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> |
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; | |
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 | |
190 | ||
fcb323cc | 191 | return dropped_lock; |
771b53d0 JA |
192 | } |
193 | ||
c5def4ab JA |
194 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
195 | struct io_wq_work *work) | |
196 | { | |
197 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
198 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
199 | ||
200 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
201 | } | |
202 | ||
203 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe, | |
204 | struct io_worker *worker) | |
205 | { | |
206 | if (worker->flags & IO_WORKER_F_BOUND) | |
207 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
208 | ||
209 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
210 | } | |
211 | ||
771b53d0 JA |
212 | static void io_worker_exit(struct io_worker *worker) |
213 | { | |
214 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab JA |
215 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); |
216 | unsigned nr_workers; | |
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 JA |
243 | acct->nr_workers--; |
244 | nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers + | |
245 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers; | |
95da8465 | 246 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
247 | |
248 | /* all workers gone, wq exit can proceed */ | |
c5def4ab | 249 | if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) |
771b53d0 JA |
250 | complete(&wqe->wq->done); |
251 | ||
364b05fd | 252 | kfree_rcu(worker, rcu); |
771b53d0 JA |
253 | } |
254 | ||
c5def4ab JA |
255 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
256 | __must_hold(wqe->lock) | |
257 | { | |
6206f0e1 JA |
258 | if (!wq_list_empty(&wqe->work_list) && |
259 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
260 | return true; |
261 | return false; | |
262 | } | |
263 | ||
264 | /* | |
265 | * Check head of free list for an available worker. If one isn't available, | |
266 | * caller must wake up the wq manager to create one. | |
267 | */ | |
268 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
269 | __must_hold(RCU) | |
270 | { | |
271 | struct hlist_nulls_node *n; | |
272 | struct io_worker *worker; | |
273 | ||
021d1cdd | 274 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
275 | if (is_a_nulls(n)) |
276 | return false; | |
277 | ||
278 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
279 | if (io_worker_get(worker)) { | |
506d95ff | 280 | wake_up_process(worker->task); |
c5def4ab JA |
281 | io_worker_release(worker); |
282 | return true; | |
283 | } | |
284 | ||
285 | return false; | |
286 | } | |
287 | ||
288 | /* | |
289 | * We need a worker. If we find a free one, we're good. If not, and we're | |
290 | * below the max number of workers, wake up the manager to create one. | |
291 | */ | |
292 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
293 | { | |
294 | bool ret; | |
295 | ||
296 | /* | |
297 | * Most likely an attempt to queue unbounded work on an io_wq that | |
298 | * wasn't setup with any unbounded workers. | |
299 | */ | |
300 | WARN_ON_ONCE(!acct->max_workers); | |
301 | ||
302 | rcu_read_lock(); | |
303 | ret = io_wqe_activate_free_worker(wqe); | |
304 | rcu_read_unlock(); | |
305 | ||
306 | if (!ret && acct->nr_workers < acct->max_workers) | |
307 | wake_up_process(wqe->wq->manager); | |
308 | } | |
309 | ||
310 | static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker) | |
311 | { | |
312 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
313 | ||
314 | atomic_inc(&acct->nr_running); | |
315 | } | |
316 | ||
317 | static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker) | |
318 | __must_hold(wqe->lock) | |
319 | { | |
320 | struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker); | |
321 | ||
322 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
323 | io_wqe_wake_worker(wqe, acct); | |
324 | } | |
325 | ||
771b53d0 JA |
326 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
327 | { | |
328 | allow_kernel_signal(SIGINT); | |
329 | ||
330 | current->flags |= PF_IO_WORKER; | |
331 | ||
332 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); | |
fcb323cc | 333 | worker->restore_files = current->files; |
9b828492 | 334 | worker->restore_nsproxy = current->nsproxy; |
9392a27d | 335 | worker->restore_fs = current->fs; |
