]>
Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
d068b506 | 14 | * through a control-dependency in io_get_cqe (smp_store_release to |
1e84b97b SB |
15 | * store head will do). Failure to do so could lead to reading invalid |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
46 | #include <linux/compat.h> | |
52de1fe1 | 47 | #include <net/compat.h> |
2b188cc1 JA |
48 | #include <linux/refcount.h> |
49 | #include <linux/uio.h> | |
6b47ee6e | 50 | #include <linux/bits.h> |
2b188cc1 JA |
51 | |
52 | #include <linux/sched/signal.h> | |
53 | #include <linux/fs.h> | |
54 | #include <linux/file.h> | |
55 | #include <linux/fdtable.h> | |
56 | #include <linux/mm.h> | |
57 | #include <linux/mman.h> | |
2b188cc1 JA |
58 | #include <linux/percpu.h> |
59 | #include <linux/slab.h> | |
2b188cc1 | 60 | #include <linux/blkdev.h> |
edafccee | 61 | #include <linux/bvec.h> |
2b188cc1 JA |
62 | #include <linux/net.h> |
63 | #include <net/sock.h> | |
64 | #include <net/af_unix.h> | |
6b06314c | 65 | #include <net/scm.h> |
2b188cc1 JA |
66 | #include <linux/anon_inodes.h> |
67 | #include <linux/sched/mm.h> | |
68 | #include <linux/uaccess.h> | |
69 | #include <linux/nospec.h> | |
edafccee JA |
70 | #include <linux/sizes.h> |
71 | #include <linux/hugetlb.h> | |
aa4c3967 | 72 | #include <linux/highmem.h> |
15b71abe JA |
73 | #include <linux/namei.h> |
74 | #include <linux/fsnotify.h> | |
4840e418 | 75 | #include <linux/fadvise.h> |
3e4827b0 | 76 | #include <linux/eventpoll.h> |
7d67af2c | 77 | #include <linux/splice.h> |
b41e9852 | 78 | #include <linux/task_work.h> |
bcf5a063 | 79 | #include <linux/pagemap.h> |
0f212204 | 80 | #include <linux/io_uring.h> |
ef98eb04 | 81 | #include <linux/tracehook.h> |
2b188cc1 | 82 | |
c826bd7a DD |
83 | #define CREATE_TRACE_POINTS |
84 | #include <trace/events/io_uring.h> | |
85 | ||
2b188cc1 JA |
86 | #include <uapi/linux/io_uring.h> |
87 | ||
88 | #include "internal.h" | |
561fb04a | 89 | #include "io-wq.h" |
2b188cc1 | 90 | |
5277deaa | 91 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 92 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
4ce8ad95 | 93 | #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8 |
65e19f54 | 94 | |
042b0d85 PB |
95 | /* 512 entries per page on 64-bit archs, 64 pages max */ |
96 | #define IORING_MAX_FIXED_FILES (1U << 15) | |
21b55dbc SG |
97 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
98 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 | 99 | |
2d091d62 PB |
100 | #define IO_RSRC_TAG_TABLE_SHIFT 9 |
101 | #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT) | |
102 | #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1) | |
103 | ||
489809e2 PB |
104 | #define IORING_MAX_REG_BUFFERS (1U << 14) |
105 | ||
b16fed66 PB |
106 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
107 | IOSQE_IO_HARDLINK | IOSQE_ASYNC | \ | |
108 | IOSQE_BUFFER_SELECT) | |
c854357b PB |
109 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
110 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS) | |
b16fed66 | 111 | |
09899b19 PB |
112 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
113 | ||
2b188cc1 JA |
114 | struct io_uring { |
115 | u32 head ____cacheline_aligned_in_smp; | |
116 | u32 tail ____cacheline_aligned_in_smp; | |
117 | }; | |
118 | ||
1e84b97b | 119 | /* |
75b28aff HV |
120 | * This data is shared with the application through the mmap at offsets |
121 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
122 | * |
123 | * The offsets to the member fields are published through struct | |
124 | * io_sqring_offsets when calling io_uring_setup. | |
125 | */ | |
75b28aff | 126 | struct io_rings { |
1e84b97b SB |
127 | /* |
128 | * Head and tail offsets into the ring; the offsets need to be | |
129 | * masked to get valid indices. | |
130 | * | |
75b28aff HV |
131 | * The kernel controls head of the sq ring and the tail of the cq ring, |
132 | * and the application controls tail of the sq ring and the head of the | |
133 | * cq ring. | |
1e84b97b | 134 | */ |
75b28aff | 135 | struct io_uring sq, cq; |
1e84b97b | 136 | /* |
75b28aff | 137 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
138 | * ring_entries - 1) |
139 | */ | |
75b28aff HV |
140 | u32 sq_ring_mask, cq_ring_mask; |
141 | /* Ring sizes (constant, power of 2) */ | |
142 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
143 | /* |
144 | * Number of invalid entries dropped by the kernel due to | |
145 | * invalid index stored in array | |
146 | * | |
147 | * Written by the kernel, shouldn't be modified by the | |
148 | * application (i.e. get number of "new events" by comparing to | |
149 | * cached value). | |
150 | * | |
151 | * After a new SQ head value was read by the application this | |
152 | * counter includes all submissions that were dropped reaching | |
153 | * the new SQ head (and possibly more). | |
154 | */ | |
75b28aff | 155 | u32 sq_dropped; |
1e84b97b | 156 | /* |
0d9b5b3a | 157 | * Runtime SQ flags |
1e84b97b SB |
158 | * |
159 | * Written by the kernel, shouldn't be modified by the | |
160 | * application. | |
161 | * | |
162 | * The application needs a full memory barrier before checking | |
163 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
164 | */ | |
75b28aff | 165 | u32 sq_flags; |
0d9b5b3a SG |
166 | /* |
167 | * Runtime CQ flags | |
168 | * | |
169 | * Written by the application, shouldn't be modified by the | |
170 | * kernel. | |
171 | */ | |
fe7e3257 | 172 | u32 cq_flags; |
1e84b97b SB |
173 | /* |
174 | * Number of completion events lost because the queue was full; | |
175 | * this should be avoided by the application by making sure | |
0b4295b5 | 176 | * there are not more requests pending than there is space in |
1e84b97b SB |
177 | * the completion queue. |
178 | * | |
179 | * Written by the kernel, shouldn't be modified by the | |
180 | * application (i.e. get number of "new events" by comparing to | |
181 | * cached value). | |
182 | * | |
183 | * As completion events come in out of order this counter is not | |
184 | * ordered with any other data. | |
185 | */ | |
75b28aff | 186 | u32 cq_overflow; |
1e84b97b SB |
187 | /* |
188 | * Ring buffer of completion events. | |
189 | * | |
190 | * The kernel writes completion events fresh every time they are | |
191 | * produced, so the application is allowed to modify pending | |
192 | * entries. | |
193 | */ | |
75b28aff | 194 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
195 | }; |
196 | ||
45d189c6 PB |
197 | enum io_uring_cmd_flags { |
198 | IO_URING_F_NONBLOCK = 1, | |
889fca73 | 199 | IO_URING_F_COMPLETE_DEFER = 2, |
45d189c6 PB |
200 | }; |
201 | ||
edafccee JA |
202 | struct io_mapped_ubuf { |
203 | u64 ubuf; | |
4751f53d | 204 | u64 ubuf_end; |
edafccee | 205 | unsigned int nr_bvecs; |
de293938 | 206 | unsigned long acct_pages; |
41edf1a5 | 207 | struct bio_vec bvec[]; |
edafccee JA |
208 | }; |
209 | ||
50238531 BM |
210 | struct io_ring_ctx; |
211 | ||
6c2450ae PB |
212 | struct io_overflow_cqe { |
213 | struct io_uring_cqe cqe; | |
214 | struct list_head list; | |
215 | }; | |
216 | ||
a04b0ac0 PB |
217 | struct io_fixed_file { |
218 | /* file * with additional FFS_* flags */ | |
219 | unsigned long file_ptr; | |
220 | }; | |
221 | ||
269bbe5f BM |
222 | struct io_rsrc_put { |
223 | struct list_head list; | |
b60c8dce | 224 | u64 tag; |
50238531 BM |
225 | union { |
226 | void *rsrc; | |
227 | struct file *file; | |
bd54b6fe | 228 | struct io_mapped_ubuf *buf; |
50238531 | 229 | }; |
269bbe5f BM |
230 | }; |
231 | ||
aeca241b | 232 | struct io_file_table { |
042b0d85 | 233 | struct io_fixed_file *files; |
31b51510 JA |
234 | }; |
235 | ||
b895c9a6 | 236 | struct io_rsrc_node { |
05589553 XW |
237 | struct percpu_ref refs; |
238 | struct list_head node; | |
269bbe5f | 239 | struct list_head rsrc_list; |
b895c9a6 | 240 | struct io_rsrc_data *rsrc_data; |
4a38aed2 | 241 | struct llist_node llist; |
e297822b | 242 | bool done; |
05589553 XW |
243 | }; |
244 | ||
40ae0ff7 PB |
245 | typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc); |
246 | ||
b895c9a6 | 247 | struct io_rsrc_data { |
05f3fb3c JA |
248 | struct io_ring_ctx *ctx; |
249 | ||
2d091d62 PB |
250 | u64 **tags; |
251 | unsigned int nr; | |
40ae0ff7 | 252 | rsrc_put_fn *do_put; |
3e942498 | 253 | atomic_t refs; |
05f3fb3c | 254 | struct completion done; |
8bad28d8 | 255 | bool quiesce; |
05f3fb3c JA |
256 | }; |
257 | ||
5a2e745d JA |
258 | struct io_buffer { |
259 | struct list_head list; | |
260 | __u64 addr; | |
d1f82808 | 261 | __u32 len; |
5a2e745d JA |
262 | __u16 bid; |
263 | }; | |
264 | ||
21b55dbc SG |
265 | struct io_restriction { |
266 | DECLARE_BITMAP(register_op, IORING_REGISTER_LAST); | |
267 | DECLARE_BITMAP(sqe_op, IORING_OP_LAST); | |
268 | u8 sqe_flags_allowed; | |
269 | u8 sqe_flags_required; | |
7e84e1c7 | 270 | bool registered; |
21b55dbc SG |
271 | }; |
272 | ||
37d1e2e3 JA |
273 | enum { |
274 | IO_SQ_THREAD_SHOULD_STOP = 0, | |
275 | IO_SQ_THREAD_SHOULD_PARK, | |
276 | }; | |
277 | ||
534ca6d6 JA |
278 | struct io_sq_data { |
279 | refcount_t refs; | |
9e138a48 | 280 | atomic_t park_pending; |
09a6f4ef | 281 | struct mutex lock; |
69fb2131 JA |
282 | |
283 | /* ctx's that are using this sqd */ | |
284 | struct list_head ctx_list; | |
69fb2131 | 285 | |
534ca6d6 JA |
286 | struct task_struct *thread; |
287 | struct wait_queue_head wait; | |
08369246 XW |
288 | |
289 | unsigned sq_thread_idle; | |
37d1e2e3 JA |
290 | int sq_cpu; |
291 | pid_t task_pid; | |
5c2469e0 | 292 | pid_t task_tgid; |
37d1e2e3 JA |
293 | |
294 | unsigned long state; | |
37d1e2e3 | 295 | struct completion exited; |
534ca6d6 JA |
296 | }; |
297 | ||
6dd0be1e | 298 | #define IO_COMPL_BATCH 32 |
6ff119a6 | 299 | #define IO_REQ_CACHE_SIZE 32 |
bf019da7 | 300 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 301 | |
a1ab7b35 PB |
302 | struct io_submit_link { |
303 | struct io_kiocb *head; | |
304 | struct io_kiocb *last; | |
305 | }; | |
306 | ||
258b29a9 PB |
307 | struct io_submit_state { |
308 | struct blk_plug plug; | |
a1ab7b35 | 309 | struct io_submit_link link; |
258b29a9 PB |
310 | |
311 | /* | |
312 | * io_kiocb alloc cache | |
313 | */ | |
bf019da7 | 314 | void *reqs[IO_REQ_CACHE_SIZE]; |
258b29a9 PB |
315 | unsigned int free_reqs; |
316 | ||
317 | bool plug_started; | |
318 | ||
319 | /* | |
320 | * Batch completion logic | |
321 | */ | |
cd0ca2e0 PB |
322 | struct io_kiocb *compl_reqs[IO_COMPL_BATCH]; |
323 | unsigned int compl_nr; | |
324 | /* inline/task_work completion list, under ->uring_lock */ | |
325 | struct list_head free_list; | |
258b29a9 | 326 | |
258b29a9 PB |
327 | unsigned int ios_left; |
328 | }; | |
329 | ||
2b188cc1 | 330 | struct io_ring_ctx { |
b52ecf8c | 331 | /* const or read-mostly hot data */ |
2b188cc1 JA |
332 | struct { |
333 | struct percpu_ref refs; | |
2b188cc1 | 334 | |
b52ecf8c | 335 | struct io_rings *rings; |
2b188cc1 | 336 | unsigned int flags; |
e1d85334 | 337 | unsigned int compat: 1; |
e1d85334 RD |
338 | unsigned int drain_next: 1; |
339 | unsigned int eventfd_async: 1; | |
21b55dbc | 340 | unsigned int restricted: 1; |
f18ee4cf | 341 | unsigned int off_timeout_used: 1; |
10c66904 | 342 | unsigned int drain_active: 1; |
b52ecf8c | 343 | } ____cacheline_aligned_in_smp; |
2b188cc1 | 344 | |
7f1129d2 | 345 | /* submission data */ |
b52ecf8c | 346 | struct { |
0499e582 PB |
347 | struct mutex uring_lock; |
348 | ||
75b28aff HV |
349 | /* |
350 | * Ring buffer of indices into array of io_uring_sqe, which is | |
351 | * mmapped by the application using the IORING_OFF_SQES offset. | |
352 | * | |
353 | * This indirection could e.g. be used to assign fixed | |
354 | * io_uring_sqe entries to operations and only submit them to | |
355 | * the queue when needed. | |
356 | * | |
357 | * The kernel modifies neither the indices array nor the entries | |
358 | * array. | |
359 | */ | |
360 | u32 *sq_array; | |
c7af47cf | 361 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
362 | unsigned cached_sq_head; |
363 | unsigned sq_entries; | |
de0617e4 | 364 | struct list_head defer_list; |
7f1129d2 PB |
365 | |
366 | /* | |
367 | * Fixed resources fast path, should be accessed only under | |
368 | * uring_lock, and updated through io_uring_register(2) | |
369 | */ | |
370 | struct io_rsrc_node *rsrc_node; | |
371 | struct io_file_table file_table; | |
372 | unsigned nr_user_files; | |
373 | unsigned nr_user_bufs; | |
374 | struct io_mapped_ubuf **user_bufs; | |
375 | ||
376 | struct io_submit_state submit_state; | |
5262f567 | 377 | struct list_head timeout_list; |
1d7bb1d5 | 378 | struct list_head cq_overflow_list; |
7f1129d2 PB |
379 | struct xarray io_buffers; |
380 | struct xarray personalities; | |
381 | u32 pers_next; | |
382 | unsigned sq_thread_idle; | |
2b188cc1 JA |
383 | } ____cacheline_aligned_in_smp; |
384 | ||
d0acdee2 PB |
385 | /* IRQ completion list, under ->completion_lock */ |
386 | struct list_head locked_free_list; | |
387 | unsigned int locked_free_nr; | |
3c1a2ead | 388 | |
7c30f36a | 389 | const struct cred *sq_creds; /* cred used for __io_sq_thread() */ |
534ca6d6 JA |
390 | struct io_sq_data *sq_data; /* if using sq thread polling */ |
391 | ||
90554200 | 392 | struct wait_queue_head sqo_sq_wait; |
69fb2131 | 393 | struct list_head sqd_list; |
75b28aff | 394 | |
5ed7a37d PB |
395 | unsigned long check_cq_overflow; |
396 | ||
206aefde JA |
397 | struct { |
398 | unsigned cached_cq_tail; | |
399 | unsigned cq_entries; | |
0499e582 | 400 | struct eventfd_ctx *cq_ev_fd; |
311997b3 | 401 | struct wait_queue_head poll_wait; |
0499e582 PB |
402 | struct wait_queue_head cq_wait; |
403 | unsigned cq_extra; | |
404 | atomic_t cq_timeouts; | |
206aefde | 405 | struct fasync_struct *cq_fasync; |
0499e582 | 406 | unsigned cq_last_tm_flush; |
206aefde | 407 | } ____cacheline_aligned_in_smp; |
2b188cc1 | 408 | |
2b188cc1 JA |
409 | struct { |
410 | spinlock_t completion_lock; | |
e94f141b | 411 | |
def596e9 | 412 | /* |
540e32a0 | 413 | * ->iopoll_list is protected by the ctx->uring_lock for |
def596e9 JA |
414 | * io_uring instances that don't use IORING_SETUP_SQPOLL. |
415 | * For SQPOLL, only the single threaded io_sq_thread() will | |
416 | * manipulate the list, hence no extra locking is needed there. | |
417 | */ | |
540e32a0 | 418 | struct list_head iopoll_list; |
78076bb6 JA |
419 | struct hlist_head *cancel_hash; |
420 | unsigned cancel_hash_bits; | |
915b3dde | 421 | bool poll_multi_queue; |
2b188cc1 | 422 | } ____cacheline_aligned_in_smp; |
85faa7b8 | 423 | |
21b55dbc | 424 | struct io_restriction restrictions; |
3c1a2ead | 425 | |
b13a8918 PB |
426 | /* slow path rsrc auxilary data, used by update/register */ |
427 | struct { | |
428 | struct io_rsrc_node *rsrc_backup_node; | |
429 | struct io_mapped_ubuf *dummy_ubuf; | |
430 | struct io_rsrc_data *file_data; | |
431 | struct io_rsrc_data *buf_data; | |
432 | ||
433 | struct delayed_work rsrc_put_work; | |
434 | struct llist_head rsrc_put_llist; | |
435 | struct list_head rsrc_ref_list; | |
436 | spinlock_t rsrc_ref_lock; | |
437 | }; | |
438 | ||
3c1a2ead | 439 | /* Keep this last, we don't need it for the fast path */ |
b986af7e PB |
440 | struct { |
441 | #if defined(CONFIG_UNIX) | |
442 | struct socket *ring_sock; | |
443 | #endif | |
444 | /* hashed buffered write serialization */ | |
445 | struct io_wq_hash *hash_map; | |
446 | ||
447 | /* Only used for accounting purposes */ | |
448 | struct user_struct *user; | |
449 | struct mm_struct *mm_account; | |
450 | ||
451 | /* ctx exit and cancelation */ | |
9011bf9a PB |
452 | struct llist_head fallback_llist; |
453 | struct delayed_work fallback_work; | |
b986af7e PB |
454 | struct work_struct exit_work; |
455 | struct list_head tctx_list; | |
456 | struct completion ref_comp; | |
457 | }; | |
2b188cc1 JA |
458 | }; |
459 | ||
53e043b2 SM |
460 | struct io_uring_task { |
461 | /* submission side */ | |
09899b19 | 462 | int cached_refs; |
53e043b2 SM |
463 | struct xarray xa; |
464 | struct wait_queue_head wait; | |
ee53fb2b SM |
465 | const struct io_ring_ctx *last; |
466 | struct io_wq *io_wq; | |
53e043b2 | 467 | struct percpu_counter inflight; |
b303fe2e | 468 | atomic_t inflight_tracked; |
53e043b2 | 469 | atomic_t in_idle; |
53e043b2 SM |
470 | |
471 | spinlock_t task_lock; | |
472 | struct io_wq_work_list task_list; | |
53e043b2 | 473 | struct callback_head task_work; |
6294f368 | 474 | bool task_running; |
53e043b2 SM |
475 | }; |
476 | ||
09bb8394 JA |
477 | /* |
478 | * First field must be the file pointer in all the | |
479 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
480 | */ | |
221c5eb2 JA |
481 | struct io_poll_iocb { |
482 | struct file *file; | |
018043be | 483 | struct wait_queue_head *head; |
221c5eb2 | 484 | __poll_t events; |
8c838788 | 485 | bool done; |
221c5eb2 | 486 | bool canceled; |
392edb45 | 487 | struct wait_queue_entry wait; |
221c5eb2 JA |
488 | }; |
489 | ||
9d805892 | 490 | struct io_poll_update { |
018043be | 491 | struct file *file; |
9d805892 PB |
492 | u64 old_user_data; |
493 | u64 new_user_data; | |
494 | __poll_t events; | |
b69de288 JA |
495 | bool update_events; |
496 | bool update_user_data; | |
018043be PB |
497 | }; |
498 | ||
b5dba59e JA |
499 | struct io_close { |
500 | struct file *file; | |
b5dba59e JA |
501 | int fd; |
502 | }; | |
503 | ||
ad8a48ac JA |
504 | struct io_timeout_data { |
505 | struct io_kiocb *req; | |
506 | struct hrtimer timer; | |
507 | struct timespec64 ts; | |
508 | enum hrtimer_mode mode; | |
509 | }; | |
510 | ||
8ed8d3c3 JA |
511 | struct io_accept { |
512 | struct file *file; | |
513 | struct sockaddr __user *addr; | |
514 | int __user *addr_len; | |
515 | int flags; | |
09952e3e | 516 | unsigned long nofile; |
8ed8d3c3 JA |
517 | }; |
518 | ||
519 | struct io_sync { | |
520 | struct file *file; | |
521 | loff_t len; | |
522 | loff_t off; | |
523 | int flags; | |
d63d1b5e | 524 | int mode; |
8ed8d3c3 JA |
525 | }; |
526 | ||
fbf23849 JA |
527 | struct io_cancel { |
528 | struct file *file; | |
529 | u64 addr; | |
530 | }; | |
531 | ||
b29472ee JA |
532 | struct io_timeout { |
533 | struct file *file; | |
bfe68a22 PB |
534 | u32 off; |
535 | u32 target_seq; | |
135fcde8 | 536 | struct list_head list; |
90cd7e42 PB |
537 | /* head of the link, used by linked timeouts only */ |
538 | struct io_kiocb *head; | |
b29472ee JA |
539 | }; |
540 | ||
0bdf7a2d PB |
541 | struct io_timeout_rem { |
542 | struct file *file; | |
543 | u64 addr; | |
9c8e11b3 PB |
544 | |
545 | /* timeout update */ | |
546 | struct timespec64 ts; | |
547 | u32 flags; | |
0bdf7a2d PB |
548 | }; |
549 | ||
9adbd45d JA |
550 | struct io_rw { |
551 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
552 | struct kiocb kiocb; | |
553 | u64 addr; | |
554 | u64 len; | |
555 | }; | |
556 | ||
3fbb51c1 JA |
557 | struct io_connect { |
558 | struct file *file; | |
559 | struct sockaddr __user *addr; | |
560 | int addr_len; | |
561 | }; | |
562 | ||
e47293fd JA |
563 | struct io_sr_msg { |
564 | struct file *file; | |
fddaface | 565 | union { |
4af3417a PB |
566 | struct compat_msghdr __user *umsg_compat; |
567 | struct user_msghdr __user *umsg; | |
568 | void __user *buf; | |
fddaface | 569 | }; |
e47293fd | 570 | int msg_flags; |
bcda7baa | 571 | int bgid; |
fddaface | 572 | size_t len; |
bcda7baa | 573 | struct io_buffer *kbuf; |
e47293fd JA |
574 | }; |
575 | ||
15b71abe JA |
576 | struct io_open { |
577 | struct file *file; | |
578 | int dfd; | |
15b71abe | 579 | struct filename *filename; |
c12cedf2 | 580 | struct open_how how; |
4022e7af | 581 | unsigned long nofile; |
15b71abe JA |
582 | }; |
583 | ||
269bbe5f | 584 | struct io_rsrc_update { |
05f3fb3c JA |
585 | struct file *file; |
586 | u64 arg; | |
587 | u32 nr_args; | |
588 | u32 offset; | |
589 | }; | |
590 | ||
4840e418 JA |
591 | struct io_fadvise { |
592 | struct file *file; | |
593 | u64 offset; | |
594 | u32 len; | |
595 | u32 advice; | |
596 | }; | |
597 | ||
c1ca757b JA |
598 | struct io_madvise { |
599 | struct file *file; | |
600 | u64 addr; | |
601 | u32 len; | |
602 | u32 advice; | |
603 | }; | |
604 | ||
3e4827b0 JA |
605 | struct io_epoll { |
606 | struct file *file; | |
607 | int epfd; | |
608 | int op; | |
609 | int fd; | |
610 | struct epoll_event event; | |
e47293fd JA |
611 | }; |
612 | ||
7d67af2c PB |
613 | struct io_splice { |
614 | struct file *file_out; | |
615 | struct file *file_in; | |
616 | loff_t off_out; | |
617 | loff_t off_in; | |
618 | u64 len; | |
619 | unsigned int flags; | |
620 | }; | |
621 | ||
ddf0322d JA |
622 | struct io_provide_buf { |
623 | struct file *file; | |
624 | __u64 addr; | |
38134ada | 625 | __u32 len; |
ddf0322d JA |
626 | __u32 bgid; |
627 | __u16 nbufs; | |
628 | __u16 bid; | |
629 | }; | |
630 | ||
1d9e1288 BM |
631 | struct io_statx { |
632 | struct file *file; | |
633 | int dfd; | |
634 | unsigned int mask; | |
635 | unsigned int flags; | |
e62753e4 | 636 | const char __user *filename; |
1d9e1288 BM |
637 | struct statx __user *buffer; |
638 | }; | |
639 | ||
36f4fa68 JA |
640 | struct io_shutdown { |
641 | struct file *file; | |
642 | int how; | |
643 | }; | |
644 | ||
80a261fd JA |
645 | struct io_rename { |
646 | struct file *file; | |
647 | int old_dfd; | |
648 | int new_dfd; | |
649 | struct filename *oldpath; | |
650 | struct filename *newpath; | |
651 | int flags; | |
652 | }; | |
653 | ||
14a1143b JA |
654 | struct io_unlink { |
655 | struct file *file; | |
656 | int dfd; | |
657 | int flags; | |
658 | struct filename *filename; | |
659 | }; | |
660 | ||
3ca405eb PB |
661 | struct io_completion { |
662 | struct file *file; | |
8c3f9cd1 | 663 | u32 cflags; |
3ca405eb PB |
664 | }; |
665 | ||
f499a021 JA |
666 | struct io_async_connect { |
667 | struct sockaddr_storage address; | |
668 | }; | |
669 | ||
03b1230c JA |
670 | struct io_async_msghdr { |
671 | struct iovec fast_iov[UIO_FASTIOV]; | |
257e84a5 PB |
672 | /* points to an allocated iov, if NULL we use fast_iov instead */ |
673 | struct iovec *free_iov; | |
03b1230c JA |
674 | struct sockaddr __user *uaddr; |
675 | struct msghdr msg; | |
b537916c | 676 | struct sockaddr_storage addr; |
03b1230c JA |
677 | }; |
678 | ||
f67676d1 JA |
679 | struct io_async_rw { |
680 | struct iovec fast_iov[UIO_FASTIOV]; | |
ff6165b2 JA |
681 | const struct iovec *free_iovec; |
682 | struct iov_iter iter; | |
227c0c96 | 683 | size_t bytes_done; |
bcf5a063 | 684 | struct wait_page_queue wpq; |
f67676d1 JA |
685 | }; |
686 | ||
6b47ee6e PB |
687 | enum { |
688 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
689 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
690 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
691 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
692 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
bcda7baa | 693 | REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT, |
6b47ee6e | 694 | |
dddca226 | 695 | /* first byte is taken by user flags, shift it to not overlap */ |
93d2bcd2 | 696 | REQ_F_FAIL_BIT = 8, |
6b47ee6e PB |
697 | REQ_F_INFLIGHT_BIT, |
698 | REQ_F_CUR_POS_BIT, | |
699 | REQ_F_NOWAIT_BIT, | |
6b47ee6e | 700 | REQ_F_LINK_TIMEOUT_BIT, |
99bc4c38 | 701 | REQ_F_NEED_CLEANUP_BIT, |
d7718a9d | 702 | REQ_F_POLLED_BIT, |
bcda7baa | 703 | REQ_F_BUFFER_SELECTED_BIT, |
900fad45 | 704 | REQ_F_LTIMEOUT_ACTIVE_BIT, |
e342c807 | 705 | REQ_F_COMPLETE_INLINE_BIT, |
230d50d4 | 706 | REQ_F_REISSUE_BIT, |
8c130827 | 707 | REQ_F_DONT_REISSUE_BIT, |
b8e64b53 | 708 | REQ_F_CREDS_BIT, |
7b29f92d | 709 | /* keep async read/write and isreg together and in order */ |
b191e2df PB |
710 | REQ_F_NOWAIT_READ_BIT, |
711 | REQ_F_NOWAIT_WRITE_BIT, | |
7b29f92d | 712 | REQ_F_ISREG_BIT, |
84557871 JA |
713 | |
714 | /* not a real bit, just to check we're not overflowing the space */ | |
715 | __REQ_F_LAST_BIT, | |
6b47ee6e PB |
716 | }; |
717 | ||
718 | enum { | |
719 | /* ctx owns file */ | |
720 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
721 | /* drain existing IO first */ | |
722 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
723 | /* linked sqes */ | |
724 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
725 | /* doesn't sever on completion < 0 */ | |
726 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
727 | /* IOSQE_ASYNC */ | |
728 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
bcda7baa JA |
729 | /* IOSQE_BUFFER_SELECT */ |
730 | REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT), | |
6b47ee6e | 731 | |
6b47ee6e | 732 | /* fail rest of links */ |
93d2bcd2 | 733 | REQ_F_FAIL = BIT(REQ_F_FAIL_BIT), |
b05a1bcd | 734 | /* on inflight list, should be cancelled and waited on exit reliably */ |
6b47ee6e PB |
735 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), |
736 | /* read/write uses file position */ | |
737 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
738 | /* must not punt to workers */ | |
739 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
900fad45 | 740 | /* has or had linked timeout */ |
6b47ee6e | 741 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), |
99bc4c38 PB |
742 | /* needs cleanup */ |
743 | REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), | |
d7718a9d JA |
744 | /* already went through poll handler */ |
745 | REQ_F_POLLED = BIT(REQ_F_POLLED_BIT), | |
bcda7baa JA |
746 | /* buffer already selected */ |
747 | REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT), | |
900fad45 PB |
748 | /* linked timeout is active, i.e. prepared by link's head */ |
749 | REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT), | |
e342c807 PB |
750 | /* completion is deferred through io_comp_state */ |
751 | REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT), | |
230d50d4 JA |
752 | /* caller should reissue async */ |
753 | REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT), | |
8c130827 PB |
754 | /* don't attempt request reissue, see io_rw_reissue() */ |
755 | REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT), | |
7b29f92d | 756 | /* supports async reads */ |
b191e2df | 757 | REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT), |
7b29f92d | 758 | /* supports async writes */ |
b191e2df | 759 | REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT), |
7b29f92d JA |
760 | /* regular file */ |
761 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
b8e64b53 PB |
762 | /* has creds assigned */ |
763 | REQ_F_CREDS = BIT(REQ_F_CREDS_BIT), | |
d7718a9d JA |
764 | }; |
765 | ||
766 | struct async_poll { | |
767 | struct io_poll_iocb poll; | |
807abcb0 | 768 | struct io_poll_iocb *double_poll; |
6b47ee6e PB |
769 | }; |
770 | ||
5b0a6acc PB |
771 | typedef void (*io_req_tw_func_t)(struct io_kiocb *req); |
772 | ||
7cbf1722 | 773 | struct io_task_work { |
5b0a6acc PB |
774 | union { |
775 | struct io_wq_work_node node; | |
776 | struct llist_node fallback_node; | |
777 | }; | |
778 | io_req_tw_func_t func; | |
7cbf1722 JA |
779 | }; |
780 | ||
992da01a PB |
781 | enum { |
782 | IORING_RSRC_FILE = 0, | |
783 | IORING_RSRC_BUFFER = 1, | |
784 | }; | |
785 | ||
09bb8394 JA |
786 | /* |
787 | * NOTE! Each of the iocb union members has the file pointer | |
788 | * as the first entry in their struct definition. So you can | |
789 | * access the file pointer through any of the sub-structs, | |
790 | * or directly as just 'ki_filp' in this struct. | |
791 | */ | |
2b188cc1 | 792 | struct io_kiocb { |
221c5eb2 | 793 | union { |
09bb8394 | 794 | struct file *file; |
9adbd45d | 795 | struct io_rw rw; |
221c5eb2 | 796 | struct io_poll_iocb poll; |
9d805892 | 797 | struct io_poll_update poll_update; |
8ed8d3c3 JA |
798 | struct io_accept accept; |
799 | struct io_sync sync; | |
fbf23849 | 800 | struct io_cancel cancel; |
b29472ee | 801 | struct io_timeout timeout; |
0bdf7a2d | 802 | struct io_timeout_rem timeout_rem; |
3fbb51c1 | 803 | struct io_connect connect; |
e47293fd | 804 | struct io_sr_msg sr_msg; |
15b71abe | 805 | struct io_open open; |
b5dba59e | 806 | struct io_close close; |
269bbe5f | 807 | struct io_rsrc_update rsrc_update; |
4840e418 | 808 | struct io_fadvise fadvise; |
c1ca757b | 809 | struct io_madvise madvise; |
3e4827b0 | 810 | struct io_epoll epoll; |
7d67af2c | 811 | struct io_splice splice; |
ddf0322d | 812 | struct io_provide_buf pbuf; |
1d9e1288 | 813 | struct io_statx statx; |
36f4fa68 | 814 | struct io_shutdown shutdown; |
80a261fd | 815 | struct io_rename rename; |
14a1143b | 816 | struct io_unlink unlink; |
3ca405eb PB |
817 | /* use only after cleaning per-op data, see io_clean_op() */ |
818 | struct io_completion compl; | |
221c5eb2 | 819 | }; |
2b188cc1 | 820 | |
e8c2bc1f JA |
821 | /* opcode allocated if it needs to store data for async defer */ |
822 | void *async_data; | |
d625c6ee | 823 | u8 opcode; |
65a6543d XW |
824 | /* polled IO has completed */ |
825 | u8 iopoll_completed; | |
2b188cc1 | 826 | |
4f4eeba8 | 827 | u16 buf_index; |
9cf7c104 | 828 | u32 result; |
4f4eeba8 | 829 | |
010e8e6b PB |
830 | struct io_ring_ctx *ctx; |
831 | unsigned int flags; | |
abc54d63 | 832 | atomic_t refs; |
010e8e6b PB |
833 | struct task_struct *task; |
834 | u64 user_data; | |
d7718a9d | 835 | |
f2f87370 | 836 | struct io_kiocb *link; |
269bbe5f | 837 | struct percpu_ref *fixed_rsrc_refs; |
fcb323cc | 838 | |
b303fe2e | 839 | /* used with ctx->iopoll_list with reads/writes */ |
010e8e6b | 840 | struct list_head inflight_entry; |
5b0a6acc | 841 | struct io_task_work io_task_work; |
010e8e6b PB |
842 | /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */ |
843 | struct hlist_node hash_node; | |
844 | struct async_poll *apoll; | |
845 | struct io_wq_work work; | |
fe7e3257 | 846 | const struct cred *creds; |
c10d1f98 | 847 | |
eae071c9 PB |
848 | /* store used ubuf, so we can prevent reloading */ |
849 | struct io_mapped_ubuf *imu; | |
2b188cc1 | 850 | }; |
05589553 | 851 | |
13bf43f5 PB |
852 | struct io_tctx_node { |
853 | struct list_head ctx_node; | |
854 | struct task_struct *task; | |
13bf43f5 PB |
855 | struct io_ring_ctx *ctx; |
856 | }; | |
857 | ||
27dc8338 PB |
858 | struct io_defer_entry { |
859 | struct list_head list; | |
860 | struct io_kiocb *req; | |
9cf7c104 | 861 | u32 seq; |
2b188cc1 JA |
862 | }; |
863 | ||
d3656344 | 864 | struct io_op_def { |
d3656344 JA |
865 | /* needs req->file assigned */ |
866 | unsigned needs_file : 1; | |
d3656344 JA |
867 | /* hash wq insertion if file is a regular file */ |
868 | unsigned hash_reg_file : 1; | |
869 | /* unbound wq insertion if file is a non-regular file */ | |
870 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
871 | /* opcode is not supported by this kernel */ |
872 | unsigned not_supported : 1; | |
8a72758c JA |
873 | /* set if opcode supports polled "wait" */ |
874 | unsigned pollin : 1; | |
875 | unsigned pollout : 1; | |
bcda7baa JA |
876 | /* op supports buffer selection */ |
877 | unsigned buffer_select : 1; | |
26f0505a PB |
878 | /* do prep async if is going to be punted */ |
879 | unsigned needs_async_setup : 1; | |
27926b68 JA |
880 | /* should block plug */ |
881 | unsigned plug : 1; | |
e8c2bc1f JA |
882 | /* size of async data needed, if any */ |
883 | unsigned short async_size; | |
d3656344 JA |
884 | }; |
885 | ||
0918682b | 886 | static const struct io_op_def io_op_defs[] = { |
0463b6c5 PB |
887 | [IORING_OP_NOP] = {}, |
888 | [IORING_OP_READV] = { | |
d3656344 JA |
889 | .needs_file = 1, |
890 | .unbound_nonreg_file = 1, | |
8a72758c | 891 | .pollin = 1, |
4d954c25 | 892 | .buffer_select = 1, |
26f0505a | 893 | .needs_async_setup = 1, |
27926b68 | 894 | .plug = 1, |
e8c2bc1f | 895 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 896 | }, |
0463b6c5 | 897 | [IORING_OP_WRITEV] = { |
d3656344 JA |
898 | .needs_file = 1, |
899 | .hash_reg_file = 1, | |
900 | .unbound_nonreg_file = 1, | |
8a72758c | 901 | .pollout = 1, |
26f0505a | 902 | .needs_async_setup = 1, |
27926b68 | 903 | .plug = 1, |
e8c2bc1f | 904 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 905 | }, |
0463b6c5 | 906 | [IORING_OP_FSYNC] = { |
d3656344 JA |
907 | .needs_file = 1, |
908 | }, | |
0463b6c5 | 909 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
910 | .needs_file = 1, |
911 | .unbound_nonreg_file = 1, | |
8a72758c | 912 | .pollin = 1, |
27926b68 | 913 | .plug = 1, |
e8c2bc1f | 914 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 915 | }, |
0463b6c5 | 916 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
917 | .needs_file = 1, |
918 | .hash_reg_file = 1, | |
919 | .unbound_nonreg_file = 1, | |
8a72758c | 920 | .pollout = 1, |
27926b68 | 921 | .plug = 1, |
e8c2bc1f | 922 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 923 | }, |
0463b6c5 | 924 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
925 | .needs_file = 1, |
926 | .unbound_nonreg_file = 1, | |
927 | }, | |
0463b6c5 PB |
928 | [IORING_OP_POLL_REMOVE] = {}, |
929 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 JA |
930 | .needs_file = 1, |
931 | }, | |
0463b6c5 | 932 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
933 | .needs_file = 1, |
934 | .unbound_nonreg_file = 1, | |
8a72758c | 935 | .pollout = 1, |
26f0505a | 936 | .needs_async_setup = 1, |
e8c2bc1f | 937 | .async_size = sizeof(struct io_async_msghdr), |
d3656344 | 938 | }, |
0463b6c5 | 939 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
940 | .needs_file = 1, |
941 | .unbound_nonreg_file = 1, | |
8a72758c | 942 | .pollin = 1, |
52de1fe1 | 943 | .buffer_select = 1, |
26f0505a | 944 | .needs_async_setup = 1, |
e8c2bc1f | 945 | .async_size = sizeof(struct io_async_msghdr), |
d3656344 | 946 | }, |
0463b6c5 | 947 | [IORING_OP_TIMEOUT] = { |
e8c2bc1f | 948 | .async_size = sizeof(struct io_timeout_data), |
d3656344 | 949 | }, |
9c8e11b3 PB |
950 | [IORING_OP_TIMEOUT_REMOVE] = { |
951 | /* used by timeout updates' prep() */ | |
9c8e11b3 | 952 | }, |
0463b6c5 | 953 | [IORING_OP_ACCEPT] = { |
d3656344 JA |
954 | .needs_file = 1, |
955 | .unbound_nonreg_file = 1, | |
8a72758c | 956 | .pollin = 1, |
d3656344 | 957 | }, |
0463b6c5 PB |
958 | [IORING_OP_ASYNC_CANCEL] = {}, |
959 | [IORING_OP_LINK_TIMEOUT] = { | |
e8c2bc1f | 960 | .async_size = sizeof(struct io_timeout_data), |
d3656344 | 961 | }, |
0463b6c5 | 962 | [IORING_OP_CONNECT] = { |
d3656344 JA |
963 | .needs_file = 1, |
964 | .unbound_nonreg_file = 1, | |
8a72758c | 965 | .pollout = 1, |
26f0505a | 966 | .needs_async_setup = 1, |
e8c2bc1f | 967 | .async_size = sizeof(struct io_async_connect), |
d3656344 | 968 | }, |
0463b6c5 | 969 | [IORING_OP_FALLOCATE] = { |
d3656344 | 970 | .needs_file = 1, |
d3656344 | 971 | }, |
44526bed JA |
972 | [IORING_OP_OPENAT] = {}, |
973 | [IORING_OP_CLOSE] = {}, | |
974 | [IORING_OP_FILES_UPDATE] = {}, | |
975 | [IORING_OP_STATX] = {}, | |
0463b6c5 | 976 | [IORING_OP_READ] = { |
3a6820f2 JA |
977 | .needs_file = 1, |
978 | .unbound_nonreg_file = 1, | |
8a72758c | 979 | .pollin = 1, |
bcda7baa | 980 | .buffer_select = 1, |
27926b68 | 981 | .plug = 1, |
e8c2bc1f | 982 | .async_size = sizeof(struct io_async_rw), |
3a6820f2 | 983 | }, |
0463b6c5 | 984 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
985 | .needs_file = 1, |
986 | .unbound_nonreg_file = 1, | |
8a72758c | 987 | .pollout = 1, |
27926b68 | 988 | .plug = 1, |
e8c2bc1f | 989 | .async_size = sizeof(struct io_async_rw), |
3a6820f2 | 990 | }, |
0463b6c5 | 991 | [IORING_OP_FADVISE] = { |
4840e418 | 992 | .needs_file = 1, |
c1ca757b | 993 | }, |
44526bed | 994 | [IORING_OP_MADVISE] = {}, |
0463b6c5 | 995 | [IORING_OP_SEND] = { |
fddaface JA |
996 | .needs_file = 1, |
997 | .unbound_nonreg_file = 1, | |
8a72758c | 998 | .pollout = 1, |
fddaface | 999 | }, |
0463b6c5 | 1000 | [IORING_OP_RECV] = { |
fddaface JA |
1001 | .needs_file = 1, |
1002 | .unbound_nonreg_file = 1, | |
8a72758c | 1003 | .pollin = 1, |
bcda7baa | 1004 | .buffer_select = 1, |
fddaface | 1005 | }, |
0463b6c5 | 1006 | [IORING_OP_OPENAT2] = { |
cebdb986 | 1007 | }, |
3e4827b0 JA |
1008 | [IORING_OP_EPOLL_CTL] = { |
1009 | .unbound_nonreg_file = 1, | |
3e4827b0 | 1010 | }, |
7d67af2c PB |
1011 | [IORING_OP_SPLICE] = { |
1012 | .needs_file = 1, | |
1013 | .hash_reg_file = 1, | |
1014 | .unbound_nonreg_file = 1, | |
ddf0322d JA |
1015 | }, |
1016 | [IORING_OP_PROVIDE_BUFFERS] = {}, | |
067524e9 | 1017 | [IORING_OP_REMOVE_BUFFERS] = {}, |
f2a8d5c7 PB |
1018 | [IORING_OP_TEE] = { |
1019 | .needs_file = 1, | |
1020 | .hash_reg_file = 1, | |
1021 | .unbound_nonreg_file = 1, | |
1022 | }, | |
36f4fa68 JA |
1023 | [IORING_OP_SHUTDOWN] = { |
1024 | .needs_file = 1, | |
1025 | }, | |
44526bed JA |
1026 | [IORING_OP_RENAMEAT] = {}, |
1027 | [IORING_OP_UNLINKAT] = {}, | |
d3656344 JA |
1028 | }; |
1029 | ||
7a612350 | 1030 | static bool io_disarm_next(struct io_kiocb *req); |
eef51daa | 1031 | static void io_uring_del_tctx_node(unsigned long index); |
9936c7c2 PB |
1032 | static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
1033 | struct task_struct *task, | |
3dd0c97a | 1034 | bool cancel_all); |
78cc687b | 1035 | static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd); |
1ffc5422 | 1036 | |
d4d19c19 PB |
1037 | static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1038 | long res, unsigned int cflags); | |
ec9c02ad | 1039 | static void io_put_req(struct io_kiocb *req); |
216578e5 | 1040 | static void io_put_req_deferred(struct io_kiocb *req, int nr); |
c7dae4ba | 1041 | static void io_dismantle_req(struct io_kiocb *req); |
94ae5e77 JA |
1042 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
1043 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
fdecb662 | 1044 | static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, |
c3bdad02 | 1045 | struct io_uring_rsrc_update2 *up, |
98f0b3b4 | 1046 | unsigned nr_args); |
68fb8979 | 1047 | static void io_clean_op(struct io_kiocb *req); |
ac177053 | 1048 | static struct file *io_file_get(struct io_ring_ctx *ctx, |
8371adf5 | 1049 | struct io_kiocb *req, int fd, bool fixed); |
c5eef2b9 | 1050 | static void __io_queue_sqe(struct io_kiocb *req); |
269bbe5f | 1051 | static void io_rsrc_put_work(struct work_struct *work); |
de0617e4 | 1052 | |
907d1df3 | 1053 | static void io_req_task_queue(struct io_kiocb *req); |
2a2758f2 | 1054 | static void io_submit_flush_completions(struct io_ring_ctx *ctx); |
179ae0d1 | 1055 | static int io_req_prep_async(struct io_kiocb *req); |
de0617e4 | 1056 | |
2b188cc1 JA |
1057 | static struct kmem_cache *req_cachep; |
1058 | ||
0918682b | 1059 | static const struct file_operations io_uring_fops; |
2b188cc1 JA |
1060 | |
1061 | struct sock *io_uring_get_socket(struct file *file) | |
1062 | { | |
1063 | #if defined(CONFIG_UNIX) | |
1064 | if (file->f_op == &io_uring_fops) { | |
1065 | struct io_ring_ctx *ctx = file->private_data; | |
1066 | ||
1067 | return ctx->ring_sock->sk; | |
1068 | } | |
1069 | #endif | |
1070 | return NULL; | |
1071 | } | |
1072 | EXPORT_SYMBOL(io_uring_get_socket); | |
1073 | ||
f2f87370 PB |
1074 | #define io_for_each_link(pos, head) \ |
1075 | for (pos = (head); pos; pos = pos->link) | |
1076 | ||
b895c9a6 | 1077 | static inline void io_req_set_rsrc_node(struct io_kiocb *req) |
36f72fe2 PB |
1078 | { |
1079 | struct io_ring_ctx *ctx = req->ctx; | |
1080 | ||
269bbe5f | 1081 | if (!req->fixed_rsrc_refs) { |
a7f0ed5a | 1082 | req->fixed_rsrc_refs = &ctx->rsrc_node->refs; |
269bbe5f | 1083 | percpu_ref_get(req->fixed_rsrc_refs); |
36f72fe2 PB |
1084 | } |
1085 | } | |
1086 | ||
f70865db PB |
1087 | static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl) |
1088 | { | |
1089 | bool got = percpu_ref_tryget(ref); | |
1090 | ||
1091 | /* already at zero, wait for ->release() */ | |
1092 | if (!got) | |
1093 | wait_for_completion(compl); | |
1094 | percpu_ref_resurrect(ref); | |
1095 | if (got) | |
1096 | percpu_ref_put(ref); | |
1097 | } | |
1098 | ||
3dd0c97a PB |
1099 | static bool io_match_task(struct io_kiocb *head, struct task_struct *task, |
1100 | bool cancel_all) | |
08d23634 PB |
1101 | { |
1102 | struct io_kiocb *req; | |
1103 | ||
68207680 | 1104 | if (task && head->task != task) |
08d23634 | 1105 | return false; |
3dd0c97a | 1106 | if (cancel_all) |
08d23634 PB |
1107 | return true; |
1108 | ||
1109 | io_for_each_link(req, head) { | |
b05a1bcd | 1110 | if (req->flags & REQ_F_INFLIGHT) |
02a13674 | 1111 | return true; |
08d23634 PB |
1112 | } |
1113 | return false; | |
1114 | } | |
1115 | ||
93d2bcd2 | 1116 | static inline void req_set_fail(struct io_kiocb *req) |
c40f6379 | 1117 | { |
93d2bcd2 | 1118 | req->flags |= REQ_F_FAIL; |
c40f6379 | 1119 | } |
4a38aed2 | 1120 | |
2b188cc1 JA |
1121 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) |
1122 | { | |
1123 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
1124 | ||
0f158b4c | 1125 | complete(&ctx->ref_comp); |
2b188cc1 JA |
1126 | } |
1127 | ||
8eb7e2d0 PB |
1128 | static inline bool io_is_timeout_noseq(struct io_kiocb *req) |
1129 | { | |
1130 | return !req->timeout.off; | |
1131 | } | |
1132 | ||
f56165e6 PB |
1133 | static void io_fallback_req_func(struct work_struct *work) |
1134 | { | |
1135 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
1136 | fallback_work.work); | |
1137 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
1138 | struct io_kiocb *req, *tmp; | |
1139 | ||
1140 | percpu_ref_get(&ctx->refs); | |
1141 | llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node) | |
1142 | req->io_task_work.func(req); | |
1143 | percpu_ref_put(&ctx->refs); | |
1144 | } | |
1145 | ||
2b188cc1 JA |
1146 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
1147 | { | |
1148 | struct io_ring_ctx *ctx; | |
78076bb6 | 1149 | int hash_bits; |
2b188cc1 JA |
1150 | |
1151 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
1152 | if (!ctx) | |
1153 | return NULL; | |
1154 | ||
78076bb6 JA |
1155 | /* |
1156 | * Use 5 bits less than the max cq entries, that should give us around | |
1157 | * 32 entries per hash list if totally full and uniformly spread. | |
1158 | */ | |
1159 | hash_bits = ilog2(p->cq_entries); | |
1160 | hash_bits -= 5; | |
1161 | if (hash_bits <= 0) | |
1162 | hash_bits = 1; | |
1163 | ctx->cancel_hash_bits = hash_bits; | |
1164 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
1165 | GFP_KERNEL); | |
1166 | if (!ctx->cancel_hash) | |
1167 | goto err; | |
1168 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
1169 | ||
6224843d PB |
1170 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
1171 | if (!ctx->dummy_ubuf) | |
1172 | goto err; | |
1173 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
1174 | ctx->dummy_ubuf->ubuf = -1UL; | |
1175 | ||
21482896 | 1176 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
1177 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
1178 | goto err; | |
2b188cc1 JA |
1179 | |
1180 | ctx->flags = p->flags; | |
90554200 | 1181 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 1182 | INIT_LIST_HEAD(&ctx->sqd_list); |
311997b3 | 1183 | init_waitqueue_head(&ctx->poll_wait); |
1d7bb1d5 | 1184 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
0f158b4c | 1185 | init_completion(&ctx->ref_comp); |
9e15c3a0 | 1186 | xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1); |
61cf9370 | 1187 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 1188 | mutex_init(&ctx->uring_lock); |
311997b3 | 1189 | init_waitqueue_head(&ctx->cq_wait); |
2b188cc1 | 1190 | spin_lock_init(&ctx->completion_lock); |
540e32a0 | 1191 | INIT_LIST_HEAD(&ctx->iopoll_list); |
de0617e4 | 1192 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 1193 | INIT_LIST_HEAD(&ctx->timeout_list); |
d67d2263 BM |
1194 | spin_lock_init(&ctx->rsrc_ref_lock); |
1195 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f BM |
1196 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
1197 | init_llist_head(&ctx->rsrc_put_llist); | |
13bf43f5 | 1198 | INIT_LIST_HEAD(&ctx->tctx_list); |
cd0ca2e0 | 1199 | INIT_LIST_HEAD(&ctx->submit_state.free_list); |
d0acdee2 | 1200 | INIT_LIST_HEAD(&ctx->locked_free_list); |
9011bf9a | 1201 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
2b188cc1 | 1202 | return ctx; |
206aefde | 1203 | err: |
6224843d | 1204 | kfree(ctx->dummy_ubuf); |
78076bb6 | 1205 | kfree(ctx->cancel_hash); |
206aefde JA |
1206 | kfree(ctx); |
1207 | return NULL; | |
2b188cc1 JA |
1208 | } |
1209 | ||
8f6ed49a PB |
1210 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
1211 | { | |
1212 | struct io_rings *r = ctx->rings; | |
1213 | ||
1214 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
1215 | ctx->cq_extra--; | |
1216 | } | |
1217 | ||
9cf7c104 | 1218 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 1219 | { |
2bc9930e JA |
1220 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
1221 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 1222 | |
8f6ed49a | 1223 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 1224 | } |
de0617e4 | 1225 | |
9d858b21 | 1226 | return false; |
de0617e4 JA |
1227 | } |
1228 | ||
c97d8a0f PB |
1229 | #define FFS_ASYNC_READ 0x1UL |
1230 | #define FFS_ASYNC_WRITE 0x2UL | |
1231 | #ifdef CONFIG_64BIT | |
1232 | #define FFS_ISREG 0x4UL | |
1233 | #else | |
1234 | #define FFS_ISREG 0x0UL | |
1235 | #endif | |
1236 | #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG) | |
1237 | ||
1238 | static inline bool io_req_ffs_set(struct io_kiocb *req) | |
1239 | { | |
1240 | return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE); | |
1241 | } | |
1242 | ||
ce3d5aae PB |
1243 | static void io_req_track_inflight(struct io_kiocb *req) |
1244 | { | |
ce3d5aae | 1245 | if (!(req->flags & REQ_F_INFLIGHT)) { |
ce3d5aae | 1246 | req->flags |= REQ_F_INFLIGHT; |
b303fe2e | 1247 | atomic_inc(¤t->io_uring->inflight_tracked); |
ce3d5aae PB |
1248 | } |
1249 | } | |
1250 | ||
1e6fa521 JA |
1251 | static void io_prep_async_work(struct io_kiocb *req) |
1252 | { | |
1253 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
1254 | struct io_ring_ctx *ctx = req->ctx; |
1255 | ||
b8e64b53 PB |
1256 | if (!(req->flags & REQ_F_CREDS)) { |
1257 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 1258 | req->creds = get_current_cred(); |
b8e64b53 | 1259 | } |
003e8dcc | 1260 | |
e1d675df PB |
1261 | req->work.list.next = NULL; |
1262 | req->work.flags = 0; | |
feaadc4f PB |
1263 | if (req->flags & REQ_F_FORCE_ASYNC) |
1264 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
1265 | ||
1e6fa521 JA |
1266 | if (req->flags & REQ_F_ISREG) { |
1267 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1268 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 1269 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
1270 | if (def->unbound_nonreg_file) |
1271 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1272 | } | |
e1d675df PB |
1273 | |
1274 | switch (req->opcode) { | |
1275 | case IORING_OP_SPLICE: | |
1276 | case IORING_OP_TEE: | |
e1d675df PB |
1277 | if (!S_ISREG(file_inode(req->splice.file_in)->i_mode)) |
1278 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1279 | break; | |
1280 | } | |
561fb04a | 1281 | } |
cccf0ee8 | 1282 | |
cbdcb435 | 1283 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 1284 | { |
cbdcb435 | 1285 | struct io_kiocb *cur; |
54a91f3b | 1286 | |
44eff40a PB |
1287 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
1288 | struct io_ring_ctx *ctx = req->ctx; | |
1289 | ||
1290 | spin_lock_irq(&ctx->completion_lock); | |
1291 | io_for_each_link(cur, req) | |
1292 | io_prep_async_work(cur); | |
1293 | spin_unlock_irq(&ctx->completion_lock); | |
1294 | } else { | |
1295 | io_for_each_link(cur, req) | |
1296 | io_prep_async_work(cur); | |
1297 | } | |
561fb04a JA |
1298 | } |
1299 | ||
ebf93667 | 1300 | static void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1301 | { |
a197f664 | 1302 | struct io_ring_ctx *ctx = req->ctx; |
cbdcb435 | 1303 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 1304 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 1305 | |
3bfe6106 JA |
1306 | BUG_ON(!tctx); |
1307 | BUG_ON(!tctx->io_wq); | |
561fb04a | 1308 | |
cbdcb435 PB |
1309 | /* init ->work of the whole link before punting */ |
1310 | io_prep_async_link(req); | |
991468dc JA |
1311 | |
1312 | /* | |
1313 | * Not expected to happen, but if we do have a bug where this _can_ | |
1314 | * happen, catch it here and ensure the request is marked as | |
1315 | * canceled. That will make io-wq go through the usual work cancel | |
1316 | * procedure rather than attempt to run this request (or create a new | |
1317 | * worker for it). | |
1318 | */ | |
1319 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
1320 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1321 | ||
d07f1e8a PB |
1322 | trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req, |
1323 | &req->work, req->flags); | |
ebf93667 | 1324 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
1325 | if (link) |
1326 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
1327 | } |
1328 | ||
1ee4160c | 1329 | static void io_kill_timeout(struct io_kiocb *req, int status) |
8c855885 | 1330 | __must_hold(&req->ctx->completion_lock) |
5262f567 | 1331 | { |
e8c2bc1f | 1332 | struct io_timeout_data *io = req->async_data; |
5262f567 | 1333 | |
fd9c7bc5 | 1334 | if (hrtimer_try_to_cancel(&io->timer) != -1) { |
