map_release(&pwq->wq->lockdep_map); | |
e159489b | 2874 | |
401a8d04 | 2875 | return true; |
4690c4ab | 2876 | already_gone: |
d565ed63 | 2877 | spin_unlock_irq(&pool->lock); |
401a8d04 | 2878 | return false; |
db700897 | 2879 | } |
baf59022 TH |
2880 | |
2881 | /** | |
2882 | * flush_work - wait for a work to finish executing the last queueing instance | |
2883 | * @work: the work to flush | |
2884 | * | |
606a5020 TH |
2885 | * Wait until @work has finished execution. @work is guaranteed to be idle |
2886 | * on return if it hasn't been requeued since flush started. | |
baf59022 TH |
2887 | * |
2888 | * RETURNS: | |
2889 | * %true if flush_work() waited for the work to finish execution, | |
2890 | * %false if it was already idle. | |
2891 | */ | |
2892 | bool flush_work(struct work_struct *work) | |
2893 | { | |
2894 | struct wq_barrier barr; | |
2895 | ||
0976dfc1 SB |
2896 | lock_map_acquire(&work->lockdep_map); |
2897 | lock_map_release(&work->lockdep_map); | |
2898 | ||
606a5020 | 2899 | if (start_flush_work(work, &barr)) { |
401a8d04 TH |
2900 | wait_for_completion(&barr.done); |
2901 | destroy_work_on_stack(&barr.work); | |
2902 | return true; | |
606a5020 | 2903 | } else { |
401a8d04 | 2904 | return false; |
6e84d644 | 2905 | } |
6e84d644 | 2906 | } |
606a5020 | 2907 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 2908 | |
01fc83d7 TH |
2909 | struct cwt_wait { |
2910 | wait_queue_t wait; | |
2911 | struct work_struct *work; | |
2912 | }; | |
2913 | ||
2914 | static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) | |
2915 | { | |
2916 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
2917 | ||
2918 | if (cwait->work != key) | |
2919 | return 0; | |
2920 | return autoremove_wake_function(wait, mode, sync, key); | |
2921 | } | |
2922 | ||
36e227d2 | 2923 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 2924 | { |
01fc83d7 | 2925 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 2926 | unsigned long flags; |
1f1f642e ON |
2927 | int ret; |
2928 | ||
2929 | do { | |
bbb68dfa TH |
2930 | ret = try_to_grab_pending(work, is_dwork, &flags); |
2931 | /* | |
01fc83d7 TH |
2932 | * If someone else is already canceling, wait for it to |
2933 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
2934 | * because we may get scheduled between @work's completion | |
2935 | * and the other canceling task resuming and clearing | |
2936 | * CANCELING - flush_work() will return false immediately | |
2937 | * as @work is no longer busy, try_to_grab_pending() will | |
2938 | * return -ENOENT as @work is still being canceled and the | |
2939 | * other canceling task won't be able to clear CANCELING as | |
2940 | * we're hogging the CPU. | |
2941 | * | |
2942 | * Let's wait for completion using a waitqueue. As this | |
2943 | * may lead to the thundering herd problem, use a custom | |
2944 | * wake function which matches @work along with exclusive | |
2945 | * wait and wakeup. | |
bbb68dfa | 2946 | */ |
01fc83d7 TH |
2947 | if (unlikely(ret == -ENOENT)) { |
2948 | struct cwt_wait cwait; | |
2949 | ||
2950 | init_wait(&cwait.wait); | |
2951 | cwait.wait.func = cwt_wakefn; | |
2952 | cwait.work = work; | |
2953 | ||
2954 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
2955 | TASK_UNINTERRUPTIBLE); | |
2956 | if (work_is_canceling(work)) | |
2957 | schedule(); | |
2958 | finish_wait(&cancel_waitq, &cwait.wait); | |
2959 | } | |
1f1f642e ON |
2960 | } while (unlikely(ret < 0)); |
2961 | ||
bbb68dfa TH |
2962 | /* tell other tasks trying to grab @work to back off */ |
2963 | mark_work_canceling(work); | |
2964 | local_irq_restore(flags); | |
2965 | ||
606a5020 | 2966 | flush_work(work); |
7a22ad75 | 2967 | clear_work_data(work); |
01fc83d7 TH |
2968 | |
2969 | /* | |
2970 | * Paired with prepare_to_wait() above so that either | |
2971 | * waitqueue_active() is visible here or !work_is_canceling() is | |
2972 | * visible there. | |
2973 | */ | |
2974 | smp_mb(); | |
2975 | if (waitqueue_active(&cancel_waitq)) | |
2976 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
2977 | ||
1f1f642e ON |
2978 | return ret; |
2979 | } | |
2980 | ||
6e84d644 | 2981 | /** |
401a8d04 TH |
2982 | * cancel_work_sync - cancel a work and wait for it to finish |
2983 | * @work: the work to cancel | |
6e84d644 | 2984 | * |
401a8d04 TH |
2985 | * Cancel @work and wait for its execution to finish. This function |
2986 | * can be used even if the work re-queues itself or migrates to | |
2987 | * another workqueue. On return from this function, @work is | |
2988 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 2989 | * |
401a8d04 TH |
2990 | * cancel_work_sync(&delayed_work->work) must not be used for |
2991 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 2992 | * |
401a8d04 | 2993 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 2994 | * queued can't be destroyed before this function returns. |
401a8d04 TH |
2995 | * |
2996 | * RETURNS: | |
2997 | * %true if @work was pending, %false otherwise. | |
6e84d644 | 2998 | */ |
401a8d04 | 2999 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3000 | { |
36e227d2 | 3001 | return __cancel_work_timer(work, false); |
b89deed3 | 3002 | } |
28e53bdd | 3003 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3004 | |
6e84d644 | 3005 | /** |
401a8d04 TH |
3006 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3007 | * @dwork: the delayed work to flush | |
6e84d644 | 3008 | * |
401a8d04 TH |
3009 | * Delayed timer is cancelled and the pending work is queued for |
3010 | * immediate execution. Like flush_work(), this function only | |
3011 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3012 | * |
401a8d04 TH |
3013 | * RETURNS: |
3014 | * %true if flush_work() waited for the work to finish execution, | |
3015 | * %false if it was already idle. | |
6e84d644 | 3016 | */ |
401a8d04 TH |
3017 | bool flush_delayed_work(struct delayed_work *dwork) |
3018 | { | |
8930caba | 3019 | local_irq_disable(); |
401a8d04 | 3020 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3021 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3022 | local_irq_enable(); |
401a8d04 TH |
3023 | return flush_work(&dwork->work); |
3024 | } | |
3025 | EXPORT_SYMBOL(flush_delayed_work); | |
3026 | ||
09383498 | 3027 | /** |
57b30ae7 TH |
3028 | * cancel_delayed_work - cancel a delayed work |
3029 | * @dwork: delayed_work to cancel | |
09383498 | 3030 | * |
57b30ae7 TH |
3031 | * Kill off a pending delayed_work. Returns %true if @dwork was pending |
3032 | * and canceled; %false if wasn't pending. Note that the work callback | |
3033 | * function may still be running on return, unless it returns %true and the | |
3034 | * work doesn't re-arm itself. Explicitly flush or use | |
3035 | * cancel_delayed_work_sync() to wait on it. | |
09383498 | 3036 | * |
57b30ae7 | 3037 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3038 | */ |
57b30ae7 | 3039 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3040 | { |
57b30ae7 TH |
3041 | unsigned long flags; |
3042 | int ret; | |
3043 | ||
3044 | do { | |
3045 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
3046 | } while (unlikely(ret == -EAGAIN)); | |
3047 | ||
3048 | if (unlikely(ret < 0)) | |
3049 | return false; | |
3050 | ||
7c3eed5c TH |
3051 | set_work_pool_and_clear_pending(&dwork->work, |
3052 | get_work_pool_id(&dwork->work)); | |
57b30ae7 | 3053 | local_irq_restore(flags); |
c0158ca6 | 3054 | return ret; |
09383498 | 3055 | } |
57b30ae7 | 3056 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3057 | |
401a8d04 TH |
3058 | /** |
3059 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3060 | * @dwork: the delayed work cancel | |
3061 | * | |
3062 | * This is cancel_work_sync() for delayed works. | |
3063 | * | |
3064 | * RETURNS: | |
3065 | * %true if @dwork was pending, %false otherwise. | |
3066 | */ | |
3067 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3068 | { |
36e227d2 | 3069 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3070 | } |
f5a421a4 | 3071 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3072 | |
b6136773 | 3073 | /** |
31ddd871 | 3074 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3075 | * @func: the function to call |
b6136773 | 3076 | * |
31ddd871 TH |
3077 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3078 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3079 | * schedule_on_each_cpu() is very slow. |
31ddd871 TH |
3080 | * |
3081 | * RETURNS: | |
3082 | * 0 on success, -errno on failure. | |
b6136773 | 3083 | */ |
65f27f38 | 3084 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3085 | { |
3086 | int cpu; | |
38f51568 | 3087 | struct work_struct __percpu *works; |
15316ba8 | 3088 | |
b6136773 AM |
3089 | works = alloc_percpu(struct work_struct); |
3090 | if (!works) | |
15316ba8 | 3091 | return -ENOMEM; |
b6136773 | 3092 | |
93981800 TH |
3093 | get_online_cpus(); |
3094 | ||
15316ba8 | 3095 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3096 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3097 | ||
3098 | INIT_WORK(work, func); | |
b71ab8c2 | 3099 | schedule_work_on(cpu, work); |
65a64464 | 3100 | } |
93981800 TH |
3101 | |
3102 | for_each_online_cpu(cpu) | |
3103 | flush_work(per_cpu_ptr(works, cpu)); | |
3104 | ||
95402b38 | 3105 | put_online_cpus(); |
b6136773 | 3106 | free_percpu(works); |
15316ba8 CL |
3107 | return 0; |
3108 | } | |
3109 | ||
eef6a7d5 AS |
3110 | /** |
3111 | * flush_scheduled_work - ensure that any scheduled work has run to completion. | |
3112 | * | |
3113 | * Forces execution of the kernel-global workqueue and blocks until its | |
3114 | * completion. | |
3115 | * | |
3116 | * Think twice before calling this function! It's very easy to get into | |
3117 | * trouble if you don't take great care. Either of the following situations | |
3118 | * will lead to deadlock: | |
3119 | * | |
3120 | * One of the work items currently on the workqueue needs to acquire | |
3121 | * a lock held by your code or its caller. | |
3122 | * | |
3123 | * Your code is running in the context of a work routine. | |
3124 | * | |
3125 | * They will be detected by lockdep when they occur, but the first might not | |
3126 | * occur very often. It depends on what work items are on the workqueue and | |
3127 | * what locks they need, which you have no control over. | |
3128 | * | |
3129 | * In most situations flushing the entire workqueue is overkill; you merely | |
3130 | * need to know that a particular work item isn't queued and isn't running. | |
3131 | * In such cases you should use cancel_delayed_work_sync() or | |
3132 | * cancel_work_sync() instead. | |
3133 | */ | |
1da177e4 LT |
3134 | void flush_scheduled_work(void) |
3135 | { | |
d320c038 | 3136 | flush_workqueue(system_wq); |
1da177e4 | 3137 | } |
ae90dd5d | 3138 | EXPORT_SYMBOL(flush_scheduled_work); |
1da177e4 | 3139 | |
1fa44eca JB |
3140 | /** |
3141 | * execute_in_process_context - reliably execute the routine with user context | |
3142 | * @fn: the function to execute | |
1fa44eca JB |
3143 | * @ew: guaranteed storage for the execute work structure (must |
3144 | * be available when the work executes) | |
3145 | * | |
