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> | |
771b53d0 | 16 | #include <linux/rculist_nulls.h> |
43c01fbe | 17 | #include <linux/cpu.h> |
3bfe6106 | 18 | #include <linux/tracehook.h> |
e4b4a13f | 19 | #include <linux/freezer.h> |
771b53d0 | 20 | |
43c01fbe | 21 | #include "../kernel/sched/sched.h" |
771b53d0 JA |
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 */ | |
145cc8c6 JA |
30 | IO_WORKER_F_FIXED = 8, /* static idle worker */ |
31 | IO_WORKER_F_BOUND = 16, /* is doing bounded work */ | |
771b53d0 JA |
32 | }; |
33 | ||
34 | enum { | |
35 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ | |
771b53d0 JA |
36 | }; |
37 | ||
38 | enum { | |
39 | IO_WQE_FLAG_STALLED = 1, /* stalled on hash */ | |
40 | }; | |
41 | ||
42 | /* | |
43 | * One for each thread in a wqe pool | |
44 | */ | |
45 | struct io_worker { | |
46 | refcount_t ref; | |
47 | unsigned flags; | |
48 | struct hlist_nulls_node nulls_node; | |
e61df66c | 49 | struct list_head all_list; |
771b53d0 | 50 | struct task_struct *task; |
771b53d0 | 51 | struct io_wqe *wqe; |
36c2f922 | 52 | |
771b53d0 | 53 | struct io_wq_work *cur_work; |
36c2f922 | 54 | spinlock_t lock; |
771b53d0 | 55 | |
eb2de941 JA |
56 | struct completion ref_done; |
57 | ||
771b53d0 | 58 | struct rcu_head rcu; |
771b53d0 JA |
59 | }; |
60 | ||
771b53d0 JA |
61 | #if BITS_PER_LONG == 64 |
62 | #define IO_WQ_HASH_ORDER 6 | |
63 | #else | |
64 | #define IO_WQ_HASH_ORDER 5 | |
65 | #endif | |
66 | ||
86f3cd1b PB |
67 | #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER) |
68 | ||
c5def4ab JA |
69 | struct io_wqe_acct { |
70 | unsigned nr_workers; | |
71 | unsigned max_workers; | |
72 | atomic_t nr_running; | |
73 | }; | |
74 | ||
75 | enum { | |
76 | IO_WQ_ACCT_BOUND, | |
77 | IO_WQ_ACCT_UNBOUND, | |
78 | }; | |
79 | ||
771b53d0 JA |
80 | /* |
81 | * Per-node worker thread pool | |
82 | */ | |
83 | struct io_wqe { | |
84 | struct { | |
95da8465 | 85 | raw_spinlock_t lock; |
6206f0e1 | 86 | struct io_wq_work_list work_list; |
771b53d0 JA |
87 | unsigned flags; |
88 | } ____cacheline_aligned_in_smp; | |
89 | ||
90 | int node; | |
c5def4ab | 91 | struct io_wqe_acct acct[2]; |
771b53d0 | 92 | |
021d1cdd | 93 | struct hlist_nulls_head free_list; |
e61df66c | 94 | struct list_head all_list; |
771b53d0 | 95 | |
e941894e JA |
96 | struct wait_queue_entry wait; |
97 | ||
771b53d0 | 98 | struct io_wq *wq; |
86f3cd1b | 99 | struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS]; |
771b53d0 JA |
100 | }; |
101 | ||
102 | /* | |
103 | * Per io_wq state | |
104 | */ | |
105 | struct io_wq { | |
106 | struct io_wqe **wqes; | |
107 | unsigned long state; | |
771b53d0 | 108 | |
e9fd9396 | 109 | free_work_fn *free_work; |
f5fa38c5 | 110 | io_wq_work_fn *do_work; |
7d723065 | 111 | |
771b53d0 | 112 | struct task_struct *manager; |
e941894e JA |
113 | |
114 | struct io_wq_hash *hash; | |
115 | ||
771b53d0 | 116 | refcount_t refs; |
d364d9e5 | 117 | struct completion exited; |
848f7e18 | 118 | |
fb3a1f6c JA |
119 | atomic_t worker_refs; |
120 | struct completion worker_done; | |
121 | ||
43c01fbe | 122 | struct hlist_node cpuhp_node; |
3bfe6106 JA |
123 | |
124 | pid_t task_pid; | |
771b53d0 JA |
125 | }; |
126 | ||
43c01fbe JA |
127 | static enum cpuhp_state io_wq_online; |
128 | ||
f0127254 JA |
129 | struct io_cb_cancel_data { |
130 | work_cancel_fn *fn; | |
131 | void *data; | |
132 | int nr_running; | |
133 | int nr_pending; | |
134 | bool cancel_all; | |
135 | }; | |
136 | ||
137 | static void io_wqe_cancel_pending_work(struct io_wqe *wqe, | |
138 | struct io_cb_cancel_data *match); | |
139 | ||
771b53d0 JA |
140 | static bool io_worker_get(struct io_worker *worker) |
141 | { | |
142 | return refcount_inc_not_zero(&worker->ref); | |
143 | } | |
144 | ||
145 | static void io_worker_release(struct io_worker *worker) | |
146 | { | |
147 | if (refcount_dec_and_test(&worker->ref)) | |
eb2de941 | 148 | complete(&worker->ref_done); |
771b53d0 JA |
149 | } |
150 | ||
c5def4ab JA |
151 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
152 | struct io_wq_work *work) | |
153 | { | |
154 | if (work->flags & IO_WQ_WORK_UNBOUND) | |
155 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
156 | ||
157 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
158 | } | |
159 | ||
958234d5 | 160 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker) |
c5def4ab | 161 | { |
958234d5 JA |
162 | struct io_wqe *wqe = worker->wqe; |
163 | ||
c5def4ab JA |
164 | if (worker->flags & IO_WORKER_F_BOUND) |
165 | return &wqe->acct[IO_WQ_ACCT_BOUND]; | |
166 | ||
167 | return &wqe->acct[IO_WQ_ACCT_UNBOUND]; | |
168 | } | |
169 | ||
771b53d0 JA |
170 | static void io_worker_exit(struct io_worker *worker) |
171 | { | |
172 | struct io_wqe *wqe = worker->wqe; | |
958234d5 | 173 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
bf1daa4b | 174 | unsigned flags; |
771b53d0 | 175 | |
eb2de941 JA |
176 | if (refcount_dec_and_test(&worker->ref)) |
177 | complete(&worker->ref_done); | |
