]>
Commit | Line | Data |
---|---|---|
771b53d0 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Basic worker thread pool for io_uring | |
4 | * | |
5 | * Copyright (C) 2019 Jens Axboe | |
6 | * | |
7 | */ | |
8 | #include <linux/kernel.h> | |
9 | #include <linux/init.h> | |
10 | #include <linux/errno.h> | |
11 | #include <linux/sched/signal.h> | |
12 | #include <linux/mm.h> | |
13 | #include <linux/mmu_context.h> | |
14 | #include <linux/sched/mm.h> | |
15 | #include <linux/percpu.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/kthread.h> | |
18 | #include <linux/rculist_nulls.h> | |
9392a27d | 19 | #include <linux/fs_struct.h> |
771b53d0 JA |
20 | |
21 | #include "io-wq.h" | |
22 | ||
23 | #define WORKER_IDLE_TIMEOUT (5 * HZ) | |
24 | ||
25 | enum { | |
26 | IO_WORKER_F_UP = 1, /* up and active */ | |
27 | IO_WORKER_F_RUNNING = 2, /* account as running */ | |
28 | IO_WORKER_F_FREE = 4, /* worker on free list */ | |
29 | IO_WORKER_F_EXITING = 8, /* worker exiting */ | |
30 | IO_WORKER_F_FIXED = 16, /* static idle worker */ | |
c5def4ab | 31 | IO_WORKER_F_BOUND = 32, /* is doing bounded work */ |
771b53d0 JA |
32 | }; |
33 | ||
34 | enum { | |
35 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ | |
36 | IO_WQ_BIT_CANCEL = 1, /* cancel work on list */ | |
b60fda60 | 37 | IO_WQ_BIT_ERROR = 2, /* error on setup */ |
771b53d0 JA |
38 | }; |
39 | ||
40 | enum { | |
41 | IO_WQE_FLAG_STALLED = 1, /* stalled on hash */ | |
42 | }; | |
43 | ||
44 | /* | |
45 | * One for each thread in a wqe pool | |
46 | */ | |
47 | struct io_worker { | |
48 | refcount_t ref; | |
49 | unsigned flags; | |
50 | struct hlist_nulls_node nulls_node; | |
e61df66c | 51 | struct list_head all_list; |
771b53d0 | 52 | struct task_struct *task; |
771b53d0 | 53 | struct io_wqe *wqe; |
36c2f922 | 54 | |
771b53d0 | 55 | struct io_wq_work *cur_work; |
36c2f922 | 56 | spinlock_t lock; |
771b53d0 JA |
57 | |
58 | struct rcu_head rcu; | |
59 | struct mm_struct *mm; | |
cccf0ee8 JA |
60 | const struct cred *cur_creds; |
61 | const struct cred *saved_creds; | |
fcb323cc | 62 | struct files_struct *restore_files; |
9392a27d | 63 | struct fs_struct *restore_fs; |
771b53d0 JA |
64 | }; |
65 | ||
771b53d0 JA |
66 | #if BITS_PER_LONG == 64 |
67 | #define IO_WQ_HASH_ORDER 6 | |
68 | #else | |
69 | #define IO_WQ_HASH_ORDER 5 | |
70 | #endif | |
71 | ||
c5def4ab JA |
72 | struct io_wqe_acct { |
73 | unsigned nr_workers; | |
74 | unsigned max_workers; | |
75 | atomic_t nr_running; | |
76 | }; | |
77 | ||
78 | enum { | |
79 | IO_WQ_ACCT_BOUND, | |
80 | IO_WQ_ACCT_UNBOUND, | |
81 | }; | |
82 | ||
771b53d0 JA |
83 | /* |
84 | * Per-node worker thread pool | |
85 | */ | |
86 | struct io_wqe { | |
87 | struct { | |
88 | spinlock_t lock; | |
6206f0e1 | 89 | struct io_wq_work_list work_list; |
771b53d0 JA |
90 | unsigned long hash_map; |
91 | unsigned flags; | |
92 | } ____cacheline_aligned_in_smp; | |
93 | ||
94 | int node; | |
c5def4ab | 95 | struct io_wqe_acct acct[2]; |
771b53d0 | 96 | |
021d1cdd | 97 | struct hlist_nulls_head free_list; |
e61df66c | 98 | struct list_head all_list; |
771b53d0 JA |
99 | |
100 | struct io_wq *wq; | |
101 | }; | |
102 | ||
103 | /* | |
104 | * Per io_wq state | |
105 | */ | |
106 | struct io_wq { | |
107 | struct io_wqe **wqes; | |
108 | unsigned long state; | |
771b53d0 | 109 | |
e9fd9396 | 110 | free_work_fn *free_work; |
7d723065 | 111 | |
771b53d0 | 112 | struct task_struct *manager; |
c5def4ab | 113 | struct user_struct *user; |
771b53d0 JA |
114 | refcount_t refs; |
115 | struct completion done; | |
848f7e18 JA |
116 | |
117 | refcount_t use_refs; | |
771b53d0 JA |
118 | }; |
119 | ||
771b53d0 JA |
120 | static bool io_worker_get(struct io_worker *worker) |
121 | { | |
122 | return refcount_inc_not_zero(&worker->ref); | |
123 | } | |
124 | ||
125 | static void io_worker_release(struct io_worker *worker) | |
126 | { | |
127 | if (refcount_dec_and_test(&worker->ref)) | |
128 | wake_up_process(worker->task); | |
129 | } | |
130 | ||
131 | /* | |
132 | * Note: drops the wqe->lock if returning true! The caller must re-acquire | |
133 | * the lock in that case. Some callers need to restart handling if this | |
134 | * happens, so we can't just re-acquire the lock on behalf of the caller. | |
135 | */ | |
136 | static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) | |
137 | { | |
fcb323cc JA |
138 | bool dropped_lock = false; |
139 | ||
cccf0ee8 JA |
140 | if (worker->saved_creds) { |
141 | revert_creds(worker->saved_creds); | |
142 | worker->cur_creds = worker->saved_creds = NULL; | |
181e448d JA |
143 | } |
144 | ||
fcb323cc JA |
145 | if (current->files != worker->restore_files) { |
146 | __acquire(&wqe->lock); | |
147 | spin_unlock_irq(&wqe->lock); | |
148 | dropped_lock = true; | |
149 | ||
150 | task_lock(current); | |
151 | current->files = worker->restore_files; | |