01cec8c1 PB |
1335 | atomic_set(&req->ctx->cq_timeouts, |
1336 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
135fcde8 | 1337 | list_del_init(&req->timeout.list); |
d4d19c19 | 1338 | io_cqring_fill_event(req->ctx, req->user_data, status, 0); |
216578e5 | 1339 | io_put_req_deferred(req, 1); |
5262f567 JA |
1340 | } |
1341 | } | |
1342 | ||
441b8a78 | 1343 | static void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 1344 | { |
441b8a78 | 1345 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
1346 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
1347 | struct io_defer_entry, list); | |
de0617e4 | 1348 | |
9cf7c104 | 1349 | if (req_need_defer(de->req, de->seq)) |
04518945 | 1350 | break; |
27dc8338 | 1351 | list_del_init(&de->list); |
907d1df3 | 1352 | io_req_task_queue(de->req); |
27dc8338 | 1353 | kfree(de); |
441b8a78 | 1354 | } |
04518945 PB |
1355 | } |
1356 | ||
360428f8 | 1357 | static void io_flush_timeouts(struct io_ring_ctx *ctx) |
de0617e4 | 1358 | { |
441b8a78 | 1359 | u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
f010505b | 1360 | |
f18ee4cf | 1361 | while (!list_empty(&ctx->timeout_list)) { |
f010505b | 1362 | u32 events_needed, events_got; |
360428f8 | 1363 | struct io_kiocb *req = list_first_entry(&ctx->timeout_list, |
135fcde8 | 1364 | struct io_kiocb, timeout.list); |
de0617e4 | 1365 | |
8eb7e2d0 | 1366 | if (io_is_timeout_noseq(req)) |
360428f8 | 1367 | break; |
f010505b MDG |
1368 | |
1369 | /* | |
1370 | * Since seq can easily wrap around over time, subtract | |
1371 | * the last seq at which timeouts were flushed before comparing. | |
1372 | * Assuming not more than 2^31-1 events have happened since, | |
1373 | * these subtractions won't have wrapped, so we can check if | |
1374 | * target is in [last_seq, current_seq] by comparing the two. | |
1375 | */ | |
1376 | events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush; | |
1377 | events_got = seq - ctx->cq_last_tm_flush; | |
1378 | if (events_got < events_needed) | |
360428f8 | 1379 | break; |
bfe68a22 | 1380 | |
135fcde8 | 1381 | list_del_init(&req->timeout.list); |
1ee4160c | 1382 | io_kill_timeout(req, 0); |
f18ee4cf | 1383 | } |
f010505b | 1384 | ctx->cq_last_tm_flush = seq; |
360428f8 | 1385 | } |
5262f567 | 1386 | |
2335f6f5 | 1387 | static void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
360428f8 | 1388 | { |
2335f6f5 PB |
1389 | if (ctx->off_timeout_used) |
1390 | io_flush_timeouts(ctx); | |
1391 | if (ctx->drain_active) | |
1392 | io_queue_deferred(ctx); | |
1393 | } | |
1394 | ||
1395 | static inline void io_commit_cqring(struct io_ring_ctx *ctx) | |
1396 | { | |
1397 | if (unlikely(ctx->off_timeout_used || ctx->drain_active)) | |
1398 | __io_commit_cqring_flush(ctx); | |
ec30e04b PB |
1399 | /* order cqe stores with ring update */ |
1400 | smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail); | |
de0617e4 JA |
1401 | } |
1402 | ||
90554200 JA |
1403 | static inline bool io_sqring_full(struct io_ring_ctx *ctx) |
1404 | { | |
1405 | struct io_rings *r = ctx->rings; | |
1406 | ||
a566c556 | 1407 | return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries; |
90554200 JA |
1408 | } |
1409 | ||
888aae2e PB |
1410 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
1411 | { | |
1412 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
1413 | } | |
1414 | ||
d068b506 | 1415 | static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) |
2b188cc1 | 1416 | { |
75b28aff | 1417 | struct io_rings *rings = ctx->rings; |
ea5ab3b5 | 1418 | unsigned tail, mask = ctx->cq_entries - 1; |
2b188cc1 | 1419 | |
115e12e5 SB |
1420 | /* |
1421 | * writes to the cq entry need to come after reading head; the | |
1422 | * control dependency is enough as we're using WRITE_ONCE to | |
1423 | * fill the cq entry | |
1424 | */ | |
a566c556 | 1425 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
2b188cc1 JA |
1426 | return NULL; |
1427 | ||
888aae2e | 1428 | tail = ctx->cached_cq_tail++; |
ea5ab3b5 | 1429 | return &rings->cqes[tail & mask]; |
2b188cc1 JA |
1430 | } |
1431 | ||
f2842ab5 JA |
1432 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1433 | { | |
44c769de | 1434 | if (likely(!ctx->cq_ev_fd)) |
f0b493e6 | 1435 | return false; |
7e55a19c SG |
1436 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
1437 | return false; | |
44c769de | 1438 | return !ctx->eventfd_async || io_wq_current_is_worker(); |
f2842ab5 JA |
1439 | } |
1440 | ||
b41e9852 | 1441 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 1442 | { |
5fd46178 JA |
1443 | /* |
1444 | * wake_up_all() may seem excessive, but io_wake_function() and | |
1445 | * io_should_wake() handle the termination of the loop and only | |
1446 | * wake as many waiters as we need to. | |
1447 | */ | |
1448 | if (wq_has_sleeper(&ctx->cq_wait)) | |
1449 | wake_up_all(&ctx->cq_wait); | |
534ca6d6 JA |
1450 | if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait)) |
1451 | wake_up(&ctx->sq_data->wait); | |
b41e9852 | 1452 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 | 1453 | eventfd_signal(ctx->cq_ev_fd, 1); |
311997b3 PB |
1454 | if (waitqueue_active(&ctx->poll_wait)) { |
1455 | wake_up_interruptible(&ctx->poll_wait); | |
4aa84f2f PB |
1456 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); |
1457 | } | |
1d7bb1d5 JA |
1458 | } |
1459 | ||
80c18e4a PB |
1460 | static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx) |
1461 | { | |
1462 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
5fd46178 JA |
1463 | if (wq_has_sleeper(&ctx->cq_wait)) |
1464 | wake_up_all(&ctx->cq_wait); | |
80c18e4a PB |
1465 | } |
1466 | if (io_should_trigger_evfd(ctx)) | |
1467 | eventfd_signal(ctx->cq_ev_fd, 1); | |
311997b3 PB |
1468 | if (waitqueue_active(&ctx->poll_wait)) { |
1469 | wake_up_interruptible(&ctx->poll_wait); | |
4aa84f2f PB |
1470 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); |
1471 | } | |
80c18e4a PB |
1472 | } |
1473 | ||
c4a2ed72 | 1474 | /* Returns true if there are no backlogged entries after the flush */ |
6c2450ae | 1475 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) |
1d7bb1d5 | 1476 | { |
1d7bb1d5 | 1477 | unsigned long flags; |
b18032bb | 1478 | bool all_flushed, posted; |
1d7bb1d5 | 1479 | |
a566c556 | 1480 | if (!force && __io_cqring_events(ctx) == ctx->cq_entries) |
e23de15f | 1481 | return false; |
1d7bb1d5 | 1482 | |
b18032bb | 1483 | posted = false; |
1d7bb1d5 | 1484 | spin_lock_irqsave(&ctx->completion_lock, flags); |
6c2450ae | 1485 | while (!list_empty(&ctx->cq_overflow_list)) { |
d068b506 | 1486 | struct io_uring_cqe *cqe = io_get_cqe(ctx); |
6c2450ae | 1487 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 1488 | |
1d7bb1d5 JA |
1489 | if (!cqe && !force) |
1490 | break; | |
6c2450ae PB |
1491 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
1492 | struct io_overflow_cqe, list); | |
1493 | if (cqe) | |
1494 | memcpy(cqe, &ocqe->cqe, sizeof(*cqe)); | |
1495 | else | |
8f6ed49a PB |
1496 | io_account_cq_overflow(ctx); |
1497 | ||
b18032bb | 1498 | posted = true; |
6c2450ae PB |
1499 | list_del(&ocqe->list); |
1500 | kfree(ocqe); | |
1d7bb1d5 JA |
1501 | } |
1502 | ||
09e88404 PB |
1503 | all_flushed = list_empty(&ctx->cq_overflow_list); |
1504 | if (all_flushed) { | |
5ed7a37d | 1505 | clear_bit(0, &ctx->check_cq_overflow); |
20c0b380 NA |
1506 | WRITE_ONCE(ctx->rings->sq_flags, |
1507 | ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW); | |
09e88404 | 1508 | } |
46930143 | 1509 | |
b18032bb JA |
1510 | if (posted) |
1511 | io_commit_cqring(ctx); | |
1d7bb1d5 | 1512 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
b18032bb JA |
1513 | if (posted) |
1514 | io_cqring_ev_posted(ctx); | |
09e88404 | 1515 | return all_flushed; |
1d7bb1d5 JA |
1516 | } |
1517 | ||
90f67366 | 1518 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 1519 | { |
ca0a2651 JA |
1520 | bool ret = true; |
1521 | ||
5ed7a37d | 1522 | if (test_bit(0, &ctx->check_cq_overflow)) { |
6c503150 PB |
1523 | /* iopoll syncs against uring_lock, not completion_lock */ |
1524 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
1525 | mutex_lock(&ctx->uring_lock); | |
90f67366 | 1526 | ret = __io_cqring_overflow_flush(ctx, false); |
6c503150 PB |
1527 | if (ctx->flags & IORING_SETUP_IOPOLL) |
1528 | mutex_unlock(&ctx->uring_lock); | |
1529 | } | |
ca0a2651 JA |
1530 | |
1531 | return ret; | |
6c503150 PB |
1532 | } |
1533 | ||
abc54d63 JA |
1534 | /* |
1535 | * Shamelessly stolen from the mm implementation of page reference checking, | |
1536 | * see commit f958d7b528b1 for details. | |
1537 | */ | |
1538 | #define req_ref_zero_or_close_to_overflow(req) \ | |
1539 | ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u) | |
1540 | ||
de9b4cca JA |
1541 | static inline bool req_ref_inc_not_zero(struct io_kiocb *req) |
1542 | { | |
abc54d63 | 1543 | return atomic_inc_not_zero(&req->refs); |
de9b4cca JA |
1544 | } |
1545 | ||
1546 | static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs) | |
1547 | { | |
abc54d63 JA |
1548 | WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req)); |
1549 | return atomic_sub_and_test(refs, &req->refs); | |
de9b4cca JA |
1550 | } |
1551 | ||
1552 | static inline bool req_ref_put_and_test(struct io_kiocb *req) | |
1553 | { | |
abc54d63 JA |
1554 | WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req)); |
1555 | return atomic_dec_and_test(&req->refs); | |
de9b4cca JA |
1556 | } |
1557 | ||
1558 | static inline void req_ref_put(struct io_kiocb *req) | |
1559 | { | |
abc54d63 | 1560 | WARN_ON_ONCE(req_ref_put_and_test(req)); |
de9b4cca JA |
1561 | } |
1562 | ||
1563 | static inline void req_ref_get(struct io_kiocb *req) | |
1564 | { | |
abc54d63 JA |
1565 | WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req)); |
1566 | atomic_inc(&req->refs); | |
de9b4cca JA |
1567 | } |
1568 | ||
6a290a14 PB |
1569 | /* must to be called somewhat shortly after putting a request */ |
1570 | static inline void io_put_task(struct task_struct *task, int nr) | |
1571 | { | |
1572 | struct io_uring_task *tctx = task->io_uring; | |
1573 | ||
1574 | percpu_counter_sub(&tctx->inflight, nr); | |
1575 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1576 | wake_up(&tctx->wait); | |
1577 | put_task_struct_many(task, nr); | |
1578 | } | |
1579 | ||
d4d19c19 PB |
1580 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
1581 | long res, unsigned int cflags) | |
2b188cc1 | 1582 | { |
cce4b8b0 | 1583 | struct io_overflow_cqe *ocqe; |
2b188cc1 | 1584 | |
cce4b8b0 PB |
1585 | ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT); |
1586 | if (!ocqe) { | |
1587 | /* | |
1588 | * If we're in ring overflow flush mode, or in task cancel mode, | |
1589 | * or cannot allocate an overflow entry, then we need to drop it | |
1590 | * on the floor. | |
1591 | */ | |
8f6ed49a | 1592 | io_account_cq_overflow(ctx); |
cce4b8b0 | 1593 | return false; |
2b188cc1 | 1594 | } |
cce4b8b0 | 1595 | if (list_empty(&ctx->cq_overflow_list)) { |
5ed7a37d | 1596 | set_bit(0, &ctx->check_cq_overflow); |
20c0b380 NA |
1597 | WRITE_ONCE(ctx->rings->sq_flags, |
1598 | ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW); | |
1599 | ||
cce4b8b0 | 1600 | } |
d4d19c19 | 1601 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
1602 | ocqe->cqe.res = res; |
1603 | ocqe->cqe.flags = cflags; | |
1604 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); | |
1605 | return true; | |
2b188cc1 JA |
1606 | } |
1607 | ||
d4d19c19 PB |
1608 | static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1609 | long res, unsigned int cflags) | |
2b188cc1 JA |
1610 | { |
1611 | struct io_uring_cqe *cqe; | |
1612 | ||
d4d19c19 | 1613 | trace_io_uring_complete(ctx, user_data, res, cflags); |
51c3ff62 | 1614 | |
2b188cc1 JA |
1615 | /* |
1616 | * If we can't get a cq entry, userspace overflowed the | |
1617 | * submission (by quite a lot). Increment the overflow count in | |
1618 | * the ring. | |
1619 | */ | |
d068b506 | 1620 | cqe = io_get_cqe(ctx); |
1d7bb1d5 | 1621 | if (likely(cqe)) { |
d4d19c19 | 1622 | WRITE_ONCE(cqe->user_data, user_data); |
2b188cc1 | 1623 | WRITE_ONCE(cqe->res, res); |
bcda7baa | 1624 | WRITE_ONCE(cqe->flags, cflags); |
8d13326e | 1625 | return true; |
2b188cc1 | 1626 | } |
d4d19c19 | 1627 | return io_cqring_event_overflow(ctx, user_data, res, cflags); |
2b188cc1 JA |
1628 | } |
1629 | ||
8d13326e | 1630 | /* not as hot to bloat with inlining */ |
d4d19c19 PB |
1631 | static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1632 | long res, unsigned int cflags) | |
bcda7baa | 1633 | { |
d4d19c19 | 1634 | return __io_cqring_fill_event(ctx, user_data, res, cflags); |
bcda7baa JA |
1635 | } |
1636 | ||
7a612350 PB |
1637 | static void io_req_complete_post(struct io_kiocb *req, long res, |
1638 | unsigned int cflags) | |
2b188cc1 | 1639 | { |
78e19bbe | 1640 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1641 | unsigned long flags; |
1642 | ||
1643 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
d4d19c19 | 1644 | __io_cqring_fill_event(ctx, req->user_data, res, cflags); |
c7dae4ba JA |
1645 | /* |
1646 | * If we're the last reference to this request, add to our locked | |
1647 | * free_list cache. | |
1648 | */ | |
de9b4cca | 1649 | if (req_ref_put_and_test(req)) { |
7a612350 | 1650 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
93d2bcd2 | 1651 | if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) |
7a612350 PB |
1652 | io_disarm_next(req); |
1653 | if (req->link) { | |
1654 | io_req_task_queue(req->link); | |
1655 | req->link = NULL; | |
1656 | } | |
1657 | } | |
c7dae4ba JA |
1658 | io_dismantle_req(req); |
1659 | io_put_task(req->task, 1); | |
bb943b82 | 1660 | list_add(&req->inflight_entry, &ctx->locked_free_list); |
d0acdee2 | 1661 | ctx->locked_free_nr++; |
180f829f PB |
1662 | } else { |
1663 | if (!percpu_ref_tryget(&ctx->refs)) | |
1664 | req = NULL; | |
1665 | } | |
7a612350 | 1666 | io_commit_cqring(ctx); |
2b188cc1 | 1667 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
7a612350 | 1668 | |
180f829f PB |
1669 | if (req) { |
1670 | io_cqring_ev_posted(ctx); | |
c7dae4ba | 1671 | percpu_ref_put(&ctx->refs); |
180f829f | 1672 | } |
229a7b63 JA |
1673 | } |
1674 | ||
4e3d9ff9 JA |
1675 | static inline bool io_req_needs_clean(struct io_kiocb *req) |
1676 | { | |
c854357b | 1677 | return req->flags & IO_REQ_CLEAN_FLAGS; |
4e3d9ff9 JA |
1678 | } |
1679 | ||
a38d68db | 1680 | static void io_req_complete_state(struct io_kiocb *req, long res, |
889fca73 | 1681 | unsigned int cflags) |
229a7b63 | 1682 | { |
4e3d9ff9 | 1683 | if (io_req_needs_clean(req)) |
68fb8979 | 1684 | io_clean_op(req); |
a38d68db PB |
1685 | req->result = res; |
1686 | req->compl.cflags = cflags; | |
e342c807 | 1687 | req->flags |= REQ_F_COMPLETE_INLINE; |
e1e16097 JA |
1688 | } |
1689 | ||
889fca73 PB |
1690 | static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags, |
1691 | long res, unsigned cflags) | |
bcda7baa | 1692 | { |
889fca73 PB |
1693 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) |
1694 | io_req_complete_state(req, res, cflags); | |
a38d68db | 1695 | else |
c7dae4ba | 1696 | io_req_complete_post(req, res, cflags); |
bcda7baa JA |
1697 | } |
1698 | ||
a38d68db | 1699 | static inline void io_req_complete(struct io_kiocb *req, long res) |
0ddf92e8 | 1700 | { |
889fca73 | 1701 | __io_req_complete(req, 0, res, 0); |
0ddf92e8 JA |
1702 | } |
1703 | ||
f41db273 PB |
1704 | static void io_req_complete_failed(struct io_kiocb *req, long res) |
1705 | { | |
93d2bcd2 | 1706 | req_set_fail(req); |
f41db273 PB |
1707 | io_put_req(req); |
1708 | io_req_complete_post(req, res, 0); | |
1709 | } | |
1710 | ||
864ea921 PB |
1711 | /* |
1712 | * Don't initialise the fields below on every allocation, but do that in | |
1713 | * advance and keep them valid across allocations. | |
1714 | */ | |
1715 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1716 | { | |
1717 | req->ctx = ctx; | |
1718 | req->link = NULL; | |
1719 | req->async_data = NULL; | |
1720 | /* not necessary, but safer to zero */ | |
1721 | req->result = 0; | |
1722 | } | |
1723 | ||
dac7a098 | 1724 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1725 | struct io_submit_state *state) |
dac7a098 PB |
1726 | { |
1727 | spin_lock_irq(&ctx->completion_lock); | |
cd0ca2e0 | 1728 | list_splice_init(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1729 | ctx->locked_free_nr = 0; |
dac7a098 PB |
1730 | spin_unlock_irq(&ctx->completion_lock); |
1731 | } | |
1732 | ||
dd78f492 | 1733 | /* Returns true IFF there are requests in the cache */ |
c7dae4ba | 1734 | static bool io_flush_cached_reqs(struct io_ring_ctx *ctx) |
0ddf92e8 | 1735 | { |
c7dae4ba | 1736 | struct io_submit_state *state = &ctx->submit_state; |
dd78f492 | 1737 | int nr; |
0ddf92e8 | 1738 | |
c7dae4ba JA |
1739 | /* |
1740 | * If we have more than a batch's worth of requests in our IRQ side | |
1741 | * locked cache, grab the lock and move them over to our submission | |
1742 | * side cache. | |
1743 | */ | |
d0acdee2 | 1744 | if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH) |
cd0ca2e0 | 1745 | io_flush_cached_locked_reqs(ctx, state); |
0ddf92e8 | 1746 | |
dd78f492 | 1747 | nr = state->free_reqs; |
cd0ca2e0 PB |
1748 | while (!list_empty(&state->free_list)) { |
1749 | struct io_kiocb *req = list_first_entry(&state->free_list, | |
bb943b82 | 1750 | struct io_kiocb, inflight_entry); |
dd78f492 | 1751 | |
bb943b82 | 1752 | list_del(&req->inflight_entry); |
dd78f492 PB |
1753 | state->reqs[nr++] = req; |
1754 | if (nr == ARRAY_SIZE(state->reqs)) | |
e5d1bc0a | 1755 | break; |
1b4c351f JA |
1756 | } |
1757 | ||
dd78f492 PB |
1758 | state->free_reqs = nr; |
1759 | return nr != 0; | |
0ddf92e8 JA |
1760 | } |
1761 | ||
e5d1bc0a | 1762 | static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx) |
2b188cc1 | 1763 | { |
e5d1bc0a | 1764 | struct io_submit_state *state = &ctx->submit_state; |
864ea921 PB |
1765 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
1766 | int ret, i; | |
e5d1bc0a | 1767 | |
fe7e3257 | 1768 | BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH); |
e5d1bc0a | 1769 | |
864ea921 PB |
1770 | if (likely(state->free_reqs || io_flush_cached_reqs(ctx))) |
1771 | goto got_req; | |
e5d1bc0a | 1772 | |
864ea921 PB |
1773 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH, |
1774 | state->reqs); | |
fd6fab2c | 1775 | |
864ea921 PB |
1776 | /* |
1777 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1778 | * retry single alloc to be on the safe side. | |
1779 | */ | |
1780 | if (unlikely(ret <= 0)) { | |
1781 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1782 | if (!state->reqs[0]) | |
1783 | return NULL; | |
1784 | ret = 1; | |
2b188cc1 | 1785 | } |
864ea921 PB |
1786 | |
1787 | for (i = 0; i < ret; i++) | |
1788 | io_preinit_req(state->reqs[i], ctx); | |
1789 | state->free_reqs = ret; | |
e5d1bc0a | 1790 | got_req: |
291b2821 PB |
1791 | state->free_reqs--; |
1792 | return state->reqs[state->free_reqs]; | |
2b188cc1 JA |
1793 | } |
1794 | ||
e1d767f0 | 1795 | static inline void io_put_file(struct file *file) |
8da11c19 | 1796 | { |
e1d767f0 | 1797 | if (file) |
8da11c19 PB |
1798 | fput(file); |
1799 | } | |
1800 | ||
4edf20f9 | 1801 | static void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1802 | { |
094bae49 | 1803 | unsigned int flags = req->flags; |
929a3af9 | 1804 | |
3a0a6902 PB |
1805 | if (io_req_needs_clean(req)) |
1806 | io_clean_op(req); | |
e1d767f0 PB |
1807 | if (!(flags & REQ_F_FIXED_FILE)) |
1808 | io_put_file(req->file); | |
269bbe5f BM |
1809 | if (req->fixed_rsrc_refs) |
1810 | percpu_ref_put(req->fixed_rsrc_refs); | |
99ebe4ef | 1811 | if (req->async_data) { |
094bae49 | 1812 | kfree(req->async_data); |
99ebe4ef PB |
1813 | req->async_data = NULL; |
1814 | } | |
e65ef56d JA |
1815 | } |
1816 | ||
216578e5 | 1817 | static void __io_free_req(struct io_kiocb *req) |
c6ca97b3 | 1818 | { |
51a4cc11 | 1819 | struct io_ring_ctx *ctx = req->ctx; |
c34b025f | 1820 | unsigned long flags; |
c6ca97b3 | 1821 | |
216578e5 | 1822 | io_dismantle_req(req); |
7c660731 | 1823 | io_put_task(req->task, 1); |
c6ca97b3 | 1824 | |
c34b025f | 1825 | spin_lock_irqsave(&ctx->completion_lock, flags); |
bb943b82 | 1826 | list_add(&req->inflight_entry, &ctx->locked_free_list); |
c34b025f PB |
1827 | ctx->locked_free_nr++; |
1828 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1829 | ||
ecfc5177 | 1830 | percpu_ref_put(&ctx->refs); |
e65ef56d JA |
1831 | } |
1832 | ||
f2f87370 PB |
1833 | static inline void io_remove_next_linked(struct io_kiocb *req) |
1834 | { | |
1835 | struct io_kiocb *nxt = req->link; | |
1836 | ||
1837 | req->link = nxt->link; | |
1838 | nxt->link = NULL; | |
1839 | } | |
1840 | ||
33cc89a9 PB |
1841 | static bool io_kill_linked_timeout(struct io_kiocb *req) |
1842 | __must_hold(&req->ctx->completion_lock) | |
2665abfd | 1843 | { |
33cc89a9 | 1844 | struct io_kiocb *link = req->link; |
f2f87370 | 1845 | |
900fad45 PB |
1846 | /* |
1847 | * Can happen if a linked timeout fired and link had been like | |
1848 | * req -> link t-out -> link t-out [-> ...] | |
1849 | */ | |
c9abd7ad PB |
1850 | if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) { |
1851 | struct io_timeout_data *io = link->async_data; | |
7c86ffee | 1852 | |
f2f87370 | 1853 | io_remove_next_linked(req); |
90cd7e42 | 1854 | link->timeout.head = NULL; |
fd9c7bc5 | 1855 | if (hrtimer_try_to_cancel(&io->timer) != -1) { |
d4d19c19 PB |
1856 | io_cqring_fill_event(link->ctx, link->user_data, |
1857 | -ECANCELED, 0); | |
33cc89a9 | 1858 | io_put_req_deferred(link, 1); |
d4729fbd | 1859 | return true; |
c9abd7ad PB |
1860 | } |
1861 | } | |
d4729fbd | 1862 | return false; |
7c86ffee PB |
1863 | } |
1864 | ||
d148ca4b | 1865 | static void io_fail_links(struct io_kiocb *req) |
33cc89a9 | 1866 | __must_hold(&req->ctx->completion_lock) |
9e645e11 | 1867 | { |
33cc89a9 | 1868 | struct io_kiocb *nxt, *link = req->link; |
9e645e11 | 1869 | |
f2f87370 | 1870 | req->link = NULL; |
f2f87370 PB |
1871 | while (link) { |
1872 | nxt = link->link; | |
1873 | link->link = NULL; | |
2665abfd | 1874 | |
f2f87370 | 1875 | trace_io_uring_fail_link(req, link); |
d4d19c19 | 1876 | io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0); |
1575f21a | 1877 | io_put_req_deferred(link, 2); |
f2f87370 | 1878 | link = nxt; |
9e645e11 | 1879 | } |
33cc89a9 | 1880 | } |
9e645e11 | 1881 | |
33cc89a9 PB |
1882 | static bool io_disarm_next(struct io_kiocb *req) |
1883 | __must_hold(&req->ctx->completion_lock) | |
1884 | { | |
1885 | bool posted = false; | |
1886 | ||
1887 | if (likely(req->flags & REQ_F_LINK_TIMEOUT)) | |
1888 | posted = io_kill_linked_timeout(req); | |
93d2bcd2 | 1889 | if (unlikely((req->flags & REQ_F_FAIL) && |
e4335ed3 | 1890 | !(req->flags & REQ_F_HARDLINK))) { |
33cc89a9 PB |
1891 | posted |= (req->link != NULL); |
1892 | io_fail_links(req); | |
1893 | } | |
1894 | return posted; | |
9e645e11 JA |
1895 | } |
1896 | ||
3fa5e0f3 | 1897 | static struct io_kiocb *__io_req_find_next(struct io_kiocb *req) |
c69f8dbe | 1898 | { |
33cc89a9 | 1899 | struct io_kiocb *nxt; |
944e58bf | 1900 | |
9e645e11 JA |
1901 | /* |
1902 | * If LINK is set, we have dependent requests in this chain. If we | |
1903 | * didn't fail this request, queue the first one up, moving any other | |
1904 | * dependencies to the next request. In case of failure, fail the rest | |
1905 | * of the chain. | |
1906 | */ | |
93d2bcd2 | 1907 | if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) { |
33cc89a9 PB |
1908 | struct io_ring_ctx *ctx = req->ctx; |
1909 | unsigned long flags; | |
1910 | bool posted; | |
1911 | ||
1912 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1913 | posted = io_disarm_next(req); | |
1914 | if (posted) | |
1915 | io_commit_cqring(req->ctx); | |
1916 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1917 | if (posted) | |
1918 | io_cqring_ev_posted(ctx); | |
f2f87370 | 1919 | } |
33cc89a9 PB |
1920 | nxt = req->link; |
1921 | req->link = NULL; | |
1922 | return nxt; | |
4d7dd462 | 1923 | } |
9e645e11 | 1924 | |
f2f87370 | 1925 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) |
3fa5e0f3 | 1926 | { |
cdbff982 | 1927 | if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK)))) |
3fa5e0f3 PB |
1928 | return NULL; |
1929 | return __io_req_find_next(req); | |
1930 | } | |
1931 | ||
2c32395d PB |
1932 | static void ctx_flush_and_put(struct io_ring_ctx *ctx) |
1933 | { | |
1934 | if (!ctx) | |
1935 | return; | |
cd0ca2e0 | 1936 | if (ctx->submit_state.compl_nr) { |
2c32395d | 1937 | mutex_lock(&ctx->uring_lock); |
2a2758f2 | 1938 | io_submit_flush_completions(ctx); |
2c32395d PB |
1939 | mutex_unlock(&ctx->uring_lock); |
1940 | } | |
1941 | percpu_ref_put(&ctx->refs); | |
1942 | } | |
1943 | ||
7cbf1722 | 1944 | static void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1945 | { |
ebd0df2e | 1946 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1947 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1948 | task_work); | |
c40f6379 | 1949 | |
16f72070 | 1950 | while (1) { |
3f18407d PB |
1951 | struct io_wq_work_node *node; |
1952 | ||
1953 | spin_lock_irq(&tctx->task_lock); | |
c6538be9 | 1954 | node = tctx->task_list.first; |
3f18407d | 1955 | INIT_WQ_LIST(&tctx->task_list); |
6294f368 PB |
1956 | if (!node) |
1957 | tctx->task_running = false; | |
3f18407d | 1958 | spin_unlock_irq(&tctx->task_lock); |
6294f368 PB |
1959 | if (!node) |
1960 | break; | |
3f18407d | 1961 | |
6294f368 | 1962 | do { |
3f18407d PB |
1963 | struct io_wq_work_node *next = node->next; |
1964 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1965 | io_task_work.node); | |
1966 | ||
1967 | if (req->ctx != ctx) { | |
1968 | ctx_flush_and_put(ctx); | |
1969 | ctx = req->ctx; | |
1970 | percpu_ref_get(&ctx->refs); | |
1971 | } | |
5b0a6acc | 1972 | req->io_task_work.func(req); |
3f18407d | 1973 | node = next; |
6294f368 PB |
1974 | } while (node); |
1975 | ||
7cbf1722 | 1976 | cond_resched(); |
3f18407d | 1977 | } |
ebd0df2e PB |
1978 | |
1979 | ctx_flush_and_put(ctx); | |
7cbf1722 JA |
1980 | } |
1981 | ||
e09ee510 | 1982 | static void io_req_task_work_add(struct io_kiocb *req) |
7cbf1722 | 1983 | { |
c15b79de | 1984 | struct task_struct *tsk = req->task; |
7cbf1722 | 1985 | struct io_uring_task *tctx = tsk->io_uring; |
c15b79de | 1986 | enum task_work_notify_mode notify; |
e09ee510 | 1987 | struct io_wq_work_node *node; |
0b81e80c | 1988 | unsigned long flags; |
6294f368 | 1989 | bool running; |
7cbf1722 JA |
1990 | |
1991 | WARN_ON_ONCE(!tctx); | |
1992 | ||
0b81e80c | 1993 | spin_lock_irqsave(&tctx->task_lock, flags); |
7cbf1722 | 1994 | wq_list_add_tail(&req->io_task_work.node, &tctx->task_list); |
6294f368 PB |
1995 | running = tctx->task_running; |
1996 | if (!running) | |
1997 | tctx->task_running = true; | |
0b81e80c | 1998 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 JA |
1999 | |
2000 | /* task_work already pending, we're done */ | |
6294f368 | 2001 | if (running) |
e09ee510 | 2002 | return; |
7cbf1722 | 2003 | |
c15b79de PB |
2004 | /* |
2005 | * SQPOLL kernel thread doesn't need notification, just a wakeup. For | |
2006 | * all other cases, use TWA_SIGNAL unconditionally to ensure we're | |
2007 | * processing task_work. There's no reliable way to tell if TWA_RESUME | |
2008 | * will do the job. | |
2009 | */ | |
2010 | notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL; | |
c15b79de PB |
2011 | if (!task_work_add(tsk, &tctx->task_work, notify)) { |
2012 | wake_up_process(tsk); | |
e09ee510 | 2013 | return; |
c15b79de | 2014 | } |
2215bed9 | 2015 | |
0b81e80c | 2016 | spin_lock_irqsave(&tctx->task_lock, flags); |
6294f368 | 2017 | tctx->task_running = false; |
e09ee510 PB |
2018 | node = tctx->task_list.first; |
2019 | INIT_WQ_LIST(&tctx->task_list); | |
0b81e80c | 2020 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 | 2021 | |
e09ee510 PB |
2022 | while (node) { |
2023 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
2024 | node = node->next; | |
2025 | if (llist_add(&req->io_task_work.fallback_node, | |
2026 | &req->ctx->fallback_llist)) | |
2027 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
2028 | } | |
eab30c4d PB |
2029 | } |
2030 | ||
5b0a6acc | 2031 | static void io_req_task_cancel(struct io_kiocb *req) |
c40f6379 | 2032 | { |
87ceb6a6 | 2033 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 | 2034 | |
e83acd7d | 2035 | /* ctx is guaranteed to stay alive while we hold uring_lock */ |
792bb6eb | 2036 | mutex_lock(&ctx->uring_lock); |
2593553a | 2037 | io_req_complete_failed(req, req->result); |
792bb6eb | 2038 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2039 | } |
2040 | ||
5b0a6acc | 2041 | static void io_req_task_submit(struct io_kiocb *req) |
c40f6379 JA |
2042 | { |
2043 | struct io_ring_ctx *ctx = req->ctx; | |
2044 | ||
04fc6c80 | 2045 | /* ctx stays valid until unlock, even if we drop all ours ctx->refs */ |
81b6d05c | 2046 | mutex_lock(&ctx->uring_lock); |
af066f31 | 2047 | if (likely(!(req->task->flags & PF_EXITING))) |
c5eef2b9 | 2048 | __io_queue_sqe(req); |
81b6d05c | 2049 | else |
2593553a | 2050 | io_req_complete_failed(req, -EFAULT); |
81b6d05c | 2051 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2052 | } |
2053 | ||
2c4b8eb6 | 2054 | static void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 2055 | { |
2c4b8eb6 | 2056 | req->result = ret; |
5b0a6acc | 2057 | req->io_task_work.func = io_req_task_cancel; |
e09ee510 | 2058 | io_req_task_work_add(req); |
c40f6379 JA |
2059 | } |
2060 | ||
2c4b8eb6 | 2061 | static void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 2062 | { |
5b0a6acc | 2063 | req->io_task_work.func = io_req_task_submit; |
e09ee510 | 2064 | io_req_task_work_add(req); |
a3df7698 PB |
2065 | } |
2066 | ||
773af691 JA |
2067 | static void io_req_task_queue_reissue(struct io_kiocb *req) |
2068 | { | |
2069 | req->io_task_work.func = io_queue_async_work; | |
2070 | io_req_task_work_add(req); | |
2071 | } | |
2072 | ||
f2f87370 | 2073 | static inline void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 2074 | { |
9b5f7bd9 | 2075 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
2076 | |
2077 | if (nxt) | |
906a8c3f | 2078 | io_req_task_queue(nxt); |
c69f8dbe JL |
2079 | } |
2080 | ||
c3524383 | 2081 | static void io_free_req(struct io_kiocb *req) |
7a743e22 | 2082 | { |
c3524383 PB |
2083 | io_queue_next(req); |
2084 | __io_free_req(req); | |
2085 | } | |
8766dd51 | 2086 | |
2d6500d4 | 2087 | struct req_batch { |
5af1d13e PB |
2088 | struct task_struct *task; |
2089 | int task_refs; | |
1b4c351f | 2090 | int ctx_refs; |
2d6500d4 PB |
2091 | }; |
2092 | ||
5af1d13e PB |
2093 | static inline void io_init_req_batch(struct req_batch *rb) |
2094 | { | |
5af1d13e | 2095 | rb->task_refs = 0; |
9ae72463 | 2096 | rb->ctx_refs = 0; |
5af1d13e PB |
2097 | rb->task = NULL; |
2098 | } | |
2099 | ||
2d6500d4 PB |
2100 | static void io_req_free_batch_finish(struct io_ring_ctx *ctx, |
2101 | struct req_batch *rb) | |
2102 | { | |
9ae72463 PB |
2103 | if (rb->ctx_refs) |
2104 | percpu_ref_put_many(&ctx->refs, rb->ctx_refs); | |
e9dbe221 PB |
2105 | if (rb->task == current) |
2106 | current->io_uring->cached_refs += rb->task_refs; | |
2107 | else if (rb->task) | |
2108 | io_put_task(rb->task, rb->task_refs); | |
2d6500d4 PB |
2109 | } |
2110 | ||
6ff119a6 PB |
2111 | static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req, |
2112 | struct io_submit_state *state) | |
2d6500d4 | 2113 | { |
f2f87370 | 2114 | io_queue_next(req); |
96670657 | 2115 | io_dismantle_req(req); |
2d6500d4 | 2116 | |
e3bc8e9d | 2117 | if (req->task != rb->task) { |
7c660731 PB |
2118 | if (rb->task) |
2119 | io_put_task(rb->task, rb->task_refs); | |
e3bc8e9d JA |
2120 | rb->task = req->task; |
2121 | rb->task_refs = 0; | |
5af1d13e | 2122 | } |
e3bc8e9d | 2123 | rb->task_refs++; |
9ae72463 | 2124 | rb->ctx_refs++; |
5af1d13e | 2125 | |
bd759045 | 2126 | if (state->free_reqs != ARRAY_SIZE(state->reqs)) |
6ff119a6 | 2127 | state->reqs[state->free_reqs++] = req; |
bd759045 | 2128 | else |
cd0ca2e0 | 2129 | list_add(&req->inflight_entry, &state->free_list); |
7a743e22 PB |
2130 | } |
2131 | ||
2a2758f2 | 2132 | static void io_submit_flush_completions(struct io_ring_ctx *ctx) |
282cdc86 | 2133 | __must_hold(&req->ctx->uring_lock) |
905c172f | 2134 | { |
cd0ca2e0 PB |
2135 | struct io_submit_state *state = &ctx->submit_state; |
2136 | int i, nr = state->compl_nr; | |
905c172f PB |
2137 | struct req_batch rb; |
2138 | ||
905c172f PB |
2139 | spin_lock_irq(&ctx->completion_lock); |
2140 | for (i = 0; i < nr; i++) { | |
cd0ca2e0 | 2141 | struct io_kiocb *req = state->compl_reqs[i]; |
5182ed2e | 2142 | |
d4d19c19 PB |
2143 | __io_cqring_fill_event(ctx, req->user_data, req->result, |
2144 | req->compl.cflags); | |
905c172f PB |
2145 | } |
2146 | io_commit_cqring(ctx); | |
2147 | spin_unlock_irq(&ctx->completion_lock); | |
905c172f | 2148 | io_cqring_ev_posted(ctx); |
5182ed2e PB |
2149 | |
2150 | io_init_req_batch(&rb); | |
905c172f | 2151 | for (i = 0; i < nr; i++) { |
cd0ca2e0 | 2152 | struct io_kiocb *req = state->compl_reqs[i]; |
905c172f PB |
2153 | |
2154 | /* submission and completion refs */ | |
de9b4cca | 2155 | if (req_ref_sub_and_test(req, 2)) |
6ff119a6 | 2156 | io_req_free_batch(&rb, req, &ctx->submit_state); |
905c172f PB |
2157 | } |
2158 | ||
2159 | io_req_free_batch_finish(ctx, &rb); | |
cd0ca2e0 | 2160 | state->compl_nr = 0; |
7a743e22 PB |
2161 | } |
2162 | ||
ba816ad6 JA |
2163 | /* |
2164 | * Drop reference to request, return next in chain (if there is one) if this | |
2165 | * was the last reference to this request. | |
2166 | */ | |
0d85035a | 2167 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 2168 | { |
9b5f7bd9 PB |
2169 | struct io_kiocb *nxt = NULL; |
2170 | ||
de9b4cca | 2171 | if (req_ref_put_and_test(req)) { |
9b5f7bd9 | 2172 | nxt = io_req_find_next(req); |
4d7dd462 | 2173 | __io_free_req(req); |
2a44f467 | 2174 | } |
9b5f7bd9 | 2175 | return nxt; |
2b188cc1 JA |
2176 | } |
2177 | ||
0d85035a | 2178 | static inline void io_put_req(struct io_kiocb *req) |
e65ef56d | 2179 | { |
de9b4cca | 2180 | if (req_ref_put_and_test(req)) |
e65ef56d | 2181 | io_free_req(req); |
2b188cc1 JA |
2182 | } |
2183 | ||
216578e5 PB |
2184 | static inline void io_put_req_deferred(struct io_kiocb *req, int refs) |
2185 | { | |
543af3a1 PB |
2186 | if (req_ref_sub_and_test(req, refs)) { |
2187 | req->io_task_work.func = io_free_req; | |
2188 | io_req_task_work_add(req); | |
2189 | } | |
216578e5 PB |
2190 | } |
2191 | ||
6c503150 | 2192 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
2193 | { |
2194 | /* See comment at the top of this file */ | |
2195 | smp_rmb(); | |
e23de15f | 2196 | return __io_cqring_events(ctx); |
a3a0e43f JA |
2197 | } |
2198 | ||
fb5ccc98 PB |
2199 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
2200 | { | |
2201 | struct io_rings *rings = ctx->rings; | |
2202 | ||
2203 | /* make sure SQ entry isn't read before tail */ | |
2204 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
2205 | } | |
2206 | ||
8ff069bf | 2207 | static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf) |
e94f141b | 2208 | { |
8ff069bf | 2209 | unsigned int cflags; |
e94f141b | 2210 | |
bcda7baa JA |
2211 | cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT; |
2212 | cflags |= IORING_CQE_F_BUFFER; | |
0e1b6fe3 | 2213 | req->flags &= ~REQ_F_BUFFER_SELECTED; |
bcda7baa JA |
2214 | kfree(kbuf); |
2215 | return cflags; | |
e94f141b JA |
2216 | } |
2217 | ||
8ff069bf | 2218 | static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req) |
bcda7baa | 2219 | { |
4d954c25 | 2220 | struct io_buffer *kbuf; |
bcda7baa | 2221 | |
4d954c25 | 2222 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; |
8ff069bf PB |
2223 | return io_put_kbuf(req, kbuf); |
2224 | } | |
2225 | ||
4c6e277c JA |
2226 | static inline bool io_run_task_work(void) |
2227 | { | |
ef98eb04 | 2228 | if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) { |
4c6e277c | 2229 | __set_current_state(TASK_RUNNING); |
ef98eb04 | 2230 | tracehook_notify_signal(); |
4c6e277c JA |
2231 | return true; |
2232 | } | |
2233 | ||
2234 | return false; | |
bcda7baa JA |
2235 | } |
2236 | ||
def596e9 JA |
2237 | /* |
2238 | * Find and free completed poll iocbs | |
2239 | */ | |
2240 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
3c30ef0f | 2241 | struct list_head *done, bool resubmit) |
def596e9 | 2242 | { |
8237e045 | 2243 | struct req_batch rb; |
def596e9 | 2244 | struct io_kiocb *req; |
bbde017a XW |
2245 | |
2246 | /* order with ->result store in io_complete_rw_iopoll() */ | |
2247 | smp_rmb(); | |
def596e9 | 2248 | |
5af1d13e | 2249 | io_init_req_batch(&rb); |
def596e9 | 2250 | while (!list_empty(done)) { |
bcda7baa JA |
2251 | int cflags = 0; |
2252 | ||
d21ffe7e | 2253 | req = list_first_entry(done, struct io_kiocb, inflight_entry); |
f161340d PB |
2254 | list_del(&req->inflight_entry); |
2255 | ||
3c30ef0f | 2256 | if (READ_ONCE(req->result) == -EAGAIN && resubmit && |
8c130827 | 2257 | !(req->flags & REQ_F_DONT_REISSUE)) { |
bbde017a | 2258 | req->iopoll_completed = 0; |
8c130827 | 2259 | req_ref_get(req); |
773af691 | 2260 | io_req_task_queue_reissue(req); |
8c130827 | 2261 | continue; |
bbde017a | 2262 | } |
def596e9 | 2263 | |
bcda7baa | 2264 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2265 | cflags = io_put_rw_kbuf(req); |
bcda7baa | 2266 | |
d4d19c19 | 2267 | __io_cqring_fill_event(ctx, req->user_data, req->result, cflags); |
def596e9 JA |
2268 | (*nr_events)++; |
2269 | ||
de9b4cca | 2270 | if (req_ref_put_and_test(req)) |
6ff119a6 | 2271 | io_req_free_batch(&rb, req, &ctx->submit_state); |
def596e9 | 2272 | } |
def596e9 | 2273 | |
09bb8394 | 2274 | io_commit_cqring(ctx); |
80c18e4a | 2275 | io_cqring_ev_posted_iopoll(ctx); |
2d6500d4 | 2276 | io_req_free_batch_finish(ctx, &rb); |
581f9810 BM |
2277 | } |
2278 | ||
def596e9 | 2279 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, |
3c30ef0f | 2280 | long min, bool resubmit) |
def596e9 JA |
2281 | { |
2282 | struct io_kiocb *req, *tmp; | |
2283 | LIST_HEAD(done); | |
2284 | bool spin; | |
def596e9 JA |
2285 | |
2286 | /* | |
2287 | * Only spin for completions if we don't have multiple devices hanging | |
2288 | * off our complete list, and we're under the requested amount. | |
2289 | */ | |
915b3dde | 2290 | spin = !ctx->poll_multi_queue && *nr_events < min; |
def596e9 | 2291 | |
d21ffe7e | 2292 | list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) { |
9adbd45d | 2293 | struct kiocb *kiocb = &req->rw.kiocb; |
a2416e1e | 2294 | int ret; |
def596e9 JA |
2295 | |
2296 | /* | |
581f9810 BM |
2297 | * Move completed and retryable entries to our local lists. |
2298 | * If we find a request that requires polling, break out | |
2299 | * and complete those lists first, if we have entries there. | |
def596e9 | 2300 | */ |
65a6543d | 2301 | if (READ_ONCE(req->iopoll_completed)) { |
d21ffe7e | 2302 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2303 | continue; |
2304 | } | |
2305 | if (!list_empty(&done)) | |
2306 | break; | |
2307 | ||
2308 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
a2416e1e PB |
2309 | if (unlikely(ret < 0)) |
2310 | return ret; | |
2311 | else if (ret) | |
2312 | spin = false; | |
def596e9 | 2313 | |
3aadc23e PB |
2314 | /* iopoll may have completed current req */ |
2315 | if (READ_ONCE(req->iopoll_completed)) | |
d21ffe7e | 2316 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2317 | } |
2318 | ||
2319 | if (!list_empty(&done)) | |
3c30ef0f | 2320 | io_iopoll_complete(ctx, nr_events, &done, resubmit); |
def596e9 | 2321 | |
a2416e1e | 2322 | return 0; |
def596e9 JA |
2323 | } |
2324 | ||
def596e9 JA |
2325 | /* |
2326 | * We can't just wait for polled events to come to us, we have to actively | |
2327 | * find and complete them. | |
2328 | */ | |
b2edc0a7 | 2329 | static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
2330 | { |
2331 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
2332 | return; | |
2333 | ||
2334 | mutex_lock(&ctx->uring_lock); | |
540e32a0 | 2335 | while (!list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2336 | unsigned int nr_events = 0; |
2337 | ||
3c30ef0f | 2338 | io_do_iopoll(ctx, &nr_events, 0, false); |
08f5439f | 2339 | |
b2edc0a7 PB |
2340 | /* let it sleep and repeat later if can't complete a request */ |
2341 | if (nr_events == 0) | |
2342 | break; | |
08f5439f JA |
2343 | /* |
2344 | * Ensure we allow local-to-the-cpu processing to take place, | |
2345 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 2346 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 2347 | */ |
3fcee5a6 PB |
2348 | if (need_resched()) { |
2349 | mutex_unlock(&ctx->uring_lock); | |
2350 | cond_resched(); | |
2351 | mutex_lock(&ctx->uring_lock); | |
2352 | } | |
def596e9 JA |
2353 | } |
2354 | mutex_unlock(&ctx->uring_lock); | |
2355 | } | |
2356 | ||
7668b92a | 2357 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 2358 | { |
7668b92a | 2359 | unsigned int nr_events = 0; |
e9979b36 | 2360 | int ret = 0; |
500f9fba | 2361 | |
c7849be9 XW |
2362 | /* |
2363 | * We disallow the app entering submit/complete with polling, but we | |
2364 | * still need to lock the ring to prevent racing with polled issue | |
2365 | * that got punted to a workqueue. | |
2366 | */ | |
2367 | mutex_lock(&ctx->uring_lock); | |
f39c8a5b PB |
2368 | /* |
2369 | * Don't enter poll loop if we already have events pending. | |
2370 | * If we do, we can potentially be spinning for commands that | |
2371 | * already triggered a CQE (eg in error). | |
2372 | */ | |
5ed7a37d | 2373 | if (test_bit(0, &ctx->check_cq_overflow)) |
f39c8a5b PB |
2374 | __io_cqring_overflow_flush(ctx, false); |
2375 | if (io_cqring_events(ctx)) | |
2376 | goto out; | |
def596e9 | 2377 | do { |
500f9fba JA |
2378 | /* |
2379 | * If a submit got punted to a workqueue, we can have the | |
2380 | * application entering polling for a command before it gets | |
2381 | * issued. That app will hold the uring_lock for the duration | |
2382 | * of the poll right here, so we need to take a breather every | |
2383 | * now and then to ensure that the issue has a chance to add | |
2384 | * the poll to the issued list. Otherwise we can spin here | |
2385 | * forever, while the workqueue is stuck trying to acquire the | |
2386 | * very same mutex. | |
2387 | */ | |
e9979b36 | 2388 | if (list_empty(&ctx->iopoll_list)) { |
8f487ef2 PB |
2389 | u32 tail = ctx->cached_cq_tail; |
2390 | ||
500f9fba | 2391 | mutex_unlock(&ctx->uring_lock); |
4c6e277c | 2392 | io_run_task_work(); |
500f9fba | 2393 | mutex_lock(&ctx->uring_lock); |
def596e9 | 2394 | |
8f487ef2 PB |
2395 | /* some requests don't go through iopoll_list */ |
2396 | if (tail != ctx->cached_cq_tail || | |
2397 | list_empty(&ctx->iopoll_list)) | |
e9979b36 | 2398 | break; |
500f9fba | 2399 | } |
3c30ef0f | 2400 | ret = io_do_iopoll(ctx, &nr_events, min, true); |
f39c8a5b PB |
2401 | } while (!ret && nr_events < min && !need_resched()); |
2402 | out: | |
500f9fba | 2403 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
2404 | return ret; |
2405 | } | |
2406 | ||
491381ce | 2407 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 2408 | { |
491381ce JA |
2409 | /* |
2410 | * Tell lockdep we inherited freeze protection from submission | |
2411 | * thread. | |
2412 | */ | |
2413 | if (req->flags & REQ_F_ISREG) { | |
1c98679d | 2414 | struct super_block *sb = file_inode(req->file)->i_sb; |
2b188cc1 | 2415 | |
1c98679d PB |
2416 | __sb_writers_acquired(sb, SB_FREEZE_WRITE); |
2417 | sb_end_write(sb); | |
2b188cc1 JA |
2418 | } |
2419 | } | |
2420 | ||
b63534c4 | 2421 | #ifdef CONFIG_BLOCK |
dc2a6e9a | 2422 | static bool io_resubmit_prep(struct io_kiocb *req) |
b63534c4 | 2423 | { |
ab454438 | 2424 | struct io_async_rw *rw = req->async_data; |
b63534c4 | 2425 | |
ab454438 PB |
2426 | if (!rw) |
2427 | return !io_req_prep_async(req); | |
2428 | /* may have left rw->iter inconsistent on -EIOCBQUEUED */ | |
2429 | iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter)); | |
2430 | return true; | |
b63534c4 | 2431 | } |
b63534c4 | 2432 | |
3e6a0d3c | 2433 | static bool io_rw_should_reissue(struct io_kiocb *req) |
b63534c4 | 2434 | { |
355afaeb | 2435 | umode_t mode = file_inode(req->file)->i_mode; |
3e6a0d3c | 2436 | struct io_ring_ctx *ctx = req->ctx; |
b63534c4 | 2437 | |
355afaeb JA |
2438 | if (!S_ISBLK(mode) && !S_ISREG(mode)) |
2439 | return false; | |
3e6a0d3c JA |
2440 | if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() && |
2441 | !(ctx->flags & IORING_SETUP_IOPOLL))) | |
b63534c4 | 2442 | return false; |
7c977a58 JA |
2443 | /* |
2444 | * If ref is dying, we might be running poll reap from the exit work. | |
2445 | * Don't attempt to reissue from that path, just let it fail with | |
2446 | * -EAGAIN. | |
2447 | */ | |
3e6a0d3c JA |
2448 | if (percpu_ref_is_dying(&ctx->refs)) |
2449 | return false; | |
ef046888 JA |
2450 | /* |
2451 | * Play it safe and assume not safe to re-import and reissue if we're | |
2452 | * not in the original thread group (or in task context). | |
2453 | */ | |
2454 | if (!same_thread_group(req->task, current) || !in_task()) | |
2455 | return false; | |
3e6a0d3c JA |
2456 | return true; |
2457 | } | |
e82ad485 | 2458 | #else |
a1ff1e3f | 2459 | static bool io_resubmit_prep(struct io_kiocb *req) |
e82ad485 JA |
2460 | { |
2461 | return false; | |
2462 | } | |
e82ad485 | 2463 | static bool io_rw_should_reissue(struct io_kiocb *req) |
3e6a0d3c | 2464 | { |
b63534c4 JA |
2465 | return false; |
2466 | } | |
3e6a0d3c | 2467 | #endif |
b63534c4 | 2468 | |
a1d7c393 | 2469 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
889fca73 | 2470 | unsigned int issue_flags) |
a1d7c393 | 2471 | { |
2f8e45f1 PB |
2472 | int cflags = 0; |
2473 | ||
b65c128f PB |
2474 | if (req->rw.kiocb.ki_flags & IOCB_WRITE) |
2475 | kiocb_end_write(req); | |
9532b99b PB |
2476 | if (res != req->result) { |
2477 | if ((res == -EAGAIN || res == -EOPNOTSUPP) && | |
2478 | io_rw_should_reissue(req)) { | |
2479 | req->flags |= REQ_F_REISSUE; | |
2480 | return; | |
2481 | } | |
93d2bcd2 | 2482 | req_set_fail(req); |
9532b99b | 2483 | } |
2f8e45f1 PB |
2484 | if (req->flags & REQ_F_BUFFER_SELECTED) |
2485 | cflags = io_put_rw_kbuf(req); | |
2486 | __io_req_complete(req, issue_flags, res, cflags); | |
ba816ad6 JA |
2487 | } |
2488 | ||
2489 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
2490 | { | |
9adbd45d | 2491 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 | 2492 | |
889fca73 | 2493 | __io_complete_rw(req, res, res2, 0); |
2b188cc1 JA |
2494 | } |
2495 | ||
def596e9 JA |
2496 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
2497 | { | |
9adbd45d | 2498 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 2499 | |
491381ce JA |
2500 | if (kiocb->ki_flags & IOCB_WRITE) |
2501 | kiocb_end_write(req); | |
9532b99b | 2502 | if (unlikely(res != req->result)) { |
a1ff1e3f JA |
2503 | if (!(res == -EAGAIN && io_rw_should_reissue(req) && |
2504 | io_resubmit_prep(req))) { | |
93d2bcd2 | 2505 | req_set_fail(req); |
9532b99b PB |
2506 | req->flags |= REQ_F_DONT_REISSUE; |
2507 | } | |
8c130827 | 2508 | } |
bbde017a XW |
2509 | |
2510 | WRITE_ONCE(req->result, res); | |
b9b0e0d3 | 2511 | /* order with io_iopoll_complete() checking ->result */ |
cd664b0e PB |
2512 | smp_wmb(); |
2513 | WRITE_ONCE(req->iopoll_completed, 1); | |
def596e9 JA |
2514 | } |
2515 | ||
2516 | /* | |
2517 | * After the iocb has been issued, it's safe to be found on the poll list. | |
2518 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 2519 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
2520 | * accessing the kiocb cookie. |
2521 | */ | |
cb3d8972 | 2522 | static void io_iopoll_req_issued(struct io_kiocb *req) |
def596e9 JA |
2523 | { |
2524 | struct io_ring_ctx *ctx = req->ctx; | |
cb3d8972 PB |
2525 | const bool in_async = io_wq_current_is_worker(); |
2526 | ||
2527 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
2528 | if (unlikely(in_async)) | |
2529 | mutex_lock(&ctx->uring_lock); | |
def596e9 JA |
2530 | |
2531 | /* | |
2532 | * Track whether we have multiple files in our lists. This will impact | |
2533 | * how we do polling eventually, not spinning if we're on potentially | |
2534 | * different devices. | |
2535 | */ | |
540e32a0 | 2536 | if (list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
2537 | ctx->poll_multi_queue = false; |
2538 | } else if (!ctx->poll_multi_queue) { | |
def596e9 | 2539 | struct io_kiocb *list_req; |
915b3dde | 2540 | unsigned int queue_num0, queue_num1; |
def596e9 | 2541 | |
540e32a0 | 2542 | list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb, |
d21ffe7e | 2543 | inflight_entry); |
915b3dde HX |
2544 | |
2545 | if (list_req->file != req->file) { | |
2546 | ctx->poll_multi_queue = true; | |
2547 | } else { | |
2548 | queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie); | |
2549 | queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie); | |
2550 | if (queue_num0 != queue_num1) | |
