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 | |
89850fce JA |
412 | spinlock_t timeout_lock; |
413 | ||
def596e9 | 414 | /* |
540e32a0 | 415 | * ->iopoll_list is protected by the ctx->uring_lock for |
def596e9 JA |
416 | * io_uring instances that don't use IORING_SETUP_SQPOLL. |
417 | * For SQPOLL, only the single threaded io_sq_thread() will | |
418 | * manipulate the list, hence no extra locking is needed there. | |
419 | */ | |
540e32a0 | 420 | struct list_head iopoll_list; |
78076bb6 JA |
421 | struct hlist_head *cancel_hash; |
422 | unsigned cancel_hash_bits; | |
915b3dde | 423 | bool poll_multi_queue; |
2b188cc1 | 424 | } ____cacheline_aligned_in_smp; |
85faa7b8 | 425 | |
21b55dbc | 426 | struct io_restriction restrictions; |
3c1a2ead | 427 | |
b13a8918 PB |
428 | /* slow path rsrc auxilary data, used by update/register */ |
429 | struct { | |
430 | struct io_rsrc_node *rsrc_backup_node; | |
431 | struct io_mapped_ubuf *dummy_ubuf; | |
432 | struct io_rsrc_data *file_data; | |
433 | struct io_rsrc_data *buf_data; | |
434 | ||
435 | struct delayed_work rsrc_put_work; | |
436 | struct llist_head rsrc_put_llist; | |
437 | struct list_head rsrc_ref_list; | |
438 | spinlock_t rsrc_ref_lock; | |
439 | }; | |
440 | ||
3c1a2ead | 441 | /* Keep this last, we don't need it for the fast path */ |
b986af7e PB |
442 | struct { |
443 | #if defined(CONFIG_UNIX) | |
444 | struct socket *ring_sock; | |
445 | #endif | |
446 | /* hashed buffered write serialization */ | |
447 | struct io_wq_hash *hash_map; | |
448 | ||
449 | /* Only used for accounting purposes */ | |
450 | struct user_struct *user; | |
451 | struct mm_struct *mm_account; | |
452 | ||
453 | /* ctx exit and cancelation */ | |
9011bf9a PB |
454 | struct llist_head fallback_llist; |
455 | struct delayed_work fallback_work; | |
b986af7e PB |
456 | struct work_struct exit_work; |
457 | struct list_head tctx_list; | |
458 | struct completion ref_comp; | |
459 | }; | |
2b188cc1 JA |
460 | }; |
461 | ||
53e043b2 SM |
462 | struct io_uring_task { |
463 | /* submission side */ | |
09899b19 | 464 | int cached_refs; |
53e043b2 SM |
465 | struct xarray xa; |
466 | struct wait_queue_head wait; | |
ee53fb2b SM |
467 | const struct io_ring_ctx *last; |
468 | struct io_wq *io_wq; | |
53e043b2 | 469 | struct percpu_counter inflight; |
b303fe2e | 470 | atomic_t inflight_tracked; |
53e043b2 | 471 | atomic_t in_idle; |
53e043b2 SM |
472 | |
473 | spinlock_t task_lock; | |
474 | struct io_wq_work_list task_list; | |
53e043b2 | 475 | struct callback_head task_work; |
6294f368 | 476 | bool task_running; |
53e043b2 SM |
477 | }; |
478 | ||
09bb8394 JA |
479 | /* |
480 | * First field must be the file pointer in all the | |
481 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
482 | */ | |
221c5eb2 JA |
483 | struct io_poll_iocb { |
484 | struct file *file; | |
018043be | 485 | struct wait_queue_head *head; |
221c5eb2 | 486 | __poll_t events; |
8c838788 | 487 | bool done; |
221c5eb2 | 488 | bool canceled; |
392edb45 | 489 | struct wait_queue_entry wait; |
221c5eb2 JA |
490 | }; |
491 | ||
9d805892 | 492 | struct io_poll_update { |
018043be | 493 | struct file *file; |
9d805892 PB |
494 | u64 old_user_data; |
495 | u64 new_user_data; | |
496 | __poll_t events; | |
b69de288 JA |
497 | bool update_events; |
498 | bool update_user_data; | |
018043be PB |
499 | }; |
500 | ||
b5dba59e JA |
501 | struct io_close { |
502 | struct file *file; | |
b5dba59e JA |
503 | int fd; |
504 | }; | |
505 | ||
ad8a48ac JA |
506 | struct io_timeout_data { |
507 | struct io_kiocb *req; | |
508 | struct hrtimer timer; | |
509 | struct timespec64 ts; | |
510 | enum hrtimer_mode mode; | |
511 | }; | |
512 | ||
8ed8d3c3 JA |
513 | struct io_accept { |
514 | struct file *file; | |
515 | struct sockaddr __user *addr; | |
516 | int __user *addr_len; | |
517 | int flags; | |
09952e3e | 518 | unsigned long nofile; |
8ed8d3c3 JA |
519 | }; |
520 | ||
521 | struct io_sync { | |
522 | struct file *file; | |
523 | loff_t len; | |
524 | loff_t off; | |
525 | int flags; | |
d63d1b5e | 526 | int mode; |
8ed8d3c3 JA |
527 | }; |
528 | ||
fbf23849 JA |
529 | struct io_cancel { |
530 | struct file *file; | |
531 | u64 addr; | |
532 | }; | |
533 | ||
b29472ee JA |
534 | struct io_timeout { |
535 | struct file *file; | |
bfe68a22 PB |
536 | u32 off; |
537 | u32 target_seq; | |
135fcde8 | 538 | struct list_head list; |
90cd7e42 PB |
539 | /* head of the link, used by linked timeouts only */ |
540 | struct io_kiocb *head; | |
89b263f6 JA |
541 | /* for linked completions */ |
542 | struct io_kiocb *prev; | |
b29472ee JA |
543 | }; |
544 | ||
0bdf7a2d PB |
545 | struct io_timeout_rem { |
546 | struct file *file; | |
547 | u64 addr; | |
9c8e11b3 PB |
548 | |
549 | /* timeout update */ | |
550 | struct timespec64 ts; | |
551 | u32 flags; | |
0bdf7a2d PB |
552 | }; |
553 | ||
9adbd45d JA |
554 | struct io_rw { |
555 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
556 | struct kiocb kiocb; | |
557 | u64 addr; | |
558 | u64 len; | |
559 | }; | |
560 | ||
3fbb51c1 JA |
561 | struct io_connect { |
562 | struct file *file; | |
563 | struct sockaddr __user *addr; | |
564 | int addr_len; | |
565 | }; | |
566 | ||
e47293fd JA |
567 | struct io_sr_msg { |
568 | struct file *file; | |
fddaface | 569 | union { |
4af3417a PB |
570 | struct compat_msghdr __user *umsg_compat; |
571 | struct user_msghdr __user *umsg; | |
572 | void __user *buf; | |
fddaface | 573 | }; |
e47293fd | 574 | int msg_flags; |
bcda7baa | 575 | int bgid; |
fddaface | 576 | size_t len; |
bcda7baa | 577 | struct io_buffer *kbuf; |
e47293fd JA |
578 | }; |
579 | ||
15b71abe JA |
580 | struct io_open { |
581 | struct file *file; | |
582 | int dfd; | |
15b71abe | 583 | struct filename *filename; |
c12cedf2 | 584 | struct open_how how; |
4022e7af | 585 | unsigned long nofile; |
15b71abe JA |
586 | }; |
587 | ||
269bbe5f | 588 | struct io_rsrc_update { |
05f3fb3c JA |
589 | struct file *file; |
590 | u64 arg; | |
591 | u32 nr_args; | |
592 | u32 offset; | |
593 | }; | |
594 | ||
4840e418 JA |
595 | struct io_fadvise { |
596 | struct file *file; | |
597 | u64 offset; | |
598 | u32 len; | |
599 | u32 advice; | |
600 | }; | |
601 | ||
c1ca757b JA |
602 | struct io_madvise { |
603 | struct file *file; | |
604 | u64 addr; | |
605 | u32 len; | |
606 | u32 advice; | |
607 | }; | |
608 | ||
3e4827b0 JA |
609 | struct io_epoll { |
610 | struct file *file; | |
611 | int epfd; | |
612 | int op; | |
613 | int fd; | |
614 | struct epoll_event event; | |
e47293fd JA |
615 | }; |
616 | ||
7d67af2c PB |
617 | struct io_splice { |
618 | struct file *file_out; | |
619 | struct file *file_in; | |
620 | loff_t off_out; | |
621 | loff_t off_in; | |
622 | u64 len; | |
623 | unsigned int flags; | |
624 | }; | |
625 | ||
ddf0322d JA |
626 | struct io_provide_buf { |
627 | struct file *file; | |
628 | __u64 addr; | |
38134ada | 629 | __u32 len; |
ddf0322d JA |
630 | __u32 bgid; |
631 | __u16 nbufs; | |
632 | __u16 bid; | |
633 | }; | |
634 | ||
1d9e1288 BM |
635 | struct io_statx { |
636 | struct file *file; | |
637 | int dfd; | |
638 | unsigned int mask; | |
639 | unsigned int flags; | |
e62753e4 | 640 | const char __user *filename; |
1d9e1288 BM |
641 | struct statx __user *buffer; |
642 | }; | |
643 | ||
36f4fa68 JA |
644 | struct io_shutdown { |
645 | struct file *file; | |
646 | int how; | |
647 | }; | |
648 | ||
80a261fd JA |
649 | struct io_rename { |
650 | struct file *file; | |
651 | int old_dfd; | |
652 | int new_dfd; | |
653 | struct filename *oldpath; | |
654 | struct filename *newpath; | |
655 | int flags; | |
656 | }; | |
657 | ||
14a1143b JA |
658 | struct io_unlink { |
659 | struct file *file; | |
660 | int dfd; | |
661 | int flags; | |
662 | struct filename *filename; | |
663 | }; | |
664 | ||
3ca405eb PB |
665 | struct io_completion { |
666 | struct file *file; | |
8c3f9cd1 | 667 | u32 cflags; |
3ca405eb PB |
668 | }; |
669 | ||
f499a021 JA |
670 | struct io_async_connect { |
671 | struct sockaddr_storage address; | |
672 | }; | |
673 | ||
03b1230c JA |
674 | struct io_async_msghdr { |
675 | struct iovec fast_iov[UIO_FASTIOV]; | |
257e84a5 PB |
676 | /* points to an allocated iov, if NULL we use fast_iov instead */ |
677 | struct iovec *free_iov; | |
03b1230c JA |
678 | struct sockaddr __user *uaddr; |
679 | struct msghdr msg; | |
b537916c | 680 | struct sockaddr_storage addr; |
03b1230c JA |
681 | }; |
682 | ||
f67676d1 JA |
683 | struct io_async_rw { |
684 | struct iovec fast_iov[UIO_FASTIOV]; | |
ff6165b2 JA |
685 | const struct iovec *free_iovec; |
686 | struct iov_iter iter; | |
227c0c96 | 687 | size_t bytes_done; |
bcf5a063 | 688 | struct wait_page_queue wpq; |
f67676d1 JA |
689 | }; |
690 | ||
6b47ee6e PB |
691 | enum { |
692 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
693 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
694 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
695 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
696 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
bcda7baa | 697 | REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT, |
6b47ee6e | 698 | |
dddca226 | 699 | /* first byte is taken by user flags, shift it to not overlap */ |
93d2bcd2 | 700 | REQ_F_FAIL_BIT = 8, |
6b47ee6e PB |
701 | REQ_F_INFLIGHT_BIT, |
702 | REQ_F_CUR_POS_BIT, | |
703 | REQ_F_NOWAIT_BIT, | |
6b47ee6e | 704 | REQ_F_LINK_TIMEOUT_BIT, |
99bc4c38 | 705 | REQ_F_NEED_CLEANUP_BIT, |
d7718a9d | 706 | REQ_F_POLLED_BIT, |
bcda7baa | 707 | REQ_F_BUFFER_SELECTED_BIT, |
e342c807 | 708 | REQ_F_COMPLETE_INLINE_BIT, |
230d50d4 | 709 | REQ_F_REISSUE_BIT, |
8c130827 | 710 | REQ_F_DONT_REISSUE_BIT, |
b8e64b53 | 711 | REQ_F_CREDS_BIT, |
20e60a38 | 712 | REQ_F_REFCOUNT_BIT, |
4d13d1a4 | 713 | REQ_F_ARM_LTIMEOUT_BIT, |
7b29f92d | 714 | /* keep async read/write and isreg together and in order */ |
b191e2df PB |
715 | REQ_F_NOWAIT_READ_BIT, |
716 | REQ_F_NOWAIT_WRITE_BIT, | |
7b29f92d | 717 | REQ_F_ISREG_BIT, |
84557871 JA |
718 | |
719 | /* not a real bit, just to check we're not overflowing the space */ | |
720 | __REQ_F_LAST_BIT, | |
6b47ee6e PB |
721 | }; |
722 | ||
723 | enum { | |
724 | /* ctx owns file */ | |
725 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
726 | /* drain existing IO first */ | |
727 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
728 | /* linked sqes */ | |
729 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
730 | /* doesn't sever on completion < 0 */ | |
731 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
732 | /* IOSQE_ASYNC */ | |
733 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
bcda7baa JA |
734 | /* IOSQE_BUFFER_SELECT */ |
735 | REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT), | |
6b47ee6e | 736 | |
6b47ee6e | 737 | /* fail rest of links */ |
93d2bcd2 | 738 | REQ_F_FAIL = BIT(REQ_F_FAIL_BIT), |
b05a1bcd | 739 | /* on inflight list, should be cancelled and waited on exit reliably */ |
6b47ee6e PB |
740 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), |
741 | /* read/write uses file position */ | |
742 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
743 | /* must not punt to workers */ | |
744 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
900fad45 | 745 | /* has or had linked timeout */ |
6b47ee6e | 746 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), |
99bc4c38 PB |
747 | /* needs cleanup */ |
748 | REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), | |
d7718a9d JA |
749 | /* already went through poll handler */ |
750 | REQ_F_POLLED = BIT(REQ_F_POLLED_BIT), | |
bcda7baa JA |
751 | /* buffer already selected */ |
752 | REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT), | |
e342c807 PB |
753 | /* completion is deferred through io_comp_state */ |
754 | REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT), | |
230d50d4 JA |
755 | /* caller should reissue async */ |
756 | REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT), | |
8c130827 PB |
757 | /* don't attempt request reissue, see io_rw_reissue() */ |
758 | REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT), | |
7b29f92d | 759 | /* supports async reads */ |
b191e2df | 760 | REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT), |
7b29f92d | 761 | /* supports async writes */ |
b191e2df | 762 | REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT), |
7b29f92d JA |
763 | /* regular file */ |
764 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
b8e64b53 PB |
765 | /* has creds assigned */ |
766 | REQ_F_CREDS = BIT(REQ_F_CREDS_BIT), | |
20e60a38 PB |
767 | /* skip refcounting if not set */ |
768 | REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT), | |
4d13d1a4 PB |
769 | /* there is a linked timeout that has to be armed */ |
770 | REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT), | |
d7718a9d JA |
771 | }; |
772 | ||
773 | struct async_poll { | |
774 | struct io_poll_iocb poll; | |
807abcb0 | 775 | struct io_poll_iocb *double_poll; |
6b47ee6e PB |
776 | }; |
777 | ||
5b0a6acc PB |
778 | typedef void (*io_req_tw_func_t)(struct io_kiocb *req); |
779 | ||
7cbf1722 | 780 | struct io_task_work { |
5b0a6acc PB |
781 | union { |
782 | struct io_wq_work_node node; | |
783 | struct llist_node fallback_node; | |
784 | }; | |
785 | io_req_tw_func_t func; | |
7cbf1722 JA |
786 | }; |
787 | ||
992da01a PB |
788 | enum { |
789 | IORING_RSRC_FILE = 0, | |
790 | IORING_RSRC_BUFFER = 1, | |
791 | }; | |
792 | ||
09bb8394 JA |
793 | /* |
794 | * NOTE! Each of the iocb union members has the file pointer | |
795 | * as the first entry in their struct definition. So you can | |
796 | * access the file pointer through any of the sub-structs, | |
797 | * or directly as just 'ki_filp' in this struct. | |
798 | */ | |
2b188cc1 | 799 | struct io_kiocb { |
221c5eb2 | 800 | union { |
09bb8394 | 801 | struct file *file; |
9adbd45d | 802 | struct io_rw rw; |
221c5eb2 | 803 | struct io_poll_iocb poll; |
9d805892 | 804 | struct io_poll_update poll_update; |
8ed8d3c3 JA |
805 | struct io_accept accept; |
806 | struct io_sync sync; | |
fbf23849 | 807 | struct io_cancel cancel; |
b29472ee | 808 | struct io_timeout timeout; |
0bdf7a2d | 809 | struct io_timeout_rem timeout_rem; |
3fbb51c1 | 810 | struct io_connect connect; |
e47293fd | 811 | struct io_sr_msg sr_msg; |
15b71abe | 812 | struct io_open open; |
b5dba59e | 813 | struct io_close close; |
269bbe5f | 814 | struct io_rsrc_update rsrc_update; |
4840e418 | 815 | struct io_fadvise fadvise; |
c1ca757b | 816 | struct io_madvise madvise; |
3e4827b0 | 817 | struct io_epoll epoll; |
7d67af2c | 818 | struct io_splice splice; |
ddf0322d | 819 | struct io_provide_buf pbuf; |
1d9e1288 | 820 | struct io_statx statx; |
36f4fa68 | 821 | struct io_shutdown shutdown; |
80a261fd | 822 | struct io_rename rename; |
14a1143b | 823 | struct io_unlink unlink; |
3ca405eb PB |
824 | /* use only after cleaning per-op data, see io_clean_op() */ |
825 | struct io_completion compl; | |
221c5eb2 | 826 | }; |
2b188cc1 | 827 | |
e8c2bc1f JA |
828 | /* opcode allocated if it needs to store data for async defer */ |
829 | void *async_data; | |
d625c6ee | 830 | u8 opcode; |
65a6543d XW |
831 | /* polled IO has completed */ |
832 | u8 iopoll_completed; | |
2b188cc1 | 833 | |
4f4eeba8 | 834 | u16 buf_index; |
9cf7c104 | 835 | u32 result; |
4f4eeba8 | 836 | |
010e8e6b PB |
837 | struct io_ring_ctx *ctx; |
838 | unsigned int flags; | |
abc54d63 | 839 | atomic_t refs; |
010e8e6b PB |
840 | struct task_struct *task; |
841 | u64 user_data; | |
d7718a9d | 842 | |
f2f87370 | 843 | struct io_kiocb *link; |
269bbe5f | 844 | struct percpu_ref *fixed_rsrc_refs; |
fcb323cc | 845 | |
b303fe2e | 846 | /* used with ctx->iopoll_list with reads/writes */ |
010e8e6b | 847 | struct list_head inflight_entry; |
5b0a6acc | 848 | struct io_task_work io_task_work; |
010e8e6b PB |
849 | /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */ |
850 | struct hlist_node hash_node; | |
851 | struct async_poll *apoll; | |
852 | struct io_wq_work work; | |
fe7e3257 | 853 | const struct cred *creds; |
c10d1f98 | 854 | |
eae071c9 PB |
855 | /* store used ubuf, so we can prevent reloading */ |
856 | struct io_mapped_ubuf *imu; | |
2b188cc1 | 857 | }; |
05589553 | 858 | |
13bf43f5 PB |
859 | struct io_tctx_node { |
860 | struct list_head ctx_node; | |
861 | struct task_struct *task; | |
13bf43f5 PB |
862 | struct io_ring_ctx *ctx; |
863 | }; | |
864 | ||
27dc8338 PB |
865 | struct io_defer_entry { |
866 | struct list_head list; | |
867 | struct io_kiocb *req; | |
9cf7c104 | 868 | u32 seq; |
2b188cc1 JA |
869 | }; |
870 | ||
d3656344 | 871 | struct io_op_def { |
d3656344 JA |
872 | /* needs req->file assigned */ |
873 | unsigned needs_file : 1; | |
d3656344 JA |
874 | /* hash wq insertion if file is a regular file */ |
875 | unsigned hash_reg_file : 1; | |
876 | /* unbound wq insertion if file is a non-regular file */ | |
877 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
878 | /* opcode is not supported by this kernel */ |
879 | unsigned not_supported : 1; | |
8a72758c JA |
880 | /* set if opcode supports polled "wait" */ |
881 | unsigned pollin : 1; | |
882 | unsigned pollout : 1; | |
bcda7baa JA |
883 | /* op supports buffer selection */ |
884 | unsigned buffer_select : 1; | |
26f0505a PB |
885 | /* do prep async if is going to be punted */ |
886 | unsigned needs_async_setup : 1; | |
27926b68 JA |
887 | /* should block plug */ |
888 | unsigned plug : 1; | |
e8c2bc1f JA |
889 | /* size of async data needed, if any */ |
890 | unsigned short async_size; | |
d3656344 JA |
891 | }; |
892 | ||
0918682b | 893 | static const struct io_op_def io_op_defs[] = { |
0463b6c5 PB |
894 | [IORING_OP_NOP] = {}, |
895 | [IORING_OP_READV] = { | |
d3656344 JA |
896 | .needs_file = 1, |
897 | .unbound_nonreg_file = 1, | |
8a72758c | 898 | .pollin = 1, |
4d954c25 | 899 | .buffer_select = 1, |
26f0505a | 900 | .needs_async_setup = 1, |
27926b68 | 901 | .plug = 1, |
e8c2bc1f | 902 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 903 | }, |
0463b6c5 | 904 | [IORING_OP_WRITEV] = { |
d3656344 JA |
905 | .needs_file = 1, |
906 | .hash_reg_file = 1, | |
907 | .unbound_nonreg_file = 1, | |
8a72758c | 908 | .pollout = 1, |
26f0505a | 909 | .needs_async_setup = 1, |
27926b68 | 910 | .plug = 1, |
e8c2bc1f | 911 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 912 | }, |
0463b6c5 | 913 | [IORING_OP_FSYNC] = { |
d3656344 JA |
914 | .needs_file = 1, |
915 | }, | |
0463b6c5 | 916 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
917 | .needs_file = 1, |
918 | .unbound_nonreg_file = 1, | |
8a72758c | 919 | .pollin = 1, |
27926b68 | 920 | .plug = 1, |
e8c2bc1f | 921 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 922 | }, |
0463b6c5 | 923 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
924 | .needs_file = 1, |
925 | .hash_reg_file = 1, | |
926 | .unbound_nonreg_file = 1, | |
8a72758c | 927 | .pollout = 1, |
27926b68 | 928 | .plug = 1, |
e8c2bc1f | 929 | .async_size = sizeof(struct io_async_rw), |
d3656344 | 930 | }, |
0463b6c5 | 931 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
932 | .needs_file = 1, |
933 | .unbound_nonreg_file = 1, | |
934 | }, | |
0463b6c5 PB |
935 | [IORING_OP_POLL_REMOVE] = {}, |
936 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 JA |
937 | .needs_file = 1, |
938 | }, | |
0463b6c5 | 939 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
940 | .needs_file = 1, |
941 | .unbound_nonreg_file = 1, | |
8a72758c | 942 | .pollout = 1, |
26f0505a | 943 | .needs_async_setup = 1, |
e8c2bc1f | 944 | .async_size = sizeof(struct io_async_msghdr), |
d3656344 | 945 | }, |
0463b6c5 | 946 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
947 | .needs_file = 1, |
948 | .unbound_nonreg_file = 1, | |
8a72758c | 949 | .pollin = 1, |
52de1fe1 | 950 | .buffer_select = 1, |
26f0505a | 951 | .needs_async_setup = 1, |
e8c2bc1f | 952 | .async_size = sizeof(struct io_async_msghdr), |
d3656344 | 953 | }, |
0463b6c5 | 954 | [IORING_OP_TIMEOUT] = { |
e8c2bc1f | 955 | .async_size = sizeof(struct io_timeout_data), |
d3656344 | 956 | }, |
9c8e11b3 PB |
957 | [IORING_OP_TIMEOUT_REMOVE] = { |
958 | /* used by timeout updates' prep() */ | |
9c8e11b3 | 959 | }, |
0463b6c5 | 960 | [IORING_OP_ACCEPT] = { |
d3656344 JA |
961 | .needs_file = 1, |
962 | .unbound_nonreg_file = 1, | |
8a72758c | 963 | .pollin = 1, |
d3656344 | 964 | }, |
0463b6c5 PB |
965 | [IORING_OP_ASYNC_CANCEL] = {}, |
966 | [IORING_OP_LINK_TIMEOUT] = { | |
e8c2bc1f | 967 | .async_size = sizeof(struct io_timeout_data), |
d3656344 | 968 | }, |
0463b6c5 | 969 | [IORING_OP_CONNECT] = { |
d3656344 JA |
970 | .needs_file = 1, |
971 | .unbound_nonreg_file = 1, | |
8a72758c | 972 | .pollout = 1, |
26f0505a | 973 | .needs_async_setup = 1, |
e8c2bc1f | 974 | .async_size = sizeof(struct io_async_connect), |
d3656344 | 975 | }, |
0463b6c5 | 976 | [IORING_OP_FALLOCATE] = { |
d3656344 | 977 | .needs_file = 1, |
d3656344 | 978 | }, |
44526bed JA |
979 | [IORING_OP_OPENAT] = {}, |
980 | [IORING_OP_CLOSE] = {}, | |
981 | [IORING_OP_FILES_UPDATE] = {}, | |
982 | [IORING_OP_STATX] = {}, | |
0463b6c5 | 983 | [IORING_OP_READ] = { |
3a6820f2 JA |
984 | .needs_file = 1, |
985 | .unbound_nonreg_file = 1, | |
8a72758c | 986 | .pollin = 1, |
bcda7baa | 987 | .buffer_select = 1, |
27926b68 | 988 | .plug = 1, |
e8c2bc1f | 989 | .async_size = sizeof(struct io_async_rw), |
3a6820f2 | 990 | }, |
0463b6c5 | 991 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
992 | .needs_file = 1, |
993 | .unbound_nonreg_file = 1, | |
8a72758c | 994 | .pollout = 1, |
27926b68 | 995 | .plug = 1, |
e8c2bc1f | 996 | .async_size = sizeof(struct io_async_rw), |
3a6820f2 | 997 | }, |
0463b6c5 | 998 | [IORING_OP_FADVISE] = { |
4840e418 | 999 | .needs_file = 1, |
c1ca757b | 1000 | }, |
44526bed | 1001 | [IORING_OP_MADVISE] = {}, |
0463b6c5 | 1002 | [IORING_OP_SEND] = { |
fddaface JA |
1003 | .needs_file = 1, |
1004 | .unbound_nonreg_file = 1, | |
8a72758c | 1005 | .pollout = 1, |
fddaface | 1006 | }, |
0463b6c5 | 1007 | [IORING_OP_RECV] = { |
fddaface JA |
1008 | .needs_file = 1, |
1009 | .unbound_nonreg_file = 1, | |
8a72758c | 1010 | .pollin = 1, |
bcda7baa | 1011 | .buffer_select = 1, |
fddaface | 1012 | }, |
0463b6c5 | 1013 | [IORING_OP_OPENAT2] = { |
cebdb986 | 1014 | }, |
3e4827b0 JA |
1015 | [IORING_OP_EPOLL_CTL] = { |
1016 | .unbound_nonreg_file = 1, | |
3e4827b0 | 1017 | }, |
7d67af2c PB |
1018 | [IORING_OP_SPLICE] = { |
1019 | .needs_file = 1, | |
1020 | .hash_reg_file = 1, | |
1021 | .unbound_nonreg_file = 1, | |
ddf0322d JA |
1022 | }, |
1023 | [IORING_OP_PROVIDE_BUFFERS] = {}, | |
067524e9 | 1024 | [IORING_OP_REMOVE_BUFFERS] = {}, |
f2a8d5c7 PB |
1025 | [IORING_OP_TEE] = { |
1026 | .needs_file = 1, | |
1027 | .hash_reg_file = 1, | |
1028 | .unbound_nonreg_file = 1, | |
1029 | }, | |
36f4fa68 JA |
1030 | [IORING_OP_SHUTDOWN] = { |
1031 | .needs_file = 1, | |
1032 | }, | |
44526bed JA |
1033 | [IORING_OP_RENAMEAT] = {}, |
1034 | [IORING_OP_UNLINKAT] = {}, | |
d3656344 JA |
1035 | }; |
1036 | ||
0756a869 PB |
1037 | /* requests with any of those set should undergo io_disarm_next() */ |
1038 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
1039 | ||
7a612350 | 1040 | static bool io_disarm_next(struct io_kiocb *req); |
eef51daa | 1041 | static void io_uring_del_tctx_node(unsigned long index); |
9936c7c2 PB |
1042 | static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
1043 | struct task_struct *task, | |
3dd0c97a | 1044 | bool cancel_all); |
78cc687b | 1045 | static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd); |
1ffc5422 | 1046 | |
d4d19c19 PB |
1047 | static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1048 | long res, unsigned int cflags); | |
ec9c02ad | 1049 | static void io_put_req(struct io_kiocb *req); |
91c2f697 | 1050 | static void io_put_req_deferred(struct io_kiocb *req); |
c7dae4ba | 1051 | static void io_dismantle_req(struct io_kiocb *req); |
