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; | |
cccf0ee8 JA |
59 | const struct cred *cur_creds; |
60 | const struct cred *saved_creds; | |
fcb323cc | 61 | struct files_struct *restore_files; |
771b53d0 JA |
62 | }; |
63 | ||
771b53d0 JA |
64 | #if BITS_PER_LONG == 64 |
65 | #define IO_WQ_HASH_ORDER 6 | |
66 | #else | |
67 | #define IO_WQ_HASH_ORDER 5 | |
68 | #endif | |
69 | ||
c5def4ab JA |
70 | struct io_wqe_acct { |
71 | unsigned nr_workers; | |
72 | unsigned max_workers; | |
73 | atomic_t nr_running; | |
74 | }; | |
75 | ||
76 | enum { | |
77 | IO_WQ_ACCT_BOUND, | |
78 | IO_WQ_ACCT_UNBOUND, | |
79 | }; | |
80 | ||
771b53d0 JA |
81 | /* |
82 | * Per-node worker thread pool | |
83 | */ | |
84 | struct io_wqe { | |
85 | struct { | |
86 | spinlock_t lock; | |
6206f0e1 | 87 | struct io_wq_work_list work_list; |
771b53d0 JA |
88 | unsigned long hash_map; |
89 | unsigned flags; | |
90 | } ____cacheline_aligned_in_smp; | |
91 | ||
92 | int node; | |
c5def4ab | 93 | struct io_wqe_acct acct[2]; |
771b53d0 | 94 | |
021d1cdd | 95 | struct hlist_nulls_head free_list; |
e61df66c | 96 | struct list_head all_list; |
771b53d0 JA |
97 | |
98 | struct io_wq *wq; | |
99 | }; | |
100 | ||
101 | /* | |
102 | * Per io_wq state | |
103 | */ | |
104 | struct io_wq { | |
105 | struct io_wqe **wqes; | |
106 | unsigned long state; | |
771b53d0 | 107 | |
7d723065 JA |
108 | get_work_fn *get_work; |
109 | put_work_fn *put_work; | |
110 | ||
771b53d0 | 111 | struct task_struct *manager; |
c5def4ab | 112 | struct user_struct *user; |
771b53d0 JA |
113 | refcount_t refs; |
114 | struct completion done; | |
848f7e18 JA |
115 | |
116 | refcount_t use_refs; | |
771b53d0 JA |
117 | }; |
118 | ||
771b53d0 JA |
119 | static bool io_worker_get(struct io_worker *worker) |
120 | { | |
121 | return refcount_inc_not_zero(&worker->ref); | |
122 | } | |
123 | ||
124 | static void io_worker_release(struct io_worker *worker) | |
125 | { | |
126 | if (refcount_dec_and_test(&worker->ref)) | |
127 | wake_up_process(worker->task); | |
128 | } | |
129 | ||
130 | /* | |
131 | * Note: drops the wqe->lock if returning true! The caller must re-acquire | |
132 | * the lock in that case. Some callers need to restart handling if this | |
133 | * happens, so we can't just re-acquire the lock on behalf of the caller. | |
134 | */ | |
135 | static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) | |
136 | { | |
fcb323cc JA |
137 | bool dropped_lock = false; |
138 | ||
cccf0ee8 JA |
139 | if (worker->saved_creds) { |
140 | revert_creds(worker->saved_creds); | |
141 | worker->cur_creds = worker->saved_creds = NULL; | |
181e448d JA |
142 | } |
143 | ||
fcb323cc JA |
144 | if (current->files != worker->restore_files) { |
145 | __acquire(&wqe->lock); | |
146 | spin_unlock_irq(&wqe->lock); | |
147 | dropped_lock = true; | |
148 | ||
149 | task_lock(current); | |
150 | current->files = worker->restore_files; | |
151 | task_unlock(current); | |
152 | } | |
153 | ||
771b53d0 JA |
154 | /* |
155 | * If we have an active mm, we need to drop the wq lock before unusing | |
156 | * it. If we do, return true and let the caller retry the idle loop. | |
157 | */ | |
158 | if (worker->mm) { | |
fcb323cc JA |
159 | if (!dropped_lock) { |
160 | __acquire(&wqe->lock); | |
161 | spin_unlock_irq(&wqe->lock); | |
162 | dropped_lock = true; | |
163 | } | |
771b53d0 JA |
164 | __set_current_state(TASK_RUNNING); |
165 | set_fs(KERNEL_DS); | |
166 | unuse_mm(worker->mm); | |
167 | mmput(worker->mm); | |
168 | worker->mm = NULL; | |
771b53d0 JA |
169 | } |
170 | ||
fcb323cc | 171 | return dropped_lock; |
771b53d0 JA |
172 | } |
173 | ||
c5def4ab JA |
174 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
175 | struct io_wq_work *work) | |
176 | { | |
177 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
178 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
179 | ||
180 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
181 | } | |
182 | ||
183 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe, | |
184 | struct io_worker *worker) | |
185 | { | |
186 | if (worker->flags & IO_WORKER_F_BOUND) | |
187 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
188 | ||
189 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
190 | } | |
191 | ||
771b53d0 JA |
192 | static void io_worker_exit(struct io_worker *worker) |
193 | { | |
194 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab JA |
195 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); |
196 | unsigned nr_workers; | |
771b53d0 JA |
197 | |
198 | /* | |
199 | * If we're not at zero, someone else is holding a brief reference | |
200 | * to the worker. Wait for that to go away. | |
201 | */ | |
202 | set_current_state(TASK_INTERRUPTIBLE); | |
203 | if (!refcount_dec_and_test(&worker->ref)) | |
