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> |
91d8f519 | 20 | #include <linux/blk-cgroup.h> |
4ea33a97 | 21 | #include <linux/audit.h> |
771b53d0 JA |
22 | |
23 | #include "io-wq.h" | |
24 | ||
25 | #define WORKER_IDLE_TIMEOUT (5 * HZ) | |
26 | ||
27 | enum { | |
28 | IO_WORKER_F_UP = 1, /* up and active */ | |
29 | IO_WORKER_F_RUNNING = 2, /* account as running */ | |
30 | IO_WORKER_F_FREE = 4, /* worker on free list */ | |
145cc8c6 JA |
31 | IO_WORKER_F_FIXED = 8, /* static idle worker */ |
32 | IO_WORKER_F_BOUND = 16, /* 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; | |
91d8f519 DZ |
61 | #ifdef CONFIG_BLK_CGROUP |
62 | struct cgroup_subsys_state *blkcg_css; | |
63 | #endif | |
cccf0ee8 JA |
64 | const struct cred *cur_creds; |
65 | const struct cred *saved_creds; | |
fcb323cc | 66 | struct files_struct *restore_files; |
9b828492 | 67 | struct nsproxy *restore_nsproxy; |
9392a27d | 68 | struct fs_struct *restore_fs; |
771b53d0 JA |
69 | }; |
70 | ||
771b53d0 JA |
71 | #if BITS_PER_LONG == 64 |
72 | #define IO_WQ_HASH_ORDER 6 | |
73 | #else | |
74 | #define IO_WQ_HASH_ORDER 5 | |
75 | #endif | |
76 | ||
86f3cd1b PB |
77 | #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER) |
78 | ||
c5def4ab JA |
79 | struct io_wqe_acct { |
80 | unsigned nr_workers; | |
81 | unsigned max_workers; | |
82 | atomic_t nr_running; | |
83 | }; | |
84 | ||
85 | enum { | |
86 | IO_WQ_ACCT_BOUND, | |
87 | IO_WQ_ACCT_UNBOUND, | |
88 | }; | |
89 | ||
771b53d0 JA |
90 | /* |
91 | * Per-node worker thread pool | |
92 | */ | |
93 | struct io_wqe { | |
94 | struct { | |
95da8465 | 95 | raw_spinlock_t lock; |
6206f0e1 | 96 | struct io_wq_work_list work_list; |
771b53d0 JA |
97 | unsigned long hash_map; |
98 | unsigned flags; | |
99 | } ____cacheline_aligned_in_smp; | |
100 | ||
101 | int node; | |
c5def4ab | 102 | struct io_wqe_acct acct[2]; |
771b53d0 | 103 | |
021d1cdd | 104 | struct hlist_nulls_head free_list; |
e61df66c | 105 | struct list_head all_list; |
771b53d0 JA |
106 | |
107 | struct io_wq *wq; | |
86f3cd1b | 108 | struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS]; |
771b53d0 JA |
109 | }; |
110 | ||
111 | /* | |
112 | * Per io_wq state | |
113 | */ | |
114 | struct io_wq { | |
115 | struct io_wqe **wqes; | |
116 | unsigned long state; | |
771b53d0 | 117 | |
e9fd9396 | 118 | free_work_fn *free_work; |
f5fa38c5 | 119 | io_wq_work_fn *do_work; |
7d723065 | 120 | |
771b53d0 | 121 | struct task_struct *manager; |
c5def4ab | 122 | struct user_struct *user; |
771b53d0 JA |
123 | refcount_t refs; |
124 | struct completion done; | |
848f7e18 JA |
125 | |
126 | refcount_t use_refs; | |
771b53d0 JA |
127 | }; |
128 | ||
771b53d0 JA |
129 | static bool io_worker_get(struct io_worker *worker) |
130 | { | |
131 | return refcount_inc_not_zero(&worker->ref); | |
132 | } | |
133 | ||
134 | static void io_worker_release(struct io_worker *worker) | |
135 | { | |
136 | if (refcount_dec_and_test(&worker->ref)) | |
137 | wake_up_process(worker->task); | |
138 | } | |
139 | ||
140 | /* | |
141 | * Note: drops the wqe->lock if returning true! The caller must re-acquire | |
142 | * the lock in that case. Some callers need to restart handling if this | |
143 | * happens, so we can't just re-acquire the lock on behalf of the caller. | |
144 | */ | |
145 | static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) | |
146 | { | |
fcb323cc JA |
147 | bool dropped_lock = false; |
148 | ||
cccf0ee8 JA |
149 | if (worker->saved_creds) { |
150 | revert_creds(worker->saved_creds); | |
151 | worker->cur_creds = worker->saved_creds = NULL; | |
181e448d JA |
152 | } |
153 | ||
fcb323cc JA |
154 | if (current->files != worker->restore_files) { |
155 | __acquire(&wqe->lock); | |
95da8465 | 156 | raw_spin_unlock_irq(&wqe->lock); |
fcb323cc JA |
157 | dropped_lock = true; |
158 | ||
159 | task_lock(current); | |
160 | current->files = worker->restore_files; | |
9b828492 | 161 | current->nsproxy = worker->restore_nsproxy; |
fcb323cc JA |
162 | task_unlock(current); |
163 | } | |
164 | ||
9392a27d JA |
165 | if (current->fs != worker->restore_fs) |
166 | current->fs = worker->restore_fs; | |
167 | ||
771b53d0 JA |
168 | /* |
169 | * If we have an active mm, we need to drop the wq lock before unusing | |
170 | * it. If we do, return true and let the caller retry the idle loop. | |
171 | */ | |
172 | if (worker->mm) { | |
fcb323cc JA |
173 | if (!dropped_lock) { |
174 | __acquire(&wqe->lock); | |
95da8465 | 175 | raw_spin_unlock_irq(&wqe->lock); |
fcb323cc JA |
176 | dropped_lock = true; |
177 | } | |
771b53d0 | 178 | __set_current_state(TASK_RUNNING); |
f5678e7f | 179 | kthread_unuse_mm(worker->mm); |
771b53d0 JA |
180 | mmput(worker->mm); |
181 | worker->mm = NULL; | |
771b53d0 JA |
182 | } |
183 | ||
91d8f519 DZ |
184 | #ifdef CONFIG_BLK_CGROUP |
185 | if (worker->blkcg_css) { | |
186 | kthread_associate_blkcg(NULL); | |
187 | worker->blkcg_css = NULL; | |
188 | } | |
189 | #endif | |
69228338 JA |
190 | if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY) |
191 | current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY; | |
fcb323cc | 192 | return dropped_lock; |
771b53d0 JA |
193 | } |
194 | ||
c5def4ab JA |
195 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
196 | struct io_wq_work *work) | |
197 | { | |
