]>
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> | |
19 | ||
20 | #include "io-wq.h" | |
21 | ||
22 | #define WORKER_IDLE_TIMEOUT (5 * HZ) | |
23 | ||
24 | enum { | |
25 | IO_WORKER_F_UP = 1, /* up and active */ | |
26 | IO_WORKER_F_RUNNING = 2, /* account as running */ | |
27 | IO_WORKER_F_FREE = 4, /* worker on free list */ | |
28 | IO_WORKER_F_EXITING = 8, /* worker exiting */ | |
29 | IO_WORKER_F_FIXED = 16, /* static idle worker */ | |
c5def4ab | 30 | IO_WORKER_F_BOUND = 32, /* is doing bounded work */ |
771b53d0 JA |
31 | }; |
32 | ||
33 | enum { | |
34 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ | |
35 | IO_WQ_BIT_CANCEL = 1, /* cancel work on list */ | |
b60fda60 | 36 | IO_WQ_BIT_ERROR = 2, /* error on setup */ |
771b53d0 JA |
37 | }; |
38 | ||
39 | enum { | |
40 | IO_WQE_FLAG_STALLED = 1, /* stalled on hash */ | |
41 | }; | |
42 | ||
43 | /* | |
44 | * One for each thread in a wqe pool | |
45 | */ | |
46 | struct io_worker { | |
47 | refcount_t ref; | |
48 | unsigned flags; | |
49 | struct hlist_nulls_node nulls_node; | |
e61df66c | 50 | struct list_head all_list; |
771b53d0 | 51 | struct task_struct *task; |
771b53d0 | 52 | struct io_wqe *wqe; |
36c2f922 | 53 | |
771b53d0 | 54 | struct io_wq_work *cur_work; |
36c2f922 | 55 | spinlock_t lock; |
771b53d0 JA |
56 | |
57 | struct rcu_head rcu; | |
58 | struct mm_struct *mm; | |
181e448d | 59 | const struct cred *creds; |
fcb323cc | 60 | struct files_struct *restore_files; |
771b53d0 JA |
61 | }; |
62 | ||
771b53d0 JA |
63 | #if BITS_PER_LONG == 64 |
64 | #define IO_WQ_HASH_ORDER 6 | |
65 | #else | |
66 | #define IO_WQ_HASH_ORDER 5 | |
67 | #endif | |
68 | ||
c5def4ab JA |
69 | struct io_wqe_acct { |
70 | unsigned nr_workers; | |
71 | unsigned max_workers; | |
72 | atomic_t nr_running; | |
73 | }; | |
74 | ||
75 | enum { | |
76 | IO_WQ_ACCT_BOUND, | |
77 | IO_WQ_ACCT_UNBOUND, | |
78 | }; | |
79 | ||
771b53d0 JA |
80 | /* |
81 | * Per-node worker thread pool | |
82 | */ | |
83 | struct io_wqe { | |
84 | struct { | |
85 | spinlock_t lock; | |
6206f0e1 | 86 | struct io_wq_work_list work_list; |
771b53d0 JA |
87 | unsigned long hash_map; |
88 | unsigned flags; | |
89 | } ____cacheline_aligned_in_smp; | |
90 | ||
91 | int node; | |
c5def4ab | 92 | struct io_wqe_acct acct[2]; |
771b53d0 | 93 | |
021d1cdd | 94 | struct hlist_nulls_head free_list; |
e61df66c | 95 | struct list_head all_list; |
771b53d0 JA |
96 | |
97 | struct io_wq *wq; | |
98 | }; | |
99 | ||
100 | /* | |
101 | * Per io_wq state | |
102 | */ | |
103 | struct io_wq { | |
104 | struct io_wqe **wqes; | |
105 | unsigned long state; | |
771b53d0 | 106 | |
7d723065 JA |
107 | get_work_fn *get_work; |
108 | put_work_fn *put_work; | |
109 | ||
771b53d0 | 110 | struct task_struct *manager; |
c5def4ab | 111 | struct user_struct *user; |
0b8c0ec7 | 112 | const struct cred *creds; |
771b53d0 JA |
113 | struct mm_struct *mm; |
114 | refcount_t refs; | |
115 | struct completion done; | |
116 | }; | |
117 | ||
771b53d0 JA |
118 | static bool io_worker_get(struct io_worker *worker) |
119 | { | |
120 | return refcount_inc_not_zero(&worker->ref); | |
121 | } | |
122 | ||
123 | static void io_worker_release(struct io_worker *worker) | |
124 | { | |
125 | if (refcount_dec_and_test(&worker->ref)) | |
126 | wake_up_process(worker->task); | |
127 | } | |
128 | ||
129 | /* | |
130 | * Note: drops the wqe->lock if returning true! The caller must re-acquire | |
131 | * the lock in that case. Some callers need to restart handling if this | |
132 | * happens, so we can't just re-acquire the lock on behalf of the caller. | |
133 | */ | |
134 | static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) | |
135 | { | |
fcb323cc JA |
136 | bool dropped_lock = false; |
137 | ||
181e448d JA |
138 | if (worker->creds) { |
139 | revert_creds(worker->creds); | |
140 | worker->creds = NULL; | |
141 | } | |
142 | ||
fcb323cc JA |
143 | if (current->files != worker->restore_files) { |
144 | __acquire(&wqe->lock); | |
145 | spin_unlock_irq(&wqe->lock); | |
146 | dropped_lock = true; | |
147 | ||
148 | task_lock(current); | |
149 | current->files = worker->restore_files; | |
150 | task_unlock(current); | |
151 | } | |
152 | ||
771b53d0 JA |
153 | /* |
154 | * If we have an active mm, we need to drop the wq lock before unusing | |
155 | * it. If we do, return true and let the caller retry the idle loop. | |
156 | */ | |
157 | if (worker->mm) { | |
fcb323cc JA |
158 | if (!dropped_lock) { |
159 | __acquire(&wqe->lock); | |
160 | spin_unlock_irq(&wqe->lock); | |
161 | dropped_lock = true; | |
162 | } | |
771b53d0 JA |
163 | __set_current_state(TASK_RUNNING); |
164 | set_fs(KERNEL_DS); | |
165 | unuse_mm(worker->mm); | |
166 | mmput(worker->mm); | |