152 | task_unlock(current); | |
153 | } | |
154 | ||
9392a27d JA |
155 | if (current->fs != worker->restore_fs) |
156 | current->fs = worker->restore_fs; | |
157 | ||
771b53d0 JA |
158 | /* |
159 | * If we have an active mm, we need to drop the wq lock before unusing | |
160 | * it. If we do, return true and let the caller retry the idle loop. | |
161 | */ | |
162 | if (worker->mm) { | |
fcb323cc JA |
163 | if (!dropped_lock) { |
164 | __acquire(&wqe->lock); | |
165 | spin_unlock_irq(&wqe->lock); | |
166 | dropped_lock = true; | |
167 | } | |
771b53d0 JA |
168 | __set_current_state(TASK_RUNNING); |
169 | set_fs(KERNEL_DS); | |
170 | unuse_mm(worker->mm); | |
171 | mmput(worker->mm); | |
172 | worker->mm = NULL; | |
771b53d0 JA |
173 | } |
174 | ||
fcb323cc | 175 | return dropped_lock; |
771b53d0 JA |
176 | } |
177 | ||
c5def4ab JA |
178 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
179 | struct io_wq_work *work) | |
180 | { | |
181 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
182 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
183 | ||
184 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
185 | } | |
186 | ||
187 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe, | |
188 | struct io_worker *worker) | |
189 | { | |
190 | if (worker->flags & IO_WORKER_F_BOUND) | |
191 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
192 | ||
193 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
194 | } | |
195 | ||
771b53d0 JA |
196 | static void io_worker_exit(struct io_worker *worker) |
197 | { | |
198 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab JA |
199 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); |
200 | unsigned nr_workers; | |
771b53d0 JA |
201 | |
202 | /* | |
203 | * If we're not at zero, someone else is holding a brief reference | |
204 | * to the worker. Wait for that to go away. | |
205 | */ | |
206 | set_current_state(TASK_INTERRUPTIBLE); | |
207 | if (!refcount_dec_and_test(&worker->ref)) | |
208 | schedule(); | |
209 | __set_current_state(TASK_RUNNING); | |
210 | ||
211 | preempt_disable(); | |
212 | current->flags &= ~PF_IO_WORKER; | |
213 | if (worker->flags & IO_WORKER_F_RUNNING) | |
c5def4ab JA |
214 | atomic_dec(&acct->nr_running); |
215 | if (!(worker->flags & IO_WORKER_F_BOUND)) | |
216 | atomic_dec(&wqe->wq->user->processes); | |
771b53d0 JA |
217 | worker->flags = 0; |
218 | preempt_enable(); | |
219 | ||
220 | spin_lock_irq(&wqe->lock); | |
221 | hlist_nulls_del_rcu(&worker->nulls_node); | |
e61df66c | 222 | list_del_rcu(&worker->all_list); |
771b53d0 JA |
223 | if (__io_worker_unuse(wqe, worker)) { |
224 | __release(&wqe->lock); | |
225 | spin_lock_irq(&wqe->lock); | |
226 | } | |
c5def4ab JA |
227 | acct->nr_workers--; |
228 | nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers + | |
229 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers; | |
771b53d0 JA |
230 | spin_unlock_irq(&wqe->lock); |
231 | ||
232 | /* all workers gone, wq exit can proceed */ | |
c5def4ab | 233 | if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) |
771b53d0 JA |
234 | complete(&wqe->wq->done); |
235 | ||
364b05fd | 236 | kfree_rcu(worker, rcu); |
771b53d0 JA |
237 | } |
238 | ||
c5def4ab JA |
239 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
240 | __must_hold(wqe->lock) | |
241 | { | |
6206f0e1 JA |
242 | if (!wq_list_empty(&wqe->work_list) && |
243 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
244 | return true; |
245 | return false; | |
246 | } | |
247 | ||
248 | /* | |
249 | * Check head of free list for an available worker. If one isn't available, | |
250 | * caller must wake up the wq manager to create one. | |
251 | */ | |
252 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
253 | __must_hold(RCU) | |
254 | { | |
255 | struct hlist_nulls_node *n; | |
256 | struct io_worker *worker; | |
257 | ||
021d1cdd | 258 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
259 | if (is_a_nulls(n)) |
260 | return false; | |
261 | ||
262 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
263 | if (io_worker_get(worker)) { | |
506d95ff | 264 | wake_up_process(worker->task); |
c5def4ab JA |
265 | io_worker_release(worker); |
266 | return true; | |
267 | } | |
268 | ||
269 | return false; | |
270 | } | |
271 | ||
272 | /* | |
273 | * We need a worker. If we find a free one, we're good. If not, and we're | |
274 | * below the max number of workers, wake up the manager to create one. | |
275 | */ | |
276 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
277 | { | |
278 | bool ret; | |
279 | ||
280 | /* | |
281 | * Most likely an attempt to queue unbounded work on an io_wq that | |
282 | * wasn't setup with any unbounded workers. | |
283 | */ | |
284 | WARN_ON_ONCE(!acct->max_workers); | |
285 | ||
286 | rcu_read_lock(); | |
287 | ret = io_wqe_activate_free_worker(wqe); | |
