]>
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); | |
1a2b192e | 143 | static void create_worker_cb(struct callback_head *cb); |
d86fed57 | 144 | static void io_wq_cancel_tw_create(struct io_wq *wq); |
f0127254 | 145 | |
771b53d0 JA |
146 | static bool io_worker_get(struct io_worker *worker) |
147 | { | |
148 | return refcount_inc_not_zero(&worker->ref); | |
149 | } | |
150 | ||
151 | static void io_worker_release(struct io_worker *worker) | |
152 | { | |
153 | if (refcount_dec_and_test(&worker->ref)) | |
eb2de941 | 154 | complete(&worker->ref_done); |
771b53d0 JA |
155 | } |
156 | ||
8418f22a PB |
157 | static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound) |
158 | { | |
159 | return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND]; | |
160 | } | |
161 | ||
c5def4ab JA |
162 | static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe, |
163 | struct io_wq_work *work) | |
164 | { | |
8418f22a | 165 | return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND)); |
c5def4ab JA |
166 | } |
167 | ||
958234d5 | 168 | static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker) |
c5def4ab | 169 | { |
8418f22a | 170 | return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND); |
c5def4ab JA |
171 | } |
172 | ||
685fe7fe JA |
173 | static void io_worker_ref_put(struct io_wq *wq) |
174 | { | |
175 | if (atomic_dec_and_test(&wq->worker_refs)) | |
176 | complete(&wq->worker_done); | |
177 | } | |
178 | ||
1a2b192e PB |
179 | static void io_worker_cancel_cb(struct io_worker *worker) |
180 | { | |
181 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); | |
182 | struct io_wqe *wqe = worker->wqe; | |
183 | struct io_wq *wq = wqe->wq; | |
184 | ||
185 | atomic_dec(&acct->nr_running); | |
186 | raw_spin_lock(&worker->wqe->lock); | |
187 | acct->nr_workers--; | |
188 | raw_spin_unlock(&worker->wqe->lock); | |
189 | io_worker_ref_put(wq); | |
190 | clear_bit_unlock(0, &worker->create_state); | |
191 | io_worker_release(worker); | |
192 | } | |
193 | ||
194 | static bool io_task_worker_match(struct callback_head *cb, void *data) | |
195 | { | |
196 | struct io_worker *worker; | |
197 | ||
198 | if (cb->func != create_worker_cb) | |
199 | return false; | |
200 | worker = container_of(cb, struct io_worker, create_work); | |
201 | return worker == data; | |
202 | } | |
203 | ||
771b53d0 JA |
204 | static void io_worker_exit(struct io_worker *worker) |
205 | { | |
206 | struct io_wqe *wqe = worker->wqe; | |
1a2b192e PB |
207 | struct io_wq *wq = wqe->wq; |
208 | ||
209 | while (1) { | |
210 | struct callback_head *cb = task_work_cancel_match(wq->task, | |
211 | io_task_worker_match, worker); | |
212 | ||
213 | if (!cb) | |
214 | break; | |
215 | io_worker_cancel_cb(worker); | |
216 | } | |
771b53d0 | 217 | |
eb2de941 JA |
218 | if (refcount_dec_and_test(&worker->ref)) |
219 | complete(&worker->ref_done); | |
220 | wait_for_completion(&worker->ref_done); | |
771b53d0 | 221 | |
a9a4aa9f | 222 | raw_spin_lock(&wqe->lock); |
83d6c393 | 223 | if (worker->flags & IO_WORKER_F_FREE) |
bf1daa4b | 224 | hlist_nulls_del_rcu(&worker->nulls_node); |
e61df66c | 225 | list_del_rcu(&worker->all_list); |
83d6c393 JA |
226 | preempt_disable(); |
227 | io_wqe_dec_running(worker); | |
228 | worker->flags = 0; | |
229 | current->flags &= ~PF_IO_WORKER; | |
230 | preempt_enable(); | |
a9a4aa9f | 231 | raw_spin_unlock(&wqe->lock); |
771b53d0 | 232 | |
364b05fd | 233 | kfree_rcu(worker, rcu); |
685fe7fe | 234 | io_worker_ref_put(wqe->wq); |
46fe18b1 | 235 | do_exit(0); |
771b53d0 JA |
236 | } |
237 | ||
f95dc207 | 238 | static inline bool io_acct_run_queue(struct io_wqe_acct *acct) |
c5def4ab | 239 | { |
f95dc207 JA |
240 | if (!wq_list_empty(&acct->work_list) && |
241 | !test_bit(IO_ACCT_STALLED_BIT, &acct->flags)) | |
c5def4ab JA |
242 | return true; |
243 | return false; | |
244 | } | |
245 | ||
246 | /* | |
247 | * Check head of free list for an available worker. If one isn't available, | |
685fe7fe | 248 | * caller must create one. |
c5def4ab | 249 | */ |
f95dc207 JA |
250 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe, |
251 | struct io_wqe_acct *acct) | |
c5def4ab JA |
252 | __must_hold(RCU) |
253 | { | |
254 | struct hlist_nulls_node *n; | |
255 | struct io_worker *worker; | |
256 | ||
83d6c393 JA |
257 | /* |
258 | * Iterate free_list and see if we can find an idle worker to | |
259 | * activate. If a given worker is on the free_list but in the process | |
260 | * of exiting, keep trying. | |
261 | */ | |
262 | hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) { | |
263 | if (!io_worker_get(worker)) | |
264 | continue; | |
f95dc207 JA |
265 | if (io_wqe_get_acct(worker) != acct) { |
266 | io_worker_release(worker); | |
267 | continue; | |
268 | } | |
83d6c393 JA |
269 | if (wake_up_process(worker->task)) { |
270 | io_worker_release(worker); | |
271 | return true; | |
272 | } | |
c5def4ab | 273 | io_worker_release(worker); |
c5def4ab JA |
274 | } |
275 | ||
276 | return false; | |
277 | } | |
278 | ||
279 | /* | |
280 | * We need a worker. If we find a free one, we're good. If not, and we're | |
685fe7fe | 281 | * below the max number of workers, create one. |
c5def4ab | 282 | */ |
3146cba9 | 283 | static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct) |
c5def4ab | 284 | { |
c5def4ab JA |
285 | /* |
286 | * Most likely an attempt to queue unbounded work on an io_wq that | |
287 | * wasn't setup with any unbounded workers. | |
288 | */ | |
e6ab8991 PB |
289 | if (unlikely(!acct->max_workers)) |
290 | pr_warn_once("io-wq is not configured for unbound workers"); | |
c5def4ab | 291 | |
94ffb0a2 | 292 | raw_spin_lock(&wqe->lock); |
bc369921 | 293 | if (acct->nr_workers >= acct->max_workers) { |
7a842fb5 HX |
294 | raw_spin_unlock(&wqe->lock); |
295 | return true; | |
94ffb0a2 | 296 | } |
7a842fb5 | 297 | acct->nr_workers++; |
94ffb0a2 | 298 | raw_spin_unlock(&wqe->lock); |
7a842fb5 HX |