198 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
199 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
200 | ||
201 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
202 | } | |
203 | ||
204 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe, | |
205 | struct io_worker *worker) | |
206 | { | |
207 | if (worker->flags & IO_WORKER_F_BOUND) | |
208 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
209 | ||
210 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
211 | } | |
212 | ||
771b53d0 JA |
213 | static void io_worker_exit(struct io_worker *worker) |
214 | { | |
215 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab | 216 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); |
771b53d0 JA |
217 | |
218 | /* | |
219 | * If we're not at zero, someone else is holding a brief reference | |
220 | * to the worker. Wait for that to go away. | |
221 | */ | |
222 | set_current_state(TASK_INTERRUPTIBLE); | |
223 | if (!refcount_dec_and_test(&worker->ref)) | |
224 | schedule(); | |
225 | __set_current_state(TASK_RUNNING); | |
226 | ||
227 | preempt_disable(); | |
228 | current->flags &= ~PF_IO_WORKER; | |
229 | if (worker->flags & IO_WORKER_F_RUNNING) | |
c5def4ab JA |
230 | atomic_dec(&acct->nr_running); |
231 | if (!(worker->flags & IO_WORKER_F_BOUND)) | |
232 | atomic_dec(&wqe->wq->user->processes); | |
771b53d0 JA |
233 | worker->flags = 0; |
234 | preempt_enable(); | |
235 | ||
95da8465 | 236 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 | 237 | hlist_nulls_del_rcu(&worker->nulls_node); |
e61df66c | 238 | list_del_rcu(&worker->all_list); |
771b53d0 JA |
239 | if (__io_worker_unuse(wqe, worker)) { |
240 | __release(&wqe->lock); | |
95da8465 | 241 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 | 242 | } |
c5def4ab | 243 | acct->nr_workers--; |
95da8465 | 244 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 | 245 | |
364b05fd | 246 | kfree_rcu(worker, rcu); |
c4068bf8 HD |
247 | if (refcount_dec_and_test(&wqe->wq->refs)) |
248 | complete(&wqe->wq->done); | |
771b53d0 JA |
249 | } |
250 | ||
c5def4ab JA |
251 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
252 | __must_hold(wqe->lock) | |
253 | { | |
6206f0e1 JA |
254 | if (!wq_list_empty(&wqe->work_list) && |
255 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
256 | return true; |
257 | return false; | |
258 | } | |
259 | ||
260 | /* | |
261 | * Check head of free list for an available worker. If one isn't available, | |
262 | * caller must wake up the wq manager to create one. | |
263 | */ | |
264 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
265 | __must_hold(RCU) | |
266 | { | |
267 | struct hlist_nulls_node *n; | |
268 | struct io_worker *worker; | |
269 | ||
021d1cdd | 270 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
271 | if (is_a_nulls(n)) |
272 | return false; | |
273 | ||
274 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
275 | if (io_worker_get(worker)) { | |
506d95ff | 276 | wake_up_process(worker->task); |
c5def4ab JA |
277 | io_worker_release(worker); |
278 | return true; | |
279 | } | |
280 | ||
281 | return false; | |
282 | } | |
283 | ||
284 | /* | |
285 | * We need a worker. If we find a free one, we're good. If not, and we're | |
286 | * below the max number of workers, wake up the manager to create one. | |
287 | */ | |
288 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
289 | { | |
290 | bool ret; | |
291 | ||
292 | /* | |
293 | * Most likely an attempt to queue unbounded work on an io_wq that | |
294 | * wasn't setup with any unbounded workers. | |
295 | */ | |
296 | WARN_ON_ONCE(!acct->max_workers); | |
297 | ||
298 | rcu_read_lock(); | |
299 | ret = io_wqe_activate_free_worker(wqe); | |
300 | rcu_read_unlock(); | |
301 | ||
302 | if (!ret && acct->nr_workers < acct->max_workers) | |
303 | wake_up_process(wqe->wq->manager); | |
304 | } | |
305 | ||
306 | static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker) | |
307 | { | |
308 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
309 | ||
310 | atomic_inc(&acct->nr_running); | |
311 | } | |
312 | ||
313 | static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker) | |
314 | __must_hold(wqe->lock) | |
315 | { | |
316 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
317 | ||
318 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
319 | io_wqe_wake_worker(wqe, acct); | |
320 | } | |
321 | ||
771b53d0 JA |
322 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
323 | { | |
324 | allow_kernel_signal(SIGINT); | |
325 | ||
326 | current->flags |= PF_IO_WORKER; | |
327 | ||
328 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); | |
fcb323cc | 329 | worker->restore_files = current->files; |
9b828492 | 330 | worker->restore_nsproxy = current->nsproxy; |
9392a27d | 331 | worker->restore_fs = current->fs; |
c5def4ab | 332 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
333 | } |
334 | ||
335 | /* | |
336 | * Worker will start processing some work. Move it to the busy list, if | |
337 | * it's currently on the freelist | |
338 | */ | |
339 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
340 | struct io_wq_work *work) | |
341 | __must_hold(wqe->lock) | |
342 | { | |
c5def4ab JA |
343 | bool worker_bound, work_bound; |
344 | ||
771b53d0 JA |
