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> | |
9392a27d | 19 | #include <linux/fs_struct.h> |
aa96bf8a | 20 | #include <linux/task_work.h> |
771b53d0 JA |
21 | |
22 | #include "io-wq.h" | |
23 | ||
24 | #define WORKER_IDLE_TIMEOUT (5 * HZ) | |
25 | ||
26 | enum { | |
27 | IO_WORKER_F_UP = 1, /* up and active */ | |
28 | IO_WORKER_F_RUNNING = 2, /* account as running */ | |
29 | IO_WORKER_F_FREE = 4, /* worker on free list */ | |
30 | IO_WORKER_F_EXITING = 8, /* worker exiting */ | |
31 | IO_WORKER_F_FIXED = 16, /* static idle worker */ | |
c5def4ab | 32 | IO_WORKER_F_BOUND = 32, /* is doing bounded work */ |
771b53d0 JA |
33 | }; |
34 | ||
35 | enum { | |
36 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ | |
37 | IO_WQ_BIT_CANCEL = 1, /* cancel work on list */ | |
b60fda60 | 38 | IO_WQ_BIT_ERROR = 2, /* error on setup */ |
771b53d0 JA |
39 | }; |
40 | ||
41 | enum { | |
42 | IO_WQE_FLAG_STALLED = 1, /* stalled on hash */ | |
43 | }; | |
44 | ||
45 | /* | |
46 | * One for each thread in a wqe pool | |
47 | */ | |
48 | struct io_worker { | |
49 | refcount_t ref; | |
50 | unsigned flags; | |
51 | struct hlist_nulls_node nulls_node; | |
e61df66c | 52 | struct list_head all_list; |
771b53d0 | 53 | struct task_struct *task; |
771b53d0 | 54 | struct io_wqe *wqe; |
36c2f922 | 55 | |
771b53d0 | 56 | struct io_wq_work *cur_work; |
36c2f922 | 57 | spinlock_t lock; |
771b53d0 JA |
58 | |
59 | struct rcu_head rcu; | |
60 | struct mm_struct *mm; | |
cccf0ee8 JA |
61 | const struct cred *cur_creds; |
62 | const struct cred *saved_creds; | |
fcb323cc | 63 | struct files_struct *restore_files; |
9392a27d | 64 | struct fs_struct *restore_fs; |
771b53d0 JA |
65 | }; |
66 | ||
771b53d0 JA |
67 | #if BITS_PER_LONG == 64 |
68 | #define IO_WQ_HASH_ORDER 6 | |
69 | #else | |
70 | #define IO_WQ_HASH_ORDER 5 | |
71 | #endif | |
72 | ||
86f3cd1b PB |
73 | #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER) |
74 | ||
c5def4ab JA |
75 | struct io_wqe_acct { |
76 | unsigned nr_workers; | |
77 | unsigned max_workers; | |
78 | atomic_t nr_running; | |
79 | }; | |
80 | ||
81 | enum { | |
82 | IO_WQ_ACCT_BOUND, | |
83 | IO_WQ_ACCT_UNBOUND, | |
84 | }; | |
85 | ||
771b53d0 JA |
86 | /* |
87 | * Per-node worker thread pool | |
88 | */ | |
89 | struct io_wqe { | |
90 | struct { | |
91 | spinlock_t lock; | |
6206f0e1 | 92 | struct io_wq_work_list work_list; |
771b53d0 JA |
93 | unsigned long hash_map; |
94 | unsigned flags; | |
95 | } ____cacheline_aligned_in_smp; | |
96 | ||
97 | int node; | |
c5def4ab | 98 | struct io_wqe_acct acct[2]; |
771b53d0 | 99 | |
021d1cdd | 100 | struct hlist_nulls_head free_list; |
e61df66c | 101 | struct list_head all_list; |
771b53d0 JA |
102 | |
103 | struct io_wq *wq; | |
86f3cd1b | 104 | struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS]; |
771b53d0 JA |
105 | }; |
106 | ||
107 | /* | |
108 | * Per io_wq state | |
109 | */ | |
110 | struct io_wq { | |
111 | struct io_wqe **wqes; | |
112 | unsigned long state; | |
771b53d0 | 113 | |
e9fd9396 | 114 | free_work_fn *free_work; |
7d723065 | 115 | |
771b53d0 | 116 | struct task_struct *manager; |
c5def4ab | 117 | struct user_struct *user; |
771b53d0 JA |
118 | refcount_t refs; |
119 | struct completion done; | |
848f7e18 JA |
120 | |
121 | refcount_t use_refs; | |
771b53d0 JA |
122 | }; |
123 | ||
771b53d0 JA |
124 | static bool io_worker_get(struct io_worker *worker) |
125 | { | |
126 | return refcount_inc_not_zero(&worker->ref); | |
127 | } | |
128 | ||
129 | static void io_worker_release(struct io_worker *worker) | |
130 | { | |
131 | if (refcount_dec_and_test(&worker->ref)) | |
132 | wake_up_process(worker->task); | |
133 | } | |
134 | ||
135 | /* | |
136 | * Note: drops the wqe->lock if returning true! The caller must re-acquire | |
137 | * the lock in that case. Some callers need to restart handling if this | |
138 | * happens, so we can't just re-acquire the lock on behalf of the caller. | |
139 | */ | |
140 | static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) | |
141 | { | |
fcb323cc JA |
142 | bool dropped_lock = false; |
143 | ||
cccf0ee8 JA |
144 | if (worker->saved_creds) { |
145 | revert_creds(worker->saved_creds); | |
146 | worker->cur_creds = worker->saved_creds = NULL; | |
181e448d JA |
147 | } |
148 | ||
fcb323cc JA |
149 | if (current->files != worker->restore_files) { |
150 | __acquire(&wqe->lock); | |
151 | spin_unlock_irq(&wqe->lock); | |
152 | dropped_lock = true; | |
153 | ||
154 | task_lock(current); | |
155 | current->files = worker->restore_files; | |
156 | task_unlock(current); | |
157 | } | |
158 | ||
9392a27d JA |
159 | if (current->fs != worker->restore_fs) |
160 | current->fs = worker->restore_fs; | |
161 | ||
771b53d0 JA |
162 | /* |
163 | * If we have an active mm, we need to drop the wq lock before unusing | |
164 | * it. If we do, return true and let the caller retry the idle loop. | |
165 | */ | |
166 | if (worker->mm) { | |
fcb323cc JA |
