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> |
e22bee78 | 47 | |
ea138446 | 48 | #include "workqueue_internal.h" |
1da177e4 | 49 | |
c8e55f36 | 50 | enum { |
24647570 TH |
51 | /* |
52 | * worker_pool flags | |
bc2ae0f5 | 53 | * |
24647570 | 54 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
55 | * While associated (!DISASSOCIATED), all workers are bound to the |
56 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
57 | * is in effect. | |
58 | * | |
59 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
60 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 61 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 TH |
62 | * |
63 | * Note that DISASSOCIATED can be flipped only while holding | |
24647570 TH |
64 | * assoc_mutex to avoid changing binding state while |
65 | * create_worker() is in progress. | |
bc2ae0f5 | 66 | */ |
11ebea50 | 67 | POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ |
24647570 | 68 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
35b6bb63 | 69 | POOL_FREEZING = 1 << 3, /* freeze in progress */ |
db7bccf4 | 70 | |
c8e55f36 TH |
71 | /* worker flags */ |
72 | WORKER_STARTED = 1 << 0, /* started */ | |
73 | WORKER_DIE = 1 << 1, /* die die die */ | |
74 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 75 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 76 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 77 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
e22bee78 | 78 | |
5f7dabfd | 79 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND | |
403c821d | 80 | WORKER_CPU_INTENSIVE, |
db7bccf4 | 81 | |
e34cdddb | 82 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 83 | |
29c91e99 | 84 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 85 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 86 | |
e22bee78 TH |
87 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
88 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
89 | ||
3233cdbd TH |
90 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
91 | /* call for help after 10ms | |
92 | (min two ticks) */ | |
e22bee78 TH |
93 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
94 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
95 | |
96 | /* | |
97 | * Rescue workers are used only on emergencies and shared by | |
98 | * all cpus. Give -20. | |
99 | */ | |
100 | RESCUER_NICE_LEVEL = -20, | |
3270476a | 101 | HIGHPRI_NICE_LEVEL = -20, |
c8e55f36 | 102 | }; |
1da177e4 LT |
103 | |
104 | /* | |
4690c4ab TH |
105 | * Structure fields follow one of the following exclusion rules. |
106 | * | |
e41e704b TH |
107 | * I: Modifiable by initialization/destruction paths and read-only for |
108 | * everyone else. | |
4690c4ab | 109 | * |
e22bee78 TH |
110 | * P: Preemption protected. Disabling preemption is enough and should |
111 | * only be modified and accessed from the local cpu. | |
112 | * | |
d565ed63 | 113 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 114 | * |
d565ed63 TH |
115 | * X: During normal operation, modification requires pool->lock and should |
116 | * be done only from local cpu. Either disabling preemption on local | |
117 | * cpu or grabbing pool->lock is enough for read access. If | |
118 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 119 | * |
73f53c4a TH |
120 | * F: wq->flush_mutex protected. |
121 | * | |
4690c4ab | 122 | * W: workqueue_lock protected. |
76af4d93 TH |
123 | * |
124 | * R: workqueue_lock protected for writes. Sched-RCU protected for reads. | |
1da177e4 | 125 | */ |
1da177e4 | 126 | |
2eaebdb3 | 127 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 128 | |
bd7bdd43 | 129 | struct worker_pool { |
d565ed63 | 130 | spinlock_t lock; /* the pool lock */ |
d84ff051 | 131 | int cpu; /* I: the associated cpu */ |
9daf9e67 | 132 | int id; /* I: pool ID */ |
11ebea50 | 133 | unsigned int flags; /* X: flags */ |
bd7bdd43 TH |
134 | |
135 | struct list_head worklist; /* L: list of pending works */ | |
136 | int nr_workers; /* L: total number of workers */ | |
ea1abd61 LJ |
137 | |
138 | /* nr_idle includes the ones off idle_list for rebinding */ | |
bd7bdd43 TH |
139 | int nr_idle; /* L: currently idle ones */ |
140 | ||
141 | struct list_head idle_list; /* X: list of idle workers */ | |
142 | struct timer_list idle_timer; /* L: worker idle timeout */ | |
143 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
144 | ||
c9e7cf27 TH |
145 | /* workers are chained either in busy_hash or idle_list */ |
146 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); | |
147 | /* L: hash of busy workers */ | |
148 | ||
34a06bd6 | 149 | struct mutex manager_arb; /* manager arbitration */ |
24647570 | 150 | struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */ |
bd7bdd43 | 151 | struct ida worker_ida; /* L: for worker IDs */ |
e19e397a | 152 | |
7a4e344c | 153 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
29c91e99 TH |
154 | struct hlist_node hash_node; /* R: unbound_pool_hash node */ |
155 | int refcnt; /* refcnt for unbound pools */ | |
7a4e344c | 156 | |
e19e397a TH |
157 | /* |
158 | * The current concurrency level. As it's likely to be accessed | |
159 | * from other CPUs during try_to_wake_up(), put it in a separate | |
160 | * cacheline. | |
161 | */ | |
162 | atomic_t nr_running ____cacheline_aligned_in_smp; | |
29c91e99 TH |
163 | |
164 | /* | |
165 | * Destruction of pool is sched-RCU protected to allow dereferences | |
166 | * from get_work_pool(). | |
167 | */ | |
168 | struct rcu_head rcu; | |
8b03ae3c TH |
169 | } ____cacheline_aligned_in_smp; |
170 | ||
1da177e4 | 171 | /* |
112202d9 TH |
172 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
173 | * of work_struct->data are used for flags and the remaining high bits | |
174 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
175 | * number of flag bits. | |
1da177e4 | 176 | */ |
112202d9 | 177 | struct pool_workqueue { |
bd7bdd43 | 178 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 179 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
180 | int work_color; /* L: current color */ |
181 | int flush_color; /* L: flushing color */ | |
182 | int nr_in_flight[WORK_NR_COLORS]; | |
183 | /* L: nr of in_flight works */ | |
1e19ffc6 | 184 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 185 | int max_active; /* L: max active works */ |
1e19ffc6 | 186 | struct list_head delayed_works; /* L: delayed works */ |
76af4d93 | 187 | struct list_head pwqs_node; /* R: node on wq->pwqs */ |
493a1724 | 188 | struct list_head mayday_node; /* W: node on wq->maydays */ |
e904e6c2 | 189 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 190 | |
73f53c4a TH |
191 | /* |
192 | * Structure used to wait for workqueue flush. | |
193 | */ | |
194 | struct wq_flusher { | |
195 | struct list_head list; /* F: list of flushers */ | |
196 | int flush_color; /* F: flush color waiting for */ | |
197 | struct completion done; /* flush completion */ | |
198 | }; | |
199 | ||
1da177e4 LT |
200 | /* |
201 | * The externally visible workqueue abstraction is an array of | |
202 | * per-CPU workqueues: | |
203 | */ | |
204 | struct workqueue_struct { | |
9c5a2ba7 | 205 | unsigned int flags; /* W: WQ_* flags */ |
420c0ddb | 206 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */ |
76af4d93 | 207 | struct list_head pwqs; /* R: all pwqs of this wq */ |
4690c4ab | 208 | struct list_head list; /* W: list of all workqueues */ |
73f53c4a TH |
209 | |
210 | struct mutex flush_mutex; /* protects wq flushing */ | |
211 | int work_color; /* F: current work color */ | |
212 | int flush_color; /* F: current flush color */ | |
112202d9 | 213 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
73f53c4a TH |
214 | struct wq_flusher *first_flusher; /* F: first flusher */ |
215 | struct list_head flusher_queue; /* F: flush waiters */ | |
216 | struct list_head flusher_overflow; /* F: flush overflow list */ | |
217 | ||
493a1724 | 218 | struct list_head maydays; /* W: pwqs requesting rescue */ |
e22bee78 TH |
219 | struct worker *rescuer; /* I: rescue worker */ |
220 | ||
9c5a2ba7 | 221 | int nr_drainers; /* W: drain in progress */ |
112202d9 | 222 | int saved_max_active; /* W: saved pwq max_active */ |
4e6045f1 | 223 | #ifdef CONFIG_LOCKDEP |
4690c4ab | 224 | struct lockdep_map lockdep_map; |
4e6045f1 | 225 | #endif |
b196be89 | 226 | char name[]; /* I: workqueue name */ |
1da177e4 LT |
227 | }; |
228 | ||
e904e6c2 TH |
229 | static struct kmem_cache *pwq_cache; |
230 | ||
29c91e99 TH |
231 | /* hash of all unbound pools keyed by pool->attrs */ |
232 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); | |
233 | ||
234 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; | |
235 | ||
d320c038 | 236 | struct workqueue_struct *system_wq __read_mostly; |
d320c038 | 237 | EXPORT_SYMBOL_GPL(system_wq); |
044c782c | 238 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 239 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 240 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 241 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 242 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 243 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 244 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 245 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
d320c038 | 246 | |
97bd2347 TH |
247 | #define CREATE_TRACE_POINTS |
248 | #include <trace/events/workqueue.h> | |
249 | ||
76af4d93 TH |
250 | #define assert_rcu_or_wq_lock() \ |
251 | rcu_lockdep_assert(rcu_read_lock_sched_held() || \ | |
252 | lockdep_is_held(&workqueue_lock), \ | |
253 | "sched RCU or workqueue lock should be held") | |
254 | ||
38db41d9 | 255 | #define for_each_std_worker_pool(pool, cpu) \ |
a60dc39c TH |
256 | for ((pool) = &std_worker_pools(cpu)[0]; \ |
257 | (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++) | |
4ce62e9e | 258 | |
b67bfe0d SL |
259 | #define for_each_busy_worker(worker, i, pool) \ |
260 | hash_for_each(pool->busy_hash, i, worker, hentry) | |
db7bccf4 | 261 | |
706026c2 TH |
262 | static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, |
263 | unsigned int sw) | |
f3421797 TH |
264 | { |
265 | if (cpu < nr_cpu_ids) { | |
266 | if (sw & 1) { | |
267 | cpu = cpumask_next(cpu, mask); | |
268 | if (cpu < nr_cpu_ids) | |
269 | return cpu; | |
270 | } | |
271 | if (sw & 2) | |
272 | return WORK_CPU_UNBOUND; | |
273 | } | |
6be19588 | 274 | return WORK_CPU_END; |
f3421797 TH |
275 | } |
276 | ||
09884951 TH |
277 | /* |
278 | * CPU iterators | |
279 | * | |
706026c2 | 280 | * An extra cpu number is defined using an invalid cpu number |
09884951 | 281 | * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any |
706026c2 TH |
282 | * specific CPU. The following iterators are similar to for_each_*_cpu() |
283 | * iterators but also considers the unbound CPU. | |
09884951 | 284 | * |
706026c2 TH |
285 | * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND |
286 | * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND | |
09884951 | 287 | */ |
706026c2 TH |
288 | #define for_each_wq_cpu(cpu) \ |
289 | for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \ | |
6be19588 | 290 | (cpu) < WORK_CPU_END; \ |
706026c2 | 291 | (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3)) |
f3421797 | 292 | |
706026c2 TH |
293 | #define for_each_online_wq_cpu(cpu) \ |
294 | for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \ | |
6be19588 | 295 | (cpu) < WORK_CPU_END; \ |
706026c2 | 296 | (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3)) |
f3421797 | 297 | |
17116969 TH |
298 | /** |
299 | * for_each_pool - iterate through all worker_pools in the system | |
300 | * @pool: iteration cursor | |
301 | * @id: integer used for iteration | |
fa1b54e6 TH |
302 | * |
303 | * This must be called either with workqueue_lock held or sched RCU read | |
304 | * locked. If the pool needs to be used beyond the locking in effect, the | |
305 | * caller is responsible for guaranteeing that the pool stays online. | |
306 | * | |
307 | * The if/else clause exists only for the lockdep assertion and can be | |
308 | * ignored. | |
17116969 TH |
309 | */ |
310 | #define for_each_pool(pool, id) \ | |
fa1b54e6 TH |
311 | idr_for_each_entry(&worker_pool_idr, pool, id) \ |
312 | if (({ assert_rcu_or_wq_lock(); false; })) { } \ | |
313 | else | |
17116969 | 314 | |
49e3cf44 TH |
315 | /** |
316 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
317 | * @pwq: iteration cursor | |
318 | * @wq: the target workqueue | |
76af4d93 TH |
319 | * |
320 | * This must be called either with workqueue_lock held or sched RCU read | |
321 | * locked. If the pwq needs to be used beyond the locking in effect, the | |
322 | * caller is responsible for guaranteeing that the pwq stays online. | |
323 | * | |
324 | * The if/else clause exists only for the lockdep assertion and can be | |
325 | * ignored. | |
49e3cf44 TH |
326 | */ |
327 | #define for_each_pwq(pwq, wq) \ | |
76af4d93 TH |
328 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ |
329 | if (({ assert_rcu_or_wq_lock(); false; })) { } \ | |
330 | else | |
f3421797 | 331 | |
dc186ad7 TG |
332 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
333 | ||
334 | static struct debug_obj_descr work_debug_descr; | |
335 | ||
99777288 SG |
336 | static void *work_debug_hint(void *addr) |
337 | { | |
338 | return ((struct work_struct *) addr)->func; | |
339 | } | |
340 | ||
dc186ad7 TG |
341 | /* |
342 | * fixup_init is called when: | |
343 | * - an active object is initialized | |
344 | */ | |
345 | static int work_fixup_init(void *addr, enum debug_obj_state state) | |
346 | { | |
347 | struct work_struct *work = addr; | |
348 | ||
349 | switch (state) { | |
350 | case ODEBUG_STATE_ACTIVE: | |
351 | cancel_work_sync(work); | |
352 | debug_object_init(work, &work_debug_descr); | |
353 | return 1; | |
354 | default: | |
355 | return 0; | |
356 | } | |
357 | } | |
358 | ||
359 | /* | |
360 | * fixup_activate is called when: | |
361 | * - an active object is activated | |
362 | * - an unknown object is activated (might be a statically initialized object) | |
363 | */ | |
364 | static int work_fixup_activate(void *addr, enum debug_obj_state state) | |
365 | { | |
366 | struct work_struct *work = addr; | |
367 | ||
368 | switch (state) { | |
369 | ||
370 | case ODEBUG_STATE_NOTAVAILABLE: | |
371 | /* | |
372 | * This is not really a fixup. The work struct was | |
373 | * statically initialized. We just make sure that it | |
374 | * is tracked in the object tracker. | |
375 | */ | |
22df02bb | 376 | if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { |
dc186ad7 TG |
377 | debug_object_init(work, &work_debug_descr); |
378 | debug_object_activate(work, &work_debug_descr); | |
379 | return 0; | |
380 | } | |
381 | WARN_ON_ONCE(1); | |
382 | return 0; | |
383 | ||
384 | case ODEBUG_STATE_ACTIVE: | |
385 | WARN_ON(1); | |
386 | ||
387 | default: | |
388 | return 0; | |
389 | } | |
390 | } | |
391 | ||
392 | /* | |
393 | * fixup_free is called when: | |
394 | * - an active object is freed | |
395 | */ | |
396 | static int work_fixup_free(void *addr, enum debug_obj_state state) | |
397 | { | |
398 | struct work_struct *work = addr; | |
399 | ||
400 | switch (state) { | |
401 | case ODEBUG_STATE_ACTIVE: | |
402 | cancel_work_sync(work); | |
403 | debug_object_free(work, &work_debug_descr); | |
404 | return 1; | |
405 | default: | |
406 | return 0; | |
407 | } | |
408 | } | |
409 | ||
410 | static struct debug_obj_descr work_debug_descr = { | |
411 | .name = "work_struct", | |
99777288 | 412 | .debug_hint = work_debug_hint, |
dc186ad7 TG |
413 | .fixup_init = work_fixup_init, |
414 | .fixup_activate = work_fixup_activate, | |
415 | .fixup_free = work_fixup_free, | |
416 | }; | |
417 | ||
418 | static inline void debug_work_activate(struct work_struct *work) | |
419 | { | |
420 | debug_object_activate(work, &work_debug_descr); | |
421 | } | |
422 | ||
423 | static inline void debug_work_deactivate(struct work_struct *work) | |
424 | { | |
425 | debug_object_deactivate(work, &work_debug_descr); | |
426 | } | |
427 | ||
428 | void __init_work(struct work_struct *work, int onstack) | |
429 | { | |
430 | if (onstack) | |
431 | debug_object_init_on_stack(work, &work_debug_descr); | |
432 | else | |
433 | debug_object_init(work, &work_debug_descr); | |
434 | } | |
435 | EXPORT_SYMBOL_GPL(__init_work); | |
436 | ||
437 | void destroy_work_on_stack(struct work_struct *work) | |
438 | { | |
439 | debug_object_free(work, &work_debug_descr); | |
440 | } | |
441 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
442 | ||
443 | #else | |
444 | static inline void debug_work_activate(struct work_struct *work) { } | |
445 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
446 | #endif | |
447 | ||
95402b38 GS |
448 | /* Serializes the accesses to the list of workqueues. */ |
449 | static DEFINE_SPINLOCK(workqueue_lock); | |
1da177e4 | 450 | static LIST_HEAD(workqueues); |
a0a1a5fd | 451 | static bool workqueue_freezing; /* W: have wqs started freezing? */ |
c34056a3 | 452 | |
e22bee78 | 453 | /* |
e19e397a TH |
454 | * The CPU and unbound standard worker pools. The unbound ones have |
455 | * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set. | |
f3421797 | 456 | */ |
e19e397a TH |
457 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], |
458 | cpu_std_worker_pools); | |
a60dc39c | 459 | static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS]; |
f3421797 | 460 | |
fa1b54e6 TH |
461 | /* |
462 | * idr of all pools. Modifications are protected by workqueue_lock. Read | |
463 | * accesses are protected by sched-RCU protected. | |
464 | */ | |
9daf9e67 TH |
465 | static DEFINE_IDR(worker_pool_idr); |
466 | ||
c34056a3 | 467 | static int worker_thread(void *__worker); |
1da177e4 | 468 | |
a60dc39c | 469 | static struct worker_pool *std_worker_pools(int cpu) |
8b03ae3c | 470 | { |
f3421797 | 471 | if (cpu != WORK_CPU_UNBOUND) |
a60dc39c | 472 | return per_cpu(cpu_std_worker_pools, cpu); |
f3421797 | 473 | else |
a60dc39c | 474 | return unbound_std_worker_pools; |
8b03ae3c TH |
475 | } |
476 | ||
4e8f0a60 TH |
477 | static int std_worker_pool_pri(struct worker_pool *pool) |
478 | { | |
a60dc39c | 479 | return pool - std_worker_pools(pool->cpu); |
4e8f0a60 TH |
480 | } |
481 | ||
9daf9e67 TH |
482 | /* allocate ID and assign it to @pool */ |
483 | static int worker_pool_assign_id(struct worker_pool *pool) | |
484 | { | |
485 | int ret; | |
486 | ||
fa1b54e6 TH |
487 | do { |
488 | if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) | |
489 | return -ENOMEM; | |
9daf9e67 | 490 | |
fa1b54e6 TH |
491 | spin_lock_irq(&workqueue_lock); |
492 | ret = idr_get_new(&worker_pool_idr, pool, &pool->id); | |
493 | spin_unlock_irq(&workqueue_lock); | |
494 | } while (ret == -EAGAIN); | |
9daf9e67 | 495 | |
fa1b54e6 | 496 | return ret; |
7c3eed5c TH |
497 | } |
498 | ||
d565ed63 TH |
499 | static struct worker_pool *get_std_worker_pool(int cpu, bool highpri) |
500 | { | |
a60dc39c | 501 | struct worker_pool *pools = std_worker_pools(cpu); |
d565ed63 | 502 | |
a60dc39c | 503 | return &pools[highpri]; |
d565ed63 TH |
504 | } |
505 | ||
76af4d93 TH |
506 | /** |
507 | * first_pwq - return the first pool_workqueue of the specified workqueue | |
508 | * @wq: the target workqueue | |
509 | * | |
510 | * This must be called either with workqueue_lock held or sched RCU read | |
511 | * locked. If the pwq needs to be used beyond the locking in effect, the | |
512 | * caller is responsible for guaranteeing that the pwq stays online. | |
513 | */ | |
7fb98ea7 | 514 | static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) |
b1f4ec17 | 515 | { |
76af4d93 TH |
516 | assert_rcu_or_wq_lock(); |
517 | return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, | |
518 | pwqs_node); | |
b1f4ec17 ON |
519 | } |
520 | ||
73f53c4a TH |
521 | static unsigned int work_color_to_flags(int color) |
522 | { | |
523 | return color << WORK_STRUCT_COLOR_SHIFT; | |
524 | } | |
525 | ||
526 | static int get_work_color(struct work_struct *work) | |
527 | { | |
528 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
529 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
530 | } | |
531 | ||
532 | static int work_next_color(int color) | |
533 | { | |
534 | return (color + 1) % WORK_NR_COLORS; | |
535 | } | |
1da177e4 | 536 | |
14441960 | 537 | /* |
112202d9 TH |
538 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
539 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 540 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 541 | * |
112202d9 TH |