204 | schedule(); | |
205 | __set_current_state(TASK_RUNNING); | |
206 | ||
207 | preempt_disable(); | |
208 | current->flags &= ~PF_IO_WORKER; | |
209 | if (worker->flags & IO_WORKER_F_RUNNING) | |
c5def4ab JA |
210 | atomic_dec(&acct->nr_running); |
211 | if (!(worker->flags & IO_WORKER_F_BOUND)) | |
212 | atomic_dec(&wqe->wq->user->processes); | |
771b53d0 JA |
213 | worker->flags = 0; |
214 | preempt_enable(); | |
215 | ||
216 | spin_lock_irq(&wqe->lock); | |
217 | hlist_nulls_del_rcu(&worker->nulls_node); | |
e61df66c | 218 | list_del_rcu(&worker->all_list); |
771b53d0 JA |
219 | if (__io_worker_unuse(wqe, worker)) { |
220 | __release(&wqe->lock); | |
221 | spin_lock_irq(&wqe->lock); | |
222 | } | |
c5def4ab JA |
223 | acct->nr_workers--; |
224 | nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers + | |
225 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers; | |
771b53d0 JA |
226 | spin_unlock_irq(&wqe->lock); |
227 | ||
228 | /* all workers gone, wq exit can proceed */ | |
c5def4ab | 229 | if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) |
771b53d0 JA |
230 | complete(&wqe->wq->done); |
231 | ||
364b05fd | 232 | kfree_rcu(worker, rcu); |
771b53d0 JA |
233 | } |
234 | ||
c5def4ab JA |
235 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
236 | __must_hold(wqe->lock) | |
237 | { | |
6206f0e1 JA |
238 | if (!wq_list_empty(&wqe->work_list) && |
239 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
240 | return true; |
241 | return false; | |
242 | } | |
243 | ||
244 | /* | |
245 | * Check head of free list for an available worker. If one isn't available, | |
246 | * caller must wake up the wq manager to create one. | |
247 | */ | |
248 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
249 | __must_hold(RCU) | |
250 | { | |
251 | struct hlist_nulls_node *n; | |
252 | struct io_worker *worker; | |
253 | ||
021d1cdd | 254 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
255 | if (is_a_nulls(n)) |
256 | return false; | |
257 | ||
258 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
259 | if (io_worker_get(worker)) { | |
506d95ff | 260 | wake_up_process(worker->task); |
c5def4ab JA |
261 | io_worker_release(worker); |
262 | return true; | |
263 | } | |
264 | ||
265 | return false; | |
266 | } | |
267 | ||
268 | /* | |
269 | * We need a worker. If we find a free one, we're good. If not, and we're | |
270 | * below the max number of workers, wake up the manager to create one. | |
271 | */ | |
272 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
273 | { | |
274 | bool ret; | |
275 | ||
276 | /* | |
277 | * Most likely an attempt to queue unbounded work on an io_wq that | |
278 | * wasn't setup with any unbounded workers. | |
279 | */ | |
280 | WARN_ON_ONCE(!acct->max_workers); | |
281 | ||
282 | rcu_read_lock(); | |
283 | ret = io_wqe_activate_free_worker(wqe); | |
284 | rcu_read_unlock(); | |
285 | ||
286 | if (!ret && acct->nr_workers < acct->max_workers) | |
287 | wake_up_process(wqe->wq->manager); | |
288 | } | |
289 | ||
290 | static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker) | |
291 | { | |
292 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
293 | ||
294 | atomic_inc(&acct->nr_running); | |
295 | } | |
296 | ||
297 | static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker) | |
298 | __must_hold(wqe->lock) | |
299 | { | |
300 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
301 | ||
302 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
303 | io_wqe_wake_worker(wqe, acct); | |
304 | } | |
305 | ||
771b53d0 JA |
306 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
307 | { | |
308 | allow_kernel_signal(SIGINT); | |
309 | ||
310 | current->flags |= PF_IO_WORKER; | |
311 | ||
312 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); | |
fcb323cc | 313 | worker->restore_files = current->files; |
c5def4ab | 314 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
315 | } |
316 | ||
317 | /* | |
318 | * Worker will start processing some work. Move it to the busy list, if | |
319 | * it's currently on the freelist | |
320 | */ | |
321 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
322 | struct io_wq_work *work) | |
323 | __must_hold(wqe->lock) | |
324 | { | |
c5def4ab JA |
325 | bool worker_bound, work_bound; |
326 | ||
771b53d0 JA |
327 | if (worker->flags & IO_WORKER_F_FREE) { |
328 | worker->flags &= ~IO_WORKER_F_FREE; | |
329 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 330 | } |
c5def4ab JA |
331 | |
332 | /* | |
333 | * If worker is moving from bound to unbound (or vice versa), then | |
334 | * ensure we update the running accounting. | |
335 | */ | |
b2e9c7d6 DC |
336 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
337 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
