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); | |
0c9d5ccd JA |
459 | /* |
460 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable work, | |
461 | * the worker function will do the right thing. | |
462 | */ | |
771b53d0 JA |
463 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) |
464 | work->flags |= IO_WQ_WORK_CANCEL; | |
465 | if (worker->mm) | |
466 | work->flags |= IO_WQ_WORK_HAS_MM; | |
467 | ||
7d723065 JA |
468 | if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) { |
469 | put_work = work; | |
470 | wq->get_work(work); | |
471 | } | |
472 | ||
771b53d0 JA |
473 | old_work = work; |
474 | work->func(&work); | |
475 | ||
36c2f922 | 476 | spin_lock_irq(&worker->lock); |
771b53d0 | 477 | worker->cur_work = NULL; |
36c2f922 JA |
478 | spin_unlock_irq(&worker->lock); |
479 | ||
480 | spin_lock_irq(&wqe->lock); | |
481 | ||
771b53d0 JA |
482 | if (hash != -1U) { |
483 | wqe->hash_map &= ~BIT_ULL(hash); | |
484 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
485 | } | |
486 | if (work && work != old_work) { | |
487 | spin_unlock_irq(&wqe->lock); | |
7d723065 JA |
488 | |
489 | if (put_work && wq->put_work) { | |
490 | wq->put_work(put_work); | |
491 | put_work = NULL; | |
492 | } | |
493 | ||
771b53d0 JA |
494 | /* dependent work not hashed */ |
495 | hash = -1U; | |
496 | goto next; | |
497 | } | |
498 | } while (1); | |
499 | } | |
500 | ||
e995d512 JA |
501 | static inline void io_worker_spin_for_work(struct io_wqe *wqe) |
502 | { | |
503 | int i = 0; | |
504 | ||
505 | while (++i < 1000) { | |
506 | if (io_wqe_run_queue(wqe)) | |
507 | break; | |
508 | if (need_resched()) | |
509 | break; | |
510 | cpu_relax(); | |
511 | } | |
512 | } | |
513 | ||
771b53d0 JA |
514 | static int io_wqe_worker(void *data) |
515 | { | |
516 | struct io_worker *worker = data; | |
517 | struct io_wqe *wqe = worker->wqe; | |
518 | struct io_wq *wq = wqe->wq; | |
e995d512 | 519 | bool did_work; |
771b53d0 JA |
520 | |
521 | io_worker_start(wqe, worker); | |
522 | ||
e995d512 | 523 | did_work = false; |
771b53d0 | 524 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
506d95ff | 525 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 JA |
526 | loop: |
527 | if (did_work) | |
528 | io_worker_spin_for_work(wqe); | |
771b53d0 JA |
529 | spin_lock_irq(&wqe->lock); |
530 | if (io_wqe_run_queue(wqe)) { | |
531 | __set_current_state(TASK_RUNNING); | |
532 | io_worker_handle_work(worker); | |
e995d512 JA |
533 | did_work = true; |
534 | goto loop; | |
771b53d0 | 535 | } |
e995d512 | 536 | did_work = false; |
771b53d0 JA |
537 | /* drops the lock on success, retry */ |
538 | if (__io_worker_idle(wqe, worker)) { | |
539 | __release(&wqe->lock); | |
e995d512 | 540 | goto loop; |
771b53d0 JA |
541 | } |
542 | spin_unlock_irq(&wqe->lock); | |
543 | if (signal_pending(current)) | |
544 | flush_signals(current); | |
545 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
546 | continue; | |
547 | /* timed out, exit unless we're the fixed worker */ | |
548 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
549 | !(worker->flags & IO_WORKER_F_FIXED)) | |
550 | break; | |
551 | } | |
552 | ||
771b53d0 JA |
553 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
554 | spin_lock_irq(&wqe->lock); | |
6206f0e1 | 555 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
556 | io_worker_handle_work(worker); |
557 | else | |
558 | spin_unlock_irq(&wqe->lock); | |
559 | } | |
560 | ||
561 | io_worker_exit(worker); | |
562 | return 0; | |
563 | } | |
564 | ||
771b53d0 JA |
565 | /* |
566 | * Called when a worker is scheduled in. Mark us as currently running. | |
567 | */ | |
568 | void io_wq_worker_running(struct task_struct *tsk) | |
569 | { | |
570 | struct io_worker *worker = kthread_data(tsk); | |
571 | struct io_wqe *wqe = worker->wqe; | |
572 | ||
573 | if (!(worker->flags & IO_WORKER_F_UP)) | |