2551 | ctx->poll_multi_queue = true; | |
2552 | } | |
def596e9 JA |
2553 | } |
2554 | ||
2555 | /* | |
2556 | * For fast devices, IO may have already completed. If it has, add | |
2557 | * it to the front so we find it first. | |
2558 | */ | |
65a6543d | 2559 | if (READ_ONCE(req->iopoll_completed)) |
d21ffe7e | 2560 | list_add(&req->inflight_entry, &ctx->iopoll_list); |
def596e9 | 2561 | else |
d21ffe7e | 2562 | list_add_tail(&req->inflight_entry, &ctx->iopoll_list); |
bdcd3eab | 2563 | |
cb3d8972 PB |
2564 | if (unlikely(in_async)) { |
2565 | /* | |
2566 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
2567 | * in sq thread task context or in io worker task context. If | |
2568 | * current task context is sq thread, we don't need to check | |
2569 | * whether should wake up sq thread. | |
2570 | */ | |
2571 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
2572 | wq_has_sleeper(&ctx->sq_data->wait)) | |
2573 | wake_up(&ctx->sq_data->wait); | |
2574 | ||
2575 | mutex_unlock(&ctx->uring_lock); | |
2576 | } | |
def596e9 JA |
2577 | } |
2578 | ||
4503b767 JA |
2579 | static bool io_bdev_nowait(struct block_device *bdev) |
2580 | { | |
9ba0d0c8 | 2581 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
2582 | } |
2583 | ||
2b188cc1 JA |
2584 | /* |
2585 | * If we tracked the file through the SCM inflight mechanism, we could support | |
2586 | * any file. For now, just ensure that anything potentially problematic is done | |
2587 | * inline. | |
2588 | */ | |
b191e2df | 2589 | static bool __io_file_supports_nowait(struct file *file, int rw) |
2b188cc1 JA |
2590 | { |
2591 | umode_t mode = file_inode(file)->i_mode; | |
2592 | ||
4503b767 | 2593 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
2594 | if (IS_ENABLED(CONFIG_BLOCK) && |
2595 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
2596 | return true; |
2597 | return false; | |
2598 | } | |
976517f1 | 2599 | if (S_ISSOCK(mode)) |
2b188cc1 | 2600 | return true; |
4503b767 | 2601 | if (S_ISREG(mode)) { |
4e7b5671 CH |
2602 | if (IS_ENABLED(CONFIG_BLOCK) && |
2603 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
4503b767 JA |
2604 | file->f_op != &io_uring_fops) |
2605 | return true; | |
2606 | return false; | |
2607 | } | |
2b188cc1 | 2608 | |
c5b85625 JA |
2609 | /* any ->read/write should understand O_NONBLOCK */ |
2610 | if (file->f_flags & O_NONBLOCK) | |
2611 | return true; | |
2612 | ||
af197f50 JA |
2613 | if (!(file->f_mode & FMODE_NOWAIT)) |
2614 | return false; | |
2615 | ||
2616 | if (rw == READ) | |
2617 | return file->f_op->read_iter != NULL; | |
2618 | ||
2619 | return file->f_op->write_iter != NULL; | |
2b188cc1 JA |
2620 | } |
2621 | ||
b191e2df | 2622 | static bool io_file_supports_nowait(struct io_kiocb *req, int rw) |
7b29f92d | 2623 | { |
b191e2df | 2624 | if (rw == READ && (req->flags & REQ_F_NOWAIT_READ)) |
7b29f92d | 2625 | return true; |
b191e2df | 2626 | else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE)) |
7b29f92d JA |
2627 | return true; |
2628 | ||
b191e2df | 2629 | return __io_file_supports_nowait(req->file, rw); |
7b29f92d JA |
2630 | } |
2631 | ||
a88fc400 | 2632 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 2633 | { |
def596e9 | 2634 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 2635 | struct kiocb *kiocb = &req->rw.kiocb; |
75c668cd | 2636 | struct file *file = req->file; |
09bb8394 JA |
2637 | unsigned ioprio; |
2638 | int ret; | |
2b188cc1 | 2639 | |
c97d8a0f | 2640 | if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode)) |
491381ce JA |
2641 | req->flags |= REQ_F_ISREG; |
2642 | ||
2b188cc1 | 2643 | kiocb->ki_pos = READ_ONCE(sqe->off); |
75c668cd | 2644 | if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) { |
ba04291e | 2645 | req->flags |= REQ_F_CUR_POS; |
75c668cd | 2646 | kiocb->ki_pos = file->f_pos; |
ba04291e | 2647 | } |
2b188cc1 | 2648 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
2649 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
2650 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
2651 | if (unlikely(ret)) | |
2652 | return ret; | |
2b188cc1 | 2653 | |
75c668cd PB |
2654 | /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */ |
2655 | if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK)) | |
2656 | req->flags |= REQ_F_NOWAIT; | |
2657 | ||
2b188cc1 JA |
2658 | ioprio = READ_ONCE(sqe->ioprio); |
2659 | if (ioprio) { | |
2660 | ret = ioprio_check_cap(ioprio); | |
2661 | if (ret) | |
09bb8394 | 2662 | return ret; |
2b188cc1 JA |
2663 | |
2664 | kiocb->ki_ioprio = ioprio; | |
2665 | } else | |
2666 | kiocb->ki_ioprio = get_current_ioprio(); | |
2667 | ||
def596e9 | 2668 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2669 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2670 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2671 | return -EOPNOTSUPP; |
2b188cc1 | 2672 | |
def596e9 JA |
2673 | kiocb->ki_flags |= IOCB_HIPRI; |
2674 | kiocb->ki_complete = io_complete_rw_iopoll; | |
65a6543d | 2675 | req->iopoll_completed = 0; |
def596e9 | 2676 | } else { |
09bb8394 JA |
2677 | if (kiocb->ki_flags & IOCB_HIPRI) |
2678 | return -EINVAL; | |
def596e9 JA |
2679 | kiocb->ki_complete = io_complete_rw; |
2680 | } | |
9adbd45d | 2681 | |
eae071c9 PB |
2682 | if (req->opcode == IORING_OP_READ_FIXED || |
2683 | req->opcode == IORING_OP_WRITE_FIXED) { | |
2684 | req->imu = NULL; | |
2685 | io_req_set_rsrc_node(req); | |
2686 | } | |
2687 | ||
3529d8c2 JA |
2688 | req->rw.addr = READ_ONCE(sqe->addr); |
2689 | req->rw.len = READ_ONCE(sqe->len); | |
4f4eeba8 | 2690 | req->buf_index = READ_ONCE(sqe->buf_index); |
2b188cc1 | 2691 | return 0; |
2b188cc1 JA |
2692 | } |
2693 | ||
2694 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2695 | { | |
2696 | switch (ret) { | |
2697 | case -EIOCBQUEUED: | |
2698 | break; | |
2699 | case -ERESTARTSYS: | |
2700 | case -ERESTARTNOINTR: | |
2701 | case -ERESTARTNOHAND: | |
2702 | case -ERESTART_RESTARTBLOCK: | |
2703 | /* | |
2704 | * We can't just restart the syscall, since previously | |
2705 | * submitted sqes may already be in progress. Just fail this | |
2706 | * IO with EINTR. | |
2707 | */ | |
2708 | ret = -EINTR; | |
df561f66 | 2709 | fallthrough; |
2b188cc1 JA |
2710 | default: |
2711 | kiocb->ki_complete(kiocb, ret, 0); | |
2712 | } | |
2713 | } | |
2714 | ||
a1d7c393 | 2715 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, |
889fca73 | 2716 | unsigned int issue_flags) |
ba816ad6 | 2717 | { |
ba04291e | 2718 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
e8c2bc1f | 2719 | struct io_async_rw *io = req->async_data; |
97284637 | 2720 | bool check_reissue = kiocb->ki_complete == io_complete_rw; |
ba04291e | 2721 | |
227c0c96 | 2722 | /* add previously done IO, if any */ |
e8c2bc1f | 2723 | if (io && io->bytes_done > 0) { |
227c0c96 | 2724 | if (ret < 0) |
e8c2bc1f | 2725 | ret = io->bytes_done; |
227c0c96 | 2726 | else |
e8c2bc1f | 2727 | ret += io->bytes_done; |
227c0c96 JA |
2728 | } |
2729 | ||
ba04291e JA |
2730 | if (req->flags & REQ_F_CUR_POS) |
2731 | req->file->f_pos = kiocb->ki_pos; | |
e149bd74 | 2732 | if (ret >= 0 && check_reissue) |
889fca73 | 2733 | __io_complete_rw(req, ret, 0, issue_flags); |
ba816ad6 JA |
2734 | else |
2735 | io_rw_done(kiocb, ret); | |
97284637 | 2736 | |
fe7e3257 | 2737 | if (check_reissue && (req->flags & REQ_F_REISSUE)) { |
97284637 | 2738 | req->flags &= ~REQ_F_REISSUE; |
a7be7c23 | 2739 | if (io_resubmit_prep(req)) { |
8c130827 | 2740 | req_ref_get(req); |
773af691 | 2741 | io_req_task_queue_reissue(req); |
8c130827 | 2742 | } else { |
97284637 PB |
2743 | int cflags = 0; |
2744 | ||
93d2bcd2 | 2745 | req_set_fail(req); |
97284637 PB |
2746 | if (req->flags & REQ_F_BUFFER_SELECTED) |
2747 | cflags = io_put_rw_kbuf(req); | |
2748 | __io_req_complete(req, issue_flags, ret, cflags); | |
2749 | } | |
2750 | } | |
ba816ad6 JA |
2751 | } |
2752 | ||
eae071c9 PB |
2753 | static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter, |
2754 | struct io_mapped_ubuf *imu) | |
edafccee | 2755 | { |
9adbd45d | 2756 | size_t len = req->rw.len; |
75769e3f | 2757 | u64 buf_end, buf_addr = req->rw.addr; |
edafccee | 2758 | size_t offset; |
edafccee | 2759 | |
75769e3f | 2760 | if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) |
edafccee JA |
2761 | return -EFAULT; |
2762 | /* not inside the mapped region */ | |
4751f53d | 2763 | if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) |
edafccee JA |
2764 | return -EFAULT; |
2765 | ||
2766 | /* | |
2767 | * May not be a start of buffer, set size appropriately | |
2768 | * and advance us to the beginning. | |
2769 | */ | |
2770 | offset = buf_addr - imu->ubuf; | |
2771 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
2772 | |
2773 | if (offset) { | |
2774 | /* | |
2775 | * Don't use iov_iter_advance() here, as it's really slow for | |
2776 | * using the latter parts of a big fixed buffer - it iterates | |
2777 | * over each segment manually. We can cheat a bit here, because | |
2778 | * we know that: | |
2779 | * | |
2780 | * 1) it's a BVEC iter, we set it up | |
2781 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
2782 | * first and last bvec | |
2783 | * | |
2784 | * So just find our index, and adjust the iterator afterwards. | |
2785 | * If the offset is within the first bvec (or the whole first | |
2786 | * bvec, just use iov_iter_advance(). This makes it easier | |
2787 | * since we can just skip the first segment, which may not | |
2788 | * be PAGE_SIZE aligned. | |
2789 | */ | |
2790 | const struct bio_vec *bvec = imu->bvec; | |
2791 | ||
2792 | if (offset <= bvec->bv_len) { | |
2793 | iov_iter_advance(iter, offset); | |
2794 | } else { | |
2795 | unsigned long seg_skip; | |
2796 | ||
2797 | /* skip first vec */ | |
2798 | offset -= bvec->bv_len; | |
2799 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
2800 | ||
2801 | iter->bvec = bvec + seg_skip; | |
2802 | iter->nr_segs -= seg_skip; | |
99c79f66 | 2803 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 2804 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
2805 | } |
2806 | } | |
2807 | ||
847595de | 2808 | return 0; |
edafccee JA |
2809 | } |
2810 | ||
eae071c9 PB |
2811 | static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter) |
2812 | { | |
2813 | struct io_ring_ctx *ctx = req->ctx; | |
2814 | struct io_mapped_ubuf *imu = req->imu; | |
2815 | u16 index, buf_index = req->buf_index; | |
2816 | ||
2817 | if (likely(!imu)) { | |
2818 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
2819 | return -EFAULT; | |
2820 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
2821 | imu = READ_ONCE(ctx->user_bufs[index]); | |
2822 | req->imu = imu; | |
2823 | } | |
2824 | return __io_import_fixed(req, rw, iter, imu); | |
2825 | } | |
2826 | ||
bcda7baa JA |
2827 | static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock) |
2828 | { | |
2829 | if (needs_lock) | |
2830 | mutex_unlock(&ctx->uring_lock); | |
2831 | } | |
2832 | ||
2833 | static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock) | |
2834 | { | |
2835 | /* | |
2836 | * "Normal" inline submissions always hold the uring_lock, since we | |
2837 | * grab it from the system call. Same is true for the SQPOLL offload. | |
2838 | * The only exception is when we've detached the request and issue it | |
2839 | * from an async worker thread, grab the lock for that case. | |
2840 | */ | |
2841 | if (needs_lock) | |
2842 | mutex_lock(&ctx->uring_lock); | |
2843 | } | |
2844 | ||
2845 | static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len, | |
2846 | int bgid, struct io_buffer *kbuf, | |
2847 | bool needs_lock) | |
2848 | { | |
2849 | struct io_buffer *head; | |
2850 | ||
2851 | if (req->flags & REQ_F_BUFFER_SELECTED) | |
2852 | return kbuf; | |
2853 | ||
2854 | io_ring_submit_lock(req->ctx, needs_lock); | |
2855 | ||
2856 | lockdep_assert_held(&req->ctx->uring_lock); | |
2857 | ||
9e15c3a0 | 2858 | head = xa_load(&req->ctx->io_buffers, bgid); |
bcda7baa JA |
2859 | if (head) { |
2860 | if (!list_empty(&head->list)) { | |
2861 | kbuf = list_last_entry(&head->list, struct io_buffer, | |
2862 | list); | |
2863 | list_del(&kbuf->list); | |
2864 | } else { | |
2865 | kbuf = head; | |
9e15c3a0 | 2866 | xa_erase(&req->ctx->io_buffers, bgid); |
bcda7baa JA |
2867 | } |
2868 | if (*len > kbuf->len) | |
2869 | *len = kbuf->len; | |
2870 | } else { | |
2871 | kbuf = ERR_PTR(-ENOBUFS); | |
2872 | } | |
2873 | ||
2874 | io_ring_submit_unlock(req->ctx, needs_lock); | |
2875 | ||
2876 | return kbuf; | |
2877 | } | |
2878 | ||
4d954c25 JA |
2879 | static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len, |
2880 | bool needs_lock) | |
2881 | { | |
2882 | struct io_buffer *kbuf; | |
4f4eeba8 | 2883 | u16 bgid; |
4d954c25 JA |
2884 | |
2885 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
4f4eeba8 | 2886 | bgid = req->buf_index; |
4d954c25 JA |
2887 | kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock); |
2888 | if (IS_ERR(kbuf)) | |
2889 | return kbuf; | |
2890 | req->rw.addr = (u64) (unsigned long) kbuf; | |
2891 | req->flags |= REQ_F_BUFFER_SELECTED; | |
2892 | return u64_to_user_ptr(kbuf->addr); | |
2893 | } | |
2894 | ||
2895 | #ifdef CONFIG_COMPAT | |
2896 | static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov, | |
2897 | bool needs_lock) | |
2898 | { | |
2899 | struct compat_iovec __user *uiov; | |
2900 | compat_ssize_t clen; | |
2901 | void __user *buf; | |
2902 | ssize_t len; | |
2903 | ||
2904 | uiov = u64_to_user_ptr(req->rw.addr); | |
2905 | if (!access_ok(uiov, sizeof(*uiov))) | |
2906 | return -EFAULT; | |
2907 | if (__get_user(clen, &uiov->iov_len)) | |
2908 | return -EFAULT; | |
2909 | if (clen < 0) | |
2910 | return -EINVAL; | |
2911 | ||
2912 | len = clen; | |
2913 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
2914 | if (IS_ERR(buf)) | |
2915 | return PTR_ERR(buf); | |
2916 | iov[0].iov_base = buf; | |
2917 | iov[0].iov_len = (compat_size_t) len; | |
2918 | return 0; | |
2919 | } | |
2920 | #endif | |
2921 | ||
2922 | static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
2923 | bool needs_lock) | |
2924 | { | |
2925 | struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr); | |
2926 | void __user *buf; | |
2927 | ssize_t len; | |
2928 | ||
2929 | if (copy_from_user(iov, uiov, sizeof(*uiov))) | |
2930 | return -EFAULT; | |
2931 | ||
2932 | len = iov[0].iov_len; | |
2933 | if (len < 0) | |
2934 | return -EINVAL; | |
2935 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
2936 | if (IS_ERR(buf)) | |
2937 | return PTR_ERR(buf); | |
2938 | iov[0].iov_base = buf; | |
2939 | iov[0].iov_len = len; | |
2940 | return 0; | |
2941 | } | |
2942 | ||
2943 | static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
2944 | bool needs_lock) | |
2945 | { | |
dddb3e26 JA |
2946 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
2947 | struct io_buffer *kbuf; | |
2948 | ||
2949 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
2950 | iov[0].iov_base = u64_to_user_ptr(kbuf->addr); | |
2951 | iov[0].iov_len = kbuf->len; | |
4d954c25 | 2952 | return 0; |
dddb3e26 | 2953 | } |
dd201662 | 2954 | if (req->rw.len != 1) |
4d954c25 JA |
2955 | return -EINVAL; |
2956 | ||
2957 | #ifdef CONFIG_COMPAT | |
2958 | if (req->ctx->compat) | |
2959 | return io_compat_import(req, iov, needs_lock); | |
2960 | #endif | |
2961 | ||
2962 | return __io_iov_buffer_select(req, iov, needs_lock); | |
2963 | } | |
2964 | ||
847595de PB |
2965 | static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec, |
2966 | struct iov_iter *iter, bool needs_lock) | |
2b188cc1 | 2967 | { |
9adbd45d JA |
2968 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
2969 | size_t sqe_len = req->rw.len; | |
847595de | 2970 | u8 opcode = req->opcode; |
4d954c25 | 2971 | ssize_t ret; |
edafccee | 2972 | |
7d009165 | 2973 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 2974 | *iovec = NULL; |
9adbd45d | 2975 | return io_import_fixed(req, rw, iter); |
edafccee | 2976 | } |
2b188cc1 | 2977 | |
bcda7baa | 2978 | /* buffer index only valid with fixed read/write, or buffer select */ |
4f4eeba8 | 2979 | if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)) |
9adbd45d JA |
2980 | return -EINVAL; |
2981 | ||
3a6820f2 | 2982 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
bcda7baa | 2983 | if (req->flags & REQ_F_BUFFER_SELECT) { |
4d954c25 | 2984 | buf = io_rw_buffer_select(req, &sqe_len, needs_lock); |
867a23ea | 2985 | if (IS_ERR(buf)) |
4d954c25 | 2986 | return PTR_ERR(buf); |
3f9d6441 | 2987 | req->rw.len = sqe_len; |
bcda7baa JA |
2988 | } |
2989 | ||
3a6820f2 JA |
2990 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); |
2991 | *iovec = NULL; | |
10fc72e4 | 2992 | return ret; |
3a6820f2 JA |
2993 | } |
2994 | ||
4d954c25 JA |
2995 | if (req->flags & REQ_F_BUFFER_SELECT) { |
2996 | ret = io_iov_buffer_select(req, *iovec, needs_lock); | |
847595de PB |
2997 | if (!ret) |
2998 | iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len); | |
4d954c25 JA |
2999 | *iovec = NULL; |
3000 | return ret; | |
3001 | } | |
3002 | ||
89cd35c5 CH |
3003 | return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter, |
3004 | req->ctx->compat); | |
2b188cc1 JA |
3005 | } |
3006 | ||
0fef9483 JA |
3007 | static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) |
3008 | { | |
5b09e37e | 3009 | return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos; |
0fef9483 JA |
3010 | } |
3011 | ||
31b51510 | 3012 | /* |
32960613 JA |
3013 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
3014 | * by looping over ->read() or ->write() manually. | |
31b51510 | 3015 | */ |
4017eb91 | 3016 | static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter) |
32960613 | 3017 | { |
4017eb91 JA |
3018 | struct kiocb *kiocb = &req->rw.kiocb; |
3019 | struct file *file = req->file; | |
32960613 JA |
3020 | ssize_t ret = 0; |
3021 | ||
3022 | /* | |
3023 | * Don't support polled IO through this interface, and we can't | |
3024 | * support non-blocking either. For the latter, this just causes | |
3025 | * the kiocb to be handled from an async context. | |
3026 | */ | |
3027 | if (kiocb->ki_flags & IOCB_HIPRI) | |
3028 | return -EOPNOTSUPP; | |
3029 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
3030 | return -EAGAIN; | |
3031 | ||
3032 | while (iov_iter_count(iter)) { | |
311ae9e1 | 3033 | struct iovec iovec; |
32960613 JA |
3034 | ssize_t nr; |
3035 | ||
311ae9e1 PB |
3036 | if (!iov_iter_is_bvec(iter)) { |
3037 | iovec = iov_iter_iovec(iter); | |
3038 | } else { | |
4017eb91 JA |
3039 | iovec.iov_base = u64_to_user_ptr(req->rw.addr); |
3040 | iovec.iov_len = req->rw.len; | |
311ae9e1 PB |
3041 | } |
3042 | ||
32960613 JA |
3043 | if (rw == READ) { |
3044 | nr = file->f_op->read(file, iovec.iov_base, | |
0fef9483 | 3045 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3046 | } else { |
3047 | nr = file->f_op->write(file, iovec.iov_base, | |
0fef9483 | 3048 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3049 | } |
3050 | ||
3051 | if (nr < 0) { | |
3052 | if (!ret) | |
3053 | ret = nr; | |
3054 | break; | |
3055 | } | |
3056 | ret += nr; | |
3057 | if (nr != iovec.iov_len) | |
3058 | break; | |
4017eb91 JA |
3059 | req->rw.len -= nr; |
3060 | req->rw.addr += nr; | |
32960613 JA |
3061 | iov_iter_advance(iter, nr); |
3062 | } | |
3063 | ||
3064 | return ret; | |
3065 | } | |
3066 | ||
ff6165b2 JA |
3067 | static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec, |
3068 | const struct iovec *fast_iov, struct iov_iter *iter) | |
f67676d1 | 3069 | { |
e8c2bc1f | 3070 | struct io_async_rw *rw = req->async_data; |
b64e3444 | 3071 | |
ff6165b2 | 3072 | memcpy(&rw->iter, iter, sizeof(*iter)); |
afb87658 | 3073 | rw->free_iovec = iovec; |
227c0c96 | 3074 | rw->bytes_done = 0; |
ff6165b2 | 3075 | /* can only be fixed buffers, no need to do anything */ |
9c3a205c | 3076 | if (iov_iter_is_bvec(iter)) |
ff6165b2 | 3077 | return; |
b64e3444 | 3078 | if (!iovec) { |
ff6165b2 JA |
3079 | unsigned iov_off = 0; |
3080 | ||
3081 | rw->iter.iov = rw->fast_iov; | |
3082 | if (iter->iov != fast_iov) { | |
3083 | iov_off = iter->iov - fast_iov; | |
3084 | rw->iter.iov += iov_off; | |
3085 | } | |
3086 | if (rw->fast_iov != fast_iov) | |
3087 | memcpy(rw->fast_iov + iov_off, fast_iov + iov_off, | |
45097dae | 3088 | sizeof(struct iovec) * iter->nr_segs); |
99bc4c38 PB |
3089 | } else { |
3090 | req->flags |= REQ_F_NEED_CLEANUP; | |
f67676d1 JA |
3091 | } |
3092 | } | |
3093 | ||
6cb78689 | 3094 | static inline int io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 3095 | { |
e8c2bc1f JA |
3096 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
3097 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
3098 | return req->async_data == NULL; | |
3d9932a8 XW |
3099 | } |
3100 | ||
ff6165b2 JA |
3101 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
3102 | const struct iovec *fast_iov, | |
227c0c96 | 3103 | struct iov_iter *iter, bool force) |
b7bb4f7d | 3104 | { |
26f0505a | 3105 | if (!force && !io_op_defs[req->opcode].needs_async_setup) |
74566df3 | 3106 | return 0; |
e8c2bc1f | 3107 | if (!req->async_data) { |
6cb78689 | 3108 | if (io_alloc_async_data(req)) { |
6bf985dc | 3109 | kfree(iovec); |
5d204bcf | 3110 | return -ENOMEM; |
6bf985dc | 3111 | } |
b7bb4f7d | 3112 | |
ff6165b2 | 3113 | io_req_map_rw(req, iovec, fast_iov, iter); |
5d204bcf | 3114 | } |
b7bb4f7d | 3115 | return 0; |
f67676d1 JA |
3116 | } |
3117 | ||
73debe68 | 3118 | static inline int io_rw_prep_async(struct io_kiocb *req, int rw) |
c3e330a4 | 3119 | { |
e8c2bc1f | 3120 | struct io_async_rw *iorw = req->async_data; |
f4bff104 | 3121 | struct iovec *iov = iorw->fast_iov; |
847595de | 3122 | int ret; |
c3e330a4 | 3123 | |
2846c481 | 3124 | ret = io_import_iovec(rw, req, &iov, &iorw->iter, false); |
c3e330a4 PB |
3125 | if (unlikely(ret < 0)) |
3126 | return ret; | |
3127 | ||
ab0b196c PB |
3128 | iorw->bytes_done = 0; |
3129 | iorw->free_iovec = iov; | |
3130 | if (iov) | |
3131 | req->flags |= REQ_F_NEED_CLEANUP; | |
c3e330a4 PB |
3132 | return 0; |
3133 | } | |
3134 | ||
73debe68 | 3135 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 3136 | { |
3529d8c2 JA |
3137 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
3138 | return -EBADF; | |
93642ef8 | 3139 | return io_prep_rw(req, sqe); |
f67676d1 JA |
3140 | } |
3141 | ||
c1dd91d1 JA |
3142 | /* |
3143 | * This is our waitqueue callback handler, registered through lock_page_async() | |
3144 | * when we initially tried to do the IO with the iocb armed our waitqueue. | |
3145 | * This gets called when the page is unlocked, and we generally expect that to | |
3146 | * happen when the page IO is completed and the page is now uptodate. This will | |
3147 | * queue a task_work based retry of the operation, attempting to copy the data | |
3148 | * again. If the latter fails because the page was NOT uptodate, then we will | |
3149 | * do a thread based blocking retry of the operation. That's the unexpected | |
3150 | * slow path. | |
3151 | */ | |
bcf5a063 JA |
3152 | static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, |
3153 | int sync, void *arg) | |
3154 | { | |
3155 | struct wait_page_queue *wpq; | |
3156 | struct io_kiocb *req = wait->private; | |
bcf5a063 | 3157 | struct wait_page_key *key = arg; |
bcf5a063 JA |
3158 | |
3159 | wpq = container_of(wait, struct wait_page_queue, wait); | |
3160 | ||
cdc8fcb4 LT |
3161 | if (!wake_page_match(wpq, key)) |
3162 | return 0; | |
3163 | ||
c8d317aa | 3164 | req->rw.kiocb.ki_flags &= ~IOCB_WAITQ; |
bcf5a063 JA |
3165 | list_del_init(&wait->entry); |
3166 | ||
bcf5a063 | 3167 | /* submit ref gets dropped, acquire a new one */ |
de9b4cca | 3168 | req_ref_get(req); |
921b9054 | 3169 | io_req_task_queue(req); |
bcf5a063 JA |
3170 | return 1; |
3171 | } | |
3172 | ||
c1dd91d1 JA |
3173 | /* |
3174 | * This controls whether a given IO request should be armed for async page | |
3175 | * based retry. If we return false here, the request is handed to the async | |
3176 | * worker threads for retry. If we're doing buffered reads on a regular file, | |
3177 | * we prepare a private wait_page_queue entry and retry the operation. This | |
3178 | * will either succeed because the page is now uptodate and unlocked, or it | |
3179 | * will register a callback when the page is unlocked at IO completion. Through | |
3180 | * that callback, io_uring uses task_work to setup a retry of the operation. | |
3181 | * That retry will attempt the buffered read again. The retry will generally | |
3182 | * succeed, or in rare cases where it fails, we then fall back to using the | |
3183 | * async worker threads for a blocking retry. | |
3184 | */ | |
227c0c96 | 3185 | static bool io_rw_should_retry(struct io_kiocb *req) |
f67676d1 | 3186 | { |
e8c2bc1f JA |
3187 | struct io_async_rw *rw = req->async_data; |
3188 | struct wait_page_queue *wait = &rw->wpq; | |
bcf5a063 | 3189 | struct kiocb *kiocb = &req->rw.kiocb; |
f67676d1 | 3190 | |
bcf5a063 JA |
3191 | /* never retry for NOWAIT, we just complete with -EAGAIN */ |
3192 | if (req->flags & REQ_F_NOWAIT) | |
3193 | return false; | |
f67676d1 | 3194 | |
227c0c96 | 3195 | /* Only for buffered IO */ |
3b2a4439 | 3196 | if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) |
bcf5a063 | 3197 | return false; |
3b2a4439 | 3198 | |
bcf5a063 JA |
3199 | /* |
3200 | * just use poll if we can, and don't attempt if the fs doesn't | |
3201 | * support callback based unlocks | |
3202 | */ | |
3203 | if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC)) | |
3204 | return false; | |
f67676d1 | 3205 | |
3b2a4439 JA |
3206 | wait->wait.func = io_async_buf_func; |
3207 | wait->wait.private = req; | |
3208 | wait->wait.flags = 0; | |
3209 | INIT_LIST_HEAD(&wait->wait.entry); | |
3210 | kiocb->ki_flags |= IOCB_WAITQ; | |
c8d317aa | 3211 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3b2a4439 | 3212 | kiocb->ki_waitq = wait; |
3b2a4439 | 3213 | return true; |
bcf5a063 JA |
3214 | } |
3215 | ||
aeab9506 | 3216 | static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter) |
bcf5a063 JA |
3217 | { |
3218 | if (req->file->f_op->read_iter) | |
3219 | return call_read_iter(req->file, &req->rw.kiocb, iter); | |
2dd2111d | 3220 | else if (req->file->f_op->read) |
4017eb91 | 3221 | return loop_rw_iter(READ, req, iter); |
2dd2111d GH |
3222 | else |
3223 | return -EINVAL; | |
f67676d1 JA |
3224 | } |
3225 | ||
889fca73 | 3226 | static int io_read(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 JA |
3227 | { |
3228 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3229 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3230 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3231 | struct io_async_rw *rw = req->async_data; |
227c0c96 | 3232 | ssize_t io_size, ret, ret2; |
45d189c6 | 3233 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
ff6165b2 | 3234 | |
2846c481 | 3235 | if (rw) { |
e8c2bc1f | 3236 | iter = &rw->iter; |
2846c481 PB |
3237 | iovec = NULL; |
3238 | } else { | |
3239 | ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock); | |
3240 | if (ret < 0) | |
3241 | return ret; | |
3242 | } | |
632546c4 | 3243 | io_size = iov_iter_count(iter); |
fa15bafb | 3244 | req->result = io_size; |
2b188cc1 | 3245 | |
fd6c2e4c JA |
3246 | /* Ensure we clear previously set non-block flag */ |
3247 | if (!force_nonblock) | |
29de5f6a | 3248 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
a88fc400 PB |
3249 | else |
3250 | kiocb->ki_flags |= IOCB_NOWAIT; | |
3251 | ||
24c74678 | 3252 | /* If the file doesn't support async, just async punt */ |
b191e2df | 3253 | if (force_nonblock && !io_file_supports_nowait(req, READ)) { |
6713e7a6 | 3254 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true); |
6bf985dc | 3255 | return ret ?: -EAGAIN; |
6713e7a6 | 3256 | } |
9e645e11 | 3257 | |
632546c4 | 3258 | ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size); |
5ea5dd45 PB |
3259 | if (unlikely(ret)) { |
3260 | kfree(iovec); | |
3261 | return ret; | |
3262 | } | |
2b188cc1 | 3263 | |
227c0c96 | 3264 | ret = io_iter_do_read(req, iter); |
32960613 | 3265 | |
230d50d4 | 3266 | if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) { |
6ad7f233 | 3267 | req->flags &= ~REQ_F_REISSUE; |
eefdf30f JA |
3268 | /* IOPOLL retry should happen for io-wq threads */ |
3269 | if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
f91daf56 | 3270 | goto done; |
75c668cd PB |
3271 | /* no retry on NONBLOCK nor RWF_NOWAIT */ |
3272 | if (req->flags & REQ_F_NOWAIT) | |
355afaeb | 3273 | goto done; |
84216315 | 3274 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3275 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
f38c7e3a | 3276 | ret = 0; |
230d50d4 JA |
3277 | } else if (ret == -EIOCBQUEUED) { |
3278 | goto out_free; | |
7335e3bf | 3279 | } else if (ret <= 0 || ret == io_size || !force_nonblock || |
75c668cd | 3280 | (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) { |
7335e3bf | 3281 | /* read all, failed, already did sync or don't want to retry */ |
00d23d51 | 3282 | goto done; |
227c0c96 JA |
3283 | } |
3284 | ||
227c0c96 | 3285 | ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true); |
6bf985dc PB |
3286 | if (ret2) |
3287 | return ret2; | |
3288 | ||
fe1cdd55 | 3289 | iovec = NULL; |
e8c2bc1f | 3290 | rw = req->async_data; |
227c0c96 | 3291 | /* now use our persistent iterator, if we aren't already */ |
e8c2bc1f | 3292 | iter = &rw->iter; |
227c0c96 | 3293 | |
b23df91b PB |
3294 | do { |
3295 | io_size -= ret; | |
3296 | rw->bytes_done += ret; | |
3297 | /* if we can retry, do so with the callbacks armed */ | |
3298 | if (!io_rw_should_retry(req)) { | |
3299 | kiocb->ki_flags &= ~IOCB_WAITQ; | |
3300 | return -EAGAIN; | |
3301 | } | |
3302 | ||
3303 | /* | |
3304 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If | |
3305 | * we get -EIOCBQUEUED, then we'll get a notification when the | |
3306 | * desired page gets unlocked. We can also get a partial read | |
3307 | * here, and if we do, then just retry at the new offset. | |
3308 | */ | |
3309 | ret = io_iter_do_read(req, iter); | |
3310 | if (ret == -EIOCBQUEUED) | |
3311 | return 0; | |
227c0c96 | 3312 | /* we got some bytes, but not all. retry. */ |
b5b0ecb7 | 3313 | kiocb->ki_flags &= ~IOCB_WAITQ; |
b23df91b | 3314 | } while (ret > 0 && ret < io_size); |
227c0c96 | 3315 | done: |
889fca73 | 3316 | kiocb_done(kiocb, ret, issue_flags); |
fe1cdd55 PB |
3317 | out_free: |
3318 | /* it's faster to check here then delegate to kfree */ | |
3319 | if (iovec) | |
3320 | kfree(iovec); | |
5ea5dd45 | 3321 | return 0; |
2b188cc1 JA |
3322 | } |
3323 | ||
73debe68 | 3324 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 3325 | { |
3529d8c2 JA |
3326 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
3327 | return -EBADF; | |
93642ef8 | 3328 | return io_prep_rw(req, sqe); |
f67676d1 JA |
3329 | } |
3330 | ||
889fca73 | 3331 | static int io_write(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 JA |
3332 | { |
3333 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3334 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3335 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3336 | struct io_async_rw *rw = req->async_data; |
fa15bafb | 3337 | ssize_t ret, ret2, io_size; |
45d189c6 | 3338 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
2b188cc1 | 3339 | |
2846c481 | 3340 | if (rw) { |
e8c2bc1f | 3341 | iter = &rw->iter; |
2846c481 PB |
3342 | iovec = NULL; |
3343 | } else { | |
3344 | ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock); | |
3345 | if (ret < 0) | |
3346 | return ret; | |
3347 | } | |
632546c4 | 3348 | io_size = iov_iter_count(iter); |
fa15bafb | 3349 | req->result = io_size; |
2b188cc1 | 3350 | |
fd6c2e4c JA |
3351 | /* Ensure we clear previously set non-block flag */ |
3352 | if (!force_nonblock) | |
a88fc400 PB |
3353 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3354 | else | |
3355 | kiocb->ki_flags |= IOCB_NOWAIT; | |
fd6c2e4c | 3356 | |
24c74678 | 3357 | /* If the file doesn't support async, just async punt */ |
b191e2df | 3358 | if (force_nonblock && !io_file_supports_nowait(req, WRITE)) |
f67676d1 | 3359 | goto copy_iov; |
31b51510 | 3360 | |
10d59345 JA |
3361 | /* file path doesn't support NOWAIT for non-direct_IO */ |
3362 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
3363 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 3364 | goto copy_iov; |
31b51510 | 3365 | |
632546c4 | 3366 | ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size); |
fa15bafb PB |
3367 | if (unlikely(ret)) |
3368 | goto out_free; | |
4ed734b0 | 3369 | |
fa15bafb PB |
3370 | /* |
3371 | * Open-code file_start_write here to grab freeze protection, | |
3372 | * which will be released by another thread in | |
3373 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
3374 | * released so that it doesn't complain about the held lock when | |
3375 | * we return to userspace. | |
3376 | */ | |
3377 | if (req->flags & REQ_F_ISREG) { | |
8a3c84b6 | 3378 | sb_start_write(file_inode(req->file)->i_sb); |
fa15bafb PB |
3379 | __sb_writers_release(file_inode(req->file)->i_sb, |
3380 | SB_FREEZE_WRITE); | |
3381 | } | |
3382 | kiocb->ki_flags |= IOCB_WRITE; | |
4ed734b0 | 3383 | |
fa15bafb | 3384 | if (req->file->f_op->write_iter) |
ff6165b2 | 3385 | ret2 = call_write_iter(req->file, kiocb, iter); |
2dd2111d | 3386 | else if (req->file->f_op->write) |
4017eb91 | 3387 | ret2 = loop_rw_iter(WRITE, req, iter); |
2dd2111d GH |
3388 | else |
3389 | ret2 = -EINVAL; | |
4ed734b0 | 3390 | |
6ad7f233 PB |
3391 | if (req->flags & REQ_F_REISSUE) { |
3392 | req->flags &= ~REQ_F_REISSUE; | |
230d50d4 | 3393 | ret2 = -EAGAIN; |
6ad7f233 | 3394 | } |
230d50d4 | 3395 | |
fa15bafb PB |
3396 | /* |
3397 | * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just | |
3398 | * retry them without IOCB_NOWAIT. | |
3399 | */ | |
3400 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
3401 | ret2 = -EAGAIN; | |
75c668cd PB |
3402 | /* no retry on NONBLOCK nor RWF_NOWAIT */ |
3403 | if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT)) | |
355afaeb | 3404 | goto done; |
fa15bafb | 3405 | if (!force_nonblock || ret2 != -EAGAIN) { |
eefdf30f JA |
3406 | /* IOPOLL retry should happen for io-wq threads */ |
3407 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN) | |
3408 | goto copy_iov; | |
355afaeb | 3409 | done: |
889fca73 | 3410 | kiocb_done(kiocb, ret2, issue_flags); |
fa15bafb | 3411 | } else { |
f67676d1 | 3412 | copy_iov: |
84216315 | 3413 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3414 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
227c0c96 | 3415 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false); |
6bf985dc | 3416 | return ret ?: -EAGAIN; |
2b188cc1 | 3417 | } |
31b51510 | 3418 | out_free: |
f261c168 | 3419 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3420 | if (iovec) |
6f2cc166 | 3421 | kfree(iovec); |
2b188cc1 JA |
3422 | return ret; |
3423 | } | |
3424 | ||
80a261fd JA |
3425 | static int io_renameat_prep(struct io_kiocb *req, |
3426 | const struct io_uring_sqe *sqe) | |
3427 | { | |
3428 | struct io_rename *ren = &req->rename; | |
3429 | const char __user *oldf, *newf; | |
3430 | ||
ed7eb259 JA |
3431 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3432 | return -EINVAL; | |
3433 | if (sqe->ioprio || sqe->buf_index) | |
3434 | return -EINVAL; | |
80a261fd JA |
3435 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
3436 | return -EBADF; | |
3437 | ||
3438 | ren->old_dfd = READ_ONCE(sqe->fd); | |
3439 | oldf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3440 | newf = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
3441 | ren->new_dfd = READ_ONCE(sqe->len); | |
3442 | ren->flags = READ_ONCE(sqe->rename_flags); | |
3443 | ||
3444 | ren->oldpath = getname(oldf); | |
3445 | if (IS_ERR(ren->oldpath)) | |
3446 | return PTR_ERR(ren->oldpath); | |
3447 | ||
3448 | ren->newpath = getname(newf); | |
3449 | if (IS_ERR(ren->newpath)) { | |
3450 | putname(ren->oldpath); | |
3451 | return PTR_ERR(ren->newpath); | |
3452 | } | |
3453 | ||
3454 | req->flags |= REQ_F_NEED_CLEANUP; | |
3455 | return 0; | |
3456 | } | |
3457 | ||
45d189c6 | 3458 | static int io_renameat(struct io_kiocb *req, unsigned int issue_flags) |
80a261fd JA |
3459 | { |
3460 | struct io_rename *ren = &req->rename; | |
3461 | int ret; | |
3462 | ||
45d189c6 | 3463 | if (issue_flags & IO_URING_F_NONBLOCK) |
80a261fd JA |
3464 | return -EAGAIN; |
3465 | ||
3466 | ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd, | |
3467 | ren->newpath, ren->flags); | |
3468 | ||
3469 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3470 | if (ret < 0) | |
93d2bcd2 | 3471 | req_set_fail(req); |
80a261fd JA |
3472 | io_req_complete(req, ret); |
3473 | return 0; | |
3474 | } | |
3475 | ||
14a1143b JA |
3476 | static int io_unlinkat_prep(struct io_kiocb *req, |
3477 | const struct io_uring_sqe *sqe) | |
3478 | { | |
3479 | struct io_unlink *un = &req->unlink; | |
3480 | const char __user *fname; | |
3481 | ||
22634bc5 JA |
3482 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3483 | return -EINVAL; | |
3484 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
3485 | return -EINVAL; | |
14a1143b JA |
3486 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
3487 | return -EBADF; | |
3488 | ||
3489 | un->dfd = READ_ONCE(sqe->fd); | |
3490 | ||
3491 | un->flags = READ_ONCE(sqe->unlink_flags); | |
3492 | if (un->flags & ~AT_REMOVEDIR) | |
3493 | return -EINVAL; | |
3494 | ||
3495 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3496 | un->filename = getname(fname); | |
3497 | if (IS_ERR(un->filename)) | |
3498 | return PTR_ERR(un->filename); | |
3499 | ||
3500 | req->flags |= REQ_F_NEED_CLEANUP; | |
3501 | return 0; | |
3502 | } | |
3503 | ||
45d189c6 | 3504 | static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags) |
14a1143b JA |
3505 | { |
3506 | struct io_unlink *un = &req->unlink; | |
3507 | int ret; | |
3508 | ||
45d189c6 | 3509 | if (issue_flags & IO_URING_F_NONBLOCK) |
14a1143b JA |
3510 | return -EAGAIN; |
3511 | ||
3512 | if (un->flags & AT_REMOVEDIR) | |
3513 | ret = do_rmdir(un->dfd, un->filename); | |
3514 | else | |
3515 | ret = do_unlinkat(un->dfd, un->filename); | |
3516 | ||
3517 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3518 | if (ret < 0) | |
93d2bcd2 | 3519 | req_set_fail(req); |
14a1143b JA |
3520 | io_req_complete(req, ret); |
3521 | return 0; | |
3522 | } | |
3523 | ||
36f4fa68 JA |
3524 | static int io_shutdown_prep(struct io_kiocb *req, |
3525 | const struct io_uring_sqe *sqe) | |
3526 | { | |
3527 | #if defined(CONFIG_NET) | |
3528 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3529 | return -EINVAL; | |
3530 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags || | |
3531 | sqe->buf_index) | |
3532 | return -EINVAL; | |
3533 | ||
3534 | req->shutdown.how = READ_ONCE(sqe->len); | |
3535 | return 0; | |
3536 | #else | |
3537 | return -EOPNOTSUPP; | |
3538 | #endif | |
3539 | } | |
3540 | ||
45d189c6 | 3541 | static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags) |
36f4fa68 JA |
3542 | { |
3543 | #if defined(CONFIG_NET) | |
3544 | struct socket *sock; | |
3545 | int ret; | |
3546 | ||
45d189c6 | 3547 | if (issue_flags & IO_URING_F_NONBLOCK) |
36f4fa68 JA |
3548 | return -EAGAIN; |
3549 | ||
48aba79b | 3550 | sock = sock_from_file(req->file); |
36f4fa68 | 3551 | if (unlikely(!sock)) |
48aba79b | 3552 | return -ENOTSOCK; |
36f4fa68 JA |
3553 | |
3554 | ret = __sys_shutdown_sock(sock, req->shutdown.how); | |
a146468d | 3555 | if (ret < 0) |
93d2bcd2 | 3556 | req_set_fail(req); |
36f4fa68 JA |
3557 | io_req_complete(req, ret); |
3558 | return 0; | |
3559 | #else | |
3560 | return -EOPNOTSUPP; | |
3561 | #endif | |
3562 | } | |
3563 | ||
f2a8d5c7 PB |
3564 | static int __io_splice_prep(struct io_kiocb *req, |
3565 | const struct io_uring_sqe *sqe) | |
7d67af2c | 3566 | { |
fe7e3257 | 3567 | struct io_splice *sp = &req->splice; |
7d67af2c | 3568 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; |
7d67af2c | 3569 | |
3232dd02 PB |
3570 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3571 | return -EINVAL; | |
7d67af2c PB |
3572 | |
3573 | sp->file_in = NULL; | |
7d67af2c PB |
3574 | sp->len = READ_ONCE(sqe->len); |
3575 | sp->flags = READ_ONCE(sqe->splice_flags); | |
3576 | ||
3577 | if (unlikely(sp->flags & ~valid_flags)) | |
3578 | return -EINVAL; | |
3579 | ||
62906e89 | 3580 | sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in), |
8371adf5 PB |
3581 | (sp->flags & SPLICE_F_FD_IN_FIXED)); |
3582 | if (!sp->file_in) | |
3583 | return -EBADF; | |
7d67af2c | 3584 | req->flags |= REQ_F_NEED_CLEANUP; |
7d67af2c PB |
3585 | return 0; |
3586 | } | |
3587 | ||
f2a8d5c7 PB |
3588 | static int io_tee_prep(struct io_kiocb *req, |
3589 | const struct io_uring_sqe *sqe) | |
3590 | { | |
3591 | if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off)) | |
3592 | return -EINVAL; | |
3593 | return __io_splice_prep(req, sqe); | |
3594 | } | |
3595 | ||
45d189c6 | 3596 | static int io_tee(struct io_kiocb *req, unsigned int issue_flags) |
f2a8d5c7 PB |
3597 | { |
3598 | struct io_splice *sp = &req->splice; | |
3599 | struct file *in = sp->file_in; | |
3600 | struct file *out = sp->file_out; | |
3601 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3602 | long ret = 0; | |
3603 | ||
45d189c6 | 3604 | if (issue_flags & IO_URING_F_NONBLOCK) |
f2a8d5c7 PB |
3605 | return -EAGAIN; |
3606 | if (sp->len) | |
3607 | ret = do_tee(in, out, sp->len, flags); | |
3608 | ||
e1d767f0 PB |
3609 | if (!(sp->flags & SPLICE_F_FD_IN_FIXED)) |
3610 | io_put_file(in); | |
f2a8d5c7 PB |
3611 | req->flags &= ~REQ_F_NEED_CLEANUP; |
3612 | ||
f2a8d5c7 | 3613 | if (ret != sp->len) |
93d2bcd2 | 3614 | req_set_fail(req); |
e1e16097 | 3615 | io_req_complete(req, ret); |
f2a8d5c7 PB |
3616 | return 0; |
3617 | } | |
3618 | ||
3619 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
3620 | { | |
fe7e3257 | 3621 | struct io_splice *sp = &req->splice; |
f2a8d5c7 PB |
3622 | |
3623 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
3624 | sp->off_out = READ_ONCE(sqe->off); | |
3625 | return __io_splice_prep(req, sqe); | |
3626 | } | |
3627 | ||
45d189c6 | 3628 | static int io_splice(struct io_kiocb *req, unsigned int issue_flags) |
7d67af2c PB |
3629 | { |
3630 | struct io_splice *sp = &req->splice; | |
3631 | struct file *in = sp->file_in; | |
3632 | struct file *out = sp->file_out; | |
3633 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3634 | loff_t *poff_in, *poff_out; | |
c9687426 | 3635 | long ret = 0; |
7d67af2c | 3636 | |
45d189c6 | 3637 | if (issue_flags & IO_URING_F_NONBLOCK) |
2fb3e822 | 3638 | return -EAGAIN; |
7d67af2c PB |
3639 | |
3640 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
3641 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
c9687426 | 3642 | |
948a7749 | 3643 | if (sp->len) |
c9687426 | 3644 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); |
7d67af2c | 3645 | |
e1d767f0 PB |
3646 | if (!(sp->flags & SPLICE_F_FD_IN_FIXED)) |
3647 | io_put_file(in); | |
7d67af2c PB |
3648 | req->flags &= ~REQ_F_NEED_CLEANUP; |
3649 | ||
7d67af2c | 3650 | if (ret != sp->len) |
93d2bcd2 | 3651 | req_set_fail(req); |
e1e16097 | 3652 | io_req_complete(req, ret); |
7d67af2c PB |
3653 | return 0; |
3654 | } | |
3655 | ||
2b188cc1 JA |
3656 | /* |
3657 | * IORING_OP_NOP just posts a completion event, nothing else. | |
3658 | */ | |
889fca73 | 3659 | static int io_nop(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 JA |
3660 | { |
3661 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 3662 | |
def596e9 JA |
3663 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
3664 | return -EINVAL; | |
3665 | ||
889fca73 | 3666 | __io_req_complete(req, issue_flags, 0, 0); |
2b188cc1 JA |
3667 | return 0; |
3668 | } | |
3669 | ||
1155c76a | 3670 | static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 3671 | { |
6b06314c | 3672 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 3673 | |
09bb8394 JA |
3674 | if (!req->file) |
3675 | return -EBADF; | |
c992fe29 | 3676 | |
6b06314c | 3677 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 3678 | return -EINVAL; |
edafccee | 3679 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
3680 | return -EINVAL; |
3681 | ||
8ed8d3c3 JA |
3682 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
3683 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
3684 | return -EINVAL; | |
3685 | ||
3686 | req->sync.off = READ_ONCE(sqe->off); | |
3687 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
3688 | return 0; |
3689 | } | |
3690 | ||
45d189c6 | 3691 | static int io_fsync(struct io_kiocb *req, unsigned int issue_flags) |
8ed8d3c3 | 3692 | { |
8ed8d3c3 | 3693 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
3694 | int ret; |
3695 | ||
ac45abc0 | 3696 | /* fsync always requires a blocking context */ |
45d189c6 | 3697 | if (issue_flags & IO_URING_F_NONBLOCK) |
ac45abc0 PB |
3698 | return -EAGAIN; |
3699 | ||
9adbd45d | 3700 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
3701 | end > 0 ? end : LLONG_MAX, |
3702 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
3703 | if (ret < 0) | |
93d2bcd2 | 3704 | req_set_fail(req); |
e1e16097 | 3705 | io_req_complete(req, ret); |
c992fe29 CH |
3706 | return 0; |
3707 | } | |
3708 | ||
d63d1b5e JA |
3709 | static int io_fallocate_prep(struct io_kiocb *req, |
3710 | const struct io_uring_sqe *sqe) | |
3711 | { | |
3712 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
3713 | return -EINVAL; | |
3232dd02 PB |
3714 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3715 | return -EINVAL; | |
d63d1b5e JA |
3716 | |
3717 | req->sync.off = READ_ONCE(sqe->off); | |
3718 | req->sync.len = READ_ONCE(sqe->addr); | |
3719 | req->sync.mode = READ_ONCE(sqe->len); | |
3720 | return 0; | |
3721 | } | |
3722 | ||
45d189c6 | 3723 | static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags) |
5d17b4a4 | 3724 | { |
ac45abc0 PB |
3725 | int ret; |
3726 | ||
d63d1b5e | 3727 | /* fallocate always requiring blocking context */ |
45d189c6 | 3728 | if (issue_flags & IO_URING_F_NONBLOCK) |
5d17b4a4 | 3729 | return -EAGAIN; |
ac45abc0 PB |
3730 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
3731 | req->sync.len); | |
ac45abc0 | 3732 | if (ret < 0) |
93d2bcd2 | 3733 | req_set_fail(req); |
e1e16097 | 3734 | io_req_complete(req, ret); |
5d17b4a4 JA |
3735 | return 0; |
3736 | } | |
3737 | ||
ec65fea5 | 3738 | static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 3739 | { |
f8748881 | 3740 | const char __user *fname; |
15b71abe | 3741 | int ret; |
b7bb4f7d | 3742 | |
d3fddf6d PB |
3743 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3744 | return -EINVAL; | |
ec65fea5 | 3745 | if (unlikely(sqe->ioprio || sqe->buf_index)) |
15b71abe | 3746 | return -EINVAL; |
ec65fea5 | 3747 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
cf3040ca | 3748 | return -EBADF; |
03b1230c | 3749 | |
ec65fea5 PB |
3750 | /* open.how should be already initialised */ |
3751 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) | |
08a1d26e | 3752 | req->open.how.flags |= O_LARGEFILE; |
3529d8c2 | 3753 | |
25e72d10 PB |
3754 | req->open.dfd = READ_ONCE(sqe->fd); |
3755 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
f8748881 | 3756 | req->open.filename = getname(fname); |
15b71abe JA |
3757 | if (IS_ERR(req->open.filename)) { |
3758 | ret = PTR_ERR(req->open.filename); | |
3759 | req->open.filename = NULL; | |
3760 | return ret; | |
3761 | } | |
4022e7af | 3762 | req->open.nofile = rlimit(RLIMIT_NOFILE); |
8fef80bf | 3763 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 3764 | return 0; |
03b1230c JA |
3765 | } |
3766 | ||
ec65fea5 PB |
3767 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3768 | { | |
d3fddf6d PB |
3769 | u64 mode = READ_ONCE(sqe->len); |
3770 | u64 flags = READ_ONCE(sqe->open_flags); | |
ec65fea5 | 3771 | |
ec65fea5 PB |
3772 | req->open.how = build_open_how(flags, mode); |
3773 | return __io_openat_prep(req, sqe); | |
3774 | } | |
3775 | ||
cebdb986 | 3776 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 3777 | { |
cebdb986 | 3778 | struct open_how __user *how; |
cebdb986 | 3779 | size_t len; |
0fa03c62 JA |
3780 | int ret; |
3781 | ||
cebdb986 JA |
3782 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
3783 | len = READ_ONCE(sqe->len); | |
cebdb986 JA |
3784 | if (len < OPEN_HOW_SIZE_VER0) |
3785 | return -EINVAL; | |
3529d8c2 | 3786 | |
cebdb986 JA |
3787 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
3788 | len); | |
3789 | if (ret) | |
3790 | return ret; | |
3529d8c2 | 3791 | |
ec65fea5 | 3792 | return __io_openat_prep(req, sqe); |
cebdb986 JA |
3793 | } |
3794 | ||
45d189c6 | 3795 | static int io_openat2(struct io_kiocb *req, unsigned int issue_flags) |
15b71abe JA |
3796 | { |
3797 | struct open_flags op; | |
15b71abe | 3798 | struct file *file; |
3a81fd02 JA |
3799 | bool nonblock_set; |
3800 | bool resolve_nonblock; | |
15b71abe JA |
3801 | int ret; |
3802 | ||
cebdb986 | 3803 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
3804 | if (ret) |
