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> |
355f841a | 16 | #include <linux/task_work.h> |
5bd2182d | 17 | #include <linux/audit.h> |
da64d6db | 18 | #include <linux/mmu_context.h> |
dd47c104 | 19 | #include <uapi/linux/io_uring.h> |
771b53d0 JA |
20 | |
21 | #include "io-wq.h" | |
a6b21fbb | 22 | #include "slist.h" |
024f15e0 | 23 | #include "io_uring.h" |
771b53d0 JA |
24 | |
25 | #define WORKER_IDLE_TIMEOUT (5 * HZ) | |
26 | ||
27 | enum { | |
28 | IO_WORKER_F_UP = 1, /* up and active */ | |
29 | IO_WORKER_F_RUNNING = 2, /* account as running */ | |
30 | IO_WORKER_F_FREE = 4, /* worker on free list */ | |
05c5f4ee | 31 | IO_WORKER_F_BOUND = 8, /* is doing bounded work */ |
771b53d0 JA |
32 | }; |
33 | ||
34 | enum { | |
35 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ | |
771b53d0 JA |
36 | }; |
37 | ||
38 | enum { | |
f95dc207 | 39 | IO_ACCT_STALLED_BIT = 0, /* stalled on hash */ |
771b53d0 JA |
40 | }; |
41 | ||
42 | /* | |
eb47943f | 43 | * One for each thread in a wq pool |
771b53d0 JA |
44 | */ |
45 | struct io_worker { | |
46 | refcount_t ref; | |
47 | unsigned flags; | |
48 | struct hlist_nulls_node nulls_node; | |
e61df66c | 49 | struct list_head all_list; |
771b53d0 | 50 | struct task_struct *task; |
eb47943f | 51 | struct io_wq *wq; |
36c2f922 | 52 | |
771b53d0 | 53 | struct io_wq_work *cur_work; |
361aee45 | 54 | struct io_wq_work *next_work; |
081b5820 | 55 | raw_spinlock_t lock; |
771b53d0 | 56 | |
eb2de941 JA |
57 | struct completion ref_done; |
58 | ||
d3e9f732 JA |
59 | unsigned long create_state; |
60 | struct callback_head create_work; | |
61 | int create_index; | |
62 | ||
3146cba9 JA |
63 | union { |
64 | struct rcu_head rcu; | |
65 | struct work_struct work; | |
66 | }; | |
771b53d0 JA |
67 | }; |
68 | ||
771b53d0 JA |
69 | #if BITS_PER_LONG == 64 |
70 | #define IO_WQ_HASH_ORDER 6 | |
71 | #else | |
72 | #define IO_WQ_HASH_ORDER 5 | |
73 | #endif | |
74 | ||
86f3cd1b PB |
75 | #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER) |
76 | ||
dfd63baf | 77 | struct io_wq_acct { |
c5def4ab JA |
78 | unsigned nr_workers; |
79 | unsigned max_workers; | |
685fe7fe | 80 | int index; |
c5def4ab | 81 | atomic_t nr_running; |
42abc95f | 82 | raw_spinlock_t lock; |
f95dc207 JA |
83 | struct io_wq_work_list work_list; |
84 | unsigned long flags; | |
c5def4ab JA |
85 | }; |
86 | ||
87 | enum { | |
88 | IO_WQ_ACCT_BOUND, | |
89 | IO_WQ_ACCT_UNBOUND, | |
f95dc207 | 90 | IO_WQ_ACCT_NR, |
c5def4ab JA |
91 | }; |
92 | ||
771b53d0 JA |
93 | /* |
94 | * Per io_wq state | |
95 | */ | |
96 | struct io_wq { | |
771b53d0 | 97 | unsigned long state; |
771b53d0 | 98 | |
e9fd9396 | 99 | free_work_fn *free_work; |
f5fa38c5 | 100 | io_wq_work_fn *do_work; |
7d723065 | 101 | |
e941894e JA |
102 | struct io_wq_hash *hash; |
103 | ||
fb3a1f6c JA |
104 | atomic_t worker_refs; |
105 | struct completion worker_done; | |
106 | ||
43c01fbe | 107 | struct hlist_node cpuhp_node; |
3bfe6106 | 108 | |
685fe7fe | 109 | struct task_struct *task; |
c7f405d6 | 110 | |
dfd63baf GKB |
111 | struct io_wq_acct acct[IO_WQ_ACCT_NR]; |
112 | ||
eb47943f GKB |
113 | /* lock protects access to elements below */ |
114 | raw_spinlock_t lock; | |
115 | ||
116 | struct hlist_nulls_head free_list; | |
117 | struct list_head all_list; | |
118 | ||
119 | struct wait_queue_entry wait; | |
120 | ||
121 | struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS]; | |
122 | ||
123 | cpumask_var_t cpu_mask; | |
771b53d0 JA |
124 | }; |
125 | ||
43c01fbe JA |
126 | static enum cpuhp_state io_wq_online; |
127 | ||
f0127254 JA |
128 | struct io_cb_cancel_data { |
129 | work_cancel_fn *fn; | |
130 | void *data; | |
131 | int nr_running; | |
132 | int nr_pending; | |
133 | bool cancel_all; | |
134 | }; | |
135 | ||
eb47943f GKB |
136 | static bool create_io_worker(struct io_wq *wq, int index); |
137 | static void io_wq_dec_running(struct io_worker *worker); | |
138 | static bool io_acct_cancel_pending_work(struct io_wq *wq, | |
dfd63baf | 139 | struct io_wq_acct *acct, |
3146cba9 | 140 | struct io_cb_cancel_data *match); |
1d5f5ea7 | 141 | static void create_worker_cb(struct callback_head *cb); |
71a85387 | 142 | static void io_wq_cancel_tw_create(struct io_wq *wq); |
f0127254 | 143 | |
771b53d0 JA |
144 | static bool io_worker_get(struct io_worker *worker) |
145 | { | |
146 | return refcount_inc_not_zero(&worker->ref); | |
147 | } | |
148 | ||
149 | static void io_worker_release(struct io_worker *worker) | |
150 | { | |
151 | if (refcount_dec_and_test(&worker->ref)) | |
eb2de941 | 152 | complete(&worker->ref_done); |
771b53d0 JA |
153 | } |
154 | ||
dfd63baf | 155 | static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound) |
8418f22a | 156 | { |
dfd63baf | 157 | return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND]; |
8418f22a PB |
158 | } |
159 | ||
dfd63baf GKB |
160 | static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq, |
161 | struct io_wq_work *work) | |
c5def4ab | 162 | { |
dfd63baf | 163 | return io_get_acct(wq, !(work->flags & IO_WQ_WORK_UNBOUND)); |
c5def4ab JA |
164 | } |
165 | ||
dfd63baf | 166 | static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker) |
c5def4ab | 167 | { |
eb47943f | 168 | return io_get_acct(worker->wq, worker->flags & IO_WORKER_F_BOUND); |
c5def4ab JA |
169 | } |
170 | ||
685fe7fe JA |
171 | static void io_worker_ref_put(struct io_wq *wq) |
172 | { | |
173 | if (atomic_dec_and_test(&wq->worker_refs)) | |
174 | complete(&wq->worker_done); | |
175 | } | |
176 | ||
45500dc4 PB |
177 | bool io_wq_worker_stopped(void) |
178 | { | |
179 | struct io_worker *worker = current->worker_private; | |
180 | ||
181 | if (WARN_ON_ONCE(!io_wq_current_is_worker())) | |
182 | return true; | |
183 | ||
184 | return test_bit(IO_WQ_BIT_EXIT, &worker->wq->state); | |
185 | } | |
186 | ||
1d5f5ea7 PB |
187 | static void io_worker_cancel_cb(struct io_worker *worker) |
188 | { | |
dfd63baf | 189 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
eb47943f | 190 | struct io_wq *wq = worker->wq; |
1d5f5ea7 PB |
191 | |
192 | atomic_dec(&acct->nr_running); | |
eb47943f | 193 | raw_spin_lock(&wq->lock); |
1d5f5ea7 | 194 | acct->nr_workers--; |
eb47943f | 195 | raw_spin_unlock(&wq->lock); |
1d5f5ea7 PB |
196 | io_worker_ref_put(wq); |
197 | clear_bit_unlock(0, &worker->create_state); | |
198 | io_worker_release(worker); | |
199 | } | |
200 | ||
201 | static bool io_task_worker_match(struct callback_head *cb, void *data) | |
202 | { | |
203 | struct io_worker *worker; | |