94ae5e77 | 1052 | static void io_queue_linked_timeout(struct io_kiocb *req); |
fdecb662 | 1053 | static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, |
c3bdad02 | 1054 | struct io_uring_rsrc_update2 *up, |
98f0b3b4 | 1055 | unsigned nr_args); |
68fb8979 | 1056 | static void io_clean_op(struct io_kiocb *req); |
ac177053 | 1057 | static struct file *io_file_get(struct io_ring_ctx *ctx, |
8371adf5 | 1058 | struct io_kiocb *req, int fd, bool fixed); |
c5eef2b9 | 1059 | static void __io_queue_sqe(struct io_kiocb *req); |
269bbe5f | 1060 | static void io_rsrc_put_work(struct work_struct *work); |
de0617e4 | 1061 | |
907d1df3 | 1062 | static void io_req_task_queue(struct io_kiocb *req); |
2a2758f2 | 1063 | static void io_submit_flush_completions(struct io_ring_ctx *ctx); |
179ae0d1 | 1064 | static int io_req_prep_async(struct io_kiocb *req); |
de0617e4 | 1065 | |
2b188cc1 JA |
1066 | static struct kmem_cache *req_cachep; |
1067 | ||
0918682b | 1068 | static const struct file_operations io_uring_fops; |
2b188cc1 JA |
1069 | |
1070 | struct sock *io_uring_get_socket(struct file *file) | |
1071 | { | |
1072 | #if defined(CONFIG_UNIX) | |
1073 | if (file->f_op == &io_uring_fops) { | |
1074 | struct io_ring_ctx *ctx = file->private_data; | |
1075 | ||
1076 | return ctx->ring_sock->sk; | |
1077 | } | |
1078 | #endif | |
1079 | return NULL; | |
1080 | } | |
1081 | EXPORT_SYMBOL(io_uring_get_socket); | |
1082 | ||
f2f87370 PB |
1083 | #define io_for_each_link(pos, head) \ |
1084 | for (pos = (head); pos; pos = pos->link) | |
1085 | ||
21c843d5 PB |
1086 | /* |
1087 | * Shamelessly stolen from the mm implementation of page reference checking, | |
1088 | * see commit f958d7b528b1 for details. | |
1089 | */ | |
1090 | #define req_ref_zero_or_close_to_overflow(req) \ | |
1091 | ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u) | |
1092 | ||
1093 | static inline bool req_ref_inc_not_zero(struct io_kiocb *req) | |
1094 | { | |
20e60a38 | 1095 | WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT)); |
21c843d5 PB |
1096 | return atomic_inc_not_zero(&req->refs); |
1097 | } | |
1098 | ||
21c843d5 PB |
1099 | static inline bool req_ref_put_and_test(struct io_kiocb *req) |
1100 | { | |
20e60a38 PB |
1101 | if (likely(!(req->flags & REQ_F_REFCOUNT))) |
1102 | return true; | |
1103 | ||
21c843d5 PB |
1104 | WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req)); |
1105 | return atomic_dec_and_test(&req->refs); | |
1106 | } | |
1107 | ||
1108 | static inline void req_ref_put(struct io_kiocb *req) | |
1109 | { | |
20e60a38 | 1110 | WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT)); |
21c843d5 PB |
1111 | WARN_ON_ONCE(req_ref_put_and_test(req)); |
1112 | } | |
1113 | ||
1114 | static inline void req_ref_get(struct io_kiocb *req) | |
1115 | { | |
20e60a38 | 1116 | WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT)); |
21c843d5 PB |
1117 | WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req)); |
1118 | atomic_inc(&req->refs); | |
1119 | } | |
1120 | ||
48dcd38d | 1121 | static inline void __io_req_set_refcount(struct io_kiocb *req, int nr) |
20e60a38 PB |
1122 | { |
1123 | if (!(req->flags & REQ_F_REFCOUNT)) { | |
1124 | req->flags |= REQ_F_REFCOUNT; | |
48dcd38d | 1125 | atomic_set(&req->refs, nr); |
20e60a38 PB |
1126 | } |
1127 | } | |
1128 | ||
48dcd38d PB |
1129 | static inline void io_req_set_refcount(struct io_kiocb *req) |
1130 | { | |
1131 | __io_req_set_refcount(req, 1); | |
1132 | } | |
1133 | ||
b895c9a6 | 1134 | static inline void io_req_set_rsrc_node(struct io_kiocb *req) |
36f72fe2 PB |
1135 | { |
1136 | struct io_ring_ctx *ctx = req->ctx; | |
1137 | ||
269bbe5f | 1138 | if (!req->fixed_rsrc_refs) { |
a7f0ed5a | 1139 | req->fixed_rsrc_refs = &ctx->rsrc_node->refs; |
269bbe5f | 1140 | percpu_ref_get(req->fixed_rsrc_refs); |
36f72fe2 PB |
1141 | } |
1142 | } | |
1143 | ||
f70865db PB |
1144 | static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl) |
1145 | { | |
1146 | bool got = percpu_ref_tryget(ref); | |
1147 | ||
1148 | /* already at zero, wait for ->release() */ | |
1149 | if (!got) | |
1150 | wait_for_completion(compl); | |
1151 | percpu_ref_resurrect(ref); | |
1152 | if (got) | |
1153 | percpu_ref_put(ref); | |
1154 | } | |
1155 | ||
3dd0c97a PB |
1156 | static bool io_match_task(struct io_kiocb *head, struct task_struct *task, |
1157 | bool cancel_all) | |
08d23634 PB |
1158 | { |
1159 | struct io_kiocb *req; | |
1160 | ||
68207680 | 1161 | if (task && head->task != task) |
08d23634 | 1162 | return false; |
3dd0c97a | 1163 | if (cancel_all) |
08d23634 PB |
1164 | return true; |
1165 | ||
1166 | io_for_each_link(req, head) { | |
b05a1bcd | 1167 | if (req->flags & REQ_F_INFLIGHT) |
02a13674 | 1168 | return true; |
08d23634 PB |
1169 | } |
1170 | return false; | |
1171 | } | |
1172 | ||
93d2bcd2 | 1173 | static inline void req_set_fail(struct io_kiocb *req) |
c40f6379 | 1174 | { |
93d2bcd2 | 1175 | req->flags |= REQ_F_FAIL; |
c40f6379 | 1176 | } |
4a38aed2 | 1177 | |
2b188cc1 JA |
1178 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) |
1179 | { | |
1180 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
1181 | ||
0f158b4c | 1182 | complete(&ctx->ref_comp); |
2b188cc1 JA |
1183 | } |
1184 | ||
8eb7e2d0 PB |
1185 | static inline bool io_is_timeout_noseq(struct io_kiocb *req) |
1186 | { | |
1187 | return !req->timeout.off; | |
1188 | } | |
1189 | ||
f56165e6 PB |
1190 | static void io_fallback_req_func(struct work_struct *work) |
1191 | { | |
1192 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
1193 | fallback_work.work); | |
1194 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
1195 | struct io_kiocb *req, *tmp; | |
1196 | ||
1197 | percpu_ref_get(&ctx->refs); | |
1198 | llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node) | |
1199 | req->io_task_work.func(req); | |
1200 | percpu_ref_put(&ctx->refs); | |
1201 | } | |
1202 | ||
2b188cc1 JA |
1203 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
1204 | { | |
1205 | struct io_ring_ctx *ctx; | |
78076bb6 | 1206 | int hash_bits; |
2b188cc1 JA |
1207 | |
1208 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
1209 | if (!ctx) | |
1210 | return NULL; | |
1211 | ||
78076bb6 JA |
1212 | /* |
1213 | * Use 5 bits less than the max cq entries, that should give us around | |
1214 | * 32 entries per hash list if totally full and uniformly spread. | |
1215 | */ | |
1216 | hash_bits = ilog2(p->cq_entries); | |
1217 | hash_bits -= 5; | |
1218 | if (hash_bits <= 0) | |
1219 | hash_bits = 1; | |
1220 | ctx->cancel_hash_bits = hash_bits; | |
1221 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
1222 | GFP_KERNEL); | |
1223 | if (!ctx->cancel_hash) | |
1224 | goto err; | |
1225 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
1226 | ||
6224843d PB |
1227 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
1228 | if (!ctx->dummy_ubuf) | |
1229 | goto err; | |
1230 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
1231 | ctx->dummy_ubuf->ubuf = -1UL; | |
1232 | ||
21482896 | 1233 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
1234 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
1235 | goto err; | |
2b188cc1 JA |
1236 | |
1237 | ctx->flags = p->flags; | |
90554200 | 1238 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 1239 | INIT_LIST_HEAD(&ctx->sqd_list); |
311997b3 | 1240 | init_waitqueue_head(&ctx->poll_wait); |
1d7bb1d5 | 1241 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
0f158b4c | 1242 | init_completion(&ctx->ref_comp); |
9e15c3a0 | 1243 | xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1); |
61cf9370 | 1244 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 1245 | mutex_init(&ctx->uring_lock); |
311997b3 | 1246 | init_waitqueue_head(&ctx->cq_wait); |
2b188cc1 | 1247 | spin_lock_init(&ctx->completion_lock); |
89850fce | 1248 | spin_lock_init(&ctx->timeout_lock); |
540e32a0 | 1249 | INIT_LIST_HEAD(&ctx->iopoll_list); |
de0617e4 | 1250 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 1251 | INIT_LIST_HEAD(&ctx->timeout_list); |
d67d2263 BM |
1252 | spin_lock_init(&ctx->rsrc_ref_lock); |
1253 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f BM |
1254 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
1255 | init_llist_head(&ctx->rsrc_put_llist); | |
13bf43f5 | 1256 | INIT_LIST_HEAD(&ctx->tctx_list); |
cd0ca2e0 | 1257 | INIT_LIST_HEAD(&ctx->submit_state.free_list); |
d0acdee2 | 1258 | INIT_LIST_HEAD(&ctx->locked_free_list); |
9011bf9a | 1259 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
2b188cc1 | 1260 | return ctx; |
206aefde | 1261 | err: |
6224843d | 1262 | kfree(ctx->dummy_ubuf); |
78076bb6 | 1263 | kfree(ctx->cancel_hash); |
206aefde JA |
1264 | kfree(ctx); |
1265 | return NULL; | |
2b188cc1 JA |
1266 | } |
1267 | ||
8f6ed49a PB |
1268 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
1269 | { | |
1270 | struct io_rings *r = ctx->rings; | |
1271 | ||
1272 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
1273 | ctx->cq_extra--; | |
1274 | } | |
1275 | ||
9cf7c104 | 1276 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 1277 | { |
2bc9930e JA |
1278 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
1279 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 1280 | |
8f6ed49a | 1281 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 1282 | } |
de0617e4 | 1283 | |
9d858b21 | 1284 | return false; |
de0617e4 JA |
1285 | } |
1286 | ||
c97d8a0f PB |
1287 | #define FFS_ASYNC_READ 0x1UL |
1288 | #define FFS_ASYNC_WRITE 0x2UL | |
1289 | #ifdef CONFIG_64BIT | |
1290 | #define FFS_ISREG 0x4UL | |
1291 | #else | |
1292 | #define FFS_ISREG 0x0UL | |
1293 | #endif | |
1294 | #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG) | |
1295 | ||
1296 | static inline bool io_req_ffs_set(struct io_kiocb *req) | |
1297 | { | |
1298 | return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE); | |
1299 | } | |
1300 | ||
ce3d5aae PB |
1301 | static void io_req_track_inflight(struct io_kiocb *req) |
1302 | { | |
ce3d5aae | 1303 | if (!(req->flags & REQ_F_INFLIGHT)) { |
ce3d5aae | 1304 | req->flags |= REQ_F_INFLIGHT; |
b303fe2e | 1305 | atomic_inc(¤t->io_uring->inflight_tracked); |
ce3d5aae PB |
1306 | } |
1307 | } | |
1308 | ||
906c6caa PB |
1309 | static inline void io_unprep_linked_timeout(struct io_kiocb *req) |
1310 | { | |
1311 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
1312 | } | |
1313 | ||
fd08e530 PB |
1314 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
1315 | { | |
906c6caa PB |
1316 | if (WARN_ON_ONCE(!req->link)) |
1317 | return NULL; | |
1318 | ||
4d13d1a4 PB |
1319 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
1320 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
1321 | |
1322 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 1323 | io_req_set_refcount(req); |
4d13d1a4 PB |
1324 | __io_req_set_refcount(req->link, 2); |
1325 | return req->link; | |
fd08e530 PB |
1326 | } |
1327 | ||
1328 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
1329 | { | |
4d13d1a4 | 1330 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
1331 | return NULL; |
1332 | return __io_prep_linked_timeout(req); | |
1333 | } | |
1334 | ||
1e6fa521 JA |
1335 | static void io_prep_async_work(struct io_kiocb *req) |
1336 | { | |
1337 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
1338 | struct io_ring_ctx *ctx = req->ctx; |
1339 | ||
b8e64b53 PB |
1340 | if (!(req->flags & REQ_F_CREDS)) { |
1341 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 1342 | req->creds = get_current_cred(); |
b8e64b53 | 1343 | } |
003e8dcc | 1344 | |
e1d675df PB |
1345 | req->work.list.next = NULL; |
1346 | req->work.flags = 0; | |
feaadc4f PB |
1347 | if (req->flags & REQ_F_FORCE_ASYNC) |
1348 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
1349 | ||
1e6fa521 JA |
1350 | if (req->flags & REQ_F_ISREG) { |
1351 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1352 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 1353 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
1354 | if (def->unbound_nonreg_file) |
1355 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1356 | } | |
e1d675df PB |
1357 | |
1358 | switch (req->opcode) { | |
1359 | case IORING_OP_SPLICE: | |
1360 | case IORING_OP_TEE: | |
e1d675df PB |
1361 | if (!S_ISREG(file_inode(req->splice.file_in)->i_mode)) |
1362 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1363 | break; | |
1364 | } | |
561fb04a | 1365 | } |
cccf0ee8 | 1366 | |
cbdcb435 | 1367 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 1368 | { |
cbdcb435 | 1369 | struct io_kiocb *cur; |
54a91f3b | 1370 | |
44eff40a PB |
1371 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
1372 | struct io_ring_ctx *ctx = req->ctx; | |
1373 | ||
79ebeaee | 1374 | spin_lock(&ctx->completion_lock); |
44eff40a PB |
1375 | io_for_each_link(cur, req) |
1376 | io_prep_async_work(cur); | |
79ebeaee | 1377 | spin_unlock(&ctx->completion_lock); |
44eff40a PB |
1378 | } else { |
1379 | io_for_each_link(cur, req) | |
1380 | io_prep_async_work(cur); | |
1381 | } | |
561fb04a JA |
1382 | } |
1383 | ||
ebf93667 | 1384 | static void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1385 | { |
a197f664 | 1386 | struct io_ring_ctx *ctx = req->ctx; |
cbdcb435 | 1387 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 1388 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 1389 | |
3bfe6106 JA |
1390 | BUG_ON(!tctx); |
1391 | BUG_ON(!tctx->io_wq); | |
561fb04a | 1392 | |
cbdcb435 PB |
1393 | /* init ->work of the whole link before punting */ |
1394 | io_prep_async_link(req); | |
991468dc JA |
1395 | |
1396 | /* | |
1397 | * Not expected to happen, but if we do have a bug where this _can_ | |
1398 | * happen, catch it here and ensure the request is marked as | |
1399 | * canceled. That will make io-wq go through the usual work cancel | |
1400 | * procedure rather than attempt to run this request (or create a new | |
1401 | * worker for it). | |
1402 | */ | |
1403 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
1404 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1405 | ||
d07f1e8a PB |
1406 | trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req, |
1407 | &req->work, req->flags); | |
ebf93667 | 1408 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
1409 | if (link) |
1410 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
1411 | } |
1412 | ||
1ee4160c | 1413 | static void io_kill_timeout(struct io_kiocb *req, int status) |
8c855885 | 1414 | __must_hold(&req->ctx->completion_lock) |
89850fce | 1415 | __must_hold(&req->ctx->timeout_lock) |
5262f567 | 1416 | { |
e8c2bc1f | 1417 | struct io_timeout_data *io = req->async_data; |
5262f567 | 1418 | |
fd9c7bc5 | 1419 | if (hrtimer_try_to_cancel(&io->timer) != -1) { |
01cec8c1 PB |
1420 | atomic_set(&req->ctx->cq_timeouts, |
1421 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
135fcde8 | 1422 | list_del_init(&req->timeout.list); |
d4d19c19 | 1423 | io_cqring_fill_event(req->ctx, req->user_data, status, 0); |
91c2f697 | 1424 | io_put_req_deferred(req); |
5262f567 JA |
1425 | } |
1426 | } | |
1427 | ||
441b8a78 | 1428 | static void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 1429 | { |
441b8a78 | 1430 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
1431 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
1432 | struct io_defer_entry, list); | |
de0617e4 | 1433 | |
9cf7c104 | 1434 | if (req_need_defer(de->req, de->seq)) |
04518945 | 1435 | break; |
27dc8338 | 1436 | list_del_init(&de->list); |
907d1df3 | 1437 | io_req_task_queue(de->req); |
27dc8338 | 1438 | kfree(de); |
441b8a78 | 1439 | } |
04518945 PB |
1440 | } |
1441 | ||
360428f8 | 1442 | static void io_flush_timeouts(struct io_ring_ctx *ctx) |
89850fce | 1443 | __must_hold(&ctx->completion_lock) |
de0617e4 | 1444 | { |
441b8a78 | 1445 | u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
f010505b | 1446 | |
79ebeaee | 1447 | spin_lock_irq(&ctx->timeout_lock); |
f18ee4cf | 1448 | while (!list_empty(&ctx->timeout_list)) { |
f010505b | 1449 | u32 events_needed, events_got; |
360428f8 | 1450 | struct io_kiocb *req = list_first_entry(&ctx->timeout_list, |
135fcde8 | 1451 | struct io_kiocb, timeout.list); |
de0617e4 | 1452 | |
8eb7e2d0 | 1453 | if (io_is_timeout_noseq(req)) |
360428f8 | 1454 | break; |
f010505b MDG |
1455 | |
1456 | /* | |
1457 | * Since seq can easily wrap around over time, subtract | |
1458 | * the last seq at which timeouts were flushed before comparing. | |
1459 | * Assuming not more than 2^31-1 events have happened since, | |
1460 | * these subtractions won't have wrapped, so we can check if | |
1461 | * target is in [last_seq, current_seq] by comparing the two. | |
1462 | */ | |
1463 | events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush; | |
1464 | events_got = seq - ctx->cq_last_tm_flush; | |
1465 | if (events_got < events_needed) | |
360428f8 | 1466 | break; |
bfe68a22 | 1467 | |
135fcde8 | 1468 | list_del_init(&req->timeout.list); |
1ee4160c | 1469 | io_kill_timeout(req, 0); |
f18ee4cf | 1470 | } |
f010505b | 1471 | ctx->cq_last_tm_flush = seq; |
79ebeaee | 1472 | spin_unlock_irq(&ctx->timeout_lock); |
360428f8 | 1473 | } |
5262f567 | 1474 | |
2335f6f5 | 1475 | static void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
360428f8 | 1476 | { |
2335f6f5 PB |
1477 | if (ctx->off_timeout_used) |
1478 | io_flush_timeouts(ctx); | |
1479 | if (ctx->drain_active) | |
1480 | io_queue_deferred(ctx); | |
1481 | } | |
1482 | ||
1483 | static inline void io_commit_cqring(struct io_ring_ctx *ctx) | |
1484 | { | |
1485 | if (unlikely(ctx->off_timeout_used || ctx->drain_active)) | |
1486 | __io_commit_cqring_flush(ctx); | |
ec30e04b PB |
1487 | /* order cqe stores with ring update */ |
1488 | smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail); | |
de0617e4 JA |
1489 | } |
1490 | ||
90554200 JA |
1491 | static inline bool io_sqring_full(struct io_ring_ctx *ctx) |
1492 | { | |
1493 | struct io_rings *r = ctx->rings; | |
1494 | ||
a566c556 | 1495 | return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries; |
90554200 JA |
1496 | } |
1497 | ||
888aae2e PB |
1498 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
1499 | { | |
1500 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
1501 | } | |
1502 | ||
d068b506 | 1503 | static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) |
2b188cc1 | 1504 | { |
75b28aff | 1505 | struct io_rings *rings = ctx->rings; |
ea5ab3b5 | 1506 | unsigned tail, mask = ctx->cq_entries - 1; |
2b188cc1 | 1507 | |
115e12e5 SB |
1508 | /* |
1509 | * writes to the cq entry need to come after reading head; the | |
1510 | * control dependency is enough as we're using WRITE_ONCE to | |
1511 | * fill the cq entry | |
1512 | */ | |
a566c556 | 1513 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
2b188cc1 JA |
1514 | return NULL; |
1515 | ||
888aae2e | 1516 | tail = ctx->cached_cq_tail++; |
ea5ab3b5 | 1517 | return &rings->cqes[tail & mask]; |
2b188cc1 JA |
1518 | } |
1519 | ||
f2842ab5 JA |
1520 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1521 | { | |
44c769de | 1522 | if (likely(!ctx->cq_ev_fd)) |
f0b493e6 | 1523 | return false; |
7e55a19c SG |
1524 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
1525 | return false; | |
44c769de | 1526 | return !ctx->eventfd_async || io_wq_current_is_worker(); |
f2842ab5 JA |
1527 | } |
1528 | ||
b41e9852 | 1529 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 1530 | { |
5fd46178 JA |
1531 | /* |
1532 | * wake_up_all() may seem excessive, but io_wake_function() and | |
1533 | * io_should_wake() handle the termination of the loop and only | |
1534 | * wake as many waiters as we need to. | |
1535 | */ | |
1536 | if (wq_has_sleeper(&ctx->cq_wait)) | |
1537 | wake_up_all(&ctx->cq_wait); | |
534ca6d6 JA |
1538 | if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait)) |
1539 | wake_up(&ctx->sq_data->wait); | |
b41e9852 | 1540 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 | 1541 | eventfd_signal(ctx->cq_ev_fd, 1); |
311997b3 PB |
1542 | if (waitqueue_active(&ctx->poll_wait)) { |
1543 | wake_up_interruptible(&ctx->poll_wait); | |
4aa84f2f PB |
1544 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); |
1545 | } | |
1d7bb1d5 JA |
1546 | } |
1547 | ||
80c18e4a PB |
1548 | static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx) |
1549 | { | |
1550 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
5fd46178 JA |
1551 | if (wq_has_sleeper(&ctx->cq_wait)) |
1552 | wake_up_all(&ctx->cq_wait); | |
80c18e4a PB |
1553 | } |
1554 | if (io_should_trigger_evfd(ctx)) | |
1555 | eventfd_signal(ctx->cq_ev_fd, 1); | |
311997b3 PB |
1556 | if (waitqueue_active(&ctx->poll_wait)) { |
1557 | wake_up_interruptible(&ctx->poll_wait); | |
4aa84f2f PB |
1558 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); |
1559 | } | |
80c18e4a PB |
1560 | } |
1561 | ||
c4a2ed72 | 1562 | /* Returns true if there are no backlogged entries after the flush */ |
6c2450ae | 1563 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) |
1d7bb1d5 | 1564 | { |
b18032bb | 1565 | bool all_flushed, posted; |
1d7bb1d5 | 1566 | |
a566c556 | 1567 | if (!force && __io_cqring_events(ctx) == ctx->cq_entries) |
e23de15f | 1568 | return false; |
1d7bb1d5 | 1569 | |
b18032bb | 1570 | posted = false; |
79ebeaee | 1571 | spin_lock(&ctx->completion_lock); |
6c2450ae | 1572 | while (!list_empty(&ctx->cq_overflow_list)) { |
d068b506 | 1573 | struct io_uring_cqe *cqe = io_get_cqe(ctx); |
6c2450ae | 1574 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 1575 | |
1d7bb1d5 JA |
1576 | if (!cqe && !force) |
1577 | break; | |
6c2450ae PB |
1578 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
1579 | struct io_overflow_cqe, list); | |
1580 | if (cqe) | |
1581 | memcpy(cqe, &ocqe->cqe, sizeof(*cqe)); | |
1582 | else | |
8f6ed49a PB |
1583 | io_account_cq_overflow(ctx); |
1584 | ||
b18032bb | 1585 | posted = true; |
6c2450ae PB |
1586 | list_del(&ocqe->list); |
1587 | kfree(ocqe); | |
1d7bb1d5 JA |
1588 | } |
1589 | ||
09e88404 PB |
1590 | all_flushed = list_empty(&ctx->cq_overflow_list); |
1591 | if (all_flushed) { | |
5ed7a37d | 1592 | clear_bit(0, &ctx->check_cq_overflow); |
20c0b380 NA |
1593 | WRITE_ONCE(ctx->rings->sq_flags, |
1594 | ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW); | |
09e88404 | 1595 | } |
46930143 | 1596 | |
b18032bb JA |
1597 | if (posted) |
1598 | io_commit_cqring(ctx); | |
79ebeaee | 1599 | spin_unlock(&ctx->completion_lock); |
b18032bb JA |
1600 | if (posted) |
1601 | io_cqring_ev_posted(ctx); | |
09e88404 | 1602 | return all_flushed; |
1d7bb1d5 JA |
1603 | } |
1604 | ||
90f67366 | 1605 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 1606 | { |
ca0a2651 JA |
1607 | bool ret = true; |
1608 | ||
5ed7a37d | 1609 | if (test_bit(0, &ctx->check_cq_overflow)) { |
6c503150 PB |
1610 | /* iopoll syncs against uring_lock, not completion_lock */ |
1611 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
1612 | mutex_lock(&ctx->uring_lock); | |
90f67366 | 1613 | ret = __io_cqring_overflow_flush(ctx, false); |
6c503150 PB |
1614 | if (ctx->flags & IORING_SETUP_IOPOLL) |
1615 | mutex_unlock(&ctx->uring_lock); | |
1616 | } | |
ca0a2651 JA |
1617 | |
1618 | return ret; | |
6c503150 PB |
1619 | } |
1620 | ||
6a290a14 PB |
1621 | /* must to be called somewhat shortly after putting a request */ |
1622 | static inline void io_put_task(struct task_struct *task, int nr) | |
1623 | { | |
1624 | struct io_uring_task *tctx = task->io_uring; | |
1625 | ||
1626 | percpu_counter_sub(&tctx->inflight, nr); | |
1627 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1628 | wake_up(&tctx->wait); | |
1629 | put_task_struct_many(task, nr); | |
1630 | } | |
1631 | ||
d4d19c19 PB |
1632 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
1633 | long res, unsigned int cflags) | |
2b188cc1 | 1634 | { |
cce4b8b0 | 1635 | struct io_overflow_cqe *ocqe; |
2b188cc1 | 1636 | |
cce4b8b0 PB |
