]>
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> | |
771b53d0 JA |
12 | #include <linux/percpu.h> |
13 | #include <linux/slab.h> | |
771b53d0 | 14 | #include <linux/rculist_nulls.h> |
43c01fbe | 15 | #include <linux/cpu.h> |
3bfe6106 | 16 | #include <linux/tracehook.h> |
dd47c104 | 17 | #include <uapi/linux/io_uring.h> |
771b53d0 JA |
18 | |
19 | #include "io-wq.h" | |
20 | ||
21 | #define WORKER_IDLE_TIMEOUT (5 * HZ) | |
22 | ||
23 | enum { | |
24 | IO_WORKER_F_UP = 1, /* up and active */ | |
25 | IO_WORKER_F_RUNNING = 2, /* account as running */ | |
26 | IO_WORKER_F_FREE = 4, /* worker on free list */ | |
05c5f4ee | 27 | IO_WORKER_F_BOUND = 8, /* is doing bounded work */ |
771b53d0 JA |
28 | }; |
29 | ||
30 | enum { | |
31 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ | |
771b53d0 JA |
32 | }; |
33 | ||
34 | enum { | |
f95dc207 | 35 | IO_ACCT_STALLED_BIT = 0, /* stalled on hash */ |
771b53d0 JA |
36 | }; |
37 | ||
38 | /* | |
39 | * One for each thread in a wqe pool | |
40 | */ | |
41 | struct io_worker { | |
42 | refcount_t ref; | |
43 | unsigned flags; | |
44 | struct hlist_nulls_node nulls_node; | |
e61df66c | 45 | struct list_head all_list; |
771b53d0 | 46 | struct task_struct *task; |
771b53d0 | 47 | struct io_wqe *wqe; |
36c2f922 | 48 | |
771b53d0 | 49 | struct io_wq_work *cur_work; |
36c2f922 | 50 | spinlock_t lock; |
771b53d0 | 51 | |
eb2de941 JA |
52 | struct completion ref_done; |
53 | ||
d3e9f732 JA |
54 | unsigned long create_state; |
55 | struct callback_head create_work; | |
56 | int create_index; | |
57 | ||
3146cba9 JA |
58 | union { |
59 | struct rcu_head rcu; | |
60 | struct work_struct work; | |
61 | }; | |
771b53d0 JA |
62 | }; |
63 | ||
771b53d0 JA |
64 | #if BITS_PER_LONG == 64 |
65 | #define IO_WQ_HASH_ORDER 6 | |
66 | #else | |
67 | #define IO_WQ_HASH_ORDER 5 | |
68 | #endif | |
69 | ||
86f3cd1b PB |
70 | #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER) |
71 | ||
c5def4ab JA |
72 | struct io_wqe_acct { |
73 | unsigned nr_workers; | |
74 | unsigned max_workers; | |
685fe7fe | 75 | int index; |
c5def4ab | 76 | atomic_t nr_running; |
f95dc207 JA |
77 | struct io_wq_work_list work_list; |
78 | unsigned long flags; | |
c5def4ab JA |
79 | }; |
80 | ||
81 | enum { | |
82 | IO_WQ_ACCT_BOUND, | |
83 | IO_WQ_ACCT_UNBOUND, | |
f95dc207 | 84 | IO_WQ_ACCT_NR, |
c5def4ab JA |
85 | }; |
86 | ||
771b53d0 JA |
87 | /* |
88 | * Per-node worker thread pool | |
89 | */ | |
90 | struct io_wqe { | |
f95dc207 JA |
91 | raw_spinlock_t lock; |
92 | struct io_wqe_acct acct[2]; | |
771b53d0 JA |
93 | |
94 | int node; | |
771b53d0 | 95 | |
021d1cdd | 96 | struct hlist_nulls_head free_list; |
e61df66c | 97 | struct list_head all_list; |
771b53d0 | 98 | |
e941894e JA |
99 | struct wait_queue_entry wait; |
100 | ||
771b53d0 | 101 | struct io_wq *wq; |
86f3cd1b | 102 | struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS]; |
0e03496d JA |
103 | |
104 | cpumask_var_t cpu_mask; | |
771b53d0 JA |
105 | }; |
106 | ||
107 | /* | |
108 | * Per io_wq state | |
109 | */ | |
110 | struct io_wq { | |
771b53d0 | 111 | unsigned long state; |
771b53d0 | 112 | |
e9fd9396 | 113 | free_work_fn *free_work; |
f5fa38c5 | 114 | io_wq_work_fn *do_work; |
7d723065 | 115 | |
e941894e JA |
116 | struct io_wq_hash *hash; |
117 | ||
fb3a1f6c JA |
118 | atomic_t worker_refs; |
119 | struct completion worker_done; | |
120 | ||
43c01fbe | 121 | struct hlist_node cpuhp_node; |
3bfe6106 | 122 | |
685fe7fe | 123 | struct task_struct *task; |
c7f405d6 PB |
124 | |
125 | struct io_wqe *wqes[]; | |
771b53d0 JA |
126 | }; |
127 | ||
43c01fbe JA |
128 | static enum cpuhp_state io_wq_online; |
129 | ||
f0127254 JA |
130 | struct io_cb_cancel_data { |
131 | work_cancel_fn *fn; | |
132 | void *data; | |
133 | int nr_running; | |
134 | int nr_pending; | |
135 | bool cancel_all; | |
136 | }; | |
137 | ||
3146cba9 | 138 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index); |
83d6c393 | 139 | static void io_wqe_dec_running(struct io_worker *worker); |
3146cba9 JA |
140 | static bool io_acct_cancel_pending_work(struct io_wqe *wqe, |
141 | struct io_wqe_acct *acct, | |
142 | struct io_cb_cancel_data *match); | |
df1389df | 143 | static void create_worker_cb(struct callback_head *cb); |
f0127254 | 144 | |
771b53d0 JA |
145 | static bool io_worker_get(struct io_worker *worker) |
146 | { | |
147 | return refcount_inc_not_zero(&worker->ref); | |
148 | } | |
149 | ||
150 | static void io_worker_release(struct io_worker *worker) | |
151 | { | |
152 | if (refcount_dec_and_test(&worker->ref)) | |
eb2de941 | 153 | complete(&worker->ref_done); |
771b53d0 JA |
154 | } |
155 | ||
8418f22a PB |
156 | static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound) |
157 | { | |
158 | return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND]; | |
159 | } | |
160 | ||
c5def4ab JA |
161 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
162 | struct io_wq_work *work) | |
163 | { | |
8418f22a | 164 | return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND)); |
c5def4ab JA |
165 | } |
166 | ||
958234d5 | 167 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker) |
c5def4ab | 168 | { |
8418f22a | 169 | return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND); |
c5def4ab JA |
170 | } |
171 | ||
685fe7fe JA |
172 | static void io_worker_ref_put(struct io_wq *wq) |
173 | { | |
174 | if (atomic_dec_and_test(&wq->worker_refs)) | |
175 | complete(&wq->worker_done); | |
176 | } | |
177 | ||
df1389df PB |
178 | static void io_worker_cancel_cb(struct io_worker *worker) |
179 | { | |
180 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); | |
181 | struct io_wqe *wqe = worker->wqe; | |
182 | struct io_wq *wq = wqe->wq; | |
183 | ||
184 | atomic_dec(&acct->nr_running); | |
185 | raw_spin_lock(&worker->wqe->lock); | |
186 | acct->nr_workers--; | |
187 | raw_spin_unlock(&worker->wqe->lock); | |
188 | io_worker_ref_put(wq); | |
189 | clear_bit_unlock(0, &worker->create_state); | |
190 | io_worker_release(worker); | |
191 | } | |
192 | ||
193 | static bool io_task_worker_match(struct callback_head *cb, void *data) | |
194 | { | |
195 | struct io_worker *worker; | |
196 | ||
197 | if (cb->func != create_worker_cb) | |
198 | return false; | |