345 | if (worker->flags & IO_WORKER_F_FREE) { |
346 | worker->flags &= ~IO_WORKER_F_FREE; | |
347 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 348 | } |
c5def4ab JA |
349 | |
350 | /* | |
351 | * If worker is moving from bound to unbound (or vice versa), then | |
352 | * ensure we update the running accounting. | |
353 | */ | |
b2e9c7d6 DC |
354 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
355 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
356 | if (worker_bound != work_bound) { | |
c5def4ab JA |
357 | io_wqe_dec_running(wqe, worker); |
358 | if (work_bound) { | |
359 | worker->flags |= IO_WORKER_F_BOUND; | |
360 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
361 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
362 | atomic_dec(&wqe->wq->user->processes); | |
363 | } else { | |
364 | worker->flags &= ~IO_WORKER_F_BOUND; | |
365 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
366 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
367 | atomic_inc(&wqe->wq->user->processes); | |
368 | } | |
369 | io_wqe_inc_running(wqe, worker); | |
370 | } | |
771b53d0 JA |
371 | } |
372 | ||
373 | /* | |
374 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
375 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
376 | * the lock in that case and return success. Since the caller has to | |
377 | * retry the loop in that case (we changed task state), we don't regrab | |
378 | * the lock if we return success. | |
379 | */ | |
380 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) | |
381 | __must_hold(wqe->lock) | |
382 | { | |
383 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
384 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 385 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 JA |
386 | } |
387 | ||
388 | return __io_worker_unuse(wqe, worker); | |
389 | } | |
390 | ||
60cf46ae PB |
391 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
392 | { | |
393 | return work->flags >> IO_WQ_HASH_SHIFT; | |
394 | } | |
395 | ||
396 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe) | |
771b53d0 JA |
397 | __must_hold(wqe->lock) |
398 | { | |
6206f0e1 | 399 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 400 | struct io_wq_work *work, *tail; |
60cf46ae | 401 | unsigned int hash; |
771b53d0 | 402 | |
6206f0e1 JA |
403 | wq_list_for_each(node, prev, &wqe->work_list) { |
404 | work = container_of(node, struct io_wq_work, list); | |
405 | ||
771b53d0 | 406 | /* not hashed, can run anytime */ |
8766dd51 | 407 | if (!io_wq_is_hashed(work)) { |
86f3cd1b | 408 | wq_list_del(&wqe->work_list, node, prev); |
771b53d0 JA |
409 | return work; |
410 | } | |
411 | ||
412 | /* hashed, can run if not already running */ | |
60cf46ae PB |
413 | hash = io_get_work_hash(work); |
414 | if (!(wqe->hash_map & BIT(hash))) { | |
415 | wqe->hash_map |= BIT(hash); | |
86f3cd1b PB |
416 | /* all items with this hash lie in [work, tail] */ |
417 | tail = wqe->hash_tail[hash]; | |
418 | wqe->hash_tail[hash] = NULL; | |
419 | wq_list_cut(&wqe->work_list, &tail->list, prev); | |
771b53d0 JA |
420 | return work; |
421 | } | |
422 | } | |
423 | ||
424 | return NULL; | |
425 | } | |
426 | ||
cccf0ee8 JA |
427 | static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work) |
428 | { | |
429 | if (worker->mm) { | |
f5678e7f | 430 | kthread_unuse_mm(worker->mm); |
cccf0ee8 JA |
431 | mmput(worker->mm); |
432 | worker->mm = NULL; | |
433 | } | |
37c54f9b | 434 | |
98447d65 JA |
435 | if (mmget_not_zero(work->identity->mm)) { |
436 | kthread_use_mm(work->identity->mm); | |
437 | worker->mm = work->identity->mm; | |
cccf0ee8 JA |
438 | return; |
439 | } | |
440 | ||
441 | /* failed grabbing mm, ensure work gets cancelled */ | |
442 | work->flags |= IO_WQ_WORK_CANCEL; | |
443 | } | |
444 | ||
91d8f519 DZ |
445 | static inline void io_wq_switch_blkcg(struct io_worker *worker, |
446 | struct io_wq_work *work) | |
447 | { | |
448 | #ifdef CONFIG_BLK_CGROUP | |
0f203765 JA |
449 | if (!(work->flags & IO_WQ_WORK_BLKCG)) |
450 | return; | |
98447d65 JA |
451 | if (work->identity->blkcg_css != worker->blkcg_css) { |
452 | kthread_associate_blkcg(work->identity->blkcg_css); | |
453 | worker->blkcg_css = work->identity->blkcg_css; | |
91d8f519 DZ |
454 | } |
455 | #endif | |
456 | } | |
457 | ||
cccf0ee8 JA |
458 | static void io_wq_switch_creds(struct io_worker *worker, |
459 | struct io_wq_work *work) | |
460 | { | |
98447d65 | 461 | const struct cred *old_creds = override_creds(work->identity->creds); |
cccf0ee8 | 462 | |
98447d65 | 463 | worker->cur_creds = work->identity->creds; |
cccf0ee8 JA |
464 | if (worker->saved_creds) |
465 | put_cred(old_creds); /* creds set by previous switch */ | |
466 | else | |
467 | worker->saved_creds = old_creds; | |
468 | } | |
469 | ||
dc026a73 PB |
470 | static void io_impersonate_work(struct io_worker *worker, |
471 | struct io_wq_work *work) | |
472 | { | |
98447d65 JA |
473 | if ((work->flags & IO_WQ_WORK_FILES) && |
474 | current->files != work->identity->files) { | |
dc026a73 | 475 | task_lock(current); |
98447d65 JA |
476 | current->files = work->identity->files; |
477 | current->nsproxy = work->identity->nsproxy; | |
dc026a73 PB |
478 | task_unlock(current); |
479 | } | |
98447d65 JA |
480 | if ((work->flags & IO_WQ_WORK_FS) && current->fs != work->identity->fs) |