338 | if (worker_bound != work_bound) { | |
c5def4ab JA |
339 | io_wqe_dec_running(wqe, worker); |
340 | if (work_bound) { | |
341 | worker->flags |= IO_WORKER_F_BOUND; | |
342 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
343 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
344 | atomic_dec(&wqe->wq->user->processes); | |
345 | } else { | |
346 | worker->flags &= ~IO_WORKER_F_BOUND; | |
347 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
348 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
349 | atomic_inc(&wqe->wq->user->processes); | |
350 | } | |
351 | io_wqe_inc_running(wqe, worker); | |
352 | } | |
771b53d0 JA |
353 | } |
354 | ||
355 | /* | |
356 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
357 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
358 | * the lock in that case and return success. Since the caller has to | |
359 | * retry the loop in that case (we changed task state), we don't regrab | |
360 | * the lock if we return success. | |
361 | */ | |
362 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) | |
363 | __must_hold(wqe->lock) | |
364 | { | |
365 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
366 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 367 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 JA |
368 | } |
369 | ||
370 | return __io_worker_unuse(wqe, worker); | |
371 | } | |
372 | ||
373 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe, unsigned *hash) | |
374 | __must_hold(wqe->lock) | |
375 | { | |
6206f0e1 | 376 | struct io_wq_work_node *node, *prev; |
771b53d0 JA |
377 | struct io_wq_work *work; |
378 | ||
6206f0e1 JA |
379 | wq_list_for_each(node, prev, &wqe->work_list) { |
380 | work = container_of(node, struct io_wq_work, list); | |
381 | ||
771b53d0 JA |
382 | /* not hashed, can run anytime */ |
383 | if (!(work->flags & IO_WQ_WORK_HASHED)) { | |
6206f0e1 | 384 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
385 | return work; |
386 | } | |
387 | ||
388 | /* hashed, can run if not already running */ | |
389 | *hash = work->flags >> IO_WQ_HASH_SHIFT; | |
390 | if (!(wqe->hash_map & BIT_ULL(*hash))) { | |
391 | wqe->hash_map |= BIT_ULL(*hash); | |
6206f0e1 | 392 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
393 | return work; |
394 | } | |
395 | } | |
396 | ||
397 | return NULL; | |
398 | } | |
399 | ||
cccf0ee8 JA |
400 | static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work) |
401 | { | |
402 | if (worker->mm) { | |
403 | unuse_mm(worker->mm); | |
404 | mmput(worker->mm); | |
405 | worker->mm = NULL; | |
406 | } | |
407 | if (!work->mm) { | |
408 | set_fs(KERNEL_DS); | |
409 | return; | |
410 | } | |
411 | if (mmget_not_zero(work->mm)) { | |
412 | use_mm(work->mm); | |
413 | if (!worker->mm) | |
414 | set_fs(USER_DS); | |
415 | worker->mm = work->mm; | |
416 | /* hang on to this mm */ | |
417 | work->mm = NULL; | |
418 | return; | |
419 | } | |
420 | ||
421 | /* failed grabbing mm, ensure work gets cancelled */ | |
422 | work->flags |= IO_WQ_WORK_CANCEL; | |
423 | } | |
424 | ||
425 | static void io_wq_switch_creds(struct io_worker *worker, | |
426 | struct io_wq_work *work) | |
427 | { | |
428 | const struct cred *old_creds = override_creds(work->creds); | |
429 | ||
430 | worker->cur_creds = work->creds; | |
431 | if (worker->saved_creds) | |
432 | put_cred(old_creds); /* creds set by previous switch */ | |
433 | else | |
434 | worker->saved_creds = old_creds; | |
435 | } | |
436 | ||
771b53d0 JA |
437 | static void io_worker_handle_work(struct io_worker *worker) |
438 | __releases(wqe->lock) | |
439 | { | |
7d723065 | 440 | struct io_wq_work *work, *old_work = NULL, *put_work = NULL; |
771b53d0 JA |
441 | struct io_wqe *wqe = worker->wqe; |
442 | struct io_wq *wq = wqe->wq; | |
443 | ||
444 | do { | |
445 | unsigned hash = -1U; | |
446 | ||
771b53d0 JA |
447 | /* |
448 | * If we got some work, mark us as busy. If we didn't, but | |
449 | * the list isn't empty, it means we stalled on hashed work. | |
450 | * Mark us stalled so we don't keep looking for work when we | |
451 | * can't make progress, any work completion or insertion will | |
452 | * clear the stalled flag. | |
453 | */ | |
454 | work = io_get_next_work(wqe, &hash); | |
455 | if (work) | |
456 | __io_worker_busy(wqe, worker, work); | |
6206f0e1 | 457 | else if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
458 | wqe->flags |= IO_WQE_FLAG_STALLED; |
459 | ||
460 | spin_unlock_irq(&wqe->lock); | |
7d723065 JA |
461 | if (put_work && wq->put_work) |
462 | wq->put_work(old_work); | |
771b53d0 JA |
463 | if (!work) |
464 | break; | |
465 | next: | |
36c2f922 JA |
466 | /* flush any pending signals before assigning new work */ |
467 | if (signal_pending(current)) | |
468 | flush_signals(current); | |
