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 | ||
432 | spin_lock_irq(&worker->lock); | |
433 | worker->cur_work = work; | |
434 | spin_unlock_irq(&worker->lock); | |
435 | ||
b76da70f JA |
436 | if (work->flags & IO_WQ_WORK_CB) |
437 | work->func(&work); | |
438 | ||
fcb323cc JA |
439 | if ((work->flags & IO_WQ_WORK_NEEDS_FILES) && |
440 | current->files != work->files) { | |
441 | task_lock(current); | |
442 | current->files = work->files; | |
443 | task_unlock(current); | |
444 | } | |
771b53d0 JA |
445 | if ((work->flags & IO_WQ_WORK_NEEDS_USER) && !worker->mm && |
446 | wq->mm && mmget_not_zero(wq->mm)) { | |
447 | use_mm(wq->mm); | |
448 | set_fs(USER_DS); | |
449 | worker->mm = wq->mm; | |
450 | } | |
181e448d JA |
451 | if (!worker->creds) |
452 | worker->creds = override_creds(wq->creds); | |
771b53d0 JA |
453 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) |
454 | work->flags |= IO_WQ_WORK_CANCEL; | |
455 | if (worker->mm) | |
456 | work->flags |= IO_WQ_WORK_HAS_MM; | |
457 | ||
7d723065 JA |
458 | if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) { |
459 | put_work = work; | |
460 | wq->get_work(work); | |
461 | } | |
462 | ||
771b53d0 JA |
463 | old_work = work; |
464 | work->func(&work); | |
465 | ||
36c2f922 | 466 | spin_lock_irq(&worker->lock); |
771b53d0 | 467 | worker->cur_work = NULL; |
36c2f922 JA |
468 | spin_unlock_irq(&worker->lock); |
469 | ||
470 | spin_lock_irq(&wqe->lock); | |
471 | ||
771b53d0 JA |
472 | if (hash != -1U) { |
473 | wqe->hash_map &= ~BIT_ULL(hash); | |
474 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
475 | } | |
476 | if (work && work != old_work) { | |
477 | spin_unlock_irq(&wqe->lock); | |
7d723065 JA |
478 | |
479 | if (put_work && wq->put_work) { | |
480 | wq->put_work(put_work); | |
481 | put_work = NULL; | |
482 | } | |
483 | ||
771b53d0 JA |
484 | /* dependent work not hashed */ |
485 | hash = -1U; | |
486 | goto next; | |
487 | } | |
488 | } while (1); | |
489 | } | |
490 | ||
e995d512 JA |
491 | static inline void io_worker_spin_for_work(struct io_wqe *wqe) |
492 | { | |
493 | int i = 0; | |
494 | ||
495 | while (++i < 1000) { | |
496 | if (io_wqe_run_queue(wqe)) | |
497 | break; | |
498 | if (need_resched()) | |
499 | break; | |
500 | cpu_relax(); | |
501 | } | |
502 | } | |
503 | ||
771b53d0 JA |
504 | static int io_wqe_worker(void *data) |
505 | { | |
506 | struct io_worker *worker = data; | |
507 | struct io_wqe *wqe = worker->wqe; | |
508 | struct io_wq *wq = wqe->wq; | |
e995d512 | 509 | bool did_work; |
771b53d0 JA |
510 | |
511 | io_worker_start(wqe, worker); | |
512 | ||
e995d512 | 513 | did_work = false; |
771b53d0 | 514 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
506d95ff | 515 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 JA |
516 | loop: |
517 | if (did_work) | |
518 | io_worker_spin_for_work(wqe); | |
771b53d0 JA |
519 | spin_lock_irq(&wqe->lock); |
520 | if (io_wqe_run_queue(wqe)) { | |
521 | __set_current_state(TASK_RUNNING); | |
522 | io_worker_handle_work(worker); | |
e995d512 JA |
523 | did_work = true; |
524 | goto loop; | |
771b53d0 | 525 | } |
e995d512 | 526 | did_work = false; |
771b53d0 JA |
527 | /* drops the lock on success, retry */ |
528 | if (__io_worker_idle(wqe, worker)) { | |
529 | __release(&wqe->lock); | |
e995d512 | 530 | goto loop; |
771b53d0 JA |
531 | } |
532 | spin_unlock_irq(&wqe->lock); | |
533 | if (signal_pending(current)) | |
534 | flush_signals(current); | |
535 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
536 | continue; | |
537 | /* timed out, exit unless we're the fixed worker */ | |
538 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
539 | !(worker->flags & IO_WORKER_F_FIXED)) | |
540 | break; | |
541 | } | |
542 | ||
771b53d0 JA |
543 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
544 | spin_lock_irq(&wqe->lock); | |
6206f0e1 | 545 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