288 | rcu_read_unlock(); | |
289 | ||
290 | if (!ret && acct->nr_workers < acct->max_workers) | |
291 | wake_up_process(wqe->wq->manager); | |
292 | } | |
293 | ||
294 | static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker) | |
295 | { | |
296 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
297 | ||
298 | atomic_inc(&acct->nr_running); | |
299 | } | |
300 | ||
301 | static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker) | |
302 | __must_hold(wqe->lock) | |
303 | { | |
304 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
305 | ||
306 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
307 | io_wqe_wake_worker(wqe, acct); | |
308 | } | |
309 | ||
771b53d0 JA |
310 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
311 | { | |
312 | allow_kernel_signal(SIGINT); | |
313 | ||
314 | current->flags |= PF_IO_WORKER; | |
315 | ||
316 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); | |
fcb323cc | 317 | worker->restore_files = current->files; |
9392a27d | 318 | worker->restore_fs = current->fs; |
c5def4ab | 319 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
320 | } |
321 | ||
322 | /* | |
323 | * Worker will start processing some work. Move it to the busy list, if | |
324 | * it's currently on the freelist | |
325 | */ | |
326 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
327 | struct io_wq_work *work) | |
328 | __must_hold(wqe->lock) | |
329 | { | |
c5def4ab JA |
330 | bool worker_bound, work_bound; |
331 | ||
771b53d0 JA |
332 | if (worker->flags & IO_WORKER_F_FREE) { |
333 | worker->flags &= ~IO_WORKER_F_FREE; | |
334 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 335 | } |
c5def4ab JA |
336 | |
337 | /* | |
338 | * If worker is moving from bound to unbound (or vice versa), then | |
339 | * ensure we update the running accounting. | |
340 | */ | |
b2e9c7d6 DC |
341 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
342 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
343 | if (worker_bound != work_bound) { | |
c5def4ab JA |
344 | io_wqe_dec_running(wqe, worker); |
345 | if (work_bound) { | |
346 | worker->flags |= IO_WORKER_F_BOUND; | |
347 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
348 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
349 | atomic_dec(&wqe->wq->user->processes); | |
350 | } else { | |
351 | worker->flags &= ~IO_WORKER_F_BOUND; | |
352 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
353 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
354 | atomic_inc(&wqe->wq->user->processes); | |
355 | } | |
356 | io_wqe_inc_running(wqe, worker); | |
357 | } | |
771b53d0 JA |
358 | } |
359 | ||
360 | /* | |
361 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
362 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
363 | * the lock in that case and return success. Since the caller has to | |
364 | * retry the loop in that case (we changed task state), we don't regrab | |
365 | * the lock if we return success. | |
366 | */ | |
367 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) | |
368 | __must_hold(wqe->lock) | |
369 | { | |
370 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
371 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 372 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 JA |
373 | } |
374 | ||
375 | return __io_worker_unuse(wqe, worker); | |
376 | } | |
377 | ||
60cf46ae PB |
378 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
379 | { | |
380 | return work->flags >> IO_WQ_HASH_SHIFT; | |
381 | } | |
382 | ||
383 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe) | |
771b53d0 JA |
384 | __must_hold(wqe->lock) |
385 | { | |
6206f0e1 | 386 | struct io_wq_work_node *node, *prev; |
771b53d0 | 387 | struct io_wq_work *work; |
60cf46ae | 388 | unsigned int hash; |
771b53d0 | 389 | |
6206f0e1 JA |
390 | wq_list_for_each(node, prev, &wqe->work_list) { |
391 | work = container_of(node, struct io_wq_work, list); | |
392 | ||
771b53d0 | 393 | /* not hashed, can run anytime */ |
8766dd51 | 394 | if (!io_wq_is_hashed(work)) { |
6206f0e1 | 395 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
396 | return work; |
397 | } | |
398 | ||
399 | /* hashed, can run if not already running */ | |
60cf46ae PB |
400 | hash = io_get_work_hash(work); |
401 | if (!(wqe->hash_map & BIT(hash))) { | |
402 | wqe->hash_map |= BIT(hash); | |
6206f0e1 | 403 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
404 | return work; |
405 | } | |
406 | } | |
407 | ||
408 | return NULL; | |
409 | } | |
410 | ||
cccf0ee8 JA |
411 | static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work) |
412 | { | |
413 | if (worker->mm) { | |
414 | unuse_mm(worker->mm); | |
415 | mmput(worker->mm); | |
416 | worker->mm = NULL; | |
417 | } | |
418 | if (!work->mm) { | |
419 | set_fs(KERNEL_DS); | |
420 | return; | |
421 | } | |
422 | if (mmget_not_zero(work->mm)) { | |