199 | worker = container_of(cb, struct io_worker, create_work); | |
200 | return worker == data; | |
201 | } | |
202 | ||
771b53d0 JA |
203 | static void io_worker_exit(struct io_worker *worker) |
204 | { | |
205 | struct io_wqe *wqe = worker->wqe; | |
df1389df PB |
206 | struct io_wq *wq = wqe->wq; |
207 | ||
208 | while (1) { | |
209 | struct callback_head *cb = task_work_cancel_match(wq->task, | |
210 | io_task_worker_match, worker); | |
211 | ||
212 | if (!cb) | |
213 | break; | |
214 | io_worker_cancel_cb(worker); | |
215 | } | |
771b53d0 | 216 | |
eb2de941 JA |
217 | if (refcount_dec_and_test(&worker->ref)) |
218 | complete(&worker->ref_done); | |
219 | wait_for_completion(&worker->ref_done); | |
771b53d0 | 220 | |
a9a4aa9f | 221 | raw_spin_lock(&wqe->lock); |
83d6c393 | 222 | if (worker->flags & IO_WORKER_F_FREE) |
bf1daa4b | 223 | hlist_nulls_del_rcu(&worker->nulls_node); |
e61df66c | 224 | list_del_rcu(&worker->all_list); |
83d6c393 JA |
225 | preempt_disable(); |
226 | io_wqe_dec_running(worker); | |
227 | worker->flags = 0; | |
228 | current->flags &= ~PF_IO_WORKER; | |
229 | preempt_enable(); | |
a9a4aa9f | 230 | raw_spin_unlock(&wqe->lock); |
771b53d0 | 231 | |
364b05fd | 232 | kfree_rcu(worker, rcu); |
685fe7fe | 233 | io_worker_ref_put(wqe->wq); |
46fe18b1 | 234 | do_exit(0); |
771b53d0 JA |
235 | } |
236 | ||
f95dc207 | 237 | static inline bool io_acct_run_queue(struct io_wqe_acct *acct) |
c5def4ab | 238 | { |
f95dc207 JA |
239 | if (!wq_list_empty(&acct->work_list) && |
240 | !test_bit(IO_ACCT_STALLED_BIT, &acct->flags)) | |
c5def4ab JA |
241 | return true; |
242 | return false; | |
243 | } | |
244 | ||
245 | /* | |
246 | * Check head of free list for an available worker. If one isn't available, | |
685fe7fe | 247 | * caller must create one. |
c5def4ab | 248 | */ |
f95dc207 JA |
249 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe, |
250 | struct io_wqe_acct *acct) | |
c5def4ab JA |
251 | __must_hold(RCU) |
252 | { | |
253 | struct hlist_nulls_node *n; | |
254 | struct io_worker *worker; | |
255 | ||
83d6c393 JA |
256 | /* |
257 | * Iterate free_list and see if we can find an idle worker to | |
258 | * activate. If a given worker is on the free_list but in the process | |
259 | * of exiting, keep trying. | |
260 | */ | |
261 | hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) { | |
262 | if (!io_worker_get(worker)) | |
263 | continue; | |
f95dc207 JA |
264 | if (io_wqe_get_acct(worker) != acct) { |
265 | io_worker_release(worker); | |
266 | continue; | |
267 | } | |
83d6c393 JA |
268 | if (wake_up_process(worker->task)) { |
269 | io_worker_release(worker); | |
270 | return true; | |
271 | } | |
c5def4ab | 272 | io_worker_release(worker); |
c5def4ab JA |
273 | } |
274 | ||
275 | return false; | |
276 | } | |
277 | ||
278 | /* | |
279 | * We need a worker. If we find a free one, we're good. If not, and we're | |
685fe7fe | 280 | * below the max number of workers, create one. |
c5def4ab | 281 | */ |
3146cba9 | 282 | static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) |
c5def4ab | 283 | { |
c5def4ab JA |
284 | /* |
285 | * Most likely an attempt to queue unbounded work on an io_wq that | |
286 | * wasn't setup with any unbounded workers. | |
287 | */ | |
e6ab8991 PB |
288 | if (unlikely(!acct->max_workers)) |
289 | pr_warn_once("io-wq is not configured for unbound workers"); | |
c5def4ab | 290 | |
94ffb0a2 | 291 | raw_spin_lock(&wqe->lock); |
bc369921 | 292 | if (acct->nr_workers >= acct->max_workers) { |
7a842fb5 HX |
293 | raw_spin_unlock(&wqe->lock); |
294 | return true; | |
94ffb0a2 | 295 | } |
7a842fb5 | 296 | acct->nr_workers++; |
94ffb0a2 | 297 | raw_spin_unlock(&wqe->lock); |
7a842fb5 HX |
298 | atomic_inc(&acct->nr_running); |
299 | atomic_inc(&wqe->wq->worker_refs); | |
300 | return create_io_worker(wqe->wq, wqe, acct->index); | |
c5def4ab JA |
301 | } |
302 | ||
958234d5 | 303 | static void io_wqe_inc_running(struct io_worker *worker) |
c5def4ab | 304 | { |
958234d5 | 305 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
c5def4ab JA |
306 | |
307 | atomic_inc(&acct->nr_running); | |
308 | } | |
309 | ||
685fe7fe JA |
310 | static void create_worker_cb(struct callback_head *cb) |
311 | { | |
d3e9f732 | 312 | struct io_worker *worker; |
685fe7fe | 313 | struct io_wq *wq; |
21698274 HX |
314 | struct io_wqe *wqe; |
315 | struct io_wqe_acct *acct; | |
05c5f4ee | 316 | bool do_create = false; |
685fe7fe | 317 | |
d3e9f732 JA |
318 | worker = container_of(cb, struct io_worker, create_work); |
319 | wqe = worker->wqe; | |
21698274 | 320 | wq = wqe->wq; |
d3e9f732 | 321 | acct = &wqe->acct[worker->create_index]; |
a9a4aa9f | 322 | raw_spin_lock(&wqe->lock); |
49e7f0c7 | 323 | if (acct->nr_workers < acct->max_workers) { |
21698274 | 324 | acct->nr_workers++; |
49e7f0c7 HX |
325 | do_create = true; |
326 | } | |
a9a4aa9f | 327 | raw_spin_unlock(&wqe->lock); |
49e7f0c7 | 328 | if (do_create) { |
05c5f4ee | 329 | create_io_worker(wq, wqe, worker->create_index); |
49e7f0c7 HX |
330 | } else { |
331 | atomic_dec(&acct->nr_running); | |
332 | io_worker_ref_put(wq); | |
333 | } | |
d3e9f732 JA |
334 | clear_bit_unlock(0, &worker->create_state); |
335 | io_worker_release(worker); | |
685fe7fe JA |
336 | } |
337 | ||
3146cba9 JA |
338 | static bool io_queue_worker_create(struct io_worker *worker, |
339 | struct io_wqe_acct *acct, | |
340 | task_work_func_t func) | |
685fe7fe | 341 | { |
3146cba9 | 342 | struct io_wqe *wqe = worker->wqe; |
685fe7fe JA |
343 | struct io_wq *wq = wqe->wq; |
344 | ||
345 | /* raced with exit, just ignore create call */ | |
346 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) | |
347 | goto fail; | |
d3e9f732 JA |
348 | if (!io_worker_get(worker)) |
349 | goto fail; | |
350 | /* | |
351 | * create_state manages ownership of create_work/index. We should | |
352 | * only need one entry per worker, as the worker going to sleep | |
353 | * will trigger the condition, and waking will clear it once it | |
354 | * runs the task_work. | |
355 | */ | |
356 | if (test_bit(0, &worker->create_state) || | |
357 | test_and_set_bit_lock(0, &worker->create_state)) | |
358 | goto fail_release; | |
685fe7fe | 359 | |
3146cba9 | 360 | init_task_work(&worker->create_work, func); |