469 | ||
fd1c4bc6 HD |
470 | cond_resched(); |
471 | ||
36c2f922 JA |
472 | spin_lock_irq(&worker->lock); |
473 | worker->cur_work = work; | |
474 | spin_unlock_irq(&worker->lock); | |
475 | ||
b76da70f JA |
476 | if (work->flags & IO_WQ_WORK_CB) |
477 | work->func(&work); | |
478 | ||
fcb323cc JA |
479 | if ((work->flags & IO_WQ_WORK_NEEDS_FILES) && |
480 | current->files != work->files) { | |
481 | task_lock(current); | |
482 | current->files = work->files; | |
483 | task_unlock(current); | |
484 | } | |
cccf0ee8 JA |
485 | if (work->mm != worker->mm) |
486 | io_wq_switch_mm(worker, work); | |
487 | if (worker->cur_creds != work->creds) | |
488 | io_wq_switch_creds(worker, work); | |
0c9d5ccd JA |
489 | /* |
490 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable work, | |
491 | * the worker function will do the right thing. | |
492 | */ | |
771b53d0 JA |
493 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) |
494 | work->flags |= IO_WQ_WORK_CANCEL; | |
495 | if (worker->mm) | |
496 | work->flags |= IO_WQ_WORK_HAS_MM; | |
497 | ||
7d723065 JA |
498 | if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) { |
499 | put_work = work; | |
500 | wq->get_work(work); | |
501 | } | |
502 | ||
771b53d0 JA |
503 | old_work = work; |
504 | work->func(&work); | |
505 | ||
36c2f922 | 506 | spin_lock_irq(&worker->lock); |
771b53d0 | 507 | worker->cur_work = NULL; |
36c2f922 JA |
508 | spin_unlock_irq(&worker->lock); |
509 | ||
510 | spin_lock_irq(&wqe->lock); | |
511 | ||
771b53d0 JA |
512 | if (hash != -1U) { |
513 | wqe->hash_map &= ~BIT_ULL(hash); | |
514 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
515 | } | |
516 | if (work && work != old_work) { | |
517 | spin_unlock_irq(&wqe->lock); | |
7d723065 JA |
518 | |
519 | if (put_work && wq->put_work) { | |
520 | wq->put_work(put_work); | |
521 | put_work = NULL; | |
522 | } | |
523 | ||
771b53d0 JA |
524 | /* dependent work not hashed */ |
525 | hash = -1U; | |
526 | goto next; | |
527 | } | |
528 | } while (1); | |
529 | } | |
530 | ||
e995d512 JA |
531 | static inline void io_worker_spin_for_work(struct io_wqe *wqe) |
532 | { | |
533 | int i = 0; | |
534 | ||
535 | while (++i < 1000) { | |
536 | if (io_wqe_run_queue(wqe)) | |
537 | break; | |
538 | if (need_resched()) | |
539 | break; | |
540 | cpu_relax(); | |
541 | } | |
542 | } | |
543 | ||
771b53d0 JA |
544 | static int io_wqe_worker(void *data) |
545 | { | |
546 | struct io_worker *worker = data; | |
547 | struct io_wqe *wqe = worker->wqe; | |
548 | struct io_wq *wq = wqe->wq; | |
e995d512 | 549 | bool did_work; |
771b53d0 JA |
550 | |
551 | io_worker_start(wqe, worker); | |
552 | ||
e995d512 | 553 | did_work = false; |
771b53d0 | 554 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
506d95ff | 555 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 JA |
556 | loop: |
557 | if (did_work) | |
558 | io_worker_spin_for_work(wqe); | |
771b53d0 JA |
559 | spin_lock_irq(&wqe->lock); |
560 | if (io_wqe_run_queue(wqe)) { | |
561 | __set_current_state(TASK_RUNNING); | |
562 | io_worker_handle_work(worker); | |
e995d512 JA |
563 | did_work = true; |
564 | goto loop; | |
771b53d0 | 565 | } |
e995d512 | 566 | did_work = false; |
771b53d0 JA |
567 | /* drops the lock on success, retry */ |
568 | if (__io_worker_idle(wqe, worker)) { | |
569 | __release(&wqe->lock); | |
e995d512 | 570 | goto loop; |
771b53d0 JA |
571 | } |
572 | spin_unlock_irq(&wqe->lock); | |
573 | if (signal_pending(current)) | |
574 | flush_signals(current); | |
575 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
576 | continue; | |
577 | /* timed out, exit unless we're the fixed worker */ | |
578 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
579 | !(worker->flags & IO_WORKER_F_FIXED)) | |
580 | break; | |
581 | } | |
582 | ||
771b53d0 JA |
583 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
584 | spin_lock_irq(&wqe->lock); | |
6206f0e1 | 585 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
586 | io_worker_handle_work(worker); |
587 | else | |
588 | spin_unlock_irq(&wqe->lock); | |
589 | } | |
590 | ||
591 | io_worker_exit(worker); | |
592 | return 0; | |
593 | } | |
594 | ||
771b53d0 JA |
595 | /* |
596 | * Called when a worker is scheduled in. Mark us as currently running. | |
597 | */ | |
598 | void io_wq_worker_running(struct task_struct *tsk) | |
599 | { | |
600 | struct io_worker *worker = kthread_data(tsk); | |
601 | struct io_wqe *wqe = worker->wqe; | |
602 | ||
603 | if (!(worker->flags & IO_WORKER_F_UP)) | |
604 | return; | |
605 | if (worker->flags & IO_WORKER_F_RUNNING) | |
606 | return; | |
607 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 608 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
609 | } |
610 | ||
611 | /* | |