204 | ||
205 | if (cb->func != create_worker_cb) | |
206 | return false; | |
207 | worker = container_of(cb, struct io_worker, create_work); | |
208 | return worker == data; | |
209 | } | |
210 | ||
771b53d0 JA |
211 | static void io_worker_exit(struct io_worker *worker) |
212 | { | |
eb47943f | 213 | struct io_wq *wq = worker->wq; |
771b53d0 | 214 | |
1d5f5ea7 PB |
215 | while (1) { |
216 | struct callback_head *cb = task_work_cancel_match(wq->task, | |
217 | io_task_worker_match, worker); | |
218 | ||
219 | if (!cb) | |
220 | break; | |
221 | io_worker_cancel_cb(worker); | |
222 | } | |
771b53d0 | 223 | |
c907e52c | 224 | io_worker_release(worker); |
eb2de941 | 225 | wait_for_completion(&worker->ref_done); |
771b53d0 | 226 | |
eb47943f | 227 | raw_spin_lock(&wq->lock); |
83d6c393 | 228 | if (worker->flags & IO_WORKER_F_FREE) |
bf1daa4b | 229 | hlist_nulls_del_rcu(&worker->nulls_node); |
e61df66c | 230 | list_del_rcu(&worker->all_list); |
eb47943f GKB |
231 | raw_spin_unlock(&wq->lock); |
232 | io_wq_dec_running(worker); | |
adeaa3f2 JA |
233 | /* |
234 | * this worker is a goner, clear ->worker_private to avoid any | |
235 | * inc/dec running calls that could happen as part of exit from | |
236 | * touching 'worker'. | |
237 | */ | |
238 | current->worker_private = NULL; | |
771b53d0 | 239 | |
364b05fd | 240 | kfree_rcu(worker, rcu); |
eb47943f | 241 | io_worker_ref_put(wq); |
46fe18b1 | 242 | do_exit(0); |
771b53d0 JA |
243 | } |
244 | ||
de36a15f | 245 | static inline bool __io_acct_run_queue(struct io_wq_acct *acct) |
c5def4ab | 246 | { |
de36a15f JA |
247 | return !test_bit(IO_ACCT_STALLED_BIT, &acct->flags) && |
248 | !wq_list_empty(&acct->work_list); | |
249 | } | |
e13fb1fe | 250 | |
de36a15f JA |
251 | /* |
252 | * If there's work to do, returns true with acct->lock acquired. If not, | |
253 | * returns false with no lock held. | |
254 | */ | |
255 | static inline bool io_acct_run_queue(struct io_wq_acct *acct) | |
256 | __acquires(&acct->lock) | |
257 | { | |
e13fb1fe | 258 | raw_spin_lock(&acct->lock); |
de36a15f JA |
259 | if (__io_acct_run_queue(acct)) |
260 | return true; | |
e13fb1fe | 261 | |
de36a15f JA |
262 | raw_spin_unlock(&acct->lock); |
263 | return false; | |
c5def4ab JA |
264 | } |
265 | ||
266 | /* | |
267 | * Check head of free list for an available worker. If one isn't available, | |
685fe7fe | 268 | * caller must create one. |
c5def4ab | 269 | */ |
eb47943f | 270 | static bool io_wq_activate_free_worker(struct io_wq *wq, |
dfd63baf | 271 | struct io_wq_acct *acct) |
c5def4ab JA |
272 | __must_hold(RCU) |
273 | { | |
274 | struct hlist_nulls_node *n; | |
275 | struct io_worker *worker; | |
276 | ||
83d6c393 JA |
277 | /* |
278 | * Iterate free_list and see if we can find an idle worker to | |
279 | * activate. If a given worker is on the free_list but in the process | |
280 | * of exiting, keep trying. | |
281 | */ | |
eb47943f | 282 | hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) { |
83d6c393 JA |
283 | if (!io_worker_get(worker)) |
284 | continue; | |
dfd63baf | 285 | if (io_wq_get_acct(worker) != acct) { |
f95dc207 JA |
286 | io_worker_release(worker); |
287 | continue; | |
288 | } | |
22f7fb80 JA |
289 | /* |
290 | * If the worker is already running, it's either already | |
291 | * starting work or finishing work. In either case, if it does | |
292 | * to go sleep, we'll kick off a new task for this work anyway. | |
293 | */ | |
294 | wake_up_process(worker->task); | |
c5def4ab | 295 | io_worker_release(worker); |
22f7fb80 | 296 | return true; |
c5def4ab JA |
297 | } |
298 | ||
299 | return false; | |
300 | } | |
301 | ||
302 | /* | |
303 | * We need a worker. If we find a free one, we're good. If not, and we're | |
685fe7fe | 304 | * below the max number of workers, create one. |
c5def4ab | 305 | */ |
eb47943f | 306 | static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct) |
c5def4ab | 307 | { |
c5def4ab JA |
308 | /* |
309 | * Most likely an attempt to queue unbounded work on an io_wq that | |
310 | * wasn't setup with any unbounded workers. | |
311 | */ | |
e6ab8991 PB |
312 | if (unlikely(!acct->max_workers)) |
313 | pr_warn_once("io-wq is not configured for unbound workers"); | |
c5def4ab | 314 | |
eb47943f | 315 | raw_spin_lock(&wq->lock); |
bc369921 | 316 | if (acct->nr_workers >= acct->max_workers) { |
eb47943f | 317 | raw_spin_unlock(&wq->lock); |
7a842fb5 | 318 | return true; |
94ffb0a2 | 319 | } |
7a842fb5 | 320 | acct->nr_workers++; |
eb47943f | 321 | raw_spin_unlock(&wq->lock); |
7a842fb5 | 322 | atomic_inc(&acct->nr_running); |
eb47943f GKB |
323 | atomic_inc(&wq->worker_refs); |
324 | return create_io_worker(wq, acct->index); | |
c5def4ab JA |
325 | } |
326 | ||
eb47943f | 327 | static void io_wq_inc_running(struct io_worker *worker) |
c5def4ab | 328 | { |
dfd63baf | 329 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
c5def4ab JA |
330 | |
331 | atomic_inc(&acct->nr_running); | |
332 | } | |
333 | ||
685fe7fe JA |
334 | static void create_worker_cb(struct callback_head *cb) |
335 | { | |
d3e9f732 | 336 | struct io_worker *worker; |
685fe7fe | 337 | struct io_wq *wq; |
eb47943f | 338 | |
dfd63baf | 339 | struct io_wq_acct *acct; |
05c5f4ee | 340 | bool do_create = false; |
685fe7fe | 341 | |
d3e9f732 | 342 | worker = container_of(cb, struct io_worker, create_work); |
eb47943f | 343 | wq = worker->wq; |
dfd63baf | 344 | acct = &wq->acct[worker->create_index]; |
eb47943f GKB |
345 | raw_spin_lock(&wq->lock); |
346 | ||
49e7f0c7 | 347 | if (acct->nr_workers < acct->max_workers) { |
21698274 | 348 | acct->nr_workers++; |
49e7f0c7 HX |
349 | do_create = true; |
350 | } | |
eb47943f | 351 | raw_spin_unlock(&wq->lock); |
49e7f0c7 | 352 | if (do_create) { |
eb47943f | 353 | create_io_worker(wq, worker->create_index); |
49e7f0c7 HX |
354 | } else { |
355 | atomic_dec(&acct->nr_running); | |
356 | io_worker_ref_put(wq); | |
357 | } | |
d3e9f732 JA |
358 | clear_bit_unlock(0, &worker->create_state); |
359 | io_worker_release(worker); | |
685fe7fe JA |
360 | } |
361 | ||
3146cba9 | 362 | static bool io_queue_worker_create(struct io_worker *worker, |
dfd63baf | 363 | struct io_wq_acct *acct, |
3146cba9 | 364 | task_work_func_t func) |
685fe7fe | 365 | { |
eb47943f | 366 | struct io_wq *wq = worker->wq; |
685fe7fe JA |
367 | |
368 | /* raced with exit, just ignore create call */ | |
369 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) | |
370 | goto fail; | |
d3e9f732 JA |
371 | if (!io_worker_get(worker)) |
