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 | ||
f86cd20c | 479 | if (work->files && current->files != work->files) { |
fcb323cc JA |
480 | task_lock(current); |
481 | current->files = work->files; | |
482 | task_unlock(current); | |
483 | } | |
cccf0ee8 JA |
484 | if (work->mm != worker->mm) |
485 | io_wq_switch_mm(worker, work); | |
486 | if (worker->cur_creds != work->creds) | |
487 | io_wq_switch_creds(worker, work); | |
0c9d5ccd JA |
488 | /* |
489 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable work, | |
490 | * the worker function will do the right thing. | |
491 | */ | |
771b53d0 JA |
492 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) |
493 | work->flags |= IO_WQ_WORK_CANCEL; | |
494 | if (worker->mm) | |
495 | work->flags |= IO_WQ_WORK_HAS_MM; | |
496 | ||
7d723065 JA |
497 | if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) { |
498 | put_work = work; | |
499 | wq->get_work(work); | |
500 | } | |
501 | ||
771b53d0 JA |
502 | old_work = work; |
503 | work->func(&work); | |
504 | ||
36c2f922 | 505 | spin_lock_irq(&worker->lock); |
771b53d0 | 506 | worker->cur_work = NULL; |
36c2f922 JA |
507 | spin_unlock_irq(&worker->lock); |
508 | ||
509 | spin_lock_irq(&wqe->lock); | |
510 | ||
771b53d0 JA |
511 | if (hash != -1U) { |
512 | wqe->hash_map &= ~BIT_ULL(hash); | |
513 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
514 | } | |
515 | if (work && work != old_work) { | |
516 | spin_unlock_irq(&wqe->lock); | |
7d723065 JA |
517 | |
518 | if (put_work && wq->put_work) { | |
519 | wq->put_work(put_work); | |
520 | put_work = NULL; | |
521 | } | |
522 | ||
771b53d0 JA |
523 | /* dependent work not hashed */ |
524 | hash = -1U; | |
525 | goto next; | |
526 | } | |
527 | } while (1); | |
528 | } | |
529 | ||
e995d512 JA |
530 | static inline void io_worker_spin_for_work(struct io_wqe *wqe) |
531 | { | |
532 | int i = 0; | |
533 | ||
534 | while (++i < 1000) { | |
535 | if (io_wqe_run_queue(wqe)) | |
536 | break; | |
537 | if (need_resched()) | |
538 | break; | |
539 | cpu_relax(); | |
540 | } | |
541 | } | |
542 | ||
771b53d0 JA |
543 | static int io_wqe_worker(void *data) |
544 | { | |
545 | struct io_worker *worker = data; | |
546 | struct io_wqe *wqe = worker->wqe; | |
547 | struct io_wq *wq = wqe->wq; | |
e995d512 | 548 | bool did_work; |
771b53d0 JA |
549 | |
550 | io_worker_start(wqe, worker); | |
551 | ||
e995d512 | 552 | did_work = false; |
771b53d0 | 553 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
506d95ff | 554 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 JA |
555 | loop: |
556 | if (did_work) | |
557 | io_worker_spin_for_work(wqe); | |
771b53d0 JA |
558 | spin_lock_irq(&wqe->lock); |
559 | if (io_wqe_run_queue(wqe)) { | |
560 | __set_current_state(TASK_RUNNING); | |
561 | io_worker_handle_work(worker); | |
e995d512 JA |
562 | did_work = true; |
563 | goto loop; | |
771b53d0 | 564 | } |
e995d512 | 565 | did_work = false; |
771b53d0 JA |
566 | /* drops the lock on success, retry */ |
567 | if (__io_worker_idle(wqe, worker)) { | |
568 | __release(&wqe->lock); | |
e995d512 | 569 | goto loop; |
771b53d0 JA |
570 | } |
571 | spin_unlock_irq(&wqe->lock); | |
572 | if (signal_pending(current)) | |
573 | flush_signals(current); | |
574 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
575 | continue; | |
576 | /* timed out, exit unless we're the fixed worker */ | |
577 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
578 | !(worker->flags & IO_WORKER_F_FIXED)) | |
579 | break; | |
580 | } | |
581 | ||
771b53d0 JA |
582 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
583 | spin_lock_irq(&wqe->lock); | |
6206f0e1 | 584 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
585 | io_worker_handle_work(worker); |
586 | else | |
587 | spin_unlock_irq(&wqe->lock); | |
588 | } | |
589 | ||
590 | io_worker_exit(worker); | |
591 | return 0; | |
592 | } | |
593 | ||
771b53d0 JA |
594 | /* |
595 | * Called when a worker is scheduled in. Mark us as currently running. | |