d3e9f732 | 361 | worker->create_index = acct->index; |
4b3626b0 BC |
362 | if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) { |
363 | clear_bit_unlock(0, &worker->create_state); | |
3146cba9 | 364 | return true; |
4b3626b0 | 365 | } |
d3e9f732 JA |
366 | clear_bit_unlock(0, &worker->create_state); |
367 | fail_release: | |
368 | io_worker_release(worker); | |
685fe7fe JA |
369 | fail: |
370 | atomic_dec(&acct->nr_running); | |
371 | io_worker_ref_put(wq); | |
3146cba9 | 372 | return false; |
685fe7fe JA |
373 | } |
374 | ||
958234d5 | 375 | static void io_wqe_dec_running(struct io_worker *worker) |
c5def4ab JA |
376 | __must_hold(wqe->lock) |
377 | { | |
958234d5 JA |
378 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
379 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab | 380 | |
685fe7fe JA |
381 | if (!(worker->flags & IO_WORKER_F_UP)) |
382 | return; | |
383 | ||
f95dc207 | 384 | if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) { |
685fe7fe JA |
385 | atomic_inc(&acct->nr_running); |
386 | atomic_inc(&wqe->wq->worker_refs); | |
3146cba9 | 387 | io_queue_worker_create(worker, acct, create_worker_cb); |
685fe7fe | 388 | } |
c5def4ab JA |
389 | } |
390 | ||
771b53d0 JA |
391 | /* |
392 | * Worker will start processing some work. Move it to the busy list, if | |
393 | * it's currently on the freelist | |
394 | */ | |
395 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
396 | struct io_wq_work *work) | |
397 | __must_hold(wqe->lock) | |
398 | { | |
399 | if (worker->flags & IO_WORKER_F_FREE) { | |
400 | worker->flags &= ~IO_WORKER_F_FREE; | |
401 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 402 | } |
771b53d0 JA |
403 | } |
404 | ||
405 | /* | |
406 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
407 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
408 | * the lock in that case and return success. Since the caller has to | |
409 | * retry the loop in that case (we changed task state), we don't regrab | |
410 | * the lock if we return success. | |
411 | */ | |
c6d77d92 | 412 | static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) |
771b53d0 JA |
413 | __must_hold(wqe->lock) |
414 | { | |
415 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
416 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 417 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 | 418 | } |
771b53d0 JA |
419 | } |
420 | ||
60cf46ae PB |
421 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
422 | { | |
423 | return work->flags >> IO_WQ_HASH_SHIFT; | |
424 | } | |
425 | ||
8b46e57d | 426 | static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash) |
e941894e JA |
427 | { |
428 | struct io_wq *wq = wqe->wq; | |
8b46e57d | 429 | bool ret = false; |
e941894e | 430 | |
08bdbd39 | 431 | spin_lock_irq(&wq->hash->wait.lock); |
e941894e JA |
432 | if (list_empty(&wqe->wait.entry)) { |
433 | __add_wait_queue(&wq->hash->wait, &wqe->wait); | |
434 | if (!test_bit(hash, &wq->hash->map)) { | |
435 | __set_current_state(TASK_RUNNING); | |
436 | list_del_init(&wqe->wait.entry); | |
8b46e57d | 437 | ret = true; |
e941894e JA |
438 | } |
439 | } | |
08bdbd39 | 440 | spin_unlock_irq(&wq->hash->wait.lock); |
8b46e57d | 441 | return ret; |
e941894e JA |
442 | } |
443 | ||
f95dc207 | 444 | static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct, |
0242f642 | 445 | struct io_worker *worker) |
771b53d0 JA |
446 | __must_hold(wqe->lock) |
447 | { | |
6206f0e1 | 448 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 449 | struct io_wq_work *work, *tail; |
e941894e | 450 | unsigned int stall_hash = -1U; |
f95dc207 | 451 | struct io_wqe *wqe = worker->wqe; |
771b53d0 | 452 | |
f95dc207 | 453 | wq_list_for_each(node, prev, &acct->work_list) { |
e941894e JA |
454 | unsigned int hash; |
455 | ||
6206f0e1 JA |
456 | work = container_of(node, struct io_wq_work, list); |
457 | ||
771b53d0 | 458 | /* not hashed, can run anytime */ |
8766dd51 | 459 | if (!io_wq_is_hashed(work)) { |
f95dc207 | 460 | wq_list_del(&acct->work_list, node, prev); |
771b53d0 JA |
461 | return work; |
462 | } | |
463 | ||
60cf46ae | 464 | hash = io_get_work_hash(work); |
e941894e JA |
465 | /* all items with this hash lie in [work, tail] */ |
466 | tail = wqe->hash_tail[hash]; | |
467 | ||
468 | /* hashed, can run if not already running */ | |
469 | if (!test_and_set_bit(hash, &wqe->wq->hash->map)) { | |
86f3cd1b | 470 | wqe->hash_tail[hash] = NULL; |
f95dc207 | 471 | wq_list_cut(&acct->work_list, &tail->list, prev); |
771b53d0 JA |
472 | return work; |
473 | } | |
e941894e JA |
474 | if (stall_hash == -1U) |
475 | stall_hash = hash; | |
476 | /* fast forward to a next hash, for-each will fix up @prev */ | |
477 | node = &tail->list; | |
478 | } | |
479 | ||
480 | if (stall_hash != -1U) { | |
8b46e57d JA |
481 | bool unstalled; |
482 | ||
0242f642 JA |
483 | /* |
484 | * Set this before dropping the lock to avoid racing with new | |
485 | * work being added and clearing the stalled bit. | |
486 | */ | |
f95dc207 | 487 | set_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
e941894e | 488 | raw_spin_unlock(&wqe->lock); |
8b46e57d | 489 | unstalled = io_wait_on_hash(wqe, stall_hash); |
e941894e | 490 | raw_spin_lock(&wqe->lock); |
8b46e57d JA |
491 | if (unstalled) { |
492 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); | |
493 | if (wq_has_sleeper(&wqe->wq->hash->wait)) | |
494 | wake_up(&wqe->wq->hash->wait); | |
495 | } | |
771b53d0 JA |
496 | } |
497 | ||
498 | return NULL; | |
499 | } | |
500 | ||
00ddff43 | 501 | static bool io_flush_signals(void) |
cccf0ee8 | 502 | { |
0b8cfa97 | 503 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) { |
00ddff43 | 504 | __set_current_state(TASK_RUNNING); |
0b8cfa97 | 505 | tracehook_notify_signal(); |
00ddff43 | 506 | return true; |
cccf0ee8 | 507 | } |
00ddff43 | 508 | return false; |
dc026a73 PB |
509 | } |
510 | ||
511 | static void io_assign_current_work(struct io_worker *worker, | |
512 | struct io_wq_work *work) | |
513 | { | |
d78298e7 | 514 | if (work) { |
3bfe6106 | 515 | io_flush_signals(); |
d78298e7 PB |
516 | cond_resched(); |
517 | } | |
dc026a73 | 518 | |
a9a4aa9f | 519 | spin_lock(&worker->lock); |
dc026a73 | 520 | worker->cur_work = work; |
a9a4aa9f | 521 | spin_unlock(&worker->lock); |
dc026a73 PB |
