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 { | |
8a565304 BL |
28 | IO_WORKER_F_UP = 0, /* up and active */ |
29 | IO_WORKER_F_RUNNING = 1, /* account as running */ | |
30 | IO_WORKER_F_FREE = 2, /* worker on free list */ | |
31 | IO_WORKER_F_BOUND = 3, /* 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; | |
8a565304 BL |
47 | int create_index; |
48 | unsigned long flags; | |
771b53d0 | 49 | struct hlist_nulls_node nulls_node; |
e61df66c | 50 | struct list_head all_list; |
771b53d0 | 51 | struct task_struct *task; |
eb47943f | 52 | struct io_wq *wq; |
36c2f922 | 53 | |
771b53d0 | 54 | struct io_wq_work *cur_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; | |
d3e9f732 | 61 | |
3146cba9 JA |
62 | union { |
63 | struct rcu_head rcu; | |
64 | struct work_struct work; | |
65 | }; | |
771b53d0 JA |
66 | }; |
67 | ||
771b53d0 JA |
68 | #if BITS_PER_LONG == 64 |
69 | #define IO_WQ_HASH_ORDER 6 | |
70 | #else | |
71 | #define IO_WQ_HASH_ORDER 5 | |
72 | #endif | |
73 | ||
86f3cd1b PB |
74 | #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER) |
75 | ||
dfd63baf | 76 | struct io_wq_acct { |
c5def4ab JA |
77 | unsigned nr_workers; |
78 | unsigned max_workers; | |
685fe7fe | 79 | int index; |
c5def4ab | 80 | atomic_t nr_running; |
42abc95f | 81 | raw_spinlock_t lock; |
f95dc207 JA |
82 | struct io_wq_work_list work_list; |
83 | unsigned long flags; | |
c5def4ab JA |
84 | }; |
85 | ||
86 | enum { | |
87 | IO_WQ_ACCT_BOUND, | |
88 | IO_WQ_ACCT_UNBOUND, | |
f95dc207 | 89 | IO_WQ_ACCT_NR, |
c5def4ab JA |
90 | }; |
91 | ||
771b53d0 JA |
92 | /* |
93 | * Per io_wq state | |
94 | */ | |
95 | struct io_wq { | |
771b53d0 | 96 | unsigned long state; |
771b53d0 | 97 | |
e9fd9396 | 98 | free_work_fn *free_work; |
f5fa38c5 | 99 | io_wq_work_fn *do_work; |
7d723065 | 100 | |
e941894e JA |
101 | struct io_wq_hash *hash; |
102 | ||
fb3a1f6c JA |
103 | atomic_t worker_refs; |
104 | struct completion worker_done; | |
105 | ||
43c01fbe | 106 | struct hlist_node cpuhp_node; |
3bfe6106 | 107 | |
685fe7fe | 108 | struct task_struct *task; |
c7f405d6 | 109 | |
dfd63baf GKB |
110 | struct io_wq_acct acct[IO_WQ_ACCT_NR]; |
111 | ||
eb47943f GKB |
112 | /* lock protects access to elements below */ |
113 | raw_spinlock_t lock; | |
114 | ||
115 | struct hlist_nulls_head free_list; | |
116 | struct list_head all_list; | |
117 | ||
118 | struct wait_queue_entry wait; | |
119 | ||
120 | struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS]; | |
121 | ||
122 | cpumask_var_t cpu_mask; | |
771b53d0 JA |
123 | }; |
124 | ||
43c01fbe JA |
125 | static enum cpuhp_state io_wq_online; |
126 | ||
f0127254 JA |
127 | struct io_cb_cancel_data { |
128 | work_cancel_fn *fn; | |
129 | void *data; | |
130 | int nr_running; | |
131 | int nr_pending; | |
132 | bool cancel_all; | |
133 | }; | |
134 | ||
eb47943f GKB |
135 | static bool create_io_worker(struct io_wq *wq, int index); |
136 | static void io_wq_dec_running(struct io_worker *worker); | |
137 | static bool io_acct_cancel_pending_work(struct io_wq *wq, | |
dfd63baf | 138 | struct io_wq_acct *acct, |
3146cba9 | 139 | struct io_cb_cancel_data *match); |
1d5f5ea7 | 140 | static void create_worker_cb(struct callback_head *cb); |
71a85387 | 141 | static void io_wq_cancel_tw_create(struct io_wq *wq); |
f0127254 | 142 | |
771b53d0 JA |
143 | static bool io_worker_get(struct io_worker *worker) |
144 | { | |
145 | return refcount_inc_not_zero(&worker->ref); | |
146 | } | |
147 | ||
148 | static void io_worker_release(struct io_worker *worker) | |
149 | { | |
150 | if (refcount_dec_and_test(&worker->ref)) | |
eb2de941 | 151 | complete(&worker->ref_done); |
771b53d0 JA |
152 | } |
153 | ||
dfd63baf | 154 | static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound) |
8418f22a | 155 | { |
dfd63baf | 156 | return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND]; |
8418f22a PB |
157 | } |
158 | ||
dfd63baf GKB |
159 | static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq, |
160 | struct io_wq_work *work) | |
c5def4ab | 161 | { |
dfd63baf | 162 | return io_get_acct(wq, !(work->flags & IO_WQ_WORK_UNBOUND)); |
c5def4ab JA |
163 | } |
164 | ||
dfd63baf | 165 | static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker) |
c5def4ab | 166 | { |
8a565304 | 167 | return io_get_acct(worker->wq, test_bit(IO_WORKER_F_BOUND, &worker->flags)); |
c5def4ab JA |
168 | } |
169 | ||
685fe7fe JA |
170 | static void io_worker_ref_put(struct io_wq *wq) |
171 | { | |
172 | if (atomic_dec_and_test(&wq->worker_refs)) | |
173 | complete(&wq->worker_done); | |
174 | } | |
175 | ||
45500dc4 PB |
176 | bool io_wq_worker_stopped(void) |
177 | { | |
178 | struct io_worker *worker = current->worker_private; | |
179 | ||
180 | if (WARN_ON_ONCE(!io_wq_current_is_worker())) | |
181 | return true; | |
182 | ||
183 | return test_bit(IO_WQ_BIT_EXIT, &worker->wq->state); | |
184 | } | |
185 | ||
1d5f5ea7 PB |
186 | static void io_worker_cancel_cb(struct io_worker *worker) |
187 | { | |
dfd63baf | 188 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
eb47943f | 189 | struct io_wq *wq = worker->wq; |
1d5f5ea7 PB |
190 | |
191 | atomic_dec(&acct->nr_running); | |
eb47943f | 192 | raw_spin_lock(&wq->lock); |
1d5f5ea7 | 193 | acct->nr_workers--; |
eb47943f | 194 | raw_spin_unlock(&wq->lock); |
1d5f5ea7 PB |
195 | io_worker_ref_put(wq); |
196 | clear_bit_unlock(0, &worker->create_state); | |
197 | io_worker_release(worker); | |
198 | } | |
199 | ||
200 | static bool io_task_worker_match(struct callback_head *cb, void *data) | |
201 | { | |
202 | struct io_worker *worker; | |
203 | ||
204 | if (cb->func != create_worker_cb) | |
205 | return false; | |
206 | worker = container_of(cb, struct io_worker, create_work); | |
207 | return worker == data; | |
208 | } | |
209 | ||
771b53d0 JA |
210 | static void io_worker_exit(struct io_worker *worker) |
211 | { | |
eb47943f | 212 | struct io_wq *wq = worker->wq; |
771b53d0 | 213 | |
1d5f5ea7 PB |
214 | while (1) { |
215 | struct callback_head *cb = task_work_cancel_match(wq->task, | |
216 | io_task_worker_match, worker); | |
217 | ||
218 | if (!cb) | |
219 | break; | |
220 | io_worker_cancel_cb(worker); | |
221 | } | |
771b53d0 | 222 | |
c907e52c | 223 | io_worker_release(worker); |
eb2de941 | 224 | wait_for_completion(&worker->ref_done); |
771b53d0 | 225 | |
eb47943f | 226 | raw_spin_lock(&wq->lock); |
8a565304 | 227 | if (test_bit(IO_WORKER_F_FREE, &worker->flags)) |