c5def4ab | 336 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
337 | } |
338 | ||
339 | /* | |
340 | * Worker will start processing some work. Move it to the busy list, if | |
341 | * it's currently on the freelist | |
342 | */ | |
343 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
344 | struct io_wq_work *work) | |
345 | __must_hold(wqe->lock) | |
346 | { | |
c5def4ab JA |
347 | bool worker_bound, work_bound; |
348 | ||
771b53d0 JA |
349 | if (worker->flags & IO_WORKER_F_FREE) { |
350 | worker->flags &= ~IO_WORKER_F_FREE; | |
351 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 352 | } |
c5def4ab JA |
353 | |
354 | /* | |
355 | * If worker is moving from bound to unbound (or vice versa), then | |
356 | * ensure we update the running accounting. | |
357 | */ | |
b2e9c7d6 DC |
358 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
359 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
360 | if (worker_bound != work_bound) { | |
c5def4ab JA |
361 | io_wqe_dec_running(wqe, worker); |
362 | if (work_bound) { | |
363 | worker->flags |= IO_WORKER_F_BOUND; | |
364 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
365 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
366 | atomic_dec(&wqe->wq->user->processes); | |
367 | } else { | |
368 | worker->flags &= ~IO_WORKER_F_BOUND; | |
369 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
370 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
371 | atomic_inc(&wqe->wq->user->processes); | |
372 | } | |
373 | io_wqe_inc_running(wqe, worker); | |
374 | } | |
771b53d0 JA |
375 | } |
376 | ||
377 | /* | |
378 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
379 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
380 | * the lock in that case and return success. Since the caller has to | |
381 | * retry the loop in that case (we changed task state), we don't regrab | |
382 | * the lock if we return success. | |
383 | */ | |
384 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) | |
385 | __must_hold(wqe->lock) | |
386 | { | |
387 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
388 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 389 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 JA |
390 | } |
391 | ||
392 | return __io_worker_unuse(wqe, worker); | |
393 | } | |
394 | ||
60cf46ae PB |
395 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
396 | { | |
397 | return work->flags >> IO_WQ_HASH_SHIFT; | |
398 | } | |
399 | ||
400 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe) | |
771b53d0 JA |
401 | __must_hold(wqe->lock) |
402 | { | |
6206f0e1 | 403 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 404 | struct io_wq_work *work, *tail; |
60cf46ae | 405 | unsigned int hash; |
771b53d0 | 406 | |
6206f0e1 JA |
407 | wq_list_for_each(node, prev, &wqe->work_list) { |
408 | work = container_of(node, struct io_wq_work, list); | |
409 | ||
771b53d0 | 410 | /* not hashed, can run anytime */ |
8766dd51 | 411 | if (!io_wq_is_hashed(work)) { |
86f3cd1b | 412 | wq_list_del(&wqe->work_list, node, prev); |
771b53d0 JA |
413 | return work; |
414 | } | |
415 | ||
416 | /* hashed, can run if not already running */ | |
60cf46ae PB |
417 | hash = io_get_work_hash(work); |
418 | if (!(wqe->hash_map & BIT(hash))) { | |
419 | wqe->hash_map |= BIT(hash); | |
86f3cd1b PB |
420 | /* all items with this hash lie in [work, tail] */ |
421 | tail = wqe->hash_tail[hash]; | |
422 | wqe->hash_tail[hash] = NULL; | |
423 | wq_list_cut(&wqe->work_list, &tail->list, prev); | |
771b53d0 JA |
424 | return work; |
425 | } | |
426 | } | |
427 | ||
428 | return NULL; | |
429 | } | |
430 | ||
cccf0ee8 JA |
431 | static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work) |
432 | { | |
433 | if (worker->mm) { | |
f5678e7f | 434 | kthread_unuse_mm(worker->mm); |
cccf0ee8 JA |
435 | mmput(worker->mm); |
436 | worker->mm = NULL; | |
437 | } | |
37c54f9b | 438 | if (!work->mm) |
cccf0ee8 | 439 | return; |
37c54f9b | 440 | |
cccf0ee8 | 441 | if (mmget_not_zero(work->mm)) { |
f5678e7f | 442 | kthread_use_mm(work->mm); |
cccf0ee8 JA |
443 | worker->mm = work->mm; |
444 | /* hang on to this mm */ | |
445 | work->mm = NULL; | |
446 | return; | |
447 | } | |
448 | ||
449 | /* failed grabbing mm, ensure work gets cancelled */ | |
450 | work->flags |= IO_WQ_WORK_CANCEL; | |
451 | } | |
452 | ||
91d8f519 DZ |
453 | static inline void io_wq_switch_blkcg(struct io_worker *worker, |
454 | struct io_wq_work *work) | |
455 | { | |
456 | #ifdef CONFIG_BLK_CGROUP | |