167 | if (!dropped_lock) { |
168 | __acquire(&wqe->lock); | |
169 | spin_unlock_irq(&wqe->lock); | |
170 | dropped_lock = true; | |
171 | } | |
771b53d0 JA |
172 | __set_current_state(TASK_RUNNING); |
173 | set_fs(KERNEL_DS); | |
174 | unuse_mm(worker->mm); | |
175 | mmput(worker->mm); | |
176 | worker->mm = NULL; | |
771b53d0 JA |
177 | } |
178 | ||
fcb323cc | 179 | return dropped_lock; |
771b53d0 JA |
180 | } |
181 | ||
c5def4ab JA |
182 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
183 | struct io_wq_work *work) | |
184 | { | |
185 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
186 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
187 | ||
188 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
189 | } | |
190 | ||
191 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe, | |
192 | struct io_worker *worker) | |
193 | { | |
194 | if (worker->flags & IO_WORKER_F_BOUND) | |
195 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
196 | ||
197 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
198 | } | |
199 | ||
771b53d0 JA |
200 | static void io_worker_exit(struct io_worker *worker) |
201 | { | |
202 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab JA |
203 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); |
204 | unsigned nr_workers; | |
771b53d0 JA |
205 | |
206 | /* | |
207 | * If we're not at zero, someone else is holding a brief reference | |
208 | * to the worker. Wait for that to go away. | |
209 | */ | |
210 | set_current_state(TASK_INTERRUPTIBLE); | |
211 | if (!refcount_dec_and_test(&worker->ref)) | |
212 | schedule(); | |
213 | __set_current_state(TASK_RUNNING); | |
214 | ||
215 | preempt_disable(); | |
216 | current->flags &= ~PF_IO_WORKER; | |
217 | if (worker->flags & IO_WORKER_F_RUNNING) | |
c5def4ab JA |
218 | atomic_dec(&acct->nr_running); |
219 | if (!(worker->flags & IO_WORKER_F_BOUND)) | |
220 | atomic_dec(&wqe->wq->user->processes); | |
771b53d0 JA |
221 | worker->flags = 0; |
222 | preempt_enable(); | |
223 | ||
224 | spin_lock_irq(&wqe->lock); | |
225 | hlist_nulls_del_rcu(&worker->nulls_node); | |
e61df66c | 226 | list_del_rcu(&worker->all_list); |
771b53d0 JA |
227 | if (__io_worker_unuse(wqe, worker)) { |
228 | __release(&wqe->lock); | |
229 | spin_lock_irq(&wqe->lock); | |
230 | } | |
c5def4ab JA |
231 | acct->nr_workers--; |
232 | nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers + | |
233 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers; | |
771b53d0 JA |
234 | spin_unlock_irq(&wqe->lock); |
235 | ||
236 | /* all workers gone, wq exit can proceed */ | |
c5def4ab | 237 | if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) |
771b53d0 JA |
238 | complete(&wqe->wq->done); |
239 | ||
364b05fd | 240 | kfree_rcu(worker, rcu); |
771b53d0 JA |
241 | } |
242 | ||
c5def4ab JA |
243 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
244 | __must_hold(wqe->lock) | |
245 | { | |
6206f0e1 JA |
246 | if (!wq_list_empty(&wqe->work_list) && |
247 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
248 | return true; |
249 | return false; | |
250 | } | |
251 | ||
252 | /* | |
253 | * Check head of free list for an available worker. If one isn't available, | |
254 | * caller must wake up the wq manager to create one. | |
255 | */ | |
256 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
257 | __must_hold(RCU) | |
258 | { | |
259 | struct hlist_nulls_node *n; | |
260 | struct io_worker *worker; | |
261 | ||
021d1cdd | 262 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
263 | if (is_a_nulls(n)) |
264 | return false; | |
265 | ||
266 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
267 | if (io_worker_get(worker)) { | |
506d95ff | 268 | wake_up_process(worker->task); |
c5def4ab JA |
269 | io_worker_release(worker); |
270 | return true; | |
271 | } | |
272 | ||
273 | return false; | |
274 | } | |
275 | ||
276 | /* | |
277 | * We need a worker. If we find a free one, we're good. If not, and we're | |
278 | * below the max number of workers, wake up the manager to create one. | |
279 | */ | |
280 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
281 | { | |
282 | bool ret; | |
283 | ||
284 | /* | |
285 | * Most likely an attempt to queue unbounded work on an io_wq that | |
286 | * wasn't setup with any unbounded workers. | |
287 | */ | |
288 | WARN_ON_ONCE(!acct->max_workers); | |
289 | ||
290 | rcu_read_lock(); | |
291 | ret = io_wqe_activate_free_worker(wqe); | |
292 | rcu_read_unlock(); | |
293 | ||
294 | if (!ret && acct->nr_workers < acct->max_workers) | |
295 | wake_up_process(wqe->wq->manager); | |
296 | } | |
297 | ||
298 | static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker) | |
299 | { | |
300 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
301 | ||
302 | atomic_inc(&acct->nr_running); | |
303 | } | |
304 | ||
305 | static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker) | |
306 | __must_hold(wqe->lock) | |
307 | { | |
308 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