bf1daa4b | 228 | hlist_nulls_del_rcu(&worker->nulls_node); |
e61df66c | 229 | list_del_rcu(&worker->all_list); |
eb47943f GKB |
230 | raw_spin_unlock(&wq->lock); |
231 | io_wq_dec_running(worker); | |
adeaa3f2 JA |
232 | /* |
233 | * this worker is a goner, clear ->worker_private to avoid any | |
234 | * inc/dec running calls that could happen as part of exit from | |
235 | * touching 'worker'. | |
236 | */ | |
237 | current->worker_private = NULL; | |
771b53d0 | 238 | |
364b05fd | 239 | kfree_rcu(worker, rcu); |
eb47943f | 240 | io_worker_ref_put(wq); |
46fe18b1 | 241 | do_exit(0); |
771b53d0 JA |
242 | } |
243 | ||
de36a15f | 244 | static inline bool __io_acct_run_queue(struct io_wq_acct *acct) |
c5def4ab | 245 | { |
de36a15f JA |
246 | return !test_bit(IO_ACCT_STALLED_BIT, &acct->flags) && |
247 | !wq_list_empty(&acct->work_list); | |
248 | } | |
e13fb1fe | 249 | |
de36a15f JA |
250 | /* |
251 | * If there's work to do, returns true with acct->lock acquired. If not, | |
252 | * returns false with no lock held. | |
253 | */ | |
254 | static inline bool io_acct_run_queue(struct io_wq_acct *acct) | |
255 | __acquires(&acct->lock) | |
256 | { | |
e13fb1fe | 257 | raw_spin_lock(&acct->lock); |
de36a15f JA |
258 | if (__io_acct_run_queue(acct)) |
259 | return true; | |
e13fb1fe | 260 | |
de36a15f JA |
261 | raw_spin_unlock(&acct->lock); |
262 | return false; | |
c5def4ab JA |
263 | } |
264 | ||
265 | /* | |
266 | * Check head of free list for an available worker. If one isn't available, | |
685fe7fe | 267 | * caller must create one. |
c5def4ab | 268 | */ |
eb47943f | 269 | static bool io_wq_activate_free_worker(struct io_wq *wq, |
dfd63baf | 270 | struct io_wq_acct *acct) |
c5def4ab JA |
271 | __must_hold(RCU) |
272 | { | |
273 | struct hlist_nulls_node *n; | |
274 | struct io_worker *worker; | |
275 | ||
83d6c393 JA |
276 | /* |
277 | * Iterate free_list and see if we can find an idle worker to | |
278 | * activate. If a given worker is on the free_list but in the process | |
279 | * of exiting, keep trying. | |
280 | */ | |
eb47943f | 281 | hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) { |
83d6c393 JA |
282 | if (!io_worker_get(worker)) |
283 | continue; | |
dfd63baf | 284 | if (io_wq_get_acct(worker) != acct) { |
f95dc207 JA |
285 | io_worker_release(worker); |
286 | continue; | |
287 | } | |
22f7fb80 JA |
288 | /* |
289 | * If the worker is already running, it's either already | |
290 | * starting work or finishing work. In either case, if it does | |
291 | * to go sleep, we'll kick off a new task for this work anyway. | |
292 | */ | |
293 | wake_up_process(worker->task); | |
c5def4ab | 294 | io_worker_release(worker); |
22f7fb80 | 295 | return true; |
c5def4ab JA |
296 | } |
297 | ||
298 | return false; | |
299 | } | |
300 | ||
301 | /* | |
302 | * We need a worker. If we find a free one, we're good. If not, and we're | |
685fe7fe | 303 | * below the max number of workers, create one. |
c5def4ab | 304 | */ |
eb47943f | 305 | static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct) |
c5def4ab | 306 | { |
c5def4ab JA |
307 | /* |
308 | * Most likely an attempt to queue unbounded work on an io_wq that | |
309 | * wasn't setup with any unbounded workers. | |
310 | */ | |
e6ab8991 PB |
311 | if (unlikely(!acct->max_workers)) |
312 | pr_warn_once("io-wq is not configured for unbound workers"); | |
c5def4ab | 313 | |
eb47943f | 314 | raw_spin_lock(&wq->lock); |
bc369921 | 315 | if (acct->nr_workers >= acct->max_workers) { |
eb47943f | 316 | raw_spin_unlock(&wq->lock); |
7a842fb5 | 317 | return true; |
94ffb0a2 | 318 | } |
7a842fb5 | 319 | acct->nr_workers++; |
eb47943f | 320 | raw_spin_unlock(&wq->lock); |
7a842fb5 | 321 | atomic_inc(&acct->nr_running); |
eb47943f GKB |
322 | atomic_inc(&wq->worker_refs); |
323 | return create_io_worker(wq, acct->index); | |
c5def4ab JA |
324 | } |
325 | ||
eb47943f | 326 | static void io_wq_inc_running(struct io_worker *worker) |
c5def4ab | 327 | { |
dfd63baf | 328 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
c5def4ab JA |
329 | |
330 | atomic_inc(&acct->nr_running); | |
331 | } | |
332 | ||
685fe7fe JA |
333 | static void create_worker_cb(struct callback_head *cb) |
334 | { | |
d3e9f732 | 335 | struct io_worker *worker; |
685fe7fe | 336 | struct io_wq *wq; |
eb47943f | 337 | |
dfd63baf | 338 | struct io_wq_acct *acct; |
05c5f4ee | 339 | bool do_create = false; |
685fe7fe | 340 | |
d3e9f732 | 341 | worker = container_of(cb, struct io_worker, create_work); |
eb47943f | 342 | wq = worker->wq; |
dfd63baf | 343 | acct = &wq->acct[worker->create_index]; |
eb47943f GKB |
344 | raw_spin_lock(&wq->lock); |
345 | ||
49e7f0c7 | 346 | if (acct->nr_workers < acct->max_workers) { |
21698274 | 347 | acct->nr_workers++; |
49e7f0c7 HX |
348 | do_create = true; |
349 | } | |
eb47943f | 350 | raw_spin_unlock(&wq->lock); |
49e7f0c7 | 351 | if (do_create) { |
eb47943f | 352 | create_io_worker(wq, worker->create_index); |
49e7f0c7 HX |
353 | } else { |
354 | atomic_dec(&acct->nr_running); | |
355 | io_worker_ref_put(wq); | |
356 | } | |
d3e9f732 JA |
357 | clear_bit_unlock(0, &worker->create_state); |
358 | io_worker_release(worker); | |
685fe7fe JA |
359 | } |
360 | ||
3146cba9 | 361 | static bool io_queue_worker_create(struct io_worker *worker, |
dfd63baf | 362 | struct io_wq_acct *acct, |
3146cba9 | 363 | task_work_func_t func) |
685fe7fe | 364 | { |
eb47943f | 365 | struct io_wq *wq = worker->wq; |
685fe7fe JA |
366 | |
367 | /* raced with exit, just ignore create call */ | |
368 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) | |
369 | goto fail; | |
d3e9f732 JA |
370 | if (!io_worker_get(worker)) |
371 | goto fail; | |
372 | /* | |
373 | * create_state manages ownership of create_work/index. We should | |
374 | * only need one entry per worker, as the worker going to sleep | |
375 | * will trigger the condition, and waking will clear it once it | |
376 | * runs the task_work. | |
377 | */ | |
378 | if (test_bit(0, &worker->create_state) || | |
379 | test_and_set_bit_lock(0, &worker->create_state)) | |
380 | goto fail_release; | |
685fe7fe | 381 | |
71a85387 | 382 | atomic_inc(&wq->worker_refs); |
3146cba9 | 383 | init_task_work(&worker->create_work, func); |
d3e9f732 | 384 | worker->create_index = acct->index; |
71a85387 JA |
385 | if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) { |
386 | /* | |
387 | * EXIT may have been set after checking it above, check after | |