457 | if (work->blkcg_css != worker->blkcg_css) { | |
458 | kthread_associate_blkcg(work->blkcg_css); | |
459 | worker->blkcg_css = work->blkcg_css; | |
460 | } | |
461 | #endif | |
462 | } | |
463 | ||
cccf0ee8 JA |
464 | static void io_wq_switch_creds(struct io_worker *worker, |
465 | struct io_wq_work *work) | |
466 | { | |
467 | const struct cred *old_creds = override_creds(work->creds); | |
468 | ||
469 | worker->cur_creds = work->creds; | |
470 | if (worker->saved_creds) | |
471 | put_cred(old_creds); /* creds set by previous switch */ | |
472 | else | |
473 | worker->saved_creds = old_creds; | |
474 | } | |
475 | ||
dc026a73 PB |
476 | static void io_impersonate_work(struct io_worker *worker, |
477 | struct io_wq_work *work) | |
478 | { | |
479 | if (work->files && current->files != work->files) { | |
480 | task_lock(current); | |
481 | current->files = work->files; | |
9b828492 | 482 | current->nsproxy = work->nsproxy; |
dc026a73 PB |
483 | task_unlock(current); |
484 | } | |
485 | if (work->fs && current->fs != work->fs) | |
486 | current->fs = work->fs; | |
487 | if (work->mm != worker->mm) | |
488 | io_wq_switch_mm(worker, work); | |
489 | if (worker->cur_creds != work->creds) | |
490 | io_wq_switch_creds(worker, work); | |
57f1a649 | 491 | current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->fsize; |
91d8f519 | 492 | io_wq_switch_blkcg(worker, work); |
dc026a73 PB |
493 | } |
494 | ||
495 | static void io_assign_current_work(struct io_worker *worker, | |
496 | struct io_wq_work *work) | |
497 | { | |
d78298e7 PB |
498 | if (work) { |
499 | /* flush pending signals before assigning new work */ | |
500 | if (signal_pending(current)) | |
501 | flush_signals(current); | |
502 | cond_resched(); | |
503 | } | |
dc026a73 PB |
504 | |
505 | spin_lock_irq(&worker->lock); | |
506 | worker->cur_work = work; | |
507 | spin_unlock_irq(&worker->lock); | |
508 | } | |
509 | ||
60cf46ae PB |
510 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
511 | ||
771b53d0 JA |
512 | static void io_worker_handle_work(struct io_worker *worker) |
513 | __releases(wqe->lock) | |
514 | { | |
771b53d0 JA |
515 | struct io_wqe *wqe = worker->wqe; |
516 | struct io_wq *wq = wqe->wq; | |
517 | ||
518 | do { | |
86f3cd1b | 519 | struct io_wq_work *work; |
f462fd36 | 520 | get_next: |
771b53d0 JA |
521 | /* |
522 | * If we got some work, mark us as busy. If we didn't, but | |
523 | * the list isn't empty, it means we stalled on hashed work. | |
524 | * Mark us stalled so we don't keep looking for work when we | |
525 | * can't make progress, any work completion or insertion will | |
526 | * clear the stalled flag. | |
527 | */ | |
60cf46ae | 528 | work = io_get_next_work(wqe); |
771b53d0 JA |
529 | if (work) |
530 | __io_worker_busy(wqe, worker, work); | |
6206f0e1 | 531 | else if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
532 | wqe->flags |= IO_WQE_FLAG_STALLED; |
533 | ||
95da8465 | 534 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
535 | if (!work) |
536 | break; | |
58e39319 | 537 | io_assign_current_work(worker, work); |
36c2f922 | 538 | |
dc026a73 PB |
539 | /* handle a whole dependent link */ |
540 | do { | |
86f3cd1b | 541 | struct io_wq_work *old_work, *next_hashed, *linked; |
b089ed39 | 542 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 543 | |
86f3cd1b | 544 | next_hashed = wq_next_work(work); |
58e39319 | 545 | io_impersonate_work(worker, work); |
dc026a73 PB |
546 | /* |
547 | * OK to set IO_WQ_WORK_CANCEL even for uncancellable | |
548 | * work, the worker function will do the right thing. | |
549 | */ | |
550 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) | |
551 | work->flags |= IO_WQ_WORK_CANCEL; | |
552 | ||
f4db7182 PB |
553 | old_work = work; |
554 | linked = wq->do_work(work); | |
86f3cd1b PB |
555 | |
556 | work = next_hashed; | |
557 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
558 | work = linked; | |
559 | linked = NULL; | |
560 | } | |
561 | io_assign_current_work(worker, work); | |
e9fd9396 | 562 | wq->free_work(old_work); |
dc026a73 | 563 | |
86f3cd1b PB |
564 | if (linked) |
565 | io_wqe_enqueue(wqe, linked); | |
566 | ||
567 | if (hash != -1U && !next_hashed) { | |
95da8465 | 568 | raw_spin_lock_irq(&wqe->lock); |
dc026a73 PB |
569 | wqe->hash_map &= ~BIT_ULL(hash); |
570 | wqe->flags &= ~IO_WQE_FLAG_STALLED; | |
f462fd36 PB |
571 | /* skip unnecessary unlock-lock wqe->lock */ |
572 | if (!work) | |
573 | goto get_next; | |