423 | use_mm(work->mm); | |
424 | if (!worker->mm) | |
425 | set_fs(USER_DS); | |
426 | worker->mm = work->mm; | |
427 | /* hang on to this mm */ | |
428 | work->mm = NULL; | |
429 | return; | |
430 | } | |
431 | ||
432 | /* failed grabbing mm, ensure work gets cancelled */ | |
433 | work->flags |= IO_WQ_WORK_CANCEL; | |
434 | } | |
435 | ||
436 | static void io_wq_switch_creds(struct io_worker *worker, | |
437 | struct io_wq_work *work) | |
438 | { | |
439 | const struct cred *old_creds = override_creds(work->creds); | |
440 | ||
441 | worker->cur_creds = work->creds; | |
442 | if (worker->saved_creds) | |
443 | put_cred(old_creds); /* creds set by previous switch */ | |
444 | else | |
445 | worker->saved_creds = old_creds; | |
446 | } | |
447 | ||
dc026a73 PB |
448 | static void io_impersonate_work(struct io_worker *worker, |
449 | struct io_wq_work *work) | |
450 | { | |
451 | if (work->files && current->files != work->files) { | |
452 | task_lock(current); | |
453 | current->files = work->files; | |
454 | task_unlock(current); | |
455 | } | |
456 | if (work->fs && current->fs != work->fs) | |
457 | current->fs = work->fs; | |
458 | if (work->mm != worker->mm) | |
459 | io_wq_switch_mm(worker, work); | |
460 | if (worker->cur_creds != work->creds) | |
461 | io_wq_switch_creds(worker, work); | |
462 | } | |
463 | ||
464 | static void io_assign_current_work(struct io_worker *worker, | |
465 | struct io_wq_work *work) | |
466 | { | |
d78298e7 PB |
467 | if (work) { |
468 | /* flush pending signals before assigning new work */ | |
469 | if (signal_pending(current)) | |
470 | flush_signals(current); | |
471 | cond_resched(); | |
472 | } | |
dc026a73 PB |
473 | |
474 | spin_lock_irq(&worker->lock); | |
475 | worker->cur_work = work; | |
476 | spin_unlock_irq(&worker->lock); | |
477 | } | |
478 | ||
60cf46ae PB |
479 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
480 | ||
771b53d0 JA |
481 | static void io_worker_handle_work(struct io_worker *worker) |
482 | __releases(wqe->lock) | |
483 | { | |
771b53d0 JA |
484 | struct io_wqe *wqe = worker->wqe; |
485 | struct io_wq *wq = wqe->wq; | |
486 | ||
487 | do { | |
f2cf1149 | 488 | struct io_wq_work *work, *assign_work; |
60cf46ae | 489 | unsigned int hash; |
f462fd36 | 490 | get_next: |
771b53d0 JA |
491 | /* |
492 | * If we got some work, mark us as busy. If we didn't, but | |
493 | * the list isn't empty, it means we stalled on hashed work. | |
494 | * Mark us stalled so we don't keep looking for work when we | |
495 | * can't make progress, any work completion or insertion will | |
496 | * clear the stalled flag. | |
497 | */ | |
60cf46ae | 498 | work = io_get_next_work(wqe); |
771b53d0 JA |
499 | if (work) |
500 | __io_worker_busy(wqe, worker, work); | |
6206f0e1 | 501 | else if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
502 | wqe->flags |= IO_WQE_FLAG_STALLED; |
503 | ||
504 | spin_unlock_irq(&wqe->lock); | |
505 | if (!work) | |
506 | break; | |
58e39319 | 507 | io_assign_current_work(worker, work); |
36c2f922 | 508 | |
dc026a73 PB |
509 | /* handle a whole dependent link */ |
510 | do { | |
58e39319 | 511 | struct io_wq_work *old_work; |
dc026a73 | 512 | |
58e39319 | 513 | io_impersonate_work(worker, work); |
dc026a73 PB |
514 | /* |
515 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable | |
516 | * work, the worker function will do the right thing. | |
517 | */ | |
518 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) | |
519 | work->flags |= IO_WQ_WORK_CANCEL; | |
520 | ||
dc026a73 | 521 | old_work = work; |
60cf46ae | 522 | hash = io_get_work_hash(work); |
dc026a73 | 523 | work->func(&work); |
58e39319 | 524 | work = (old_work == work) ? NULL : work; |
f2cf1149 PB |
525 | |
526 | assign_work = work; | |
527 | if (work && io_wq_is_hashed(work)) | |
528 | assign_work = NULL; | |
529 | io_assign_current_work(worker, assign_work); | |
e9fd9396 | 530 | wq->free_work(old_work); |
dc026a73 | 531 | |
f2cf1149 | 532 | if (work && !assign_work) { |
60cf46ae PB |
533 | io_wqe_enqueue(wqe, work); |
534 | work = NULL; | |
535 | } | |
dc026a73 PB |
536 | if (hash != -1U) { |
537 | spin_lock_irq(&wqe->lock); | |
538 | wqe->hash_map &= ~BIT_ULL(hash); | |
539 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
dc026a73 PB |
540 | /* dependent work is not hashed */ |
541 | hash = -1U; | |
f462fd36 PB |
542 | /* skip unnecessary unlock-lock wqe->lock */ |
543 | if (!work) | |
544 | goto get_next; | |
545 | spin_unlock_irq(&wqe->lock); | |
7d723065 | 546 | } |
58e39319 | 547 | } while (work); |
7d723065 | 548 | |
dc026a73 | 549 | spin_lock_irq(&wqe->lock); |
771b53d0 JA |
550 | } while (1); |
551 | } | |
552 | ||
771b53d0 JA |
553 | static int io_wqe_worker(void *data) |
554 | { | |
555 | struct io_worker *worker = data; | |
556 | struct io_wqe *wqe = worker->wqe; | |
557 | struct io_wq *wq = wqe->wq; | |