546 | io_worker_handle_work(worker); |
547 | else | |
548 | spin_unlock_irq(&wqe->lock); | |
549 | } | |
550 | ||
551 | io_worker_exit(worker); | |
552 | return 0; | |
553 | } | |
554 | ||
771b53d0 JA |
555 | /* |
556 | * Called when a worker is scheduled in. Mark us as currently running. | |
557 | */ | |
558 | void io_wq_worker_running(struct task_struct *tsk) | |
559 | { | |
560 | struct io_worker *worker = kthread_data(tsk); | |
561 | struct io_wqe *wqe = worker->wqe; | |
562 | ||
563 | if (!(worker->flags & IO_WORKER_F_UP)) | |
564 | return; | |
565 | if (worker->flags & IO_WORKER_F_RUNNING) | |
566 | return; | |
567 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 568 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
569 | } |
570 | ||
571 | /* | |
572 | * Called when worker is going to sleep. If there are no workers currently | |
573 | * running and we have work pending, wake up a free one or have the manager | |
574 | * set one up. | |
575 | */ | |
576 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
577 | { | |
578 | struct io_worker *worker = kthread_data(tsk); | |
579 | struct io_wqe *wqe = worker->wqe; | |
580 | ||
581 | if (!(worker->flags & IO_WORKER_F_UP)) | |
582 | return; | |
583 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
584 | return; | |
585 | ||
586 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
587 | ||
588 | spin_lock_irq(&wqe->lock); | |
c5def4ab | 589 | io_wqe_dec_running(wqe, worker); |
771b53d0 JA |
590 | spin_unlock_irq(&wqe->lock); |
591 | } | |
592 | ||
b60fda60 | 593 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 594 | { |
c5def4ab | 595 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
596 | struct io_worker *worker; |
597 | ||
ad6e005c | 598 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 599 | if (!worker) |
b60fda60 | 600 | return false; |
771b53d0 JA |
601 | |
602 | refcount_set(&worker->ref, 1); | |
603 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 604 | worker->wqe = wqe; |
36c2f922 | 605 | spin_lock_init(&worker->lock); |
771b53d0 JA |
606 | |
607 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 608 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
609 | if (IS_ERR(worker->task)) { |
610 | kfree(worker); | |
b60fda60 | 611 | return false; |
771b53d0 JA |
612 | } |
613 | ||
614 | spin_lock_irq(&wqe->lock); | |
021d1cdd | 615 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 616 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 617 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
618 | if (index == IO_WQ_ACCT_BOUND) |
619 | worker->flags |= IO_WORKER_F_BOUND; | |
620 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 621 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 622 | acct->nr_workers++; |
771b53d0 JA |
623 | spin_unlock_irq(&wqe->lock); |
624 | ||
c5def4ab JA |
625 | if (index == IO_WQ_ACCT_UNBOUND) |
626 | atomic_inc(&wq->user->processes); | |
627 | ||
771b53d0 | 628 | wake_up_process(worker->task); |
b60fda60 | 629 | return true; |
771b53d0 JA |
630 | } |
631 | ||
c5def4ab | 632 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
633 | __must_hold(wqe->lock) |
634 | { | |
c5def4ab | 635 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 636 | |
c5def4ab | 637 | /* if we have available workers or no work, no need */ |
021d1cdd | 638 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
639 | return false; |
640 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
641 | } |
642 | ||
643 | /* | |
644 | * Manager thread. Tasked with creating new workers, if we need them. | |
645 | */ | |
646 | static int io_wq_manager(void *data) | |
647 | { | |
648 | struct io_wq *wq = data; | |
3fc50ab5 JH |
649 | int workers_to_create = num_possible_nodes(); |
650 | int node; | |
771b53d0 | 651 | |
b60fda60 | 652 | /* create fixed workers */ |
3fc50ab5 JH |