372 | goto fail; | |
373 | /* | |
374 | * create_state manages ownership of create_work/index. We should | |
375 | * only need one entry per worker, as the worker going to sleep | |
376 | * will trigger the condition, and waking will clear it once it | |
377 | * runs the task_work. | |
378 | */ | |
379 | if (test_bit(0, &worker->create_state) || | |
380 | test_and_set_bit_lock(0, &worker->create_state)) | |
381 | goto fail_release; | |
685fe7fe | 382 | |
71a85387 | 383 | atomic_inc(&wq->worker_refs); |
3146cba9 | 384 | init_task_work(&worker->create_work, func); |
d3e9f732 | 385 | worker->create_index = acct->index; |
71a85387 JA |
386 | if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) { |
387 | /* | |
388 | * EXIT may have been set after checking it above, check after | |
389 | * adding the task_work and remove any creation item if it is | |
390 | * now set. wq exit does that too, but we can have added this | |
391 | * work item after we canceled in io_wq_exit_workers(). | |
392 | */ | |
393 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) | |
394 | io_wq_cancel_tw_create(wq); | |
395 | io_worker_ref_put(wq); | |
3146cba9 | 396 | return true; |
71a85387 JA |
397 | } |
398 | io_worker_ref_put(wq); | |
d3e9f732 JA |
399 | clear_bit_unlock(0, &worker->create_state); |
400 | fail_release: | |
401 | io_worker_release(worker); | |
685fe7fe JA |
402 | fail: |
403 | atomic_dec(&acct->nr_running); | |
404 | io_worker_ref_put(wq); | |
3146cba9 | 405 | return false; |
685fe7fe JA |
406 | } |
407 | ||
eb47943f | 408 | static void io_wq_dec_running(struct io_worker *worker) |
c5def4ab | 409 | { |
dfd63baf | 410 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
eb47943f | 411 | struct io_wq *wq = worker->wq; |
c5def4ab | 412 | |
685fe7fe JA |
413 | if (!(worker->flags & IO_WORKER_F_UP)) |
414 | return; | |
415 | ||
42abc95f HX |
416 | if (!atomic_dec_and_test(&acct->nr_running)) |
417 | return; | |
e13fb1fe | 418 | if (!io_acct_run_queue(acct)) |
42abc95f | 419 | return; |
42abc95f | 420 | |
de36a15f | 421 | raw_spin_unlock(&acct->lock); |
42abc95f | 422 | atomic_inc(&acct->nr_running); |
eb47943f | 423 | atomic_inc(&wq->worker_refs); |
42abc95f | 424 | io_queue_worker_create(worker, acct, create_worker_cb); |
c5def4ab JA |
425 | } |
426 | ||
771b53d0 JA |
427 | /* |
428 | * Worker will start processing some work. Move it to the busy list, if | |
429 | * it's currently on the freelist | |
430 | */ | |
eb47943f | 431 | static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker) |
771b53d0 JA |
432 | { |
433 | if (worker->flags & IO_WORKER_F_FREE) { | |
434 | worker->flags &= ~IO_WORKER_F_FREE; | |
eb47943f | 435 | raw_spin_lock(&wq->lock); |
771b53d0 | 436 | hlist_nulls_del_init_rcu(&worker->nulls_node); |
eb47943f | 437 | raw_spin_unlock(&wq->lock); |
771b53d0 | 438 | } |
771b53d0 JA |
439 | } |
440 | ||
441 | /* | |
07d99096 | 442 | * No work, worker going to sleep. Move to freelist. |
771b53d0 | 443 | */ |
eb47943f GKB |
444 | static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker) |
445 | __must_hold(wq->lock) | |
771b53d0 JA |
446 | { |
447 | if (!(worker->flags & IO_WORKER_F_FREE)) { | |
448 | worker->flags |= IO_WORKER_F_FREE; | |
eb47943f | 449 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); |
771b53d0 | 450 | } |
771b53d0 JA |
451 | } |
452 | ||
60cf46ae PB |
453 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
454 | { | |
455 | return work->flags >> IO_WQ_HASH_SHIFT; | |
456 | } | |
457 | ||
eb47943f | 458 | static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash) |
e941894e | 459 | { |
d3e3c102 | 460 | bool ret = false; |
e941894e | 461 | |
08bdbd39 | 462 | spin_lock_irq(&wq->hash->wait.lock); |
eb47943f GKB |
463 | if (list_empty(&wq->wait.entry)) { |
464 | __add_wait_queue(&wq->hash->wait, &wq->wait); | |
e941894e JA |
465 | if (!test_bit(hash, &wq->hash->map)) { |
466 | __set_current_state(TASK_RUNNING); | |
eb47943f | 467 | list_del_init(&wq->wait.entry); |
d3e3c102 | 468 | ret = true; |
e941894e JA |
469 | } |
470 | } | |
08bdbd39 | 471 | spin_unlock_irq(&wq->hash->wait.lock); |
d3e3c102 | 472 | return ret; |
e941894e JA |
473 | } |
474 | ||
dfd63baf | 475 | static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct, |
0242f642 | 476 | struct io_worker *worker) |
42abc95f | 477 | __must_hold(acct->lock) |
771b53d0 | 478 | { |
6206f0e1 | 479 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 480 | struct io_wq_work *work, *tail; |
e941894e | 481 | unsigned int stall_hash = -1U; |
eb47943f | 482 | struct io_wq *wq = worker->wq; |
771b53d0 | 483 | |
f95dc207 | 484 | wq_list_for_each(node, prev, &acct->work_list) { |
e941894e JA |
485 | unsigned int hash; |
486 | ||
6206f0e1 JA |
487 | work = container_of(node, struct io_wq_work, list); |
488 | ||
771b53d0 | 489 | /* not hashed, can run anytime */ |
8766dd51 | 490 | if (!io_wq_is_hashed(work)) { |
f95dc207 | 491 | wq_list_del(&acct->work_list, node, prev); |
771b53d0 JA |
492 | return work; |
493 | } | |
494 | ||
60cf46ae | 495 | hash = io_get_work_hash(work); |
e941894e | 496 | /* all items with this hash lie in [work, tail] */ |
eb47943f | 497 | tail = wq->hash_tail[hash]; |
e941894e JA |
498 | |
499 | /* hashed, can run if not already running */ | |
eb47943f GKB |
500 | if (!test_and_set_bit(hash, &wq->hash->map)) { |
501 | wq->hash_tail[hash] = NULL; | |
f95dc207 | 502 | wq_list_cut(&acct->work_list, &tail->list, prev); |
771b53d0 JA |
503 | return work; |
504 | } | |
e941894e JA |
505 | if (stall_hash == -1U) |
506 | stall_hash = hash; | |
507 | /* fast forward to a next hash, for-each will fix up @prev */ | |
508 | node = &tail->list; | |
509 | } | |
510 | ||
511 | if (stall_hash != -1U) { | |
d3e3c102 JA |
512 | bool unstalled; |
513 | ||
0242f642 JA |
514 | /* |
515 | * Set this before dropping the lock to avoid racing with new | |
516 | * work being added and clearing the stalled bit. | |
517 | */ | |
f95dc207 | 518 | set_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
42abc95f | 519 | raw_spin_unlock(&acct->lock); |
eb47943f | 520 | unstalled = io_wait_on_hash(wq, stall_hash); |
42abc95f | 521 | raw_spin_lock(&acct->lock); |
d3e3c102 JA |
522 | if (unstalled) { |
523 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); | |
eb47943f GKB |
524 | if (wq_has_sleeper(&wq->hash->wait)) |
525 | wake_up(&wq->hash->wait); | |
d3e3c102 | 526 | } |
771b53d0 JA |
527 | } |
528 | ||
529 | return NULL; | |
530 | } | |
531 | ||
dc026a73 PB |
532 | static void io_assign_current_work(struct io_worker *worker, |