95da8465 | 574 | raw_spin_unlock_irq(&wqe->lock); |
7d723065 | 575 | } |
58e39319 | 576 | } while (work); |
7d723065 | 577 | |
95da8465 | 578 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 JA |
579 | } while (1); |
580 | } | |
581 | ||
771b53d0 JA |
582 | static int io_wqe_worker(void *data) |
583 | { | |
584 | struct io_worker *worker = data; | |
585 | struct io_wqe *wqe = worker->wqe; | |
586 | struct io_wq *wq = wqe->wq; | |
771b53d0 JA |
587 | |
588 | io_worker_start(wqe, worker); | |
589 | ||
590 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
506d95ff | 591 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 592 | loop: |
95da8465 | 593 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 JA |
594 | if (io_wqe_run_queue(wqe)) { |
595 | __set_current_state(TASK_RUNNING); | |
596 | io_worker_handle_work(worker); | |
e995d512 | 597 | goto loop; |
771b53d0 JA |
598 | } |
599 | /* drops the lock on success, retry */ | |
600 | if (__io_worker_idle(wqe, worker)) { | |
601 | __release(&wqe->lock); | |
e995d512 | 602 | goto loop; |
771b53d0 | 603 | } |
95da8465 | 604 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
605 | if (signal_pending(current)) |
606 | flush_signals(current); | |
607 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) | |
608 | continue; | |
609 | /* timed out, exit unless we're the fixed worker */ | |
610 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
611 | !(worker->flags & IO_WORKER_F_FIXED)) | |
612 | break; | |
613 | } | |
614 | ||
771b53d0 | 615 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
95da8465 | 616 | raw_spin_lock_irq(&wqe->lock); |
6206f0e1 | 617 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
618 | io_worker_handle_work(worker); |
619 | else | |
95da8465 | 620 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
621 | } |
622 | ||
623 | io_worker_exit(worker); | |
624 | return 0; | |
625 | } | |
626 | ||
771b53d0 JA |
627 | /* |
628 | * Called when a worker is scheduled in. Mark us as currently running. | |
629 | */ | |
630 | void io_wq_worker_running(struct task_struct *tsk) | |
631 | { | |
632 | struct io_worker *worker = kthread_data(tsk); | |
633 | struct io_wqe *wqe = worker->wqe; | |
634 | ||
635 | if (!(worker->flags & IO_WORKER_F_UP)) | |
636 | return; | |
637 | if (worker->flags & IO_WORKER_F_RUNNING) | |
638 | return; | |
639 | worker->flags |= IO_WORKER_F_RUNNING; | |
c5def4ab | 640 | io_wqe_inc_running(wqe, worker); |
771b53d0 JA |
641 | } |
642 | ||
643 | /* | |
644 | * Called when worker is going to sleep. If there are no workers currently | |
645 | * running and we have work pending, wake up a free one or have the manager | |
646 | * set one up. | |
647 | */ | |
648 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
649 | { | |
650 | struct io_worker *worker = kthread_data(tsk); | |
651 | struct io_wqe *wqe = worker->wqe; | |
652 | ||
653 | if (!(worker->flags & IO_WORKER_F_UP)) | |
654 | return; | |
655 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
656 | return; | |
657 | ||
658 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
659 | ||
95da8465 | 660 | raw_spin_lock_irq(&wqe->lock); |
c5def4ab | 661 | io_wqe_dec_running(wqe, worker); |
95da8465 | 662 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
663 | } |
664 | ||
b60fda60 | 665 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
771b53d0 | 666 | { |
c5def4ab | 667 | struct io_wqe_acct *acct =&wqe->acct[index]; |
771b53d0 JA |
668 | struct io_worker *worker; |
669 | ||
ad6e005c | 670 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
771b53d0 | 671 | if (!worker) |
b60fda60 | 672 | return false; |
771b53d0 JA |
673 | |
674 | refcount_set(&worker->ref, 1); | |
675 | worker->nulls_node.pprev = NULL; | |
771b53d0 | 676 | worker->wqe = wqe; |
36c2f922 | 677 | spin_lock_init(&worker->lock); |
771b53d0 JA |
678 | |
679 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, | |
c5def4ab | 680 | "io_wqe_worker-%d/%d", index, wqe->node); |
771b53d0 JA |
681 | if (IS_ERR(worker->task)) { |
682 | kfree(worker); | |
b60fda60 | 683 | return false; |
771b53d0 JA |
684 | } |
685 | ||
95da8465 | 686 | raw_spin_lock_irq(&wqe->lock); |
021d1cdd | 687 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
e61df66c | 688 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); |
771b53d0 | 689 | worker->flags |= IO_WORKER_F_FREE; |
c5def4ab JA |
690 | if (index == IO_WQ_ACCT_BOUND) |
691 | worker->flags |= IO_WORKER_F_BOUND; | |