596 | */ | |
597 | void io_wq_worker_running(struct task_struct *tsk) | |
598 | { | |
599 | struct io_worker *worker = kthread_data(tsk); | |
600 | struct io_wqe *wqe = worker->wqe; | |
601 | ||
602 | if (!(worker->flags & IO_WORKER_F_UP)) | |
603 | return; | |
604 | if (worker->flags & IO_WORKER_F_RUNNING) | |
605 | return; | |
606 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 607 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
608 | } |
609 | ||
610 | /* | |
611 | * Called when worker is going to sleep. If there are no workers currently | |
612 | * running and we have work pending, wake up a free one or have the manager | |
613 | * set one up. | |
614 | */ | |
615 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
616 | { | |
617 | struct io_worker *worker = kthread_data(tsk); | |
618 | struct io_wqe *wqe = worker->wqe; | |
619 | ||
620 | if (!(worker->flags & IO_WORKER_F_UP)) | |
621 | return; | |
622 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
623 | return; | |
624 | ||
625 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
626 | ||
627 | spin_lock_irq(&wqe->lock); | |
c5def4ab | 628 | io_wqe_dec_running(wqe, worker); |
771b53d0 JA |
629 | spin_unlock_irq(&wqe->lock); |
630 | } | |
631 | ||
b60fda60 | 632 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 633 | { |
c5def4ab | 634 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
635 | struct io_worker *worker; |
636 | ||
ad6e005c | 637 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 638 | if (!worker) |
b60fda60 | 639 | return false; |
771b53d0 JA |
640 | |
641 | refcount_set(&worker->ref, 1); | |
642 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 643 | worker->wqe = wqe; |
36c2f922 | 644 | spin_lock_init(&worker->lock); |
771b53d0 JA |
645 | |
646 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 647 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
648 | if (IS_ERR(worker->task)) { |
649 | kfree(worker); | |
b60fda60 | 650 | return false; |
771b53d0 JA |
651 | } |
652 | ||
653 | spin_lock_irq(&wqe->lock); | |
021d1cdd | 654 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 655 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 656 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
657 | if (index == IO_WQ_ACCT_BOUND) |
658 | worker->flags |= IO_WORKER_F_BOUND; | |
659 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 660 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 661 | acct->nr_workers++; |
771b53d0 JA |
662 | spin_unlock_irq(&wqe->lock); |
663 | ||
c5def4ab JA |
664 | if (index == IO_WQ_ACCT_UNBOUND) |
665 | atomic_inc(&wq->user->processes); | |
666 | ||
771b53d0 | 667 | wake_up_process(worker->task); |
b60fda60 | 668 | return true; |
771b53d0 JA |
669 | } |
670 | ||
c5def4ab | 671 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
672 | __must_hold(wqe->lock) |
673 | { | |
c5def4ab | 674 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 675 | |
c5def4ab | 676 | /* if we have available workers or no work, no need */ |
021d1cdd | 677 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
678 | return false; |
679 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
680 | } |
681 | ||
682 | /* | |
683 | * Manager thread. Tasked with creating new workers, if we need them. | |
684 | */ | |
685 | static int io_wq_manager(void *data) | |
686 | { | |
687 | struct io_wq *wq = data; | |
3fc50ab5 JH |
688 | int workers_to_create = num_possible_nodes(); |
689 | int node; | |
771b53d0 | 690 | |
b60fda60 | 691 | /* create fixed workers */ |
3fc50ab5 JH |
692 | refcount_set(&wq->refs, workers_to_create); |
693 | for_each_node(node) { | |
694 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) | |
695 | goto err; | |
696 | workers_to_create--; | |
b60fda60 | 697 | } |
771b53d0 | 698 | |
b60fda60 JA |
699 | complete(&wq->done); |
700 | ||
701 | while (!kthread_should_stop()) { | |
3fc50ab5 JH |
702 | for_each_node(node) { |