533 | struct io_wq_work *work) | |
534 | { | |
d78298e7 | 535 | if (work) { |
024f15e0 | 536 | io_run_task_work(); |
d78298e7 PB |
537 | cond_resched(); |
538 | } | |
dc026a73 | 539 | |
081b5820 | 540 | raw_spin_lock(&worker->lock); |
dc026a73 | 541 | worker->cur_work = work; |
361aee45 | 542 | worker->next_work = NULL; |
081b5820 | 543 | raw_spin_unlock(&worker->lock); |
dc026a73 PB |
544 | } |
545 | ||
de36a15f JA |
546 | /* |
547 | * Called with acct->lock held, drops it before returning | |
548 | */ | |
549 | static void io_worker_handle_work(struct io_wq_acct *acct, | |
550 | struct io_worker *worker) | |
551 | __releases(&acct->lock) | |
771b53d0 | 552 | { |
eb47943f | 553 | struct io_wq *wq = worker->wq; |
c60eb049 | 554 | bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state); |
771b53d0 JA |
555 | |
556 | do { | |
86f3cd1b | 557 | struct io_wq_work *work; |
73031f76 | 558 | |
771b53d0 JA |
559 | /* |
560 | * If we got some work, mark us as busy. If we didn't, but | |
561 | * the list isn't empty, it means we stalled on hashed work. | |
562 | * Mark us stalled so we don't keep looking for work when we | |
563 | * can't make progress, any work completion or insertion will | |
564 | * clear the stalled flag. | |
565 | */ | |
f95dc207 | 566 | work = io_get_next_work(acct, worker); |
42abc95f | 567 | raw_spin_unlock(&acct->lock); |
361aee45 | 568 | if (work) { |
eb47943f | 569 | __io_worker_busy(wq, worker); |
771b53d0 | 570 | |
361aee45 JA |
571 | /* |
572 | * Make sure cancelation can find this, even before | |
573 | * it becomes the active work. That avoids a window | |
574 | * where the work has been removed from our general | |
575 | * work list, but isn't yet discoverable as the | |
576 | * current work item for this worker. | |
577 | */ | |
578 | raw_spin_lock(&worker->lock); | |
579 | worker->next_work = work; | |
580 | raw_spin_unlock(&worker->lock); | |
42abc95f | 581 | } else { |
771b53d0 | 582 | break; |
42abc95f | 583 | } |
58e39319 | 584 | io_assign_current_work(worker, work); |
e941894e | 585 | __set_current_state(TASK_RUNNING); |
36c2f922 | 586 | |
dc026a73 PB |
587 | /* handle a whole dependent link */ |
588 | do { | |
5280f7e5 | 589 | struct io_wq_work *next_hashed, *linked; |
b089ed39 | 590 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 591 | |
86f3cd1b | 592 | next_hashed = wq_next_work(work); |
c60eb049 PB |
593 | |
594 | if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND)) | |
595 | work->flags |= IO_WQ_WORK_CANCEL; | |
5280f7e5 PB |
596 | wq->do_work(work); |
597 | io_assign_current_work(worker, NULL); | |
dc026a73 | 598 | |
5280f7e5 | 599 | linked = wq->free_work(work); |
86f3cd1b PB |
600 | work = next_hashed; |
601 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
602 | work = linked; | |
603 | linked = NULL; | |
604 | } | |
605 | io_assign_current_work(worker, work); | |
86f3cd1b | 606 | if (linked) |
eb47943f | 607 | io_wq_enqueue(wq, linked); |
86f3cd1b PB |
608 | |
609 | if (hash != -1U && !next_hashed) { | |
d3e3c102 JA |
610 | /* serialize hash clear with wake_up() */ |
611 | spin_lock_irq(&wq->hash->wait.lock); | |
e941894e | 612 | clear_bit(hash, &wq->hash->map); |
f95dc207 | 613 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
d3e3c102 | 614 | spin_unlock_irq(&wq->hash->wait.lock); |
e941894e JA |
615 | if (wq_has_sleeper(&wq->hash->wait)) |
616 | wake_up(&wq->hash->wait); | |
7d723065 | 617 | } |
58e39319 | 618 | } while (work); |
de36a15f JA |
619 | |
620 | if (!__io_acct_run_queue(acct)) | |
621 | break; | |
622 | raw_spin_lock(&acct->lock); | |
771b53d0 JA |
623 | } while (1); |
624 | } | |
625 | ||
eb47943f | 626 | static int io_wq_worker(void *data) |
771b53d0 JA |
627 | { |
628 | struct io_worker *worker = data; | |
dfd63baf | 629 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
eb47943f | 630 | struct io_wq *wq = worker->wq; |
01e68ce0 | 631 | bool exit_mask = false, last_timeout = false; |
46fe18b1 | 632 | char buf[TASK_COMM_LEN]; |
771b53d0 | 633 | |
46fe18b1 | 634 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); |
46fe18b1 | 635 | |
685fe7fe | 636 | snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid); |
46fe18b1 | 637 | set_task_comm(current, buf); |
771b53d0 JA |
638 | |
639 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
16efa4fc JA |
640 | long ret; |
641 | ||
506d95ff | 642 | set_current_state(TASK_INTERRUPTIBLE); |
de36a15f JA |
643 | |
644 | /* | |
645 | * If we have work to do, io_acct_run_queue() returns with | |
646 | * the acct->lock held. If not, it will drop it. | |
647 | */ | |
e13fb1fe | 648 | while (io_acct_run_queue(acct)) |
de36a15f | 649 | io_worker_handle_work(acct, worker); |
e13fb1fe | 650 | |
eb47943f | 651 | raw_spin_lock(&wq->lock); |
01e68ce0 JA |
652 | /* |
653 | * Last sleep timed out. Exit if we're not the last worker, | |
654 | * or if someone modified our affinity. | |
655 | */ | |
656 | if (last_timeout && (exit_mask || acct->nr_workers > 1)) { | |
767a65e9 | 657 | acct->nr_workers--; |
eb47943f | 658 | raw_spin_unlock(&wq->lock); |
05c5f4ee JA |
659 | __set_current_state(TASK_RUNNING); |
660 | break; | |
661 | } | |
662 | last_timeout = false; | |
eb47943f GKB |
663 | __io_worker_idle(wq, worker); |
664 | raw_spin_unlock(&wq->lock); | |
024f15e0 | 665 | if (io_run_task_work()) |
00ddff43 | 666 | continue; |
16efa4fc | 667 | ret = schedule_timeout(WORKER_IDLE_TIMEOUT); |
dbe1bdbb JA |
668 | if (signal_pending(current)) { |
669 | struct ksignal ksig; | |
670 | ||
671 | if (!get_signal(&ksig)) | |
672 | continue; | |
78f8876c | 673 | break; |
dbe1bdbb | 674 | } |
01e68ce0 JA |
675 | if (!ret) { |
676 | last_timeout = true; | |
677 | exit_mask = !cpumask_test_cpu(raw_smp_processor_id(), | |
eb47943f | 678 | wq->cpu_mask); |
01e68ce0 | 679 | } |
771b53d0 JA |
680 | } |
681 | ||
de36a15f JA |
682 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct)) |
683 | io_worker_handle_work(acct, worker); | |
771b53d0 JA |
684 | |
685 | io_worker_exit(worker); | |
686 | return 0; | |
687 | } | |
688 | ||
771b53d0 JA |
689 | /* |
690 | * Called when a worker is scheduled in. Mark us as currently running. | |
691 | */ | |
692 | void io_wq_worker_running(struct task_struct *tsk) | |
693 | { | |
e32cf5df | 694 | struct io_worker *worker = tsk->worker_private; |
771b53d0 | 695 | |
3bfe6106 JA |
696 | if (!worker) |
697 | return; | |
771b53d0 JA |
698 | if (!(worker->flags & IO_WORKER_F_UP)) |
699 | return; | |
700 | if (worker->flags & IO_WORKER_F_RUNNING) | |