3146 | * Executes the function immediately if process context is available, | |
3147 | * otherwise schedules the function for delayed execution. | |
3148 | * | |
3149 | * Returns: 0 - function was executed | |
3150 | * 1 - function was scheduled for execution | |
3151 | */ | |
65f27f38 | 3152 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3153 | { |
3154 | if (!in_interrupt()) { | |
65f27f38 | 3155 | fn(&ew->work); |
1fa44eca JB |
3156 | return 0; |
3157 | } | |
3158 | ||
65f27f38 | 3159 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3160 | schedule_work(&ew->work); |
3161 | ||
3162 | return 1; | |
3163 | } | |
3164 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3165 | ||
226223ab TH |
3166 | #ifdef CONFIG_SYSFS |
3167 | /* | |
3168 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
3169 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
3170 | * following attributes. | |
3171 | * | |
3172 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
3173 | * max_active RW int : maximum number of in-flight work items | |
3174 | * | |
3175 | * Unbound workqueues have the following extra attributes. | |
3176 | * | |
3177 | * id RO int : the associated pool ID | |
3178 | * nice RW int : nice value of the workers | |
3179 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
3180 | */ | |
3181 | struct wq_device { | |
3182 | struct workqueue_struct *wq; | |
3183 | struct device dev; | |
3184 | }; | |
3185 | ||
3186 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
3187 | { | |
3188 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
3189 | ||
3190 | return wq_dev->wq; | |
3191 | } | |
3192 | ||
3193 | static ssize_t wq_per_cpu_show(struct device *dev, | |
3194 | struct device_attribute *attr, char *buf) | |
3195 | { | |
3196 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3197 | ||
3198 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
3199 | } | |
3200 | ||
3201 | static ssize_t wq_max_active_show(struct device *dev, | |
3202 | struct device_attribute *attr, char *buf) | |
3203 | { | |
3204 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3205 | ||
3206 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
3207 | } | |
3208 | ||
3209 | static ssize_t wq_max_active_store(struct device *dev, | |
3210 | struct device_attribute *attr, | |
3211 | const char *buf, size_t count) | |
3212 | { | |
3213 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3214 | int val; | |
3215 | ||
3216 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
3217 | return -EINVAL; | |
3218 | ||
3219 | workqueue_set_max_active(wq, val); | |
3220 | return count; | |
3221 | } | |
3222 | ||
3223 | static struct device_attribute wq_sysfs_attrs[] = { | |
3224 | __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), | |
3225 | __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), | |
3226 | __ATTR_NULL, | |
3227 | }; | |
3228 | ||
d55262c4 TH |
3229 | static ssize_t wq_pool_ids_show(struct device *dev, |
3230 | struct device_attribute *attr, char *buf) | |
226223ab TH |
3231 | { |
3232 | struct workqueue_struct *wq = dev_to_wq(dev); | |
d55262c4 TH |
3233 | const char *delim = ""; |
3234 | int node, written = 0; | |
226223ab TH |
3235 | |
3236 | rcu_read_lock_sched(); | |
d55262c4 TH |
3237 | for_each_node(node) { |
3238 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
3239 | "%s%d:%d", delim, node, | |
3240 | unbound_pwq_by_node(wq, node)->pool->id); | |
3241 | delim = " "; | |
3242 | } | |
3243 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
226223ab TH |
3244 | rcu_read_unlock_sched(); |
3245 | ||
3246 | return written; | |
3247 | } | |
3248 | ||
3249 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
3250 | char *buf) | |
3251 | { | |
3252 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3253 | int written; | |
3254 | ||
6029a918 TH |
3255 | mutex_lock(&wq->mutex); |
3256 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
3257 | mutex_unlock(&wq->mutex); | |
226223ab TH |
3258 | |
3259 | return written; | |
3260 | } | |
3261 | ||
3262 | /* prepare workqueue_attrs for sysfs store operations */ | |
3263 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
3264 | { | |
3265 | struct workqueue_attrs *attrs; | |
3266 | ||
3267 | attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3268 | if (!attrs) | |
3269 | return NULL; | |
3270 | ||
6029a918 TH |
3271 | mutex_lock(&wq->mutex); |
3272 | copy_workqueue_attrs(attrs, wq->unbound_attrs); | |
3273 | mutex_unlock(&wq->mutex); | |
226223ab TH |
3274 | return attrs; |
3275 | } | |
3276 | ||
3277 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
3278 | const char *buf, size_t count) | |
3279 | { | |
3280 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3281 | struct workqueue_attrs *attrs; | |
3282 | int ret; | |
3283 | ||
3284 | attrs = wq_sysfs_prep_attrs(wq); | |
3285 | if (!attrs) | |
3286 | return -ENOMEM; | |
3287 | ||
3288 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
3289 | attrs->nice >= -20 && attrs->nice <= 19) | |
3290 | ret = apply_workqueue_attrs(wq, attrs); | |
3291 | else | |
3292 | ret = -EINVAL; | |
3293 | ||
3294 | free_workqueue_attrs(attrs); | |
3295 | return ret ?: count; | |
3296 | } | |
3297 | ||
3298 | static ssize_t wq_cpumask_show(struct device *dev, | |
3299 | struct device_attribute *attr, char *buf) | |
3300 | { | |
3301 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3302 | int written; | |
3303 | ||
6029a918 TH |
3304 | mutex_lock(&wq->mutex); |
3305 | written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); | |
3306 | mutex_unlock(&wq->mutex); | |
226223ab TH |
3307 | |
3308 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
3309 | return written; | |
3310 | } | |
3311 | ||
3312 | static ssize_t wq_cpumask_store(struct device *dev, | |
3313 | struct device_attribute *attr, | |
3314 | const char *buf, size_t count) | |
3315 | { | |
3316 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3317 | struct workqueue_attrs *attrs; | |
3318 | int ret; | |
3319 | ||
3320 | attrs = wq_sysfs_prep_attrs(wq); | |
3321 | if (!attrs) | |
3322 | return -ENOMEM; | |
3323 | ||
3324 | ret = cpumask_parse(buf, attrs->cpumask); | |
3325 | if (!ret) | |
3326 | ret = apply_workqueue_attrs(wq, attrs); | |
3327 | ||
3328 | free_workqueue_attrs(attrs); | |
3329 | return ret ?: count; | |
3330 | } | |
3331 | ||
d55262c4 TH |
3332 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, |
3333 | char *buf) | |
3334 | { | |
3335 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3336 | int written; | |
3337 | ||
3338 | mutex_lock(&wq->mutex); | |
3339 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
3340 | !wq->unbound_attrs->no_numa); | |
3341 | mutex_unlock(&wq->mutex); | |
3342 | ||
3343 | return written; | |
3344 | } | |
3345 | ||
3346 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, | |
3347 | const char *buf, size_t count) | |
3348 | { | |
3349 | struct workqueue_struct *wq = dev_to_wq(dev); | |
3350 | struct workqueue_attrs *attrs; | |
3351 | int v, ret; | |
3352 | ||
3353 | attrs = wq_sysfs_prep_attrs(wq); | |
3354 | if (!attrs) | |
3355 | return -ENOMEM; | |
3356 | ||
3357 | ret = -EINVAL; | |
3358 | if (sscanf(buf, "%d", &v) == 1) { | |
3359 | attrs->no_numa = !v; | |
3360 | ret = apply_workqueue_attrs(wq, attrs); | |
3361 | } | |
3362 | ||
3363 | free_workqueue_attrs(attrs); | |
3364 | return ret ?: count; | |
3365 | } | |
3366 | ||
226223ab | 3367 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
d55262c4 | 3368 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), |
226223ab TH |
3369 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
3370 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
d55262c4 | 3371 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), |
226223ab TH |
3372 | __ATTR_NULL, |
3373 | }; | |
3374 | ||
3375 | static struct bus_type wq_subsys = { | |
3376 | .name = "workqueue", | |
3377 | .dev_attrs = wq_sysfs_attrs, | |
3378 | }; | |
3379 | ||
3380 | static int __init wq_sysfs_init(void) | |
3381 | { | |
3382 | return subsys_virtual_register(&wq_subsys, NULL); | |
3383 | } | |
3384 | core_initcall(wq_sysfs_init); | |
3385 | ||
3386 | static void wq_device_release(struct device *dev) | |
3387 | { | |
3388 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
3389 | ||
3390 | kfree(wq_dev); | |
3391 | } | |
3392 | ||
3393 | /** | |
3394 | * workqueue_sysfs_register - make a workqueue visible in sysfs | |
3395 | * @wq: the workqueue to register | |
3396 | * | |
3397 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. | |
3398 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
3399 | * which is the preferred method. | |
3400 | * | |
3401 | * Workqueue user should use this function directly iff it wants to apply | |
3402 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
3403 | * apply_workqueue_attrs() may race against userland updating the | |
3404 | * attributes. | |
3405 | * | |
3406 | * Returns 0 on success, -errno on failure. | |
3407 | */ | |
3408 | int workqueue_sysfs_register(struct workqueue_struct *wq) | |
3409 | { | |
3410 | struct wq_device *wq_dev; | |
3411 | int ret; | |
3412 | ||
3413 | /* | |
3414 | * Adjusting max_active or creating new pwqs by applyting | |
3415 | * attributes breaks ordering guarantee. Disallow exposing ordered | |
3416 | * workqueues. | |
3417 | */ | |
f3538ebd | 3418 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
226223ab TH |
3419 | return -EINVAL; |
3420 | ||
3421 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); | |
3422 | if (!wq_dev) | |
3423 | return -ENOMEM; | |
3424 | ||
3425 | wq_dev->wq = wq; | |
3426 | wq_dev->dev.bus = &wq_subsys; | |
3427 | wq_dev->dev.init_name = wq->name; | |
3428 | wq_dev->dev.release = wq_device_release; | |
3429 | ||
3430 | /* | |
3431 | * unbound_attrs are created separately. Suppress uevent until | |
3432 | * everything is ready. | |
3433 | */ | |
3434 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
3435 | ||
3436 | ret = device_register(&wq_dev->dev); | |
3437 | if (ret) { | |
3438 | kfree(wq_dev); | |
3439 | wq->wq_dev = NULL; | |
3440 | return ret; | |
3441 | } | |
3442 | ||
3443 | if (wq->flags & WQ_UNBOUND) { | |
3444 | struct device_attribute *attr; | |
3445 | ||
3446 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { | |
3447 | ret = device_create_file(&wq_dev->dev, attr); | |
3448 | if (ret) { | |
3449 | device_unregister(&wq_dev->dev); | |
3450 | wq->wq_dev = NULL; | |
3451 | return ret; | |
3452 | } | |
3453 | } | |
3454 | } | |
3455 | ||
7c36f88c | 3456 | dev_set_uevent_suppress(&wq_dev->dev, false); |
226223ab TH |
3457 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); |
3458 | return 0; | |
3459 | } | |
3460 | ||
3461 | /** | |
3462 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() | |
3463 | * @wq: the workqueue to unregister | |
3464 | * | |
3465 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. | |
3466 | */ | |
3467 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) | |
3468 | { | |
3469 | struct wq_device *wq_dev = wq->wq_dev; | |
3470 | ||
3471 | if (!wq->wq_dev) | |
3472 | return; | |
3473 | ||
3474 | wq->wq_dev = NULL; | |
3475 | device_unregister(&wq_dev->dev); | |