167 | worker->mm = NULL; | |
771b53d0 JA |
168 | } |
169 | ||
fcb323cc | 170 | return dropped_lock; |
771b53d0 JA |
171 | } |
172 | ||
c5def4ab JA |
173 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
174 | struct io_wq_work *work) | |
175 | { | |
176 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
177 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
178 | ||
179 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
180 | } | |
181 | ||
182 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe, | |
183 | struct io_worker *worker) | |
184 | { | |
185 | if (worker->flags & IO_WORKER_F_BOUND) | |
186 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
187 | ||
188 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
189 | } | |
190 | ||
771b53d0 JA |
191 | static void io_worker_exit(struct io_worker *worker) |
192 | { | |
193 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab JA |
194 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); |
195 | unsigned nr_workers; | |
771b53d0 JA |
196 | |
197 | /* | |
198 | * If we're not at zero, someone else is holding a brief reference | |
199 | * to the worker. Wait for that to go away. | |
200 | */ | |
201 | set_current_state(TASK_INTERRUPTIBLE); | |
202 | if (!refcount_dec_and_test(&worker->ref)) | |
203 | schedule(); | |
204 | __set_current_state(TASK_RUNNING); | |
205 | ||
206 | preempt_disable(); | |
207 | current->flags &= ~PF_IO_WORKER; | |
208 | if (worker->flags & IO_WORKER_F_RUNNING) | |
c5def4ab JA |
209 | atomic_dec(&acct->nr_running); |
210 | if (!(worker->flags & IO_WORKER_F_BOUND)) | |
211 | atomic_dec(&wqe->wq->user->processes); | |
771b53d0 JA |
212 | worker->flags = 0; |
213 | preempt_enable(); | |
214 | ||
215 | spin_lock_irq(&wqe->lock); | |
216 | hlist_nulls_del_rcu(&worker->nulls_node); | |
e61df66c | 217 | list_del_rcu(&worker->all_list); |
771b53d0 JA |
218 | if (__io_worker_unuse(wqe, worker)) { |
219 | __release(&wqe->lock); | |
220 | spin_lock_irq(&wqe->lock); | |
221 | } | |
c5def4ab JA |
222 | acct->nr_workers--; |
223 | nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers + | |
224 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers; | |
771b53d0 JA |
225 | spin_unlock_irq(&wqe->lock); |
226 | ||
227 | /* all workers gone, wq exit can proceed */ | |
c5def4ab | 228 | if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) |
771b53d0 JA |
229 | complete(&wqe->wq->done); |
230 | ||
364b05fd | 231 | kfree_rcu(worker, rcu); |
771b53d0 JA |
232 | } |
233 | ||
c5def4ab JA |
234 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
235 | __must_hold(wqe->lock) | |
236 | { | |
6206f0e1 JA |
237 | if (!wq_list_empty(&wqe->work_list) && |
238 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
239 | return true; |
240 | return false; | |
241 | } | |
242 | ||
243 | /* | |
244 | * Check head of free list for an available worker. If one isn't available, | |
245 | * caller must wake up the wq manager to create one. | |
246 | */ | |
247 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
248 | __must_hold(RCU) | |
249 | { | |
250 | struct hlist_nulls_node *n; | |
251 | struct io_worker *worker; | |
252 | ||
021d1cdd | 253 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
254 | if (is_a_nulls(n)) |
255 | return false; | |
256 | ||
257 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
258 | if (io_worker_get(worker)) { | |
506d95ff | 259 | wake_up_process(worker->task); |
c5def4ab JA |
260 | io_worker_release(worker); |
261 | return true; | |
262 | } | |
263 | ||
264 | return false; | |
265 | } | |
266 | ||
267 | /* | |
268 | * We need a worker. If we find a free one, we're good. If not, and we're | |
269 | * below the max number of workers, wake up the manager to create one. | |
270 | */ | |
271 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
272 | { | |
273 | bool ret; | |
274 | ||
275 | /* | |
276 | * Most likely an attempt to queue unbounded work on an io_wq that | |
277 | * wasn't setup with any unbounded workers. | |
278 | */ | |
279 | WARN_ON_ONCE(!acct->max_workers); | |
280 | ||
281 | rcu_read_lock(); | |
282 | ret = io_wqe_activate_free_worker(wqe); | |
283 | rcu_read_unlock(); | |
284 | ||
285 | if (!ret && acct->nr_workers < acct->max_workers) | |
286 | wake_up_process(wqe->wq->manager); | |
287 | } | |
288 | ||
289 | static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker) | |
290 | { | |
291 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
292 | ||
293 | atomic_inc(&acct->nr_running); | |
294 | } | |
295 | ||
296 | static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker) | |
297 | __must_hold(wqe->lock) | |
298 | { | |
299 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
300 | ||
301 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