703 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 704 | bool fork_worker[2] = { false, false }; |
771b53d0 JA |
705 | |
706 | spin_lock_irq(&wqe->lock); | |
c5def4ab JA |
707 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
708 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
709 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
710 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
771b53d0 | 711 | spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
712 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
713 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
714 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
715 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
716 | } |
717 | set_current_state(TASK_INTERRUPTIBLE); | |
718 | schedule_timeout(HZ); | |
719 | } | |
720 | ||
721 | return 0; | |
b60fda60 JA |
722 | err: |
723 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
724 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 725 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
726 | complete(&wq->done); |
727 | return 0; | |
771b53d0 JA |
728 | } |
729 | ||
c5def4ab JA |
730 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
731 | struct io_wq_work *work) | |
732 | { | |
733 | bool free_worker; | |
734 | ||
735 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
736 | return true; | |
737 | if (atomic_read(&acct->nr_running)) | |
738 | return true; | |
739 | ||
740 | rcu_read_lock(); | |
021d1cdd | 741 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
742 | rcu_read_unlock(); |
743 | if (free_worker) | |
744 | return true; | |
745 | ||
746 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
747 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
748 | return false; | |
749 | ||
750 | return true; | |
751 | } | |
752 | ||
771b53d0 JA |
753 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
754 | { | |
c5def4ab | 755 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 756 | int work_flags; |
771b53d0 JA |
757 | unsigned long flags; |
758 | ||
c5def4ab JA |
759 | /* |
760 | * Do early check to see if we need a new unbound worker, and if we do, | |
761 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
762 | * gap between this check and incrementing the value, but that's OK. | |
763 | * It's close enough to not be an issue, fork() has the same delay. | |
764 | */ | |
765 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
766 | work->flags |= IO_WQ_WORK_CANCEL; | |
767 | work->func(&work); | |
768 | return; | |
769 | } | |
770 | ||
895e2ca0 | 771 | work_flags = work->flags; |
771b53d0 | 772 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 | 773 | wq_list_add_tail(&work->list, &wqe->work_list); |
771b53d0 JA |
774 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
775 | spin_unlock_irqrestore(&wqe->lock, flags); | |
776 | ||
895e2ca0 JA |
777 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
778 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 779 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
780 | } |
781 | ||
782 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
783 | { | |
784 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
785 | ||
786 | io_wqe_enqueue(wqe, work); | |
787 | } | |
788 | ||
789 | /* | |
790 | * Enqueue work, hashed by some key. Work items that hash to the same value | |
791 | * will not be done in parallel. Used to limit concurrent writes, generally | |
792 | * hashed by inode. | |
793 | */ | |
794 | void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val) | |
795 | { | |
796 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
797 | unsigned bit; | |
798 | ||
799 | ||
800 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
801 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
802 | io_wqe_enqueue(wqe, work); | |
803 | } | |
804 | ||
805 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
806 | { | |
807 | send_sig(SIGINT, worker->task, 1); | |
808 | return false; | |
809 | } | |
810 | ||
811 | /* | |