309 | ||
310 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
311 | io_wqe_wake_worker(wqe, acct); | |
312 | } | |
313 | ||
771b53d0 JA |
314 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
315 | { | |
316 | allow_kernel_signal(SIGINT); | |
317 | ||
318 | current->flags |= PF_IO_WORKER; | |
319 | ||
320 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); | |
fcb323cc | 321 | worker->restore_files = current->files; |
9392a27d | 322 | worker->restore_fs = current->fs; |
c5def4ab | 323 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
324 | } |
325 | ||
326 | /* | |
327 | * Worker will start processing some work. Move it to the busy list, if | |
328 | * it's currently on the freelist | |
329 | */ | |
330 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
331 | struct io_wq_work *work) | |
332 | __must_hold(wqe->lock) | |
333 | { | |
c5def4ab JA |
334 | bool worker_bound, work_bound; |
335 | ||
771b53d0 JA |
336 | if (worker->flags & IO_WORKER_F_FREE) { |
337 | worker->flags &= ~IO_WORKER_F_FREE; | |
338 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 339 | } |
c5def4ab JA |
340 | |
341 | /* | |
342 | * If worker is moving from bound to unbound (or vice versa), then | |
343 | * ensure we update the running accounting. | |
344 | */ | |
b2e9c7d6 DC |
345 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
346 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
347 | if (worker_bound != work_bound) { | |
c5def4ab JA |
348 | io_wqe_dec_running(wqe, worker); |
349 | if (work_bound) { | |
350 | worker->flags |= IO_WORKER_F_BOUND; | |
351 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
352 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
353 | atomic_dec(&wqe->wq->user->processes); | |
354 | } else { | |
355 | worker->flags &= ~IO_WORKER_F_BOUND; | |
356 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
357 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
358 | atomic_inc(&wqe->wq->user->processes); | |
359 | } | |
360 | io_wqe_inc_running(wqe, worker); | |
361 | } | |
771b53d0 JA |
362 | } |
363 | ||
364 | /* | |
365 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
366 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
367 | * the lock in that case and return success. Since the caller has to | |
368 | * retry the loop in that case (we changed task state), we don't regrab | |
369 | * the lock if we return success. | |
370 | */ | |
371 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) | |
372 | __must_hold(wqe->lock) | |
373 | { | |
374 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
375 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 376 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 JA |
377 | } |
378 | ||
379 | return __io_worker_unuse(wqe, worker); | |
380 | } | |
381 | ||
60cf46ae PB |
382 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
383 | { | |
384 | return work->flags >> IO_WQ_HASH_SHIFT; | |
385 | } | |
386 | ||
387 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe) | |
771b53d0 JA |
388 | __must_hold(wqe->lock) |
389 | { | |
6206f0e1 | 390 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 391 | struct io_wq_work *work, *tail; |
60cf46ae | 392 | unsigned int hash; |
771b53d0 | 393 | |
6206f0e1 JA |
394 | wq_list_for_each(node, prev, &wqe->work_list) { |
395 | work = container_of(node, struct io_wq_work, list); | |
396 | ||
771b53d0 | 397 | /* not hashed, can run anytime */ |
8766dd51 | 398 | if (!io_wq_is_hashed(work)) { |
86f3cd1b | 399 | wq_list_del(&wqe->work_list, node, prev); |
771b53d0 JA |
400 | return work; |
401 | } | |
402 | ||
403 | /* hashed, can run if not already running */ | |
60cf46ae PB |
404 | hash = io_get_work_hash(work); |
405 | if (!(wqe->hash_map & BIT(hash))) { | |
406 | wqe->hash_map |= BIT(hash); | |
86f3cd1b PB |
407 | /* all items with this hash lie in [work, tail] */ |
408 | tail = wqe->hash_tail[hash]; | |
409 | wqe->hash_tail[hash] = NULL; | |
410 | wq_list_cut(&wqe->work_list, &tail->list, prev); | |
771b53d0 JA |
411 | return work; |
412 | } | |
413 | } | |
414 | ||
415 | return NULL; | |
416 | } | |
417 | ||
cccf0ee8 JA |
418 | static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work) |
419 | { | |
420 | if (worker->mm) { | |
421 | unuse_mm(worker->mm); | |
422 | mmput(worker->mm); | |
423 | worker->mm = NULL; | |
424 | } | |
425 | if (!work->mm) { | |
426 | set_fs(KERNEL_DS); | |
427 | return; | |
428 | } | |
429 | if (mmget_not_zero(work->mm)) { | |
430 | use_mm(work->mm); | |
431 | if (!worker->mm) | |
432 | set_fs(USER_DS); | |
433 | worker->mm = work->mm; | |
434 | /* hang on to this mm */ | |
435 | work->mm = NULL; | |
436 | return; | |
437 | } | |
438 | ||
439 | /* failed grabbing mm, ensure work gets cancelled */ | |
440 | work->flags |= IO_WQ_WORK_CANCEL; | |
441 | } | |
442 | ||
443 | static void io_wq_switch_creds(struct io_worker *worker, | |
444 | struct io_wq_work *work) | |
445 | { | |