522 | } |
523 | ||
60cf46ae PB |
524 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
525 | ||
771b53d0 JA |
526 | static void io_worker_handle_work(struct io_worker *worker) |
527 | __releases(wqe->lock) | |
528 | { | |
f95dc207 | 529 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
771b53d0 JA |
530 | struct io_wqe *wqe = worker->wqe; |
531 | struct io_wq *wq = wqe->wq; | |
c60eb049 | 532 | bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state); |
771b53d0 JA |
533 | |
534 | do { | |
86f3cd1b | 535 | struct io_wq_work *work; |
f462fd36 | 536 | get_next: |
771b53d0 JA |
537 | /* |
538 | * If we got some work, mark us as busy. If we didn't, but | |
539 | * the list isn't empty, it means we stalled on hashed work. | |
540 | * Mark us stalled so we don't keep looking for work when we | |
541 | * can't make progress, any work completion or insertion will | |
542 | * clear the stalled flag. | |
543 | */ | |
f95dc207 | 544 | work = io_get_next_work(acct, worker); |
771b53d0 JA |
545 | if (work) |
546 | __io_worker_busy(wqe, worker, work); | |
771b53d0 | 547 | |
a9a4aa9f | 548 | raw_spin_unlock(&wqe->lock); |
771b53d0 JA |
549 | if (!work) |
550 | break; | |
58e39319 | 551 | io_assign_current_work(worker, work); |
e941894e | 552 | __set_current_state(TASK_RUNNING); |
36c2f922 | 553 | |
dc026a73 PB |
554 | /* handle a whole dependent link */ |
555 | do { | |
5280f7e5 | 556 | struct io_wq_work *next_hashed, *linked; |
b089ed39 | 557 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 558 | |
86f3cd1b | 559 | next_hashed = wq_next_work(work); |
c60eb049 PB |
560 | |
561 | if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND)) | |
562 | work->flags |= IO_WQ_WORK_CANCEL; | |
5280f7e5 PB |
563 | wq->do_work(work); |
564 | io_assign_current_work(worker, NULL); | |
dc026a73 | 565 | |
5280f7e5 | 566 | linked = wq->free_work(work); |
86f3cd1b PB |
567 | work = next_hashed; |
568 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
569 | work = linked; | |
570 | linked = NULL; | |
571 | } | |
572 | io_assign_current_work(worker, work); | |
86f3cd1b PB |
573 | if (linked) |
574 | io_wqe_enqueue(wqe, linked); | |
575 | ||
576 | if (hash != -1U && !next_hashed) { | |
8b46e57d JA |
577 | /* serialize hash clear with wake_up() */ |
578 | spin_lock_irq(&wq->hash->wait.lock); | |
e941894e | 579 | clear_bit(hash, &wq->hash->map); |
f95dc207 | 580 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
8b46e57d | 581 | spin_unlock_irq(&wq->hash->wait.lock); |
e941894e JA |
582 | if (wq_has_sleeper(&wq->hash->wait)) |
583 | wake_up(&wq->hash->wait); | |
a9a4aa9f | 584 | raw_spin_lock(&wqe->lock); |
f462fd36 PB |
585 | /* skip unnecessary unlock-lock wqe->lock */ |
586 | if (!work) | |
587 | goto get_next; | |
a9a4aa9f | 588 | raw_spin_unlock(&wqe->lock); |
7d723065 | 589 | } |
58e39319 | 590 | } while (work); |
7d723065 | 591 | |
a9a4aa9f | 592 | raw_spin_lock(&wqe->lock); |
771b53d0 JA |
593 | } while (1); |
594 | } | |
595 | ||
771b53d0 JA |
596 | static int io_wqe_worker(void *data) |
597 | { | |
598 | struct io_worker *worker = data; | |
f95dc207 | 599 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
771b53d0 JA |
600 | struct io_wqe *wqe = worker->wqe; |
601 | struct io_wq *wq = wqe->wq; | |
05c5f4ee | 602 | bool last_timeout = false; |
46fe18b1 | 603 | char buf[TASK_COMM_LEN]; |
771b53d0 | 604 | |
46fe18b1 | 605 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); |
46fe18b1 | 606 | |
685fe7fe | 607 | snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid); |
46fe18b1 | 608 | set_task_comm(current, buf); |
771b53d0 JA |
609 | |
610 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
16efa4fc JA |
611 | long ret; |
612 | ||
506d95ff | 613 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 614 | loop: |
a9a4aa9f | 615 | raw_spin_lock(&wqe->lock); |
f95dc207 | 616 | if (io_acct_run_queue(acct)) { |
771b53d0 | 617 | io_worker_handle_work(worker); |
e995d512 | 618 | goto loop; |
771b53d0 | 619 | } |
05c5f4ee JA |
620 | /* timed out, exit unless we're the last worker */ |
621 | if (last_timeout && acct->nr_workers > 1) { | |
767a65e9 | 622 | acct->nr_workers--; |
05c5f4ee JA |
623 | raw_spin_unlock(&wqe->lock); |
624 | __set_current_state(TASK_RUNNING); | |
625 | break; | |
626 | } | |
627 | last_timeout = false; | |
c6d77d92 | 628 | __io_worker_idle(wqe, worker); |
a9a4aa9f | 629 | raw_spin_unlock(&wqe->lock); |
00ddff43 JA |
630 | if (io_flush_signals()) |
631 | continue; | |
16efa4fc | 632 | ret = schedule_timeout(WORKER_IDLE_TIMEOUT); |
dbe1bdbb JA |
633 | if (signal_pending(current)) { |
634 | struct ksignal ksig; | |
635 | ||
636 | if (!get_signal(&ksig)) | |
637 | continue; | |
78f8876c | 638 | break; |
dbe1bdbb | 639 | } |
05c5f4ee | 640 | last_timeout = !ret; |
771b53d0 JA |
641 | } |
642 | ||
771b53d0 | 643 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
a9a4aa9f | 644 | raw_spin_lock(&wqe->lock); |
e587227b | 645 | io_worker_handle_work(worker); |
771b53d0 JA |
646 | } |
647 | ||
648 | io_worker_exit(worker); | |
649 | return 0; | |
650 | } | |
651 | ||
771b53d0 JA |
652 | /* |
653 | * Called when a worker is scheduled in. Mark us as currently running. | |
654 | */ | |
655 | void io_wq_worker_running(struct task_struct *tsk) | |
656 | { | |
3bfe6106 | 657 | struct io_worker *worker = tsk->pf_io_worker; |
771b53d0 | 658 | |
3bfe6106 JA |
659 | if (!worker) |
660 | return; | |
771b53d0 JA |
661 | if (!(worker->flags & IO_WORKER_F_UP)) |
662 | return; | |
663 | if (worker->flags & IO_WORKER_F_RUNNING) | |
664 | return; | |
665 | worker->flags |= IO_WORKER_F_RUNNING; | |
958234d5 | 666 | io_wqe_inc_running(worker); |
771b53d0 JA |
667 | } |
668 | ||
669 | /* | |
670 | * Called when worker is going to sleep. If there are no workers currently | |
685fe7fe | 671 | * running and we have work pending, wake up a free one or create a new one. |
771b53d0 JA |
672 | */ |
673 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
674 | { | |
3bfe6106 | 675 | struct io_worker *worker = tsk->pf_io_worker; |
771b53d0 | 676 | |
3bfe6106 JA |
677 | if (!worker) |
678 | return; | |
771b53d0 JA |
679 | if (!(worker->flags & IO_WORKER_F_UP)) |
680 | return; | |
681 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