1637 | ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT); |
1638 | if (!ocqe) { | |
1639 | /* | |
1640 | * If we're in ring overflow flush mode, or in task cancel mode, | |
1641 | * or cannot allocate an overflow entry, then we need to drop it | |
1642 | * on the floor. | |
1643 | */ | |
8f6ed49a | 1644 | io_account_cq_overflow(ctx); |
cce4b8b0 | 1645 | return false; |
2b188cc1 | 1646 | } |
cce4b8b0 | 1647 | if (list_empty(&ctx->cq_overflow_list)) { |
5ed7a37d | 1648 | set_bit(0, &ctx->check_cq_overflow); |
20c0b380 NA |
1649 | WRITE_ONCE(ctx->rings->sq_flags, |
1650 | ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW); | |
1651 | ||
cce4b8b0 | 1652 | } |
d4d19c19 | 1653 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
1654 | ocqe->cqe.res = res; |
1655 | ocqe->cqe.flags = cflags; | |
1656 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); | |
1657 | return true; | |
2b188cc1 JA |
1658 | } |
1659 | ||
d4d19c19 PB |
1660 | static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1661 | long res, unsigned int cflags) | |
2b188cc1 JA |
1662 | { |
1663 | struct io_uring_cqe *cqe; | |
1664 | ||
d4d19c19 | 1665 | trace_io_uring_complete(ctx, user_data, res, cflags); |
51c3ff62 | 1666 | |
2b188cc1 JA |
1667 | /* |
1668 | * If we can't get a cq entry, userspace overflowed the | |
1669 | * submission (by quite a lot). Increment the overflow count in | |
1670 | * the ring. | |
1671 | */ | |
d068b506 | 1672 | cqe = io_get_cqe(ctx); |
1d7bb1d5 | 1673 | if (likely(cqe)) { |
d4d19c19 | 1674 | WRITE_ONCE(cqe->user_data, user_data); |
2b188cc1 | 1675 | WRITE_ONCE(cqe->res, res); |
bcda7baa | 1676 | WRITE_ONCE(cqe->flags, cflags); |
8d13326e | 1677 | return true; |
2b188cc1 | 1678 | } |
d4d19c19 | 1679 | return io_cqring_event_overflow(ctx, user_data, res, cflags); |
2b188cc1 JA |
1680 | } |
1681 | ||
8d13326e | 1682 | /* not as hot to bloat with inlining */ |
d4d19c19 PB |
1683 | static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1684 | long res, unsigned int cflags) | |
bcda7baa | 1685 | { |
d4d19c19 | 1686 | return __io_cqring_fill_event(ctx, user_data, res, cflags); |
bcda7baa JA |
1687 | } |
1688 | ||
7a612350 PB |
1689 | static void io_req_complete_post(struct io_kiocb *req, long res, |
1690 | unsigned int cflags) | |
2b188cc1 | 1691 | { |
78e19bbe | 1692 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1693 | |
79ebeaee | 1694 | spin_lock(&ctx->completion_lock); |
d4d19c19 | 1695 | __io_cqring_fill_event(ctx, req->user_data, res, cflags); |
c7dae4ba JA |
1696 | /* |
1697 | * If we're the last reference to this request, add to our locked | |
1698 | * free_list cache. | |
1699 | */ | |
de9b4cca | 1700 | if (req_ref_put_and_test(req)) { |
7a612350 | 1701 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
0756a869 | 1702 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
1703 | io_disarm_next(req); |
1704 | if (req->link) { | |
1705 | io_req_task_queue(req->link); | |
1706 | req->link = NULL; | |
1707 | } | |
1708 | } | |
c7dae4ba JA |
1709 | io_dismantle_req(req); |
1710 | io_put_task(req->task, 1); | |
bb943b82 | 1711 | list_add(&req->inflight_entry, &ctx->locked_free_list); |
d0acdee2 | 1712 | ctx->locked_free_nr++; |
180f829f PB |
1713 | } else { |
1714 | if (!percpu_ref_tryget(&ctx->refs)) | |
1715 | req = NULL; | |
1716 | } | |
7a612350 | 1717 | io_commit_cqring(ctx); |
79ebeaee | 1718 | spin_unlock(&ctx->completion_lock); |
7a612350 | 1719 | |
180f829f PB |
1720 | if (req) { |
1721 | io_cqring_ev_posted(ctx); | |
c7dae4ba | 1722 | percpu_ref_put(&ctx->refs); |
180f829f | 1723 | } |
229a7b63 JA |
1724 | } |
1725 | ||
4e3d9ff9 JA |
1726 | static inline bool io_req_needs_clean(struct io_kiocb *req) |
1727 | { | |
c854357b | 1728 | return req->flags & IO_REQ_CLEAN_FLAGS; |
4e3d9ff9 JA |
1729 | } |
1730 | ||
a38d68db | 1731 | static void io_req_complete_state(struct io_kiocb *req, long res, |
889fca73 | 1732 | unsigned int cflags) |
229a7b63 | 1733 | { |
4e3d9ff9 | 1734 | if (io_req_needs_clean(req)) |
68fb8979 | 1735 | io_clean_op(req); |
a38d68db PB |
1736 | req->result = res; |
1737 | req->compl.cflags = cflags; | |
e342c807 | 1738 | req->flags |= REQ_F_COMPLETE_INLINE; |
e1e16097 JA |
1739 | } |
1740 | ||
889fca73 PB |
1741 | static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags, |
1742 | long res, unsigned cflags) | |
bcda7baa | 1743 | { |
889fca73 PB |
1744 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) |
1745 | io_req_complete_state(req, res, cflags); | |
a38d68db | 1746 | else |
c7dae4ba | 1747 | io_req_complete_post(req, res, cflags); |
bcda7baa JA |
1748 | } |
1749 | ||
a38d68db | 1750 | static inline void io_req_complete(struct io_kiocb *req, long res) |
0ddf92e8 | 1751 | { |
889fca73 | 1752 | __io_req_complete(req, 0, res, 0); |
0ddf92e8 JA |
1753 | } |
1754 | ||
f41db273 PB |
1755 | static void io_req_complete_failed(struct io_kiocb *req, long res) |
1756 | { | |
93d2bcd2 | 1757 | req_set_fail(req); |
f41db273 PB |
1758 | io_req_complete_post(req, res, 0); |
1759 | } | |
1760 | ||
864ea921 PB |
1761 | /* |
1762 | * Don't initialise the fields below on every allocation, but do that in | |
1763 | * advance and keep them valid across allocations. | |
1764 | */ | |
1765 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1766 | { | |
1767 | req->ctx = ctx; | |
1768 | req->link = NULL; | |
1769 | req->async_data = NULL; | |
1770 | /* not necessary, but safer to zero */ | |
1771 | req->result = 0; | |
1772 | } | |
1773 | ||
dac7a098 | 1774 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1775 | struct io_submit_state *state) |
dac7a098 | 1776 | { |
79ebeaee | 1777 | spin_lock(&ctx->completion_lock); |
cd0ca2e0 | 1778 | list_splice_init(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1779 | ctx->locked_free_nr = 0; |
79ebeaee | 1780 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1781 | } |
1782 | ||
dd78f492 | 1783 | /* Returns true IFF there are requests in the cache */ |
c7dae4ba | 1784 | static bool io_flush_cached_reqs(struct io_ring_ctx *ctx) |
0ddf92e8 | 1785 | { |
c7dae4ba | 1786 | struct io_submit_state *state = &ctx->submit_state; |
dd78f492 | 1787 | int nr; |
0ddf92e8 | 1788 | |
c7dae4ba JA |
1789 | /* |
1790 | * If we have more than a batch's worth of requests in our IRQ side | |
1791 | * locked cache, grab the lock and move them over to our submission | |
1792 | * side cache. | |
1793 | */ | |
d0acdee2 | 1794 | if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH) |
cd0ca2e0 | 1795 | io_flush_cached_locked_reqs(ctx, state); |
0ddf92e8 | 1796 | |
dd78f492 | 1797 | nr = state->free_reqs; |
cd0ca2e0 PB |
1798 | while (!list_empty(&state->free_list)) { |
1799 | struct io_kiocb *req = list_first_entry(&state->free_list, | |
bb943b82 | 1800 | struct io_kiocb, inflight_entry); |
dd78f492 | 1801 | |
bb943b82 | 1802 | list_del(&req->inflight_entry); |
dd78f492 PB |
1803 | state->reqs[nr++] = req; |
1804 | if (nr == ARRAY_SIZE(state->reqs)) | |
e5d1bc0a | 1805 | break; |
1b4c351f JA |
1806 | } |
1807 | ||
dd78f492 PB |
1808 | state->free_reqs = nr; |
1809 | return nr != 0; | |
0ddf92e8 JA |
1810 | } |
1811 | ||
5d5901a3 PB |
1812 | /* |
1813 | * A request might get retired back into the request caches even before opcode | |
1814 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1815 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1816 | * and with extra caution to not get a request that is still worked on. | |
1817 | */ | |
e5d1bc0a | 1818 | static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx) |
5d5901a3 | 1819 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1820 | { |
e5d1bc0a | 1821 | struct io_submit_state *state = &ctx->submit_state; |
864ea921 PB |
1822 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
1823 | int ret, i; | |
e5d1bc0a | 1824 | |
fe7e3257 | 1825 | BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH); |
e5d1bc0a | 1826 | |
864ea921 PB |
1827 | if (likely(state->free_reqs || io_flush_cached_reqs(ctx))) |
1828 | goto got_req; | |
e5d1bc0a | 1829 | |
864ea921 PB |
1830 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH, |
1831 | state->reqs); | |
fd6fab2c | 1832 | |
864ea921 PB |
1833 | /* |
1834 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1835 | * retry single alloc to be on the safe side. | |
1836 | */ | |
1837 | if (unlikely(ret <= 0)) { | |
1838 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1839 | if (!state->reqs[0]) | |
1840 | return NULL; | |
1841 | ret = 1; | |
2b188cc1 | 1842 | } |
864ea921 PB |
1843 | |
1844 | for (i = 0; i < ret; i++) | |
1845 | io_preinit_req(state->reqs[i], ctx); | |
1846 | state->free_reqs = ret; | |
e5d1bc0a | 1847 | got_req: |
291b2821 PB |
1848 | state->free_reqs--; |
1849 | return state->reqs[state->free_reqs]; | |
2b188cc1 JA |
1850 | } |
1851 | ||
e1d767f0 | 1852 | static inline void io_put_file(struct file *file) |
8da11c19 | 1853 | { |
e1d767f0 | 1854 | if (file) |
8da11c19 PB |
1855 | fput(file); |
1856 | } | |
1857 | ||
4edf20f9 | 1858 | static void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1859 | { |
094bae49 | 1860 | unsigned int flags = req->flags; |
929a3af9 | 1861 | |
3a0a6902 PB |
1862 | if (io_req_needs_clean(req)) |
1863 | io_clean_op(req); | |
e1d767f0 PB |
1864 | if (!(flags & REQ_F_FIXED_FILE)) |
1865 | io_put_file(req->file); | |
269bbe5f BM |
1866 | if (req->fixed_rsrc_refs) |
1867 | percpu_ref_put(req->fixed_rsrc_refs); | |
99ebe4ef | 1868 | if (req->async_data) { |
094bae49 | 1869 | kfree(req->async_data); |
99ebe4ef PB |
1870 | req->async_data = NULL; |
1871 | } | |
e65ef56d JA |
1872 | } |
1873 | ||
216578e5 | 1874 | static void __io_free_req(struct io_kiocb *req) |
c6ca97b3 | 1875 | { |
51a4cc11 | 1876 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1877 | |
216578e5 | 1878 | io_dismantle_req(req); |
7c660731 | 1879 | io_put_task(req->task, 1); |
c6ca97b3 | 1880 | |
79ebeaee | 1881 | spin_lock(&ctx->completion_lock); |
bb943b82 | 1882 | list_add(&req->inflight_entry, &ctx->locked_free_list); |
c34b025f | 1883 | ctx->locked_free_nr++; |
79ebeaee | 1884 | spin_unlock(&ctx->completion_lock); |
c34b025f | 1885 | |
ecfc5177 | 1886 | percpu_ref_put(&ctx->refs); |
e65ef56d JA |
1887 | } |
1888 | ||
f2f87370 PB |
1889 | static inline void io_remove_next_linked(struct io_kiocb *req) |
1890 | { | |
1891 | struct io_kiocb *nxt = req->link; | |
1892 | ||
1893 | req->link = nxt->link; | |
1894 | nxt->link = NULL; | |
1895 | } | |
1896 | ||
33cc89a9 PB |
1897 | static bool io_kill_linked_timeout(struct io_kiocb *req) |
1898 | __must_hold(&req->ctx->completion_lock) | |
89b263f6 | 1899 | __must_hold(&req->ctx->timeout_lock) |
2665abfd | 1900 | { |
33cc89a9 | 1901 | struct io_kiocb *link = req->link; |
f2f87370 | 1902 | |
b97e736a | 1903 | if (link && link->opcode == IORING_OP_LINK_TIMEOUT) { |
c9abd7ad | 1904 | struct io_timeout_data *io = link->async_data; |
7c86ffee | 1905 | |
f2f87370 | 1906 | io_remove_next_linked(req); |
90cd7e42 | 1907 | link->timeout.head = NULL; |
fd9c7bc5 | 1908 | if (hrtimer_try_to_cancel(&io->timer) != -1) { |
d4d19c19 PB |
1909 | io_cqring_fill_event(link->ctx, link->user_data, |
1910 | -ECANCELED, 0); | |
91c2f697 | 1911 | io_put_req_deferred(link); |
d4729fbd | 1912 | return true; |
c9abd7ad PB |
1913 | } |
1914 | } | |
d4729fbd | 1915 | return false; |
7c86ffee PB |
1916 | } |
1917 | ||
d148ca4b | 1918 | static void io_fail_links(struct io_kiocb *req) |
33cc89a9 | 1919 | __must_hold(&req->ctx->completion_lock) |
9e645e11 | 1920 | { |
33cc89a9 | 1921 | struct io_kiocb *nxt, *link = req->link; |
9e645e11 | 1922 | |
f2f87370 | 1923 | req->link = NULL; |
f2f87370 PB |
1924 | while (link) { |
1925 | nxt = link->link; | |
1926 | link->link = NULL; | |
2665abfd | 1927 | |
f2f87370 | 1928 | trace_io_uring_fail_link(req, link); |
d4d19c19 | 1929 | io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0); |
91c2f697 | 1930 | io_put_req_deferred(link); |
f2f87370 | 1931 | link = nxt; |
9e645e11 | 1932 | } |
33cc89a9 | 1933 | } |
9e645e11 | 1934 | |
33cc89a9 PB |
1935 | static bool io_disarm_next(struct io_kiocb *req) |
1936 | __must_hold(&req->ctx->completion_lock) | |
1937 | { | |
1938 | bool posted = false; | |
1939 | ||
0756a869 PB |
1940 | if (req->flags & REQ_F_ARM_LTIMEOUT) { |
1941 | struct io_kiocb *link = req->link; | |
1942 | ||
906c6caa | 1943 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
0756a869 PB |
1944 | if (link && link->opcode == IORING_OP_LINK_TIMEOUT) { |
1945 | io_remove_next_linked(req); | |
1946 | io_cqring_fill_event(link->ctx, link->user_data, | |
1947 | -ECANCELED, 0); | |
1948 | io_put_req_deferred(link); | |
1949 | posted = true; | |
1950 | } | |
1951 | } else if (req->flags & REQ_F_LINK_TIMEOUT) { | |
89b263f6 JA |
1952 | struct io_ring_ctx *ctx = req->ctx; |
1953 | ||
1954 | spin_lock_irq(&ctx->timeout_lock); | |
33cc89a9 | 1955 | posted = io_kill_linked_timeout(req); |
89b263f6 JA |
1956 | spin_unlock_irq(&ctx->timeout_lock); |
1957 | } | |
93d2bcd2 | 1958 | if (unlikely((req->flags & REQ_F_FAIL) && |
e4335ed3 | 1959 | !(req->flags & REQ_F_HARDLINK))) { |
33cc89a9 PB |
1960 | posted |= (req->link != NULL); |
1961 | io_fail_links(req); | |
1962 | } | |
1963 | return posted; | |
9e645e11 JA |
1964 | } |
1965 | ||
3fa5e0f3 | 1966 | static struct io_kiocb *__io_req_find_next(struct io_kiocb *req) |
c69f8dbe | 1967 | { |
33cc89a9 | 1968 | struct io_kiocb *nxt; |
944e58bf | 1969 | |
9e645e11 JA |
1970 | /* |
1971 | * If LINK is set, we have dependent requests in this chain. If we | |
1972 | * didn't fail this request, queue the first one up, moving any other | |
1973 | * dependencies to the next request. In case of failure, fail the rest | |
1974 | * of the chain. | |
1975 | */ | |
0756a869 | 1976 | if (req->flags & IO_DISARM_MASK) { |
33cc89a9 | 1977 | struct io_ring_ctx *ctx = req->ctx; |
33cc89a9 PB |
1978 | bool posted; |
1979 | ||
79ebeaee | 1980 | spin_lock(&ctx->completion_lock); |
33cc89a9 PB |
1981 | posted = io_disarm_next(req); |
1982 | if (posted) | |
1983 | io_commit_cqring(req->ctx); | |
79ebeaee | 1984 | spin_unlock(&ctx->completion_lock); |
33cc89a9 PB |
1985 | if (posted) |
1986 | io_cqring_ev_posted(ctx); | |
f2f87370 | 1987 | } |
33cc89a9 PB |
1988 | nxt = req->link; |
1989 | req->link = NULL; | |
1990 | return nxt; | |
4d7dd462 | 1991 | } |
9e645e11 | 1992 | |
f2f87370 | 1993 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) |
3fa5e0f3 | 1994 | { |
cdbff982 | 1995 | if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK)))) |
3fa5e0f3 PB |
1996 | return NULL; |
1997 | return __io_req_find_next(req); | |
1998 | } | |
1999 | ||
2c32395d PB |
2000 | static void ctx_flush_and_put(struct io_ring_ctx *ctx) |
2001 | { | |
2002 | if (!ctx) | |
2003 | return; | |
cd0ca2e0 | 2004 | if (ctx->submit_state.compl_nr) { |
2c32395d | 2005 | mutex_lock(&ctx->uring_lock); |
2a2758f2 | 2006 | io_submit_flush_completions(ctx); |
2c32395d PB |
2007 | mutex_unlock(&ctx->uring_lock); |
2008 | } | |
2009 | percpu_ref_put(&ctx->refs); | |
2010 | } | |
2011 | ||
7cbf1722 | 2012 | static void tctx_task_work(struct callback_head *cb) |
c40f6379 | 2013 | { |
ebd0df2e | 2014 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
2015 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
2016 | task_work); | |
c40f6379 | 2017 | |
16f72070 | 2018 | while (1) { |
3f18407d PB |
2019 | struct io_wq_work_node *node; |
2020 | ||
2021 | spin_lock_irq(&tctx->task_lock); | |
c6538be9 | 2022 | node = tctx->task_list.first; |
3f18407d | 2023 | INIT_WQ_LIST(&tctx->task_list); |
6294f368 PB |
2024 | if (!node) |
2025 | tctx->task_running = false; | |
3f18407d | 2026 | spin_unlock_irq(&tctx->task_lock); |
6294f368 PB |
2027 | if (!node) |
2028 | break; | |
3f18407d | 2029 | |
6294f368 | 2030 | do { |
3f18407d PB |
2031 | struct io_wq_work_node *next = node->next; |
2032 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
2033 | io_task_work.node); | |
2034 | ||
2035 | if (req->ctx != ctx) { | |
2036 | ctx_flush_and_put(ctx); | |
2037 | ctx = req->ctx; | |
2038 | percpu_ref_get(&ctx->refs); | |
2039 | } | |
5b0a6acc | 2040 | req->io_task_work.func(req); |
3f18407d | 2041 | node = next; |
6294f368 PB |
2042 | } while (node); |
2043 | ||
7cbf1722 | 2044 | cond_resched(); |
3f18407d | 2045 | } |
ebd0df2e PB |
2046 | |
2047 | ctx_flush_and_put(ctx); | |
7cbf1722 JA |
2048 | } |
2049 | ||
e09ee510 | 2050 | static void io_req_task_work_add(struct io_kiocb *req) |
7cbf1722 | 2051 | { |
c15b79de | 2052 | struct task_struct *tsk = req->task; |
7cbf1722 | 2053 | struct io_uring_task *tctx = tsk->io_uring; |
c15b79de | 2054 | enum task_work_notify_mode notify; |
e09ee510 | 2055 | struct io_wq_work_node *node; |
0b81e80c | 2056 | unsigned long flags; |
6294f368 | 2057 | bool running; |
7cbf1722 JA |
2058 | |
2059 | WARN_ON_ONCE(!tctx); | |
2060 | ||
0b81e80c | 2061 | spin_lock_irqsave(&tctx->task_lock, flags); |
7cbf1722 | 2062 | wq_list_add_tail(&req->io_task_work.node, &tctx->task_list); |
6294f368 PB |
2063 | running = tctx->task_running; |
2064 | if (!running) | |
2065 | tctx->task_running = true; | |
0b81e80c | 2066 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 JA |
2067 | |
2068 | /* task_work already pending, we're done */ | |
6294f368 | 2069 | if (running) |
e09ee510 | 2070 | return; |
7cbf1722 | 2071 | |
c15b79de PB |
2072 | /* |
2073 | * SQPOLL kernel thread doesn't need notification, just a wakeup. For | |
2074 | * all other cases, use TWA_SIGNAL unconditionally to ensure we're | |
2075 | * processing task_work. There's no reliable way to tell if TWA_RESUME | |
2076 | * will do the job. | |
2077 | */ | |
2078 | notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL; | |
c15b79de PB |
2079 | if (!task_work_add(tsk, &tctx->task_work, notify)) { |
2080 | wake_up_process(tsk); | |
e09ee510 | 2081 | return; |
c15b79de | 2082 | } |
2215bed9 | 2083 | |
0b81e80c | 2084 | spin_lock_irqsave(&tctx->task_lock, flags); |
6294f368 | 2085 | tctx->task_running = false; |
e09ee510 PB |
2086 | node = tctx->task_list.first; |
2087 | INIT_WQ_LIST(&tctx->task_list); | |
0b81e80c | 2088 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 | 2089 | |
e09ee510 PB |
2090 | while (node) { |
2091 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
2092 | node = node->next; | |
2093 | if (llist_add(&req->io_task_work.fallback_node, | |
2094 | &req->ctx->fallback_llist)) | |
2095 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
2096 | } | |
eab30c4d PB |
2097 | } |
2098 | ||
5b0a6acc | 2099 | static void io_req_task_cancel(struct io_kiocb *req) |
c40f6379 | 2100 | { |
87ceb6a6 | 2101 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 | 2102 | |
e83acd7d | 2103 | /* ctx is guaranteed to stay alive while we hold uring_lock */ |
792bb6eb | 2104 | mutex_lock(&ctx->uring_lock); |
2593553a | 2105 | io_req_complete_failed(req, req->result); |
792bb6eb | 2106 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2107 | } |
2108 | ||
5b0a6acc | 2109 | static void io_req_task_submit(struct io_kiocb *req) |
c40f6379 JA |
2110 | { |
2111 | struct io_ring_ctx *ctx = req->ctx; | |
2112 | ||
04fc6c80 | 2113 | /* ctx stays valid until unlock, even if we drop all ours ctx->refs */ |
81b6d05c | 2114 | mutex_lock(&ctx->uring_lock); |
af066f31 | 2115 | if (likely(!(req->task->flags & PF_EXITING))) |
c5eef2b9 | 2116 | __io_queue_sqe(req); |
81b6d05c | 2117 | else |
2593553a | 2118 | io_req_complete_failed(req, -EFAULT); |
81b6d05c | 2119 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2120 | } |
2121 | ||
2c4b8eb6 | 2122 | static void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 2123 | { |
2c4b8eb6 | 2124 | req->result = ret; |
5b0a6acc | 2125 | req->io_task_work.func = io_req_task_cancel; |
e09ee510 | 2126 | io_req_task_work_add(req); |
c40f6379 JA |
2127 | } |
2128 | ||
2c4b8eb6 | 2129 | static void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 2130 | { |
5b0a6acc | 2131 | req->io_task_work.func = io_req_task_submit; |
e09ee510 | 2132 | io_req_task_work_add(req); |
a3df7698 PB |
2133 | } |
2134 | ||
773af691 JA |
2135 | static void io_req_task_queue_reissue(struct io_kiocb *req) |
2136 | { | |
2137 | req->io_task_work.func = io_queue_async_work; | |
2138 | io_req_task_work_add(req); | |
2139 | } | |
2140 | ||
f2f87370 | 2141 | static inline void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 2142 | { |
9b5f7bd9 | 2143 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
2144 | |
2145 | if (nxt) | |
906a8c3f | 2146 | io_req_task_queue(nxt); |
c69f8dbe JL |
2147 | } |
2148 | ||
c3524383 | 2149 | static void io_free_req(struct io_kiocb *req) |
7a743e22 | 2150 | { |
c3524383 PB |
2151 | io_queue_next(req); |
2152 | __io_free_req(req); | |
2153 | } | |
8766dd51 | 2154 | |
2d6500d4 | 2155 | struct req_batch { |
5af1d13e PB |
2156 | struct task_struct *task; |
2157 | int task_refs; | |
1b4c351f | 2158 | int ctx_refs; |
2d6500d4 PB |
2159 | }; |
2160 | ||
5af1d13e PB |
2161 | static inline void io_init_req_batch(struct req_batch *rb) |
2162 | { | |
5af1d13e | 2163 | rb->task_refs = 0; |
9ae72463 | 2164 | rb->ctx_refs = 0; |
5af1d13e PB |
2165 | rb->task = NULL; |
2166 | } | |
2167 | ||
2d6500d4 PB |
2168 | static void io_req_free_batch_finish(struct io_ring_ctx *ctx, |
2169 | struct req_batch *rb) | |
2170 | { | |
9ae72463 PB |
2171 | if (rb->ctx_refs) |
2172 | percpu_ref_put_many(&ctx->refs, rb->ctx_refs); | |
e9dbe221 PB |
2173 | if (rb->task == current) |
2174 | current->io_uring->cached_refs += rb->task_refs; | |
2175 | else if (rb->task) | |
2176 | io_put_task(rb->task, rb->task_refs); | |
2d6500d4 PB |
2177 | } |
2178 | ||
6ff119a6 PB |
2179 | static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req, |
2180 | struct io_submit_state *state) | |
2d6500d4 | 2181 | { |
f2f87370 | 2182 | io_queue_next(req); |
96670657 | 2183 | io_dismantle_req(req); |
2d6500d4 | 2184 | |
e3bc8e9d | 2185 | if (req->task != rb->task) { |
7c660731 PB |
2186 | if (rb->task) |
2187 | io_put_task(rb->task, rb->task_refs); | |
e3bc8e9d JA |
2188 | rb->task = req->task; |
2189 | rb->task_refs = 0; | |
5af1d13e | 2190 | } |
e3bc8e9d | 2191 | rb->task_refs++; |
9ae72463 | 2192 | rb->ctx_refs++; |
5af1d13e | 2193 | |
bd759045 | 2194 | if (state->free_reqs != ARRAY_SIZE(state->reqs)) |
6ff119a6 | 2195 | state->reqs[state->free_reqs++] = req; |
bd759045 | 2196 | else |
cd0ca2e0 | 2197 | list_add(&req->inflight_entry, &state->free_list); |
7a743e22 PB |
2198 | } |
2199 | ||
2a2758f2 | 2200 | static void io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 2201 | __must_hold(&ctx->uring_lock) |
905c172f | 2202 | { |
cd0ca2e0 PB |
2203 | struct io_submit_state *state = &ctx->submit_state; |
2204 | int i, nr = state->compl_nr; | |
905c172f PB |
2205 | struct req_batch rb; |
2206 | ||
79ebeaee | 2207 | spin_lock(&ctx->completion_lock); |
905c172f | 2208 | for (i = 0; i < nr; i++) { |
cd0ca2e0 | 2209 | struct io_kiocb *req = state->compl_reqs[i]; |
5182ed2e | 2210 | |
d4d19c19 PB |
2211 | __io_cqring_fill_event(ctx, req->user_data, req->result, |
2212 | req->compl.cflags); | |
905c172f PB |
2213 | } |
2214 | io_commit_cqring(ctx); | |
79ebeaee | 2215 | spin_unlock(&ctx->completion_lock); |
905c172f | 2216 | io_cqring_ev_posted(ctx); |
5182ed2e PB |
2217 | |
2218 | io_init_req_batch(&rb); | |
905c172f | 2219 | for (i = 0; i < nr; i++) { |
cd0ca2e0 | 2220 | struct io_kiocb *req = state->compl_reqs[i]; |
905c172f | 2221 | |
91c2f697 | 2222 | if (req_ref_put_and_test(req)) |
6ff119a6 | 2223 | io_req_free_batch(&rb, req, &ctx->submit_state); |
905c172f PB |
2224 | } |
2225 | ||
2226 | io_req_free_batch_finish(ctx, &rb); | |
cd0ca2e0 | 2227 | state->compl_nr = 0; |
7a743e22 PB |
2228 | } |
2229 | ||
ba816ad6 JA |
2230 | /* |
2231 | * Drop reference to request, return next in chain (if there is one) if this | |
2232 | * was the last reference to this request. | |
2233 | */ | |
0d85035a | 2234 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 2235 | { |
9b5f7bd9 PB |