446 | const struct cred *old_creds = override_creds(work->creds); | |
447 | ||
448 | worker->cur_creds = work->creds; | |
449 | if (worker->saved_creds) | |
450 | put_cred(old_creds); /* creds set by previous switch */ | |
451 | else | |
452 | worker->saved_creds = old_creds; | |
453 | } | |
454 | ||
dc026a73 PB |
455 | static void io_impersonate_work(struct io_worker *worker, |
456 | struct io_wq_work *work) | |
457 | { | |
458 | if (work->files && current->files != work->files) { | |
459 | task_lock(current); | |
460 | current->files = work->files; | |
461 | task_unlock(current); | |
462 | } | |
463 | if (work->fs && current->fs != work->fs) | |
464 | current->fs = work->fs; | |
465 | if (work->mm != worker->mm) | |
466 | io_wq_switch_mm(worker, work); | |
467 | if (worker->cur_creds != work->creds) | |
468 | io_wq_switch_creds(worker, work); | |
469 | } | |
470 | ||
471 | static void io_assign_current_work(struct io_worker *worker, | |
472 | struct io_wq_work *work) | |
473 | { | |
d78298e7 PB |
474 | if (work) { |
475 | /* flush pending signals before assigning new work */ | |
476 | if (signal_pending(current)) | |
477 | flush_signals(current); | |
478 | cond_resched(); | |
479 | } | |
dc026a73 PB |
480 | |
481 | spin_lock_irq(&worker->lock); | |
482 | worker->cur_work = work; | |
483 | spin_unlock_irq(&worker->lock); | |
484 | } | |
485 | ||
60cf46ae PB |
486 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
487 | ||
771b53d0 JA |
488 | static void io_worker_handle_work(struct io_worker *worker) |
489 | __releases(wqe->lock) | |
490 | { | |
771b53d0 JA |
491 | struct io_wqe *wqe = worker->wqe; |
492 | struct io_wq *wq = wqe->wq; | |
493 | ||
494 | do { | |
86f3cd1b | 495 | struct io_wq_work *work; |
60cf46ae | 496 | unsigned int hash; |
f462fd36 | 497 | get_next: |
771b53d0 JA |
498 | /* |
499 | * If we got some work, mark us as busy. If we didn't, but | |
500 | * the list isn't empty, it means we stalled on hashed work. | |
501 | * Mark us stalled so we don't keep looking for work when we | |
502 | * can't make progress, any work completion or insertion will | |
503 | * clear the stalled flag. | |
504 | */ | |
60cf46ae | 505 | work = io_get_next_work(wqe); |
771b53d0 JA |
506 | if (work) |
507 | __io_worker_busy(wqe, worker, work); | |
6206f0e1 | 508 | else if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
509 | wqe->flags |= IO_WQE_FLAG_STALLED; |
510 | ||
511 | spin_unlock_irq(&wqe->lock); | |
512 | if (!work) | |
513 | break; | |
58e39319 | 514 | io_assign_current_work(worker, work); |
36c2f922 | 515 | |
dc026a73 PB |
516 | /* handle a whole dependent link */ |
517 | do { | |
86f3cd1b | 518 | struct io_wq_work *old_work, *next_hashed, *linked; |
dc026a73 | 519 | |
86f3cd1b | 520 | next_hashed = wq_next_work(work); |
58e39319 | 521 | io_impersonate_work(worker, work); |
dc026a73 PB |
522 | /* |
523 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable | |
524 | * work, the worker function will do the right thing. | |
525 | */ | |
526 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) | |
527 | work->flags |= IO_WQ_WORK_CANCEL; | |
528 | ||
60cf46ae | 529 | hash = io_get_work_hash(work); |
86f3cd1b PB |
530 | linked = old_work = work; |
531 | linked->func(&linked); | |
532 | linked = (old_work == linked) ? NULL : linked; | |
533 | ||
534 | work = next_hashed; | |
535 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
536 | work = linked; | |
537 | linked = NULL; | |
538 | } | |
539 | io_assign_current_work(worker, work); | |
e9fd9396 | 540 | wq->free_work(old_work); |
dc026a73 | 541 | |
86f3cd1b PB |
542 | if (linked) |
543 | io_wqe_enqueue(wqe, linked); | |
544 | ||
545 | if (hash != -1U && !next_hashed) { | |
dc026a73 PB |
546 | spin_lock_irq(&wqe->lock); |
547 | wqe->hash_map &= ~BIT_ULL(hash); | |
548 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
dc026a73 PB |
549 | /* dependent work is not hashed */ |
550 | hash = -1U; | |
f462fd36 PB |
551 | /* skip unnecessary unlock-lock wqe->lock */ |
552 | if (!work) | |
553 | goto get_next; | |
554 | spin_unlock_irq(&wqe->lock); | |
7d723065 | 555 | } |
58e39319 | 556 | } while (work); |
7d723065 | 557 | |
dc026a73 | 558 | spin_lock_irq(&wqe->lock); |
771b53d0 JA |
559 | } while (1); |
560 | } | |
561 | ||
771b53d0 JA |
562 | static int io_wqe_worker(void *data) |
563 | { | |
564 | struct io_worker *worker = data; | |
565 | struct io_wqe *wqe = worker->wqe; | |
566 | struct io_wq *wq = wqe->wq; | |
771b53d0 JA |
567 | |
568 | io_worker_start(wqe, worker); | |
569 | ||
570 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
506d95ff | 571 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 572 | loop: |
771b53d0 JA |
573 | spin_lock_irq(&wqe->lock); |
574 | if (io_wqe_run_queue(wqe)) { | |
575 | __set_current_state(TASK_RUNNING); | |
576 | io_worker_handle_work(worker); | |
e995d512 | 577 | goto loop; |
771b53d0 JA |
578 | } |
579 | /* drops the lock on success, retry */ | |