812 | * Iterate the passed in list and call the specific function for each | |
813 | * worker that isn't exiting | |
814 | */ | |
815 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
816 | bool (*func)(struct io_worker *, void *), |
817 | void *data) | |
818 | { | |
771b53d0 JA |
819 | struct io_worker *worker; |
820 | bool ret = false; | |
821 | ||
e61df66c | 822 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 JA |
823 | if (io_worker_get(worker)) { |
824 | ret = func(worker, data); | |
825 | io_worker_release(worker); | |
826 | if (ret) | |
827 | break; | |
828 | } | |
829 | } | |
e61df66c | 830 | |
771b53d0 JA |
831 | return ret; |
832 | } | |
833 | ||
834 | void io_wq_cancel_all(struct io_wq *wq) | |
835 | { | |
3fc50ab5 | 836 | int node; |
771b53d0 JA |
837 | |
838 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
839 | ||
771b53d0 | 840 | rcu_read_lock(); |
3fc50ab5 JH |
841 | for_each_node(node) { |
842 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 843 | |
e61df66c | 844 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
845 | } |
846 | rcu_read_unlock(); | |
847 | } | |
848 | ||
62755e35 JA |
849 | struct io_cb_cancel_data { |
850 | struct io_wqe *wqe; | |
851 | work_cancel_fn *cancel; | |
852 | void *caller_data; | |
853 | }; | |
854 | ||
855 | static bool io_work_cancel(struct io_worker *worker, void *cancel_data) | |
856 | { | |
857 | struct io_cb_cancel_data *data = cancel_data; | |
6f72653e | 858 | unsigned long flags; |
62755e35 JA |
859 | bool ret = false; |
860 | ||
861 | /* | |
862 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 863 | * may dereference the passed in work. |
62755e35 | 864 | */ |
36c2f922 | 865 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 866 | if (worker->cur_work && |
0c9d5ccd | 867 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
62755e35 JA |
868 | data->cancel(worker->cur_work, data->caller_data)) { |
869 | send_sig(SIGINT, worker->task, 1); | |
870 | ret = true; | |
871 | } | |
36c2f922 | 872 | spin_unlock_irqrestore(&worker->lock, flags); |
62755e35 JA |
873 | |
874 | return ret; | |
875 | } | |
876 | ||
877 | static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe, | |
878 | work_cancel_fn *cancel, | |
879 | void *cancel_data) | |
880 | { | |
881 | struct io_cb_cancel_data data = { | |
882 | .wqe = wqe, | |
883 | .cancel = cancel, | |
884 | .caller_data = cancel_data, | |
885 | }; | |
6206f0e1 | 886 | struct io_wq_work_node *node, *prev; |
62755e35 | 887 | struct io_wq_work *work; |
6f72653e | 888 | unsigned long flags; |
62755e35 JA |
889 | bool found = false; |
890 | ||
6f72653e | 891 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
892 | wq_list_for_each(node, prev, &wqe->work_list) { |
893 | work = container_of(node, struct io_wq_work, list); | |
894 | ||
62755e35 | 895 | if (cancel(work, cancel_data)) { |
6206f0e1 | 896 | wq_node_del(&wqe->work_list, node, prev); |
62755e35 JA |
897 | found = true; |
898 | break; | |
899 | } | |
900 | } | |
6f72653e | 901 | spin_unlock_irqrestore(&wqe->lock, flags); |
62755e35 JA |
902 | |
903 | if (found) { | |
904 | work->flags |= IO_WQ_WORK_CANCEL; | |
905 | work->func(&work); | |
906 | return IO_WQ_CANCEL_OK; | |
907 | } | |
908 | ||
909 | rcu_read_lock(); | |
e61df66c | 910 | found = io_wq_for_each_worker(wqe, io_work_cancel, &data); |
62755e35 JA |
911 | rcu_read_unlock(); |
912 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
913 | } | |
914 | ||
915 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, | |
916 | void *data) | |
917 | { | |
918 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 919 | int node; |
62755e35 | 920 | |
3fc50ab5 JH |
921 | for_each_node(node) { |
922 | struct io_wqe *wqe = wq->wqes[node]; | |
62755e35 JA |
923 | |
924 | ret = io_wqe_cancel_cb_work(wqe, cancel, data); | |
925 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
926 | break; | |
927 | } | |
928 | ||
929 | return ret; | |
930 | } | |
931 | ||
771b53d0 JA |