612 | * Called when worker is going to sleep. If there are no workers currently | |
613 | * running and we have work pending, wake up a free one or have the manager | |
614 | * set one up. | |
615 | */ | |
616 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
617 | { | |
618 | struct io_worker *worker = kthread_data(tsk); | |
619 | struct io_wqe *wqe = worker->wqe; | |
620 | ||
621 | if (!(worker->flags & IO_WORKER_F_UP)) | |
622 | return; | |
623 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
624 | return; | |
625 | ||
626 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
627 | ||
628 | spin_lock_irq(&wqe->lock); | |
c5def4ab | 629 | io_wqe_dec_running(wqe, worker); |
771b53d0 JA |
630 | spin_unlock_irq(&wqe->lock); |
631 | } | |
632 | ||
b60fda60 | 633 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 634 | { |
c5def4ab | 635 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
636 | struct io_worker *worker; |
637 | ||
ad6e005c | 638 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 639 | if (!worker) |
b60fda60 | 640 | return false; |
771b53d0 JA |
641 | |
642 | refcount_set(&worker->ref, 1); | |
643 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 644 | worker->wqe = wqe; |
36c2f922 | 645 | spin_lock_init(&worker->lock); |
771b53d0 JA |
646 | |
647 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 648 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
649 | if (IS_ERR(worker->task)) { |
650 | kfree(worker); | |
b60fda60 | 651 | return false; |
771b53d0 JA |
652 | } |
653 | ||
654 | spin_lock_irq(&wqe->lock); | |
021d1cdd | 655 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 656 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 657 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
658 | if (index == IO_WQ_ACCT_BOUND) |
659 | worker->flags |= IO_WORKER_F_BOUND; | |
660 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 661 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 662 | acct->nr_workers++; |
771b53d0 JA |
663 | spin_unlock_irq(&wqe->lock); |
664 | ||
c5def4ab JA |
665 | if (index == IO_WQ_ACCT_UNBOUND) |
666 | atomic_inc(&wq->user->processes); | |
667 | ||
771b53d0 | 668 | wake_up_process(worker->task); |
b60fda60 | 669 | return true; |
771b53d0 JA |
670 | } |
671 | ||
c5def4ab | 672 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
673 | __must_hold(wqe->lock) |
674 | { | |
c5def4ab | 675 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 676 | |
c5def4ab | 677 | /* if we have available workers or no work, no need */ |
021d1cdd | 678 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
679 | return false; |
680 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
681 | } |
682 | ||
683 | /* | |
684 | * Manager thread. Tasked with creating new workers, if we need them. | |
685 | */ | |
686 | static int io_wq_manager(void *data) | |
687 | { | |
688 | struct io_wq *wq = data; | |
3fc50ab5 JH |
689 | int workers_to_create = num_possible_nodes(); |
690 | int node; | |
771b53d0 | 691 | |
b60fda60 | 692 | /* create fixed workers */ |
3fc50ab5 JH |
693 | refcount_set(&wq->refs, workers_to_create); |
694 | for_each_node(node) { | |
695 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) | |
696 | goto err; | |
697 | workers_to_create--; | |
b60fda60 | 698 | } |
771b53d0 | 699 | |
b60fda60 JA |
700 | complete(&wq->done); |
701 | ||
702 | while (!kthread_should_stop()) { | |
3fc50ab5 JH |
703 | for_each_node(node) { |
704 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 705 | bool fork_worker[2] = { false, false }; |
771b53d0 JA |
706 | |
707 | spin_lock_irq(&wqe->lock); | |
c5def4ab JA |
708 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
709 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
710 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
711 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
771b53d0 | 712 | spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
713 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
714 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
715 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
716 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
717 | } |
718 | set_current_state(TASK_INTERRUPTIBLE); | |
719 | schedule_timeout(HZ); | |
720 | } | |
721 | ||
722 | return 0; | |
b60fda60 JA |
723 | err: |
724 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
725 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 726 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
727 | complete(&wq->done); |
728 | return 0; | |
771b53d0 JA |
729 | } |
730 | ||
c5def4ab JA |
731 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
732 | struct io_wq_work *work) | |
733 | { | |
734 | bool free_worker; | |
735 | ||
736 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