580 | if (__io_worker_idle(wqe, worker)) { | |
581 | __release(&wqe->lock); | |
e995d512 | 582 | goto loop; |
771b53d0 JA |
583 | } |
584 | spin_unlock_irq(&wqe->lock); | |
585 | if (signal_pending(current)) | |
586 | flush_signals(current); | |
587 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
588 | continue; | |
589 | /* timed out, exit unless we're the fixed worker */ | |
590 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
591 | !(worker->flags & IO_WORKER_F_FIXED)) | |
592 | break; | |
593 | } | |
594 | ||
771b53d0 JA |
595 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
596 | spin_lock_irq(&wqe->lock); | |
6206f0e1 | 597 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
598 | io_worker_handle_work(worker); |
599 | else | |
600 | spin_unlock_irq(&wqe->lock); | |
601 | } | |
602 | ||
603 | io_worker_exit(worker); | |
604 | return 0; | |
605 | } | |
606 | ||
771b53d0 JA |
607 | /* |
608 | * Called when a worker is scheduled in. Mark us as currently running. | |
609 | */ | |
610 | void io_wq_worker_running(struct task_struct *tsk) | |
611 | { | |
612 | struct io_worker *worker = kthread_data(tsk); | |
613 | struct io_wqe *wqe = worker->wqe; | |
614 | ||
615 | if (!(worker->flags & IO_WORKER_F_UP)) | |
616 | return; | |
617 | if (worker->flags & IO_WORKER_F_RUNNING) | |
618 | return; | |
619 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 620 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
621 | } |
622 | ||
623 | /* | |
624 | * Called when worker is going to sleep. If there are no workers currently | |
625 | * running and we have work pending, wake up a free one or have the manager | |
626 | * set one up. | |
627 | */ | |
628 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
629 | { | |
630 | struct io_worker *worker = kthread_data(tsk); | |
631 | struct io_wqe *wqe = worker->wqe; | |
632 | ||
633 | if (!(worker->flags & IO_WORKER_F_UP)) | |
634 | return; | |
635 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
636 | return; | |
637 | ||
638 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
639 | ||
640 | spin_lock_irq(&wqe->lock); | |
c5def4ab | 641 | io_wqe_dec_running(wqe, worker); |
771b53d0 JA |
642 | spin_unlock_irq(&wqe->lock); |
643 | } | |
644 | ||
b60fda60 | 645 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 646 | { |
c5def4ab | 647 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
648 | struct io_worker *worker; |
649 | ||
ad6e005c | 650 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 651 | if (!worker) |
b60fda60 | 652 | return false; |
771b53d0 JA |
653 | |
654 | refcount_set(&worker->ref, 1); | |
655 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 656 | worker->wqe = wqe; |
36c2f922 | 657 | spin_lock_init(&worker->lock); |
771b53d0 JA |
658 | |
659 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 660 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
661 | if (IS_ERR(worker->task)) { |
662 | kfree(worker); | |
b60fda60 | 663 | return false; |
771b53d0 JA |
664 | } |
665 | ||
666 | spin_lock_irq(&wqe->lock); | |
021d1cdd | 667 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 668 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 669 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
670 | if (index == IO_WQ_ACCT_BOUND) |
671 | worker->flags |= IO_WORKER_F_BOUND; | |
672 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 673 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 674 | acct->nr_workers++; |
771b53d0 JA |
675 | spin_unlock_irq(&wqe->lock); |
676 | ||
c5def4ab JA |
677 | if (index == IO_WQ_ACCT_UNBOUND) |
678 | atomic_inc(&wq->user->processes); | |
679 | ||
771b53d0 | 680 | wake_up_process(worker->task); |
b60fda60 | 681 | return true; |
771b53d0 JA |
682 | } |
683 | ||
c5def4ab | 684 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
685 | __must_hold(wqe->lock) |
686 | { | |
c5def4ab | 687 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 688 | |
c5def4ab | 689 | /* if we have available workers or no work, no need */ |
021d1cdd | 690 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
691 | return false; |
692 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
693 | } |
694 | ||
695 | /* | |
696 | * Manager thread. Tasked with creating new workers, if we need them. | |
697 | */ | |
698 | static int io_wq_manager(void *data) | |
699 | { | |
700 | struct io_wq *wq = data; | |
3fc50ab5 JH |
701 | int workers_to_create = num_possible_nodes(); |
702 | int node; | |
771b53d0 | 703 | |
b60fda60 | 704 | /* create fixed workers */ |
3fc50ab5 JH |
705 | refcount_set(&wq->refs, workers_to_create); |
706 | for_each_node(node) { | |
7563439a JA |
707 | if (!node_online(node)) |
708 | continue; | |
3fc50ab5 JH |
709 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) |
710 | goto err; | |
711 | workers_to_create--; | |
b60fda60 | 712 | } |
771b53d0 | 713 | |
7563439a JA |
714 | while (workers_to_create--) |
715 | refcount_dec(&wq->refs); | |
716 | ||
b60fda60 JA |