701 | return; | |
702 | worker->flags |= IO_WORKER_F_RUNNING; | |
eb47943f | 703 | io_wq_inc_running(worker); |
771b53d0 JA |
704 | } |
705 | ||
706 | /* | |
707 | * Called when worker is going to sleep. If there are no workers currently | |
685fe7fe | 708 | * running and we have work pending, wake up a free one or create a new one. |
771b53d0 JA |
709 | */ |
710 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
711 | { | |
e32cf5df | 712 | struct io_worker *worker = tsk->worker_private; |
771b53d0 | 713 | |
3bfe6106 JA |
714 | if (!worker) |
715 | return; | |
771b53d0 JA |
716 | if (!(worker->flags & IO_WORKER_F_UP)) |
717 | return; | |
718 | if (!(worker->flags & IO_WORKER_F_RUNNING)) | |
719 | return; | |
720 | ||
721 | worker->flags &= ~IO_WORKER_F_RUNNING; | |
eb47943f | 722 | io_wq_dec_running(worker); |
771b53d0 JA |
723 | } |
724 | ||
eb47943f | 725 | static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker, |
3146cba9 JA |
726 | struct task_struct *tsk) |
727 | { | |
e32cf5df | 728 | tsk->worker_private = worker; |
3146cba9 | 729 | worker->task = tsk; |
eb47943f | 730 | set_cpus_allowed_ptr(tsk, wq->cpu_mask); |
3146cba9 | 731 | |
eb47943f GKB |
732 | raw_spin_lock(&wq->lock); |
733 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); | |
734 | list_add_tail_rcu(&worker->all_list, &wq->all_list); | |
3146cba9 | 735 | worker->flags |= IO_WORKER_F_FREE; |
eb47943f | 736 | raw_spin_unlock(&wq->lock); |
3146cba9 JA |
737 | wake_up_new_task(tsk); |
738 | } | |
739 | ||
740 | static bool io_wq_work_match_all(struct io_wq_work *work, void *data) | |
741 | { | |
742 | return true; | |
743 | } | |
744 | ||
745 | static inline bool io_should_retry_thread(long err) | |
746 | { | |
a226abcd JA |
747 | /* |
748 | * Prevent perpetual task_work retry, if the task (or its group) is | |
749 | * exiting. | |
750 | */ | |
751 | if (fatal_signal_pending(current)) | |
752 | return false; | |
753 | ||
3146cba9 JA |
754 | switch (err) { |
755 | case -EAGAIN: | |
756 | case -ERESTARTSYS: | |
757 | case -ERESTARTNOINTR: | |
758 | case -ERESTARTNOHAND: | |
759 | return true; | |
760 | default: | |
761 | return false; | |
762 | } | |
763 | } | |
764 | ||
765 | static void create_worker_cont(struct callback_head *cb) | |
766 | { | |
767 | struct io_worker *worker; | |
768 | struct task_struct *tsk; | |
eb47943f | 769 | struct io_wq *wq; |
3146cba9 JA |
770 | |
771 | worker = container_of(cb, struct io_worker, create_work); | |
772 | clear_bit_unlock(0, &worker->create_state); | |
eb47943f GKB |
773 | wq = worker->wq; |
774 | tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); | |
3146cba9 | 775 | if (!IS_ERR(tsk)) { |
eb47943f | 776 | io_init_new_worker(wq, worker, tsk); |
3146cba9 JA |
777 | io_worker_release(worker); |
778 | return; | |
779 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { | |
dfd63baf | 780 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
3146cba9 JA |
781 | |
782 | atomic_dec(&acct->nr_running); | |
eb47943f | 783 | raw_spin_lock(&wq->lock); |
3146cba9 JA |
784 | acct->nr_workers--; |
785 | if (!acct->nr_workers) { | |
786 | struct io_cb_cancel_data match = { | |
787 | .fn = io_wq_work_match_all, | |
788 | .cancel_all = true, | |
789 | }; | |
790 | ||
eb47943f GKB |
791 | raw_spin_unlock(&wq->lock); |
792 | while (io_acct_cancel_pending_work(wq, acct, &match)) | |
42abc95f HX |
793 | ; |
794 | } else { | |
eb47943f | 795 | raw_spin_unlock(&wq->lock); |
3146cba9 | 796 | } |
eb47943f | 797 | io_worker_ref_put(wq); |
66e70be7 | 798 | kfree(worker); |
3146cba9 JA |
799 | return; |
800 | } | |
801 | ||
802 | /* re-create attempts grab a new worker ref, drop the existing one */ | |
803 | io_worker_release(worker); | |
804 | schedule_work(&worker->work); | |
805 | } | |
806 | ||
807 | static void io_workqueue_create(struct work_struct *work) | |
808 | { | |
809 | struct io_worker *worker = container_of(work, struct io_worker, work); | |
dfd63baf | 810 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
3146cba9 | 811 | |
71e1cef2 | 812 | if (!io_queue_worker_create(worker, acct, create_worker_cont)) |
66e70be7 | 813 | kfree(worker); |
3146cba9 JA |
814 | } |
815 | ||
eb47943f | 816 | static bool create_io_worker(struct io_wq *wq, int index) |
3bfe6106 | 817 | { |
dfd63baf | 818 | struct io_wq_acct *acct = &wq->acct[index]; |
3bfe6106 | 819 | struct io_worker *worker; |
46fe18b1 | 820 | struct task_struct *tsk; |
3bfe6106 | 821 | |
8b3e78b5 JA |
822 | __set_current_state(TASK_RUNNING); |
823 | ||
da64d6db | 824 | worker = kzalloc(sizeof(*worker), GFP_KERNEL); |
3146cba9 | 825 | if (!worker) { |
685fe7fe JA |
826 | fail: |
827 | atomic_dec(&acct->nr_running); | |
eb47943f | 828 | raw_spin_lock(&wq->lock); |
3d4e4fac | 829 | acct->nr_workers--; |
eb47943f | 830 | raw_spin_unlock(&wq->lock); |
685fe7fe | 831 | io_worker_ref_put(wq); |
3146cba9 | 832 | return false; |
3bfe6106 | 833 | } |
46fe18b1 | 834 | |
3146cba9 | 835 | refcount_set(&worker->ref, 1); |
eb47943f | 836 | worker->wq = wq; |
081b5820 | 837 | raw_spin_lock_init(&worker->lock); |
3146cba9 | 838 | init_completion(&worker->ref_done); |
46fe18b1 | 839 | |
46fe18b1 JA |
840 | if (index == IO_WQ_ACCT_BOUND) |
841 | worker->flags |= IO_WORKER_F_BOUND; | |
3146cba9 | 842 | |
eb47943f | 843 | tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); |
3146cba9 | 844 | if (!IS_ERR(tsk)) { |
eb47943f | 845 | io_init_new_worker(wq, worker, tsk); |
3146cba9 | 846 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { |
66e70be7 | 847 | kfree(worker); |
3146cba9 JA |
848 | goto fail; |
849 | } else { | |
850 | INIT_WORK(&worker->work, io_workqueue_create); | |
851 | schedule_work(&worker->work); | |
852 | } | |
853 | ||
854 | return true; | |
771b53d0 JA |
855 | } |
856 | ||
c4068bf8 HD |
857 | /* |
858 | * Iterate the passed in list and call the specific function for each | |
859 | * worker that isn't exiting | |
860 | */ | |
eb47943f | 861 | static bool io_wq_for_each_worker(struct io_wq *wq, |
c4068bf8 HD |
862 | bool (*func)(struct io_worker *, void *), |
863 | void *data) | |
864 | { | |
865 | struct io_worker *worker; | |
866 | bool ret = false; | |
867 | ||
eb47943f | 868 | list_for_each_entry_rcu(worker, &wq->all_list, all_list) { |
c4068bf8 HD |
869 | if (io_worker_get(worker)) { |
870 | /* no task if node is/was offline */ | |
871 | if (worker->task) | |
872 | ret = func(worker, data); | |
873 | io_worker_release(worker); | |
874 | if (ret) | |
875 | break; | |
876 | } | |