3476 | } | |
3477 | #else /* CONFIG_SYSFS */ | |
3478 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
3479 | #endif /* CONFIG_SYSFS */ | |
3480 | ||
7a4e344c TH |
3481 | /** |
3482 | * free_workqueue_attrs - free a workqueue_attrs | |
3483 | * @attrs: workqueue_attrs to free | |
3484 | * | |
3485 | * Undo alloc_workqueue_attrs(). | |
3486 | */ | |
3487 | void free_workqueue_attrs(struct workqueue_attrs *attrs) | |
3488 | { | |
3489 | if (attrs) { | |
3490 | free_cpumask_var(attrs->cpumask); | |
3491 | kfree(attrs); | |
3492 | } | |
3493 | } | |
3494 | ||
3495 | /** | |
3496 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
3497 | * @gfp_mask: allocation mask to use | |
3498 | * | |
3499 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3500 | * return it. Returns NULL on failure. | |
3501 | */ | |
3502 | struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) | |
3503 | { | |
3504 | struct workqueue_attrs *attrs; | |
3505 | ||
3506 | attrs = kzalloc(sizeof(*attrs), gfp_mask); | |
3507 | if (!attrs) | |
3508 | goto fail; | |
3509 | if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) | |
3510 | goto fail; | |
3511 | ||
13e2e556 | 3512 | cpumask_copy(attrs->cpumask, cpu_possible_mask); |
7a4e344c TH |
3513 | return attrs; |
3514 | fail: | |
3515 | free_workqueue_attrs(attrs); | |
3516 | return NULL; | |
3517 | } | |
3518 | ||
29c91e99 TH |
3519 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3520 | const struct workqueue_attrs *from) | |
3521 | { | |
3522 | to->nice = from->nice; | |
3523 | cpumask_copy(to->cpumask, from->cpumask); | |
73b8bd6d SL |
3524 | /* |
3525 | * Unlike hash and equality test, this function doesn't ignore | |
3526 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3527 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3528 | */ | |
3529 | to->no_numa = from->no_numa; | |
29c91e99 TH |
3530 | } |
3531 | ||
29c91e99 TH |
3532 | /* hash value of the content of @attr */ |
3533 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
3534 | { | |
3535 | u32 hash = 0; | |
3536 | ||
3537 | hash = jhash_1word(attrs->nice, hash); | |
13e2e556 TH |
3538 | hash = jhash(cpumask_bits(attrs->cpumask), |
3539 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
29c91e99 TH |
3540 | return hash; |
3541 | } | |
3542 | ||
3543 | /* content equality test */ | |
3544 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3545 | const struct workqueue_attrs *b) | |
3546 | { | |
3547 | if (a->nice != b->nice) | |
3548 | return false; | |
3549 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3550 | return false; | |
3551 | return true; | |
3552 | } | |
3553 | ||
7a4e344c TH |
3554 | /** |
3555 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3556 | * @pool: worker_pool to initialize | |
3557 | * | |
3558 | * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. | |
29c91e99 TH |
3559 | * Returns 0 on success, -errno on failure. Even on failure, all fields |
3560 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3561 | * on @pool safely to release it. | |
7a4e344c TH |
3562 | */ |
3563 | static int init_worker_pool(struct worker_pool *pool) | |
4e1a1f9a TH |
3564 | { |
3565 | spin_lock_init(&pool->lock); | |
29c91e99 TH |
3566 | pool->id = -1; |
3567 | pool->cpu = -1; | |
f3f90ad4 | 3568 | pool->node = NUMA_NO_NODE; |
4e1a1f9a TH |
3569 | pool->flags |= POOL_DISASSOCIATED; |
3570 | INIT_LIST_HEAD(&pool->worklist); | |
3571 | INIT_LIST_HEAD(&pool->idle_list); | |
3572 | hash_init(pool->busy_hash); | |
3573 | ||
3574 | init_timer_deferrable(&pool->idle_timer); | |
3575 | pool->idle_timer.function = idle_worker_timeout; | |
3576 | pool->idle_timer.data = (unsigned long)pool; | |
3577 | ||
3578 | setup_timer(&pool->mayday_timer, pool_mayday_timeout, | |
3579 | (unsigned long)pool); | |
3580 | ||
3581 | mutex_init(&pool->manager_arb); | |
bc3a1afc | 3582 | mutex_init(&pool->manager_mutex); |
822d8405 | 3583 | idr_init(&pool->worker_idr); |
7a4e344c | 3584 | |
29c91e99 TH |
3585 | INIT_HLIST_NODE(&pool->hash_node); |
3586 | pool->refcnt = 1; | |
3587 | ||
3588 | /* shouldn't fail above this point */ | |
7a4e344c TH |
3589 | pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); |
3590 | if (!pool->attrs) | |
3591 | return -ENOMEM; | |
3592 | return 0; | |
4e1a1f9a TH |
3593 | } |
3594 | ||
29c91e99 TH |
3595 | static void rcu_free_pool(struct rcu_head *rcu) |
3596 | { | |
3597 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); | |
3598 | ||
822d8405 | 3599 | idr_destroy(&pool->worker_idr); |
29c91e99 TH |
3600 | free_workqueue_attrs(pool->attrs); |
3601 | kfree(pool); | |
3602 | } | |
3603 | ||
3604 | /** | |
3605 | * put_unbound_pool - put a worker_pool | |
3606 | * @pool: worker_pool to put | |
3607 | * | |
3608 | * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU | |
c5aa87bb TH |
3609 | * safe manner. get_unbound_pool() calls this function on its failure path |
3610 | * and this function should be able to release pools which went through, | |
3611 | * successfully or not, init_worker_pool(). | |
a892cacc TH |
3612 | * |
3613 | * Should be called with wq_pool_mutex held. | |
29c91e99 TH |
3614 | */ |
3615 | static void put_unbound_pool(struct worker_pool *pool) | |
3616 | { | |
3617 | struct worker *worker; | |
3618 | ||
a892cacc TH |
3619 | lockdep_assert_held(&wq_pool_mutex); |
3620 | ||
3621 | if (--pool->refcnt) | |
29c91e99 | 3622 | return; |
29c91e99 TH |
3623 | |
3624 | /* sanity checks */ | |
3625 | if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || | |
a892cacc | 3626 | WARN_ON(!list_empty(&pool->worklist))) |
29c91e99 | 3627 | return; |
29c91e99 TH |
3628 | |
3629 | /* release id and unhash */ | |
3630 | if (pool->id >= 0) | |
3631 | idr_remove(&worker_pool_idr, pool->id); | |
3632 | hash_del(&pool->hash_node); | |
3633 | ||
c5aa87bb TH |
3634 | /* |
3635 | * Become the manager and destroy all workers. Grabbing | |
3636 | * manager_arb prevents @pool's workers from blocking on | |
3637 | * manager_mutex. | |
3638 | */ | |
29c91e99 | 3639 | mutex_lock(&pool->manager_arb); |
cd549687 | 3640 | mutex_lock(&pool->manager_mutex); |
29c91e99 TH |
3641 | spin_lock_irq(&pool->lock); |
3642 | ||
3643 | while ((worker = first_worker(pool))) | |
3644 | destroy_worker(worker); | |
3645 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
3646 | ||
3647 | spin_unlock_irq(&pool->lock); | |
cd549687 | 3648 | mutex_unlock(&pool->manager_mutex); |
29c91e99 TH |
3649 | mutex_unlock(&pool->manager_arb); |
3650 | ||
3651 | /* shut down the timers */ | |
3652 | del_timer_sync(&pool->idle_timer); | |
3653 | del_timer_sync(&pool->mayday_timer); | |
3654 | ||
3655 | /* sched-RCU protected to allow dereferences from get_work_pool() */ | |
3656 | call_rcu_sched(&pool->rcu, rcu_free_pool); | |
3657 | } | |
3658 | ||
3659 | /** | |
3660 | * get_unbound_pool - get a worker_pool with the specified attributes | |
3661 | * @attrs: the attributes of the worker_pool to get | |
3662 | * | |
3663 | * Obtain a worker_pool which has the same attributes as @attrs, bump the | |
3664 | * reference count and return it. If there already is a matching | |
3665 | * worker_pool, it will be used; otherwise, this function attempts to | |
3666 | * create a new one. On failure, returns NULL. | |
a892cacc TH |
3667 | * |
3668 | * Should be called with wq_pool_mutex held. | |
29c91e99 TH |
3669 | */ |
3670 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) | |
3671 | { | |
29c91e99 TH |
3672 | u32 hash = wqattrs_hash(attrs); |
3673 | struct worker_pool *pool; | |
f3f90ad4 | 3674 | int node; |
29c91e99 | 3675 | |
a892cacc | 3676 | lockdep_assert_held(&wq_pool_mutex); |
29c91e99 TH |
3677 | |
3678 | /* do we already have a matching pool? */ | |
29c91e99 TH |
3679 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { |
3680 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3681 | pool->refcnt++; | |
3682 | goto out_unlock; | |
3683 | } | |
3684 | } | |
29c91e99 TH |
3685 | |
3686 | /* nope, create a new one */ | |
3687 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); | |
3688 | if (!pool || init_worker_pool(pool) < 0) | |
3689 | goto fail; | |
3690 | ||
12ee4fc6 LJ |
3691 | if (workqueue_freezing) |
3692 | pool->flags |= POOL_FREEZING; | |
3693 | ||
8864b4e5 | 3694 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ |
29c91e99 TH |
3695 | copy_workqueue_attrs(pool->attrs, attrs); |
3696 | ||
73b8bd6d SL |
3697 | /* |
3698 | * no_numa isn't a worker_pool attribute, always clear it. See | |
3699 | * 'struct workqueue_attrs' comments for detail. | |
3700 | */ | |
3701 | pool->attrs->no_numa = false; | |
3702 | ||
f3f90ad4 TH |
3703 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3704 | if (wq_numa_enabled) { | |
3705 | for_each_node(node) { | |
3706 | if (cpumask_subset(pool->attrs->cpumask, | |
3707 | wq_numa_possible_cpumask[node])) { | |
3708 | pool->node = node; | |
3709 | break; | |
3710 | } | |
3711 | } | |
3712 | } | |
3713 | ||
29c91e99 TH |
3714 | if (worker_pool_assign_id(pool) < 0) |
3715 | goto fail; | |
3716 | ||
3717 | /* create and start the initial worker */ | |
ebf44d16 | 3718 | if (create_and_start_worker(pool) < 0) |
29c91e99 TH |
3719 | goto fail; |
3720 | ||
29c91e99 | 3721 | /* install */ |
29c91e99 TH |
3722 | hash_add(unbound_pool_hash, &pool->hash_node, hash); |
3723 | out_unlock: | |
29c91e99 TH |
3724 | return pool; |
3725 | fail: | |
29c91e99 TH |
3726 | if (pool) |
3727 | put_unbound_pool(pool); | |
3728 | return NULL; | |
3729 | } | |
3730 | ||
8864b4e5 TH |
3731 | static void rcu_free_pwq(struct rcu_head *rcu) |
3732 | { | |
3733 | kmem_cache_free(pwq_cache, | |
3734 | container_of(rcu, struct pool_workqueue, rcu)); | |
3735 | } | |
3736 | ||
3737 | /* | |
3738 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3739 | * and needs to be destroyed. | |
3740 | */ | |
3741 | static void pwq_unbound_release_workfn(struct work_struct *work) | |
3742 | { | |
3743 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, | |
3744 | unbound_release_work); | |
3745 | struct workqueue_struct *wq = pwq->wq; | |
3746 | struct worker_pool *pool = pwq->pool; | |
bc0caf09 | 3747 | bool is_last; |
8864b4e5 TH |
3748 | |
3749 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) | |
3750 | return; | |
3751 | ||
75ccf595 | 3752 | /* |
3c25a55d | 3753 | * Unlink @pwq. Synchronization against wq->mutex isn't strictly |
75ccf595 TH |
3754 | * necessary on release but do it anyway. It's easier to verify |
3755 | * and consistent with the linking path. | |
3756 | */ | |
3c25a55d | 3757 | mutex_lock(&wq->mutex); |
8864b4e5 | 3758 | list_del_rcu(&pwq->pwqs_node); |
bc0caf09 | 3759 | is_last = list_empty(&wq->pwqs); |
3c25a55d | 3760 | mutex_unlock(&wq->mutex); |
8864b4e5 | 3761 | |
a892cacc | 3762 | mutex_lock(&wq_pool_mutex); |
8864b4e5 | 3763 | put_unbound_pool(pool); |
a892cacc TH |
3764 | mutex_unlock(&wq_pool_mutex); |
3765 | ||
8864b4e5 TH |
3766 | call_rcu_sched(&pwq->rcu, rcu_free_pwq); |
3767 | ||
3768 | /* | |
3769 | * If we're the last pwq going away, @wq is already dead and no one | |
3770 | * is gonna access it anymore. Free it. | |
3771 | */ | |
6029a918 TH |
3772 | if (is_last) { |