481 | current->fs = work->identity->fs; | |
482 | if ((work->flags & IO_WQ_WORK_MM) && work->identity->mm != worker->mm) | |
dc026a73 | 483 | io_wq_switch_mm(worker, work); |
98447d65 JA |
484 | if ((work->flags & IO_WQ_WORK_CREDS) && |
485 | worker->cur_creds != work->identity->creds) | |
dc026a73 | 486 | io_wq_switch_creds(worker, work); |
69228338 JA |
487 | if (work->flags & IO_WQ_WORK_FSIZE) |
488 | current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->identity->fsize; | |
489 | else if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY) | |
490 | current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY; | |
91d8f519 | 491 | io_wq_switch_blkcg(worker, work); |
4ea33a97 JA |
492 | #ifdef CONFIG_AUDIT |
493 | current->loginuid = work->identity->loginuid; | |
494 | current->sessionid = work->identity->sessionid; | |
495 | #endif | |
dc026a73 PB |
496 | } |
497 | ||
498 | static void io_assign_current_work(struct io_worker *worker, | |
499 | struct io_wq_work *work) | |
500 | { | |
d78298e7 PB |
501 | if (work) { |
502 | /* flush pending signals before assigning new work */ | |
503 | if (signal_pending(current)) | |
504 | flush_signals(current); | |
505 | cond_resched(); | |
506 | } | |
dc026a73 | 507 | |
4ea33a97 JA |
508 | #ifdef CONFIG_AUDIT |
509 | current->loginuid = KUIDT_INIT(AUDIT_UID_UNSET); | |
510 | current->sessionid = AUDIT_SID_UNSET; | |
511 | #endif | |
512 | ||
dc026a73 PB |
513 | spin_lock_irq(&worker->lock); |
514 | worker->cur_work = work; | |
515 | spin_unlock_irq(&worker->lock); | |
516 | } | |
517 | ||
60cf46ae PB |
518 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
519 | ||
771b53d0 JA |
520 | static void io_worker_handle_work(struct io_worker *worker) |
521 | __releases(wqe->lock) | |
522 | { | |
771b53d0 JA |
523 | struct io_wqe *wqe = worker->wqe; |
524 | struct io_wq *wq = wqe->wq; | |
525 | ||
526 | do { | |
86f3cd1b | 527 | struct io_wq_work *work; |
f462fd36 | 528 | get_next: |
771b53d0 JA |
529 | /* |
530 | * If we got some work, mark us as busy. If we didn't, but | |
531 | * the list isn't empty, it means we stalled on hashed work. | |
532 | * Mark us stalled so we don't keep looking for work when we | |
533 | * can't make progress, any work completion or insertion will | |
534 | * clear the stalled flag. | |
535 | */ | |
60cf46ae | 536 | work = io_get_next_work(wqe); |
771b53d0 JA |
537 | if (work) |
538 | __io_worker_busy(wqe, worker, work); | |
6206f0e1 | 539 | else if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
540 | wqe->flags |= IO_WQE_FLAG_STALLED; |
541 | ||
95da8465 | 542 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
543 | if (!work) |
544 | break; | |
58e39319 | 545 | io_assign_current_work(worker, work); |
36c2f922 | 546 | |
dc026a73 PB |
547 | /* handle a whole dependent link */ |
548 | do { | |
86f3cd1b | 549 | struct io_wq_work *old_work, *next_hashed, *linked; |
b089ed39 | 550 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 551 | |
86f3cd1b | 552 | next_hashed = wq_next_work(work); |
58e39319 | 553 | io_impersonate_work(worker, work); |
dc026a73 PB |
554 | /* |
555 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable | |
556 | * work, the worker function will do the right thing. | |
557 | */ | |
558 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) | |
559 | work->flags |= IO_WQ_WORK_CANCEL; | |
560 | ||
f4db7182 PB |
561 | old_work = work; |
562 | linked = wq->do_work(work); | |
86f3cd1b PB |
563 | |
564 | work = next_hashed; | |
565 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
566 | work = linked; | |
567 | linked = NULL; | |
568 | } | |
569 | io_assign_current_work(worker, work); | |
e9fd9396 | 570 | wq->free_work(old_work); |
dc026a73 | 571 | |
86f3cd1b PB |
572 | if (linked) |
573 | io_wqe_enqueue(wqe, linked); | |
574 | ||
575 | if (hash != -1U && !next_hashed) { | |
95da8465 | 576 | raw_spin_lock_irq(&wqe->lock); |
dc026a73 PB |
577 | wqe->hash_map &= ~BIT_ULL(hash); |
578 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
f462fd36 PB |
579 | /* skip unnecessary unlock-lock wqe->lock */ |
580 | if (!work) | |
581 | goto get_next; | |
95da8465 | 582 | raw_spin_unlock_irq(&wqe->lock); |
7d723065 | 583 | } |
58e39319 | 584 | } while (work); |
7d723065 | 585 | |
95da8465 | 586 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 JA |
587 | } while (1); |
588 | } | |
589 | ||
771b53d0 JA |
590 | static int io_wqe_worker(void *data) |
591 | { | |
592 | struct io_worker *worker = data; | |
593 | struct io_wqe *wqe = worker->wqe; | |
594 | struct io_wq *wq = wqe->wq; | |
771b53d0 JA |
595 | |
596 | io_worker_start(wqe, worker); | |
597 | ||
598 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
506d95ff | 599 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 600 | loop: |
95da8465 | 601 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 JA |
602 | if (io_wqe_run_queue(wqe)) { |
603 | __set_current_state(TASK_RUNNING); | |
604 | io_worker_handle_work(worker); | |
e995d512 | 605 | goto loop; |
771b53d0 JA |
606 | } |
607 | /* drops the lock on success, retry */ | |
608 | if (__io_worker_idle(wqe, worker)) { | |
609 | __release(&wqe->lock); | |
e995d512 | 610 | goto loop; |
771b53d0 | 611 | } |
95da8465 | 612 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