737 | return true; | |
738 | if (atomic_read(&acct->nr_running)) | |
739 | return true; | |
740 | ||
741 | rcu_read_lock(); | |
021d1cdd | 742 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
743 | rcu_read_unlock(); |
744 | if (free_worker) | |
745 | return true; | |
746 | ||
747 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
748 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
749 | return false; | |
750 | ||
751 | return true; | |
752 | } | |
753 | ||
771b53d0 JA |
754 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
755 | { | |
c5def4ab | 756 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 757 | int work_flags; |
771b53d0 JA |
758 | unsigned long flags; |
759 | ||
c5def4ab JA |
760 | /* |
761 | * Do early check to see if we need a new unbound worker, and if we do, | |
762 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
763 | * gap between this check and incrementing the value, but that's OK. | |
764 | * It's close enough to not be an issue, fork() has the same delay. | |
765 | */ | |
766 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
767 | work->flags |= IO_WQ_WORK_CANCEL; | |
768 | work->func(&work); | |
769 | return; | |
770 | } | |
771 | ||
895e2ca0 | 772 | work_flags = work->flags; |
771b53d0 | 773 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 | 774 | wq_list_add_tail(&work->list, &wqe->work_list); |
771b53d0 JA |
775 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
776 | spin_unlock_irqrestore(&wqe->lock, flags); | |
777 | ||
895e2ca0 JA |
778 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
779 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 780 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
781 | } |
782 | ||
783 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
784 | { | |
785 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
786 | ||
787 | io_wqe_enqueue(wqe, work); | |
788 | } | |
789 | ||
790 | /* | |
791 | * Enqueue work, hashed by some key. Work items that hash to the same value | |
792 | * will not be done in parallel. Used to limit concurrent writes, generally | |
793 | * hashed by inode. | |
794 | */ | |
795 | void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val) | |
796 | { | |
797 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
798 | unsigned bit; | |
799 | ||
800 | ||
801 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
802 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
803 | io_wqe_enqueue(wqe, work); | |
804 | } | |
805 | ||
806 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
807 | { | |
808 | send_sig(SIGINT, worker->task, 1); | |
809 | return false; | |
810 | } | |
811 | ||
812 | /* | |
813 | * Iterate the passed in list and call the specific function for each | |
814 | * worker that isn't exiting | |
815 | */ | |
816 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
817 | bool (*func)(struct io_worker *, void *), |
818 | void *data) | |
819 | { | |
771b53d0 JA |
820 | struct io_worker *worker; |
821 | bool ret = false; | |
822 | ||
e61df66c | 823 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 JA |
824 | if (io_worker_get(worker)) { |
825 | ret = func(worker, data); | |
826 | io_worker_release(worker); | |
827 | if (ret) | |
828 | break; | |
829 | } | |
830 | } | |
e61df66c | 831 | |
771b53d0 JA |
832 | return ret; |
833 | } | |
834 | ||
835 | void io_wq_cancel_all(struct io_wq *wq) | |
836 | { | |
3fc50ab5 | 837 | int node; |
771b53d0 JA |
838 | |
839 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
840 | ||
771b53d0 | 841 | rcu_read_lock(); |
3fc50ab5 JH |
842 | for_each_node(node) { |
843 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 844 | |
e61df66c | 845 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
846 | } |
847 | rcu_read_unlock(); | |
848 | } | |
849 | ||
62755e35 JA |
850 | struct io_cb_cancel_data { |
851 | struct io_wqe *wqe; | |
852 | work_cancel_fn *cancel; | |
853 | void *caller_data; | |
854 | }; | |
855 | ||
856 | static bool io_work_cancel(struct io_worker *worker, void *cancel_data) | |
857 | { | |
858 | struct io_cb_cancel_data *data = cancel_data; | |
6f72653e | 859 | unsigned long flags; |
62755e35 JA |
860 | bool ret = false; |
861 | ||
862 | /* | |
863 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 864 | * may dereference the passed in work. |
62755e35 | 865 | */ |
36c2f922 | 866 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 867 | if (worker->cur_work && |
0c9d5ccd | 868 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
62755e35 JA |
869 | data->cancel(worker->cur_work, data->caller_data)) { |
870 | send_sig(SIGINT, worker->task, 1); | |
871 | ret = true; | |
872 | } | |
36c2f922 | 873 | spin_unlock_irqrestore(&worker->lock, flags); |
62755e35 JA |
874 | |
875 | return ret; | |