574 | return; | |
575 | if (worker->flags & IO_WORKER_F_RUNNING) | |
576 | return; | |
577 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 578 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
579 | } |
580 | ||
581 | /* | |
582 | * Called when worker is going to sleep. If there are no workers currently | |
583 | * running and we have work pending, wake up a free one or have the manager | |
584 | * set one up. | |
585 | */ | |
586 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
587 | { | |
588 | struct io_worker *worker = kthread_data(tsk); | |
589 | struct io_wqe *wqe = worker->wqe; | |
590 | ||
591 | if (!(worker->flags & IO_WORKER_F_UP)) | |
592 | return; | |
593 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
594 | return; | |
595 | ||
596 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
597 | ||
598 | spin_lock_irq(&wqe->lock); | |
c5def4ab | 599 | io_wqe_dec_running(wqe, worker); |
771b53d0 JA |
600 | spin_unlock_irq(&wqe->lock); |
601 | } | |
602 | ||
b60fda60 | 603 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 604 | { |
c5def4ab | 605 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
606 | struct io_worker *worker; |
607 | ||
ad6e005c | 608 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 609 | if (!worker) |
b60fda60 | 610 | return false; |
771b53d0 JA |
611 | |
612 | refcount_set(&worker->ref, 1); | |
613 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 614 | worker->wqe = wqe; |
36c2f922 | 615 | spin_lock_init(&worker->lock); |
771b53d0 JA |
616 | |
617 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 618 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
619 | if (IS_ERR(worker->task)) { |
620 | kfree(worker); | |
b60fda60 | 621 | return false; |
771b53d0 JA |
622 | } |
623 | ||
624 | spin_lock_irq(&wqe->lock); | |
021d1cdd | 625 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 626 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 627 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
628 | if (index == IO_WQ_ACCT_BOUND) |
629 | worker->flags |= IO_WORKER_F_BOUND; | |
630 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 631 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 632 | acct->nr_workers++; |
771b53d0 JA |
633 | spin_unlock_irq(&wqe->lock); |
634 | ||
c5def4ab JA |
635 | if (index == IO_WQ_ACCT_UNBOUND) |
636 | atomic_inc(&wq->user->processes); | |
637 | ||
771b53d0 | 638 | wake_up_process(worker->task); |
b60fda60 | 639 | return true; |
771b53d0 JA |
640 | } |
641 | ||
c5def4ab | 642 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
643 | __must_hold(wqe->lock) |
644 | { | |
c5def4ab | 645 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 646 | |
c5def4ab | 647 | /* if we have available workers or no work, no need */ |
021d1cdd | 648 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
649 | return false; |
650 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
651 | } |
652 | ||
653 | /* | |
654 | * Manager thread. Tasked with creating new workers, if we need them. | |
655 | */ | |
656 | static int io_wq_manager(void *data) | |
657 | { | |
658 | struct io_wq *wq = data; | |
3fc50ab5 JH |
659 | int workers_to_create = num_possible_nodes(); |
660 | int node; | |
771b53d0 | 661 | |
b60fda60 | 662 | /* create fixed workers */ |
3fc50ab5 JH |
663 | refcount_set(&wq->refs, workers_to_create); |
664 | for_each_node(node) { | |
665 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) | |
666 | goto err; | |
667 | workers_to_create--; | |
b60fda60 | 668 | } |
771b53d0 | 669 | |
b60fda60 JA |
670 | complete(&wq->done); |
671 | ||
672 | while (!kthread_should_stop()) { | |
3fc50ab5 JH |
673 | for_each_node(node) { |
674 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 675 | bool fork_worker[2] = { false, false }; |
771b53d0 JA |
676 | |