3773 | free_workqueue_attrs(wq->unbound_attrs); | |
8864b4e5 | 3774 | kfree(wq); |
6029a918 | 3775 | } |
8864b4e5 TH |
3776 | } |
3777 | ||
0fbd95aa | 3778 | /** |
699ce097 | 3779 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
0fbd95aa | 3780 | * @pwq: target pool_workqueue |
0fbd95aa | 3781 | * |
699ce097 TH |
3782 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
3783 | * workqueue's saved_max_active and activate delayed work items | |
3784 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. | |
0fbd95aa | 3785 | */ |
699ce097 | 3786 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
0fbd95aa | 3787 | { |
699ce097 TH |
3788 | struct workqueue_struct *wq = pwq->wq; |
3789 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3790 | ||
3791 | /* for @wq->saved_max_active */ | |
a357fc03 | 3792 | lockdep_assert_held(&wq->mutex); |
699ce097 TH |
3793 | |
3794 | /* fast exit for non-freezable wqs */ | |
3795 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
3796 | return; | |
3797 | ||
a357fc03 | 3798 | spin_lock_irq(&pwq->pool->lock); |
699ce097 TH |
3799 | |
3800 | if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { | |
3801 | pwq->max_active = wq->saved_max_active; | |
0fbd95aa | 3802 | |
699ce097 TH |
3803 | while (!list_empty(&pwq->delayed_works) && |
3804 | pwq->nr_active < pwq->max_active) | |
3805 | pwq_activate_first_delayed(pwq); | |
951a078a LJ |
3806 | |
3807 | /* | |
3808 | * Need to kick a worker after thawed or an unbound wq's | |
3809 | * max_active is bumped. It's a slow path. Do it always. | |
3810 | */ | |
3811 | wake_up_worker(pwq->pool); | |
699ce097 TH |
3812 | } else { |
3813 | pwq->max_active = 0; | |
3814 | } | |
3815 | ||
a357fc03 | 3816 | spin_unlock_irq(&pwq->pool->lock); |
0fbd95aa TH |
3817 | } |
3818 | ||
e50aba9a | 3819 | /* initialize newly alloced @pwq which is associated with @wq and @pool */ |
f147f29e TH |
3820 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
3821 | struct worker_pool *pool) | |
d2c1d404 TH |
3822 | { |
3823 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); | |
3824 | ||
e50aba9a TH |
3825 | memset(pwq, 0, sizeof(*pwq)); |
3826 | ||
d2c1d404 TH |
3827 | pwq->pool = pool; |
3828 | pwq->wq = wq; | |
3829 | pwq->flush_color = -1; | |
8864b4e5 | 3830 | pwq->refcnt = 1; |
d2c1d404 | 3831 | INIT_LIST_HEAD(&pwq->delayed_works); |
1befcf30 | 3832 | INIT_LIST_HEAD(&pwq->pwqs_node); |
d2c1d404 | 3833 | INIT_LIST_HEAD(&pwq->mayday_node); |
8864b4e5 | 3834 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); |
f147f29e | 3835 | } |
d2c1d404 | 3836 | |
f147f29e | 3837 | /* sync @pwq with the current state of its associated wq and link it */ |
1befcf30 | 3838 | static void link_pwq(struct pool_workqueue *pwq) |
f147f29e TH |
3839 | { |
3840 | struct workqueue_struct *wq = pwq->wq; | |
3841 | ||
3842 | lockdep_assert_held(&wq->mutex); | |
75ccf595 | 3843 | |
1befcf30 TH |
3844 | /* may be called multiple times, ignore if already linked */ |
3845 | if (!list_empty(&pwq->pwqs_node)) | |
3846 | return; | |
3847 | ||
983ca25e TH |
3848 | /* |
3849 | * Set the matching work_color. This is synchronized with | |
3c25a55d | 3850 | * wq->mutex to avoid confusing flush_workqueue(). |
983ca25e | 3851 | */ |
75ccf595 | 3852 | pwq->work_color = wq->work_color; |
983ca25e TH |
3853 | |
3854 | /* sync max_active to the current setting */ | |
3855 | pwq_adjust_max_active(pwq); | |
3856 | ||
3857 | /* link in @pwq */ | |
9e8cd2f5 | 3858 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); |
f147f29e | 3859 | } |
a357fc03 | 3860 | |
f147f29e TH |
3861 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3862 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3863 | const struct workqueue_attrs *attrs) | |
3864 | { | |
3865 | struct worker_pool *pool; | |
3866 | struct pool_workqueue *pwq; | |
3867 | ||
3868 | lockdep_assert_held(&wq_pool_mutex); | |
3869 | ||
3870 | pool = get_unbound_pool(attrs); | |
3871 | if (!pool) | |
3872 | return NULL; | |
3873 | ||
e50aba9a | 3874 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
f147f29e TH |
3875 | if (!pwq) { |
3876 | put_unbound_pool(pool); | |
3877 | return NULL; | |
df2d5ae4 | 3878 | } |
6029a918 | 3879 | |
f147f29e TH |
3880 | init_pwq(pwq, wq, pool); |
3881 | return pwq; | |
d2c1d404 TH |
3882 | } |
3883 | ||
4c16bd32 TH |
3884 | /* undo alloc_unbound_pwq(), used only in the error path */ |
3885 | static void free_unbound_pwq(struct pool_workqueue *pwq) | |
3886 | { | |
3887 | lockdep_assert_held(&wq_pool_mutex); | |
3888 | ||
3889 | if (pwq) { | |
3890 | put_unbound_pool(pwq->pool); | |
cece95df | 3891 | kmem_cache_free(pwq_cache, pwq); |
4c16bd32 TH |
3892 | } |
3893 | } | |
3894 | ||
3895 | /** | |
3896 | * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node | |
3897 | * @attrs: the wq_attrs of interest | |
3898 | * @node: the target NUMA node | |
3899 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3900 | * @cpumask: outarg, the resulting cpumask | |
3901 | * | |
3902 | * Calculate the cpumask a workqueue with @attrs should use on @node. If | |
3903 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
3904 | * calculation. The result is stored in @cpumask. This function returns | |
3905 | * %true if the resulting @cpumask is different from @attrs->cpumask, | |
3906 | * %false if equal. | |
3907 | * | |
3908 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If | |
3909 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3910 | * cpumask is the intersection of the possible CPUs of @node and | |
3911 | * @attrs->cpumask. | |
3912 | * | |
3913 | * The caller is responsible for ensuring that the cpumask of @node stays | |
3914 | * stable. | |
3915 | */ | |
3916 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, | |
3917 | int cpu_going_down, cpumask_t *cpumask) | |
3918 | { | |
d55262c4 | 3919 | if (!wq_numa_enabled || attrs->no_numa) |
4c16bd32 TH |
3920 | goto use_dfl; |
3921 | ||
3922 | /* does @node have any online CPUs @attrs wants? */ | |
3923 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
3924 | if (cpu_going_down >= 0) | |
3925 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
3926 | ||
3927 | if (cpumask_empty(cpumask)) | |
3928 | goto use_dfl; | |
3929 | ||
3930 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
3931 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
3932 | return !cpumask_equal(cpumask, attrs->cpumask); | |
3933 | ||
3934 | use_dfl: | |
3935 | cpumask_copy(cpumask, attrs->cpumask); | |
3936 | return false; | |
3937 | } | |
3938 | ||
1befcf30 TH |
3939 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
3940 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
3941 | int node, | |
3942 | struct pool_workqueue *pwq) | |
3943 | { | |
3944 | struct pool_workqueue *old_pwq; | |
3945 | ||
3946 | lockdep_assert_held(&wq->mutex); | |
3947 | ||
3948 | /* link_pwq() can handle duplicate calls */ | |
3949 | link_pwq(pwq); | |
3950 | ||
3951 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
3952 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
3953 | return old_pwq; | |
3954 | } | |
3955 | ||
9e8cd2f5 TH |
3956 | /** |
3957 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
3958 | * @wq: the target workqueue | |
3959 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
3960 | * | |
4c16bd32 TH |
3961 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA |
3962 | * machines, this function maps a separate pwq to each NUMA node with | |
3963 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
3964 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
3965 | * items finish. Note that a work item which repeatedly requeues itself | |
3966 | * back-to-back will stay on its current pwq. | |
9e8cd2f5 TH |
3967 | * |
3968 | * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on | |
3969 | * failure. | |
3970 | */ | |
3971 | int apply_workqueue_attrs(struct workqueue_struct *wq, | |
3972 | const struct workqueue_attrs *attrs) | |
3973 | { | |
4c16bd32 TH |
3974 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
3975 | struct pool_workqueue **pwq_tbl, *dfl_pwq; | |
f147f29e | 3976 | int node, ret; |
9e8cd2f5 | 3977 | |
8719dcea | 3978 | /* only unbound workqueues can change attributes */ |
9e8cd2f5 TH |
3979 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) |
3980 | return -EINVAL; | |
3981 | ||
8719dcea | 3982 | /* creating multiple pwqs breaks ordering guarantee */ |
f3538ebd TH |
3983 | if (!list_empty(&wq->pwqs)) { |
3984 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
3985 | return -EINVAL; | |
3986 | ||
3987 | wq->flags &= ~__WQ_ORDERED; | |
3988 | } | |
8719dcea | 3989 | |
4c16bd32 | 3990 | pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); |
13e2e556 | 3991 | new_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
4c16bd32 TH |
3992 | tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
3993 | if (!pwq_tbl || !new_attrs || !tmp_attrs) | |
13e2e556 TH |
3994 | goto enomem; |
3995 | ||
4c16bd32 | 3996 | /* make a copy of @attrs and sanitize it */ |
13e2e556 TH |
3997 | copy_workqueue_attrs(new_attrs, attrs); |
3998 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
3999 | ||
4c16bd32 TH |
4000 | /* |
4001 | * We may create multiple pwqs with differing cpumasks. Make a | |
4002 | * copy of @new_attrs which will be modified and used to obtain | |
4003 | * pools. | |
4004 | */ | |
4005 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
4006 | ||
4007 | /* | |
4008 | * CPUs should stay stable across pwq creations and installations. | |
4009 | * Pin CPUs, determine the target cpumask for each node and create | |
4010 | * pwqs accordingly. | |
4011 | */ | |
4012 | get_online_cpus(); | |
4013 | ||
a892cacc | 4014 | mutex_lock(&wq_pool_mutex); |
4c16bd32 TH |
4015 | |
4016 | /* | |
4017 | * If something goes wrong during CPU up/down, we'll fall back to | |
4018 | * the default pwq covering whole @attrs->cpumask. Always create | |
4019 | * it even if we don't use it immediately. | |
4020 | */ | |
4021 | dfl_pwq = alloc_unbound_pwq(wq, new_attrs); | |
4022 | if (!dfl_pwq) | |
4023 | goto enomem_pwq; | |
4024 | ||
4025 | for_each_node(node) { | |
4026 | if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { | |
4027 | pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); | |
4028 | if (!pwq_tbl[node]) | |
4029 | goto enomem_pwq; | |
4030 | } else { | |
4031 | dfl_pwq->refcnt++; | |
4032 | pwq_tbl[node] = dfl_pwq; | |
4033 | } | |
4034 | } | |
4035 | ||
f147f29e | 4036 | mutex_unlock(&wq_pool_mutex); |
9e8cd2f5 | 4037 | |
4c16bd32 | 4038 | /* all pwqs have been created successfully, let's install'em */ |
f147f29e | 4039 | mutex_lock(&wq->mutex); |
a892cacc | 4040 | |
f147f29e | 4041 | copy_workqueue_attrs(wq->unbound_attrs, new_attrs); |
4c16bd32 TH |
4042 | |
4043 | /* save the previous pwq and install the new one */ | |
f147f29e | 4044 | for_each_node(node) |
4c16bd32 TH |
4045 | pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); |
4046 | ||
4047 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
4048 | link_pwq(dfl_pwq); | |
4049 | swap(wq->dfl_pwq, dfl_pwq); | |
f147f29e TH |
4050 | |
4051 | mutex_unlock(&wq->mutex); | |
9e8cd2f5 | 4052 | |
4c16bd32 TH |