876 | } | |
877 | ||
878 | static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe, | |
879 | work_cancel_fn *cancel, | |
880 | void *cancel_data) | |
881 | { | |
882 | struct io_cb_cancel_data data = { | |
883 | .wqe = wqe, | |
884 | .cancel = cancel, | |
885 | .caller_data = cancel_data, | |
886 | }; | |
6206f0e1 | 887 | struct io_wq_work_node *node, *prev; |
62755e35 | 888 | struct io_wq_work *work; |
6f72653e | 889 | unsigned long flags; |
62755e35 JA |
890 | bool found = false; |
891 | ||
6f72653e | 892 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
893 | wq_list_for_each(node, prev, &wqe->work_list) { |
894 | work = container_of(node, struct io_wq_work, list); | |
895 | ||
62755e35 | 896 | if (cancel(work, cancel_data)) { |
6206f0e1 | 897 | wq_node_del(&wqe->work_list, node, prev); |
62755e35 JA |
898 | found = true; |
899 | break; | |
900 | } | |
901 | } | |
6f72653e | 902 | spin_unlock_irqrestore(&wqe->lock, flags); |
62755e35 JA |
903 | |
904 | if (found) { | |
905 | work->flags |= IO_WQ_WORK_CANCEL; | |
906 | work->func(&work); | |
907 | return IO_WQ_CANCEL_OK; | |
908 | } | |
909 | ||
910 | rcu_read_lock(); | |
e61df66c | 911 | found = io_wq_for_each_worker(wqe, io_work_cancel, &data); |
62755e35 JA |
912 | rcu_read_unlock(); |
913 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
914 | } | |
915 | ||
916 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, | |
917 | void *data) | |
918 | { | |
919 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 920 | int node; |
62755e35 | 921 | |
3fc50ab5 JH |
922 | for_each_node(node) { |
923 | struct io_wqe *wqe = wq->wqes[node]; | |
62755e35 JA |
924 | |
925 | ret = io_wqe_cancel_cb_work(wqe, cancel, data); | |
926 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
927 | break; | |
928 | } | |
929 | ||
930 | return ret; | |
931 | } | |
932 | ||
771b53d0 JA |
933 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
934 | { | |
935 | struct io_wq_work *work = data; | |
36c2f922 JA |
936 | unsigned long flags; |
937 | bool ret = false; | |
771b53d0 | 938 | |
36c2f922 JA |
939 | if (worker->cur_work != work) |
940 | return false; | |
941 | ||
942 | spin_lock_irqsave(&worker->lock, flags); | |
0c9d5ccd JA |
943 | if (worker->cur_work == work && |
944 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL)) { | |
771b53d0 | 945 | send_sig(SIGINT, worker->task, 1); |
36c2f922 | 946 | ret = true; |
771b53d0 | 947 | } |
36c2f922 | 948 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 949 | |
36c2f922 | 950 | return ret; |
771b53d0 JA |
951 | } |
952 | ||
953 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, | |
954 | struct io_wq_work *cwork) | |
955 | { | |
6206f0e1 | 956 | struct io_wq_work_node *node, *prev; |
771b53d0 | 957 | struct io_wq_work *work; |
6f72653e | 958 | unsigned long flags; |
771b53d0 JA |
959 | bool found = false; |
960 | ||
961 | cwork->flags |= IO_WQ_WORK_CANCEL; | |
962 | ||
963 | /* | |
964 | * First check pending list, if we're lucky we can just remove it | |
965 | * from there. CANCEL_OK means that the work is returned as-new, | |
966 | * no completion will be posted for it. | |
967 | */ | |
6f72653e | 968 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
969 | wq_list_for_each(node, prev, &wqe->work_list) { |
970 | work = container_of(node, struct io_wq_work, list); | |
971 | ||
771b53d0 | 972 | if (work == cwork) { |
6206f0e1 | 973 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
974 | found = true; |
975 | break; | |
976 | } | |
977 | } | |
6f72653e | 978 | spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 JA |
979 | |
980 | if (found) { | |
981 | work->flags |= IO_WQ_WORK_CANCEL; | |
982 | work->func(&work); | |
983 | return IO_WQ_CANCEL_OK; | |
984 | } | |
985 | ||
986 | /* | |
987 | * Now check if a free (going busy) or busy worker has the work | |
988 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
d195a66e | 989 | * as an indication that we attempt to signal cancellation. The |
771b53d0 JA |
990 | * completion will run normally in this case. |
991 | */ | |
992 | rcu_read_lock(); | |
e61df66c | 993 | found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, cwork); |
771b53d0 JA |
994 | rcu_read_unlock(); |
995 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
996 | } | |
997 | ||
998 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
999 | { | |
1000 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 1001 | int node; |
771b53d0 | 1002 | |
3fc50ab5 JH |
1003 | for_each_node(node) { |
1004 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
1005 | |
1006 | ret = io_wqe_cancel_work(wqe, cwork); | |
1007 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
1008 | break; | |
1009 | } | |
1010 | ||
1011 | return ret; | |
1012 | } | |