3805 | goto err; | |
3a81fd02 JA |
3806 | nonblock_set = op.open_flag & O_NONBLOCK; |
3807 | resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED; | |
45d189c6 | 3808 | if (issue_flags & IO_URING_F_NONBLOCK) { |
3a81fd02 JA |
3809 | /* |
3810 | * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open, | |
3811 | * it'll always -EAGAIN | |
3812 | */ | |
3813 | if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE)) | |
3814 | return -EAGAIN; | |
3815 | op.lookup_flags |= LOOKUP_CACHED; | |
3816 | op.open_flag |= O_NONBLOCK; | |
3817 | } | |
15b71abe | 3818 | |
4022e7af | 3819 | ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile); |
15b71abe JA |
3820 | if (ret < 0) |
3821 | goto err; | |
3822 | ||
3823 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
12dcb58a | 3824 | if (IS_ERR(file)) { |
944d1444 | 3825 | /* |
12dcb58a PB |
3826 | * We could hang on to this 'fd' on retrying, but seems like |
3827 | * marginal gain for something that is now known to be a slower | |
3828 | * path. So just put it, and we'll get a new one when we retry. | |
944d1444 | 3829 | */ |
3a81fd02 | 3830 | put_unused_fd(ret); |
3a81fd02 | 3831 | |
15b71abe | 3832 | ret = PTR_ERR(file); |
12dcb58a PB |
3833 | /* only retry if RESOLVE_CACHED wasn't already set by application */ |
3834 | if (ret == -EAGAIN && | |
3835 | (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK))) | |
3836 | return -EAGAIN; | |
3837 | goto err; | |
15b71abe | 3838 | } |
12dcb58a PB |
3839 | |
3840 | if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set) | |
3841 | file->f_flags &= ~O_NONBLOCK; | |
3842 | fsnotify_open(file); | |
3843 | fd_install(ret, file); | |
15b71abe JA |
3844 | err: |
3845 | putname(req->open.filename); | |
8fef80bf | 3846 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe | 3847 | if (ret < 0) |
93d2bcd2 | 3848 | req_set_fail(req); |
0bdf3398 | 3849 | __io_req_complete(req, issue_flags, ret, 0); |
15b71abe JA |
3850 | return 0; |
3851 | } | |
3852 | ||
45d189c6 | 3853 | static int io_openat(struct io_kiocb *req, unsigned int issue_flags) |
cebdb986 | 3854 | { |
e45cff58 | 3855 | return io_openat2(req, issue_flags); |
cebdb986 JA |
3856 | } |
3857 | ||
067524e9 JA |
3858 | static int io_remove_buffers_prep(struct io_kiocb *req, |
3859 | const struct io_uring_sqe *sqe) | |
3860 | { | |
3861 | struct io_provide_buf *p = &req->pbuf; | |
3862 | u64 tmp; | |
3863 | ||
3864 | if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off) | |
3865 | return -EINVAL; | |
3866 | ||
3867 | tmp = READ_ONCE(sqe->fd); | |
3868 | if (!tmp || tmp > USHRT_MAX) | |
3869 | return -EINVAL; | |
3870 | ||
3871 | memset(p, 0, sizeof(*p)); | |
3872 | p->nbufs = tmp; | |
3873 | p->bgid = READ_ONCE(sqe->buf_group); | |
3874 | return 0; | |
3875 | } | |
3876 | ||
3877 | static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf, | |
3878 | int bgid, unsigned nbufs) | |
3879 | { | |
3880 | unsigned i = 0; | |
3881 | ||
3882 | /* shouldn't happen */ | |
3883 | if (!nbufs) | |
3884 | return 0; | |
3885 | ||
3886 | /* the head kbuf is the list itself */ | |
3887 | while (!list_empty(&buf->list)) { | |
3888 | struct io_buffer *nxt; | |
3889 | ||
3890 | nxt = list_first_entry(&buf->list, struct io_buffer, list); | |
3891 | list_del(&nxt->list); | |
3892 | kfree(nxt); | |
3893 | if (++i == nbufs) | |
3894 | return i; | |
3895 | } | |
3896 | i++; | |
3897 | kfree(buf); | |
9e15c3a0 | 3898 | xa_erase(&ctx->io_buffers, bgid); |
067524e9 JA |
3899 | |
3900 | return i; | |
3901 | } | |
3902 | ||
889fca73 | 3903 | static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags) |
067524e9 JA |
3904 | { |
3905 | struct io_provide_buf *p = &req->pbuf; | |
3906 | struct io_ring_ctx *ctx = req->ctx; | |
3907 | struct io_buffer *head; | |
3908 | int ret = 0; | |
45d189c6 | 3909 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
067524e9 JA |
3910 | |
3911 | io_ring_submit_lock(ctx, !force_nonblock); | |
3912 | ||
3913 | lockdep_assert_held(&ctx->uring_lock); | |
3914 | ||
3915 | ret = -ENOENT; | |
9e15c3a0 | 3916 | head = xa_load(&ctx->io_buffers, p->bgid); |
067524e9 JA |
3917 | if (head) |
3918 | ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs); | |
067524e9 | 3919 | if (ret < 0) |
93d2bcd2 | 3920 | req_set_fail(req); |
067524e9 | 3921 | |
9fb8cb49 PB |
3922 | /* complete before unlock, IOPOLL may need the lock */ |
3923 | __io_req_complete(req, issue_flags, ret, 0); | |
3924 | io_ring_submit_unlock(ctx, !force_nonblock); | |
067524e9 JA |
3925 | return 0; |
3926 | } | |
3927 | ||
ddf0322d JA |
3928 | static int io_provide_buffers_prep(struct io_kiocb *req, |
3929 | const struct io_uring_sqe *sqe) | |
3930 | { | |
38134ada | 3931 | unsigned long size, tmp_check; |
ddf0322d JA |
3932 | struct io_provide_buf *p = &req->pbuf; |
3933 | u64 tmp; | |
3934 | ||
3935 | if (sqe->ioprio || sqe->rw_flags) | |
3936 | return -EINVAL; | |
3937 | ||
3938 | tmp = READ_ONCE(sqe->fd); | |
3939 | if (!tmp || tmp > USHRT_MAX) | |
3940 | return -E2BIG; | |
3941 | p->nbufs = tmp; | |
3942 | p->addr = READ_ONCE(sqe->addr); | |
3943 | p->len = READ_ONCE(sqe->len); | |
3944 | ||
38134ada PB |
3945 | if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs, |
3946 | &size)) | |
3947 | return -EOVERFLOW; | |
3948 | if (check_add_overflow((unsigned long)p->addr, size, &tmp_check)) | |
3949 | return -EOVERFLOW; | |
3950 | ||
d81269fe PB |
3951 | size = (unsigned long)p->len * p->nbufs; |
3952 | if (!access_ok(u64_to_user_ptr(p->addr), size)) | |
ddf0322d JA |
3953 | return -EFAULT; |
3954 | ||
3955 | p->bgid = READ_ONCE(sqe->buf_group); | |
3956 | tmp = READ_ONCE(sqe->off); | |
3957 | if (tmp > USHRT_MAX) | |
3958 | return -E2BIG; | |
3959 | p->bid = tmp; | |
3960 | return 0; | |
3961 | } | |
3962 | ||
3963 | static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head) | |
3964 | { | |
3965 | struct io_buffer *buf; | |
3966 | u64 addr = pbuf->addr; | |
3967 | int i, bid = pbuf->bid; | |
3968 | ||
3969 | for (i = 0; i < pbuf->nbufs; i++) { | |
3970 | buf = kmalloc(sizeof(*buf), GFP_KERNEL); | |
3971 | if (!buf) | |
3972 | break; | |
3973 | ||
3974 | buf->addr = addr; | |
d1f82808 | 3975 | buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT); |
ddf0322d JA |
3976 | buf->bid = bid; |
3977 | addr += pbuf->len; | |
3978 | bid++; | |
3979 | if (!*head) { | |
3980 | INIT_LIST_HEAD(&buf->list); | |
3981 | *head = buf; | |
3982 | } else { | |
3983 | list_add_tail(&buf->list, &(*head)->list); | |
3984 | } | |
3985 | } | |
3986 | ||
3987 | return i ? i : -ENOMEM; | |
3988 | } | |
3989 | ||
889fca73 | 3990 | static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags) |
ddf0322d JA |
3991 | { |
3992 | struct io_provide_buf *p = &req->pbuf; | |
3993 | struct io_ring_ctx *ctx = req->ctx; | |
3994 | struct io_buffer *head, *list; | |
3995 | int ret = 0; | |
45d189c6 | 3996 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
ddf0322d JA |
3997 | |
3998 | io_ring_submit_lock(ctx, !force_nonblock); | |
3999 | ||
4000 | lockdep_assert_held(&ctx->uring_lock); | |
4001 | ||
9e15c3a0 | 4002 | list = head = xa_load(&ctx->io_buffers, p->bgid); |
ddf0322d JA |
4003 | |
4004 | ret = io_add_buffers(p, &head); | |
9e15c3a0 JA |
4005 | if (ret >= 0 && !list) { |
4006 | ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL); | |
4007 | if (ret < 0) | |
067524e9 | 4008 | __io_remove_buffers(ctx, head, p->bgid, -1U); |
ddf0322d | 4009 | } |
ddf0322d | 4010 | if (ret < 0) |
93d2bcd2 | 4011 | req_set_fail(req); |
9fb8cb49 PB |
4012 | /* complete before unlock, IOPOLL may need the lock */ |
4013 | __io_req_complete(req, issue_flags, ret, 0); | |
4014 | io_ring_submit_unlock(ctx, !force_nonblock); | |
ddf0322d | 4015 | return 0; |
cebdb986 JA |
4016 | } |
4017 | ||
3e4827b0 JA |
4018 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
4019 | const struct io_uring_sqe *sqe) | |
4020 | { | |
4021 | #if defined(CONFIG_EPOLL) | |
4022 | if (sqe->ioprio || sqe->buf_index) | |
4023 | return -EINVAL; | |
2d74d042 | 4024 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4025 | return -EINVAL; |
3e4827b0 JA |
4026 | |
4027 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
4028 | req->epoll.op = READ_ONCE(sqe->len); | |
4029 | req->epoll.fd = READ_ONCE(sqe->off); | |
4030 | ||
4031 | if (ep_op_has_event(req->epoll.op)) { | |
4032 | struct epoll_event __user *ev; | |
4033 | ||
4034 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
4035 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
4036 | return -EFAULT; | |
4037 | } | |
4038 | ||
4039 | return 0; | |
4040 | #else | |
4041 | return -EOPNOTSUPP; | |
4042 | #endif | |
4043 | } | |
4044 | ||
889fca73 | 4045 | static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags) |
3e4827b0 JA |
4046 | { |
4047 | #if defined(CONFIG_EPOLL) | |
4048 | struct io_epoll *ie = &req->epoll; | |
4049 | int ret; | |
45d189c6 | 4050 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
3e4827b0 JA |
4051 | |
4052 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
4053 | if (force_nonblock && ret == -EAGAIN) | |
4054 | return -EAGAIN; | |
4055 | ||
4056 | if (ret < 0) | |
93d2bcd2 | 4057 | req_set_fail(req); |
889fca73 | 4058 | __io_req_complete(req, issue_flags, ret, 0); |
3e4827b0 JA |
4059 | return 0; |
4060 | #else | |
4061 | return -EOPNOTSUPP; | |
4062 | #endif | |
4063 | } | |
4064 | ||
c1ca757b JA |
4065 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4066 | { | |
4067 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4068 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
4069 | return -EINVAL; | |
3232dd02 PB |
4070 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4071 | return -EINVAL; | |
c1ca757b JA |
4072 | |
4073 | req->madvise.addr = READ_ONCE(sqe->addr); | |
4074 | req->madvise.len = READ_ONCE(sqe->len); | |
4075 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4076 | return 0; | |
4077 | #else | |
4078 | return -EOPNOTSUPP; | |
4079 | #endif | |
4080 | } | |
4081 | ||
45d189c6 | 4082 | static int io_madvise(struct io_kiocb *req, unsigned int issue_flags) |
c1ca757b JA |
4083 | { |
4084 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4085 | struct io_madvise *ma = &req->madvise; | |
4086 | int ret; | |
4087 | ||
45d189c6 | 4088 | if (issue_flags & IO_URING_F_NONBLOCK) |
c1ca757b JA |
4089 | return -EAGAIN; |
4090 | ||
0726b01e | 4091 | ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice); |
c1ca757b | 4092 | if (ret < 0) |
93d2bcd2 | 4093 | req_set_fail(req); |
e1e16097 | 4094 | io_req_complete(req, ret); |
c1ca757b JA |
4095 | return 0; |
4096 | #else | |
4097 | return -EOPNOTSUPP; | |
4098 | #endif | |
4099 | } | |
4100 | ||
4840e418 JA |
4101 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4102 | { | |
4103 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
4104 | return -EINVAL; | |
3232dd02 PB |
4105 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4106 | return -EINVAL; | |
4840e418 JA |
4107 | |
4108 | req->fadvise.offset = READ_ONCE(sqe->off); | |
4109 | req->fadvise.len = READ_ONCE(sqe->len); | |
4110 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4111 | return 0; | |
4112 | } | |
4113 | ||
45d189c6 | 4114 | static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags) |
4840e418 JA |
4115 | { |
4116 | struct io_fadvise *fa = &req->fadvise; | |
4117 | int ret; | |
4118 | ||
45d189c6 | 4119 | if (issue_flags & IO_URING_F_NONBLOCK) { |
3e69426d JA |
4120 | switch (fa->advice) { |
4121 | case POSIX_FADV_NORMAL: | |
4122 | case POSIX_FADV_RANDOM: | |
4123 | case POSIX_FADV_SEQUENTIAL: | |
4124 | break; | |
4125 | default: | |
4126 | return -EAGAIN; | |
4127 | } | |
4128 | } | |
4840e418 JA |
4129 | |
4130 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
4131 | if (ret < 0) | |
93d2bcd2 | 4132 | req_set_fail(req); |
0bdf3398 | 4133 | __io_req_complete(req, issue_flags, ret, 0); |
4840e418 JA |
4134 | return 0; |
4135 | } | |
4136 | ||
eddc7ef5 JA |
4137 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4138 | { | |
2d74d042 | 4139 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4140 | return -EINVAL; |
eddc7ef5 JA |
4141 | if (sqe->ioprio || sqe->buf_index) |
4142 | return -EINVAL; | |
9c280f90 | 4143 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4144 | return -EBADF; |
eddc7ef5 | 4145 | |
1d9e1288 BM |
4146 | req->statx.dfd = READ_ONCE(sqe->fd); |
4147 | req->statx.mask = READ_ONCE(sqe->len); | |
e62753e4 | 4148 | req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1d9e1288 BM |
4149 | req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
4150 | req->statx.flags = READ_ONCE(sqe->statx_flags); | |
eddc7ef5 JA |
4151 | |
4152 | return 0; | |
4153 | } | |
4154 | ||
45d189c6 | 4155 | static int io_statx(struct io_kiocb *req, unsigned int issue_flags) |
eddc7ef5 | 4156 | { |
1d9e1288 | 4157 | struct io_statx *ctx = &req->statx; |
eddc7ef5 JA |
4158 | int ret; |
4159 | ||
59d70013 | 4160 | if (issue_flags & IO_URING_F_NONBLOCK) |
eddc7ef5 JA |
4161 | return -EAGAIN; |
4162 | ||
e62753e4 BM |
4163 | ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask, |
4164 | ctx->buffer); | |
eddc7ef5 | 4165 | |
eddc7ef5 | 4166 | if (ret < 0) |
93d2bcd2 | 4167 | req_set_fail(req); |
e1e16097 | 4168 | io_req_complete(req, ret); |
eddc7ef5 JA |
4169 | return 0; |
4170 | } | |
4171 | ||
b5dba59e JA |
4172 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4173 | { | |
14587a46 | 4174 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4175 | return -EINVAL; |
b5dba59e JA |
4176 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || |
4177 | sqe->rw_flags || sqe->buf_index) | |
4178 | return -EINVAL; | |
9c280f90 | 4179 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4180 | return -EBADF; |
b5dba59e JA |
4181 | |
4182 | req->close.fd = READ_ONCE(sqe->fd); | |
b5dba59e | 4183 | return 0; |
b5dba59e JA |
4184 | } |
4185 | ||
889fca73 | 4186 | static int io_close(struct io_kiocb *req, unsigned int issue_flags) |
b5dba59e | 4187 | { |
9eac1904 | 4188 | struct files_struct *files = current->files; |
3af73b28 | 4189 | struct io_close *close = &req->close; |
9eac1904 | 4190 | struct fdtable *fdt; |
a1fde923 PB |
4191 | struct file *file = NULL; |
4192 | int ret = -EBADF; | |
b5dba59e | 4193 | |
9eac1904 JA |
4194 | spin_lock(&files->file_lock); |
4195 | fdt = files_fdtable(files); | |
4196 | if (close->fd >= fdt->max_fds) { | |
4197 | spin_unlock(&files->file_lock); | |
4198 | goto err; | |
4199 | } | |
4200 | file = fdt->fd[close->fd]; | |
a1fde923 | 4201 | if (!file || file->f_op == &io_uring_fops) { |
9eac1904 JA |
4202 | spin_unlock(&files->file_lock); |
4203 | file = NULL; | |
4204 | goto err; | |
3af73b28 | 4205 | } |
b5dba59e JA |
4206 | |
4207 | /* if the file has a flush method, be safe and punt to async */ | |
45d189c6 | 4208 | if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) { |
9eac1904 | 4209 | spin_unlock(&files->file_lock); |
0bf0eefd | 4210 | return -EAGAIN; |
a2100672 | 4211 | } |
b5dba59e | 4212 | |
9eac1904 JA |
4213 | ret = __close_fd_get_file(close->fd, &file); |
4214 | spin_unlock(&files->file_lock); | |
4215 | if (ret < 0) { | |
4216 | if (ret == -ENOENT) | |
4217 | ret = -EBADF; | |
4218 | goto err; | |
4219 | } | |
4220 | ||
3af73b28 | 4221 | /* No ->flush() or already async, safely close from here */ |
9eac1904 JA |
4222 | ret = filp_close(file, current->files); |
4223 | err: | |
3af73b28 | 4224 | if (ret < 0) |
93d2bcd2 | 4225 | req_set_fail(req); |
9eac1904 JA |
4226 | if (file) |
4227 | fput(file); | |
889fca73 | 4228 | __io_req_complete(req, issue_flags, ret, 0); |
1a417f4e | 4229 | return 0; |
b5dba59e JA |
4230 | } |
4231 | ||
1155c76a | 4232 | static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
4233 | { |
4234 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 | 4235 | |
5d17b4a4 JA |
4236 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
4237 | return -EINVAL; | |
4238 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
4239 | return -EINVAL; | |
4240 | ||
8ed8d3c3 JA |
4241 | req->sync.off = READ_ONCE(sqe->off); |
4242 | req->sync.len = READ_ONCE(sqe->len); | |
4243 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
4244 | return 0; |
4245 | } | |
4246 | ||
45d189c6 | 4247 | static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags) |
8ed8d3c3 | 4248 | { |
8ed8d3c3 JA |
4249 | int ret; |
4250 | ||
ac45abc0 | 4251 | /* sync_file_range always requires a blocking context */ |
45d189c6 | 4252 | if (issue_flags & IO_URING_F_NONBLOCK) |
ac45abc0 PB |
4253 | return -EAGAIN; |
4254 | ||
9adbd45d | 4255 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
4256 | req->sync.flags); |
4257 | if (ret < 0) | |
93d2bcd2 | 4258 | req_set_fail(req); |
e1e16097 | 4259 | io_req_complete(req, ret); |
5d17b4a4 JA |
4260 | return 0; |
4261 | } | |
4262 | ||
469956e8 | 4263 | #if defined(CONFIG_NET) |
02d27d89 PB |
4264 | static int io_setup_async_msg(struct io_kiocb *req, |
4265 | struct io_async_msghdr *kmsg) | |
4266 | { | |
e8c2bc1f JA |
4267 | struct io_async_msghdr *async_msg = req->async_data; |
4268 | ||
4269 | if (async_msg) | |
02d27d89 | 4270 | return -EAGAIN; |
e8c2bc1f | 4271 | if (io_alloc_async_data(req)) { |
257e84a5 | 4272 | kfree(kmsg->free_iov); |
02d27d89 PB |
4273 | return -ENOMEM; |
4274 | } | |
e8c2bc1f | 4275 | async_msg = req->async_data; |
02d27d89 | 4276 | req->flags |= REQ_F_NEED_CLEANUP; |
e8c2bc1f | 4277 | memcpy(async_msg, kmsg, sizeof(*kmsg)); |
2a780802 | 4278 | async_msg->msg.msg_name = &async_msg->addr; |
257e84a5 PB |
4279 | /* if were using fast_iov, set it to the new one */ |
4280 | if (!async_msg->free_iov) | |
4281 | async_msg->msg.msg_iter.iov = async_msg->fast_iov; | |
4282 | ||
02d27d89 PB |
4283 | return -EAGAIN; |
4284 | } | |
4285 | ||
2ae523ed PB |
4286 | static int io_sendmsg_copy_hdr(struct io_kiocb *req, |
4287 | struct io_async_msghdr *iomsg) | |
4288 | { | |
2ae523ed | 4289 | iomsg->msg.msg_name = &iomsg->addr; |
257e84a5 | 4290 | iomsg->free_iov = iomsg->fast_iov; |
2ae523ed | 4291 | return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg, |
257e84a5 | 4292 | req->sr_msg.msg_flags, &iomsg->free_iov); |
2ae523ed PB |
4293 | } |
4294 | ||
93642ef8 PB |
4295 | static int io_sendmsg_prep_async(struct io_kiocb *req) |
4296 | { | |
4297 | int ret; | |
4298 | ||
93642ef8 PB |
4299 | ret = io_sendmsg_copy_hdr(req, req->async_data); |
4300 | if (!ret) | |
4301 | req->flags |= REQ_F_NEED_CLEANUP; | |
4302 | return ret; | |
4303 | } | |
4304 | ||
3529d8c2 | 4305 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 4306 | { |
e47293fd | 4307 | struct io_sr_msg *sr = &req->sr_msg; |
03b1230c | 4308 | |
d2b6f48b PB |
4309 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4310 | return -EINVAL; | |
4311 | ||
270a5940 | 4312 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
fddaface | 4313 | sr->len = READ_ONCE(sqe->len); |
04411806 PB |
4314 | sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL; |
4315 | if (sr->msg_flags & MSG_DONTWAIT) | |
4316 | req->flags |= REQ_F_NOWAIT; | |
3529d8c2 | 4317 | |
d8768362 JA |
4318 | #ifdef CONFIG_COMPAT |
4319 | if (req->ctx->compat) | |
4320 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4321 | #endif | |
93642ef8 | 4322 | return 0; |
03b1230c JA |
4323 | } |
4324 | ||
889fca73 | 4325 | static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags) |
aa1fa28f | 4326 | { |
6b754c8b | 4327 | struct io_async_msghdr iomsg, *kmsg; |
0fa03c62 | 4328 | struct socket *sock; |
7a7cacba | 4329 | unsigned flags; |
0031275d | 4330 | int min_ret = 0; |
0fa03c62 JA |
4331 | int ret; |
4332 | ||
dba4a925 | 4333 | sock = sock_from_file(req->file); |
7a7cacba | 4334 | if (unlikely(!sock)) |
dba4a925 | 4335 | return -ENOTSOCK; |
3529d8c2 | 4336 | |
257e84a5 PB |
4337 | kmsg = req->async_data; |
4338 | if (!kmsg) { | |
7a7cacba PB |
4339 | ret = io_sendmsg_copy_hdr(req, &iomsg); |
4340 | if (ret) | |
4341 | return ret; | |
4342 | kmsg = &iomsg; | |
0fa03c62 | 4343 | } |
0fa03c62 | 4344 | |
04411806 PB |
4345 | flags = req->sr_msg.msg_flags; |
4346 | if (issue_flags & IO_URING_F_NONBLOCK) | |
7a7cacba | 4347 | flags |= MSG_DONTWAIT; |
0031275d SM |
4348 | if (flags & MSG_WAITALL) |
4349 | min_ret = iov_iter_count(&kmsg->msg.msg_iter); | |
4350 | ||
7a7cacba | 4351 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
45d189c6 | 4352 | if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN) |
7a7cacba PB |
4353 | return io_setup_async_msg(req, kmsg); |
4354 | if (ret == -ERESTARTSYS) | |
4355 | ret = -EINTR; | |
0fa03c62 | 4356 | |
257e84a5 PB |
4357 | /* fast path, check for non-NULL to avoid function call */ |
4358 | if (kmsg->free_iov) | |
4359 | kfree(kmsg->free_iov); | |
99bc4c38 | 4360 | req->flags &= ~REQ_F_NEED_CLEANUP; |
0031275d | 4361 | if (ret < min_ret) |
93d2bcd2 | 4362 | req_set_fail(req); |
889fca73 | 4363 | __io_req_complete(req, issue_flags, ret, 0); |
5d17b4a4 | 4364 | return 0; |
03b1230c | 4365 | } |
aa1fa28f | 4366 | |
889fca73 | 4367 | static int io_send(struct io_kiocb *req, unsigned int issue_flags) |
fddaface | 4368 | { |
7a7cacba PB |
4369 | struct io_sr_msg *sr = &req->sr_msg; |
4370 | struct msghdr msg; | |
4371 | struct iovec iov; | |
fddaface | 4372 | struct socket *sock; |
7a7cacba | 4373 | unsigned flags; |
0031275d | 4374 | int min_ret = 0; |
fddaface JA |
4375 | int ret; |
4376 | ||
dba4a925 | 4377 | sock = sock_from_file(req->file); |
7a7cacba | 4378 | if (unlikely(!sock)) |
dba4a925 | 4379 | return -ENOTSOCK; |
fddaface | 4380 | |
7a7cacba PB |
4381 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter); |
4382 | if (unlikely(ret)) | |
14db8411 | 4383 | return ret; |
fddaface | 4384 | |
7a7cacba PB |
4385 | msg.msg_name = NULL; |
4386 | msg.msg_control = NULL; | |
4387 | msg.msg_controllen = 0; | |
4388 | msg.msg_namelen = 0; | |
fddaface | 4389 | |
04411806 PB |
4390 | flags = req->sr_msg.msg_flags; |
4391 | if (issue_flags & IO_URING_F_NONBLOCK) | |
7a7cacba | 4392 | flags |= MSG_DONTWAIT; |
0031275d SM |
4393 | if (flags & MSG_WAITALL) |
4394 | min_ret = iov_iter_count(&msg.msg_iter); | |
4395 | ||
7a7cacba PB |
4396 | msg.msg_flags = flags; |
4397 | ret = sock_sendmsg(sock, &msg); | |
45d189c6 | 4398 | if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN) |
7a7cacba PB |
4399 | return -EAGAIN; |
4400 | if (ret == -ERESTARTSYS) | |
4401 | ret = -EINTR; | |
fddaface | 4402 | |
0031275d | 4403 | if (ret < min_ret) |
93d2bcd2 | 4404 | req_set_fail(req); |
889fca73 | 4405 | __io_req_complete(req, issue_flags, ret, 0); |
fddaface | 4406 | return 0; |
fddaface JA |
4407 | } |
4408 | ||
1400e697 PB |
4409 | static int __io_recvmsg_copy_hdr(struct io_kiocb *req, |
4410 | struct io_async_msghdr *iomsg) | |
52de1fe1 JA |
4411 | { |
4412 | struct io_sr_msg *sr = &req->sr_msg; | |
4413 | struct iovec __user *uiov; | |
4414 | size_t iov_len; | |
4415 | int ret; | |
4416 | ||
1400e697 PB |
4417 | ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg, |
4418 | &iomsg->uaddr, &uiov, &iov_len); | |
52de1fe1 JA |
4419 | if (ret) |
4420 | return ret; | |
4421 | ||
4422 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4423 | if (iov_len > 1) | |
4424 | return -EINVAL; | |
5476dfed | 4425 | if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov))) |
52de1fe1 | 4426 | return -EFAULT; |
5476dfed | 4427 | sr->len = iomsg->fast_iov[0].iov_len; |
257e84a5 | 4428 | iomsg->free_iov = NULL; |
52de1fe1 | 4429 | } else { |
257e84a5 | 4430 | iomsg->free_iov = iomsg->fast_iov; |
89cd35c5 | 4431 | ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV, |
257e84a5 | 4432 | &iomsg->free_iov, &iomsg->msg.msg_iter, |
89cd35c5 | 4433 | false); |
52de1fe1 JA |
4434 | if (ret > 0) |
4435 | ret = 0; | |
4436 | } | |
4437 | ||
4438 | return ret; | |
4439 | } | |
4440 | ||
4441 | #ifdef CONFIG_COMPAT | |
4442 | static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, | |
1400e697 | 4443 | struct io_async_msghdr *iomsg) |
52de1fe1 | 4444 | { |
52de1fe1 JA |
4445 | struct io_sr_msg *sr = &req->sr_msg; |
4446 | struct compat_iovec __user *uiov; | |
4447 | compat_uptr_t ptr; | |
4448 | compat_size_t len; | |
4449 | int ret; | |
4450 | ||
4af3417a PB |
4451 | ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr, |
4452 | &ptr, &len); | |
52de1fe1 JA |
4453 | if (ret) |
4454 | return ret; | |
4455 | ||
4456 | uiov = compat_ptr(ptr); | |
4457 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4458 | compat_ssize_t clen; | |
4459 | ||
4460 | if (len > 1) | |
4461 | return -EINVAL; | |
4462 | if (!access_ok(uiov, sizeof(*uiov))) | |
4463 | return -EFAULT; | |
4464 | if (__get_user(clen, &uiov->iov_len)) | |
4465 | return -EFAULT; | |
4466 | if (clen < 0) | |
4467 | return -EINVAL; | |
2d280bc8 | 4468 | sr->len = clen; |
257e84a5 | 4469 | iomsg->free_iov = NULL; |
52de1fe1 | 4470 | } else { |
257e84a5 | 4471 | iomsg->free_iov = iomsg->fast_iov; |
89cd35c5 | 4472 | ret = __import_iovec(READ, (struct iovec __user *)uiov, len, |
257e84a5 | 4473 | UIO_FASTIOV, &iomsg->free_iov, |
89cd35c5 | 4474 | &iomsg->msg.msg_iter, true); |
52de1fe1 JA |
4475 | if (ret < 0) |
4476 | return ret; | |
4477 | } | |
4478 | ||
4479 | return 0; | |
4480 | } | |
4481 | #endif | |
4482 | ||
1400e697 PB |
4483 | static int io_recvmsg_copy_hdr(struct io_kiocb *req, |
4484 | struct io_async_msghdr *iomsg) | |
52de1fe1 | 4485 | { |
1400e697 | 4486 | iomsg->msg.msg_name = &iomsg->addr; |
52de1fe1 JA |
4487 | |
4488 | #ifdef CONFIG_COMPAT | |
4489 | if (req->ctx->compat) | |
1400e697 | 4490 | return __io_compat_recvmsg_copy_hdr(req, iomsg); |
fddaface | 4491 | #endif |
52de1fe1 | 4492 | |
1400e697 | 4493 | return __io_recvmsg_copy_hdr(req, iomsg); |
52de1fe1 JA |
4494 | } |
4495 | ||
bcda7baa | 4496 | static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req, |
7fbb1b54 | 4497 | bool needs_lock) |
bcda7baa JA |
4498 | { |
4499 | struct io_sr_msg *sr = &req->sr_msg; | |
4500 | struct io_buffer *kbuf; | |
4501 | ||
bcda7baa JA |
4502 | kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock); |
4503 | if (IS_ERR(kbuf)) | |
4504 | return kbuf; | |
4505 | ||
4506 | sr->kbuf = kbuf; | |
4507 | req->flags |= REQ_F_BUFFER_SELECTED; | |
bcda7baa | 4508 | return kbuf; |
fddaface JA |
4509 | } |
4510 | ||
7fbb1b54 PB |
4511 | static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req) |
4512 | { | |
4513 | return io_put_kbuf(req, req->sr_msg.kbuf); | |
4514 | } | |
4515 | ||
93642ef8 | 4516 | static int io_recvmsg_prep_async(struct io_kiocb *req) |
aa1fa28f | 4517 | { |
99bc4c38 | 4518 | int ret; |
3529d8c2 | 4519 | |
93642ef8 PB |
4520 | ret = io_recvmsg_copy_hdr(req, req->async_data); |
4521 | if (!ret) | |
4522 | req->flags |= REQ_F_NEED_CLEANUP; | |
4523 | return ret; | |
4524 | } | |
4525 | ||
4526 | static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
4527 | { | |
4528 | struct io_sr_msg *sr = &req->sr_msg; | |
4529 | ||
d2b6f48b PB |
4530 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4531 | return -EINVAL; | |
4532 | ||
270a5940 | 4533 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
0b7b21e4 | 4534 | sr->len = READ_ONCE(sqe->len); |
bcda7baa | 4535 | sr->bgid = READ_ONCE(sqe->buf_group); |
04411806 PB |
4536 | sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL; |
4537 | if (sr->msg_flags & MSG_DONTWAIT) | |
4538 | req->flags |= REQ_F_NOWAIT; | |
06b76d44 | 4539 | |
d8768362 JA |
4540 | #ifdef CONFIG_COMPAT |
4541 | if (req->ctx->compat) | |
4542 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4543 | #endif | |
93642ef8 | 4544 | return 0; |
aa1fa28f JA |
4545 | } |
4546 | ||
889fca73 | 4547 | static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags) |
aa1fa28f | 4548 | { |
6b754c8b | 4549 | struct io_async_msghdr iomsg, *kmsg; |
03b1230c | 4550 | struct socket *sock; |
7fbb1b54 | 4551 | struct io_buffer *kbuf; |
7a7cacba | 4552 | unsigned flags; |
0031275d | 4553 | int min_ret = 0; |
52de1fe1 | 4554 | int ret, cflags = 0; |
45d189c6 | 4555 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
03b1230c | 4556 | |
dba4a925 | 4557 | sock = sock_from_file(req->file); |
7a7cacba | 4558 | if (unlikely(!sock)) |
dba4a925 | 4559 | return -ENOTSOCK; |
3529d8c2 | 4560 | |
257e84a5 PB |
4561 | kmsg = req->async_data; |
4562 | if (!kmsg) { | |
7a7cacba PB |
4563 | ret = io_recvmsg_copy_hdr(req, &iomsg); |
4564 | if (ret) | |
681fda8d | 4565 | return ret; |
7a7cacba PB |
4566 | kmsg = &iomsg; |
4567 | } | |
03b1230c | 4568 | |
bc02ef33 | 4569 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4570 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bc02ef33 | 4571 | if (IS_ERR(kbuf)) |
52de1fe1 | 4572 | return PTR_ERR(kbuf); |
7a7cacba | 4573 | kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr); |
5476dfed PB |
4574 | kmsg->fast_iov[0].iov_len = req->sr_msg.len; |
4575 | iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov, | |
7a7cacba PB |
4576 | 1, req->sr_msg.len); |
4577 | } | |
52de1fe1 | 4578 | |
04411806 PB |
4579 | flags = req->sr_msg.msg_flags; |
4580 | if (force_nonblock) | |
7a7cacba | 4581 | flags |= MSG_DONTWAIT; |
0031275d SM |
4582 | if (flags & MSG_WAITALL) |
4583 | min_ret = iov_iter_count(&kmsg->msg.msg_iter); | |
4584 | ||
7a7cacba PB |
4585 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg, |
4586 | kmsg->uaddr, flags); | |
0e1b6fe3 PB |
4587 | if (force_nonblock && ret == -EAGAIN) |
4588 | return io_setup_async_msg(req, kmsg); | |
7a7cacba PB |
4589 | if (ret == -ERESTARTSYS) |
4590 | ret = -EINTR; | |
03b1230c | 4591 | |
7fbb1b54 PB |
4592 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4593 | cflags = io_put_recv_kbuf(req); | |
257e84a5 PB |
4594 | /* fast path, check for non-NULL to avoid function call */ |
4595 | if (kmsg->free_iov) | |
4596 | kfree(kmsg->free_iov); | |
99bc4c38 | 4597 | req->flags &= ~REQ_F_NEED_CLEANUP; |
0031275d | 4598 | if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC)))) |
93d2bcd2 | 4599 | req_set_fail(req); |
889fca73 | 4600 | __io_req_complete(req, issue_flags, ret, cflags); |
03b1230c | 4601 | return 0; |
0fa03c62 | 4602 | } |
5d17b4a4 | 4603 | |
889fca73 | 4604 | static int io_recv(struct io_kiocb *req, unsigned int issue_flags) |
fddaface | 4605 | { |
6b754c8b | 4606 | struct io_buffer *kbuf; |
7a7cacba PB |
4607 | struct io_sr_msg *sr = &req->sr_msg; |
4608 | struct msghdr msg; | |
4609 | void __user *buf = sr->buf; | |
fddaface | 4610 | struct socket *sock; |
7a7cacba PB |
4611 | struct iovec iov; |
4612 | unsigned flags; | |
0031275d | 4613 | int min_ret = 0; |
bcda7baa | 4614 | int ret, cflags = 0; |
45d189c6 | 4615 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
fddaface | 4616 | |
dba4a925 | 4617 | sock = sock_from_file(req->file); |
7a7cacba | 4618 | if (unlikely(!sock)) |
dba4a925 | 4619 | return -ENOTSOCK; |
fddaface | 4620 | |
bc02ef33 | 4621 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4622 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bcda7baa JA |
4623 | if (IS_ERR(kbuf)) |
4624 | return PTR_ERR(kbuf); | |
7a7cacba | 4625 | buf = u64_to_user_ptr(kbuf->addr); |
bc02ef33 | 4626 | } |
bcda7baa | 4627 | |
7a7cacba | 4628 | ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter); |
14c32eee PB |
4629 | if (unlikely(ret)) |
4630 | goto out_free; | |
fddaface | 4631 | |
7a7cacba PB |
4632 | msg.msg_name = NULL; |
4633 | msg.msg_control = NULL; | |
4634 | msg.msg_controllen = 0; | |
4635 | msg.msg_namelen = 0; | |
4636 | msg.msg_iocb = NULL; | |
4637 | msg.msg_flags = 0; | |
fddaface | 4638 | |
04411806 PB |
4639 | flags = req->sr_msg.msg_flags; |
4640 | if (force_nonblock) | |
7a7cacba | 4641 | flags |= MSG_DONTWAIT; |
0031275d SM |
4642 | if (flags & MSG_WAITALL) |
4643 | min_ret = iov_iter_count(&msg.msg_iter); | |
4644 | ||
7a7cacba PB |
4645 | ret = sock_recvmsg(sock, &msg, flags); |
4646 | if (force_nonblock && ret == -EAGAIN) | |
4647 | return -EAGAIN; | |
4648 | if (ret == -ERESTARTSYS) | |
4649 | ret = -EINTR; | |
14c32eee | 4650 | out_free: |
7fbb1b54 PB |
4651 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4652 | cflags = io_put_recv_kbuf(req); | |
0031275d | 4653 | if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC)))) |
93d2bcd2 | 4654 | req_set_fail(req); |
889fca73 | 4655 | __io_req_complete(req, issue_flags, ret, cflags); |
fddaface | 4656 | return 0; |
fddaface JA |
4657 | } |
4658 | ||
3529d8c2 | 4659 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 | 4660 | { |
8ed8d3c3 JA |
4661 | struct io_accept *accept = &req->accept; |
4662 | ||
14587a46 | 4663 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
17f2fe35 | 4664 | return -EINVAL; |
8042d6ce | 4665 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
4666 | return -EINVAL; |
4667 | ||
d55e5f5b JA |
4668 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4669 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 4670 | accept->flags = READ_ONCE(sqe->accept_flags); |
09952e3e | 4671 | accept->nofile = rlimit(RLIMIT_NOFILE); |
8ed8d3c3 | 4672 | return 0; |
8ed8d3c3 | 4673 | } |
17f2fe35 | 4674 | |
889fca73 | 4675 | static int io_accept(struct io_kiocb *req, unsigned int issue_flags) |
8ed8d3c3 JA |
4676 | { |
4677 | struct io_accept *accept = &req->accept; | |
45d189c6 | 4678 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
ac45abc0 | 4679 | unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0; |
8ed8d3c3 JA |
4680 | int ret; |
4681 | ||
e697deed JX |
4682 | if (req->file->f_flags & O_NONBLOCK) |
4683 | req->flags |= REQ_F_NOWAIT; | |
4684 | ||
8ed8d3c3 | 4685 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, |
09952e3e JA |
4686 | accept->addr_len, accept->flags, |
4687 | accept->nofile); | |
8ed8d3c3 | 4688 | if (ret == -EAGAIN && force_nonblock) |
17f2fe35 | 4689 | return -EAGAIN; |
ac45abc0 PB |
4690 | if (ret < 0) { |
4691 | if (ret == -ERESTARTSYS) | |
4692 | ret = -EINTR; | |
93d2bcd2 | 4693 | req_set_fail(req); |
ac45abc0 | 4694 | } |
889fca73 | 4695 | __io_req_complete(req, issue_flags, ret, 0); |
17f2fe35 | 4696 | return 0; |
8ed8d3c3 JA |
4697 | } |
4698 | ||
93642ef8 PB |
4699 | static int io_connect_prep_async(struct io_kiocb *req) |
4700 | { | |
4701 | struct io_async_connect *io = req->async_data; | |
4702 | struct io_connect *conn = &req->connect; | |
4703 | ||
4704 | return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address); | |
4705 | } | |
4706 | ||
3529d8c2 | 4707 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 | 4708 | { |
3529d8c2 | 4709 | struct io_connect *conn = &req->connect; |
f499a021 | 4710 | |
14587a46 | 4711 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3fbb51c1 JA |
4712 | return -EINVAL; |
4713 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
4714 | return -EINVAL; | |
4715 | ||
3529d8c2 JA |
4716 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4717 | conn->addr_len = READ_ONCE(sqe->addr2); | |
93642ef8 | 4718 | return 0; |
f499a021 JA |
4719 | } |
4720 | ||
889fca73 | 4721 | static int io_connect(struct io_kiocb *req, unsigned int issue_flags) |
f8e85cf2 | 4722 | { |
e8c2bc1f | 4723 | struct io_async_connect __io, *io; |
f8e85cf2 | 4724 | unsigned file_flags; |
3fbb51c1 | 4725 | int ret; |
45d189c6 | 4726 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
f8e85cf2 | 4727 | |
e8c2bc1f JA |
4728 | if (req->async_data) { |
4729 | io = req->async_data; | |
f499a021 | 4730 | } else { |
3529d8c2 JA |
4731 | ret = move_addr_to_kernel(req->connect.addr, |
4732 | req->connect.addr_len, | |
e8c2bc1f | 4733 | &__io.address); |
f499a021 JA |
4734 | if (ret) |
4735 | goto out; | |
4736 | io = &__io; | |
4737 | } | |
4738 | ||
3fbb51c1 JA |
4739 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
4740 | ||
e8c2bc1f | 4741 | ret = __sys_connect_file(req->file, &io->address, |
3fbb51c1 | 4742 | req->connect.addr_len, file_flags); |
87f80d62 | 4743 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
e8c2bc1f | 4744 | if (req->async_data) |
b7bb4f7d | 4745 | return -EAGAIN; |
e8c2bc1f | 4746 | if (io_alloc_async_data(req)) { |
f499a021 JA |
4747 | ret = -ENOMEM; |
4748 | goto out; | |
4749 | } | |
e8c2bc1f | 4750 | memcpy(req->async_data, &__io, sizeof(__io)); |
f8e85cf2 | 4751 | return -EAGAIN; |
f499a021 | 4752 | } |
f8e85cf2 JA |
4753 | if (ret == -ERESTARTSYS) |
4754 | ret = -EINTR; | |
f499a021 | 4755 | out: |
4e88d6e7 | 4756 | if (ret < 0) |
93d2bcd2 | 4757 | req_set_fail(req); |
889fca73 | 4758 | __io_req_complete(req, issue_flags, ret, 0); |
f8e85cf2 | 4759 | return 0; |
469956e8 Y |
4760 | } |
4761 | #else /* !CONFIG_NET */ | |
99a10081 JA |
4762 | #define IO_NETOP_FN(op) \ |
4763 | static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \ | |
4764 | { \ | |
4765 | return -EOPNOTSUPP; \ | |
4766 | } | |
4767 | ||
4768 | #define IO_NETOP_PREP(op) \ | |
4769 | IO_NETOP_FN(op) \ | |
4770 | static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \ | |
4771 | { \ | |
4772 | return -EOPNOTSUPP; \ | |
4773 | } \ | |
4774 | ||
4775 | #define IO_NETOP_PREP_ASYNC(op) \ | |
4776 | IO_NETOP_PREP(op) \ | |
4777 | static int io_##op##_prep_async(struct io_kiocb *req) \ | |
4778 | { \ | |
4779 | return -EOPNOTSUPP; \ | |
4780 | } | |
4781 | ||
4782 | IO_NETOP_PREP_ASYNC(sendmsg); | |
4783 | IO_NETOP_PREP_ASYNC(recvmsg); | |
4784 | IO_NETOP_PREP_ASYNC(connect); | |
4785 | IO_NETOP_PREP(accept); | |
4786 | IO_NETOP_FN(send); | |
4787 | IO_NETOP_FN(recv); | |
469956e8 | 4788 | #endif /* CONFIG_NET */ |
f8e85cf2 | 4789 | |
d7718a9d JA |
4790 | struct io_poll_table { |
4791 | struct poll_table_struct pt; | |
4792 | struct io_kiocb *req; | |
68b11e8b | 4793 | int nr_entries; |
d7718a9d JA |
4794 | int error; |
4795 | }; | |
ce593a6c | 4796 | |
d7718a9d | 4797 | static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll, |
5b0a6acc | 4798 | __poll_t mask, io_req_tw_func_t func) |
d7718a9d | 4799 | { |
d7718a9d JA |
4800 | /* for instances that support it check for an event match first: */ |
4801 | if (mask && !(mask & poll->events)) | |
4802 | return 0; | |
4803 | ||
4804 | trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask); | |
4805 | ||
4806 | list_del_init(&poll->wait.entry); | |
4807 | ||
d7718a9d | 4808 | req->result = mask; |
5b0a6acc | 4809 | req->io_task_work.func = func; |
6d816e08 | 4810 | |
d7718a9d | 4811 | /* |
e3aabf95 JA |
4812 | * If this fails, then the task is exiting. When a task exits, the |
4813 | * work gets canceled, so just cancel this request as well instead | |
4814 | * of executing it. We can't safely execute it anyway, as we may not | |
4815 | * have the needed state needed for it anyway. | |
d7718a9d | 4816 | */ |
e09ee510 | 4817 | io_req_task_work_add(req); |
d7718a9d JA |
4818 | return 1; |
4819 | } | |
4820 | ||
74ce6ce4 JA |
4821 | static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll) |
4822 | __acquires(&req->ctx->completion_lock) | |
4823 | { | |
4824 | struct io_ring_ctx *ctx = req->ctx; | |
4825 | ||
e09ee510 PB |
4826 | if (unlikely(req->task->flags & PF_EXITING)) |
4827 | WRITE_ONCE(poll->canceled, true); | |
4828 | ||
74ce6ce4 JA |
4829 | if (!req->result && !READ_ONCE(poll->canceled)) { |
4830 | struct poll_table_struct pt = { ._key = poll->events }; | |
4831 | ||
4832 | req->result = vfs_poll(req->file, &pt) & poll->events; | |
4833 | } | |
4834 | ||
4835 | spin_lock_irq(&ctx->completion_lock); | |
4836 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
4837 | add_wait_queue(poll->head, &poll->wait); | |
4838 | return true; | |
4839 | } | |
4840 | ||
4841 | return false; | |
4842 | } | |
4843 | ||
d4e7cd36 | 4844 | static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req) |
18bceab1 | 4845 | { |
e8c2bc1f | 4846 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ |
d4e7cd36 | 4847 | if (req->opcode == IORING_OP_POLL_ADD) |
e8c2bc1f | 4848 | return req->async_data; |
d4e7cd36 JA |
4849 | return req->apoll->double_poll; |
4850 | } | |
4851 | ||
4852 | static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req) | |
4853 | { | |
4854 | if (req->opcode == IORING_OP_POLL_ADD) | |
4855 | return &req->poll; | |
4856 | return &req->apoll->poll; | |
4857 | } | |
4858 | ||
4859 | static void io_poll_remove_double(struct io_kiocb *req) | |
e07785b0 | 4860 | __must_hold(&req->ctx->completion_lock) |
d4e7cd36 JA |
4861 | { |
4862 | struct io_poll_iocb *poll = io_poll_get_double(req); | |
18bceab1 JA |
4863 | |
4864 | lockdep_assert_held(&req->ctx->completion_lock); | |
4865 | ||
4866 | if (poll && poll->head) { | |
4867 | struct wait_queue_head *head = poll->head; | |
4868 | ||
4869 | spin_lock(&head->lock); | |
4870 | list_del_init(&poll->wait.entry); | |
4871 | if (poll->wait.private) | |
de9b4cca | 4872 | req_ref_put(req); |
18bceab1 JA |
4873 | poll->head = NULL; |
4874 | spin_unlock(&head->lock); | |
4875 | } | |
4876 | } | |
4877 | ||
e27414be | 4878 | static bool io_poll_complete(struct io_kiocb *req, __poll_t mask) |
e07785b0 | 4879 | __must_hold(&req->ctx->completion_lock) |
18bceab1 JA |
4880 | { |
4881 | struct io_ring_ctx *ctx = req->ctx; | |
88e41cf9 | 4882 | unsigned flags = IORING_CQE_F_MORE; |
e27414be | 4883 | int error; |
18bceab1 | 4884 | |
e27414be | 4885 | if (READ_ONCE(req->poll.canceled)) { |
45ab03b1 | 4886 | error = -ECANCELED; |
88e41cf9 | 4887 | req->poll.events |= EPOLLONESHOT; |
e27414be | 4888 | } else { |
5082620f | 4889 | error = mangle_poll(mask); |
e27414be | 4890 | } |
b69de288 JA |
4891 | if (req->poll.events & EPOLLONESHOT) |
4892 | flags = 0; | |
d4d19c19 | 4893 | if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) { |
88e41cf9 JA |
4894 | req->poll.done = true; |
4895 | flags = 0; | |
4896 | } | |
7b289c38 HX |
4897 | if (flags & IORING_CQE_F_MORE) |
4898 | ctx->cq_extra++; | |
18bceab1 | 4899 | |
18bceab1 | 4900 | io_commit_cqring(ctx); |
88e41cf9 | 4901 | return !(flags & IORING_CQE_F_MORE); |
18bceab1 JA |
4902 | } |
4903 | ||
5b0a6acc | 4904 | static void io_poll_task_func(struct io_kiocb *req) |
18bceab1 JA |
4905 | { |
4906 | struct io_ring_ctx *ctx = req->ctx; | |
dd221f46 | 4907 | struct io_kiocb *nxt; |
18bceab1 JA |
4908 | |
4909 | if (io_poll_rewait(req, &req->poll)) { | |
4910 | spin_unlock_irq(&ctx->completion_lock); | |
dd221f46 | 4911 | } else { |
f40b964a | 4912 | bool done; |
18bceab1 | 4913 | |
e27414be | 4914 | done = io_poll_complete(req, req->result); |
88e41cf9 | 4915 | if (done) { |
a890d01e | 4916 | io_poll_remove_double(req); |
88e41cf9 | 4917 | hash_del(&req->hash_node); |
f40b964a | 4918 | } else { |
88e41cf9 JA |
4919 | req->result = 0; |
4920 | add_wait_queue(req->poll.head, &req->poll.wait); | |
4921 | } | |
dd221f46 | 4922 | spin_unlock_irq(&ctx->completion_lock); |
dd221f46 | 4923 | io_cqring_ev_posted(ctx); |
18bceab1 | 4924 | |
88e41cf9 JA |
4925 | if (done) { |
4926 | nxt = io_put_req_find_next(req); | |
4927 | if (nxt) | |
5b0a6acc | 4928 | io_req_task_submit(nxt); |
88e41cf9 | 4929 | } |
dd221f46 | 4930 | } |
18bceab1 JA |
4931 | } |
4932 | ||
4933 | static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode, | |
4934 | int sync, void *key) | |
4935 | { | |
4936 | struct io_kiocb *req = wait->private; | |
d4e7cd36 | 4937 | struct io_poll_iocb *poll = io_poll_get_single(req); |
18bceab1 JA |
4938 | __poll_t mask = key_to_poll(key); |
4939 | ||
4940 | /* for instances that support it check for an event match first: */ | |
4941 | if (mask && !(mask & poll->events)) | |
4942 | return 0; | |
88e41cf9 JA |
4943 | if (!(poll->events & EPOLLONESHOT)) |
4944 | return poll->wait.func(&poll->wait, mode, sync, key); | |
18bceab1 | 4945 | |
8706e04e JA |
4946 | list_del_init(&wait->entry); |
4947 | ||
9ce85ef2 | 4948 | if (poll->head) { |
18bceab1 JA |
4949 | bool done; |
4950 | ||
807abcb0 JA |
4951 | spin_lock(&poll->head->lock); |
4952 | done = list_empty(&poll->wait.entry); | |
18bceab1 | 4953 | if (!done) |
807abcb0 | 4954 | list_del_init(&poll->wait.entry); |
d4e7cd36 JA |
4955 | /* make sure double remove sees this as being gone */ |
4956 | wait->private = NULL; | |
807abcb0 | 4957 | spin_unlock(&poll->head->lock); |
c8b5e260 JA |
4958 | if (!done) { |
4959 | /* use wait func handler, so it matches the rq type */ | |
4960 | poll->wait.func(&poll->wait, mode, sync, key); | |
4961 | } | |
18bceab1 | 4962 | } |
de9b4cca | 4963 | req_ref_put(req); |
18bceab1 JA |
4964 | return 1; |
4965 | } | |
4966 | ||
4967 | static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events, | |
4968 | wait_queue_func_t wake_func) | |
4969 | { | |
4970 | poll->head = NULL; | |
4971 | poll->done = false; | |
4972 | poll->canceled = false; | |
464dca61 JA |
4973 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) |
4974 | /* mask in events that we always want/need */ | |
4975 | poll->events = events | IO_POLL_UNMASK; | |
18bceab1 JA |
4976 | INIT_LIST_HEAD(&poll->wait.entry); |
4977 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
4978 | } | |
4979 | ||
4980 | static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt, | |
807abcb0 JA |
4981 | struct wait_queue_head *head, |
4982 | struct io_poll_iocb **poll_ptr) | |
18bceab1 JA |
4983 | { |
4984 | struct io_kiocb *req = pt->req; | |
4985 | ||
4986 | /* | |
68b11e8b PB |
4987 | * The file being polled uses multiple waitqueues for poll handling |
4988 | * (e.g. one for read, one for write). Setup a separate io_poll_iocb | |
4989 | * if this happens. | |
18bceab1 | 4990 | */ |
68b11e8b | 4991 | if (unlikely(pt->nr_entries)) { |
58852d4d PB |
4992 | struct io_poll_iocb *poll_one = poll; |
4993 | ||
18bceab1 | 4994 | /* already have a 2nd entry, fail a third attempt */ |
807abcb0 | 4995 | if (*poll_ptr) { |
18bceab1 JA |
4996 | pt->error = -EINVAL; |
4997 | return; | |
4998 | } | |
ea6a693d JA |
4999 | /* |
5000 | * Can't handle multishot for double wait for now, turn it | |
5001 | * into one-shot mode. | |
5002 | */ | |
7a274727 PB |
5003 | if (!(poll_one->events & EPOLLONESHOT)) |
5004 | poll_one->events |= EPOLLONESHOT; | |
1c3b3e65 | 5005 | /* double add on the same waitqueue head, ignore */ |
7a274727 | 5006 | if (poll_one->head == head) |
1c3b3e65 | 5007 | return; |
18bceab1 JA |
5008 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); |
5009 | if (!poll) { | |
5010 | pt->error = -ENOMEM; | |
5011 | return; | |
5012 | } | |
58852d4d | 5013 | io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake); |
de9b4cca | 5014 | req_ref_get(req); |
18bceab1 | 5015 | poll->wait.private = req; |
807abcb0 | 5016 | *poll_ptr = poll; |
18bceab1 JA |
5017 | } |
5018 | ||
68b11e8b | 5019 | pt->nr_entries++; |
18bceab1 | 5020 | poll->head = head; |
a31eb4a2 JX |
5021 | |
5022 | if (poll->events & EPOLLEXCLUSIVE) | |
5023 | add_wait_queue_exclusive(head, &poll->wait); | |
5024 | else | |
5025 | add_wait_queue(head, &poll->wait); | |
18bceab1 JA |
5026 | } |
5027 | ||
5028 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
5029 | struct poll_table_struct *p) | |
5030 | { | |
5031 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
807abcb0 | 5032 | struct async_poll *apoll = pt->req->apoll; |
18bceab1 | 5033 | |
807abcb0 | 5034 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
18bceab1 JA |
5035 | } |
5036 | ||
5b0a6acc | 5037 | static void io_async_task_func(struct io_kiocb *req) |
d7718a9d | 5038 | { |
d7718a9d JA |
5039 | struct async_poll *apoll = req->apoll; |
5040 | struct io_ring_ctx *ctx = req->ctx; | |
5041 | ||
236daeae | 5042 | trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data); |
d7718a9d | 5043 | |
74ce6ce4 | 5044 | if (io_poll_rewait(req, &apoll->poll)) { |
d7718a9d | 5045 | spin_unlock_irq(&ctx->completion_lock); |
74ce6ce4 | 5046 | return; |
d7718a9d JA |
5047 | } |
5048 | ||
0ea13b44 | 5049 | hash_del(&req->hash_node); |
d4e7cd36 | 5050 | io_poll_remove_double(req); |
74ce6ce4 JA |
5051 | spin_unlock_irq(&ctx->completion_lock); |
5052 | ||
0be0b0e3 | 5053 | if (!READ_ONCE(apoll->poll.canceled)) |
5b0a6acc | 5054 | io_req_task_submit(req); |
0be0b0e3 | 5055 | else |
2593553a | 5056 | io_req_complete_failed(req, -ECANCELED); |
d7718a9d JA |
5057 | } |
5058 | ||
5059 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
5060 | void *key) | |
5061 | { | |
5062 | struct io_kiocb *req = wait->private; | |
5063 | struct io_poll_iocb *poll = &req->apoll->poll; | |
5064 | ||
5065 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
5066 | key_to_poll(key)); | |
5067 | ||
5068 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
5069 | } | |
5070 | ||
5071 | static void io_poll_req_insert(struct io_kiocb *req) | |
5072 | { | |
5073 | struct io_ring_ctx *ctx = req->ctx; | |
5074 | struct hlist_head *list; | |
5075 | ||
5076 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
5077 | hlist_add_head(&req->hash_node, list); | |
5078 | } | |
5079 | ||
5080 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
5081 | struct io_poll_iocb *poll, | |
5082 | struct io_poll_table *ipt, __poll_t mask, | |
5083 | wait_queue_func_t wake_func) | |
5084 | __acquires(&ctx->completion_lock) | |
5085 | { | |
5086 | struct io_ring_ctx *ctx = req->ctx; | |
5087 | bool cancel = false; | |
5088 | ||
4d52f338 | 5089 | INIT_HLIST_NODE(&req->hash_node); |
18bceab1 | 5090 | io_init_poll_iocb(poll, mask, wake_func); |
b90cd197 | 5091 | poll->file = req->file; |