692 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
771b53d0 | 693 | worker->flags |= IO_WORKER_F_FIXED; |
c5def4ab | 694 | acct->nr_workers++; |
95da8465 | 695 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 | 696 | |
c5def4ab JA |
697 | if (index == IO_WQ_ACCT_UNBOUND) |
698 | atomic_inc(&wq->user->processes); | |
699 | ||
771b53d0 | 700 | wake_up_process(worker->task); |
b60fda60 | 701 | return true; |
771b53d0 JA |
702 | } |
703 | ||
c5def4ab | 704 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
705 | __must_hold(wqe->lock) |
706 | { | |
c5def4ab | 707 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 708 | |
c5def4ab | 709 | /* if we have available workers or no work, no need */ |
021d1cdd | 710 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
711 | return false; |
712 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
713 | } |
714 | ||
715 | /* | |
716 | * Manager thread. Tasked with creating new workers, if we need them. | |
717 | */ | |
718 | static int io_wq_manager(void *data) | |
719 | { | |
720 | struct io_wq *wq = data; | |
3fc50ab5 JH |
721 | int workers_to_create = num_possible_nodes(); |
722 | int node; | |
771b53d0 | 723 | |
b60fda60 | 724 | /* create fixed workers */ |
3fc50ab5 JH |
725 | refcount_set(&wq->refs, workers_to_create); |
726 | for_each_node(node) { | |
7563439a JA |
727 | if (!node_online(node)) |
728 | continue; | |
3fc50ab5 JH |
729 | if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND)) |
730 | goto err; | |
731 | workers_to_create--; | |
b60fda60 | 732 | } |
771b53d0 | 733 | |
7563439a JA |
734 | while (workers_to_create--) |
735 | refcount_dec(&wq->refs); | |
736 | ||
b60fda60 JA |
737 | complete(&wq->done); |
738 | ||
739 | while (!kthread_should_stop()) { | |
aa96bf8a JA |
740 | if (current->task_works) |
741 | task_work_run(); | |
742 | ||
3fc50ab5 JH |
743 | for_each_node(node) { |
744 | struct io_wqe *wqe = wq->wqes[node]; | |
c5def4ab | 745 | bool fork_worker[2] = { false, false }; |
771b53d0 | 746 | |
7563439a JA |
747 | if (!node_online(node)) |
748 | continue; | |
749 | ||
95da8465 | 750 | raw_spin_lock_irq(&wqe->lock); |
c5def4ab JA |
751 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) |
752 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
753 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
754 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
95da8465 | 755 | raw_spin_unlock_irq(&wqe->lock); |
c5def4ab JA |
756 | if (fork_worker[IO_WQ_ACCT_BOUND]) |
757 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
758 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
759 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
771b53d0 JA |
760 | } |
761 | set_current_state(TASK_INTERRUPTIBLE); | |
762 | schedule_timeout(HZ); | |
763 | } | |
764 | ||
aa96bf8a JA |
765 | if (current->task_works) |
766 | task_work_run(); | |
767 | ||
771b53d0 | 768 | return 0; |
b60fda60 JA |
769 | err: |
770 | set_bit(IO_WQ_BIT_ERROR, &wq->state); | |
771 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
3fc50ab5 | 772 | if (refcount_sub_and_test(workers_to_create, &wq->refs)) |
b60fda60 JA |
773 | complete(&wq->done); |
774 | return 0; | |
771b53d0 JA |
775 | } |
776 | ||
c5def4ab JA |
777 | static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct, |
778 | struct io_wq_work *work) | |
779 | { | |
780 | bool free_worker; | |
781 | ||
782 | if (!(work->flags & IO_WQ_WORK_UNBOUND)) | |
783 | return true; | |
784 | if (atomic_read(&acct->nr_running)) | |
785 | return true; | |
786 | ||
787 | rcu_read_lock(); | |
021d1cdd | 788 | free_worker = !hlist_nulls_empty(&wqe->free_list); |
c5def4ab JA |
789 | rcu_read_unlock(); |
790 | if (free_worker) | |
791 | return true; | |
792 | ||
793 | if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers && | |
794 | !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN))) | |
795 | return false; | |
796 | ||
797 | return true; | |
798 | } | |
799 | ||
e9fd9396 | 800 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 801 | { |
e9fd9396 PB |
802 | struct io_wq *wq = wqe->wq; |
803 | ||
fc04c39b PB |
804 | do { |
805 | struct io_wq_work *old_work = work; | |
806 | ||
807 | work->flags |= IO_WQ_WORK_CANCEL; | |
f4db7182 | 808 | work = wq->do_work(work); |
e9fd9396 | 809 | wq->free_work(old_work); |
fc04c39b PB |
810 | } while (work); |
811 | } | |
812 | ||
86f3cd1b PB |
813 | static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work) |
814 | { | |
815 | unsigned int hash; | |