682 | return; | |
683 | ||
684 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
685 | ||
a9a4aa9f | 686 | raw_spin_lock(&worker->wqe->lock); |
958234d5 | 687 | io_wqe_dec_running(worker); |
a9a4aa9f | 688 | raw_spin_unlock(&worker->wqe->lock); |
771b53d0 JA |
689 | } |
690 | ||
3146cba9 JA |
691 | static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker, |
692 | struct task_struct *tsk) | |
693 | { | |
694 | tsk->pf_io_worker = worker; | |
695 | worker->task = tsk; | |
696 | set_cpus_allowed_ptr(tsk, wqe->cpu_mask); | |
697 | tsk->flags |= PF_NO_SETAFFINITY; | |
698 | ||
699 | raw_spin_lock(&wqe->lock); | |
700 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); | |
701 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); | |
702 | worker->flags |= IO_WORKER_F_FREE; | |
703 | raw_spin_unlock(&wqe->lock); | |
704 | wake_up_new_task(tsk); | |
705 | } | |
706 | ||
707 | static bool io_wq_work_match_all(struct io_wq_work *work, void *data) | |
708 | { | |
709 | return true; | |
710 | } | |
711 | ||
712 | static inline bool io_should_retry_thread(long err) | |
713 | { | |
714 | switch (err) { | |
715 | case -EAGAIN: | |
716 | case -ERESTARTSYS: | |
717 | case -ERESTARTNOINTR: | |
718 | case -ERESTARTNOHAND: | |
719 | return true; | |
720 | default: | |
721 | return false; | |
722 | } | |
723 | } | |
724 | ||
725 | static void create_worker_cont(struct callback_head *cb) | |
726 | { | |
727 | struct io_worker *worker; | |
728 | struct task_struct *tsk; | |
729 | struct io_wqe *wqe; | |
730 | ||
731 | worker = container_of(cb, struct io_worker, create_work); | |
732 | clear_bit_unlock(0, &worker->create_state); | |
733 | wqe = worker->wqe; | |
734 | tsk = create_io_thread(io_wqe_worker, worker, wqe->node); | |
735 | if (!IS_ERR(tsk)) { | |
736 | io_init_new_worker(wqe, worker, tsk); | |
737 | io_worker_release(worker); | |
738 | return; | |
739 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { | |
740 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); | |
741 | ||
742 | atomic_dec(&acct->nr_running); | |
743 | raw_spin_lock(&wqe->lock); | |
744 | acct->nr_workers--; | |
745 | if (!acct->nr_workers) { | |
746 | struct io_cb_cancel_data match = { | |
747 | .fn = io_wq_work_match_all, | |
748 | .cancel_all = true, | |
749 | }; | |
750 | ||
751 | while (io_acct_cancel_pending_work(wqe, acct, &match)) | |
752 | raw_spin_lock(&wqe->lock); | |
753 | } | |
754 | raw_spin_unlock(&wqe->lock); | |
755 | io_worker_ref_put(wqe->wq); | |
66e70be7 | 756 | kfree(worker); |
3146cba9 JA |
757 | return; |
758 | } | |
759 | ||
760 | /* re-create attempts grab a new worker ref, drop the existing one */ | |
761 | io_worker_release(worker); | |
762 | schedule_work(&worker->work); | |
763 | } | |
764 | ||
765 | static void io_workqueue_create(struct work_struct *work) | |
766 | { | |
767 | struct io_worker *worker = container_of(work, struct io_worker, work); | |
768 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); | |
769 | ||
4b3626b0 | 770 | if (!io_queue_worker_create(worker, acct, create_worker_cont)) |
66e70be7 | 771 | kfree(worker); |
3146cba9 JA |
772 | } |
773 | ||
774 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) | |
3bfe6106 | 775 | { |
46fe18b1 | 776 | struct io_wqe_acct *acct = &wqe->acct[index]; |
3bfe6106 | 777 | struct io_worker *worker; |
46fe18b1 | 778 | struct task_struct *tsk; |
3bfe6106 | 779 | |
8b3e78b5 JA |
780 | __set_current_state(TASK_RUNNING); |
781 | ||
3bfe6106 | 782 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
3146cba9 | 783 | if (!worker) { |
685fe7fe JA |
784 | fail: |
785 | atomic_dec(&acct->nr_running); | |
a9a4aa9f | 786 | raw_spin_lock(&wqe->lock); |
3d4e4fac | 787 | acct->nr_workers--; |
a9a4aa9f | 788 | raw_spin_unlock(&wqe->lock); |
685fe7fe | 789 | io_worker_ref_put(wq); |
3146cba9 | 790 | return false; |
3bfe6106 | 791 | } |
46fe18b1 | 792 | |
3146cba9 JA |
793 | refcount_set(&worker->ref, 1); |
794 | worker->wqe = wqe; | |
795 | spin_lock_init(&worker->lock); | |
796 | init_completion(&worker->ref_done); | |
46fe18b1 | 797 | |
46fe18b1 JA |
798 | if (index == IO_WQ_ACCT_BOUND) |
799 | worker->flags |= IO_WORKER_F_BOUND; | |
3146cba9 JA |
800 | |
801 | tsk = create_io_thread(io_wqe_worker, worker, wqe->node); | |
802 | if (!IS_ERR(tsk)) { | |
803 | io_init_new_worker(wqe, worker, tsk); | |
804 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { | |
66e70be7 | 805 | kfree(worker); |
3146cba9 JA |
806 | goto fail; |
807 | } else { | |
808 | INIT_WORK(&worker->work, io_workqueue_create); | |
809 | schedule_work(&worker->work); | |
810 | } | |
811 | ||
812 | return true; | |
771b53d0 JA |
813 | } |
814 | ||
c4068bf8 HD |
815 | /* |
816 | * Iterate the passed in list and call the specific function for each | |
817 | * worker that isn't exiting | |
818 | */ | |
819 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
820 | bool (*func)(struct io_worker *, void *), | |
821 | void *data) | |
822 | { | |
823 | struct io_worker *worker; | |
824 | bool ret = false; | |
825 | ||
826 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { | |
827 | if (io_worker_get(worker)) { | |
828 | /* no task if node is/was offline */ | |
829 | if (worker->task) | |
830 | ret = func(worker, data); | |
831 | io_worker_release(worker); | |
832 | if (ret) | |
833 | break; | |
834 | } | |
835 | } | |
836 | ||
837 | return ret; | |
838 | } | |
839 | ||
840 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
841 | { | |
46fe18b1 | 842 | set_notify_signal(worker->task); |
c4068bf8 HD |
843 | wake_up_process(worker->task); |
844 | return false; | |
845 | } | |
846 | ||
e9fd9396 | 847 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 848 | { |
e9fd9396 PB |
849 | struct io_wq *wq = wqe->wq; |
850 | ||
fc04c39b | 851 | do { |
fc04c39b | 852 | work->flags |= IO_WQ_WORK_CANCEL; |
5280f7e5 PB |
853 | wq->do_work(work); |
854 | work = wq->free_work(work); | |
fc04c39b PB |
855 | } while (work); |
856 | } | |
857 | ||
86f3cd1b PB |
858 | static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work) |
859 | { | |
f95dc207 | 860 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
86f3cd1b PB |
861 | unsigned int hash; |
862 | struct io_wq_work *tail; | |
863 | ||
864 | if (!io_wq_is_hashed(work)) { | |