2236 | struct io_kiocb *nxt = NULL; |
2237 | ||
de9b4cca | 2238 | if (req_ref_put_and_test(req)) { |
9b5f7bd9 | 2239 | nxt = io_req_find_next(req); |
4d7dd462 | 2240 | __io_free_req(req); |
2a44f467 | 2241 | } |
9b5f7bd9 | 2242 | return nxt; |
2b188cc1 JA |
2243 | } |
2244 | ||
0d85035a | 2245 | static inline void io_put_req(struct io_kiocb *req) |
e65ef56d | 2246 | { |
de9b4cca | 2247 | if (req_ref_put_and_test(req)) |
e65ef56d | 2248 | io_free_req(req); |
2b188cc1 JA |
2249 | } |
2250 | ||
91c2f697 | 2251 | static inline void io_put_req_deferred(struct io_kiocb *req) |
216578e5 | 2252 | { |
91c2f697 | 2253 | if (req_ref_put_and_test(req)) { |
543af3a1 PB |
2254 | req->io_task_work.func = io_free_req; |
2255 | io_req_task_work_add(req); | |
2256 | } | |
216578e5 PB |
2257 | } |
2258 | ||
6c503150 | 2259 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
2260 | { |
2261 | /* See comment at the top of this file */ | |
2262 | smp_rmb(); | |
e23de15f | 2263 | return __io_cqring_events(ctx); |
a3a0e43f JA |
2264 | } |
2265 | ||
fb5ccc98 PB |
2266 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
2267 | { | |
2268 | struct io_rings *rings = ctx->rings; | |
2269 | ||
2270 | /* make sure SQ entry isn't read before tail */ | |
2271 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
2272 | } | |
2273 | ||
8ff069bf | 2274 | static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf) |
e94f141b | 2275 | { |
8ff069bf | 2276 | unsigned int cflags; |
e94f141b | 2277 | |
bcda7baa JA |
2278 | cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT; |
2279 | cflags |= IORING_CQE_F_BUFFER; | |
0e1b6fe3 | 2280 | req->flags &= ~REQ_F_BUFFER_SELECTED; |
bcda7baa JA |
2281 | kfree(kbuf); |
2282 | return cflags; | |
e94f141b JA |
2283 | } |
2284 | ||
8ff069bf | 2285 | static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req) |
bcda7baa | 2286 | { |
4d954c25 | 2287 | struct io_buffer *kbuf; |
bcda7baa | 2288 | |
ae421d93 PB |
2289 | if (likely(!(req->flags & REQ_F_BUFFER_SELECTED))) |
2290 | return 0; | |
4d954c25 | 2291 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; |
8ff069bf PB |
2292 | return io_put_kbuf(req, kbuf); |
2293 | } | |
2294 | ||
4c6e277c JA |
2295 | static inline bool io_run_task_work(void) |
2296 | { | |
ef98eb04 | 2297 | if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) { |
4c6e277c | 2298 | __set_current_state(TASK_RUNNING); |
ef98eb04 | 2299 | tracehook_notify_signal(); |
4c6e277c JA |
2300 | return true; |
2301 | } | |
2302 | ||
2303 | return false; | |
bcda7baa JA |
2304 | } |
2305 | ||
def596e9 JA |
2306 | /* |
2307 | * Find and free completed poll iocbs | |
2308 | */ | |
2309 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
a8576af9 | 2310 | struct list_head *done) |
def596e9 | 2311 | { |
8237e045 | 2312 | struct req_batch rb; |
def596e9 | 2313 | struct io_kiocb *req; |
bbde017a XW |
2314 | |
2315 | /* order with ->result store in io_complete_rw_iopoll() */ | |
2316 | smp_rmb(); | |
def596e9 | 2317 | |
5af1d13e | 2318 | io_init_req_batch(&rb); |
def596e9 | 2319 | while (!list_empty(done)) { |
d21ffe7e | 2320 | req = list_first_entry(done, struct io_kiocb, inflight_entry); |
f161340d PB |
2321 | list_del(&req->inflight_entry); |
2322 | ||
a8576af9 | 2323 | if (READ_ONCE(req->result) == -EAGAIN && |
8c130827 | 2324 | !(req->flags & REQ_F_DONT_REISSUE)) { |
bbde017a | 2325 | req->iopoll_completed = 0; |
773af691 | 2326 | io_req_task_queue_reissue(req); |
8c130827 | 2327 | continue; |
bbde017a | 2328 | } |
def596e9 | 2329 | |
ae421d93 PB |
2330 | __io_cqring_fill_event(ctx, req->user_data, req->result, |
2331 | io_put_rw_kbuf(req)); | |
def596e9 JA |
2332 | (*nr_events)++; |
2333 | ||
de9b4cca | 2334 | if (req_ref_put_and_test(req)) |
6ff119a6 | 2335 | io_req_free_batch(&rb, req, &ctx->submit_state); |
def596e9 | 2336 | } |
def596e9 | 2337 | |
09bb8394 | 2338 | io_commit_cqring(ctx); |
80c18e4a | 2339 | io_cqring_ev_posted_iopoll(ctx); |
2d6500d4 | 2340 | io_req_free_batch_finish(ctx, &rb); |
581f9810 BM |
2341 | } |
2342 | ||
def596e9 | 2343 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, |
a8576af9 | 2344 | long min) |
def596e9 JA |
2345 | { |
2346 | struct io_kiocb *req, *tmp; | |
2347 | LIST_HEAD(done); | |
2348 | bool spin; | |
def596e9 JA |
2349 | |
2350 | /* | |
2351 | * Only spin for completions if we don't have multiple devices hanging | |
2352 | * off our complete list, and we're under the requested amount. | |
2353 | */ | |
915b3dde | 2354 | spin = !ctx->poll_multi_queue && *nr_events < min; |
def596e9 | 2355 | |
d21ffe7e | 2356 | list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) { |
9adbd45d | 2357 | struct kiocb *kiocb = &req->rw.kiocb; |
a2416e1e | 2358 | int ret; |
def596e9 JA |
2359 | |
2360 | /* | |
581f9810 BM |
2361 | * Move completed and retryable entries to our local lists. |
2362 | * If we find a request that requires polling, break out | |
2363 | * and complete those lists first, if we have entries there. | |
def596e9 | 2364 | */ |
65a6543d | 2365 | if (READ_ONCE(req->iopoll_completed)) { |
d21ffe7e | 2366 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2367 | continue; |
2368 | } | |
2369 | if (!list_empty(&done)) | |
2370 | break; | |
2371 | ||
2372 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
a2416e1e PB |
2373 | if (unlikely(ret < 0)) |
2374 | return ret; | |
2375 | else if (ret) | |
2376 | spin = false; | |
def596e9 | 2377 | |
3aadc23e PB |
2378 | /* iopoll may have completed current req */ |
2379 | if (READ_ONCE(req->iopoll_completed)) | |
d21ffe7e | 2380 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2381 | } |
2382 | ||
2383 | if (!list_empty(&done)) | |
a8576af9 | 2384 | io_iopoll_complete(ctx, nr_events, &done); |
def596e9 | 2385 | |
a2416e1e | 2386 | return 0; |
def596e9 JA |
2387 | } |
2388 | ||
def596e9 JA |
2389 | /* |
2390 | * We can't just wait for polled events to come to us, we have to actively | |
2391 | * find and complete them. | |
2392 | */ | |
b2edc0a7 | 2393 | static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
2394 | { |
2395 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
2396 | return; | |
2397 | ||
2398 | mutex_lock(&ctx->uring_lock); | |
540e32a0 | 2399 | while (!list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2400 | unsigned int nr_events = 0; |
2401 | ||
a8576af9 | 2402 | io_do_iopoll(ctx, &nr_events, 0); |
08f5439f | 2403 | |
b2edc0a7 PB |
2404 | /* let it sleep and repeat later if can't complete a request */ |
2405 | if (nr_events == 0) | |
2406 | break; | |
08f5439f JA |
2407 | /* |
2408 | * Ensure we allow local-to-the-cpu processing to take place, | |
2409 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 2410 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 2411 | */ |
3fcee5a6 PB |
2412 | if (need_resched()) { |
2413 | mutex_unlock(&ctx->uring_lock); | |
2414 | cond_resched(); | |
2415 | mutex_lock(&ctx->uring_lock); | |
2416 | } | |
def596e9 JA |
2417 | } |
2418 | mutex_unlock(&ctx->uring_lock); | |
2419 | } | |
2420 | ||
7668b92a | 2421 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 2422 | { |
7668b92a | 2423 | unsigned int nr_events = 0; |
e9979b36 | 2424 | int ret = 0; |
500f9fba | 2425 | |
c7849be9 XW |
2426 | /* |
2427 | * We disallow the app entering submit/complete with polling, but we | |
2428 | * still need to lock the ring to prevent racing with polled issue | |
2429 | * that got punted to a workqueue. | |
2430 | */ | |
2431 | mutex_lock(&ctx->uring_lock); | |
f39c8a5b PB |
2432 | /* |
2433 | * Don't enter poll loop if we already have events pending. | |
2434 | * If we do, we can potentially be spinning for commands that | |
2435 | * already triggered a CQE (eg in error). | |
2436 | */ | |
5ed7a37d | 2437 | if (test_bit(0, &ctx->check_cq_overflow)) |
f39c8a5b PB |
2438 | __io_cqring_overflow_flush(ctx, false); |
2439 | if (io_cqring_events(ctx)) | |
2440 | goto out; | |
def596e9 | 2441 | do { |
500f9fba JA |
2442 | /* |
2443 | * If a submit got punted to a workqueue, we can have the | |
2444 | * application entering polling for a command before it gets | |
2445 | * issued. That app will hold the uring_lock for the duration | |
2446 | * of the poll right here, so we need to take a breather every | |
2447 | * now and then to ensure that the issue has a chance to add | |
2448 | * the poll to the issued list. Otherwise we can spin here | |
2449 | * forever, while the workqueue is stuck trying to acquire the | |
2450 | * very same mutex. | |
2451 | */ | |
e9979b36 | 2452 | if (list_empty(&ctx->iopoll_list)) { |
8f487ef2 PB |
2453 | u32 tail = ctx->cached_cq_tail; |
2454 | ||
500f9fba | 2455 | mutex_unlock(&ctx->uring_lock); |
4c6e277c | 2456 | io_run_task_work(); |
500f9fba | 2457 | mutex_lock(&ctx->uring_lock); |
def596e9 | 2458 | |
8f487ef2 PB |
2459 | /* some requests don't go through iopoll_list */ |
2460 | if (tail != ctx->cached_cq_tail || | |
2461 | list_empty(&ctx->iopoll_list)) | |
e9979b36 | 2462 | break; |
500f9fba | 2463 | } |
a8576af9 | 2464 | ret = io_do_iopoll(ctx, &nr_events, min); |
f39c8a5b PB |
2465 | } while (!ret && nr_events < min && !need_resched()); |
2466 | out: | |
500f9fba | 2467 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
2468 | return ret; |
2469 | } | |
2470 | ||
491381ce | 2471 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 2472 | { |
491381ce JA |
2473 | /* |
2474 | * Tell lockdep we inherited freeze protection from submission | |
2475 | * thread. | |
2476 | */ | |
2477 | if (req->flags & REQ_F_ISREG) { | |
1c98679d | 2478 | struct super_block *sb = file_inode(req->file)->i_sb; |
2b188cc1 | 2479 | |
1c98679d PB |
2480 | __sb_writers_acquired(sb, SB_FREEZE_WRITE); |
2481 | sb_end_write(sb); | |
2b188cc1 JA |
2482 | } |
2483 | } | |
2484 | ||
b63534c4 | 2485 | #ifdef CONFIG_BLOCK |
dc2a6e9a | 2486 | static bool io_resubmit_prep(struct io_kiocb *req) |
b63534c4 | 2487 | { |
ab454438 | 2488 | struct io_async_rw *rw = req->async_data; |
b63534c4 | 2489 | |
ab454438 PB |
2490 | if (!rw) |
2491 | return !io_req_prep_async(req); | |
2492 | /* may have left rw->iter inconsistent on -EIOCBQUEUED */ | |
2493 | iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter)); | |
2494 | return true; | |
b63534c4 | 2495 | } |
b63534c4 | 2496 | |
3e6a0d3c | 2497 | static bool io_rw_should_reissue(struct io_kiocb *req) |
b63534c4 | 2498 | { |
355afaeb | 2499 | umode_t mode = file_inode(req->file)->i_mode; |
3e6a0d3c | 2500 | struct io_ring_ctx *ctx = req->ctx; |
b63534c4 | 2501 | |
355afaeb JA |
2502 | if (!S_ISBLK(mode) && !S_ISREG(mode)) |
2503 | return false; | |
3e6a0d3c JA |
2504 | if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() && |
2505 | !(ctx->flags & IORING_SETUP_IOPOLL))) | |
b63534c4 | 2506 | return false; |
7c977a58 JA |
2507 | /* |
2508 | * If ref is dying, we might be running poll reap from the exit work. | |
2509 | * Don't attempt to reissue from that path, just let it fail with | |
2510 | * -EAGAIN. | |
2511 | */ | |
3e6a0d3c JA |
2512 | if (percpu_ref_is_dying(&ctx->refs)) |
2513 | return false; | |
ef046888 JA |
2514 | /* |
2515 | * Play it safe and assume not safe to re-import and reissue if we're | |
2516 | * not in the original thread group (or in task context). | |
2517 | */ | |
2518 | if (!same_thread_group(req->task, current) || !in_task()) | |
2519 | return false; | |
3e6a0d3c JA |
2520 | return true; |
2521 | } | |
e82ad485 | 2522 | #else |
a1ff1e3f | 2523 | static bool io_resubmit_prep(struct io_kiocb *req) |
e82ad485 JA |
2524 | { |
2525 | return false; | |
2526 | } | |
e82ad485 | 2527 | static bool io_rw_should_reissue(struct io_kiocb *req) |
3e6a0d3c | 2528 | { |
b63534c4 JA |
2529 | return false; |
2530 | } | |
3e6a0d3c | 2531 | #endif |
b63534c4 | 2532 | |
8ef12efe | 2533 | static bool __io_complete_rw_common(struct io_kiocb *req, long res) |
a1d7c393 | 2534 | { |
b65c128f PB |
2535 | if (req->rw.kiocb.ki_flags & IOCB_WRITE) |
2536 | kiocb_end_write(req); | |
9532b99b PB |
2537 | if (res != req->result) { |
2538 | if ((res == -EAGAIN || res == -EOPNOTSUPP) && | |
2539 | io_rw_should_reissue(req)) { | |
2540 | req->flags |= REQ_F_REISSUE; | |
8ef12efe | 2541 | return true; |
9532b99b | 2542 | } |
93d2bcd2 | 2543 | req_set_fail(req); |
8ef12efe | 2544 | req->result = res; |
9532b99b | 2545 | } |
8ef12efe JA |
2546 | return false; |
2547 | } | |
2548 | ||
2549 | static void io_req_task_complete(struct io_kiocb *req) | |
2550 | { | |
ae421d93 | 2551 | __io_req_complete(req, 0, req->result, io_put_rw_kbuf(req)); |
8ef12efe JA |
2552 | } |
2553 | ||
2554 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, | |
2555 | unsigned int issue_flags) | |
2556 | { | |
2557 | if (__io_complete_rw_common(req, res)) | |
2558 | return; | |
2559 | io_req_task_complete(req); | |
ba816ad6 JA |
2560 | } |
2561 | ||
2562 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
2563 | { | |
9adbd45d | 2564 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 | 2565 | |
8ef12efe JA |
2566 | if (__io_complete_rw_common(req, res)) |
2567 | return; | |
2568 | req->result = res; | |
2569 | req->io_task_work.func = io_req_task_complete; | |
2570 | io_req_task_work_add(req); | |
2b188cc1 JA |
2571 | } |
2572 | ||
def596e9 JA |
2573 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
2574 | { | |
9adbd45d | 2575 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 2576 | |
491381ce JA |
2577 | if (kiocb->ki_flags & IOCB_WRITE) |
2578 | kiocb_end_write(req); | |
9532b99b | 2579 | if (unlikely(res != req->result)) { |
a1ff1e3f JA |
2580 | if (!(res == -EAGAIN && io_rw_should_reissue(req) && |
2581 | io_resubmit_prep(req))) { | |
93d2bcd2 | 2582 | req_set_fail(req); |
9532b99b PB |
2583 | req->flags |= REQ_F_DONT_REISSUE; |
2584 | } | |
8c130827 | 2585 | } |
bbde017a XW |
2586 | |
2587 | WRITE_ONCE(req->result, res); | |
b9b0e0d3 | 2588 | /* order with io_iopoll_complete() checking ->result */ |
cd664b0e PB |
2589 | smp_wmb(); |
2590 | WRITE_ONCE(req->iopoll_completed, 1); | |
def596e9 JA |
2591 | } |
2592 | ||
2593 | /* | |
2594 | * After the iocb has been issued, it's safe to be found on the poll list. | |
2595 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 2596 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
2597 | * accessing the kiocb cookie. |
2598 | */ | |
cb3d8972 | 2599 | static void io_iopoll_req_issued(struct io_kiocb *req) |
def596e9 JA |
2600 | { |
2601 | struct io_ring_ctx *ctx = req->ctx; | |
cb3d8972 PB |
2602 | const bool in_async = io_wq_current_is_worker(); |
2603 | ||
2604 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
2605 | if (unlikely(in_async)) | |
2606 | mutex_lock(&ctx->uring_lock); | |
def596e9 JA |
2607 | |
2608 | /* | |
2609 | * Track whether we have multiple files in our lists. This will impact | |
2610 | * how we do polling eventually, not spinning if we're on potentially | |
2611 | * different devices. | |
2612 | */ | |
540e32a0 | 2613 | if (list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
2614 | ctx->poll_multi_queue = false; |
2615 | } else if (!ctx->poll_multi_queue) { | |
def596e9 | 2616 | struct io_kiocb *list_req; |
915b3dde | 2617 | unsigned int queue_num0, queue_num1; |
def596e9 | 2618 | |
540e32a0 | 2619 | list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb, |
d21ffe7e | 2620 | inflight_entry); |
915b3dde HX |
2621 | |
2622 | if (list_req->file != req->file) { | |
2623 | ctx->poll_multi_queue = true; | |
2624 | } else { | |
2625 | queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie); | |
2626 | queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie); | |
2627 | if (queue_num0 != queue_num1) | |
2628 | ctx->poll_multi_queue = true; | |
2629 | } | |
def596e9 JA |
2630 | } |
2631 | ||
2632 | /* | |
2633 | * For fast devices, IO may have already completed. If it has, add | |
2634 | * it to the front so we find it first. | |
2635 | */ | |
65a6543d | 2636 | if (READ_ONCE(req->iopoll_completed)) |
d21ffe7e | 2637 | list_add(&req->inflight_entry, &ctx->iopoll_list); |
def596e9 | 2638 | else |
d21ffe7e | 2639 | list_add_tail(&req->inflight_entry, &ctx->iopoll_list); |
bdcd3eab | 2640 | |
cb3d8972 PB |
2641 | if (unlikely(in_async)) { |
2642 | /* | |
2643 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
2644 | * in sq thread task context or in io worker task context. If | |
2645 | * current task context is sq thread, we don't need to check | |
2646 | * whether should wake up sq thread. | |
2647 | */ | |
2648 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
2649 | wq_has_sleeper(&ctx->sq_data->wait)) | |
2650 | wake_up(&ctx->sq_data->wait); | |
2651 | ||
2652 | mutex_unlock(&ctx->uring_lock); | |
2653 | } | |
def596e9 JA |
2654 | } |
2655 | ||
4503b767 JA |
2656 | static bool io_bdev_nowait(struct block_device *bdev) |
2657 | { | |
9ba0d0c8 | 2658 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
2659 | } |
2660 | ||
2b188cc1 JA |
2661 | /* |
2662 | * If we tracked the file through the SCM inflight mechanism, we could support | |
2663 | * any file. For now, just ensure that anything potentially problematic is done | |
2664 | * inline. | |
2665 | */ | |
b191e2df | 2666 | static bool __io_file_supports_nowait(struct file *file, int rw) |
2b188cc1 JA |
2667 | { |
2668 | umode_t mode = file_inode(file)->i_mode; | |
2669 | ||
4503b767 | 2670 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
2671 | if (IS_ENABLED(CONFIG_BLOCK) && |
2672 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
2673 | return true; |
2674 | return false; | |
2675 | } | |
976517f1 | 2676 | if (S_ISSOCK(mode)) |
2b188cc1 | 2677 | return true; |
4503b767 | 2678 | if (S_ISREG(mode)) { |
4e7b5671 CH |
2679 | if (IS_ENABLED(CONFIG_BLOCK) && |
2680 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
4503b767 JA |
2681 | file->f_op != &io_uring_fops) |
2682 | return true; | |
2683 | return false; | |
2684 | } | |
2b188cc1 | 2685 | |
c5b85625 JA |
2686 | /* any ->read/write should understand O_NONBLOCK */ |
2687 | if (file->f_flags & O_NONBLOCK) | |
2688 | return true; | |
2689 | ||
af197f50 JA |
2690 | if (!(file->f_mode & FMODE_NOWAIT)) |
2691 | return false; | |
2692 | ||
2693 | if (rw == READ) | |
2694 | return file->f_op->read_iter != NULL; | |
2695 | ||
2696 | return file->f_op->write_iter != NULL; | |
2b188cc1 JA |
2697 | } |
2698 | ||
b191e2df | 2699 | static bool io_file_supports_nowait(struct io_kiocb *req, int rw) |
7b29f92d | 2700 | { |
b191e2df | 2701 | if (rw == READ && (req->flags & REQ_F_NOWAIT_READ)) |
7b29f92d | 2702 | return true; |
b191e2df | 2703 | else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE)) |
7b29f92d JA |
2704 | return true; |
2705 | ||
b191e2df | 2706 | return __io_file_supports_nowait(req->file, rw); |
7b29f92d JA |
2707 | } |
2708 | ||
a88fc400 | 2709 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 2710 | { |
def596e9 | 2711 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 2712 | struct kiocb *kiocb = &req->rw.kiocb; |
75c668cd | 2713 | struct file *file = req->file; |
09bb8394 JA |
2714 | unsigned ioprio; |
2715 | int ret; | |
2b188cc1 | 2716 | |
c97d8a0f | 2717 | if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode)) |
491381ce JA |
2718 | req->flags |= REQ_F_ISREG; |
2719 | ||
2b188cc1 | 2720 | kiocb->ki_pos = READ_ONCE(sqe->off); |
75c668cd | 2721 | if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) { |
ba04291e | 2722 | req->flags |= REQ_F_CUR_POS; |
75c668cd | 2723 | kiocb->ki_pos = file->f_pos; |
ba04291e | 2724 | } |
2b188cc1 | 2725 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
2726 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
2727 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
2728 | if (unlikely(ret)) | |
2729 | return ret; | |
2b188cc1 | 2730 | |
75c668cd PB |
2731 | /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */ |
2732 | if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK)) | |
2733 | req->flags |= REQ_F_NOWAIT; | |
2734 | ||
2b188cc1 JA |
2735 | ioprio = READ_ONCE(sqe->ioprio); |
2736 | if (ioprio) { | |
2737 | ret = ioprio_check_cap(ioprio); | |
2738 | if (ret) | |
09bb8394 | 2739 | return ret; |
2b188cc1 JA |
2740 | |
2741 | kiocb->ki_ioprio = ioprio; | |
2742 | } else | |
2743 | kiocb->ki_ioprio = get_current_ioprio(); | |
2744 | ||
def596e9 | 2745 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2746 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2747 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2748 | return -EOPNOTSUPP; |
2b188cc1 | 2749 | |
def596e9 JA |
2750 | kiocb->ki_flags |= IOCB_HIPRI; |
2751 | kiocb->ki_complete = io_complete_rw_iopoll; | |
65a6543d | 2752 | req->iopoll_completed = 0; |
def596e9 | 2753 | } else { |
09bb8394 JA |
2754 | if (kiocb->ki_flags & IOCB_HIPRI) |
2755 | return -EINVAL; | |
def596e9 JA |
2756 | kiocb->ki_complete = io_complete_rw; |
2757 | } | |
9adbd45d | 2758 | |
eae071c9 PB |
2759 | if (req->opcode == IORING_OP_READ_FIXED || |
2760 | req->opcode == IORING_OP_WRITE_FIXED) { | |
2761 | req->imu = NULL; | |
2762 | io_req_set_rsrc_node(req); | |
2763 | } | |
2764 | ||
3529d8c2 JA |
2765 | req->rw.addr = READ_ONCE(sqe->addr); |
2766 | req->rw.len = READ_ONCE(sqe->len); | |
4f4eeba8 | 2767 | req->buf_index = READ_ONCE(sqe->buf_index); |
2b188cc1 | 2768 | return 0; |
2b188cc1 JA |
2769 | } |
2770 | ||
2771 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2772 | { | |
2773 | switch (ret) { | |
2774 | case -EIOCBQUEUED: | |
2775 | break; | |
2776 | case -ERESTARTSYS: | |
2777 | case -ERESTARTNOINTR: | |
2778 | case -ERESTARTNOHAND: | |
2779 | case -ERESTART_RESTARTBLOCK: | |
2780 | /* | |
2781 | * We can't just restart the syscall, since previously | |
2782 | * submitted sqes may already be in progress. Just fail this | |
2783 | * IO with EINTR. | |
2784 | */ | |
2785 | ret = -EINTR; | |
df561f66 | 2786 | fallthrough; |
2b188cc1 JA |
2787 | default: |
2788 | kiocb->ki_complete(kiocb, ret, 0); | |
2789 | } | |
2790 | } | |
2791 | ||
a1d7c393 | 2792 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, |
889fca73 | 2793 | unsigned int issue_flags) |
ba816ad6 | 2794 | { |
ba04291e | 2795 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
e8c2bc1f | 2796 | struct io_async_rw *io = req->async_data; |
97284637 | 2797 | bool check_reissue = kiocb->ki_complete == io_complete_rw; |
ba04291e | 2798 | |
227c0c96 | 2799 | /* add previously done IO, if any */ |
e8c2bc1f | 2800 | if (io && io->bytes_done > 0) { |
227c0c96 | 2801 | if (ret < 0) |
e8c2bc1f | 2802 | ret = io->bytes_done; |
227c0c96 | 2803 | else |
e8c2bc1f | 2804 | ret += io->bytes_done; |
227c0c96 JA |
2805 | } |
2806 | ||
ba04291e JA |
2807 | if (req->flags & REQ_F_CUR_POS) |
2808 | req->file->f_pos = kiocb->ki_pos; | |
e149bd74 | 2809 | if (ret >= 0 && check_reissue) |
889fca73 | 2810 | __io_complete_rw(req, ret, 0, issue_flags); |
ba816ad6 JA |
2811 | else |
2812 | io_rw_done(kiocb, ret); | |
97284637 | 2813 | |
fe7e3257 | 2814 | if (check_reissue && (req->flags & REQ_F_REISSUE)) { |
97284637 | 2815 | req->flags &= ~REQ_F_REISSUE; |
a7be7c23 | 2816 | if (io_resubmit_prep(req)) { |
773af691 | 2817 | io_req_task_queue_reissue(req); |
8c130827 | 2818 | } else { |
93d2bcd2 | 2819 | req_set_fail(req); |
ae421d93 PB |
2820 | __io_req_complete(req, issue_flags, ret, |
2821 | io_put_rw_kbuf(req)); | |
97284637 PB |
2822 | } |
2823 | } | |
ba816ad6 JA |
2824 | } |
2825 | ||
eae071c9 PB |
2826 | static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter, |
2827 | struct io_mapped_ubuf *imu) | |
edafccee | 2828 | { |
9adbd45d | 2829 | size_t len = req->rw.len; |
75769e3f | 2830 | u64 buf_end, buf_addr = req->rw.addr; |
edafccee | 2831 | size_t offset; |
edafccee | 2832 | |
75769e3f | 2833 | if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) |
edafccee JA |
2834 | return -EFAULT; |
2835 | /* not inside the mapped region */ | |
4751f53d | 2836 | if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) |
edafccee JA |
2837 | return -EFAULT; |