388 | * adding the task_work and remove any creation item if it is | |
389 | * now set. wq exit does that too, but we can have added this | |
390 | * work item after we canceled in io_wq_exit_workers(). | |
391 | */ | |
392 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) | |
393 | io_wq_cancel_tw_create(wq); | |
394 | io_worker_ref_put(wq); | |
3146cba9 | 395 | return true; |
71a85387 JA |
396 | } |
397 | io_worker_ref_put(wq); | |
d3e9f732 JA |
398 | clear_bit_unlock(0, &worker->create_state); |
399 | fail_release: | |
400 | io_worker_release(worker); | |
685fe7fe JA |
401 | fail: |
402 | atomic_dec(&acct->nr_running); | |
403 | io_worker_ref_put(wq); | |
3146cba9 | 404 | return false; |
685fe7fe JA |
405 | } |
406 | ||
eb47943f | 407 | static void io_wq_dec_running(struct io_worker *worker) |
c5def4ab | 408 | { |
dfd63baf | 409 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
eb47943f | 410 | struct io_wq *wq = worker->wq; |
c5def4ab | 411 | |
8a565304 | 412 | if (!test_bit(IO_WORKER_F_UP, &worker->flags)) |
685fe7fe JA |
413 | return; |
414 | ||
42abc95f HX |
415 | if (!atomic_dec_and_test(&acct->nr_running)) |
416 | return; | |
e13fb1fe | 417 | if (!io_acct_run_queue(acct)) |
42abc95f | 418 | return; |
42abc95f | 419 | |
de36a15f | 420 | raw_spin_unlock(&acct->lock); |
42abc95f | 421 | atomic_inc(&acct->nr_running); |
eb47943f | 422 | atomic_inc(&wq->worker_refs); |
42abc95f | 423 | io_queue_worker_create(worker, acct, create_worker_cb); |
c5def4ab JA |
424 | } |
425 | ||
771b53d0 JA |
426 | /* |
427 | * Worker will start processing some work. Move it to the busy list, if | |
428 | * it's currently on the freelist | |
429 | */ | |
eb47943f | 430 | static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker) |
771b53d0 | 431 | { |
8a565304 BL |
432 | if (test_bit(IO_WORKER_F_FREE, &worker->flags)) { |
433 | clear_bit(IO_WORKER_F_FREE, &worker->flags); | |
eb47943f | 434 | raw_spin_lock(&wq->lock); |
771b53d0 | 435 | hlist_nulls_del_init_rcu(&worker->nulls_node); |
eb47943f | 436 | raw_spin_unlock(&wq->lock); |
771b53d0 | 437 | } |
771b53d0 JA |
438 | } |
439 | ||
440 | /* | |
07d99096 | 441 | * No work, worker going to sleep. Move to freelist. |
771b53d0 | 442 | */ |
eb47943f GKB |
443 | static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker) |
444 | __must_hold(wq->lock) | |
771b53d0 | 445 | { |
8a565304 BL |
446 | if (!test_bit(IO_WORKER_F_FREE, &worker->flags)) { |
447 | set_bit(IO_WORKER_F_FREE, &worker->flags); | |
eb47943f | 448 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); |
771b53d0 | 449 | } |
771b53d0 JA |
450 | } |
451 | ||
60cf46ae PB |
452 | static inline unsigned int io_get_work_hash(struct io_wq_work *work) |
453 | { | |
454 | return work->flags >> IO_WQ_HASH_SHIFT; | |
455 | } | |
456 | ||
eb47943f | 457 | static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash) |
e941894e | 458 | { |
d3e3c102 | 459 | bool ret = false; |
e941894e | 460 | |
08bdbd39 | 461 | spin_lock_irq(&wq->hash->wait.lock); |
eb47943f GKB |
462 | if (list_empty(&wq->wait.entry)) { |
463 | __add_wait_queue(&wq->hash->wait, &wq->wait); | |
e941894e JA |
464 | if (!test_bit(hash, &wq->hash->map)) { |
465 | __set_current_state(TASK_RUNNING); | |
eb47943f | 466 | list_del_init(&wq->wait.entry); |
d3e3c102 | 467 | ret = true; |
e941894e JA |
468 | } |
469 | } | |
08bdbd39 | 470 | spin_unlock_irq(&wq->hash->wait.lock); |
d3e3c102 | 471 | return ret; |
e941894e JA |
472 | } |
473 | ||
dfd63baf | 474 | static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct, |
0242f642 | 475 | struct io_worker *worker) |
42abc95f | 476 | __must_hold(acct->lock) |
771b53d0 | 477 | { |
6206f0e1 | 478 | struct io_wq_work_node *node, *prev; |
86f3cd1b | 479 | struct io_wq_work *work, *tail; |
e941894e | 480 | unsigned int stall_hash = -1U; |
eb47943f | 481 | struct io_wq *wq = worker->wq; |
771b53d0 | 482 | |
f95dc207 | 483 | wq_list_for_each(node, prev, &acct->work_list) { |
e941894e JA |
484 | unsigned int hash; |
485 | ||
6206f0e1 JA |
486 | work = container_of(node, struct io_wq_work, list); |
487 | ||
771b53d0 | 488 | /* not hashed, can run anytime */ |
8766dd51 | 489 | if (!io_wq_is_hashed(work)) { |
f95dc207 | 490 | wq_list_del(&acct->work_list, node, prev); |
771b53d0 JA |
491 | return work; |
492 | } | |
493 | ||
60cf46ae | 494 | hash = io_get_work_hash(work); |
e941894e | 495 | /* all items with this hash lie in [work, tail] */ |
eb47943f | 496 | tail = wq->hash_tail[hash]; |
e941894e JA |
497 | |
498 | /* hashed, can run if not already running */ | |
eb47943f GKB |
499 | if (!test_and_set_bit(hash, &wq->hash->map)) { |
500 | wq->hash_tail[hash] = NULL; | |
f95dc207 | 501 | wq_list_cut(&acct->work_list, &tail->list, prev); |
771b53d0 JA |
502 | return work; |
503 | } | |
e941894e JA |
504 | if (stall_hash == -1U) |
505 | stall_hash = hash; | |
506 | /* fast forward to a next hash, for-each will fix up @prev */ | |
507 | node = &tail->list; | |
508 | } | |
509 | ||
510 | if (stall_hash != -1U) { | |
d3e3c102 JA |
511 | bool unstalled; |
512 | ||
0242f642 JA |
513 | /* |
514 | * Set this before dropping the lock to avoid racing with new | |
515 | * work being added and clearing the stalled bit. | |
516 | */ | |
f95dc207 | 517 | set_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
42abc95f | 518 | raw_spin_unlock(&acct->lock); |
eb47943f | 519 | unstalled = io_wait_on_hash(wq, stall_hash); |
42abc95f | 520 | raw_spin_lock(&acct->lock); |
d3e3c102 JA |
521 | if (unstalled) { |
522 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); | |
eb47943f GKB |
523 | if (wq_has_sleeper(&wq->hash->wait)) |
524 | wake_up(&wq->hash->wait); | |
d3e3c102 | 525 | } |
771b53d0 JA |
526 | } |
527 | ||
528 | return NULL; | |
529 | } | |
530 | ||
dc026a73 PB |
531 | static void io_assign_current_work(struct io_worker *worker, |
532 | struct io_wq_work *work) | |
533 | { | |
d78298e7 | 534 | if (work) { |
024f15e0 | 535 | io_run_task_work(); |
d78298e7 PB |
536 | cond_resched(); |
537 | } | |
dc026a73 | 538 | |
081b5820 | 539 | raw_spin_lock(&worker->lock); |
dc026a73 | 540 | worker->cur_work = work; |
081b5820 | 541 | raw_spin_unlock(&worker->lock); |
dc026a73 PB |
542 | } |
543 | ||
de36a15f JA |
544 | /* |
545 | * Called with acct->lock held, drops it before returning | |
546 | */ | |
547 | static void io_worker_handle_work(struct io_wq_acct *acct, | |
548 | struct io_worker *worker) | |
549 | __releases(&acct->lock) | |
771b53d0 | 550 | { |