299 | atomic_inc(&acct->nr_running); |
300 | atomic_inc(&wqe->wq->worker_refs); | |
301 | return create_io_worker(wqe->wq, wqe, acct->index); | |
c5def4ab JA |
302 | } |
303 | ||
958234d5 | 304 | static void io_wqe_inc_running(struct io_worker *worker) |
c5def4ab | 305 | { |
958234d5 | 306 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
c5def4ab JA |
307 | |
308 | atomic_inc(&acct->nr_running); | |
309 | } | |
310 | ||
685fe7fe JA |
311 | static void create_worker_cb(struct callback_head *cb) |
312 | { | |
d3e9f732 | 313 | struct io_worker *worker; |
685fe7fe | 314 | struct io_wq *wq; |
21698274 HX |
315 | struct io_wqe *wqe; |
316 | struct io_wqe_acct *acct; | |
05c5f4ee | 317 | bool do_create = false; |
685fe7fe | 318 | |
d3e9f732 JA |
319 | worker = container_of(cb, struct io_worker, create_work); |
320 | wqe = worker->wqe; | |
21698274 | 321 | wq = wqe->wq; |
d3e9f732 | 322 | acct = &wqe->acct[worker->create_index]; |
a9a4aa9f | 323 | raw_spin_lock(&wqe->lock); |
49e7f0c7 | 324 | if (acct->nr_workers < acct->max_workers) { |
21698274 | 325 | acct->nr_workers++; |
49e7f0c7 HX |
326 | do_create = true; |
327 | } | |
a9a4aa9f | 328 | raw_spin_unlock(&wqe->lock); |
49e7f0c7 | 329 | if (do_create) { |
05c5f4ee | 330 | create_io_worker(wq, wqe, worker->create_index); |
49e7f0c7 HX |
331 | } else { |
332 | atomic_dec(&acct->nr_running); | |
333 | io_worker_ref_put(wq); | |
334 | } | |
d3e9f732 JA |
335 | clear_bit_unlock(0, &worker->create_state); |
336 | io_worker_release(worker); | |
685fe7fe JA |
337 | } |
338 | ||
3146cba9 JA |
339 | static bool io_queue_worker_create(struct io_worker *worker, |
340 | struct io_wqe_acct *acct, | |
341 | task_work_func_t func) | |
685fe7fe | 342 | { |
3146cba9 | 343 | struct io_wqe *wqe = worker->wqe; |
685fe7fe JA |
344 | struct io_wq *wq = wqe->wq; |
345 | ||
346 | /* raced with exit, just ignore create call */ | |
347 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) | |
348 | goto fail; | |
d3e9f732 JA |
349 | if (!io_worker_get(worker)) |
350 | goto fail; | |
351 | /* | |
352 | * create_state manages ownership of create_work/index. We should | |
353 | * only need one entry per worker, as the worker going to sleep | |
354 | * will trigger the condition, and waking will clear it once it | |
355 | * runs the task_work. | |
356 | */ | |
357 | if (test_bit(0, &worker->create_state) || | |
358 | test_and_set_bit_lock(0, &worker->create_state)) | |
359 | goto fail_release; | |
685fe7fe | 360 | |
d86fed57 | 361 | atomic_inc(&wq->worker_refs); |
3146cba9 | 362 | init_task_work(&worker->create_work, func); |
d3e9f732 | 363 | worker->create_index = acct->index; |
d86fed57 JA |
364 | if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) { |
365 | /* | |
366 | * EXIT may have been set after checking it above, check after | |
367 | * adding the task_work and remove any creation item if it is | |
368 | * now set. wq exit does that too, but we can have added this | |
369 | * work item after we canceled in io_wq_exit_workers(). | |
370 | */ | |
371 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) | |
372 | io_wq_cancel_tw_create(wq); | |
373 | io_worker_ref_put(wq); | |
3146cba9 | 374 | return true; |
d86fed57 JA |
375 | } |
376 | io_worker_ref_put(wq); | |
d3e9f732 JA |
377 | clear_bit_unlock(0, &worker->create_state); |
378 | fail_release: | |
379 | io_worker_release(worker); | |
685fe7fe JA |
380 | fail: |
381 | atomic_dec(&acct->nr_running); | |
382 | io_worker_ref_put(wq); | |
3146cba9 | 383 | return false; |
685fe7fe JA |
384 | } |
385 | ||
958234d5 | 386 | static void io_wqe_dec_running(struct io_worker *worker) |
c5def4ab JA |
387 | __must_hold(wqe->lock) |
388 | { | |
958234d5 JA |
389 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
390 | struct io_wqe *wqe = worker->wqe; | |
c5def4ab | 391 | |
685fe7fe JA |
392 | if (!(worker->flags & IO_WORKER_F_UP)) |
393 | return; | |
394 | ||
f95dc207 | 395 | if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) { |
685fe7fe JA |
396 | atomic_inc(&acct->nr_running); |
397 | atomic_inc(&wqe->wq->worker_refs); | |
63414a5b | 398 | raw_spin_unlock(&wqe->lock); |
3146cba9 | 399 | io_queue_worker_create(worker, acct, create_worker_cb); |
63414a5b | 400 | raw_spin_lock(&wqe->lock); |
685fe7fe | 401 | } |
c5def4ab JA |
402 | } |
403 | ||
771b53d0 JA |
404 | /* |
405 | * Worker will start processing some work. Move it to the busy list, if | |
406 | * it's currently on the freelist | |
407 | */ | |
408 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, | |
409 | struct io_wq_work *work) | |
410 | __must_hold(wqe->lock) | |
411 | { | |
412 | if (worker->flags & IO_WORKER_F_FREE) { | |
413 | worker->flags &= ~IO_WORKER_F_FREE; | |
414 | hlist_nulls_del_init_rcu(&worker->nulls_node); | |
771b53d0 | 415 | } |
771b53d0 JA |
416 | } |
417 | ||
418 | /* | |
419 | * No work, worker going to sleep. Move to freelist, and unuse mm if we | |
420 | * have one attached. Dropping the mm may potentially sleep, so we drop | |
421 | * the lock in that case and return success. Since the caller has to | |
422 | * retry the loop in that case (we changed task state), we don't regrab | |
423 | * the lock if we return success. | |
424 | */ | |
c6d77d92 | 425 | static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) |
771b53d0 JA |
426 | __must_hold(wqe->lock) |
427 | { | |
428 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
429 | worker->flags |= IO_WORKER_F_FREE; | |
021d1cdd | 430 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); |
771b53d0 | 431 | } |
771b53d0 JA |
432 | } |
433 | ||
60cf46ae PB |
434 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
435 | { | |
436 | return work->flags >> IO_WQ_HASH_SHIFT; | |
437 | } | |
438 | ||
0e199d11 | 439 | static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash) |
e941894e JA |
440 | { |
441 | struct io_wq *wq = wqe->wq; | |
0e199d11 | 442 | bool ret = false; |