178 | wait_for_completion(&worker->ref_done); | |
771b53d0 JA |
179 | |
180 | preempt_disable(); | |
181 | current->flags &= ~PF_IO_WORKER; | |
bf1daa4b JA |
182 | flags = worker->flags; |
183 | worker->flags = 0; | |
184 | if (flags & IO_WORKER_F_RUNNING) | |
c5def4ab | 185 | atomic_dec(&acct->nr_running); |
771b53d0 JA |
186 | worker->flags = 0; |
187 | preempt_enable(); | |
188 | ||
95da8465 | 189 | raw_spin_lock_irq(&wqe->lock); |
bf1daa4b JA |
190 | if (flags & IO_WORKER_F_FREE) |
191 | hlist_nulls_del_rcu(&worker->nulls_node); | |
e61df66c | 192 | list_del_rcu(&worker->all_list); |
c5def4ab | 193 | acct->nr_workers--; |
95da8465 | 194 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 | 195 | |
364b05fd | 196 | kfree_rcu(worker, rcu); |
fb3a1f6c JA |
197 | if (atomic_dec_and_test(&wqe->wq->worker_refs)) |
198 | complete(&wqe->wq->worker_done); | |
46fe18b1 | 199 | do_exit(0); |
771b53d0 JA |
200 | } |
201 | ||
c5def4ab JA |
202 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
203 | __must_hold(wqe->lock) | |
204 | { | |
6206f0e1 JA |
205 | if (!wq_list_empty(&wqe->work_list) && |
206 | !(wqe->flags & IO_WQE_FLAG_STALLED)) | |
c5def4ab JA |
207 | return true; |
208 | return false; | |
209 | } | |
210 | ||
211 | /* | |
212 | * Check head of free list for an available worker. If one isn't available, | |
213 | * caller must wake up the wq manager to create one. | |
214 | */ | |
215 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) | |
216 | __must_hold(RCU) | |
217 | { | |
218 | struct hlist_nulls_node *n; | |
219 | struct io_worker *worker; | |
220 | ||
021d1cdd | 221 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list)); |
c5def4ab JA |
222 | if (is_a_nulls(n)) |
223 | return false; | |
224 | ||
225 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); | |
226 | if (io_worker_get(worker)) { | |
506d95ff | 227 | wake_up_process(worker->task); |
c5def4ab JA |
228 | io_worker_release(worker); |
229 | return true; | |
230 | } | |
231 | ||
232 | return false; | |
233 | } | |
234 | ||
235 | /* | |
236 | * We need a worker. If we find a free one, we're good. If not, and we're | |
237 | * below the max number of workers, wake up the manager to create one. | |
238 | */ | |
239 | static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) | |
240 | { | |
241 | bool ret; | |
242 | ||
243 | /* | |
244 | * Most likely an attempt to queue unbounded work on an io_wq that | |
245 | * wasn't setup with any unbounded workers. | |
246 | */ | |
247 | WARN_ON_ONCE(!acct->max_workers); | |
248 | ||
249 | rcu_read_lock(); | |
250 | ret = io_wqe_activate_free_worker(wqe); | |
251 | rcu_read_unlock(); | |
252 | ||
253 | if (!ret && acct->nr_workers < acct->max_workers) | |
254 | wake_up_process(wqe->wq->manager); | |
255 | } | |
256 | ||
958234d5 | 257 | static void io_wqe_inc_running(struct io_worker *worker) |
c5def4ab | 258 | { |
958234d5 | 259 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
c5def4ab JA |
260 | |
261 | atomic_inc(&acct->nr_running); | |
262 | } | |
263 | ||
958234d5 | 264 | static void io_wqe_dec_running(struct io_worker *worker) |
c5def4ab JA |
265 | __must_hold(wqe->lock) |
266 | { | |
958234d5 JA |
267 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
268 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab JA |
269 | |
270 | if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) | |
271 | io_wqe_wake_worker(wqe, acct); | |
272 | } | |
273 | ||
771b53d0 JA |
274 | /* |
275 | * Worker will start processing some work. Move it to the busy list, if | |
276 | * it's currently on the freelist | |
277 | */ | |
278 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
279 | struct io_wq_work *work) | |
280 | __must_hold(wqe->lock) | |
281 | { | |
c5def4ab JA |
282 | bool worker_bound, work_bound; |
283 | ||
771b53d0 JA |
284 | if (worker->flags & IO_WORKER_F_FREE) { |
285 | worker->flags &= ~IO_WORKER_F_FREE; | |
286 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 287 | } |
c5def4ab JA |
288 | |
289 | /* | |
290 | * If worker is moving from bound to unbound (or vice versa), then | |
291 | * ensure we update the running accounting. | |
292 | */ | |
b2e9c7d6 DC |
293 | worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0; |
294 | work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0; | |
295 | if (worker_bound != work_bound) { | |
958234d5 | 296 | io_wqe_dec_running(worker); |
c5def4ab JA |
297 | if (work_bound) { |
298 | worker->flags |= IO_WORKER_F_BOUND; | |
299 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--; | |
300 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++; | |
c5def4ab JA |
301 | } else { |
302 | worker->flags &= ~IO_WORKER_F_BOUND; | |
303 | wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++; | |
304 | wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--; | |
c5def4ab | 305 | } |
958234d5 | 306 | io_wqe_inc_running(worker); |
c5def4ab | 307 | } |
771b53d0 JA |
308 | } |
309 | ||
310 | /* | |
311 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