18bceab1 | 5092 | poll->wait.private = req; |
d7718a9d JA |
5093 | |
5094 | ipt->pt._key = mask; | |
5095 | ipt->req = req; | |
68b11e8b PB |
5096 | ipt->error = 0; |
5097 | ipt->nr_entries = 0; | |
d7718a9d | 5098 | |
d7718a9d | 5099 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
68b11e8b PB |
5100 | if (unlikely(!ipt->nr_entries) && !ipt->error) |
5101 | ipt->error = -EINVAL; | |
d7718a9d JA |
5102 | |
5103 | spin_lock_irq(&ctx->completion_lock); | |
a890d01e | 5104 | if (ipt->error || (mask && (poll->events & EPOLLONESHOT))) |
46fee9ab | 5105 | io_poll_remove_double(req); |
d7718a9d JA |
5106 | if (likely(poll->head)) { |
5107 | spin_lock(&poll->head->lock); | |
5108 | if (unlikely(list_empty(&poll->wait.entry))) { | |
5109 | if (ipt->error) | |
5110 | cancel = true; | |
5111 | ipt->error = 0; | |
5112 | mask = 0; | |
5113 | } | |
88e41cf9 | 5114 | if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error) |
d7718a9d JA |
5115 | list_del_init(&poll->wait.entry); |
5116 | else if (cancel) | |
5117 | WRITE_ONCE(poll->canceled, true); | |
5118 | else if (!poll->done) /* actually waiting for an event */ | |
5119 | io_poll_req_insert(req); | |
5120 | spin_unlock(&poll->head->lock); | |
5121 | } | |
5122 | ||
5123 | return mask; | |
5124 | } | |
5125 | ||
59b735ae OL |
5126 | enum { |
5127 | IO_APOLL_OK, | |
5128 | IO_APOLL_ABORTED, | |
5129 | IO_APOLL_READY | |
5130 | }; | |
5131 | ||
5132 | static int io_arm_poll_handler(struct io_kiocb *req) | |
d7718a9d JA |
5133 | { |
5134 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
5135 | struct io_ring_ctx *ctx = req->ctx; | |
5136 | struct async_poll *apoll; | |
5137 | struct io_poll_table ipt; | |
b2d9c3da | 5138 | __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI; |
9dab14b8 | 5139 | int rw; |
d7718a9d JA |
5140 | |
5141 | if (!req->file || !file_can_poll(req->file)) | |
59b735ae | 5142 | return IO_APOLL_ABORTED; |
24c74678 | 5143 | if (req->flags & REQ_F_POLLED) |
59b735ae | 5144 | return IO_APOLL_ABORTED; |
b2d9c3da PB |
5145 | if (!def->pollin && !def->pollout) |
5146 | return IO_APOLL_ABORTED; | |
5147 | ||
5148 | if (def->pollin) { | |
9dab14b8 | 5149 | rw = READ; |
b2d9c3da PB |
5150 | mask |= POLLIN | POLLRDNORM; |
5151 | ||
5152 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
5153 | if ((req->opcode == IORING_OP_RECVMSG) && | |
5154 | (req->sr_msg.msg_flags & MSG_ERRQUEUE)) | |
5155 | mask &= ~POLLIN; | |
5156 | } else { | |
9dab14b8 | 5157 | rw = WRITE; |
b2d9c3da PB |
5158 | mask |= POLLOUT | POLLWRNORM; |
5159 | } | |
5160 | ||
9dab14b8 | 5161 | /* if we can't nonblock try, then no point in arming a poll handler */ |
b191e2df | 5162 | if (!io_file_supports_nowait(req, rw)) |
59b735ae | 5163 | return IO_APOLL_ABORTED; |
d7718a9d JA |
5164 | |
5165 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
5166 | if (unlikely(!apoll)) | |
59b735ae | 5167 | return IO_APOLL_ABORTED; |
807abcb0 | 5168 | apoll->double_poll = NULL; |
d7718a9d | 5169 | req->apoll = apoll; |
b2d9c3da | 5170 | req->flags |= REQ_F_POLLED; |
d7718a9d JA |
5171 | ipt.pt._qproc = io_async_queue_proc; |
5172 | ||
5173 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
5174 | io_async_wake); | |
a36da65c | 5175 | if (ret || ipt.error) { |
d7718a9d | 5176 | spin_unlock_irq(&ctx->completion_lock); |
59b735ae OL |
5177 | if (ret) |
5178 | return IO_APOLL_READY; | |
5179 | return IO_APOLL_ABORTED; | |
d7718a9d JA |
5180 | } |
5181 | spin_unlock_irq(&ctx->completion_lock); | |
236daeae OL |
5182 | trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data, |
5183 | mask, apoll->poll.events); | |
59b735ae | 5184 | return IO_APOLL_OK; |
d7718a9d JA |
5185 | } |
5186 | ||
5187 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
b2e720ac | 5188 | struct io_poll_iocb *poll, bool do_cancel) |
e07785b0 | 5189 | __must_hold(&req->ctx->completion_lock) |
221c5eb2 | 5190 | { |
b41e9852 | 5191 | bool do_complete = false; |
221c5eb2 | 5192 | |
5082620f JA |
5193 | if (!poll->head) |
5194 | return false; | |
221c5eb2 | 5195 | spin_lock(&poll->head->lock); |
b2e720ac JA |
5196 | if (do_cancel) |
5197 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
5198 | if (!list_empty(&poll->wait.entry)) { |
5199 | list_del_init(&poll->wait.entry); | |
b41e9852 | 5200 | do_complete = true; |
221c5eb2 JA |
5201 | } |
5202 | spin_unlock(&poll->head->lock); | |
3bfa5bcb | 5203 | hash_del(&req->hash_node); |
d7718a9d JA |
5204 | return do_complete; |
5205 | } | |
5206 | ||
5d709043 | 5207 | static bool io_poll_remove_one(struct io_kiocb *req) |
e07785b0 | 5208 | __must_hold(&req->ctx->completion_lock) |
d7718a9d | 5209 | { |
5d709043 | 5210 | int refs; |
d7718a9d JA |
5211 | bool do_complete; |
5212 | ||
d4e7cd36 | 5213 | io_poll_remove_double(req); |
e31001a3 | 5214 | do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true); |
d4e7cd36 | 5215 | |
b41e9852 | 5216 | if (do_complete) { |
d4d19c19 | 5217 | io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0); |
b41e9852 | 5218 | io_commit_cqring(req->ctx); |
93d2bcd2 | 5219 | req_set_fail(req); |
b41e9852 | 5220 | |
5d709043 PB |
5221 | /* non-poll requests have submit ref still */ |
5222 | refs = 1 + (req->opcode != IORING_OP_POLL_ADD); | |
5223 | io_put_req_deferred(req, refs); | |
5224 | } | |
b41e9852 | 5225 | return do_complete; |
221c5eb2 JA |
5226 | } |
5227 | ||
76e1b642 JA |
5228 | /* |
5229 | * Returns true if we found and killed one or more poll requests | |
5230 | */ | |
6b81928d | 5231 | static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, |
3dd0c97a | 5232 | bool cancel_all) |
221c5eb2 | 5233 | { |
78076bb6 | 5234 | struct hlist_node *tmp; |
221c5eb2 | 5235 | struct io_kiocb *req; |
8e2e1faf | 5236 | int posted = 0, i; |
221c5eb2 JA |
5237 | |
5238 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
5239 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
5240 | struct hlist_head *list; | |
5241 | ||
5242 | list = &ctx->cancel_hash[i]; | |
f3606e3a | 5243 | hlist_for_each_entry_safe(req, tmp, list, hash_node) { |
3dd0c97a | 5244 | if (io_match_task(req, tsk, cancel_all)) |
f3606e3a JA |
5245 | posted += io_poll_remove_one(req); |
5246 | } | |
221c5eb2 JA |
5247 | } |
5248 | spin_unlock_irq(&ctx->completion_lock); | |
b41e9852 | 5249 | |
8e2e1faf JA |
5250 | if (posted) |
5251 | io_cqring_ev_posted(ctx); | |
76e1b642 JA |
5252 | |
5253 | return posted != 0; | |
221c5eb2 JA |
5254 | } |
5255 | ||
9ba5fac8 PB |
5256 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr, |
5257 | bool poll_only) | |
e07785b0 | 5258 | __must_hold(&ctx->completion_lock) |
47f46768 | 5259 | { |
78076bb6 | 5260 | struct hlist_head *list; |
47f46768 JA |
5261 | struct io_kiocb *req; |
5262 | ||
78076bb6 JA |
5263 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
5264 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
5265 | if (sqe_addr != req->user_data) |
5266 | continue; | |
9ba5fac8 PB |
5267 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) |
5268 | continue; | |
b2cb805f | 5269 | return req; |
47f46768 | 5270 | } |
b2cb805f JA |
5271 | return NULL; |
5272 | } | |
5273 | ||
9ba5fac8 PB |
5274 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr, |
5275 | bool poll_only) | |
e07785b0 | 5276 | __must_hold(&ctx->completion_lock) |
b2cb805f JA |
5277 | { |
5278 | struct io_kiocb *req; | |
5279 | ||
9ba5fac8 | 5280 | req = io_poll_find(ctx, sqe_addr, poll_only); |
b2cb805f JA |
5281 | if (!req) |
5282 | return -ENOENT; | |
5283 | if (io_poll_remove_one(req)) | |
5284 | return 0; | |
5285 | ||
5286 | return -EALREADY; | |
47f46768 JA |
5287 | } |
5288 | ||
9096af3e PB |
5289 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
5290 | unsigned int flags) | |
5291 | { | |
5292 | u32 events; | |
47f46768 | 5293 | |
9096af3e PB |
5294 | events = READ_ONCE(sqe->poll32_events); |
5295 | #ifdef __BIG_ENDIAN | |
5296 | events = swahw32(events); | |
5297 | #endif | |
5298 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
5299 | events |= EPOLLONESHOT; | |
5300 | return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT)); | |
47f46768 JA |
5301 | } |
5302 | ||
c5de0036 | 5303 | static int io_poll_update_prep(struct io_kiocb *req, |
3529d8c2 | 5304 | const struct io_uring_sqe *sqe) |
0969e783 | 5305 | { |
c5de0036 PB |
5306 | struct io_poll_update *upd = &req->poll_update; |
5307 | u32 flags; | |
5308 | ||
0969e783 JA |
5309 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5310 | return -EINVAL; | |
c5de0036 PB |
5311 | if (sqe->ioprio || sqe->buf_index) |
5312 | return -EINVAL; | |
5313 | flags = READ_ONCE(sqe->len); | |
5314 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
5315 | IORING_POLL_ADD_MULTI)) | |
5316 | return -EINVAL; | |
5317 | /* meaningless without update */ | |
5318 | if (flags == IORING_POLL_ADD_MULTI) | |
0969e783 JA |
5319 | return -EINVAL; |
5320 | ||
c5de0036 PB |
5321 | upd->old_user_data = READ_ONCE(sqe->addr); |
5322 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
5323 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
221c5eb2 | 5324 | |
c5de0036 PB |
5325 | upd->new_user_data = READ_ONCE(sqe->off); |
5326 | if (!upd->update_user_data && upd->new_user_data) | |
5327 | return -EINVAL; | |
5328 | if (upd->update_events) | |
5329 | upd->events = io_poll_parse_events(sqe, flags); | |
5330 | else if (sqe->poll32_events) | |
5331 | return -EINVAL; | |
221c5eb2 | 5332 | |
221c5eb2 JA |
5333 | return 0; |
5334 | } | |
5335 | ||
221c5eb2 JA |
5336 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
5337 | void *key) | |
5338 | { | |
c2f2eb7d JA |
5339 | struct io_kiocb *req = wait->private; |
5340 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 5341 | |
d7718a9d | 5342 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
5343 | } |
5344 | ||
221c5eb2 JA |
5345 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
5346 | struct poll_table_struct *p) | |
5347 | { | |
5348 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
5349 | ||
e8c2bc1f | 5350 | __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data); |
eac406c6 JA |
5351 | } |
5352 | ||
3529d8c2 | 5353 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
5354 | { |
5355 | struct io_poll_iocb *poll = &req->poll; | |
c5de0036 | 5356 | u32 flags; |
221c5eb2 JA |
5357 | |
5358 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
5359 | return -EINVAL; | |
c5de0036 | 5360 | if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr) |
88e41cf9 JA |
5361 | return -EINVAL; |
5362 | flags = READ_ONCE(sqe->len); | |
c5de0036 | 5363 | if (flags & ~IORING_POLL_ADD_MULTI) |
221c5eb2 JA |
5364 | return -EINVAL; |
5365 | ||
c5de0036 | 5366 | poll->events = io_poll_parse_events(sqe, flags); |
0969e783 JA |
5367 | return 0; |
5368 | } | |
5369 | ||
61e98203 | 5370 | static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) |
0969e783 JA |
5371 | { |
5372 | struct io_poll_iocb *poll = &req->poll; | |
5373 | struct io_ring_ctx *ctx = req->ctx; | |
5374 | struct io_poll_table ipt; | |
0969e783 | 5375 | __poll_t mask; |
0969e783 | 5376 | |
d7718a9d | 5377 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 5378 | |
d7718a9d JA |
5379 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
5380 | io_poll_wake); | |
221c5eb2 | 5381 | |
8c838788 | 5382 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 5383 | ipt.error = 0; |
e27414be | 5384 | io_poll_complete(req, mask); |
221c5eb2 | 5385 | } |
221c5eb2 JA |
5386 | spin_unlock_irq(&ctx->completion_lock); |
5387 | ||
8c838788 JA |
5388 | if (mask) { |
5389 | io_cqring_ev_posted(ctx); | |
88e41cf9 JA |
5390 | if (poll->events & EPOLLONESHOT) |
5391 | io_put_req(req); | |
221c5eb2 | 5392 | } |
8c838788 | 5393 | return ipt.error; |
221c5eb2 JA |
5394 | } |
5395 | ||
c5de0036 | 5396 | static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags) |
b69de288 JA |
5397 | { |
5398 | struct io_ring_ctx *ctx = req->ctx; | |
5399 | struct io_kiocb *preq; | |
cb3b200e | 5400 | bool completing; |
b69de288 JA |
5401 | int ret; |
5402 | ||
5403 | spin_lock_irq(&ctx->completion_lock); | |
9ba5fac8 | 5404 | preq = io_poll_find(ctx, req->poll_update.old_user_data, true); |
b69de288 JA |
5405 | if (!preq) { |
5406 | ret = -ENOENT; | |
5407 | goto err; | |
b69de288 | 5408 | } |
cb3b200e | 5409 | |
c5de0036 PB |
5410 | if (!req->poll_update.update_events && !req->poll_update.update_user_data) { |
5411 | completing = true; | |
5412 | ret = io_poll_remove_one(preq) ? 0 : -EALREADY; | |
5413 | goto err; | |
5414 | } | |
5415 | ||
cb3b200e JA |
5416 | /* |
5417 | * Don't allow racy completion with singleshot, as we cannot safely | |
5418 | * update those. For multishot, if we're racing with completion, just | |
5419 | * let completion re-add it. | |
5420 | */ | |
5421 | completing = !__io_poll_remove_one(preq, &preq->poll, false); | |
5422 | if (completing && (preq->poll.events & EPOLLONESHOT)) { | |
5423 | ret = -EALREADY; | |
5424 | goto err; | |
b69de288 JA |
5425 | } |
5426 | /* we now have a detached poll request. reissue. */ | |
5427 | ret = 0; | |
5428 | err: | |
b69de288 | 5429 | if (ret < 0) { |
cb3b200e | 5430 | spin_unlock_irq(&ctx->completion_lock); |
93d2bcd2 | 5431 | req_set_fail(req); |
b69de288 JA |
5432 | io_req_complete(req, ret); |
5433 | return 0; | |
5434 | } | |
5435 | /* only mask one event flags, keep behavior flags */ | |
9d805892 | 5436 | if (req->poll_update.update_events) { |
b69de288 | 5437 | preq->poll.events &= ~0xffff; |
9d805892 | 5438 | preq->poll.events |= req->poll_update.events & 0xffff; |
b69de288 JA |
5439 | preq->poll.events |= IO_POLL_UNMASK; |
5440 | } | |
9d805892 PB |
5441 | if (req->poll_update.update_user_data) |
5442 | preq->user_data = req->poll_update.new_user_data; | |
cb3b200e JA |
5443 | spin_unlock_irq(&ctx->completion_lock); |
5444 | ||
b69de288 JA |
5445 | /* complete update request, we're done with it */ |
5446 | io_req_complete(req, ret); | |
5447 | ||
cb3b200e | 5448 | if (!completing) { |
c5de0036 | 5449 | ret = io_poll_add(preq, issue_flags); |
cb3b200e | 5450 | if (ret < 0) { |
93d2bcd2 | 5451 | req_set_fail(preq); |
cb3b200e JA |
5452 | io_req_complete(preq, ret); |
5453 | } | |
b69de288 JA |
5454 | } |
5455 | return 0; | |
5456 | } | |
5457 | ||
5262f567 JA |
5458 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
5459 | { | |
ad8a48ac JA |
5460 | struct io_timeout_data *data = container_of(timer, |
5461 | struct io_timeout_data, timer); | |
5462 | struct io_kiocb *req = data->req; | |
5463 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
5464 | unsigned long flags; |
5465 | ||
5262f567 | 5466 | spin_lock_irqsave(&ctx->completion_lock, flags); |
a71976f3 | 5467 | list_del_init(&req->timeout.list); |
01cec8c1 PB |
5468 | atomic_set(&req->ctx->cq_timeouts, |
5469 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
5470 | ||
d4d19c19 | 5471 | io_cqring_fill_event(ctx, req->user_data, -ETIME, 0); |
5262f567 JA |
5472 | io_commit_cqring(ctx); |
5473 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5474 | ||
5475 | io_cqring_ev_posted(ctx); | |
93d2bcd2 | 5476 | req_set_fail(req); |
5262f567 JA |
5477 | io_put_req(req); |
5478 | return HRTIMER_NORESTART; | |
5479 | } | |
5480 | ||
fbd15848 PB |
5481 | static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx, |
5482 | __u64 user_data) | |
e07785b0 | 5483 | __must_hold(&ctx->completion_lock) |
f254ac04 | 5484 | { |
fbd15848 | 5485 | struct io_timeout_data *io; |
47f46768 | 5486 | struct io_kiocb *req; |
fd9c7bc5 | 5487 | bool found = false; |
f254ac04 | 5488 | |
135fcde8 | 5489 | list_for_each_entry(req, &ctx->timeout_list, timeout.list) { |
fd9c7bc5 PB |
5490 | found = user_data == req->user_data; |
5491 | if (found) | |
47f46768 | 5492 | break; |
47f46768 | 5493 | } |
fd9c7bc5 PB |
5494 | if (!found) |
5495 | return ERR_PTR(-ENOENT); | |
fbd15848 PB |
5496 | |
5497 | io = req->async_data; | |
fd9c7bc5 | 5498 | if (hrtimer_try_to_cancel(&io->timer) == -1) |
fbd15848 | 5499 | return ERR_PTR(-EALREADY); |
a71976f3 | 5500 | list_del_init(&req->timeout.list); |
fbd15848 PB |
5501 | return req; |
5502 | } | |
47f46768 | 5503 | |
fbd15848 | 5504 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
e07785b0 | 5505 | __must_hold(&ctx->completion_lock) |
fbd15848 PB |
5506 | { |
5507 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); | |
5508 | ||
5509 | if (IS_ERR(req)) | |
5510 | return PTR_ERR(req); | |
f254ac04 | 5511 | |
93d2bcd2 | 5512 | req_set_fail(req); |
d4d19c19 | 5513 | io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0); |
216578e5 | 5514 | io_put_req_deferred(req, 1); |
f254ac04 JA |
5515 | return 0; |
5516 | } | |
5517 | ||
9c8e11b3 PB |
5518 | static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data, |
5519 | struct timespec64 *ts, enum hrtimer_mode mode) | |
e07785b0 | 5520 | __must_hold(&ctx->completion_lock) |
47f46768 | 5521 | { |
9c8e11b3 PB |
5522 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); |
5523 | struct io_timeout_data *data; | |
47f46768 | 5524 | |
9c8e11b3 PB |
5525 | if (IS_ERR(req)) |
5526 | return PTR_ERR(req); | |
47f46768 | 5527 | |
9c8e11b3 PB |
5528 | req->timeout.off = 0; /* noseq */ |
5529 | data = req->async_data; | |
5530 | list_add_tail(&req->timeout.list, &ctx->timeout_list); | |
5531 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode); | |
5532 | data->timer.function = io_timeout_fn; | |
5533 | hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode); | |
5534 | return 0; | |
47f46768 JA |
5535 | } |
5536 | ||
3529d8c2 JA |
5537 | static int io_timeout_remove_prep(struct io_kiocb *req, |
5538 | const struct io_uring_sqe *sqe) | |
b29472ee | 5539 | { |
9c8e11b3 PB |
5540 | struct io_timeout_rem *tr = &req->timeout_rem; |
5541 | ||
b29472ee JA |
5542 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5543 | return -EINVAL; | |
61710e43 DA |
5544 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5545 | return -EINVAL; | |
9c8e11b3 | 5546 | if (sqe->ioprio || sqe->buf_index || sqe->len) |
b29472ee JA |
5547 | return -EINVAL; |
5548 | ||
9c8e11b3 PB |
5549 | tr->addr = READ_ONCE(sqe->addr); |
5550 | tr->flags = READ_ONCE(sqe->timeout_flags); | |
5551 | if (tr->flags & IORING_TIMEOUT_UPDATE) { | |
5552 | if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS)) | |
5553 | return -EINVAL; | |
5554 | if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2))) | |
5555 | return -EFAULT; | |
5556 | } else if (tr->flags) { | |
5557 | /* timeout removal doesn't support flags */ | |
b29472ee | 5558 | return -EINVAL; |
9c8e11b3 | 5559 | } |
b29472ee | 5560 | |
b29472ee JA |
5561 | return 0; |
5562 | } | |
5563 | ||
8662daec PB |
5564 | static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags) |
5565 | { | |
5566 | return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS | |
5567 | : HRTIMER_MODE_REL; | |
5568 | } | |
5569 | ||
11365043 JA |
5570 | /* |
5571 | * Remove or update an existing timeout command | |
5572 | */ | |
61e98203 | 5573 | static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags) |
11365043 | 5574 | { |
9c8e11b3 | 5575 | struct io_timeout_rem *tr = &req->timeout_rem; |
11365043 | 5576 | struct io_ring_ctx *ctx = req->ctx; |
47f46768 | 5577 | int ret; |
11365043 | 5578 | |
11365043 | 5579 | spin_lock_irq(&ctx->completion_lock); |
8662daec | 5580 | if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) |
9c8e11b3 | 5581 | ret = io_timeout_cancel(ctx, tr->addr); |
8662daec PB |
5582 | else |
5583 | ret = io_timeout_update(ctx, tr->addr, &tr->ts, | |
5584 | io_translate_timeout_mode(tr->flags)); | |
11365043 | 5585 | |
d4d19c19 | 5586 | io_cqring_fill_event(ctx, req->user_data, ret, 0); |
11365043 JA |
5587 | io_commit_cqring(ctx); |
5588 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 5589 | io_cqring_ev_posted(ctx); |
4e88d6e7 | 5590 | if (ret < 0) |
93d2bcd2 | 5591 | req_set_fail(req); |
ec9c02ad | 5592 | io_put_req(req); |
11365043 | 5593 | return 0; |
5262f567 JA |
5594 | } |
5595 | ||
3529d8c2 | 5596 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 5597 | bool is_timeout_link) |
5262f567 | 5598 | { |
ad8a48ac | 5599 | struct io_timeout_data *data; |
a41525ab | 5600 | unsigned flags; |
56080b02 | 5601 | u32 off = READ_ONCE(sqe->off); |
5262f567 | 5602 | |
ad8a48ac | 5603 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 5604 | return -EINVAL; |
ad8a48ac | 5605 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 5606 | return -EINVAL; |
56080b02 | 5607 | if (off && is_timeout_link) |
2d28390a | 5608 | return -EINVAL; |
a41525ab JA |
5609 | flags = READ_ONCE(sqe->timeout_flags); |
5610 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 5611 | return -EINVAL; |
bdf20073 | 5612 | |
bfe68a22 | 5613 | req->timeout.off = off; |
f18ee4cf PB |
5614 | if (unlikely(off && !req->ctx->off_timeout_used)) |
5615 | req->ctx->off_timeout_used = true; | |
26a61679 | 5616 | |
e8c2bc1f | 5617 | if (!req->async_data && io_alloc_async_data(req)) |
26a61679 JA |
5618 | return -ENOMEM; |
5619 | ||
e8c2bc1f | 5620 | data = req->async_data; |
ad8a48ac | 5621 | data->req = req; |
ad8a48ac JA |
5622 | |
5623 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
5624 | return -EFAULT; |
5625 | ||
8662daec | 5626 | data->mode = io_translate_timeout_mode(flags); |
ad8a48ac | 5627 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
2482b58f PB |
5628 | if (is_timeout_link) |
5629 | io_req_track_inflight(req); | |
ad8a48ac JA |
5630 | return 0; |
5631 | } | |
5632 | ||
61e98203 | 5633 | static int io_timeout(struct io_kiocb *req, unsigned int issue_flags) |
ad8a48ac | 5634 | { |
ad8a48ac | 5635 | struct io_ring_ctx *ctx = req->ctx; |
e8c2bc1f | 5636 | struct io_timeout_data *data = req->async_data; |
ad8a48ac | 5637 | struct list_head *entry; |
bfe68a22 | 5638 | u32 tail, off = req->timeout.off; |
ad8a48ac | 5639 | |
733f5c95 | 5640 | spin_lock_irq(&ctx->completion_lock); |
93bd25bb | 5641 | |
5262f567 JA |
5642 | /* |
5643 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
5644 | * timeout event to be satisfied. If it isn't set, then this is |
5645 | * a pure timeout request, sequence isn't used. | |
5262f567 | 5646 | */ |
8eb7e2d0 | 5647 | if (io_is_timeout_noseq(req)) { |
93bd25bb JA |
5648 | entry = ctx->timeout_list.prev; |
5649 | goto add; | |
5650 | } | |
5262f567 | 5651 | |
bfe68a22 PB |
5652 | tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
5653 | req->timeout.target_seq = tail + off; | |
5262f567 | 5654 | |
f010505b MDG |
5655 | /* Update the last seq here in case io_flush_timeouts() hasn't. |
5656 | * This is safe because ->completion_lock is held, and submissions | |
5657 | * and completions are never mixed in the same ->completion_lock section. | |
5658 | */ | |
5659 | ctx->cq_last_tm_flush = tail; | |
5660 | ||
5262f567 JA |
5661 | /* |
5662 | * Insertion sort, ensuring the first entry in the list is always | |
5663 | * the one we need first. | |
5664 | */ | |
5262f567 | 5665 | list_for_each_prev(entry, &ctx->timeout_list) { |
135fcde8 PB |
5666 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, |
5667 | timeout.list); | |
5262f567 | 5668 | |
8eb7e2d0 | 5669 | if (io_is_timeout_noseq(nxt)) |
93bd25bb | 5670 | continue; |
bfe68a22 PB |
5671 | /* nxt.seq is behind @tail, otherwise would've been completed */ |
5672 | if (off >= nxt->timeout.target_seq - tail) | |
5262f567 JA |
5673 | break; |
5674 | } | |
93bd25bb | 5675 | add: |
135fcde8 | 5676 | list_add(&req->timeout.list, entry); |
ad8a48ac JA |
5677 | data->timer.function = io_timeout_fn; |
5678 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 5679 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
5680 | return 0; |
5681 | } | |
5262f567 | 5682 | |
f458dd84 PB |
5683 | struct io_cancel_data { |
5684 | struct io_ring_ctx *ctx; | |
5685 | u64 user_data; | |
5686 | }; | |
5687 | ||
62755e35 JA |
5688 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
5689 | { | |
5690 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
f458dd84 | 5691 | struct io_cancel_data *cd = data; |
62755e35 | 5692 | |
f458dd84 | 5693 | return req->ctx == cd->ctx && req->user_data == cd->user_data; |
62755e35 JA |
5694 | } |
5695 | ||
f458dd84 PB |
5696 | static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data, |
5697 | struct io_ring_ctx *ctx) | |
62755e35 | 5698 | { |
f458dd84 | 5699 | struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, }; |
62755e35 | 5700 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
5701 | int ret = 0; |
5702 | ||
f458dd84 | 5703 | if (!tctx || !tctx->io_wq) |
5aa75ed5 JA |
5704 | return -ENOENT; |
5705 | ||
f458dd84 | 5706 | cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false); |
62755e35 JA |
5707 | switch (cancel_ret) { |
5708 | case IO_WQ_CANCEL_OK: | |
5709 | ret = 0; | |
5710 | break; | |
5711 | case IO_WQ_CANCEL_RUNNING: | |
5712 | ret = -EALREADY; | |
5713 | break; | |
5714 | case IO_WQ_CANCEL_NOTFOUND: | |
5715 | ret = -ENOENT; | |
5716 | break; | |
5717 | } | |
5718 | ||
e977d6d3 JA |
5719 | return ret; |
5720 | } | |
5721 | ||
47f46768 JA |
5722 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
5723 | struct io_kiocb *req, __u64 sqe_addr, | |
014db007 | 5724 | int success_ret) |
47f46768 JA |
5725 | { |
5726 | unsigned long flags; | |
5727 | int ret; | |
5728 | ||
f458dd84 | 5729 | ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx); |
47f46768 | 5730 | spin_lock_irqsave(&ctx->completion_lock, flags); |
df9727af PB |
5731 | if (ret != -ENOENT) |
5732 | goto done; | |
47f46768 JA |
5733 | ret = io_timeout_cancel(ctx, sqe_addr); |
5734 | if (ret != -ENOENT) | |
5735 | goto done; | |
9ba5fac8 | 5736 | ret = io_poll_cancel(ctx, sqe_addr, false); |
47f46768 | 5737 | done: |
b0dd8a41 JA |
5738 | if (!ret) |
5739 | ret = success_ret; | |
d4d19c19 | 5740 | io_cqring_fill_event(ctx, req->user_data, ret, 0); |
47f46768 JA |
5741 | io_commit_cqring(ctx); |
5742 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5743 | io_cqring_ev_posted(ctx); | |
5744 | ||
4e88d6e7 | 5745 | if (ret < 0) |
93d2bcd2 | 5746 | req_set_fail(req); |
47f46768 JA |
5747 | } |
5748 | ||
3529d8c2 JA |
5749 | static int io_async_cancel_prep(struct io_kiocb *req, |
5750 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 5751 | { |
fbf23849 | 5752 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 | 5753 | return -EINVAL; |
61710e43 DA |
5754 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5755 | return -EINVAL; | |
5756 | if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags) | |
e977d6d3 JA |
5757 | return -EINVAL; |
5758 | ||
fbf23849 JA |
5759 | req->cancel.addr = READ_ONCE(sqe->addr); |
5760 | return 0; | |
5761 | } | |
5762 | ||
61e98203 | 5763 | static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags) |
fbf23849 JA |
5764 | { |
5765 | struct io_ring_ctx *ctx = req->ctx; | |
58f99373 PB |
5766 | u64 sqe_addr = req->cancel.addr; |
5767 | struct io_tctx_node *node; | |
5768 | int ret; | |
5769 | ||
5770 | /* tasks should wait for their io-wq threads, so safe w/o sync */ | |
5771 | ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx); | |
5772 | spin_lock_irq(&ctx->completion_lock); | |
5773 | if (ret != -ENOENT) | |
5774 | goto done; | |
5775 | ret = io_timeout_cancel(ctx, sqe_addr); | |
5776 | if (ret != -ENOENT) | |
5777 | goto done; | |
9ba5fac8 | 5778 | ret = io_poll_cancel(ctx, sqe_addr, false); |
58f99373 PB |
5779 | if (ret != -ENOENT) |
5780 | goto done; | |
5781 | spin_unlock_irq(&ctx->completion_lock); | |
5782 | ||
5783 | /* slow path, try all io-wq's */ | |
5784 | io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK)); | |
5785 | ret = -ENOENT; | |
5786 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
5787 | struct io_uring_task *tctx = node->task->io_uring; | |
fbf23849 | 5788 | |
58f99373 PB |
5789 | ret = io_async_cancel_one(tctx, req->cancel.addr, ctx); |
5790 | if (ret != -ENOENT) | |
5791 | break; | |
5792 | } | |
5793 | io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK)); | |
5794 | ||
5795 | spin_lock_irq(&ctx->completion_lock); | |
5796 | done: | |
d4d19c19 | 5797 | io_cqring_fill_event(ctx, req->user_data, ret, 0); |
58f99373 PB |
5798 | io_commit_cqring(ctx); |
5799 | spin_unlock_irq(&ctx->completion_lock); | |
5800 | io_cqring_ev_posted(ctx); | |
5801 | ||
5802 | if (ret < 0) | |
93d2bcd2 | 5803 | req_set_fail(req); |
58f99373 | 5804 | io_put_req(req); |
5262f567 JA |
5805 | return 0; |
5806 | } | |
5807 | ||
269bbe5f | 5808 | static int io_rsrc_update_prep(struct io_kiocb *req, |
05f3fb3c JA |
5809 | const struct io_uring_sqe *sqe) |
5810 | { | |
61710e43 DA |
5811 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5812 | return -EINVAL; | |
5813 | if (sqe->ioprio || sqe->rw_flags) | |
05f3fb3c JA |
5814 | return -EINVAL; |
5815 | ||
269bbe5f BM |
5816 | req->rsrc_update.offset = READ_ONCE(sqe->off); |
5817 | req->rsrc_update.nr_args = READ_ONCE(sqe->len); | |
5818 | if (!req->rsrc_update.nr_args) | |
05f3fb3c | 5819 | return -EINVAL; |
269bbe5f | 5820 | req->rsrc_update.arg = READ_ONCE(sqe->addr); |
05f3fb3c JA |
5821 | return 0; |
5822 | } | |
5823 | ||
889fca73 | 5824 | static int io_files_update(struct io_kiocb *req, unsigned int issue_flags) |
fbf23849 JA |
5825 | { |
5826 | struct io_ring_ctx *ctx = req->ctx; | |
c3bdad02 | 5827 | struct io_uring_rsrc_update2 up; |
05f3fb3c | 5828 | int ret; |
fbf23849 | 5829 | |
45d189c6 | 5830 | if (issue_flags & IO_URING_F_NONBLOCK) |
05f3fb3c | 5831 | return -EAGAIN; |
05f3fb3c | 5832 | |
269bbe5f BM |
5833 | up.offset = req->rsrc_update.offset; |
5834 | up.data = req->rsrc_update.arg; | |
c3bdad02 PB |
5835 | up.nr = 0; |
5836 | up.tags = 0; | |
615cee49 | 5837 | up.resv = 0; |
05f3fb3c JA |
5838 | |
5839 | mutex_lock(&ctx->uring_lock); | |
fdecb662 | 5840 | ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, |
98f0b3b4 | 5841 | &up, req->rsrc_update.nr_args); |
05f3fb3c JA |
5842 | mutex_unlock(&ctx->uring_lock); |
5843 | ||
5844 | if (ret < 0) | |
93d2bcd2 | 5845 | req_set_fail(req); |
889fca73 | 5846 | __io_req_complete(req, issue_flags, ret, 0); |
5262f567 JA |
5847 | return 0; |
5848 | } | |
5849 | ||
bfe76559 | 5850 | static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 5851 | { |
d625c6ee | 5852 | switch (req->opcode) { |
e781573e | 5853 | case IORING_OP_NOP: |
bfe76559 | 5854 | return 0; |
f67676d1 JA |
5855 | case IORING_OP_READV: |
5856 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 5857 | case IORING_OP_READ: |
bfe76559 | 5858 | return io_read_prep(req, sqe); |
f67676d1 JA |
5859 | case IORING_OP_WRITEV: |
5860 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 5861 | case IORING_OP_WRITE: |
bfe76559 | 5862 | return io_write_prep(req, sqe); |
0969e783 | 5863 | case IORING_OP_POLL_ADD: |
bfe76559 | 5864 | return io_poll_add_prep(req, sqe); |
0969e783 | 5865 | case IORING_OP_POLL_REMOVE: |
c5de0036 | 5866 | return io_poll_update_prep(req, sqe); |
8ed8d3c3 | 5867 | case IORING_OP_FSYNC: |
1155c76a | 5868 | return io_fsync_prep(req, sqe); |
8ed8d3c3 | 5869 | case IORING_OP_SYNC_FILE_RANGE: |
1155c76a | 5870 | return io_sfr_prep(req, sqe); |
03b1230c | 5871 | case IORING_OP_SENDMSG: |
fddaface | 5872 | case IORING_OP_SEND: |
bfe76559 | 5873 | return io_sendmsg_prep(req, sqe); |
03b1230c | 5874 | case IORING_OP_RECVMSG: |
fddaface | 5875 | case IORING_OP_RECV: |
bfe76559 | 5876 | return io_recvmsg_prep(req, sqe); |
f499a021 | 5877 | case IORING_OP_CONNECT: |
bfe76559 | 5878 | return io_connect_prep(req, sqe); |
2d28390a | 5879 | case IORING_OP_TIMEOUT: |
bfe76559 | 5880 | return io_timeout_prep(req, sqe, false); |
b29472ee | 5881 | case IORING_OP_TIMEOUT_REMOVE: |
bfe76559 | 5882 | return io_timeout_remove_prep(req, sqe); |
fbf23849 | 5883 | case IORING_OP_ASYNC_CANCEL: |
bfe76559 | 5884 | return io_async_cancel_prep(req, sqe); |
2d28390a | 5885 | case IORING_OP_LINK_TIMEOUT: |
bfe76559 | 5886 | return io_timeout_prep(req, sqe, true); |
8ed8d3c3 | 5887 | case IORING_OP_ACCEPT: |
bfe76559 | 5888 | return io_accept_prep(req, sqe); |
d63d1b5e | 5889 | case IORING_OP_FALLOCATE: |
bfe76559 | 5890 | return io_fallocate_prep(req, sqe); |
15b71abe | 5891 | case IORING_OP_OPENAT: |
bfe76559 | 5892 | return io_openat_prep(req, sqe); |
b5dba59e | 5893 | case IORING_OP_CLOSE: |
bfe76559 | 5894 | return io_close_prep(req, sqe); |
05f3fb3c | 5895 | case IORING_OP_FILES_UPDATE: |
269bbe5f | 5896 | return io_rsrc_update_prep(req, sqe); |
eddc7ef5 | 5897 | case IORING_OP_STATX: |
bfe76559 | 5898 | return io_statx_prep(req, sqe); |
4840e418 | 5899 | case IORING_OP_FADVISE: |
bfe76559 | 5900 | return io_fadvise_prep(req, sqe); |
c1ca757b | 5901 | case IORING_OP_MADVISE: |
bfe76559 | 5902 | return io_madvise_prep(req, sqe); |
cebdb986 | 5903 | case IORING_OP_OPENAT2: |
bfe76559 | 5904 | return io_openat2_prep(req, sqe); |
3e4827b0 | 5905 | case IORING_OP_EPOLL_CTL: |
bfe76559 | 5906 | return io_epoll_ctl_prep(req, sqe); |
7d67af2c | 5907 | case IORING_OP_SPLICE: |
bfe76559 | 5908 | return io_splice_prep(req, sqe); |
ddf0322d | 5909 | case IORING_OP_PROVIDE_BUFFERS: |
bfe76559 | 5910 | return io_provide_buffers_prep(req, sqe); |
067524e9 | 5911 | case IORING_OP_REMOVE_BUFFERS: |
bfe76559 | 5912 | return io_remove_buffers_prep(req, sqe); |
f2a8d5c7 | 5913 | case IORING_OP_TEE: |
bfe76559 | 5914 | return io_tee_prep(req, sqe); |
36f4fa68 JA |
5915 | case IORING_OP_SHUTDOWN: |
5916 | return io_shutdown_prep(req, sqe); | |
80a261fd JA |
5917 | case IORING_OP_RENAMEAT: |
5918 | return io_renameat_prep(req, sqe); | |
14a1143b JA |
5919 | case IORING_OP_UNLINKAT: |
5920 | return io_unlinkat_prep(req, sqe); | |
f67676d1 JA |
5921 | } |
5922 | ||
bfe76559 PB |
5923 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
5924 | req->opcode); | |
bd54b6fe | 5925 | return -EINVAL; |
bfe76559 PB |
5926 | } |
5927 | ||
93642ef8 | 5928 | static int io_req_prep_async(struct io_kiocb *req) |
bfe76559 | 5929 | { |
b7e298d2 PB |
5930 | if (!io_op_defs[req->opcode].needs_async_setup) |
5931 | return 0; | |
5932 | if (WARN_ON_ONCE(req->async_data)) | |
5933 | return -EFAULT; | |
5934 | if (io_alloc_async_data(req)) | |
5935 | return -EAGAIN; | |
5936 | ||
93642ef8 PB |
5937 | switch (req->opcode) { |
5938 | case IORING_OP_READV: | |
93642ef8 PB |
5939 | return io_rw_prep_async(req, READ); |
5940 | case IORING_OP_WRITEV: | |
93642ef8 PB |
5941 | return io_rw_prep_async(req, WRITE); |
5942 | case IORING_OP_SENDMSG: | |
93642ef8 PB |
5943 | return io_sendmsg_prep_async(req); |
5944 | case IORING_OP_RECVMSG: | |
93642ef8 PB |
5945 | return io_recvmsg_prep_async(req); |
5946 | case IORING_OP_CONNECT: | |
5947 | return io_connect_prep_async(req); | |
5948 | } | |
b7e298d2 PB |
5949 | printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n", |
5950 | req->opcode); | |
5951 | return -EFAULT; | |
f67676d1 JA |
5952 | } |
5953 | ||
9cf7c104 PB |
5954 | static u32 io_get_sequence(struct io_kiocb *req) |
5955 | { | |
a3dbdf54 | 5956 | u32 seq = req->ctx->cached_sq_head; |
9cf7c104 | 5957 | |
a3dbdf54 PB |
5958 | /* need original cached_sq_head, but it was increased for each req */ |
5959 | io_for_each_link(req, req) | |
5960 | seq--; | |
5961 | return seq; | |
9cf7c104 PB |
5962 | } |
5963 | ||
76cc33d7 | 5964 | static bool io_drain_req(struct io_kiocb *req) |
de0617e4 | 5965 | { |
3c19966d | 5966 | struct io_kiocb *pos; |
a197f664 | 5967 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 5968 | struct io_defer_entry *de; |
f67676d1 | 5969 | int ret; |
9cf7c104 | 5970 | u32 seq; |
de0617e4 | 5971 | |
3c19966d PB |
5972 | /* |
5973 | * If we need to drain a request in the middle of a link, drain the | |
5974 | * head request and the next request/link after the current link. | |
5975 | * Considering sequential execution of links, IOSQE_IO_DRAIN will be | |
5976 | * maintained for every request of our link. | |
5977 | */ | |
5978 | if (ctx->drain_next) { | |
5979 | req->flags |= REQ_F_IO_DRAIN; | |
5980 | ctx->drain_next = false; | |
5981 | } | |
5982 | /* not interested in head, start from the first linked */ | |
5983 | io_for_each_link(pos, req->link) { | |
5984 | if (pos->flags & REQ_F_IO_DRAIN) { | |
5985 | ctx->drain_next = true; | |
5986 | req->flags |= REQ_F_IO_DRAIN; | |
5987 | break; | |
5988 | } | |
5989 | } | |
5990 | ||
9d858b21 | 5991 | /* Still need defer if there is pending req in defer list. */ |
9cf7c104 | 5992 | if (likely(list_empty_careful(&ctx->defer_list) && |
10c66904 PB |
5993 | !(req->flags & REQ_F_IO_DRAIN))) { |
5994 | ctx->drain_active = false; | |
76cc33d7 | 5995 | return false; |
10c66904 | 5996 | } |
9cf7c104 PB |
5997 | |
5998 | seq = io_get_sequence(req); | |
5999 | /* Still a chance to pass the sequence check */ | |
6000 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) | |
76cc33d7 | 6001 | return false; |
de0617e4 | 6002 | |
b7e298d2 | 6003 | ret = io_req_prep_async(req); |
be7053b7 | 6004 | if (ret) |
1b48773f | 6005 | goto fail; |
cbdcb435 | 6006 | io_prep_async_link(req); |
27dc8338 | 6007 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 6008 | if (!de) { |
1b48773f PB |
6009 | ret = -ENOMEM; |
6010 | fail: | |
6011 | io_req_complete_failed(req, ret); | |
76cc33d7 PB |
6012 | return true; |
6013 | } | |
2d28390a | 6014 | |
de0617e4 | 6015 | spin_lock_irq(&ctx->completion_lock); |
9cf7c104 | 6016 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
de0617e4 | 6017 | spin_unlock_irq(&ctx->completion_lock); |
27dc8338 | 6018 | kfree(de); |
ae34817b | 6019 | io_queue_async_work(req); |
76cc33d7 | 6020 | return true; |
de0617e4 JA |
6021 | } |
6022 | ||
915967f6 | 6023 | trace_io_uring_defer(ctx, req, req->user_data); |
27dc8338 | 6024 | de->req = req; |
9cf7c104 | 6025 | de->seq = seq; |
27dc8338 | 6026 | list_add_tail(&de->list, &ctx->defer_list); |
de0617e4 | 6027 | spin_unlock_irq(&ctx->completion_lock); |
76cc33d7 | 6028 | return true; |
de0617e4 JA |
6029 | } |
6030 | ||
68fb8979 | 6031 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 6032 | { |
0e1b6fe3 PB |
6033 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
6034 | switch (req->opcode) { | |
6035 | case IORING_OP_READV: | |
6036 | case IORING_OP_READ_FIXED: | |
6037 | case IORING_OP_READ: | |
bcda7baa | 6038 | kfree((void *)(unsigned long)req->rw.addr); |
0e1b6fe3 PB |
6039 | break; |
6040 | case IORING_OP_RECVMSG: | |
6041 | case IORING_OP_RECV: | |
bcda7baa | 6042 | kfree(req->sr_msg.kbuf); |
0e1b6fe3 PB |
6043 | break; |
6044 | } | |
99bc4c38 PB |
6045 | } |
6046 | ||
0e1b6fe3 PB |
6047 | if (req->flags & REQ_F_NEED_CLEANUP) { |
6048 | switch (req->opcode) { | |
6049 | case IORING_OP_READV: | |
6050 | case IORING_OP_READ_FIXED: | |
6051 | case IORING_OP_READ: | |
6052 | case IORING_OP_WRITEV: | |
6053 | case IORING_OP_WRITE_FIXED: | |
e8c2bc1f JA |
6054 | case IORING_OP_WRITE: { |
6055 | struct io_async_rw *io = req->async_data; | |
1dacb4df PB |
6056 | |
6057 | kfree(io->free_iovec); | |
0e1b6fe3 | 6058 | break; |
e8c2bc1f | 6059 | } |
0e1b6fe3 | 6060 | case IORING_OP_RECVMSG: |
e8c2bc1f JA |
6061 | case IORING_OP_SENDMSG: { |
6062 | struct io_async_msghdr *io = req->async_data; | |
257e84a5 PB |
6063 | |
6064 | kfree(io->free_iov); | |
0e1b6fe3 | 6065 | break; |
e8c2bc1f | 6066 | } |
0e1b6fe3 PB |
6067 | case IORING_OP_SPLICE: |
6068 | case IORING_OP_TEE: | |
e1d767f0 PB |
6069 | if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED)) |
6070 | io_put_file(req->splice.file_in); | |
0e1b6fe3 | 6071 | break; |
f3cd4850 JA |
6072 | case IORING_OP_OPENAT: |
6073 | case IORING_OP_OPENAT2: | |
6074 | if (req->open.filename) | |
6075 | putname(req->open.filename); | |
6076 | break; | |
80a261fd JA |
6077 | case IORING_OP_RENAMEAT: |
6078 | putname(req->rename.oldpath); | |
6079 | putname(req->rename.newpath); | |
6080 | break; | |
14a1143b JA |
6081 | case IORING_OP_UNLINKAT: |
6082 | putname(req->unlink.filename); | |
6083 | break; | |
0e1b6fe3 | 6084 | } |
99bc4c38 | 6085 | } |
75652a30 JA |
6086 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
6087 | kfree(req->apoll->double_poll); | |
6088 | kfree(req->apoll); | |
6089 | req->apoll = NULL; | |
6090 | } | |
3a0a6902 PB |
6091 | if (req->flags & REQ_F_INFLIGHT) { |
6092 | struct io_uring_task *tctx = req->task->io_uring; | |
6093 | ||
6094 | atomic_dec(&tctx->inflight_tracked); | |
3a0a6902 | 6095 | } |
c854357b | 6096 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 6097 | put_cred(req->creds); |
c854357b PB |
6098 | |
6099 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
99bc4c38 PB |
6100 | } |
6101 | ||
889fca73 | 6102 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 6103 | { |
a197f664 | 6104 | struct io_ring_ctx *ctx = req->ctx; |
5730b27e | 6105 | const struct cred *creds = NULL; |
d625c6ee | 6106 | int ret; |
2b188cc1 | 6107 | |
b8e64b53 | 6108 | if ((req->flags & REQ_F_CREDS) && req->creds != current_cred()) |
c10d1f98 | 6109 | creds = override_creds(req->creds); |
5730b27e | 6110 | |
d625c6ee | 6111 | switch (req->opcode) { |
2b188cc1 | 6112 | case IORING_OP_NOP: |
889fca73 | 6113 | ret = io_nop(req, issue_flags); |
2b188cc1 JA |
6114 | break; |
6115 | case IORING_OP_READV: | |
edafccee | 6116 | case IORING_OP_READ_FIXED: |
3a6820f2 | 6117 | case IORING_OP_READ: |
889fca73 | 6118 | ret = io_read(req, issue_flags); |
edafccee | 6119 | break; |
3529d8c2 | 6120 | case IORING_OP_WRITEV: |
edafccee | 6121 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 6122 | case IORING_OP_WRITE: |
889fca73 | 6123 | ret = io_write(req, issue_flags); |
2b188cc1 | 6124 | break; |
c992fe29 | 6125 | case IORING_OP_FSYNC: |
45d189c6 | 6126 | ret = io_fsync(req, issue_flags); |
c992fe29 | 6127 | break; |
221c5eb2 | 6128 | case IORING_OP_POLL_ADD: |
61e98203 | 6129 | ret = io_poll_add(req, issue_flags); |
221c5eb2 JA |
6130 | break; |
6131 | case IORING_OP_POLL_REMOVE: | |
c5de0036 | 6132 | ret = io_poll_update(req, issue_flags); |
221c5eb2 | 6133 | break; |
5d17b4a4 | 6134 | case IORING_OP_SYNC_FILE_RANGE: |
45d189c6 | 6135 | ret = io_sync_file_range(req, issue_flags); |
5d17b4a4 | 6136 | break; |
0fa03c62 | 6137 | case IORING_OP_SENDMSG: |
889fca73 | 6138 | ret = io_sendmsg(req, issue_flags); |
062d04d7 | 6139 | break; |
fddaface | 6140 | case IORING_OP_SEND: |
889fca73 | 6141 | ret = io_send(req, issue_flags); |
0fa03c62 | 6142 | break; |
aa1fa28f | 6143 | case IORING_OP_RECVMSG: |
889fca73 | 6144 | ret = io_recvmsg(req, issue_flags); |
062d04d7 | 6145 | break; |
fddaface | 6146 | case IORING_OP_RECV: |
889fca73 | 6147 | ret = io_recv(req, issue_flags); |
aa1fa28f | 6148 | break; |
5262f567 | 6149 | case IORING_OP_TIMEOUT: |
61e98203 | 6150 | ret = io_timeout(req, issue_flags); |
5262f567 | 6151 | break; |
11365043 | 6152 | case IORING_OP_TIMEOUT_REMOVE: |
61e98203 | 6153 | ret = io_timeout_remove(req, issue_flags); |
11365043 | 6154 | break; |
17f2fe35 | 6155 | case IORING_OP_ACCEPT: |
889fca73 | 6156 | ret = io_accept(req, issue_flags); |
17f2fe35 | 6157 | break; |
f8e85cf2 | 6158 | case IORING_OP_CONNECT: |
889fca73 | 6159 | ret = io_connect(req, issue_flags); |
f8e85cf2 | 6160 | break; |
62755e35 | 6161 | case IORING_OP_ASYNC_CANCEL: |
61e98203 | 6162 | ret = io_async_cancel(req, issue_flags); |
62755e35 | 6163 | break; |
d63d1b5e | 6164 | case IORING_OP_FALLOCATE: |
45d189c6 | 6165 | ret = io_fallocate(req, issue_flags); |
d63d1b5e | 6166 | break; |
15b71abe | 6167 | case IORING_OP_OPENAT: |
45d189c6 | 6168 | ret = io_openat(req, issue_flags); |
15b71abe | 6169 | break; |
b5dba59e | 6170 | case IORING_OP_CLOSE: |
889fca73 | 6171 | ret = io_close(req, issue_flags); |
b5dba59e | 6172 | break; |
05f3fb3c | 6173 | case IORING_OP_FILES_UPDATE: |
889fca73 | 6174 | ret = io_files_update(req, issue_flags); |
05f3fb3c | 6175 | break; |
eddc7ef5 | 6176 | case IORING_OP_STATX: |
45d189c6 | 6177 | ret = io_statx(req, issue_flags); |
eddc7ef5 | 6178 | break; |
4840e418 | 6179 | case IORING_OP_FADVISE: |
45d189c6 | 6180 | ret = io_fadvise(req, issue_flags); |
4840e418 | 6181 | break; |
c1ca757b | 6182 | case IORING_OP_MADVISE: |
45d189c6 | 6183 | ret = io_madvise(req, issue_flags); |
c1ca757b | 6184 | break; |
cebdb986 | 6185 | case IORING_OP_OPENAT2: |
45d189c6 | 6186 | ret = io_openat2(req, issue_flags); |
cebdb986 | 6187 | break; |
3e4827b0 | 6188 | case IORING_OP_EPOLL_CTL: |
889fca73 | 6189 | ret = io_epoll_ctl(req, issue_flags); |
3e4827b0 | 6190 | break; |
7d67af2c | 6191 | case IORING_OP_SPLICE: |
45d189c6 | 6192 | ret = io_splice(req, issue_flags); |
7d67af2c | 6193 | break; |
ddf0322d | 6194 | case IORING_OP_PROVIDE_BUFFERS: |
889fca73 | 6195 | ret = io_provide_buffers(req, issue_flags); |
ddf0322d | 6196 | break; |
067524e9 | 6197 | case IORING_OP_REMOVE_BUFFERS: |
889fca73 | 6198 | ret = io_remove_buffers(req, issue_flags); |
3e4827b0 | 6199 | break; |
f2a8d5c7 | 6200 | case IORING_OP_TEE: |
45d189c6 | 6201 | ret = io_tee(req, issue_flags); |
f2a8d5c7 | 6202 | break; |
36f4fa68 | 6203 | case IORING_OP_SHUTDOWN: |
45d189c6 | 6204 | ret = io_shutdown(req, issue_flags); |
36f4fa68 | 6205 | break; |
80a261fd | 6206 | case IORING_OP_RENAMEAT: |
45d189c6 | 6207 | ret = io_renameat(req, issue_flags); |
80a261fd | 6208 | break; |
14a1143b | 6209 | case IORING_OP_UNLINKAT: |
45d189c6 | 6210 | ret = io_unlinkat(req, issue_flags); |
14a1143b | 6211 | break; |
2b188cc1 JA |
6212 | default: |
6213 | ret = -EINVAL; | |
6214 | break; | |
6215 | } | |
6216 | ||
5730b27e JA |
6217 | if (creds) |
6218 | revert_creds(creds); | |
def596e9 JA |
6219 | if (ret) |
6220 | return ret; | |
b532576e | 6221 | /* If the op doesn't have a file, we're not polling for it */ |
cb3d8972 PB |
6222 | if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) |
6223 | io_iopoll_req_issued(req); | |
def596e9 JA |
6224 | |
6225 | return 0; | |
2b188cc1 JA |
6226 | } |
6227 | ||
ebc11b6c PB |
6228 | static struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
6229 | { | |
6230 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6231 | ||
6232 | req = io_put_req_find_next(req); | |
6233 | return req ? &req->work : NULL; | |
6234 | } | |
6235 | ||
5280f7e5 | 6236 | static void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
6237 | { |
6238 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6df1db6b | 6239 | struct io_kiocb *timeout; |
561fb04a | 6240 | int ret = 0; |
2b188cc1 | 6241 | |
6df1db6b PB |
6242 | timeout = io_prep_linked_timeout(req); |
6243 | if (timeout) | |
6244 | io_queue_linked_timeout(timeout); | |
d4c81f38 | 6245 | |
4014d943 | 6246 | if (work->flags & IO_WQ_WORK_CANCEL) |
561fb04a | 6247 | ret = -ECANCELED; |
31b51510 | 6248 | |
561fb04a | 6249 | if (!ret) { |
561fb04a | 6250 | do { |
889fca73 | 6251 | ret = io_issue_sqe(req, 0); |
561fb04a JA |
6252 | /* |
6253 | * We can get EAGAIN for polled IO even though we're | |
6254 | * forcing a sync submission from here, since we can't | |
6255 | * wait for request slots on the block side. | |
6256 | */ | |
6257 | if (ret != -EAGAIN) | |
6258 | break; | |
6259 | cond_resched(); | |
6260 | } while (1); | |
6261 | } | |
31b51510 | 6262 | |
a3df7698 | 6263 | /* avoid locking problems by failing it from a clean context */ |
561fb04a | 6264 | if (ret) { |
a3df7698 | 6265 | /* io-wq is going to take one down */ |
de9b4cca | 6266 | req_ref_get(req); |
a3df7698 | 6267 | io_req_task_queue_fail(req, ret); |
edafccee | 6268 | } |
2b188cc1 JA |
6269 | } |
6270 | ||
aeca241b | 6271 | static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table, |
042b0d85 | 6272 | unsigned i) |
65e19f54 | 6273 | { |
042b0d85 | 6274 | return &table->files[i]; |
dafecf19 PB |
6275 | } |
6276 | ||
65e19f54 JA |
6277 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
6278 | int index) | |
6279 | { | |
aeca241b | 6280 | struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index); |
65e19f54 | 6281 | |
a04b0ac0 | 6282 | return (struct file *) (slot->file_ptr & FFS_MASK); |
65e19f54 JA |
6283 | } |
6284 | ||
a04b0ac0 | 6285 | static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file) |
9a321c98 PB |
6286 | { |
6287 | unsigned long file_ptr = (unsigned long) file; | |
6288 | ||
b191e2df | 6289 | if (__io_file_supports_nowait(file, READ)) |
9a321c98 | 6290 | file_ptr |= FFS_ASYNC_READ; |
b191e2df | 6291 | if (__io_file_supports_nowait(file, WRITE)) |
9a321c98 PB |
6292 | file_ptr |= FFS_ASYNC_WRITE; |
6293 | if (S_ISREG(file_inode(file)->i_mode)) | |
6294 | file_ptr |= FFS_ISREG; | |
a04b0ac0 | 6295 | file_slot->file_ptr = file_ptr; |
65e19f54 JA |
6296 | } |
6297 | ||
ac177053 PB |
6298 | static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx, |
6299 | struct io_kiocb *req, int fd) | |
09bb8394 | 6300 | { |
8da11c19 | 6301 | struct file *file; |
ac177053 | 6302 | unsigned long file_ptr; |
09bb8394 | 6303 | |
ac177053 PB |
6304 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
6305 | return NULL; | |
6306 | fd = array_index_nospec(fd, ctx->nr_user_files); | |
6307 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
6308 | file = (struct file *) (file_ptr & FFS_MASK); | |
6309 | file_ptr &= ~FFS_MASK; | |
6310 | /* mask in overlapping REQ_F and FFS bits */ | |
b191e2df | 6311 | req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT); |
ac177053 PB |
6312 | io_req_set_rsrc_node(req); |