865 | append: | |
f95dc207 | 866 | wq_list_add_tail(&work->list, &acct->work_list); |
86f3cd1b PB |
867 | return; |
868 | } | |
869 | ||
870 | hash = io_get_work_hash(work); | |
871 | tail = wqe->hash_tail[hash]; | |
872 | wqe->hash_tail[hash] = work; | |
873 | if (!tail) | |
874 | goto append; | |
875 | ||
f95dc207 | 876 | wq_list_add_after(&work->list, &tail->list, &acct->work_list); |
86f3cd1b PB |
877 | } |
878 | ||
713b9825 PB |
879 | static bool io_wq_work_match_item(struct io_wq_work *work, void *data) |
880 | { | |
881 | return work == data; | |
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); |
94ffb0a2 JA |
887 | unsigned work_flags = work->flags; |
888 | bool do_create; | |
771b53d0 | 889 | |
991468dc JA |
890 | /* |
891 | * If io-wq is exiting for this task, or if the request has explicitly | |
892 | * been marked as one that should not get executed, cancel it here. | |
893 | */ | |
894 | if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) || | |
895 | (work->flags & IO_WQ_WORK_CANCEL)) { | |
70e35125 | 896 | io_run_cancel(work, wqe); |
4fb6ac32 JA |
897 | return; |
898 | } | |
899 | ||
a9a4aa9f | 900 | raw_spin_lock(&wqe->lock); |
86f3cd1b | 901 | io_wqe_insert_work(wqe, work); |
f95dc207 | 902 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
94ffb0a2 JA |
903 | |
904 | rcu_read_lock(); | |
f95dc207 | 905 | do_create = !io_wqe_activate_free_worker(wqe, acct); |
94ffb0a2 JA |
906 | rcu_read_unlock(); |
907 | ||
a9a4aa9f | 908 | raw_spin_unlock(&wqe->lock); |
771b53d0 | 909 | |
94ffb0a2 | 910 | if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) || |
3146cba9 JA |
911 | !atomic_read(&acct->nr_running))) { |
912 | bool did_create; | |
913 | ||
914 | did_create = io_wqe_create_worker(wqe, acct); | |
713b9825 PB |
915 | if (likely(did_create)) |
916 | return; | |
917 | ||
918 | raw_spin_lock(&wqe->lock); | |
919 | /* fatal condition, failed to create the first worker */ | |
920 | if (!acct->nr_workers) { | |
921 | struct io_cb_cancel_data match = { | |
922 | .fn = io_wq_work_match_item, | |
923 | .data = work, | |
924 | .cancel_all = false, | |
925 | }; | |
926 | ||
927 | if (io_acct_cancel_pending_work(wqe, acct, &match)) | |
928 | raw_spin_lock(&wqe->lock); | |
3146cba9 | 929 | } |
713b9825 | 930 | raw_spin_unlock(&wqe->lock); |
3146cba9 | 931 | } |
771b53d0 JA |
932 | } |
933 | ||
934 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
935 | { | |
936 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
937 | ||
938 | io_wqe_enqueue(wqe, work); | |
939 | } | |
940 | ||
941 | /* | |
8766dd51 PB |
942 | * Work items that hash to the same value will not be done in parallel. |
943 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 944 | */ |
8766dd51 | 945 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 946 | { |
8766dd51 | 947 | unsigned int bit; |
771b53d0 JA |
948 | |
949 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
950 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
951 | } |
952 | ||
2293b419 | 953 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 954 | { |
2293b419 | 955 | struct io_cb_cancel_data *match = data; |
62755e35 JA |
956 | |
957 | /* | |
958 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 959 | * may dereference the passed in work. |
62755e35 | 960 | */ |
a9a4aa9f | 961 | spin_lock(&worker->lock); |
62755e35 | 962 | if (worker->cur_work && |
2293b419 | 963 | match->fn(worker->cur_work, match->data)) { |
3bfe6106 | 964 | set_notify_signal(worker->task); |
4f26bda1 | 965 | match->nr_running++; |
771b53d0 | 966 | } |
a9a4aa9f | 967 | spin_unlock(&worker->lock); |
771b53d0 | 968 | |
4f26bda1 | 969 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
970 | } |
971 | ||
204361a7 PB |
972 | static inline void io_wqe_remove_pending(struct io_wqe *wqe, |
973 | struct io_wq_work *work, | |
974 | struct io_wq_work_node *prev) | |
975 | { | |
f95dc207 | 976 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
204361a7 PB |
977 | unsigned int hash = io_get_work_hash(work); |
978 | struct io_wq_work *prev_work = NULL; | |
979 | ||
980 | if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) { | |
981 | if (prev) | |
982 | prev_work = container_of(prev, struct io_wq_work, list); | |
983 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
984 | wqe->hash_tail[hash] = prev_work; | |
985 | else | |
986 | wqe->hash_tail[hash] = NULL; | |
987 | } | |
f95dc207 | 988 | wq_list_del(&acct->work_list, &work->list, prev); |
204361a7 PB |
989 | } |
990 | ||
3146cba9 JA |
991 | static bool io_acct_cancel_pending_work(struct io_wqe *wqe, |
992 | struct io_wqe_acct *acct, | |
993 | struct io_cb_cancel_data *match) | |
994 | __releases(wqe->lock) | |
771b53d0 | 995 | { |
6206f0e1 | 996 | struct io_wq_work_node *node, *prev; |
771b53d0 | 997 | struct io_wq_work *work; |
771b53d0 | 998 | |
3146cba9 JA |
999 | wq_list_for_each(node, prev, &acct->work_list) { |
1000 | work = container_of(node, struct io_wq_work, list); | |
1001 | if (!match->fn(work, match->data)) | |
1002 | continue; | |
1003 | io_wqe_remove_pending(wqe, work, prev); | |
1004 | raw_spin_unlock(&wqe->lock); | |
1005 | io_run_cancel(work, wqe); | |
1006 | match->nr_pending++; | |
1007 | /* not safe to continue after unlock */ | |
1008 | return true; | |
1009 | } | |
1010 | ||
1011 | return false; | |
1012 | } | |
1013 | ||
1014 | static void io_wqe_cancel_pending_work(struct io_wqe *wqe, | |
1015 | struct io_cb_cancel_data *match) | |
1016 | { | |
1017 | int i; | |
4f26bda1 | 1018 | retry: |
a9a4aa9f | 1019 | raw_spin_lock(&wqe->lock); |
f95dc207 JA |
1020 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
1021 | struct io_wqe_acct *acct = io_get_acct(wqe, i == 0); | |
4f26bda1 | 1022 | |
3146cba9 JA |
1023 | if (io_acct_cancel_pending_work(wqe, acct, match)) { |
1024 | if (match->cancel_all) | |
1025 | goto retry; | |
1026 | return; | |
f95dc207 | 1027 | } |
771b53d0 | 1028 | } |
a9a4aa9f | 1029 | raw_spin_unlock(&wqe->lock); |
f4c2665e PB |
1030 | } |
1031 | ||
4f26bda1 | 1032 | static void io_wqe_cancel_running_work(struct io_wqe *wqe, |
f4c2665e PB |
1033 | struct io_cb_cancel_data *match) |
1034 | { | |
771b53d0 | 1035 | rcu_read_lock(); |
4f26bda1 | 1036 | io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 | 1037 | rcu_read_unlock(); |