312 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
313 | * the lock in that case and return success. Since the caller has to | |
314 | * retry the loop in that case (we changed task state), we don't regrab | |
315 | * the lock if we return success. | |
316 | */ | |
c6d77d92 | 317 | static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) |
771b53d0 JA |
318 | __must_hold(wqe->lock) |
319 | { | |
320 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
321 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 322 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 | 323 | } |
771b53d0 JA |
324 | } |
325 | ||
60cf46ae PB |
326 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
327 | { | |
328 | return work->flags >> IO_WQ_HASH_SHIFT; | |
329 | } | |
330 | ||
e941894e JA |
331 | static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash) |
332 | { | |
333 | struct io_wq *wq = wqe->wq; | |
334 | ||
335 | spin_lock(&wq->hash->wait.lock); | |
336 | if (list_empty(&wqe->wait.entry)) { | |
337 | __add_wait_queue(&wq->hash->wait, &wqe->wait); | |
338 | if (!test_bit(hash, &wq->hash->map)) { | |
339 | __set_current_state(TASK_RUNNING); | |
340 | list_del_init(&wqe->wait.entry); | |
341 | } | |
342 | } | |
343 | spin_unlock(&wq->hash->wait.lock); | |
344 | } | |
345 | ||
60cf46ae | 346 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe) |
771b53d0 JA |
347 | __must_hold(wqe->lock) |
348 | { | |
6206f0e1 | 349 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 350 | struct io_wq_work *work, *tail; |
e941894e | 351 | unsigned int stall_hash = -1U; |
771b53d0 | 352 | |
6206f0e1 | 353 | wq_list_for_each(node, prev, &wqe->work_list) { |
e941894e JA |
354 | unsigned int hash; |
355 | ||
6206f0e1 JA |
356 | work = container_of(node, struct io_wq_work, list); |
357 | ||
771b53d0 | 358 | /* not hashed, can run anytime */ |
8766dd51 | 359 | if (!io_wq_is_hashed(work)) { |
86f3cd1b | 360 | wq_list_del(&wqe->work_list, node, prev); |
771b53d0 JA |
361 | return work; |
362 | } | |
363 | ||
60cf46ae | 364 | hash = io_get_work_hash(work); |
e941894e JA |
365 | /* all items with this hash lie in [work, tail] */ |
366 | tail = wqe->hash_tail[hash]; | |
367 | ||
368 | /* hashed, can run if not already running */ | |
369 | if (!test_and_set_bit(hash, &wqe->wq->hash->map)) { | |
86f3cd1b PB |
370 | wqe->hash_tail[hash] = NULL; |
371 | wq_list_cut(&wqe->work_list, &tail->list, prev); | |
771b53d0 JA |
372 | return work; |
373 | } | |
e941894e JA |
374 | if (stall_hash == -1U) |
375 | stall_hash = hash; | |
376 | /* fast forward to a next hash, for-each will fix up @prev */ | |
377 | node = &tail->list; | |
378 | } | |
379 | ||
380 | if (stall_hash != -1U) { | |
381 | raw_spin_unlock(&wqe->lock); | |
382 | io_wait_on_hash(wqe, stall_hash); | |
383 | raw_spin_lock(&wqe->lock); | |
771b53d0 JA |
384 | } |
385 | ||
386 | return NULL; | |
387 | } | |
388 | ||
00ddff43 | 389 | static bool io_flush_signals(void) |
cccf0ee8 | 390 | { |
3bfe6106 | 391 | if (unlikely(test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))) { |
00ddff43 | 392 | __set_current_state(TASK_RUNNING); |
3bfe6106 JA |
393 | if (current->task_works) |
394 | task_work_run(); | |
395 | clear_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL); | |
00ddff43 | 396 | return true; |
cccf0ee8 | 397 | } |
00ddff43 | 398 | return false; |
dc026a73 PB |
399 | } |
400 | ||
401 | static void io_assign_current_work(struct io_worker *worker, | |
402 | struct io_wq_work *work) | |
403 | { | |
d78298e7 | 404 | if (work) { |
3bfe6106 | 405 | io_flush_signals(); |
d78298e7 PB |
406 | cond_resched(); |
407 | } | |
dc026a73 PB |
408 | |
409 | spin_lock_irq(&worker->lock); | |
410 | worker->cur_work = work; | |
411 | spin_unlock_irq(&worker->lock); | |
412 | } | |
413 | ||
60cf46ae PB |
414 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
415 | ||
771b53d0 JA |
416 | static void io_worker_handle_work(struct io_worker *worker) |
417 | __releases(wqe->lock) | |
418 | { | |
771b53d0 JA |
419 | struct io_wqe *wqe = worker->wqe; |
420 | struct io_wq *wq = wqe->wq; | |
421 | ||
422 | do { | |
86f3cd1b | 423 | struct io_wq_work *work; |
f462fd36 | 424 | get_next: |
771b53d0 JA |
425 | /* |
426 | * If we got some work, mark us as busy. If we didn't, but | |
427 | * the list isn't empty, it means we stalled on hashed work. | |
428 | * Mark us stalled so we don't keep looking for work when we | |
429 | * can't make progress, any work completion or insertion will | |
430 | * clear the stalled flag. | |
431 | */ | |
60cf46ae | 432 | work = io_get_next_work(wqe); |
771b53d0 JA |
433 | if (work) |
434 | __io_worker_busy(wqe, worker, work); | |
6206f0e1 | 435 | else if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
436 | wqe->flags |= IO_WQE_FLAG_STALLED; |
437 | ||
95da8465 | 438 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
439 | if (!work) |
440 | break; | |
58e39319 | 441 | io_assign_current_work(worker, work); |
e941894e | 442 | __set_current_state(TASK_RUNNING); |
36c2f922 | 443 | |