653 | refcount_set(&wq->refs, workers_to_create); |
654 | for_each_node(node) { | |
655 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) | |
656 | goto err; | |
657 | workers_to_create--; | |
b60fda60 | 658 | } |
771b53d0 | 659 | |
b60fda60 JA |
660 | complete(&wq->done); |
661 | ||
662 | while (!kthread_should_stop()) { | |
3fc50ab5 JH |
663 | for_each_node(node) { |
664 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 665 | bool fork_worker[2] = { false, false }; |
771b53d0 JA |
666 | |
667 | spin_lock_irq(&wqe->lock); | |
c5def4ab JA |
668 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
669 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
670 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
671 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
771b53d0 | 672 | spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
673 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
674 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
675 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
676 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
677 | } |
678 | set_current_state(TASK_INTERRUPTIBLE); | |
679 | schedule_timeout(HZ); | |
680 | } | |
681 | ||
682 | return 0; | |
b60fda60 JA |
683 | err: |
684 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
685 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 686 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
687 | complete(&wq->done); |
688 | return 0; | |
771b53d0 JA |
689 | } |
690 | ||
c5def4ab JA |
691 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
692 | struct io_wq_work *work) | |
693 | { | |
694 | bool free_worker; | |
695 | ||
696 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
697 | return true; | |
698 | if (atomic_read(&acct->nr_running)) | |
699 | return true; | |
700 | ||
701 | rcu_read_lock(); | |
021d1cdd | 702 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
703 | rcu_read_unlock(); |
704 | if (free_worker) | |
705 | return true; | |
706 | ||
707 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
708 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
709 | return false; | |
710 | ||
711 | return true; | |
712 | } | |
713 | ||
771b53d0 JA |
714 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
715 | { | |
c5def4ab | 716 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
771b53d0 JA |
717 | unsigned long flags; |
718 | ||
c5def4ab JA |
719 | /* |
720 | * Do early check to see if we need a new unbound worker, and if we do, | |
721 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
722 | * gap between this check and incrementing the value, but that's OK. | |
723 | * It's close enough to not be an issue, fork() has the same delay. | |
724 | */ | |
725 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
726 | work->flags |= IO_WQ_WORK_CANCEL; | |
727 | work->func(&work); | |
728 | return; | |
729 | } | |
730 | ||
771b53d0 | 731 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 | 732 | wq_list_add_tail(&work->list, &wqe->work_list); |
771b53d0 JA |
733 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
734 | spin_unlock_irqrestore(&wqe->lock, flags); | |
735 | ||
c5def4ab JA |
736 | if (!atomic_read(&acct->nr_running)) |
737 | io_wqe_wake_worker(wqe, acct); | |
771b53d0 JA |
738 | } |
739 | ||
740 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
741 | { | |
742 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
743 | ||
744 | io_wqe_enqueue(wqe, work); | |
745 | } | |
746 | ||
747 | /* | |
748 | * Enqueue work, hashed by some key. Work items that hash to the same value | |
749 | * will not be done in parallel. Used to limit concurrent writes, generally | |
750 | * hashed by inode. | |
751 | */ | |
752 | void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val) | |
753 | { | |
754 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
755 | unsigned bit; | |
756 | ||
757 | ||
