]>
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> | |
771b53d0 JA |
13 | #include <linux/sched/mm.h> |
14 | #include <linux/percpu.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/kthread.h> | |
17 | #include <linux/rculist_nulls.h> | |
9392a27d | 18 | #include <linux/fs_struct.h> |
aa96bf8a | 19 | #include <linux/task_work.h> |
771b53d0 JA |
20 | |
21 | #include "io-wq.h" | |
22 | ||
23 | #define WORKER_IDLE_TIMEOUT (5 * HZ) | |
24 | ||
25 | enum { | |
26 | IO_WORKER_F_UP = 1, /* up and active */ | |
27 | IO_WORKER_F_RUNNING = 2, /* account as running */ | |
28 | IO_WORKER_F_FREE = 4, /* worker on free list */ | |
29 | IO_WORKER_F_EXITING = 8, /* worker exiting */ | |
30 | IO_WORKER_F_FIXED = 16, /* static idle worker */ | |
c5def4ab | 31 | IO_WORKER_F_BOUND = 32, /* is doing bounded work */ |
771b53d0 JA |
32 | }; |
33 | ||
34 | enum { | |
35 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ | |
36 | IO_WQ_BIT_CANCEL = 1, /* cancel work on list */ | |
b60fda60 | 37 | IO_WQ_BIT_ERROR = 2, /* error on setup */ |
771b53d0 JA |
38 | }; |
39 | ||
40 | enum { | |
41 | IO_WQE_FLAG_STALLED = 1, /* stalled on hash */ | |
42 | }; | |
43 | ||
44 | /* | |
45 | * One for each thread in a wqe pool | |
46 | */ | |
47 | struct io_worker { | |
48 | refcount_t ref; | |
49 | unsigned flags; | |
50 | struct hlist_nulls_node nulls_node; | |
e61df66c | 51 | struct list_head all_list; |
771b53d0 | 52 | struct task_struct *task; |
771b53d0 | 53 | struct io_wqe *wqe; |
36c2f922 | 54 | |
771b53d0 | 55 | struct io_wq_work *cur_work; |
36c2f922 | 56 | spinlock_t lock; |
771b53d0 JA |
57 | |
58 | struct rcu_head rcu; | |
59 | struct mm_struct *mm; | |
cccf0ee8 JA |
60 | const struct cred *cur_creds; |
61 | const struct cred *saved_creds; | |
fcb323cc | 62 | struct files_struct *restore_files; |
9b828492 | 63 | struct nsproxy *restore_nsproxy; |
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 { | |
95da8465 | 91 | raw_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; |
f5fa38c5 | 115 | io_wq_work_fn *do_work; |
7d723065 | 116 | |
771b53d0 | 117 | struct task_struct *manager; |
c5def4ab | 118 | struct user_struct *user; |
771b53d0 JA |
119 | refcount_t refs; |
120 | struct completion done; | |
848f7e18 JA |
121 | |
122 | refcount_t use_refs; | |
771b53d0 JA |
123 | }; |
124 | ||
771b53d0 JA |
125 | static bool io_worker_get(struct io_worker *worker) |
126 | { | |
127 | return refcount_inc_not_zero(&worker->ref); | |
128 | } | |
129 | ||
130 | static void io_worker_release(struct io_worker *worker) | |
131 | { | |
132 | if (refcount_dec_and_test(&worker->ref)) | |
133 | wake_up_process(worker->task); | |
134 | } | |
135 | ||
136 | /* | |
137 | * Note: drops the wqe->lock if returning true! The caller must re-acquire | |
138 | * the lock in that case. Some callers need to restart handling if this | |
139 | * happens, so we can't just re-acquire the lock on behalf of the caller. | |
140 | */ | |
141 | static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) | |
142 | { | |
fcb323cc JA |
143 | bool dropped_lock = false; |
144 | ||
cccf0ee8 JA |
145 | if (worker->saved_creds) { |
146 | revert_creds(worker->saved_creds); | |
147 | worker->cur_creds = worker->saved_creds = NULL; | |
181e448d JA |
148 | } |
149 | ||
fcb323cc JA |
150 | if (current->files != worker->restore_files) { |
151 | __acquire(&wqe->lock); | |
95da8465 | 152 | raw_spin_unlock_irq(&wqe->lock); |
fcb323cc JA |
153 | dropped_lock = true; |
154 | ||
155 | task_lock(current); | |
156 | current->files = worker->restore_files; | |
9b828492 | 157 | current->nsproxy = worker->restore_nsproxy; |
fcb323cc JA |
158 | task_unlock(current); |
159 | } | |
160 | ||
9392a27d JA |
161 | if (current->fs != worker->restore_fs) |
162 | current->fs = worker->restore_fs; | |
163 | ||
771b53d0 JA |
164 | /* |
165 | * If we have an active mm, we need to drop the wq lock before unusing | |
166 | * it. If we do, return true and let the caller retry the idle loop. | |
167 | */ | |
168 | if (worker->mm) { | |
fcb323cc JA |
169 | if (!dropped_lock) { |
170 | __acquire(&wqe->lock); | |
95da8465 | 171 | raw_spin_unlock_irq(&wqe->lock); |
fcb323cc JA |
172 | dropped_lock = true; |
173 | } | |
771b53d0 | 174 | __set_current_state(TASK_RUNNING); |
f5678e7f | 175 | kthread_unuse_mm(worker->mm); |
771b53d0 JA |
176 | mmput(worker->mm); |
177 | worker->mm = NULL; | |
771b53d0 JA |
178 | } |
179 | ||
fcb323cc | 180 | return dropped_lock; |
771b53d0 JA |
181 | } |
182 | ||
c5def4ab JA |
183 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
184 | struct io_wq_work *work) | |
185 | { | |
186 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
187 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
188 | ||
189 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
190 | } | |
191 | ||
192 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe, | |
193 | struct io_worker *worker) | |