302 | io_wqe_wake_worker(wqe, acct); | |
303 | } | |
304 | ||
771b53d0 JA |
305 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
306 | { | |
307 | allow_kernel_signal(SIGINT); | |
308 | ||
309 | current->flags |= PF_IO_WORKER; | |
310 | ||
311 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); | |
fcb323cc | 312 | worker->restore_files = current->files; |
c5def4ab | 313 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
314 | } |
315 | ||
316 | /* | |
317 | * Worker will start processing some work. Move it to the busy list, if | |
318 | * it's currently on the freelist | |
319 | */ | |
320 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
321 | struct io_wq_work *work) | |
322 | __must_hold(wqe->lock) | |
323 | { | |
c5def4ab JA |
324 | bool worker_bound, work_bound; |
325 | ||
771b53d0 JA |
326 | if (worker->flags & IO_WORKER_F_FREE) { |
327 | worker->flags &= ~IO_WORKER_F_FREE; | |
328 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 329 | } |
c5def4ab JA |
330 | |
331 | /* | |
332 | * If worker is moving from bound to unbound (or vice versa), then | |
333 | * ensure we update the running accounting. | |
334 | */ | |
b2e9c7d6 DC |
335 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
336 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
337 | if (worker_bound != work_bound) { | |
c5def4ab JA |
338 | io_wqe_dec_running(wqe, worker); |
339 | if (work_bound) { | |
340 | worker->flags |= IO_WORKER_F_BOUND; | |
341 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
342 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
343 | atomic_dec(&wqe->wq->user->processes); | |
344 | } else { | |
345 | worker->flags &= ~IO_WORKER_F_BOUND; | |
346 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
347 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
348 | atomic_inc(&wqe->wq->user->processes); | |
349 | } | |
350 | io_wqe_inc_running(wqe, worker); | |
351 | } | |
771b53d0 JA |
352 | } |
353 | ||
354 | /* | |
355 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
356 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
357 | * the lock in that case and return success. Since the caller has to | |
358 | * retry the loop in that case (we changed task state), we don't regrab | |
359 | * the lock if we return success. | |
360 | */ | |
361 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) | |
362 | __must_hold(wqe->lock) | |
363 | { | |
364 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
365 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 366 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 JA |
367 | } |
368 | ||
369 | return __io_worker_unuse(wqe, worker); | |
370 | } | |
371 | ||
372 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe, unsigned *hash) | |
373 | __must_hold(wqe->lock) | |
374 | { | |
6206f0e1 | 375 | struct io_wq_work_node *node, *prev; |
771b53d0 JA |
376 | struct io_wq_work *work; |
377 | ||
6206f0e1 JA |
378 | wq_list_for_each(node, prev, &wqe->work_list) { |
379 | work = container_of(node, struct io_wq_work, list); | |
380 | ||
771b53d0 JA |
381 | /* not hashed, can run anytime */ |
382 | if (!(work->flags & IO_WQ_WORK_HASHED)) { | |
6206f0e1 | 383 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
384 | return work; |
385 | } | |
386 | ||
387 | /* hashed, can run if not already running */ | |
388 | *hash = work->flags >> IO_WQ_HASH_SHIFT; | |
389 | if (!(wqe->hash_map & BIT_ULL(*hash))) { | |
390 | wqe->hash_map |= BIT_ULL(*hash); | |
6206f0e1 | 391 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
392 | return work; |
393 | } | |
394 | } | |
395 | ||
396 | return NULL; | |
397 | } | |
398 | ||
399 | static void io_worker_handle_work(struct io_worker *worker) | |
400 | __releases(wqe->lock) | |
401 | { | |
7d723065 | 402 | struct io_wq_work *work, *old_work = NULL, *put_work = NULL; |
771b53d0 JA |
403 | struct io_wqe *wqe = worker->wqe; |
404 | struct io_wq *wq = wqe->wq; | |
405 | ||
406 | do { | |
407 | unsigned hash = -1U; | |
408 | ||
771b53d0 JA |
409 | /* |
410 | * If we got some work, mark us as busy. If we didn't, but | |
411 | * the list isn't empty, it means we stalled on hashed work. | |
412 | * Mark us stalled so we don't keep looking for work when we | |
413 | * can't make progress, any work completion or insertion will | |
414 | * clear the stalled flag. | |
415 | */ | |
416 | work = io_get_next_work(wqe, &hash); | |
417 | if (work) | |
418 | __io_worker_busy(wqe, worker, work); | |
6206f0e1 | 419 | else if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
420 | wqe->flags |= IO_WQE_FLAG_STALLED; |
421 | ||
422 | spin_unlock_irq(&wqe->lock); | |
7d723065 JA |
423 | if (put_work && wq->put_work) |
424 | wq->put_work(old_work); | |
771b53d0 JA |