932 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
933 | { | |
934 | struct io_wq_work *work = data; | |
36c2f922 JA |
935 | unsigned long flags; |
936 | bool ret = false; | |
771b53d0 | 937 | |
36c2f922 JA |
938 | if (worker->cur_work != work) |
939 | return false; | |
940 | ||
941 | spin_lock_irqsave(&worker->lock, flags); | |
0c9d5ccd JA |
942 | if (worker->cur_work == work && |
943 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL)) { | |
771b53d0 | 944 | send_sig(SIGINT, worker->task, 1); |
36c2f922 | 945 | ret = true; |
771b53d0 | 946 | } |
36c2f922 | 947 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 948 | |
36c2f922 | 949 | return ret; |
771b53d0 JA |
950 | } |
951 | ||
952 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, | |
953 | struct io_wq_work *cwork) | |
954 | { | |
6206f0e1 | 955 | struct io_wq_work_node *node, *prev; |
771b53d0 | 956 | struct io_wq_work *work; |
6f72653e | 957 | unsigned long flags; |
771b53d0 JA |
958 | bool found = false; |
959 | ||
960 | cwork->flags |= IO_WQ_WORK_CANCEL; | |
961 | ||
962 | /* | |
963 | * First check pending list, if we're lucky we can just remove it | |
964 | * from there. CANCEL_OK means that the work is returned as-new, | |
965 | * no completion will be posted for it. | |
966 | */ | |
6f72653e | 967 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
968 | wq_list_for_each(node, prev, &wqe->work_list) { |
969 | work = container_of(node, struct io_wq_work, list); | |
970 | ||
771b53d0 | 971 | if (work == cwork) { |
6206f0e1 | 972 | wq_node_del(&wqe->work_list, node, prev); |
771b53d0 JA |
973 | found = true; |
974 | break; | |
975 | } | |
976 | } | |
6f72653e | 977 | spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 JA |
978 | |
979 | if (found) { | |
980 | work->flags |= IO_WQ_WORK_CANCEL; | |
981 | work->func(&work); | |
982 | return IO_WQ_CANCEL_OK; | |
983 | } | |
984 | ||
985 | /* | |
986 | * Now check if a free (going busy) or busy worker has the work | |
987 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
d195a66e | 988 | * as an indication that we attempt to signal cancellation. The |
771b53d0 JA |
989 | * completion will run normally in this case. |
990 | */ | |
991 | rcu_read_lock(); | |
e61df66c | 992 | found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, cwork); |
771b53d0 JA |
993 | rcu_read_unlock(); |
994 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
995 | } | |
996 | ||
997 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
998 | { | |
999 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
3fc50ab5 | 1000 | int node; |
771b53d0 | 1001 | |
3fc50ab5 JH |
1002 | for_each_node(node) { |
1003 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
1004 | |
1005 | ret = io_wqe_cancel_work(wqe, cwork); | |
1006 | if (ret != IO_WQ_CANCEL_NOTFOUND) | |
1007 | break; | |
1008 | } | |
1009 | ||
1010 | return ret; | |
1011 | } | |
1012 | ||
1013 | struct io_wq_flush_data { | |
1014 | struct io_wq_work work; | |
1015 | struct completion done; | |
1016 | }; | |
1017 | ||
1018 | static void io_wq_flush_func(struct io_wq_work **workptr) | |
1019 | { | |
1020 | struct io_wq_work *work = *workptr; | |
1021 | struct io_wq_flush_data *data; | |
1022 | ||
1023 | data = container_of(work, struct io_wq_flush_data, work); | |
1024 | complete(&data->done); | |
1025 | } | |
1026 | ||
1027 | /* | |
1028 | * Doesn't wait for previously queued work to finish. When this completes, | |
1029 | * it just means that previously queued work was started. | |
1030 | */ | |
1031 | void io_wq_flush(struct io_wq *wq) | |
1032 | { | |
1033 | struct io_wq_flush_data data; | |
3fc50ab5 | 1034 | int node; |
771b53d0 | 1035 | |
3fc50ab5 JH |
1036 | for_each_node(node) { |
1037 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 JA |
1038 | |
1039 | init_completion(&data.done); | |
1040 | INIT_IO_WORK(&data.work, io_wq_flush_func); | |
7d723065 | 1041 | data.work.flags |= IO_WQ_WORK_INTERNAL; |
771b53d0 JA |
1042 | io_wqe_enqueue(wqe, &data.work); |