194 | { | |
195 | if (worker->flags & IO_WORKER_F_BOUND) | |
196 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
197 | ||
198 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
199 | } | |
200 | ||
771b53d0 JA |
201 | static void io_worker_exit(struct io_worker *worker) |
202 | { | |
203 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab JA |
204 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); |
205 | unsigned nr_workers; | |
771b53d0 JA |
206 | |
207 | /* | |
208 | * If we're not at zero, someone else is holding a brief reference | |
209 | * to the worker. Wait for that to go away. | |
210 | */ | |
211 | set_current_state(TASK_INTERRUPTIBLE); | |
212 | if (!refcount_dec_and_test(&worker->ref)) | |
213 | schedule(); | |
214 | __set_current_state(TASK_RUNNING); | |
215 | ||
216 | preempt_disable(); | |
217 | current->flags &= ~PF_IO_WORKER; | |
218 | if (worker->flags & IO_WORKER_F_RUNNING) | |
c5def4ab JA |
219 | atomic_dec(&acct->nr_running); |
220 | if (!(worker->flags & IO_WORKER_F_BOUND)) | |
221 | atomic_dec(&wqe->wq->user->processes); | |
771b53d0 JA |
222 | worker->flags = 0; |
223 | preempt_enable(); | |
224 | ||
95da8465 | 225 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 | 226 | hlist_nulls_del_rcu(&worker->nulls_node); |
e61df66c | 227 | list_del_rcu(&worker->all_list); |
771b53d0 JA |
228 | if (__io_worker_unuse(wqe, worker)) { |
229 | __release(&wqe->lock); | |
95da8465 | 230 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 | 231 | } |
c5def4ab JA |
232 | acct->nr_workers--; |
233 | nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers + | |
234 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers; | |
95da8465 | 235 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
236 | |
237 | /* all workers gone, wq exit can proceed */ | |
c5def4ab | 238 | if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) |
771b53d0 JA |
239 | complete(&wqe->wq->done); |
240 | ||
364b05fd | 241 | kfree_rcu(worker, rcu); |
771b53d0 JA |
242 | } |
243 | ||
c5def4ab JA |
244 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
245 | __must_hold(wqe->lock) | |
246 | { | |
6206f0e1 JA |
247 | if (!wq_list_empty(&wqe->work_list) && |
248 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
249 | return true; |
250 | return false; | |
251 | } | |
252 | ||
253 | /* | |
254 | * Check head of free list for an available worker. If one isn't available, | |
255 | * caller must wake up the wq manager to create one. | |
256 | */ | |
257 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
258 | __must_hold(RCU) | |
259 | { | |
260 | struct hlist_nulls_node *n; | |
261 | struct io_worker *worker; | |
262 | ||
021d1cdd | 263 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
264 | if (is_a_nulls(n)) |
265 | return false; | |
266 | ||
267 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
268 | if (io_worker_get(worker)) { | |
506d95ff | 269 | wake_up_process(worker->task); |
c5def4ab JA |
270 | io_worker_release(worker); |
271 | return true; | |
272 | } | |
273 | ||
274 | return false; | |
275 | } | |
276 | ||
277 | /* | |
278 | * We need a worker. If we find a free one, we're good. If not, and we're | |
279 | * below the max number of workers, wake up the manager to create one. | |
280 | */ | |
281 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
282 | { | |
283 | bool ret; | |
284 | ||
285 | /* | |
286 | * Most likely an attempt to queue unbounded work on an io_wq that | |
287 | * wasn't setup with any unbounded workers. | |
288 | */ | |
289 | WARN_ON_ONCE(!acct->max_workers); | |
290 | ||
291 | rcu_read_lock(); | |
292 | ret = io_wqe_activate_free_worker(wqe); | |
293 | rcu_read_unlock(); | |
294 | ||
295 | if (!ret && acct->nr_workers < acct->max_workers) | |
296 | wake_up_process(wqe->wq->manager); | |
297 | } | |
298 | ||
299 | static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker) | |
300 | { | |
301 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
302 | ||
303 | atomic_inc(&acct->nr_running); | |
304 | } | |
305 | ||
306 | static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker) | |
307 | __must_hold(wqe->lock) | |
308 | { | |
309 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
310 | ||
311 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
312 | io_wqe_wake_worker(wqe, acct); | |
313 | } | |
314 | ||
771b53d0 JA |
315 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
316 | { | |
317 | allow_kernel_signal(SIGINT); | |
318 | ||
319 | current->flags |= PF_IO_WORKER; | |
320 | ||
321 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); | |
fcb323cc | 322 | worker->restore_files = current->files; |
9b828492 | 323 | worker->restore_nsproxy = current->nsproxy; |
9392a27d | 324 | worker->restore_fs = current->fs; |
c5def4ab | 325 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
326 | } |
327 | ||
328 | /* | |