425 | if (!work) |
426 | break; | |
427 | next: | |
36c2f922 JA |
428 | /* flush any pending signals before assigning new work */ |
429 | if (signal_pending(current)) | |
430 | flush_signals(current); | |
431 | ||
fd1c4bc6 HD |
432 | cond_resched(); |
433 | ||
36c2f922 JA |
434 | spin_lock_irq(&worker->lock); |
435 | worker->cur_work = work; | |
436 | spin_unlock_irq(&worker->lock); | |
437 | ||
b76da70f JA |
438 | if (work->flags & IO_WQ_WORK_CB) |
439 | work->func(&work); | |
440 | ||
fcb323cc JA |
441 | if ((work->flags & IO_WQ_WORK_NEEDS_FILES) && |
442 | current->files != work->files) { | |
443 | task_lock(current); | |
444 | current->files = work->files; | |
445 | task_unlock(current); | |
446 | } | |
771b53d0 | 447 | if ((work->flags & IO_WQ_WORK_NEEDS_USER) && !worker->mm && |
e0bbb346 JA |
448 | wq->mm) { |
449 | if (mmget_not_zero(wq->mm)) { | |
450 | use_mm(wq->mm); | |
451 | set_fs(USER_DS); | |
452 | worker->mm = wq->mm; | |
453 | } else { | |
454 | work->flags |= IO_WQ_WORK_CANCEL; | |
455 | } | |
771b53d0 | 456 | } |
181e448d JA |
457 | if (!worker->creds) |
458 | worker->creds = override_creds(wq->creds); | |
771b53d0 JA |
459 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) |
460 | work->flags |= IO_WQ_WORK_CANCEL; | |
461 | if (worker->mm) | |
462 | work->flags |= IO_WQ_WORK_HAS_MM; | |
463 | ||
7d723065 JA |
464 | if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) { |
465 | put_work = work; | |
466 | wq->get_work(work); | |
467 | } | |
468 | ||
771b53d0 JA |
469 | old_work = work; |
470 | work->func(&work); | |
471 | ||
36c2f922 | 472 | spin_lock_irq(&worker->lock); |
771b53d0 | 473 | worker->cur_work = NULL; |
36c2f922 JA |
474 | spin_unlock_irq(&worker->lock); |
475 | ||
476 | spin_lock_irq(&wqe->lock); | |
477 | ||
771b53d0 JA |
478 | if (hash != -1U) { |
479 | wqe->hash_map &= ~BIT_ULL(hash); | |
480 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
481 | } | |
482 | if (work && work != old_work) { | |
483 | spin_unlock_irq(&wqe->lock); | |
7d723065 JA |
484 | |
485 | if (put_work && wq->put_work) { | |
486 | wq->put_work(put_work); | |
487 | put_work = NULL; | |
488 | } | |
489 | ||
771b53d0 JA |
490 | /* dependent work not hashed */ |
491 | hash = -1U; | |
492 | goto next; | |
493 | } | |
494 | } while (1); | |
495 | } | |
496 | ||
e995d512 JA |
497 | static inline void io_worker_spin_for_work(struct io_wqe *wqe) |
498 | { | |
499 | int i = 0; | |
500 | ||
501 | while (++i < 1000) { | |
502 | if (io_wqe_run_queue(wqe)) | |
503 | break; | |
504 | if (need_resched()) | |
505 | break; | |
506 | cpu_relax(); | |
507 | } | |
508 | } | |
509 | ||
771b53d0 JA |
510 | static int io_wqe_worker(void *data) |
511 | { | |
512 | struct io_worker *worker = data; | |
513 | struct io_wqe *wqe = worker->wqe; | |
514 | struct io_wq *wq = wqe->wq; | |
e995d512 | 515 | bool did_work; |
771b53d0 JA |
516 | |
517 | io_worker_start(wqe, worker); | |
518 | ||
e995d512 | 519 | did_work = false; |
771b53d0 | 520 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
506d95ff | 521 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 JA |
522 | loop: |
523 | if (did_work) | |
524 | io_worker_spin_for_work(wqe); | |
771b53d0 JA |
525 | spin_lock_irq(&wqe->lock); |
526 | if (io_wqe_run_queue(wqe)) { | |
527 | __set_current_state(TASK_RUNNING); | |
528 | io_worker_handle_work(worker); | |
e995d512 JA |
529 | did_work = true; |
530 | goto loop; | |
771b53d0 | 531 | } |
e995d512 | 532 | did_work = false; |
771b53d0 JA |
533 | /* drops the lock on success, retry */ |
534 | if (__io_worker_idle(wqe, worker)) { | |
535 | __release(&wqe->lock); | |
e995d512 | 536 | goto loop; |
771b53d0 JA |
537 | } |
538 | spin_unlock_irq(&wqe->lock); | |
539 | if (signal_pending(current)) | |
540 | flush_signals(current); | |
541 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
542 | continue; | |
543 | /* timed out, exit unless we're the fixed worker */ | |
544 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
545 | !(worker->flags & IO_WORKER_F_FIXED)) | |
546 | break; | |
547 | } | |
548 | ||
771b53d0 JA |
549 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
550 | spin_lock_irq(&wqe->lock); | |
6206f0e1 | 551 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
552 | io_worker_handle_work(worker); |
553 | else | |
554 | spin_unlock_irq(&wqe->lock); | |
555 | } | |
556 | ||
557 | io_worker_exit(worker); | |
558 | return 0; | |
559 | } | |
560 | ||
771b53d0 JA |
561 | /* |
562 | * Called when a worker is scheduled in. Mark us as currently running. | |
563 | */ | |
564 | void io_wq_worker_running(struct task_struct *tsk) | |
565 | { | |
566 | struct io_worker *worker = kthread_data(tsk); | |
567 | struct io_wqe *wqe = worker->wqe; | |
568 | ||