dc026a73 PB |
444 | /* handle a whole dependent link */ |
445 | do { | |
5280f7e5 | 446 | struct io_wq_work *next_hashed, *linked; |
b089ed39 | 447 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 448 | |
86f3cd1b | 449 | next_hashed = wq_next_work(work); |
5280f7e5 PB |
450 | wq->do_work(work); |
451 | io_assign_current_work(worker, NULL); | |
dc026a73 | 452 | |
5280f7e5 | 453 | linked = wq->free_work(work); |
86f3cd1b PB |
454 | work = next_hashed; |
455 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
456 | work = linked; | |
457 | linked = NULL; | |
458 | } | |
459 | io_assign_current_work(worker, work); | |
86f3cd1b PB |
460 | if (linked) |
461 | io_wqe_enqueue(wqe, linked); | |
462 | ||
463 | if (hash != -1U && !next_hashed) { | |
e941894e JA |
464 | clear_bit(hash, &wq->hash->map); |
465 | if (wq_has_sleeper(&wq->hash->wait)) | |
466 | wake_up(&wq->hash->wait); | |
95da8465 | 467 | raw_spin_lock_irq(&wqe->lock); |
dc026a73 | 468 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
f462fd36 PB |
469 | /* skip unnecessary unlock-lock wqe->lock */ |
470 | if (!work) | |
471 | goto get_next; | |
95da8465 | 472 | raw_spin_unlock_irq(&wqe->lock); |
7d723065 | 473 | } |
58e39319 | 474 | } while (work); |
7d723065 | 475 | |
95da8465 | 476 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 JA |
477 | } while (1); |
478 | } | |
479 | ||
771b53d0 JA |
480 | static int io_wqe_worker(void *data) |
481 | { | |
482 | struct io_worker *worker = data; | |
483 | struct io_wqe *wqe = worker->wqe; | |
484 | struct io_wq *wq = wqe->wq; | |
46fe18b1 | 485 | char buf[TASK_COMM_LEN]; |
771b53d0 | 486 | |
46fe18b1 JA |
487 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); |
488 | io_wqe_inc_running(worker); | |
489 | ||
490 | sprintf(buf, "iou-wrk-%d", wq->task_pid); | |
491 | set_task_comm(current, buf); | |
771b53d0 JA |
492 | |
493 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
16efa4fc JA |
494 | long ret; |
495 | ||
506d95ff | 496 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 497 | loop: |
95da8465 | 498 | raw_spin_lock_irq(&wqe->lock); |
771b53d0 | 499 | if (io_wqe_run_queue(wqe)) { |
771b53d0 | 500 | io_worker_handle_work(worker); |
e995d512 | 501 | goto loop; |
771b53d0 | 502 | } |
c6d77d92 | 503 | __io_worker_idle(wqe, worker); |
95da8465 | 504 | raw_spin_unlock_irq(&wqe->lock); |
00ddff43 JA |
505 | if (io_flush_signals()) |
506 | continue; | |
16efa4fc JA |
507 | ret = schedule_timeout(WORKER_IDLE_TIMEOUT); |
508 | if (try_to_freeze() || ret) | |
771b53d0 | 509 | continue; |
3bfe6106 JA |
510 | if (fatal_signal_pending(current)) |
511 | break; | |
771b53d0 JA |
512 | /* timed out, exit unless we're the fixed worker */ |
513 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || | |
514 | !(worker->flags & IO_WORKER_F_FIXED)) | |
515 | break; | |
516 | } | |
517 | ||
771b53d0 | 518 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
95da8465 | 519 | raw_spin_lock_irq(&wqe->lock); |
6206f0e1 | 520 | if (!wq_list_empty(&wqe->work_list)) |
771b53d0 JA |
521 | io_worker_handle_work(worker); |
522 | else | |
95da8465 | 523 | raw_spin_unlock_irq(&wqe->lock); |
771b53d0 JA |
524 | } |
525 | ||
526 | io_worker_exit(worker); | |
527 | return 0; | |
528 | } | |
529 | ||
771b53d0 JA |
530 | /* |
531 | * Called when a worker is scheduled in. Mark us as currently running. | |
532 | */ | |
533 | void io_wq_worker_running(struct task_struct *tsk) | |
534 | { | |
3bfe6106 | 535 | struct io_worker *worker = tsk->pf_io_worker; |
771b53d0 | 536 | |
3bfe6106 JA |
537 | if (!worker) |
538 | return; | |
771b53d0 JA |
539 | if (!(worker->flags & IO_WORKER_F_UP)) |
540 | return; | |
541 | if (worker->flags & IO_WORKER_F_RUNNING) | |
542 | return; | |
543 | worker->flags |= IO_WORKER_F_RUNNING; | |
958234d5 | 544 | io_wqe_inc_running(worker); |
771b53d0 JA |
545 | } |
546 | ||
547 | /* | |
548 | * Called when worker is going to sleep. If there are no workers currently | |
549 | * running and we have work pending, wake up a free one or have the manager | |
550 | * set one up. | |
551 | */ | |
552 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
553 | { | |
3bfe6106 | 554 | struct io_worker *worker = tsk->pf_io_worker; |
771b53d0 | 555 | |
3bfe6106 JA |
556 | if (!worker) |
557 | return; | |
771b53d0 JA |
558 | if (!(worker->flags & IO_WORKER_F_UP)) |
559 | return; | |
560 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
561 | return; | |
562 | ||
563 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
564 | ||
3bfe6106 | 565 | raw_spin_lock_irq(&worker->wqe->lock); |
958234d5 | 566 | io_wqe_dec_running(worker); |
3bfe6106 | 567 | raw_spin_unlock_irq(&worker->wqe->lock); |
771b53d0 JA |
568 | } |
569 | ||
3bfe6106 JA |
570 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) |
571 | { | |
46fe18b1 | 572 | struct io_wqe_acct *acct = &wqe->acct[index]; |
3bfe6106 | 573 | struct io_worker *worker; |
46fe18b1 | 574 | struct task_struct *tsk; |
3bfe6106 | 575 | |
8b3e78b5 JA |
576 | __set_current_state(TASK_RUNNING); |