e941894e | 443 | |
08bdbd39 | 444 | spin_lock_irq(&wq->hash->wait.lock); |
e941894e JA |
445 | if (list_empty(&wqe->wait.entry)) { |
446 | __add_wait_queue(&wq->hash->wait, &wqe->wait); | |
447 | if (!test_bit(hash, &wq->hash->map)) { | |
448 | __set_current_state(TASK_RUNNING); | |
449 | list_del_init(&wqe->wait.entry); | |
0e199d11 | 450 | ret = true; |
e941894e JA |
451 | } |
452 | } | |
08bdbd39 | 453 | spin_unlock_irq(&wq->hash->wait.lock); |
0e199d11 | 454 | return ret; |
e941894e JA |
455 | } |
456 | ||
f95dc207 | 457 | static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct, |
0242f642 | 458 | struct io_worker *worker) |
771b53d0 JA |
459 | __must_hold(wqe->lock) |
460 | { | |
6206f0e1 | 461 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 462 | struct io_wq_work *work, *tail; |
e941894e | 463 | unsigned int stall_hash = -1U; |
f95dc207 | 464 | struct io_wqe *wqe = worker->wqe; |
771b53d0 | 465 | |
f95dc207 | 466 | wq_list_for_each(node, prev, &acct->work_list) { |
e941894e JA |
467 | unsigned int hash; |
468 | ||
6206f0e1 JA |
469 | work = container_of(node, struct io_wq_work, list); |
470 | ||
771b53d0 | 471 | /* not hashed, can run anytime */ |
8766dd51 | 472 | if (!io_wq_is_hashed(work)) { |
f95dc207 | 473 | wq_list_del(&acct->work_list, node, prev); |
771b53d0 JA |
474 | return work; |
475 | } | |
476 | ||
60cf46ae | 477 | hash = io_get_work_hash(work); |
e941894e JA |
478 | /* all items with this hash lie in [work, tail] */ |
479 | tail = wqe->hash_tail[hash]; | |
480 | ||
481 | /* hashed, can run if not already running */ | |
482 | if (!test_and_set_bit(hash, &wqe->wq->hash->map)) { | |
86f3cd1b | 483 | wqe->hash_tail[hash] = NULL; |
f95dc207 | 484 | wq_list_cut(&acct->work_list, &tail->list, prev); |
771b53d0 JA |
485 | return work; |
486 | } | |
e941894e JA |
487 | if (stall_hash == -1U) |
488 | stall_hash = hash; | |
489 | /* fast forward to a next hash, for-each will fix up @prev */ | |
490 | node = &tail->list; | |
491 | } | |
492 | ||
493 | if (stall_hash != -1U) { | |
0e199d11 JA |
494 | bool unstalled; |
495 | ||
0242f642 JA |
496 | /* |
497 | * Set this before dropping the lock to avoid racing with new | |
498 | * work being added and clearing the stalled bit. | |
499 | */ | |
f95dc207 | 500 | set_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
e941894e | 501 | raw_spin_unlock(&wqe->lock); |
0e199d11 | 502 | unstalled = io_wait_on_hash(wqe, stall_hash); |
e941894e | 503 | raw_spin_lock(&wqe->lock); |
0e199d11 JA |
504 | if (unstalled) { |
505 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); | |
506 | if (wq_has_sleeper(&wqe->wq->hash->wait)) | |
507 | wake_up(&wqe->wq->hash->wait); | |
508 | } | |
771b53d0 JA |
509 | } |
510 | ||
511 | return NULL; | |
512 | } | |
513 | ||
00ddff43 | 514 | static bool io_flush_signals(void) |
cccf0ee8 | 515 | { |
0b8cfa97 | 516 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) { |
00ddff43 | 517 | __set_current_state(TASK_RUNNING); |
0b8cfa97 | 518 | tracehook_notify_signal(); |
00ddff43 | 519 | return true; |
cccf0ee8 | 520 | } |
00ddff43 | 521 | return false; |
dc026a73 PB |
522 | } |
523 | ||
524 | static void io_assign_current_work(struct io_worker *worker, | |
525 | struct io_wq_work *work) | |
526 | { | |
d78298e7 | 527 | if (work) { |
3bfe6106 | 528 | io_flush_signals(); |
d78298e7 PB |
529 | cond_resched(); |
530 | } | |
dc026a73 | 531 | |
a9a4aa9f | 532 | spin_lock(&worker->lock); |
dc026a73 | 533 | worker->cur_work = work; |
a9a4aa9f | 534 | spin_unlock(&worker->lock); |
dc026a73 PB |
535 | } |
536 | ||
60cf46ae PB |
537 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work); |
538 | ||
771b53d0 JA |
539 | static void io_worker_handle_work(struct io_worker *worker) |
540 | __releases(wqe->lock) | |
541 | { | |
f95dc207 | 542 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
771b53d0 JA |
543 | struct io_wqe *wqe = worker->wqe; |
544 | struct io_wq *wq = wqe->wq; | |
c60eb049 | 545 | bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state); |
771b53d0 JA |
546 | |
547 | do { | |
86f3cd1b | 548 | struct io_wq_work *work; |
f462fd36 | 549 | get_next: |
771b53d0 JA |
550 | /* |
551 | * If we got some work, mark us as busy. If we didn't, but | |
552 | * the list isn't empty, it means we stalled on hashed work. | |
553 | * Mark us stalled so we don't keep looking for work when we | |
554 | * can't make progress, any work completion or insertion will | |
555 | * clear the stalled flag. | |
556 | */ | |
f95dc207 | 557 | work = io_get_next_work(acct, worker); |
771b53d0 JA |
558 | if (work) |
559 | __io_worker_busy(wqe, worker, work); | |
771b53d0 | 560 | |
a9a4aa9f | 561 | raw_spin_unlock(&wqe->lock); |
771b53d0 JA |
562 | if (!work) |
563 | break; | |
58e39319 | 564 | io_assign_current_work(worker, work); |
e941894e | 565 | __set_current_state(TASK_RUNNING); |
36c2f922 | 566 | |
dc026a73 PB |
567 | /* handle a whole dependent link */ |
568 | do { | |
5280f7e5 | 569 | struct io_wq_work *next_hashed, *linked; |
b089ed39 | 570 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 571 | |
86f3cd1b | 572 | next_hashed = wq_next_work(work); |
c60eb049 PB |
573 | |
574 | if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND)) | |
575 | work->flags |= IO_WQ_WORK_CANCEL; | |
5280f7e5 PB |
576 | wq->do_work(work); |
577 | io_assign_current_work(worker, NULL); | |
dc026a73 | 578 | |
5280f7e5 | 579 | linked = wq->free_work(work); |
86f3cd1b PB |
580 | work = next_hashed; |
581 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
582 | work = linked; | |
583 | linked = NULL; | |
584 | } | |
585 | io_assign_current_work(worker, work); | |
86f3cd1b PB |
586 | if (linked) |
587 | io_wqe_enqueue(wqe, linked); | |
588 | ||
589 | if (hash != -1U && !next_hashed) { | |
0e199d11 JA |
590 | /* serialize hash clear with wake_up() */ |
591 | spin_lock_irq(&wq->hash->wait.lock); | |
e941894e | 592 | clear_bit(hash, &wq->hash->map); |