569 | if (!(worker->flags & IO_WORKER_F_UP)) | |
570 | return; | |
571 | if (worker->flags & IO_WORKER_F_RUNNING) | |
572 | return; | |
573 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 574 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
575 | } |
576 | ||
577 | /* | |
578 | * Called when worker is going to sleep. If there are no workers currently | |
579 | * running and we have work pending, wake up a free one or have the manager | |
580 | * set one up. | |
581 | */ | |
582 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
583 | { | |
584 | struct io_worker *worker = kthread_data(tsk); | |
585 | struct io_wqe *wqe = worker->wqe; | |
586 | ||
587 | if (!(worker->flags & IO_WORKER_F_UP)) | |
588 | return; | |
589 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
590 | return; | |
591 | ||
592 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
593 | ||
594 | spin_lock_irq(&wqe->lock); | |
c5def4ab | 595 | io_wqe_dec_running(wqe, worker); |
771b53d0 JA |
596 | spin_unlock_irq(&wqe->lock); |
597 | } | |
598 | ||
b60fda60 | 599 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 600 | { |
c5def4ab | 601 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
602 | struct io_worker *worker; |
603 | ||
ad6e005c | 604 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 605 | if (!worker) |
b60fda60 | 606 | return false; |
771b53d0 JA |
607 | |
608 | refcount_set(&worker->ref, 1); | |
609 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 610 | worker->wqe = wqe; |
36c2f922 | 611 | spin_lock_init(&worker->lock); |
771b53d0 JA |
612 | |
613 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 614 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
615 | if (IS_ERR(worker->task)) { |
616 | kfree(worker); | |
b60fda60 | 617 | return false; |
771b53d0 JA |
618 | } |
619 | ||
620 | spin_lock_irq(&wqe->lock); | |
021d1cdd | 621 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 622 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 623 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
624 | if (index == IO_WQ_ACCT_BOUND) |
625 | worker->flags |= IO_WORKER_F_BOUND; | |
626 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 627 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 628 | acct->nr_workers++; |
771b53d0 JA |
629 | spin_unlock_irq(&wqe->lock); |
630 | ||
c5def4ab JA |
631 | if (index == IO_WQ_ACCT_UNBOUND) |
632 | atomic_inc(&wq->user->processes); | |
633 | ||
771b53d0 | 634 | wake_up_process(worker->task); |
b60fda60 | 635 | return true; |
771b53d0 JA |
636 | } |
637 | ||
c5def4ab | 638 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
639 | __must_hold(wqe->lock) |
640 | { | |
c5def4ab | 641 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 642 | |
c5def4ab | 643 | /* if we have available workers or no work, no need */ |
021d1cdd | 644 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
645 | return false; |
646 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
647 | } |
648 | ||
649 | /* | |
650 | * Manager thread. Tasked with creating new workers, if we need them. | |
651 | */ | |
652 | static int io_wq_manager(void *data) | |
653 | { | |
654 | struct io_wq *wq = data; | |
3fc50ab5 JH |
655 | int workers_to_create = num_possible_nodes(); |
656 | int node; | |
771b53d0 | 657 | |
b60fda60 | 658 | /* create fixed workers */ |
3fc50ab5 JH |
659 | refcount_set(&wq->refs, workers_to_create); |
660 | for_each_node(node) { | |
661 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) | |
662 | goto err; | |
663 | workers_to_create--; | |
b60fda60 | 664 | } |
771b53d0 | 665 | |
b60fda60 JA |
666 | complete(&wq->done); |
667 | ||
668 | while (!kthread_should_stop()) { | |
3fc50ab5 JH |
669 | for_each_node(node) { |
670 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 671 | bool fork_worker[2] = { false, false }; |
771b53d0 JA |
672 | |
673 | spin_lock_irq(&wqe->lock); | |
c5def4ab JA |
674 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
675 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
676 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
677 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
771b53d0 | 678 | spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
679 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
680 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
681 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
682 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
683 | } |
684 | set_current_state(TASK_INTERRUPTIBLE); | |
685 | schedule_timeout(HZ); | |
686 | } | |
687 | ||
688 | return 0; | |
b60fda60 JA |
689 | err: |
690 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
691 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 692 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