577 | ||
3bfe6106 JA |
578 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
579 | if (!worker) | |
580 | return false; | |
581 | ||
582 | refcount_set(&worker->ref, 1); | |
583 | worker->nulls_node.pprev = NULL; | |
584 | worker->wqe = wqe; | |
585 | spin_lock_init(&worker->lock); | |
eb2de941 | 586 | init_completion(&worker->ref_done); |
3bfe6106 | 587 | |
fb3a1f6c | 588 | atomic_inc(&wq->worker_refs); |
8b3e78b5 | 589 | |
46fe18b1 JA |
590 | tsk = create_io_thread(io_wqe_worker, worker, wqe->node); |
591 | if (IS_ERR(tsk)) { | |
fb3a1f6c JA |
592 | if (atomic_dec_and_test(&wq->worker_refs)) |
593 | complete(&wq->worker_done); | |
3bfe6106 JA |
594 | kfree(worker); |
595 | return false; | |
596 | } | |
46fe18b1 JA |
597 | |
598 | tsk->pf_io_worker = worker; | |
599 | worker->task = tsk; | |
600 | set_cpus_allowed_ptr(tsk, cpumask_of_node(wqe->node)); | |
e22bc9b4 | 601 | tsk->flags |= PF_NO_SETAFFINITY; |
46fe18b1 JA |
602 | |
603 | raw_spin_lock_irq(&wqe->lock); | |
604 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); | |
605 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); | |
606 | worker->flags |= IO_WORKER_F_FREE; | |
607 | if (index == IO_WQ_ACCT_BOUND) | |
608 | worker->flags |= IO_WORKER_F_BOUND; | |
609 | if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND)) | |
610 | worker->flags |= IO_WORKER_F_FIXED; | |
611 | acct->nr_workers++; | |
612 | raw_spin_unlock_irq(&wqe->lock); | |
613 | wake_up_new_task(tsk); | |
b60fda60 | 614 | return true; |
771b53d0 JA |
615 | } |
616 | ||
c5def4ab | 617 | static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index) |
771b53d0 JA |
618 | __must_hold(wqe->lock) |
619 | { | |
c5def4ab | 620 | struct io_wqe_acct *acct = &wqe->acct[index]; |
771b53d0 | 621 | |
613eeb60 JA |
622 | if (acct->nr_workers && test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state)) |
623 | return false; | |
c5def4ab | 624 | /* if we have available workers or no work, no need */ |
021d1cdd | 625 | if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe)) |
c5def4ab JA |
626 | return false; |
627 | return acct->nr_workers < acct->max_workers; | |
771b53d0 JA |
628 | } |
629 | ||
c4068bf8 HD |
630 | /* |
631 | * Iterate the passed in list and call the specific function for each | |
632 | * worker that isn't exiting | |
633 | */ | |
634 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
635 | bool (*func)(struct io_worker *, void *), | |
636 | void *data) | |
637 | { | |
638 | struct io_worker *worker; | |
639 | bool ret = false; | |
640 | ||
641 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { | |
642 | if (io_worker_get(worker)) { | |
643 | /* no task if node is/was offline */ | |
644 | if (worker->task) | |
645 | ret = func(worker, data); | |
646 | io_worker_release(worker); | |
647 | if (ret) | |
648 | break; | |
649 | } | |
650 | } | |
651 | ||
652 | return ret; | |
653 | } | |
654 | ||
655 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
656 | { | |
46fe18b1 | 657 | set_notify_signal(worker->task); |
c4068bf8 HD |
658 | wake_up_process(worker->task); |
659 | return false; | |
660 | } | |
661 | ||
8b3e78b5 JA |
662 | static void io_wq_check_workers(struct io_wq *wq) |
663 | { | |
664 | int node; | |
665 | ||
666 | for_each_node(node) { | |
667 | struct io_wqe *wqe = wq->wqes[node]; | |
668 | bool fork_worker[2] = { false, false }; | |
669 | ||
670 | if (!node_online(node)) | |
671 | continue; | |
672 | ||
673 | raw_spin_lock_irq(&wqe->lock); | |
674 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND)) | |
675 | fork_worker[IO_WQ_ACCT_BOUND] = true; | |
676 | if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND)) | |
677 | fork_worker[IO_WQ_ACCT_UNBOUND] = true; | |
678 | raw_spin_unlock_irq(&wqe->lock); | |
679 | if (fork_worker[IO_WQ_ACCT_BOUND]) | |
680 | create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND); | |
681 | if (fork_worker[IO_WQ_ACCT_UNBOUND]) | |
682 | create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND); | |
683 | } | |
684 | } | |
685 | ||
f0127254 JA |
686 | static bool io_wq_work_match_all(struct io_wq_work *work, void *data) |
687 | { | |
688 | return true; | |
689 | } | |
690 | ||
691 | static void io_wq_cancel_pending(struct io_wq *wq) | |
692 | { | |
693 | struct io_cb_cancel_data match = { | |
694 | .fn = io_wq_work_match_all, | |
695 | .cancel_all = true, | |
696 | }; | |
697 | int node; | |
698 | ||
699 | for_each_node(node) | |
700 | io_wqe_cancel_pending_work(wq->wqes[node], &match); | |
701 | } | |
702 | ||
771b53d0 JA |
703 | /* |
704 | * Manager thread. Tasked with creating new workers, if we need them. | |
705 | */ | |
706 | static int io_wq_manager(void *data) | |
707 | { | |
708 | struct io_wq *wq = data; | |
3bfe6106 | 709 | char buf[TASK_COMM_LEN]; |
fb3a1f6c | 710 | int node; |
771b53d0 | 711 | |
3bfe6106 JA |
712 | sprintf(buf, "iou-mgr-%d", wq->task_pid); |
713 | set_task_comm(current, buf); | |
b60fda60 | 714 | |
8b3e78b5 | 715 | do { |
771b53d0 | 716 | set_current_state(TASK_INTERRUPTIBLE); |