613 | if (signal_pending(current)) |
614 | flush_signals(current); | |
615 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
616 | continue; | |
617 | /* timed out, exit unless we're the fixed worker */ | |
618 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
619 | !(worker->flags & IO_WORKER_F_FIXED)) | |
620 | break; | |
621 | } | |
622 | ||
771b53d0 | 623 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
95da8465 | 624 | raw_spin_lock_irq(&wqe->lock); |
6206f0e1 | 625 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
626 | io_worker_handle_work(worker); |
627 | else | |
95da8465 | 628 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
629 | } |
630 | ||
631 | io_worker_exit(worker); | |
632 | return 0; | |
633 | } | |
634 | ||
771b53d0 JA |
635 | /* |
636 | * Called when a worker is scheduled in. Mark us as currently running. | |
637 | */ | |
638 | void io_wq_worker_running(struct task_struct *tsk) | |
639 | { | |
640 | struct io_worker *worker = kthread_data(tsk); | |
641 | struct io_wqe *wqe = worker->wqe; | |
642 | ||
643 | if (!(worker->flags & IO_WORKER_F_UP)) | |
644 | return; | |
645 | if (worker->flags & IO_WORKER_F_RUNNING) | |
646 | return; | |
647 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 648 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
649 | } |
650 | ||
651 | /* | |
652 | * Called when worker is going to sleep. If there are no workers currently | |
653 | * running and we have work pending, wake up a free one or have the manager | |
654 | * set one up. | |
655 | */ | |
656 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
657 | { | |
658 | struct io_worker *worker = kthread_data(tsk); | |
659 | struct io_wqe *wqe = worker->wqe; | |
660 | ||
661 | if (!(worker->flags & IO_WORKER_F_UP)) | |
662 | return; | |
663 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
664 | return; | |
665 | ||
666 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
667 | ||
95da8465 | 668 | raw_spin_lock_irq(&wqe->lock); |
c5def4ab | 669 | io_wqe_dec_running(wqe, worker); |
95da8465 | 670 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
671 | } |
672 | ||
b60fda60 | 673 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 674 | { |
c4068bf8 | 675 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 JA |
676 | struct io_worker *worker; |
677 | ||
ad6e005c | 678 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 679 | if (!worker) |
b60fda60 | 680 | return false; |
771b53d0 JA |
681 | |
682 | refcount_set(&worker->ref, 1); | |
683 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 684 | worker->wqe = wqe; |
36c2f922 | 685 | spin_lock_init(&worker->lock); |
771b53d0 JA |
686 | |
687 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 688 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
689 | if (IS_ERR(worker->task)) { |
690 | kfree(worker); | |
b60fda60 | 691 | return false; |
771b53d0 | 692 | } |
a8b595b2 | 693 | kthread_bind_mask(worker->task, cpumask_of_node(wqe->node)); |
771b53d0 | 694 | |
95da8465 | 695 | raw_spin_lock_irq(&wqe->lock); |
021d1cdd | 696 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 697 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 698 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
699 | if (index == IO_WQ_ACCT_BOUND) |
700 | worker->flags |= IO_WORKER_F_BOUND; | |
701 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 702 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 703 | acct->nr_workers++; |
95da8465 | 704 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 | 705 | |
c5def4ab JA |
706 | if (index == IO_WQ_ACCT_UNBOUND) |
707 | atomic_inc(&wq->user->processes); | |
708 | ||
c4068bf8 | 709 | refcount_inc(&wq->refs); |
771b53d0 | 710 | wake_up_process(worker->task); |
b60fda60 | 711 | return true; |
771b53d0 JA |
712 | } |
713 | ||
c5def4ab | 714 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
715 | __must_hold(wqe->lock) |
716 | { | |
c5def4ab | 717 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 718 | |
c5def4ab | 719 | /* if we have available workers or no work, no need */ |
021d1cdd | 720 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
721 | return false; |
722 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
723 | } |
724 | ||
c4068bf8 HD |
725 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) |
726 | { | |
727 | send_sig(SIGINT, worker->task, 1); | |
728 | return false; | |
729 | } | |
730 | ||
731 | /* | |
732 | * Iterate the passed in list and call the specific function for each | |
733 | * worker that isn't exiting | |
734 | */ | |
735 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
736 | bool (*func)(struct io_worker *, void *), | |
737 | void *data) | |
738 | { | |
739 | struct io_worker *worker; | |
740 | bool ret = false; | |
741 | ||
742 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { | |
743 | if (io_worker_get(worker)) { | |
744 | /* no task if node is/was offline */ | |
745 | if (worker->task) | |
746 | ret = func(worker, data); | |
747 | io_worker_release(worker); | |
748 | if (ret) | |
749 | break; | |
750 | } | |
751 | } | |
752 | ||
753 | return ret; | |
754 | } | |
755 | ||
756 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
757 | { | |
758 | wake_up_process(worker->task); | |
759 | return false; | |
760 | } | |
761 | ||
771b53d0 JA |