693 | complete(&wq->done); |
694 | return 0; | |
771b53d0 JA |
695 | } |
696 | ||
c5def4ab JA |
697 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
698 | struct io_wq_work *work) | |
699 | { | |
700 | bool free_worker; | |
701 | ||
702 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
703 | return true; | |
704 | if (atomic_read(&acct->nr_running)) | |
705 | return true; | |
706 | ||
707 | rcu_read_lock(); | |
021d1cdd | 708 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
709 | rcu_read_unlock(); |
710 | if (free_worker) | |
711 | return true; | |
712 | ||
713 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
714 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
715 | return false; | |
716 | ||
717 | return true; | |
718 | } | |
719 | ||
771b53d0 JA |
720 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
721 | { | |
c5def4ab | 722 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
771b53d0 JA |
723 | unsigned long flags; |
724 | ||
c5def4ab JA |
725 | /* |
726 | * Do early check to see if we need a new unbound worker, and if we do, | |
727 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
728 | * gap between this check and incrementing the value, but that's OK. | |
729 | * It's close enough to not be an issue, fork() has the same delay. | |
730 | */ | |
731 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
732 | work->flags |= IO_WQ_WORK_CANCEL; | |
733 | work->func(&work); | |
734 | return; | |
735 | } | |
736 | ||
771b53d0 | 737 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 | 738 | wq_list_add_tail(&work->list, &wqe->work_list); |
771b53d0 JA |
739 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
740 | spin_unlock_irqrestore(&wqe->lock, flags); | |
741 | ||
c5def4ab JA |
742 | if (!atomic_read(&acct->nr_running)) |
743 | io_wqe_wake_worker(wqe, acct); | |
771b53d0 JA |
744 | } |
745 | ||
746 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
747 | { | |
748 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
749 | ||
750 | io_wqe_enqueue(wqe, work); | |
751 | } | |
752 | ||
753 | /* | |
754 | * Enqueue work, hashed by some key. Work items that hash to the same value | |
755 | * will not be done in parallel. Used to limit concurrent writes, generally | |
756 | * hashed by inode. | |
757 | */ | |
758 | void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val) | |
759 | { | |
760 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
761 | unsigned bit; | |
762 | ||
763 | ||
764 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
765 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
766 | io_wqe_enqueue(wqe, work); | |
767 | } | |
768 | ||
769 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
770 | { | |
771 | send_sig(SIGINT, worker->task, 1); | |
772 | return false; | |
773 | } | |
774 | ||
775 | /* | |
776 | * Iterate the passed in list and call the specific function for each | |
777 | * worker that isn't exiting | |
778 | */ | |
779 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
780 | bool (*func)(struct io_worker *, void *), |
781 | void *data) | |
782 | { | |
771b53d0 JA |
783 | struct io_worker *worker; |
784 | bool ret = false; | |
785 | ||
e61df66c | 786 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 JA |
787 | if (io_worker_get(worker)) { |
788 | ret = func(worker, data); | |
789 | io_worker_release(worker); | |
790 | if (ret) | |
791 | break; | |
792 | } | |
793 | } | |
e61df66c | 794 | |
771b53d0 JA |
795 | return ret; |
796 | } | |
797 | ||
798 | void io_wq_cancel_all(struct io_wq *wq) | |
799 | { | |
3fc50ab5 | 800 | int node; |
771b53d0 JA |
801 | |
802 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
803 | ||
771b53d0 | 804 | rcu_read_lock(); |
3fc50ab5 JH |
805 | for_each_node(node) { |
806 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 807 | |
e61df66c | 808 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
809 | } |
810 | rcu_read_unlock(); | |
811 | } | |
812 | ||
62755e35 JA |
813 | struct io_cb_cancel_data { |
814 | struct io_wqe *wqe; | |
815 | work_cancel_fn *cancel; | |
816 | void *caller_data; | |
817 | }; | |
818 | ||
819 | static bool io_work_cancel(struct io_worker *worker, void *cancel_data) | |
820 | { | |
821 | struct io_cb_cancel_data *data = cancel_data; | |
6f72653e | 822 | unsigned long flags; |
62755e35 JA |
823 | bool ret = false; |
824 | ||
825 | /* | |
826 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 827 | * may dereference the passed in work. |
62755e35 | 828 | */ |
36c2f922 | 829 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 JA |
830 | if (worker->cur_work && |
831 | data->cancel(worker->cur_work, data->caller_data)) { | |
832 | send_sig(SIGINT, worker->task, 1); | |