771b53d0 JA |
1038 | } |
1039 | ||
2293b419 | 1040 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 1041 | void *data, bool cancel_all) |
771b53d0 | 1042 | { |
2293b419 | 1043 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
1044 | .fn = cancel, |
1045 | .data = data, | |
1046 | .cancel_all = cancel_all, | |
00bcda13 | 1047 | }; |
3fc50ab5 | 1048 | int node; |
771b53d0 | 1049 | |
f4c2665e PB |
1050 | /* |
1051 | * First check pending list, if we're lucky we can just remove it | |
1052 | * from there. CANCEL_OK means that the work is returned as-new, | |
1053 | * no completion will be posted for it. | |
1054 | */ | |
3fc50ab5 JH |
1055 | for_each_node(node) { |
1056 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 1057 | |
4f26bda1 PB |
1058 | io_wqe_cancel_pending_work(wqe, &match); |
1059 | if (match.nr_pending && !match.cancel_all) | |
f4c2665e | 1060 | return IO_WQ_CANCEL_OK; |
771b53d0 JA |
1061 | } |
1062 | ||
f4c2665e PB |
1063 | /* |
1064 | * Now check if a free (going busy) or busy worker has the work | |
1065 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
1066 | * as an indication that we attempt to signal cancellation. The | |
1067 | * completion will run normally in this case. | |
1068 | */ | |
1069 | for_each_node(node) { | |
1070 | struct io_wqe *wqe = wq->wqes[node]; | |
1071 | ||
4f26bda1 PB |
1072 | io_wqe_cancel_running_work(wqe, &match); |
1073 | if (match.nr_running && !match.cancel_all) | |
f4c2665e PB |
1074 | return IO_WQ_CANCEL_RUNNING; |
1075 | } | |
1076 | ||
4f26bda1 PB |
1077 | if (match.nr_running) |
1078 | return IO_WQ_CANCEL_RUNNING; | |
1079 | if (match.nr_pending) | |
1080 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 1081 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
1082 | } |
1083 | ||
e941894e JA |
1084 | static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode, |
1085 | int sync, void *key) | |
1086 | { | |
1087 | struct io_wqe *wqe = container_of(wait, struct io_wqe, wait); | |
f95dc207 | 1088 | int i; |
e941894e JA |
1089 | |
1090 | list_del_init(&wait->entry); | |
1091 | ||
1092 | rcu_read_lock(); | |
f95dc207 JA |
1093 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
1094 | struct io_wqe_acct *acct = &wqe->acct[i]; | |
1095 | ||
1096 | if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags)) | |
1097 | io_wqe_activate_free_worker(wqe, acct); | |
1098 | } | |
e941894e | 1099 | rcu_read_unlock(); |
e941894e JA |
1100 | return 1; |
1101 | } | |
1102 | ||
576a347b | 1103 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1104 | { |
f95dc207 | 1105 | int ret, node, i; |
771b53d0 JA |
1106 | struct io_wq *wq; |
1107 | ||
f5fa38c5 | 1108 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 | 1109 | return ERR_PTR(-EINVAL); |
e6ab8991 PB |
1110 | if (WARN_ON_ONCE(!bounded)) |
1111 | return ERR_PTR(-EINVAL); | |
e9fd9396 | 1112 | |
c7f405d6 | 1113 | wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL); |
771b53d0 JA |
1114 | if (!wq) |
1115 | return ERR_PTR(-ENOMEM); | |
43c01fbe JA |
1116 | ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
1117 | if (ret) | |
c7f405d6 | 1118 | goto err_wq; |
771b53d0 | 1119 | |
e941894e JA |
1120 | refcount_inc(&data->hash->refs); |
1121 | wq->hash = data->hash; | |
e9fd9396 | 1122 | wq->free_work = data->free_work; |
f5fa38c5 | 1123 | wq->do_work = data->do_work; |
7d723065 | 1124 | |
43c01fbe | 1125 | ret = -ENOMEM; |
3fc50ab5 | 1126 | for_each_node(node) { |
771b53d0 | 1127 | struct io_wqe *wqe; |
7563439a | 1128 | int alloc_node = node; |
771b53d0 | 1129 | |
7563439a JA |
1130 | if (!node_online(alloc_node)) |
1131 | alloc_node = NUMA_NO_NODE; | |
1132 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1133 | if (!wqe) |
3fc50ab5 | 1134 | goto err; |
0e03496d JA |
1135 | if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL)) |
1136 | goto err; | |
1137 | cpumask_copy(wqe->cpu_mask, cpumask_of_node(node)); | |
3fc50ab5 | 1138 | wq->wqes[node] = wqe; |
7563439a | 1139 | wqe->node = alloc_node; |
c5def4ab | 1140 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
728f13e7 | 1141 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
c5def4ab | 1142 | task_rlimit(current, RLIMIT_NPROC); |
e941894e | 1143 | INIT_LIST_HEAD(&wqe->wait.entry); |
f95dc207 JA |
1144 | wqe->wait.func = io_wqe_hash_wake; |
1145 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { | |
1146 | struct io_wqe_acct *acct = &wqe->acct[i]; | |
1147 | ||
1148 | acct->index = i; | |
1149 | atomic_set(&acct->nr_running, 0); | |
1150 | INIT_WQ_LIST(&acct->work_list); | |
1151 | } | |
771b53d0 | 1152 | wqe->wq = wq; |
95da8465 | 1153 | raw_spin_lock_init(&wqe->lock); |
021d1cdd | 1154 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1155 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1156 | } |
1157 | ||
685fe7fe | 1158 | wq->task = get_task_struct(data->task); |
685fe7fe JA |
1159 | atomic_set(&wq->worker_refs, 1); |
1160 | init_completion(&wq->worker_done); | |
1161 | return wq; | |
b60fda60 | 1162 | err: |
dc7bbc9e | 1163 | io_wq_put_hash(data->hash); |
43c01fbe | 1164 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
0e03496d JA |
1165 | for_each_node(node) { |
1166 | if (!wq->wqes[node]) | |
1167 | continue; | |
1168 | free_cpumask_var(wq->wqes[node]->cpu_mask); | |
3fc50ab5 | 1169 | kfree(wq->wqes[node]); |
0e03496d | 1170 | } |
43c01fbe | 1171 | err_wq: |
b60fda60 | 1172 | kfree(wq); |
771b53d0 JA |
1173 | return ERR_PTR(ret); |
1174 | } | |
1175 | ||
c80ca470 JA |
1176 | static bool io_task_work_match(struct callback_head *cb, void *data) |
1177 | { | |
d3e9f732 | 1178 | struct io_worker *worker; |
c80ca470 | 1179 | |
3b33e3f4 | 1180 | if (cb->func != create_worker_cb && cb->func != create_worker_cont) |
c80ca470 | 1181 | return false; |
d3e9f732 JA |
1182 | worker = container_of(cb, struct io_worker, create_work); |
1183 | return worker->wqe->wq == data; | |
c80ca470 JA |
1184 | } |
1185 | ||
17a91051 PB |
1186 | void io_wq_exit_start(struct io_wq *wq) |
1187 | { | |
1188 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
1189 | } | |
1190 | ||
685fe7fe | 1191 | static void io_wq_exit_workers(struct io_wq *wq) |
afcc4015 | 1192 | { |
685fe7fe JA |
1193 | struct callback_head *cb; |