2838 | ||
2839 | /* | |
2840 | * May not be a start of buffer, set size appropriately | |
2841 | * and advance us to the beginning. | |
2842 | */ | |
2843 | offset = buf_addr - imu->ubuf; | |
2844 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
2845 | |
2846 | if (offset) { | |
2847 | /* | |
2848 | * Don't use iov_iter_advance() here, as it's really slow for | |
2849 | * using the latter parts of a big fixed buffer - it iterates | |
2850 | * over each segment manually. We can cheat a bit here, because | |
2851 | * we know that: | |
2852 | * | |
2853 | * 1) it's a BVEC iter, we set it up | |
2854 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
2855 | * first and last bvec | |
2856 | * | |
2857 | * So just find our index, and adjust the iterator afterwards. | |
2858 | * If the offset is within the first bvec (or the whole first | |
2859 | * bvec, just use iov_iter_advance(). This makes it easier | |
2860 | * since we can just skip the first segment, which may not | |
2861 | * be PAGE_SIZE aligned. | |
2862 | */ | |
2863 | const struct bio_vec *bvec = imu->bvec; | |
2864 | ||
2865 | if (offset <= bvec->bv_len) { | |
2866 | iov_iter_advance(iter, offset); | |
2867 | } else { | |
2868 | unsigned long seg_skip; | |
2869 | ||
2870 | /* skip first vec */ | |
2871 | offset -= bvec->bv_len; | |
2872 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
2873 | ||
2874 | iter->bvec = bvec + seg_skip; | |
2875 | iter->nr_segs -= seg_skip; | |
99c79f66 | 2876 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 2877 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
2878 | } |
2879 | } | |
2880 | ||
847595de | 2881 | return 0; |
edafccee JA |
2882 | } |
2883 | ||
eae071c9 PB |
2884 | static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter) |
2885 | { | |
2886 | struct io_ring_ctx *ctx = req->ctx; | |
2887 | struct io_mapped_ubuf *imu = req->imu; | |
2888 | u16 index, buf_index = req->buf_index; | |
2889 | ||
2890 | if (likely(!imu)) { | |
2891 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
2892 | return -EFAULT; | |
2893 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
2894 | imu = READ_ONCE(ctx->user_bufs[index]); | |
2895 | req->imu = imu; | |
2896 | } | |
2897 | return __io_import_fixed(req, rw, iter, imu); | |
2898 | } | |
2899 | ||
bcda7baa JA |
2900 | static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock) |
2901 | { | |
2902 | if (needs_lock) | |
2903 | mutex_unlock(&ctx->uring_lock); | |
2904 | } | |
2905 | ||
2906 | static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock) | |
2907 | { | |
2908 | /* | |
2909 | * "Normal" inline submissions always hold the uring_lock, since we | |
2910 | * grab it from the system call. Same is true for the SQPOLL offload. | |
2911 | * The only exception is when we've detached the request and issue it | |
2912 | * from an async worker thread, grab the lock for that case. | |
2913 | */ | |
2914 | if (needs_lock) | |
2915 | mutex_lock(&ctx->uring_lock); | |
2916 | } | |
2917 | ||
2918 | static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len, | |
2919 | int bgid, struct io_buffer *kbuf, | |
2920 | bool needs_lock) | |
2921 | { | |
2922 | struct io_buffer *head; | |
2923 | ||
2924 | if (req->flags & REQ_F_BUFFER_SELECTED) | |
2925 | return kbuf; | |
2926 | ||
2927 | io_ring_submit_lock(req->ctx, needs_lock); | |
2928 | ||
2929 | lockdep_assert_held(&req->ctx->uring_lock); | |
2930 | ||
9e15c3a0 | 2931 | head = xa_load(&req->ctx->io_buffers, bgid); |
bcda7baa JA |
2932 | if (head) { |
2933 | if (!list_empty(&head->list)) { | |
2934 | kbuf = list_last_entry(&head->list, struct io_buffer, | |
2935 | list); | |
2936 | list_del(&kbuf->list); | |
2937 | } else { | |
2938 | kbuf = head; | |
9e15c3a0 | 2939 | xa_erase(&req->ctx->io_buffers, bgid); |
bcda7baa JA |
2940 | } |
2941 | if (*len > kbuf->len) | |
2942 | *len = kbuf->len; | |
2943 | } else { | |
2944 | kbuf = ERR_PTR(-ENOBUFS); | |
2945 | } | |
2946 | ||
2947 | io_ring_submit_unlock(req->ctx, needs_lock); | |
2948 | ||
2949 | return kbuf; | |
2950 | } | |
2951 | ||
4d954c25 JA |
2952 | static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len, |
2953 | bool needs_lock) | |
2954 | { | |
2955 | struct io_buffer *kbuf; | |
4f4eeba8 | 2956 | u16 bgid; |
4d954c25 JA |
2957 | |
2958 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
4f4eeba8 | 2959 | bgid = req->buf_index; |
4d954c25 JA |
2960 | kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock); |
2961 | if (IS_ERR(kbuf)) | |
2962 | return kbuf; | |
2963 | req->rw.addr = (u64) (unsigned long) kbuf; | |
2964 | req->flags |= REQ_F_BUFFER_SELECTED; | |
2965 | return u64_to_user_ptr(kbuf->addr); | |
2966 | } | |
2967 | ||
2968 | #ifdef CONFIG_COMPAT | |
2969 | static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov, | |
2970 | bool needs_lock) | |
2971 | { | |
2972 | struct compat_iovec __user *uiov; | |
2973 | compat_ssize_t clen; | |
2974 | void __user *buf; | |
2975 | ssize_t len; | |
2976 | ||
2977 | uiov = u64_to_user_ptr(req->rw.addr); | |
2978 | if (!access_ok(uiov, sizeof(*uiov))) | |
2979 | return -EFAULT; | |
2980 | if (__get_user(clen, &uiov->iov_len)) | |
2981 | return -EFAULT; | |
2982 | if (clen < 0) | |
2983 | return -EINVAL; | |
2984 | ||
2985 | len = clen; | |
2986 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
2987 | if (IS_ERR(buf)) | |
2988 | return PTR_ERR(buf); | |
2989 | iov[0].iov_base = buf; | |
2990 | iov[0].iov_len = (compat_size_t) len; | |
2991 | return 0; | |
2992 | } | |
2993 | #endif | |
2994 | ||
2995 | static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
2996 | bool needs_lock) | |
2997 | { | |
2998 | struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr); | |
2999 | void __user *buf; | |
3000 | ssize_t len; | |
3001 | ||
3002 | if (copy_from_user(iov, uiov, sizeof(*uiov))) | |
3003 | return -EFAULT; | |
3004 | ||
3005 | len = iov[0].iov_len; | |
3006 | if (len < 0) | |
3007 | return -EINVAL; | |
3008 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
3009 | if (IS_ERR(buf)) | |
3010 | return PTR_ERR(buf); | |
3011 | iov[0].iov_base = buf; | |
3012 | iov[0].iov_len = len; | |
3013 | return 0; | |
3014 | } | |
3015 | ||
3016 | static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
3017 | bool needs_lock) | |
3018 | { | |
dddb3e26 JA |
3019 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
3020 | struct io_buffer *kbuf; | |
3021 | ||
3022 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
3023 | iov[0].iov_base = u64_to_user_ptr(kbuf->addr); | |
3024 | iov[0].iov_len = kbuf->len; | |
4d954c25 | 3025 | return 0; |
dddb3e26 | 3026 | } |
dd201662 | 3027 | if (req->rw.len != 1) |
4d954c25 JA |
3028 | return -EINVAL; |
3029 | ||
3030 | #ifdef CONFIG_COMPAT | |
3031 | if (req->ctx->compat) | |
3032 | return io_compat_import(req, iov, needs_lock); | |
3033 | #endif | |
3034 | ||
3035 | return __io_iov_buffer_select(req, iov, needs_lock); | |
3036 | } | |
3037 | ||
847595de PB |
3038 | static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec, |
3039 | struct iov_iter *iter, bool needs_lock) | |
2b188cc1 | 3040 | { |
9adbd45d JA |
3041 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
3042 | size_t sqe_len = req->rw.len; | |
847595de | 3043 | u8 opcode = req->opcode; |
4d954c25 | 3044 | ssize_t ret; |
edafccee | 3045 | |
7d009165 | 3046 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 3047 | *iovec = NULL; |
9adbd45d | 3048 | return io_import_fixed(req, rw, iter); |
edafccee | 3049 | } |
2b188cc1 | 3050 | |
bcda7baa | 3051 | /* buffer index only valid with fixed read/write, or buffer select */ |
4f4eeba8 | 3052 | if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)) |
9adbd45d JA |
3053 | return -EINVAL; |
3054 | ||
3a6820f2 | 3055 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
bcda7baa | 3056 | if (req->flags & REQ_F_BUFFER_SELECT) { |
4d954c25 | 3057 | buf = io_rw_buffer_select(req, &sqe_len, needs_lock); |
867a23ea | 3058 | if (IS_ERR(buf)) |
4d954c25 | 3059 | return PTR_ERR(buf); |
3f9d6441 | 3060 | req->rw.len = sqe_len; |
bcda7baa JA |
3061 | } |
3062 | ||
3a6820f2 JA |
3063 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); |
3064 | *iovec = NULL; | |
10fc72e4 | 3065 | return ret; |
3a6820f2 JA |
3066 | } |
3067 | ||
4d954c25 JA |
3068 | if (req->flags & REQ_F_BUFFER_SELECT) { |
3069 | ret = io_iov_buffer_select(req, *iovec, needs_lock); | |
847595de PB |
3070 | if (!ret) |
3071 | iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len); | |
4d954c25 JA |
3072 | *iovec = NULL; |
3073 | return ret; | |
3074 | } | |
3075 | ||
89cd35c5 CH |
3076 | return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter, |
3077 | req->ctx->compat); | |
2b188cc1 JA |
3078 | } |
3079 | ||
0fef9483 JA |
3080 | static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) |
3081 | { | |
5b09e37e | 3082 | return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos; |
0fef9483 JA |
3083 | } |
3084 | ||
31b51510 | 3085 | /* |
32960613 JA |
3086 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
3087 | * by looping over ->read() or ->write() manually. | |
31b51510 | 3088 | */ |
4017eb91 | 3089 | static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter) |
32960613 | 3090 | { |
4017eb91 JA |
3091 | struct kiocb *kiocb = &req->rw.kiocb; |
3092 | struct file *file = req->file; | |
32960613 JA |
3093 | ssize_t ret = 0; |
3094 | ||
3095 | /* | |
3096 | * Don't support polled IO through this interface, and we can't | |
3097 | * support non-blocking either. For the latter, this just causes | |
3098 | * the kiocb to be handled from an async context. | |
3099 | */ | |
3100 | if (kiocb->ki_flags & IOCB_HIPRI) | |
3101 | return -EOPNOTSUPP; | |
3102 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
3103 | return -EAGAIN; | |
3104 | ||
3105 | while (iov_iter_count(iter)) { | |
311ae9e1 | 3106 | struct iovec iovec; |
32960613 JA |
3107 | ssize_t nr; |
3108 | ||
311ae9e1 PB |
3109 | if (!iov_iter_is_bvec(iter)) { |
3110 | iovec = iov_iter_iovec(iter); | |
3111 | } else { | |
4017eb91 JA |
3112 | iovec.iov_base = u64_to_user_ptr(req->rw.addr); |
3113 | iovec.iov_len = req->rw.len; | |
311ae9e1 PB |
3114 | } |
3115 | ||
32960613 JA |
3116 | if (rw == READ) { |
3117 | nr = file->f_op->read(file, iovec.iov_base, | |
0fef9483 | 3118 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3119 | } else { |
3120 | nr = file->f_op->write(file, iovec.iov_base, | |
0fef9483 | 3121 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3122 | } |
3123 | ||
3124 | if (nr < 0) { | |
3125 | if (!ret) | |
3126 | ret = nr; | |
3127 | break; | |
3128 | } | |
3129 | ret += nr; | |
3130 | if (nr != iovec.iov_len) | |
3131 | break; | |
4017eb91 JA |
3132 | req->rw.len -= nr; |
3133 | req->rw.addr += nr; | |
32960613 JA |
3134 | iov_iter_advance(iter, nr); |
3135 | } | |
3136 | ||
3137 | return ret; | |
3138 | } | |
3139 | ||
ff6165b2 JA |
3140 | static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec, |
3141 | const struct iovec *fast_iov, struct iov_iter *iter) | |
f67676d1 | 3142 | { |
e8c2bc1f | 3143 | struct io_async_rw *rw = req->async_data; |
b64e3444 | 3144 | |
ff6165b2 | 3145 | memcpy(&rw->iter, iter, sizeof(*iter)); |
afb87658 | 3146 | rw->free_iovec = iovec; |
227c0c96 | 3147 | rw->bytes_done = 0; |
ff6165b2 | 3148 | /* can only be fixed buffers, no need to do anything */ |
9c3a205c | 3149 | if (iov_iter_is_bvec(iter)) |
ff6165b2 | 3150 | return; |
b64e3444 | 3151 | if (!iovec) { |
ff6165b2 JA |
3152 | unsigned iov_off = 0; |
3153 | ||
3154 | rw->iter.iov = rw->fast_iov; | |
3155 | if (iter->iov != fast_iov) { | |
3156 | iov_off = iter->iov - fast_iov; | |
3157 | rw->iter.iov += iov_off; | |
3158 | } | |
3159 | if (rw->fast_iov != fast_iov) | |
3160 | memcpy(rw->fast_iov + iov_off, fast_iov + iov_off, | |
45097dae | 3161 | sizeof(struct iovec) * iter->nr_segs); |
99bc4c38 PB |
3162 | } else { |
3163 | req->flags |= REQ_F_NEED_CLEANUP; | |
f67676d1 JA |
3164 | } |
3165 | } | |
3166 | ||
6cb78689 | 3167 | static inline int io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 3168 | { |
e8c2bc1f JA |
3169 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
3170 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
3171 | return req->async_data == NULL; | |
3d9932a8 XW |
3172 | } |
3173 | ||
ff6165b2 JA |
3174 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
3175 | const struct iovec *fast_iov, | |
227c0c96 | 3176 | struct iov_iter *iter, bool force) |
b7bb4f7d | 3177 | { |
26f0505a | 3178 | if (!force && !io_op_defs[req->opcode].needs_async_setup) |
74566df3 | 3179 | return 0; |
e8c2bc1f | 3180 | if (!req->async_data) { |
6cb78689 | 3181 | if (io_alloc_async_data(req)) { |
6bf985dc | 3182 | kfree(iovec); |
5d204bcf | 3183 | return -ENOMEM; |
6bf985dc | 3184 | } |
b7bb4f7d | 3185 | |
ff6165b2 | 3186 | io_req_map_rw(req, iovec, fast_iov, iter); |
5d204bcf | 3187 | } |
b7bb4f7d | 3188 | return 0; |
f67676d1 JA |
3189 | } |
3190 | ||
73debe68 | 3191 | static inline int io_rw_prep_async(struct io_kiocb *req, int rw) |
c3e330a4 | 3192 | { |
e8c2bc1f | 3193 | struct io_async_rw *iorw = req->async_data; |
f4bff104 | 3194 | struct iovec *iov = iorw->fast_iov; |
847595de | 3195 | int ret; |
c3e330a4 | 3196 | |
2846c481 | 3197 | ret = io_import_iovec(rw, req, &iov, &iorw->iter, false); |
c3e330a4 PB |
3198 | if (unlikely(ret < 0)) |
3199 | return ret; | |
3200 | ||
ab0b196c PB |
3201 | iorw->bytes_done = 0; |
3202 | iorw->free_iovec = iov; | |
3203 | if (iov) | |
3204 | req->flags |= REQ_F_NEED_CLEANUP; | |
c3e330a4 PB |
3205 | return 0; |
3206 | } | |
3207 | ||
73debe68 | 3208 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 3209 | { |
3529d8c2 JA |
3210 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
3211 | return -EBADF; | |
93642ef8 | 3212 | return io_prep_rw(req, sqe); |
f67676d1 JA |
3213 | } |
3214 | ||
c1dd91d1 JA |
3215 | /* |
3216 | * This is our waitqueue callback handler, registered through lock_page_async() | |
3217 | * when we initially tried to do the IO with the iocb armed our waitqueue. | |
3218 | * This gets called when the page is unlocked, and we generally expect that to | |
3219 | * happen when the page IO is completed and the page is now uptodate. This will | |
3220 | * queue a task_work based retry of the operation, attempting to copy the data | |
3221 | * again. If the latter fails because the page was NOT uptodate, then we will | |
3222 | * do a thread based blocking retry of the operation. That's the unexpected | |
3223 | * slow path. | |
3224 | */ | |
bcf5a063 JA |
3225 | static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, |
3226 | int sync, void *arg) | |
3227 | { | |
3228 | struct wait_page_queue *wpq; | |
3229 | struct io_kiocb *req = wait->private; | |
bcf5a063 | 3230 | struct wait_page_key *key = arg; |
bcf5a063 JA |
3231 | |
3232 | wpq = container_of(wait, struct wait_page_queue, wait); | |
3233 | ||
cdc8fcb4 LT |
3234 | if (!wake_page_match(wpq, key)) |
3235 | return 0; | |
3236 | ||
c8d317aa | 3237 | req->rw.kiocb.ki_flags &= ~IOCB_WAITQ; |
bcf5a063 | 3238 | list_del_init(&wait->entry); |
921b9054 | 3239 | io_req_task_queue(req); |
bcf5a063 JA |
3240 | return 1; |
3241 | } | |
3242 | ||
c1dd91d1 JA |
3243 | /* |
3244 | * This controls whether a given IO request should be armed for async page | |
3245 | * based retry. If we return false here, the request is handed to the async | |
3246 | * worker threads for retry. If we're doing buffered reads on a regular file, | |
3247 | * we prepare a private wait_page_queue entry and retry the operation. This | |
3248 | * will either succeed because the page is now uptodate and unlocked, or it | |
3249 | * will register a callback when the page is unlocked at IO completion. Through | |
3250 | * that callback, io_uring uses task_work to setup a retry of the operation. | |
3251 | * That retry will attempt the buffered read again. The retry will generally | |
3252 | * succeed, or in rare cases where it fails, we then fall back to using the | |
3253 | * async worker threads for a blocking retry. | |
3254 | */ | |
227c0c96 | 3255 | static bool io_rw_should_retry(struct io_kiocb *req) |
f67676d1 | 3256 | { |
e8c2bc1f JA |
3257 | struct io_async_rw *rw = req->async_data; |
3258 | struct wait_page_queue *wait = &rw->wpq; | |
bcf5a063 | 3259 | struct kiocb *kiocb = &req->rw.kiocb; |
f67676d1 | 3260 | |
bcf5a063 JA |
3261 | /* never retry for NOWAIT, we just complete with -EAGAIN */ |
3262 | if (req->flags & REQ_F_NOWAIT) | |
3263 | return false; | |
f67676d1 | 3264 | |
227c0c96 | 3265 | /* Only for buffered IO */ |
3b2a4439 | 3266 | if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) |
bcf5a063 | 3267 | return false; |
3b2a4439 | 3268 | |
bcf5a063 JA |
3269 | /* |
3270 | * just use poll if we can, and don't attempt if the fs doesn't | |
3271 | * support callback based unlocks | |
3272 | */ | |
3273 | if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC)) | |
3274 | return false; | |
f67676d1 | 3275 | |
3b2a4439 JA |
3276 | wait->wait.func = io_async_buf_func; |
3277 | wait->wait.private = req; | |
3278 | wait->wait.flags = 0; | |
3279 | INIT_LIST_HEAD(&wait->wait.entry); | |
3280 | kiocb->ki_flags |= IOCB_WAITQ; | |
c8d317aa | 3281 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3b2a4439 | 3282 | kiocb->ki_waitq = wait; |
3b2a4439 | 3283 | return true; |
bcf5a063 JA |
3284 | } |
3285 | ||
aeab9506 | 3286 | static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter) |
bcf5a063 JA |
3287 | { |
3288 | if (req->file->f_op->read_iter) | |
3289 | return call_read_iter(req->file, &req->rw.kiocb, iter); | |
2dd2111d | 3290 | else if (req->file->f_op->read) |
4017eb91 | 3291 | return loop_rw_iter(READ, req, iter); |
2dd2111d GH |
3292 | else |
3293 | return -EINVAL; | |
f67676d1 JA |
3294 | } |
3295 | ||
889fca73 | 3296 | static int io_read(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 JA |
3297 | { |
3298 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3299 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3300 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3301 | struct io_async_rw *rw = req->async_data; |
227c0c96 | 3302 | ssize_t io_size, ret, ret2; |
45d189c6 | 3303 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
ff6165b2 | 3304 | |
2846c481 | 3305 | if (rw) { |
e8c2bc1f | 3306 | iter = &rw->iter; |
2846c481 PB |
3307 | iovec = NULL; |
3308 | } else { | |
3309 | ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock); | |
3310 | if (ret < 0) | |
3311 | return ret; | |
3312 | } | |
632546c4 | 3313 | io_size = iov_iter_count(iter); |
fa15bafb | 3314 | req->result = io_size; |
2b188cc1 | 3315 | |
fd6c2e4c JA |
3316 | /* Ensure we clear previously set non-block flag */ |
3317 | if (!force_nonblock) | |
29de5f6a | 3318 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
a88fc400 PB |
3319 | else |
3320 | kiocb->ki_flags |= IOCB_NOWAIT; | |
3321 | ||
24c74678 | 3322 | /* If the file doesn't support async, just async punt */ |
b191e2df | 3323 | if (force_nonblock && !io_file_supports_nowait(req, READ)) { |
6713e7a6 | 3324 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true); |
6bf985dc | 3325 | return ret ?: -EAGAIN; |
6713e7a6 | 3326 | } |
9e645e11 | 3327 | |
632546c4 | 3328 | ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size); |
5ea5dd45 PB |
3329 | if (unlikely(ret)) { |
3330 | kfree(iovec); | |
3331 | return ret; | |
3332 | } | |
2b188cc1 | 3333 | |
227c0c96 | 3334 | ret = io_iter_do_read(req, iter); |
32960613 | 3335 | |
230d50d4 | 3336 | if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) { |
6ad7f233 | 3337 | req->flags &= ~REQ_F_REISSUE; |
eefdf30f JA |
3338 | /* IOPOLL retry should happen for io-wq threads */ |
3339 | if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
f91daf56 | 3340 | goto done; |
75c668cd PB |
3341 | /* no retry on NONBLOCK nor RWF_NOWAIT */ |
3342 | if (req->flags & REQ_F_NOWAIT) | |
355afaeb | 3343 | goto done; |
84216315 | 3344 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3345 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
f38c7e3a | 3346 | ret = 0; |
230d50d4 JA |
3347 | } else if (ret == -EIOCBQUEUED) { |
3348 | goto out_free; | |
7335e3bf | 3349 | } else if (ret <= 0 || ret == io_size || !force_nonblock || |
75c668cd | 3350 | (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) { |
7335e3bf | 3351 | /* read all, failed, already did sync or don't want to retry */ |
00d23d51 | 3352 | goto done; |
227c0c96 JA |
3353 | } |
3354 | ||
227c0c96 | 3355 | ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true); |
6bf985dc PB |
3356 | if (ret2) |
3357 | return ret2; | |
3358 | ||
fe1cdd55 | 3359 | iovec = NULL; |
e8c2bc1f | 3360 | rw = req->async_data; |
227c0c96 | 3361 | /* now use our persistent iterator, if we aren't already */ |
e8c2bc1f | 3362 | iter = &rw->iter; |
227c0c96 | 3363 | |
b23df91b PB |
3364 | do { |
3365 | io_size -= ret; | |
3366 | rw->bytes_done += ret; | |
3367 | /* if we can retry, do so with the callbacks armed */ | |
3368 | if (!io_rw_should_retry(req)) { | |
3369 | kiocb->ki_flags &= ~IOCB_WAITQ; | |
3370 | return -EAGAIN; | |
3371 | } | |
3372 | ||
3373 | /* | |
3374 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If | |
3375 | * we get -EIOCBQUEUED, then we'll get a notification when the | |
3376 | * desired page gets unlocked. We can also get a partial read | |
3377 | * here, and if we do, then just retry at the new offset. | |
3378 | */ | |
3379 | ret = io_iter_do_read(req, iter); | |
3380 | if (ret == -EIOCBQUEUED) | |
3381 | return 0; | |
227c0c96 | 3382 | /* we got some bytes, but not all. retry. */ |
b5b0ecb7 | 3383 | kiocb->ki_flags &= ~IOCB_WAITQ; |
b23df91b | 3384 | } while (ret > 0 && ret < io_size); |
227c0c96 | 3385 | done: |
889fca73 | 3386 | kiocb_done(kiocb, ret, issue_flags); |
fe1cdd55 PB |
3387 | out_free: |
3388 | /* it's faster to check here then delegate to kfree */ | |
3389 | if (iovec) | |
3390 | kfree(iovec); | |
5ea5dd45 | 3391 | return 0; |
2b188cc1 JA |
3392 | } |
3393 | ||
73debe68 | 3394 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 3395 | { |
3529d8c2 JA |
3396 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
3397 | return -EBADF; | |
93642ef8 | 3398 | return io_prep_rw(req, sqe); |
f67676d1 JA |
3399 | } |
3400 | ||
889fca73 | 3401 | static int io_write(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 JA |
3402 | { |
3403 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3404 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3405 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3406 | struct io_async_rw *rw = req->async_data; |
fa15bafb | 3407 | ssize_t ret, ret2, io_size; |
45d189c6 | 3408 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
2b188cc1 | 3409 | |
2846c481 | 3410 | if (rw) { |
e8c2bc1f | 3411 | iter = &rw->iter; |
2846c481 PB |
3412 | iovec = NULL; |
3413 | } else { | |
3414 | ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock); | |
3415 | if (ret < 0) | |
3416 | return ret; | |
3417 | } | |
632546c4 | 3418 | io_size = iov_iter_count(iter); |
fa15bafb | 3419 | req->result = io_size; |
2b188cc1 | 3420 | |
fd6c2e4c JA |
3421 | /* Ensure we clear previously set non-block flag */ |
3422 | if (!force_nonblock) | |
a88fc400 PB |
3423 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3424 | else | |
3425 | kiocb->ki_flags |= IOCB_NOWAIT; | |
fd6c2e4c | 3426 | |
24c74678 | 3427 | /* If the file doesn't support async, just async punt */ |
b191e2df | 3428 | if (force_nonblock && !io_file_supports_nowait(req, WRITE)) |
f67676d1 | 3429 | goto copy_iov; |
31b51510 | 3430 | |
10d59345 JA |
3431 | /* file path doesn't support NOWAIT for non-direct_IO */ |
3432 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
3433 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 3434 | goto copy_iov; |
31b51510 | 3435 | |
632546c4 | 3436 | ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size); |