833 | ret = true; | |
834 | } | |
36c2f922 | 835 | spin_unlock_irqrestore(&worker->lock, flags); |
62755e35 JA |
836 | |
837 | return ret; | |
838 | } | |
839 | ||
840 | static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe, | |
841 | work_cancel_fn *cancel, | |
842 | void *cancel_data) | |
843 | { | |
844 | struct io_cb_cancel_data data = { | |
845 | .wqe = wqe, | |
846 | .cancel = cancel, | |
847 | .caller_data = cancel_data, | |
848 | }; | |
6206f0e1 | 849 | struct io_wq_work_node *node, *prev; |
62755e35 | 850 | struct io_wq_work *work; |
6f72653e | 851 | unsigned long flags; |
62755e35 JA |
852 | bool found = false; |
853 | ||
6f72653e | 854 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
855 | wq_list_for_each(node, prev, &wqe->work_list) { |
856 | work = container_of(node, struct io_wq_work, list); | |
857 | ||
62755e35 | 858 | if (cancel(work, cancel_data)) { |
6206f0e1 | 859 | wq_node_del(&wqe->work_list, node, prev); |
62755e35 JA |
860 | found = true; |
861 | break; | |
862 | } | |
863 | } | |
6f72653e | 864 | spin_unlock_irqrestore(&wqe->lock, flags); |
62755e35 JA |
865 | |
866 | if (found) { | |
867 | work->flags |= IO_WQ_WORK_CANCEL; | |
868 | work->func(&work); | |
869 | return IO_WQ_CANCEL_OK; | |
870 | } | |
871 | ||
872 | rcu_read_lock(); | |
e61df66c | 873 | found = io_wq_for_each_worker(wqe, io_work_cancel, &data); |
62755e35 JA |
874 | rcu_read_unlock(); |
875 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
876 | } | |
877 | ||
878 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, | |
879 | void *data) | |
880 | { | |
881 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 882 | int node; |
62755e35 | 883 | |
3fc50ab5 JH |
884 | for_each_node(node) { |
885 | struct io_wqe *wqe = wq->wqes[node]; | |
62755e35 JA |
886 | |
887 | ret = io_wqe_cancel_cb_work(wqe, cancel, data); | |
888 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
889 | break; | |
890 | } | |
891 | ||
892 | return ret; | |
893 | } | |
894 | ||
771b53d0 JA |
895 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
896 | { | |
897 | struct io_wq_work *work = data; | |
36c2f922 JA |
898 | unsigned long flags; |
899 | bool ret = false; | |
771b53d0 | 900 | |
36c2f922 JA |
901 | if (worker->cur_work != work) |
902 | return false; | |
903 | ||
904 | spin_lock_irqsave(&worker->lock, flags); | |
771b53d0 JA |
905 | if (worker->cur_work == work) { |
906 | send_sig(SIGINT, worker->task, 1); | |
36c2f922 | 907 | ret = true; |
771b53d0 | 908 | } |
36c2f922 | 909 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 910 | |
36c2f922 | 911 | return ret; |
771b53d0 JA |
912 | } |
913 | ||
914 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, | |
915 | struct io_wq_work *cwork) | |
916 | { | |
6206f0e1 | 917 | struct io_wq_work_node *node, *prev; |
771b53d0 | 918 | struct io_wq_work *work; |
6f72653e | 919 | unsigned long flags; |
771b53d0 JA |
920 | bool found = false; |
921 | ||
922 | cwork->flags |= IO_WQ_WORK_CANCEL; | |
923 | ||
924 | /* | |
925 | * First check pending list, if we're lucky we can just remove it | |
926 | * from there. CANCEL_OK means that the work is returned as-new, | |
927 | * no completion will be posted for it. | |
928 | */ | |
6f72653e | 929 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
930 | wq_list_for_each(node, prev, &wqe->work_list) { |
931 | work = container_of(node, struct io_wq_work, list); | |
932 | ||
771b53d0 | 933 | if (work == cwork) { |
6206f0e1 | 934 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
935 | found = true; |
936 | break; | |
937 | } | |
938 | } | |
6f72653e | 939 | spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 JA |
940 | |
941 | if (found) { | |
942 | work->flags |= IO_WQ_WORK_CANCEL; | |
943 | work->func(&work); | |
944 | return IO_WQ_CANCEL_OK; | |
945 | } | |
946 | ||
947 | /* | |
948 | * Now check if a free (going busy) or busy worker has the work | |
949 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
d195a66e | 950 | * as an indication that we attempt to signal cancellation. The |
771b53d0 JA |
951 | * completion will run normally in this case. |
952 | */ | |
953 | rcu_read_lock(); | |
e61df66c | 954 | found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, cwork); |
771b53d0 JA |
955 | rcu_read_unlock(); |
956 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
957 | } | |
958 | ||
959 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
960 | { | |
961 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 962 | int node; |
771b53d0 | 963 | |
3fc50ab5 JH |
964 | for_each_node(node) { |
965 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
966 | |
967 | ret = io_wqe_cancel_work(wqe, cwork); | |
968 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