677 | spin_lock_irq(&wqe->lock); | |
c5def4ab JA |
678 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
679 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
680 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
681 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
771b53d0 | 682 | spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
683 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
684 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
685 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
686 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
687 | } |
688 | set_current_state(TASK_INTERRUPTIBLE); | |
689 | schedule_timeout(HZ); | |
690 | } | |
691 | ||
692 | return 0; | |
b60fda60 JA |
693 | err: |
694 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
695 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 696 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
697 | complete(&wq->done); |
698 | return 0; | |
771b53d0 JA |
699 | } |
700 | ||
c5def4ab JA |
701 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
702 | struct io_wq_work *work) | |
703 | { | |
704 | bool free_worker; | |
705 | ||
706 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
707 | return true; | |
708 | if (atomic_read(&acct->nr_running)) | |
709 | return true; | |
710 | ||
711 | rcu_read_lock(); | |
021d1cdd | 712 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
713 | rcu_read_unlock(); |
714 | if (free_worker) | |
715 | return true; | |
716 | ||
717 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
718 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
719 | return false; | |
720 | ||
721 | return true; | |
722 | } | |
723 | ||
771b53d0 JA |
724 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
725 | { | |
c5def4ab | 726 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 727 | int work_flags; |
771b53d0 JA |
728 | unsigned long flags; |
729 | ||
c5def4ab JA |
730 | /* |
731 | * Do early check to see if we need a new unbound worker, and if we do, | |
732 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
733 | * gap between this check and incrementing the value, but that's OK. | |
734 | * It's close enough to not be an issue, fork() has the same delay. | |
735 | */ | |
736 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
737 | work->flags |= IO_WQ_WORK_CANCEL; | |
738 | work->func(&work); | |
739 | return; | |
740 | } | |
741 | ||
895e2ca0 | 742 | work_flags = work->flags; |
771b53d0 | 743 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 | 744 | wq_list_add_tail(&work->list, &wqe->work_list); |
771b53d0 JA |
745 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
746 | spin_unlock_irqrestore(&wqe->lock, flags); | |
747 | ||
895e2ca0 JA |
748 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
749 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 750 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
751 | } |
752 | ||
753 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
754 | { | |
755 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
756 | ||
757 | io_wqe_enqueue(wqe, work); | |
758 | } | |
759 | ||
760 | /* | |
761 | * Enqueue work, hashed by some key. Work items that hash to the same value | |
762 | * will not be done in parallel. Used to limit concurrent writes, generally | |
763 | * hashed by inode. | |
764 | */ | |
765 | void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val) | |
766 | { | |
767 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
768 | unsigned bit; | |
769 | ||
770 | ||
771 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
772 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
773 | io_wqe_enqueue(wqe, work); | |
774 | } | |
775 | ||
776 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
777 | { | |
778 | send_sig(SIGINT, worker->task, 1); | |
779 | return false; | |