762 | /* |
763 | * Manager thread. Tasked with creating new workers, if we need them. | |
764 | */ | |
765 | static int io_wq_manager(void *data) | |
766 | { | |
767 | struct io_wq *wq = data; | |
3fc50ab5 | 768 | int node; |
771b53d0 | 769 | |
b60fda60 | 770 | /* create fixed workers */ |
c4068bf8 | 771 | refcount_set(&wq->refs, 1); |
3fc50ab5 | 772 | for_each_node(node) { |
7563439a JA |
773 | if (!node_online(node)) |
774 | continue; | |
c4068bf8 HD |
775 | if (create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) |
776 | continue; | |
777 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
778 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
779 | goto out; | |
b60fda60 | 780 | } |
771b53d0 | 781 | |
b60fda60 JA |
782 | complete(&wq->done); |
783 | ||
784 | while (!kthread_should_stop()) { | |
aa96bf8a JA |
785 | if (current->task_works) |
786 | task_work_run(); | |
787 | ||
3fc50ab5 JH |
788 | for_each_node(node) { |
789 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 790 | bool fork_worker[2] = { false, false }; |
771b53d0 | 791 | |
7563439a JA |
792 | if (!node_online(node)) |
793 | continue; | |
794 | ||
95da8465 | 795 | raw_spin_lock_irq(&wqe->lock); |
c5def4ab JA |
796 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
797 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
798 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
799 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
95da8465 | 800 | raw_spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
801 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
802 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
803 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
804 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
805 | } |
806 | set_current_state(TASK_INTERRUPTIBLE); | |
807 | schedule_timeout(HZ); | |
808 | } | |
809 | ||
aa96bf8a JA |
810 | if (current->task_works) |
811 | task_work_run(); | |
812 | ||
c4068bf8 HD |
813 | out: |
814 | if (refcount_dec_and_test(&wq->refs)) { | |
b60fda60 | 815 | complete(&wq->done); |
c4068bf8 HD |
816 | return 0; |
817 | } | |
818 | /* if ERROR is set and we get here, we have workers to wake */ | |
819 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
820 | rcu_read_lock(); | |
821 | for_each_node(node) | |
822 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
823 | rcu_read_unlock(); | |
824 | } | |
b60fda60 | 825 | return 0; |
771b53d0 JA |
826 | } |
827 | ||
c5def4ab JA |
828 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
829 | struct io_wq_work *work) | |
830 | { | |
831 | bool free_worker; | |
832 | ||
833 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
834 | return true; | |
835 | if (atomic_read(&acct->nr_running)) | |
836 | return true; | |
837 | ||
838 | rcu_read_lock(); | |
021d1cdd | 839 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
840 | rcu_read_unlock(); |
841 | if (free_worker) | |
842 | return true; | |
843 | ||
844 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
845 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
846 | return false; | |
847 | ||
848 | return true; | |
849 | } | |
850 | ||
e9fd9396 | 851 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 852 | { |
e9fd9396 PB |
853 | struct io_wq *wq = wqe->wq; |
854 | ||
fc04c39b PB |
855 | do { |
856 | struct io_wq_work *old_work = work; | |
857 | ||
858 | work->flags |= IO_WQ_WORK_CANCEL; | |
f4db7182 | 859 | work = wq->do_work(work); |
e9fd9396 | 860 | wq->free_work(old_work); |
fc04c39b PB |
861 | } while (work); |
862 | } | |
863 | ||
86f3cd1b PB |
864 | static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work) |
865 | { | |
866 | unsigned int hash; | |
867 | struct io_wq_work *tail; | |
868 | ||
869 | if (!io_wq_is_hashed(work)) { | |
870 | append: | |
871 | wq_list_add_tail(&work->list, &wqe->work_list); | |
872 | return; | |
873 | } | |
874 | ||
875 | hash = io_get_work_hash(work); | |
876 | tail = wqe->hash_tail[hash]; | |
877 | wqe->hash_tail[hash] = work; | |
878 | if (!tail) | |
879 | goto append; | |
880 | ||
881 | wq_list_add_after(&work->list, &tail->list, &wqe->work_list); | |
882 | } | |
883 | ||
771b53d0 JA |
884 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
885 | { | |
c5def4ab | 886 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 887 | int work_flags; |
771b53d0 JA |
888 | unsigned long flags; |
889 | ||
c5def4ab JA |
890 | /* |
891 | * Do early check to see if we need a new unbound worker, and if we do, | |
892 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
893 | * gap between this check and incrementing the value, but that's OK. | |
894 | * It's close enough to not be an issue, fork() has the same delay. | |
895 | */ | |
896 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
e9fd9396 | 897 | io_run_cancel(work, wqe); |
c5def4ab JA |
898 | return; |
899 | } | |
900 | ||
895e2ca0 | 901 | work_flags = work->flags; |
95da8465 | 902 | raw_spin_lock_irqsave(&wqe->lock, flags); |
86f3cd1b | 903 | io_wqe_insert_work(wqe, work); |
771b53d0 | 904 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
95da8465 | 905 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 | 906 | |
895e2ca0 JA |
907 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