4053 | /* put the old pwqs */ |
4054 | for_each_node(node) | |
4055 | put_pwq_unlocked(pwq_tbl[node]); | |
4056 | put_pwq_unlocked(dfl_pwq); | |
4057 | ||
4058 | put_online_cpus(); | |
4862125b TH |
4059 | ret = 0; |
4060 | /* fall through */ | |
4061 | out_free: | |
4c16bd32 | 4062 | free_workqueue_attrs(tmp_attrs); |
4862125b | 4063 | free_workqueue_attrs(new_attrs); |
4c16bd32 | 4064 | kfree(pwq_tbl); |
4862125b | 4065 | return ret; |
13e2e556 | 4066 | |
4c16bd32 TH |
4067 | enomem_pwq: |
4068 | free_unbound_pwq(dfl_pwq); | |
4069 | for_each_node(node) | |
4070 | if (pwq_tbl && pwq_tbl[node] != dfl_pwq) | |
4071 | free_unbound_pwq(pwq_tbl[node]); | |
4072 | mutex_unlock(&wq_pool_mutex); | |
4073 | put_online_cpus(); | |
13e2e556 | 4074 | enomem: |
4862125b TH |
4075 | ret = -ENOMEM; |
4076 | goto out_free; | |
9e8cd2f5 TH |
4077 | } |
4078 | ||
4c16bd32 TH |
4079 | /** |
4080 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
4081 | * @wq: the target workqueue | |
4082 | * @cpu: the CPU coming up or going down | |
4083 | * @online: whether @cpu is coming up or going down | |
4084 | * | |
4085 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
4086 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
4087 | * @wq accordingly. | |
4088 | * | |
4089 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
4090 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
4091 | * correct. | |
4092 | * | |
4093 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
4094 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
4095 | * already executing the work items for the workqueue will lose their CPU | |
4096 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
4097 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
4098 | * affinity, it's the user's responsibility to flush the work item from | |
4099 | * CPU_DOWN_PREPARE. | |
4100 | */ | |
4101 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
4102 | bool online) | |
4103 | { | |
4104 | int node = cpu_to_node(cpu); | |
4105 | int cpu_off = online ? -1 : cpu; | |
4106 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
4107 | struct workqueue_attrs *target_attrs; | |
4108 | cpumask_t *cpumask; | |
4109 | ||
4110 | lockdep_assert_held(&wq_pool_mutex); | |
4111 | ||
4112 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) | |
4113 | return; | |
4114 | ||
4115 | /* | |
4116 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4117 | * Let's use a preallocated one. The following buf is protected by | |
4118 | * CPU hotplug exclusion. | |
4119 | */ | |
4120 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
4121 | cpumask = target_attrs->cpumask; | |
4122 | ||
4123 | mutex_lock(&wq->mutex); | |
d55262c4 TH |
4124 | if (wq->unbound_attrs->no_numa) |
4125 | goto out_unlock; | |
4c16bd32 TH |
4126 | |
4127 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); | |
4128 | pwq = unbound_pwq_by_node(wq, node); | |
4129 | ||
4130 | /* | |
4131 | * Let's determine what needs to be done. If the target cpumask is | |
4132 | * different from wq's, we need to compare it to @pwq's and create | |
4133 | * a new one if they don't match. If the target cpumask equals | |
4134 | * wq's, the default pwq should be used. If @pwq is already the | |
4135 | * default one, nothing to do; otherwise, install the default one. | |
4136 | */ | |
4137 | if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { | |
4138 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) | |
4139 | goto out_unlock; | |
4140 | } else { | |
4141 | if (pwq == wq->dfl_pwq) | |
4142 | goto out_unlock; | |
4143 | else | |
4144 | goto use_dfl_pwq; | |
4145 | } | |
4146 | ||
4147 | mutex_unlock(&wq->mutex); | |
4148 | ||
4149 | /* create a new pwq */ | |
4150 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4151 | if (!pwq) { | |
4152 | pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", | |
4153 | wq->name); | |
55a3dfcc DY |
4154 | mutex_lock(&wq->mutex); |
4155 | goto use_dfl_pwq; | |
4c16bd32 TH |
4156 | } |
4157 | ||
4158 | /* | |
4159 | * Install the new pwq. As this function is called only from CPU | |
4160 | * hotplug callbacks and applying a new attrs is wrapped with | |
4161 | * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed | |
4162 | * inbetween. | |
4163 | */ | |
4164 | mutex_lock(&wq->mutex); | |
4165 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
4166 | goto out_unlock; | |
4167 | ||
4168 | use_dfl_pwq: | |
4169 | spin_lock_irq(&wq->dfl_pwq->pool->lock); | |
4170 | get_pwq(wq->dfl_pwq); | |
4171 | spin_unlock_irq(&wq->dfl_pwq->pool->lock); | |
4172 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); | |
4173 | out_unlock: | |
4174 | mutex_unlock(&wq->mutex); | |
4175 | put_pwq_unlocked(old_pwq); | |
4176 | } | |
4177 | ||
30cdf249 | 4178 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4179 | { |
49e3cf44 | 4180 | bool highpri = wq->flags & WQ_HIGHPRI; |
ced4ac92 | 4181 | int cpu, ret; |
30cdf249 TH |
4182 | |
4183 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
4184 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
4185 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
4186 | return -ENOMEM; |
4187 | ||
4188 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
4189 | struct pool_workqueue *pwq = |
4190 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 4191 | struct worker_pool *cpu_pools = |
f02ae73a | 4192 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 4193 | |
f147f29e TH |
4194 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
4195 | ||
4196 | mutex_lock(&wq->mutex); | |
1befcf30 | 4197 | link_pwq(pwq); |
f147f29e | 4198 | mutex_unlock(&wq->mutex); |
30cdf249 | 4199 | } |
9e8cd2f5 | 4200 | return 0; |
ced4ac92 TH |
4201 | } else if (wq->flags & __WQ_ORDERED) { |
4202 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); | |
4203 | /* there should only be single pwq for ordering guarantee */ | |
4204 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
4205 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
4206 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
4207 | return ret; | |
30cdf249 | 4208 | } else { |
9e8cd2f5 | 4209 | return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 4210 | } |
0f900049 TH |
4211 | } |
4212 | ||
f3421797 TH |
4213 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
4214 | const char *name) | |
b71ab8c2 | 4215 | { |
f3421797 TH |
4216 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
4217 | ||
4218 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
4219 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
4220 | max_active, name, 1, lim); | |
b71ab8c2 | 4221 | |
f3421797 | 4222 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
4223 | } |
4224 | ||
b196be89 | 4225 | struct workqueue_struct *__alloc_workqueue_key(const char *fmt, |
d320c038 TH |
4226 | unsigned int flags, |
4227 | int max_active, | |
4228 | struct lock_class_key *key, | |
b196be89 | 4229 | const char *lock_name, ...) |
1da177e4 | 4230 | { |
df2d5ae4 | 4231 | size_t tbl_size = 0; |
ecf6881f | 4232 | va_list args; |
1da177e4 | 4233 | struct workqueue_struct *wq; |
49e3cf44 | 4234 | struct pool_workqueue *pwq; |
b196be89 | 4235 | |
7a1dc940 TH |
4236 | /* |
4237 | * Unbound && max_active == 1 used to imply ordered, which is no | |
4238 | * longer the case on NUMA machines due to per-node pools. While | |
4239 | * alloc_ordered_workqueue() is the right way to create an ordered | |
4240 | * workqueue, keep the previous behavior to avoid subtle breakages | |
4241 | * on NUMA. | |
4242 | */ | |
4243 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
4244 | flags |= __WQ_ORDERED; | |
4245 | ||
3c2a0909 S |
4246 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4247 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4248 | flags |= WQ_UNBOUND; | |
4249 | ||
ecf6881f | 4250 | /* allocate wq and format name */ |
df2d5ae4 TH |
4251 | if (flags & WQ_UNBOUND) |
4252 | tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); | |
4253 | ||
4254 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 4255 | if (!wq) |
d2c1d404 | 4256 | return NULL; |
b196be89 | 4257 | |
6029a918 TH |
4258 | if (flags & WQ_UNBOUND) { |
4259 | wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
4260 | if (!wq->unbound_attrs) | |
4261 | goto err_free_wq; | |
4262 | } | |
4263 | ||
ecf6881f TH |
4264 | va_start(args, lock_name); |
4265 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); | |
b196be89 | 4266 | va_end(args); |
1da177e4 | 4267 | |
d320c038 | 4268 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4269 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4270 | |
b196be89 | 4271 | /* init wq */ |
97e37d7b | 4272 | wq->flags = flags; |
a0a1a5fd | 4273 | wq->saved_max_active = max_active; |
3c25a55d | 4274 | mutex_init(&wq->mutex); |
112202d9 | 4275 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4276 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4277 | INIT_LIST_HEAD(&wq->flusher_queue); |
4278 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4279 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4280 | |
eb13ba87 | 4281 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
cce1a165 | 4282 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4283 | |
30cdf249 | 4284 | if (alloc_and_link_pwqs(wq) < 0) |
d2c1d404 | 4285 | goto err_free_wq; |
1537663f | 4286 | |
493008a8 TH |
4287 | /* |
4288 | * Workqueues which may be used during memory reclaim should | |
4289 | * have a rescuer to guarantee forward progress. | |
4290 | */ | |
4291 | if (flags & WQ_MEM_RECLAIM) { | |
e22bee78 TH |
4292 | struct worker *rescuer; |
4293 | ||
d2c1d404 | 4294 | rescuer = alloc_worker(); |
e22bee78 | 4295 | if (!rescuer) |
d2c1d404 | 4296 | goto err_destroy; |
e22bee78 | 4297 | |
111c225a TH |
4298 | rescuer->rescue_wq = wq; |
4299 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", | |
b196be89 | 4300 | wq->name); |
d2c1d404 TH |
4301 | if (IS_ERR(rescuer->task)) { |
4302 | kfree(rescuer); | |
4303 | goto err_destroy; | |
4304 | } | |
e22bee78 | 4305 | |
d2c1d404 | 4306 | wq->rescuer = rescuer; |
14a40ffc | 4307 | rescuer->task->flags |= PF_NO_SETAFFINITY; |
e22bee78 | 4308 | wake_up_process(rescuer->task); |
3af24433 ON |
4309 | } |
4310 | ||
226223ab TH |
4311 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4312 | goto err_destroy; | |
4313 | ||
a0a1a5fd | 4314 | /* |
68e13a67 LJ |
4315 | * wq_pool_mutex protects global freeze state and workqueues list. |
4316 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4317 | * list. | |
a0a1a5fd | 4318 | */ |
68e13a67 | 4319 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4320 | |
a357fc03 | 4321 | mutex_lock(&wq->mutex); |
699ce097 TH |
4322 | for_each_pwq(pwq, wq) |
4323 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4324 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4325 | |
1537663f | 4326 | list_add(&wq->list, &workqueues); |
a0a1a5fd | 4327 | |
68e13a67 | 4328 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4329 | |
3af24433 | 4330 | return wq; |
d2c1d404 TH |
4331 | |
4332 | err_free_wq: | |
6029a918 | 4333 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4334 | kfree(wq); |