780 | } | |
781 | ||
782 | /* | |
783 | * Iterate the passed in list and call the specific function for each | |
784 | * worker that isn't exiting | |
785 | */ | |
786 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
787 | bool (*func)(struct io_worker *, void *), |
788 | void *data) | |
789 | { | |
771b53d0 JA |
790 | struct io_worker *worker; |
791 | bool ret = false; | |
792 | ||
e61df66c | 793 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 JA |
794 | if (io_worker_get(worker)) { |
795 | ret = func(worker, data); | |
796 | io_worker_release(worker); | |
797 | if (ret) | |
798 | break; | |
799 | } | |
800 | } | |
e61df66c | 801 | |
771b53d0 JA |
802 | return ret; |
803 | } | |
804 | ||
805 | void io_wq_cancel_all(struct io_wq *wq) | |
806 | { | |
3fc50ab5 | 807 | int node; |
771b53d0 JA |
808 | |
809 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
810 | ||
771b53d0 | 811 | rcu_read_lock(); |
3fc50ab5 JH |
812 | for_each_node(node) { |
813 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 814 | |
e61df66c | 815 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
816 | } |
817 | rcu_read_unlock(); | |
818 | } | |
819 | ||
62755e35 JA |
820 | struct io_cb_cancel_data { |
821 | struct io_wqe *wqe; | |
822 | work_cancel_fn *cancel; | |
823 | void *caller_data; | |
824 | }; | |
825 | ||
826 | static bool io_work_cancel(struct io_worker *worker, void *cancel_data) | |
827 | { | |
828 | struct io_cb_cancel_data *data = cancel_data; | |
6f72653e | 829 | unsigned long flags; |
62755e35 JA |
830 | bool ret = false; |
831 | ||
832 | /* | |
833 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 834 | * may dereference the passed in work. |
62755e35 | 835 | */ |
36c2f922 | 836 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 837 | if (worker->cur_work && |
0c9d5ccd | 838 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
62755e35 JA |
839 | data->cancel(worker->cur_work, data->caller_data)) { |
840 | send_sig(SIGINT, worker->task, 1); | |
841 | ret = true; | |
842 | } | |
36c2f922 | 843 | spin_unlock_irqrestore(&worker->lock, flags); |
62755e35 JA |
844 | |
845 | return ret; | |
846 | } | |
847 | ||
848 | static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe, | |
849 | work_cancel_fn *cancel, | |
850 | void *cancel_data) | |
851 | { | |
852 | struct io_cb_cancel_data data = { | |
853 | .wqe = wqe, | |
854 | .cancel = cancel, | |
855 | .caller_data = cancel_data, | |
856 | }; | |
6206f0e1 | 857 | struct io_wq_work_node *node, *prev; |
62755e35 | 858 | struct io_wq_work *work; |
6f72653e | 859 | unsigned long flags; |
62755e35 JA |
860 | bool found = false; |
861 | ||
6f72653e | 862 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
863 | wq_list_for_each(node, prev, &wqe->work_list) { |
864 | work = container_of(node, struct io_wq_work, list); | |
865 | ||
62755e35 | 866 | if (cancel(work, cancel_data)) { |
6206f0e1 | 867 | wq_node_del(&wqe->work_list, node, prev); |
62755e35 JA |
868 | found = true; |
869 | break; | |
870 | } | |
871 | } | |
6f72653e | 872 | spin_unlock_irqrestore(&wqe->lock, flags); |
62755e35 JA |
873 | |
874 | if (found) { | |
875 | work->flags |= IO_WQ_WORK_CANCEL; | |
876 | work->func(&work); | |
877 | return IO_WQ_CANCEL_OK; | |
878 | } | |
879 | ||
880 | rcu_read_lock(); | |
e61df66c | 881 | found = io_wq_for_each_worker(wqe, io_work_cancel, &data); |
62755e35 JA |
882 | rcu_read_unlock(); |
883 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
884 | } | |
885 | ||
886 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, | |
887 | void *data) | |
888 | { | |
889 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 890 | int node; |
62755e35 | 891 | |
3fc50ab5 JH |
892 | for_each_node(node) { |
893 | struct io_wqe *wqe = wq->wqes[node]; | |
62755e35 JA |
894 | |
895 | ret = io_wqe_cancel_cb_work(wqe, cancel, data); | |