f95dc207 | 593 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
0e199d11 | 594 | spin_unlock_irq(&wq->hash->wait.lock); |
e941894e JA |
595 | if (wq_has_sleeper(&wq->hash->wait)) |
596 | wake_up(&wq->hash->wait); | |
a9a4aa9f | 597 | raw_spin_lock(&wqe->lock); |
f462fd36 PB |
598 | /* skip unnecessary unlock-lock wqe->lock */ |
599 | if (!work) | |
600 | goto get_next; | |
a9a4aa9f | 601 | raw_spin_unlock(&wqe->lock); |
7d723065 | 602 | } |
58e39319 | 603 | } while (work); |
7d723065 | 604 | |
a9a4aa9f | 605 | raw_spin_lock(&wqe->lock); |
771b53d0 JA |
606 | } while (1); |
607 | } | |
608 | ||
771b53d0 JA |
609 | static int io_wqe_worker(void *data) |
610 | { | |
611 | struct io_worker *worker = data; | |
f95dc207 | 612 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); |
771b53d0 JA |
613 | struct io_wqe *wqe = worker->wqe; |
614 | struct io_wq *wq = wqe->wq; | |
05c5f4ee | 615 | bool last_timeout = false; |
46fe18b1 | 616 | char buf[TASK_COMM_LEN]; |
771b53d0 | 617 | |
46fe18b1 | 618 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); |
46fe18b1 | 619 | |
685fe7fe | 620 | snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid); |
46fe18b1 | 621 | set_task_comm(current, buf); |
771b53d0 JA |
622 | |
623 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
16efa4fc JA |
624 | long ret; |
625 | ||
506d95ff | 626 | set_current_state(TASK_INTERRUPTIBLE); |
e995d512 | 627 | loop: |
a9a4aa9f | 628 | raw_spin_lock(&wqe->lock); |
f95dc207 | 629 | if (io_acct_run_queue(acct)) { |
771b53d0 | 630 | io_worker_handle_work(worker); |
e995d512 | 631 | goto loop; |
771b53d0 | 632 | } |
05c5f4ee JA |
633 | /* timed out, exit unless we're the last worker */ |
634 | if (last_timeout && acct->nr_workers > 1) { | |
767a65e9 | 635 | acct->nr_workers--; |
05c5f4ee JA |
636 | raw_spin_unlock(&wqe->lock); |
637 | __set_current_state(TASK_RUNNING); | |
638 | break; | |
639 | } | |
640 | last_timeout = false; | |
c6d77d92 | 641 | __io_worker_idle(wqe, worker); |
a9a4aa9f | 642 | raw_spin_unlock(&wqe->lock); |
00ddff43 JA |
643 | if (io_flush_signals()) |
644 | continue; | |
16efa4fc | 645 | ret = schedule_timeout(WORKER_IDLE_TIMEOUT); |
dbe1bdbb JA |
646 | if (signal_pending(current)) { |
647 | struct ksignal ksig; | |
648 | ||
649 | if (!get_signal(&ksig)) | |
650 | continue; | |
78f8876c | 651 | break; |
dbe1bdbb | 652 | } |
05c5f4ee | 653 | last_timeout = !ret; |
771b53d0 JA |
654 | } |
655 | ||
771b53d0 | 656 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
a9a4aa9f | 657 | raw_spin_lock(&wqe->lock); |
e587227b | 658 | io_worker_handle_work(worker); |
771b53d0 JA |
659 | } |
660 | ||
661 | io_worker_exit(worker); | |
662 | return 0; | |
663 | } | |
664 | ||
771b53d0 JA |
665 | /* |
666 | * Called when a worker is scheduled in. Mark us as currently running. | |
667 | */ | |
668 | void io_wq_worker_running(struct task_struct *tsk) | |
669 | { | |
3bfe6106 | 670 | struct io_worker *worker = tsk->pf_io_worker; |
771b53d0 | 671 | |
3bfe6106 JA |
672 | if (!worker) |
673 | return; | |
771b53d0 JA |
674 | if (!(worker->flags & IO_WORKER_F_UP)) |
675 | return; | |
676 | if (worker->flags & IO_WORKER_F_RUNNING) | |
677 | return; | |
678 | worker->flags |= IO_WORKER_F_RUNNING; | |
958234d5 | 679 | io_wqe_inc_running(worker); |
771b53d0 JA |
680 | } |
681 | ||
682 | /* | |
683 | * Called when worker is going to sleep. If there are no workers currently | |
685fe7fe | 684 | * running and we have work pending, wake up a free one or create a new one. |
771b53d0 JA |
685 | */ |
686 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
687 | { | |
3bfe6106 | 688 | struct io_worker *worker = tsk->pf_io_worker; |
771b53d0 | 689 | |
3bfe6106 JA |
690 | if (!worker) |
691 | return; | |
771b53d0 JA |
692 | if (!(worker->flags & IO_WORKER_F_UP)) |
693 | return; | |
694 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
695 | return; | |
696 | ||
697 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
698 | ||
a9a4aa9f | 699 | raw_spin_lock(&worker->wqe->lock); |
958234d5 | 700 | io_wqe_dec_running(worker); |
a9a4aa9f | 701 | raw_spin_unlock(&worker->wqe->lock); |
771b53d0 JA |
702 | } |
703 | ||
3146cba9 JA |
704 | static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker, |
705 | struct task_struct *tsk) | |
706 | { | |
707 | tsk->pf_io_worker = worker; | |
708 | worker->task = tsk; | |
709 | set_cpus_allowed_ptr(tsk, wqe->cpu_mask); | |
710 | tsk->flags |= PF_NO_SETAFFINITY; | |
711 | ||
712 | raw_spin_lock(&wqe->lock); | |
713 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list); | |
714 | list_add_tail_rcu(&worker->all_list, &wqe->all_list); | |
715 | worker->flags |= IO_WORKER_F_FREE; | |
716 | raw_spin_unlock(&wqe->lock); | |
717 | wake_up_new_task(tsk); | |
718 | } | |
719 | ||
720 | static bool io_wq_work_match_all(struct io_wq_work *work, void *data) | |
721 | { | |
722 | return true; | |
723 | } | |
724 | ||
725 | static inline bool io_should_retry_thread(long err) | |
726 | { | |
b2fa968d JA |
727 | /* |
728 | * Prevent perpetual task_work retry, if the task (or its group) is | |
729 | * exiting. | |
730 | */ | |
731 | if (fatal_signal_pending(current)) | |
732 | return false; | |
733 | ||
3146cba9 JA |
734 | switch (err) { |
735 | case -EAGAIN: | |
736 | case -ERESTARTSYS: | |
737 | case -ERESTARTNOINTR: | |
738 | case -ERESTARTNOHAND: | |
739 | return true; | |
740 | default: | |
741 | return false; | |
742 | } | |
743 | } | |
744 | ||
745 | static void create_worker_cont(struct callback_head *cb) | |
746 | { | |
747 | struct io_worker *worker; | |
748 | struct task_struct *tsk; | |
749 | struct io_wqe *wqe; | |
750 | ||
751 | worker = container_of(cb, struct io_worker, create_work); | |
752 | clear_bit_unlock(0, &worker->create_state); | |
753 | wqe = worker->wqe; | |
754 | tsk = create_io_thread(io_wqe_worker, worker, wqe->node); | |
755 | if (!IS_ERR(tsk)) { | |
756 | io_init_new_worker(wqe, worker, tsk); | |