fa15bafb PB |
3437 | if (unlikely(ret)) |
3438 | goto out_free; | |
4ed734b0 | 3439 | |
fa15bafb PB |
3440 | /* |
3441 | * Open-code file_start_write here to grab freeze protection, | |
3442 | * which will be released by another thread in | |
3443 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
3444 | * released so that it doesn't complain about the held lock when | |
3445 | * we return to userspace. | |
3446 | */ | |
3447 | if (req->flags & REQ_F_ISREG) { | |
8a3c84b6 | 3448 | sb_start_write(file_inode(req->file)->i_sb); |
fa15bafb PB |
3449 | __sb_writers_release(file_inode(req->file)->i_sb, |
3450 | SB_FREEZE_WRITE); | |
3451 | } | |
3452 | kiocb->ki_flags |= IOCB_WRITE; | |
4ed734b0 | 3453 | |
fa15bafb | 3454 | if (req->file->f_op->write_iter) |
ff6165b2 | 3455 | ret2 = call_write_iter(req->file, kiocb, iter); |
2dd2111d | 3456 | else if (req->file->f_op->write) |
4017eb91 | 3457 | ret2 = loop_rw_iter(WRITE, req, iter); |
2dd2111d GH |
3458 | else |
3459 | ret2 = -EINVAL; | |
4ed734b0 | 3460 | |
6ad7f233 PB |
3461 | if (req->flags & REQ_F_REISSUE) { |
3462 | req->flags &= ~REQ_F_REISSUE; | |
230d50d4 | 3463 | ret2 = -EAGAIN; |
6ad7f233 | 3464 | } |
230d50d4 | 3465 | |
fa15bafb PB |
3466 | /* |
3467 | * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just | |
3468 | * retry them without IOCB_NOWAIT. | |
3469 | */ | |
3470 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
3471 | ret2 = -EAGAIN; | |
75c668cd PB |
3472 | /* no retry on NONBLOCK nor RWF_NOWAIT */ |
3473 | if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT)) | |
355afaeb | 3474 | goto done; |
fa15bafb | 3475 | if (!force_nonblock || ret2 != -EAGAIN) { |
eefdf30f JA |
3476 | /* IOPOLL retry should happen for io-wq threads */ |
3477 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN) | |
3478 | goto copy_iov; | |
355afaeb | 3479 | done: |
889fca73 | 3480 | kiocb_done(kiocb, ret2, issue_flags); |
fa15bafb | 3481 | } else { |
f67676d1 | 3482 | copy_iov: |
84216315 | 3483 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3484 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
227c0c96 | 3485 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false); |
6bf985dc | 3486 | return ret ?: -EAGAIN; |
2b188cc1 | 3487 | } |
31b51510 | 3488 | out_free: |
f261c168 | 3489 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3490 | if (iovec) |
6f2cc166 | 3491 | kfree(iovec); |
2b188cc1 JA |
3492 | return ret; |
3493 | } | |
3494 | ||
80a261fd JA |
3495 | static int io_renameat_prep(struct io_kiocb *req, |
3496 | const struct io_uring_sqe *sqe) | |
3497 | { | |
3498 | struct io_rename *ren = &req->rename; | |
3499 | const char __user *oldf, *newf; | |
3500 | ||
ed7eb259 JA |
3501 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3502 | return -EINVAL; | |
3503 | if (sqe->ioprio || sqe->buf_index) | |
3504 | return -EINVAL; | |
80a261fd JA |
3505 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
3506 | return -EBADF; | |
3507 | ||
3508 | ren->old_dfd = READ_ONCE(sqe->fd); | |
3509 | oldf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3510 | newf = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
3511 | ren->new_dfd = READ_ONCE(sqe->len); | |
3512 | ren->flags = READ_ONCE(sqe->rename_flags); | |
3513 | ||
3514 | ren->oldpath = getname(oldf); | |
3515 | if (IS_ERR(ren->oldpath)) | |
3516 | return PTR_ERR(ren->oldpath); | |
3517 | ||
3518 | ren->newpath = getname(newf); | |
3519 | if (IS_ERR(ren->newpath)) { | |
3520 | putname(ren->oldpath); | |
3521 | return PTR_ERR(ren->newpath); | |
3522 | } | |
3523 | ||
3524 | req->flags |= REQ_F_NEED_CLEANUP; | |
3525 | return 0; | |
3526 | } | |
3527 | ||
45d189c6 | 3528 | static int io_renameat(struct io_kiocb *req, unsigned int issue_flags) |
80a261fd JA |
3529 | { |
3530 | struct io_rename *ren = &req->rename; | |
3531 | int ret; | |
3532 | ||
45d189c6 | 3533 | if (issue_flags & IO_URING_F_NONBLOCK) |
80a261fd JA |
3534 | return -EAGAIN; |
3535 | ||
3536 | ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd, | |
3537 | ren->newpath, ren->flags); | |
3538 | ||
3539 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3540 | if (ret < 0) | |
93d2bcd2 | 3541 | req_set_fail(req); |
80a261fd JA |
3542 | io_req_complete(req, ret); |
3543 | return 0; | |
3544 | } | |
3545 | ||
14a1143b JA |
3546 | static int io_unlinkat_prep(struct io_kiocb *req, |
3547 | const struct io_uring_sqe *sqe) | |
3548 | { | |
3549 | struct io_unlink *un = &req->unlink; | |
3550 | const char __user *fname; | |
3551 | ||
22634bc5 JA |
3552 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3553 | return -EINVAL; | |
3554 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
3555 | return -EINVAL; | |
14a1143b JA |
3556 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
3557 | return -EBADF; | |
3558 | ||
3559 | un->dfd = READ_ONCE(sqe->fd); | |
3560 | ||
3561 | un->flags = READ_ONCE(sqe->unlink_flags); | |
3562 | if (un->flags & ~AT_REMOVEDIR) | |
3563 | return -EINVAL; | |
3564 | ||
3565 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3566 | un->filename = getname(fname); | |
3567 | if (IS_ERR(un->filename)) | |
3568 | return PTR_ERR(un->filename); | |
3569 | ||
3570 | req->flags |= REQ_F_NEED_CLEANUP; | |
3571 | return 0; | |
3572 | } | |
3573 | ||
45d189c6 | 3574 | static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags) |
14a1143b JA |
3575 | { |
3576 | struct io_unlink *un = &req->unlink; | |
3577 | int ret; | |
3578 | ||
45d189c6 | 3579 | if (issue_flags & IO_URING_F_NONBLOCK) |
14a1143b JA |
3580 | return -EAGAIN; |
3581 | ||
3582 | if (un->flags & AT_REMOVEDIR) | |
3583 | ret = do_rmdir(un->dfd, un->filename); | |
3584 | else | |
3585 | ret = do_unlinkat(un->dfd, un->filename); | |
3586 | ||
3587 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3588 | if (ret < 0) | |
93d2bcd2 | 3589 | req_set_fail(req); |
14a1143b JA |
3590 | io_req_complete(req, ret); |
3591 | return 0; | |
3592 | } | |
3593 | ||
36f4fa68 JA |
3594 | static int io_shutdown_prep(struct io_kiocb *req, |
3595 | const struct io_uring_sqe *sqe) | |
3596 | { | |
3597 | #if defined(CONFIG_NET) | |
3598 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3599 | return -EINVAL; | |
3600 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags || | |
3601 | sqe->buf_index) | |
3602 | return -EINVAL; | |
3603 | ||
3604 | req->shutdown.how = READ_ONCE(sqe->len); | |
3605 | return 0; | |
3606 | #else | |
3607 | return -EOPNOTSUPP; | |
3608 | #endif | |
3609 | } | |
3610 | ||
45d189c6 | 3611 | static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags) |
36f4fa68 JA |
3612 | { |
3613 | #if defined(CONFIG_NET) | |
3614 | struct socket *sock; | |
3615 | int ret; | |
3616 | ||
45d189c6 | 3617 | if (issue_flags & IO_URING_F_NONBLOCK) |
36f4fa68 JA |
3618 | return -EAGAIN; |
3619 | ||
48aba79b | 3620 | sock = sock_from_file(req->file); |
36f4fa68 | 3621 | if (unlikely(!sock)) |
48aba79b | 3622 | return -ENOTSOCK; |
36f4fa68 JA |
3623 | |
3624 | ret = __sys_shutdown_sock(sock, req->shutdown.how); | |
a146468d | 3625 | if (ret < 0) |
93d2bcd2 | 3626 | req_set_fail(req); |
36f4fa68 JA |
3627 | io_req_complete(req, ret); |
3628 | return 0; | |
3629 | #else | |
3630 | return -EOPNOTSUPP; | |
3631 | #endif | |
3632 | } | |
3633 | ||
f2a8d5c7 PB |
3634 | static int __io_splice_prep(struct io_kiocb *req, |
3635 | const struct io_uring_sqe *sqe) | |
7d67af2c | 3636 | { |
fe7e3257 | 3637 | struct io_splice *sp = &req->splice; |
7d67af2c | 3638 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; |
7d67af2c | 3639 | |
3232dd02 PB |
3640 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3641 | return -EINVAL; | |
7d67af2c PB |
3642 | |
3643 | sp->file_in = NULL; | |
7d67af2c PB |
3644 | sp->len = READ_ONCE(sqe->len); |
3645 | sp->flags = READ_ONCE(sqe->splice_flags); | |
3646 | ||
3647 | if (unlikely(sp->flags & ~valid_flags)) | |
3648 | return -EINVAL; | |
3649 | ||
62906e89 | 3650 | sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in), |
8371adf5 PB |
3651 | (sp->flags & SPLICE_F_FD_IN_FIXED)); |
3652 | if (!sp->file_in) | |
3653 | return -EBADF; | |
7d67af2c | 3654 | req->flags |= REQ_F_NEED_CLEANUP; |
7d67af2c PB |
3655 | return 0; |
3656 | } | |
3657 | ||
f2a8d5c7 PB |
3658 | static int io_tee_prep(struct io_kiocb *req, |
3659 | const struct io_uring_sqe *sqe) | |
3660 | { | |
3661 | if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off)) | |
3662 | return -EINVAL; | |
3663 | return __io_splice_prep(req, sqe); | |
3664 | } | |
3665 | ||
45d189c6 | 3666 | static int io_tee(struct io_kiocb *req, unsigned int issue_flags) |
f2a8d5c7 PB |
3667 | { |
3668 | struct io_splice *sp = &req->splice; | |
3669 | struct file *in = sp->file_in; | |
3670 | struct file *out = sp->file_out; | |
3671 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3672 | long ret = 0; | |
3673 | ||
45d189c6 | 3674 | if (issue_flags & IO_URING_F_NONBLOCK) |
f2a8d5c7 PB |
3675 | return -EAGAIN; |
3676 | if (sp->len) | |
3677 | ret = do_tee(in, out, sp->len, flags); | |
3678 | ||
e1d767f0 PB |
3679 | if (!(sp->flags & SPLICE_F_FD_IN_FIXED)) |
3680 | io_put_file(in); | |
f2a8d5c7 PB |
3681 | req->flags &= ~REQ_F_NEED_CLEANUP; |
3682 | ||
f2a8d5c7 | 3683 | if (ret != sp->len) |
93d2bcd2 | 3684 | req_set_fail(req); |
e1e16097 | 3685 | io_req_complete(req, ret); |
f2a8d5c7 PB |
3686 | return 0; |
3687 | } | |
3688 | ||
3689 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
3690 | { | |
fe7e3257 | 3691 | struct io_splice *sp = &req->splice; |
f2a8d5c7 PB |
3692 | |
3693 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
3694 | sp->off_out = READ_ONCE(sqe->off); | |
3695 | return __io_splice_prep(req, sqe); | |
3696 | } | |
3697 | ||
45d189c6 | 3698 | static int io_splice(struct io_kiocb *req, unsigned int issue_flags) |
7d67af2c PB |
3699 | { |
3700 | struct io_splice *sp = &req->splice; | |
3701 | struct file *in = sp->file_in; | |
3702 | struct file *out = sp->file_out; | |
3703 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3704 | loff_t *poff_in, *poff_out; | |
c9687426 | 3705 | long ret = 0; |
7d67af2c | 3706 | |
45d189c6 | 3707 | if (issue_flags & IO_URING_F_NONBLOCK) |
2fb3e822 | 3708 | return -EAGAIN; |
7d67af2c PB |
3709 | |
3710 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
3711 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
c9687426 | 3712 | |
948a7749 | 3713 | if (sp->len) |
c9687426 | 3714 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); |
7d67af2c | 3715 | |
e1d767f0 PB |
3716 | if (!(sp->flags & SPLICE_F_FD_IN_FIXED)) |
3717 | io_put_file(in); | |
7d67af2c PB |
3718 | req->flags &= ~REQ_F_NEED_CLEANUP; |
3719 | ||
7d67af2c | 3720 | if (ret != sp->len) |
93d2bcd2 | 3721 | req_set_fail(req); |
e1e16097 | 3722 | io_req_complete(req, ret); |
7d67af2c PB |
3723 | return 0; |
3724 | } | |
3725 | ||
2b188cc1 JA |
3726 | /* |
3727 | * IORING_OP_NOP just posts a completion event, nothing else. | |
3728 | */ | |
889fca73 | 3729 | static int io_nop(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 JA |
3730 | { |
3731 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 3732 | |
def596e9 JA |
3733 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
3734 | return -EINVAL; | |
3735 | ||
889fca73 | 3736 | __io_req_complete(req, issue_flags, 0, 0); |
2b188cc1 JA |
3737 | return 0; |
3738 | } | |
3739 | ||
1155c76a | 3740 | static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 3741 | { |
6b06314c | 3742 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 3743 | |
09bb8394 JA |
3744 | if (!req->file) |
3745 | return -EBADF; | |
c992fe29 | 3746 | |
6b06314c | 3747 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 3748 | return -EINVAL; |
edafccee | 3749 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
3750 | return -EINVAL; |
3751 | ||
8ed8d3c3 JA |
3752 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
3753 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
3754 | return -EINVAL; | |
3755 | ||
3756 | req->sync.off = READ_ONCE(sqe->off); | |
3757 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
3758 | return 0; |
3759 | } | |
3760 | ||
45d189c6 | 3761 | static int io_fsync(struct io_kiocb *req, unsigned int issue_flags) |
8ed8d3c3 | 3762 | { |
8ed8d3c3 | 3763 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
3764 | int ret; |
3765 | ||
ac45abc0 | 3766 | /* fsync always requires a blocking context */ |
45d189c6 | 3767 | if (issue_flags & IO_URING_F_NONBLOCK) |
ac45abc0 PB |
3768 | return -EAGAIN; |
3769 | ||
9adbd45d | 3770 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
3771 | end > 0 ? end : LLONG_MAX, |
3772 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
3773 | if (ret < 0) | |
93d2bcd2 | 3774 | req_set_fail(req); |
e1e16097 | 3775 | io_req_complete(req, ret); |
c992fe29 CH |
3776 | return 0; |
3777 | } | |
3778 | ||
d63d1b5e JA |
3779 | static int io_fallocate_prep(struct io_kiocb *req, |
3780 | const struct io_uring_sqe *sqe) | |
3781 | { | |
3782 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
3783 | return -EINVAL; | |
3232dd02 PB |
3784 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3785 | return -EINVAL; | |
d63d1b5e JA |
3786 | |
3787 | req->sync.off = READ_ONCE(sqe->off); | |
3788 | req->sync.len = READ_ONCE(sqe->addr); | |
3789 | req->sync.mode = READ_ONCE(sqe->len); | |
3790 | return 0; | |
3791 | } | |
3792 | ||
45d189c6 | 3793 | static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags) |
5d17b4a4 | 3794 | { |
ac45abc0 PB |
3795 | int ret; |
3796 | ||
d63d1b5e | 3797 | /* fallocate always requiring blocking context */ |
45d189c6 | 3798 | if (issue_flags & IO_URING_F_NONBLOCK) |
5d17b4a4 | 3799 | return -EAGAIN; |
ac45abc0 PB |
3800 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
3801 | req->sync.len); | |
ac45abc0 | 3802 | if (ret < 0) |
93d2bcd2 | 3803 | req_set_fail(req); |
e1e16097 | 3804 | io_req_complete(req, ret); |
5d17b4a4 JA |
3805 | return 0; |
3806 | } | |
3807 | ||
ec65fea5 | 3808 | static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 3809 | { |
f8748881 | 3810 | const char __user *fname; |
15b71abe | 3811 | int ret; |
b7bb4f7d | 3812 | |
d3fddf6d PB |
3813 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3814 | return -EINVAL; | |
ec65fea5 | 3815 | if (unlikely(sqe->ioprio || sqe->buf_index)) |
15b71abe | 3816 | return -EINVAL; |
ec65fea5 | 3817 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
cf3040ca | 3818 | return -EBADF; |
03b1230c | 3819 | |
ec65fea5 PB |
3820 | /* open.how should be already initialised */ |
3821 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) | |
08a1d26e | 3822 | req->open.how.flags |= O_LARGEFILE; |
3529d8c2 | 3823 | |
25e72d10 PB |
3824 | req->open.dfd = READ_ONCE(sqe->fd); |
3825 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
f8748881 | 3826 | req->open.filename = getname(fname); |
15b71abe JA |
3827 | if (IS_ERR(req->open.filename)) { |
3828 | ret = PTR_ERR(req->open.filename); | |
3829 | req->open.filename = NULL; | |
3830 | return ret; | |
3831 | } | |
4022e7af | 3832 | req->open.nofile = rlimit(RLIMIT_NOFILE); |
8fef80bf | 3833 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 3834 | return 0; |
03b1230c JA |
3835 | } |
3836 | ||
ec65fea5 PB |
3837 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3838 | { | |
d3fddf6d PB |
3839 | u64 mode = READ_ONCE(sqe->len); |
3840 | u64 flags = READ_ONCE(sqe->open_flags); | |
ec65fea5 | 3841 | |
ec65fea5 PB |
3842 | req->open.how = build_open_how(flags, mode); |
3843 | return __io_openat_prep(req, sqe); | |
3844 | } | |
3845 | ||
cebdb986 | 3846 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 3847 | { |
cebdb986 | 3848 | struct open_how __user *how; |
cebdb986 | 3849 | size_t len; |
0fa03c62 JA |
3850 | int ret; |
3851 | ||
cebdb986 JA |
3852 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
3853 | len = READ_ONCE(sqe->len); | |
cebdb986 JA |
3854 | if (len < OPEN_HOW_SIZE_VER0) |
3855 | return -EINVAL; | |
3529d8c2 | 3856 | |
cebdb986 JA |
3857 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
3858 | len); | |
3859 | if (ret) | |
3860 | return ret; | |
3529d8c2 | 3861 | |
ec65fea5 | 3862 | return __io_openat_prep(req, sqe); |
cebdb986 JA |
3863 | } |
3864 | ||
45d189c6 | 3865 | static int io_openat2(struct io_kiocb *req, unsigned int issue_flags) |
15b71abe JA |
3866 | { |
3867 | struct open_flags op; | |
15b71abe | 3868 | struct file *file; |
3a81fd02 JA |
3869 | bool nonblock_set; |
3870 | bool resolve_nonblock; | |
15b71abe JA |
3871 | int ret; |
3872 | ||
cebdb986 | 3873 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
3874 | if (ret) |
3875 | goto err; | |
3a81fd02 JA |
3876 | nonblock_set = op.open_flag & O_NONBLOCK; |
3877 | resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED; | |
45d189c6 | 3878 | if (issue_flags & IO_URING_F_NONBLOCK) { |
3a81fd02 JA |
3879 | /* |
3880 | * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open, | |
3881 | * it'll always -EAGAIN | |
3882 | */ | |
3883 | if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE)) | |
3884 | return -EAGAIN; | |
3885 | op.lookup_flags |= LOOKUP_CACHED; | |
3886 | op.open_flag |= O_NONBLOCK; | |
3887 | } | |
15b71abe | 3888 | |
4022e7af | 3889 | ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile); |
15b71abe JA |
3890 | if (ret < 0) |
3891 | goto err; | |
3892 | ||
3893 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
12dcb58a | 3894 | if (IS_ERR(file)) { |
944d1444 | 3895 | /* |
12dcb58a PB |
3896 | * We could hang on to this 'fd' on retrying, but seems like |
3897 | * marginal gain for something that is now known to be a slower | |
3898 | * path. So just put it, and we'll get a new one when we retry. | |
944d1444 | 3899 | */ |
3a81fd02 | 3900 | put_unused_fd(ret); |
3a81fd02 | 3901 | |
15b71abe | 3902 | ret = PTR_ERR(file); |
12dcb58a PB |
3903 | /* only retry if RESOLVE_CACHED wasn't already set by application */ |
3904 | if (ret == -EAGAIN && | |
3905 | (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK))) | |
3906 | return -EAGAIN; | |
3907 | goto err; | |
15b71abe | 3908 | } |
12dcb58a PB |
3909 | |
3910 | if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set) | |
3911 | file->f_flags &= ~O_NONBLOCK; | |
3912 | fsnotify_open(file); | |
3913 | fd_install(ret, file); | |
15b71abe JA |
3914 | err: |
3915 | putname(req->open.filename); | |
8fef80bf | 3916 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe | 3917 | if (ret < 0) |
93d2bcd2 | 3918 | req_set_fail(req); |
0bdf3398 | 3919 | __io_req_complete(req, issue_flags, ret, 0); |
15b71abe JA |
3920 | return 0; |
3921 | } | |
3922 | ||
45d189c6 | 3923 | static int io_openat(struct io_kiocb *req, unsigned int issue_flags) |
cebdb986 | 3924 | { |
e45cff58 | 3925 | return io_openat2(req, issue_flags); |
cebdb986 JA |
3926 | } |
3927 | ||
067524e9 JA |
3928 | static int io_remove_buffers_prep(struct io_kiocb *req, |
3929 | const struct io_uring_sqe *sqe) | |
3930 | { | |
3931 | struct io_provide_buf *p = &req->pbuf; | |
3932 | u64 tmp; | |
3933 | ||
3934 | if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off) | |
3935 | return -EINVAL; | |
3936 | ||
3937 | tmp = READ_ONCE(sqe->fd); | |
3938 | if (!tmp || tmp > USHRT_MAX) | |
3939 | return -EINVAL; | |
3940 | ||
3941 | memset(p, 0, sizeof(*p)); | |
3942 | p->nbufs = tmp; | |
3943 | p->bgid = READ_ONCE(sqe->buf_group); | |
3944 | return 0; | |
3945 | } | |
3946 | ||
3947 | static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf, | |
3948 | int bgid, unsigned nbufs) | |
3949 | { | |
3950 | unsigned i = 0; | |
3951 | ||
3952 | /* shouldn't happen */ | |
3953 | if (!nbufs) | |
3954 | return 0; | |
3955 | ||
3956 | /* the head kbuf is the list itself */ | |
3957 | while (!list_empty(&buf->list)) { | |
3958 | struct io_buffer *nxt; | |
3959 | ||
3960 | nxt = list_first_entry(&buf->list, struct io_buffer, list); | |
3961 | list_del(&nxt->list); | |
3962 | kfree(nxt); | |
3963 | if (++i == nbufs) | |
3964 | return i; | |
3965 | } | |
3966 | i++; | |
3967 | kfree(buf); | |
9e15c3a0 | 3968 | xa_erase(&ctx->io_buffers, bgid); |
067524e9 JA |
3969 | |
3970 | return i; | |
3971 | } | |
3972 | ||
889fca73 | 3973 | static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags) |
067524e9 JA |
3974 | { |
3975 | struct io_provide_buf *p = &req->pbuf; | |
3976 | struct io_ring_ctx *ctx = req->ctx; | |
3977 | struct io_buffer *head; | |
3978 | int ret = 0; | |
45d189c6 | 3979 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
067524e9 JA |
3980 | |
3981 | io_ring_submit_lock(ctx, !force_nonblock); | |
3982 | ||
3983 | lockdep_assert_held(&ctx->uring_lock); | |
3984 | ||
3985 | ret = -ENOENT; | |
9e15c3a0 | 3986 | head = xa_load(&ctx->io_buffers, p->bgid); |
067524e9 JA |
3987 | if (head) |
3988 | ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs); | |
067524e9 | 3989 | if (ret < 0) |
93d2bcd2 | 3990 | req_set_fail(req); |
067524e9 | 3991 | |
9fb8cb49 PB |
3992 | /* complete before unlock, IOPOLL may need the lock */ |
3993 | __io_req_complete(req, issue_flags, ret, 0); | |
3994 | io_ring_submit_unlock(ctx, !force_nonblock); | |
067524e9 JA |
3995 | return 0; |
3996 | } | |
3997 | ||
ddf0322d JA |
3998 | static int io_provide_buffers_prep(struct io_kiocb *req, |
3999 | const struct io_uring_sqe *sqe) | |
4000 | { | |
38134ada | 4001 | unsigned long size, tmp_check; |
ddf0322d JA |
4002 | struct io_provide_buf *p = &req->pbuf; |
4003 | u64 tmp; | |
4004 | ||
4005 | if (sqe->ioprio || sqe->rw_flags) | |
4006 | return -EINVAL; | |
4007 | ||
4008 | tmp = READ_ONCE(sqe->fd); | |
4009 | if (!tmp || tmp > USHRT_MAX) | |
4010 | return -E2BIG; | |
4011 | p->nbufs = tmp; | |
4012 | p->addr = READ_ONCE(sqe->addr); | |
4013 | p->len = READ_ONCE(sqe->len); | |
4014 | ||
38134ada PB |
4015 | if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs, |
4016 | &size)) | |
4017 | return -EOVERFLOW; | |
4018 | if (check_add_overflow((unsigned long)p->addr, size, &tmp_check)) | |
4019 | return -EOVERFLOW; | |
4020 | ||
d81269fe PB |
4021 | size = (unsigned long)p->len * p->nbufs; |
4022 | if (!access_ok(u64_to_user_ptr(p->addr), size)) | |
ddf0322d JA |
4023 | return -EFAULT; |
4024 | ||
4025 | p->bgid = READ_ONCE(sqe->buf_group); | |
4026 | tmp = READ_ONCE(sqe->off); | |
4027 | if (tmp > USHRT_MAX) | |
4028 | return -E2BIG; | |
4029 | p->bid = tmp; | |
4030 | return 0; | |
4031 | } | |
4032 | ||
4033 | static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head) | |
4034 | { | |
4035 | struct io_buffer *buf; | |
4036 | u64 addr = pbuf->addr; | |
4037 | int i, bid = pbuf->bid; | |
4038 | ||
4039 | for (i = 0; i < pbuf->nbufs; i++) { | |
4040 | buf = kmalloc(sizeof(*buf), GFP_KERNEL); | |
4041 | if (!buf) | |
4042 | break; | |
4043 | ||
4044 | buf->addr = addr; | |
d1f82808 | 4045 | buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT); |
ddf0322d JA |
4046 | buf->bid = bid; |
4047 | addr += pbuf->len; | |
4048 | bid++; | |
4049 | if (!*head) { | |
4050 | INIT_LIST_HEAD(&buf->list); | |
4051 | *head = buf; | |
4052 | } else { | |
4053 | list_add_tail(&buf->list, &(*head)->list); | |
4054 | } | |
4055 | } | |
4056 | ||
4057 | return i ? i : -ENOMEM; | |
4058 | } | |
4059 | ||
889fca73 | 4060 | static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags) |
ddf0322d JA |
4061 | { |
4062 | struct io_provide_buf *p = &req->pbuf; | |
4063 | struct io_ring_ctx *ctx = req->ctx; | |
4064 | struct io_buffer *head, *list; | |
4065 | int ret = 0; | |
45d189c6 | 4066 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
ddf0322d JA |
4067 | |
4068 | io_ring_submit_lock(ctx, !force_nonblock); | |
4069 | ||
4070 | lockdep_assert_held(&ctx->uring_lock); | |
4071 | ||
9e15c3a0 | 4072 | list = head = xa_load(&ctx->io_buffers, p->bgid); |
ddf0322d JA |
4073 | |
4074 | ret = io_add_buffers(p, &head); | |
9e15c3a0 JA |
4075 | if (ret >= 0 && !list) { |
4076 | ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL); | |
4077 | if (ret < 0) | |