4335 | return NULL; | |
4336 | err_destroy: | |
4337 | destroy_workqueue(wq); | |
4690c4ab | 4338 | return NULL; |
3af24433 | 4339 | } |
d320c038 | 4340 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
1da177e4 | 4341 | |
3af24433 ON |
4342 | /** |
4343 | * destroy_workqueue - safely terminate a workqueue | |
4344 | * @wq: target workqueue | |
4345 | * | |
4346 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4347 | */ | |
4348 | void destroy_workqueue(struct workqueue_struct *wq) | |
4349 | { | |
49e3cf44 | 4350 | struct pool_workqueue *pwq; |
4c16bd32 | 4351 | int node; |
3af24433 | 4352 | |
9c5a2ba7 TH |
4353 | /* drain it before proceeding with destruction */ |
4354 | drain_workqueue(wq); | |
c8efcc25 | 4355 | |
6183c009 | 4356 | /* sanity checks */ |
b09f4fd3 | 4357 | mutex_lock(&wq->mutex); |
49e3cf44 | 4358 | for_each_pwq(pwq, wq) { |
6183c009 TH |
4359 | int i; |
4360 | ||
76af4d93 TH |
4361 | for (i = 0; i < WORK_NR_COLORS; i++) { |
4362 | if (WARN_ON(pwq->nr_in_flight[i])) { | |
b09f4fd3 | 4363 | mutex_unlock(&wq->mutex); |
6183c009 | 4364 | return; |
76af4d93 TH |
4365 | } |
4366 | } | |
4367 | ||
5c529597 | 4368 | if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || |
8864b4e5 | 4369 | WARN_ON(pwq->nr_active) || |
76af4d93 | 4370 | WARN_ON(!list_empty(&pwq->delayed_works))) { |
b09f4fd3 | 4371 | mutex_unlock(&wq->mutex); |
6183c009 | 4372 | return; |
76af4d93 | 4373 | } |
6183c009 | 4374 | } |
b09f4fd3 | 4375 | mutex_unlock(&wq->mutex); |
6183c009 | 4376 | |
a0a1a5fd TH |
4377 | /* |
4378 | * wq list is used to freeze wq, remove from list after | |
4379 | * flushing is complete in case freeze races us. | |
4380 | */ | |
68e13a67 | 4381 | mutex_lock(&wq_pool_mutex); |
d2c1d404 | 4382 | list_del_init(&wq->list); |
68e13a67 | 4383 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4384 | |
226223ab TH |
4385 | workqueue_sysfs_unregister(wq); |
4386 | ||
493008a8 | 4387 | if (wq->rescuer) { |
e22bee78 | 4388 | kthread_stop(wq->rescuer->task); |
8d9df9f0 | 4389 | kfree(wq->rescuer); |
493008a8 | 4390 | wq->rescuer = NULL; |
e22bee78 TH |
4391 | } |
4392 | ||
8864b4e5 TH |
4393 | if (!(wq->flags & WQ_UNBOUND)) { |
4394 | /* | |
4395 | * The base ref is never dropped on per-cpu pwqs. Directly | |
4396 | * free the pwqs and wq. | |
4397 | */ | |
4398 | free_percpu(wq->cpu_pwqs); | |
4399 | kfree(wq); | |
4400 | } else { | |
4401 | /* | |
4402 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4403 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4404 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4405 | */ |
4c16bd32 TH |
4406 | for_each_node(node) { |
4407 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4408 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4409 | put_pwq_unlocked(pwq); | |
4410 | } | |
4411 | ||
4412 | /* | |
4413 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4414 | * put. Don't access it afterwards. | |
4415 | */ | |
4416 | pwq = wq->dfl_pwq; | |
4417 | wq->dfl_pwq = NULL; | |
dce90d47 | 4418 | put_pwq_unlocked(pwq); |
29c91e99 | 4419 | } |
3af24433 ON |
4420 | } |
4421 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4422 | ||
dcd989cb TH |
4423 | /** |
4424 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4425 | * @wq: target workqueue | |
4426 | * @max_active: new max_active value. | |
4427 | * | |
4428 | * Set max_active of @wq to @max_active. | |
4429 | * | |
4430 | * CONTEXT: | |
4431 | * Don't call from IRQ context. | |
4432 | */ | |
4433 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4434 | { | |
49e3cf44 | 4435 | struct pool_workqueue *pwq; |
dcd989cb | 4436 | |
8719dcea | 4437 | /* disallow meddling with max_active for ordered workqueues */ |
f3538ebd | 4438 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
4439 | return; |
4440 | ||
f3421797 | 4441 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4442 | |
a357fc03 | 4443 | mutex_lock(&wq->mutex); |
dcd989cb | 4444 | |
f3538ebd | 4445 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb TH |
4446 | wq->saved_max_active = max_active; |
4447 | ||
699ce097 TH |
4448 | for_each_pwq(pwq, wq) |
4449 | pwq_adjust_max_active(pwq); | |
93981800 | 4450 | |
a357fc03 | 4451 | mutex_unlock(&wq->mutex); |
15316ba8 | 4452 | } |
dcd989cb | 4453 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4454 | |
e6267616 TH |
4455 | /** |
4456 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4457 | * | |
4458 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4459 | * work functions to determine whether it's being run off the rescuer task. | |
4460 | */ | |
4461 | bool current_is_workqueue_rescuer(void) | |
4462 | { | |
4463 | struct worker *worker = current_wq_worker(); | |
4464 | ||
6a092dfd | 4465 | return worker && worker->rescue_wq; |
e6267616 TH |
4466 | } |
4467 | ||
eef6a7d5 | 4468 | /** |
dcd989cb TH |
4469 | * workqueue_congested - test whether a workqueue is congested |
4470 | * @cpu: CPU in question | |
4471 | * @wq: target workqueue | |
eef6a7d5 | 4472 | * |
dcd989cb TH |
4473 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4474 | * no synchronization around this function and the test result is | |
4475 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4476 | * |
d3251859 TH |
4477 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4478 | * Note that both per-cpu and unbound workqueues may be associated with | |
4479 | * multiple pool_workqueues which have separate congested states. A | |
4480 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4481 | * contested on other CPUs / NUMA nodes. | |
4482 | * | |
dcd989cb TH |
4483 | * RETURNS: |
4484 | * %true if congested, %false otherwise. | |
eef6a7d5 | 4485 | */ |
d84ff051 | 4486 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4487 | { |
7fb98ea7 | 4488 | struct pool_workqueue *pwq; |
76af4d93 TH |
4489 | bool ret; |
4490 | ||
88109453 | 4491 | rcu_read_lock_sched(); |
7fb98ea7 | 4492 | |
d3251859 TH |
4493 | if (cpu == WORK_CPU_UNBOUND) |
4494 | cpu = smp_processor_id(); | |
4495 | ||
7fb98ea7 TH |
4496 | if (!(wq->flags & WQ_UNBOUND)) |
4497 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4498 | else | |
df2d5ae4 | 4499 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4500 | |
76af4d93 | 4501 | ret = !list_empty(&pwq->delayed_works); |
88109453 | 4502 | rcu_read_unlock_sched(); |
76af4d93 TH |
4503 | |
4504 | return ret; | |
1da177e4 | 4505 | } |
dcd989cb | 4506 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4507 | |
dcd989cb TH |
4508 | /** |
4509 | * work_busy - test whether a work is currently pending or running | |
4510 | * @work: the work to be tested | |
4511 | * | |
4512 | * Test whether @work is currently pending or running. There is no | |
4513 | * synchronization around this function and the test result is | |
4514 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb TH |
4515 | * |
4516 | * RETURNS: | |
4517 | * OR'd bitmask of WORK_BUSY_* bits. | |
4518 | */ | |
4519 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4520 | { |
fa1b54e6 | 4521 | struct worker_pool *pool; |
dcd989cb TH |
4522 | unsigned long flags; |
4523 | unsigned int ret = 0; | |
1da177e4 | 4524 | |
dcd989cb TH |
4525 | if (work_pending(work)) |
4526 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4527 | |
fa1b54e6 TH |
4528 | local_irq_save(flags); |
4529 | pool = get_work_pool(work); | |
038366c5 | 4530 | if (pool) { |
fa1b54e6 | 4531 | spin_lock(&pool->lock); |
038366c5 LJ |
4532 | if (find_worker_executing_work(pool, work)) |
4533 | ret |= WORK_BUSY_RUNNING; | |
fa1b54e6 | 4534 | spin_unlock(&pool->lock); |
038366c5 | 4535 | } |
fa1b54e6 | 4536 | local_irq_restore(flags); |
1da177e4 | 4537 | |
dcd989cb | 4538 | return ret; |
1da177e4 | 4539 | } |
dcd989cb | 4540 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4541 | |
3d1cb205 TH |
4542 | /** |
4543 | * set_worker_desc - set description for the current work item | |
4544 | * @fmt: printf-style format string | |
4545 | * @...: arguments for the format string | |
4546 | * | |
4547 | * This function can be called by a running work function to describe what | |
4548 | * the work item is about. If the worker task gets dumped, this | |
4549 | * information will be printed out together to help debugging. The | |
4550 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4551 | */ | |
4552 | void set_worker_desc(const char *fmt, ...) | |
4553 | { | |
4554 | struct worker *worker = current_wq_worker(); | |
4555 | va_list args; | |
4556 | ||
4557 | if (worker) { | |
4558 | va_start(args, fmt); | |
4559 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4560 | va_end(args); | |
4561 | worker->desc_valid = true; | |
4562 | } | |
4563 | } | |
4564 | ||
4565 | /** | |
4566 | * print_worker_info - print out worker information and description | |
4567 | * @log_lvl: the log level to use when printing | |
4568 | * @task: target task | |
4569 | * | |
4570 | * If @task is a worker and currently executing a work item, print out the | |
4571 | * name of the workqueue being serviced and worker description set with | |
4572 | * set_worker_desc() by the currently executing work item. | |
4573 | * | |
4574 | * This function can be safely called on any task as long as the | |
4575 | * task_struct itself is accessible. While safe, this function isn't | |
4576 | * synchronized and may print out mixups or garbages of limited length. | |
4577 | */ | |
4578 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4579 | { | |
4580 | work_func_t *fn = NULL; | |
4581 | char name[WQ_NAME_LEN] = { }; | |
4582 | char desc[WORKER_DESC_LEN] = { }; | |
4583 | struct pool_workqueue *pwq = NULL; | |
4584 | struct workqueue_struct *wq = NULL; | |
4585 | bool desc_valid = false; | |
4586 | struct worker *worker; | |
4587 | ||
4588 | if (!(task->flags & PF_WQ_WORKER)) | |
4589 | return; | |
4590 | ||
4591 | /* | |
4592 | * This function is called without any synchronization and @task | |
4593 | * could be in any state. Be careful with dereferences. | |
4594 | */ | |
4595 | worker = probe_kthread_data(task); | |
4596 | ||
4597 | /* | |
4598 | * Carefully copy the associated workqueue's workfn and name. Keep | |
4599 | * the original last '\0' in case the original contains garbage. | |
4600 | */ | |
4601 | probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); | |
4602 | probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4603 | probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); | |
4604 | probe_kernel_read(name, wq->name, sizeof(name) - 1); | |
4605 | ||
4606 | /* copy worker description */ | |
4607 | probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); | |
4608 | if (desc_valid) | |
4609 | probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); | |
4610 | ||
4611 | if (fn || name[0] || desc[0]) { | |
4612 | printk("%sWorkqueue: %s %pf", log_lvl, name, fn); | |
4613 | if (desc[0]) | |
4614 | pr_cont(" (%s)", desc); | |
4615 | pr_cont("\n"); | |
4616 | } | |
4617 | } | |
4618 | ||
db7bccf4 TH |
4619 | /* |
4620 | * CPU hotplug. | |
4621 | * | |
e22bee78 | 4622 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 4623 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 4624 | * pool which make migrating pending and scheduled works very |