757 | io_worker_release(worker); | |
758 | return; | |
759 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { | |
760 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); | |
761 | ||
762 | atomic_dec(&acct->nr_running); | |
763 | raw_spin_lock(&wqe->lock); | |
764 | acct->nr_workers--; | |
765 | if (!acct->nr_workers) { | |
766 | struct io_cb_cancel_data match = { | |
767 | .fn = io_wq_work_match_all, | |
768 | .cancel_all = true, | |
769 | }; | |
770 | ||
771 | while (io_acct_cancel_pending_work(wqe, acct, &match)) | |
772 | raw_spin_lock(&wqe->lock); | |
773 | } | |
774 | raw_spin_unlock(&wqe->lock); | |
775 | io_worker_ref_put(wqe->wq); | |
66e70be7 | 776 | kfree(worker); |
3146cba9 JA |
777 | return; |
778 | } | |
779 | ||
780 | /* re-create attempts grab a new worker ref, drop the existing one */ | |
781 | io_worker_release(worker); | |
782 | schedule_work(&worker->work); | |
783 | } | |
784 | ||
785 | static void io_workqueue_create(struct work_struct *work) | |
786 | { | |
787 | struct io_worker *worker = container_of(work, struct io_worker, work); | |
788 | struct io_wqe_acct *acct = io_wqe_get_acct(worker); | |
789 | ||
b9245d3a | 790 | if (!io_queue_worker_create(worker, acct, create_worker_cont)) |
66e70be7 | 791 | kfree(worker); |
3146cba9 JA |
792 | } |
793 | ||
794 | static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index) | |
3bfe6106 | 795 | { |
46fe18b1 | 796 | struct io_wqe_acct *acct = &wqe->acct[index]; |
3bfe6106 | 797 | struct io_worker *worker; |
46fe18b1 | 798 | struct task_struct *tsk; |
3bfe6106 | 799 | |
8b3e78b5 JA |
800 | __set_current_state(TASK_RUNNING); |
801 | ||
3bfe6106 | 802 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node); |
3146cba9 | 803 | if (!worker) { |
685fe7fe JA |
804 | fail: |
805 | atomic_dec(&acct->nr_running); | |
a9a4aa9f | 806 | raw_spin_lock(&wqe->lock); |
3d4e4fac | 807 | acct->nr_workers--; |
a9a4aa9f | 808 | raw_spin_unlock(&wqe->lock); |
685fe7fe | 809 | io_worker_ref_put(wq); |
3146cba9 | 810 | return false; |
3bfe6106 | 811 | } |
46fe18b1 | 812 | |
3146cba9 JA |
813 | refcount_set(&worker->ref, 1); |
814 | worker->wqe = wqe; | |
815 | spin_lock_init(&worker->lock); | |
816 | init_completion(&worker->ref_done); | |
46fe18b1 | 817 | |
46fe18b1 JA |
818 | if (index == IO_WQ_ACCT_BOUND) |
819 | worker->flags |= IO_WORKER_F_BOUND; | |
3146cba9 JA |
820 | |
821 | tsk = create_io_thread(io_wqe_worker, worker, wqe->node); | |
822 | if (!IS_ERR(tsk)) { | |
823 | io_init_new_worker(wqe, worker, tsk); | |
824 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { | |
66e70be7 | 825 | kfree(worker); |
3146cba9 JA |
826 | goto fail; |
827 | } else { | |
828 | INIT_WORK(&worker->work, io_workqueue_create); | |
829 | schedule_work(&worker->work); | |
830 | } | |
831 | ||
832 | return true; | |
771b53d0 JA |
833 | } |
834 | ||
c4068bf8 HD |
835 | /* |
836 | * Iterate the passed in list and call the specific function for each | |
837 | * worker that isn't exiting | |
838 | */ | |
839 | static bool io_wq_for_each_worker(struct io_wqe *wqe, | |
840 | bool (*func)(struct io_worker *, void *), | |
841 | void *data) | |
842 | { | |
843 | struct io_worker *worker; | |
844 | bool ret = false; | |
845 | ||
846 | list_for_each_entry_rcu(worker, &wqe->all_list, all_list) { | |
847 | if (io_worker_get(worker)) { | |
848 | /* no task if node is/was offline */ | |
849 | if (worker->task) | |
850 | ret = func(worker, data); | |
851 | io_worker_release(worker); | |
852 | if (ret) | |
853 | break; | |
854 | } | |
855 | } | |
856 | ||
857 | return ret; | |
858 | } | |
859 | ||
860 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
861 | { | |
46fe18b1 | 862 | set_notify_signal(worker->task); |
c4068bf8 HD |
863 | wake_up_process(worker->task); |
864 | return false; | |
865 | } | |
866 | ||
e9fd9396 | 867 | static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe) |
fc04c39b | 868 | { |
e9fd9396 PB |
869 | struct io_wq *wq = wqe->wq; |
870 | ||
fc04c39b | 871 | do { |
fc04c39b | 872 | work->flags |= IO_WQ_WORK_CANCEL; |
5280f7e5 PB |
873 | wq->do_work(work); |
874 | work = wq->free_work(work); | |
fc04c39b PB |
875 | } while (work); |
876 | } | |
877 | ||
86f3cd1b PB |
878 | static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work) |
879 | { | |
f95dc207 | 880 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
86f3cd1b PB |
881 | unsigned int hash; |
882 | struct io_wq_work *tail; | |
883 | ||
884 | if (!io_wq_is_hashed(work)) { | |
885 | append: | |
f95dc207 | 886 | wq_list_add_tail(&work->list, &acct->work_list); |
86f3cd1b PB |
887 | return; |
888 | } | |
889 | ||
890 | hash = io_get_work_hash(work); | |
891 | tail = wqe->hash_tail[hash]; | |
892 | wqe->hash_tail[hash] = work; | |
893 | if (!tail) | |
894 | goto append; | |
895 | ||
f95dc207 | 896 | wq_list_add_after(&work->list, &tail->list, &acct->work_list); |
86f3cd1b PB |
897 | } |
898 | ||
713b9825 PB |
899 | static bool io_wq_work_match_item(struct io_wq_work *work, void *data) |
900 | { | |
901 | return work == data; | |
902 | } | |
903 | ||
771b53d0 JA |
904 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
905 | { | |
c5def4ab | 906 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
94ffb0a2 JA |
907 | unsigned work_flags = work->flags; |
908 | bool do_create; | |
771b53d0 | 909 | |
991468dc JA |
910 | /* |
911 | * If io-wq is exiting for this task, or if the request has explicitly | |
912 | * been marked as one that should not get executed, cancel it here. | |
913 | */ | |
914 | if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) || | |
915 | (work->flags & IO_WQ_WORK_CANCEL)) { | |
70e35125 | 916 | io_run_cancel(work, wqe); |
4fb6ac32 JA |
917 | return; |
918 | } | |
919 | ||
a9a4aa9f | 920 | raw_spin_lock(&wqe->lock); |
86f3cd1b | 921 | io_wqe_insert_work(wqe, work); |
f95dc207 | 922 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
94ffb0a2 JA |