067524e9 | 4078 | __io_remove_buffers(ctx, head, p->bgid, -1U); |
ddf0322d | 4079 | } |
ddf0322d | 4080 | if (ret < 0) |
93d2bcd2 | 4081 | req_set_fail(req); |
9fb8cb49 PB |
4082 | /* complete before unlock, IOPOLL may need the lock */ |
4083 | __io_req_complete(req, issue_flags, ret, 0); | |
4084 | io_ring_submit_unlock(ctx, !force_nonblock); | |
ddf0322d | 4085 | return 0; |
cebdb986 JA |
4086 | } |
4087 | ||
3e4827b0 JA |
4088 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
4089 | const struct io_uring_sqe *sqe) | |
4090 | { | |
4091 | #if defined(CONFIG_EPOLL) | |
4092 | if (sqe->ioprio || sqe->buf_index) | |
4093 | return -EINVAL; | |
2d74d042 | 4094 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4095 | return -EINVAL; |
3e4827b0 JA |
4096 | |
4097 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
4098 | req->epoll.op = READ_ONCE(sqe->len); | |
4099 | req->epoll.fd = READ_ONCE(sqe->off); | |
4100 | ||
4101 | if (ep_op_has_event(req->epoll.op)) { | |
4102 | struct epoll_event __user *ev; | |
4103 | ||
4104 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
4105 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
4106 | return -EFAULT; | |
4107 | } | |
4108 | ||
4109 | return 0; | |
4110 | #else | |
4111 | return -EOPNOTSUPP; | |
4112 | #endif | |
4113 | } | |
4114 | ||
889fca73 | 4115 | static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags) |
3e4827b0 JA |
4116 | { |
4117 | #if defined(CONFIG_EPOLL) | |
4118 | struct io_epoll *ie = &req->epoll; | |
4119 | int ret; | |
45d189c6 | 4120 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
3e4827b0 JA |
4121 | |
4122 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
4123 | if (force_nonblock && ret == -EAGAIN) | |
4124 | return -EAGAIN; | |
4125 | ||
4126 | if (ret < 0) | |
93d2bcd2 | 4127 | req_set_fail(req); |
889fca73 | 4128 | __io_req_complete(req, issue_flags, ret, 0); |
3e4827b0 JA |
4129 | return 0; |
4130 | #else | |
4131 | return -EOPNOTSUPP; | |
4132 | #endif | |
4133 | } | |
4134 | ||
c1ca757b JA |
4135 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4136 | { | |
4137 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4138 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
4139 | return -EINVAL; | |
3232dd02 PB |
4140 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4141 | return -EINVAL; | |
c1ca757b JA |
4142 | |
4143 | req->madvise.addr = READ_ONCE(sqe->addr); | |
4144 | req->madvise.len = READ_ONCE(sqe->len); | |
4145 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4146 | return 0; | |
4147 | #else | |
4148 | return -EOPNOTSUPP; | |
4149 | #endif | |
4150 | } | |
4151 | ||
45d189c6 | 4152 | static int io_madvise(struct io_kiocb *req, unsigned int issue_flags) |
c1ca757b JA |
4153 | { |
4154 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4155 | struct io_madvise *ma = &req->madvise; | |
4156 | int ret; | |
4157 | ||
45d189c6 | 4158 | if (issue_flags & IO_URING_F_NONBLOCK) |
c1ca757b JA |
4159 | return -EAGAIN; |
4160 | ||
0726b01e | 4161 | ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice); |
c1ca757b | 4162 | if (ret < 0) |
93d2bcd2 | 4163 | req_set_fail(req); |
e1e16097 | 4164 | io_req_complete(req, ret); |
c1ca757b JA |
4165 | return 0; |
4166 | #else | |
4167 | return -EOPNOTSUPP; | |
4168 | #endif | |
4169 | } | |
4170 | ||
4840e418 JA |
4171 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4172 | { | |
4173 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
4174 | return -EINVAL; | |
3232dd02 PB |
4175 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4176 | return -EINVAL; | |
4840e418 JA |
4177 | |
4178 | req->fadvise.offset = READ_ONCE(sqe->off); | |
4179 | req->fadvise.len = READ_ONCE(sqe->len); | |
4180 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4181 | return 0; | |
4182 | } | |
4183 | ||
45d189c6 | 4184 | static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags) |
4840e418 JA |
4185 | { |
4186 | struct io_fadvise *fa = &req->fadvise; | |
4187 | int ret; | |
4188 | ||
45d189c6 | 4189 | if (issue_flags & IO_URING_F_NONBLOCK) { |
3e69426d JA |
4190 | switch (fa->advice) { |
4191 | case POSIX_FADV_NORMAL: | |
4192 | case POSIX_FADV_RANDOM: | |
4193 | case POSIX_FADV_SEQUENTIAL: | |
4194 | break; | |
4195 | default: | |
4196 | return -EAGAIN; | |
4197 | } | |
4198 | } | |
4840e418 JA |
4199 | |
4200 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
4201 | if (ret < 0) | |
93d2bcd2 | 4202 | req_set_fail(req); |
0bdf3398 | 4203 | __io_req_complete(req, issue_flags, ret, 0); |
4840e418 JA |
4204 | return 0; |
4205 | } | |
4206 | ||
eddc7ef5 JA |
4207 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4208 | { | |
2d74d042 | 4209 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4210 | return -EINVAL; |
eddc7ef5 JA |
4211 | if (sqe->ioprio || sqe->buf_index) |
4212 | return -EINVAL; | |
9c280f90 | 4213 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4214 | return -EBADF; |
eddc7ef5 | 4215 | |
1d9e1288 BM |
4216 | req->statx.dfd = READ_ONCE(sqe->fd); |
4217 | req->statx.mask = READ_ONCE(sqe->len); | |
e62753e4 | 4218 | req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1d9e1288 BM |
4219 | req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
4220 | req->statx.flags = READ_ONCE(sqe->statx_flags); | |
eddc7ef5 JA |
4221 | |
4222 | return 0; | |
4223 | } | |
4224 | ||
45d189c6 | 4225 | static int io_statx(struct io_kiocb *req, unsigned int issue_flags) |
eddc7ef5 | 4226 | { |
1d9e1288 | 4227 | struct io_statx *ctx = &req->statx; |
eddc7ef5 JA |
4228 | int ret; |
4229 | ||
59d70013 | 4230 | if (issue_flags & IO_URING_F_NONBLOCK) |
eddc7ef5 JA |
4231 | return -EAGAIN; |
4232 | ||
e62753e4 BM |
4233 | ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask, |
4234 | ctx->buffer); | |
eddc7ef5 | 4235 | |
eddc7ef5 | 4236 | if (ret < 0) |
93d2bcd2 | 4237 | req_set_fail(req); |
e1e16097 | 4238 | io_req_complete(req, ret); |
eddc7ef5 JA |
4239 | return 0; |
4240 | } | |
4241 | ||
b5dba59e JA |
4242 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4243 | { | |
14587a46 | 4244 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4245 | return -EINVAL; |
b5dba59e JA |
4246 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || |
4247 | sqe->rw_flags || sqe->buf_index) | |
4248 | return -EINVAL; | |
9c280f90 | 4249 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4250 | return -EBADF; |
b5dba59e JA |
4251 | |
4252 | req->close.fd = READ_ONCE(sqe->fd); | |
b5dba59e | 4253 | return 0; |
b5dba59e JA |
4254 | } |
4255 | ||
889fca73 | 4256 | static int io_close(struct io_kiocb *req, unsigned int issue_flags) |
b5dba59e | 4257 | { |
9eac1904 | 4258 | struct files_struct *files = current->files; |
3af73b28 | 4259 | struct io_close *close = &req->close; |
9eac1904 | 4260 | struct fdtable *fdt; |
a1fde923 PB |
4261 | struct file *file = NULL; |
4262 | int ret = -EBADF; | |
b5dba59e | 4263 | |
9eac1904 JA |
4264 | spin_lock(&files->file_lock); |
4265 | fdt = files_fdtable(files); | |
4266 | if (close->fd >= fdt->max_fds) { | |
4267 | spin_unlock(&files->file_lock); | |
4268 | goto err; | |
4269 | } | |
4270 | file = fdt->fd[close->fd]; | |
a1fde923 | 4271 | if (!file || file->f_op == &io_uring_fops) { |
9eac1904 JA |
4272 | spin_unlock(&files->file_lock); |
4273 | file = NULL; | |
4274 | goto err; | |
3af73b28 | 4275 | } |
b5dba59e JA |
4276 | |
4277 | /* if the file has a flush method, be safe and punt to async */ | |
45d189c6 | 4278 | if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) { |
9eac1904 | 4279 | spin_unlock(&files->file_lock); |
0bf0eefd | 4280 | return -EAGAIN; |
a2100672 | 4281 | } |
b5dba59e | 4282 | |
9eac1904 JA |
4283 | ret = __close_fd_get_file(close->fd, &file); |
4284 | spin_unlock(&files->file_lock); | |
4285 | if (ret < 0) { | |
4286 | if (ret == -ENOENT) | |
4287 | ret = -EBADF; | |
4288 | goto err; | |
4289 | } | |
4290 | ||
3af73b28 | 4291 | /* No ->flush() or already async, safely close from here */ |
9eac1904 JA |
4292 | ret = filp_close(file, current->files); |
4293 | err: | |
3af73b28 | 4294 | if (ret < 0) |
93d2bcd2 | 4295 | req_set_fail(req); |
9eac1904 JA |
4296 | if (file) |
4297 | fput(file); | |
889fca73 | 4298 | __io_req_complete(req, issue_flags, ret, 0); |
1a417f4e | 4299 | return 0; |
b5dba59e JA |
4300 | } |
4301 | ||
1155c76a | 4302 | static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
4303 | { |
4304 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 | 4305 | |
5d17b4a4 JA |
4306 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
4307 | return -EINVAL; | |
4308 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
4309 | return -EINVAL; | |
4310 | ||
8ed8d3c3 JA |
4311 | req->sync.off = READ_ONCE(sqe->off); |
4312 | req->sync.len = READ_ONCE(sqe->len); | |
4313 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
4314 | return 0; |
4315 | } | |
4316 | ||
45d189c6 | 4317 | static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags) |
8ed8d3c3 | 4318 | { |
8ed8d3c3 JA |
4319 | int ret; |
4320 | ||
ac45abc0 | 4321 | /* sync_file_range always requires a blocking context */ |
45d189c6 | 4322 | if (issue_flags & IO_URING_F_NONBLOCK) |
ac45abc0 PB |
4323 | return -EAGAIN; |
4324 | ||
9adbd45d | 4325 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
4326 | req->sync.flags); |
4327 | if (ret < 0) | |
93d2bcd2 | 4328 | req_set_fail(req); |
e1e16097 | 4329 | io_req_complete(req, ret); |
5d17b4a4 JA |
4330 | return 0; |
4331 | } | |
4332 | ||
469956e8 | 4333 | #if defined(CONFIG_NET) |
02d27d89 PB |
4334 | static int io_setup_async_msg(struct io_kiocb *req, |
4335 | struct io_async_msghdr *kmsg) | |
4336 | { | |
e8c2bc1f JA |
4337 | struct io_async_msghdr *async_msg = req->async_data; |
4338 | ||
4339 | if (async_msg) | |
02d27d89 | 4340 | return -EAGAIN; |
e8c2bc1f | 4341 | if (io_alloc_async_data(req)) { |
257e84a5 | 4342 | kfree(kmsg->free_iov); |
02d27d89 PB |
4343 | return -ENOMEM; |
4344 | } | |
e8c2bc1f | 4345 | async_msg = req->async_data; |
02d27d89 | 4346 | req->flags |= REQ_F_NEED_CLEANUP; |
e8c2bc1f | 4347 | memcpy(async_msg, kmsg, sizeof(*kmsg)); |
2a780802 | 4348 | async_msg->msg.msg_name = &async_msg->addr; |
257e84a5 PB |
4349 | /* if were using fast_iov, set it to the new one */ |
4350 | if (!async_msg->free_iov) | |
4351 | async_msg->msg.msg_iter.iov = async_msg->fast_iov; | |
4352 | ||
02d27d89 PB |
4353 | return -EAGAIN; |
4354 | } | |
4355 | ||
2ae523ed PB |
4356 | static int io_sendmsg_copy_hdr(struct io_kiocb *req, |
4357 | struct io_async_msghdr *iomsg) | |
4358 | { | |
2ae523ed | 4359 | iomsg->msg.msg_name = &iomsg->addr; |
257e84a5 | 4360 | iomsg->free_iov = iomsg->fast_iov; |
2ae523ed | 4361 | return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg, |
257e84a5 | 4362 | req->sr_msg.msg_flags, &iomsg->free_iov); |
2ae523ed PB |
4363 | } |
4364 | ||
93642ef8 PB |
4365 | static int io_sendmsg_prep_async(struct io_kiocb *req) |
4366 | { | |
4367 | int ret; | |
4368 | ||
93642ef8 PB |
4369 | ret = io_sendmsg_copy_hdr(req, req->async_data); |
4370 | if (!ret) | |
4371 | req->flags |= REQ_F_NEED_CLEANUP; | |
4372 | return ret; | |
4373 | } | |
4374 | ||
3529d8c2 | 4375 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 4376 | { |
e47293fd | 4377 | struct io_sr_msg *sr = &req->sr_msg; |
03b1230c | 4378 | |
d2b6f48b PB |
4379 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4380 | return -EINVAL; | |
4381 | ||
270a5940 | 4382 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
fddaface | 4383 | sr->len = READ_ONCE(sqe->len); |
04411806 PB |
4384 | sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL; |
4385 | if (sr->msg_flags & MSG_DONTWAIT) | |
4386 | req->flags |= REQ_F_NOWAIT; | |
3529d8c2 | 4387 | |
d8768362 JA |
4388 | #ifdef CONFIG_COMPAT |
4389 | if (req->ctx->compat) | |
4390 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4391 | #endif | |
93642ef8 | 4392 | return 0; |
03b1230c JA |
4393 | } |
4394 | ||
889fca73 | 4395 | static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags) |
aa1fa28f | 4396 | { |
6b754c8b | 4397 | struct io_async_msghdr iomsg, *kmsg; |
0fa03c62 | 4398 | struct socket *sock; |
7a7cacba | 4399 | unsigned flags; |
0031275d | 4400 | int min_ret = 0; |
0fa03c62 JA |
4401 | int ret; |
4402 | ||
dba4a925 | 4403 | sock = sock_from_file(req->file); |
7a7cacba | 4404 | if (unlikely(!sock)) |
dba4a925 | 4405 | return -ENOTSOCK; |
3529d8c2 | 4406 | |
257e84a5 PB |
4407 | kmsg = req->async_data; |
4408 | if (!kmsg) { | |
7a7cacba PB |
4409 | ret = io_sendmsg_copy_hdr(req, &iomsg); |
4410 | if (ret) | |
4411 | return ret; | |
4412 | kmsg = &iomsg; | |
0fa03c62 | 4413 | } |
0fa03c62 | 4414 | |
04411806 PB |
4415 | flags = req->sr_msg.msg_flags; |
4416 | if (issue_flags & IO_URING_F_NONBLOCK) | |
7a7cacba | 4417 | flags |= MSG_DONTWAIT; |
0031275d SM |
4418 | if (flags & MSG_WAITALL) |
4419 | min_ret = iov_iter_count(&kmsg->msg.msg_iter); | |
4420 | ||
7a7cacba | 4421 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
45d189c6 | 4422 | if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN) |
7a7cacba PB |
4423 | return io_setup_async_msg(req, kmsg); |
4424 | if (ret == -ERESTARTSYS) | |
4425 | ret = -EINTR; | |
0fa03c62 | 4426 | |
257e84a5 PB |
4427 | /* fast path, check for non-NULL to avoid function call */ |
4428 | if (kmsg->free_iov) | |
4429 | kfree(kmsg->free_iov); | |
99bc4c38 | 4430 | req->flags &= ~REQ_F_NEED_CLEANUP; |
0031275d | 4431 | if (ret < min_ret) |
93d2bcd2 | 4432 | req_set_fail(req); |
889fca73 | 4433 | __io_req_complete(req, issue_flags, ret, 0); |
5d17b4a4 | 4434 | return 0; |
03b1230c | 4435 | } |
aa1fa28f | 4436 | |
889fca73 | 4437 | static int io_send(struct io_kiocb *req, unsigned int issue_flags) |
fddaface | 4438 | { |
7a7cacba PB |
4439 | struct io_sr_msg *sr = &req->sr_msg; |
4440 | struct msghdr msg; | |
4441 | struct iovec iov; | |
fddaface | 4442 | struct socket *sock; |
7a7cacba | 4443 | unsigned flags; |
0031275d | 4444 | int min_ret = 0; |
fddaface JA |
4445 | int ret; |
4446 | ||
dba4a925 | 4447 | sock = sock_from_file(req->file); |
7a7cacba | 4448 | if (unlikely(!sock)) |
dba4a925 | 4449 | return -ENOTSOCK; |
fddaface | 4450 | |
7a7cacba PB |
4451 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter); |
4452 | if (unlikely(ret)) | |
14db8411 | 4453 | return ret; |
fddaface | 4454 | |
7a7cacba PB |
4455 | msg.msg_name = NULL; |
4456 | msg.msg_control = NULL; | |
4457 | msg.msg_controllen = 0; | |
4458 | msg.msg_namelen = 0; | |
fddaface | 4459 | |
04411806 PB |
4460 | flags = req->sr_msg.msg_flags; |
4461 | if (issue_flags & IO_URING_F_NONBLOCK) | |
7a7cacba | 4462 | flags |= MSG_DONTWAIT; |
0031275d SM |
4463 | if (flags & MSG_WAITALL) |
4464 | min_ret = iov_iter_count(&msg.msg_iter); | |
4465 | ||
7a7cacba PB |
4466 | msg.msg_flags = flags; |
4467 | ret = sock_sendmsg(sock, &msg); | |
45d189c6 | 4468 | if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN) |
7a7cacba PB |
4469 | return -EAGAIN; |
4470 | if (ret == -ERESTARTSYS) | |
4471 | ret = -EINTR; | |
fddaface | 4472 | |
0031275d | 4473 | if (ret < min_ret) |
93d2bcd2 | 4474 | req_set_fail(req); |
889fca73 | 4475 | __io_req_complete(req, issue_flags, ret, 0); |
fddaface | 4476 | return 0; |
fddaface JA |
4477 | } |
4478 | ||
1400e697 PB |
4479 | static int __io_recvmsg_copy_hdr(struct io_kiocb *req, |
4480 | struct io_async_msghdr *iomsg) | |
52de1fe1 JA |
4481 | { |
4482 | struct io_sr_msg *sr = &req->sr_msg; | |
4483 | struct iovec __user *uiov; | |
4484 | size_t iov_len; | |
4485 | int ret; | |
4486 | ||
1400e697 PB |
4487 | ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg, |
4488 | &iomsg->uaddr, &uiov, &iov_len); | |
52de1fe1 JA |
4489 | if (ret) |
4490 | return ret; | |
4491 | ||
4492 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4493 | if (iov_len > 1) | |
4494 | return -EINVAL; | |
5476dfed | 4495 | if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov))) |
52de1fe1 | 4496 | return -EFAULT; |
5476dfed | 4497 | sr->len = iomsg->fast_iov[0].iov_len; |
257e84a5 | 4498 | iomsg->free_iov = NULL; |
52de1fe1 | 4499 | } else { |
257e84a5 | 4500 | iomsg->free_iov = iomsg->fast_iov; |
89cd35c5 | 4501 | ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV, |
257e84a5 | 4502 | &iomsg->free_iov, &iomsg->msg.msg_iter, |
89cd35c5 | 4503 | false); |
52de1fe1 JA |
4504 | if (ret > 0) |
4505 | ret = 0; | |
4506 | } | |
4507 | ||
4508 | return ret; | |
4509 | } | |
4510 | ||
4511 | #ifdef CONFIG_COMPAT | |
4512 | static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, | |
1400e697 | 4513 | struct io_async_msghdr *iomsg) |
52de1fe1 | 4514 | { |
52de1fe1 JA |
4515 | struct io_sr_msg *sr = &req->sr_msg; |
4516 | struct compat_iovec __user *uiov; | |
4517 | compat_uptr_t ptr; | |
4518 | compat_size_t len; | |
4519 | int ret; | |
4520 | ||
4af3417a PB |
4521 | ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr, |
4522 | &ptr, &len); | |
52de1fe1 JA |
4523 | if (ret) |
4524 | return ret; | |
4525 | ||
4526 | uiov = compat_ptr(ptr); | |
4527 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4528 | compat_ssize_t clen; | |
4529 | ||
4530 | if (len > 1) | |
4531 | return -EINVAL; | |
4532 | if (!access_ok(uiov, sizeof(*uiov))) | |
4533 | return -EFAULT; | |
4534 | if (__get_user(clen, &uiov->iov_len)) | |
4535 | return -EFAULT; | |
4536 | if (clen < 0) | |
4537 | return -EINVAL; | |
2d280bc8 | 4538 | sr->len = clen; |
257e84a5 | 4539 | iomsg->free_iov = NULL; |
52de1fe1 | 4540 | } else { |
257e84a5 | 4541 | iomsg->free_iov = iomsg->fast_iov; |
89cd35c5 | 4542 | ret = __import_iovec(READ, (struct iovec __user *)uiov, len, |
257e84a5 | 4543 | UIO_FASTIOV, &iomsg->free_iov, |
89cd35c5 | 4544 | &iomsg->msg.msg_iter, true); |
52de1fe1 JA |
4545 | if (ret < 0) |
4546 | return ret; | |
4547 | } | |
4548 | ||
4549 | return 0; | |
4550 | } | |
4551 | #endif | |
4552 | ||
1400e697 PB |
4553 | static int io_recvmsg_copy_hdr(struct io_kiocb *req, |
4554 | struct io_async_msghdr *iomsg) | |
52de1fe1 | 4555 | { |
1400e697 | 4556 | iomsg->msg.msg_name = &iomsg->addr; |
52de1fe1 JA |
4557 | |
4558 | #ifdef CONFIG_COMPAT | |
4559 | if (req->ctx->compat) | |
1400e697 | 4560 | return __io_compat_recvmsg_copy_hdr(req, iomsg); |
fddaface | 4561 | #endif |
52de1fe1 | 4562 | |
1400e697 | 4563 | return __io_recvmsg_copy_hdr(req, iomsg); |
52de1fe1 JA |
4564 | } |
4565 | ||
bcda7baa | 4566 | static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req, |
7fbb1b54 | 4567 | bool needs_lock) |
bcda7baa JA |
4568 | { |
4569 | struct io_sr_msg *sr = &req->sr_msg; | |
4570 | struct io_buffer *kbuf; | |
4571 | ||
bcda7baa JA |
4572 | kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock); |
4573 | if (IS_ERR(kbuf)) | |
4574 | return kbuf; | |
4575 | ||
4576 | sr->kbuf = kbuf; | |
4577 | req->flags |= REQ_F_BUFFER_SELECTED; | |
bcda7baa | 4578 | return kbuf; |
fddaface JA |
4579 | } |
4580 | ||
7fbb1b54 PB |
4581 | static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req) |
4582 | { | |
4583 | return io_put_kbuf(req, req->sr_msg.kbuf); | |
4584 | } | |
4585 | ||
93642ef8 | 4586 | static int io_recvmsg_prep_async(struct io_kiocb *req) |
aa1fa28f | 4587 | { |
99bc4c38 | 4588 | int ret; |
3529d8c2 | 4589 | |
93642ef8 PB |
4590 | ret = io_recvmsg_copy_hdr(req, req->async_data); |
4591 | if (!ret) | |
4592 | req->flags |= REQ_F_NEED_CLEANUP; | |
4593 | return ret; | |
4594 | } | |
4595 | ||
4596 | static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
4597 | { | |
4598 | struct io_sr_msg *sr = &req->sr_msg; | |
4599 | ||
d2b6f48b PB |
4600 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4601 | return -EINVAL; | |
4602 | ||
270a5940 | 4603 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
0b7b21e4 | 4604 | sr->len = READ_ONCE(sqe->len); |
bcda7baa | 4605 | sr->bgid = READ_ONCE(sqe->buf_group); |
04411806 PB |
4606 | sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL; |
4607 | if (sr->msg_flags & MSG_DONTWAIT) | |
4608 | req->flags |= REQ_F_NOWAIT; | |
06b76d44 | 4609 | |
d8768362 JA |
4610 | #ifdef CONFIG_COMPAT |
4611 | if (req->ctx->compat) | |
4612 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4613 | #endif | |
93642ef8 | 4614 | return 0; |
aa1fa28f JA |
4615 | } |
4616 | ||
889fca73 | 4617 | static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags) |
aa1fa28f | 4618 | { |
6b754c8b | 4619 | struct io_async_msghdr iomsg, *kmsg; |
03b1230c | 4620 | struct socket *sock; |
7fbb1b54 | 4621 | struct io_buffer *kbuf; |
7a7cacba | 4622 | unsigned flags; |
0031275d | 4623 | int min_ret = 0; |
52de1fe1 | 4624 | int ret, cflags = 0; |
45d189c6 | 4625 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
03b1230c | 4626 | |
dba4a925 | 4627 | sock = sock_from_file(req->file); |
7a7cacba | 4628 | if (unlikely(!sock)) |
dba4a925 | 4629 | return -ENOTSOCK; |
3529d8c2 | 4630 | |
257e84a5 PB |
4631 | kmsg = req->async_data; |
4632 | if (!kmsg) { | |
7a7cacba PB |
4633 | ret = io_recvmsg_copy_hdr(req, &iomsg); |
4634 | if (ret) | |
681fda8d | 4635 | return ret; |
7a7cacba PB |
4636 | kmsg = &iomsg; |
4637 | } | |
03b1230c | 4638 | |
bc02ef33 | 4639 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4640 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bc02ef33 | 4641 | if (IS_ERR(kbuf)) |
52de1fe1 | 4642 | return PTR_ERR(kbuf); |
7a7cacba | 4643 | kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr); |
5476dfed PB |
4644 | kmsg->fast_iov[0].iov_len = req->sr_msg.len; |
4645 | iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov, | |
7a7cacba PB |
4646 | 1, req->sr_msg.len); |
4647 | } | |
52de1fe1 | 4648 | |
04411806 PB |
4649 | flags = req->sr_msg.msg_flags; |
4650 | if (force_nonblock) | |
7a7cacba | 4651 | flags |= MSG_DONTWAIT; |
0031275d SM |
4652 | if (flags & MSG_WAITALL) |
4653 | min_ret = iov_iter_count(&kmsg->msg.msg_iter); | |
4654 | ||
7a7cacba PB |
4655 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg, |
4656 | kmsg->uaddr, flags); | |
0e1b6fe3 PB |
4657 | if (force_nonblock && ret == -EAGAIN) |
4658 | return io_setup_async_msg(req, kmsg); | |
7a7cacba PB |
4659 | if (ret == -ERESTARTSYS) |
4660 | ret = -EINTR; | |
03b1230c | 4661 | |
7fbb1b54 PB |
4662 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4663 | cflags = io_put_recv_kbuf(req); | |
257e84a5 PB |
4664 | /* fast path, check for non-NULL to avoid function call */ |
4665 | if (kmsg->free_iov) | |
4666 | kfree(kmsg->free_iov); | |
99bc4c38 | 4667 | req->flags &= ~REQ_F_NEED_CLEANUP; |
0031275d | 4668 | if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC)))) |
93d2bcd2 | 4669 | req_set_fail(req); |
889fca73 | 4670 | __io_req_complete(req, issue_flags, ret, cflags); |
03b1230c | 4671 | return 0; |
0fa03c62 | 4672 | } |
5d17b4a4 | 4673 | |
889fca73 | 4674 | static int io_recv(struct io_kiocb *req, unsigned int issue_flags) |
fddaface | 4675 | { |
6b754c8b | 4676 | struct io_buffer *kbuf; |
7a7cacba PB |
4677 | struct io_sr_msg *sr = &req->sr_msg; |
4678 | struct msghdr msg; | |
4679 | void __user *buf = sr->buf; | |
fddaface | 4680 | struct socket *sock; |
7a7cacba PB |
4681 | struct iovec iov; |
4682 | unsigned flags; | |
0031275d | 4683 | int min_ret = 0; |
bcda7baa | 4684 | int ret, cflags = 0; |
45d189c6 | 4685 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
fddaface | 4686 | |
dba4a925 | 4687 | sock = sock_from_file(req->file); |
7a7cacba | 4688 | if (unlikely(!sock)) |
dba4a925 | 4689 | return -ENOTSOCK; |
fddaface | 4690 | |
bc02ef33 | 4691 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4692 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bcda7baa JA |
4693 | if (IS_ERR(kbuf)) |
4694 | return PTR_ERR(kbuf); | |
7a7cacba | 4695 | buf = u64_to_user_ptr(kbuf->addr); |
bc02ef33 | 4696 | } |
bcda7baa | 4697 | |
7a7cacba | 4698 | ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter); |
14c32eee PB |
4699 | if (unlikely(ret)) |
4700 | goto out_free; | |
fddaface | 4701 | |
7a7cacba PB |
4702 | msg.msg_name = NULL; |