329 | * Worker will start processing some work. Move it to the busy list, if | |
330 | * it's currently on the freelist | |
331 | */ | |
332 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
333 | struct io_wq_work *work) | |
334 | __must_hold(wqe->lock) | |
335 | { | |
c5def4ab JA |
336 | bool worker_bound, work_bound; |
337 | ||
771b53d0 JA |
338 | if (worker->flags & IO_WORKER_F_FREE) { |
339 | worker->flags &= ~IO_WORKER_F_FREE; | |
340 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 341 | } |
c5def4ab JA |
342 | |
343 | /* | |
344 | * If worker is moving from bound to unbound (or vice versa), then | |
345 | * ensure we update the running accounting. | |
346 | */ | |
b2e9c7d6 DC |
347 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
348 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
349 | if (worker_bound != work_bound) { | |
c5def4ab JA |
350 | io_wqe_dec_running(wqe, worker); |
351 | if (work_bound) { | |
352 | worker->flags |= IO_WORKER_F_BOUND; | |
353 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
354 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
355 | atomic_dec(&wqe->wq->user->processes); | |
356 | } else { | |
357 | worker->flags &= ~IO_WORKER_F_BOUND; | |
358 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
359 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
360 | atomic_inc(&wqe->wq->user->processes); | |
361 | } | |
362 | io_wqe_inc_running(wqe, worker); | |
363 | } | |
771b53d0 JA |
364 | } |
365 | ||
366 | /* | |
367 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
368 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
369 | * the lock in that case and return success. Since the caller has to | |
370 | * retry the loop in that case (we changed task state), we don't regrab | |
371 | * the lock if we return success. | |
372 | */ | |
373 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) | |
374 | __must_hold(wqe->lock) | |
375 | { | |
376 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
377 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 378 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 JA |
379 | } |
380 | ||
381 | return __io_worker_unuse(wqe, worker); | |
382 | } | |
383 | ||
60cf46ae PB |
384 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
385 | { | |
386 | return work->flags >> IO_WQ_HASH_SHIFT; | |
387 | } | |
388 | ||
389 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe) | |
771b53d0 JA |
390 | __must_hold(wqe->lock) |
391 | { | |
6206f0e1 | 392 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 393 | struct io_wq_work *work, *tail; |
60cf46ae | 394 | unsigned int hash; |
771b53d0 | 395 | |
6206f0e1 JA |
396 | wq_list_for_each(node, prev, &wqe->work_list) { |
397 | work = container_of(node, struct io_wq_work, list); | |
398 | ||
771b53d0 | 399 | /* not hashed, can run anytime */ |
8766dd51 | 400 | if (!io_wq_is_hashed(work)) { |
86f3cd1b | 401 | wq_list_del(&wqe->work_list, node, prev); |
771b53d0 JA |
402 | return work; |
403 | } | |
404 | ||
405 | /* hashed, can run if not already running */ | |
60cf46ae PB |
406 | hash = io_get_work_hash(work); |
407 | if (!(wqe->hash_map & BIT(hash))) { | |
408 | wqe->hash_map |= BIT(hash); | |
86f3cd1b PB |
409 | /* all items with this hash lie in [work, tail] */ |
410 | tail = wqe->hash_tail[hash]; | |
411 | wqe->hash_tail[hash] = NULL; | |
412 | wq_list_cut(&wqe->work_list, &tail->list, prev); | |
771b53d0 JA |
413 | return work; |
414 | } | |
415 | } | |
416 | ||
417 | return NULL; | |
418 | } | |
419 | ||
cccf0ee8 JA |
420 | static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work) |
421 | { | |
422 | if (worker->mm) { | |
f5678e7f | 423 | kthread_unuse_mm(worker->mm); |
cccf0ee8 JA |
424 | mmput(worker->mm); |
425 | worker->mm = NULL; | |
426 | } | |
37c54f9b | 427 | if (!work->mm) |
cccf0ee8 | 428 | return; |
37c54f9b | 429 | |
cccf0ee8 | 430 | if (mmget_not_zero(work->mm)) { |
f5678e7f | 431 | kthread_use_mm(work->mm); |
cccf0ee8 JA |
432 | worker->mm = work->mm; |
433 | /* hang on to this mm */ | |
434 | work->mm = NULL; | |
435 | return; | |
436 | } | |
437 | ||
438 | /* failed grabbing mm, ensure work gets cancelled */ | |
439 | work->flags |= IO_WQ_WORK_CANCEL; | |
440 | } | |
441 | ||
442 | static void io_wq_switch_creds(struct io_worker *worker, | |
443 | struct io_wq_work *work) | |
444 | { | |
445 | const struct cred *old_creds = override_creds(work->creds); | |
446 | ||
447 | worker->cur_creds = work->creds; | |
448 | if (worker->saved_creds) | |
449 | put_cred(old_creds); /* creds set by previous switch */ | |
450 | else | |
451 | worker->saved_creds = old_creds; | |
452 | } | |
453 | ||
dc026a73 PB |
454 | static void io_impersonate_work(struct io_worker *worker, |
455 | struct io_wq_work *work) | |
456 | { | |
457 | if (work->files && current->files != work->files) { | |
458 | task_lock(current); | |
459 | current->files = work->files; | |
9b828492 | 460 | current->nsproxy = work->nsproxy; |