771b53d0 JA |
558 | |
559 | io_worker_start(wqe, worker); | |
560 | ||
561 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
506d95ff | 562 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 563 | loop: |
771b53d0 JA |
564 | spin_lock_irq(&wqe->lock); |
565 | if (io_wqe_run_queue(wqe)) { | |
566 | __set_current_state(TASK_RUNNING); | |
567 | io_worker_handle_work(worker); | |
e995d512 | 568 | goto loop; |
771b53d0 JA |
569 | } |
570 | /* drops the lock on success, retry */ | |
571 | if (__io_worker_idle(wqe, worker)) { | |
572 | __release(&wqe->lock); | |
e995d512 | 573 | goto loop; |
771b53d0 JA |
574 | } |
575 | spin_unlock_irq(&wqe->lock); | |
576 | if (signal_pending(current)) | |
577 | flush_signals(current); | |
578 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
579 | continue; | |
580 | /* timed out, exit unless we're the fixed worker */ | |
581 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
582 | !(worker->flags & IO_WORKER_F_FIXED)) | |
583 | break; | |
584 | } | |
585 | ||
771b53d0 JA |
586 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
587 | spin_lock_irq(&wqe->lock); | |
6206f0e1 | 588 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
589 | io_worker_handle_work(worker); |
590 | else | |
591 | spin_unlock_irq(&wqe->lock); | |
592 | } | |
593 | ||
594 | io_worker_exit(worker); | |
595 | return 0; | |
596 | } | |
597 | ||
771b53d0 JA |
598 | /* |
599 | * Called when a worker is scheduled in. Mark us as currently running. | |
600 | */ | |
601 | void io_wq_worker_running(struct task_struct *tsk) | |
602 | { | |
603 | struct io_worker *worker = kthread_data(tsk); | |
604 | struct io_wqe *wqe = worker->wqe; | |
605 | ||
606 | if (!(worker->flags & IO_WORKER_F_UP)) | |
607 | return; | |
608 | if (worker->flags & IO_WORKER_F_RUNNING) | |
609 | return; | |
610 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 611 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
612 | } |
613 | ||
614 | /* | |
615 | * Called when worker is going to sleep. If there are no workers currently | |
616 | * running and we have work pending, wake up a free one or have the manager | |
617 | * set one up. | |
618 | */ | |
619 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
620 | { | |
621 | struct io_worker *worker = kthread_data(tsk); | |
622 | struct io_wqe *wqe = worker->wqe; | |
623 | ||
624 | if (!(worker->flags & IO_WORKER_F_UP)) | |
625 | return; | |
626 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
627 | return; | |
628 | ||
629 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
630 | ||
631 | spin_lock_irq(&wqe->lock); | |
c5def4ab | 632 | io_wqe_dec_running(wqe, worker); |
771b53d0 JA |
633 | spin_unlock_irq(&wqe->lock); |
634 | } | |
635 | ||
b60fda60 | 636 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 637 | { |
c5def4ab | 638 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
639 | struct io_worker *worker; |
640 | ||
ad6e005c | 641 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 642 | if (!worker) |
b60fda60 | 643 | return false; |
771b53d0 JA |
644 | |
645 | refcount_set(&worker->ref, 1); | |
646 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 647 | worker->wqe = wqe; |
36c2f922 | 648 | spin_lock_init(&worker->lock); |
771b53d0 JA |
649 | |
650 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 651 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
652 | if (IS_ERR(worker->task)) { |
653 | kfree(worker); | |
b60fda60 | 654 | return false; |
771b53d0 JA |
655 | } |
656 | ||
657 | spin_lock_irq(&wqe->lock); | |
021d1cdd | 658 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 659 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 660 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
661 | if (index == IO_WQ_ACCT_BOUND) |
662 | worker->flags |= IO_WORKER_F_BOUND; | |
663 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 664 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 665 | acct->nr_workers++; |
771b53d0 JA |
666 | spin_unlock_irq(&wqe->lock); |
667 | ||
c5def4ab JA |
668 | if (index == IO_WQ_ACCT_UNBOUND) |
669 | atomic_inc(&wq->user->processes); | |
670 | ||
771b53d0 | 671 | wake_up_process(worker->task); |
b60fda60 | 672 | return true; |
771b53d0 JA |
673 | } |
674 | ||
c5def4ab | 675 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
676 | __must_hold(wqe->lock) |
677 | { | |
c5def4ab | 678 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 679 | |
c5def4ab | 680 | /* if we have available workers or no work, no need */ |
021d1cdd | 681 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
682 | return false; |
683 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
684 | } |
685 | ||
686 | /* | |
687 | * Manager thread. Tasked with creating new workers, if we need them. | |
688 | */ | |
689 | static int io_wq_manager(void *data) | |
690 | { | |