923 | |
924 | rcu_read_lock(); | |
f95dc207 | 925 | do_create = !io_wqe_activate_free_worker(wqe, acct); |
94ffb0a2 JA |
926 | rcu_read_unlock(); |
927 | ||
a9a4aa9f | 928 | raw_spin_unlock(&wqe->lock); |
771b53d0 | 929 | |
94ffb0a2 | 930 | if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) || |
3146cba9 JA |
931 | !atomic_read(&acct->nr_running))) { |
932 | bool did_create; | |
933 | ||
934 | did_create = io_wqe_create_worker(wqe, acct); | |
713b9825 PB |
935 | if (likely(did_create)) |
936 | return; | |
937 | ||
938 | raw_spin_lock(&wqe->lock); | |
939 | /* fatal condition, failed to create the first worker */ | |
940 | if (!acct->nr_workers) { | |
941 | struct io_cb_cancel_data match = { | |
942 | .fn = io_wq_work_match_item, | |
943 | .data = work, | |
944 | .cancel_all = false, | |
945 | }; | |
946 | ||
947 | if (io_acct_cancel_pending_work(wqe, acct, &match)) | |
948 | raw_spin_lock(&wqe->lock); | |
3146cba9 | 949 | } |
713b9825 | 950 | raw_spin_unlock(&wqe->lock); |
3146cba9 | 951 | } |
771b53d0 JA |
952 | } |
953 | ||
954 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) | |
955 | { | |
956 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; | |
957 | ||
958 | io_wqe_enqueue(wqe, work); | |
959 | } | |
960 | ||
961 | /* | |
8766dd51 PB |
962 | * Work items that hash to the same value will not be done in parallel. |
963 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 964 | */ |
8766dd51 | 965 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 966 | { |
8766dd51 | 967 | unsigned int bit; |
771b53d0 JA |
968 | |
969 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
970 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
971 | } |
972 | ||
2293b419 | 973 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 974 | { |
2293b419 | 975 | struct io_cb_cancel_data *match = data; |
62755e35 JA |
976 | |
977 | /* | |
978 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 979 | * may dereference the passed in work. |
62755e35 | 980 | */ |
a9a4aa9f | 981 | spin_lock(&worker->lock); |
62755e35 | 982 | if (worker->cur_work && |
2293b419 | 983 | match->fn(worker->cur_work, match->data)) { |
3bfe6106 | 984 | set_notify_signal(worker->task); |
4f26bda1 | 985 | match->nr_running++; |
771b53d0 | 986 | } |
a9a4aa9f | 987 | spin_unlock(&worker->lock); |
771b53d0 | 988 | |
4f26bda1 | 989 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
990 | } |
991 | ||
204361a7 PB |
992 | static inline void io_wqe_remove_pending(struct io_wqe *wqe, |
993 | struct io_wq_work *work, | |
994 | struct io_wq_work_node *prev) | |
995 | { | |
f95dc207 | 996 | struct io_wqe_acct *acct = io_work_get_acct(wqe, work); |
204361a7 PB |
997 | unsigned int hash = io_get_work_hash(work); |
998 | struct io_wq_work *prev_work = NULL; | |
999 | ||
1000 | if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) { | |
1001 | if (prev) | |
1002 | prev_work = container_of(prev, struct io_wq_work, list); | |
1003 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
1004 | wqe->hash_tail[hash] = prev_work; | |
1005 | else | |
1006 | wqe->hash_tail[hash] = NULL; | |
1007 | } | |
f95dc207 | 1008 | wq_list_del(&acct->work_list, &work->list, prev); |
204361a7 PB |
1009 | } |
1010 | ||
3146cba9 JA |
1011 | static bool io_acct_cancel_pending_work(struct io_wqe *wqe, |
1012 | struct io_wqe_acct *acct, | |
1013 | struct io_cb_cancel_data *match) | |
1014 | __releases(wqe->lock) | |
771b53d0 | 1015 | { |
6206f0e1 | 1016 | struct io_wq_work_node *node, *prev; |
771b53d0 | 1017 | struct io_wq_work *work; |
771b53d0 | 1018 | |
3146cba9 JA |
1019 | wq_list_for_each(node, prev, &acct->work_list) { |
1020 | work = container_of(node, struct io_wq_work, list); | |
1021 | if (!match->fn(work, match->data)) | |
1022 | continue; | |
1023 | io_wqe_remove_pending(wqe, work, prev); | |
1024 | raw_spin_unlock(&wqe->lock); | |
1025 | io_run_cancel(work, wqe); | |
1026 | match->nr_pending++; | |
1027 | /* not safe to continue after unlock */ | |
1028 | return true; | |
1029 | } | |
1030 | ||
1031 | return false; | |
1032 | } | |
1033 | ||
1034 | static void io_wqe_cancel_pending_work(struct io_wqe *wqe, | |
1035 | struct io_cb_cancel_data *match) | |
1036 | { | |
1037 | int i; | |
4f26bda1 | 1038 | retry: |
a9a4aa9f | 1039 | raw_spin_lock(&wqe->lock); |
f95dc207 JA |
1040 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
1041 | struct io_wqe_acct *acct = io_get_acct(wqe, i == 0); | |
4f26bda1 | 1042 | |
3146cba9 JA |
1043 | if (io_acct_cancel_pending_work(wqe, acct, match)) { |
1044 | if (match->cancel_all) | |
1045 | goto retry; | |
1046 | return; | |
f95dc207 | 1047 | } |
771b53d0 | 1048 | } |
a9a4aa9f | 1049 | raw_spin_unlock(&wqe->lock); |
f4c2665e PB |
1050 | } |
1051 | ||
4f26bda1 | 1052 | static void io_wqe_cancel_running_work(struct io_wqe *wqe, |
f4c2665e PB |
1053 | struct io_cb_cancel_data *match) |
1054 | { | |
771b53d0 | 1055 | rcu_read_lock(); |
4f26bda1 | 1056 | io_wq_for_each_worker(wqe, io_wq_worker_cancel, match); |
771b53d0 | 1057 | rcu_read_unlock(); |
771b53d0 JA |
1058 | } |
1059 | ||
2293b419 | 1060 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 1061 | void *data, bool cancel_all) |
771b53d0 | 1062 | { |
2293b419 | 1063 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
1064 | .fn = cancel, |
1065 | .data = data, | |
1066 | .cancel_all = cancel_all, | |
00bcda13 | 1067 | }; |
3fc50ab5 | 1068 | int node; |
771b53d0 | 1069 | |
f4c2665e PB |
1070 | /* |
1071 | * First check pending list, if we're lucky we can just remove it | |
1072 | * from there. CANCEL_OK means that the work is returned as-new, | |
1073 | * no completion will be posted for it. | |
1074 | */ | |
3fc50ab5 JH |
1075 | for_each_node(node) { |
1076 | struct io_wqe *wqe = wq->wqes[node]; | |
771b53d0 | 1077 | |
4f26bda1 PB |
1078 | io_wqe_cancel_pending_work(wqe, &match); |
1079 | if (match.nr_pending && !match.cancel_all) | |
f4c2665e | 1080 | return IO_WQ_CANCEL_OK; |