dc026a73 PB |
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); | |
57f1a649 | 469 | current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->fsize; |
dc026a73 PB |
470 | } |
471 | ||
472 | static void io_assign_current_work(struct io_worker *worker, | |
473 | struct io_wq_work *work) | |
474 | { | |
d78298e7 PB |
475 | if (work) { |
476 | /* flush pending signals before assigning new work */ | |
477 | if (signal_pending(current)) | |
478 | flush_signals(current); | |
479 | cond_resched(); | |
480 | } | |
dc026a73 PB |
481 | |
482 | spin_lock_irq(&worker->lock); | |
483 | worker->cur_work = work; | |
484 | spin_unlock_irq(&worker->lock); | |
485 | } | |
486 | ||
60cf46ae PB |
487 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
488 | ||
771b53d0 JA |
489 | static void io_worker_handle_work(struct io_worker *worker) |
490 | __releases(wqe->lock) | |
491 | { | |
771b53d0 JA |
492 | struct io_wqe *wqe = worker->wqe; |
493 | struct io_wq *wq = wqe->wq; | |
494 | ||
495 | do { | |
86f3cd1b | 496 | struct io_wq_work *work; |
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 | ||
95da8465 | 511 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
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; |
b089ed39 | 519 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 520 | |
86f3cd1b | 521 | next_hashed = wq_next_work(work); |
58e39319 | 522 | io_impersonate_work(worker, work); |
dc026a73 PB |
523 | /* |
524 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable | |
525 | * work, the worker function will do the right thing. | |
526 | */ | |
527 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) | |
528 | work->flags |= IO_WQ_WORK_CANCEL; | |
529 | ||
f4db7182 PB |
530 | old_work = work; |
531 | linked = wq->do_work(work); | |
86f3cd1b PB |
532 | |
533 | work = next_hashed; | |
534 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
535 | work = linked; | |
536 | linked = NULL; | |
537 | } | |
538 | io_assign_current_work(worker, work); | |
e9fd9396 | 539 | wq->free_work(old_work); |
dc026a73 | 540 | |
86f3cd1b PB |
541 | if (linked) |
542 | io_wqe_enqueue(wqe, linked); | |
543 | ||
544 | if (hash != -1U && !next_hashed) { | |
95da8465 | 545 | raw_spin_lock_irq(&wqe->lock); |
dc026a73 PB |
546 | wqe->hash_map &= ~BIT_ULL(hash); |
547 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
f462fd36 PB |
548 | /* skip unnecessary unlock-lock wqe->lock */ |
549 | if (!work) | |
550 | goto get_next; | |
95da8465 | 551 | raw_spin_unlock_irq(&wqe->lock); |
7d723065 | 552 | } |
58e39319 | 553 | } while (work); |
7d723065 | 554 | |
95da8465 | 555 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 JA |
556 | } while (1); |
557 | } | |
558 | ||
771b53d0 JA |
559 | static int io_wqe_worker(void *data) |
560 | { | |
561 | struct io_worker *worker = data; | |
562 | struct io_wqe *wqe = worker->wqe; | |
563 | struct io_wq *wq = wqe->wq; | |
771b53d0 JA |
564 | |
565 | io_worker_start(wqe, worker); | |
566 | ||
567 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
506d95ff | 568 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 569 | loop: |
95da8465 | 570 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 JA |
571 | if (io_wqe_run_queue(wqe)) { |
572 | __set_current_state(TASK_RUNNING); | |
573 | io_worker_handle_work(worker); | |
e995d512 | 574 | goto loop; |
771b53d0 JA |
575 | } |
576 | /* drops the lock on success, retry */ | |
577 | if (__io_worker_idle(wqe, worker)) { | |
578 | __release(&wqe->lock); | |
e995d512 | 579 | goto loop; |
771b53d0 | 580 | } |
95da8465 | 581 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
582 | if (signal_pending(current)) |
583 | flush_signals(current); | |
584 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
585 | continue; | |
586 | /* timed out, exit unless we're the fixed worker */ | |
587 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
588 | !(worker->flags & IO_WORKER_F_FIXED)) | |
589 | break; | |
590 | } | |
591 | ||
771b53d0 | 592 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
95da8465 | 593 | raw_spin_lock_irq(&wqe->lock); |
6206f0e1 | 594 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
595 | io_worker_handle_work(worker); |
596 | else | |
95da8465 | 597 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
598 | } |
599 | ||
600 | io_worker_exit(worker); | |
601 | return 0; | |
602 | } | |
603 | ||
771b53d0 JA |
604 | /* |
605 | * Called when a worker is scheduled in. Mark us as currently running. | |
606 | */ | |
607 | void io_wq_worker_running(struct task_struct *tsk) | |
608 | { | |
609 | struct io_worker *worker = kthread_data(tsk); | |
610 | struct io_wqe *wqe = worker->wqe; | |
611 | ||
612 | if (!(worker->flags & IO_WORKER_F_UP)) | |
613 | return; | |
614 | if (worker->flags & IO_WORKER_F_RUNNING) | |
615 | return; | |
616 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 617 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