542 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
543 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
544 | * work->data. These functions should only be called while the work is |
545 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 546 | * |
112202d9 | 547 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 548 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 549 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 550 | * available only while the work item is queued. |
7a22ad75 | 551 | * |
bbb68dfa TH |
552 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
553 | * canceled. While being canceled, a work item may have its PENDING set | |
554 | * but stay off timer and worklist for arbitrarily long and nobody should | |
555 | * try to steal the PENDING bit. | |
14441960 | 556 | */ |
7a22ad75 TH |
557 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
558 | unsigned long flags) | |
365970a1 | 559 | { |
6183c009 | 560 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
561 | atomic_long_set(&work->data, data | flags | work_static(work)); |
562 | } | |
365970a1 | 563 | |
112202d9 | 564 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
565 | unsigned long extra_flags) |
566 | { | |
112202d9 TH |
567 | set_work_data(work, (unsigned long)pwq, |
568 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
569 | } |
570 | ||
4468a00f LJ |
571 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
572 | int pool_id) | |
573 | { | |
574 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
575 | WORK_STRUCT_PENDING); | |
576 | } | |
577 | ||
7c3eed5c TH |
578 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
579 | int pool_id) | |
7a22ad75 | 580 | { |
23657bb1 TH |
581 | /* |
582 | * The following wmb is paired with the implied mb in | |
583 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
584 | * here are visible to and precede any updates by the next PENDING | |
585 | * owner. | |
586 | */ | |
587 | smp_wmb(); | |
7c3eed5c | 588 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
7a22ad75 | 589 | } |
f756d5e2 | 590 | |
7a22ad75 | 591 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 592 | { |
7c3eed5c TH |
593 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
594 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
595 | } |
596 | ||
112202d9 | 597 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 598 | { |
e120153d | 599 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 600 | |
112202d9 | 601 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
602 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
603 | else | |
604 | return NULL; | |
4d707b9f ON |
605 | } |
606 | ||
7c3eed5c TH |
607 | /** |
608 | * get_work_pool - return the worker_pool a given work was associated with | |
609 | * @work: the work item of interest | |
610 | * | |
611 | * Return the worker_pool @work was last associated with. %NULL if none. | |
fa1b54e6 TH |
612 | * |
613 | * Pools are created and destroyed under workqueue_lock, and allows read | |
614 | * access under sched-RCU read lock. As such, this function should be | |
615 | * called under workqueue_lock or with preemption disabled. | |
616 | * | |
617 | * All fields of the returned pool are accessible as long as the above | |
618 | * mentioned locking is in effect. If the returned pool needs to be used | |
619 | * beyond the critical section, the caller is responsible for ensuring the | |
620 | * returned pool is and stays online. | |
7c3eed5c TH |
621 | */ |
622 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 623 | { |
e120153d | 624 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 625 | int pool_id; |
7a22ad75 | 626 | |
fa1b54e6 TH |
627 | assert_rcu_or_wq_lock(); |
628 | ||
112202d9 TH |
629 | if (data & WORK_STRUCT_PWQ) |
630 | return ((struct pool_workqueue *) | |
7c3eed5c | 631 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 632 | |
7c3eed5c TH |
633 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
634 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
635 | return NULL; |
636 | ||
fa1b54e6 | 637 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
638 | } |
639 | ||
640 | /** | |
641 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
642 | * @work: the work item of interest | |
643 | * | |
644 | * Return the worker_pool ID @work was last associated with. | |
645 | * %WORK_OFFQ_POOL_NONE if none. | |
646 | */ | |
647 | static int get_work_pool_id(struct work_struct *work) | |
648 | { | |
54d5b7d0 LJ |
649 | unsigned long data = atomic_long_read(&work->data); |
650 | ||
112202d9 TH |
651 | if (data & WORK_STRUCT_PWQ) |
652 | return ((struct pool_workqueue *) | |
54d5b7d0 | 653 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 654 | |
54d5b7d0 | 655 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
656 | } |
657 | ||
bbb68dfa TH |
658 | static void mark_work_canceling(struct work_struct *work) |
659 | { | |
7c3eed5c | 660 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 661 | |
7c3eed5c TH |
662 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
663 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
664 | } |
665 | ||
666 | static bool work_is_canceling(struct work_struct *work) | |
667 | { | |
668 | unsigned long data = atomic_long_read(&work->data); | |
669 | ||
112202d9 | 670 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
671 | } |
672 | ||
e22bee78 | 673 | /* |
3270476a TH |
674 | * Policy functions. These define the policies on how the global worker |
675 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 676 | * they're being called with pool->lock held. |
e22bee78 TH |
677 | */ |
678 | ||
63d95a91 | 679 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 680 | { |
e19e397a | 681 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
682 | } |
683 | ||
4594bf15 | 684 | /* |
e22bee78 TH |
685 | * Need to wake up a worker? Called from anything but currently |
686 | * running workers. | |
974271c4 TH |
687 | * |
688 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 689 | * function will always return %true for unbound pools as long as the |
974271c4 | 690 | * worklist isn't empty. |
4594bf15 | 691 | */ |
63d95a91 | 692 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 693 | { |
63d95a91 | 694 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 695 | } |
4594bf15 | 696 | |
e22bee78 | 697 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 698 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 699 | { |
63d95a91 | 700 | return pool->nr_idle; |
e22bee78 TH |
701 | } |
702 | ||
703 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 704 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 705 | { |
e19e397a TH |
706 | return !list_empty(&pool->worklist) && |
707 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
708 | } |
709 | ||
710 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 711 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 712 | { |
63d95a91 | 713 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 714 | } |
365970a1 | 715 | |
e22bee78 | 716 | /* Do I need to be the manager? */ |
63d95a91 | 717 | static bool need_to_manage_workers(struct worker_pool *pool) |
e22bee78 | 718 | { |
63d95a91 | 719 | return need_to_create_worker(pool) || |
11ebea50 | 720 | (pool->flags & POOL_MANAGE_WORKERS); |
e22bee78 TH |
721 | } |
722 | ||
723 | /* Do we have too many workers and should some go away? */ | |
63d95a91 | 724 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 725 | { |
34a06bd6 | 726 | bool managing = mutex_is_locked(&pool->manager_arb); |
63d95a91 TH |
727 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
728 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 | 729 | |
ea1abd61 LJ |
730 | /* |
731 | * nr_idle and idle_list may disagree if idle rebinding is in | |
732 | * progress. Never return %true if idle_list is empty. | |
733 | */ | |
734 | if (list_empty(&pool->idle_list)) | |
735 | return false; | |
736 | ||
e22bee78 | 737 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; |
365970a1 DH |
738 | } |
739 | ||
4d707b9f | 740 | /* |
e22bee78 TH |
741 | * Wake up functions. |
742 | */ | |
743 | ||
7e11629d | 744 | /* Return the first worker. Safe with preemption disabled */ |
63d95a91 | 745 | static struct worker *first_worker(struct worker_pool *pool) |
7e11629d | 746 | { |
63d95a91 | 747 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
748 | return NULL; |
749 | ||
63d95a91 | 750 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
751 | } |
752 | ||
753 | /** | |
754 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 755 | * @pool: worker pool to wake worker from |
7e11629d | 756 | * |
63d95a91 | 757 | * Wake up the first idle worker of @pool. |
7e11629d TH |
758 | * |
759 | * CONTEXT: | |
d565ed63 | 760 | * spin_lock_irq(pool->lock). |
7e11629d | 761 | */ |
63d95a91 | 762 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 763 | { |
63d95a91 | 764 | struct worker *worker = first_worker(pool); |
7e11629d TH |
765 | |
766 | if (likely(worker)) | |
767 | wake_up_process(worker->task); | |
768 | } | |
769 | ||
d302f017 | 770 | /** |
e22bee78 TH |
771 | * wq_worker_waking_up - a worker is waking up |
772 | * @task: task waking up | |
773 | * @cpu: CPU @task is waking up to | |
774 | * | |
775 | * This function is called during try_to_wake_up() when a worker is | |
776 | * being awoken. | |
777 | * | |
778 | * CONTEXT: | |
779 | * spin_lock_irq(rq->lock) | |
780 | */ | |
d84ff051 | 781 | void wq_worker_waking_up(struct task_struct *task, int cpu) |
e22bee78 TH |
782 | { |
783 | struct worker *worker = kthread_data(task); | |
784 | ||
36576000 | 785 | if (!(worker->flags & WORKER_NOT_RUNNING)) { |
ec22ca5e | 786 | WARN_ON_ONCE(worker->pool->cpu != cpu); |
e19e397a | 787 | atomic_inc(&worker->pool->nr_running); |
36576000 | 788 | } |
e22bee78 TH |
789 | } |
790 | ||
791 | /** | |
792 | * wq_worker_sleeping - a worker is going to sleep | |
793 | * @task: task going to sleep | |
794 | * @cpu: CPU in question, must be the current CPU number | |
795 | * | |
796 | * This function is called during schedule() when a busy worker is | |
797 | * going to sleep. Worker on the same cpu can be woken up by | |
798 | * returning pointer to its task. | |
799 | * | |
800 | * CONTEXT: | |
801 | * spin_lock_irq(rq->lock) | |
802 | * | |
803 | * RETURNS: | |
804 | * Worker task on @cpu to wake up, %NULL if none. | |
805 | */ | |
d84ff051 | 806 | struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) |
e22bee78 TH |
807 | { |
808 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
111c225a | 809 | struct worker_pool *pool; |
e22bee78 | 810 | |
111c225a TH |
811 | /* |
812 | * Rescuers, which may not have all the fields set up like normal | |
813 | * workers, also reach here, let's not access anything before | |
814 | * checking NOT_RUNNING. | |
815 | */ | |
2d64672e | 816 | if (worker->flags & WORKER_NOT_RUNNING) |
e22bee78 TH |
817 | return NULL; |
818 | ||
111c225a | 819 | pool = worker->pool; |
111c225a | 820 | |
e22bee78 | 821 | /* this can only happen on the local cpu */ |
6183c009 TH |
822 | if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) |
823 | return NULL; | |
e22bee78 TH |
824 | |
825 | /* | |
826 | * The counterpart of the following dec_and_test, implied mb, | |
827 | * worklist not empty test sequence is in insert_work(). | |
828 | * Please read comment there. | |
829 | * | |
628c78e7 TH |
830 | * NOT_RUNNING is clear. This means that we're bound to and |
831 | * running on the local cpu w/ rq lock held and preemption | |
832 | * disabled, which in turn means that none else could be | |
d565ed63 | 833 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 834 | * lock is safe. |
e22bee78 | 835 | */ |
e19e397a TH |
836 | if (atomic_dec_and_test(&pool->nr_running) && |
837 | !list_empty(&pool->worklist)) | |
63d95a91 | 838 | to_wakeup = first_worker(pool); |
e22bee78 TH |
839 | return to_wakeup ? to_wakeup->task : NULL; |
840 | } | |
841 | ||
842 | /** | |
843 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 844 | * @worker: self |
d302f017 TH |
845 | * @flags: flags to set |
846 | * @wakeup: wakeup an idle worker if necessary | |
847 | * | |
e22bee78 TH |
848 | * Set @flags in @worker->flags and adjust nr_running accordingly. If |
849 | * nr_running becomes zero and @wakeup is %true, an idle worker is | |
850 | * woken up. | |
d302f017 | 851 | * |
cb444766 | 852 | * CONTEXT: |
d565ed63 | 853 | * spin_lock_irq(pool->lock) |
d302f017 TH |
854 | */ |
855 | static inline void worker_set_flags(struct worker *worker, unsigned int flags, | |
856 | bool wakeup) | |
857 | { | |
bd7bdd43 | 858 | struct worker_pool *pool = worker->pool; |
e22bee78 | 859 | |
cb444766 TH |
860 | WARN_ON_ONCE(worker->task != current); |
861 | ||
e22bee78 TH |
862 | /* |
863 | * If transitioning into NOT_RUNNING, adjust nr_running and | |
864 | * wake up an idle worker as necessary if requested by | |
865 | * @wakeup. | |
866 | */ | |
867 | if ((flags & WORKER_NOT_RUNNING) && | |
868 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
e22bee78 | 869 | if (wakeup) { |
e19e397a | 870 | if (atomic_dec_and_test(&pool->nr_running) && |
bd7bdd43 | 871 | !list_empty(&pool->worklist)) |
63d95a91 | 872 | wake_up_worker(pool); |
e22bee78 | 873 | } else |
e19e397a | 874 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
875 | } |
876 | ||
d302f017 TH |
877 | worker->flags |= flags; |
878 | } | |
879 | ||
880 | /** | |
e22bee78 | 881 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 882 | * @worker: self |
d302f017 TH |
883 | * @flags: flags to clear |
884 | * | |
e22bee78 | 885 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 886 | * |
cb444766 | 887 | * CONTEXT: |
d565ed63 | 888 | * spin_lock_irq(pool->lock) |
d302f017 TH |
889 | */ |
890 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
891 | { | |
63d95a91 | 892 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
893 | unsigned int oflags = worker->flags; |
894 | ||
cb444766 TH |
895 | WARN_ON_ONCE(worker->task != current); |
896 | ||
d302f017 | 897 | worker->flags &= ~flags; |
e22bee78 | 898 | |
42c025f3 TH |
899 | /* |
900 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
901 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
902 | * of multiple flags, not a single flag. | |
903 | */ | |
e22bee78 TH |
904 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
905 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 906 | atomic_inc(&pool->nr_running); |
d302f017 TH |
907 | } |
908 | ||
8cca0eea TH |
909 | /** |
910 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 911 | * @pool: pool of interest |
8cca0eea TH |
912 | * @work: work to find worker for |
913 | * | |
c9e7cf27 TH |
914 | * Find a worker which is executing @work on @pool by searching |
915 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
916 | * to match, its current execution should match the address of @work and |
917 | * its work function. This is to avoid unwanted dependency between | |
918 | * unrelated work executions through a work item being recycled while still | |
919 | * being executed. | |
920 | * | |
921 | * This is a bit tricky. A work item may be freed once its execution | |
922 | * starts and nothing prevents the freed area from being recycled for | |
923 | * another work item. If the same work item address ends up being reused | |
924 | * before the original execution finishes, workqueue will identify the | |
925 | * recycled work item as currently executing and make it wait until the | |
926 | * current execution finishes, introducing an unwanted dependency. | |
927 | * | |
928 | * This function checks the work item address, work function and workqueue | |
929 | * to avoid false positives. Note that this isn't complete as one may | |
930 | * construct a work function which can introduce dependency onto itself | |
931 | * through a recycled work item. Well, if somebody wants to shoot oneself | |
932 | * in the foot that badly, there's only so much we can do, and if such | |
933 | * deadlock actually occurs, it should be easy to locate the culprit work | |
934 | * function. | |
8cca0eea TH |
935 | * |
936 | * CONTEXT: | |
d565ed63 | 937 | * spin_lock_irq(pool->lock). |
8cca0eea TH |
938 | * |
939 | * RETURNS: | |
940 | * Pointer to worker which is executing @work if found, NULL | |
941 | * otherwise. | |
4d707b9f | 942 | */ |
c9e7cf27 | 943 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 944 | struct work_struct *work) |
4d707b9f | 945 | { |
42f8570f | 946 | struct worker *worker; |
42f8570f | 947 | |
b67bfe0d | 948 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
949 | (unsigned long)work) |
950 | if (worker->current_work == work && | |
951 | worker->current_func == work->func) | |
42f8570f SL |
952 | return worker; |
953 | ||
954 | return NULL; | |
4d707b9f ON |
955 | } |
956 | ||
bf4ede01 TH |
957 | /** |
958 | * move_linked_works - move linked works to a list | |
959 | * @work: start of series of works to be scheduled | |
960 | * @head: target list to append @work to | |
961 | * @nextp: out paramter for nested worklist walking | |
962 | * | |
963 | * Schedule linked works starting from @work to @head. Work series to | |
964 | * be scheduled starts at @work and includes any consecutive work with | |
965 | * WORK_STRUCT_LINKED set in its predecessor. | |
966 | * | |
967 | * If @nextp is not NULL, it's updated to point to the next work of | |
968 | * the last scheduled work. This allows move_linked_works() to be | |
969 | * nested inside outer list_for_each_entry_safe(). | |
970 | * | |
971 | * CONTEXT: | |
d565ed63 | 972 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
973 | */ |
974 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
975 | struct work_struct **nextp) | |
976 | { | |
977 | struct work_struct *n; | |
978 | ||
979 | /* | |
980 | * Linked worklist will always end before the end of the list, | |
981 | * use NULL for list head. | |
982 | */ | |
983 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
984 | list_move_tail(&work->entry, head); | |
985 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
986 | break; | |
987 | } | |
988 | ||
989 | /* | |
990 | * If we're already inside safe list traversal and have moved | |
991 | * multiple works to the scheduled queue, the next position | |
992 | * needs to be updated. | |
993 | */ | |
994 | if (nextp) | |
995 | *nextp = n; | |
996 | } | |
997 | ||
112202d9 | 998 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 999 | { |
112202d9 | 1000 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1001 | |
1002 | trace_workqueue_activate_work(work); | |
112202d9 | 1003 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1004 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1005 | pwq->nr_active++; |
bf4ede01 TH |
1006 | } |
1007 | ||
112202d9 | 1008 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1009 | { |
112202d9 | 1010 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1011 | struct work_struct, entry); |
1012 | ||
112202d9 | 1013 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1014 | } |
1015 | ||
bf4ede01 | 1016 | /** |
112202d9 TH |
1017 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1018 | * @pwq: pwq of interest | |
bf4ede01 | 1019 | * @color: color of work which left the queue |
bf4ede01 TH |
1020 | * |
1021 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1022 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1023 | * |
1024 | * CONTEXT: | |
d565ed63 | 1025 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1026 | */ |
112202d9 | 1027 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 TH |
1028 | { |
1029 | /* ignore uncolored works */ | |
1030 | if (color == WORK_NO_COLOR) | |
1031 | return; | |
1032 | ||
112202d9 | 1033 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1034 | |
112202d9 TH |
1035 | pwq->nr_active--; |
1036 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1037 | /* one down, submit a delayed one */ |
112202d9 TH |
1038 | if (pwq->nr_active < pwq->max_active) |
1039 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1040 | } |
1041 | ||
1042 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1043 | if (likely(pwq->flush_color != color)) |
bf4ede01 TH |
1044 | return; |
1045 | ||
1046 | /* are there still in-flight works? */ | |
112202d9 | 1047 | if (pwq->nr_in_flight[color]) |
bf4ede01 TH |
1048 | return; |
1049 | ||
112202d9 TH |
1050 | /* this pwq is done, clear flush_color */ |
1051 | pwq->flush_color = -1; | |
bf4ede01 TH |
1052 | |
1053 | /* | |