771b53d0 JA |
1081 | } |
1082 | ||
f4c2665e PB |
1083 | /* |
1084 | * Now check if a free (going busy) or busy worker has the work | |
1085 | * currently running. If we find it there, we'll return CANCEL_RUNNING | |
1086 | * as an indication that we attempt to signal cancellation. The | |
1087 | * completion will run normally in this case. | |
1088 | */ | |
1089 | for_each_node(node) { | |
1090 | struct io_wqe *wqe = wq->wqes[node]; | |
1091 | ||
4f26bda1 PB |
1092 | io_wqe_cancel_running_work(wqe, &match); |
1093 | if (match.nr_running && !match.cancel_all) | |
f4c2665e PB |
1094 | return IO_WQ_CANCEL_RUNNING; |
1095 | } | |
1096 | ||
4f26bda1 PB |
1097 | if (match.nr_running) |
1098 | return IO_WQ_CANCEL_RUNNING; | |
1099 | if (match.nr_pending) | |
1100 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 1101 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
1102 | } |
1103 | ||
e941894e JA |
1104 | static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode, |
1105 | int sync, void *key) | |
1106 | { | |
1107 | struct io_wqe *wqe = container_of(wait, struct io_wqe, wait); | |
f95dc207 | 1108 | int i; |
e941894e JA |
1109 | |
1110 | list_del_init(&wait->entry); | |
1111 | ||
1112 | rcu_read_lock(); | |
f95dc207 JA |
1113 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
1114 | struct io_wqe_acct *acct = &wqe->acct[i]; | |
1115 | ||
1116 | if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags)) | |
1117 | io_wqe_activate_free_worker(wqe, acct); | |
1118 | } | |
e941894e | 1119 | rcu_read_unlock(); |
e941894e JA |
1120 | return 1; |
1121 | } | |
1122 | ||
576a347b | 1123 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1124 | { |
f95dc207 | 1125 | int ret, node, i; |
771b53d0 JA |
1126 | struct io_wq *wq; |
1127 | ||
f5fa38c5 | 1128 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 | 1129 | return ERR_PTR(-EINVAL); |
e6ab8991 PB |
1130 | if (WARN_ON_ONCE(!bounded)) |
1131 | return ERR_PTR(-EINVAL); | |
e9fd9396 | 1132 | |
c7f405d6 | 1133 | wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL); |
771b53d0 JA |
1134 | if (!wq) |
1135 | return ERR_PTR(-ENOMEM); | |
43c01fbe JA |
1136 | ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
1137 | if (ret) | |
c7f405d6 | 1138 | goto err_wq; |
771b53d0 | 1139 | |
e941894e JA |
1140 | refcount_inc(&data->hash->refs); |
1141 | wq->hash = data->hash; | |
e9fd9396 | 1142 | wq->free_work = data->free_work; |
f5fa38c5 | 1143 | wq->do_work = data->do_work; |
7d723065 | 1144 | |
43c01fbe | 1145 | ret = -ENOMEM; |
3fc50ab5 | 1146 | for_each_node(node) { |
771b53d0 | 1147 | struct io_wqe *wqe; |
7563439a | 1148 | int alloc_node = node; |
771b53d0 | 1149 | |
7563439a JA |
1150 | if (!node_online(alloc_node)) |
1151 | alloc_node = NUMA_NO_NODE; | |
1152 | wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node); | |
771b53d0 | 1153 | if (!wqe) |
3fc50ab5 | 1154 | goto err; |
0e03496d JA |
1155 | if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL)) |
1156 | goto err; | |
1157 | cpumask_copy(wqe->cpu_mask, cpumask_of_node(node)); | |
3fc50ab5 | 1158 | wq->wqes[node] = wqe; |
7563439a | 1159 | wqe->node = alloc_node; |
c5def4ab | 1160 | wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
728f13e7 | 1161 | wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers = |
c5def4ab | 1162 | task_rlimit(current, RLIMIT_NPROC); |
e941894e | 1163 | INIT_LIST_HEAD(&wqe->wait.entry); |
f95dc207 JA |
1164 | wqe->wait.func = io_wqe_hash_wake; |
1165 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { | |
1166 | struct io_wqe_acct *acct = &wqe->acct[i]; | |
1167 | ||
1168 | acct->index = i; | |
1169 | atomic_set(&acct->nr_running, 0); | |
1170 | INIT_WQ_LIST(&acct->work_list); | |
1171 | } | |
771b53d0 | 1172 | wqe->wq = wq; |
95da8465 | 1173 | raw_spin_lock_init(&wqe->lock); |
021d1cdd | 1174 | INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0); |
e61df66c | 1175 | INIT_LIST_HEAD(&wqe->all_list); |
771b53d0 JA |
1176 | } |
1177 | ||
685fe7fe | 1178 | wq->task = get_task_struct(data->task); |
685fe7fe JA |
1179 | atomic_set(&wq->worker_refs, 1); |
1180 | init_completion(&wq->worker_done); | |
1181 | return wq; | |
b60fda60 | 1182 | err: |
dc7bbc9e | 1183 | io_wq_put_hash(data->hash); |
43c01fbe | 1184 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
0e03496d JA |
1185 | for_each_node(node) { |
1186 | if (!wq->wqes[node]) | |
1187 | continue; | |
1188 | free_cpumask_var(wq->wqes[node]->cpu_mask); | |
3fc50ab5 | 1189 | kfree(wq->wqes[node]); |
0e03496d | 1190 | } |
43c01fbe | 1191 | err_wq: |
b60fda60 | 1192 | kfree(wq); |
771b53d0 JA |
1193 | return ERR_PTR(ret); |
1194 | } | |
1195 | ||
c80ca470 JA |
1196 | static bool io_task_work_match(struct callback_head *cb, void *data) |
1197 | { | |
d3e9f732 | 1198 | struct io_worker *worker; |
c80ca470 | 1199 | |
3b33e3f4 | 1200 | if (cb->func != create_worker_cb && cb->func != create_worker_cont) |
c80ca470 | 1201 | return false; |
d3e9f732 JA |
1202 | worker = container_of(cb, struct io_worker, create_work); |
1203 | return worker->wqe->wq == data; | |
c80ca470 JA |
1204 | } |
1205 | ||
17a91051 PB |
1206 | void io_wq_exit_start(struct io_wq *wq) |
1207 | { | |
1208 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
1209 | } | |
1210 | ||
d86fed57 | 1211 | static void io_wq_cancel_tw_create(struct io_wq *wq) |
afcc4015 | 1212 | { |
685fe7fe | 1213 | struct callback_head *cb; |
685fe7fe | 1214 | |
c80ca470 | 1215 | while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) { |
d3e9f732 | 1216 | struct io_worker *worker; |
685fe7fe | 1217 | |
d3e9f732 | 1218 | worker = container_of(cb, struct io_worker, create_work); |
1a2b192e | 1219 | io_worker_cancel_cb(worker); |
685fe7fe | 1220 | } |
d86fed57 JA |
1221 | } |
1222 | ||
1223 | static void io_wq_exit_workers(struct io_wq *wq) | |
1224 | { | |
1225 | int node; | |
1226 | ||
1227 | if (!wq->task) | |
1228 | return; | |