908 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 909 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
910 | } |
911 | ||
912 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
913 | { | |
914 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
915 | ||
916 | io_wqe_enqueue(wqe, work); | |
917 | } | |
918 | ||
919 | /* | |
8766dd51 PB |
920 | * Work items that hash to the same value will not be done in parallel. |
921 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 922 | */ |
8766dd51 | 923 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 924 | { |
8766dd51 | 925 | unsigned int bit; |
771b53d0 JA |
926 | |
927 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
928 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
929 | } |
930 | ||
771b53d0 JA |
931 | void io_wq_cancel_all(struct io_wq *wq) |
932 | { | |
3fc50ab5 | 933 | int node; |
771b53d0 JA |
934 | |
935 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
936 | ||
771b53d0 | 937 | rcu_read_lock(); |
3fc50ab5 JH |
938 | for_each_node(node) { |
939 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 940 | |
e61df66c | 941 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
942 | } |
943 | rcu_read_unlock(); | |
944 | } | |
945 | ||
62755e35 | 946 | struct io_cb_cancel_data { |
2293b419 PB |
947 | work_cancel_fn *fn; |
948 | void *data; | |
4f26bda1 PB |
949 | int nr_running; |
950 | int nr_pending; | |
951 | bool cancel_all; | |
62755e35 JA |
952 | }; |
953 | ||
2293b419 | 954 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 955 | { |
2293b419 | 956 | struct io_cb_cancel_data *match = data; |
6f72653e | 957 | unsigned long flags; |
62755e35 JA |
958 | |
959 | /* | |
960 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 961 | * may dereference the passed in work. |
62755e35 | 962 | */ |
36c2f922 | 963 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 964 | if (worker->cur_work && |
0c9d5ccd | 965 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
2293b419 | 966 | match->fn(worker->cur_work, match->data)) { |
771b53d0 | 967 | send_sig(SIGINT, worker->task, 1); |
4f26bda1 | 968 | match->nr_running++; |
771b53d0 | 969 | } |
36c2f922 | 970 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 971 | |
4f26bda1 | 972 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
973 | } |
974 | ||
204361a7 PB |
975 | static inline void io_wqe_remove_pending(struct io_wqe *wqe, |
976 | struct io_wq_work *work, | |
977 | struct io_wq_work_node *prev) | |
978 | { | |
979 | unsigned int hash = io_get_work_hash(work); | |
980 | struct io_wq_work *prev_work = NULL; | |
981 | ||
982 | if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) { | |
983 | if (prev) | |
984 | prev_work = container_of(prev, struct io_wq_work, list); | |
985 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
986 | wqe->hash_tail[hash] = prev_work; | |
987 | else | |
988 | wqe->hash_tail[hash] = NULL; | |
989 | } | |
990 | wq_list_del(&wqe->work_list, &work->list, prev); | |
991 | } | |
992 | ||
4f26bda1 | 993 | static void io_wqe_cancel_pending_work(struct io_wqe *wqe, |
f4c2665e | 994 | struct io_cb_cancel_data *match) |
771b53d0 | 995 | { |
6206f0e1 | 996 | struct io_wq_work_node *node, *prev; |
771b53d0 | 997 | struct io_wq_work *work; |
6f72653e | 998 | unsigned long flags; |
771b53d0 | 999 | |
4f26bda1 | 1000 | retry: |
95da8465 | 1001 | raw_spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
1002 | wq_list_for_each(node, prev, &wqe->work_list) { |
1003 | work = container_of(node, struct io_wq_work, list); | |
4f26bda1 PB |
1004 | if (!match->fn(work, match->data)) |
1005 | continue; | |
204361a7 | 1006 | io_wqe_remove_pending(wqe, work, prev); |
95da8465 | 1007 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
4f26bda1 PB |
1008 | io_run_cancel(work, wqe); |
1009 | match->nr_pending++; | |
1010 | if (!match->cancel_all) | |
1011 | return; | |
1012 | ||
1013 | /* not safe to continue after unlock */ | |
1014 | goto retry; | |
771b53d0 | 1015 | } |
95da8465 | 1016 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
f4c2665e PB |
1017 | } |
1018 | ||
4f26bda1 | 1019 | static void io_wqe_cancel_running_work(struct io_wqe *wqe, |
f4c2665e PB |
1020 | struct io_cb_cancel_data *match) |
1021 | { | |
771b53d0 | 1022 | rcu_read_lock(); |
4f26bda1 | 1023 | io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 | 1024 | rcu_read_unlock(); |
771b53d0 JA |
1025 | } |
1026 | ||
2293b419 | 1027 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 1028 | void *data, bool cancel_all) |
771b53d0 | 1029 | { |
2293b419 | 1030 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
1031 | .fn = cancel, |
1032 | .data = data, | |
1033 | .cancel_all = cancel_all, | |
00bcda13 | 1034 | }; |
3fc50ab5 | 1035 | int node; |
771b53d0 | 1036 | |
f4c2665e PB |
1037 | /* |
1038 | * First check pending list, if we're lucky we can just remove it | |
1039 | * from there. CANCEL_OK means that the work is returned as-new, | |
1040 | * no completion will be posted for it. | |
1041 | */ | |
3fc50ab5 JH |
1042 | for_each_node(node) { |
1043 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 1044 | |
4f26bda1 PB |
1045 | io_wqe_cancel_pending_work(wqe, &match); |
1046 | if (match.nr_pending && !match.cancel_all) | |
f4c2665e | 1047 | return IO_WQ_CANCEL_OK; |
771b53d0 JA |
1048 | } |
1049 | ||
f4c2665e PB |