1194 | int node; | |
1195 | ||
685fe7fe JA |
1196 | if (!wq->task) |
1197 | return; | |
1198 | ||
c80ca470 | 1199 | while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) { |
d3e9f732 | 1200 | struct io_worker *worker; |
685fe7fe | 1201 | |
d3e9f732 | 1202 | worker = container_of(cb, struct io_worker, create_work); |
df1389df | 1203 | io_worker_cancel_cb(worker); |
685fe7fe JA |
1204 | } |
1205 | ||
1206 | rcu_read_lock(); | |
1207 | for_each_node(node) { | |
1208 | struct io_wqe *wqe = wq->wqes[node]; | |
1209 | ||
1210 | io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL); | |
afcc4015 | 1211 | } |
685fe7fe JA |
1212 | rcu_read_unlock(); |
1213 | io_worker_ref_put(wq); | |
1214 | wait_for_completion(&wq->worker_done); | |
3743c172 Z |
1215 | |
1216 | for_each_node(node) { | |
1217 | spin_lock_irq(&wq->hash->wait.lock); | |
1218 | list_del_init(&wq->wqes[node]->wait.entry); | |
1219 | spin_unlock_irq(&wq->hash->wait.lock); | |
1220 | } | |
685fe7fe JA |
1221 | put_task_struct(wq->task); |
1222 | wq->task = NULL; | |
afcc4015 JA |
1223 | } |
1224 | ||
4fb6ac32 | 1225 | static void io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1226 | { |
3fc50ab5 | 1227 | int node; |
771b53d0 | 1228 | |
43c01fbe JA |
1229 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
1230 | ||
e941894e JA |
1231 | for_each_node(node) { |
1232 | struct io_wqe *wqe = wq->wqes[node]; | |
f5d2d23b JA |
1233 | struct io_cb_cancel_data match = { |
1234 | .fn = io_wq_work_match_all, | |
1235 | .cancel_all = true, | |
1236 | }; | |
1237 | io_wqe_cancel_pending_work(wqe, &match); | |
0e03496d | 1238 | free_cpumask_var(wqe->cpu_mask); |
e941894e JA |
1239 | kfree(wqe); |
1240 | } | |
e941894e | 1241 | io_wq_put_hash(wq->hash); |
771b53d0 | 1242 | kfree(wq); |
4fb6ac32 JA |
1243 | } |
1244 | ||
afcc4015 JA |
1245 | void io_wq_put_and_exit(struct io_wq *wq) |
1246 | { | |
17a91051 PB |
1247 | WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state)); |
1248 | ||
685fe7fe | 1249 | io_wq_exit_workers(wq); |
382cb030 | 1250 | io_wq_destroy(wq); |
afcc4015 JA |
1251 | } |
1252 | ||
0e03496d JA |
1253 | struct online_data { |
1254 | unsigned int cpu; | |
1255 | bool online; | |
1256 | }; | |
1257 | ||
43c01fbe JA |
1258 | static bool io_wq_worker_affinity(struct io_worker *worker, void *data) |
1259 | { | |
0e03496d | 1260 | struct online_data *od = data; |
e0051d7d | 1261 | |
0e03496d JA |
1262 | if (od->online) |
1263 | cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask); | |
1264 | else | |
1265 | cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask); | |
43c01fbe JA |
1266 | return false; |
1267 | } | |
1268 | ||
0e03496d | 1269 | static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online) |
43c01fbe | 1270 | { |
0e03496d JA |
1271 | struct online_data od = { |
1272 | .cpu = cpu, | |
1273 | .online = online | |
1274 | }; | |
43c01fbe JA |
1275 | int i; |
1276 | ||
1277 | rcu_read_lock(); | |
1278 | for_each_node(i) | |
0e03496d | 1279 | io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od); |
43c01fbe JA |
1280 | rcu_read_unlock(); |
1281 | return 0; | |
1282 | } | |
1283 | ||
0e03496d JA |
1284 | static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node) |
1285 | { | |
1286 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1287 | ||
1288 | return __io_wq_cpu_online(wq, cpu, true); | |
1289 | } | |
1290 | ||
1291 | static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node) | |
1292 | { | |
1293 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1294 | ||
1295 | return __io_wq_cpu_online(wq, cpu, false); | |
1296 | } | |
1297 | ||
fe76421d JA |
1298 | int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask) |
1299 | { | |
1300 | int i; | |
1301 | ||
1302 | rcu_read_lock(); | |
1303 | for_each_node(i) { | |
1304 | struct io_wqe *wqe = wq->wqes[i]; | |
1305 | ||
1306 | if (mask) | |
1307 | cpumask_copy(wqe->cpu_mask, mask); | |
1308 | else | |
1309 | cpumask_copy(wqe->cpu_mask, cpumask_of_node(i)); | |
1310 | } | |
1311 | rcu_read_unlock(); | |
1312 | return 0; | |
1313 | } | |
1314 | ||
2e480058 JA |
1315 | /* |
1316 | * Set max number of unbounded workers, returns old value. If new_count is 0, | |
1317 | * then just return the old value. | |
1318 | */ | |
1319 | int io_wq_max_workers(struct io_wq *wq, int *new_count) | |
1320 | { | |
2093ae1d BZ |
1321 | int prev[IO_WQ_ACCT_NR]; |
1322 | bool first_node = true; | |
1323 | int i, node; | |
2e480058 | 1324 | |
dd47c104 ES |
1325 | BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND); |
1326 | BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND); | |
1327 | BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2); | |
1328 | ||
2e480058 JA |
1329 | for (i = 0; i < 2; i++) { |
1330 | if (new_count[i] > task_rlimit(current, RLIMIT_NPROC)) | |
1331 | new_count[i] = task_rlimit(current, RLIMIT_NPROC); | |
1332 | } | |
1333 | ||
2093ae1d BZ |
1334 | for (i = 0; i < IO_WQ_ACCT_NR; i++) |
1335 | prev[i] = 0; | |
1336 | ||
2e480058 JA |
1337 | rcu_read_lock(); |
1338 | for_each_node(node) { | |
bc369921 | 1339 | struct io_wqe *wqe = wq->wqes[node]; |
2e480058 JA |
1340 | struct io_wqe_acct *acct; |
1341 | ||
bc369921 | 1342 | raw_spin_lock(&wqe->lock); |
f95dc207 | 1343 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
bc369921 | 1344 | acct = &wqe->acct[i]; |
2093ae1d BZ |
1345 | if (first_node) |
1346 | prev[i] = max_t(int, acct->max_workers, prev[i]); | |
2e480058 JA |
1347 | if (new_count[i]) |
1348 | acct->max_workers = new_count[i]; | |
2e480058 | 1349 | } |
bc369921 | 1350 | raw_spin_unlock(&wqe->lock); |
2093ae1d | 1351 | first_node = false; |
2e480058 JA |
1352 | } |
1353 | rcu_read_unlock(); | |
2093ae1d BZ |
1354 | |
1355 | for (i = 0; i < IO_WQ_ACCT_NR; i++) | |
1356 | new_count[i] = prev[i]; | |
1357 | ||
2e480058 JA |
1358 | return 0; |
1359 | } | |
1360 | ||
43c01fbe JA |
1361 | static __init int io_wq_init(void) |
1362 | { | |
1363 | int ret; | |
1364 | ||
1365 | ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online", | |
0e03496d | 1366 | io_wq_cpu_online, io_wq_cpu_offline); |
43c01fbe JA |
1367 | if (ret < 0) |
1368 | return ret; | |
1369 | io_wq_online = ret; | |
1370 | return 0; | |
1371 | } | |
1372 | subsys_initcall(io_wq_init); |