896 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
897 | break; | |
898 | } | |
899 | ||
900 | return ret; | |
901 | } | |
902 | ||
771b53d0 JA |
903 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
904 | { | |
905 | struct io_wq_work *work = data; | |
36c2f922 JA |
906 | unsigned long flags; |
907 | bool ret = false; | |
771b53d0 | 908 | |
36c2f922 JA |
909 | if (worker->cur_work != work) |
910 | return false; | |
911 | ||
912 | spin_lock_irqsave(&worker->lock, flags); | |
0c9d5ccd JA |
913 | if (worker->cur_work == work && |
914 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL)) { | |
771b53d0 | 915 | send_sig(SIGINT, worker->task, 1); |
36c2f922 | 916 | ret = true; |
771b53d0 | 917 | } |
36c2f922 | 918 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 919 | |
36c2f922 | 920 | return ret; |
771b53d0 JA |
921 | } |
922 | ||
923 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, | |
924 | struct io_wq_work *cwork) | |
925 | { | |
6206f0e1 | 926 | struct io_wq_work_node *node, *prev; |
771b53d0 | 927 | struct io_wq_work *work; |
6f72653e | 928 | unsigned long flags; |
771b53d0 JA |
929 | bool found = false; |
930 | ||
931 | cwork->flags |= IO_WQ_WORK_CANCEL; | |
932 | ||
933 | /* | |
934 | * First check pending list, if we're lucky we can just remove it | |
935 | * from there. CANCEL_OK means that the work is returned as-new, | |
936 | * no completion will be posted for it. | |
937 | */ | |
6f72653e | 938 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
939 | wq_list_for_each(node, prev, &wqe->work_list) { |
940 | work = container_of(node, struct io_wq_work, list); | |
941 | ||
771b53d0 | 942 | if (work == cwork) { |
6206f0e1 | 943 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
944 | found = true; |
945 | break; | |
946 | } | |
947 | } | |
6f72653e | 948 | spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 JA |
949 | |
950 | if (found) { | |
951 | work->flags |= IO_WQ_WORK_CANCEL; | |
952 | work->func(&work); | |
953 | return IO_WQ_CANCEL_OK; | |
954 | } | |
955 | ||
956 | /* | |
957 | * Now check if a free (going busy) or busy worker has the work | |
958 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
d195a66e | 959 | * as an indication that we attempt to signal cancellation. The |
771b53d0 JA |
960 | * completion will run normally in this case. |
961 | */ | |
962 | rcu_read_lock(); | |
e61df66c | 963 | found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, cwork); |
771b53d0 JA |
964 | rcu_read_unlock(); |
965 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
966 | } | |
967 | ||
968 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
969 | { | |
970 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 971 | int node; |
771b53d0 | 972 | |
3fc50ab5 JH |
973 | for_each_node(node) { |
974 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
975 | |
976 | ret = io_wqe_cancel_work(wqe, cwork); | |
977 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
978 | break; | |
979 | } | |
980 | ||
981 | return ret; | |
982 | } | |
983 | ||
984 | struct io_wq_flush_data { | |
985 | struct io_wq_work work; | |
986 | struct completion done; | |
987 | }; | |
988 | ||
989 | static void io_wq_flush_func(struct io_wq_work **workptr) | |
990 | { | |
991 | struct io_wq_work *work = *workptr; | |
992 | struct io_wq_flush_data *data; | |
993 | ||
994 | data = container_of(work, struct io_wq_flush_data, work); | |
995 | complete(&data->done); | |
996 | } | |
997 | ||
998 | /* | |
999 | * Doesn't wait for previously queued work to finish. When this completes, | |
1000 | * it just means that previously queued work was started. | |
1001 | */ | |
1002 | void io_wq_flush(struct io_wq *wq) | |
1003 | { | |
1004 | struct io_wq_flush_data data; | |
3fc50ab5 | 1005 | int node; |
771b53d0 | 1006 | |
3fc50ab5 JH |
1007 | for_each_node(node) { |