1050 | /* |
1051 | * Now check if a free (going busy) or busy worker has the work | |
1052 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
1053 | * as an indication that we attempt to signal cancellation. The | |
1054 | * completion will run normally in this case. | |
1055 | */ | |
1056 | for_each_node(node) { | |
1057 | struct io_wqe *wqe = wq->wqes[node]; | |
1058 | ||
4f26bda1 PB |
1059 | io_wqe_cancel_running_work(wqe, &match); |
1060 | if (match.nr_running && !match.cancel_all) | |
f4c2665e PB |
1061 | return IO_WQ_CANCEL_RUNNING; |
1062 | } | |
1063 | ||
4f26bda1 PB |
1064 | if (match.nr_running) |
1065 | return IO_WQ_CANCEL_RUNNING; | |
1066 | if (match.nr_pending) | |
1067 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 1068 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
1069 | } |
1070 | ||
2293b419 PB |
1071 | static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data) |
1072 | { | |
1073 | return work == data; | |
1074 | } | |
1075 | ||
1076 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
1077 | { | |
4f26bda1 | 1078 | return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork, false); |
2293b419 PB |
1079 | } |
1080 | ||
576a347b | 1081 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1082 | { |
3fc50ab5 | 1083 | int ret = -ENOMEM, node; |
771b53d0 JA |
1084 | struct io_wq *wq; |
1085 | ||
f5fa38c5 | 1086 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 PB |
1087 | return ERR_PTR(-EINVAL); |
1088 | ||
ad6e005c | 1089 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1090 | if (!wq) |
1091 | return ERR_PTR(-ENOMEM); | |
1092 | ||
3fc50ab5 | 1093 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1094 | if (!wq->wqes) { |
1095 | kfree(wq); | |
1096 | return ERR_PTR(-ENOMEM); | |
1097 | } | |
1098 | ||
e9fd9396 | 1099 | wq->free_work = data->free_work; |
f5fa38c5 | 1100 | wq->do_work = data->do_work; |
7d723065 | 1101 | |
c5def4ab | 1102 | /* caller must already hold a reference to this */ |
576a347b | 1103 | wq->user = data->user; |
c5def4ab | 1104 | |
3fc50ab5 | 1105 | for_each_node(node) { |
771b53d0 | 1106 | struct io_wqe *wqe; |
7563439a | 1107 | int alloc_node = node; |
771b53d0 | 1108 | |
7563439a JA |
1109 | if (!node_online(alloc_node)) |
1110 | alloc_node = NUMA_NO_NODE; | |
1111 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1112 | if (!wqe) |
3fc50ab5 JH |
1113 | goto err; |
1114 | wq->wqes[node] = wqe; | |
7563439a | 1115 | wqe->node = alloc_node; |
c5def4ab JA |
1116 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1117 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1118 | if (wq->user) { |
c5def4ab JA |
1119 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1120 | task_rlimit(current, RLIMIT_NPROC); | |
1121 | } | |
1122 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 | 1123 | wqe->wq = wq; |
95da8465 | 1124 | raw_spin_lock_init(&wqe->lock); |
6206f0e1 | 1125 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1126 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1127 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1128 | } |
1129 | ||
1130 | init_completion(&wq->done); | |
1131 | ||
771b53d0 JA |
1132 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
1133 | if (!IS_ERR(wq->manager)) { | |
1134 | wake_up_process(wq->manager); | |
b60fda60 JA |
1135 | wait_for_completion(&wq->done); |
1136 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1137 | ret = -ENOMEM; | |
1138 | goto err; | |
1139 | } | |
848f7e18 | 1140 | refcount_set(&wq->use_refs, 1); |
b60fda60 | 1141 | reinit_completion(&wq->done); |
771b53d0 JA |
1142 | return wq; |
1143 | } | |
1144 | ||
1145 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1146 | complete(&wq->done); |
b60fda60 | 1147 | err: |
3fc50ab5 JH |
1148 | for_each_node(node) |
1149 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1150 | kfree(wq->wqes); |
1151 | kfree(wq); | |
771b53d0 JA |
1152 | return ERR_PTR(ret); |
1153 | } | |
1154 | ||
eba6f5a3 PB |
1155 | bool io_wq_get(struct io_wq *wq, struct io_wq_data *data) |
1156 | { | |
f5fa38c5 | 1157 | if (data->free_work != wq->free_work || data->do_work != wq->do_work) |
eba6f5a3 PB |
1158 | return false; |
1159 | ||
1160 | return refcount_inc_not_zero(&wq->use_refs); | |
1161 | } | |
1162 | ||
848f7e18 | 1163 | static void __io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1164 | { |
3fc50ab5 | 1165 | int node; |
771b53d0 | 1166 | |
b60fda60 JA |
1167 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1168 | if (wq->manager) | |
771b53d0 | 1169 | kthread_stop(wq->manager); |
771b53d0 JA |
1170 | |
1171 | rcu_read_lock(); | |
3fc50ab5 JH |
1172 | for_each_node(node) |
1173 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1174 | rcu_read_unlock(); |
1175 | ||
1176 | wait_for_completion(&wq->done); | |
1177 | ||
3fc50ab5 JH |
1178 | for_each_node(node) |
1179 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1180 | kfree(wq->wqes); |
1181 | kfree(wq); | |
1182 | } | |
848f7e18 JA |
1183 | |
1184 | void io_wq_destroy(struct io_wq *wq) | |
1185 | { | |
1186 | if (refcount_dec_and_test(&wq->use_refs)) | |
1187 | __io_wq_destroy(wq); | |
1188 | } | |
aa96bf8a JA |
1189 | |
1190 | struct task_struct *io_wq_get_task(struct io_wq *wq) | |
1191 | { | |
1192 | return wq->manager; | |
1193 | } |