1013 | ||
1014 | struct io_wq_flush_data { | |
1015 | struct io_wq_work work; | |
1016 | struct completion done; | |
1017 | }; | |
1018 | ||
1019 | static void io_wq_flush_func(struct io_wq_work **workptr) | |
1020 | { | |
1021 | struct io_wq_work *work = *workptr; | |
1022 | struct io_wq_flush_data *data; | |
1023 | ||
1024 | data = container_of(work, struct io_wq_flush_data, work); | |
1025 | complete(&data->done); | |
1026 | } | |
1027 | ||
1028 | /* | |
1029 | * Doesn't wait for previously queued work to finish. When this completes, | |
1030 | * it just means that previously queued work was started. | |
1031 | */ | |
1032 | void io_wq_flush(struct io_wq *wq) | |
1033 | { | |
1034 | struct io_wq_flush_data data; | |
3fc50ab5 | 1035 | int node; |
771b53d0 | 1036 | |
3fc50ab5 JH |
1037 | for_each_node(node) { |
1038 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
1039 | |
1040 | init_completion(&data.done); | |
1041 | INIT_IO_WORK(&data.work, io_wq_flush_func); | |
7d723065 | 1042 | data.work.flags |= IO_WQ_WORK_INTERNAL; |
771b53d0 JA |
1043 | io_wqe_enqueue(wqe, &data.work); |
1044 | wait_for_completion(&data.done); | |
1045 | } | |
1046 | } | |
1047 | ||
576a347b | 1048 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1049 | { |
3fc50ab5 | 1050 | int ret = -ENOMEM, node; |
771b53d0 JA |
1051 | struct io_wq *wq; |
1052 | ||
ad6e005c | 1053 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1054 | if (!wq) |
1055 | return ERR_PTR(-ENOMEM); | |
1056 | ||
3fc50ab5 | 1057 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1058 | if (!wq->wqes) { |
1059 | kfree(wq); | |
1060 | return ERR_PTR(-ENOMEM); | |
1061 | } | |
1062 | ||
576a347b JA |
1063 | wq->get_work = data->get_work; |
1064 | wq->put_work = data->put_work; | |
7d723065 | 1065 | |
c5def4ab | 1066 | /* caller must already hold a reference to this */ |
576a347b | 1067 | wq->user = data->user; |
c5def4ab | 1068 | |
3fc50ab5 | 1069 | for_each_node(node) { |
771b53d0 JA |
1070 | struct io_wqe *wqe; |
1071 | ||
ad6e005c | 1072 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node); |
771b53d0 | 1073 | if (!wqe) |
3fc50ab5 JH |
1074 | goto err; |
1075 | wq->wqes[node] = wqe; | |
771b53d0 | 1076 | wqe->node = node; |
c5def4ab JA |
1077 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1078 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1079 | if (wq->user) { |
c5def4ab JA |
1080 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1081 | task_rlimit(current, RLIMIT_NPROC); | |
1082 | } | |
1083 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 JA |
1084 | wqe->node = node; |
1085 | wqe->wq = wq; | |
1086 | spin_lock_init(&wqe->lock); | |
6206f0e1 | 1087 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1088 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1089 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1090 | } |
1091 | ||
1092 | init_completion(&wq->done); | |
1093 | ||
771b53d0 JA |
1094 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
1095 | if (!IS_ERR(wq->manager)) { | |
1096 | wake_up_process(wq->manager); | |
b60fda60 JA |
1097 | wait_for_completion(&wq->done); |
1098 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1099 | ret = -ENOMEM; | |
1100 | goto err; | |
1101 | } | |
848f7e18 | 1102 | refcount_set(&wq->use_refs, 1); |
b60fda60 | 1103 | reinit_completion(&wq->done); |
771b53d0 JA |
1104 | return wq; |
1105 | } | |
1106 | ||
1107 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1108 | complete(&wq->done); |
b60fda60 | 1109 | err: |
3fc50ab5 JH |
1110 | for_each_node(node) |
1111 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1112 | kfree(wq->wqes); |
1113 | kfree(wq); | |
771b53d0 JA |
1114 | return ERR_PTR(ret); |
1115 | } | |
1116 | ||
1117 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
1118 | { | |
1119 | wake_up_process(worker->task); | |
1120 | return false; | |
1121 | } | |
1122 | ||
848f7e18 | 1123 | static void __io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1124 | { |
3fc50ab5 | 1125 | int node; |
771b53d0 | 1126 | |
b60fda60 JA |
1127 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1128 | if (wq->manager) | |
771b53d0 | 1129 | kthread_stop(wq->manager); |
771b53d0 JA |
1130 | |
1131 | rcu_read_lock(); | |
3fc50ab5 JH |
1132 | for_each_node(node) |
1133 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1134 | rcu_read_unlock(); |
1135 | ||
1136 | wait_for_completion(&wq->done); | |
1137 | ||
3fc50ab5 JH |
1138 | for_each_node(node) |
1139 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1140 | kfree(wq->wqes); |
1141 | kfree(wq); | |
1142 | } | |
848f7e18 JA |
1143 | |
1144 | void io_wq_destroy(struct io_wq *wq) | |
1145 | { | |
1146 | if (refcount_dec_and_test(&wq->use_refs)) | |
1147 | __io_wq_destroy(wq); | |
1148 | } |