eb47943f | 551 | struct io_wq *wq = worker->wq; |
c60eb049 | 552 | bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state); |
771b53d0 JA |
553 | |
554 | do { | |
86f3cd1b | 555 | struct io_wq_work *work; |
73031f76 | 556 | |
771b53d0 JA |
557 | /* |
558 | * If we got some work, mark us as busy. If we didn't, but | |
559 | * the list isn't empty, it means we stalled on hashed work. | |
560 | * Mark us stalled so we don't keep looking for work when we | |
561 | * can't make progress, any work completion or insertion will | |
562 | * clear the stalled flag. | |
563 | */ | |
f95dc207 | 564 | work = io_get_next_work(acct, worker); |
361aee45 | 565 | if (work) { |
361aee45 JA |
566 | /* |
567 | * Make sure cancelation can find this, even before | |
568 | * it becomes the active work. That avoids a window | |
569 | * where the work has been removed from our general | |
570 | * work list, but isn't yet discoverable as the | |
571 | * current work item for this worker. | |
572 | */ | |
573 | raw_spin_lock(&worker->lock); | |
24c3fc5c | 574 | worker->cur_work = work; |
361aee45 | 575 | raw_spin_unlock(&worker->lock); |
42abc95f | 576 | } |
068c27e3 GKB |
577 | |
578 | raw_spin_unlock(&acct->lock); | |
579 | ||
580 | if (!work) | |
581 | break; | |
582 | ||
583 | __io_worker_busy(wq, worker); | |
584 | ||
58e39319 | 585 | io_assign_current_work(worker, work); |
e941894e | 586 | __set_current_state(TASK_RUNNING); |
36c2f922 | 587 | |
dc026a73 PB |
588 | /* handle a whole dependent link */ |
589 | do { | |
5280f7e5 | 590 | struct io_wq_work *next_hashed, *linked; |
b089ed39 | 591 | unsigned int hash = io_get_work_hash(work); |
dc026a73 | 592 | |
86f3cd1b | 593 | next_hashed = wq_next_work(work); |
c60eb049 PB |
594 | |
595 | if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND)) | |
596 | work->flags |= IO_WQ_WORK_CANCEL; | |
5280f7e5 PB |
597 | wq->do_work(work); |
598 | io_assign_current_work(worker, NULL); | |
dc026a73 | 599 | |
5280f7e5 | 600 | linked = wq->free_work(work); |
86f3cd1b PB |
601 | work = next_hashed; |
602 | if (!work && linked && !io_wq_is_hashed(linked)) { | |
603 | work = linked; | |
604 | linked = NULL; | |
605 | } | |
606 | io_assign_current_work(worker, work); | |
86f3cd1b | 607 | if (linked) |
eb47943f | 608 | io_wq_enqueue(wq, linked); |
86f3cd1b PB |
609 | |
610 | if (hash != -1U && !next_hashed) { | |
d3e3c102 JA |
611 | /* serialize hash clear with wake_up() */ |
612 | spin_lock_irq(&wq->hash->wait.lock); | |
e941894e | 613 | clear_bit(hash, &wq->hash->map); |
f95dc207 | 614 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
d3e3c102 | 615 | spin_unlock_irq(&wq->hash->wait.lock); |
e941894e JA |
616 | if (wq_has_sleeper(&wq->hash->wait)) |
617 | wake_up(&wq->hash->wait); | |
7d723065 | 618 | } |
58e39319 | 619 | } while (work); |
de36a15f JA |
620 | |
621 | if (!__io_acct_run_queue(acct)) | |
622 | break; | |
623 | raw_spin_lock(&acct->lock); | |
771b53d0 JA |
624 | } while (1); |
625 | } | |
626 | ||
eb47943f | 627 | static int io_wq_worker(void *data) |
771b53d0 JA |
628 | { |
629 | struct io_worker *worker = data; | |
dfd63baf | 630 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
eb47943f | 631 | struct io_wq *wq = worker->wq; |
01e68ce0 | 632 | bool exit_mask = false, last_timeout = false; |
46fe18b1 | 633 | char buf[TASK_COMM_LEN]; |
771b53d0 | 634 | |
8a565304 BL |
635 | set_mask_bits(&worker->flags, 0, |
636 | BIT(IO_WORKER_F_UP) | BIT(IO_WORKER_F_RUNNING)); | |
46fe18b1 | 637 | |
685fe7fe | 638 | snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid); |
46fe18b1 | 639 | set_task_comm(current, buf); |
771b53d0 JA |
640 | |
641 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { | |
16efa4fc JA |
642 | long ret; |
643 | ||
506d95ff | 644 | set_current_state(TASK_INTERRUPTIBLE); |
de36a15f JA |
645 | |
646 | /* | |
647 | * If we have work to do, io_acct_run_queue() returns with | |
648 | * the acct->lock held. If not, it will drop it. | |
649 | */ | |
e13fb1fe | 650 | while (io_acct_run_queue(acct)) |
de36a15f | 651 | io_worker_handle_work(acct, worker); |
e13fb1fe | 652 | |
eb47943f | 653 | raw_spin_lock(&wq->lock); |
01e68ce0 JA |
654 | /* |
655 | * Last sleep timed out. Exit if we're not the last worker, | |
656 | * or if someone modified our affinity. | |
657 | */ | |
658 | if (last_timeout && (exit_mask || acct->nr_workers > 1)) { | |
767a65e9 | 659 | acct->nr_workers--; |
eb47943f | 660 | raw_spin_unlock(&wq->lock); |
05c5f4ee JA |
661 | __set_current_state(TASK_RUNNING); |
662 | break; | |
663 | } | |
664 | last_timeout = false; | |
eb47943f GKB |
665 | __io_worker_idle(wq, worker); |
666 | raw_spin_unlock(&wq->lock); | |
024f15e0 | 667 | if (io_run_task_work()) |
00ddff43 | 668 | continue; |
16efa4fc | 669 | ret = schedule_timeout(WORKER_IDLE_TIMEOUT); |
dbe1bdbb JA |
670 | if (signal_pending(current)) { |
671 | struct ksignal ksig; | |
672 | ||
673 | if (!get_signal(&ksig)) | |
674 | continue; | |
78f8876c | 675 | break; |
dbe1bdbb | 676 | } |
01e68ce0 JA |
677 | if (!ret) { |
678 | last_timeout = true; | |
679 | exit_mask = !cpumask_test_cpu(raw_smp_processor_id(), | |
eb47943f | 680 | wq->cpu_mask); |
01e68ce0 | 681 | } |
771b53d0 JA |
682 | } |
683 | ||
de36a15f JA |
684 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct)) |
685 | io_worker_handle_work(acct, worker); | |
771b53d0 JA |
686 | |
687 | io_worker_exit(worker); | |
688 | return 0; | |
689 | } | |
690 | ||
771b53d0 JA |
691 | /* |
692 | * Called when a worker is scheduled in. Mark us as currently running. | |
693 | */ | |
694 | void io_wq_worker_running(struct task_struct *tsk) | |
695 | { | |
e32cf5df | 696 | struct io_worker *worker = tsk->worker_private; |
771b53d0 | 697 | |
3bfe6106 JA |
698 | if (!worker) |
699 | return; | |
8a565304 | 700 | if (!test_bit(IO_WORKER_F_UP, &worker->flags)) |
771b53d0 | 701 | return; |
8a565304 | 702 | if (test_bit(IO_WORKER_F_RUNNING, &worker->flags)) |
771b53d0 | 703 | return; |
8a565304 | 704 | set_bit(IO_WORKER_F_RUNNING, &worker->flags); |
eb47943f | 705 | io_wq_inc_running(worker); |
771b53d0 JA |
706 | } |
707 | ||
708 | /* | |
709 | * Called when worker is going to sleep. If there are no workers currently | |
685fe7fe | 710 | * running and we have work pending, wake up a free one or create a new one. |
771b53d0 JA |
711 | */ |
712 | void io_wq_worker_sleeping(struct task_struct *tsk) | |
713 | { | |
e32cf5df | 714 | struct io_worker *worker = tsk->worker_private; |