717 | complete(&wq->done); |
718 | ||
719 | while (!kthread_should_stop()) { | |
aa96bf8a JA |
720 | if (current->task_works) |
721 | task_work_run(); | |
722 | ||
3fc50ab5 JH |
723 | for_each_node(node) { |
724 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 725 | bool fork_worker[2] = { false, false }; |
771b53d0 | 726 | |
7563439a JA |
727 | if (!node_online(node)) |
728 | continue; | |
729 | ||
771b53d0 | 730 | spin_lock_irq(&wqe->lock); |
c5def4ab JA |
731 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
732 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
733 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
734 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
771b53d0 | 735 | spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
736 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
737 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
738 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
739 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
740 | } |
741 | set_current_state(TASK_INTERRUPTIBLE); | |
742 | schedule_timeout(HZ); | |
743 | } | |
744 | ||
aa96bf8a JA |
745 | if (current->task_works) |
746 | task_work_run(); | |
747 | ||
771b53d0 | 748 | return 0; |
b60fda60 JA |
749 | err: |
750 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
751 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 752 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
753 | complete(&wq->done); |
754 | return 0; | |
771b53d0 JA |
755 | } |
756 | ||
c5def4ab JA |
757 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
758 | struct io_wq_work *work) | |
759 | { | |
760 | bool free_worker; | |
761 | ||
762 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
763 | return true; | |
764 | if (atomic_read(&acct->nr_running)) | |
765 | return true; | |
766 | ||
767 | rcu_read_lock(); | |
021d1cdd | 768 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
769 | rcu_read_unlock(); |
770 | if (free_worker) | |
771 | return true; | |
772 | ||
773 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
774 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
775 | return false; | |
776 | ||
777 | return true; | |
778 | } | |
779 | ||
e9fd9396 | 780 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 781 | { |
e9fd9396 PB |
782 | struct io_wq *wq = wqe->wq; |
783 | ||
fc04c39b PB |
784 | do { |
785 | struct io_wq_work *old_work = work; | |
786 | ||
787 | work->flags |= IO_WQ_WORK_CANCEL; | |
788 | work->func(&work); | |
789 | work = (work == old_work) ? NULL : work; | |
e9fd9396 | 790 | wq->free_work(old_work); |
fc04c39b PB |
791 | } while (work); |
792 | } | |
793 | ||
86f3cd1b PB |
794 | static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work) |
795 | { | |
796 | unsigned int hash; | |
797 | struct io_wq_work *tail; | |
798 | ||
799 | if (!io_wq_is_hashed(work)) { | |
800 | append: | |
801 | wq_list_add_tail(&work->list, &wqe->work_list); | |
802 | return; | |
803 | } | |
804 | ||
805 | hash = io_get_work_hash(work); | |
806 | tail = wqe->hash_tail[hash]; | |
807 | wqe->hash_tail[hash] = work; | |
808 | if (!tail) | |
809 | goto append; | |
810 | ||
811 | wq_list_add_after(&work->list, &tail->list, &wqe->work_list); | |
812 | } | |
813 | ||
771b53d0 JA |
814 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
815 | { | |
c5def4ab | 816 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 817 | int work_flags; |
771b53d0 JA |
818 | unsigned long flags; |
819 | ||
c5def4ab JA |
820 | /* |
821 | * Do early check to see if we need a new unbound worker, and if we do, | |
822 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
823 | * gap between this check and incrementing the value, but that's OK. | |
824 | * It's close enough to not be an issue, fork() has the same delay. | |
825 | */ | |
826 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
e9fd9396 | 827 | io_run_cancel(work, wqe); |
c5def4ab JA |
828 | return; |
829 | } | |
830 | ||
895e2ca0 | 831 | work_flags = work->flags; |
771b53d0 | 832 | spin_lock_irqsave(&wqe->lock, flags); |
86f3cd1b | 833 | io_wqe_insert_work(wqe, work); |
771b53d0 JA |
834 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
835 | spin_unlock_irqrestore(&wqe->lock, flags); | |
836 | ||
895e2ca0 JA |
837 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
838 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 839 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
840 | } |
841 | ||
842 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
843 | { | |
844 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
845 | ||
846 | io_wqe_enqueue(wqe, work); | |
847 | } | |
848 | ||
849 | /* | |
8766dd51 PB |
850 | * Work items that hash to the same value will not be done in parallel. |
851 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 852 | */ |
8766dd51 | 853 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 854 | { |
8766dd51 | 855 | unsigned int bit; |
771b53d0 JA |
856 | |