618 | } |
619 | ||
620 | /* | |
621 | * Called when worker is going to sleep. If there are no workers currently | |
622 | * running and we have work pending, wake up a free one or have the manager | |
623 | * set one up. | |
624 | */ | |
625 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
626 | { | |
627 | struct io_worker *worker = kthread_data(tsk); | |
628 | struct io_wqe *wqe = worker->wqe; | |
629 | ||
630 | if (!(worker->flags & IO_WORKER_F_UP)) | |
631 | return; | |
632 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
633 | return; | |
634 | ||
635 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
636 | ||
95da8465 | 637 | raw_spin_lock_irq(&wqe->lock); |
c5def4ab | 638 | io_wqe_dec_running(wqe, worker); |
95da8465 | 639 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
640 | } |
641 | ||
b60fda60 | 642 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 643 | { |
c5def4ab | 644 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
645 | struct io_worker *worker; |
646 | ||
ad6e005c | 647 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 648 | if (!worker) |
b60fda60 | 649 | return false; |
771b53d0 JA |
650 | |
651 | refcount_set(&worker->ref, 1); | |
652 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 653 | worker->wqe = wqe; |
36c2f922 | 654 | spin_lock_init(&worker->lock); |
771b53d0 JA |
655 | |
656 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 657 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
658 | if (IS_ERR(worker->task)) { |
659 | kfree(worker); | |
b60fda60 | 660 | return false; |
771b53d0 JA |
661 | } |
662 | ||
95da8465 | 663 | raw_spin_lock_irq(&wqe->lock); |
021d1cdd | 664 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 665 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 666 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
667 | if (index == IO_WQ_ACCT_BOUND) |
668 | worker->flags |= IO_WORKER_F_BOUND; | |
669 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 670 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 671 | acct->nr_workers++; |
95da8465 | 672 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 | 673 | |
c5def4ab JA |
674 | if (index == IO_WQ_ACCT_UNBOUND) |
675 | atomic_inc(&wq->user->processes); | |
676 | ||
771b53d0 | 677 | wake_up_process(worker->task); |
b60fda60 | 678 | return true; |
771b53d0 JA |
679 | } |
680 | ||
c5def4ab | 681 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
682 | __must_hold(wqe->lock) |
683 | { | |
c5def4ab | 684 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 685 | |
c5def4ab | 686 | /* if we have available workers or no work, no need */ |
021d1cdd | 687 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
688 | return false; |
689 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
690 | } |
691 | ||
692 | /* | |
693 | * Manager thread. Tasked with creating new workers, if we need them. | |
694 | */ | |
695 | static int io_wq_manager(void *data) | |
696 | { | |
697 | struct io_wq *wq = data; | |
3fc50ab5 JH |
698 | int workers_to_create = num_possible_nodes(); |
699 | int node; | |
771b53d0 | 700 | |
b60fda60 | 701 | /* create fixed workers */ |
3fc50ab5 JH |
702 | refcount_set(&wq->refs, workers_to_create); |
703 | for_each_node(node) { | |
7563439a JA |
704 | if (!node_online(node)) |
705 | continue; | |
3fc50ab5 JH |
706 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) |
707 | goto err; | |
708 | workers_to_create--; | |
b60fda60 | 709 | } |
771b53d0 | 710 | |
7563439a JA |
711 | while (workers_to_create--) |
712 | refcount_dec(&wq->refs); | |
713 | ||
b60fda60 JA |
714 | complete(&wq->done); |
715 | ||
716 | while (!kthread_should_stop()) { | |
aa96bf8a JA |
717 | if (current->task_works) |
718 | task_work_run(); | |
719 | ||
3fc50ab5 JH |
720 | for_each_node(node) { |
721 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 722 | bool fork_worker[2] = { false, false }; |
771b53d0 | 723 | |
7563439a JA |
724 | if (!node_online(node)) |
725 | continue; | |
726 | ||
95da8465 | 727 | raw_spin_lock_irq(&wqe->lock); |
c5def4ab JA |
728 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
729 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
730 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
731 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
95da8465 | 732 | raw_spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
733 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
734 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
735 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
736 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
737 | } |
738 | set_current_state(TASK_INTERRUPTIBLE); | |
739 | schedule_timeout(HZ); | |
740 | } | |
741 | ||
aa96bf8a JA |
742 | if (current->task_works) |
743 | task_work_run(); | |
744 | ||
771b53d0 | 745 | return 0; |
b60fda60 JA |
746 | err: |