691 | struct io_wq *wq = data; | |
3fc50ab5 JH |
692 | int workers_to_create = num_possible_nodes(); |
693 | int node; | |
771b53d0 | 694 | |
b60fda60 | 695 | /* create fixed workers */ |
3fc50ab5 JH |
696 | refcount_set(&wq->refs, workers_to_create); |
697 | for_each_node(node) { | |
7563439a JA |
698 | if (!node_online(node)) |
699 | continue; | |
3fc50ab5 JH |
700 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) |
701 | goto err; | |
702 | workers_to_create--; | |
b60fda60 | 703 | } |
771b53d0 | 704 | |
7563439a JA |
705 | while (workers_to_create--) |
706 | refcount_dec(&wq->refs); | |
707 | ||
b60fda60 JA |
708 | complete(&wq->done); |
709 | ||
710 | while (!kthread_should_stop()) { | |
3fc50ab5 JH |
711 | for_each_node(node) { |
712 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 713 | bool fork_worker[2] = { false, false }; |
771b53d0 | 714 | |
7563439a JA |
715 | if (!node_online(node)) |
716 | continue; | |
717 | ||
771b53d0 | 718 | spin_lock_irq(&wqe->lock); |
c5def4ab JA |
719 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
720 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
721 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
722 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
771b53d0 | 723 | spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
724 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
725 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
726 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
727 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
728 | } |
729 | set_current_state(TASK_INTERRUPTIBLE); | |
730 | schedule_timeout(HZ); | |
731 | } | |
732 | ||
733 | return 0; | |
b60fda60 JA |
734 | err: |
735 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
736 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 737 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
738 | complete(&wq->done); |
739 | return 0; | |
771b53d0 JA |
740 | } |
741 | ||
c5def4ab JA |
742 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
743 | struct io_wq_work *work) | |
744 | { | |
745 | bool free_worker; | |
746 | ||
747 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
748 | return true; | |
749 | if (atomic_read(&acct->nr_running)) | |
750 | return true; | |
751 | ||
752 | rcu_read_lock(); | |
021d1cdd | 753 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
754 | rcu_read_unlock(); |
755 | if (free_worker) | |
756 | return true; | |
757 | ||
758 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
759 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
760 | return false; | |
761 | ||
762 | return true; | |
763 | } | |
764 | ||
e9fd9396 | 765 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 766 | { |
e9fd9396 PB |
767 | struct io_wq *wq = wqe->wq; |
768 | ||
fc04c39b PB |
769 | do { |
770 | struct io_wq_work *old_work = work; | |
771 | ||
772 | work->flags |= IO_WQ_WORK_CANCEL; | |
773 | work->func(&work); | |
774 | work = (work == old_work) ? NULL : work; | |
e9fd9396 | 775 | wq->free_work(old_work); |
fc04c39b PB |
776 | } while (work); |
777 | } | |
778 | ||
771b53d0 JA |
779 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
780 | { | |
c5def4ab | 781 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 782 | int work_flags; |
771b53d0 JA |
783 | unsigned long flags; |
784 | ||
c5def4ab JA |
785 | /* |
786 | * Do early check to see if we need a new unbound worker, and if we do, | |
787 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
788 | * gap between this check and incrementing the value, but that's OK. | |
789 | * It's close enough to not be an issue, fork() has the same delay. | |
790 | */ | |
791 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
e9fd9396 | 792 | io_run_cancel(work, wqe); |
c5def4ab JA |
793 | return; |
794 | } | |
795 | ||
895e2ca0 | 796 | work_flags = work->flags; |
771b53d0 | 797 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 | 798 | wq_list_add_tail(&work->list, &wqe->work_list); |
771b53d0 JA |
799 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
800 | spin_unlock_irqrestore(&wqe->lock, flags); | |
801 | ||
895e2ca0 JA |
802 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
803 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 804 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
805 | } |
806 | ||
807 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
808 | { | |
809 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
810 | ||
811 | io_wqe_enqueue(wqe, work); | |
812 | } | |
813 | ||
814 | /* | |
8766dd51 PB |
815 | * Work items that hash to the same value will not be done in parallel. |
816 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 817 | */ |
8766dd51 | 818 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 819 | { |
8766dd51 | 820 | unsigned int bit; |
771b53d0 JA |
821 | |