e22bee78 | 4625 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 4626 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
4627 | * blocked draining impractical. |
4628 | * | |
24647570 | 4629 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
4630 | * running as an unbound one and allowing it to be reattached later if the |
4631 | * cpu comes back online. | |
db7bccf4 | 4632 | */ |
1da177e4 | 4633 | |
706026c2 | 4634 | static void wq_unbind_fn(struct work_struct *work) |
3af24433 | 4635 | { |
38db41d9 | 4636 | int cpu = smp_processor_id(); |
4ce62e9e | 4637 | struct worker_pool *pool; |
db7bccf4 | 4638 | struct worker *worker; |
a9ab775b | 4639 | int wi; |
3af24433 | 4640 | |
f02ae73a | 4641 | for_each_cpu_worker_pool(pool, cpu) { |
6183c009 | 4642 | WARN_ON_ONCE(cpu != smp_processor_id()); |
db7bccf4 | 4643 | |
bc3a1afc | 4644 | mutex_lock(&pool->manager_mutex); |
94cf58bb | 4645 | spin_lock_irq(&pool->lock); |
3af24433 | 4646 | |
94cf58bb | 4647 | /* |
bc3a1afc | 4648 | * We've blocked all manager operations. Make all workers |
94cf58bb TH |
4649 | * unbound and set DISASSOCIATED. Before this, all workers |
4650 | * except for the ones which are still executing works from | |
4651 | * before the last CPU down must be on the cpu. After | |
4652 | * this, they may become diasporas. | |
4653 | */ | |
a9ab775b | 4654 | for_each_pool_worker(worker, wi, pool) |
c9e7cf27 | 4655 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 4656 | |
24647570 | 4657 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 4658 | |
94cf58bb | 4659 | spin_unlock_irq(&pool->lock); |
bc3a1afc | 4660 | mutex_unlock(&pool->manager_mutex); |
628c78e7 | 4661 | |
eb283428 LJ |
4662 | /* |
4663 | * Call schedule() so that we cross rq->lock and thus can | |
4664 | * guarantee sched callbacks see the %WORKER_UNBOUND flag. | |
4665 | * This is necessary as scheduler callbacks may be invoked | |
4666 | * from other cpus. | |
4667 | */ | |
4668 | schedule(); | |
06ba38a9 | 4669 | |
eb283428 LJ |
4670 | /* |
4671 | * Sched callbacks are disabled now. Zap nr_running. | |
4672 | * After this, nr_running stays zero and need_more_worker() | |
4673 | * and keep_working() are always true as long as the | |
4674 | * worklist is not empty. This pool now behaves as an | |
4675 | * unbound (in terms of concurrency management) pool which | |
4676 | * are served by workers tied to the pool. | |
4677 | */ | |
e19e397a | 4678 | atomic_set(&pool->nr_running, 0); |
eb283428 LJ |
4679 | |
4680 | /* | |
4681 | * With concurrency management just turned off, a busy | |
4682 | * worker blocking could lead to lengthy stalls. Kick off | |
4683 | * unbound chain execution of currently pending work items. | |
4684 | */ | |
4685 | spin_lock_irq(&pool->lock); | |
4686 | wake_up_worker(pool); | |
4687 | spin_unlock_irq(&pool->lock); | |
4688 | } | |
3af24433 | 4689 | } |
3af24433 | 4690 | |
bd7c089e TH |
4691 | /** |
4692 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
4693 | * @pool: pool of interest | |
4694 | * | |
a9ab775b | 4695 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
4696 | */ |
4697 | static void rebind_workers(struct worker_pool *pool) | |
4698 | { | |
a9ab775b TH |
4699 | struct worker *worker; |
4700 | int wi; | |
bd7c089e TH |
4701 | |
4702 | lockdep_assert_held(&pool->manager_mutex); | |
bd7c089e | 4703 | |
a9ab775b TH |
4704 | /* |
4705 | * Restore CPU affinity of all workers. As all idle workers should | |
4706 | * be on the run-queue of the associated CPU before any local | |
4707 | * wake-ups for concurrency management happen, restore CPU affinty | |
4708 | * of all workers first and then clear UNBOUND. As we're called | |
4709 | * from CPU_ONLINE, the following shouldn't fail. | |
4710 | */ | |
4711 | for_each_pool_worker(worker, wi, pool) | |
4712 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
4713 | pool->attrs->cpumask) < 0); | |
bd7c089e | 4714 | |
a9ab775b | 4715 | spin_lock_irq(&pool->lock); |
bd7c089e | 4716 | |
a9ab775b TH |
4717 | for_each_pool_worker(worker, wi, pool) { |
4718 | unsigned int worker_flags = worker->flags; | |
bd7c089e TH |
4719 | |
4720 | /* | |
a9ab775b TH |
4721 | * A bound idle worker should actually be on the runqueue |
4722 | * of the associated CPU for local wake-ups targeting it to | |
4723 | * work. Kick all idle workers so that they migrate to the | |
4724 | * associated CPU. Doing this in the same loop as | |
4725 | * replacing UNBOUND with REBOUND is safe as no worker will | |
4726 | * be bound before @pool->lock is released. | |
bd7c089e | 4727 | */ |
a9ab775b TH |
4728 | if (worker_flags & WORKER_IDLE) |
4729 | wake_up_process(worker->task); | |
bd7c089e | 4730 | |
a9ab775b TH |
4731 | /* |
4732 | * We want to clear UNBOUND but can't directly call | |
4733 | * worker_clr_flags() or adjust nr_running. Atomically | |
4734 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
4735 | * @worker will clear REBOUND using worker_clr_flags() when | |
4736 | * it initiates the next execution cycle thus restoring | |
4737 | * concurrency management. Note that when or whether | |
4738 | * @worker clears REBOUND doesn't affect correctness. | |
4739 | * | |
4740 | * ACCESS_ONCE() is necessary because @worker->flags may be | |
4741 | * tested without holding any lock in | |
4742 | * wq_worker_waking_up(). Without it, NOT_RUNNING test may | |
4743 | * fail incorrectly leading to premature concurrency | |
4744 | * management operations. | |
4745 | */ | |
4746 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
4747 | worker_flags |= WORKER_REBOUND; | |
4748 | worker_flags &= ~WORKER_UNBOUND; | |
4749 | ACCESS_ONCE(worker->flags) = worker_flags; | |
bd7c089e | 4750 | } |
a9ab775b TH |
4751 | |
4752 | spin_unlock_irq(&pool->lock); | |
bd7c089e TH |
4753 | } |
4754 | ||
7dbc725e TH |
4755 | /** |
4756 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
4757 | * @pool: unbound pool of interest | |
4758 | * @cpu: the CPU which is coming up | |
4759 | * | |
4760 | * An unbound pool may end up with a cpumask which doesn't have any online | |
4761 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
4762 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
4763 | * online CPU before, cpus_allowed of all its workers should be restored. | |
4764 | */ | |
4765 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
4766 | { | |
4767 | static cpumask_t cpumask; | |
4768 | struct worker *worker; | |
4769 | int wi; | |
4770 | ||
4771 | lockdep_assert_held(&pool->manager_mutex); | |
4772 | ||
4773 | /* is @cpu allowed for @pool? */ | |
4774 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
4775 | return; | |
4776 | ||
4777 | /* is @cpu the only online CPU? */ | |
4778 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); | |
4779 | if (cpumask_weight(&cpumask) != 1) | |
4780 | return; | |
4781 | ||
4782 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
4783 | for_each_pool_worker(worker, wi, pool) | |
4784 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
4785 | pool->attrs->cpumask) < 0); | |
4786 | } | |
4787 | ||
8db25e78 TH |
4788 | /* |
4789 | * Workqueues should be brought up before normal priority CPU notifiers. | |
4790 | * This will be registered high priority CPU notifier. | |
4791 | */ | |
9fdf9b73 | 4792 | static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, |
8db25e78 TH |
4793 | unsigned long action, |
4794 | void *hcpu) | |
3af24433 | 4795 | { |
d84ff051 | 4796 | int cpu = (unsigned long)hcpu; |
4ce62e9e | 4797 | struct worker_pool *pool; |
4c16bd32 | 4798 | struct workqueue_struct *wq; |
7dbc725e | 4799 | int pi; |
3ce63377 | 4800 | |
8db25e78 | 4801 | switch (action & ~CPU_TASKS_FROZEN) { |
3af24433 | 4802 | case CPU_UP_PREPARE: |
f02ae73a | 4803 | for_each_cpu_worker_pool(pool, cpu) { |
3ce63377 TH |
4804 | if (pool->nr_workers) |
4805 | continue; | |
ebf44d16 | 4806 | if (create_and_start_worker(pool) < 0) |
3ce63377 | 4807 | return NOTIFY_BAD; |
3af24433 | 4808 | } |
8db25e78 | 4809 | break; |
3af24433 | 4810 | |
db7bccf4 TH |
4811 | case CPU_DOWN_FAILED: |
4812 | case CPU_ONLINE: | |
68e13a67 | 4813 | mutex_lock(&wq_pool_mutex); |
7dbc725e TH |
4814 | |
4815 | for_each_pool(pool, pi) { | |
bc3a1afc | 4816 | mutex_lock(&pool->manager_mutex); |
94cf58bb | 4817 | |
7dbc725e TH |
4818 | if (pool->cpu == cpu) { |
4819 | spin_lock_irq(&pool->lock); | |
4820 | pool->flags &= ~POOL_DISASSOCIATED; | |
4821 | spin_unlock_irq(&pool->lock); | |
a9ab775b | 4822 | |
7dbc725e TH |
4823 | rebind_workers(pool); |
4824 | } else if (pool->cpu < 0) { | |
4825 | restore_unbound_workers_cpumask(pool, cpu); | |
4826 | } | |
94cf58bb | 4827 | |
bc3a1afc | 4828 | mutex_unlock(&pool->manager_mutex); |
94cf58bb | 4829 | } |
7dbc725e | 4830 | |
4c16bd32 TH |
4831 | /* update NUMA affinity of unbound workqueues */ |
4832 | list_for_each_entry(wq, &workqueues, list) | |
4833 | wq_update_unbound_numa(wq, cpu, true); | |
4834 | ||
68e13a67 | 4835 | mutex_unlock(&wq_pool_mutex); |
db7bccf4 | 4836 | break; |
00dfcaf7 | 4837 | } |
65758202 TH |
4838 | return NOTIFY_OK; |
4839 | } | |
4840 | ||
4841 | /* | |
4842 | * Workqueues should be brought down after normal priority CPU notifiers. | |
4843 | * This will be registered as low priority CPU notifier. | |
4844 | */ | |
9fdf9b73 | 4845 | static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, |
65758202 TH |
4846 | unsigned long action, |
4847 | void *hcpu) | |
4848 | { | |
d84ff051 | 4849 | int cpu = (unsigned long)hcpu; |
8db25e78 | 4850 | struct work_struct unbind_work; |
4c16bd32 | 4851 | struct workqueue_struct *wq; |
8db25e78 | 4852 | |
65758202 TH |
4853 | switch (action & ~CPU_TASKS_FROZEN) { |
4854 | case CPU_DOWN_PREPARE: | |
4c16bd32 | 4855 | /* unbinding per-cpu workers should happen on the local CPU */ |
706026c2 | 4856 | INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); |
7635d2fd | 4857 | queue_work_on(cpu, system_highpri_wq, &unbind_work); |
4c16bd32 TH |
4858 | |
4859 | /* update NUMA affinity of unbound workqueues */ | |
4860 | mutex_lock(&wq_pool_mutex); | |
4861 | list_for_each_entry(wq, &workqueues, list) | |
4862 | wq_update_unbound_numa(wq, cpu, false); | |
4863 | mutex_unlock(&wq_pool_mutex); | |
4864 | ||
4865 | /* wait for per-cpu unbinding to finish */ | |
8db25e78 TH |
4866 | flush_work(&unbind_work); |
4867 | break; | |
65758202 TH |
4868 | } |
4869 | return NOTIFY_OK; | |
4870 | } | |
4871 | ||
2d3854a3 | 4872 | #ifdef CONFIG_SMP |
8ccad40d | 4873 | |
2d3854a3 | 4874 | struct work_for_cpu { |
ed48ece2 | 4875 | struct work_struct work; |
2d3854a3 RR |
4876 | long (*fn)(void *); |
4877 | void *arg; | |
4878 | long ret; | |
4879 | }; | |
4880 | ||
ed48ece2 | 4881 | static void work_for_cpu_fn(struct work_struct *work) |
2d3854a3 | 4882 | { |
ed48ece2 TH |
4883 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); |
4884 | ||
2d3854a3 RR |
4885 | wfc->ret = wfc->fn(wfc->arg); |
4886 | } | |
4887 | ||
4888 | /** | |
4889 | * work_on_cpu - run a function in user context on a particular cpu | |
4890 | * @cpu: the cpu to run on | |
4891 | * @fn: the function to run | |
4892 | * @arg: the function arg | |
4893 | * | |