758 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
759 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
760 | io_wqe_enqueue(wqe, work); | |
761 | } | |
762 | ||
763 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
764 | { | |
765 | send_sig(SIGINT, worker->task, 1); | |
766 | return false; | |
767 | } | |
768 | ||
769 | /* | |
770 | * Iterate the passed in list and call the specific function for each | |
771 | * worker that isn't exiting | |
772 | */ | |
773 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
774 | bool (*func)(struct io_worker *, void *), |
775 | void *data) | |
776 | { | |
771b53d0 JA |
777 | struct io_worker *worker; |
778 | bool ret = false; | |
779 | ||
e61df66c | 780 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 JA |
781 | if (io_worker_get(worker)) { |
782 | ret = func(worker, data); | |
783 | io_worker_release(worker); | |
784 | if (ret) | |
785 | break; | |
786 | } | |
787 | } | |
e61df66c | 788 | |
771b53d0 JA |
789 | return ret; |
790 | } | |
791 | ||
792 | void io_wq_cancel_all(struct io_wq *wq) | |
793 | { | |
3fc50ab5 | 794 | int node; |
771b53d0 JA |
795 | |
796 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
797 | ||
771b53d0 | 798 | rcu_read_lock(); |
3fc50ab5 JH |
799 | for_each_node(node) { |
800 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 801 | |
e61df66c | 802 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
803 | } |
804 | rcu_read_unlock(); | |
805 | } | |
806 | ||
62755e35 JA |
807 | struct io_cb_cancel_data { |
808 | struct io_wqe *wqe; | |
809 | work_cancel_fn *cancel; | |
810 | void *caller_data; | |
811 | }; | |
812 | ||
813 | static bool io_work_cancel(struct io_worker *worker, void *cancel_data) | |
814 | { | |
815 | struct io_cb_cancel_data *data = cancel_data; | |
6f72653e | 816 | unsigned long flags; |
62755e35 JA |
817 | bool ret = false; |
818 | ||
819 | /* | |
820 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 821 | * may dereference the passed in work. |
62755e35 | 822 | */ |
36c2f922 | 823 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 JA |
824 | if (worker->cur_work && |
825 | data->cancel(worker->cur_work, data->caller_data)) { | |
826 | send_sig(SIGINT, worker->task, 1); | |
827 | ret = true; | |
828 | } | |
36c2f922 | 829 | spin_unlock_irqrestore(&worker->lock, flags); |
62755e35 JA |
830 | |
831 | return ret; | |
832 | } | |
833 | ||
834 | static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe, | |
835 | work_cancel_fn *cancel, | |
836 | void *cancel_data) | |
837 | { | |
838 | struct io_cb_cancel_data data = { | |
839 | .wqe = wqe, | |
840 | .cancel = cancel, | |
841 | .caller_data = cancel_data, | |
842 | }; | |
6206f0e1 | 843 | struct io_wq_work_node *node, *prev; |
62755e35 | 844 | struct io_wq_work *work; |
6f72653e | 845 | unsigned long flags; |
62755e35 JA |
846 | bool found = false; |
847 | ||
6f72653e | 848 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
849 | wq_list_for_each(node, prev, &wqe->work_list) { |
850 | work = container_of(node, struct io_wq_work, list); | |
851 | ||
62755e35 | 852 | if (cancel(work, cancel_data)) { |
6206f0e1 | 853 | wq_node_del(&wqe->work_list, node, prev); |
62755e35 JA |
854 | found = true; |
855 | break; | |
856 | } | |
857 | } | |
6f72653e | 858 | spin_unlock_irqrestore(&wqe->lock, flags); |
62755e35 JA |
859 | |
860 | if (found) { | |
861 | work->flags |= IO_WQ_WORK_CANCEL; | |
862 | work->func(&work); | |
863 | return IO_WQ_CANCEL_OK; | |
864 | } | |
865 | ||
866 | rcu_read_lock(); | |
e61df66c | 867 | found = io_wq_for_each_worker(wqe, io_work_cancel, &data); |
62755e35 JA |
868 | rcu_read_unlock(); |
869 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
870 | } | |
871 | ||
872 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, | |
873 | void *data) | |
874 | { | |
875 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 876 | int node; |