1008 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
1009 | |
1010 | init_completion(&data.done); | |
1011 | INIT_IO_WORK(&data.work, io_wq_flush_func); | |
7d723065 | 1012 | data.work.flags |= IO_WQ_WORK_INTERNAL; |
771b53d0 JA |
1013 | io_wqe_enqueue(wqe, &data.work); |
1014 | wait_for_completion(&data.done); | |
1015 | } | |
1016 | } | |
1017 | ||
576a347b | 1018 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1019 | { |
3fc50ab5 | 1020 | int ret = -ENOMEM, node; |
771b53d0 JA |
1021 | struct io_wq *wq; |
1022 | ||
ad6e005c | 1023 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1024 | if (!wq) |
1025 | return ERR_PTR(-ENOMEM); | |
1026 | ||
3fc50ab5 | 1027 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1028 | if (!wq->wqes) { |
1029 | kfree(wq); | |
1030 | return ERR_PTR(-ENOMEM); | |
1031 | } | |
1032 | ||
576a347b JA |
1033 | wq->get_work = data->get_work; |
1034 | wq->put_work = data->put_work; | |
7d723065 | 1035 | |
c5def4ab | 1036 | /* caller must already hold a reference to this */ |
576a347b | 1037 | wq->user = data->user; |
181e448d | 1038 | wq->creds = data->creds; |
c5def4ab | 1039 | |
3fc50ab5 | 1040 | for_each_node(node) { |
771b53d0 JA |
1041 | struct io_wqe *wqe; |
1042 | ||
ad6e005c | 1043 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node); |
771b53d0 | 1044 | if (!wqe) |
3fc50ab5 JH |
1045 | goto err; |
1046 | wq->wqes[node] = wqe; | |
771b53d0 | 1047 | wqe->node = node; |
c5def4ab JA |
1048 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1049 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1050 | if (wq->user) { |
c5def4ab JA |
1051 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1052 | task_rlimit(current, RLIMIT_NPROC); | |
1053 | } | |
1054 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 JA |
1055 | wqe->node = node; |
1056 | wqe->wq = wq; | |
1057 | spin_lock_init(&wqe->lock); | |
6206f0e1 | 1058 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1059 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1060 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1061 | } |
1062 | ||
1063 | init_completion(&wq->done); | |
1064 | ||
771b53d0 | 1065 | /* caller must have already done mmgrab() on this mm */ |
576a347b | 1066 | wq->mm = data->mm; |
771b53d0 JA |
1067 | |
1068 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); | |
1069 | if (!IS_ERR(wq->manager)) { | |
1070 | wake_up_process(wq->manager); | |
b60fda60 JA |
1071 | wait_for_completion(&wq->done); |
1072 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1073 | ret = -ENOMEM; | |
1074 | goto err; | |
1075 | } | |
1076 | reinit_completion(&wq->done); | |
771b53d0 JA |
1077 | return wq; |
1078 | } | |
1079 | ||
1080 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1081 | complete(&wq->done); |
b60fda60 | 1082 | err: |
3fc50ab5 JH |
1083 | for_each_node(node) |
1084 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1085 | kfree(wq->wqes); |
1086 | kfree(wq); | |
771b53d0 JA |
1087 | return ERR_PTR(ret); |
1088 | } | |
1089 | ||
1090 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
1091 | { | |
1092 | wake_up_process(worker->task); | |
1093 | return false; | |
1094 | } | |
1095 | ||
1096 | void io_wq_destroy(struct io_wq *wq) | |
1097 | { | |
3fc50ab5 | 1098 | int node; |
771b53d0 | 1099 | |
b60fda60 JA |
1100 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1101 | if (wq->manager) | |
771b53d0 | 1102 | kthread_stop(wq->manager); |
771b53d0 JA |
1103 | |
1104 | rcu_read_lock(); | |
3fc50ab5 JH |
1105 | for_each_node(node) |
1106 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1107 | rcu_read_unlock(); |
1108 | ||
1109 | wait_for_completion(&wq->done); | |
1110 | ||
3fc50ab5 JH |
1111 | for_each_node(node) |
1112 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1113 | kfree(wq->wqes); |
1114 | kfree(wq); | |
1115 | } |