822 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
823 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
824 | } |
825 | ||
826 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
827 | { | |
828 | send_sig(SIGINT, worker->task, 1); | |
829 | return false; | |
830 | } | |
831 | ||
832 | /* | |
833 | * Iterate the passed in list and call the specific function for each | |
834 | * worker that isn't exiting | |
835 | */ | |
836 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
837 | bool (*func)(struct io_worker *, void *), |
838 | void *data) | |
839 | { | |
771b53d0 JA |
840 | struct io_worker *worker; |
841 | bool ret = false; | |
842 | ||
e61df66c | 843 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 | 844 | if (io_worker_get(worker)) { |
7563439a JA |
845 | /* no task if node is/was offline */ |
846 | if (worker->task) | |
847 | ret = func(worker, data); | |
771b53d0 JA |
848 | io_worker_release(worker); |
849 | if (ret) | |
850 | break; | |
851 | } | |
852 | } | |
e61df66c | 853 | |
771b53d0 JA |
854 | return ret; |
855 | } | |
856 | ||
857 | void io_wq_cancel_all(struct io_wq *wq) | |
858 | { | |
3fc50ab5 | 859 | int node; |
771b53d0 JA |
860 | |
861 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
862 | ||
771b53d0 | 863 | rcu_read_lock(); |
3fc50ab5 JH |
864 | for_each_node(node) { |
865 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 866 | |
e61df66c | 867 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
868 | } |
869 | rcu_read_unlock(); | |
870 | } | |
871 | ||
62755e35 | 872 | struct io_cb_cancel_data { |
2293b419 PB |
873 | work_cancel_fn *fn; |
874 | void *data; | |
62755e35 JA |
875 | }; |
876 | ||
2293b419 | 877 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 878 | { |
2293b419 | 879 | struct io_cb_cancel_data *match = data; |
6f72653e | 880 | unsigned long flags; |
62755e35 JA |
881 | bool ret = false; |
882 | ||
883 | /* | |
884 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 885 | * may dereference the passed in work. |
62755e35 | 886 | */ |
36c2f922 | 887 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 888 | if (worker->cur_work && |
0c9d5ccd | 889 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
2293b419 | 890 | match->fn(worker->cur_work, match->data)) { |
771b53d0 | 891 | send_sig(SIGINT, worker->task, 1); |
36c2f922 | 892 | ret = true; |
771b53d0 | 893 | } |
36c2f922 | 894 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 895 | |
36c2f922 | 896 | return ret; |
771b53d0 JA |
897 | } |
898 | ||
899 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, | |
2293b419 | 900 | struct io_cb_cancel_data *match) |
771b53d0 | 901 | { |
6206f0e1 | 902 | struct io_wq_work_node *node, *prev; |
771b53d0 | 903 | struct io_wq_work *work; |
6f72653e | 904 | unsigned long flags; |
771b53d0 JA |
905 | bool found = false; |
906 | ||
771b53d0 JA |
907 | /* |
908 | * First check pending list, if we're lucky we can just remove it | |
909 | * from there. CANCEL_OK means that the work is returned as-new, | |
910 | * no completion will be posted for it. | |
911 | */ | |
6f72653e | 912 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
913 | wq_list_for_each(node, prev, &wqe->work_list) { |
914 | work = container_of(node, struct io_wq_work, list); | |
915 | ||
00bcda13 | 916 | if (match->fn(work, match->data)) { |
6206f0e1 | 917 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
918 | found = true; |
919 | break; | |
920 | } | |
921 | } | |
6f72653e | 922 | spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 JA |
923 | |
924 | if (found) { | |
e9fd9396 | 925 | io_run_cancel(work, wqe); |
771b53d0 JA |
926 | return IO_WQ_CANCEL_OK; |
927 | } | |
928 | ||
929 | /* | |
930 | * Now check if a free (going busy) or busy worker has the work | |
931 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
d195a66e | 932 | * as an indication that we attempt to signal cancellation. The |
771b53d0 JA |
933 | * completion will run normally in this case. |
934 | */ | |
935 | rcu_read_lock(); | |
00bcda13 | 936 | found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 JA |
937 | rcu_read_unlock(); |
938 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
939 | } | |
940 | ||
2293b419 PB |
941 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
942 | void *data) | |
771b53d0 | 943 | { |
2293b419 PB |
944 | struct io_cb_cancel_data match = { |
945 | .fn = cancel, | |
946 | .data = data, | |
00bcda13 | 947 | }; |
771b53d0 | 948 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; |
3fc50ab5 | 949 | int node; |
771b53d0 | 950 | |
3fc50ab5 JH |
951 | for_each_node(node) { |
952 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 953 | |
00bcda13 | 954 | ret = io_wqe_cancel_work(wqe, &match); |
771b53d0 JA |
955 | if (ret != IO_WQ_CANCEL_NOTFOUND) |
956 | break; | |
957 | } | |
958 | ||
959 | return ret; | |
960 | } | |
961 | ||