112202d9 | 1054 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1055 | * will handle the rest. |
1056 | */ | |
112202d9 TH |
1057 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1058 | complete(&pwq->wq->first_flusher->done); | |
bf4ede01 TH |
1059 | } |
1060 | ||
36e227d2 | 1061 | /** |
bbb68dfa | 1062 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1063 | * @work: work item to steal |
1064 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1065 | * @flags: place to store irq state |
36e227d2 TH |
1066 | * |
1067 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
1068 | * stable state - idle, on timer or on worklist. Return values are | |
1069 | * | |
1070 | * 1 if @work was pending and we successfully stole PENDING | |
1071 | * 0 if @work was idle and we claimed PENDING | |
1072 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1073 | * -ENOENT if someone else is canceling @work, this state may persist |
1074 | * for arbitrarily long | |
36e227d2 | 1075 | * |
bbb68dfa | 1076 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1077 | * interrupted while holding PENDING and @work off queue, irq must be |
1078 | * disabled on entry. This, combined with delayed_work->timer being | |
1079 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1080 | * |
1081 | * On successful return, >= 0, irq is disabled and the caller is | |
1082 | * responsible for releasing it using local_irq_restore(*@flags). | |
1083 | * | |
e0aecdd8 | 1084 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1085 | */ |
bbb68dfa TH |
1086 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1087 | unsigned long *flags) | |
bf4ede01 | 1088 | { |
d565ed63 | 1089 | struct worker_pool *pool; |
112202d9 | 1090 | struct pool_workqueue *pwq; |
bf4ede01 | 1091 | |
bbb68dfa TH |
1092 | local_irq_save(*flags); |
1093 | ||
36e227d2 TH |
1094 | /* try to steal the timer if it exists */ |
1095 | if (is_dwork) { | |
1096 | struct delayed_work *dwork = to_delayed_work(work); | |
1097 | ||
e0aecdd8 TH |
1098 | /* |
1099 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1100 | * guaranteed that the timer is not queued anywhere and not | |
1101 | * running on the local CPU. | |
1102 | */ | |
36e227d2 TH |
1103 | if (likely(del_timer(&dwork->timer))) |
1104 | return 1; | |
1105 | } | |
1106 | ||
1107 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1108 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1109 | return 0; | |
1110 | ||
1111 | /* | |
1112 | * The queueing is in progress, or it is already queued. Try to | |
1113 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1114 | */ | |
d565ed63 TH |
1115 | pool = get_work_pool(work); |
1116 | if (!pool) | |
bbb68dfa | 1117 | goto fail; |
bf4ede01 | 1118 | |
d565ed63 | 1119 | spin_lock(&pool->lock); |
0b3dae68 | 1120 | /* |
112202d9 TH |
1121 | * work->data is guaranteed to point to pwq only while the work |
1122 | * item is queued on pwq->wq, and both updating work->data to point | |
1123 | * to pwq on queueing and to pool on dequeueing are done under | |
1124 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1125 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1126 | * item is currently queued on that pool. |
1127 | */ | |
112202d9 TH |
1128 | pwq = get_work_pwq(work); |
1129 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1130 | debug_work_deactivate(work); |
1131 | ||
1132 | /* | |
1133 | * A delayed work item cannot be grabbed directly because | |
1134 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1135 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1136 | * management later on and cause stall. Make sure the work |
1137 | * item is activated before grabbing. | |
1138 | */ | |
1139 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1140 | pwq_activate_delayed_work(work); |
16062836 TH |
1141 | |
1142 | list_del_init(&work->entry); | |
112202d9 | 1143 | pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); |
16062836 | 1144 | |
112202d9 | 1145 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1146 | set_work_pool_and_keep_pending(work, pool->id); |
1147 | ||
1148 | spin_unlock(&pool->lock); | |
1149 | return 1; | |
bf4ede01 | 1150 | } |
d565ed63 | 1151 | spin_unlock(&pool->lock); |
bbb68dfa TH |
1152 | fail: |
1153 | local_irq_restore(*flags); | |
1154 | if (work_is_canceling(work)) | |
1155 | return -ENOENT; | |
1156 | cpu_relax(); | |
36e227d2 | 1157 | return -EAGAIN; |
bf4ede01 TH |
1158 | } |
1159 | ||
4690c4ab | 1160 | /** |
706026c2 | 1161 | * insert_work - insert a work into a pool |
112202d9 | 1162 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1163 | * @work: work to insert |
1164 | * @head: insertion point | |
1165 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1166 | * | |
112202d9 | 1167 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1168 | * work_struct flags. |
4690c4ab TH |
1169 | * |
1170 | * CONTEXT: | |
d565ed63 | 1171 | * spin_lock_irq(pool->lock). |
4690c4ab | 1172 | */ |
112202d9 TH |
1173 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1174 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1175 | { |
112202d9 | 1176 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1177 | |
4690c4ab | 1178 | /* we own @work, set data and link */ |
112202d9 | 1179 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1180 | list_add_tail(&work->entry, head); |
e22bee78 TH |
1181 | |
1182 | /* | |
1183 | * Ensure either worker_sched_deactivated() sees the above | |
1184 | * list_add_tail() or we see zero nr_running to avoid workers | |
1185 | * lying around lazily while there are works to be processed. | |
1186 | */ | |
1187 | smp_mb(); | |
1188 | ||
63d95a91 TH |
1189 | if (__need_more_worker(pool)) |
1190 | wake_up_worker(pool); | |
b89deed3 ON |
1191 | } |
1192 | ||
c8efcc25 TH |
1193 | /* |
1194 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1195 | * same workqueue. |
c8efcc25 TH |
1196 | */ |
1197 | static bool is_chained_work(struct workqueue_struct *wq) | |
1198 | { | |
8d03ecfe TH |
1199 | struct worker *worker; |
1200 | ||
1201 | worker = current_wq_worker(); | |
1202 | /* | |
1203 | * Return %true iff I'm a worker execuing a work item on @wq. If | |
1204 | * I'm @worker, it's safe to dereference it without locking. | |
1205 | */ | |
112202d9 | 1206 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1207 | } |
1208 | ||
d84ff051 | 1209 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1210 | struct work_struct *work) |
1211 | { | |
112202d9 | 1212 | struct pool_workqueue *pwq; |
1e19ffc6 | 1213 | struct list_head *worklist; |
8a2e8e5d | 1214 | unsigned int work_flags; |
b75cac93 | 1215 | unsigned int req_cpu = cpu; |
8930caba TH |
1216 | |
1217 | /* | |
1218 | * While a work item is PENDING && off queue, a task trying to | |
1219 | * steal the PENDING will busy-loop waiting for it to either get | |
1220 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1221 | * happen with IRQ disabled. | |
1222 | */ | |
1223 | WARN_ON_ONCE(!irqs_disabled()); | |
1da177e4 | 1224 | |
dc186ad7 | 1225 | debug_work_activate(work); |
1e19ffc6 | 1226 | |
c8efcc25 | 1227 | /* if dying, only works from the same workqueue are allowed */ |
9c5a2ba7 | 1228 | if (unlikely(wq->flags & WQ_DRAINING) && |
c8efcc25 | 1229 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b TH |
1230 | return; |
1231 | ||
112202d9 | 1232 | /* determine the pwq to use */ |
c7fc77f7 | 1233 | if (!(wq->flags & WQ_UNBOUND)) { |
c9e7cf27 | 1234 | struct worker_pool *last_pool; |
18aa9eff | 1235 | |
57469821 | 1236 | if (cpu == WORK_CPU_UNBOUND) |
c7fc77f7 TH |
1237 | cpu = raw_smp_processor_id(); |
1238 | ||
18aa9eff | 1239 | /* |
dbf2576e TH |
1240 | * It's multi cpu. If @work was previously on a different |
1241 | * cpu, it might still be running there, in which case the | |
1242 | * work needs to be queued on that cpu to guarantee | |
1243 | * non-reentrancy. | |
18aa9eff | 1244 | */ |
7fb98ea7 | 1245 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
c9e7cf27 | 1246 | last_pool = get_work_pool(work); |
dbf2576e | 1247 | |
112202d9 | 1248 | if (last_pool && last_pool != pwq->pool) { |
18aa9eff TH |
1249 | struct worker *worker; |
1250 | ||
d565ed63 | 1251 | spin_lock(&last_pool->lock); |
18aa9eff | 1252 | |
c9e7cf27 | 1253 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1254 | |
112202d9 | 1255 | if (worker && worker->current_pwq->wq == wq) { |
7fb98ea7 | 1256 | pwq = per_cpu_ptr(wq->cpu_pwqs, last_pool->cpu); |
8594fade | 1257 | } else { |
18aa9eff | 1258 | /* meh... not running there, queue here */ |
d565ed63 | 1259 | spin_unlock(&last_pool->lock); |
112202d9 | 1260 | spin_lock(&pwq->pool->lock); |
18aa9eff | 1261 | } |
8930caba | 1262 | } else { |
112202d9 | 1263 | spin_lock(&pwq->pool->lock); |
8930caba | 1264 | } |
f3421797 | 1265 | } else { |
7fb98ea7 | 1266 | pwq = first_pwq(wq); |
112202d9 | 1267 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1268 | } |
1269 | ||
112202d9 TH |
1270 | /* pwq determined, queue */ |
1271 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1272 | |
f5b2552b | 1273 | if (WARN_ON(!list_empty(&work->entry))) { |
112202d9 | 1274 | spin_unlock(&pwq->pool->lock); |
f5b2552b DC |
1275 | return; |
1276 | } | |
1e19ffc6 | 1277 | |
112202d9 TH |
1278 | pwq->nr_in_flight[pwq->work_color]++; |
1279 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1280 | |
112202d9 | 1281 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1282 | trace_workqueue_activate_work(work); |
112202d9 TH |
1283 | pwq->nr_active++; |
1284 | worklist = &pwq->pool->worklist; | |
8a2e8e5d TH |
1285 | } else { |
1286 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1287 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1288 | } |
1e19ffc6 | 1289 | |
112202d9 | 1290 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1291 | |
112202d9 | 1292 | spin_unlock(&pwq->pool->lock); |
1da177e4 LT |
1293 | } |
1294 | ||
0fcb78c2 | 1295 | /** |
c1a220e7 ZR |
1296 | * queue_work_on - queue work on specific cpu |
1297 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1298 | * @wq: workqueue to use |
1299 | * @work: work to queue | |
1300 | * | |
d4283e93 | 1301 | * Returns %false if @work was already on a queue, %true otherwise. |
1da177e4 | 1302 | * |
c1a220e7 ZR |
1303 | * We queue the work to a specific CPU, the caller must ensure it |
1304 | * can't go away. | |
1da177e4 | 1305 | */ |
d4283e93 TH |
1306 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1307 | struct work_struct *work) | |
1da177e4 | 1308 | { |
d4283e93 | 1309 | bool ret = false; |
8930caba | 1310 | unsigned long flags; |
ef1ca236 | 1311 | |
8930caba | 1312 | local_irq_save(flags); |
c1a220e7 | 1313 | |
22df02bb | 1314 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1315 | __queue_work(cpu, wq, work); |
d4283e93 | 1316 | ret = true; |
c1a220e7 | 1317 | } |
ef1ca236 | 1318 | |
8930caba | 1319 | local_irq_restore(flags); |
1da177e4 LT |
1320 | return ret; |
1321 | } | |
c1a220e7 | 1322 | EXPORT_SYMBOL_GPL(queue_work_on); |
1da177e4 | 1323 | |
c1a220e7 | 1324 | /** |
0a13c00e | 1325 | * queue_work - queue work on a workqueue |
c1a220e7 ZR |
1326 | * @wq: workqueue to use |
1327 | * @work: work to queue | |
1328 | * | |
d4283e93 | 1329 | * Returns %false if @work was already on a queue, %true otherwise. |
c1a220e7 | 1330 | * |
0a13c00e TH |
1331 | * We queue the work to the CPU on which it was submitted, but if the CPU dies |
1332 | * it can be processed by another CPU. | |
c1a220e7 | 1333 | */ |
d4283e93 | 1334 | bool queue_work(struct workqueue_struct *wq, struct work_struct *work) |
c1a220e7 | 1335 | { |
57469821 | 1336 | return queue_work_on(WORK_CPU_UNBOUND, wq, work); |
c1a220e7 | 1337 | } |
0a13c00e | 1338 | EXPORT_SYMBOL_GPL(queue_work); |
c1a220e7 | 1339 | |
d8e794df | 1340 | void delayed_work_timer_fn(unsigned long __data) |
1da177e4 | 1341 | { |
52bad64d | 1342 | struct delayed_work *dwork = (struct delayed_work *)__data; |
1da177e4 | 1343 | |
e0aecdd8 | 1344 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1345 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1346 | } |
1438ade5 | 1347 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1348 | |
7beb2edf TH |
1349 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1350 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1351 | { |
7beb2edf TH |
1352 | struct timer_list *timer = &dwork->timer; |
1353 | struct work_struct *work = &dwork->work; | |
7beb2edf TH |
1354 | |
1355 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn || | |
1356 | timer->data != (unsigned long)dwork); | |
fc4b514f TH |
1357 | WARN_ON_ONCE(timer_pending(timer)); |
1358 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1359 | |
8852aac2 TH |
1360 | /* |
1361 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1362 | * both optimization and correctness. The earliest @timer can | |
1363 | * expire is on the closest next tick and delayed_work users depend | |
1364 | * on that there's no such delay when @delay is 0. | |
1365 | */ | |
1366 | if (!delay) { | |
1367 | __queue_work(cpu, wq, &dwork->work); | |
1368 | return; | |
1369 | } | |
1370 | ||
7beb2edf | 1371 | timer_stats_timer_set_start_info(&dwork->timer); |
1da177e4 | 1372 | |
60c057bc | 1373 | dwork->wq = wq; |
1265057f | 1374 | dwork->cpu = cpu; |
7beb2edf TH |
1375 | timer->expires = jiffies + delay; |
1376 | ||
1377 | if (unlikely(cpu != WORK_CPU_UNBOUND)) | |
1378 | add_timer_on(timer, cpu); | |
1379 | else | |
1380 | add_timer(timer); | |
1da177e4 LT |
1381 | } |
1382 | ||
0fcb78c2 REB |
1383 | /** |
1384 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1385 | * @cpu: CPU number to execute work on | |
1386 | * @wq: workqueue to use | |
af9997e4 | 1387 | * @dwork: work to queue |
0fcb78c2 REB |
1388 | * @delay: number of jiffies to wait before queueing |
1389 | * | |
715f1300 TH |
1390 | * Returns %false if @work was already on a queue, %true otherwise. If |
1391 | * @delay is zero and @dwork is idle, it will be scheduled for immediate | |
1392 | * execution. | |
0fcb78c2 | 1393 | */ |
d4283e93 TH |
1394 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1395 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1396 | { |
52bad64d | 1397 | struct work_struct *work = &dwork->work; |
d4283e93 | 1398 | bool ret = false; |
8930caba | 1399 | unsigned long flags; |
7a6bc1cd | 1400 | |
8930caba TH |
1401 | /* read the comment in __queue_work() */ |
1402 | local_irq_save(flags); | |
7a6bc1cd | 1403 | |
22df02bb | 1404 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1405 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1406 | ret = true; |
7a6bc1cd | 1407 | } |
8a3e77cc | 1408 | |
8930caba | 1409 | local_irq_restore(flags); |
7a6bc1cd VP |
1410 | return ret; |
1411 | } | |
ae90dd5d | 1412 | EXPORT_SYMBOL_GPL(queue_delayed_work_on); |
c7fc77f7 | 1413 | |
0a13c00e TH |
1414 | /** |
1415 | * queue_delayed_work - queue work on a workqueue after delay | |
1416 | * @wq: workqueue to use | |
1417 | * @dwork: delayable work to queue | |
1418 | * @delay: number of jiffies to wait before queueing | |
1419 | * | |
715f1300 | 1420 | * Equivalent to queue_delayed_work_on() but tries to use the local CPU. |
0a13c00e | 1421 | */ |
d4283e93 | 1422 | bool queue_delayed_work(struct workqueue_struct *wq, |
0a13c00e TH |
1423 | struct delayed_work *dwork, unsigned long delay) |
1424 | { | |
57469821 | 1425 | return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); |
0a13c00e TH |
1426 | } |
1427 | EXPORT_SYMBOL_GPL(queue_delayed_work); | |
c7fc77f7 | 1428 | |
8376fe22 TH |
1429 | /** |
1430 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1431 | * @cpu: CPU number to execute work on | |
1432 | * @wq: workqueue to use | |
1433 | * @dwork: work to queue | |
1434 | * @delay: number of jiffies to wait before queueing | |
1435 | * | |
1436 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1437 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1438 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1439 | * current state. | |
1440 | * | |
1441 | * Returns %false if @dwork was idle and queued, %true if @dwork was | |
1442 | * pending and its timer was modified. | |
1443 | * | |
e0aecdd8 | 1444 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1445 | * See try_to_grab_pending() for details. |
1446 | */ | |
1447 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1448 | struct delayed_work *dwork, unsigned long delay) | |
1449 | { | |
1450 | unsigned long flags; | |
1451 | int ret; | |
c7fc77f7 | 1452 | |
8376fe22 TH |
1453 | do { |
1454 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1455 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1456 | |
8376fe22 TH |
1457 | if (likely(ret >= 0)) { |
1458 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1459 | local_irq_restore(flags); | |
7a6bc1cd | 1460 | } |
8376fe22 TH |
1461 | |
1462 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1463 | return ret; |
1464 | } | |
8376fe22 TH |
1465 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1466 | ||
1467 | /** | |
1468 | * mod_delayed_work - modify delay of or queue a delayed work | |
1469 | * @wq: workqueue to use | |
1470 | * @dwork: work to queue | |
1471 | * @delay: number of jiffies to wait before queueing | |
1472 | * | |
1473 | * mod_delayed_work_on() on local CPU. | |
1474 | */ | |
1475 | bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, | |
1476 | unsigned long delay) | |
1477 | { | |
1478 | return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); | |
1479 | } | |
1480 | EXPORT_SYMBOL_GPL(mod_delayed_work); | |
1da177e4 | 1481 | |
c8e55f36 TH |
1482 | /** |
1483 | * worker_enter_idle - enter idle state | |
1484 | * @worker: worker which is entering idle state | |
1485 | * | |
1486 | * @worker is entering idle state. Update stats and idle timer if | |
1487 | * necessary. | |
1488 | * | |
1489 | * LOCKING: | |
d565ed63 | 1490 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1491 | */ |
1492 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1493 | { |
bd7bdd43 | 1494 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1495 | |
6183c009 TH |
1496 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1497 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1498 | (worker->hentry.next || worker->hentry.pprev))) | |
1499 | return; | |
c8e55f36 | 1500 | |
cb444766 TH |
1501 | /* can't use worker_set_flags(), also called from start_worker() */ |
1502 | worker->flags |= WORKER_IDLE; | |
bd7bdd43 | 1503 | pool->nr_idle++; |
e22bee78 | 1504 | worker->last_active = jiffies; |
c8e55f36 TH |
1505 | |
1506 | /* idle_list is LIFO */ | |
bd7bdd43 | 1507 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1508 | |
628c78e7 TH |
1509 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1510 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1511 | |
544ecf31 | 1512 | /* |
706026c2 | 1513 | * Sanity check nr_running. Because wq_unbind_fn() releases |
d565ed63 | 1514 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1515 | * nr_running, the warning may trigger spuriously. Check iff |
1516 | * unbind is not in progress. | |
544ecf31 | 1517 | */ |
24647570 | 1518 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1519 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1520 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1521 | } |
1522 | ||
1523 | /** | |
1524 | * worker_leave_idle - leave idle state | |
1525 | * @worker: worker which is leaving idle state | |
1526 | * | |
1527 | * @worker is leaving idle state. Update stats. | |
1528 | * | |
1529 | * LOCKING: | |
d565ed63 | 1530 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1531 | */ |
1532 | static void worker_leave_idle(struct worker *worker) | |
1533 | { | |
bd7bdd43 | 1534 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1535 | |
6183c009 TH |
1536 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1537 | return; | |
d302f017 | 1538 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1539 | pool->nr_idle--; |
c8e55f36 TH |
1540 | list_del_init(&worker->entry); |
1541 | } | |
1542 | ||
e22bee78 | 1543 | /** |
f36dc67b LJ |
1544 | * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it |
1545 | * @pool: target worker_pool | |
1546 | * | |
1547 | * Bind %current to the cpu of @pool if it is associated and lock @pool. | |
e22bee78 TH |
1548 | * |
1549 | * Works which are scheduled while the cpu is online must at least be | |
1550 | * scheduled to a worker which is bound to the cpu so that if they are | |
1551 | * flushed from cpu callbacks while cpu is going down, they are | |
1552 | * guaranteed to execute on the cpu. | |
1553 | * | |