747 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
748 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 749 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
750 | complete(&wq->done); |
751 | return 0; | |
771b53d0 JA |
752 | } |
753 | ||
c5def4ab JA |
754 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
755 | struct io_wq_work *work) | |
756 | { | |
757 | bool free_worker; | |
758 | ||
759 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
760 | return true; | |
761 | if (atomic_read(&acct->nr_running)) | |
762 | return true; | |
763 | ||
764 | rcu_read_lock(); | |
021d1cdd | 765 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
766 | rcu_read_unlock(); |
767 | if (free_worker) | |
768 | return true; | |
769 | ||
770 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
771 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
772 | return false; | |
773 | ||
774 | return true; | |
775 | } | |
776 | ||
e9fd9396 | 777 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 778 | { |
e9fd9396 PB |
779 | struct io_wq *wq = wqe->wq; |
780 | ||
fc04c39b PB |
781 | do { |
782 | struct io_wq_work *old_work = work; | |
783 | ||
784 | work->flags |= IO_WQ_WORK_CANCEL; | |
f4db7182 | 785 | work = wq->do_work(work); |
e9fd9396 | 786 | wq->free_work(old_work); |
fc04c39b PB |
787 | } while (work); |
788 | } | |
789 | ||
86f3cd1b PB |
790 | static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work) |
791 | { | |
792 | unsigned int hash; | |
793 | struct io_wq_work *tail; | |
794 | ||
795 | if (!io_wq_is_hashed(work)) { | |
796 | append: | |
797 | wq_list_add_tail(&work->list, &wqe->work_list); | |
798 | return; | |
799 | } | |
800 | ||
801 | hash = io_get_work_hash(work); | |
802 | tail = wqe->hash_tail[hash]; | |
803 | wqe->hash_tail[hash] = work; | |
804 | if (!tail) | |
805 | goto append; | |
806 | ||
807 | wq_list_add_after(&work->list, &tail->list, &wqe->work_list); | |
808 | } | |
809 | ||
771b53d0 JA |
810 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
811 | { | |
c5def4ab | 812 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 813 | int work_flags; |
771b53d0 JA |
814 | unsigned long flags; |
815 | ||
c5def4ab JA |
816 | /* |
817 | * Do early check to see if we need a new unbound worker, and if we do, | |
818 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
819 | * gap between this check and incrementing the value, but that's OK. | |
820 | * It's close enough to not be an issue, fork() has the same delay. | |
821 | */ | |
822 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
e9fd9396 | 823 | io_run_cancel(work, wqe); |
c5def4ab JA |
824 | return; |
825 | } | |
826 | ||
895e2ca0 | 827 | work_flags = work->flags; |
95da8465 | 828 | raw_spin_lock_irqsave(&wqe->lock, flags); |
86f3cd1b | 829 | io_wqe_insert_work(wqe, work); |
771b53d0 | 830 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
95da8465 | 831 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 | 832 | |
895e2ca0 JA |
833 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
834 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 835 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
836 | } |
837 | ||
838 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
839 | { | |
840 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
841 | ||
842 | io_wqe_enqueue(wqe, work); | |
843 | } | |
844 | ||
845 | /* | |
8766dd51 PB |
846 | * Work items that hash to the same value will not be done in parallel. |
847 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 848 | */ |
8766dd51 | 849 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 850 | { |
8766dd51 | 851 | unsigned int bit; |
771b53d0 JA |
852 | |
853 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
854 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
855 | } |
856 | ||
857 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
858 | { | |
859 | send_sig(SIGINT, worker->task, 1); | |
860 | return false; | |
861 | } | |
862 | ||
863 | /* | |
864 | * Iterate the passed in list and call the specific function for each | |
865 | * worker that isn't exiting | |
866 | */ | |
867 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
868 | bool (*func)(struct io_worker *, void *), |
869 | void *data) | |
870 | { | |
771b53d0 JA |
871 | struct io_worker *worker; |
872 | bool ret = false; | |
873 | ||
e61df66c | 874 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 | 875 | if (io_worker_get(worker)) { |
7563439a JA |
876 | /* no task if node is/was offline */ |
877 | if (worker->task) | |
878 | ret = func(worker, data); | |
771b53d0 JA |
879 | io_worker_release(worker); |
880 | if (ret) | |
881 | break; | |
882 | } | |
883 | } | |
e61df66c | 884 | |
771b53d0 JA |
885 | return ret; |
886 | } | |
887 | ||
888 | void io_wq_cancel_all(struct io_wq *wq) | |
889 | { | |
3fc50ab5 | 890 | int node; |
771b53d0 JA |
891 | |
892 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