816 | struct io_wq_work *tail; | |
817 | ||
818 | if (!io_wq_is_hashed(work)) { | |
819 | append: | |
820 | wq_list_add_tail(&work->list, &wqe->work_list); | |
821 | return; | |
822 | } | |
823 | ||
824 | hash = io_get_work_hash(work); | |
825 | tail = wqe->hash_tail[hash]; | |
826 | wqe->hash_tail[hash] = work; | |
827 | if (!tail) | |
828 | goto append; | |
829 | ||
830 | wq_list_add_after(&work->list, &tail->list, &wqe->work_list); | |
831 | } | |
832 | ||
771b53d0 JA |
833 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
834 | { | |
c5def4ab | 835 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 836 | int work_flags; |
771b53d0 JA |
837 | unsigned long flags; |
838 | ||
c5def4ab JA |
839 | /* |
840 | * Do early check to see if we need a new unbound worker, and if we do, | |
841 | * if we're allowed to do so. This isn't 100% accurate as there's a | |
842 | * gap between this check and incrementing the value, but that's OK. | |
843 | * It's close enough to not be an issue, fork() has the same delay. | |
844 | */ | |
845 | if (unlikely(!io_wq_can_queue(wqe, acct, work))) { | |
e9fd9396 | 846 | io_run_cancel(work, wqe); |
c5def4ab JA |
847 | return; |
848 | } | |
849 | ||
895e2ca0 | 850 | work_flags = work->flags; |
95da8465 | 851 | raw_spin_lock_irqsave(&wqe->lock, flags); |
86f3cd1b | 852 | io_wqe_insert_work(wqe, work); |
771b53d0 | 853 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
95da8465 | 854 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 | 855 | |
895e2ca0 JA |
856 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
857 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 858 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
859 | } |
860 | ||
861 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
862 | { | |
863 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
864 | ||
865 | io_wqe_enqueue(wqe, work); | |
866 | } | |
867 | ||
868 | /* | |
8766dd51 PB |
869 | * Work items that hash to the same value will not be done in parallel. |
870 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 871 | */ |
8766dd51 | 872 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 873 | { |
8766dd51 | 874 | unsigned int bit; |
771b53d0 JA |
875 | |
876 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
877 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
878 | } |
879 | ||
880 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) | |
881 | { | |
882 | send_sig(SIGINT, worker->task, 1); | |
883 | return false; | |
884 | } | |
885 | ||
886 | /* | |
887 | * Iterate the passed in list and call the specific function for each | |
888 | * worker that isn't exiting | |
889 | */ | |
890 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
771b53d0 JA |
891 | bool (*func)(struct io_worker *, void *), |
892 | void *data) | |
893 | { | |
771b53d0 JA |
894 | struct io_worker *worker; |
895 | bool ret = false; | |
896 | ||
e61df66c | 897 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { |
771b53d0 | 898 | if (io_worker_get(worker)) { |
7563439a JA |
899 | /* no task if node is/was offline */ |
900 | if (worker->task) | |
901 | ret = func(worker, data); | |
771b53d0 JA |
902 | io_worker_release(worker); |
903 | if (ret) | |
904 | break; | |
905 | } | |
906 | } | |
e61df66c | 907 | |
771b53d0 JA |
908 | return ret; |
909 | } | |
910 | ||
911 | void io_wq_cancel_all(struct io_wq *wq) | |
912 | { | |
3fc50ab5 | 913 | int node; |
771b53d0 JA |
914 | |
915 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); | |
916 | ||
771b53d0 | 917 | rcu_read_lock(); |
3fc50ab5 JH |
918 | for_each_node(node) { |
919 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 920 | |
e61df66c | 921 | io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL); |
771b53d0 JA |
922 | } |
923 | rcu_read_unlock(); | |
924 | } | |
925 | ||
62755e35 | 926 | struct io_cb_cancel_data { |
2293b419 PB |
927 | work_cancel_fn *fn; |
928 | void *data; | |
4f26bda1 PB |
929 | int nr_running; |
930 | int nr_pending; | |
931 | bool cancel_all; | |
62755e35 JA |
932 | }; |
933 | ||
2293b419 | 934 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 935 | { |
2293b419 | 936 | struct io_cb_cancel_data *match = data; |
6f72653e | 937 | unsigned long flags; |
62755e35 JA |
938 | |
939 | /* | |
940 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 941 | * may dereference the passed in work. |
62755e35 | 942 | */ |
36c2f922 | 943 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 944 | if (worker->cur_work && |