f5faa077 | 1554 | * This function is to be used by unbound workers and rescuers to bind |
e22bee78 TH |
1555 | * themselves to the target cpu and may race with cpu going down or |
1556 | * coming online. kthread_bind() can't be used because it may put the | |
1557 | * worker to already dead cpu and set_cpus_allowed_ptr() can't be used | |
706026c2 | 1558 | * verbatim as it's best effort and blocking and pool may be |
e22bee78 TH |
1559 | * [dis]associated in the meantime. |
1560 | * | |
706026c2 | 1561 | * This function tries set_cpus_allowed() and locks pool and verifies the |
24647570 | 1562 | * binding against %POOL_DISASSOCIATED which is set during |
f2d5a0ee TH |
1563 | * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker |
1564 | * enters idle state or fetches works without dropping lock, it can | |
1565 | * guarantee the scheduling requirement described in the first paragraph. | |
e22bee78 TH |
1566 | * |
1567 | * CONTEXT: | |
d565ed63 | 1568 | * Might sleep. Called without any lock but returns with pool->lock |
e22bee78 TH |
1569 | * held. |
1570 | * | |
1571 | * RETURNS: | |
706026c2 | 1572 | * %true if the associated pool is online (@worker is successfully |
e22bee78 TH |
1573 | * bound), %false if offline. |
1574 | */ | |
f36dc67b | 1575 | static bool worker_maybe_bind_and_lock(struct worker_pool *pool) |
d565ed63 | 1576 | __acquires(&pool->lock) |
e22bee78 | 1577 | { |
e22bee78 | 1578 | while (true) { |
4e6045f1 | 1579 | /* |
e22bee78 TH |
1580 | * The following call may fail, succeed or succeed |
1581 | * without actually migrating the task to the cpu if | |
1582 | * it races with cpu hotunplug operation. Verify | |
24647570 | 1583 | * against POOL_DISASSOCIATED. |
4e6045f1 | 1584 | */ |
24647570 | 1585 | if (!(pool->flags & POOL_DISASSOCIATED)) |
7a4e344c | 1586 | set_cpus_allowed_ptr(current, pool->attrs->cpumask); |
e22bee78 | 1587 | |
d565ed63 | 1588 | spin_lock_irq(&pool->lock); |
24647570 | 1589 | if (pool->flags & POOL_DISASSOCIATED) |
e22bee78 | 1590 | return false; |
f5faa077 | 1591 | if (task_cpu(current) == pool->cpu && |
7a4e344c | 1592 | cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) |
e22bee78 | 1593 | return true; |
d565ed63 | 1594 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1595 | |
5035b20f TH |
1596 | /* |
1597 | * We've raced with CPU hot[un]plug. Give it a breather | |
1598 | * and retry migration. cond_resched() is required here; | |
1599 | * otherwise, we might deadlock against cpu_stop trying to | |
1600 | * bring down the CPU on non-preemptive kernel. | |
1601 | */ | |
e22bee78 | 1602 | cpu_relax(); |
5035b20f | 1603 | cond_resched(); |
e22bee78 TH |
1604 | } |
1605 | } | |
1606 | ||
25511a47 | 1607 | /* |
ea1abd61 | 1608 | * Rebind an idle @worker to its CPU. worker_thread() will test |
5f7dabfd | 1609 | * list_empty(@worker->entry) before leaving idle and call this function. |
25511a47 TH |
1610 | */ |
1611 | static void idle_worker_rebind(struct worker *worker) | |
1612 | { | |
5f7dabfd | 1613 | /* CPU may go down again inbetween, clear UNBOUND only on success */ |
f36dc67b | 1614 | if (worker_maybe_bind_and_lock(worker->pool)) |
5f7dabfd | 1615 | worker_clr_flags(worker, WORKER_UNBOUND); |
25511a47 | 1616 | |
ea1abd61 LJ |
1617 | /* rebind complete, become available again */ |
1618 | list_add(&worker->entry, &worker->pool->idle_list); | |
d565ed63 | 1619 | spin_unlock_irq(&worker->pool->lock); |
25511a47 TH |
1620 | } |
1621 | ||
e22bee78 | 1622 | /* |
25511a47 | 1623 | * Function for @worker->rebind.work used to rebind unbound busy workers to |
403c821d TH |
1624 | * the associated cpu which is coming back online. This is scheduled by |
1625 | * cpu up but can race with other cpu hotplug operations and may be | |
1626 | * executed twice without intervening cpu down. | |
e22bee78 | 1627 | */ |
25511a47 | 1628 | static void busy_worker_rebind_fn(struct work_struct *work) |
e22bee78 TH |
1629 | { |
1630 | struct worker *worker = container_of(work, struct worker, rebind_work); | |
e22bee78 | 1631 | |
f36dc67b | 1632 | if (worker_maybe_bind_and_lock(worker->pool)) |
eab6d828 | 1633 | worker_clr_flags(worker, WORKER_UNBOUND); |
e22bee78 | 1634 | |
d565ed63 | 1635 | spin_unlock_irq(&worker->pool->lock); |
e22bee78 TH |
1636 | } |
1637 | ||
25511a47 | 1638 | /** |
94cf58bb TH |
1639 | * rebind_workers - rebind all workers of a pool to the associated CPU |
1640 | * @pool: pool of interest | |
25511a47 | 1641 | * |
94cf58bb | 1642 | * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding |
25511a47 TH |
1643 | * is different for idle and busy ones. |
1644 | * | |
ea1abd61 LJ |
1645 | * Idle ones will be removed from the idle_list and woken up. They will |
1646 | * add themselves back after completing rebind. This ensures that the | |
1647 | * idle_list doesn't contain any unbound workers when re-bound busy workers | |
1648 | * try to perform local wake-ups for concurrency management. | |
25511a47 | 1649 | * |
ea1abd61 LJ |
1650 | * Busy workers can rebind after they finish their current work items. |
1651 | * Queueing the rebind work item at the head of the scheduled list is | |
1652 | * enough. Note that nr_running will be properly bumped as busy workers | |
1653 | * rebind. | |
25511a47 | 1654 | * |
ea1abd61 LJ |
1655 | * On return, all non-manager workers are scheduled for rebind - see |
1656 | * manage_workers() for the manager special case. Any idle worker | |
1657 | * including the manager will not appear on @idle_list until rebind is | |
1658 | * complete, making local wake-ups safe. | |
25511a47 | 1659 | */ |
94cf58bb | 1660 | static void rebind_workers(struct worker_pool *pool) |
25511a47 | 1661 | { |
ea1abd61 | 1662 | struct worker *worker, *n; |
25511a47 TH |
1663 | int i; |
1664 | ||
94cf58bb TH |
1665 | lockdep_assert_held(&pool->assoc_mutex); |
1666 | lockdep_assert_held(&pool->lock); | |
25511a47 | 1667 | |
5f7dabfd | 1668 | /* dequeue and kick idle ones */ |
94cf58bb TH |
1669 | list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { |
1670 | /* | |
1671 | * idle workers should be off @pool->idle_list until rebind | |
1672 | * is complete to avoid receiving premature local wake-ups. | |
1673 | */ | |
1674 | list_del_init(&worker->entry); | |
25511a47 | 1675 | |
94cf58bb TH |
1676 | /* |
1677 | * worker_thread() will see the above dequeuing and call | |
1678 | * idle_worker_rebind(). | |
1679 | */ | |
1680 | wake_up_process(worker->task); | |
1681 | } | |
25511a47 | 1682 | |
94cf58bb | 1683 | /* rebind busy workers */ |
b67bfe0d | 1684 | for_each_busy_worker(worker, i, pool) { |
94cf58bb TH |
1685 | struct work_struct *rebind_work = &worker->rebind_work; |
1686 | struct workqueue_struct *wq; | |
25511a47 | 1687 | |
94cf58bb TH |
1688 | if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, |
1689 | work_data_bits(rebind_work))) | |
1690 | continue; | |
25511a47 | 1691 | |
94cf58bb | 1692 | debug_work_activate(rebind_work); |
90beca5d | 1693 | |
94cf58bb TH |
1694 | /* |
1695 | * wq doesn't really matter but let's keep @worker->pool | |
112202d9 | 1696 | * and @pwq->pool consistent for sanity. |
94cf58bb | 1697 | */ |
7a4e344c | 1698 | if (worker->pool->attrs->nice < 0) |
94cf58bb TH |
1699 | wq = system_highpri_wq; |
1700 | else | |
1701 | wq = system_wq; | |
1702 | ||
7fb98ea7 | 1703 | insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work, |
94cf58bb TH |
1704 | worker->scheduled.next, |
1705 | work_color_to_flags(WORK_NO_COLOR)); | |
ec58815a | 1706 | } |
25511a47 TH |
1707 | } |
1708 | ||
c34056a3 TH |
1709 | static struct worker *alloc_worker(void) |
1710 | { | |
1711 | struct worker *worker; | |
1712 | ||
1713 | worker = kzalloc(sizeof(*worker), GFP_KERNEL); | |
c8e55f36 TH |
1714 | if (worker) { |
1715 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1716 | INIT_LIST_HEAD(&worker->scheduled); |
25511a47 | 1717 | INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); |
e22bee78 TH |
1718 | /* on creation a worker is in !idle && prep state */ |
1719 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1720 | } |
c34056a3 TH |
1721 | return worker; |
1722 | } | |
1723 | ||
1724 | /** | |
1725 | * create_worker - create a new workqueue worker | |
63d95a91 | 1726 | * @pool: pool the new worker will belong to |
c34056a3 | 1727 | * |
63d95a91 | 1728 | * Create a new worker which is bound to @pool. The returned worker |
c34056a3 TH |
1729 | * can be started by calling start_worker() or destroyed using |
1730 | * destroy_worker(). | |
1731 | * | |
1732 | * CONTEXT: | |
1733 | * Might sleep. Does GFP_KERNEL allocations. | |
1734 | * | |
1735 | * RETURNS: | |
1736 | * Pointer to the newly created worker. | |
1737 | */ | |
bc2ae0f5 | 1738 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1739 | { |
7a4e344c | 1740 | const char *pri = pool->attrs->nice < 0 ? "H" : ""; |
c34056a3 | 1741 | struct worker *worker = NULL; |
f3421797 | 1742 | int id = -1; |
c34056a3 | 1743 | |
d565ed63 | 1744 | spin_lock_irq(&pool->lock); |
bd7bdd43 | 1745 | while (ida_get_new(&pool->worker_ida, &id)) { |
d565ed63 | 1746 | spin_unlock_irq(&pool->lock); |
bd7bdd43 | 1747 | if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) |
c34056a3 | 1748 | goto fail; |
d565ed63 | 1749 | spin_lock_irq(&pool->lock); |
c34056a3 | 1750 | } |
d565ed63 | 1751 | spin_unlock_irq(&pool->lock); |
c34056a3 TH |
1752 | |
1753 | worker = alloc_worker(); | |
1754 | if (!worker) | |
1755 | goto fail; | |
1756 | ||
bd7bdd43 | 1757 | worker->pool = pool; |
c34056a3 TH |
1758 | worker->id = id; |
1759 | ||
29c91e99 | 1760 | if (pool->cpu >= 0) |
94dcf29a | 1761 | worker->task = kthread_create_on_node(worker_thread, |
ec22ca5e | 1762 | worker, cpu_to_node(pool->cpu), |
d84ff051 | 1763 | "kworker/%d:%d%s", pool->cpu, id, pri); |
f3421797 TH |
1764 | else |
1765 | worker->task = kthread_create(worker_thread, worker, | |
3270476a | 1766 | "kworker/u:%d%s", id, pri); |
c34056a3 TH |
1767 | if (IS_ERR(worker->task)) |
1768 | goto fail; | |
1769 | ||
7a4e344c TH |
1770 | set_user_nice(worker->task, pool->attrs->nice); |
1771 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
3270476a | 1772 | |
db7bccf4 | 1773 | /* |
7a4e344c TH |
1774 | * %PF_THREAD_BOUND is used to prevent userland from meddling with |
1775 | * cpumask of workqueue workers. This is an abuse. We need | |
1776 | * %PF_NO_SETAFFINITY. | |
db7bccf4 | 1777 | */ |
7a4e344c TH |
1778 | worker->task->flags |= PF_THREAD_BOUND; |
1779 | ||
1780 | /* | |
1781 | * The caller is responsible for ensuring %POOL_DISASSOCIATED | |
1782 | * remains stable across this function. See the comments above the | |
1783 | * flag definition for details. | |
1784 | */ | |
1785 | if (pool->flags & POOL_DISASSOCIATED) | |
bc2ae0f5 | 1786 | worker->flags |= WORKER_UNBOUND; |
c34056a3 TH |
1787 | |
1788 | return worker; | |
1789 | fail: | |
1790 | if (id >= 0) { | |
d565ed63 | 1791 | spin_lock_irq(&pool->lock); |
bd7bdd43 | 1792 | ida_remove(&pool->worker_ida, id); |
d565ed63 | 1793 | spin_unlock_irq(&pool->lock); |
c34056a3 TH |
1794 | } |
1795 | kfree(worker); | |
1796 | return NULL; | |
1797 | } | |
1798 | ||
1799 | /** | |
1800 | * start_worker - start a newly created worker | |
1801 | * @worker: worker to start | |
1802 | * | |
706026c2 | 1803 | * Make the pool aware of @worker and start it. |
c34056a3 TH |
1804 | * |
1805 | * CONTEXT: | |
d565ed63 | 1806 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1807 | */ |
1808 | static void start_worker(struct worker *worker) | |
1809 | { | |
cb444766 | 1810 | worker->flags |= WORKER_STARTED; |
bd7bdd43 | 1811 | worker->pool->nr_workers++; |
c8e55f36 | 1812 | worker_enter_idle(worker); |
c34056a3 TH |
1813 | wake_up_process(worker->task); |
1814 | } | |
1815 | ||
1816 | /** | |
1817 | * destroy_worker - destroy a workqueue worker | |
1818 | * @worker: worker to be destroyed | |
1819 | * | |
706026c2 | 1820 | * Destroy @worker and adjust @pool stats accordingly. |
c8e55f36 TH |
1821 | * |
1822 | * CONTEXT: | |
d565ed63 | 1823 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
c34056a3 TH |
1824 | */ |
1825 | static void destroy_worker(struct worker *worker) | |
1826 | { | |
bd7bdd43 | 1827 | struct worker_pool *pool = worker->pool; |
c34056a3 TH |
1828 | int id = worker->id; |
1829 | ||
1830 | /* sanity check frenzy */ | |
6183c009 TH |
1831 | if (WARN_ON(worker->current_work) || |
1832 | WARN_ON(!list_empty(&worker->scheduled))) | |
1833 | return; | |
c34056a3 | 1834 | |
c8e55f36 | 1835 | if (worker->flags & WORKER_STARTED) |
bd7bdd43 | 1836 | pool->nr_workers--; |
c8e55f36 | 1837 | if (worker->flags & WORKER_IDLE) |
bd7bdd43 | 1838 | pool->nr_idle--; |
c8e55f36 TH |
1839 | |
1840 | list_del_init(&worker->entry); | |
cb444766 | 1841 | worker->flags |= WORKER_DIE; |
c8e55f36 | 1842 | |
d565ed63 | 1843 | spin_unlock_irq(&pool->lock); |
c8e55f36 | 1844 | |
c34056a3 TH |
1845 | kthread_stop(worker->task); |
1846 | kfree(worker); | |
1847 | ||
d565ed63 | 1848 | spin_lock_irq(&pool->lock); |
bd7bdd43 | 1849 | ida_remove(&pool->worker_ida, id); |
c34056a3 TH |
1850 | } |
1851 | ||
63d95a91 | 1852 | static void idle_worker_timeout(unsigned long __pool) |
e22bee78 | 1853 | { |
63d95a91 | 1854 | struct worker_pool *pool = (void *)__pool; |
e22bee78 | 1855 | |
d565ed63 | 1856 | spin_lock_irq(&pool->lock); |
e22bee78 | 1857 | |
63d95a91 | 1858 | if (too_many_workers(pool)) { |
e22bee78 TH |
1859 | struct worker *worker; |
1860 | unsigned long expires; | |
1861 | ||
1862 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1863 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1864 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1865 | ||
1866 | if (time_before(jiffies, expires)) | |
63d95a91 | 1867 | mod_timer(&pool->idle_timer, expires); |
e22bee78 TH |
1868 | else { |
1869 | /* it's been idle for too long, wake up manager */ | |
11ebea50 | 1870 | pool->flags |= POOL_MANAGE_WORKERS; |
63d95a91 | 1871 | wake_up_worker(pool); |
d5abe669 | 1872 | } |
e22bee78 TH |
1873 | } |
1874 | ||
d565ed63 | 1875 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1876 | } |
d5abe669 | 1877 | |
493a1724 | 1878 | static void send_mayday(struct work_struct *work) |
e22bee78 | 1879 | { |
112202d9 TH |
1880 | struct pool_workqueue *pwq = get_work_pwq(work); |
1881 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 TH |
1882 | |
1883 | lockdep_assert_held(&workqueue_lock); | |
e22bee78 TH |
1884 | |
1885 | if (!(wq->flags & WQ_RESCUER)) | |
493a1724 | 1886 | return; |
e22bee78 TH |
1887 | |
1888 | /* mayday mayday mayday */ | |
493a1724 TH |
1889 | if (list_empty(&pwq->mayday_node)) { |
1890 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
e22bee78 | 1891 | wake_up_process(wq->rescuer->task); |
493a1724 | 1892 | } |
e22bee78 TH |
1893 | } |
1894 | ||
706026c2 | 1895 | static void pool_mayday_timeout(unsigned long __pool) |
e22bee78 | 1896 | { |
63d95a91 | 1897 | struct worker_pool *pool = (void *)__pool; |
e22bee78 TH |
1898 | struct work_struct *work; |
1899 | ||
493a1724 TH |
1900 | spin_lock_irq(&workqueue_lock); /* for wq->maydays */ |
1901 | spin_lock(&pool->lock); | |
e22bee78 | 1902 | |
63d95a91 | 1903 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
1904 | /* |
1905 | * We've been trying to create a new worker but | |
1906 | * haven't been successful. We might be hitting an | |
1907 | * allocation deadlock. Send distress signals to | |
1908 | * rescuers. | |
1909 | */ | |
63d95a91 | 1910 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 1911 | send_mayday(work); |
1da177e4 | 1912 | } |
e22bee78 | 1913 | |
493a1724 TH |
1914 | spin_unlock(&pool->lock); |
1915 | spin_unlock_irq(&workqueue_lock); | |
e22bee78 | 1916 | |
63d95a91 | 1917 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
1918 | } |
1919 | ||
e22bee78 TH |
1920 | /** |
1921 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 1922 | * @pool: pool to create a new worker for |
e22bee78 | 1923 | * |
63d95a91 | 1924 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
1925 | * have at least one idle worker on return from this function. If |
1926 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 1927 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
1928 | * possible allocation deadlock. |
1929 | * | |
1930 | * On return, need_to_create_worker() is guaranteed to be false and | |
1931 | * may_start_working() true. | |
1932 | * | |
1933 | * LOCKING: | |
d565ed63 | 1934 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1935 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1936 | * manager. | |
1937 | * | |
1938 | * RETURNS: | |
d565ed63 | 1939 | * false if no action was taken and pool->lock stayed locked, true |
e22bee78 TH |
1940 | * otherwise. |
1941 | */ | |
63d95a91 | 1942 | static bool maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
1943 | __releases(&pool->lock) |
1944 | __acquires(&pool->lock) | |
1da177e4 | 1945 | { |
63d95a91 | 1946 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
1947 | return false; |
1948 | restart: | |
d565ed63 | 1949 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 1950 | |
e22bee78 | 1951 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 1952 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
1953 | |
1954 | while (true) { | |
1955 | struct worker *worker; | |
1956 | ||
bc2ae0f5 | 1957 | worker = create_worker(pool); |
e22bee78 | 1958 | if (worker) { |
63d95a91 | 1959 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1960 | spin_lock_irq(&pool->lock); |
e22bee78 | 1961 | start_worker(worker); |
6183c009 TH |
1962 | if (WARN_ON_ONCE(need_to_create_worker(pool))) |
1963 | goto restart; | |
e22bee78 TH |
1964 | return true; |
1965 | } | |
1966 | ||
63d95a91 | 1967 | if (!need_to_create_worker(pool)) |
e22bee78 | 1968 | break; |
1da177e4 | 1969 | |
e22bee78 TH |
1970 | __set_current_state(TASK_INTERRUPTIBLE); |
1971 | schedule_timeout(CREATE_COOLDOWN); | |
9f9c2364 | 1972 | |
63d95a91 | 1973 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
1974 | break; |
1975 | } | |
1976 | ||
63d95a91 | 1977 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1978 | spin_lock_irq(&pool->lock); |
63d95a91 | 1979 | if (need_to_create_worker(pool)) |
e22bee78 TH |
1980 | goto restart; |
1981 | return true; | |
1982 | } | |
1983 | ||
1984 | /** | |
1985 | * maybe_destroy_worker - destroy workers which have been idle for a while | |
63d95a91 | 1986 | * @pool: pool to destroy workers for |
e22bee78 | 1987 | * |
63d95a91 | 1988 | * Destroy @pool workers which have been idle for longer than |
e22bee78 TH |
1989 | * IDLE_WORKER_TIMEOUT. |
1990 | * | |
1991 | * LOCKING: | |
d565ed63 | 1992 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1993 | * multiple times. Called only from manager. |
1994 | * | |
1995 | * RETURNS: | |
d565ed63 | 1996 | * false if no action was taken and pool->lock stayed locked, true |
e22bee78 TH |
1997 | * otherwise. |
1998 | */ | |
63d95a91 | 1999 | static bool maybe_destroy_workers(struct worker_pool *pool) |
e22bee78 TH |
2000 | { |
2001 | bool ret = false; | |
1da177e4 | 2002 | |
63d95a91 | 2003 | while (too_many_workers(pool)) { |
e22bee78 TH |
2004 | struct worker *worker; |
2005 | unsigned long expires; | |
3af24433 | 2006 | |
63d95a91 | 2007 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 | 2008 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
85f4186a | 2009 | |
e22bee78 | 2010 | if (time_before(jiffies, expires)) { |
63d95a91 | 2011 | mod_timer(&pool->idle_timer, expires); |
3af24433 | 2012 | break; |
e22bee78 | 2013 | } |
1da177e4 | 2014 | |
e22bee78 TH |
2015 | destroy_worker(worker); |
2016 | ret = true; | |
1da177e4 | 2017 | } |
1e19ffc6 | 2018 | |
e22bee78 | 2019 | return ret; |
1e19ffc6 TH |
2020 | } |
2021 | ||
73f53c4a | 2022 | /** |
e22bee78 TH |
2023 | * manage_workers - manage worker pool |
2024 | * @worker: self | |
73f53c4a | 2025 | * |
706026c2 | 2026 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2027 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2028 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2029 | * |
2030 | * The caller can safely start processing works on false return. On | |
2031 | * true return, it's guaranteed that need_to_create_worker() is false | |
2032 | * and may_start_working() is true. | |
73f53c4a TH |
2033 | * |
2034 | * CONTEXT: | |
d565ed63 | 2035 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2036 | * multiple times. Does GFP_KERNEL allocations. |
2037 | * | |
2038 | * RETURNS: | |
d565ed63 TH |
2039 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
2040 | * multiple times. Does GFP_KERNEL allocations. | |
73f53c4a | 2041 | */ |