771b53d0 | 715 | |
3bfe6106 JA |
716 | if (!worker) |
717 | return; | |
8a565304 | 718 | if (!test_bit(IO_WORKER_F_UP, &worker->flags)) |
771b53d0 | 719 | return; |
8a565304 | 720 | if (!test_bit(IO_WORKER_F_RUNNING, &worker->flags)) |
771b53d0 JA |
721 | return; |
722 | ||
8a565304 | 723 | clear_bit(IO_WORKER_F_RUNNING, &worker->flags); |
eb47943f | 724 | io_wq_dec_running(worker); |
771b53d0 JA |
725 | } |
726 | ||
eb47943f | 727 | static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker, |
3146cba9 JA |
728 | struct task_struct *tsk) |
729 | { | |
e32cf5df | 730 | tsk->worker_private = worker; |
3146cba9 | 731 | worker->task = tsk; |
eb47943f | 732 | set_cpus_allowed_ptr(tsk, wq->cpu_mask); |
3146cba9 | 733 | |
eb47943f GKB |
734 | raw_spin_lock(&wq->lock); |
735 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); | |
736 | list_add_tail_rcu(&worker->all_list, &wq->all_list); | |
8a565304 | 737 | set_bit(IO_WORKER_F_FREE, &worker->flags); |
eb47943f | 738 | raw_spin_unlock(&wq->lock); |
3146cba9 JA |
739 | wake_up_new_task(tsk); |
740 | } | |
741 | ||
742 | static bool io_wq_work_match_all(struct io_wq_work *work, void *data) | |
743 | { | |
744 | return true; | |
745 | } | |
746 | ||
747 | static inline bool io_should_retry_thread(long err) | |
748 | { | |
a226abcd JA |
749 | /* |
750 | * Prevent perpetual task_work retry, if the task (or its group) is | |
751 | * exiting. | |
752 | */ | |
753 | if (fatal_signal_pending(current)) | |
754 | return false; | |
755 | ||
3146cba9 JA |
756 | switch (err) { |
757 | case -EAGAIN: | |
758 | case -ERESTARTSYS: | |
759 | case -ERESTARTNOINTR: | |
760 | case -ERESTARTNOHAND: | |
761 | return true; | |
762 | default: | |
763 | return false; | |
764 | } | |
765 | } | |
766 | ||
767 | static void create_worker_cont(struct callback_head *cb) | |
768 | { | |
769 | struct io_worker *worker; | |
770 | struct task_struct *tsk; | |
eb47943f | 771 | struct io_wq *wq; |
3146cba9 JA |
772 | |
773 | worker = container_of(cb, struct io_worker, create_work); | |
774 | clear_bit_unlock(0, &worker->create_state); | |
eb47943f GKB |
775 | wq = worker->wq; |
776 | tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); | |
3146cba9 | 777 | if (!IS_ERR(tsk)) { |
eb47943f | 778 | io_init_new_worker(wq, worker, tsk); |
3146cba9 JA |
779 | io_worker_release(worker); |
780 | return; | |
781 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { | |
dfd63baf | 782 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
3146cba9 JA |
783 | |
784 | atomic_dec(&acct->nr_running); | |
eb47943f | 785 | raw_spin_lock(&wq->lock); |
3146cba9 JA |
786 | acct->nr_workers--; |
787 | if (!acct->nr_workers) { | |
788 | struct io_cb_cancel_data match = { | |
789 | .fn = io_wq_work_match_all, | |
790 | .cancel_all = true, | |
791 | }; | |
792 | ||
eb47943f GKB |
793 | raw_spin_unlock(&wq->lock); |
794 | while (io_acct_cancel_pending_work(wq, acct, &match)) | |
42abc95f HX |
795 | ; |
796 | } else { | |
eb47943f | 797 | raw_spin_unlock(&wq->lock); |
3146cba9 | 798 | } |
eb47943f | 799 | io_worker_ref_put(wq); |
66e70be7 | 800 | kfree(worker); |
3146cba9 JA |
801 | return; |
802 | } | |
803 | ||
804 | /* re-create attempts grab a new worker ref, drop the existing one */ | |
805 | io_worker_release(worker); | |
806 | schedule_work(&worker->work); | |
807 | } | |
808 | ||
809 | static void io_workqueue_create(struct work_struct *work) | |
810 | { | |
811 | struct io_worker *worker = container_of(work, struct io_worker, work); | |
dfd63baf | 812 | struct io_wq_acct *acct = io_wq_get_acct(worker); |
3146cba9 | 813 | |
71e1cef2 | 814 | if (!io_queue_worker_create(worker, acct, create_worker_cont)) |
66e70be7 | 815 | kfree(worker); |
3146cba9 JA |
816 | } |
817 | ||
eb47943f | 818 | static bool create_io_worker(struct io_wq *wq, int index) |
3bfe6106 | 819 | { |
dfd63baf | 820 | struct io_wq_acct *acct = &wq->acct[index]; |
3bfe6106 | 821 | struct io_worker *worker; |
46fe18b1 | 822 | struct task_struct *tsk; |
3bfe6106 | 823 | |
8b3e78b5 JA |
824 | __set_current_state(TASK_RUNNING); |
825 | ||
da64d6db | 826 | worker = kzalloc(sizeof(*worker), GFP_KERNEL); |
3146cba9 | 827 | if (!worker) { |
685fe7fe JA |
828 | fail: |
829 | atomic_dec(&acct->nr_running); | |
eb47943f | 830 | raw_spin_lock(&wq->lock); |
3d4e4fac | 831 | acct->nr_workers--; |
eb47943f | 832 | raw_spin_unlock(&wq->lock); |
685fe7fe | 833 | io_worker_ref_put(wq); |
3146cba9 | 834 | return false; |
3bfe6106 | 835 | } |
46fe18b1 | 836 | |
3146cba9 | 837 | refcount_set(&worker->ref, 1); |
eb47943f | 838 | worker->wq = wq; |
081b5820 | 839 | raw_spin_lock_init(&worker->lock); |
3146cba9 | 840 | init_completion(&worker->ref_done); |
46fe18b1 | 841 | |
46fe18b1 | 842 | if (index == IO_WQ_ACCT_BOUND) |
8a565304 | 843 | set_bit(IO_WORKER_F_BOUND, &worker->flags); |
3146cba9 | 844 | |
eb47943f | 845 | tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); |
3146cba9 | 846 | if (!IS_ERR(tsk)) { |
eb47943f | 847 | io_init_new_worker(wq, worker, tsk); |
3146cba9 | 848 | } else if (!io_should_retry_thread(PTR_ERR(tsk))) { |
66e70be7 | 849 | kfree(worker); |
3146cba9 JA |
850 | goto fail; |
851 | } else { | |
852 | INIT_WORK(&worker->work, io_workqueue_create); | |
853 | schedule_work(&worker->work); | |
854 | } | |
855 | ||
856 | return true; | |
771b53d0 JA |
857 | } |
858 | ||
c4068bf8 HD |
859 | /* |
860 | * Iterate the passed in list and call the specific function for each | |
861 | * worker that isn't exiting | |
862 | */ | |
eb47943f | 863 | static bool io_wq_for_each_worker(struct io_wq *wq, |
c4068bf8 HD |
864 | bool (*func)(struct io_worker *, void *), |
865 | void *data) | |
866 | { | |
867 | struct io_worker *worker; | |
868 | bool ret = false; | |
869 | ||
eb47943f | 870 | list_for_each_entry_rcu(worker, &wq->all_list, all_list) { |
c4068bf8 HD |
871 | if (io_worker_get(worker)) { |
872 | /* no task if node is/was offline */ | |
873 | if (worker->task) | |
874 | ret = func(worker, data); | |
875 | io_worker_release(worker); | |
876 | if (ret) | |
877 | break; | |
878 | } | |
879 | } | |
880 | ||
881 | return ret; | |
882 | } | |
883 | ||
884 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) | |
885 | { | |
6cf5862e | 886 | __set_notify_signal(worker->task); |
c4068bf8 HD |
887 | wake_up_process(worker->task); |
888 | return false; | |
889 | } | |
890 | ||
eb47943f | 891 | static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq) |
fc04c39b PB |