62755e35 | 877 | |
3fc50ab5 JH |
878 | for_each_node(node) { |
879 | struct io_wqe *wqe = wq->wqes[node]; | |
62755e35 JA |
880 | |
881 | ret = io_wqe_cancel_cb_work(wqe, cancel, data); | |
882 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
883 | break; | |
884 | } | |
885 | ||
886 | return ret; | |
887 | } | |
888 | ||
771b53d0 JA |
889 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
890 | { | |
891 | struct io_wq_work *work = data; | |
36c2f922 JA |
892 | unsigned long flags; |
893 | bool ret = false; | |
771b53d0 | 894 | |
36c2f922 JA |
895 | if (worker->cur_work != work) |
896 | return false; | |
897 | ||
898 | spin_lock_irqsave(&worker->lock, flags); | |
771b53d0 JA |
899 | if (worker->cur_work == work) { |
900 | send_sig(SIGINT, worker->task, 1); | |
36c2f922 | 901 | ret = true; |
771b53d0 | 902 | } |
36c2f922 | 903 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 904 | |
36c2f922 | 905 | return ret; |
771b53d0 JA |
906 | } |
907 | ||
908 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, | |
909 | struct io_wq_work *cwork) | |
910 | { | |
6206f0e1 | 911 | struct io_wq_work_node *node, *prev; |
771b53d0 | 912 | struct io_wq_work *work; |
6f72653e | 913 | unsigned long flags; |
771b53d0 JA |
914 | bool found = false; |
915 | ||
916 | cwork->flags |= IO_WQ_WORK_CANCEL; | |
917 | ||
918 | /* | |
919 | * First check pending list, if we're lucky we can just remove it | |
920 | * from there. CANCEL_OK means that the work is returned as-new, | |
921 | * no completion will be posted for it. | |
922 | */ | |
6f72653e | 923 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
924 | wq_list_for_each(node, prev, &wqe->work_list) { |
925 | work = container_of(node, struct io_wq_work, list); | |
926 | ||
771b53d0 | 927 | if (work == cwork) { |
6206f0e1 | 928 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
929 | found = true; |
930 | break; | |
931 | } | |
932 | } | |
6f72653e | 933 | spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 JA |
934 | |
935 | if (found) { | |
936 | work->flags |= IO_WQ_WORK_CANCEL; | |
937 | work->func(&work); | |
938 | return IO_WQ_CANCEL_OK; | |
939 | } | |
940 | ||
941 | /* | |
942 | * Now check if a free (going busy) or busy worker has the work | |
943 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
d195a66e | 944 | * as an indication that we attempt to signal cancellation. The |
771b53d0 JA |
945 | * completion will run normally in this case. |
946 | */ | |
947 | rcu_read_lock(); | |
e61df66c | 948 | found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, cwork); |
771b53d0 JA |
949 | rcu_read_unlock(); |
950 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
951 | } | |
952 | ||
953 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
954 | { | |
955 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 956 | int node; |
771b53d0 | 957 | |
3fc50ab5 JH |
958 | for_each_node(node) { |
959 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
960 | |
961 | ret = io_wqe_cancel_work(wqe, cwork); | |
962 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
963 | break; | |
964 | } | |
965 | ||
966 | return ret; | |
967 | } | |
968 | ||
969 | struct io_wq_flush_data { | |
970 | struct io_wq_work work; | |
971 | struct completion done; | |
972 | }; | |
973 | ||
974 | static void io_wq_flush_func(struct io_wq_work **workptr) | |
975 | { | |
976 | struct io_wq_work *work = *workptr; | |
977 | struct io_wq_flush_data *data; | |
978 | ||
979 | data = container_of(work, struct io_wq_flush_data, work); | |
980 | complete(&data->done); | |
981 | } | |
982 | ||
983 | /* | |
984 | * Doesn't wait for previously queued work to finish. When this completes, | |
985 | * it just means that previously queued work was started. | |
986 | */ | |
987 | void io_wq_flush(struct io_wq *wq) | |
988 | { | |
989 | struct io_wq_flush_data data; | |
3fc50ab5 | 990 | int node; |
771b53d0 | 991 | |