8b3e78b5 | 717 | io_wq_check_workers(wq); |
771b53d0 | 718 | schedule_timeout(HZ); |
16efa4fc | 719 | try_to_freeze(); |
3bfe6106 JA |
720 | if (fatal_signal_pending(current)) |
721 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
8b3e78b5 JA |
722 | } while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)); |
723 | ||
724 | io_wq_check_workers(wq); | |
fb3a1f6c JA |
725 | |
726 | rcu_read_lock(); | |
727 | for_each_node(node) | |
728 | io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL); | |
729 | rcu_read_unlock(); | |
730 | ||
886d0137 JA |
731 | if (atomic_dec_and_test(&wq->worker_refs)) |
732 | complete(&wq->worker_done); | |
733 | wait_for_completion(&wq->worker_done); | |
f0127254 | 734 | |
09ca6c40 JA |
735 | spin_lock_irq(&wq->hash->wait.lock); |
736 | for_each_node(node) | |
737 | list_del_init(&wq->wqes[node]->wait.entry); | |
738 | spin_unlock_irq(&wq->hash->wait.lock); | |
739 | ||
f0127254 | 740 | io_wq_cancel_pending(wq); |
d364d9e5 | 741 | complete(&wq->exited); |
3bfe6106 | 742 | do_exit(0); |
771b53d0 JA |
743 | } |
744 | ||
e9fd9396 | 745 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 746 | { |
e9fd9396 PB |
747 | struct io_wq *wq = wqe->wq; |
748 | ||
fc04c39b | 749 | do { |
fc04c39b | 750 | work->flags |= IO_WQ_WORK_CANCEL; |
5280f7e5 PB |
751 | wq->do_work(work); |
752 | work = wq->free_work(work); | |
fc04c39b PB |
753 | } while (work); |
754 | } | |
755 | ||
86f3cd1b PB |
756 | static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work) |
757 | { | |
758 | unsigned int hash; | |
759 | struct io_wq_work *tail; | |
760 | ||
761 | if (!io_wq_is_hashed(work)) { | |
762 | append: | |
763 | wq_list_add_tail(&work->list, &wqe->work_list); | |
764 | return; | |
765 | } | |
766 | ||
767 | hash = io_get_work_hash(work); | |
768 | tail = wqe->hash_tail[hash]; | |
769 | wqe->hash_tail[hash] = work; | |
770 | if (!tail) | |
771 | goto append; | |
772 | ||
773 | wq_list_add_after(&work->list, &tail->list, &wqe->work_list); | |
774 | } | |
775 | ||
4fb6ac32 JA |
776 | static int io_wq_fork_manager(struct io_wq *wq) |
777 | { | |
46fe18b1 | 778 | struct task_struct *tsk; |
4fb6ac32 JA |
779 | |
780 | if (wq->manager) | |
781 | return 0; | |
782 | ||
678eeba4 PB |
783 | WARN_ON_ONCE(test_bit(IO_WQ_BIT_EXIT, &wq->state)); |
784 | ||
886d0137 JA |
785 | init_completion(&wq->worker_done); |
786 | atomic_set(&wq->worker_refs, 1); | |
46fe18b1 JA |
787 | tsk = create_io_thread(io_wq_manager, wq, NUMA_NO_NODE); |
788 | if (!IS_ERR(tsk)) { | |
789 | wq->manager = get_task_struct(tsk); | |
790 | wake_up_new_task(tsk); | |
4fb6ac32 JA |
791 | return 0; |
792 | } | |
793 | ||
886d0137 JA |
794 | if (atomic_dec_and_test(&wq->worker_refs)) |
795 | complete(&wq->worker_done); | |
796 | ||
46fe18b1 | 797 | return PTR_ERR(tsk); |
4fb6ac32 JA |
798 | } |
799 | ||
771b53d0 JA |
800 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
801 | { | |
c5def4ab | 802 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
895e2ca0 | 803 | int work_flags; |
771b53d0 JA |
804 | unsigned long flags; |
805 | ||
4fb6ac32 | 806 | /* Can only happen if manager creation fails after exec */ |
f0127254 JA |
807 | if (io_wq_fork_manager(wqe->wq) || |
808 | test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state)) { | |
70e35125 | 809 | io_run_cancel(work, wqe); |
4fb6ac32 JA |
810 | return; |
811 | } | |
812 | ||
895e2ca0 | 813 | work_flags = work->flags; |
95da8465 | 814 | raw_spin_lock_irqsave(&wqe->lock, flags); |
86f3cd1b | 815 | io_wqe_insert_work(wqe, work); |
771b53d0 | 816 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
95da8465 | 817 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
771b53d0 | 818 | |
895e2ca0 JA |
819 | if ((work_flags & IO_WQ_WORK_CONCURRENT) || |
820 | !atomic_read(&acct->nr_running)) | |
c5def4ab | 821 | io_wqe_wake_worker(wqe, acct); |
771b53d0 JA |
822 | } |
823 | ||
824 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
825 | { | |
826 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
827 | ||
828 | io_wqe_enqueue(wqe, work); | |
829 | } | |
830 | ||
831 | /* | |
8766dd51 PB |
832 | * Work items that hash to the same value will not be done in parallel. |
833 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 834 | */ |
8766dd51 | 835 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 836 | { |
8766dd51 | 837 | unsigned int bit; |
771b53d0 JA |
838 | |
839 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
840 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
841 | } |
842 | ||
2293b419 | 843 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 844 | { |
2293b419 | 845 | struct io_cb_cancel_data *match = data; |
6f72653e | 846 | unsigned long flags; |
62755e35 JA |
847 | |
848 | /* | |
849 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 850 | * may dereference the passed in work. |
62755e35 | 851 | */ |
36c2f922 | 852 | spin_lock_irqsave(&worker->lock, flags); |
62755e35 | 853 | if (worker->cur_work && |