2293b419 PB |
962 | static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data) |
963 | { | |
964 | return work == data; | |
965 | } | |
966 | ||
967 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
968 | { | |
969 | return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork); | |
970 | } | |
971 | ||
36282881 JA |
972 | static bool io_wq_pid_match(struct io_wq_work *work, void *data) |
973 | { | |
974 | pid_t pid = (pid_t) (unsigned long) data; | |
975 | ||
2293b419 | 976 | return work->task_pid == pid; |
36282881 JA |
977 | } |
978 | ||
979 | enum io_wq_cancel io_wq_cancel_pid(struct io_wq *wq, pid_t pid) | |
980 | { | |
2293b419 | 981 | void *data = (void *) (unsigned long) pid; |
36282881 | 982 | |
2293b419 | 983 | return io_wq_cancel_cb(wq, io_wq_pid_match, data); |
36282881 JA |
984 | } |
985 | ||
576a347b | 986 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 987 | { |
3fc50ab5 | 988 | int ret = -ENOMEM, node; |
771b53d0 JA |
989 | struct io_wq *wq; |
990 | ||
e9fd9396 PB |
991 | if (WARN_ON_ONCE(!data->free_work)) |
992 | return ERR_PTR(-EINVAL); | |
993 | ||
ad6e005c | 994 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
995 | if (!wq) |
996 | return ERR_PTR(-ENOMEM); | |
997 | ||
3fc50ab5 | 998 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
999 | if (!wq->wqes) { |
1000 | kfree(wq); | |
1001 | return ERR_PTR(-ENOMEM); | |
1002 | } | |
1003 | ||
e9fd9396 | 1004 | wq->free_work = data->free_work; |
7d723065 | 1005 | |
c5def4ab | 1006 | /* caller must already hold a reference to this */ |
576a347b | 1007 | wq->user = data->user; |
c5def4ab | 1008 | |
3fc50ab5 | 1009 | for_each_node(node) { |
771b53d0 | 1010 | struct io_wqe *wqe; |
7563439a | 1011 | int alloc_node = node; |
771b53d0 | 1012 | |
7563439a JA |
1013 | if (!node_online(alloc_node)) |
1014 | alloc_node = NUMA_NO_NODE; | |
1015 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1016 | if (!wqe) |
3fc50ab5 JH |
1017 | goto err; |
1018 | wq->wqes[node] = wqe; | |
7563439a | 1019 | wqe->node = alloc_node; |
c5def4ab JA |
1020 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1021 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1022 | if (wq->user) { |
c5def4ab JA |
1023 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1024 | task_rlimit(current, RLIMIT_NPROC); | |
1025 | } | |
1026 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 JA |
1027 | wqe->wq = wq; |
1028 | spin_lock_init(&wqe->lock); | |
6206f0e1 | 1029 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1030 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1031 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1032 | } |
1033 | ||
1034 | init_completion(&wq->done); | |
1035 | ||
771b53d0 JA |
1036 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
1037 | if (!IS_ERR(wq->manager)) { | |
1038 | wake_up_process(wq->manager); | |
b60fda60 JA |
1039 | wait_for_completion(&wq->done); |
1040 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1041 | ret = -ENOMEM; | |
1042 | goto err; | |
1043 | } | |
848f7e18 | 1044 | refcount_set(&wq->use_refs, 1); |
b60fda60 | 1045 | reinit_completion(&wq->done); |
771b53d0 JA |
1046 | return wq; |
1047 | } | |
1048 | ||
1049 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1050 | complete(&wq->done); |
b60fda60 | 1051 | err: |
3fc50ab5 JH |
1052 | for_each_node(node) |
1053 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1054 | kfree(wq->wqes); |
1055 | kfree(wq); | |
771b53d0 JA |
1056 | return ERR_PTR(ret); |
1057 | } | |
1058 | ||
eba6f5a3 PB |
1059 | bool io_wq_get(struct io_wq *wq, struct io_wq_data *data) |
1060 | { | |
e9fd9396 | 1061 | if (data->free_work != wq->free_work) |
eba6f5a3 PB |
1062 | return false; |
1063 | ||
1064 | return refcount_inc_not_zero(&wq->use_refs); | |
1065 | } | |
1066 | ||
771b53d0 JA |
1067 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) |
1068 | { | |
1069 | wake_up_process(worker->task); | |
1070 | return false; | |
1071 | } | |
1072 | ||
848f7e18 | 1073 | static void __io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1074 | { |
3fc50ab5 | 1075 | int node; |
771b53d0 | 1076 | |
b60fda60 JA |
1077 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1078 | if (wq->manager) | |
771b53d0 | 1079 | kthread_stop(wq->manager); |
771b53d0 JA |
1080 | |
1081 | rcu_read_lock(); | |
3fc50ab5 JH |
1082 | for_each_node(node) |
1083 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1084 | rcu_read_unlock(); |
1085 | ||
1086 | wait_for_completion(&wq->done); | |
1087 | ||
3fc50ab5 JH |
1088 | for_each_node(node) |
1089 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1090 | kfree(wq->wqes); |
1091 | kfree(wq); | |
1092 | } | |
848f7e18 JA |
1093 | |
1094 | void io_wq_destroy(struct io_wq *wq) | |
1095 | { | |
1096 | if (refcount_dec_and_test(&wq->use_refs)) | |
1097 | __io_wq_destroy(wq); | |
1098 | } |