893 | ||
771b53d0 | 894 | rcu_read_lock(); |
3fc50ab5 JH |
895 | for_each_node(node) { |
896 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 897 | |
e61df66c | 898 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
899 | } |
900 | rcu_read_unlock(); | |
901 | } | |
902 | ||
62755e35 | 903 | struct io_cb_cancel_data { |
2293b419 PB |
904 | work_cancel_fn *fn; |
905 | void *data; | |
4f26bda1 PB |
906 | int nr_running; |
907 | int nr_pending; | |
908 | bool cancel_all; | |
62755e35 JA |
909 | }; |
910 | ||
2293b419 | 911 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 912 | { |
2293b419 | 913 | struct io_cb_cancel_data *match = data; |
6f72653e | 914 | unsigned long flags; |
62755e35 JA |
915 | |
916 | /* | |
917 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 918 | * may dereference the passed in work. |
62755e35 | 919 | */ |
36c2f922 | 920 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 921 | if (worker->cur_work && |
0c9d5ccd | 922 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
2293b419 | 923 | match->fn(worker->cur_work, match->data)) { |
771b53d0 | 924 | send_sig(SIGINT, worker->task, 1); |
4f26bda1 | 925 | match->nr_running++; |
771b53d0 | 926 | } |
36c2f922 | 927 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 928 | |
4f26bda1 | 929 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
930 | } |
931 | ||
204361a7 PB |
932 | static inline void io_wqe_remove_pending(struct io_wqe *wqe, |
933 | struct io_wq_work *work, | |
934 | struct io_wq_work_node *prev) | |
935 | { | |
936 | unsigned int hash = io_get_work_hash(work); | |
937 | struct io_wq_work *prev_work = NULL; | |
938 | ||
939 | if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) { | |
940 | if (prev) | |
941 | prev_work = container_of(prev, struct io_wq_work, list); | |
942 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
943 | wqe->hash_tail[hash] = prev_work; | |
944 | else | |
945 | wqe->hash_tail[hash] = NULL; | |
946 | } | |
947 | wq_list_del(&wqe->work_list, &work->list, prev); | |
948 | } | |
949 | ||
4f26bda1 | 950 | static void io_wqe_cancel_pending_work(struct io_wqe *wqe, |
f4c2665e | 951 | struct io_cb_cancel_data *match) |
771b53d0 | 952 | { |
6206f0e1 | 953 | struct io_wq_work_node *node, *prev; |
771b53d0 | 954 | struct io_wq_work *work; |
6f72653e | 955 | unsigned long flags; |
771b53d0 | 956 | |
4f26bda1 | 957 | retry: |
95da8465 | 958 | raw_spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
959 | wq_list_for_each(node, prev, &wqe->work_list) { |
960 | work = container_of(node, struct io_wq_work, list); | |
4f26bda1 PB |
961 | if (!match->fn(work, match->data)) |
962 | continue; | |
204361a7 | 963 | io_wqe_remove_pending(wqe, work, prev); |
95da8465 | 964 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
4f26bda1 PB |
965 | io_run_cancel(work, wqe); |
966 | match->nr_pending++; | |
967 | if (!match->cancel_all) | |
968 | return; | |
969 | ||
970 | /* not safe to continue after unlock */ | |
971 | goto retry; | |
771b53d0 | 972 | } |
95da8465 | 973 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
f4c2665e PB |
974 | } |
975 | ||
4f26bda1 | 976 | static void io_wqe_cancel_running_work(struct io_wqe *wqe, |
f4c2665e PB |
977 | struct io_cb_cancel_data *match) |
978 | { | |
771b53d0 | 979 | rcu_read_lock(); |
4f26bda1 | 980 | io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 | 981 | rcu_read_unlock(); |
771b53d0 JA |
982 | } |
983 | ||
2293b419 | 984 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 985 | void *data, bool cancel_all) |
771b53d0 | 986 | { |
2293b419 | 987 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
988 | .fn = cancel, |
989 | .data = data, | |
990 | .cancel_all = cancel_all, | |
00bcda13 | 991 | }; |
3fc50ab5 | 992 | int node; |
771b53d0 | 993 | |
f4c2665e PB |
994 | /* |
995 | * First check pending list, if we're lucky we can just remove it | |
996 | * from there. CANCEL_OK means that the work is returned as-new, | |
997 | * no completion will be posted for it. | |
998 | */ | |
3fc50ab5 JH |
999 | for_each_node(node) { |
1000 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 1001 | |
4f26bda1 PB |
1002 | io_wqe_cancel_pending_work(wqe, &match); |
1003 | if (match.nr_pending && !match.cancel_all) | |
f4c2665e | 1004 | return IO_WQ_CANCEL_OK; |
771b53d0 JA |
1005 | } |
1006 | ||
f4c2665e PB |
1007 | /* |
1008 | * Now check if a free (going busy) or busy worker has the work | |
1009 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
1010 | * as an indication that we attempt to signal cancellation. The | |
1011 | * completion will run normally in this case. | |
1012 | */ | |
1013 | for_each_node(node) { | |
1014 | struct io_wqe *wqe = wq->wqes[node]; | |
1015 | ||
4f26bda1 PB |
1016 | io_wqe_cancel_running_work(wqe, &match); |
1017 | if (match.nr_running && !match.cancel_all) | |
f4c2665e PB |
1018 | return IO_WQ_CANCEL_RUNNING; |
1019 | } | |