1229 | ||
1230 | io_wq_cancel_tw_create(wq); | |
685fe7fe JA |
1231 | |
1232 | rcu_read_lock(); | |
1233 | for_each_node(node) { | |
1234 | struct io_wqe *wqe = wq->wqes[node]; | |
1235 | ||
1236 | io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL); | |
afcc4015 | 1237 | } |
685fe7fe JA |
1238 | rcu_read_unlock(); |
1239 | io_worker_ref_put(wq); | |
1240 | wait_for_completion(&wq->worker_done); | |
3743c172 Z |
1241 | |
1242 | for_each_node(node) { | |
1243 | spin_lock_irq(&wq->hash->wait.lock); | |
1244 | list_del_init(&wq->wqes[node]->wait.entry); | |
1245 | spin_unlock_irq(&wq->hash->wait.lock); | |
1246 | } | |
685fe7fe JA |
1247 | put_task_struct(wq->task); |
1248 | wq->task = NULL; | |
afcc4015 JA |
1249 | } |
1250 | ||
4fb6ac32 | 1251 | static void io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1252 | { |
3fc50ab5 | 1253 | int node; |
771b53d0 | 1254 | |
43c01fbe JA |
1255 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
1256 | ||
e941894e JA |
1257 | for_each_node(node) { |
1258 | struct io_wqe *wqe = wq->wqes[node]; | |
f5d2d23b JA |
1259 | struct io_cb_cancel_data match = { |
1260 | .fn = io_wq_work_match_all, | |
1261 | .cancel_all = true, | |
1262 | }; | |
1263 | io_wqe_cancel_pending_work(wqe, &match); | |
0e03496d | 1264 | free_cpumask_var(wqe->cpu_mask); |
e941894e JA |
1265 | kfree(wqe); |
1266 | } | |
e941894e | 1267 | io_wq_put_hash(wq->hash); |
771b53d0 | 1268 | kfree(wq); |
4fb6ac32 JA |
1269 | } |
1270 | ||
afcc4015 JA |
1271 | void io_wq_put_and_exit(struct io_wq *wq) |
1272 | { | |
17a91051 PB |
1273 | WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state)); |
1274 | ||
685fe7fe | 1275 | io_wq_exit_workers(wq); |
382cb030 | 1276 | io_wq_destroy(wq); |
afcc4015 JA |
1277 | } |
1278 | ||
0e03496d JA |
1279 | struct online_data { |
1280 | unsigned int cpu; | |
1281 | bool online; | |
1282 | }; | |
1283 | ||
43c01fbe JA |
1284 | static bool io_wq_worker_affinity(struct io_worker *worker, void *data) |
1285 | { | |
0e03496d | 1286 | struct online_data *od = data; |
e0051d7d | 1287 | |
0e03496d JA |
1288 | if (od->online) |
1289 | cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask); | |
1290 | else | |
1291 | cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask); | |
43c01fbe JA |
1292 | return false; |
1293 | } | |
1294 | ||
0e03496d | 1295 | static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online) |
43c01fbe | 1296 | { |
0e03496d JA |
1297 | struct online_data od = { |
1298 | .cpu = cpu, | |
1299 | .online = online | |
1300 | }; | |
43c01fbe JA |
1301 | int i; |
1302 | ||
1303 | rcu_read_lock(); | |
1304 | for_each_node(i) | |
0e03496d | 1305 | io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od); |
43c01fbe JA |
1306 | rcu_read_unlock(); |
1307 | return 0; | |
1308 | } | |
1309 | ||
0e03496d JA |
1310 | static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node) |
1311 | { | |
1312 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1313 | ||
1314 | return __io_wq_cpu_online(wq, cpu, true); | |
1315 | } | |
1316 | ||
1317 | static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node) | |
1318 | { | |
1319 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1320 | ||
1321 | return __io_wq_cpu_online(wq, cpu, false); | |
1322 | } | |
1323 | ||
fe76421d JA |
1324 | int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask) |
1325 | { | |
1326 | int i; | |
1327 | ||
1328 | rcu_read_lock(); | |
1329 | for_each_node(i) { | |
1330 | struct io_wqe *wqe = wq->wqes[i]; | |
1331 | ||
1332 | if (mask) | |
1333 | cpumask_copy(wqe->cpu_mask, mask); | |
1334 | else | |
1335 | cpumask_copy(wqe->cpu_mask, cpumask_of_node(i)); | |
1336 | } | |
1337 | rcu_read_unlock(); | |
1338 | return 0; | |
1339 | } | |
1340 | ||
2e480058 JA |
1341 | /* |
1342 | * Set max number of unbounded workers, returns old value. If new_count is 0, | |
1343 | * then just return the old value. | |
1344 | */ | |
1345 | int io_wq_max_workers(struct io_wq *wq, int *new_count) | |
1346 | { | |
56da89bd BZ |
1347 | int prev[IO_WQ_ACCT_NR]; |
1348 | bool first_node = true; | |
1349 | int i, node; | |
2e480058 | 1350 | |
dd47c104 ES |
1351 | BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND); |
1352 | BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND); | |
1353 | BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2); | |
1354 | ||
2e480058 JA |
1355 | for (i = 0; i < 2; i++) { |
1356 | if (new_count[i] > task_rlimit(current, RLIMIT_NPROC)) | |
1357 | new_count[i] = task_rlimit(current, RLIMIT_NPROC); | |
1358 | } | |
1359 | ||
56da89bd BZ |
1360 | for (i = 0; i < IO_WQ_ACCT_NR; i++) |
1361 | prev[i] = 0; | |
1362 | ||
2e480058 JA |
1363 | rcu_read_lock(); |
1364 | for_each_node(node) { | |
bc369921 | 1365 | struct io_wqe *wqe = wq->wqes[node]; |
2e480058 JA |
1366 | struct io_wqe_acct *acct; |
1367 | ||
bc369921 | 1368 | raw_spin_lock(&wqe->lock); |
f95dc207 | 1369 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
bc369921 | 1370 | acct = &wqe->acct[i]; |
56da89bd BZ |
1371 | if (first_node) |
1372 | prev[i] = max_t(int, acct->max_workers, prev[i]); | |
2e480058 JA |
1373 | if (new_count[i]) |
1374 | acct->max_workers = new_count[i]; | |
2e480058 | 1375 | } |
bc369921 | 1376 | raw_spin_unlock(&wqe->lock); |
56da89bd | 1377 | first_node = false; |
2e480058 JA |
1378 | } |
1379 | rcu_read_unlock(); | |
56da89bd BZ |
1380 | |
1381 | for (i = 0; i < IO_WQ_ACCT_NR; i++) | |
1382 | new_count[i] = prev[i]; | |
1383 | ||
2e480058 JA |
1384 | return 0; |
1385 | } | |
1386 | ||
43c01fbe JA |
1387 | static __init int io_wq_init(void) |
1388 | { | |
1389 | int ret; | |
1390 | ||
1391 | ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online", | |
0e03496d | 1392 | io_wq_cpu_online, io_wq_cpu_offline); |
43c01fbe JA |
1393 | if (ret < 0) |
1394 | return ret; | |
1395 | io_wq_online = ret; | |
1396 | return 0; | |
1397 | } | |
1398 | subsys_initcall(io_wq_init); |