6313 | return file; | |
6314 | } | |
d44f554e | 6315 | |
ac177053 | 6316 | static struct file *io_file_get_normal(struct io_ring_ctx *ctx, |
ac177053 PB |
6317 | struct io_kiocb *req, int fd) |
6318 | { | |
62906e89 | 6319 | struct file *file = fget(fd); |
ac177053 PB |
6320 | |
6321 | trace_io_uring_file_get(ctx, fd); | |
09bb8394 | 6322 | |
ac177053 PB |
6323 | /* we don't allow fixed io_uring files */ |
6324 | if (file && unlikely(file->f_op == &io_uring_fops)) | |
6325 | io_req_track_inflight(req); | |
8371adf5 | 6326 | return file; |
09bb8394 JA |
6327 | } |
6328 | ||
ac177053 | 6329 | static inline struct file *io_file_get(struct io_ring_ctx *ctx, |
ac177053 PB |
6330 | struct io_kiocb *req, int fd, bool fixed) |
6331 | { | |
6332 | if (fixed) | |
6333 | return io_file_get_fixed(ctx, req, fd); | |
6334 | else | |
62906e89 | 6335 | return io_file_get_normal(ctx, req, fd); |
ac177053 PB |
6336 | } |
6337 | ||
2665abfd | 6338 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 6339 | { |
ad8a48ac JA |
6340 | struct io_timeout_data *data = container_of(timer, |
6341 | struct io_timeout_data, timer); | |
90cd7e42 | 6342 | struct io_kiocb *prev, *req = data->req; |
2665abfd | 6343 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 6344 | unsigned long flags; |
2665abfd JA |
6345 | |
6346 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
90cd7e42 PB |
6347 | prev = req->timeout.head; |
6348 | req->timeout.head = NULL; | |
2665abfd JA |
6349 | |
6350 | /* | |
6351 | * We don't expect the list to be empty, that will only happen if we | |
6352 | * race with the completion of the linked work. | |
6353 | */ | |
447c19f3 | 6354 | if (prev) { |
f2f87370 | 6355 | io_remove_next_linked(prev); |
447c19f3 PB |
6356 | if (!req_ref_inc_not_zero(prev)) |
6357 | prev = NULL; | |
6358 | } | |
2665abfd JA |
6359 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
6360 | ||
6361 | if (prev) { | |
014db007 | 6362 | io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME); |
9ae1f8dd | 6363 | io_put_req_deferred(prev, 1); |
a298232e | 6364 | io_put_req_deferred(req, 1); |
47f46768 | 6365 | } else { |
9ae1f8dd | 6366 | io_req_complete_post(req, -ETIME, 0); |
2665abfd | 6367 | } |
2665abfd JA |
6368 | return HRTIMER_NORESTART; |
6369 | } | |
6370 | ||
de968c18 | 6371 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 6372 | { |
de968c18 PB |
6373 | struct io_ring_ctx *ctx = req->ctx; |
6374 | ||
6375 | spin_lock_irq(&ctx->completion_lock); | |
76a46e06 | 6376 | /* |
f2f87370 PB |
6377 | * If the back reference is NULL, then our linked request finished |
6378 | * before we got a chance to setup the timer | |
76a46e06 | 6379 | */ |
90cd7e42 | 6380 | if (req->timeout.head) { |
e8c2bc1f | 6381 | struct io_timeout_data *data = req->async_data; |
94ae5e77 | 6382 | |
ad8a48ac JA |
6383 | data->timer.function = io_link_timeout_fn; |
6384 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
6385 | data->mode); | |
2665abfd | 6386 | } |
76a46e06 | 6387 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 6388 | /* drop submission reference */ |
76a46e06 JA |
6389 | io_put_req(req); |
6390 | } | |
2665abfd | 6391 | |
ad8a48ac | 6392 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd | 6393 | { |
f2f87370 | 6394 | struct io_kiocb *nxt = req->link; |
2665abfd | 6395 | |
f2f87370 PB |
6396 | if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) || |
6397 | nxt->opcode != IORING_OP_LINK_TIMEOUT) | |
76a46e06 | 6398 | return NULL; |
2665abfd | 6399 | |
90cd7e42 | 6400 | nxt->timeout.head = req; |
900fad45 | 6401 | nxt->flags |= REQ_F_LTIMEOUT_ACTIVE; |
76a46e06 | 6402 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 6403 | return nxt; |
2665abfd JA |
6404 | } |
6405 | ||
c5eef2b9 | 6406 | static void __io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 6407 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 6408 | { |
d3d7298d | 6409 | struct io_kiocb *linked_timeout = io_prep_linked_timeout(req); |
e0c5c576 | 6410 | int ret; |
2b188cc1 | 6411 | |
59b735ae | 6412 | issue_sqe: |
c5eef2b9 | 6413 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 6414 | |
491381ce JA |
6415 | /* |
6416 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
6417 | * doesn't support non-blocking read/write attempts | |
6418 | */ | |
1840038e | 6419 | if (likely(!ret)) { |
0d63c148 | 6420 | /* drop submission reference */ |
e342c807 | 6421 | if (req->flags & REQ_F_COMPLETE_INLINE) { |
c5eef2b9 | 6422 | struct io_ring_ctx *ctx = req->ctx; |
cd0ca2e0 | 6423 | struct io_submit_state *state = &ctx->submit_state; |
e65ef56d | 6424 | |
cd0ca2e0 PB |
6425 | state->compl_reqs[state->compl_nr++] = req; |
6426 | if (state->compl_nr == ARRAY_SIZE(state->compl_reqs)) | |
2a2758f2 | 6427 | io_submit_flush_completions(ctx); |
9affd664 | 6428 | } else { |
d3d7298d | 6429 | io_put_req(req); |
0d63c148 | 6430 | } |
1840038e | 6431 | } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
59b735ae OL |
6432 | switch (io_arm_poll_handler(req)) { |
6433 | case IO_APOLL_READY: | |
6434 | goto issue_sqe; | |
6435 | case IO_APOLL_ABORTED: | |
1840038e PB |
6436 | /* |
6437 | * Queued up for async execution, worker will release | |
6438 | * submit reference when the iocb is actually submitted. | |
6439 | */ | |
6440 | io_queue_async_work(req); | |
59b735ae | 6441 | break; |
1840038e | 6442 | } |
0d63c148 | 6443 | } else { |
f41db273 | 6444 | io_req_complete_failed(req, ret); |
9e645e11 | 6445 | } |
d3d7298d PB |
6446 | if (linked_timeout) |
6447 | io_queue_linked_timeout(linked_timeout); | |
2b188cc1 JA |
6448 | } |
6449 | ||
441b8a78 | 6450 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 6451 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 6452 | { |
10c66904 | 6453 | if (unlikely(req->ctx->drain_active) && io_drain_req(req)) |
76cc33d7 | 6454 | return; |
4fe2c963 | 6455 | |
76cc33d7 | 6456 | if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) { |
c5eef2b9 | 6457 | __io_queue_sqe(req); |
76cc33d7 PB |
6458 | } else { |
6459 | int ret = io_req_prep_async(req); | |
6460 | ||
6461 | if (unlikely(ret)) | |
6462 | io_req_complete_failed(req, ret); | |
6463 | else | |
6464 | io_queue_async_work(req); | |
ce35a47a | 6465 | } |
4fe2c963 JL |
6466 | } |
6467 | ||
b16fed66 PB |
6468 | /* |
6469 | * Check SQE restrictions (opcode and flags). | |
6470 | * | |
6471 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
6472 | */ | |
6473 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
6474 | struct io_kiocb *req, | |
6475 | unsigned int sqe_flags) | |
4fe2c963 | 6476 | { |
4cfb25bf | 6477 | if (likely(!ctx->restricted)) |
b16fed66 PB |
6478 | return true; |
6479 | ||
6480 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) | |
6481 | return false; | |
6482 | ||
6483 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
6484 | ctx->restrictions.sqe_flags_required) | |
6485 | return false; | |
6486 | ||
6487 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
6488 | ctx->restrictions.sqe_flags_required)) | |
6489 | return false; | |
6490 | ||
6491 | return true; | |
4fe2c963 JL |
6492 | } |
6493 | ||
b16fed66 PB |
6494 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
6495 | const struct io_uring_sqe *sqe) | |
282cdc86 | 6496 | __must_hold(&ctx->uring_lock) |
b16fed66 PB |
6497 | { |
6498 | struct io_submit_state *state; | |
6499 | unsigned int sqe_flags; | |
003e8dcc | 6500 | int personality, ret = 0; |
b16fed66 | 6501 | |
864ea921 | 6502 | /* req is partially pre-initialised, see io_preinit_req() */ |
b16fed66 PB |
6503 | req->opcode = READ_ONCE(sqe->opcode); |
6504 | /* same numerical values with corresponding REQ_F_*, safe to copy */ | |
6505 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
6506 | req->user_data = READ_ONCE(sqe->user_data); | |
b16fed66 | 6507 | req->file = NULL; |
b16fed66 PB |
6508 | req->fixed_rsrc_refs = NULL; |
6509 | /* one is dropped after submission, the other at completion */ | |
abc54d63 | 6510 | atomic_set(&req->refs, 2); |
b16fed66 | 6511 | req->task = current; |
b16fed66 PB |
6512 | |
6513 | /* enforce forwards compatibility on users */ | |
dddca226 | 6514 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) |
b16fed66 | 6515 | return -EINVAL; |
b16fed66 PB |
6516 | if (unlikely(req->opcode >= IORING_OP_LAST)) |
6517 | return -EINVAL; | |
4cfb25bf | 6518 | if (!io_check_restriction(ctx, req, sqe_flags)) |
b16fed66 PB |
6519 | return -EACCES; |
6520 | ||
6521 | if ((sqe_flags & IOSQE_BUFFER_SELECT) && | |
6522 | !io_op_defs[req->opcode].buffer_select) | |
6523 | return -EOPNOTSUPP; | |
3c19966d PB |
6524 | if (unlikely(sqe_flags & IOSQE_IO_DRAIN)) |
6525 | ctx->drain_active = true; | |
863e0560 | 6526 | |
003e8dcc JA |
6527 | personality = READ_ONCE(sqe->personality); |
6528 | if (personality) { | |
c10d1f98 PB |
6529 | req->creds = xa_load(&ctx->personalities, personality); |
6530 | if (!req->creds) | |
003e8dcc | 6531 | return -EINVAL; |
c10d1f98 | 6532 | get_cred(req->creds); |
b8e64b53 | 6533 | req->flags |= REQ_F_CREDS; |
003e8dcc | 6534 | } |
b16fed66 PB |
6535 | state = &ctx->submit_state; |
6536 | ||
6537 | /* | |
6538 | * Plug now if we have more than 1 IO left after this, and the target | |
6539 | * is potentially a read/write to block based storage. | |
6540 | */ | |
6541 | if (!state->plug_started && state->ios_left > 1 && | |
6542 | io_op_defs[req->opcode].plug) { | |
6543 | blk_start_plug(&state->plug); | |
6544 | state->plug_started = true; | |
6545 | } | |
6546 | ||
6547 | if (io_op_defs[req->opcode].needs_file) { | |
62906e89 | 6548 | req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd), |
ac177053 | 6549 | (sqe_flags & IOSQE_FIXED_FILE)); |
b16fed66 PB |
6550 | if (unlikely(!req->file)) |
6551 | ret = -EBADF; | |
6552 | } | |
6553 | ||
6554 | state->ios_left--; | |
6555 | return ret; | |
6556 | } | |
6557 | ||
a6b8cadc | 6558 | static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
a1ab7b35 | 6559 | const struct io_uring_sqe *sqe) |
282cdc86 | 6560 | __must_hold(&ctx->uring_lock) |
9e645e11 | 6561 | { |
a1ab7b35 | 6562 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 6563 | int ret; |
9e645e11 | 6564 | |
a6b8cadc PB |
6565 | ret = io_init_req(ctx, req, sqe); |
6566 | if (unlikely(ret)) { | |
6567 | fail_req: | |
de59bc10 PB |
6568 | if (link->head) { |
6569 | /* fail even hard links since we don't submit */ | |
93d2bcd2 | 6570 | req_set_fail(link->head); |
f41db273 | 6571 | io_req_complete_failed(link->head, -ECANCELED); |
de59bc10 PB |
6572 | link->head = NULL; |
6573 | } | |
f41db273 | 6574 | io_req_complete_failed(req, ret); |
a6b8cadc PB |
6575 | return ret; |
6576 | } | |
441b8a78 | 6577 | |
be7053b7 PB |
6578 | ret = io_req_prep(req, sqe); |
6579 | if (unlikely(ret)) | |
6580 | goto fail_req; | |
a6b8cadc | 6581 | |
be7053b7 | 6582 | /* don't need @sqe from now on */ |
236daeae OL |
6583 | trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data, |
6584 | req->flags, true, | |
6585 | ctx->flags & IORING_SETUP_SQPOLL); | |
a6b8cadc | 6586 | |
9e645e11 JA |
6587 | /* |
6588 | * If we already have a head request, queue this one for async | |
6589 | * submittal once the head completes. If we don't have a head but | |
6590 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
6591 | * submitted sync once the chain is complete. If none of those | |
6592 | * conditions are true (normal request), then just queue it. | |
6593 | */ | |
863e0560 PB |
6594 | if (link->head) { |
6595 | struct io_kiocb *head = link->head; | |
4e88d6e7 | 6596 | |
b7e298d2 | 6597 | ret = io_req_prep_async(req); |
cf109604 | 6598 | if (unlikely(ret)) |
a6b8cadc | 6599 | goto fail_req; |
9d76377f | 6600 | trace_io_uring_link(ctx, req, head); |
f2f87370 | 6601 | link->last->link = req; |
863e0560 | 6602 | link->last = req; |
32fe525b PB |
6603 | |
6604 | /* last request of a link, enqueue the link */ | |
ef4ff581 | 6605 | if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) { |
863e0560 | 6606 | link->head = NULL; |
5e159204 | 6607 | io_queue_sqe(head); |
32fe525b | 6608 | } |
9e645e11 | 6609 | } else { |
ef4ff581 | 6610 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
863e0560 PB |
6611 | link->head = req; |
6612 | link->last = req; | |
711be031 | 6613 | } else { |
be7053b7 | 6614 | io_queue_sqe(req); |
711be031 | 6615 | } |
9e645e11 | 6616 | } |
2e6e1fde | 6617 | |
1d4240cc | 6618 | return 0; |
9e645e11 JA |
6619 | } |
6620 | ||
9a56a232 JA |
6621 | /* |
6622 | * Batched submission is done, ensure local IO is flushed out. | |
6623 | */ | |
ba88ff11 PB |
6624 | static void io_submit_state_end(struct io_submit_state *state, |
6625 | struct io_ring_ctx *ctx) | |
9a56a232 | 6626 | { |
a1ab7b35 | 6627 | if (state->link.head) |
de59bc10 | 6628 | io_queue_sqe(state->link.head); |
cd0ca2e0 | 6629 | if (state->compl_nr) |
2a2758f2 | 6630 | io_submit_flush_completions(ctx); |
27926b68 JA |
6631 | if (state->plug_started) |
6632 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
6633 | } |
6634 | ||
6635 | /* | |
6636 | * Start submission side cache. | |
6637 | */ | |
6638 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 6639 | unsigned int max_ios) |
9a56a232 | 6640 | { |
27926b68 | 6641 | state->plug_started = false; |
9a56a232 | 6642 | state->ios_left = max_ios; |
a1ab7b35 PB |
6643 | /* set only head, no need to init link_last in advance */ |
6644 | state->link.head = NULL; | |
9a56a232 JA |
6645 | } |
6646 | ||
2b188cc1 JA |
6647 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
6648 | { | |
75b28aff | 6649 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 6650 | |
caf582c6 PB |
6651 | /* |
6652 | * Ensure any loads from the SQEs are done at this point, | |
6653 | * since once we write the new head, the application could | |
6654 | * write new data to them. | |
6655 | */ | |
6656 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
6657 | } |
6658 | ||
2b188cc1 | 6659 | /* |
dd9ae8a0 | 6660 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
6661 | * that is mapped by userspace. This means that care needs to be taken to |
6662 | * ensure that reads are stable, as we cannot rely on userspace always | |
6663 | * being a good citizen. If members of the sqe are validated and then later | |
6664 | * used, it's important that those reads are done through READ_ONCE() to | |
6665 | * prevent a re-load down the line. | |
6666 | */ | |
709b302f | 6667 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 6668 | { |
ea5ab3b5 | 6669 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 6670 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
6671 | |
6672 | /* | |
6673 | * The cached sq head (or cq tail) serves two purposes: | |
6674 | * | |
6675 | * 1) allows us to batch the cost of updating the user visible | |
6676 | * head updates. | |
6677 | * 2) allows the kernel side to track the head on its own, even | |
6678 | * though the application is the one updating it. | |
6679 | */ | |
17d3aeb3 | 6680 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
709b302f PB |
6681 | if (likely(head < ctx->sq_entries)) |
6682 | return &ctx->sq_sqes[head]; | |
2b188cc1 JA |
6683 | |
6684 | /* drop invalid entries */ | |
15641e42 PB |
6685 | ctx->cq_extra--; |
6686 | WRITE_ONCE(ctx->rings->sq_dropped, | |
6687 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
6688 | return NULL; |
6689 | } | |
6690 | ||
0f212204 | 6691 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 6692 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 6693 | { |
09899b19 | 6694 | struct io_uring_task *tctx; |
46c4e16a | 6695 | int submitted = 0; |
6c271ce2 | 6696 | |
ee7d46d9 PB |
6697 | /* make sure SQ entry isn't read before tail */ |
6698 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
2b85edfc PB |
6699 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
6700 | return -EAGAIN; | |
6c271ce2 | 6701 | |
09899b19 PB |
6702 | tctx = current->io_uring; |
6703 | tctx->cached_refs -= nr; | |
6704 | if (unlikely(tctx->cached_refs < 0)) { | |
6705 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
6706 | ||
6707 | percpu_counter_add(&tctx->inflight, refill); | |
6708 | refcount_add(refill, ¤t->usage); | |
6709 | tctx->cached_refs += refill; | |
6710 | } | |
ba88ff11 | 6711 | io_submit_state_start(&ctx->submit_state, nr); |
b14cca0c | 6712 | |
46c4e16a | 6713 | while (submitted < nr) { |
3529d8c2 | 6714 | const struct io_uring_sqe *sqe; |
196be95c | 6715 | struct io_kiocb *req; |
fb5ccc98 | 6716 | |
258b29a9 | 6717 | req = io_alloc_req(ctx); |
196be95c PB |
6718 | if (unlikely(!req)) { |
6719 | if (!submitted) | |
6720 | submitted = -EAGAIN; | |
fb5ccc98 | 6721 | break; |
196be95c | 6722 | } |
4fccfcbb PB |
6723 | sqe = io_get_sqe(ctx); |
6724 | if (unlikely(!sqe)) { | |
6725 | kmem_cache_free(req_cachep, req); | |
6726 | break; | |
6727 | } | |
d3656344 JA |
6728 | /* will complete beyond this point, count as submitted */ |
6729 | submitted++; | |
a1ab7b35 | 6730 | if (io_submit_sqe(ctx, req, sqe)) |
196be95c | 6731 | break; |
6c271ce2 JA |
6732 | } |
6733 | ||
9466f437 PB |
6734 | if (unlikely(submitted != nr)) { |
6735 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
d8a6df10 | 6736 | int unused = nr - ref_used; |
9466f437 | 6737 | |
09899b19 | 6738 | current->io_uring->cached_refs += unused; |
d8a6df10 | 6739 | percpu_ref_put_many(&ctx->refs, unused); |
9466f437 | 6740 | } |
6c271ce2 | 6741 | |
a1ab7b35 | 6742 | io_submit_state_end(&ctx->submit_state, ctx); |
ae9428ca PB |
6743 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
6744 | io_commit_sqring(ctx); | |
6745 | ||
6c271ce2 JA |
6746 | return submitted; |
6747 | } | |
6748 | ||
e4b6d902 PB |
6749 | static inline bool io_sqd_events_pending(struct io_sq_data *sqd) |
6750 | { | |
6751 | return READ_ONCE(sqd->state); | |
6752 | } | |
6753 | ||
23b3628e XW |
6754 | static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx) |
6755 | { | |
6756 | /* Tell userspace we may need a wakeup call */ | |
6757 | spin_lock_irq(&ctx->completion_lock); | |
20c0b380 NA |
6758 | WRITE_ONCE(ctx->rings->sq_flags, |
6759 | ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP); | |
23b3628e XW |
6760 | spin_unlock_irq(&ctx->completion_lock); |
6761 | } | |
6762 | ||
6763 | static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx) | |
6764 | { | |
6765 | spin_lock_irq(&ctx->completion_lock); | |
20c0b380 NA |
6766 | WRITE_ONCE(ctx->rings->sq_flags, |
6767 | ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP); | |
23b3628e XW |
6768 | spin_unlock_irq(&ctx->completion_lock); |
6769 | } | |
6770 | ||
08369246 | 6771 | static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) |
6c271ce2 | 6772 | { |
c8d1ba58 | 6773 | unsigned int to_submit; |
bdcd3eab | 6774 | int ret = 0; |
6c271ce2 | 6775 | |
c8d1ba58 | 6776 | to_submit = io_sqring_entries(ctx); |
e95eee2d | 6777 | /* if we're handling multiple rings, cap submit size for fairness */ |
4ce8ad95 OL |
6778 | if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE) |
6779 | to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE; | |
e95eee2d | 6780 | |
906a3c6f | 6781 | if (!list_empty(&ctx->iopoll_list) || to_submit) { |
c8d1ba58 | 6782 | unsigned nr_events = 0; |
948e1947 PB |
6783 | const struct cred *creds = NULL; |
6784 | ||
6785 | if (ctx->sq_creds != current_cred()) | |
6786 | creds = override_creds(ctx->sq_creds); | |
a4c0b3de | 6787 | |
c8d1ba58 | 6788 | mutex_lock(&ctx->uring_lock); |
906a3c6f | 6789 | if (!list_empty(&ctx->iopoll_list)) |
3c30ef0f | 6790 | io_do_iopoll(ctx, &nr_events, 0, true); |
906a3c6f | 6791 | |
3b763ba1 PB |
6792 | /* |
6793 | * Don't submit if refs are dying, good for io_uring_register(), | |
6794 | * but also it is relied upon by io_ring_exit_work() | |
6795 | */ | |
0298ef96 PB |
6796 | if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) && |
6797 | !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
08369246 | 6798 | ret = io_submit_sqes(ctx, to_submit); |
c8d1ba58 | 6799 | mutex_unlock(&ctx->uring_lock); |
6c271ce2 | 6800 | |
acfb381d PB |
6801 | if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait)) |
6802 | wake_up(&ctx->sqo_sq_wait); | |
948e1947 PB |
6803 | if (creds) |
6804 | revert_creds(creds); | |
acfb381d | 6805 | } |
6c271ce2 | 6806 | |
08369246 XW |
6807 | return ret; |
6808 | } | |
6c271ce2 | 6809 | |
08369246 XW |
6810 | static void io_sqd_update_thread_idle(struct io_sq_data *sqd) |
6811 | { | |
6812 | struct io_ring_ctx *ctx; | |
6813 | unsigned sq_thread_idle = 0; | |
6c271ce2 | 6814 | |
c9dca27d PB |
6815 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
6816 | sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle); | |
08369246 | 6817 | sqd->sq_thread_idle = sq_thread_idle; |
c8d1ba58 | 6818 | } |
6c271ce2 | 6819 | |
e4b6d902 PB |
6820 | static bool io_sqd_handle_event(struct io_sq_data *sqd) |
6821 | { | |
6822 | bool did_sig = false; | |
6823 | struct ksignal ksig; | |
6824 | ||
6825 | if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) || | |
6826 | signal_pending(current)) { | |
6827 | mutex_unlock(&sqd->lock); | |
6828 | if (signal_pending(current)) | |
6829 | did_sig = get_signal(&ksig); | |
6830 | cond_resched(); | |
6831 | mutex_lock(&sqd->lock); | |
6832 | } | |
e4b6d902 PB |
6833 | return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
6834 | } | |
6835 | ||
c8d1ba58 JA |
6836 | static int io_sq_thread(void *data) |
6837 | { | |
69fb2131 JA |
6838 | struct io_sq_data *sqd = data; |
6839 | struct io_ring_ctx *ctx; | |
a0d9205f | 6840 | unsigned long timeout = 0; |
37d1e2e3 | 6841 | char buf[TASK_COMM_LEN]; |
08369246 | 6842 | DEFINE_WAIT(wait); |
6c271ce2 | 6843 | |
696ee88a | 6844 | snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid); |
37d1e2e3 | 6845 | set_task_comm(current, buf); |
37d1e2e3 JA |
6846 | |
6847 | if (sqd->sq_cpu != -1) | |
6848 | set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu)); | |
6849 | else | |
6850 | set_cpus_allowed_ptr(current, cpu_online_mask); | |
6851 | current->flags |= PF_NO_SETAFFINITY; | |
6852 | ||
09a6f4ef | 6853 | mutex_lock(&sqd->lock); |
e4b6d902 | 6854 | while (1) { |
1a924a80 | 6855 | bool cap_entries, sqt_spin = false; |
c1edbf5f | 6856 | |
e4b6d902 PB |
6857 | if (io_sqd_events_pending(sqd) || signal_pending(current)) { |
6858 | if (io_sqd_handle_event(sqd)) | |
c7d95613 | 6859 | break; |
08369246 XW |
6860 | timeout = jiffies + sqd->sq_thread_idle; |
6861 | } | |
e4b6d902 | 6862 | |
e95eee2d | 6863 | cap_entries = !list_is_singular(&sqd->ctx_list); |
69fb2131 | 6864 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
948e1947 | 6865 | int ret = __io_sq_thread(ctx, cap_entries); |
7c30f36a | 6866 | |
08369246 XW |
6867 | if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list))) |
6868 | sqt_spin = true; | |
69fb2131 | 6869 | } |
dd432ea5 PB |
6870 | if (io_run_task_work()) |
6871 | sqt_spin = true; | |
6c271ce2 | 6872 | |
08369246 | 6873 | if (sqt_spin || !time_after(jiffies, timeout)) { |
c8d1ba58 | 6874 | cond_resched(); |
08369246 XW |
6875 | if (sqt_spin) |
6876 | timeout = jiffies + sqd->sq_thread_idle; | |
6877 | continue; | |
6878 | } | |
6879 | ||
08369246 | 6880 | prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE); |
dd432ea5 | 6881 | if (!io_sqd_events_pending(sqd) && !current->task_works) { |
1a924a80 PB |
6882 | bool needs_sched = true; |
6883 | ||
724cb4f9 | 6884 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
aaa9f0f4 PB |
6885 | io_ring_set_wakeup_flag(ctx); |
6886 | ||
724cb4f9 HX |
6887 | if ((ctx->flags & IORING_SETUP_IOPOLL) && |
6888 | !list_empty_careful(&ctx->iopoll_list)) { | |
6889 | needs_sched = false; | |
6890 | break; | |
6891 | } | |
6892 | if (io_sqring_entries(ctx)) { | |
6893 | needs_sched = false; | |
6894 | break; | |
6895 | } | |
6896 | } | |
6897 | ||
6898 | if (needs_sched) { | |
6899 | mutex_unlock(&sqd->lock); | |
6900 | schedule(); | |
6901 | mutex_lock(&sqd->lock); | |
6902 | } | |
69fb2131 JA |
6903 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
6904 | io_ring_clear_wakeup_flag(ctx); | |
6c271ce2 | 6905 | } |
08369246 XW |
6906 | |
6907 | finish_wait(&sqd->wait, &wait); | |
6908 | timeout = jiffies + sqd->sq_thread_idle; | |
6c271ce2 | 6909 | } |
28cea78a | 6910 | |
78cc687b | 6911 | io_uring_cancel_generic(true, sqd); |
37d1e2e3 | 6912 | sqd->thread = NULL; |
05962f95 | 6913 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
5f3f26f9 | 6914 | io_ring_set_wakeup_flag(ctx); |
521d6a73 | 6915 | io_run_task_work(); |
734551df PB |
6916 | mutex_unlock(&sqd->lock); |
6917 | ||
37d1e2e3 JA |
6918 | complete(&sqd->exited); |
6919 | do_exit(0); | |
6c271ce2 JA |
6920 | } |
6921 | ||
bda52162 JA |
6922 | struct io_wait_queue { |
6923 | struct wait_queue_entry wq; | |
6924 | struct io_ring_ctx *ctx; | |
5fd46178 | 6925 | unsigned cq_tail; |
bda52162 JA |
6926 | unsigned nr_timeouts; |
6927 | }; | |
6928 | ||
6c503150 | 6929 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
6930 | { |
6931 | struct io_ring_ctx *ctx = iowq->ctx; | |
5fd46178 | 6932 | int dist = ctx->cached_cq_tail - (int) iowq->cq_tail; |
bda52162 JA |
6933 | |
6934 | /* | |
d195a66e | 6935 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
6936 | * started waiting. For timeouts, we always want to return to userspace, |
6937 | * regardless of event count. | |
6938 | */ | |
5fd46178 | 6939 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
6940 | } |
6941 | ||
6942 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
6943 | int wake_flags, void *key) | |
6944 | { | |
6945 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
6946 | wq); | |
6947 | ||
6c503150 PB |
6948 | /* |
6949 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
6950 | * the task, and the next invocation will do it. | |
6951 | */ | |
5ed7a37d | 6952 | if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow)) |
6c503150 PB |
6953 | return autoremove_wake_function(curr, mode, wake_flags, key); |
6954 | return -1; | |
bda52162 JA |
6955 | } |
6956 | ||
af9c1a44 JA |
6957 | static int io_run_task_work_sig(void) |
6958 | { | |
6959 | if (io_run_task_work()) | |
6960 | return 1; | |
6961 | if (!signal_pending(current)) | |
6962 | return 0; | |
0b8cfa97 | 6963 | if (test_thread_flag(TIF_NOTIFY_SIGNAL)) |
792ee0f6 | 6964 | return -ERESTARTSYS; |
af9c1a44 JA |
6965 | return -EINTR; |
6966 | } | |
6967 | ||
eeb60b9a PB |
6968 | /* when returns >0, the caller should retry */ |
6969 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
6970 | struct io_wait_queue *iowq, | |
6971 | signed long *timeout) | |
6972 | { | |
6973 | int ret; | |
6974 | ||
6975 | /* make sure we run task_work before checking for signals */ | |
6976 | ret = io_run_task_work_sig(); | |
6977 | if (ret || io_should_wake(iowq)) | |
6978 | return ret; | |
6979 | /* let the caller flush overflows, retry */ | |
5ed7a37d | 6980 | if (test_bit(0, &ctx->check_cq_overflow)) |
eeb60b9a PB |
6981 | return 1; |
6982 | ||
6983 | *timeout = schedule_timeout(*timeout); | |
6984 | return !*timeout ? -ETIME : 1; | |
6985 | } | |
6986 | ||
2b188cc1 JA |
6987 | /* |
6988 | * Wait until events become available, if we don't already have some. The | |
6989 | * application must reap them itself, as they reside on the shared cq ring. | |
6990 | */ | |
6991 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
6992 | const sigset_t __user *sig, size_t sigsz, |
6993 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 6994 | { |
90291099 | 6995 | struct io_wait_queue iowq; |
75b28aff | 6996 | struct io_rings *rings = ctx->rings; |
c1d5a224 PB |
6997 | signed long timeout = MAX_SCHEDULE_TIMEOUT; |
6998 | int ret; | |
2b188cc1 | 6999 | |
b41e9852 | 7000 | do { |
90f67366 | 7001 | io_cqring_overflow_flush(ctx); |
6c503150 | 7002 | if (io_cqring_events(ctx) >= min_events) |
b41e9852 | 7003 | return 0; |
4c6e277c | 7004 | if (!io_run_task_work()) |
b41e9852 | 7005 | break; |
b41e9852 | 7006 | } while (1); |
2b188cc1 JA |
7007 | |
7008 | if (sig) { | |
9e75ad5d AB |
7009 | #ifdef CONFIG_COMPAT |
7010 | if (in_compat_syscall()) | |
7011 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 7012 | sigsz); |
9e75ad5d AB |
7013 | else |
7014 | #endif | |
b772434b | 7015 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 7016 | |
2b188cc1 JA |
7017 | if (ret) |
7018 | return ret; | |
7019 | } | |
7020 | ||
c73ebb68 | 7021 | if (uts) { |
c1d5a224 PB |
7022 | struct timespec64 ts; |
7023 | ||
c73ebb68 HX |
7024 | if (get_timespec64(&ts, uts)) |
7025 | return -EFAULT; | |
7026 | timeout = timespec64_to_jiffies(&ts); | |
7027 | } | |
7028 | ||
90291099 PB |
7029 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
7030 | iowq.wq.private = current; | |
7031 | INIT_LIST_HEAD(&iowq.wq.entry); | |
7032 | iowq.ctx = ctx; | |
bda52162 | 7033 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 7034 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
90291099 | 7035 | |
c826bd7a | 7036 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 7037 | do { |
ca0a2651 | 7038 | /* if we can't even flush overflow, don't wait for more */ |
90f67366 | 7039 | if (!io_cqring_overflow_flush(ctx)) { |
ca0a2651 JA |
7040 | ret = -EBUSY; |
7041 | break; | |
7042 | } | |
311997b3 | 7043 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, |
bda52162 | 7044 | TASK_INTERRUPTIBLE); |
eeb60b9a | 7045 | ret = io_cqring_wait_schedule(ctx, &iowq, &timeout); |
311997b3 | 7046 | finish_wait(&ctx->cq_wait, &iowq.wq); |
ca0a2651 | 7047 | cond_resched(); |
eeb60b9a | 7048 | } while (ret > 0); |
bda52162 | 7049 | |
b7db41c9 | 7050 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 7051 | |
75b28aff | 7052 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
7053 | } |
7054 | ||
9123c8ff | 7055 | static void io_free_page_table(void **table, size_t size) |
05f3fb3c | 7056 | { |
9123c8ff | 7057 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); |
05f3fb3c | 7058 | |
846a4ef2 | 7059 | for (i = 0; i < nr_tables; i++) |
9123c8ff PB |
7060 | kfree(table[i]); |
7061 | kfree(table); | |
7062 | } | |
7063 | ||
7064 | static void **io_alloc_page_table(size_t size) | |
7065 | { | |
7066 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); | |
7067 | size_t init_size = size; | |
7068 | void **table; | |
7069 | ||
7070 | table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL); | |
7071 | if (!table) | |
7072 | return NULL; | |
7073 | ||
7074 | for (i = 0; i < nr_tables; i++) { | |
27f6b318 | 7075 | unsigned int this_size = min_t(size_t, size, PAGE_SIZE); |
9123c8ff PB |
7076 | |
7077 | table[i] = kzalloc(this_size, GFP_KERNEL); | |
7078 | if (!table[i]) { | |
7079 | io_free_page_table(table, init_size); | |
7080 | return NULL; | |
7081 | } | |
7082 | size -= this_size; | |
7083 | } | |
7084 | return table; | |
05f3fb3c JA |
7085 | } |
7086 | ||
28a9fe25 | 7087 | static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node) |
1642b445 | 7088 | { |
28a9fe25 PB |
7089 | percpu_ref_exit(&ref_node->refs); |
7090 | kfree(ref_node); | |
1642b445 PB |
7091 | } |
7092 | ||
b9bd2bea PB |
7093 | static void io_rsrc_node_ref_zero(struct percpu_ref *ref) |
7094 | { | |
7095 | struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs); | |
7096 | struct io_ring_ctx *ctx = node->rsrc_data->ctx; | |
7097 | unsigned long flags; | |
7098 | bool first_add = false; | |
7099 | ||
7100 | spin_lock_irqsave(&ctx->rsrc_ref_lock, flags); | |
7101 | node->done = true; | |
7102 | ||
7103 | while (!list_empty(&ctx->rsrc_ref_list)) { | |
7104 | node = list_first_entry(&ctx->rsrc_ref_list, | |
7105 | struct io_rsrc_node, node); | |
7106 | /* recycle ref nodes in order */ | |
7107 | if (!node->done) | |
7108 | break; | |
7109 | list_del(&node->node); | |
7110 | first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist); | |
7111 | } | |
7112 | spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags); | |
7113 | ||
7114 | if (first_add) | |
7115 | mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ); | |
7116 | } | |
7117 | ||
7118 | static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx) | |
7119 | { | |
7120 | struct io_rsrc_node *ref_node; | |
7121 | ||
7122 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); | |
7123 | if (!ref_node) | |
7124 | return NULL; | |
7125 | ||
7126 | if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero, | |
7127 | 0, GFP_KERNEL)) { | |
7128 | kfree(ref_node); | |
7129 | return NULL; | |
7130 | } | |
7131 | INIT_LIST_HEAD(&ref_node->node); | |
7132 | INIT_LIST_HEAD(&ref_node->rsrc_list); | |
7133 | ref_node->done = false; | |
7134 | return ref_node; | |
7135 | } | |
7136 | ||
a7f0ed5a PB |
7137 | static void io_rsrc_node_switch(struct io_ring_ctx *ctx, |
7138 | struct io_rsrc_data *data_to_kill) | |
6b06314c | 7139 | { |
a7f0ed5a PB |
7140 | WARN_ON_ONCE(!ctx->rsrc_backup_node); |
7141 | WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node); | |
6b06314c | 7142 | |
a7f0ed5a PB |
7143 | if (data_to_kill) { |
7144 | struct io_rsrc_node *rsrc_node = ctx->rsrc_node; | |
82fbcfa9 | 7145 | |
a7f0ed5a | 7146 | rsrc_node->rsrc_data = data_to_kill; |
4956b9ea | 7147 | spin_lock_irq(&ctx->rsrc_ref_lock); |
a7f0ed5a | 7148 | list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list); |
4956b9ea | 7149 | spin_unlock_irq(&ctx->rsrc_ref_lock); |
82fbcfa9 | 7150 | |
3e942498 | 7151 | atomic_inc(&data_to_kill->refs); |
a7f0ed5a PB |
7152 | percpu_ref_kill(&rsrc_node->refs); |
7153 | ctx->rsrc_node = NULL; | |
7154 | } | |
6b06314c | 7155 | |
a7f0ed5a PB |
7156 | if (!ctx->rsrc_node) { |
7157 | ctx->rsrc_node = ctx->rsrc_backup_node; | |
7158 | ctx->rsrc_backup_node = NULL; | |
7159 | } | |
8bad28d8 HX |
7160 | } |
7161 | ||
a7f0ed5a | 7162 | static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx) |
8dd03afe PB |
7163 | { |
7164 | if (ctx->rsrc_backup_node) | |
7165 | return 0; | |
b895c9a6 | 7166 | ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx); |
8dd03afe | 7167 | return ctx->rsrc_backup_node ? 0 : -ENOMEM; |
8bad28d8 HX |
7168 | } |
7169 | ||
40ae0ff7 | 7170 | static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx) |
8bad28d8 HX |
7171 | { |
7172 | int ret; | |
05589553 | 7173 | |
215c3902 | 7174 | /* As we may drop ->uring_lock, other task may have started quiesce */ |
8bad28d8 HX |
7175 | if (data->quiesce) |
7176 | return -ENXIO; | |
05589553 | 7177 | |
8bad28d8 | 7178 | data->quiesce = true; |
1ffc5422 | 7179 | do { |
a7f0ed5a | 7180 | ret = io_rsrc_node_switch_start(ctx); |
8dd03afe | 7181 | if (ret) |
f2303b1f | 7182 | break; |
a7f0ed5a | 7183 | io_rsrc_node_switch(ctx, data); |
f2303b1f | 7184 | |
3e942498 PB |
7185 | /* kill initial ref, already quiesced if zero */ |
7186 | if (atomic_dec_and_test(&data->refs)) | |
7187 | break; | |
c018db4a | 7188 | mutex_unlock(&ctx->uring_lock); |
8bad28d8 | 7189 | flush_delayed_work(&ctx->rsrc_put_work); |
1ffc5422 | 7190 | ret = wait_for_completion_interruptible(&data->done); |
c018db4a JA |
7191 | if (!ret) { |
7192 | mutex_lock(&ctx->uring_lock); | |
1ffc5422 | 7193 | break; |
c018db4a | 7194 | } |
8bad28d8 | 7195 | |
3e942498 PB |
7196 | atomic_inc(&data->refs); |
7197 | /* wait for all works potentially completing data->done */ | |
7198 | flush_delayed_work(&ctx->rsrc_put_work); | |
cb5e1b81 | 7199 | reinit_completion(&data->done); |
8dd03afe | 7200 | |
1ffc5422 | 7201 | ret = io_run_task_work_sig(); |
8bad28d8 | 7202 | mutex_lock(&ctx->uring_lock); |
f2303b1f | 7203 | } while (ret >= 0); |
8bad28d8 | 7204 | data->quiesce = false; |
05f3fb3c | 7205 | |
8bad28d8 | 7206 | return ret; |
d7954b2b BM |
7207 | } |
7208 | ||
2d091d62 PB |
7209 | static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx) |
7210 | { | |
7211 | unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK; | |
7212 | unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT; | |
7213 | ||
7214 | return &data->tags[table_idx][off]; | |
7215 | } | |
7216 | ||
44b31f2f | 7217 | static void io_rsrc_data_free(struct io_rsrc_data *data) |
1ad555c6 | 7218 | { |
2d091d62 PB |
7219 | size_t size = data->nr * sizeof(data->tags[0][0]); |
7220 | ||
7221 | if (data->tags) | |
7222 | io_free_page_table((void **)data->tags, size); | |
44b31f2f PB |
7223 | kfree(data); |
7224 | } | |
7225 | ||
d878c816 PB |
7226 | static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put, |
7227 | u64 __user *utags, unsigned nr, | |
7228 | struct io_rsrc_data **pdata) | |
1ad555c6 | 7229 | { |
b895c9a6 | 7230 | struct io_rsrc_data *data; |
2d091d62 | 7231 | int ret = -ENOMEM; |
d878c816 | 7232 | unsigned i; |
1ad555c6 BM |
7233 | |
7234 | data = kzalloc(sizeof(*data), GFP_KERNEL); | |
7235 | if (!data) | |
d878c816 | 7236 | return -ENOMEM; |
2d091d62 | 7237 | data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); |
b60c8dce | 7238 | if (!data->tags) { |
1ad555c6 | 7239 | kfree(data); |
d878c816 PB |
7240 | return -ENOMEM; |
7241 | } | |
2d091d62 PB |
7242 | |
7243 | data->nr = nr; | |
7244 | data->ctx = ctx; | |
7245 | data->do_put = do_put; | |
d878c816 | 7246 | if (utags) { |
2d091d62 | 7247 | ret = -EFAULT; |
d878c816 | 7248 | for (i = 0; i < nr; i++) { |
fdd1dc31 CIK |
7249 | u64 *tag_slot = io_get_tag_slot(data, i); |
7250 | ||
7251 | if (copy_from_user(tag_slot, &utags[i], | |
7252 | sizeof(*tag_slot))) | |
2d091d62 | 7253 | goto fail; |
d878c816 | 7254 | } |
1ad555c6 | 7255 | } |
b60c8dce | 7256 | |
3e942498 | 7257 | atomic_set(&data->refs, 1); |
1ad555c6 | 7258 | init_completion(&data->done); |
d878c816 PB |
7259 | *pdata = data; |
7260 | return 0; | |
2d091d62 PB |
7261 | fail: |
7262 | io_rsrc_data_free(data); | |
7263 | return ret; | |
1ad555c6 BM |
7264 | } |
7265 | ||
9123c8ff PB |
7266 | static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files) |
7267 | { | |
042b0d85 | 7268 | table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL); |
9123c8ff PB |
7269 | return !!table->files; |
7270 | } | |
7271 | ||
042b0d85 | 7272 | static void io_free_file_tables(struct io_file_table *table) |
9123c8ff | 7273 | { |
042b0d85 | 7274 | kvfree(table->files); |
9123c8ff PB |
7275 | table->files = NULL; |
7276 | } | |
7277 | ||
fff4db76 | 7278 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 7279 | { |
fff4db76 PB |
7280 | #if defined(CONFIG_UNIX) |
7281 | if (ctx->ring_sock) { | |
7282 | struct sock *sock = ctx->ring_sock->sk; | |
7283 | struct sk_buff *skb; | |
7284 | ||
7285 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
7286 | kfree_skb(skb); | |
7287 | } | |
7288 | #else | |
7289 | int i; | |
7290 | ||
7291 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7292 | struct file *file; | |
7293 | ||
7294 | file = io_file_from_index(ctx, i); | |
7295 | if (file) | |
7296 | fput(file); | |
7297 | } | |
7298 | #endif | |
042b0d85 | 7299 | io_free_file_tables(&ctx->file_table); |
44b31f2f | 7300 | io_rsrc_data_free(ctx->file_data); |
fff4db76 PB |
7301 | ctx->file_data = NULL; |
7302 | ctx->nr_user_files = 0; | |
1ad555c6 BM |
7303 | } |
7304 | ||
d7954b2b BM |
7305 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
7306 | { | |
d7954b2b BM |
7307 | int ret; |
7308 | ||
08480400 | 7309 | if (!ctx->file_data) |
d7954b2b | 7310 | return -ENXIO; |
08480400 PB |
7311 | ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); |
7312 | if (!ret) | |
7313 | __io_sqe_files_unregister(ctx); | |
7314 | return ret; | |
6b06314c JA |
7315 | } |
7316 | ||
37d1e2e3 | 7317 | static void io_sq_thread_unpark(struct io_sq_data *sqd) |
09a6f4ef | 7318 | __releases(&sqd->lock) |
37d1e2e3 | 7319 | { |
521d6a73 PB |
7320 | WARN_ON_ONCE(sqd->thread == current); |
7321 | ||
9e138a48 PB |
7322 | /* |
7323 | * Do the dance but not conditional clear_bit() because it'd race with | |
7324 | * other threads incrementing park_pending and setting the bit. | |
7325 | */ | |
37d1e2e3 | 7326 | clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
9e138a48 PB |
7327 | if (atomic_dec_return(&sqd->park_pending)) |
7328 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); | |
09a6f4ef | 7329 | mutex_unlock(&sqd->lock); |
37d1e2e3 JA |
7330 | } |
7331 | ||
86e0d676 | 7332 | static void io_sq_thread_park(struct io_sq_data *sqd) |
09a6f4ef | 7333 | __acquires(&sqd->lock) |
37d1e2e3 | 7334 | { |
521d6a73 PB |
7335 | WARN_ON_ONCE(sqd->thread == current); |
7336 | ||
9e138a48 | 7337 | atomic_inc(&sqd->park_pending); |
86e0d676 | 7338 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
09a6f4ef | 7339 | mutex_lock(&sqd->lock); |
05962f95 | 7340 | if (sqd->thread) |
86e0d676 | 7341 | wake_up_process(sqd->thread); |
37d1e2e3 JA |
7342 | } |
7343 | ||
7344 | static void io_sq_thread_stop(struct io_sq_data *sqd) | |
7345 | { | |
521d6a73 | 7346 | WARN_ON_ONCE(sqd->thread == current); |
88885f66 | 7347 | WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)); |
521d6a73 | 7348 | |
05962f95 | 7349 | set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
88885f66 | 7350 | mutex_lock(&sqd->lock); |
e8f98f24 JA |
7351 | if (sqd->thread) |
7352 | wake_up_process(sqd->thread); | |
09a6f4ef | 7353 | mutex_unlock(&sqd->lock); |
05962f95 | 7354 | wait_for_completion(&sqd->exited); |
37d1e2e3 JA |
7355 | } |
7356 | ||
534ca6d6 | 7357 | static void io_put_sq_data(struct io_sq_data *sqd) |
6c271ce2 | 7358 | { |
534ca6d6 | 7359 | if (refcount_dec_and_test(&sqd->refs)) { |
9e138a48 PB |
7360 | WARN_ON_ONCE(atomic_read(&sqd->park_pending)); |
7361 | ||
37d1e2e3 JA |
7362 | io_sq_thread_stop(sqd); |
7363 | kfree(sqd); | |
7364 | } | |
7365 | } | |
7366 | ||
7367 | static void io_sq_thread_finish(struct io_ring_ctx *ctx) | |
7368 | { | |
7369 | struct io_sq_data *sqd = ctx->sq_data; | |
7370 | ||
7371 | if (sqd) { | |
05962f95 | 7372 | io_sq_thread_park(sqd); |
521d6a73 | 7373 | list_del_init(&ctx->sqd_list); |
37d1e2e3 | 7374 | io_sqd_update_thread_idle(sqd); |
05962f95 | 7375 | io_sq_thread_unpark(sqd); |
37d1e2e3 JA |
7376 | |
7377 | io_put_sq_data(sqd); | |
7378 | ctx->sq_data = NULL; | |
534ca6d6 JA |
7379 | } |
7380 | } | |
7381 | ||
aa06165d JA |
7382 | static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p) |
7383 | { | |
7384 | struct io_ring_ctx *ctx_attach; | |
7385 | struct io_sq_data *sqd; | |
7386 | struct fd f; | |
7387 | ||
7388 | f = fdget(p->wq_fd); | |
7389 | if (!f.file) | |
7390 | return ERR_PTR(-ENXIO); | |
7391 | if (f.file->f_op != &io_uring_fops) { | |
7392 | fdput(f); | |
7393 | return ERR_PTR(-EINVAL); | |
7394 | } | |
7395 | ||
7396 | ctx_attach = f.file->private_data; | |
7397 | sqd = ctx_attach->sq_data; | |
7398 | if (!sqd) { | |
7399 | fdput(f); | |
7400 | return ERR_PTR(-EINVAL); | |
7401 | } | |
5c2469e0 JA |
7402 | if (sqd->task_tgid != current->tgid) { |
7403 | fdput(f); | |
7404 | return ERR_PTR(-EPERM); | |
7405 | } | |
aa06165d JA |
7406 | |
7407 | refcount_inc(&sqd->refs); | |
7408 | fdput(f); | |
7409 | return sqd; | |
7410 | } | |
7411 | ||
26984fbf PB |
7412 | static struct io_sq_data *io_get_sq_data(struct io_uring_params *p, |
7413 | bool *attached) | |
534ca6d6 JA |
7414 | { |
7415 | struct io_sq_data *sqd; | |
7416 | ||
26984fbf | 7417 | *attached = false; |
5c2469e0 JA |
7418 | if (p->flags & IORING_SETUP_ATTACH_WQ) { |
7419 | sqd = io_attach_sq_data(p); | |
26984fbf PB |
7420 | if (!IS_ERR(sqd)) { |
7421 | *attached = true; | |
5c2469e0 | 7422 | return sqd; |
26984fbf | 7423 | } |
5c2469e0 JA |
7424 | /* fall through for EPERM case, setup new sqd/task */ |
7425 | if (PTR_ERR(sqd) != -EPERM) | |
7426 | return sqd; | |
7427 | } | |
aa06165d | 7428 | |
534ca6d6 JA |
7429 | sqd = kzalloc(sizeof(*sqd), GFP_KERNEL); |
7430 | if (!sqd) | |
7431 | return ERR_PTR(-ENOMEM); | |
7432 | ||
9e138a48 | 7433 | atomic_set(&sqd->park_pending, 0); |
534ca6d6 | 7434 | refcount_set(&sqd->refs, 1); |
69fb2131 | 7435 | INIT_LIST_HEAD(&sqd->ctx_list); |
09a6f4ef | 7436 | mutex_init(&sqd->lock); |
534ca6d6 | 7437 | init_waitqueue_head(&sqd->wait); |
37d1e2e3 | 7438 | init_completion(&sqd->exited); |
534ca6d6 JA |
7439 | return sqd; |
7440 | } | |
7441 | ||
6b06314c | 7442 | #if defined(CONFIG_UNIX) |
6b06314c JA |
7443 | /* |
7444 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
7445 | * the io_uring can be safely unregistered on process exit, even if we have | |
7446 | * loops in the file referencing. | |
7447 | */ | |
7448 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
7449 | { | |
7450 | struct sock *sk = ctx->ring_sock->sk; | |
7451 | struct scm_fp_list *fpl; | |
7452 | struct sk_buff *skb; | |
08a45173 | 7453 | int i, nr_files; |
6b06314c | 7454 | |
6b06314c JA |
7455 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); |
7456 | if (!fpl) | |
7457 | return -ENOMEM; | |
7458 | ||
7459 | skb = alloc_skb(0, GFP_KERNEL); | |
7460 | if (!skb) { | |
7461 | kfree(fpl); | |
7462 | return -ENOMEM; | |
7463 | } | |
7464 | ||
7465 | skb->sk = sk; | |
6b06314c | 7466 | |
08a45173 | 7467 | nr_files = 0; |
62e398be | 7468 | fpl->user = get_uid(current_user()); |
6b06314c | 7469 | for (i = 0; i < nr; i++) { |
65e19f54 JA |
7470 | struct file *file = io_file_from_index(ctx, i + offset); |
7471 | ||
7472 | if (!file) | |
08a45173 | 7473 | continue; |
65e19f54 | 7474 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
7475 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
7476 | nr_files++; | |
6b06314c JA |
7477 | } |
7478 | ||
08a45173 JA |
7479 | if (nr_files) { |
7480 | fpl->max = SCM_MAX_FD; | |
7481 | fpl->count = nr_files; | |
7482 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 7483 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
7484 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
7485 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 7486 | |
08a45173 JA |
7487 | for (i = 0; i < nr_files; i++) |
7488 | fput(fpl->fp[i]); | |
7489 | } else { | |
7490 | kfree_skb(skb); | |
7491 | kfree(fpl); | |
7492 | } | |
6b06314c JA |
7493 | |
7494 | return 0; | |
7495 | } | |
7496 | ||
7497 | /* | |
7498 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
7499 | * causes regular reference counting to break down. We rely on the UNIX | |
7500 | * garbage collection to take care of this problem for us. | |
7501 | */ | |
7502 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7503 | { | |
7504 | unsigned left, total; | |
7505 | int ret = 0; | |
7506 | ||
7507 | total = 0; | |
7508 | left = ctx->nr_user_files; | |
7509 | while (left) { | |
7510 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
7511 | |
7512 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
7513 | if (ret) | |
7514 | break; | |
7515 | left -= this_files; | |
7516 | total += this_files; | |
7517 | } | |
7518 | ||
7519 | if (!ret) | |
7520 | return 0; | |
7521 | ||
7522 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
7523 | struct file *file = io_file_from_index(ctx, total); |
7524 | ||
7525 | if (file) | |
7526 | fput(file); | |
6b06314c JA |
7527 | total++; |
7528 | } | |
7529 | ||
7530 | return ret; | |
7531 | } | |
7532 | #else | |
7533 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7534 | { | |
7535 | return 0; | |
7536 | } | |
7537 | #endif | |
7538 | ||
47e90392 | 7539 | static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) |
05f3fb3c | 7540 | { |
50238531 | 7541 | struct file *file = prsrc->file; |
05f3fb3c JA |
7542 | #if defined(CONFIG_UNIX) |
7543 | struct sock *sock = ctx->ring_sock->sk; | |
7544 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
7545 | struct sk_buff *skb; | |
7546 | int i; | |
7547 | ||
7548 | __skb_queue_head_init(&list); | |
7549 | ||
7550 | /* | |
7551 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
7552 | * remove this entry and rearrange the file array. | |
7553 | */ | |
7554 | skb = skb_dequeue(head); | |
7555 | while (skb) { | |
7556 | struct scm_fp_list *fp; | |
7557 | ||
7558 | fp = UNIXCB(skb).fp; | |
7559 | for (i = 0; i < fp->count; i++) { | |
7560 | int left; | |
7561 | ||
7562 | if (fp->fp[i] != file) | |
7563 | continue; | |
7564 | ||
7565 | unix_notinflight(fp->user, fp->fp[i]); | |
7566 | left = fp->count - 1 - i; | |
7567 | if (left) { | |
7568 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
7569 | left * sizeof(struct file *)); | |
7570 | } | |
7571 | fp->count--; | |
7572 | if (!fp->count) { | |
7573 | kfree_skb(skb); | |
7574 | skb = NULL; | |
7575 | } else { | |
7576 | __skb_queue_tail(&list, skb); | |
7577 | } | |
7578 | fput(file); | |
7579 | file = NULL; | |
7580 | break; | |
7581 | } | |
7582 | ||
7583 | if (!file) | |
7584 | break; | |
7585 | ||
7586 | __skb_queue_tail(&list, skb); | |
7587 | ||
7588 | skb = skb_dequeue(head); | |
7589 | } | |
7590 | ||
7591 | if (skb_peek(&list)) { | |
7592 | spin_lock_irq(&head->lock); | |
7593 | while ((skb = __skb_dequeue(&list)) != NULL) | |
7594 | __skb_queue_tail(head, skb); | |
7595 | spin_unlock_irq(&head->lock); | |
7596 | } | |
7597 | #else | |
7598 | fput(file); | |
7599 | #endif | |
7600 | } | |
7601 | ||
b895c9a6 | 7602 | static void __io_rsrc_put_work(struct io_rsrc_node *ref_node) |
65e19f54 | 7603 | { |
b895c9a6 | 7604 | struct io_rsrc_data *rsrc_data = ref_node->rsrc_data; |
269bbe5f BM |
7605 | struct io_ring_ctx *ctx = rsrc_data->ctx; |
7606 | struct io_rsrc_put *prsrc, *tmp; | |
05589553 | 7607 | |
269bbe5f BM |