31ad9081 RR |
4894 | * This will return the value @fn returns. |
4895 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
6b44003e | 4896 | * The caller must not hold any locks which would prevent @fn from completing. |
2d3854a3 | 4897 | */ |
d84ff051 | 4898 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) |
2d3854a3 | 4899 | { |
ed48ece2 | 4900 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; |
6b44003e | 4901 | |
ed48ece2 TH |
4902 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); |
4903 | schedule_work_on(cpu, &wfc.work); | |
4904 | flush_work(&wfc.work); | |
2d3854a3 RR |
4905 | return wfc.ret; |
4906 | } | |
4907 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
4908 | #endif /* CONFIG_SMP */ | |
4909 | ||
a0a1a5fd TH |
4910 | #ifdef CONFIG_FREEZER |
4911 | ||
4912 | /** | |
4913 | * freeze_workqueues_begin - begin freezing workqueues | |
4914 | * | |
58a69cb4 | 4915 | * Start freezing workqueues. After this function returns, all freezable |
c5aa87bb | 4916 | * workqueues will queue new works to their delayed_works list instead of |
706026c2 | 4917 | * pool->worklist. |
a0a1a5fd TH |
4918 | * |
4919 | * CONTEXT: | |
a357fc03 | 4920 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. |
a0a1a5fd TH |
4921 | */ |
4922 | void freeze_workqueues_begin(void) | |
4923 | { | |
17116969 | 4924 | struct worker_pool *pool; |
24b8a847 TH |
4925 | struct workqueue_struct *wq; |
4926 | struct pool_workqueue *pwq; | |
611c92a0 | 4927 | int pi; |
a0a1a5fd | 4928 | |
68e13a67 | 4929 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4930 | |
6183c009 | 4931 | WARN_ON_ONCE(workqueue_freezing); |
a0a1a5fd TH |
4932 | workqueue_freezing = true; |
4933 | ||
24b8a847 | 4934 | /* set FREEZING */ |
611c92a0 | 4935 | for_each_pool(pool, pi) { |
5bcab335 | 4936 | spin_lock_irq(&pool->lock); |
17116969 TH |
4937 | WARN_ON_ONCE(pool->flags & POOL_FREEZING); |
4938 | pool->flags |= POOL_FREEZING; | |
5bcab335 | 4939 | spin_unlock_irq(&pool->lock); |
24b8a847 | 4940 | } |
a0a1a5fd | 4941 | |
24b8a847 | 4942 | list_for_each_entry(wq, &workqueues, list) { |
a357fc03 | 4943 | mutex_lock(&wq->mutex); |
699ce097 TH |
4944 | for_each_pwq(pwq, wq) |
4945 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4946 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4947 | } |
5bcab335 | 4948 | |
68e13a67 | 4949 | mutex_unlock(&wq_pool_mutex); |
a0a1a5fd TH |
4950 | } |
4951 | ||
4952 | /** | |
58a69cb4 | 4953 | * freeze_workqueues_busy - are freezable workqueues still busy? |
a0a1a5fd TH |
4954 | * |
4955 | * Check whether freezing is complete. This function must be called | |
4956 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
4957 | * | |
4958 | * CONTEXT: | |
68e13a67 | 4959 | * Grabs and releases wq_pool_mutex. |
a0a1a5fd TH |
4960 | * |
4961 | * RETURNS: | |
58a69cb4 TH |
4962 | * %true if some freezable workqueues are still busy. %false if freezing |
4963 | * is complete. | |
a0a1a5fd TH |
4964 | */ |
4965 | bool freeze_workqueues_busy(void) | |
4966 | { | |
a0a1a5fd | 4967 | bool busy = false; |
24b8a847 TH |
4968 | struct workqueue_struct *wq; |
4969 | struct pool_workqueue *pwq; | |
a0a1a5fd | 4970 | |
68e13a67 | 4971 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4972 | |
6183c009 | 4973 | WARN_ON_ONCE(!workqueue_freezing); |
a0a1a5fd | 4974 | |
24b8a847 TH |
4975 | list_for_each_entry(wq, &workqueues, list) { |
4976 | if (!(wq->flags & WQ_FREEZABLE)) | |
4977 | continue; | |
a0a1a5fd TH |
4978 | /* |
4979 | * nr_active is monotonically decreasing. It's safe | |
4980 | * to peek without lock. | |
4981 | */ | |
88109453 | 4982 | rcu_read_lock_sched(); |
24b8a847 | 4983 | for_each_pwq(pwq, wq) { |
6183c009 | 4984 | WARN_ON_ONCE(pwq->nr_active < 0); |
112202d9 | 4985 | if (pwq->nr_active) { |
a0a1a5fd | 4986 | busy = true; |
88109453 | 4987 | rcu_read_unlock_sched(); |
a0a1a5fd TH |
4988 | goto out_unlock; |
4989 | } | |
4990 | } | |
88109453 | 4991 | rcu_read_unlock_sched(); |
a0a1a5fd TH |
4992 | } |
4993 | out_unlock: | |
68e13a67 | 4994 | mutex_unlock(&wq_pool_mutex); |
a0a1a5fd TH |
4995 | return busy; |
4996 | } | |
4997 | ||
4998 | /** | |
4999 | * thaw_workqueues - thaw workqueues | |
5000 | * | |
5001 | * Thaw workqueues. Normal queueing is restored and all collected | |
706026c2 | 5002 | * frozen works are transferred to their respective pool worklists. |
a0a1a5fd TH |
5003 | * |
5004 | * CONTEXT: | |
a357fc03 | 5005 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. |
a0a1a5fd TH |
5006 | */ |
5007 | void thaw_workqueues(void) | |
5008 | { | |
24b8a847 TH |
5009 | struct workqueue_struct *wq; |
5010 | struct pool_workqueue *pwq; | |
5011 | struct worker_pool *pool; | |
611c92a0 | 5012 | int pi; |
a0a1a5fd | 5013 | |
68e13a67 | 5014 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd TH |
5015 | |
5016 | if (!workqueue_freezing) | |
5017 | goto out_unlock; | |
5018 | ||
24b8a847 | 5019 | /* clear FREEZING */ |
611c92a0 | 5020 | for_each_pool(pool, pi) { |
5bcab335 | 5021 | spin_lock_irq(&pool->lock); |
24b8a847 TH |
5022 | WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); |
5023 | pool->flags &= ~POOL_FREEZING; | |
5bcab335 | 5024 | spin_unlock_irq(&pool->lock); |
24b8a847 | 5025 | } |
8b03ae3c | 5026 | |
24b8a847 TH |
5027 | /* restore max_active and repopulate worklist */ |
5028 | list_for_each_entry(wq, &workqueues, list) { | |
a357fc03 | 5029 | mutex_lock(&wq->mutex); |
699ce097 TH |
5030 | for_each_pwq(pwq, wq) |
5031 | pwq_adjust_max_active(pwq); | |
a357fc03 | 5032 | mutex_unlock(&wq->mutex); |
a0a1a5fd TH |
5033 | } |
5034 | ||
5035 | workqueue_freezing = false; | |
5036 | out_unlock: | |
68e13a67 | 5037 | mutex_unlock(&wq_pool_mutex); |
a0a1a5fd TH |
5038 | } |
5039 | #endif /* CONFIG_FREEZER */ | |
5040 | ||
bce90380 TH |
5041 | static void __init wq_numa_init(void) |
5042 | { | |
5043 | cpumask_var_t *tbl; | |
5044 | int node, cpu; | |
5045 | ||
5046 | /* determine NUMA pwq table len - highest node id + 1 */ | |
5047 | for_each_node(node) | |
5048 | wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); | |
5049 | ||
5050 | if (num_possible_nodes() <= 1) | |
5051 | return; | |
5052 | ||
d55262c4 TH |
5053 | if (wq_disable_numa) { |
5054 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
5055 | return; | |
5056 | } | |
5057 | ||
4c16bd32 TH |
5058 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); |
5059 | BUG_ON(!wq_update_unbound_numa_attrs_buf); | |
5060 | ||
bce90380 TH |
5061 | /* |
5062 | * We want masks of possible CPUs of each node which isn't readily | |
5063 | * available. Build one from cpu_to_node() which should have been | |
5064 | * fully initialized by now. | |
5065 | */ | |
5066 | tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); | |
5067 | BUG_ON(!tbl); | |
5068 | ||
5069 | for_each_node(node) | |
3e24998c | 5070 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 5071 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
5072 | |
5073 | for_each_possible_cpu(cpu) { | |
5074 | node = cpu_to_node(cpu); | |
5075 | if (WARN_ON(node == NUMA_NO_NODE)) { | |
5076 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
5077 | /* happens iff arch is bonkers, let's just proceed */ | |
5078 | return; | |
5079 | } | |
5080 | cpumask_set_cpu(cpu, tbl[node]); | |
5081 | } | |
5082 | ||
5083 | wq_numa_possible_cpumask = tbl; | |
5084 | wq_numa_enabled = true; | |
5085 | } | |
5086 | ||
6ee0578b | 5087 | static int __init init_workqueues(void) |
1da177e4 | 5088 | { |
7a4e344c TH |
5089 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
5090 | int i, cpu; | |
c34056a3 | 5091 | |
7c3eed5c TH |
5092 | /* make sure we have enough bits for OFFQ pool ID */ |
5093 | BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < | |
6be19588 | 5094 | WORK_CPU_END * NR_STD_WORKER_POOLS); |
b5490077 | 5095 | |
e904e6c2 TH |
5096 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
5097 | ||
5098 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); | |
5099 | ||
65758202 | 5100 | cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); |
a5b4e57d | 5101 | hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); |
8b03ae3c | 5102 | |
bce90380 TH |
5103 | wq_numa_init(); |
5104 | ||
706026c2 | 5105 | /* initialize CPU pools */ |
29c91e99 | 5106 | for_each_possible_cpu(cpu) { |
4ce62e9e | 5107 | struct worker_pool *pool; |
8b03ae3c | 5108 | |
7a4e344c | 5109 | i = 0; |
f02ae73a | 5110 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 5111 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 5112 | pool->cpu = cpu; |
29c91e99 | 5113 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 5114 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 5115 | pool->node = cpu_to_node(cpu); |
7a4e344c | 5116 | |
9daf9e67 | 5117 | /* alloc pool ID */ |
68e13a67 | 5118 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 5119 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 5120 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 5121 | } |
8b03ae3c TH |
5122 | } |
5123 | ||
e22bee78 | 5124 | /* create the initial worker */ |
29c91e99 | 5125 | for_each_online_cpu(cpu) { |
4ce62e9e | 5126 | struct worker_pool *pool; |
e22bee78 | 5127 | |
f02ae73a | 5128 | for_each_cpu_worker_pool(pool, cpu) { |
29c91e99 | 5129 | pool->flags &= ~POOL_DISASSOCIATED; |
ebf44d16 | 5130 | BUG_ON(create_and_start_worker(pool) < 0); |
4ce62e9e | 5131 | } |
e22bee78 TH |
5132 | } |
5133 | ||
ced4ac92 | 5134 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
5135 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
5136 | struct workqueue_attrs *attrs; | |
5137 | ||
5138 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
29c91e99 | 5139 | attrs->nice = std_nice[i]; |
29c91e99 | 5140 | unbound_std_wq_attrs[i] = attrs; |
ced4ac92 TH |
5141 | |
5142 | /* | |
5143 | * An ordered wq should have only one pwq as ordering is | |
5144 | * guaranteed by max_active which is enforced by pwqs. | |
5145 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
5146 | */ | |
5147 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
5148 | attrs->nice = std_nice[i]; | |
5149 | attrs->no_numa = true; | |
5150 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
5151 | } |
5152 | ||
d320c038 | 5153 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 5154 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 5155 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
5156 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
5157 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
5158 | system_freezable_wq = alloc_workqueue("events_freezable", |
5159 | WQ_FREEZABLE, 0); | |
3c2a0909 S |
5160 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
5161 | WQ_POWER_EFFICIENT, 0); | |
5162 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
5163 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
5164 | 0); | |
1aabe902 | 5165 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
3c2a0909 S |
5166 | !system_unbound_wq || !system_freezable_wq || |
5167 | !system_power_efficient_wq || | |
5168 | !system_freezable_power_efficient_wq); | |
6ee0578b | 5169 | return 0; |
1da177e4 | 5170 | } |
6ee0578b | 5171 | early_initcall(init_workqueues); |