892 | { |
893 | do { | |
fc04c39b | 894 | work->flags |= IO_WQ_WORK_CANCEL; |
5280f7e5 PB |
895 | wq->do_work(work); |
896 | work = wq->free_work(work); | |
fc04c39b PB |
897 | } while (work); |
898 | } | |
899 | ||
eb47943f | 900 | static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work) |
86f3cd1b | 901 | { |
eb47943f | 902 | struct io_wq_acct *acct = io_work_get_acct(wq, work); |
86f3cd1b PB |
903 | unsigned int hash; |
904 | struct io_wq_work *tail; | |
905 | ||
906 | if (!io_wq_is_hashed(work)) { | |
907 | append: | |
f95dc207 | 908 | wq_list_add_tail(&work->list, &acct->work_list); |
86f3cd1b PB |
909 | return; |
910 | } | |
911 | ||
912 | hash = io_get_work_hash(work); | |
eb47943f GKB |
913 | tail = wq->hash_tail[hash]; |
914 | wq->hash_tail[hash] = work; | |
86f3cd1b PB |
915 | if (!tail) |
916 | goto append; | |
917 | ||
f95dc207 | 918 | wq_list_add_after(&work->list, &tail->list, &acct->work_list); |
86f3cd1b PB |
919 | } |
920 | ||
713b9825 PB |
921 | static bool io_wq_work_match_item(struct io_wq_work *work, void *data) |
922 | { | |
923 | return work == data; | |
924 | } | |
925 | ||
eb47943f | 926 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) |
771b53d0 | 927 | { |
eb47943f | 928 | struct io_wq_acct *acct = io_work_get_acct(wq, work); |
8a565304 | 929 | unsigned long work_flags = work->flags; |
91215f70 SH |
930 | struct io_cb_cancel_data match = { |
931 | .fn = io_wq_work_match_item, | |
932 | .data = work, | |
933 | .cancel_all = false, | |
934 | }; | |
94ffb0a2 | 935 | bool do_create; |
771b53d0 | 936 | |
991468dc JA |
937 | /* |
938 | * If io-wq is exiting for this task, or if the request has explicitly | |
939 | * been marked as one that should not get executed, cancel it here. | |
940 | */ | |
eb47943f | 941 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || |
991468dc | 942 | (work->flags & IO_WQ_WORK_CANCEL)) { |
eb47943f | 943 | io_run_cancel(work, wq); |
4fb6ac32 JA |
944 | return; |
945 | } | |
946 | ||
42abc95f | 947 | raw_spin_lock(&acct->lock); |
eb47943f | 948 | io_wq_insert_work(wq, work); |
f95dc207 | 949 | clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); |
42abc95f | 950 | raw_spin_unlock(&acct->lock); |
94ffb0a2 JA |
951 | |
952 | rcu_read_lock(); | |
eb47943f | 953 | do_create = !io_wq_activate_free_worker(wq, acct); |
94ffb0a2 JA |
954 | rcu_read_unlock(); |
955 | ||
94ffb0a2 | 956 | if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) || |
3146cba9 JA |
957 | !atomic_read(&acct->nr_running))) { |
958 | bool did_create; | |
959 | ||
eb47943f | 960 | did_create = io_wq_create_worker(wq, acct); |
713b9825 PB |
961 | if (likely(did_create)) |
962 | return; | |
963 | ||
eb47943f | 964 | raw_spin_lock(&wq->lock); |
42abc95f | 965 | if (acct->nr_workers) { |
eb47943f | 966 | raw_spin_unlock(&wq->lock); |
42abc95f | 967 | return; |
3146cba9 | 968 | } |
eb47943f | 969 | raw_spin_unlock(&wq->lock); |
42abc95f HX |
970 | |
971 | /* fatal condition, failed to create the first worker */ | |
eb47943f | 972 | io_acct_cancel_pending_work(wq, acct, &match); |
3146cba9 | 973 | } |
771b53d0 JA |
974 | } |
975 | ||
771b53d0 | 976 | /* |
8766dd51 PB |
977 | * Work items that hash to the same value will not be done in parallel. |
978 | * Used to limit concurrent writes, generally hashed by inode. | |
771b53d0 | 979 | */ |
8766dd51 | 980 | void io_wq_hash_work(struct io_wq_work *work, void *val) |
771b53d0 | 981 | { |
8766dd51 | 982 | unsigned int bit; |
771b53d0 JA |
983 | |
984 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); | |
985 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); | |
771b53d0 JA |
986 | } |
987 | ||
361aee45 JA |
988 | static bool __io_wq_worker_cancel(struct io_worker *worker, |
989 | struct io_cb_cancel_data *match, | |
990 | struct io_wq_work *work) | |
991 | { | |
992 | if (work && match->fn(work, match->data)) { | |
993 | work->flags |= IO_WQ_WORK_CANCEL; | |
6cf5862e | 994 | __set_notify_signal(worker->task); |
361aee45 JA |
995 | return true; |
996 | } | |
997 | ||
998 | return false; | |
999 | } | |
1000 | ||
2293b419 | 1001 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
62755e35 | 1002 | { |
2293b419 | 1003 | struct io_cb_cancel_data *match = data; |
62755e35 JA |
1004 | |
1005 | /* | |
1006 | * Hold the lock to avoid ->cur_work going out of scope, caller | |
36c2f922 | 1007 | * may dereference the passed in work. |
62755e35 | 1008 | */ |
081b5820 | 1009 | raw_spin_lock(&worker->lock); |
24c3fc5c | 1010 | if (__io_wq_worker_cancel(worker, match, worker->cur_work)) |
4f26bda1 | 1011 | match->nr_running++; |
081b5820 | 1012 | raw_spin_unlock(&worker->lock); |
771b53d0 | 1013 | |
4f26bda1 | 1014 | return match->nr_running && !match->cancel_all; |
771b53d0 JA |
1015 | } |
1016 | ||
eb47943f | 1017 | static inline void io_wq_remove_pending(struct io_wq *wq, |
204361a7 PB |
1018 | struct io_wq_work *work, |
1019 | struct io_wq_work_node *prev) | |
1020 | { | |
eb47943f | 1021 | struct io_wq_acct *acct = io_work_get_acct(wq, work); |
204361a7 PB |
1022 | unsigned int hash = io_get_work_hash(work); |
1023 | struct io_wq_work *prev_work = NULL; | |
1024 | ||
eb47943f | 1025 | if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) { |
204361a7 PB |
1026 | if (prev) |
1027 | prev_work = container_of(prev, struct io_wq_work, list); | |
1028 | if (prev_work && io_get_work_hash(prev_work) == hash) | |
eb47943f | 1029 | wq->hash_tail[hash] = prev_work; |
204361a7 | 1030 | else |
eb47943f | 1031 | wq->hash_tail[hash] = NULL; |
204361a7 | 1032 | } |
f95dc207 | 1033 | wq_list_del(&acct->work_list, &work->list, prev); |
204361a7 PB |
1034 | } |
1035 | ||
eb47943f | 1036 | static bool io_acct_cancel_pending_work(struct io_wq *wq, |
dfd63baf | 1037 | struct io_wq_acct *acct, |
3146cba9 | 1038 | struct io_cb_cancel_data *match) |
771b53d0 | 1039 | { |
6206f0e1 | 1040 | struct io_wq_work_node *node, *prev; |
771b53d0 | 1041 | struct io_wq_work *work; |
771b53d0 | 1042 | |
42abc95f | 1043 | raw_spin_lock(&acct->lock); |
3146cba9 JA |
1044 | wq_list_for_each(node, prev, &acct->work_list) { |
1045 | work = container_of(node, struct io_wq_work, list); | |
1046 | if (!match->fn(work, match->data)) | |
1047 | continue; | |
eb47943f | 1048 | io_wq_remove_pending(wq, work, prev); |
42abc95f | 1049 | raw_spin_unlock(&acct->lock); |
eb47943f | 1050 | io_run_cancel(work, wq); |
3146cba9 JA |
1051 | match->nr_pending++; |
1052 | /* not safe to continue after unlock */ | |