7608 | list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) { |
7609 | list_del(&prsrc->list); | |
b60c8dce PB |
7610 | |
7611 | if (prsrc->tag) { | |
7612 | bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL; | |
b60c8dce PB |
7613 | |
7614 | io_ring_submit_lock(ctx, lock_ring); | |
157d257f | 7615 | spin_lock_irq(&ctx->completion_lock); |
b60c8dce | 7616 | io_cqring_fill_event(ctx, prsrc->tag, 0, 0); |
2840f710 | 7617 | ctx->cq_extra++; |
b60c8dce | 7618 | io_commit_cqring(ctx); |
157d257f | 7619 | spin_unlock_irq(&ctx->completion_lock); |
b60c8dce PB |
7620 | io_cqring_ev_posted(ctx); |
7621 | io_ring_submit_unlock(ctx, lock_ring); | |
7622 | } | |
7623 | ||
40ae0ff7 | 7624 | rsrc_data->do_put(ctx, prsrc); |
269bbe5f | 7625 | kfree(prsrc); |
65e19f54 | 7626 | } |
05589553 | 7627 | |
28a9fe25 | 7628 | io_rsrc_node_destroy(ref_node); |
3e942498 PB |
7629 | if (atomic_dec_and_test(&rsrc_data->refs)) |
7630 | complete(&rsrc_data->done); | |
2faf852d | 7631 | } |
65e19f54 | 7632 | |
269bbe5f | 7633 | static void io_rsrc_put_work(struct work_struct *work) |
4a38aed2 JA |
7634 | { |
7635 | struct io_ring_ctx *ctx; | |
7636 | struct llist_node *node; | |
7637 | ||
269bbe5f BM |
7638 | ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work); |
7639 | node = llist_del_all(&ctx->rsrc_put_llist); | |
4a38aed2 JA |
7640 | |
7641 | while (node) { | |
b895c9a6 | 7642 | struct io_rsrc_node *ref_node; |
4a38aed2 JA |
7643 | struct llist_node *next = node->next; |
7644 | ||
b895c9a6 | 7645 | ref_node = llist_entry(node, struct io_rsrc_node, llist); |
269bbe5f | 7646 | __io_rsrc_put_work(ref_node); |
4a38aed2 JA |
7647 | node = next; |
7648 | } | |
7649 | } | |
7650 | ||
6b06314c | 7651 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
792e3582 | 7652 | unsigned nr_args, u64 __user *tags) |
6b06314c JA |
7653 | { |
7654 | __s32 __user *fds = (__s32 __user *) arg; | |
05f3fb3c | 7655 | struct file *file; |
f3baed39 | 7656 | int fd, ret; |
846a4ef2 | 7657 | unsigned i; |
6b06314c | 7658 | |
05f3fb3c | 7659 | if (ctx->file_data) |
6b06314c JA |
7660 | return -EBUSY; |
7661 | if (!nr_args) | |
7662 | return -EINVAL; | |
7663 | if (nr_args > IORING_MAX_FIXED_FILES) | |
7664 | return -EMFILE; | |
a7f0ed5a | 7665 | ret = io_rsrc_node_switch_start(ctx); |
f3baed39 PB |
7666 | if (ret) |
7667 | return ret; | |
d878c816 PB |
7668 | ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, |
7669 | &ctx->file_data); | |
7670 | if (ret) | |
7671 | return ret; | |
6b06314c | 7672 | |
f3baed39 | 7673 | ret = -ENOMEM; |
aeca241b | 7674 | if (!io_alloc_file_tables(&ctx->file_table, nr_args)) |
1ad555c6 | 7675 | goto out_free; |
65e19f54 | 7676 | |
08a45173 | 7677 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
d878c816 | 7678 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) { |
600cf3f8 PB |
7679 | ret = -EFAULT; |
7680 | goto out_fput; | |
7681 | } | |
08a45173 | 7682 | /* allow sparse sets */ |
792e3582 PB |
7683 | if (fd == -1) { |
7684 | ret = -EINVAL; | |
2d091d62 | 7685 | if (unlikely(*io_get_tag_slot(ctx->file_data, i))) |
792e3582 | 7686 | goto out_fput; |
08a45173 | 7687 | continue; |
792e3582 | 7688 | } |
6b06314c | 7689 | |
05f3fb3c | 7690 | file = fget(fd); |
6b06314c | 7691 | ret = -EBADF; |
792e3582 | 7692 | if (unlikely(!file)) |
600cf3f8 | 7693 | goto out_fput; |
05f3fb3c | 7694 | |
6b06314c JA |
7695 | /* |
7696 | * Don't allow io_uring instances to be registered. If UNIX | |
7697 | * isn't enabled, then this causes a reference cycle and this | |
7698 | * instance can never get freed. If UNIX is enabled we'll | |
7699 | * handle it just fine, but there's still no point in allowing | |
7700 | * a ring fd as it doesn't support regular read/write anyway. | |
7701 | */ | |
05f3fb3c JA |
7702 | if (file->f_op == &io_uring_fops) { |
7703 | fput(file); | |
600cf3f8 | 7704 | goto out_fput; |
6b06314c | 7705 | } |
aeca241b | 7706 | io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file); |
6b06314c JA |
7707 | } |
7708 | ||
6b06314c | 7709 | ret = io_sqe_files_scm(ctx); |
05589553 | 7710 | if (ret) { |
08480400 | 7711 | __io_sqe_files_unregister(ctx); |
05589553 XW |
7712 | return ret; |
7713 | } | |
6b06314c | 7714 | |
a7f0ed5a | 7715 | io_rsrc_node_switch(ctx, NULL); |
6b06314c | 7716 | return ret; |
600cf3f8 PB |
7717 | out_fput: |
7718 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7719 | file = io_file_from_index(ctx, i); | |
7720 | if (file) | |
7721 | fput(file); | |
7722 | } | |
042b0d85 | 7723 | io_free_file_tables(&ctx->file_table); |
600cf3f8 | 7724 | ctx->nr_user_files = 0; |
600cf3f8 | 7725 | out_free: |
44b31f2f | 7726 | io_rsrc_data_free(ctx->file_data); |
55cbc256 | 7727 | ctx->file_data = NULL; |
6b06314c JA |
7728 | return ret; |
7729 | } | |
7730 | ||
c3a31e60 JA |
7731 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
7732 | int index) | |
7733 | { | |
7734 | #if defined(CONFIG_UNIX) | |
7735 | struct sock *sock = ctx->ring_sock->sk; | |
7736 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
7737 | struct sk_buff *skb; | |
7738 | ||
7739 | /* | |
7740 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
7741 | * file set. If there's no room, fall back to allocating a new skb | |
7742 | * and filling it in. | |
7743 | */ | |
7744 | spin_lock_irq(&head->lock); | |
7745 | skb = skb_peek(head); | |
7746 | if (skb) { | |
7747 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
7748 | ||
7749 | if (fpl->count < SCM_MAX_FD) { | |
7750 | __skb_unlink(skb, head); | |
7751 | spin_unlock_irq(&head->lock); | |
7752 | fpl->fp[fpl->count] = get_file(file); | |
7753 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
7754 | fpl->count++; | |
7755 | spin_lock_irq(&head->lock); | |
7756 | __skb_queue_head(head, skb); | |
7757 | } else { | |
7758 | skb = NULL; | |
7759 | } | |
7760 | } | |
7761 | spin_unlock_irq(&head->lock); | |
7762 | ||
7763 | if (skb) { | |
7764 | fput(file); | |
7765 | return 0; | |
7766 | } | |
7767 | ||
7768 | return __io_sqe_files_scm(ctx, 1, index); | |
7769 | #else | |
7770 | return 0; | |
7771 | #endif | |
7772 | } | |
7773 | ||
b60c8dce | 7774 | static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, |
e7c78371 | 7775 | struct io_rsrc_node *node, void *rsrc) |
05f3fb3c | 7776 | { |
269bbe5f | 7777 | struct io_rsrc_put *prsrc; |
05f3fb3c | 7778 | |
269bbe5f BM |
7779 | prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL); |
7780 | if (!prsrc) | |
a5318d3c | 7781 | return -ENOMEM; |
05f3fb3c | 7782 | |
2d091d62 | 7783 | prsrc->tag = *io_get_tag_slot(data, idx); |
50238531 | 7784 | prsrc->rsrc = rsrc; |
e7c78371 | 7785 | list_add(&prsrc->list, &node->rsrc_list); |
a5318d3c | 7786 | return 0; |
05f3fb3c JA |
7787 | } |
7788 | ||
7789 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
c3bdad02 | 7790 | struct io_uring_rsrc_update2 *up, |
05f3fb3c JA |
7791 | unsigned nr_args) |
7792 | { | |
c3bdad02 | 7793 | u64 __user *tags = u64_to_user_ptr(up->tags); |
98f0b3b4 | 7794 | __s32 __user *fds = u64_to_user_ptr(up->data); |
b895c9a6 | 7795 | struct io_rsrc_data *data = ctx->file_data; |
a04b0ac0 PB |
7796 | struct io_fixed_file *file_slot; |
7797 | struct file *file; | |
98f0b3b4 PB |
7798 | int fd, i, err = 0; |
7799 | unsigned int done; | |
05589553 | 7800 | bool needs_switch = false; |
c3a31e60 | 7801 | |
98f0b3b4 PB |
7802 | if (!ctx->file_data) |
7803 | return -ENXIO; | |
7804 | if (up->offset + nr_args > ctx->nr_user_files) | |
c3a31e60 JA |
7805 | return -EINVAL; |
7806 | ||
67973b93 | 7807 | for (done = 0; done < nr_args; done++) { |
c3bdad02 PB |
7808 | u64 tag = 0; |
7809 | ||
7810 | if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || | |
7811 | copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
c3a31e60 JA |
7812 | err = -EFAULT; |
7813 | break; | |
7814 | } | |
c3bdad02 PB |
7815 | if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { |
7816 | err = -EINVAL; | |
7817 | break; | |
7818 | } | |
4e0377a1 | 7819 | if (fd == IORING_REGISTER_FILES_SKIP) |
7820 | continue; | |
7821 | ||
67973b93 | 7822 | i = array_index_nospec(up->offset + done, ctx->nr_user_files); |
aeca241b | 7823 | file_slot = io_fixed_file_slot(&ctx->file_table, i); |
ea64ec02 | 7824 | |
a04b0ac0 PB |
7825 | if (file_slot->file_ptr) { |
7826 | file = (struct file *)(file_slot->file_ptr & FFS_MASK); | |
b60c8dce PB |
7827 | err = io_queue_rsrc_removal(data, up->offset + done, |
7828 | ctx->rsrc_node, file); | |
a5318d3c HD |
7829 | if (err) |
7830 | break; | |
a04b0ac0 | 7831 | file_slot->file_ptr = 0; |
05589553 | 7832 | needs_switch = true; |
c3a31e60 JA |
7833 | } |
7834 | if (fd != -1) { | |
c3a31e60 JA |
7835 | file = fget(fd); |
7836 | if (!file) { | |
7837 | err = -EBADF; | |
7838 | break; | |
7839 | } | |
7840 | /* | |
7841 | * Don't allow io_uring instances to be registered. If | |
7842 | * UNIX isn't enabled, then this causes a reference | |
7843 | * cycle and this instance can never get freed. If UNIX | |
7844 | * is enabled we'll handle it just fine, but there's | |
7845 | * still no point in allowing a ring fd as it doesn't | |
7846 | * support regular read/write anyway. | |
7847 | */ | |
7848 | if (file->f_op == &io_uring_fops) { | |
7849 | fput(file); | |
7850 | err = -EBADF; | |
7851 | break; | |
7852 | } | |
2d091d62 | 7853 | *io_get_tag_slot(data, up->offset + done) = tag; |
9a321c98 | 7854 | io_fixed_file_set(file_slot, file); |
c3a31e60 | 7855 | err = io_sqe_file_register(ctx, file, i); |
f3bd9dae | 7856 | if (err) { |
a04b0ac0 | 7857 | file_slot->file_ptr = 0; |
f3bd9dae | 7858 | fput(file); |
c3a31e60 | 7859 | break; |
f3bd9dae | 7860 | } |
c3a31e60 | 7861 | } |
05f3fb3c JA |
7862 | } |
7863 | ||
a7f0ed5a PB |
7864 | if (needs_switch) |
7865 | io_rsrc_node_switch(ctx, data); | |
c3a31e60 JA |
7866 | return done ? done : err; |
7867 | } | |
05589553 | 7868 | |
685fe7fe JA |
7869 | static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx, |
7870 | struct task_struct *task) | |
24369c2e | 7871 | { |
e941894e | 7872 | struct io_wq_hash *hash; |
24369c2e | 7873 | struct io_wq_data data; |
24369c2e | 7874 | unsigned int concurrency; |
24369c2e | 7875 | |
362a9e65 | 7876 | mutex_lock(&ctx->uring_lock); |
e941894e JA |
7877 | hash = ctx->hash_map; |
7878 | if (!hash) { | |
7879 | hash = kzalloc(sizeof(*hash), GFP_KERNEL); | |
362a9e65 YY |
7880 | if (!hash) { |
7881 | mutex_unlock(&ctx->uring_lock); | |
e941894e | 7882 | return ERR_PTR(-ENOMEM); |
362a9e65 | 7883 | } |
e941894e JA |
7884 | refcount_set(&hash->refs, 1); |
7885 | init_waitqueue_head(&hash->wait); | |
7886 | ctx->hash_map = hash; | |
24369c2e | 7887 | } |
362a9e65 | 7888 | mutex_unlock(&ctx->uring_lock); |
24369c2e | 7889 | |
e941894e | 7890 | data.hash = hash; |
685fe7fe | 7891 | data.task = task; |
ebc11b6c | 7892 | data.free_work = io_wq_free_work; |
f5fa38c5 | 7893 | data.do_work = io_wq_submit_work; |
24369c2e | 7894 | |
d25e3a3d JA |
7895 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
7896 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
24369c2e | 7897 | |
5aa75ed5 | 7898 | return io_wq_create(concurrency, &data); |
24369c2e PB |
7899 | } |
7900 | ||
5aa75ed5 JA |
7901 | static int io_uring_alloc_task_context(struct task_struct *task, |
7902 | struct io_ring_ctx *ctx) | |
0f212204 JA |
7903 | { |
7904 | struct io_uring_task *tctx; | |
d8a6df10 | 7905 | int ret; |
0f212204 | 7906 | |
09899b19 | 7907 | tctx = kzalloc(sizeof(*tctx), GFP_KERNEL); |
0f212204 JA |
7908 | if (unlikely(!tctx)) |
7909 | return -ENOMEM; | |
7910 | ||
d8a6df10 JA |
7911 | ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL); |
7912 | if (unlikely(ret)) { | |
7913 | kfree(tctx); | |
7914 | return ret; | |
7915 | } | |
7916 | ||
685fe7fe | 7917 | tctx->io_wq = io_init_wq_offload(ctx, task); |
5aa75ed5 JA |
7918 | if (IS_ERR(tctx->io_wq)) { |
7919 | ret = PTR_ERR(tctx->io_wq); | |
7920 | percpu_counter_destroy(&tctx->inflight); | |
7921 | kfree(tctx); | |
7922 | return ret; | |
7923 | } | |
7924 | ||
0f212204 JA |
7925 | xa_init(&tctx->xa); |
7926 | init_waitqueue_head(&tctx->wait); | |
fdaf083c | 7927 | atomic_set(&tctx->in_idle, 0); |
b303fe2e | 7928 | atomic_set(&tctx->inflight_tracked, 0); |
0f212204 | 7929 | task->io_uring = tctx; |
7cbf1722 JA |
7930 | spin_lock_init(&tctx->task_lock); |
7931 | INIT_WQ_LIST(&tctx->task_list); | |
7cbf1722 | 7932 | init_task_work(&tctx->task_work, tctx_task_work); |
0f212204 JA |
7933 | return 0; |
7934 | } | |
7935 | ||
7936 | void __io_uring_free(struct task_struct *tsk) | |
7937 | { | |
7938 | struct io_uring_task *tctx = tsk->io_uring; | |
7939 | ||
7940 | WARN_ON_ONCE(!xa_empty(&tctx->xa)); | |
ef8eaa4e | 7941 | WARN_ON_ONCE(tctx->io_wq); |
09899b19 | 7942 | WARN_ON_ONCE(tctx->cached_refs); |
ef8eaa4e | 7943 | |
d8a6df10 | 7944 | percpu_counter_destroy(&tctx->inflight); |
0f212204 JA |
7945 | kfree(tctx); |
7946 | tsk->io_uring = NULL; | |
7947 | } | |
7948 | ||
7e84e1c7 SG |
7949 | static int io_sq_offload_create(struct io_ring_ctx *ctx, |
7950 | struct io_uring_params *p) | |
2b188cc1 JA |
7951 | { |
7952 | int ret; | |
7953 | ||
d25e3a3d JA |
7954 | /* Retain compatibility with failing for an invalid attach attempt */ |
7955 | if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) == | |
7956 | IORING_SETUP_ATTACH_WQ) { | |
7957 | struct fd f; | |
7958 | ||
7959 | f = fdget(p->wq_fd); | |
7960 | if (!f.file) | |
7961 | return -ENXIO; | |
0cc936f7 JA |
7962 | if (f.file->f_op != &io_uring_fops) { |
7963 | fdput(f); | |
f2a48dd0 | 7964 | return -EINVAL; |
0cc936f7 JA |
7965 | } |
7966 | fdput(f); | |
d25e3a3d | 7967 | } |
6c271ce2 | 7968 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
46fe18b1 | 7969 | struct task_struct *tsk; |
534ca6d6 | 7970 | struct io_sq_data *sqd; |
26984fbf | 7971 | bool attached; |
534ca6d6 | 7972 | |
26984fbf | 7973 | sqd = io_get_sq_data(p, &attached); |
534ca6d6 JA |
7974 | if (IS_ERR(sqd)) { |
7975 | ret = PTR_ERR(sqd); | |
7976 | goto err; | |
7977 | } | |
69fb2131 | 7978 | |
7c30f36a | 7979 | ctx->sq_creds = get_current_cred(); |
534ca6d6 | 7980 | ctx->sq_data = sqd; |
917257da JA |
7981 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
7982 | if (!ctx->sq_thread_idle) | |
7983 | ctx->sq_thread_idle = HZ; | |
7984 | ||
78d7f6ba | 7985 | io_sq_thread_park(sqd); |
de75a3d3 PB |
7986 | list_add(&ctx->sqd_list, &sqd->ctx_list); |
7987 | io_sqd_update_thread_idle(sqd); | |
26984fbf | 7988 | /* don't attach to a dying SQPOLL thread, would be racy */ |
f2a48dd0 | 7989 | ret = (attached && !sqd->thread) ? -ENXIO : 0; |
78d7f6ba PB |
7990 | io_sq_thread_unpark(sqd); |
7991 | ||
de75a3d3 PB |
7992 | if (ret < 0) |
7993 | goto err; | |
7994 | if (attached) | |
5aa75ed5 | 7995 | return 0; |
aa06165d | 7996 | |
6c271ce2 | 7997 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 7998 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 7999 | |
917257da | 8000 | ret = -EINVAL; |
f2a48dd0 | 8001 | if (cpu >= nr_cpu_ids || !cpu_online(cpu)) |
e8f98f24 | 8002 | goto err_sqpoll; |
37d1e2e3 | 8003 | sqd->sq_cpu = cpu; |
6c271ce2 | 8004 | } else { |
37d1e2e3 | 8005 | sqd->sq_cpu = -1; |
6c271ce2 | 8006 | } |
37d1e2e3 JA |
8007 | |
8008 | sqd->task_pid = current->pid; | |
5c2469e0 | 8009 | sqd->task_tgid = current->tgid; |
46fe18b1 JA |
8010 | tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE); |
8011 | if (IS_ERR(tsk)) { | |
8012 | ret = PTR_ERR(tsk); | |
e8f98f24 | 8013 | goto err_sqpoll; |
6c271ce2 | 8014 | } |
97a73a0f | 8015 | |
46fe18b1 | 8016 | sqd->thread = tsk; |
97a73a0f | 8017 | ret = io_uring_alloc_task_context(tsk, ctx); |
46fe18b1 | 8018 | wake_up_new_task(tsk); |
0f212204 JA |
8019 | if (ret) |
8020 | goto err; | |
6c271ce2 JA |
8021 | } else if (p->flags & IORING_SETUP_SQ_AFF) { |
8022 | /* Can't have SQ_AFF without SQPOLL */ | |
8023 | ret = -EINVAL; | |
8024 | goto err; | |
8025 | } | |
8026 | ||
2b188cc1 | 8027 | return 0; |
f2a48dd0 PB |
8028 | err_sqpoll: |
8029 | complete(&ctx->sq_data->exited); | |
2b188cc1 | 8030 | err: |
37d1e2e3 | 8031 | io_sq_thread_finish(ctx); |
2b188cc1 JA |
8032 | return ret; |
8033 | } | |
8034 | ||
a087e2b5 BM |
8035 | static inline void __io_unaccount_mem(struct user_struct *user, |
8036 | unsigned long nr_pages) | |
2b188cc1 JA |
8037 | { |
8038 | atomic_long_sub(nr_pages, &user->locked_vm); | |
8039 | } | |
8040 | ||
a087e2b5 BM |
8041 | static inline int __io_account_mem(struct user_struct *user, |
8042 | unsigned long nr_pages) | |
2b188cc1 JA |
8043 | { |
8044 | unsigned long page_limit, cur_pages, new_pages; | |
8045 | ||
8046 | /* Don't allow more pages than we can safely lock */ | |
8047 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
8048 | ||
8049 | do { | |
8050 | cur_pages = atomic_long_read(&user->locked_vm); | |
8051 | new_pages = cur_pages + nr_pages; | |
8052 | if (new_pages > page_limit) | |
8053 | return -ENOMEM; | |
8054 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
8055 | new_pages) != cur_pages); | |
8056 | ||
8057 | return 0; | |
8058 | } | |
8059 | ||
26bfa89e | 8060 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) |
a087e2b5 | 8061 | { |
62e398be | 8062 | if (ctx->user) |
a087e2b5 | 8063 | __io_unaccount_mem(ctx->user, nr_pages); |
30975825 | 8064 | |
26bfa89e JA |
8065 | if (ctx->mm_account) |
8066 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); | |
a087e2b5 BM |
8067 | } |
8068 | ||
26bfa89e | 8069 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) |
a087e2b5 | 8070 | { |
30975825 BM |
8071 | int ret; |
8072 | ||
62e398be | 8073 | if (ctx->user) { |
30975825 BM |
8074 | ret = __io_account_mem(ctx->user, nr_pages); |
8075 | if (ret) | |
8076 | return ret; | |
8077 | } | |
8078 | ||
26bfa89e JA |
8079 | if (ctx->mm_account) |
8080 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); | |
a087e2b5 BM |
8081 | |
8082 | return 0; | |
8083 | } | |
8084 | ||
2b188cc1 JA |
8085 | static void io_mem_free(void *ptr) |
8086 | { | |
52e04ef4 MR |
8087 | struct page *page; |
8088 | ||
8089 | if (!ptr) | |
8090 | return; | |
2b188cc1 | 8091 | |
52e04ef4 | 8092 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
8093 | if (put_page_testzero(page)) |
8094 | free_compound_page(page); | |
8095 | } | |
8096 | ||
8097 | static void *io_mem_alloc(size_t size) | |
8098 | { | |
8099 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
26bfa89e | 8100 | __GFP_NORETRY | __GFP_ACCOUNT; |
2b188cc1 JA |
8101 | |
8102 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
8103 | } | |
8104 | ||
75b28aff HV |
8105 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
8106 | size_t *sq_offset) | |
8107 | { | |
8108 | struct io_rings *rings; | |
8109 | size_t off, sq_array_size; | |
8110 | ||
8111 | off = struct_size(rings, cqes, cq_entries); | |
8112 | if (off == SIZE_MAX) | |
8113 | return SIZE_MAX; | |
8114 | ||
8115 | #ifdef CONFIG_SMP | |
8116 | off = ALIGN(off, SMP_CACHE_BYTES); | |
8117 | if (off == 0) | |
8118 | return SIZE_MAX; | |
8119 | #endif | |
8120 | ||
b36200f5 DV |
8121 | if (sq_offset) |
8122 | *sq_offset = off; | |
8123 | ||
75b28aff HV |
8124 | sq_array_size = array_size(sizeof(u32), sq_entries); |
8125 | if (sq_array_size == SIZE_MAX) | |
8126 | return SIZE_MAX; | |
8127 | ||
8128 | if (check_add_overflow(off, sq_array_size, &off)) | |
8129 | return SIZE_MAX; | |
8130 | ||
75b28aff HV |
8131 | return off; |
8132 | } | |
8133 | ||
41edf1a5 | 8134 | static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot) |
7f61a1e9 | 8135 | { |
41edf1a5 | 8136 | struct io_mapped_ubuf *imu = *slot; |
7f61a1e9 PB |
8137 | unsigned int i; |
8138 | ||
6224843d PB |
8139 | if (imu != ctx->dummy_ubuf) { |
8140 | for (i = 0; i < imu->nr_bvecs; i++) | |
8141 | unpin_user_page(imu->bvec[i].bv_page); | |
8142 | if (imu->acct_pages) | |
8143 | io_unaccount_mem(ctx, imu->acct_pages); | |
8144 | kvfree(imu); | |
8145 | } | |
41edf1a5 | 8146 | *slot = NULL; |
7f61a1e9 PB |
8147 | } |
8148 | ||
bd54b6fe | 8149 | static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) |
edafccee | 8150 | { |
634d00df PB |
8151 | io_buffer_unmap(ctx, &prsrc->buf); |
8152 | prsrc->buf = NULL; | |
bd54b6fe | 8153 | } |
edafccee | 8154 | |
bd54b6fe BM |
8155 | static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) |
8156 | { | |
8157 | unsigned int i; | |
edafccee | 8158 | |
7f61a1e9 PB |
8159 | for (i = 0; i < ctx->nr_user_bufs; i++) |
8160 | io_buffer_unmap(ctx, &ctx->user_bufs[i]); | |
edafccee | 8161 | kfree(ctx->user_bufs); |
bb6659cc | 8162 | io_rsrc_data_free(ctx->buf_data); |
edafccee | 8163 | ctx->user_bufs = NULL; |
bd54b6fe | 8164 | ctx->buf_data = NULL; |
edafccee | 8165 | ctx->nr_user_bufs = 0; |
bd54b6fe BM |
8166 | } |
8167 | ||
0a96bbe4 | 8168 | static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) |
edafccee | 8169 | { |
bd54b6fe | 8170 | int ret; |
edafccee | 8171 | |
bd54b6fe | 8172 | if (!ctx->buf_data) |
edafccee JA |
8173 | return -ENXIO; |
8174 | ||
bd54b6fe BM |
8175 | ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); |
8176 | if (!ret) | |
8177 | __io_sqe_buffers_unregister(ctx); | |
8178 | return ret; | |
edafccee JA |
8179 | } |
8180 | ||
8181 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
8182 | void __user *arg, unsigned index) | |
8183 | { | |
8184 | struct iovec __user *src; | |
8185 | ||
8186 | #ifdef CONFIG_COMPAT | |
8187 | if (ctx->compat) { | |
8188 | struct compat_iovec __user *ciovs; | |
8189 | struct compat_iovec ciov; | |
8190 | ||
8191 | ciovs = (struct compat_iovec __user *) arg; | |
8192 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
8193 | return -EFAULT; | |
8194 | ||
d55e5f5b | 8195 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
8196 | dst->iov_len = ciov.iov_len; |
8197 | return 0; | |
8198 | } | |
8199 | #endif | |
8200 | src = (struct iovec __user *) arg; | |
8201 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
8202 | return -EFAULT; | |
8203 | return 0; | |
8204 | } | |
8205 | ||
de293938 JA |
8206 | /* |
8207 | * Not super efficient, but this is just a registration time. And we do cache | |
8208 | * the last compound head, so generally we'll only do a full search if we don't | |
8209 | * match that one. | |
8210 | * | |
8211 | * We check if the given compound head page has already been accounted, to | |
8212 | * avoid double accounting it. This allows us to account the full size of the | |
8213 | * page, not just the constituent pages of a huge page. | |
8214 | */ | |
8215 | static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, | |
8216 | int nr_pages, struct page *hpage) | |
8217 | { | |
8218 | int i, j; | |
8219 | ||
8220 | /* check current page array */ | |
8221 | for (i = 0; i < nr_pages; i++) { | |
8222 | if (!PageCompound(pages[i])) | |
8223 | continue; | |
8224 | if (compound_head(pages[i]) == hpage) | |
8225 | return true; | |
8226 | } | |
8227 | ||
8228 | /* check previously registered pages */ | |
8229 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
41edf1a5 | 8230 | struct io_mapped_ubuf *imu = ctx->user_bufs[i]; |
de293938 JA |
8231 | |
8232 | for (j = 0; j < imu->nr_bvecs; j++) { | |
8233 | if (!PageCompound(imu->bvec[j].bv_page)) | |
8234 | continue; | |
8235 | if (compound_head(imu->bvec[j].bv_page) == hpage) | |
8236 | return true; | |
8237 | } | |
8238 | } | |
8239 | ||
8240 | return false; | |
8241 | } | |
8242 | ||
8243 | static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, | |
8244 | int nr_pages, struct io_mapped_ubuf *imu, | |
8245 | struct page **last_hpage) | |
8246 | { | |
8247 | int i, ret; | |
8248 | ||
216e5835 | 8249 | imu->acct_pages = 0; |
de293938 JA |
8250 | for (i = 0; i < nr_pages; i++) { |
8251 | if (!PageCompound(pages[i])) { | |
8252 | imu->acct_pages++; | |
8253 | } else { | |
8254 | struct page *hpage; | |
8255 | ||
8256 | hpage = compound_head(pages[i]); | |
8257 | if (hpage == *last_hpage) | |
8258 | continue; | |
8259 | *last_hpage = hpage; | |
8260 | if (headpage_already_acct(ctx, pages, i, hpage)) | |
8261 | continue; | |
8262 | imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; | |
8263 | } | |
8264 | } | |
8265 | ||
8266 | if (!imu->acct_pages) | |
8267 | return 0; | |
8268 | ||
26bfa89e | 8269 | ret = io_account_mem(ctx, imu->acct_pages); |
de293938 JA |
8270 | if (ret) |
8271 | imu->acct_pages = 0; | |
8272 | return ret; | |
8273 | } | |
8274 | ||
0a96bbe4 | 8275 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, |
41edf1a5 | 8276 | struct io_mapped_ubuf **pimu, |
0a96bbe4 | 8277 | struct page **last_hpage) |
edafccee | 8278 | { |
41edf1a5 | 8279 | struct io_mapped_ubuf *imu = NULL; |
edafccee JA |
8280 | struct vm_area_struct **vmas = NULL; |
8281 | struct page **pages = NULL; | |
0a96bbe4 BM |
8282 | unsigned long off, start, end, ubuf; |
8283 | size_t size; | |
8284 | int ret, pret, nr_pages, i; | |
8285 | ||
6224843d PB |
8286 | if (!iov->iov_base) { |
8287 | *pimu = ctx->dummy_ubuf; | |
8288 | return 0; | |
8289 | } | |
8290 | ||
0a96bbe4 BM |
8291 | ubuf = (unsigned long) iov->iov_base; |
8292 | end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
8293 | start = ubuf >> PAGE_SHIFT; | |
8294 | nr_pages = end - start; | |
8295 | ||
41edf1a5 | 8296 | *pimu = NULL; |
0a96bbe4 BM |
8297 | ret = -ENOMEM; |
8298 | ||
8299 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
8300 | if (!pages) | |
8301 | goto done; | |
8302 | ||
8303 | vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), | |
8304 | GFP_KERNEL); | |
8305 | if (!vmas) | |
8306 | goto done; | |
edafccee | 8307 | |
41edf1a5 | 8308 | imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); |
a2b4198c | 8309 | if (!imu) |
0a96bbe4 BM |
8310 | goto done; |
8311 | ||
8312 | ret = 0; | |
8313 | mmap_read_lock(current->mm); | |
8314 | pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
8315 | pages, vmas); | |
8316 | if (pret == nr_pages) { | |
8317 | /* don't support file backed memory */ | |
8318 | for (i = 0; i < nr_pages; i++) { | |
8319 | struct vm_area_struct *vma = vmas[i]; | |
8320 | ||
40dad765 PB |
8321 | if (vma_is_shmem(vma)) |
8322 | continue; | |
0a96bbe4 BM |
8323 | if (vma->vm_file && |
8324 | !is_file_hugepages(vma->vm_file)) { | |
8325 | ret = -EOPNOTSUPP; | |
8326 | break; | |
8327 | } | |
8328 | } | |
8329 | } else { | |
8330 | ret = pret < 0 ? pret : -EFAULT; | |
8331 | } | |
8332 | mmap_read_unlock(current->mm); | |
8333 | if (ret) { | |
8334 | /* | |
8335 | * if we did partial map, or found file backed vmas, | |
8336 | * release any pages we did get | |
8337 | */ | |
8338 | if (pret > 0) | |
8339 | unpin_user_pages(pages, pret); | |
0a96bbe4 BM |
8340 | goto done; |
8341 | } | |
8342 | ||
8343 | ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage); | |
8344 | if (ret) { | |
8345 | unpin_user_pages(pages, pret); | |
0a96bbe4 BM |
8346 | goto done; |
8347 | } | |
8348 | ||
8349 | off = ubuf & ~PAGE_MASK; | |
8350 | size = iov->iov_len; | |
8351 | for (i = 0; i < nr_pages; i++) { | |
8352 | size_t vec_len; | |
8353 | ||
8354 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
8355 | imu->bvec[i].bv_page = pages[i]; | |
8356 | imu->bvec[i].bv_len = vec_len; | |
8357 | imu->bvec[i].bv_offset = off; | |
8358 | off = 0; | |
8359 | size -= vec_len; | |
8360 | } | |
8361 | /* store original address for later verification */ | |
8362 | imu->ubuf = ubuf; | |
4751f53d | 8363 | imu->ubuf_end = ubuf + iov->iov_len; |
0a96bbe4 | 8364 | imu->nr_bvecs = nr_pages; |
41edf1a5 | 8365 | *pimu = imu; |
0a96bbe4 BM |
8366 | ret = 0; |
8367 | done: | |
41edf1a5 PB |
8368 | if (ret) |
8369 | kvfree(imu); | |
0a96bbe4 BM |
8370 | kvfree(pages); |
8371 | kvfree(vmas); | |
8372 | return ret; | |
8373 | } | |
8374 | ||
2b358604 | 8375 | static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) |
0a96bbe4 | 8376 | { |
87094465 PB |
8377 | ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); |
8378 | return ctx->user_bufs ? 0 : -ENOMEM; | |
2b358604 | 8379 | } |
edafccee | 8380 | |
2b358604 BM |
8381 | static int io_buffer_validate(struct iovec *iov) |
8382 | { | |
50e96989 PB |
8383 | unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1); |
8384 | ||
2b358604 BM |
8385 | /* |
8386 | * Don't impose further limits on the size and buffer | |
8387 | * constraints here, we'll -EINVAL later when IO is | |
8388 | * submitted if they are wrong. | |
8389 | */ | |
6224843d PB |
8390 | if (!iov->iov_base) |
8391 | return iov->iov_len ? -EFAULT : 0; | |
8392 | if (!iov->iov_len) | |
2b358604 | 8393 | return -EFAULT; |
edafccee | 8394 | |
2b358604 BM |
8395 | /* arbitrary limit, but we need something */ |
8396 | if (iov->iov_len > SZ_1G) | |
8397 | return -EFAULT; | |
edafccee | 8398 | |
50e96989 PB |
8399 | if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp)) |
8400 | return -EOVERFLOW; | |
8401 | ||
2b358604 BM |
8402 | return 0; |
8403 | } | |
edafccee | 8404 | |
2b358604 | 8405 | static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, |
634d00df | 8406 | unsigned int nr_args, u64 __user *tags) |
2b358604 | 8407 | { |
bd54b6fe BM |
8408 | struct page *last_hpage = NULL; |
8409 | struct io_rsrc_data *data; | |
2b358604 BM |
8410 | int i, ret; |
8411 | struct iovec iov; | |
edafccee | 8412 | |
87094465 PB |
8413 | if (ctx->user_bufs) |
8414 | return -EBUSY; | |
489809e2 | 8415 | if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) |
87094465 | 8416 | return -EINVAL; |
bd54b6fe | 8417 | ret = io_rsrc_node_switch_start(ctx); |
2b358604 BM |
8418 | if (ret) |
8419 | return ret; | |
d878c816 PB |
8420 | ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); |
8421 | if (ret) | |
8422 | return ret; | |
bd54b6fe BM |
8423 | ret = io_buffers_map_alloc(ctx, nr_args); |
8424 | if (ret) { | |
bb6659cc | 8425 | io_rsrc_data_free(data); |
bd54b6fe BM |
8426 | return ret; |
8427 | } | |
edafccee | 8428 | |
87094465 | 8429 | for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { |
edafccee JA |
8430 | ret = io_copy_iov(ctx, &iov, arg, i); |
8431 | if (ret) | |
0a96bbe4 | 8432 | break; |
2b358604 BM |
8433 | ret = io_buffer_validate(&iov); |
8434 | if (ret) | |
0a96bbe4 | 8435 | break; |
2d091d62 | 8436 | if (!iov.iov_base && *io_get_tag_slot(data, i)) { |
cf3770e7 CIK |
8437 | ret = -EINVAL; |
8438 | break; | |
8439 | } | |
edafccee | 8440 | |
41edf1a5 PB |
8441 | ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], |
8442 | &last_hpage); | |
0a96bbe4 BM |
8443 | if (ret) |
8444 | break; | |
edafccee | 8445 | } |
0a96bbe4 | 8446 | |
bd54b6fe | 8447 | WARN_ON_ONCE(ctx->buf_data); |
0a96bbe4 | 8448 | |
bd54b6fe BM |
8449 | ctx->buf_data = data; |
8450 | if (ret) | |
8451 | __io_sqe_buffers_unregister(ctx); | |
8452 | else | |
8453 | io_rsrc_node_switch(ctx, NULL); | |
edafccee JA |
8454 | return ret; |
8455 | } | |
8456 | ||
634d00df PB |
8457 | static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, |
8458 | struct io_uring_rsrc_update2 *up, | |
8459 | unsigned int nr_args) | |
8460 | { | |
8461 | u64 __user *tags = u64_to_user_ptr(up->tags); | |
8462 | struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); | |
634d00df PB |
8463 | struct page *last_hpage = NULL; |
8464 | bool needs_switch = false; | |
8465 | __u32 done; | |
8466 | int i, err; | |
8467 | ||
8468 | if (!ctx->buf_data) | |
8469 | return -ENXIO; | |
8470 | if (up->offset + nr_args > ctx->nr_user_bufs) | |
8471 | return -EINVAL; | |
8472 | ||
8473 | for (done = 0; done < nr_args; done++) { | |
0b8c0e7c PB |
8474 | struct io_mapped_ubuf *imu; |
8475 | int offset = up->offset + done; | |
634d00df PB |
8476 | u64 tag = 0; |
8477 | ||
8478 | err = io_copy_iov(ctx, &iov, iovs, done); | |
8479 | if (err) | |
8480 | break; | |
8481 | if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { | |
8482 | err = -EFAULT; | |
8483 | break; | |
8484 | } | |
0b8c0e7c PB |
8485 | err = io_buffer_validate(&iov); |
8486 | if (err) | |
8487 | break; | |
cf3770e7 CIK |
8488 | if (!iov.iov_base && tag) { |
8489 | err = -EINVAL; | |
8490 | break; | |
8491 | } | |
0b8c0e7c PB |
8492 | err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); |
8493 | if (err) | |
8494 | break; | |
634d00df | 8495 | |
0b8c0e7c | 8496 | i = array_index_nospec(offset, ctx->nr_user_bufs); |
6224843d | 8497 | if (ctx->user_bufs[i] != ctx->dummy_ubuf) { |
0b8c0e7c PB |
8498 | err = io_queue_rsrc_removal(ctx->buf_data, offset, |
8499 | ctx->rsrc_node, ctx->user_bufs[i]); | |
8500 | if (unlikely(err)) { | |
8501 | io_buffer_unmap(ctx, &imu); | |
634d00df | 8502 | break; |
0b8c0e7c | 8503 | } |
634d00df PB |
8504 | ctx->user_bufs[i] = NULL; |
8505 | needs_switch = true; | |
8506 | } | |
8507 | ||
0b8c0e7c | 8508 | ctx->user_bufs[i] = imu; |
2d091d62 | 8509 | *io_get_tag_slot(ctx->buf_data, offset) = tag; |
634d00df PB |
8510 | } |
8511 | ||
8512 | if (needs_switch) | |
8513 | io_rsrc_node_switch(ctx, ctx->buf_data); | |
8514 | return done ? done : err; | |
8515 | } | |
8516 | ||
9b402849 JA |
8517 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
8518 | { | |
8519 | __s32 __user *fds = arg; | |
8520 | int fd; | |
8521 | ||
8522 | if (ctx->cq_ev_fd) | |
8523 | return -EBUSY; | |
8524 | ||
8525 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
8526 | return -EFAULT; | |
8527 | ||
8528 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
8529 | if (IS_ERR(ctx->cq_ev_fd)) { | |
8530 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
fe7e3257 | 8531 | |
9b402849 JA |
8532 | ctx->cq_ev_fd = NULL; |
8533 | return ret; | |
8534 | } | |
8535 | ||
8536 | return 0; | |
8537 | } | |
8538 | ||
8539 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
8540 | { | |
8541 | if (ctx->cq_ev_fd) { | |
8542 | eventfd_ctx_put(ctx->cq_ev_fd); | |
8543 | ctx->cq_ev_fd = NULL; | |
8544 | return 0; | |
8545 | } | |
8546 | ||
8547 | return -ENXIO; | |
8548 | } | |
8549 | ||
5a2e745d JA |
8550 | static void io_destroy_buffers(struct io_ring_ctx *ctx) |
8551 | { | |
9e15c3a0 JA |
8552 | struct io_buffer *buf; |
8553 | unsigned long index; | |
8554 | ||
8555 | xa_for_each(&ctx->io_buffers, index, buf) | |
8556 | __io_remove_buffers(ctx, buf, index, -1U); | |
5a2e745d JA |
8557 | } |
8558 | ||
7255834e | 8559 | static void io_req_cache_free(struct list_head *list) |
1b4c351f | 8560 | { |
68e68ee6 | 8561 | struct io_kiocb *req, *nxt; |
1b4c351f | 8562 | |
bb943b82 PB |
8563 | list_for_each_entry_safe(req, nxt, list, inflight_entry) { |
8564 | list_del(&req->inflight_entry); | |
1b4c351f JA |
8565 | kmem_cache_free(req_cachep, req); |
8566 | } | |
8567 | } | |
8568 | ||
4010fec4 | 8569 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 8570 | { |
cd0ca2e0 | 8571 | struct io_submit_state *state = &ctx->submit_state; |
bf019da7 | 8572 | |
9a4fdbd8 JA |
8573 | mutex_lock(&ctx->uring_lock); |
8574 | ||
cd0ca2e0 PB |
8575 | if (state->free_reqs) { |
8576 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); | |
8577 | state->free_reqs = 0; | |
8e5c66c4 | 8578 | } |
9a4fdbd8 | 8579 | |
cd0ca2e0 PB |
8580 | io_flush_cached_locked_reqs(ctx, state); |
8581 | io_req_cache_free(&state->free_list); | |
9a4fdbd8 JA |
8582 | mutex_unlock(&ctx->uring_lock); |
8583 | } | |
8584 | ||
43597aac | 8585 | static void io_wait_rsrc_data(struct io_rsrc_data *data) |
2b188cc1 | 8586 | { |
43597aac | 8587 | if (data && !atomic_dec_and_test(&data->refs)) |
bd54b6fe | 8588 | wait_for_completion(&data->done); |
bd54b6fe | 8589 | } |
04fc6c80 | 8590 | |
2b188cc1 JA |
8591 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
8592 | { | |
37d1e2e3 | 8593 | io_sq_thread_finish(ctx); |
2aede0e4 | 8594 | |
37d1e2e3 | 8595 | if (ctx->mm_account) { |
2aede0e4 JA |
8596 | mmdrop(ctx->mm_account); |
8597 | ctx->mm_account = NULL; | |
30975825 | 8598 | } |
def596e9 | 8599 | |
43597aac PB |
8600 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
8601 | io_wait_rsrc_data(ctx->buf_data); | |
8602 | io_wait_rsrc_data(ctx->file_data); | |
8603 | ||
8bad28d8 | 8604 | mutex_lock(&ctx->uring_lock); |
43597aac | 8605 | if (ctx->buf_data) |
bd54b6fe | 8606 | __io_sqe_buffers_unregister(ctx); |
43597aac | 8607 | if (ctx->file_data) |
08480400 | 8608 | __io_sqe_files_unregister(ctx); |
c4ea060e PB |
8609 | if (ctx->rings) |
8610 | __io_cqring_overflow_flush(ctx, true); | |
8bad28d8 | 8611 | mutex_unlock(&ctx->uring_lock); |
9b402849 | 8612 | io_eventfd_unregister(ctx); |
5a2e745d | 8613 | io_destroy_buffers(ctx); |
07db298a PB |
8614 | if (ctx->sq_creds) |
8615 | put_cred(ctx->sq_creds); | |
def596e9 | 8616 | |
a7f0ed5a PB |
8617 | /* there are no registered resources left, nobody uses it */ |
8618 | if (ctx->rsrc_node) | |
8619 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 8620 | if (ctx->rsrc_backup_node) |
b895c9a6 | 8621 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a PB |
8622 | flush_delayed_work(&ctx->rsrc_put_work); |
8623 | ||
8624 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
8625 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 8626 | |
2b188cc1 | 8627 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
8628 | if (ctx->ring_sock) { |
8629 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 8630 | sock_release(ctx->ring_sock); |
355e8d26 | 8631 | } |
2b188cc1 JA |
8632 | #endif |
8633 | ||
75b28aff | 8634 | io_mem_free(ctx->rings); |
2b188cc1 | 8635 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
8636 | |
8637 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 8638 | free_uid(ctx->user); |
4010fec4 | 8639 | io_req_caches_free(ctx); |
e941894e JA |
8640 | if (ctx->hash_map) |
8641 | io_wq_put_hash(ctx->hash_map); | |
78076bb6 | 8642 | kfree(ctx->cancel_hash); |
6224843d | 8643 | kfree(ctx->dummy_ubuf); |
2b188cc1 JA |
8644 | kfree(ctx); |
8645 | } | |
8646 | ||
8647 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
8648 | { | |
8649 | struct io_ring_ctx *ctx = file->private_data; | |
8650 | __poll_t mask = 0; | |
8651 | ||
311997b3 | 8652 | poll_wait(file, &ctx->poll_wait, wait); |
4f7067c3 SB |
8653 | /* |
8654 | * synchronizes with barrier from wq_has_sleeper call in | |
8655 | * io_commit_cqring | |
8656 | */ | |
2b188cc1 | 8657 | smp_rmb(); |
90554200 | 8658 | if (!io_sqring_full(ctx)) |
2b188cc1 | 8659 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
8660 | |
8661 | /* | |
8662 | * Don't flush cqring overflow list here, just do a simple check. | |
8663 | * Otherwise there could possible be ABBA deadlock: | |
8664 | * CPU0 CPU1 | |
8665 | * ---- ---- | |
8666 | * lock(&ctx->uring_lock); | |
8667 | * lock(&ep->mtx); | |
8668 | * lock(&ctx->uring_lock); | |
8669 | * lock(&ep->mtx); | |
8670 | * | |
8671 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
8672 | * pushs them to do the flush. | |
8673 | */ | |
5ed7a37d | 8674 | if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow)) |
2b188cc1 JA |
8675 | mask |= EPOLLIN | EPOLLRDNORM; |
8676 | ||
8677 | return mask; | |
8678 | } | |
8679 | ||
8680 | static int io_uring_fasync(int fd, struct file *file, int on) | |
8681 | { | |
8682 | struct io_ring_ctx *ctx = file->private_data; | |
8683 | ||
8684 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
8685 | } | |
8686 | ||
0bead8cd | 8687 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 8688 | { |
4379bf8b | 8689 | const struct cred *creds; |
071698e1 | 8690 | |
61cf9370 | 8691 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
8692 | if (creds) { |
8693 | put_cred(creds); | |
0bead8cd | 8694 | return 0; |
1e6fa521 | 8695 | } |
0bead8cd YD |
8696 | |
8697 | return -EINVAL; | |
8698 | } | |
8699 | ||
d56d938b PB |
8700 | struct io_tctx_exit { |
8701 | struct callback_head task_work; | |
8702 | struct completion completion; | |
baf186c4 | 8703 | struct io_ring_ctx *ctx; |
d56d938b PB |
8704 | }; |
8705 | ||
8706 | static void io_tctx_exit_cb(struct callback_head *cb) | |
8707 | { | |
8708 | struct io_uring_task *tctx = current->io_uring; | |
8709 | struct io_tctx_exit *work; | |
8710 | ||
8711 | work = container_of(cb, struct io_tctx_exit, task_work); | |
8712 | /* | |
8713 | * When @in_idle, we're in cancellation and it's racy to remove the | |
8714 | * node. It'll be removed by the end of cancellation, just ignore it. | |
8715 | */ | |
8716 | if (!atomic_read(&tctx->in_idle)) | |
eef51daa | 8717 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
8718 | complete(&work->completion); |
8719 | } | |
8720 | ||
28090c13 PB |
8721 | static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
8722 | { | |
8723 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
8724 | ||
8725 | return req->ctx == data; | |
8726 | } | |
8727 | ||
85faa7b8 JA |
8728 | static void io_ring_exit_work(struct work_struct *work) |
8729 | { | |
d56d938b | 8730 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 8731 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 8732 | unsigned long interval = HZ / 20; |
d56d938b PB |
8733 | struct io_tctx_exit exit; |
8734 | struct io_tctx_node *node; | |
8735 | int ret; | |
85faa7b8 | 8736 | |
56952e91 JA |
8737 | /* |
8738 | * If we're doing polled IO and end up having requests being | |
8739 | * submitted async (out-of-line), then completions can come in while | |
8740 | * we're waiting for refs to drop. We need to reap these manually, | |
8741 | * as nobody else will be looking for them. | |
8742 | */ | |
b2edc0a7 | 8743 | do { |
3dd0c97a | 8744 | io_uring_try_cancel_requests(ctx, NULL, true); |
28090c13 PB |
8745 | if (ctx->sq_data) { |
8746 | struct io_sq_data *sqd = ctx->sq_data; | |
8747 | struct task_struct *tsk; | |
8748 | ||
8749 | io_sq_thread_park(sqd); | |
8750 | tsk = sqd->thread; | |
8751 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
8752 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
8753 | io_cancel_ctx_cb, ctx, true); | |
8754 | io_sq_thread_unpark(sqd); | |
8755 | } | |
b5bb3a24 | 8756 | |
58d3be2c PB |
8757 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
8758 | /* there is little hope left, don't run it too often */ | |
8759 | interval = HZ * 60; | |
8760 | } | |
8761 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 8762 | |
7f00651a PB |
8763 | init_completion(&exit.completion); |
8764 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
8765 | exit.ctx = ctx; | |
89b5066e PB |
8766 | /* |
8767 | * Some may use context even when all refs and requests have been put, | |
8768 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 8769 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
8770 | * this lock/unlock section also waits them to finish. |
8771 | */ | |
d56d938b PB |
8772 | mutex_lock(&ctx->uring_lock); |
8773 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
8774 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
8775 | ||
d56d938b PB |
8776 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
8777 | ctx_node); | |
7f00651a PB |
8778 | /* don't spin on a single task if cancellation failed */ |
8779 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
8780 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
8781 | if (WARN_ON_ONCE(ret)) | |
8782 | continue; | |
8783 | wake_up_process(node->task); | |
8784 | ||
8785 | mutex_unlock(&ctx->uring_lock); | |
8786 | wait_for_completion(&exit.completion); | |
d56d938b PB |
8787 | mutex_lock(&ctx->uring_lock); |
8788 | } | |
8789 | mutex_unlock(&ctx->uring_lock); | |
89b5066e PB |
8790 | spin_lock_irq(&ctx->completion_lock); |
8791 | spin_unlock_irq(&ctx->completion_lock); | |
d56d938b | 8792 | |
85faa7b8 JA |
8793 | io_ring_ctx_free(ctx); |
8794 | } | |
8795 | ||
80c4cbdb PB |
8796 | /* Returns true if we found and killed one or more timeouts */ |
8797 | static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
3dd0c97a | 8798 | bool cancel_all) |
80c4cbdb PB |
8799 | { |
8800 | struct io_kiocb *req, *tmp; | |
8801 | int canceled = 0; | |
8802 | ||
8803 | spin_lock_irq(&ctx->completion_lock); | |
8804 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) { | |
3dd0c97a | 8805 | if (io_match_task(req, tsk, cancel_all)) { |
80c4cbdb PB |
8806 | io_kill_timeout(req, -ECANCELED); |
8807 | canceled++; | |
8808 | } | |
8809 | } | |
51520426 PB |
8810 | if (canceled != 0) |
8811 | io_commit_cqring(ctx); | |
80c4cbdb | 8812 | spin_unlock_irq(&ctx->completion_lock); |
80c4cbdb PB |
8813 | if (canceled != 0) |
8814 | io_cqring_ev_posted(ctx); | |
8815 | return canceled != 0; | |
8816 | } | |
8817 | ||
2b188cc1 JA |
8818 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
8819 | { | |
61cf9370 MWO |
8820 | unsigned long index; |
8821 | struct creds *creds; | |
8822 | ||
2b188cc1 JA |
8823 | mutex_lock(&ctx->uring_lock); |
8824 | percpu_ref_kill(&ctx->refs); | |
634578f8 | 8825 | if (ctx->rings) |
6c2450ae | 8826 | __io_cqring_overflow_flush(ctx, true); |
61cf9370 MWO |
8827 | xa_for_each(&ctx->personalities, index, creds) |
8828 | io_unregister_personality(ctx, index); | |
2b188cc1 JA |
8829 | mutex_unlock(&ctx->uring_lock); |
8830 | ||
3dd0c97a PB |
8831 | io_kill_timeouts(ctx, NULL, true); |
8832 | io_poll_remove_all(ctx, NULL, true); | |
561fb04a | 8833 | |
15dff286 | 8834 | /* if we failed setting up the ctx, we might not have any rings */ |
b2edc0a7 | 8835 | io_iopoll_try_reap_events(ctx); |
309fc03a | 8836 | |
85faa7b8 | 8837 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
8838 | /* |
8839 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
8840 | * if we're exiting a ton of rings at the same time. It just adds | |
8841 | * noise and overhead, there's no discernable change in runtime | |
8842 | * over using system_wq. | |
8843 | */ | |
8844 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
8845 | } |
8846 | ||
8847 | static int io_uring_release(struct inode *inode, struct file *file) | |
8848 | { | |
8849 | struct io_ring_ctx *ctx = file->private_data; | |
8850 | ||
8851 | file->private_data = NULL; | |
8852 | io_ring_ctx_wait_and_kill(ctx); | |
8853 | return 0; | |
8854 | } | |
8855 | ||
f6edbabb PB |
8856 | struct io_task_cancel { |
8857 | struct task_struct *task; | |
3dd0c97a | 8858 | bool all; |
f6edbabb | 8859 | }; |
f254ac04 | 8860 | |
f6edbabb | 8861 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 8862 | { |
9a472ef7 | 8863 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 8864 | struct io_task_cancel *cancel = data; |
9a472ef7 PB |
8865 | bool ret; |
8866 | ||
3dd0c97a | 8867 | if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) { |
9a472ef7 PB |
8868 | unsigned long flags; |
8869 | struct io_ring_ctx *ctx = req->ctx; | |
8870 | ||
8871 | /* protect against races with linked timeouts */ | |
8872 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3dd0c97a | 8873 | ret = io_match_task(req, cancel->task, cancel->all); |
9a472ef7 PB |
8874 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
8875 | } else { | |
3dd0c97a | 8876 | ret = io_match_task(req, cancel->task, cancel->all); |
9a472ef7 PB |
8877 | } |
8878 | return ret; | |
b711d4ea JA |
8879 | } |
8880 | ||
e1915f76 | 8881 | static bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3dd0c97a | 8882 | struct task_struct *task, bool cancel_all) |
b7ddce3c | 8883 | { |
e1915f76 | 8884 | struct io_defer_entry *de; |
b7ddce3c PB |
8885 | LIST_HEAD(list); |
8886 | ||
8887 | spin_lock_irq(&ctx->completion_lock); | |
8888 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { | |
3dd0c97a | 8889 | if (io_match_task(de->req, task, cancel_all)) { |
b7ddce3c PB |
8890 | list_cut_position(&list, &ctx->defer_list, &de->list); |
8891 | break; | |
8892 | } | |
8893 | } | |
8894 | spin_unlock_irq(&ctx->completion_lock); | |
e1915f76 PB |
8895 | if (list_empty(&list)) |
8896 | return false; | |
b7ddce3c PB |
8897 | |
8898 | while (!list_empty(&list)) { | |
8899 | de = list_first_entry(&list, struct io_defer_entry, list); | |
8900 | list_del_init(&de->list); | |
f41db273 | 8901 | io_req_complete_failed(de->req, -ECANCELED); |
b7ddce3c PB |
8902 | kfree(de); |
8903 | } | |
e1915f76 | 8904 | return true; |
b7ddce3c PB |
8905 | } |
8906 | ||
1b00764f PB |
8907 | static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
8908 | { | |
8909 | struct io_tctx_node *node; | |
8910 | enum io_wq_cancel cret; | |
8911 | bool ret = false; | |
8912 | ||
8913 | mutex_lock(&ctx->uring_lock); | |
8914 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
8915 | struct io_uring_task *tctx = node->task->io_uring; | |
8916 | ||
8917 | /* | |
8918 | * io_wq will stay alive while we hold uring_lock, because it's | |
8919 | * killed after ctx nodes, which requires to take the lock. | |
8920 | */ | |
8921 | if (!tctx || !tctx->io_wq) | |
8922 | continue; | |
8923 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
8924 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
8925 | } | |
8926 | mutex_unlock(&ctx->uring_lock); | |
8927 | ||
8928 | return ret; | |
8929 | } | |
8930 | ||
9936c7c2 PB |
8931 | static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
8932 | struct task_struct *task, | |
3dd0c97a | 8933 | bool cancel_all) |
9936c7c2 | 8934 | { |
3dd0c97a | 8935 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 8936 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
9936c7c2 PB |
8937 | |
8938 | while (1) { | |
8939 | enum io_wq_cancel cret; | |
8940 | bool ret = false; | |
8941 | ||
1b00764f PB |
8942 | if (!task) { |
8943 | ret |= io_uring_try_cancel_iowq(ctx); | |