e22bee78 | 2042 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2043 | { |
63d95a91 | 2044 | struct worker_pool *pool = worker->pool; |
e22bee78 | 2045 | bool ret = false; |
73f53c4a | 2046 | |
34a06bd6 | 2047 | if (!mutex_trylock(&pool->manager_arb)) |
e22bee78 | 2048 | return ret; |
1e19ffc6 | 2049 | |
ee378aa4 LJ |
2050 | /* |
2051 | * To simplify both worker management and CPU hotplug, hold off | |
2052 | * management while hotplug is in progress. CPU hotplug path can't | |
34a06bd6 TH |
2053 | * grab @pool->manager_arb to achieve this because that can lead to |
2054 | * idle worker depletion (all become busy thinking someone else is | |
2055 | * managing) which in turn can result in deadlock under extreme | |
2056 | * circumstances. Use @pool->assoc_mutex to synchronize manager | |
2057 | * against CPU hotplug. | |
ee378aa4 | 2058 | * |
b2eb83d1 | 2059 | * assoc_mutex would always be free unless CPU hotplug is in |
d565ed63 | 2060 | * progress. trylock first without dropping @pool->lock. |
ee378aa4 | 2061 | */ |
b2eb83d1 | 2062 | if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { |
d565ed63 | 2063 | spin_unlock_irq(&pool->lock); |
b2eb83d1 | 2064 | mutex_lock(&pool->assoc_mutex); |
ee378aa4 LJ |
2065 | /* |
2066 | * CPU hotplug could have happened while we were waiting | |
b2eb83d1 | 2067 | * for assoc_mutex. Hotplug itself can't handle us |
ee378aa4 | 2068 | * because manager isn't either on idle or busy list, and |
706026c2 | 2069 | * @pool's state and ours could have deviated. |
ee378aa4 | 2070 | * |
b2eb83d1 | 2071 | * As hotplug is now excluded via assoc_mutex, we can |
ee378aa4 | 2072 | * simply try to bind. It will succeed or fail depending |
706026c2 | 2073 | * on @pool's current state. Try it and adjust |
ee378aa4 LJ |
2074 | * %WORKER_UNBOUND accordingly. |
2075 | */ | |
f36dc67b | 2076 | if (worker_maybe_bind_and_lock(pool)) |
ee378aa4 LJ |
2077 | worker->flags &= ~WORKER_UNBOUND; |
2078 | else | |
2079 | worker->flags |= WORKER_UNBOUND; | |
73f53c4a | 2080 | |
ee378aa4 LJ |
2081 | ret = true; |
2082 | } | |
73f53c4a | 2083 | |
11ebea50 | 2084 | pool->flags &= ~POOL_MANAGE_WORKERS; |
73f53c4a TH |
2085 | |
2086 | /* | |
e22bee78 TH |
2087 | * Destroy and then create so that may_start_working() is true |
2088 | * on return. | |
73f53c4a | 2089 | */ |
63d95a91 TH |
2090 | ret |= maybe_destroy_workers(pool); |
2091 | ret |= maybe_create_worker(pool); | |
e22bee78 | 2092 | |
b2eb83d1 | 2093 | mutex_unlock(&pool->assoc_mutex); |
34a06bd6 | 2094 | mutex_unlock(&pool->manager_arb); |
e22bee78 | 2095 | return ret; |
73f53c4a TH |
2096 | } |
2097 | ||
a62428c0 TH |
2098 | /** |
2099 | * process_one_work - process single work | |
c34056a3 | 2100 | * @worker: self |
a62428c0 TH |
2101 | * @work: work to process |
2102 | * | |
2103 | * Process @work. This function contains all the logics necessary to | |
2104 | * process a single work including synchronization against and | |
2105 | * interaction with other workers on the same cpu, queueing and | |
2106 | * flushing. As long as context requirement is met, any worker can | |
2107 | * call this function to process a work. | |
2108 | * | |
2109 | * CONTEXT: | |
d565ed63 | 2110 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2111 | */ |
c34056a3 | 2112 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2113 | __releases(&pool->lock) |
2114 | __acquires(&pool->lock) | |
a62428c0 | 2115 | { |
112202d9 | 2116 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2117 | struct worker_pool *pool = worker->pool; |
112202d9 | 2118 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2119 | int work_color; |
7e11629d | 2120 | struct worker *collision; |
a62428c0 TH |
2121 | #ifdef CONFIG_LOCKDEP |
2122 | /* | |
2123 | * It is permissible to free the struct work_struct from | |
2124 | * inside the function that is called from it, this we need to | |
2125 | * take into account for lockdep too. To avoid bogus "held | |
2126 | * lock freed" warnings as well as problems when looking into | |
2127 | * work->lockdep_map, make a copy and use that here. | |
2128 | */ | |
4d82a1de PZ |
2129 | struct lockdep_map lockdep_map; |
2130 | ||
2131 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2132 | #endif |
6fec10a1 TH |
2133 | /* |
2134 | * Ensure we're on the correct CPU. DISASSOCIATED test is | |
2135 | * necessary to avoid spurious warnings from rescuers servicing the | |
24647570 | 2136 | * unbound or a disassociated pool. |
6fec10a1 | 2137 | */ |
5f7dabfd | 2138 | WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && |
24647570 | 2139 | !(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2140 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2141 | |
7e11629d TH |
2142 | /* |
2143 | * A single work shouldn't be executed concurrently by | |
2144 | * multiple workers on a single cpu. Check whether anyone is | |
2145 | * already processing the work. If so, defer the work to the | |
2146 | * currently executing one. | |
2147 | */ | |
c9e7cf27 | 2148 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2149 | if (unlikely(collision)) { |
2150 | move_linked_works(work, &collision->scheduled, NULL); | |
2151 | return; | |
2152 | } | |
2153 | ||
8930caba | 2154 | /* claim and dequeue */ |
a62428c0 | 2155 | debug_work_deactivate(work); |
c9e7cf27 | 2156 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2157 | worker->current_work = work; |
a2c1c57b | 2158 | worker->current_func = work->func; |
112202d9 | 2159 | worker->current_pwq = pwq; |
73f53c4a | 2160 | work_color = get_work_color(work); |
7a22ad75 | 2161 | |
a62428c0 TH |
2162 | list_del_init(&work->entry); |
2163 | ||
fb0e7beb TH |
2164 | /* |
2165 | * CPU intensive works don't participate in concurrency | |
2166 | * management. They're the scheduler's responsibility. | |
2167 | */ | |
2168 | if (unlikely(cpu_intensive)) | |
2169 | worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); | |
2170 | ||
974271c4 | 2171 | /* |
d565ed63 | 2172 | * Unbound pool isn't concurrency managed and work items should be |
974271c4 TH |
2173 | * executed ASAP. Wake up another worker if necessary. |
2174 | */ | |
63d95a91 TH |
2175 | if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) |
2176 | wake_up_worker(pool); | |
974271c4 | 2177 | |
8930caba | 2178 | /* |
7c3eed5c | 2179 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2180 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2181 | * PENDING and queued state changes happen together while IRQ is |
2182 | * disabled. | |
8930caba | 2183 | */ |
7c3eed5c | 2184 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2185 | |
d565ed63 | 2186 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2187 | |
112202d9 | 2188 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
a62428c0 | 2189 | lock_map_acquire(&lockdep_map); |
e36c886a | 2190 | trace_workqueue_execute_start(work); |
a2c1c57b | 2191 | worker->current_func(work); |
e36c886a AV |
2192 | /* |
2193 | * While we must be careful to not use "work" after this, the trace | |
2194 | * point will only record its address. | |
2195 | */ | |
2196 | trace_workqueue_execute_end(work); | |
a62428c0 | 2197 | lock_map_release(&lockdep_map); |
112202d9 | 2198 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2199 | |
2200 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c VI |
2201 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
2202 | " last function: %pf\n", | |
a2c1c57b TH |
2203 | current->comm, preempt_count(), task_pid_nr(current), |
2204 | worker->current_func); | |
a62428c0 TH |
2205 | debug_show_held_locks(current); |
2206 | dump_stack(); | |
2207 | } | |
2208 | ||
d565ed63 | 2209 | spin_lock_irq(&pool->lock); |
a62428c0 | 2210 | |
fb0e7beb TH |
2211 | /* clear cpu intensive status */ |
2212 | if (unlikely(cpu_intensive)) | |
2213 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2214 | ||
a62428c0 | 2215 | /* we're done with it, release */ |
42f8570f | 2216 | hash_del(&worker->hentry); |
c34056a3 | 2217 | worker->current_work = NULL; |
a2c1c57b | 2218 | worker->current_func = NULL; |
112202d9 TH |
2219 | worker->current_pwq = NULL; |
2220 | pwq_dec_nr_in_flight(pwq, work_color); | |
a62428c0 TH |
2221 | } |
2222 | ||
affee4b2 TH |
2223 | /** |
2224 | * process_scheduled_works - process scheduled works | |
2225 | * @worker: self | |
2226 | * | |
2227 | * Process all scheduled works. Please note that the scheduled list | |
2228 | * may change while processing a work, so this function repeatedly | |
2229 | * fetches a work from the top and executes it. | |
2230 | * | |
2231 | * CONTEXT: | |
d565ed63 | 2232 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2233 | * multiple times. |
2234 | */ | |
2235 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2236 | { |
affee4b2 TH |
2237 | while (!list_empty(&worker->scheduled)) { |
2238 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2239 | struct work_struct, entry); |
c34056a3 | 2240 | process_one_work(worker, work); |
1da177e4 | 2241 | } |
1da177e4 LT |
2242 | } |
2243 | ||
4690c4ab TH |
2244 | /** |
2245 | * worker_thread - the worker thread function | |
c34056a3 | 2246 | * @__worker: self |
4690c4ab | 2247 | * |
706026c2 TH |
2248 | * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools |
2249 | * of these per each cpu. These workers process all works regardless of | |
e22bee78 TH |
2250 | * their specific target workqueue. The only exception is works which |
2251 | * belong to workqueues with a rescuer which will be explained in | |
2252 | * rescuer_thread(). | |
4690c4ab | 2253 | */ |
c34056a3 | 2254 | static int worker_thread(void *__worker) |
1da177e4 | 2255 | { |
c34056a3 | 2256 | struct worker *worker = __worker; |
bd7bdd43 | 2257 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2258 | |
e22bee78 TH |
2259 | /* tell the scheduler that this is a workqueue worker */ |
2260 | worker->task->flags |= PF_WQ_WORKER; | |
c8e55f36 | 2261 | woke_up: |
d565ed63 | 2262 | spin_lock_irq(&pool->lock); |
1da177e4 | 2263 | |
5f7dabfd LJ |
2264 | /* we are off idle list if destruction or rebind is requested */ |
2265 | if (unlikely(list_empty(&worker->entry))) { | |
d565ed63 | 2266 | spin_unlock_irq(&pool->lock); |
25511a47 | 2267 | |
5f7dabfd | 2268 | /* if DIE is set, destruction is requested */ |
25511a47 TH |
2269 | if (worker->flags & WORKER_DIE) { |
2270 | worker->task->flags &= ~PF_WQ_WORKER; | |
2271 | return 0; | |
2272 | } | |
2273 | ||
5f7dabfd | 2274 | /* otherwise, rebind */ |
25511a47 TH |
2275 | idle_worker_rebind(worker); |
2276 | goto woke_up; | |
c8e55f36 | 2277 | } |
affee4b2 | 2278 | |
c8e55f36 | 2279 | worker_leave_idle(worker); |
db7bccf4 | 2280 | recheck: |
e22bee78 | 2281 | /* no more worker necessary? */ |
63d95a91 | 2282 | if (!need_more_worker(pool)) |
e22bee78 TH |
2283 | goto sleep; |
2284 | ||
2285 | /* do we need to manage? */ | |
63d95a91 | 2286 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2287 | goto recheck; |
2288 | ||
c8e55f36 TH |
2289 | /* |
2290 | * ->scheduled list can only be filled while a worker is | |
2291 | * preparing to process a work or actually processing it. | |
2292 | * Make sure nobody diddled with it while I was sleeping. | |
2293 | */ | |
6183c009 | 2294 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2295 | |
e22bee78 TH |
2296 | /* |
2297 | * When control reaches this point, we're guaranteed to have | |
2298 | * at least one idle worker or that someone else has already | |
2299 | * assumed the manager role. | |
2300 | */ | |
2301 | worker_clr_flags(worker, WORKER_PREP); | |
2302 | ||
2303 | do { | |
c8e55f36 | 2304 | struct work_struct *work = |
bd7bdd43 | 2305 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2306 | struct work_struct, entry); |
2307 | ||
2308 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { | |
2309 | /* optimization path, not strictly necessary */ | |
2310 | process_one_work(worker, work); | |
2311 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2312 | process_scheduled_works(worker); |
c8e55f36 TH |
2313 | } else { |
2314 | move_linked_works(work, &worker->scheduled, NULL); | |
2315 | process_scheduled_works(worker); | |
affee4b2 | 2316 | } |
63d95a91 | 2317 | } while (keep_working(pool)); |
e22bee78 TH |
2318 | |
2319 | worker_set_flags(worker, WORKER_PREP, false); | |
d313dd85 | 2320 | sleep: |
63d95a91 | 2321 | if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) |
e22bee78 | 2322 | goto recheck; |
d313dd85 | 2323 | |
c8e55f36 | 2324 | /* |
d565ed63 TH |
2325 | * pool->lock is held and there's no work to process and no need to |
2326 | * manage, sleep. Workers are woken up only while holding | |
2327 | * pool->lock or from local cpu, so setting the current state | |
2328 | * before releasing pool->lock is enough to prevent losing any | |
2329 | * event. | |
c8e55f36 TH |
2330 | */ |
2331 | worker_enter_idle(worker); | |
2332 | __set_current_state(TASK_INTERRUPTIBLE); | |
d565ed63 | 2333 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2334 | schedule(); |
2335 | goto woke_up; | |
1da177e4 LT |
2336 | } |
2337 | ||
e22bee78 TH |
2338 | /** |
2339 | * rescuer_thread - the rescuer thread function | |
111c225a | 2340 | * @__rescuer: self |
e22bee78 TH |
2341 | * |
2342 | * Workqueue rescuer thread function. There's one rescuer for each | |
2343 | * workqueue which has WQ_RESCUER set. | |
2344 | * | |
706026c2 | 2345 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2346 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2347 | * developing into deadlock if some works currently on the same queue | |
2348 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2349 | * the problem rescuer solves. | |
2350 | * | |
706026c2 TH |
2351 | * When such condition is possible, the pool summons rescuers of all |
2352 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2353 | * those works so that forward progress can be guaranteed. |
2354 | * | |
2355 | * This should happen rarely. | |
2356 | */ | |
111c225a | 2357 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2358 | { |
111c225a TH |
2359 | struct worker *rescuer = __rescuer; |
2360 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2361 | struct list_head *scheduled = &rescuer->scheduled; |
e22bee78 TH |
2362 | |
2363 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2364 | |
2365 | /* | |
2366 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2367 | * doesn't participate in concurrency management. | |
2368 | */ | |
2369 | rescuer->task->flags |= PF_WQ_WORKER; | |
e22bee78 TH |
2370 | repeat: |
2371 | set_current_state(TASK_INTERRUPTIBLE); | |
2372 | ||
412d32e6 MG |
2373 | if (kthread_should_stop()) { |
2374 | __set_current_state(TASK_RUNNING); | |
111c225a | 2375 | rescuer->task->flags &= ~PF_WQ_WORKER; |
e22bee78 | 2376 | return 0; |
412d32e6 | 2377 | } |
e22bee78 | 2378 | |
493a1724 TH |
2379 | /* see whether any pwq is asking for help */ |
2380 | spin_lock_irq(&workqueue_lock); | |
2381 | ||
2382 | while (!list_empty(&wq->maydays)) { | |
2383 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2384 | struct pool_workqueue, mayday_node); | |
112202d9 | 2385 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
2386 | struct work_struct *work, *n; |
2387 | ||
2388 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2389 | list_del_init(&pwq->mayday_node); |
2390 | ||
2391 | spin_unlock_irq(&workqueue_lock); | |
e22bee78 TH |
2392 | |
2393 | /* migrate to the target cpu if possible */ | |
f36dc67b | 2394 | worker_maybe_bind_and_lock(pool); |
b3104104 | 2395 | rescuer->pool = pool; |
e22bee78 TH |
2396 | |
2397 | /* | |
2398 | * Slurp in all works issued via this workqueue and | |
2399 | * process'em. | |
2400 | */ | |
6183c009 | 2401 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
bd7bdd43 | 2402 | list_for_each_entry_safe(work, n, &pool->worklist, entry) |
112202d9 | 2403 | if (get_work_pwq(work) == pwq) |
e22bee78 TH |
2404 | move_linked_works(work, scheduled, &n); |
2405 | ||
2406 | process_scheduled_works(rescuer); | |
7576958a TH |
2407 | |
2408 | /* | |
d565ed63 | 2409 | * Leave this pool. If keep_working() is %true, notify a |
7576958a TH |
2410 | * regular worker; otherwise, we end up with 0 concurrency |
2411 | * and stalling the execution. | |
2412 | */ | |
63d95a91 TH |
2413 | if (keep_working(pool)) |
2414 | wake_up_worker(pool); | |
7576958a | 2415 | |
b3104104 | 2416 | rescuer->pool = NULL; |
493a1724 TH |
2417 | spin_unlock(&pool->lock); |
2418 | spin_lock(&workqueue_lock); | |
e22bee78 TH |
2419 | } |
2420 | ||
493a1724 TH |
2421 | spin_unlock_irq(&workqueue_lock); |
2422 | ||
111c225a TH |
2423 | /* rescuers should never participate in concurrency management */ |
2424 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2425 | schedule(); |
2426 | goto repeat; | |
1da177e4 LT |
2427 | } |
2428 | ||
fc2e4d70 ON |
2429 | struct wq_barrier { |
2430 | struct work_struct work; | |
2431 | struct completion done; | |
2432 | }; | |
2433 | ||
2434 | static void wq_barrier_func(struct work_struct *work) | |
2435 | { | |
2436 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2437 | complete(&barr->done); | |
2438 | } | |
2439 | ||
4690c4ab TH |
2440 | /** |
2441 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2442 | * @pwq: pwq to insert barrier into |
4690c4ab | 2443 | * @barr: wq_barrier to insert |
affee4b2 TH |
2444 | * @target: target work to attach @barr to |
2445 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2446 | * |
affee4b2 TH |
2447 | * @barr is linked to @target such that @barr is completed only after |
2448 | * @target finishes execution. Please note that the ordering | |
2449 | * guarantee is observed only with respect to @target and on the local | |
2450 | * cpu. | |
2451 | * | |
2452 | * Currently, a queued barrier can't be canceled. This is because | |
2453 | * try_to_grab_pending() can't determine whether the work to be | |
2454 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2455 | * flag of the previous work while there must be a valid next work | |
2456 | * after a work with LINKED flag set. | |
2457 | * | |
2458 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2459 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2460 | * |
2461 | * CONTEXT: | |
d565ed63 | 2462 | * spin_lock_irq(pool->lock). |
4690c4ab | 2463 | */ |
112202d9 | 2464 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2465 | struct wq_barrier *barr, |
2466 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2467 | { |
affee4b2 TH |
2468 | struct list_head *head; |
2469 | unsigned int linked = 0; | |
2470 | ||
dc186ad7 | 2471 | /* |
d565ed63 | 2472 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2473 | * as we know for sure that this will not trigger any of the |
2474 | * checks and call back into the fixup functions where we | |
2475 | * might deadlock. | |
2476 | */ | |
ca1cab37 | 2477 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2478 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
fc2e4d70 | 2479 | init_completion(&barr->done); |
83c22520 | 2480 | |
affee4b2 TH |
2481 | /* |
2482 | * If @target is currently being executed, schedule the | |
2483 | * barrier to the worker; otherwise, put it after @target. | |
2484 | */ | |
2485 | if (worker) | |
2486 | head = worker->scheduled.next; | |
2487 | else { | |
2488 | unsigned long *bits = work_data_bits(target); | |
2489 | ||
2490 | head = target->entry.next; | |
2491 | /* there can already be other linked works, inherit and set */ | |
2492 | linked = *bits & WORK_STRUCT_LINKED; | |
2493 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2494 | } | |
2495 | ||
dc186ad7 | 2496 | debug_work_activate(&barr->work); |
112202d9 | 2497 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2498 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2499 | } |
2500 | ||
73f53c4a | 2501 | /** |
112202d9 | 2502 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2503 | * @wq: workqueue being flushed |
2504 | * @flush_color: new flush color, < 0 for no-op | |
2505 | * @work_color: new work color, < 0 for no-op | |
2506 | * | |
112202d9 | 2507 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2508 | * |
112202d9 TH |
2509 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2510 | * -1. If no pwq has in-flight commands at the specified color, all | |
2511 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2512 | * has in flight commands, its pwq->flush_color is set to | |
2513 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2514 | * wakeup logic is armed and %true is returned. |
2515 | * | |
2516 | * The caller should have initialized @wq->first_flusher prior to | |