877 | } | |
878 | ||
879 | return ret; | |
880 | } | |
881 | ||
882 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
883 | { | |
6cf5862e | 884 | __set_notify_signal(worker->task); |
c4068bf8 HD |
885 | wake_up_process(worker->task); |
886 | return false; | |
887 | } | |
888 | ||
eb47943f | 889 | static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq) |
fc04c39b PB |
890 | { |
891 | do { | |
fc04c39b | 892 | work->flags |= IO_WQ_WORK_CANCEL; |
5280f7e5 PB |
893 | wq->do_work(work); |
894 | work = wq->free_work(work); | |
fc04c39b PB |
895 | } while (work); |
896 | } | |
897 | ||
eb47943f | 898 | static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work) |
86f3cd1b | 899 | { |
eb47943f | 900 | struct io_wq_acct *acct = io_work_get_acct(wq, work); |
86f3cd1b PB |
901 | unsigned int hash; |
902 | struct io_wq_work *tail; | |
903 | ||
904 | if (!io_wq_is_hashed(work)) { | |
905 | append: | |
f95dc207 | 906 | wq_list_add_tail(&work->list, &acct->work_list); |
86f3cd1b PB |
907 | return; |
908 | } | |
909 | ||
910 | hash = io_get_work_hash(work); | |
eb47943f GKB |
911 | tail = wq->hash_tail[hash]; |
912 | wq->hash_tail[hash] = work; | |
86f3cd1b PB |
913 | if (!tail) |
914 | goto append; | |
915 | ||
f95dc207 | 916 | wq_list_add_after(&work->list, &tail->list, &acct->work_list); |
86f3cd1b PB |
917 | } |
918 | ||
713b9825 PB |
919 | static bool io_wq_work_match_item(struct io_wq_work *work, void *data) |
920 | { | |
921 | return work == data; | |
922 | } | |
923 | ||
eb47943f | 924 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) |
771b53d0 | 925 | { |
eb47943f | 926 | struct io_wq_acct *acct = io_work_get_acct(wq, work); |
42abc95f | 927 | struct io_cb_cancel_data match; |
94ffb0a2 JA |
928 | unsigned work_flags = work->flags; |
929 | bool do_create; | |
771b53d0 | 930 | |
991468dc JA |
931 | /* |
932 | * If io-wq is exiting for this task, or if the request has explicitly | |
933 | * been marked as one that should not get executed, cancel it here. | |
934 | */ | |
eb47943f | 935 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || |
991468dc | 936 | (work->flags & IO_WQ_WORK_CANCEL)) { |
eb47943f | 937 | io_run_cancel(work, wq); |
4fb6ac32 JA |
938 | return; |
939 | } | |
940 | ||
42abc95f | 941 | raw_spin_lock(&acct->lock); |
eb47943f | 942 | io_wq_insert_work(wq, work); |
f95dc207 | 943 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
42abc95f | 944 | raw_spin_unlock(&acct->lock); |
94ffb0a2 JA |
945 | |
946 | rcu_read_lock(); | |
eb47943f | 947 | do_create = !io_wq_activate_free_worker(wq, acct); |
94ffb0a2 JA |
948 | rcu_read_unlock(); |
949 | ||
94ffb0a2 | 950 | if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) || |
3146cba9 JA |
951 | !atomic_read(&acct->nr_running))) { |
952 | bool did_create; | |
953 | ||
eb47943f | 954 | did_create = io_wq_create_worker(wq, acct); |
713b9825 PB |
955 | if (likely(did_create)) |
956 | return; | |
957 | ||
eb47943f | 958 | raw_spin_lock(&wq->lock); |
42abc95f | 959 | if (acct->nr_workers) { |
eb47943f | 960 | raw_spin_unlock(&wq->lock); |
42abc95f | 961 | return; |
3146cba9 | 962 | } |
eb47943f | 963 | raw_spin_unlock(&wq->lock); |
42abc95f HX |
964 | |
965 | /* fatal condition, failed to create the first worker */ | |
966 | match.fn = io_wq_work_match_item, | |
967 | match.data = work, | |
968 | match.cancel_all = false, | |
969 | ||
eb47943f | 970 | io_acct_cancel_pending_work(wq, acct, &match); |
3146cba9 | 971 | } |
771b53d0 JA |
972 | } |
973 | ||
771b53d0 | 974 | /* |
8766dd51 PB |
975 | * Work items that hash to the same value will not be done in parallel. |
976 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 977 | */ |
8766dd51 | 978 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 979 | { |
8766dd51 | 980 | unsigned int bit; |
771b53d0 JA |
981 | |
982 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
983 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
984 | } |
985 | ||
361aee45 JA |
986 | static bool __io_wq_worker_cancel(struct io_worker *worker, |
987 | struct io_cb_cancel_data *match, | |
988 | struct io_wq_work *work) | |
989 | { | |
990 | if (work && match->fn(work, match->data)) { | |
991 | work->flags |= IO_WQ_WORK_CANCEL; | |
6cf5862e | 992 | __set_notify_signal(worker->task); |
361aee45 JA |
993 | return true; |
994 | } | |
995 | ||
996 | return false; | |
997 | } | |
998 | ||
2293b419 | 999 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 1000 | { |
2293b419 | 1001 | struct io_cb_cancel_data *match = data; |
62755e35 JA |
1002 | |
1003 | /* | |
1004 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 1005 | * may dereference the passed in work. |
62755e35 | 1006 | */ |
081b5820 | 1007 | raw_spin_lock(&worker->lock); |
361aee45 JA |
1008 | if (__io_wq_worker_cancel(worker, match, worker->cur_work) || |
1009 | __io_wq_worker_cancel(worker, match, worker->next_work)) | |
4f26bda1 | 1010 | match->nr_running++; |
081b5820 | 1011 | raw_spin_unlock(&worker->lock); |
771b53d0 | 1012 | |
4f26bda1 | 1013 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
1014 | } |
1015 | ||
eb47943f | 1016 | static inline void io_wq_remove_pending(struct io_wq *wq, |
204361a7 PB |
1017 | struct io_wq_work *work, |
1018 | struct io_wq_work_node *prev) | |
1019 | { | |
eb47943f | 1020 | struct io_wq_acct *acct = io_work_get_acct(wq, work); |
204361a7 PB |
1021 | unsigned int hash = io_get_work_hash(work); |
1022 | struct io_wq_work *prev_work = NULL; | |
1023 | ||
eb47943f | 1024 | if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) { |
204361a7 PB |
1025 | if (prev) |
1026 | prev_work = container_of(prev, struct io_wq_work, list); | |
1027 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
eb47943f | 1028 | wq->hash_tail[hash] = prev_work; |
204361a7 | 1029 | else |
eb47943f | 1030 | wq->hash_tail[hash] = NULL; |
204361a7 | 1031 | } |
f95dc207 | 1032 | wq_list_del(&acct->work_list, &work->list, prev); |
204361a7 PB |
1033 | } |
1034 | ||
eb47943f | 1035 | static bool io_acct_cancel_pending_work(struct io_wq *wq, |
dfd63baf | 1036 | struct io_wq_acct *acct, |
3146cba9 | 1037 | struct io_cb_cancel_data *match) |
771b53d0 | 1038 | { |
6206f0e1 | 1039 | struct io_wq_work_node *node, *prev; |
771b53d0 | 1040 | struct io_wq_work *work; |
771b53d0 | 1041 | |
42abc95f | 1042 | raw_spin_lock(&acct->lock); |
3146cba9 JA |
1043 | wq_list_for_each(node, prev, &acct->work_list) { |
1044 | work = container_of(node, struct io_wq_work, list); | |
1045 | if (!match->fn(work, match->data)) | |