0c9d5ccd | 945 | !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) && |
2293b419 | 946 | match->fn(worker->cur_work, match->data)) { |
771b53d0 | 947 | send_sig(SIGINT, worker->task, 1); |
4f26bda1 | 948 | match->nr_running++; |
771b53d0 | 949 | } |
36c2f922 | 950 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 951 | |
4f26bda1 | 952 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
953 | } |
954 | ||
204361a7 PB |
955 | static inline void io_wqe_remove_pending(struct io_wqe *wqe, |
956 | struct io_wq_work *work, | |
957 | struct io_wq_work_node *prev) | |
958 | { | |
959 | unsigned int hash = io_get_work_hash(work); | |
960 | struct io_wq_work *prev_work = NULL; | |
961 | ||
962 | if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) { | |
963 | if (prev) | |
964 | prev_work = container_of(prev, struct io_wq_work, list); | |
965 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
966 | wqe->hash_tail[hash] = prev_work; | |
967 | else | |
968 | wqe->hash_tail[hash] = NULL; | |
969 | } | |
970 | wq_list_del(&wqe->work_list, &work->list, prev); | |
971 | } | |
972 | ||
4f26bda1 | 973 | static void io_wqe_cancel_pending_work(struct io_wqe *wqe, |
f4c2665e | 974 | struct io_cb_cancel_data *match) |
771b53d0 | 975 | { |
6206f0e1 | 976 | struct io_wq_work_node *node, *prev; |
771b53d0 | 977 | struct io_wq_work *work; |
6f72653e | 978 | unsigned long flags; |
771b53d0 | 979 | |
4f26bda1 | 980 | retry: |
95da8465 | 981 | raw_spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
982 | wq_list_for_each(node, prev, &wqe->work_list) { |
983 | work = container_of(node, struct io_wq_work, list); | |
4f26bda1 PB |
984 | if (!match->fn(work, match->data)) |
985 | continue; | |
204361a7 | 986 | io_wqe_remove_pending(wqe, work, prev); |
95da8465 | 987 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
4f26bda1 PB |
988 | io_run_cancel(work, wqe); |
989 | match->nr_pending++; | |
990 | if (!match->cancel_all) | |
991 | return; | |
992 | ||
993 | /* not safe to continue after unlock */ | |
994 | goto retry; | |
771b53d0 | 995 | } |
95da8465 | 996 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
f4c2665e PB |
997 | } |
998 | ||
4f26bda1 | 999 | static void io_wqe_cancel_running_work(struct io_wqe *wqe, |
f4c2665e PB |
1000 | struct io_cb_cancel_data *match) |
1001 | { | |
771b53d0 | 1002 | rcu_read_lock(); |
4f26bda1 | 1003 | io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 | 1004 | rcu_read_unlock(); |
771b53d0 JA |
1005 | } |
1006 | ||
2293b419 | 1007 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 1008 | void *data, bool cancel_all) |
771b53d0 | 1009 | { |
2293b419 | 1010 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
1011 | .fn = cancel, |
1012 | .data = data, | |
1013 | .cancel_all = cancel_all, | |
00bcda13 | 1014 | }; |
3fc50ab5 | 1015 | int node; |
771b53d0 | 1016 | |
f4c2665e PB |
1017 | /* |
1018 | * First check pending list, if we're lucky we can just remove it | |
1019 | * from there. CANCEL_OK means that the work is returned as-new, | |
1020 | * no completion will be posted for it. | |
1021 | */ | |
3fc50ab5 JH |
1022 | for_each_node(node) { |
1023 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 1024 | |
4f26bda1 PB |
1025 | io_wqe_cancel_pending_work(wqe, &match); |
1026 | if (match.nr_pending && !match.cancel_all) | |
f4c2665e | 1027 | return IO_WQ_CANCEL_OK; |
771b53d0 JA |
1028 | } |
1029 | ||
f4c2665e PB |
1030 | /* |
1031 | * Now check if a free (going busy) or busy worker has the work | |
1032 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
1033 | * as an indication that we attempt to signal cancellation. The | |
1034 | * completion will run normally in this case. | |
1035 | */ | |
1036 | for_each_node(node) { | |
1037 | struct io_wqe *wqe = wq->wqes[node]; | |
1038 | ||
4f26bda1 PB |
1039 | io_wqe_cancel_running_work(wqe, &match); |
1040 | if (match.nr_running && !match.cancel_all) | |
f4c2665e PB |
1041 | return IO_WQ_CANCEL_RUNNING; |
1042 | } | |
1043 | ||
4f26bda1 PB |
1044 | if (match.nr_running) |
1045 | return IO_WQ_CANCEL_RUNNING; | |
1046 | if (match.nr_pending) | |
1047 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 1048 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
1049 | } |
1050 | ||
2293b419 PB |
1051 | static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data) |
1052 | { | |
1053 | return work == data; | |
1054 | } | |
1055 | ||
1056 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) | |