2293b419 | 854 | match->fn(worker->cur_work, match->data)) { |
3bfe6106 | 855 | set_notify_signal(worker->task); |
4f26bda1 | 856 | match->nr_running++; |
771b53d0 | 857 | } |
36c2f922 | 858 | spin_unlock_irqrestore(&worker->lock, flags); |
771b53d0 | 859 | |
4f26bda1 | 860 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
861 | } |
862 | ||
204361a7 PB |
863 | static inline void io_wqe_remove_pending(struct io_wqe *wqe, |
864 | struct io_wq_work *work, | |
865 | struct io_wq_work_node *prev) | |
866 | { | |
867 | unsigned int hash = io_get_work_hash(work); | |
868 | struct io_wq_work *prev_work = NULL; | |
869 | ||
870 | if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) { | |
871 | if (prev) | |
872 | prev_work = container_of(prev, struct io_wq_work, list); | |
873 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
874 | wqe->hash_tail[hash] = prev_work; | |
875 | else | |
876 | wqe->hash_tail[hash] = NULL; | |
877 | } | |
878 | wq_list_del(&wqe->work_list, &work->list, prev); | |
879 | } | |
880 | ||
4f26bda1 | 881 | static void io_wqe_cancel_pending_work(struct io_wqe *wqe, |
f4c2665e | 882 | struct io_cb_cancel_data *match) |
771b53d0 | 883 | { |
6206f0e1 | 884 | struct io_wq_work_node *node, *prev; |
771b53d0 | 885 | struct io_wq_work *work; |
6f72653e | 886 | unsigned long flags; |
771b53d0 | 887 | |
4f26bda1 | 888 | retry: |
95da8465 | 889 | raw_spin_lock_irqsave(&wqe->lock, flags); |
6206f0e1 JA |
890 | wq_list_for_each(node, prev, &wqe->work_list) { |
891 | work = container_of(node, struct io_wq_work, list); | |
4f26bda1 PB |
892 | if (!match->fn(work, match->data)) |
893 | continue; | |
204361a7 | 894 | io_wqe_remove_pending(wqe, work, prev); |
95da8465 | 895 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
4f26bda1 PB |
896 | io_run_cancel(work, wqe); |
897 | match->nr_pending++; | |
898 | if (!match->cancel_all) | |
899 | return; | |
900 | ||
901 | /* not safe to continue after unlock */ | |
902 | goto retry; | |
771b53d0 | 903 | } |
95da8465 | 904 | raw_spin_unlock_irqrestore(&wqe->lock, flags); |
f4c2665e PB |
905 | } |
906 | ||
4f26bda1 | 907 | static void io_wqe_cancel_running_work(struct io_wqe *wqe, |
f4c2665e PB |
908 | struct io_cb_cancel_data *match) |
909 | { | |
771b53d0 | 910 | rcu_read_lock(); |
4f26bda1 | 911 | io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 | 912 | rcu_read_unlock(); |
771b53d0 JA |
913 | } |
914 | ||
2293b419 | 915 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 916 | void *data, bool cancel_all) |
771b53d0 | 917 | { |
2293b419 | 918 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
919 | .fn = cancel, |
920 | .data = data, | |
921 | .cancel_all = cancel_all, | |
00bcda13 | 922 | }; |
3fc50ab5 | 923 | int node; |
771b53d0 | 924 | |
f4c2665e PB |
925 | /* |
926 | * First check pending list, if we're lucky we can just remove it | |
927 | * from there. CANCEL_OK means that the work is returned as-new, | |
928 | * no completion will be posted for it. | |
929 | */ | |
3fc50ab5 JH |
930 | for_each_node(node) { |
931 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 932 | |
4f26bda1 PB |
933 | io_wqe_cancel_pending_work(wqe, &match); |
934 | if (match.nr_pending && !match.cancel_all) | |
f4c2665e | 935 | return IO_WQ_CANCEL_OK; |
771b53d0 JA |
936 | } |
937 | ||
f4c2665e PB |
938 | /* |
939 | * Now check if a free (going busy) or busy worker has the work | |
940 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
941 | * as an indication that we attempt to signal cancellation. The | |
942 | * completion will run normally in this case. | |
943 | */ | |
944 | for_each_node(node) { | |
945 | struct io_wqe *wqe = wq->wqes[node]; | |
946 | ||
4f26bda1 PB |
947 | io_wqe_cancel_running_work(wqe, &match); |
948 | if (match.nr_running && !match.cancel_all) | |
f4c2665e PB |
949 | return IO_WQ_CANCEL_RUNNING; |
950 | } | |
951 | ||
4f26bda1 PB |
952 | if (match.nr_running) |
953 | return IO_WQ_CANCEL_RUNNING; | |
954 | if (match.nr_pending) | |
955 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 956 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
957 | } |
958 | ||
e941894e JA |
959 | static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode, |
960 | int sync, void *key) | |
961 | { | |
962 | struct io_wqe *wqe = container_of(wait, struct io_wqe, wait); | |
963 | int ret; | |
964 | ||
965 | list_del_init(&wait->entry); | |
966 | ||
967 | rcu_read_lock(); | |
968 | ret = io_wqe_activate_free_worker(wqe); | |
969 | rcu_read_unlock(); | |
970 | ||
971 | if (!ret) | |
972 | wake_up_process(wqe->wq->manager); | |
973 | ||
974 | return 1; | |
975 | } | |
976 | ||
576a347b | 977 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 978 | { |
3fc50ab5 | 979 | int ret = -ENOMEM, node; |
771b53d0 JA |
980 | struct io_wq *wq; |
981 | ||
f5fa38c5 | 982 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 PB |
983 | return ERR_PTR(-EINVAL); |
984 | ||
ad6e005c | 985 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
771b53d0 JA |
986 | if (!wq) |
987 | return ERR_PTR(-ENOMEM); | |