857 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
858 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
859 | } |
860 | ||
861 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
862 | { | |
863 | send_sig(SIGINT, worker->task, 1); | |
864 | return false; | |
865 | } | |
866 | ||
867 | /* | |
868 | * Iterate the passed in list and call the specific function for each | |
869 | * worker that isn't exiting | |
870 | */ | |
871 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
872 | bool (*func)(struct io_worker *, void *), |
873 | void *data) | |
874 | { | |
771b53d0 JA |
875 | struct io_worker *worker; |
876 | bool ret = false; | |
877 | ||
e61df66c | 878 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 | 879 | if (io_worker_get(worker)) { |
7563439a JA |
880 | /* no task if node is/was offline */ |
881 | if (worker->task) | |
882 | ret = func(worker, data); | |
771b53d0 JA |
883 | io_worker_release(worker); |
884 | if (ret) | |
885 | break; | |
886 | } | |
887 | } | |
e61df66c | 888 | |
771b53d0 JA |
889 | return ret; |
890 | } | |
891 | ||
892 | void io_wq_cancel_all(struct io_wq *wq) | |
893 | { | |
3fc50ab5 | 894 | int node; |
771b53d0 JA |
895 | |
896 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
897 | ||
771b53d0 | 898 | rcu_read_lock(); |
3fc50ab5 JH |
899 | for_each_node(node) { |
900 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 901 | |
e61df66c | 902 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
903 | } |
904 | rcu_read_unlock(); | |
905 | } | |
906 | ||
62755e35 | 907 | struct io_cb_cancel_data { |
2293b419 PB |
908 | work_cancel_fn *fn; |
909 | void *data; | |
62755e35 JA |
910 | }; |
911 | ||
2293b419 | 912 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 913 | { |
2293b419 | 914 | struct io_cb_cancel_data *match = data; |
6f72653e | 915 | unsigned long flags; |
62755e35 JA |
916 | bool ret = false; |
917 | ||
918 | /* | |
919 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 920 | * may dereference the passed in work. |
62755e35 | 921 | */ |
36c2f922 | 922 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 923 | if (worker->cur_work && |
0c9d5ccd | 924 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
2293b419 | 925 | match->fn(worker->cur_work, match->data)) { |
771b53d0 | 926 | send_sig(SIGINT, worker->task, 1); |
36c2f922 | 927 | ret = true; |
771b53d0 | 928 | } |
36c2f922 | 929 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 930 | |
36c2f922 | 931 | return ret; |
771b53d0 JA |
932 | } |
933 | ||
934 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, | |
2293b419 | 935 | struct io_cb_cancel_data *match) |
771b53d0 | 936 | { |
6206f0e1 | 937 | struct io_wq_work_node *node, *prev; |
771b53d0 | 938 | struct io_wq_work *work; |
6f72653e | 939 | unsigned long flags; |
771b53d0 JA |
940 | bool found = false; |
941 | ||
771b53d0 JA |
942 | /* |
943 | * First check pending list, if we're lucky we can just remove it | |
944 | * from there. CANCEL_OK means that the work is returned as-new, | |
945 | * no completion will be posted for it. | |
946 | */ | |
6f72653e | 947 | spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
948 | wq_list_for_each(node, prev, &wqe->work_list) { |
949 | work = container_of(node, struct io_wq_work, list); | |
950 | ||
00bcda13 | 951 | if (match->fn(work, match->data)) { |
86f3cd1b | 952 | wq_list_del(&wqe->work_list, node, prev); |
771b53d0 JA |
953 | found = true; |
954 | break; | |
955 | } | |
956 | } | |
6f72653e | 957 | spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 JA |
958 | |
959 | if (found) { | |
e9fd9396 | 960 | io_run_cancel(work, wqe); |
771b53d0 JA |
961 | return IO_WQ_CANCEL_OK; |
962 | } | |
963 | ||
964 | /* | |
965 | * Now check if a free (going busy) or busy worker has the work | |
966 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
d195a66e | 967 | * as an indication that we attempt to signal cancellation. The |
771b53d0 JA |
968 | * completion will run normally in this case. |
969 | */ | |
970 | rcu_read_lock(); | |
00bcda13 | 971 | found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 JA |
972 | rcu_read_unlock(); |
973 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; | |
974 | } | |
975 | ||
2293b419 PB |
976 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
977 | void *data) | |
771b53d0 | 978 | { |
2293b419 PB |
979 | struct io_cb_cancel_data match = { |
980 | .fn = cancel, | |
981 | .data = data, | |
00bcda13 | 982 | }; |
771b53d0 | 983 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; |
3fc50ab5 | 984 | int node; |
771b53d0 | 985 | |
3fc50ab5 JH |
986 | for_each_node(node) { |
987 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 988 | |
00bcda13 | 989 | ret = io_wqe_cancel_work(wqe, &match); |
771b53d0 JA |
990 | if (ret != IO_WQ_CANCEL_NOTFOUND) |
991 | break; | |
992 | } | |
993 | ||
994 | return ret; | |
995 | } | |
996 | ||
2293b419 PB |
997 | static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data) |