1046 | continue; | |
eb47943f | 1047 | io_wq_remove_pending(wq, work, prev); |
42abc95f | 1048 | raw_spin_unlock(&acct->lock); |
eb47943f | 1049 | io_run_cancel(work, wq); |
3146cba9 JA |
1050 | match->nr_pending++; |
1051 | /* not safe to continue after unlock */ | |
1052 | return true; | |
1053 | } | |
42abc95f | 1054 | raw_spin_unlock(&acct->lock); |
3146cba9 JA |
1055 | |
1056 | return false; | |
1057 | } | |
1058 | ||
eb47943f | 1059 | static void io_wq_cancel_pending_work(struct io_wq *wq, |
dfd63baf | 1060 | struct io_cb_cancel_data *match) |
3146cba9 JA |
1061 | { |
1062 | int i; | |
4f26bda1 | 1063 | retry: |
f95dc207 | 1064 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
eb47943f | 1065 | struct io_wq_acct *acct = io_get_acct(wq, i == 0); |
4f26bda1 | 1066 | |
eb47943f | 1067 | if (io_acct_cancel_pending_work(wq, acct, match)) { |
3146cba9 JA |
1068 | if (match->cancel_all) |
1069 | goto retry; | |
36e4c58b | 1070 | break; |
f95dc207 | 1071 | } |
771b53d0 | 1072 | } |
f4c2665e PB |
1073 | } |
1074 | ||
eb47943f | 1075 | static void io_wq_cancel_running_work(struct io_wq *wq, |
f4c2665e PB |
1076 | struct io_cb_cancel_data *match) |
1077 | { | |
771b53d0 | 1078 | rcu_read_lock(); |
eb47943f | 1079 | io_wq_for_each_worker(wq, io_wq_worker_cancel, match); |
771b53d0 | 1080 | rcu_read_unlock(); |
771b53d0 JA |
1081 | } |
1082 | ||
2293b419 | 1083 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 1084 | void *data, bool cancel_all) |
771b53d0 | 1085 | { |
2293b419 | 1086 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
1087 | .fn = cancel, |
1088 | .data = data, | |
1089 | .cancel_all = cancel_all, | |
00bcda13 | 1090 | }; |
771b53d0 | 1091 | |
f4c2665e PB |
1092 | /* |
1093 | * First check pending list, if we're lucky we can just remove it | |
1094 | * from there. CANCEL_OK means that the work is returned as-new, | |
1095 | * no completion will be posted for it. | |
efdf5184 JA |
1096 | * |
1097 | * Then check if a free (going busy) or busy worker has the work | |
f4c2665e PB |
1098 | * currently running. If we find it there, we'll return CANCEL_RUNNING |
1099 | * as an indication that we attempt to signal cancellation. The | |
1100 | * completion will run normally in this case. | |
efdf5184 | 1101 | * |
eb47943f | 1102 | * Do both of these while holding the wq->lock, to ensure that |
efdf5184 | 1103 | * we'll find a work item regardless of state. |
f4c2665e | 1104 | */ |
eb47943f | 1105 | io_wq_cancel_pending_work(wq, &match); |
da64d6db BL |
1106 | if (match.nr_pending && !match.cancel_all) |
1107 | return IO_WQ_CANCEL_OK; | |
efdf5184 | 1108 | |
eb47943f GKB |
1109 | raw_spin_lock(&wq->lock); |
1110 | io_wq_cancel_running_work(wq, &match); | |
1111 | raw_spin_unlock(&wq->lock); | |
da64d6db BL |
1112 | if (match.nr_running && !match.cancel_all) |
1113 | return IO_WQ_CANCEL_RUNNING; | |
f4c2665e | 1114 | |
4f26bda1 PB |
1115 | if (match.nr_running) |
1116 | return IO_WQ_CANCEL_RUNNING; | |
1117 | if (match.nr_pending) | |
1118 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 1119 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
1120 | } |
1121 | ||
eb47943f | 1122 | static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode, |
e941894e JA |
1123 | int sync, void *key) |
1124 | { | |
eb47943f | 1125 | struct io_wq *wq = container_of(wait, struct io_wq, wait); |
f95dc207 | 1126 | int i; |
e941894e JA |
1127 | |
1128 | list_del_init(&wait->entry); | |
1129 | ||
1130 | rcu_read_lock(); | |
f95dc207 | 1131 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
dfd63baf | 1132 | struct io_wq_acct *acct = &wq->acct[i]; |
f95dc207 JA |
1133 | |
1134 | if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags)) | |
eb47943f | 1135 | io_wq_activate_free_worker(wq, acct); |
f95dc207 | 1136 | } |
e941894e | 1137 | rcu_read_unlock(); |
e941894e JA |
1138 | return 1; |
1139 | } | |
1140 | ||
576a347b | 1141 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1142 | { |
da64d6db | 1143 | int ret, i; |
771b53d0 JA |
1144 | struct io_wq *wq; |
1145 | ||
f5fa38c5 | 1146 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 | 1147 | return ERR_PTR(-EINVAL); |
e6ab8991 PB |
1148 | if (WARN_ON_ONCE(!bounded)) |
1149 | return ERR_PTR(-EINVAL); | |
e9fd9396 | 1150 | |
da64d6db | 1151 | wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL); |
771b53d0 JA |
1152 | if (!wq) |
1153 | return ERR_PTR(-ENOMEM); | |
771b53d0 | 1154 | |
e941894e JA |
1155 | refcount_inc(&data->hash->refs); |
1156 | wq->hash = data->hash; | |
e9fd9396 | 1157 | wq->free_work = data->free_work; |
f5fa38c5 | 1158 | wq->do_work = data->do_work; |
7d723065 | 1159 | |
43c01fbe | 1160 | ret = -ENOMEM; |
da64d6db | 1161 | |
eb47943f | 1162 | if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL)) |
da64d6db | 1163 | goto err; |
eb47943f | 1164 | cpumask_copy(wq->cpu_mask, cpu_possible_mask); |
dfd63baf GKB |
1165 | wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1166 | wq->acct[IO_WQ_ACCT_UNBOUND].max_workers = | |
da64d6db | 1167 | task_rlimit(current, RLIMIT_NPROC); |
eb47943f GKB |
1168 | INIT_LIST_HEAD(&wq->wait.entry); |
1169 | wq->wait.func = io_wq_hash_wake; | |
da64d6db | 1170 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
dfd63baf | 1171 | struct io_wq_acct *acct = &wq->acct[i]; |
da64d6db BL |
1172 | |
1173 | acct->index = i; | |
1174 | atomic_set(&acct->nr_running, 0); | |
1175 | INIT_WQ_LIST(&acct->work_list); | |
1176 | raw_spin_lock_init(&acct->lock); | |
771b53d0 | 1177 | } |
eb47943f GKB |
1178 | |
1179 | raw_spin_lock_init(&wq->lock); | |
1180 | INIT_HLIST_NULLS_HEAD(&wq->free_list, 0); | |
1181 | INIT_LIST_HEAD(&wq->all_list); | |
771b53d0 | 1182 | |
685fe7fe | 1183 | wq->task = get_task_struct(data->task); |
685fe7fe JA |
1184 | atomic_set(&wq->worker_refs, 1); |
1185 | init_completion(&wq->worker_done); | |
0f8baa3c JM |
1186 | ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
1187 | if (ret) | |
1188 | goto err; | |
1189 | ||
685fe7fe | 1190 | return wq; |
b60fda60 | 1191 | err: |
dc7bbc9e | 1192 | io_wq_put_hash(data->hash); |
eb47943f | 1193 | free_cpumask_var(wq->cpu_mask); |
b60fda60 | 1194 | kfree(wq); |
771b53d0 JA |
1195 | return ERR_PTR(ret); |
1196 | } | |
1197 | ||
c80ca470 JA |
1198 | static bool io_task_work_match(struct callback_head *cb, void *data) |
1199 | { | |
d3e9f732 | 1200 | struct io_worker *worker; |
c80ca470 | 1201 | |
3b33e3f4 | 1202 | if (cb->func != create_worker_cb && cb->func != create_worker_cont) |
c80ca470 | 1203 | return false; |