988 | ||
3fc50ab5 | 989 | wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL); |
43c01fbe JA |
990 | if (!wq->wqes) |
991 | goto err_wq; | |
992 | ||
993 | ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node); | |
994 | if (ret) | |
995 | goto err_wqes; | |
771b53d0 | 996 | |
e941894e JA |
997 | refcount_inc(&data->hash->refs); |
998 | wq->hash = data->hash; | |
e9fd9396 | 999 | wq->free_work = data->free_work; |
f5fa38c5 | 1000 | wq->do_work = data->do_work; |
7d723065 | 1001 | |
43c01fbe | 1002 | ret = -ENOMEM; |
3fc50ab5 | 1003 | for_each_node(node) { |
771b53d0 | 1004 | struct io_wqe *wqe; |
7563439a | 1005 | int alloc_node = node; |
771b53d0 | 1006 | |
7563439a JA |
1007 | if (!node_online(alloc_node)) |
1008 | alloc_node = NUMA_NO_NODE; | |
1009 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1010 | if (!wqe) |
3fc50ab5 JH |
1011 | goto err; |
1012 | wq->wqes[node] = wqe; | |
7563439a | 1013 | wqe->node = alloc_node; |
c5def4ab JA |
1014 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1015 | atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0); | |
728f13e7 | 1016 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
c5def4ab | 1017 | task_rlimit(current, RLIMIT_NPROC); |
c5def4ab | 1018 | atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0); |
e941894e JA |
1019 | wqe->wait.func = io_wqe_hash_wake; |
1020 | INIT_LIST_HEAD(&wqe->wait.entry); | |
771b53d0 | 1021 | wqe->wq = wq; |
95da8465 | 1022 | raw_spin_lock_init(&wqe->lock); |
6206f0e1 | 1023 | INIT_WQ_LIST(&wqe->work_list); |
021d1cdd | 1024 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1025 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1026 | } |
1027 | ||
3bfe6106 | 1028 | wq->task_pid = current->pid; |
d364d9e5 | 1029 | init_completion(&wq->exited); |
3bfe6106 | 1030 | refcount_set(&wq->refs, 1); |
771b53d0 | 1031 | |
4fb6ac32 JA |
1032 | ret = io_wq_fork_manager(wq); |
1033 | if (!ret) | |
771b53d0 | 1034 | return wq; |
b60fda60 | 1035 | err: |
dc7bbc9e | 1036 | io_wq_put_hash(data->hash); |
43c01fbe | 1037 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
3fc50ab5 JH |
1038 | for_each_node(node) |
1039 | kfree(wq->wqes[node]); | |
43c01fbe | 1040 | err_wqes: |
b60fda60 | 1041 | kfree(wq->wqes); |
43c01fbe | 1042 | err_wq: |
b60fda60 | 1043 | kfree(wq); |
771b53d0 JA |
1044 | return ERR_PTR(ret); |
1045 | } | |
1046 | ||
afcc4015 JA |
1047 | static void io_wq_destroy_manager(struct io_wq *wq) |
1048 | { | |
1049 | if (wq->manager) { | |
1050 | wake_up_process(wq->manager); | |
1051 | wait_for_completion(&wq->exited); | |
1052 | put_task_struct(wq->manager); | |
1053 | wq->manager = NULL; | |
1054 | } | |
1055 | } | |
1056 | ||
4fb6ac32 | 1057 | static void io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1058 | { |
3fc50ab5 | 1059 | int node; |
771b53d0 | 1060 | |
43c01fbe JA |
1061 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
1062 | ||
b60fda60 | 1063 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
afcc4015 | 1064 | io_wq_destroy_manager(wq); |
771b53d0 | 1065 | |
e941894e JA |
1066 | for_each_node(node) { |
1067 | struct io_wqe *wqe = wq->wqes[node]; | |
f0127254 | 1068 | WARN_ON_ONCE(!wq_list_empty(&wqe->work_list)); |
e941894e JA |
1069 | kfree(wqe); |
1070 | } | |
e941894e | 1071 | io_wq_put_hash(wq->hash); |
771b53d0 JA |
1072 | kfree(wq->wqes); |
1073 | kfree(wq); | |
4fb6ac32 JA |
1074 | } |
1075 | ||
1076 | void io_wq_put(struct io_wq *wq) | |
1077 | { | |
1078 | if (refcount_dec_and_test(&wq->refs)) | |
1079 | io_wq_destroy(wq); | |
771b53d0 | 1080 | } |
848f7e18 | 1081 | |
afcc4015 JA |
1082 | void io_wq_put_and_exit(struct io_wq *wq) |
1083 | { | |
1084 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
1085 | io_wq_destroy_manager(wq); | |
1086 | io_wq_put(wq); | |
1087 | } | |
1088 | ||
43c01fbe JA |
1089 | static bool io_wq_worker_affinity(struct io_worker *worker, void *data) |
1090 | { | |
1091 | struct task_struct *task = worker->task; | |
1092 | struct rq_flags rf; | |
1093 | struct rq *rq; | |
1094 | ||
1095 | rq = task_rq_lock(task, &rf); | |
1096 | do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node)); | |
1097 | task->flags |= PF_NO_SETAFFINITY; | |
1098 | task_rq_unlock(rq, task, &rf); | |
1099 | return false; | |
1100 | } | |
1101 | ||
1102 | static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node) | |
1103 | { | |
1104 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1105 | int i; | |
1106 | ||
1107 | rcu_read_lock(); | |
1108 | for_each_node(i) | |
1109 | io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL); | |
1110 | rcu_read_unlock(); | |
1111 | return 0; | |
1112 | } | |
1113 | ||
1114 | static __init int io_wq_init(void) | |
1115 | { | |
1116 | int ret; | |
1117 | ||
1118 | ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online", | |
1119 | io_wq_cpu_online, NULL); | |
1120 | if (ret < 0) | |
1121 | return ret; | |
1122 | io_wq_online = ret; | |
1123 | return 0; | |
1124 | } | |
1125 | subsys_initcall(io_wq_init); |