1020 | ||
4f26bda1 PB |
1021 | if (match.nr_running) |
1022 | return IO_WQ_CANCEL_RUNNING; | |
1023 | if (match.nr_pending) | |
1024 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 1025 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
1026 | } |
1027 | ||
2293b419 PB |
1028 | static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data) |
1029 | { | |
1030 | return work == data; | |
1031 | } | |
1032 | ||
1033 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
1034 | { | |
4f26bda1 | 1035 | return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork, false); |
2293b419 PB |
1036 | } |
1037 | ||
576a347b | 1038 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1039 | { |
3fc50ab5 | 1040 | int ret = -ENOMEM, node; |
771b53d0 JA |
1041 | struct io_wq *wq; |
1042 | ||
f5fa38c5 | 1043 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 PB |
1044 | return ERR_PTR(-EINVAL); |
1045 | ||
ad6e005c | 1046 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1047 | if (!wq) |
1048 | return ERR_PTR(-ENOMEM); | |
1049 | ||
3fc50ab5 | 1050 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1051 | if (!wq->wqes) { |
1052 | kfree(wq); | |
1053 | return ERR_PTR(-ENOMEM); | |
1054 | } | |
1055 | ||
e9fd9396 | 1056 | wq->free_work = data->free_work; |
f5fa38c5 | 1057 | wq->do_work = data->do_work; |
7d723065 | 1058 | |
c5def4ab | 1059 | /* caller must already hold a reference to this */ |
576a347b | 1060 | wq->user = data->user; |
c5def4ab | 1061 | |
3fc50ab5 | 1062 | for_each_node(node) { |
771b53d0 | 1063 | struct io_wqe *wqe; |
7563439a | 1064 | int alloc_node = node; |
771b53d0 | 1065 | |
7563439a JA |
1066 | if (!node_online(alloc_node)) |
1067 | alloc_node = NUMA_NO_NODE; | |
1068 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1069 | if (!wqe) |
3fc50ab5 JH |
1070 | goto err; |
1071 | wq->wqes[node] = wqe; | |
7563439a | 1072 | wqe->node = alloc_node; |
c5def4ab JA |
1073 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1074 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1075 | if (wq->user) { |
c5def4ab JA |
1076 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1077 | task_rlimit(current, RLIMIT_NPROC); | |
1078 | } | |
1079 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 | 1080 | wqe->wq = wq; |
95da8465 | 1081 | raw_spin_lock_init(&wqe->lock); |
6206f0e1 | 1082 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1083 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1084 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1085 | } |
1086 | ||
1087 | init_completion(&wq->done); | |
1088 | ||
771b53d0 JA |
1089 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
1090 | if (!IS_ERR(wq->manager)) { | |
1091 | wake_up_process(wq->manager); | |
b60fda60 JA |
1092 | wait_for_completion(&wq->done); |
1093 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1094 | ret = -ENOMEM; | |
1095 | goto err; | |
1096 | } | |
848f7e18 | 1097 | refcount_set(&wq->use_refs, 1); |
b60fda60 | 1098 | reinit_completion(&wq->done); |
771b53d0 JA |
1099 | return wq; |
1100 | } | |
1101 | ||
1102 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1103 | complete(&wq->done); |
b60fda60 | 1104 | err: |
3fc50ab5 JH |
1105 | for_each_node(node) |
1106 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1107 | kfree(wq->wqes); |
1108 | kfree(wq); | |
771b53d0 JA |
1109 | return ERR_PTR(ret); |
1110 | } | |
1111 | ||
eba6f5a3 PB |
1112 | bool io_wq_get(struct io_wq *wq, struct io_wq_data *data) |
1113 | { | |
f5fa38c5 | 1114 | if (data->free_work != wq->free_work || data->do_work != wq->do_work) |
eba6f5a3 PB |
1115 | return false; |
1116 | ||
1117 | return refcount_inc_not_zero(&wq->use_refs); | |
1118 | } | |
1119 | ||
771b53d0 JA |
1120 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) |
1121 | { | |
1122 | wake_up_process(worker->task); | |
1123 | return false; | |
1124 | } | |
1125 | ||
848f7e18 | 1126 | static void __io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1127 | { |
3fc50ab5 | 1128 | int node; |
771b53d0 | 1129 | |
b60fda60 JA |
1130 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1131 | if (wq->manager) | |
771b53d0 | 1132 | kthread_stop(wq->manager); |
771b53d0 JA |
1133 | |
1134 | rcu_read_lock(); | |
3fc50ab5 JH |
1135 | for_each_node(node) |
1136 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1137 | rcu_read_unlock(); |
1138 | ||
1139 | wait_for_completion(&wq->done); | |
1140 | ||
3fc50ab5 JH |
1141 | for_each_node(node) |
1142 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1143 | kfree(wq->wqes); |
1144 | kfree(wq); | |
1145 | } | |
848f7e18 JA |
1146 | |
1147 | void io_wq_destroy(struct io_wq *wq) | |
1148 | { | |
1149 | if (refcount_dec_and_test(&wq->use_refs)) | |
1150 | __io_wq_destroy(wq); | |
1151 | } | |
aa96bf8a JA |
1152 | |
1153 | struct task_struct *io_wq_get_task(struct io_wq *wq) | |
1154 | { | |
1155 | return wq->manager; | |
1156 | } |