d3e9f732 | 1204 | worker = container_of(cb, struct io_worker, create_work); |
eb47943f | 1205 | return worker->wq == data; |
c80ca470 JA |
1206 | } |
1207 | ||
17a91051 PB |
1208 | void io_wq_exit_start(struct io_wq *wq) |
1209 | { | |
1210 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
1211 | } | |
1212 | ||
71a85387 | 1213 | static void io_wq_cancel_tw_create(struct io_wq *wq) |
afcc4015 | 1214 | { |
685fe7fe | 1215 | struct callback_head *cb; |
685fe7fe | 1216 | |
c80ca470 | 1217 | while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) { |
d3e9f732 | 1218 | struct io_worker *worker; |
685fe7fe | 1219 | |
d3e9f732 | 1220 | worker = container_of(cb, struct io_worker, create_work); |
1d5f5ea7 | 1221 | io_worker_cancel_cb(worker); |
e6db6f93 JA |
1222 | /* |
1223 | * Only the worker continuation helper has worker allocated and | |
1224 | * hence needs freeing. | |
1225 | */ | |
1226 | if (cb->func == create_worker_cont) | |
1227 | kfree(worker); | |
685fe7fe | 1228 | } |
71a85387 JA |
1229 | } |
1230 | ||
1231 | static void io_wq_exit_workers(struct io_wq *wq) | |
1232 | { | |
71a85387 JA |
1233 | if (!wq->task) |
1234 | return; | |
1235 | ||
1236 | io_wq_cancel_tw_create(wq); | |
685fe7fe JA |
1237 | |
1238 | rcu_read_lock(); | |
eb47943f | 1239 | io_wq_for_each_worker(wq, io_wq_worker_wake, NULL); |
685fe7fe JA |
1240 | rcu_read_unlock(); |
1241 | io_worker_ref_put(wq); | |
1242 | wait_for_completion(&wq->worker_done); | |
3743c172 | 1243 | |
da64d6db | 1244 | spin_lock_irq(&wq->hash->wait.lock); |
eb47943f | 1245 | list_del_init(&wq->wait.entry); |
da64d6db BL |
1246 | spin_unlock_irq(&wq->hash->wait.lock); |
1247 | ||
685fe7fe JA |
1248 | put_task_struct(wq->task); |
1249 | wq->task = NULL; | |
afcc4015 JA |
1250 | } |
1251 | ||
4fb6ac32 | 1252 | static void io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1253 | { |
da64d6db BL |
1254 | struct io_cb_cancel_data match = { |
1255 | .fn = io_wq_work_match_all, | |
1256 | .cancel_all = true, | |
1257 | }; | |
771b53d0 | 1258 | |
43c01fbe | 1259 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
eb47943f GKB |
1260 | io_wq_cancel_pending_work(wq, &match); |
1261 | free_cpumask_var(wq->cpu_mask); | |
e941894e | 1262 | io_wq_put_hash(wq->hash); |
771b53d0 | 1263 | kfree(wq); |
4fb6ac32 JA |
1264 | } |
1265 | ||
afcc4015 JA |
1266 | void io_wq_put_and_exit(struct io_wq *wq) |
1267 | { | |
17a91051 PB |
1268 | WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state)); |
1269 | ||
685fe7fe | 1270 | io_wq_exit_workers(wq); |
382cb030 | 1271 | io_wq_destroy(wq); |
afcc4015 JA |
1272 | } |
1273 | ||
0e03496d JA |
1274 | struct online_data { |
1275 | unsigned int cpu; | |
1276 | bool online; | |
1277 | }; | |
1278 | ||
43c01fbe JA |
1279 | static bool io_wq_worker_affinity(struct io_worker *worker, void *data) |
1280 | { | |
0e03496d | 1281 | struct online_data *od = data; |
e0051d7d | 1282 | |
0e03496d | 1283 | if (od->online) |
eb47943f | 1284 | cpumask_set_cpu(od->cpu, worker->wq->cpu_mask); |
0e03496d | 1285 | else |
eb47943f | 1286 | cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask); |
43c01fbe JA |
1287 | return false; |
1288 | } | |
1289 | ||
0e03496d | 1290 | static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online) |
43c01fbe | 1291 | { |
0e03496d JA |
1292 | struct online_data od = { |
1293 | .cpu = cpu, | |
1294 | .online = online | |
1295 | }; | |
43c01fbe JA |
1296 | |
1297 | rcu_read_lock(); | |
eb47943f | 1298 | io_wq_for_each_worker(wq, io_wq_worker_affinity, &od); |
43c01fbe JA |
1299 | rcu_read_unlock(); |
1300 | return 0; | |
1301 | } | |
1302 | ||
0e03496d JA |
1303 | static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node) |
1304 | { | |
1305 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1306 | ||
1307 | return __io_wq_cpu_online(wq, cpu, true); | |
1308 | } | |
1309 | ||
1310 | static int io_wq_cpu_offline(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, false); | |
1315 | } | |
1316 | ||
ebdfefc0 | 1317 | int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask) |
fe76421d | 1318 | { |
ebdfefc0 JA |
1319 | if (!tctx || !tctx->io_wq) |
1320 | return -EINVAL; | |
1321 | ||
fe76421d | 1322 | rcu_read_lock(); |
da64d6db | 1323 | if (mask) |
ebdfefc0 | 1324 | cpumask_copy(tctx->io_wq->cpu_mask, mask); |
da64d6db | 1325 | else |
ebdfefc0 | 1326 | cpumask_copy(tctx->io_wq->cpu_mask, cpu_possible_mask); |
fe76421d | 1327 | rcu_read_unlock(); |
da64d6db | 1328 | |
fe76421d JA |
1329 | return 0; |
1330 | } | |
1331 | ||
2e480058 JA |
1332 | /* |
1333 | * Set max number of unbounded workers, returns old value. If new_count is 0, | |
1334 | * then just return the old value. | |
1335 | */ | |
1336 | int io_wq_max_workers(struct io_wq *wq, int *new_count) | |
1337 | { | |
dfd63baf | 1338 | struct io_wq_acct *acct; |
71c9ce27 | 1339 | int prev[IO_WQ_ACCT_NR]; |
da64d6db | 1340 | int i; |
2e480058 | 1341 | |
dd47c104 ES |
1342 | BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND); |
1343 | BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND); | |
1344 | BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2); | |
1345 | ||
86127bb1 | 1346 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
2e480058 JA |
1347 | if (new_count[i] > task_rlimit(current, RLIMIT_NPROC)) |
1348 | new_count[i] = task_rlimit(current, RLIMIT_NPROC); | |
1349 | } | |
1350 | ||
71c9ce27 BZ |
1351 | for (i = 0; i < IO_WQ_ACCT_NR; i++) |
1352 | prev[i] = 0; | |
1353 | ||
2e480058 | 1354 | rcu_read_lock(); |
2e480058 | 1355 | |
eb47943f | 1356 | raw_spin_lock(&wq->lock); |
da64d6db | 1357 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
dfd63baf | 1358 | acct = &wq->acct[i]; |
da64d6db BL |
1359 | prev[i] = max_t(int, acct->max_workers, prev[i]); |
1360 | if (new_count[i]) | |
1361 | acct->max_workers = new_count[i]; | |
2e480058 | 1362 | } |
eb47943f | 1363 | raw_spin_unlock(&wq->lock); |
2e480058 | 1364 | rcu_read_unlock(); |
71c9ce27 BZ |
1365 | |
1366 | for (i = 0; i < IO_WQ_ACCT_NR; i++) | |
1367 | new_count[i] = prev[i]; | |
1368 | ||
2e480058 JA |
1369 | return 0; |
1370 | } | |
1371 | ||
43c01fbe JA |
1372 | static __init int io_wq_init(void) |
1373 | { | |
1374 | int ret; | |
1375 | ||
1376 | ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online", | |
0e03496d | 1377 | io_wq_cpu_online, io_wq_cpu_offline); |
43c01fbe JA |
1378 | if (ret < 0) |
1379 | return ret; | |
1380 | io_wq_online = ret; | |
1381 | return 0; | |
1382 | } | |
1383 | subsys_initcall(io_wq_init); |