2517 | * calling this function with non-negative @flush_color. If | |
2518 | * @flush_color is negative, no flush color update is done and %false | |
2519 | * is returned. | |
2520 | * | |
112202d9 | 2521 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2522 | * work_color which is previous to @work_color and all will be |
2523 | * advanced to @work_color. | |
2524 | * | |
2525 | * CONTEXT: | |
2526 | * mutex_lock(wq->flush_mutex). | |
2527 | * | |
2528 | * RETURNS: | |
2529 | * %true if @flush_color >= 0 and there's something to flush. %false | |
2530 | * otherwise. | |
2531 | */ | |
112202d9 | 2532 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2533 | int flush_color, int work_color) |
1da177e4 | 2534 | { |
73f53c4a | 2535 | bool wait = false; |
49e3cf44 | 2536 | struct pool_workqueue *pwq; |
1da177e4 | 2537 | |
73f53c4a | 2538 | if (flush_color >= 0) { |
6183c009 | 2539 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2540 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2541 | } |
2355b70f | 2542 | |
76af4d93 TH |
2543 | local_irq_disable(); |
2544 | ||
49e3cf44 | 2545 | for_each_pwq(pwq, wq) { |
112202d9 | 2546 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2547 | |
76af4d93 | 2548 | spin_lock(&pool->lock); |
83c22520 | 2549 | |
73f53c4a | 2550 | if (flush_color >= 0) { |
6183c009 | 2551 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2552 | |
112202d9 TH |
2553 | if (pwq->nr_in_flight[flush_color]) { |
2554 | pwq->flush_color = flush_color; | |
2555 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2556 | wait = true; |
2557 | } | |
2558 | } | |
1da177e4 | 2559 | |
73f53c4a | 2560 | if (work_color >= 0) { |
6183c009 | 2561 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2562 | pwq->work_color = work_color; |
73f53c4a | 2563 | } |
1da177e4 | 2564 | |
76af4d93 | 2565 | spin_unlock(&pool->lock); |
1da177e4 | 2566 | } |
2355b70f | 2567 | |
76af4d93 TH |
2568 | local_irq_enable(); |
2569 | ||
112202d9 | 2570 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2571 | complete(&wq->first_flusher->done); |
14441960 | 2572 | |
73f53c4a | 2573 | return wait; |
1da177e4 LT |
2574 | } |
2575 | ||
0fcb78c2 | 2576 | /** |
1da177e4 | 2577 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2578 | * @wq: workqueue to flush |
1da177e4 LT |
2579 | * |
2580 | * Forces execution of the workqueue and blocks until its completion. | |
2581 | * This is typically used in driver shutdown handlers. | |
2582 | * | |
fc2e4d70 ON |
2583 | * We sleep until all works which were queued on entry have been handled, |
2584 | * but we are not livelocked by new incoming ones. | |
1da177e4 | 2585 | */ |
7ad5b3a5 | 2586 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2587 | { |
73f53c4a TH |
2588 | struct wq_flusher this_flusher = { |
2589 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2590 | .flush_color = -1, | |
2591 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), | |
2592 | }; | |
2593 | int next_color; | |
1da177e4 | 2594 | |
3295f0ef IM |
2595 | lock_map_acquire(&wq->lockdep_map); |
2596 | lock_map_release(&wq->lockdep_map); | |
73f53c4a TH |
2597 | |
2598 | mutex_lock(&wq->flush_mutex); | |
2599 | ||
2600 | /* | |
2601 | * Start-to-wait phase | |
2602 | */ | |
2603 | next_color = work_next_color(wq->work_color); | |
2604 | ||
2605 | if (next_color != wq->flush_color) { | |
2606 | /* | |
2607 | * Color space is not full. The current work_color | |
2608 | * becomes our flush_color and work_color is advanced | |
2609 | * by one. | |
2610 | */ | |
6183c009 | 2611 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2612 | this_flusher.flush_color = wq->work_color; |
2613 | wq->work_color = next_color; | |
2614 | ||
2615 | if (!wq->first_flusher) { | |
2616 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2617 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2618 | |
2619 | wq->first_flusher = &this_flusher; | |
2620 | ||
112202d9 | 2621 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2622 | wq->work_color)) { |
2623 | /* nothing to flush, done */ | |
2624 | wq->flush_color = next_color; | |
2625 | wq->first_flusher = NULL; | |
2626 | goto out_unlock; | |
2627 | } | |
2628 | } else { | |
2629 | /* wait in queue */ | |
6183c009 | 2630 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2631 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2632 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2633 | } |
2634 | } else { | |
2635 | /* | |
2636 | * Oops, color space is full, wait on overflow queue. | |
2637 | * The next flush completion will assign us | |
2638 | * flush_color and transfer to flusher_queue. | |
2639 | */ | |
2640 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2641 | } | |
2642 | ||
2643 | mutex_unlock(&wq->flush_mutex); | |
2644 | ||
2645 | wait_for_completion(&this_flusher.done); | |
2646 | ||
2647 | /* | |
2648 | * Wake-up-and-cascade phase | |
2649 | * | |
2650 | * First flushers are responsible for cascading flushes and | |
2651 | * handling overflow. Non-first flushers can simply return. | |
2652 | */ | |
2653 | if (wq->first_flusher != &this_flusher) | |
2654 | return; | |
2655 | ||
2656 | mutex_lock(&wq->flush_mutex); | |
2657 | ||
4ce48b37 TH |
2658 | /* we might have raced, check again with mutex held */ |
2659 | if (wq->first_flusher != &this_flusher) | |
2660 | goto out_unlock; | |
2661 | ||
73f53c4a TH |
2662 | wq->first_flusher = NULL; |
2663 | ||
6183c009 TH |
2664 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2665 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2666 | |
2667 | while (true) { | |
2668 | struct wq_flusher *next, *tmp; | |
2669 | ||
2670 | /* complete all the flushers sharing the current flush color */ | |
2671 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2672 | if (next->flush_color != wq->flush_color) | |
2673 | break; | |
2674 | list_del_init(&next->list); | |
2675 | complete(&next->done); | |
2676 | } | |
2677 | ||
6183c009 TH |
2678 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2679 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2680 | |
2681 | /* this flush_color is finished, advance by one */ | |
2682 | wq->flush_color = work_next_color(wq->flush_color); | |
2683 | ||
2684 | /* one color has been freed, handle overflow queue */ | |
2685 | if (!list_empty(&wq->flusher_overflow)) { | |
2686 | /* | |
2687 | * Assign the same color to all overflowed | |
2688 | * flushers, advance work_color and append to | |
2689 | * flusher_queue. This is the start-to-wait | |
2690 | * phase for these overflowed flushers. | |
2691 | */ | |
2692 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2693 | tmp->flush_color = wq->work_color; | |
2694 | ||
2695 | wq->work_color = work_next_color(wq->work_color); | |
2696 | ||
2697 | list_splice_tail_init(&wq->flusher_overflow, | |
2698 | &wq->flusher_queue); | |
112202d9 | 2699 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2700 | } |
2701 | ||
2702 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2703 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2704 | break; |
2705 | } | |
2706 | ||
2707 | /* | |
2708 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2709 | * the new first flusher and arm pwqs. |
73f53c4a | 2710 | */ |
6183c009 TH |
2711 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2712 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2713 | |
2714 | list_del_init(&next->list); | |
2715 | wq->first_flusher = next; | |
2716 | ||
112202d9 | 2717 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2718 | break; |
2719 | ||
2720 | /* | |
2721 | * Meh... this color is already done, clear first | |
2722 | * flusher and repeat cascading. | |
2723 | */ | |
2724 | wq->first_flusher = NULL; | |
2725 | } | |
2726 | ||
2727 | out_unlock: | |
2728 | mutex_unlock(&wq->flush_mutex); | |
1da177e4 | 2729 | } |
ae90dd5d | 2730 | EXPORT_SYMBOL_GPL(flush_workqueue); |
1da177e4 | 2731 | |
9c5a2ba7 TH |
2732 | /** |
2733 | * drain_workqueue - drain a workqueue | |
2734 | * @wq: workqueue to drain | |
2735 | * | |
2736 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2737 | * only chain queueing is allowed. IOW, only currently pending or running | |
2738 | * work items on @wq can queue further work items on it. @wq is flushed | |
2739 | * repeatedly until it becomes empty. The number of flushing is detemined | |
2740 | * by the depth of chaining and should be relatively short. Whine if it | |
2741 | * takes too long. | |
2742 | */ | |
2743 | void drain_workqueue(struct workqueue_struct *wq) | |
2744 | { | |
2745 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2746 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2747 | |
2748 | /* | |
2749 | * __queue_work() needs to test whether there are drainers, is much | |
2750 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
2751 | * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. | |
2752 | */ | |
e98d5b16 | 2753 | spin_lock_irq(&workqueue_lock); |
9c5a2ba7 TH |
2754 | if (!wq->nr_drainers++) |
2755 | wq->flags |= WQ_DRAINING; | |
e98d5b16 | 2756 | spin_unlock_irq(&workqueue_lock); |
9c5a2ba7 TH |
2757 | reflush: |
2758 | flush_workqueue(wq); | |
2759 | ||
76af4d93 TH |
2760 | local_irq_disable(); |
2761 | ||
49e3cf44 | 2762 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2763 | bool drained; |
9c5a2ba7 | 2764 | |
76af4d93 | 2765 | spin_lock(&pwq->pool->lock); |
112202d9 | 2766 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
76af4d93 | 2767 | spin_unlock(&pwq->pool->lock); |
fa2563e4 TT |
2768 | |
2769 | if (drained) | |
9c5a2ba7 TH |
2770 | continue; |
2771 | ||
2772 | if (++flush_cnt == 10 || | |
2773 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
044c782c VI |
2774 | pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", |
2775 | wq->name, flush_cnt); | |
76af4d93 TH |
2776 | |
2777 | local_irq_enable(); | |
9c5a2ba7 TH |
2778 | goto reflush; |
2779 | } | |
2780 | ||
76af4d93 | 2781 | spin_lock(&workqueue_lock); |
9c5a2ba7 TH |
2782 | if (!--wq->nr_drainers) |
2783 | wq->flags &= ~WQ_DRAINING; | |
76af4d93 TH |
2784 | spin_unlock(&workqueue_lock); |
2785 | ||
2786 | local_irq_enable(); | |
9c5a2ba7 TH |
2787 | } |
2788 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2789 | ||
606a5020 | 2790 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) |
db700897 | 2791 | { |
affee4b2 | 2792 | struct worker *worker = NULL; |
c9e7cf27 | 2793 | struct worker_pool *pool; |
112202d9 | 2794 | struct pool_workqueue *pwq; |
db700897 ON |
2795 | |
2796 | might_sleep(); | |
fa1b54e6 TH |
2797 | |
2798 | local_irq_disable(); | |
c9e7cf27 | 2799 | pool = get_work_pool(work); |
fa1b54e6 TH |
2800 | if (!pool) { |
2801 | local_irq_enable(); | |
baf59022 | 2802 | return false; |
fa1b54e6 | 2803 | } |
db700897 | 2804 | |
fa1b54e6 | 2805 | spin_lock(&pool->lock); |
0b3dae68 | 2806 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2807 | pwq = get_work_pwq(work); |
2808 | if (pwq) { | |
2809 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2810 | goto already_gone; |
606a5020 | 2811 | } else { |
c9e7cf27 | 2812 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2813 | if (!worker) |
4690c4ab | 2814 | goto already_gone; |
112202d9 | 2815 | pwq = worker->current_pwq; |
606a5020 | 2816 | } |
db700897 | 2817 | |
112202d9 | 2818 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 2819 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 2820 | |
e159489b TH |
2821 | /* |
2822 | * If @max_active is 1 or rescuer is in use, flushing another work | |
2823 | * item on the same workqueue may lead to deadlock. Make sure the | |
2824 | * flusher is not running on the same workqueue by verifying write | |
2825 | * access. | |
2826 | */ | |
112202d9 TH |
2827 | if (pwq->wq->saved_max_active == 1 || pwq->wq->flags & WQ_RESCUER) |
2828 | lock_map_acquire(&pwq->wq->lockdep_map); | |
e159489b | 2829 | else |
112202d9 TH |
2830 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
2831 | lock_map_release(&pwq->wq->lockdep_map); | |
e159489b | 2832 | |
401a8d04 | 2833 | return true; |
4690c4ab | 2834 | already_gone: |
d565ed63 | 2835 | spin_unlock_irq(&pool->lock); |
401a8d04 | 2836 | return false; |
db700897 | 2837 | } |
baf59022 TH |
2838 | |
2839 | /** | |
2840 | * flush_work - wait for a work to finish executing the last queueing instance | |
2841 | * @work: the work to flush | |
2842 | * | |
606a5020 TH |
2843 | * Wait until @work has finished execution. @work is guaranteed to be idle |
2844 | * on return if it hasn't been requeued since flush started. | |
baf59022 TH |
2845 | * |
2846 | * RETURNS: | |
2847 | * %true if flush_work() waited for the work to finish execution, | |
2848 | * %false if it was already idle. | |
2849 | */ | |
2850 | bool flush_work(struct work_struct *work) | |
2851 | { | |
2852 | struct wq_barrier barr; | |
2853 | ||
0976dfc1 SB |
2854 | lock_map_acquire(&work->lockdep_map); |
2855 | lock_map_release(&work->lockdep_map); | |
2856 | ||
606a5020 | 2857 | if (start_flush_work(work, &barr)) { |
401a8d04 TH |
2858 | wait_for_completion(&barr.done); |
2859 | destroy_work_on_stack(&barr.work); | |
2860 | return true; | |
606a5020 | 2861 | } else { |
401a8d04 | 2862 | return false; |
6e84d644 | 2863 | } |
6e84d644 | 2864 | } |
606a5020 | 2865 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 2866 | |
36e227d2 | 2867 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 2868 | { |
bbb68dfa | 2869 | unsigned long flags; |
1f1f642e ON |
2870 | int ret; |
2871 | ||
2872 | do { | |
bbb68dfa TH |
2873 | ret = try_to_grab_pending(work, is_dwork, &flags); |
2874 | /* | |
2875 | * If someone else is canceling, wait for the same event it | |
2876 | * would be waiting for before retrying. | |
2877 | */ | |
2878 | if (unlikely(ret == -ENOENT)) | |
606a5020 | 2879 | flush_work(work); |
1f1f642e ON |
2880 | } while (unlikely(ret < 0)); |
2881 | ||
bbb68dfa TH |
2882 | /* tell other tasks trying to grab @work to back off */ |
2883 | mark_work_canceling(work); | |
2884 | local_irq_restore(flags); | |
2885 | ||
606a5020 | 2886 | flush_work(work); |
7a22ad75 | 2887 | clear_work_data(work); |
1f1f642e ON |
2888 | return ret; |
2889 | } | |
2890 | ||
6e84d644 | 2891 | /** |
401a8d04 TH |
2892 | * cancel_work_sync - cancel a work and wait for it to finish |
2893 | * @work: the work to cancel | |
6e84d644 | 2894 | * |
401a8d04 TH |
2895 | * Cancel @work and wait for its execution to finish. This function |
2896 | * can be used even if the work re-queues itself or migrates to | |
2897 | * another workqueue. On return from this function, @work is | |
2898 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 2899 | * |
401a8d04 TH |
2900 | * cancel_work_sync(&delayed_work->work) must not be used for |
2901 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 2902 | * |
401a8d04 | 2903 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 2904 | * queued can't be destroyed before this function returns. |
401a8d04 TH |
2905 | * |
2906 | * RETURNS: | |
2907 | * %true if @work was pending, %false otherwise. | |
6e84d644 | 2908 | */ |
401a8d04 | 2909 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 2910 | { |
36e227d2 | 2911 | return __cancel_work_timer(work, false); |
b89deed3 | 2912 | } |
28e53bdd | 2913 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 2914 | |
6e84d644 | 2915 | /** |
401a8d04 TH |
2916 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
2917 | * @dwork: the delayed work to flush | |
6e84d644 | 2918 | * |
401a8d04 TH |
2919 | * Delayed timer is cancelled and the pending work is queued for |
2920 | * immediate execution. Like flush_work(), this function only | |
2921 | * considers the last queueing instance of @dwork. | |
1f1f642e | 2922 | * |
401a8d04 TH |
2923 | * RETURNS: |
2924 | * %true if flush_work() waited for the work to finish execution, | |
2925 | * %false if it was already idle. | |
6e84d644 | 2926 | */ |
401a8d04 TH |
2927 | bool flush_delayed_work(struct delayed_work *dwork) |
2928 | { | |
8930caba | 2929 | local_irq_disable(); |
401a8d04 | 2930 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 2931 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 2932 | local_irq_enable(); |
401a8d04 TH |
2933 | return flush_work(&dwork->work); |
2934 | } | |
2935 | EXPORT_SYMBOL(flush_delayed_work); | |
2936 | ||
09383498 | 2937 | /** |
57b30ae7 TH |
2938 | * cancel_delayed_work - cancel a delayed work |
2939 | * @dwork: delayed_work to cancel | |
09383498 | 2940 | * |
57b30ae7 TH |
2941 | * Kill off a pending delayed_work. Returns %true if @dwork was pending |
2942 | * and canceled; %false if wasn't pending. Note that the work callback | |
2943 | * function may still be running on return, unless it returns %true and the | |
2944 | * work doesn't re-arm itself. Explicitly flush or use | |
2945 | * cancel_delayed_work_sync() to wait on it. | |
09383498 | 2946 | * |
57b30ae7 | 2947 | * This function is safe to call from any context including IRQ handler. |
09383498 | 2948 | */ |
57b30ae7 | 2949 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 2950 | { |
57b30ae7 TH |
2951 | unsigned long flags; |
2952 | int ret; | |
2953 | ||
2954 | do { | |
2955 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
2956 | } while (unlikely(ret == -EAGAIN)); | |
2957 | ||
2958 | if (unlikely(ret < 0)) | |
2959 | return false; | |
2960 | ||
7c3eed5c TH |
2961 | set_work_pool_and_clear_pending(&dwork->work, |
2962 | get_work_pool_id(&dwork->work)); | |
57b30ae7 | 2963 | local_irq_restore(flags); |
c0158ca6 | 2964 | return ret; |
09383498 | 2965 | } |
57b30ae7 | 2966 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 2967 | |
401a8d04 TH |
2968 | /** |
2969 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
2970 | * @dwork: the delayed work cancel | |
2971 | * | |
2972 | * This is cancel_work_sync() for delayed works. | |
2973 | * | |
2974 | * RETURNS: | |
2975 | * %true if @dwork was pending, %false otherwise. | |
2976 | */ | |
2977 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 2978 | { |
36e227d2 | 2979 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 2980 | } |
f5a421a4 | 2981 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 2982 | |
0fcb78c2 | 2983 | /** |
c1a220e7 ZR |
2984 | * schedule_work_on - put work task on a specific cpu |
2985 | * @cpu: cpu to put the work task on | |
2986 | * @work: job to be done | |
2987 | * | |
2988 | * This puts a job on a specific cpu | |
2989 | */ | |
d4283e93 | 2990 | bool schedule_work_on(int cpu, struct work_struct *work) |
c1a220e7 | 2991 | { |
d320c038 | 2992 | return queue_work_on(cpu, system_wq, work); |
c1a220e7 ZR |
2993 | } |
2994 | EXPORT_SYMBOL(schedule_work_on); | |
2995 | ||
0fcb78c2 | 2996 | /** |
0fcb78c2 REB |
2997 | * schedule_work - put work task in global workqueue |
2998 | * @work: job to be done | |
0fcb78c2 | 2999 | * |
d4283e93 TH |
3000 | * Returns %false if @work was already on the kernel-global workqueue and |
3001 | * %true otherwise. | |
5b0f437d BVA |
3002 | * |
3003 | * This puts a job in the kernel-global workqueue if it was not already | |
3004 | * queued and leaves it in the same position on the kernel-global | |
3005 | * workqueue otherwise. | |
0fcb78c2 | 3006 | */ |
d4283e93 | 3007 | bool schedule_work(struct work_struct *work) |
1da177e4 | 3008 | { |
d320c038 | 3009 | return queue_work(system_wq, work); |
1da177e4 | 3010 | } |
ae90dd5d | 3011 | EXPORT_SYMBOL(schedule_work); |
1da177e4 | 3012 | |
0fcb78c2 REB |
3013 | /** |
3014 | * schedule_delayed_work_on - queue work in global workqueue on CPU after delay | |
3015 | * @cpu: cpu to use | |
52bad64d | 3016 | * @dwork: job to be done |
0fcb78c2 REB |
3017 | * @delay: number of jiffies to wait |
3018 | * | |
3019 | * After waiting for a given time this puts a job in the kernel-global | |
3020 | * workqueue on the specified CPU. | |
3021 | */ | |
d4283e93 TH |
3022 | bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, |
3023 | unsigned long delay) | |
1da177e4 | 3024 | { |
d320c038 | 3025 | return queue_delayed_work_on(cpu, system_wq, dwork, delay); |
1da177e4 | 3026 | } |
ae90dd5d | 3027 | EXPORT_SYMBOL(schedule_delayed_work_on); |
1da177e4 | 3028 | |
0fcb78c2 REB |
3029 | /** |
3030 | * schedule_delayed_work - put work task in global workqueue after delay | |
52bad64d DH |
3031 | * @dwork: job to be done |
3032 | * @delay: number of jiffies to wait or 0 for immediate execution | |
0fcb78c2 REB |
3033 | * |
3034 | * After waiting for a given time this puts a job in the kernel-global | |
3035 | * workqueue. | |
3036 | */ | |