1053 | return true; | |
1054 | } | |
42abc95f | 1055 | raw_spin_unlock(&acct->lock); |
3146cba9 JA |
1056 | |
1057 | return false; | |
1058 | } | |
1059 | ||
eb47943f | 1060 | static void io_wq_cancel_pending_work(struct io_wq *wq, |
dfd63baf | 1061 | struct io_cb_cancel_data *match) |
3146cba9 JA |
1062 | { |
1063 | int i; | |
4f26bda1 | 1064 | retry: |
f95dc207 | 1065 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
eb47943f | 1066 | struct io_wq_acct *acct = io_get_acct(wq, i == 0); |
4f26bda1 | 1067 | |
eb47943f | 1068 | if (io_acct_cancel_pending_work(wq, acct, match)) { |
3146cba9 JA |
1069 | if (match->cancel_all) |
1070 | goto retry; | |
36e4c58b | 1071 | break; |
f95dc207 | 1072 | } |
771b53d0 | 1073 | } |
f4c2665e PB |
1074 | } |
1075 | ||
eb47943f | 1076 | static void io_wq_cancel_running_work(struct io_wq *wq, |
f4c2665e PB |
1077 | struct io_cb_cancel_data *match) |
1078 | { | |
771b53d0 | 1079 | rcu_read_lock(); |
eb47943f | 1080 | io_wq_for_each_worker(wq, io_wq_worker_cancel, match); |
771b53d0 | 1081 | rcu_read_unlock(); |
771b53d0 JA |
1082 | } |
1083 | ||
2293b419 | 1084 | enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, |
4f26bda1 | 1085 | void *data, bool cancel_all) |
771b53d0 | 1086 | { |
2293b419 | 1087 | struct io_cb_cancel_data match = { |
4f26bda1 PB |
1088 | .fn = cancel, |
1089 | .data = data, | |
1090 | .cancel_all = cancel_all, | |
00bcda13 | 1091 | }; |
771b53d0 | 1092 | |
f4c2665e PB |
1093 | /* |
1094 | * First check pending list, if we're lucky we can just remove it | |
1095 | * from there. CANCEL_OK means that the work is returned as-new, | |
1096 | * no completion will be posted for it. | |
efdf5184 JA |
1097 | * |
1098 | * Then check if a free (going busy) or busy worker has the work | |
f4c2665e PB |
1099 | * currently running. If we find it there, we'll return CANCEL_RUNNING |
1100 | * as an indication that we attempt to signal cancellation. The | |
1101 | * completion will run normally in this case. | |
efdf5184 | 1102 | * |
eb47943f | 1103 | * Do both of these while holding the wq->lock, to ensure that |
efdf5184 | 1104 | * we'll find a work item regardless of state. |
f4c2665e | 1105 | */ |
eb47943f | 1106 | io_wq_cancel_pending_work(wq, &match); |
da64d6db BL |
1107 | if (match.nr_pending && !match.cancel_all) |
1108 | return IO_WQ_CANCEL_OK; | |
efdf5184 | 1109 | |
eb47943f GKB |
1110 | raw_spin_lock(&wq->lock); |
1111 | io_wq_cancel_running_work(wq, &match); | |
1112 | raw_spin_unlock(&wq->lock); | |
da64d6db BL |
1113 | if (match.nr_running && !match.cancel_all) |
1114 | return IO_WQ_CANCEL_RUNNING; | |
f4c2665e | 1115 | |
4f26bda1 PB |
1116 | if (match.nr_running) |
1117 | return IO_WQ_CANCEL_RUNNING; | |
1118 | if (match.nr_pending) | |
1119 | return IO_WQ_CANCEL_OK; | |
f4c2665e | 1120 | return IO_WQ_CANCEL_NOTFOUND; |
771b53d0 JA |
1121 | } |
1122 | ||
eb47943f | 1123 | static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode, |
e941894e JA |
1124 | int sync, void *key) |
1125 | { | |
eb47943f | 1126 | struct io_wq *wq = container_of(wait, struct io_wq, wait); |
f95dc207 | 1127 | int i; |
e941894e JA |
1128 | |
1129 | list_del_init(&wait->entry); | |
1130 | ||
1131 | rcu_read_lock(); | |
f95dc207 | 1132 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
dfd63baf | 1133 | struct io_wq_acct *acct = &wq->acct[i]; |
f95dc207 JA |
1134 | |
1135 | if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags)) | |
eb47943f | 1136 | io_wq_activate_free_worker(wq, acct); |
f95dc207 | 1137 | } |
e941894e | 1138 | rcu_read_unlock(); |
e941894e JA |
1139 | return 1; |
1140 | } | |
1141 | ||
576a347b | 1142 | struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) |
771b53d0 | 1143 | { |
da64d6db | 1144 | int ret, i; |
771b53d0 JA |
1145 | struct io_wq *wq; |
1146 | ||
f5fa38c5 | 1147 | if (WARN_ON_ONCE(!data->free_work || !data->do_work)) |
e9fd9396 | 1148 | return ERR_PTR(-EINVAL); |
e6ab8991 PB |
1149 | if (WARN_ON_ONCE(!bounded)) |
1150 | return ERR_PTR(-EINVAL); | |
e9fd9396 | 1151 | |
da64d6db | 1152 | wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL); |
771b53d0 JA |
1153 | if (!wq) |
1154 | return ERR_PTR(-ENOMEM); | |
771b53d0 | 1155 | |
e941894e JA |
1156 | refcount_inc(&data->hash->refs); |
1157 | wq->hash = data->hash; | |
e9fd9396 | 1158 | wq->free_work = data->free_work; |
f5fa38c5 | 1159 | wq->do_work = data->do_work; |
7d723065 | 1160 | |
43c01fbe | 1161 | ret = -ENOMEM; |
da64d6db | 1162 | |
eb47943f | 1163 | if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL)) |
da64d6db | 1164 | goto err; |
eb47943f | 1165 | cpumask_copy(wq->cpu_mask, cpu_possible_mask); |
dfd63baf GKB |
1166 | wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; |
1167 | wq->acct[IO_WQ_ACCT_UNBOUND].max_workers = | |
da64d6db | 1168 | task_rlimit(current, RLIMIT_NPROC); |
eb47943f GKB |
1169 | INIT_LIST_HEAD(&wq->wait.entry); |
1170 | wq->wait.func = io_wq_hash_wake; | |
da64d6db | 1171 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
dfd63baf | 1172 | struct io_wq_acct *acct = &wq->acct[i]; |
da64d6db BL |
1173 | |
1174 | acct->index = i; | |
1175 | atomic_set(&acct->nr_running, 0); | |
1176 | INIT_WQ_LIST(&acct->work_list); | |
1177 | raw_spin_lock_init(&acct->lock); | |
771b53d0 | 1178 | } |
eb47943f GKB |
1179 | |
1180 | raw_spin_lock_init(&wq->lock); | |
1181 | INIT_HLIST_NULLS_HEAD(&wq->free_list, 0); | |
1182 | INIT_LIST_HEAD(&wq->all_list); | |
771b53d0 | 1183 | |
685fe7fe | 1184 | wq->task = get_task_struct(data->task); |
685fe7fe JA |
1185 | atomic_set(&wq->worker_refs, 1); |
1186 | init_completion(&wq->worker_done); | |
0f8baa3c JM |
1187 | ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
1188 | if (ret) | |
1189 | goto err; | |
1190 | ||
685fe7fe | 1191 | return wq; |
b60fda60 | 1192 | err: |
dc7bbc9e | 1193 | io_wq_put_hash(data->hash); |
eb47943f | 1194 | free_cpumask_var(wq->cpu_mask); |
b60fda60 | 1195 | kfree(wq); |
771b53d0 JA |
1196 | return ERR_PTR(ret); |
1197 | } | |
1198 | ||
c80ca470 JA |
1199 | static bool io_task_work_match(struct callback_head *cb, void *data) |
1200 | { | |
d3e9f732 | 1201 | struct io_worker *worker; |
c80ca470 | 1202 | |
3b33e3f4 | 1203 | if (cb->func != create_worker_cb && cb->func != create_worker_cont) |
c80ca470 | 1204 | return false; |
d3e9f732 | 1205 | worker = container_of(cb, struct io_worker, create_work); |
eb47943f | 1206 | return worker->wq == data; |
c80ca470 JA |
1207 | } |