3fc50ab5 JH |
992 | for_each_node(node) { |
993 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
994 | |
995 | init_completion(&data.done); | |
996 | INIT_IO_WORK(&data.work, io_wq_flush_func); | |
7d723065 | 997 | data.work.flags |= IO_WQ_WORK_INTERNAL; |
771b53d0 JA |
998 | io_wqe_enqueue(wqe, &data.work); |
999 | wait_for_completion(&data.done); | |
1000 | } | |
1001 | } | |
1002 | ||
576a347b | 1003 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1004 | { |
3fc50ab5 | 1005 | int ret = -ENOMEM, node; |
771b53d0 JA |
1006 | struct io_wq *wq; |
1007 | ||
ad6e005c | 1008 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1009 | if (!wq) |
1010 | return ERR_PTR(-ENOMEM); | |
1011 | ||
3fc50ab5 | 1012 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1013 | if (!wq->wqes) { |
1014 | kfree(wq); | |
1015 | return ERR_PTR(-ENOMEM); | |
1016 | } | |
1017 | ||
576a347b JA |
1018 | wq->get_work = data->get_work; |
1019 | wq->put_work = data->put_work; | |
7d723065 | 1020 | |
c5def4ab | 1021 | /* caller must already hold a reference to this */ |
576a347b | 1022 | wq->user = data->user; |
181e448d | 1023 | wq->creds = data->creds; |
c5def4ab | 1024 | |
3fc50ab5 | 1025 | for_each_node(node) { |
771b53d0 JA |
1026 | struct io_wqe *wqe; |
1027 | ||
ad6e005c | 1028 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node); |
771b53d0 | 1029 | if (!wqe) |
3fc50ab5 JH |
1030 | goto err; |
1031 | wq->wqes[node] = wqe; | |
771b53d0 | 1032 | wqe->node = node; |
c5def4ab JA |
1033 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1034 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1035 | if (wq->user) { |
c5def4ab JA |
1036 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1037 | task_rlimit(current, RLIMIT_NPROC); | |
1038 | } | |
1039 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 JA |
1040 | wqe->node = node; |
1041 | wqe->wq = wq; | |
1042 | spin_lock_init(&wqe->lock); | |
6206f0e1 | 1043 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1044 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1045 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1046 | } |
1047 | ||
1048 | init_completion(&wq->done); | |
1049 | ||
771b53d0 | 1050 | /* caller must have already done mmgrab() on this mm */ |
576a347b | 1051 | wq->mm = data->mm; |
771b53d0 JA |
1052 | |
1053 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); | |
1054 | if (!IS_ERR(wq->manager)) { | |
1055 | wake_up_process(wq->manager); | |
b60fda60 JA |
1056 | wait_for_completion(&wq->done); |
1057 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1058 | ret = -ENOMEM; | |
1059 | goto err; | |
1060 | } | |
1061 | reinit_completion(&wq->done); | |
771b53d0 JA |
1062 | return wq; |
1063 | } | |
1064 | ||
1065 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1066 | complete(&wq->done); |
b60fda60 | 1067 | err: |
3fc50ab5 JH |
1068 | for_each_node(node) |
1069 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1070 | kfree(wq->wqes); |
1071 | kfree(wq); | |
771b53d0 JA |
1072 | return ERR_PTR(ret); |
1073 | } | |
1074 | ||
1075 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
1076 | { | |
1077 | wake_up_process(worker->task); | |
1078 | return false; | |
1079 | } | |
1080 | ||
1081 | void io_wq_destroy(struct io_wq *wq) | |
1082 | { | |
3fc50ab5 | 1083 | int node; |
771b53d0 | 1084 | |
b60fda60 JA |
1085 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1086 | if (wq->manager) | |
771b53d0 | 1087 | kthread_stop(wq->manager); |
771b53d0 JA |
1088 | |
1089 | rcu_read_lock(); | |
3fc50ab5 JH |
1090 | for_each_node(node) |
1091 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1092 | rcu_read_unlock(); |
1093 | ||
1094 | wait_for_completion(&wq->done); | |
1095 | ||
3fc50ab5 JH |
1096 | for_each_node(node) |
1097 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1098 | kfree(wq->wqes); |
1099 | kfree(wq); | |
1100 | } |