8944 | } else if (tctx && tctx->io_wq) { | |
8945 | /* | |
8946 | * Cancels requests of all rings, not only @ctx, but | |
8947 | * it's fine as the task is in exit/exec. | |
8948 | */ | |
5aa75ed5 | 8949 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, |
9936c7c2 PB |
8950 | &cancel, true); |
8951 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
8952 | } | |
8953 | ||
8954 | /* SQPOLL thread does its own polling */ | |
3dd0c97a | 8955 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || |
d052d1d6 | 8956 | (ctx->sq_data && ctx->sq_data->thread == current)) { |
9936c7c2 PB |
8957 | while (!list_empty_careful(&ctx->iopoll_list)) { |
8958 | io_iopoll_try_reap_events(ctx); | |
8959 | ret = true; | |
8960 | } | |
8961 | } | |
8962 | ||
3dd0c97a PB |
8963 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
8964 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
8965 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
e5dc480d PB |
8966 | if (task) |
8967 | ret |= io_run_task_work(); | |
9936c7c2 PB |
8968 | if (!ret) |
8969 | break; | |
8970 | cond_resched(); | |
8971 | } | |
8972 | } | |
8973 | ||
eef51daa | 8974 | static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx) |
0f212204 | 8975 | { |
236434c3 | 8976 | struct io_uring_task *tctx = current->io_uring; |
13bf43f5 | 8977 | struct io_tctx_node *node; |
a528b04e | 8978 | int ret; |
236434c3 MWO |
8979 | |
8980 | if (unlikely(!tctx)) { | |
5aa75ed5 | 8981 | ret = io_uring_alloc_task_context(current, ctx); |
0f212204 JA |
8982 | if (unlikely(ret)) |
8983 | return ret; | |
236434c3 | 8984 | tctx = current->io_uring; |
0f212204 | 8985 | } |
cf27f3b1 PB |
8986 | if (!xa_load(&tctx->xa, (unsigned long)ctx)) { |
8987 | node = kmalloc(sizeof(*node), GFP_KERNEL); | |
8988 | if (!node) | |
8989 | return -ENOMEM; | |
8990 | node->ctx = ctx; | |
8991 | node->task = current; | |
13bf43f5 | 8992 | |
cf27f3b1 PB |
8993 | ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx, |
8994 | node, GFP_KERNEL)); | |
8995 | if (ret) { | |
8996 | kfree(node); | |
8997 | return ret; | |
0f212204 | 8998 | } |
cf27f3b1 PB |
8999 | |
9000 | mutex_lock(&ctx->uring_lock); | |
9001 | list_add(&node->ctx_node, &ctx->tctx_list); | |
9002 | mutex_unlock(&ctx->uring_lock); | |
0f212204 | 9003 | } |
cf27f3b1 | 9004 | tctx->last = ctx; |
0f212204 JA |
9005 | return 0; |
9006 | } | |
9007 | ||
cf27f3b1 PB |
9008 | /* |
9009 | * Note that this task has used io_uring. We use it for cancelation purposes. | |
9010 | */ | |
eef51daa | 9011 | static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx) |
cf27f3b1 PB |
9012 | { |
9013 | struct io_uring_task *tctx = current->io_uring; | |
9014 | ||
9015 | if (likely(tctx && tctx->last == ctx)) | |
9016 | return 0; | |
eef51daa | 9017 | return __io_uring_add_tctx_node(ctx); |
cf27f3b1 PB |
9018 | } |
9019 | ||
0f212204 JA |
9020 | /* |
9021 | * Remove this io_uring_file -> task mapping. | |
9022 | */ | |
eef51daa | 9023 | static void io_uring_del_tctx_node(unsigned long index) |
0f212204 JA |
9024 | { |
9025 | struct io_uring_task *tctx = current->io_uring; | |
13bf43f5 | 9026 | struct io_tctx_node *node; |
2941267b | 9027 | |
eebd2e37 PB |
9028 | if (!tctx) |
9029 | return; | |
13bf43f5 PB |
9030 | node = xa_erase(&tctx->xa, index); |
9031 | if (!node) | |
2941267b | 9032 | return; |
0f212204 | 9033 | |
13bf43f5 PB |
9034 | WARN_ON_ONCE(current != node->task); |
9035 | WARN_ON_ONCE(list_empty(&node->ctx_node)); | |
9036 | ||
9037 | mutex_lock(&node->ctx->uring_lock); | |
9038 | list_del(&node->ctx_node); | |
9039 | mutex_unlock(&node->ctx->uring_lock); | |
9040 | ||
baf186c4 | 9041 | if (tctx->last == node->ctx) |
0f212204 | 9042 | tctx->last = NULL; |
13bf43f5 | 9043 | kfree(node); |
0f212204 JA |
9044 | } |
9045 | ||
8452d4a6 | 9046 | static void io_uring_clean_tctx(struct io_uring_task *tctx) |
de7f1d9e | 9047 | { |
ba5ef6dc | 9048 | struct io_wq *wq = tctx->io_wq; |
13bf43f5 | 9049 | struct io_tctx_node *node; |
de7f1d9e PB |
9050 | unsigned long index; |
9051 | ||
13bf43f5 | 9052 | xa_for_each(&tctx->xa, index, node) |
eef51daa | 9053 | io_uring_del_tctx_node(index); |
b16ef427 ME |
9054 | if (wq) { |
9055 | /* | |
9056 | * Must be after io_uring_del_task_file() (removes nodes under | |
9057 | * uring_lock) to avoid race with io_uring_try_cancel_iowq(). | |
9058 | */ | |
9059 | tctx->io_wq = NULL; | |
ba5ef6dc | 9060 | io_wq_put_and_exit(wq); |
b16ef427 | 9061 | } |
de7f1d9e PB |
9062 | } |
9063 | ||
3f48cf18 | 9064 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 9065 | { |
3f48cf18 PB |
9066 | if (tracked) |
9067 | return atomic_read(&tctx->inflight_tracked); | |
521d6a73 PB |
9068 | return percpu_counter_sum(&tctx->inflight); |
9069 | } | |
9070 | ||
09899b19 PB |
9071 | static void io_uring_drop_tctx_refs(struct task_struct *task) |
9072 | { | |
9073 | struct io_uring_task *tctx = task->io_uring; | |
9074 | unsigned int refs = tctx->cached_refs; | |
9075 | ||
e9dbe221 PB |
9076 | if (refs) { |
9077 | tctx->cached_refs = 0; | |
9078 | percpu_counter_sub(&tctx->inflight, refs); | |
9079 | put_task_struct_many(task, refs); | |
9080 | } | |
09899b19 PB |
9081 | } |
9082 | ||
78cc687b PB |
9083 | /* |
9084 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
9085 | * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation. | |
9086 | */ | |
9087 | static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) | |
0e9ddb39 | 9088 | { |
521d6a73 | 9089 | struct io_uring_task *tctx = current->io_uring; |
734551df | 9090 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
9091 | s64 inflight; |
9092 | DEFINE_WAIT(wait); | |
fdaf083c | 9093 | |
78cc687b PB |
9094 | WARN_ON_ONCE(sqd && sqd->thread != current); |
9095 | ||
6d042ffb PO |
9096 | if (!current->io_uring) |
9097 | return; | |
17a91051 PB |
9098 | if (tctx->io_wq) |
9099 | io_wq_exit_start(tctx->io_wq); | |
9100 | ||
0e9ddb39 PB |
9101 | atomic_inc(&tctx->in_idle); |
9102 | do { | |
e9dbe221 | 9103 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 9104 | /* read completions before cancelations */ |
78cc687b | 9105 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
9106 | if (!inflight) |
9107 | break; | |
fdaf083c | 9108 | |
78cc687b PB |
9109 | if (!sqd) { |
9110 | struct io_tctx_node *node; | |
9111 | unsigned long index; | |
0f212204 | 9112 | |
78cc687b PB |
9113 | xa_for_each(&tctx->xa, index, node) { |
9114 | /* sqpoll task will cancel all its requests */ | |
9115 | if (node->ctx->sq_data) | |
9116 | continue; | |
9117 | io_uring_try_cancel_requests(node->ctx, current, | |
9118 | cancel_all); | |
9119 | } | |
9120 | } else { | |
9121 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
9122 | io_uring_try_cancel_requests(ctx, current, | |
9123 | cancel_all); | |
9124 | } | |
17a91051 | 9125 | |
0f212204 | 9126 | prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE); |
e9dbe221 | 9127 | io_uring_drop_tctx_refs(current); |
0f212204 | 9128 | /* |
a1bb3cd5 PB |
9129 | * If we've seen completions, retry without waiting. This |
9130 | * avoids a race where a completion comes in before we did | |
9131 | * prepare_to_wait(). | |
0f212204 | 9132 | */ |
3dd0c97a | 9133 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 9134 | schedule(); |
f57555ed | 9135 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 9136 | } while (1); |
fdaf083c | 9137 | atomic_dec(&tctx->in_idle); |
de7f1d9e | 9138 | |
8452d4a6 | 9139 | io_uring_clean_tctx(tctx); |
3dd0c97a | 9140 | if (cancel_all) { |
3f48cf18 PB |
9141 | /* for exec all current's requests should be gone, kill tctx */ |
9142 | __io_uring_free(current); | |
9143 | } | |
44e728b8 PB |
9144 | } |
9145 | ||
78cc687b PB |
9146 | void __io_uring_cancel(struct files_struct *files) |
9147 | { | |
9148 | io_uring_cancel_generic(!files, NULL); | |
9149 | } | |
9150 | ||
6c5c240e RP |
9151 | static void *io_uring_validate_mmap_request(struct file *file, |
9152 | loff_t pgoff, size_t sz) | |
2b188cc1 | 9153 | { |
2b188cc1 | 9154 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 9155 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
9156 | struct page *page; |
9157 | void *ptr; | |
9158 | ||
9159 | switch (offset) { | |
9160 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
9161 | case IORING_OFF_CQ_RING: |
9162 | ptr = ctx->rings; | |
2b188cc1 JA |
9163 | break; |
9164 | case IORING_OFF_SQES: | |
9165 | ptr = ctx->sq_sqes; | |
9166 | break; | |
2b188cc1 | 9167 | default: |
6c5c240e | 9168 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
9169 | } |
9170 | ||
9171 | page = virt_to_head_page(ptr); | |
a50b854e | 9172 | if (sz > page_size(page)) |
6c5c240e RP |
9173 | return ERR_PTR(-EINVAL); |
9174 | ||
9175 | return ptr; | |
9176 | } | |
9177 | ||
9178 | #ifdef CONFIG_MMU | |
9179 | ||
9180 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9181 | { | |
9182 | size_t sz = vma->vm_end - vma->vm_start; | |
9183 | unsigned long pfn; | |
9184 | void *ptr; | |
9185 | ||
9186 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
9187 | if (IS_ERR(ptr)) | |
9188 | return PTR_ERR(ptr); | |
2b188cc1 JA |
9189 | |
9190 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
9191 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
9192 | } | |
9193 | ||
6c5c240e RP |
9194 | #else /* !CONFIG_MMU */ |
9195 | ||
9196 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9197 | { | |
9198 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
9199 | } | |
9200 | ||
9201 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
9202 | { | |
9203 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
9204 | } | |
9205 | ||
9206 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
9207 | unsigned long addr, unsigned long len, | |
9208 | unsigned long pgoff, unsigned long flags) | |
9209 | { | |
9210 | void *ptr; | |
9211 | ||
9212 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
9213 | if (IS_ERR(ptr)) | |
9214 | return PTR_ERR(ptr); | |
9215 | ||
9216 | return (unsigned long) ptr; | |
9217 | } | |
9218 | ||
9219 | #endif /* !CONFIG_MMU */ | |
9220 | ||
d9d05217 | 9221 | static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx) |
90554200 JA |
9222 | { |
9223 | DEFINE_WAIT(wait); | |
9224 | ||
9225 | do { | |
9226 | if (!io_sqring_full(ctx)) | |
9227 | break; | |
90554200 JA |
9228 | prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE); |
9229 | ||
9230 | if (!io_sqring_full(ctx)) | |
9231 | break; | |
90554200 JA |
9232 | schedule(); |
9233 | } while (!signal_pending(current)); | |
9234 | ||
9235 | finish_wait(&ctx->sqo_sq_wait, &wait); | |
5199328a | 9236 | return 0; |
90554200 JA |
9237 | } |
9238 | ||
c73ebb68 HX |
9239 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
9240 | struct __kernel_timespec __user **ts, | |
9241 | const sigset_t __user **sig) | |
9242 | { | |
9243 | struct io_uring_getevents_arg arg; | |
9244 | ||
9245 | /* | |
9246 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
9247 | * is just a pointer to the sigset_t. | |
9248 | */ | |
9249 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
9250 | *sig = (const sigset_t __user *) argp; | |
9251 | *ts = NULL; | |
9252 | return 0; | |
9253 | } | |
9254 | ||
9255 | /* | |
9256 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
9257 | * timespec and sigset_t pointers if good. | |
9258 | */ | |
9259 | if (*argsz != sizeof(arg)) | |
9260 | return -EINVAL; | |
9261 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
9262 | return -EFAULT; | |
9263 | *sig = u64_to_user_ptr(arg.sigmask); | |
9264 | *argsz = arg.sigmask_sz; | |
9265 | *ts = u64_to_user_ptr(arg.ts); | |
9266 | return 0; | |
9267 | } | |
9268 | ||
2b188cc1 | 9269 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
9270 | u32, min_complete, u32, flags, const void __user *, argp, |
9271 | size_t, argsz) | |
2b188cc1 JA |
9272 | { |
9273 | struct io_ring_ctx *ctx; | |
2b188cc1 JA |
9274 | int submitted = 0; |
9275 | struct fd f; | |
33f993da | 9276 | long ret; |
2b188cc1 | 9277 | |
4c6e277c | 9278 | io_run_task_work(); |
b41e9852 | 9279 | |
33f993da PB |
9280 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
9281 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))) | |
2b188cc1 JA |
9282 | return -EINVAL; |
9283 | ||
9284 | f = fdget(fd); | |
33f993da | 9285 | if (unlikely(!f.file)) |
2b188cc1 JA |
9286 | return -EBADF; |
9287 | ||
9288 | ret = -EOPNOTSUPP; | |
33f993da | 9289 | if (unlikely(f.file->f_op != &io_uring_fops)) |
2b188cc1 JA |
9290 | goto out_fput; |
9291 | ||
9292 | ret = -ENXIO; | |
9293 | ctx = f.file->private_data; | |
33f993da | 9294 | if (unlikely(!percpu_ref_tryget(&ctx->refs))) |
2b188cc1 JA |
9295 | goto out_fput; |
9296 | ||
7e84e1c7 | 9297 | ret = -EBADFD; |
33f993da | 9298 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
9299 | goto out; |
9300 | ||
6c271ce2 JA |
9301 | /* |
9302 | * For SQ polling, the thread will do all submissions and completions. | |
9303 | * Just return the requested submit count, and wake the thread if | |
9304 | * we were asked to. | |
9305 | */ | |
b2a9eada | 9306 | ret = 0; |
6c271ce2 | 9307 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 9308 | io_cqring_overflow_flush(ctx); |
89448c47 | 9309 | |
21f96522 JA |
9310 | if (unlikely(ctx->sq_data->thread == NULL)) { |
9311 | ret = -EOWNERDEAD; | |
04147488 | 9312 | goto out; |
21f96522 | 9313 | } |
6c271ce2 | 9314 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 9315 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
9316 | if (flags & IORING_ENTER_SQ_WAIT) { |
9317 | ret = io_sqpoll_wait_sq(ctx); | |
9318 | if (ret) | |
9319 | goto out; | |
9320 | } | |
6c271ce2 | 9321 | submitted = to_submit; |
b2a9eada | 9322 | } else if (to_submit) { |
eef51daa | 9323 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
9324 | if (unlikely(ret)) |
9325 | goto out; | |
2b188cc1 | 9326 | mutex_lock(&ctx->uring_lock); |
0f212204 | 9327 | submitted = io_submit_sqes(ctx, to_submit); |
2b188cc1 | 9328 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
9329 | |
9330 | if (submitted != to_submit) | |
9331 | goto out; | |
2b188cc1 JA |
9332 | } |
9333 | if (flags & IORING_ENTER_GETEVENTS) { | |
c73ebb68 HX |
9334 | const sigset_t __user *sig; |
9335 | struct __kernel_timespec __user *ts; | |
9336 | ||
9337 | ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); | |
9338 | if (unlikely(ret)) | |
9339 | goto out; | |
9340 | ||
2b188cc1 JA |
9341 | min_complete = min(min_complete, ctx->cq_entries); |
9342 | ||
32b2244a XW |
9343 | /* |
9344 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
9345 | * space applications don't need to do io completion events | |
9346 | * polling again, they can rely on io_sq_thread to do polling | |
9347 | * work, which can reduce cpu usage and uring_lock contention. | |
9348 | */ | |
9349 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
9350 | !(ctx->flags & IORING_SETUP_SQPOLL)) { | |
7668b92a | 9351 | ret = io_iopoll_check(ctx, min_complete); |
def596e9 | 9352 | } else { |
c73ebb68 | 9353 | ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts); |
def596e9 | 9354 | } |
2b188cc1 JA |
9355 | } |
9356 | ||
7c504e65 | 9357 | out: |
6805b32e | 9358 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
9359 | out_fput: |
9360 | fdput(f); | |
9361 | return submitted ? submitted : ret; | |
9362 | } | |
9363 | ||
bebdb65e | 9364 | #ifdef CONFIG_PROC_FS |
61cf9370 MWO |
9365 | static int io_uring_show_cred(struct seq_file *m, unsigned int id, |
9366 | const struct cred *cred) | |
87ce955b | 9367 | { |
87ce955b JA |
9368 | struct user_namespace *uns = seq_user_ns(m); |
9369 | struct group_info *gi; | |
9370 | kernel_cap_t cap; | |
9371 | unsigned __capi; | |
9372 | int g; | |
9373 | ||
9374 | seq_printf(m, "%5d\n", id); | |
9375 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
9376 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
9377 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
9378 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
9379 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
9380 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
9381 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
9382 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
9383 | seq_puts(m, "\n\tGroups:\t"); | |
9384 | gi = cred->group_info; | |
9385 | for (g = 0; g < gi->ngroups; g++) { | |
9386 | seq_put_decimal_ull(m, g ? " " : "", | |
9387 | from_kgid_munged(uns, gi->gid[g])); | |
9388 | } | |
9389 | seq_puts(m, "\n\tCapEff:\t"); | |
9390 | cap = cred->cap_effective; | |
9391 | CAP_FOR_EACH_U32(__capi) | |
9392 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
9393 | seq_putc(m, '\n'); | |
9394 | return 0; | |
9395 | } | |
9396 | ||
9397 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
9398 | { | |
dbbe9c64 | 9399 | struct io_sq_data *sq = NULL; |
fad8e0de | 9400 | bool has_lock; |
87ce955b JA |
9401 | int i; |
9402 | ||
fad8e0de JA |
9403 | /* |
9404 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, | |
9405 | * since fdinfo case grabs it in the opposite direction of normal use | |
9406 | * cases. If we fail to get the lock, we just don't iterate any | |
9407 | * structures that could be going away outside the io_uring mutex. | |
9408 | */ | |
9409 | has_lock = mutex_trylock(&ctx->uring_lock); | |
9410 | ||
5f3f26f9 | 9411 | if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) { |
dbbe9c64 | 9412 | sq = ctx->sq_data; |
5f3f26f9 JA |
9413 | if (!sq->thread) |
9414 | sq = NULL; | |
9415 | } | |
dbbe9c64 JQ |
9416 | |
9417 | seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1); | |
9418 | seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1); | |
87ce955b | 9419 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
fad8e0de | 9420 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
7b29f92d | 9421 | struct file *f = io_file_from_index(ctx, i); |
87ce955b | 9422 | |
87ce955b JA |
9423 | if (f) |
9424 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
9425 | else | |
9426 | seq_printf(m, "%5u: <none>\n", i); | |
9427 | } | |
9428 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
fad8e0de | 9429 | for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) { |
41edf1a5 | 9430 | struct io_mapped_ubuf *buf = ctx->user_bufs[i]; |
4751f53d | 9431 | unsigned int len = buf->ubuf_end - buf->ubuf; |
87ce955b | 9432 | |
4751f53d | 9433 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len); |
87ce955b | 9434 | } |
61cf9370 MWO |
9435 | if (has_lock && !xa_empty(&ctx->personalities)) { |
9436 | unsigned long index; | |
9437 | const struct cred *cred; | |
9438 | ||
87ce955b | 9439 | seq_printf(m, "Personalities:\n"); |
61cf9370 MWO |
9440 | xa_for_each(&ctx->personalities, index, cred) |
9441 | io_uring_show_cred(m, index, cred); | |
87ce955b | 9442 | } |
d7718a9d JA |
9443 | seq_printf(m, "PollList:\n"); |
9444 | spin_lock_irq(&ctx->completion_lock); | |
9445 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
9446 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
9447 | struct io_kiocb *req; | |
9448 | ||
9449 | hlist_for_each_entry(req, list, hash_node) | |
9450 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
9451 | req->task->task_works != NULL); | |
9452 | } | |
9453 | spin_unlock_irq(&ctx->completion_lock); | |
fad8e0de JA |
9454 | if (has_lock) |
9455 | mutex_unlock(&ctx->uring_lock); | |
87ce955b JA |
9456 | } |
9457 | ||
9458 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
9459 | { | |
9460 | struct io_ring_ctx *ctx = f->private_data; | |
9461 | ||
9462 | if (percpu_ref_tryget(&ctx->refs)) { | |
9463 | __io_uring_show_fdinfo(ctx, m); | |
9464 | percpu_ref_put(&ctx->refs); | |
9465 | } | |
9466 | } | |
bebdb65e | 9467 | #endif |
87ce955b | 9468 | |
2b188cc1 JA |
9469 | static const struct file_operations io_uring_fops = { |
9470 | .release = io_uring_release, | |
9471 | .mmap = io_uring_mmap, | |
6c5c240e RP |
9472 | #ifndef CONFIG_MMU |
9473 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
9474 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
9475 | #endif | |
2b188cc1 JA |
9476 | .poll = io_uring_poll, |
9477 | .fasync = io_uring_fasync, | |
bebdb65e | 9478 | #ifdef CONFIG_PROC_FS |
87ce955b | 9479 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 9480 | #endif |
2b188cc1 JA |
9481 | }; |
9482 | ||
9483 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
9484 | struct io_uring_params *p) | |
9485 | { | |
75b28aff HV |
9486 | struct io_rings *rings; |
9487 | size_t size, sq_array_offset; | |
2b188cc1 | 9488 | |
bd740481 JA |
9489 | /* make sure these are sane, as we already accounted them */ |
9490 | ctx->sq_entries = p->sq_entries; | |
9491 | ctx->cq_entries = p->cq_entries; | |
9492 | ||
75b28aff HV |
9493 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
9494 | if (size == SIZE_MAX) | |
9495 | return -EOVERFLOW; | |
9496 | ||
9497 | rings = io_mem_alloc(size); | |
9498 | if (!rings) | |
2b188cc1 JA |
9499 | return -ENOMEM; |
9500 | ||
75b28aff HV |
9501 | ctx->rings = rings; |
9502 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
9503 | rings->sq_ring_mask = p->sq_entries - 1; | |
9504 | rings->cq_ring_mask = p->cq_entries - 1; | |
9505 | rings->sq_ring_entries = p->sq_entries; | |
9506 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 JA |
9507 | |
9508 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
9509 | if (size == SIZE_MAX) { |
9510 | io_mem_free(ctx->rings); | |
9511 | ctx->rings = NULL; | |
2b188cc1 | 9512 | return -EOVERFLOW; |
eb065d30 | 9513 | } |
2b188cc1 JA |
9514 | |
9515 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
9516 | if (!ctx->sq_sqes) { |
9517 | io_mem_free(ctx->rings); | |
9518 | ctx->rings = NULL; | |
2b188cc1 | 9519 | return -ENOMEM; |
eb065d30 | 9520 | } |
2b188cc1 | 9521 | |
2b188cc1 JA |
9522 | return 0; |
9523 | } | |
9524 | ||
9faadcc8 PB |
9525 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
9526 | { | |
9527 | int ret, fd; | |
9528 | ||
9529 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
9530 | if (fd < 0) | |
9531 | return fd; | |
9532 | ||
eef51daa | 9533 | ret = io_uring_add_tctx_node(ctx); |
9faadcc8 PB |
9534 | if (ret) { |
9535 | put_unused_fd(fd); | |
9536 | return ret; | |
9537 | } | |
9538 | fd_install(fd, file); | |
9539 | return fd; | |
9540 | } | |
9541 | ||
2b188cc1 JA |
9542 | /* |
9543 | * Allocate an anonymous fd, this is what constitutes the application | |
9544 | * visible backing of an io_uring instance. The application mmaps this | |
9545 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
9546 | * we have to tie this fd to a socket for file garbage collection purposes. | |
9547 | */ | |
9faadcc8 | 9548 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
9549 | { |
9550 | struct file *file; | |
9faadcc8 | 9551 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
9552 | int ret; |
9553 | ||
2b188cc1 JA |
9554 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
9555 | &ctx->ring_sock); | |
9556 | if (ret) | |
9faadcc8 | 9557 | return ERR_PTR(ret); |
2b188cc1 JA |
9558 | #endif |
9559 | ||
2b188cc1 JA |
9560 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, |
9561 | O_RDWR | O_CLOEXEC); | |
2b188cc1 | 9562 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
9563 | if (IS_ERR(file)) { |
9564 | sock_release(ctx->ring_sock); | |
9565 | ctx->ring_sock = NULL; | |
9566 | } else { | |
9567 | ctx->ring_sock->file = file; | |
0f212204 | 9568 | } |
2b188cc1 | 9569 | #endif |
9faadcc8 | 9570 | return file; |
2b188cc1 JA |
9571 | } |
9572 | ||
7f13657d XW |
9573 | static int io_uring_create(unsigned entries, struct io_uring_params *p, |
9574 | struct io_uring_params __user *params) | |
2b188cc1 | 9575 | { |
2b188cc1 | 9576 | struct io_ring_ctx *ctx; |
9faadcc8 | 9577 | struct file *file; |
2b188cc1 JA |
9578 | int ret; |
9579 | ||
8110c1a6 | 9580 | if (!entries) |
2b188cc1 | 9581 | return -EINVAL; |
8110c1a6 JA |
9582 | if (entries > IORING_MAX_ENTRIES) { |
9583 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9584 | return -EINVAL; | |
9585 | entries = IORING_MAX_ENTRIES; | |
9586 | } | |
2b188cc1 JA |
9587 | |
9588 | /* | |
9589 | * Use twice as many entries for the CQ ring. It's possible for the | |
9590 | * application to drive a higher depth than the size of the SQ ring, | |
9591 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
9592 | * some flexibility in overcommitting a bit. If the application has |
9593 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
9594 | * of CQ ring entries manually. | |
2b188cc1 JA |
9595 | */ |
9596 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
9597 | if (p->flags & IORING_SETUP_CQSIZE) { |
9598 | /* | |
9599 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
9600 | * to a power-of-two, if it isn't already. We do NOT impose | |
9601 | * any cq vs sq ring sizing. | |
9602 | */ | |
eb2667b3 | 9603 | if (!p->cq_entries) |
33a107f0 | 9604 | return -EINVAL; |
8110c1a6 JA |
9605 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
9606 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9607 | return -EINVAL; | |
9608 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
9609 | } | |
eb2667b3 JQ |
9610 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
9611 | if (p->cq_entries < p->sq_entries) | |
9612 | return -EINVAL; | |
33a107f0 JA |
9613 | } else { |
9614 | p->cq_entries = 2 * p->sq_entries; | |
9615 | } | |
2b188cc1 | 9616 | |
2b188cc1 | 9617 | ctx = io_ring_ctx_alloc(p); |
62e398be | 9618 | if (!ctx) |
2b188cc1 | 9619 | return -ENOMEM; |
2b188cc1 | 9620 | ctx->compat = in_compat_syscall(); |
62e398be JA |
9621 | if (!capable(CAP_IPC_LOCK)) |
9622 | ctx->user = get_uid(current_user()); | |
2aede0e4 JA |
9623 | |
9624 | /* | |
9625 | * This is just grabbed for accounting purposes. When a process exits, | |
9626 | * the mm is exited and dropped before the files, hence we need to hang | |
9627 | * on to this mm purely for the purposes of being able to unaccount | |
9628 | * memory (locked/pinned vm). It's not used for anything else. | |
9629 | */ | |
6b7898eb | 9630 | mmgrab(current->mm); |
2aede0e4 | 9631 | ctx->mm_account = current->mm; |
6b7898eb | 9632 | |
2b188cc1 JA |
9633 | ret = io_allocate_scq_urings(ctx, p); |
9634 | if (ret) | |
9635 | goto err; | |
9636 | ||
7e84e1c7 | 9637 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
9638 | if (ret) |
9639 | goto err; | |
eae071c9 | 9640 | /* always set a rsrc node */ |
47b228ce PB |
9641 | ret = io_rsrc_node_switch_start(ctx); |
9642 | if (ret) | |
9643 | goto err; | |
eae071c9 | 9644 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 9645 | |
2b188cc1 | 9646 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
9647 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
9648 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
9649 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
9650 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
9651 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
9652 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
9653 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
9654 | |
9655 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
9656 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
9657 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
9658 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
9659 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
9660 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
9661 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 9662 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 9663 | |
7f13657d XW |
9664 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
9665 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 9666 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 9667 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e PB |
9668 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
9669 | IORING_FEAT_RSRC_TAGS; | |
7f13657d XW |
9670 | |
9671 | if (copy_to_user(params, p, sizeof(*p))) { | |
9672 | ret = -EFAULT; | |
9673 | goto err; | |
9674 | } | |
d1719f70 | 9675 | |
9faadcc8 PB |
9676 | file = io_uring_get_file(ctx); |
9677 | if (IS_ERR(file)) { | |
9678 | ret = PTR_ERR(file); | |
9679 | goto err; | |
9680 | } | |
9681 | ||
044c1ab3 JA |
9682 | /* |
9683 | * Install ring fd as the very last thing, so we don't risk someone | |
9684 | * having closed it before we finish setup | |
9685 | */ | |
9faadcc8 PB |
9686 | ret = io_uring_install_fd(ctx, file); |
9687 | if (ret < 0) { | |
9688 | /* fput will clean it up */ | |
9689 | fput(file); | |
9690 | return ret; | |
9691 | } | |
044c1ab3 | 9692 | |
c826bd7a | 9693 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
9694 | return ret; |
9695 | err: | |
9696 | io_ring_ctx_wait_and_kill(ctx); | |
9697 | return ret; | |
9698 | } | |
9699 | ||
9700 | /* | |
9701 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
9702 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
9703 | * params structure passed in. | |
9704 | */ | |
9705 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
9706 | { | |
9707 | struct io_uring_params p; | |
2b188cc1 JA |
9708 | int i; |
9709 | ||
9710 | if (copy_from_user(&p, params, sizeof(p))) | |
9711 | return -EFAULT; | |
9712 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
9713 | if (p.resv[i]) | |
9714 | return -EINVAL; | |
9715 | } | |
9716 | ||
6c271ce2 | 9717 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 9718 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 SG |
9719 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
9720 | IORING_SETUP_R_DISABLED)) | |
2b188cc1 JA |
9721 | return -EINVAL; |
9722 | ||
7f13657d | 9723 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
9724 | } |
9725 | ||
9726 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
9727 | struct io_uring_params __user *, params) | |
9728 | { | |
9729 | return io_uring_setup(entries, params); | |
9730 | } | |
9731 | ||
66f4af93 JA |
9732 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
9733 | { | |
9734 | struct io_uring_probe *p; | |
9735 | size_t size; | |
9736 | int i, ret; | |
9737 | ||
9738 | size = struct_size(p, ops, nr_args); | |
9739 | if (size == SIZE_MAX) | |
9740 | return -EOVERFLOW; | |
9741 | p = kzalloc(size, GFP_KERNEL); | |
9742 | if (!p) | |
9743 | return -ENOMEM; | |
9744 | ||
9745 | ret = -EFAULT; | |
9746 | if (copy_from_user(p, arg, size)) | |
9747 | goto out; | |
9748 | ret = -EINVAL; | |
9749 | if (memchr_inv(p, 0, size)) | |
9750 | goto out; | |
9751 | ||
9752 | p->last_op = IORING_OP_LAST - 1; | |
9753 | if (nr_args > IORING_OP_LAST) | |
9754 | nr_args = IORING_OP_LAST; | |
9755 | ||
9756 | for (i = 0; i < nr_args; i++) { | |
9757 | p->ops[i].op = i; | |
9758 | if (!io_op_defs[i].not_supported) | |
9759 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
9760 | } | |
9761 | p->ops_len = i; | |
9762 | ||
9763 | ret = 0; | |
9764 | if (copy_to_user(arg, p, size)) | |
9765 | ret = -EFAULT; | |
9766 | out: | |
9767 | kfree(p); | |
9768 | return ret; | |
9769 | } | |
9770 | ||
071698e1 JA |
9771 | static int io_register_personality(struct io_ring_ctx *ctx) |
9772 | { | |
4379bf8b | 9773 | const struct cred *creds; |
61cf9370 | 9774 | u32 id; |
1e6fa521 | 9775 | int ret; |
071698e1 | 9776 | |
4379bf8b | 9777 | creds = get_current_cred(); |
1e6fa521 | 9778 | |
61cf9370 MWO |
9779 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
9780 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
9781 | if (ret < 0) { |
9782 | put_cred(creds); | |
9783 | return ret; | |
9784 | } | |
9785 | return id; | |
071698e1 JA |
9786 | } |
9787 | ||
21b55dbc SG |
9788 | static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg, |
9789 | unsigned int nr_args) | |
9790 | { | |
9791 | struct io_uring_restriction *res; | |
9792 | size_t size; | |
9793 | int i, ret; | |
9794 | ||
7e84e1c7 SG |
9795 | /* Restrictions allowed only if rings started disabled */ |
9796 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9797 | return -EBADFD; | |
9798 | ||
21b55dbc | 9799 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 9800 | if (ctx->restrictions.registered) |
21b55dbc SG |
9801 | return -EBUSY; |
9802 | ||
9803 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
9804 | return -EINVAL; | |
9805 | ||
9806 | size = array_size(nr_args, sizeof(*res)); | |
9807 | if (size == SIZE_MAX) | |
9808 | return -EOVERFLOW; | |
9809 | ||
9810 | res = memdup_user(arg, size); | |
9811 | if (IS_ERR(res)) | |
9812 | return PTR_ERR(res); | |
9813 | ||
9814 | ret = 0; | |
9815 | ||
9816 | for (i = 0; i < nr_args; i++) { | |
9817 | switch (res[i].opcode) { | |
9818 | case IORING_RESTRICTION_REGISTER_OP: | |
9819 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
9820 | ret = -EINVAL; | |
9821 | goto out; | |
9822 | } | |
9823 | ||
9824 | __set_bit(res[i].register_op, | |
9825 | ctx->restrictions.register_op); | |
9826 | break; | |
9827 | case IORING_RESTRICTION_SQE_OP: | |
9828 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
9829 | ret = -EINVAL; | |
9830 | goto out; | |
9831 | } | |
9832 | ||
9833 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
9834 | break; | |
9835 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
9836 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
9837 | break; | |
9838 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
9839 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
9840 | break; | |
9841 | default: | |
9842 | ret = -EINVAL; | |
9843 | goto out; | |
9844 | } | |
9845 | } | |
9846 | ||
9847 | out: | |
9848 | /* Reset all restrictions if an error happened */ | |
9849 | if (ret != 0) | |
9850 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
9851 | else | |
7e84e1c7 | 9852 | ctx->restrictions.registered = true; |
21b55dbc SG |
9853 | |
9854 | kfree(res); | |
9855 | return ret; | |
9856 | } | |
9857 | ||
7e84e1c7 SG |
9858 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
9859 | { | |
9860 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9861 | return -EBADFD; | |
9862 | ||
9863 | if (ctx->restrictions.registered) | |
9864 | ctx->restricted = 1; | |
9865 | ||
0298ef96 PB |
9866 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
9867 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
9868 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
9869 | return 0; |
9870 | } | |
9871 | ||
fdecb662 | 9872 | static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, |
c3bdad02 | 9873 | struct io_uring_rsrc_update2 *up, |
98f0b3b4 PB |
9874 | unsigned nr_args) |
9875 | { | |
9876 | __u32 tmp; | |
9877 | int err; | |
9878 | ||
c3bdad02 PB |
9879 | if (up->resv) |
9880 | return -EINVAL; | |
98f0b3b4 PB |
9881 | if (check_add_overflow(up->offset, nr_args, &tmp)) |
9882 | return -EOVERFLOW; | |
9883 | err = io_rsrc_node_switch_start(ctx); | |
9884 | if (err) | |
9885 | return err; | |
9886 | ||
fdecb662 PB |
9887 | switch (type) { |
9888 | case IORING_RSRC_FILE: | |
98f0b3b4 | 9889 | return __io_sqe_files_update(ctx, up, nr_args); |
634d00df PB |
9890 | case IORING_RSRC_BUFFER: |
9891 | return __io_sqe_buffers_update(ctx, up, nr_args); | |
98f0b3b4 PB |
9892 | } |
9893 | return -EINVAL; | |
9894 | } | |
9895 | ||
c3bdad02 PB |
9896 | static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, |
9897 | unsigned nr_args) | |
98f0b3b4 | 9898 | { |
c3bdad02 | 9899 | struct io_uring_rsrc_update2 up; |
98f0b3b4 PB |
9900 | |
9901 | if (!nr_args) | |
9902 | return -EINVAL; | |
c3bdad02 PB |
9903 | memset(&up, 0, sizeof(up)); |
9904 | if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) | |
9905 | return -EFAULT; | |
9906 | return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); | |
9907 | } | |
9908 | ||
9909 | static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, | |
992da01a | 9910 | unsigned size, unsigned type) |
c3bdad02 PB |
9911 | { |
9912 | struct io_uring_rsrc_update2 up; | |
9913 | ||
9914 | if (size != sizeof(up)) | |
9915 | return -EINVAL; | |
98f0b3b4 PB |
9916 | if (copy_from_user(&up, arg, sizeof(up))) |
9917 | return -EFAULT; | |
992da01a | 9918 | if (!up.nr || up.resv) |
98f0b3b4 | 9919 | return -EINVAL; |
992da01a | 9920 | return __io_register_rsrc_update(ctx, type, &up, up.nr); |
98f0b3b4 PB |
9921 | } |
9922 | ||
792e3582 | 9923 | static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, |
992da01a | 9924 | unsigned int size, unsigned int type) |
792e3582 PB |
9925 | { |
9926 | struct io_uring_rsrc_register rr; | |
9927 | ||
9928 | /* keep it extendible */ | |
9929 | if (size != sizeof(rr)) | |
9930 | return -EINVAL; | |
9931 | ||
9932 | memset(&rr, 0, sizeof(rr)); | |
9933 | if (copy_from_user(&rr, arg, size)) | |
9934 | return -EFAULT; | |
992da01a | 9935 | if (!rr.nr || rr.resv || rr.resv2) |
792e3582 PB |
9936 | return -EINVAL; |
9937 | ||
992da01a | 9938 | switch (type) { |
792e3582 PB |
9939 | case IORING_RSRC_FILE: |
9940 | return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), | |
9941 | rr.nr, u64_to_user_ptr(rr.tags)); | |
634d00df PB |
9942 | case IORING_RSRC_BUFFER: |
9943 | return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), | |
9944 | rr.nr, u64_to_user_ptr(rr.tags)); | |
792e3582 PB |
9945 | } |
9946 | return -EINVAL; | |
9947 | } | |
9948 | ||
fe76421d JA |
9949 | static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg, |
9950 | unsigned len) | |
9951 | { | |
9952 | struct io_uring_task *tctx = current->io_uring; | |
9953 | cpumask_var_t new_mask; | |
9954 | int ret; | |
9955 | ||
9956 | if (!tctx || !tctx->io_wq) | |
9957 | return -EINVAL; | |
9958 | ||
9959 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
9960 | return -ENOMEM; | |
9961 | ||
9962 | cpumask_clear(new_mask); | |
9963 | if (len > cpumask_size()) | |
9964 | len = cpumask_size(); | |
9965 | ||
9966 | if (copy_from_user(new_mask, arg, len)) { | |
9967 | free_cpumask_var(new_mask); | |
9968 | return -EFAULT; | |
9969 | } | |
9970 | ||
9971 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
9972 | free_cpumask_var(new_mask); | |
9973 | return ret; | |
9974 | } | |
9975 | ||
9976 | static int io_unregister_iowq_aff(struct io_ring_ctx *ctx) | |
9977 | { | |
9978 | struct io_uring_task *tctx = current->io_uring; | |
9979 | ||
9980 | if (!tctx || !tctx->io_wq) | |
9981 | return -EINVAL; | |
9982 | ||
9983 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
9984 | } | |
9985 | ||
071698e1 JA |
9986 | static bool io_register_op_must_quiesce(int op) |
9987 | { | |
9988 | switch (op) { | |
bd54b6fe BM |
9989 | case IORING_REGISTER_BUFFERS: |
9990 | case IORING_UNREGISTER_BUFFERS: | |
f4f7d21c | 9991 | case IORING_REGISTER_FILES: |
071698e1 JA |
9992 | case IORING_UNREGISTER_FILES: |
9993 | case IORING_REGISTER_FILES_UPDATE: | |
9994 | case IORING_REGISTER_PROBE: | |
9995 | case IORING_REGISTER_PERSONALITY: | |
9996 | case IORING_UNREGISTER_PERSONALITY: | |
992da01a PB |
9997 | case IORING_REGISTER_FILES2: |
9998 | case IORING_REGISTER_FILES_UPDATE2: | |
9999 | case IORING_REGISTER_BUFFERS2: | |
10000 | case IORING_REGISTER_BUFFERS_UPDATE: | |
fe76421d JA |
10001 | case IORING_REGISTER_IOWQ_AFF: |
10002 | case IORING_UNREGISTER_IOWQ_AFF: | |
071698e1 JA |
10003 | return false; |
10004 | default: | |
10005 | return true; | |
10006 | } | |
10007 | } | |
10008 | ||
e73c5c7c PB |
10009 | static int io_ctx_quiesce(struct io_ring_ctx *ctx) |
10010 | { | |
10011 | long ret; | |
10012 | ||
10013 | percpu_ref_kill(&ctx->refs); | |
10014 | ||
10015 | /* | |
10016 | * Drop uring mutex before waiting for references to exit. If another | |
10017 | * thread is currently inside io_uring_enter() it might need to grab the | |
10018 | * uring_lock to make progress. If we hold it here across the drain | |
10019 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
10020 | * no new references will come in after we've killed the percpu ref. | |
10021 | */ | |
10022 | mutex_unlock(&ctx->uring_lock); | |
10023 | do { | |
10024 | ret = wait_for_completion_interruptible(&ctx->ref_comp); | |
10025 | if (!ret) | |
10026 | break; | |
10027 | ret = io_run_task_work_sig(); | |
10028 | } while (ret >= 0); | |
10029 | mutex_lock(&ctx->uring_lock); | |
10030 | ||
10031 | if (ret) | |
10032 | io_refs_resurrect(&ctx->refs, &ctx->ref_comp); | |
10033 | return ret; | |
10034 | } | |
10035 | ||
edafccee JA |
10036 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
10037 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
10038 | __releases(ctx->uring_lock) |
10039 | __acquires(ctx->uring_lock) | |
edafccee JA |
10040 | { |
10041 | int ret; | |
10042 | ||
35fa71a0 JA |
10043 | /* |
10044 | * We're inside the ring mutex, if the ref is already dying, then | |
10045 | * someone else killed the ctx or is already going through | |
10046 | * io_uring_register(). | |
10047 | */ | |
10048 | if (percpu_ref_is_dying(&ctx->refs)) | |
10049 | return -ENXIO; | |
10050 | ||
75c4021a PB |
10051 | if (ctx->restricted) { |
10052 | if (opcode >= IORING_REGISTER_LAST) | |
10053 | return -EINVAL; | |
10054 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); | |
10055 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
10056 | return -EACCES; | |
10057 | } | |
10058 | ||
071698e1 | 10059 | if (io_register_op_must_quiesce(opcode)) { |
e73c5c7c PB |
10060 | ret = io_ctx_quiesce(ctx); |
10061 | if (ret) | |
f70865db | 10062 | return ret; |
05f3fb3c | 10063 | } |
edafccee JA |
10064 | |
10065 | switch (opcode) { | |
10066 | case IORING_REGISTER_BUFFERS: | |
634d00df | 10067 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
10068 | break; |
10069 | case IORING_UNREGISTER_BUFFERS: | |
10070 | ret = -EINVAL; | |
10071 | if (arg || nr_args) | |
10072 | break; | |
0a96bbe4 | 10073 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 10074 | break; |
6b06314c | 10075 | case IORING_REGISTER_FILES: |
792e3582 | 10076 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
10077 | break; |
10078 | case IORING_UNREGISTER_FILES: | |
10079 | ret = -EINVAL; | |
10080 | if (arg || nr_args) | |
10081 | break; | |
10082 | ret = io_sqe_files_unregister(ctx); | |
10083 | break; | |
c3a31e60 | 10084 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 10085 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 10086 | break; |
9b402849 | 10087 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 10088 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
10089 | ret = -EINVAL; |
10090 | if (nr_args != 1) | |
10091 | break; | |
10092 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
10093 | if (ret) |
10094 | break; | |
10095 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
10096 | ctx->eventfd_async = 1; | |
10097 | else | |
10098 | ctx->eventfd_async = 0; | |
9b402849 JA |
10099 | break; |
10100 | case IORING_UNREGISTER_EVENTFD: | |
10101 | ret = -EINVAL; | |
10102 | if (arg || nr_args) | |
10103 | break; | |
10104 | ret = io_eventfd_unregister(ctx); | |
10105 | break; | |
66f4af93 JA |
10106 | case IORING_REGISTER_PROBE: |
10107 | ret = -EINVAL; | |
10108 | if (!arg || nr_args > 256) | |
10109 | break; | |
10110 | ret = io_probe(ctx, arg, nr_args); | |
10111 | break; | |
071698e1 JA |
10112 | case IORING_REGISTER_PERSONALITY: |
10113 | ret = -EINVAL; | |
10114 | if (arg || nr_args) | |
10115 | break; | |
10116 | ret = io_register_personality(ctx); | |
10117 | break; | |
10118 | case IORING_UNREGISTER_PERSONALITY: | |
10119 | ret = -EINVAL; | |
10120 | if (arg) | |
10121 | break; | |
10122 | ret = io_unregister_personality(ctx, nr_args); | |
10123 | break; | |
7e84e1c7 SG |
10124 | case IORING_REGISTER_ENABLE_RINGS: |
10125 | ret = -EINVAL; | |
10126 | if (arg || nr_args) | |
10127 | break; | |
10128 | ret = io_register_enable_rings(ctx); | |
10129 | break; | |
21b55dbc SG |
10130 | case IORING_REGISTER_RESTRICTIONS: |
10131 | ret = io_register_restrictions(ctx, arg, nr_args); | |
10132 | break; | |
992da01a PB |
10133 | case IORING_REGISTER_FILES2: |
10134 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
10135 | break; | |
10136 | case IORING_REGISTER_FILES_UPDATE2: | |
10137 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
10138 | IORING_RSRC_FILE); | |
10139 | break; | |
10140 | case IORING_REGISTER_BUFFERS2: | |
10141 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 10142 | break; |
992da01a PB |
10143 | case IORING_REGISTER_BUFFERS_UPDATE: |
10144 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
10145 | IORING_RSRC_BUFFER); | |
c3bdad02 | 10146 | break; |
fe76421d JA |
10147 | case IORING_REGISTER_IOWQ_AFF: |
10148 | ret = -EINVAL; | |
10149 | if (!arg || !nr_args) | |
10150 | break; | |
10151 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
10152 | break; | |
10153 | case IORING_UNREGISTER_IOWQ_AFF: | |
10154 | ret = -EINVAL; | |
10155 | if (arg || nr_args) | |
10156 | break; | |
10157 | ret = io_unregister_iowq_aff(ctx); | |
10158 | break; | |
edafccee JA |
10159 | default: |
10160 | ret = -EINVAL; | |
10161 | break; | |
10162 | } | |
10163 | ||
071698e1 | 10164 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 10165 | /* bring the ctx back to life */ |
05f3fb3c | 10166 | percpu_ref_reinit(&ctx->refs); |
0f158b4c | 10167 | reinit_completion(&ctx->ref_comp); |
05f3fb3c | 10168 | } |
edafccee JA |
10169 | return ret; |
10170 | } | |
10171 | ||
10172 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
10173 | void __user *, arg, unsigned int, nr_args) | |
10174 | { | |
10175 | struct io_ring_ctx *ctx; | |
10176 | long ret = -EBADF; | |
10177 | struct fd f; | |
10178 | ||
10179 | f = fdget(fd); | |
10180 | if (!f.file) | |
10181 | return -EBADF; | |
10182 | ||
10183 | ret = -EOPNOTSUPP; | |
10184 | if (f.file->f_op != &io_uring_fops) | |
10185 | goto out_fput; | |
10186 | ||
10187 | ctx = f.file->private_data; | |
10188 | ||
b6c23dd5 PB |
10189 | io_run_task_work(); |
10190 | ||
edafccee JA |
10191 | mutex_lock(&ctx->uring_lock); |
10192 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
10193 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
10194 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
10195 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
10196 | out_fput: |
10197 | fdput(f); | |
10198 | return ret; | |
10199 | } | |
10200 | ||
2b188cc1 JA |
10201 | static int __init io_uring_init(void) |
10202 | { | |
d7f62e82 SM |
10203 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
10204 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
10205 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
10206 | } while (0) | |
10207 | ||
10208 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
10209 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
10210 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
10211 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
10212 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
10213 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
10214 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
10215 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
10216 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
10217 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 10218 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
10219 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
10220 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
10221 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
10222 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
10223 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
10224 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
10225 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
10226 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
10227 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
10228 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
10229 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
10230 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
10231 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
10232 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
10233 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 10234 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
10235 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
10236 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 10237 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 10238 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 10239 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 10240 | |
b0d658ec PB |
10241 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
10242 | sizeof(struct io_uring_rsrc_update)); | |
10243 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
10244 | sizeof(struct io_uring_rsrc_update2)); | |
10245 | /* should fit into one byte */ | |
10246 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
10247 | ||
d3656344 | 10248 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
84557871 | 10249 | BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int)); |
16340eab | 10250 | |
91f245d5 JA |
10251 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
10252 | SLAB_ACCOUNT); | |
2b188cc1 JA |
10253 | return 0; |
10254 | }; | |
10255 | __initcall(io_uring_init); |