1057 | { | |
4f26bda1 | 1058 | return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork, false); |
2293b419 PB |
1059 | } |
1060 | ||
576a347b | 1061 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1062 | { |
3fc50ab5 | 1063 | int ret = -ENOMEM, node; |
771b53d0 JA |
1064 | struct io_wq *wq; |
1065 | ||
f5fa38c5 | 1066 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 PB |
1067 | return ERR_PTR(-EINVAL); |
1068 | ||
ad6e005c | 1069 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
1070 | if (!wq) |
1071 | return ERR_PTR(-ENOMEM); | |
1072 | ||
3fc50ab5 | 1073 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
771b53d0 JA |
1074 | if (!wq->wqes) { |
1075 | kfree(wq); | |
1076 | return ERR_PTR(-ENOMEM); | |
1077 | } | |
1078 | ||
e9fd9396 | 1079 | wq->free_work = data->free_work; |
f5fa38c5 | 1080 | wq->do_work = data->do_work; |
7d723065 | 1081 | |
c5def4ab | 1082 | /* caller must already hold a reference to this */ |
576a347b | 1083 | wq->user = data->user; |
c5def4ab | 1084 | |
3fc50ab5 | 1085 | for_each_node(node) { |
771b53d0 | 1086 | struct io_wqe *wqe; |
7563439a | 1087 | int alloc_node = node; |
771b53d0 | 1088 | |
7563439a JA |
1089 | if (!node_online(alloc_node)) |
1090 | alloc_node = NUMA_NO_NODE; | |
1091 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1092 | if (!wqe) |
3fc50ab5 JH |
1093 | goto err; |
1094 | wq->wqes[node] = wqe; | |
7563439a | 1095 | wqe->node = alloc_node; |
c5def4ab JA |
1096 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1097 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
576a347b | 1098 | if (wq->user) { |
c5def4ab JA |
1099 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
1100 | task_rlimit(current, RLIMIT_NPROC); | |
1101 | } | |
1102 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); | |
771b53d0 | 1103 | wqe->wq = wq; |
95da8465 | 1104 | raw_spin_lock_init(&wqe->lock); |
6206f0e1 | 1105 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1106 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1107 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1108 | } |
1109 | ||
1110 | init_completion(&wq->done); | |
1111 | ||
771b53d0 JA |
1112 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
1113 | if (!IS_ERR(wq->manager)) { | |
1114 | wake_up_process(wq->manager); | |
b60fda60 JA |
1115 | wait_for_completion(&wq->done); |
1116 | if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) { | |
1117 | ret = -ENOMEM; | |
1118 | goto err; | |
1119 | } | |
848f7e18 | 1120 | refcount_set(&wq->use_refs, 1); |
b60fda60 | 1121 | reinit_completion(&wq->done); |
771b53d0 JA |
1122 | return wq; |
1123 | } | |
1124 | ||
1125 | ret = PTR_ERR(wq->manager); | |
771b53d0 | 1126 | complete(&wq->done); |
b60fda60 | 1127 | err: |
3fc50ab5 JH |
1128 | for_each_node(node) |
1129 | kfree(wq->wqes[node]); | |
b60fda60 JA |
1130 | kfree(wq->wqes); |
1131 | kfree(wq); | |
771b53d0 JA |
1132 | return ERR_PTR(ret); |
1133 | } | |
1134 | ||
eba6f5a3 PB |
1135 | bool io_wq_get(struct io_wq *wq, struct io_wq_data *data) |
1136 | { | |
f5fa38c5 | 1137 | if (data->free_work != wq->free_work || data->do_work != wq->do_work) |
eba6f5a3 PB |
1138 | return false; |
1139 | ||
1140 | return refcount_inc_not_zero(&wq->use_refs); | |
1141 | } | |
1142 | ||
771b53d0 JA |
1143 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) |
1144 | { | |
1145 | wake_up_process(worker->task); | |
1146 | return false; | |
1147 | } | |
1148 | ||
848f7e18 | 1149 | static void __io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1150 | { |
3fc50ab5 | 1151 | int node; |
771b53d0 | 1152 | |
b60fda60 JA |
1153 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
1154 | if (wq->manager) | |
771b53d0 | 1155 | kthread_stop(wq->manager); |
771b53d0 JA |
1156 | |
1157 | rcu_read_lock(); | |
3fc50ab5 JH |
1158 | for_each_node(node) |
1159 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
771b53d0 JA |
1160 | rcu_read_unlock(); |
1161 | ||
1162 | wait_for_completion(&wq->done); | |
1163 | ||
3fc50ab5 JH |
1164 | for_each_node(node) |
1165 | kfree(wq->wqes[node]); | |
771b53d0 JA |
1166 | kfree(wq->wqes); |
1167 | kfree(wq); | |
1168 | } | |
848f7e18 JA |
1169 | |
1170 | void io_wq_destroy(struct io_wq *wq) | |
1171 | { | |
1172 | if (refcount_dec_and_test(&wq->use_refs)) | |
1173 | __io_wq_destroy(wq); | |
1174 | } | |
aa96bf8a JA |
1175 | |
1176 | struct task_struct *io_wq_get_task(struct io_wq *wq) | |
1177 | { | |
1178 | return wq->manager; | |
1179 | } |