1043 | wait_for_completion(&data.done); | |
1044 | } | |
1045 | } | |
1046 | ||
576a347b | 1047 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1048 | { |
3fc50ab5 | 1049 | int ret = -ENOMEM, node; |
771b53d0 JA |
1050 | struct io_wq *wq; |
1051 | ||
ad6e005c | 1052 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1053 | if (!wq) |
1054 | return ERR_PTR(-ENOMEM); | |
1055 | ||
3fc50ab5 | 1056 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1057 | if (!wq->wqes) { |
1058 | kfree(wq); | |
1059 | return ERR_PTR(-ENOMEM); | |
1060 | } | |
1061 | ||
576a347b JA |
1062 | wq->get_work = data->get_work; |
1063 | wq->put_work = data->put_work; | |
7d723065 | 1064 | |
c5def4ab | 1065 | /* caller must already hold a reference to this */ |
576a347b | 1066 | wq->user = data->user; |
c5def4ab | 1067 | |
3fc50ab5 | 1068 | for_each_node(node) { |
771b53d0 JA |
1069 | struct io_wqe *wqe; |
1070 | ||
ad6e005c | 1071 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node); |
771b53d0 | 1072 | if (!wqe) |
3fc50ab5 JH |
1073 | goto err; |
1074 | wq->wqes[node] = wqe; | |
771b53d0 | 1075 | wqe->node = node; |
c5def4ab JA |
1076 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1077 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1078 | if (wq->user) { |
c5def4ab JA |
1079 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1080 | task_rlimit(current, RLIMIT_NPROC); | |
1081 | } | |
1082 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 JA |
1083 | wqe->node = node; |
1084 | wqe->wq = wq; | |
1085 | spin_lock_init(&wqe->lock); | |
6206f0e1 | 1086 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1087 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1088 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1089 | } |
1090 | ||
1091 | init_completion(&wq->done); | |
1092 | ||
771b53d0 JA |
1093 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
1094 | if (!IS_ERR(wq->manager)) { | |
1095 | wake_up_process(wq->manager); | |
b60fda60 JA |
1096 | wait_for_completion(&wq->done); |
1097 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1098 | ret = -ENOMEM; | |
1099 | goto err; | |
1100 | } | |
848f7e18 | 1101 | refcount_set(&wq->use_refs, 1); |
b60fda60 | 1102 | reinit_completion(&wq->done); |
771b53d0 JA |
1103 | return wq; |
1104 | } | |
1105 | ||
1106 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1107 | complete(&wq->done); |
b60fda60 | 1108 | err: |
3fc50ab5 JH |
1109 | for_each_node(node) |
1110 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1111 | kfree(wq->wqes); |
1112 | kfree(wq); | |
771b53d0 JA |
1113 | return ERR_PTR(ret); |
1114 | } | |
1115 | ||
eba6f5a3 PB |
1116 | bool io_wq_get(struct io_wq *wq, struct io_wq_data *data) |
1117 | { | |
1118 | if (data->get_work != wq->get_work || data->put_work != wq->put_work) | |
1119 | return false; | |
1120 | ||
1121 | return refcount_inc_not_zero(&wq->use_refs); | |
1122 | } | |
1123 | ||
771b53d0 JA |
1124 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) |
1125 | { | |
1126 | wake_up_process(worker->task); | |
1127 | return false; | |
1128 | } | |
1129 | ||
848f7e18 | 1130 | static void __io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1131 | { |
3fc50ab5 | 1132 | int node; |
771b53d0 | 1133 | |
b60fda60 JA |
1134 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1135 | if (wq->manager) | |
771b53d0 | 1136 | kthread_stop(wq->manager); |
771b53d0 JA |
1137 | |
1138 | rcu_read_lock(); | |
3fc50ab5 JH |
1139 | for_each_node(node) |
1140 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1141 | rcu_read_unlock(); |
1142 | ||
1143 | wait_for_completion(&wq->done); | |
1144 | ||
3fc50ab5 JH |
1145 | for_each_node(node) |
1146 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1147 | kfree(wq->wqes); |
1148 | kfree(wq); | |
1149 | } | |
848f7e18 JA |
1150 | |
1151 | void io_wq_destroy(struct io_wq *wq) | |
1152 | { | |
1153 | if (refcount_dec_and_test(&wq->use_refs)) | |
1154 | __io_wq_destroy(wq); | |
1155 | } |