1208 | ||
17a91051 PB |
1209 | void io_wq_exit_start(struct io_wq *wq) |
1210 | { | |
1211 | set_bit(IO_WQ_BIT_EXIT, &wq->state); | |
1212 | } | |
1213 | ||
71a85387 | 1214 | static void io_wq_cancel_tw_create(struct io_wq *wq) |
afcc4015 | 1215 | { |
685fe7fe | 1216 | struct callback_head *cb; |
685fe7fe | 1217 | |
c80ca470 | 1218 | while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) { |
d3e9f732 | 1219 | struct io_worker *worker; |
685fe7fe | 1220 | |
d3e9f732 | 1221 | worker = container_of(cb, struct io_worker, create_work); |
1d5f5ea7 | 1222 | io_worker_cancel_cb(worker); |
e6db6f93 JA |
1223 | /* |
1224 | * Only the worker continuation helper has worker allocated and | |
1225 | * hence needs freeing. | |
1226 | */ | |
1227 | if (cb->func == create_worker_cont) | |
1228 | kfree(worker); | |
685fe7fe | 1229 | } |
71a85387 JA |
1230 | } |
1231 | ||
1232 | static void io_wq_exit_workers(struct io_wq *wq) | |
1233 | { | |
71a85387 JA |
1234 | if (!wq->task) |
1235 | return; | |
1236 | ||
1237 | io_wq_cancel_tw_create(wq); | |
685fe7fe JA |
1238 | |
1239 | rcu_read_lock(); | |
eb47943f | 1240 | io_wq_for_each_worker(wq, io_wq_worker_wake, NULL); |
685fe7fe JA |
1241 | rcu_read_unlock(); |
1242 | io_worker_ref_put(wq); | |
1243 | wait_for_completion(&wq->worker_done); | |
3743c172 | 1244 | |
da64d6db | 1245 | spin_lock_irq(&wq->hash->wait.lock); |
eb47943f | 1246 | list_del_init(&wq->wait.entry); |
da64d6db BL |
1247 | spin_unlock_irq(&wq->hash->wait.lock); |
1248 | ||
685fe7fe JA |
1249 | put_task_struct(wq->task); |
1250 | wq->task = NULL; | |
afcc4015 JA |
1251 | } |
1252 | ||
4fb6ac32 | 1253 | static void io_wq_destroy(struct io_wq *wq) |
771b53d0 | 1254 | { |
da64d6db BL |
1255 | struct io_cb_cancel_data match = { |
1256 | .fn = io_wq_work_match_all, | |
1257 | .cancel_all = true, | |
1258 | }; | |
771b53d0 | 1259 | |
43c01fbe | 1260 | cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); |
eb47943f GKB |
1261 | io_wq_cancel_pending_work(wq, &match); |
1262 | free_cpumask_var(wq->cpu_mask); | |
e941894e | 1263 | io_wq_put_hash(wq->hash); |
771b53d0 | 1264 | kfree(wq); |
4fb6ac32 JA |
1265 | } |
1266 | ||
afcc4015 JA |
1267 | void io_wq_put_and_exit(struct io_wq *wq) |
1268 | { | |
17a91051 PB |
1269 | WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state)); |
1270 | ||
685fe7fe | 1271 | io_wq_exit_workers(wq); |
382cb030 | 1272 | io_wq_destroy(wq); |
afcc4015 JA |
1273 | } |
1274 | ||
0e03496d JA |
1275 | struct online_data { |
1276 | unsigned int cpu; | |
1277 | bool online; | |
1278 | }; | |
1279 | ||
43c01fbe JA |
1280 | static bool io_wq_worker_affinity(struct io_worker *worker, void *data) |
1281 | { | |
0e03496d | 1282 | struct online_data *od = data; |
e0051d7d | 1283 | |
0e03496d | 1284 | if (od->online) |
eb47943f | 1285 | cpumask_set_cpu(od->cpu, worker->wq->cpu_mask); |
0e03496d | 1286 | else |
eb47943f | 1287 | cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask); |
43c01fbe JA |
1288 | return false; |
1289 | } | |
1290 | ||
0e03496d | 1291 | static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online) |
43c01fbe | 1292 | { |
0e03496d JA |
1293 | struct online_data od = { |
1294 | .cpu = cpu, | |
1295 | .online = online | |
1296 | }; | |
43c01fbe JA |
1297 | |
1298 | rcu_read_lock(); | |
eb47943f | 1299 | io_wq_for_each_worker(wq, io_wq_worker_affinity, &od); |
43c01fbe JA |
1300 | rcu_read_unlock(); |
1301 | return 0; | |
1302 | } | |
1303 | ||
0e03496d JA |
1304 | static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node) |
1305 | { | |
1306 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1307 | ||
1308 | return __io_wq_cpu_online(wq, cpu, true); | |
1309 | } | |
1310 | ||
1311 | static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node) | |
1312 | { | |
1313 | struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); | |
1314 | ||
1315 | return __io_wq_cpu_online(wq, cpu, false); | |
1316 | } | |
1317 | ||
ebdfefc0 | 1318 | int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask) |
fe76421d | 1319 | { |
ebdfefc0 JA |
1320 | if (!tctx || !tctx->io_wq) |
1321 | return -EINVAL; | |
1322 | ||
fe76421d | 1323 | rcu_read_lock(); |
da64d6db | 1324 | if (mask) |
ebdfefc0 | 1325 | cpumask_copy(tctx->io_wq->cpu_mask, mask); |
da64d6db | 1326 | else |
ebdfefc0 | 1327 | cpumask_copy(tctx->io_wq->cpu_mask, cpu_possible_mask); |
fe76421d | 1328 | rcu_read_unlock(); |
da64d6db | 1329 | |
fe76421d JA |
1330 | return 0; |
1331 | } | |
1332 | ||
2e480058 JA |
1333 | /* |
1334 | * Set max number of unbounded workers, returns old value. If new_count is 0, | |
1335 | * then just return the old value. | |
1336 | */ | |
1337 | int io_wq_max_workers(struct io_wq *wq, int *new_count) | |
1338 | { | |
dfd63baf | 1339 | struct io_wq_acct *acct; |
71c9ce27 | 1340 | int prev[IO_WQ_ACCT_NR]; |
da64d6db | 1341 | int i; |
2e480058 | 1342 | |
dd47c104 ES |
1343 | BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND); |
1344 | BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND); | |
1345 | BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2); | |
1346 | ||
86127bb1 | 1347 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
2e480058 JA |
1348 | if (new_count[i] > task_rlimit(current, RLIMIT_NPROC)) |
1349 | new_count[i] = task_rlimit(current, RLIMIT_NPROC); | |
1350 | } | |
1351 | ||
71c9ce27 BZ |
1352 | for (i = 0; i < IO_WQ_ACCT_NR; i++) |
1353 | prev[i] = 0; | |
1354 | ||
2e480058 | 1355 | rcu_read_lock(); |
2e480058 | 1356 | |
eb47943f | 1357 | raw_spin_lock(&wq->lock); |
da64d6db | 1358 | for (i = 0; i < IO_WQ_ACCT_NR; i++) { |
dfd63baf | 1359 | acct = &wq->acct[i]; |
da64d6db BL |
1360 | prev[i] = max_t(int, acct->max_workers, prev[i]); |
1361 | if (new_count[i]) | |
1362 | acct->max_workers = new_count[i]; | |
2e480058 | 1363 | } |
eb47943f | 1364 | raw_spin_unlock(&wq->lock); |
2e480058 | 1365 | rcu_read_unlock(); |
71c9ce27 BZ |
1366 | |
1367 | for (i = 0; i < IO_WQ_ACCT_NR; i++) | |
1368 | new_count[i] = prev[i]; | |
1369 | ||
2e480058 JA |
1370 | return 0; |
1371 | } | |
1372 | ||
43c01fbe JA |
1373 | static __init int io_wq_init(void) |
1374 | { | |
1375 | int ret; | |
1376 | ||
1377 | ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online", | |
0e03496d | 1378 | io_wq_cpu_online, io_wq_cpu_offline); |
43c01fbe JA |
1379 | if (ret < 0) |
1380 | return ret; | |
1381 | io_wq_online = ret; | |
1382 | return 0; | |
1383 | } | |
1384 | subsys_initcall(io_wq_init); |