d4283e93 | 3037 | bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) |
1da177e4 | 3038 | { |
d320c038 | 3039 | return queue_delayed_work(system_wq, dwork, delay); |
1da177e4 | 3040 | } |
ae90dd5d | 3041 | EXPORT_SYMBOL(schedule_delayed_work); |
1da177e4 | 3042 | |
b6136773 | 3043 | /** |
31ddd871 | 3044 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3045 | * @func: the function to call |
b6136773 | 3046 | * |
31ddd871 TH |
3047 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3048 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3049 | * schedule_on_each_cpu() is very slow. |
31ddd871 TH |
3050 | * |
3051 | * RETURNS: | |
3052 | * 0 on success, -errno on failure. | |
b6136773 | 3053 | */ |
65f27f38 | 3054 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3055 | { |
3056 | int cpu; | |
38f51568 | 3057 | struct work_struct __percpu *works; |
15316ba8 | 3058 | |
b6136773 AM |
3059 | works = alloc_percpu(struct work_struct); |
3060 | if (!works) | |
15316ba8 | 3061 | return -ENOMEM; |
b6136773 | 3062 | |
93981800 TH |
3063 | get_online_cpus(); |
3064 | ||
15316ba8 | 3065 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3066 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3067 | ||
3068 | INIT_WORK(work, func); | |
b71ab8c2 | 3069 | schedule_work_on(cpu, work); |
65a64464 | 3070 | } |
93981800 TH |
3071 | |
3072 | for_each_online_cpu(cpu) | |
3073 | flush_work(per_cpu_ptr(works, cpu)); | |
3074 | ||
95402b38 | 3075 | put_online_cpus(); |
b6136773 | 3076 | free_percpu(works); |
15316ba8 CL |
3077 | return 0; |
3078 | } | |
3079 | ||
eef6a7d5 AS |
3080 | /** |
3081 | * flush_scheduled_work - ensure that any scheduled work has run to completion. | |
3082 | * | |
3083 | * Forces execution of the kernel-global workqueue and blocks until its | |
3084 | * completion. | |
3085 | * | |
3086 | * Think twice before calling this function! It's very easy to get into | |
3087 | * trouble if you don't take great care. Either of the following situations | |
3088 | * will lead to deadlock: | |
3089 | * | |
3090 | * One of the work items currently on the workqueue needs to acquire | |
3091 | * a lock held by your code or its caller. | |
3092 | * | |
3093 | * Your code is running in the context of a work routine. | |
3094 | * | |
3095 | * They will be detected by lockdep when they occur, but the first might not | |
3096 | * occur very often. It depends on what work items are on the workqueue and | |
3097 | * what locks they need, which you have no control over. | |
3098 | * | |
3099 | * In most situations flushing the entire workqueue is overkill; you merely | |
3100 | * need to know that a particular work item isn't queued and isn't running. | |
3101 | * In such cases you should use cancel_delayed_work_sync() or | |
3102 | * cancel_work_sync() instead. | |
3103 | */ | |
1da177e4 LT |
3104 | void flush_scheduled_work(void) |
3105 | { | |
d320c038 | 3106 | flush_workqueue(system_wq); |
1da177e4 | 3107 | } |
ae90dd5d | 3108 | EXPORT_SYMBOL(flush_scheduled_work); |
1da177e4 | 3109 | |
1fa44eca JB |
3110 | /** |
3111 | * execute_in_process_context - reliably execute the routine with user context | |
3112 | * @fn: the function to execute | |
1fa44eca JB |
3113 | * @ew: guaranteed storage for the execute work structure (must |
3114 | * be available when the work executes) | |
3115 | * | |
3116 | * Executes the function immediately if process context is available, | |
3117 | * otherwise schedules the function for delayed execution. | |
3118 | * | |
3119 | * Returns: 0 - function was executed | |
3120 | * 1 - function was scheduled for execution | |
3121 | */ | |
65f27f38 | 3122 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3123 | { |
3124 | if (!in_interrupt()) { | |
65f27f38 | 3125 | fn(&ew->work); |
1fa44eca JB |
3126 | return 0; |
3127 | } | |
3128 | ||
65f27f38 | 3129 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3130 | schedule_work(&ew->work); |
3131 | ||
3132 | return 1; | |
3133 | } | |
3134 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3135 | ||
1da177e4 LT |
3136 | int keventd_up(void) |
3137 | { | |
d320c038 | 3138 | return system_wq != NULL; |
1da177e4 LT |
3139 | } |
3140 | ||
7a4e344c TH |
3141 | /** |
3142 | * free_workqueue_attrs - free a workqueue_attrs | |
3143 | * @attrs: workqueue_attrs to free | |
3144 | * | |
3145 | * Undo alloc_workqueue_attrs(). | |
3146 | */ | |
3147 | void free_workqueue_attrs(struct workqueue_attrs *attrs) | |
3148 | { | |
3149 | if (attrs) { | |
3150 | free_cpumask_var(attrs->cpumask); | |
3151 | kfree(attrs); | |
3152 | } | |
3153 | } | |
3154 | ||
3155 | /** | |
3156 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
3157 | * @gfp_mask: allocation mask to use | |
3158 | * | |
3159 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3160 | * return it. Returns NULL on failure. | |
3161 | */ | |
3162 | struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) | |
3163 | { | |
3164 | struct workqueue_attrs *attrs; | |
3165 | ||
3166 | attrs = kzalloc(sizeof(*attrs), gfp_mask); | |
3167 | if (!attrs) | |
3168 | goto fail; | |
3169 | if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) | |
3170 | goto fail; | |
3171 | ||
3172 | cpumask_setall(attrs->cpumask); | |
3173 | return attrs; | |
3174 | fail: | |
3175 | free_workqueue_attrs(attrs); | |
3176 | return NULL; | |
3177 | } | |
3178 | ||
29c91e99 TH |
3179 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3180 | const struct workqueue_attrs *from) | |
3181 | { | |
3182 | to->nice = from->nice; | |
3183 | cpumask_copy(to->cpumask, from->cpumask); | |
3184 | } | |
3185 | ||
3186 | /* | |
3187 | * Hacky implementation of jhash of bitmaps which only considers the | |
3188 | * specified number of bits. We probably want a proper implementation in | |
3189 | * include/linux/jhash.h. | |
3190 | */ | |
3191 | static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash) | |
3192 | { | |
3193 | int nr_longs = bits / BITS_PER_LONG; | |
3194 | int nr_leftover = bits % BITS_PER_LONG; | |
3195 | unsigned long leftover = 0; | |
3196 | ||
3197 | if (nr_longs) | |
3198 | hash = jhash(bitmap, nr_longs * sizeof(long), hash); | |
3199 | if (nr_leftover) { | |
3200 | bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover); | |
3201 | hash = jhash(&leftover, sizeof(long), hash); | |
3202 | } | |
3203 | return hash; | |
3204 | } | |
3205 | ||
3206 | /* hash value of the content of @attr */ | |
3207 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
3208 | { | |
3209 | u32 hash = 0; | |
3210 | ||
3211 | hash = jhash_1word(attrs->nice, hash); | |
3212 | hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash); | |
3213 | return hash; | |
3214 | } | |
3215 | ||
3216 | /* content equality test */ | |
3217 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3218 | const struct workqueue_attrs *b) | |
3219 | { | |
3220 | if (a->nice != b->nice) | |
3221 | return false; | |
3222 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3223 | return false; | |
3224 | return true; | |
3225 | } | |
3226 | ||
7a4e344c TH |
3227 | /** |
3228 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3229 | * @pool: worker_pool to initialize | |
3230 | * | |
3231 | * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. | |
29c91e99 TH |
3232 | * Returns 0 on success, -errno on failure. Even on failure, all fields |
3233 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3234 | * on @pool safely to release it. | |
7a4e344c TH |
3235 | */ |
3236 | static int init_worker_pool(struct worker_pool *pool) | |
4e1a1f9a TH |
3237 | { |
3238 | spin_lock_init(&pool->lock); | |
29c91e99 TH |
3239 | pool->id = -1; |
3240 | pool->cpu = -1; | |
4e1a1f9a TH |
3241 | pool->flags |= POOL_DISASSOCIATED; |
3242 | INIT_LIST_HEAD(&pool->worklist); | |
3243 | INIT_LIST_HEAD(&pool->idle_list); | |
3244 | hash_init(pool->busy_hash); | |
3245 | ||
3246 | init_timer_deferrable(&pool->idle_timer); | |
3247 | pool->idle_timer.function = idle_worker_timeout; | |
3248 | pool->idle_timer.data = (unsigned long)pool; | |
3249 | ||
3250 | setup_timer(&pool->mayday_timer, pool_mayday_timeout, | |
3251 | (unsigned long)pool); | |
3252 | ||
3253 | mutex_init(&pool->manager_arb); | |
3254 | mutex_init(&pool->assoc_mutex); | |
3255 | ida_init(&pool->worker_ida); | |
7a4e344c | 3256 | |
29c91e99 TH |
3257 | INIT_HLIST_NODE(&pool->hash_node); |
3258 | pool->refcnt = 1; | |
3259 | ||
3260 | /* shouldn't fail above this point */ | |
7a4e344c TH |
3261 | pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); |
3262 | if (!pool->attrs) | |
3263 | return -ENOMEM; | |
3264 | return 0; | |
4e1a1f9a TH |
3265 | } |
3266 | ||
29c91e99 TH |
3267 | static void rcu_free_pool(struct rcu_head *rcu) |
3268 | { | |
3269 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); | |
3270 | ||
3271 | ida_destroy(&pool->worker_ida); | |
3272 | free_workqueue_attrs(pool->attrs); | |
3273 | kfree(pool); | |
3274 | } | |
3275 | ||
3276 | /** | |
3277 | * put_unbound_pool - put a worker_pool | |
3278 | * @pool: worker_pool to put | |
3279 | * | |
3280 | * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU | |
3281 | * safe manner. | |
3282 | */ | |
3283 | static void put_unbound_pool(struct worker_pool *pool) | |
3284 | { | |
3285 | struct worker *worker; | |
3286 | ||
3287 | spin_lock_irq(&workqueue_lock); | |
3288 | if (--pool->refcnt) { | |
3289 | spin_unlock_irq(&workqueue_lock); | |
3290 | return; | |
3291 | } | |
3292 | ||
3293 | /* sanity checks */ | |
3294 | if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || | |
3295 | WARN_ON(!list_empty(&pool->worklist))) { | |
3296 | spin_unlock_irq(&workqueue_lock); | |
3297 | return; | |
3298 | } | |
3299 | ||
3300 | /* release id and unhash */ | |
3301 | if (pool->id >= 0) | |
3302 | idr_remove(&worker_pool_idr, pool->id); | |
3303 | hash_del(&pool->hash_node); | |
3304 | ||
3305 | spin_unlock_irq(&workqueue_lock); | |
3306 | ||
3307 | /* lock out manager and destroy all workers */ | |
3308 | mutex_lock(&pool->manager_arb); | |
3309 | spin_lock_irq(&pool->lock); | |
3310 | ||
3311 | while ((worker = first_worker(pool))) | |
3312 | destroy_worker(worker); | |
3313 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
3314 | ||
3315 | spin_unlock_irq(&pool->lock); | |
3316 | mutex_unlock(&pool->manager_arb); | |
3317 | ||
3318 | /* shut down the timers */ | |
3319 | del_timer_sync(&pool->idle_timer); | |
3320 | del_timer_sync(&pool->mayday_timer); | |
3321 | ||
3322 | /* sched-RCU protected to allow dereferences from get_work_pool() */ | |
3323 | call_rcu_sched(&pool->rcu, rcu_free_pool); | |
3324 | } | |
3325 | ||
3326 | /** | |
3327 | * get_unbound_pool - get a worker_pool with the specified attributes | |
3328 | * @attrs: the attributes of the worker_pool to get | |
3329 | * | |
3330 | * Obtain a worker_pool which has the same attributes as @attrs, bump the | |
3331 | * reference count and return it. If there already is a matching | |
3332 | * worker_pool, it will be used; otherwise, this function attempts to | |
3333 | * create a new one. On failure, returns NULL. | |
3334 | */ | |
3335 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) | |
3336 | { | |
3337 | static DEFINE_MUTEX(create_mutex); | |
3338 | u32 hash = wqattrs_hash(attrs); | |
3339 | struct worker_pool *pool; | |
3340 | struct worker *worker; | |
3341 | ||
3342 | mutex_lock(&create_mutex); | |
3343 | ||
3344 | /* do we already have a matching pool? */ | |
3345 | spin_lock_irq(&workqueue_lock); | |
3346 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3347 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3348 | pool->refcnt++; | |
3349 | goto out_unlock; | |
3350 | } | |
3351 | } | |
3352 | spin_unlock_irq(&workqueue_lock); | |
3353 | ||
3354 | /* nope, create a new one */ | |
3355 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); | |
3356 | if (!pool || init_worker_pool(pool) < 0) | |
3357 | goto fail; | |
3358 | ||
3359 | copy_workqueue_attrs(pool->attrs, attrs); | |
3360 | ||
3361 | if (worker_pool_assign_id(pool) < 0) | |
3362 | goto fail; | |
3363 | ||
3364 | /* create and start the initial worker */ | |
3365 | worker = create_worker(pool); | |
3366 | if (!worker) | |
3367 | goto fail; | |
3368 | ||
3369 | spin_lock_irq(&pool->lock); | |
3370 | start_worker(worker); | |
3371 | spin_unlock_irq(&pool->lock); | |
3372 | ||
3373 | /* install */ | |
3374 | spin_lock_irq(&workqueue_lock); | |
3375 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
3376 | out_unlock: | |
3377 | spin_unlock_irq(&workqueue_lock); | |
3378 | mutex_unlock(&create_mutex); | |
3379 | return pool; | |
3380 | fail: | |
3381 | mutex_unlock(&create_mutex); | |
3382 | if (pool) | |
3383 | put_unbound_pool(pool); | |
3384 | return NULL; | |
3385 | } | |
3386 | ||
30cdf249 | 3387 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 3388 | { |
49e3cf44 | 3389 | bool highpri = wq->flags & WQ_HIGHPRI; |
30cdf249 TH |
3390 | int cpu; |
3391 | ||
3392 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
3393 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
3394 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
3395 | return -ENOMEM; |
3396 | ||
3397 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
3398 | struct pool_workqueue *pwq = |
3399 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
f3421797 | 3400 | |
49e3cf44 | 3401 | pwq->pool = get_std_worker_pool(cpu, highpri); |
76af4d93 | 3402 | list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); |
30cdf249 TH |
3403 | } |
3404 | } else { | |
3405 | struct pool_workqueue *pwq; | |
3406 | ||
3407 | pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); | |
3408 | if (!pwq) | |
3409 | return -ENOMEM; | |
3410 | ||
29c91e99 TH |
3411 | pwq->pool = get_unbound_pool(unbound_std_wq_attrs[highpri]); |
3412 | if (!pwq->pool) { | |
3413 | kmem_cache_free(pwq_cache, pwq); | |
3414 | return -ENOMEM; | |
3415 | } | |
3416 | ||
76af4d93 | 3417 | list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); |
30cdf249 TH |
3418 | } |
3419 | ||
3420 | return 0; | |
0f900049 TH |
3421 | } |
3422 | ||
112202d9 | 3423 | static void free_pwqs(struct workqueue_struct *wq) |
0f900049 | 3424 | { |
e06ffa1e | 3425 | if (!(wq->flags & WQ_UNBOUND)) |
420c0ddb TH |
3426 | free_percpu(wq->cpu_pwqs); |
3427 | else if (!list_empty(&wq->pwqs)) | |
3428 | kmem_cache_free(pwq_cache, list_first_entry(&wq->pwqs, | |
3429 | struct pool_workqueue, pwqs_node)); | |
0f900049 TH |
3430 | } |
3431 | ||
f3421797 TH |
3432 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
3433 | const char *name) | |
b71ab8c2 | 3434 | { |
f3421797 TH |
3435 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
3436 | ||
3437 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
3438 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
3439 | max_active, name, 1, lim); | |
b71ab8c2 | 3440 | |
f3421797 | 3441 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
3442 | } |
3443 | ||
b196be89 | 3444 | struct workqueue_struct *__alloc_workqueue_key(const char *fmt, |
d320c038 TH |
3445 | unsigned int flags, |
3446 | int max_active, | |
3447 | struct lock_class_key *key, | |
b196be89 | 3448 | const char *lock_name, ...) |
1da177e4 | 3449 | { |
b196be89 | 3450 | va_list args, args1; |
1da177e4 | 3451 | struct workqueue_struct *wq; |
49e3cf44 | 3452 | struct pool_workqueue *pwq; |
b196be89 TH |
3453 | size_t namelen; |
3454 | ||
3455 | /* determine namelen, allocate wq and format name */ | |
3456 | va_start(args, lock_name); | |
3457 | va_copy(args1, args); | |
3458 | namelen = vsnprintf(NULL, 0, fmt, args) + 1; | |
3459 | ||
3460 | wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); | |
3461 | if (!wq) | |
3462 | goto err; | |
3463 | ||
3464 | vsnprintf(wq->name, namelen, fmt, args1); | |
3465 | va_end(args); | |
3466 | va_end(args1); | |
1da177e4 | 3467 | |
6370a6ad TH |
3468 | /* |
3469 | * Workqueues which may be used during memory reclaim should | |
3470 | * have a rescuer to guarantee forward progress. | |
3471 | */ | |
3472 | if (flags & WQ_MEM_RECLAIM) | |
3473 | flags |= WQ_RESCUER; | |
3474 | ||
d320c038 | 3475 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 3476 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 3477 | |
b196be89 | 3478 | /* init wq */ |
97e37d7b | 3479 | wq->flags = flags; |
a0a1a5fd | 3480 | wq->saved_max_active = max_active; |
73f53c4a | 3481 | mutex_init(&wq->flush_mutex); |
112202d9 | 3482 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 3483 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
3484 | INIT_LIST_HEAD(&wq->flusher_queue); |
3485 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 3486 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 3487 | |
eb13ba87 | 3488 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
cce1a165 | 3489 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 3490 | |
30cdf249 | 3491 | if (alloc_and_link_pwqs(wq) < 0) |
bdbc5dd7 TH |
3492 | goto err; |
3493 | ||
76af4d93 | 3494 | local_irq_disable(); |
49e3cf44 | 3495 | for_each_pwq(pwq, wq) { |
112202d9 | 3496 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
112202d9 TH |
3497 | pwq->wq = wq; |
3498 | pwq->flush_color = -1; | |
3499 | pwq->max_active = max_active; | |
3500 | INIT_LIST_HEAD(&pwq->delayed_works); | |
493a1724 | 3501 | INIT_LIST_HEAD(&pwq->mayday_node); |
e22bee78 | 3502 | } |
76af4d93 | 3503 | local_irq_enable(); |
1537663f | 3504 | |
e22bee78 TH |
3505 | if (flags & WQ_RESCUER) { |
3506 | struct worker *rescuer; | |
3507 | ||
e22bee78 TH |
3508 | wq->rescuer = rescuer = alloc_worker(); |
3509 | if (!rescuer) | |
3510 | goto err; | |
3511 | ||
111c225a TH |
3512 | rescuer->rescue_wq = wq; |
3513 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", | |
b196be89 | 3514 | wq->name); |
e22bee78 TH |
3515 | if (IS_ERR(rescuer->task)) |
3516 | goto err; | |
3517 | ||
e22bee78 TH |
3518 | rescuer->task->flags |= PF_THREAD_BOUND; |
3519 | wake_up_process(rescuer->task); | |
3af24433 ON |
3520 | } |
3521 | ||
a0a1a5fd TH |
3522 | /* |
3523 | * workqueue_lock protects global freeze state and workqueues | |
3524 | * list. Grab it, set max_active accordingly and add the new | |
3525 | * workqueue to workqueues list. | |
3526 | */ | |
e98d5b16 | 3527 | spin_lock_irq(&workqueue_lock); |
a0a1a5fd | 3528 | |
58a69cb4 | 3529 | if (workqueue_freezing && wq->flags & WQ_FREEZABLE) |
49e3cf44 TH |
3530 | for_each_pwq(pwq, wq) |
3531 | pwq->max_active = 0; | |
a0a1a5fd | 3532 | |
1537663f | 3533 | list_add(&wq->list, &workqueues); |
a0a1a5fd | 3534 | |
e98d5b16 | 3535 | spin_unlock_irq(&workqueue_lock); |
1537663f | 3536 | |
3af24433 | 3537 | return wq; |
4690c4ab TH |
3538 | err: |
3539 | if (wq) { | |
112202d9 | 3540 | free_pwqs(wq); |
e22bee78 | 3541 | kfree(wq->rescuer); |
4690c4ab TH |
3542 | kfree(wq); |
3543 | } | |
3544 | return NULL; | |
3af24433 | 3545 | } |
d320c038 | 3546 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
1da177e4 | 3547 | |
3af24433 ON |
3548 | /** |
3549 | * destroy_workqueue - safely terminate a workqueue | |
3550 | * @wq: target workqueue | |
3551 | * | |
3552 | * Safely destroy a workqueue. All work currently pending will be done first. | |
3553 | */ | |
3554 | void destroy_workqueue(struct workqueue_struct *wq) | |
3555 | { | |
49e3cf44 | 3556 | struct pool_workqueue *pwq; |
3af24433 | 3557 | |
9c5a2ba7 TH |
3558 | /* drain it before proceeding with destruction */ |
3559 | drain_workqueue(wq); | |
c8efcc25 | 3560 | |
76af4d93 TH |
3561 | spin_lock_irq(&workqueue_lock); |
3562 | ||
6183c009 | 3563 | /* sanity checks */ |
49e3cf44 | 3564 | for_each_pwq(pwq, wq) { |
6183c009 TH |
3565 | int i; |
3566 | ||
76af4d93 TH |
3567 | for (i = 0; i < WORK_NR_COLORS; i++) { |
3568 | if (WARN_ON(pwq->nr_in_flight[i])) { | |
3569 | spin_unlock_irq(&workqueue_lock); | |
6183c009 | 3570 | return; |
76af4d93 TH |
3571 | } |
3572 | } | |
3573 | ||
6183c009 | 3574 | if (WARN_ON(pwq->nr_active) || |
76af4d93 TH |
3575 | WARN_ON(!list_empty(&pwq->delayed_works))) { |
3576 | spin_unlock_irq(&workqueue_lock); | |
6183c009 | 3577 | return; |
76af4d93 | 3578 | } |
6183c009 TH |
3579 | } |
3580 | ||
a0a1a5fd TH |
3581 | /* |
3582 | * wq list is used to freeze wq, remove from list after | |
3583 | * flushing is complete in case freeze races us. | |
3584 | */ | |
b1f4ec17 | 3585 | list_del(&wq->list); |
76af4d93 | 3586 | |
e98d5b16 | 3587 | spin_unlock_irq(&workqueue_lock); |
3af24433 | 3588 | |
e22bee78 TH |
3589 | if (wq->flags & WQ_RESCUER) { |
3590 | kthread_stop(wq->rescuer->task); | |
8d9df9f0 | 3591 | kfree(wq->rescuer); |
e22bee78 TH |
3592 | } |
3593 | ||
29c91e99 TH |
3594 | /* |