969 | break; | |
970 | } | |
971 | ||
972 | return ret; | |
973 | } | |
974 | ||
975 | struct io_wq_flush_data { | |
976 | struct io_wq_work work; | |
977 | struct completion done; | |
978 | }; | |
979 | ||
980 | static void io_wq_flush_func(struct io_wq_work **workptr) | |
981 | { | |
982 | struct io_wq_work *work = *workptr; | |
983 | struct io_wq_flush_data *data; | |
984 | ||
985 | data = container_of(work, struct io_wq_flush_data, work); | |
986 | complete(&data->done); | |
987 | } | |
988 | ||
989 | /* | |
990 | * Doesn't wait for previously queued work to finish. When this completes, | |
991 | * it just means that previously queued work was started. | |
992 | */ | |
993 | void io_wq_flush(struct io_wq *wq) | |
994 | { | |
995 | struct io_wq_flush_data data; | |
3fc50ab5 | 996 | int node; |
771b53d0 | 997 | |
3fc50ab5 JH |
998 | for_each_node(node) { |
999 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
1000 | |
1001 | init_completion(&data.done); | |
1002 | INIT_IO_WORK(&data.work, io_wq_flush_func); | |
7d723065 | 1003 | data.work.flags |= IO_WQ_WORK_INTERNAL; |
771b53d0 JA |
1004 | io_wqe_enqueue(wqe, &data.work); |
1005 | wait_for_completion(&data.done); | |
1006 | } | |
1007 | } | |
1008 | ||
576a347b | 1009 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1010 | { |
3fc50ab5 | 1011 | int ret = -ENOMEM, node; |
771b53d0 JA |
1012 | struct io_wq *wq; |
1013 | ||
ad6e005c | 1014 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1015 | if (!wq) |
1016 | return ERR_PTR(-ENOMEM); | |
1017 | ||
3fc50ab5 | 1018 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1019 | if (!wq->wqes) { |
1020 | kfree(wq); | |
1021 | return ERR_PTR(-ENOMEM); | |
1022 | } | |
1023 | ||
576a347b JA |
1024 | wq->get_work = data->get_work; |
1025 | wq->put_work = data->put_work; | |
7d723065 | 1026 | |
c5def4ab | 1027 | /* caller must already hold a reference to this */ |
576a347b | 1028 | wq->user = data->user; |
181e448d | 1029 | wq->creds = data->creds; |
c5def4ab | 1030 | |
3fc50ab5 | 1031 | for_each_node(node) { |
771b53d0 JA |
1032 | struct io_wqe *wqe; |
1033 | ||
ad6e005c | 1034 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node); |
771b53d0 | 1035 | if (!wqe) |
3fc50ab5 JH |
1036 | goto err; |
1037 | wq->wqes[node] = wqe; | |
771b53d0 | 1038 | wqe->node = node; |
c5def4ab JA |
1039 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1040 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1041 | if (wq->user) { |
c5def4ab JA |
1042 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1043 | task_rlimit(current, RLIMIT_NPROC); | |
1044 | } | |
1045 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 JA |
1046 | wqe->node = node; |
1047 | wqe->wq = wq; | |
1048 | spin_lock_init(&wqe->lock); | |
6206f0e1 | 1049 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1050 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1051 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1052 | } |
1053 | ||
1054 | init_completion(&wq->done); | |
1055 | ||
771b53d0 | 1056 | /* caller must have already done mmgrab() on this mm */ |
576a347b | 1057 | wq->mm = data->mm; |
771b53d0 JA |
1058 | |
1059 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); | |
1060 | if (!IS_ERR(wq->manager)) { | |
1061 | wake_up_process(wq->manager); | |
b60fda60 JA |
1062 | wait_for_completion(&wq->done); |
1063 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1064 | ret = -ENOMEM; | |
1065 | goto err; | |
1066 | } | |
1067 | reinit_completion(&wq->done); | |
771b53d0 JA |
1068 | return wq; |
1069 | } | |
1070 | ||
1071 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1072 | complete(&wq->done); |
b60fda60 | 1073 | err: |
3fc50ab5 JH |
1074 | for_each_node(node) |
1075 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1076 | kfree(wq->wqes); |
1077 | kfree(wq); | |
771b53d0 JA |
1078 | return ERR_PTR(ret); |
1079 | } | |
1080 | ||
1081 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
1082 | { | |
1083 | wake_up_process(worker->task); | |
1084 | return false; | |
1085 | } | |
1086 | ||
1087 | void io_wq_destroy(struct io_wq *wq) | |
1088 | { | |
3fc50ab5 | 1089 | int node; |
771b53d0 | 1090 | |
b60fda60 JA |
1091 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1092 | if (wq->manager) | |
771b53d0 | 1093 | kthread_stop(wq->manager); |
771b53d0 JA |
1094 | |
1095 | rcu_read_lock(); | |
3fc50ab5 JH |
1096 | for_each_node(node) |
1097 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1098 | rcu_read_unlock(); |
1099 | ||
1100 | wait_for_completion(&wq->done); | |
1101 | ||
3fc50ab5 JH |
1102 | for_each_node(node) |
1103 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1104 | kfree(wq->wqes); |
1105 | kfree(wq); | |
1106 | } |