998 | { | |
999 | return work == data; | |
1000 | } | |
1001 | ||
1002 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
1003 | { | |
1004 | return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork); | |
1005 | } | |
1006 | ||
36282881 JA |
1007 | static bool io_wq_pid_match(struct io_wq_work *work, void *data) |
1008 | { | |
1009 | pid_t pid = (pid_t) (unsigned long) data; | |
1010 | ||
2293b419 | 1011 | return work->task_pid == pid; |
36282881 JA |
1012 | } |
1013 | ||
1014 | enum io_wq_cancel io_wq_cancel_pid(struct io_wq *wq, pid_t pid) | |
1015 | { | |
2293b419 | 1016 | void *data = (void *) (unsigned long) pid; |
36282881 | 1017 | |
2293b419 | 1018 | return io_wq_cancel_cb(wq, io_wq_pid_match, data); |
36282881 JA |
1019 | } |
1020 | ||
576a347b | 1021 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1022 | { |
3fc50ab5 | 1023 | int ret = -ENOMEM, node; |
771b53d0 JA |
1024 | struct io_wq *wq; |
1025 | ||
e9fd9396 PB |
1026 | if (WARN_ON_ONCE(!data->free_work)) |
1027 | return ERR_PTR(-EINVAL); | |
1028 | ||
ad6e005c | 1029 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1030 | if (!wq) |
1031 | return ERR_PTR(-ENOMEM); | |
1032 | ||
3fc50ab5 | 1033 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1034 | if (!wq->wqes) { |
1035 | kfree(wq); | |
1036 | return ERR_PTR(-ENOMEM); | |
1037 | } | |
1038 | ||
e9fd9396 | 1039 | wq->free_work = data->free_work; |
7d723065 | 1040 | |
c5def4ab | 1041 | /* caller must already hold a reference to this */ |
576a347b | 1042 | wq->user = data->user; |
c5def4ab | 1043 | |
3fc50ab5 | 1044 | for_each_node(node) { |
771b53d0 | 1045 | struct io_wqe *wqe; |
7563439a | 1046 | int alloc_node = node; |
771b53d0 | 1047 | |
7563439a JA |
1048 | if (!node_online(alloc_node)) |
1049 | alloc_node = NUMA_NO_NODE; | |
1050 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1051 | if (!wqe) |
3fc50ab5 JH |
1052 | goto err; |
1053 | wq->wqes[node] = wqe; | |
7563439a | 1054 | wqe->node = alloc_node; |
c5def4ab JA |
1055 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1056 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1057 | if (wq->user) { |
c5def4ab JA |
1058 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1059 | task_rlimit(current, RLIMIT_NPROC); | |
1060 | } | |
1061 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 JA |
1062 | wqe->wq = wq; |
1063 | spin_lock_init(&wqe->lock); | |
6206f0e1 | 1064 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1065 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1066 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1067 | } |
1068 | ||
1069 | init_completion(&wq->done); | |
1070 | ||
771b53d0 JA |
1071 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
1072 | if (!IS_ERR(wq->manager)) { | |
1073 | wake_up_process(wq->manager); | |
b60fda60 JA |
1074 | wait_for_completion(&wq->done); |
1075 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1076 | ret = -ENOMEM; | |
1077 | goto err; | |
1078 | } | |
848f7e18 | 1079 | refcount_set(&wq->use_refs, 1); |
b60fda60 | 1080 | reinit_completion(&wq->done); |
771b53d0 JA |
1081 | return wq; |
1082 | } | |
1083 | ||
1084 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1085 | complete(&wq->done); |
b60fda60 | 1086 | err: |
3fc50ab5 JH |
1087 | for_each_node(node) |
1088 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1089 | kfree(wq->wqes); |
1090 | kfree(wq); | |
771b53d0 JA |
1091 | return ERR_PTR(ret); |
1092 | } | |
1093 | ||
eba6f5a3 PB |
1094 | bool io_wq_get(struct io_wq *wq, struct io_wq_data *data) |
1095 | { | |
e9fd9396 | 1096 | if (data->free_work != wq->free_work) |
eba6f5a3 PB |
1097 | return false; |
1098 | ||
1099 | return refcount_inc_not_zero(&wq->use_refs); | |
1100 | } | |
1101 | ||
771b53d0 JA |
1102 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) |
1103 | { | |
1104 | wake_up_process(worker->task); | |
1105 | return false; | |
1106 | } | |
1107 | ||
848f7e18 | 1108 | static void __io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1109 | { |
3fc50ab5 | 1110 | int node; |
771b53d0 | 1111 | |
b60fda60 JA |
1112 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1113 | if (wq->manager) | |
771b53d0 | 1114 | kthread_stop(wq->manager); |
771b53d0 JA |
1115 | |
1116 | rcu_read_lock(); | |
3fc50ab5 JH |
1117 | for_each_node(node) |
1118 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1119 | rcu_read_unlock(); |
1120 | ||
1121 | wait_for_completion(&wq->done); | |
1122 | ||
3fc50ab5 JH |
1123 | for_each_node(node) |
1124 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1125 | kfree(wq->wqes); |
1126 | kfree(wq); | |
1127 | } | |
848f7e18 JA |
1128 | |
1129 | void io_wq_destroy(struct io_wq *wq) | |
1130 | { | |
1131 | if (refcount_dec_and_test(&wq->use_refs)) | |
1132 | __io_wq_destroy(wq); | |
1133 | } | |
aa96bf8a JA |
1134 | |
1135 | struct task_struct *io_wq_get_task(struct io_wq *wq) | |
1136 | { | |
1137 | return wq->manager; | |
1138 | } |