3595 | * We're the sole accessor of @wq at this point. Directly access | |
3596 | * the first pwq and put its pool. | |
3597 | */ | |
3598 | if (wq->flags & WQ_UNBOUND) { | |
3599 | pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, | |
3600 | pwqs_node); | |
3601 | put_unbound_pool(pwq->pool); | |
3602 | } | |
112202d9 | 3603 | free_pwqs(wq); |
3af24433 ON |
3604 | kfree(wq); |
3605 | } | |
3606 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
3607 | ||
9f4bd4cd | 3608 | /** |
112202d9 TH |
3609 | * pwq_set_max_active - adjust max_active of a pwq |
3610 | * @pwq: target pool_workqueue | |
9f4bd4cd LJ |
3611 | * @max_active: new max_active value. |
3612 | * | |
112202d9 | 3613 | * Set @pwq->max_active to @max_active and activate delayed works if |
9f4bd4cd LJ |
3614 | * increased. |
3615 | * | |
3616 | * CONTEXT: | |
d565ed63 | 3617 | * spin_lock_irq(pool->lock). |
9f4bd4cd | 3618 | */ |
112202d9 | 3619 | static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active) |
9f4bd4cd | 3620 | { |
112202d9 | 3621 | pwq->max_active = max_active; |
9f4bd4cd | 3622 | |
112202d9 TH |
3623 | while (!list_empty(&pwq->delayed_works) && |
3624 | pwq->nr_active < pwq->max_active) | |
3625 | pwq_activate_first_delayed(pwq); | |
9f4bd4cd LJ |
3626 | } |
3627 | ||
dcd989cb TH |
3628 | /** |
3629 | * workqueue_set_max_active - adjust max_active of a workqueue | |
3630 | * @wq: target workqueue | |
3631 | * @max_active: new max_active value. | |
3632 | * | |
3633 | * Set max_active of @wq to @max_active. | |
3634 | * | |
3635 | * CONTEXT: | |
3636 | * Don't call from IRQ context. | |
3637 | */ | |
3638 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
3639 | { | |
49e3cf44 | 3640 | struct pool_workqueue *pwq; |
dcd989cb | 3641 | |
f3421797 | 3642 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 3643 | |
e98d5b16 | 3644 | spin_lock_irq(&workqueue_lock); |
dcd989cb TH |
3645 | |
3646 | wq->saved_max_active = max_active; | |
3647 | ||
49e3cf44 | 3648 | for_each_pwq(pwq, wq) { |
112202d9 | 3649 | struct worker_pool *pool = pwq->pool; |
dcd989cb | 3650 | |
e98d5b16 | 3651 | spin_lock(&pool->lock); |
dcd989cb | 3652 | |
58a69cb4 | 3653 | if (!(wq->flags & WQ_FREEZABLE) || |
35b6bb63 | 3654 | !(pool->flags & POOL_FREEZING)) |
112202d9 | 3655 | pwq_set_max_active(pwq, max_active); |
9bfb1839 | 3656 | |
e98d5b16 | 3657 | spin_unlock(&pool->lock); |
65a64464 | 3658 | } |
93981800 | 3659 | |
e98d5b16 | 3660 | spin_unlock_irq(&workqueue_lock); |
15316ba8 | 3661 | } |
dcd989cb | 3662 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 3663 | |
eef6a7d5 | 3664 | /** |
dcd989cb TH |
3665 | * workqueue_congested - test whether a workqueue is congested |
3666 | * @cpu: CPU in question | |
3667 | * @wq: target workqueue | |
eef6a7d5 | 3668 | * |
dcd989cb TH |
3669 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
3670 | * no synchronization around this function and the test result is | |
3671 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 3672 | * |
dcd989cb TH |
3673 | * RETURNS: |
3674 | * %true if congested, %false otherwise. | |
eef6a7d5 | 3675 | */ |
d84ff051 | 3676 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 3677 | { |
7fb98ea7 | 3678 | struct pool_workqueue *pwq; |
76af4d93 TH |
3679 | bool ret; |
3680 | ||
3681 | preempt_disable(); | |
7fb98ea7 TH |
3682 | |
3683 | if (!(wq->flags & WQ_UNBOUND)) | |
3684 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
3685 | else | |
3686 | pwq = first_pwq(wq); | |
dcd989cb | 3687 | |
76af4d93 TH |
3688 | ret = !list_empty(&pwq->delayed_works); |
3689 | preempt_enable(); | |
3690 | ||
3691 | return ret; | |
1da177e4 | 3692 | } |
dcd989cb | 3693 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 3694 | |
dcd989cb TH |
3695 | /** |
3696 | * work_busy - test whether a work is currently pending or running | |
3697 | * @work: the work to be tested | |
3698 | * | |
3699 | * Test whether @work is currently pending or running. There is no | |
3700 | * synchronization around this function and the test result is | |
3701 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb TH |
3702 | * |
3703 | * RETURNS: | |
3704 | * OR'd bitmask of WORK_BUSY_* bits. | |
3705 | */ | |
3706 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 3707 | { |
fa1b54e6 | 3708 | struct worker_pool *pool; |
dcd989cb TH |
3709 | unsigned long flags; |
3710 | unsigned int ret = 0; | |
1da177e4 | 3711 | |
dcd989cb TH |
3712 | if (work_pending(work)) |
3713 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 3714 | |
fa1b54e6 TH |
3715 | local_irq_save(flags); |
3716 | pool = get_work_pool(work); | |
038366c5 | 3717 | if (pool) { |
fa1b54e6 | 3718 | spin_lock(&pool->lock); |
038366c5 LJ |
3719 | if (find_worker_executing_work(pool, work)) |
3720 | ret |= WORK_BUSY_RUNNING; | |
fa1b54e6 | 3721 | spin_unlock(&pool->lock); |
038366c5 | 3722 | } |
fa1b54e6 | 3723 | local_irq_restore(flags); |
1da177e4 | 3724 | |
dcd989cb | 3725 | return ret; |
1da177e4 | 3726 | } |
dcd989cb | 3727 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 3728 | |
db7bccf4 TH |
3729 | /* |
3730 | * CPU hotplug. | |
3731 | * | |
e22bee78 | 3732 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 3733 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 3734 | * pool which make migrating pending and scheduled works very |
e22bee78 | 3735 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 3736 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
3737 | * blocked draining impractical. |
3738 | * | |
24647570 | 3739 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
3740 | * running as an unbound one and allowing it to be reattached later if the |
3741 | * cpu comes back online. | |
db7bccf4 | 3742 | */ |
1da177e4 | 3743 | |
706026c2 | 3744 | static void wq_unbind_fn(struct work_struct *work) |
3af24433 | 3745 | { |
38db41d9 | 3746 | int cpu = smp_processor_id(); |
4ce62e9e | 3747 | struct worker_pool *pool; |
db7bccf4 | 3748 | struct worker *worker; |
db7bccf4 | 3749 | int i; |
3af24433 | 3750 | |
38db41d9 | 3751 | for_each_std_worker_pool(pool, cpu) { |
6183c009 | 3752 | WARN_ON_ONCE(cpu != smp_processor_id()); |
db7bccf4 | 3753 | |
94cf58bb TH |
3754 | mutex_lock(&pool->assoc_mutex); |
3755 | spin_lock_irq(&pool->lock); | |
3af24433 | 3756 | |
94cf58bb TH |
3757 | /* |
3758 | * We've claimed all manager positions. Make all workers | |
3759 | * unbound and set DISASSOCIATED. Before this, all workers | |
3760 | * except for the ones which are still executing works from | |
3761 | * before the last CPU down must be on the cpu. After | |
3762 | * this, they may become diasporas. | |
3763 | */ | |
4ce62e9e | 3764 | list_for_each_entry(worker, &pool->idle_list, entry) |
403c821d | 3765 | worker->flags |= WORKER_UNBOUND; |
3af24433 | 3766 | |
b67bfe0d | 3767 | for_each_busy_worker(worker, i, pool) |
c9e7cf27 | 3768 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 3769 | |
24647570 | 3770 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 3771 | |
94cf58bb TH |
3772 | spin_unlock_irq(&pool->lock); |
3773 | mutex_unlock(&pool->assoc_mutex); | |
3774 | } | |
628c78e7 | 3775 | |
e22bee78 | 3776 | /* |
403c821d | 3777 | * Call schedule() so that we cross rq->lock and thus can guarantee |
628c78e7 TH |
3778 | * sched callbacks see the %WORKER_UNBOUND flag. This is necessary |
3779 | * as scheduler callbacks may be invoked from other cpus. | |
e22bee78 | 3780 | */ |
e22bee78 | 3781 | schedule(); |
06ba38a9 | 3782 | |
e22bee78 | 3783 | /* |
628c78e7 TH |
3784 | * Sched callbacks are disabled now. Zap nr_running. After this, |
3785 | * nr_running stays zero and need_more_worker() and keep_working() | |
38db41d9 TH |
3786 | * are always true as long as the worklist is not empty. Pools on |
3787 | * @cpu now behave as unbound (in terms of concurrency management) | |
3788 | * pools which are served by workers tied to the CPU. | |
628c78e7 TH |
3789 | * |
3790 | * On return from this function, the current worker would trigger | |
3791 | * unbound chain execution of pending work items if other workers | |
3792 | * didn't already. | |
e22bee78 | 3793 | */ |
38db41d9 | 3794 | for_each_std_worker_pool(pool, cpu) |
e19e397a | 3795 | atomic_set(&pool->nr_running, 0); |
3af24433 | 3796 | } |
3af24433 | 3797 | |
8db25e78 TH |
3798 | /* |
3799 | * Workqueues should be brought up before normal priority CPU notifiers. | |
3800 | * This will be registered high priority CPU notifier. | |
3801 | */ | |
9fdf9b73 | 3802 | static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, |
8db25e78 TH |
3803 | unsigned long action, |
3804 | void *hcpu) | |
3af24433 | 3805 | { |
d84ff051 | 3806 | int cpu = (unsigned long)hcpu; |
4ce62e9e | 3807 | struct worker_pool *pool; |
3ce63377 | 3808 | |
8db25e78 | 3809 | switch (action & ~CPU_TASKS_FROZEN) { |
3af24433 | 3810 | case CPU_UP_PREPARE: |
38db41d9 | 3811 | for_each_std_worker_pool(pool, cpu) { |
3ce63377 TH |
3812 | struct worker *worker; |
3813 | ||
3814 | if (pool->nr_workers) | |
3815 | continue; | |
3816 | ||
3817 | worker = create_worker(pool); | |
3818 | if (!worker) | |
3819 | return NOTIFY_BAD; | |
3820 | ||
d565ed63 | 3821 | spin_lock_irq(&pool->lock); |
3ce63377 | 3822 | start_worker(worker); |
d565ed63 | 3823 | spin_unlock_irq(&pool->lock); |
3af24433 | 3824 | } |
8db25e78 | 3825 | break; |
3af24433 | 3826 | |
db7bccf4 TH |
3827 | case CPU_DOWN_FAILED: |
3828 | case CPU_ONLINE: | |
38db41d9 | 3829 | for_each_std_worker_pool(pool, cpu) { |
94cf58bb TH |
3830 | mutex_lock(&pool->assoc_mutex); |
3831 | spin_lock_irq(&pool->lock); | |
3832 | ||
24647570 | 3833 | pool->flags &= ~POOL_DISASSOCIATED; |
94cf58bb TH |
3834 | rebind_workers(pool); |
3835 | ||
3836 | spin_unlock_irq(&pool->lock); | |
3837 | mutex_unlock(&pool->assoc_mutex); | |
3838 | } | |
db7bccf4 | 3839 | break; |
00dfcaf7 | 3840 | } |
65758202 TH |
3841 | return NOTIFY_OK; |
3842 | } | |
3843 | ||
3844 | /* | |
3845 | * Workqueues should be brought down after normal priority CPU notifiers. | |
3846 | * This will be registered as low priority CPU notifier. | |
3847 | */ | |
9fdf9b73 | 3848 | static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, |
65758202 TH |
3849 | unsigned long action, |
3850 | void *hcpu) | |
3851 | { | |
d84ff051 | 3852 | int cpu = (unsigned long)hcpu; |
8db25e78 TH |
3853 | struct work_struct unbind_work; |
3854 | ||
65758202 TH |
3855 | switch (action & ~CPU_TASKS_FROZEN) { |
3856 | case CPU_DOWN_PREPARE: | |
8db25e78 | 3857 | /* unbinding should happen on the local CPU */ |
706026c2 | 3858 | INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); |
7635d2fd | 3859 | queue_work_on(cpu, system_highpri_wq, &unbind_work); |
8db25e78 TH |
3860 | flush_work(&unbind_work); |
3861 | break; | |
65758202 TH |
3862 | } |
3863 | return NOTIFY_OK; | |
3864 | } | |
3865 | ||
2d3854a3 | 3866 | #ifdef CONFIG_SMP |
8ccad40d | 3867 | |
2d3854a3 | 3868 | struct work_for_cpu { |
ed48ece2 | 3869 | struct work_struct work; |
2d3854a3 RR |
3870 | long (*fn)(void *); |
3871 | void *arg; | |
3872 | long ret; | |
3873 | }; | |
3874 | ||
ed48ece2 | 3875 | static void work_for_cpu_fn(struct work_struct *work) |
2d3854a3 | 3876 | { |
ed48ece2 TH |
3877 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); |
3878 | ||
2d3854a3 RR |
3879 | wfc->ret = wfc->fn(wfc->arg); |
3880 | } | |
3881 | ||
3882 | /** | |
3883 | * work_on_cpu - run a function in user context on a particular cpu | |
3884 | * @cpu: the cpu to run on | |
3885 | * @fn: the function to run | |
3886 | * @arg: the function arg | |
3887 | * | |
31ad9081 RR |
3888 | * This will return the value @fn returns. |
3889 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
6b44003e | 3890 | * The caller must not hold any locks which would prevent @fn from completing. |
2d3854a3 | 3891 | */ |
d84ff051 | 3892 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) |
2d3854a3 | 3893 | { |
ed48ece2 | 3894 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; |
6b44003e | 3895 | |
ed48ece2 TH |
3896 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); |
3897 | schedule_work_on(cpu, &wfc.work); | |
3898 | flush_work(&wfc.work); | |
2d3854a3 RR |
3899 | return wfc.ret; |
3900 | } | |
3901 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
3902 | #endif /* CONFIG_SMP */ | |
3903 | ||
a0a1a5fd TH |
3904 | #ifdef CONFIG_FREEZER |
3905 | ||
3906 | /** | |
3907 | * freeze_workqueues_begin - begin freezing workqueues | |
3908 | * | |
58a69cb4 TH |
3909 | * Start freezing workqueues. After this function returns, all freezable |
3910 | * workqueues will queue new works to their frozen_works list instead of | |
706026c2 | 3911 | * pool->worklist. |
a0a1a5fd TH |
3912 | * |
3913 | * CONTEXT: | |
d565ed63 | 3914 | * Grabs and releases workqueue_lock and pool->lock's. |
a0a1a5fd TH |
3915 | */ |
3916 | void freeze_workqueues_begin(void) | |
3917 | { | |
17116969 | 3918 | struct worker_pool *pool; |
24b8a847 TH |
3919 | struct workqueue_struct *wq; |
3920 | struct pool_workqueue *pwq; | |
17116969 | 3921 | int id; |
a0a1a5fd | 3922 | |
e98d5b16 | 3923 | spin_lock_irq(&workqueue_lock); |
a0a1a5fd | 3924 | |
6183c009 | 3925 | WARN_ON_ONCE(workqueue_freezing); |
a0a1a5fd TH |
3926 | workqueue_freezing = true; |
3927 | ||
24b8a847 | 3928 | /* set FREEZING */ |
17116969 | 3929 | for_each_pool(pool, id) { |
17116969 | 3930 | spin_lock(&pool->lock); |
17116969 TH |
3931 | WARN_ON_ONCE(pool->flags & POOL_FREEZING); |
3932 | pool->flags |= POOL_FREEZING; | |
24b8a847 TH |
3933 | spin_unlock(&pool->lock); |
3934 | } | |
a0a1a5fd | 3935 | |
24b8a847 TH |
3936 | /* suppress further executions by setting max_active to zero */ |
3937 | list_for_each_entry(wq, &workqueues, list) { | |
3938 | if (!(wq->flags & WQ_FREEZABLE)) | |
3939 | continue; | |
8b03ae3c | 3940 | |
24b8a847 TH |
3941 | for_each_pwq(pwq, wq) { |
3942 | spin_lock(&pwq->pool->lock); | |
3943 | pwq->max_active = 0; | |
3944 | spin_unlock(&pwq->pool->lock); | |
a1056305 | 3945 | } |
a0a1a5fd TH |
3946 | } |
3947 | ||
e98d5b16 | 3948 | spin_unlock_irq(&workqueue_lock); |
a0a1a5fd TH |
3949 | } |
3950 | ||
3951 | /** | |
58a69cb4 | 3952 | * freeze_workqueues_busy - are freezable workqueues still busy? |
a0a1a5fd TH |
3953 | * |
3954 | * Check whether freezing is complete. This function must be called | |
3955 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
3956 | * | |
3957 | * CONTEXT: | |
3958 | * Grabs and releases workqueue_lock. | |
3959 | * | |
3960 | * RETURNS: | |
58a69cb4 TH |
3961 | * %true if some freezable workqueues are still busy. %false if freezing |
3962 | * is complete. | |
a0a1a5fd TH |
3963 | */ |
3964 | bool freeze_workqueues_busy(void) | |
3965 | { | |
a0a1a5fd | 3966 | bool busy = false; |
24b8a847 TH |
3967 | struct workqueue_struct *wq; |
3968 | struct pool_workqueue *pwq; | |
a0a1a5fd | 3969 | |
e98d5b16 | 3970 | spin_lock_irq(&workqueue_lock); |
a0a1a5fd | 3971 | |
6183c009 | 3972 | WARN_ON_ONCE(!workqueue_freezing); |
a0a1a5fd | 3973 | |
24b8a847 TH |
3974 | list_for_each_entry(wq, &workqueues, list) { |
3975 | if (!(wq->flags & WQ_FREEZABLE)) | |
3976 | continue; | |
a0a1a5fd TH |
3977 | /* |
3978 | * nr_active is monotonically decreasing. It's safe | |
3979 | * to peek without lock. | |
3980 | */ | |
24b8a847 | 3981 | for_each_pwq(pwq, wq) { |
6183c009 | 3982 | WARN_ON_ONCE(pwq->nr_active < 0); |
112202d9 | 3983 | if (pwq->nr_active) { |
a0a1a5fd TH |
3984 | busy = true; |
3985 | goto out_unlock; | |
3986 | } | |
3987 | } | |
3988 | } | |
3989 | out_unlock: | |
e98d5b16 | 3990 | spin_unlock_irq(&workqueue_lock); |
a0a1a5fd TH |
3991 | return busy; |
3992 | } | |
3993 | ||
3994 | /** | |
3995 | * thaw_workqueues - thaw workqueues | |
3996 | * | |
3997 | * Thaw workqueues. Normal queueing is restored and all collected | |
706026c2 | 3998 | * frozen works are transferred to their respective pool worklists. |
a0a1a5fd TH |
3999 | * |
4000 | * CONTEXT: | |
d565ed63 | 4001 | * Grabs and releases workqueue_lock and pool->lock's. |
a0a1a5fd TH |
4002 | */ |
4003 | void thaw_workqueues(void) | |
4004 | { | |
24b8a847 TH |
4005 | struct workqueue_struct *wq; |
4006 | struct pool_workqueue *pwq; | |
4007 | struct worker_pool *pool; | |
4008 | int id; | |
a0a1a5fd | 4009 | |
e98d5b16 | 4010 | spin_lock_irq(&workqueue_lock); |
a0a1a5fd TH |
4011 | |
4012 | if (!workqueue_freezing) | |
4013 | goto out_unlock; | |
4014 | ||
24b8a847 TH |
4015 | /* clear FREEZING */ |
4016 | for_each_pool(pool, id) { | |
4017 | spin_lock(&pool->lock); | |
4018 | WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); | |
4019 | pool->flags &= ~POOL_FREEZING; | |
4020 | spin_unlock(&pool->lock); | |
4021 | } | |
8b03ae3c | 4022 | |
24b8a847 TH |
4023 | /* restore max_active and repopulate worklist */ |
4024 | list_for_each_entry(wq, &workqueues, list) { | |
4025 | if (!(wq->flags & WQ_FREEZABLE)) | |
4026 | continue; | |
a1056305 | 4027 | |
24b8a847 TH |
4028 | for_each_pwq(pwq, wq) { |
4029 | spin_lock(&pwq->pool->lock); | |
4030 | pwq_set_max_active(pwq, wq->saved_max_active); | |
4031 | spin_unlock(&pwq->pool->lock); | |
d565ed63 | 4032 | } |
a0a1a5fd TH |
4033 | } |
4034 | ||
24b8a847 TH |
4035 | /* kick workers */ |
4036 | for_each_pool(pool, id) { | |
4037 | spin_lock(&pool->lock); | |
4038 | wake_up_worker(pool); | |
4039 | spin_unlock(&pool->lock); | |
4040 | } | |
4041 | ||
a0a1a5fd TH |
4042 | workqueue_freezing = false; |
4043 | out_unlock: | |
e98d5b16 | 4044 | spin_unlock_irq(&workqueue_lock); |
a0a1a5fd TH |
4045 | } |
4046 | #endif /* CONFIG_FREEZER */ | |
4047 | ||
6ee0578b | 4048 | static int __init init_workqueues(void) |
1da177e4 | 4049 | { |
7a4e344c TH |
4050 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
4051 | int i, cpu; | |
c34056a3 | 4052 | |
7c3eed5c TH |
4053 | /* make sure we have enough bits for OFFQ pool ID */ |
4054 | BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < | |
6be19588 | 4055 | WORK_CPU_END * NR_STD_WORKER_POOLS); |
b5490077 | 4056 | |
e904e6c2 TH |
4057 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
4058 | ||
4059 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); | |
4060 | ||
65758202 | 4061 | cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); |
a5b4e57d | 4062 | hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); |
8b03ae3c | 4063 | |
706026c2 | 4064 | /* initialize CPU pools */ |
29c91e99 | 4065 | for_each_possible_cpu(cpu) { |
4ce62e9e | 4066 | struct worker_pool *pool; |
8b03ae3c | 4067 | |
7a4e344c | 4068 | i = 0; |
38db41d9 | 4069 | for_each_std_worker_pool(pool, cpu) { |
7a4e344c | 4070 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 4071 | pool->cpu = cpu; |
29c91e99 | 4072 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c TH |
4073 | pool->attrs->nice = std_nice[i++]; |
4074 | ||
9daf9e67 TH |
4075 | /* alloc pool ID */ |
4076 | BUG_ON(worker_pool_assign_id(pool)); | |
4ce62e9e | 4077 | } |
8b03ae3c TH |
4078 | } |
4079 | ||
e22bee78 | 4080 | /* create the initial worker */ |
29c91e99 | 4081 | for_each_online_cpu(cpu) { |
4ce62e9e | 4082 | struct worker_pool *pool; |
e22bee78 | 4083 | |
38db41d9 | 4084 | for_each_std_worker_pool(pool, cpu) { |
4ce62e9e TH |
4085 | struct worker *worker; |
4086 | ||
29c91e99 | 4087 | pool->flags &= ~POOL_DISASSOCIATED; |
24647570 | 4088 | |
bc2ae0f5 | 4089 | worker = create_worker(pool); |
4ce62e9e | 4090 | BUG_ON(!worker); |
d565ed63 | 4091 | spin_lock_irq(&pool->lock); |
4ce62e9e | 4092 | start_worker(worker); |
d565ed63 | 4093 | spin_unlock_irq(&pool->lock); |
4ce62e9e | 4094 | } |
e22bee78 TH |
4095 | } |
4096 | ||
29c91e99 TH |
4097 | /* create default unbound wq attrs */ |
4098 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { | |
4099 | struct workqueue_attrs *attrs; | |
4100 | ||
4101 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
4102 | ||
4103 | attrs->nice = std_nice[i]; | |
4104 | cpumask_setall(attrs->cpumask); | |
4105 | ||
4106 | unbound_std_wq_attrs[i] = attrs; | |
4107 | } | |
4108 | ||
d320c038 | 4109 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 4110 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 4111 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
4112 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
4113 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
4114 | system_freezable_wq = alloc_workqueue("events_freezable", |
4115 | WQ_FREEZABLE, 0); | |
1aabe902 | 4116 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
ae930e0f | 4117 | !system_unbound_wq || !system_freezable_wq); |
6ee0578b | 4118 | return 0; |
1da177e4 | 4119 | } |
6ee0578b | 4120 | early_initcall(init_workqueues); |