4703 | msg.msg_control = NULL; | |
4704 | msg.msg_controllen = 0; | |
4705 | msg.msg_namelen = 0; | |
4706 | msg.msg_iocb = NULL; | |
4707 | msg.msg_flags = 0; | |
fddaface | 4708 | |
04411806 PB |
4709 | flags = req->sr_msg.msg_flags; |
4710 | if (force_nonblock) | |
7a7cacba | 4711 | flags |= MSG_DONTWAIT; |
0031275d SM |
4712 | if (flags & MSG_WAITALL) |
4713 | min_ret = iov_iter_count(&msg.msg_iter); | |
4714 | ||
7a7cacba PB |
4715 | ret = sock_recvmsg(sock, &msg, flags); |
4716 | if (force_nonblock && ret == -EAGAIN) | |
4717 | return -EAGAIN; | |
4718 | if (ret == -ERESTARTSYS) | |
4719 | ret = -EINTR; | |
14c32eee | 4720 | out_free: |
7fbb1b54 PB |
4721 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4722 | cflags = io_put_recv_kbuf(req); | |
0031275d | 4723 | if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC)))) |
93d2bcd2 | 4724 | req_set_fail(req); |
889fca73 | 4725 | __io_req_complete(req, issue_flags, ret, cflags); |
fddaface | 4726 | return 0; |
fddaface JA |
4727 | } |
4728 | ||
3529d8c2 | 4729 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 | 4730 | { |
8ed8d3c3 JA |
4731 | struct io_accept *accept = &req->accept; |
4732 | ||
14587a46 | 4733 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
17f2fe35 | 4734 | return -EINVAL; |
8042d6ce | 4735 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
4736 | return -EINVAL; |
4737 | ||
d55e5f5b JA |
4738 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4739 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 4740 | accept->flags = READ_ONCE(sqe->accept_flags); |
09952e3e | 4741 | accept->nofile = rlimit(RLIMIT_NOFILE); |
8ed8d3c3 | 4742 | return 0; |
8ed8d3c3 | 4743 | } |
17f2fe35 | 4744 | |
889fca73 | 4745 | static int io_accept(struct io_kiocb *req, unsigned int issue_flags) |
8ed8d3c3 JA |
4746 | { |
4747 | struct io_accept *accept = &req->accept; | |
45d189c6 | 4748 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
ac45abc0 | 4749 | unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0; |
8ed8d3c3 JA |
4750 | int ret; |
4751 | ||
e697deed JX |
4752 | if (req->file->f_flags & O_NONBLOCK) |
4753 | req->flags |= REQ_F_NOWAIT; | |
4754 | ||
8ed8d3c3 | 4755 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, |
09952e3e JA |
4756 | accept->addr_len, accept->flags, |
4757 | accept->nofile); | |
8ed8d3c3 | 4758 | if (ret == -EAGAIN && force_nonblock) |
17f2fe35 | 4759 | return -EAGAIN; |
ac45abc0 PB |
4760 | if (ret < 0) { |
4761 | if (ret == -ERESTARTSYS) | |
4762 | ret = -EINTR; | |
93d2bcd2 | 4763 | req_set_fail(req); |
ac45abc0 | 4764 | } |
889fca73 | 4765 | __io_req_complete(req, issue_flags, ret, 0); |
17f2fe35 | 4766 | return 0; |
8ed8d3c3 JA |
4767 | } |
4768 | ||
93642ef8 PB |
4769 | static int io_connect_prep_async(struct io_kiocb *req) |
4770 | { | |
4771 | struct io_async_connect *io = req->async_data; | |
4772 | struct io_connect *conn = &req->connect; | |
4773 | ||
4774 | return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address); | |
4775 | } | |
4776 | ||
3529d8c2 | 4777 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 | 4778 | { |
3529d8c2 | 4779 | struct io_connect *conn = &req->connect; |
f499a021 | 4780 | |
14587a46 | 4781 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3fbb51c1 JA |
4782 | return -EINVAL; |
4783 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
4784 | return -EINVAL; | |
4785 | ||
3529d8c2 JA |
4786 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4787 | conn->addr_len = READ_ONCE(sqe->addr2); | |
93642ef8 | 4788 | return 0; |
f499a021 JA |
4789 | } |
4790 | ||
889fca73 | 4791 | static int io_connect(struct io_kiocb *req, unsigned int issue_flags) |
f8e85cf2 | 4792 | { |
e8c2bc1f | 4793 | struct io_async_connect __io, *io; |
f8e85cf2 | 4794 | unsigned file_flags; |
3fbb51c1 | 4795 | int ret; |
45d189c6 | 4796 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; |
f8e85cf2 | 4797 | |
e8c2bc1f JA |
4798 | if (req->async_data) { |
4799 | io = req->async_data; | |
f499a021 | 4800 | } else { |
3529d8c2 JA |
4801 | ret = move_addr_to_kernel(req->connect.addr, |
4802 | req->connect.addr_len, | |
e8c2bc1f | 4803 | &__io.address); |
f499a021 JA |
4804 | if (ret) |
4805 | goto out; | |
4806 | io = &__io; | |
4807 | } | |
4808 | ||
3fbb51c1 JA |
4809 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
4810 | ||
e8c2bc1f | 4811 | ret = __sys_connect_file(req->file, &io->address, |
3fbb51c1 | 4812 | req->connect.addr_len, file_flags); |
87f80d62 | 4813 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
e8c2bc1f | 4814 | if (req->async_data) |
b7bb4f7d | 4815 | return -EAGAIN; |
e8c2bc1f | 4816 | if (io_alloc_async_data(req)) { |
f499a021 JA |
4817 | ret = -ENOMEM; |
4818 | goto out; | |
4819 | } | |
e8c2bc1f | 4820 | memcpy(req->async_data, &__io, sizeof(__io)); |
f8e85cf2 | 4821 | return -EAGAIN; |
f499a021 | 4822 | } |
f8e85cf2 JA |
4823 | if (ret == -ERESTARTSYS) |
4824 | ret = -EINTR; | |
f499a021 | 4825 | out: |
4e88d6e7 | 4826 | if (ret < 0) |
93d2bcd2 | 4827 | req_set_fail(req); |
889fca73 | 4828 | __io_req_complete(req, issue_flags, ret, 0); |
f8e85cf2 | 4829 | return 0; |
469956e8 Y |
4830 | } |
4831 | #else /* !CONFIG_NET */ | |
99a10081 JA |
4832 | #define IO_NETOP_FN(op) \ |
4833 | static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \ | |
4834 | { \ | |
4835 | return -EOPNOTSUPP; \ | |
4836 | } | |
4837 | ||
4838 | #define IO_NETOP_PREP(op) \ | |
4839 | IO_NETOP_FN(op) \ | |
4840 | static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \ | |
4841 | { \ | |
4842 | return -EOPNOTSUPP; \ | |
4843 | } \ | |
4844 | ||
4845 | #define IO_NETOP_PREP_ASYNC(op) \ | |
4846 | IO_NETOP_PREP(op) \ | |
4847 | static int io_##op##_prep_async(struct io_kiocb *req) \ | |
4848 | { \ | |
4849 | return -EOPNOTSUPP; \ | |
4850 | } | |
4851 | ||
4852 | IO_NETOP_PREP_ASYNC(sendmsg); | |
4853 | IO_NETOP_PREP_ASYNC(recvmsg); | |
4854 | IO_NETOP_PREP_ASYNC(connect); | |
4855 | IO_NETOP_PREP(accept); | |
4856 | IO_NETOP_FN(send); | |
4857 | IO_NETOP_FN(recv); | |
469956e8 | 4858 | #endif /* CONFIG_NET */ |
f8e85cf2 | 4859 | |
d7718a9d JA |
4860 | struct io_poll_table { |
4861 | struct poll_table_struct pt; | |
4862 | struct io_kiocb *req; | |
68b11e8b | 4863 | int nr_entries; |
d7718a9d JA |
4864 | int error; |
4865 | }; | |
ce593a6c | 4866 | |
d7718a9d | 4867 | static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll, |
5b0a6acc | 4868 | __poll_t mask, io_req_tw_func_t func) |
d7718a9d | 4869 | { |
d7718a9d JA |
4870 | /* for instances that support it check for an event match first: */ |
4871 | if (mask && !(mask & poll->events)) | |
4872 | return 0; | |
4873 | ||
4874 | trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask); | |
4875 | ||
4876 | list_del_init(&poll->wait.entry); | |
4877 | ||
d7718a9d | 4878 | req->result = mask; |
5b0a6acc | 4879 | req->io_task_work.func = func; |
6d816e08 | 4880 | |
d7718a9d | 4881 | /* |
e3aabf95 JA |
4882 | * If this fails, then the task is exiting. When a task exits, the |
4883 | * work gets canceled, so just cancel this request as well instead | |
4884 | * of executing it. We can't safely execute it anyway, as we may not | |
4885 | * have the needed state needed for it anyway. | |
d7718a9d | 4886 | */ |
e09ee510 | 4887 | io_req_task_work_add(req); |
d7718a9d JA |
4888 | return 1; |
4889 | } | |
4890 | ||
74ce6ce4 JA |
4891 | static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll) |
4892 | __acquires(&req->ctx->completion_lock) | |
4893 | { | |
4894 | struct io_ring_ctx *ctx = req->ctx; | |
4895 | ||
e09ee510 PB |
4896 | if (unlikely(req->task->flags & PF_EXITING)) |
4897 | WRITE_ONCE(poll->canceled, true); | |
4898 | ||
74ce6ce4 JA |
4899 | if (!req->result && !READ_ONCE(poll->canceled)) { |
4900 | struct poll_table_struct pt = { ._key = poll->events }; | |
4901 | ||
4902 | req->result = vfs_poll(req->file, &pt) & poll->events; | |
4903 | } | |
4904 | ||
79ebeaee | 4905 | spin_lock(&ctx->completion_lock); |
74ce6ce4 JA |
4906 | if (!req->result && !READ_ONCE(poll->canceled)) { |
4907 | add_wait_queue(poll->head, &poll->wait); | |
4908 | return true; | |
4909 | } | |
4910 | ||
4911 | return false; | |
4912 | } | |
4913 | ||
d4e7cd36 | 4914 | static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req) |
18bceab1 | 4915 | { |
e8c2bc1f | 4916 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ |
d4e7cd36 | 4917 | if (req->opcode == IORING_OP_POLL_ADD) |
e8c2bc1f | 4918 | return req->async_data; |
d4e7cd36 JA |
4919 | return req->apoll->double_poll; |
4920 | } | |
4921 | ||
4922 | static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req) | |
4923 | { | |
4924 | if (req->opcode == IORING_OP_POLL_ADD) | |
4925 | return &req->poll; | |
4926 | return &req->apoll->poll; | |
4927 | } | |
4928 | ||
4929 | static void io_poll_remove_double(struct io_kiocb *req) | |
e07785b0 | 4930 | __must_hold(&req->ctx->completion_lock) |
d4e7cd36 JA |
4931 | { |
4932 | struct io_poll_iocb *poll = io_poll_get_double(req); | |
18bceab1 JA |
4933 | |
4934 | lockdep_assert_held(&req->ctx->completion_lock); | |
4935 | ||
4936 | if (poll && poll->head) { | |
4937 | struct wait_queue_head *head = poll->head; | |
4938 | ||
79ebeaee | 4939 | spin_lock_irq(&head->lock); |
18bceab1 JA |
4940 | list_del_init(&poll->wait.entry); |
4941 | if (poll->wait.private) | |
de9b4cca | 4942 | req_ref_put(req); |
18bceab1 | 4943 | poll->head = NULL; |
79ebeaee | 4944 | spin_unlock_irq(&head->lock); |
18bceab1 JA |
4945 | } |
4946 | } | |
4947 | ||
e27414be | 4948 | static bool io_poll_complete(struct io_kiocb *req, __poll_t mask) |
e07785b0 | 4949 | __must_hold(&req->ctx->completion_lock) |
18bceab1 JA |
4950 | { |
4951 | struct io_ring_ctx *ctx = req->ctx; | |
88e41cf9 | 4952 | unsigned flags = IORING_CQE_F_MORE; |
e27414be | 4953 | int error; |
18bceab1 | 4954 | |
e27414be | 4955 | if (READ_ONCE(req->poll.canceled)) { |
45ab03b1 | 4956 | error = -ECANCELED; |
88e41cf9 | 4957 | req->poll.events |= EPOLLONESHOT; |
e27414be | 4958 | } else { |
5082620f | 4959 | error = mangle_poll(mask); |
e27414be | 4960 | } |
b69de288 JA |
4961 | if (req->poll.events & EPOLLONESHOT) |
4962 | flags = 0; | |
d4d19c19 | 4963 | if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) { |
88e41cf9 JA |
4964 | req->poll.done = true; |
4965 | flags = 0; | |
4966 | } | |
7b289c38 HX |
4967 | if (flags & IORING_CQE_F_MORE) |
4968 | ctx->cq_extra++; | |
18bceab1 | 4969 | |
18bceab1 | 4970 | io_commit_cqring(ctx); |
88e41cf9 | 4971 | return !(flags & IORING_CQE_F_MORE); |
18bceab1 JA |
4972 | } |
4973 | ||
5b0a6acc | 4974 | static void io_poll_task_func(struct io_kiocb *req) |
18bceab1 JA |
4975 | { |
4976 | struct io_ring_ctx *ctx = req->ctx; | |
dd221f46 | 4977 | struct io_kiocb *nxt; |
18bceab1 JA |
4978 | |
4979 | if (io_poll_rewait(req, &req->poll)) { | |
79ebeaee | 4980 | spin_unlock(&ctx->completion_lock); |
dd221f46 | 4981 | } else { |
f40b964a | 4982 | bool done; |
18bceab1 | 4983 | |
e27414be | 4984 | done = io_poll_complete(req, req->result); |
88e41cf9 | 4985 | if (done) { |
a890d01e | 4986 | io_poll_remove_double(req); |
88e41cf9 | 4987 | hash_del(&req->hash_node); |
f40b964a | 4988 | } else { |
88e41cf9 JA |
4989 | req->result = 0; |
4990 | add_wait_queue(req->poll.head, &req->poll.wait); | |
4991 | } | |
79ebeaee | 4992 | spin_unlock(&ctx->completion_lock); |
dd221f46 | 4993 | io_cqring_ev_posted(ctx); |
18bceab1 | 4994 | |
88e41cf9 JA |
4995 | if (done) { |
4996 | nxt = io_put_req_find_next(req); | |
4997 | if (nxt) | |
5b0a6acc | 4998 | io_req_task_submit(nxt); |
88e41cf9 | 4999 | } |
dd221f46 | 5000 | } |
18bceab1 JA |
5001 | } |
5002 | ||
5003 | static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode, | |
5004 | int sync, void *key) | |
5005 | { | |
5006 | struct io_kiocb *req = wait->private; | |
d4e7cd36 | 5007 | struct io_poll_iocb *poll = io_poll_get_single(req); |
18bceab1 | 5008 | __poll_t mask = key_to_poll(key); |
79ebeaee | 5009 | unsigned long flags; |
18bceab1 JA |
5010 | |
5011 | /* for instances that support it check for an event match first: */ | |
5012 | if (mask && !(mask & poll->events)) | |
5013 | return 0; | |
88e41cf9 JA |
5014 | if (!(poll->events & EPOLLONESHOT)) |
5015 | return poll->wait.func(&poll->wait, mode, sync, key); | |
18bceab1 | 5016 | |
8706e04e JA |
5017 | list_del_init(&wait->entry); |
5018 | ||
9ce85ef2 | 5019 | if (poll->head) { |
18bceab1 JA |
5020 | bool done; |
5021 | ||
79ebeaee | 5022 | spin_lock_irqsave(&poll->head->lock, flags); |
807abcb0 | 5023 | done = list_empty(&poll->wait.entry); |
18bceab1 | 5024 | if (!done) |
807abcb0 | 5025 | list_del_init(&poll->wait.entry); |
d4e7cd36 JA |
5026 | /* make sure double remove sees this as being gone */ |
5027 | wait->private = NULL; | |
79ebeaee | 5028 | spin_unlock_irqrestore(&poll->head->lock, flags); |
c8b5e260 JA |
5029 | if (!done) { |
5030 | /* use wait func handler, so it matches the rq type */ | |
5031 | poll->wait.func(&poll->wait, mode, sync, key); | |
5032 | } | |
18bceab1 | 5033 | } |
de9b4cca | 5034 | req_ref_put(req); |
18bceab1 JA |
5035 | return 1; |
5036 | } | |
5037 | ||
5038 | static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events, | |
5039 | wait_queue_func_t wake_func) | |
5040 | { | |
5041 | poll->head = NULL; | |
5042 | poll->done = false; | |
5043 | poll->canceled = false; | |
464dca61 JA |
5044 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) |
5045 | /* mask in events that we always want/need */ | |
5046 | poll->events = events | IO_POLL_UNMASK; | |
18bceab1 JA |
5047 | INIT_LIST_HEAD(&poll->wait.entry); |
5048 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
5049 | } | |
5050 | ||
5051 | static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt, | |
807abcb0 JA |
5052 | struct wait_queue_head *head, |
5053 | struct io_poll_iocb **poll_ptr) | |
18bceab1 JA |
5054 | { |
5055 | struct io_kiocb *req = pt->req; | |
5056 | ||
5057 | /* | |
68b11e8b PB |
5058 | * The file being polled uses multiple waitqueues for poll handling |
5059 | * (e.g. one for read, one for write). Setup a separate io_poll_iocb | |
5060 | * if this happens. | |
18bceab1 | 5061 | */ |
68b11e8b | 5062 | if (unlikely(pt->nr_entries)) { |
58852d4d PB |
5063 | struct io_poll_iocb *poll_one = poll; |
5064 | ||
23a65db8 PB |
5065 | /* double add on the same waitqueue head, ignore */ |
5066 | if (poll_one->head == head) | |
5067 | return; | |
18bceab1 | 5068 | /* already have a 2nd entry, fail a third attempt */ |
807abcb0 | 5069 | if (*poll_ptr) { |
23a65db8 PB |
5070 | if ((*poll_ptr)->head == head) |
5071 | return; | |
18bceab1 JA |
5072 | pt->error = -EINVAL; |
5073 | return; | |
5074 | } | |
ea6a693d JA |
5075 | /* |
5076 | * Can't handle multishot for double wait for now, turn it | |
5077 | * into one-shot mode. | |
5078 | */ | |
7a274727 PB |
5079 | if (!(poll_one->events & EPOLLONESHOT)) |
5080 | poll_one->events |= EPOLLONESHOT; | |
18bceab1 JA |
5081 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); |
5082 | if (!poll) { | |
5083 | pt->error = -ENOMEM; | |
5084 | return; | |
5085 | } | |
58852d4d | 5086 | io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake); |
de9b4cca | 5087 | req_ref_get(req); |
18bceab1 | 5088 | poll->wait.private = req; |
807abcb0 | 5089 | *poll_ptr = poll; |
18bceab1 JA |
5090 | } |
5091 | ||
68b11e8b | 5092 | pt->nr_entries++; |
18bceab1 | 5093 | poll->head = head; |
a31eb4a2 JX |
5094 | |
5095 | if (poll->events & EPOLLEXCLUSIVE) | |
5096 | add_wait_queue_exclusive(head, &poll->wait); | |
5097 | else | |
5098 | add_wait_queue(head, &poll->wait); | |
18bceab1 JA |
5099 | } |
5100 | ||
5101 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
5102 | struct poll_table_struct *p) | |
5103 | { | |
5104 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
807abcb0 | 5105 | struct async_poll *apoll = pt->req->apoll; |
18bceab1 | 5106 | |
807abcb0 | 5107 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
18bceab1 JA |
5108 | } |
5109 | ||
5b0a6acc | 5110 | static void io_async_task_func(struct io_kiocb *req) |
d7718a9d | 5111 | { |
d7718a9d JA |
5112 | struct async_poll *apoll = req->apoll; |
5113 | struct io_ring_ctx *ctx = req->ctx; | |
5114 | ||
236daeae | 5115 | trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data); |
d7718a9d | 5116 | |
74ce6ce4 | 5117 | if (io_poll_rewait(req, &apoll->poll)) { |
79ebeaee | 5118 | spin_unlock(&ctx->completion_lock); |
74ce6ce4 | 5119 | return; |
d7718a9d JA |
5120 | } |
5121 | ||
0ea13b44 | 5122 | hash_del(&req->hash_node); |
d4e7cd36 | 5123 | io_poll_remove_double(req); |
79ebeaee | 5124 | spin_unlock(&ctx->completion_lock); |
74ce6ce4 | 5125 | |
0be0b0e3 | 5126 | if (!READ_ONCE(apoll->poll.canceled)) |
5b0a6acc | 5127 | io_req_task_submit(req); |
0be0b0e3 | 5128 | else |
2593553a | 5129 | io_req_complete_failed(req, -ECANCELED); |
d7718a9d JA |
5130 | } |
5131 | ||
5132 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
5133 | void *key) | |
5134 | { | |
5135 | struct io_kiocb *req = wait->private; | |
5136 | struct io_poll_iocb *poll = &req->apoll->poll; | |
5137 | ||
5138 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
5139 | key_to_poll(key)); | |
5140 | ||
5141 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
5142 | } | |
5143 | ||
5144 | static void io_poll_req_insert(struct io_kiocb *req) | |
5145 | { | |
5146 | struct io_ring_ctx *ctx = req->ctx; | |
5147 | struct hlist_head *list; | |
5148 | ||
5149 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
5150 | hlist_add_head(&req->hash_node, list); | |
5151 | } | |
5152 | ||
5153 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
5154 | struct io_poll_iocb *poll, | |
5155 | struct io_poll_table *ipt, __poll_t mask, | |
5156 | wait_queue_func_t wake_func) | |
5157 | __acquires(&ctx->completion_lock) | |
5158 | { | |
5159 | struct io_ring_ctx *ctx = req->ctx; | |
5160 | bool cancel = false; | |
5161 | ||
4d52f338 | 5162 | INIT_HLIST_NODE(&req->hash_node); |
18bceab1 | 5163 | io_init_poll_iocb(poll, mask, wake_func); |
b90cd197 | 5164 | poll->file = req->file; |
18bceab1 | 5165 | poll->wait.private = req; |
d7718a9d JA |
5166 | |
5167 | ipt->pt._key = mask; | |
5168 | ipt->req = req; | |
68b11e8b PB |
5169 | ipt->error = 0; |
5170 | ipt->nr_entries = 0; | |
d7718a9d | 5171 | |
d7718a9d | 5172 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
68b11e8b PB |
5173 | if (unlikely(!ipt->nr_entries) && !ipt->error) |
5174 | ipt->error = -EINVAL; | |
d7718a9d | 5175 | |
79ebeaee | 5176 | spin_lock(&ctx->completion_lock); |
a890d01e | 5177 | if (ipt->error || (mask && (poll->events & EPOLLONESHOT))) |
46fee9ab | 5178 | io_poll_remove_double(req); |
d7718a9d | 5179 | if (likely(poll->head)) { |
79ebeaee | 5180 | spin_lock_irq(&poll->head->lock); |
d7718a9d JA |
5181 | if (unlikely(list_empty(&poll->wait.entry))) { |
5182 | if (ipt->error) | |
5183 | cancel = true; | |
5184 | ipt->error = 0; | |
5185 | mask = 0; | |
5186 | } | |
88e41cf9 | 5187 | if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error) |
d7718a9d JA |
5188 | list_del_init(&poll->wait.entry); |
5189 | else if (cancel) | |
5190 | WRITE_ONCE(poll->canceled, true); | |
5191 | else if (!poll->done) /* actually waiting for an event */ | |
5192 | io_poll_req_insert(req); | |
79ebeaee | 5193 | spin_unlock_irq(&poll->head->lock); |
d7718a9d JA |
5194 | } |
5195 | ||
5196 | return mask; | |
5197 | } | |
5198 | ||
59b735ae OL |
5199 | enum { |
5200 | IO_APOLL_OK, | |
5201 | IO_APOLL_ABORTED, | |
5202 | IO_APOLL_READY | |
5203 | }; | |
5204 | ||
5205 | static int io_arm_poll_handler(struct io_kiocb *req) | |
d7718a9d JA |
5206 | { |
5207 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
5208 | struct io_ring_ctx *ctx = req->ctx; | |
5209 | struct async_poll *apoll; | |
5210 | struct io_poll_table ipt; | |
b2d9c3da | 5211 | __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI; |
9dab14b8 | 5212 | int rw; |
d7718a9d JA |
5213 | |
5214 | if (!req->file || !file_can_poll(req->file)) | |
59b735ae | 5215 | return IO_APOLL_ABORTED; |
24c74678 | 5216 | if (req->flags & REQ_F_POLLED) |
59b735ae | 5217 | return IO_APOLL_ABORTED; |
b2d9c3da PB |
5218 | if (!def->pollin && !def->pollout) |
5219 | return IO_APOLL_ABORTED; | |
5220 | ||
5221 | if (def->pollin) { | |
9dab14b8 | 5222 | rw = READ; |
b2d9c3da PB |
5223 | mask |= POLLIN | POLLRDNORM; |
5224 | ||
5225 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
5226 | if ((req->opcode == IORING_OP_RECVMSG) && | |
5227 | (req->sr_msg.msg_flags & MSG_ERRQUEUE)) | |
5228 | mask &= ~POLLIN; | |
5229 | } else { | |
9dab14b8 | 5230 | rw = WRITE; |
b2d9c3da PB |
5231 | mask |= POLLOUT | POLLWRNORM; |
5232 | } | |
5233 | ||
9dab14b8 | 5234 | /* if we can't nonblock try, then no point in arming a poll handler */ |
b191e2df | 5235 | if (!io_file_supports_nowait(req, rw)) |
59b735ae | 5236 | return IO_APOLL_ABORTED; |
d7718a9d JA |
5237 | |
5238 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
5239 | if (unlikely(!apoll)) | |
59b735ae | 5240 | return IO_APOLL_ABORTED; |
807abcb0 | 5241 | apoll->double_poll = NULL; |
d7718a9d | 5242 | req->apoll = apoll; |
b2d9c3da | 5243 | req->flags |= REQ_F_POLLED; |
d7718a9d | 5244 | ipt.pt._qproc = io_async_queue_proc; |
48dcd38d | 5245 | io_req_set_refcount(req); |
d7718a9d JA |
5246 | |
5247 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
5248 | io_async_wake); | |
79ebeaee | 5249 | spin_unlock(&ctx->completion_lock); |
41a5169c HX |
5250 | if (ret || ipt.error) |
5251 | return ret ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
5252 | ||
236daeae OL |
5253 | trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data, |
5254 | mask, apoll->poll.events); | |
59b735ae | 5255 | return IO_APOLL_OK; |
d7718a9d JA |
5256 | } |
5257 | ||
5258 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
b2e720ac | 5259 | struct io_poll_iocb *poll, bool do_cancel) |
e07785b0 | 5260 | __must_hold(&req->ctx->completion_lock) |
221c5eb2 | 5261 | { |
b41e9852 | 5262 | bool do_complete = false; |
221c5eb2 | 5263 | |
5082620f JA |
5264 | if (!poll->head) |
5265 | return false; | |
79ebeaee | 5266 | spin_lock_irq(&poll->head->lock); |
b2e720ac JA |
5267 | if (do_cancel) |
5268 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
5269 | if (!list_empty(&poll->wait.entry)) { |
5270 | list_del_init(&poll->wait.entry); | |
b41e9852 | 5271 | do_complete = true; |
221c5eb2 | 5272 | } |
79ebeaee | 5273 | spin_unlock_irq(&poll->head->lock); |
3bfa5bcb | 5274 | hash_del(&req->hash_node); |
d7718a9d JA |
5275 | return do_complete; |
5276 | } | |
5277 | ||
5d709043 | 5278 | static bool io_poll_remove_one(struct io_kiocb *req) |
e07785b0 | 5279 | __must_hold(&req->ctx->completion_lock) |
d7718a9d JA |
5280 | { |
5281 | bool do_complete; | |
5282 | ||
d4e7cd36 | 5283 | io_poll_remove_double(req); |
e31001a3 | 5284 | do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true); |
d4e7cd36 | 5285 | |
b41e9852 | 5286 | if (do_complete) { |
d4d19c19 | 5287 | io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0); |
b41e9852 | 5288 | io_commit_cqring(req->ctx); |
93d2bcd2 | 5289 | req_set_fail(req); |
91c2f697 | 5290 | io_put_req_deferred(req); |
5d709043 | 5291 | } |
b41e9852 | 5292 | return do_complete; |
221c5eb2 JA |
5293 | } |
5294 | ||
76e1b642 JA |
5295 | /* |
5296 | * Returns true if we found and killed one or more poll requests | |
5297 | */ | |
6b81928d | 5298 | static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, |
3dd0c97a | 5299 | bool cancel_all) |
221c5eb2 | 5300 | { |
78076bb6 | 5301 | struct hlist_node *tmp; |
221c5eb2 | 5302 | struct io_kiocb *req; |
8e2e1faf | 5303 | int posted = 0, i; |
221c5eb2 | 5304 | |
79ebeaee | 5305 | spin_lock(&ctx->completion_lock); |
78076bb6 JA |
5306 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
5307 | struct hlist_head *list; | |
5308 | ||
5309 | list = &ctx->cancel_hash[i]; | |
f3606e3a | 5310 | hlist_for_each_entry_safe(req, tmp, list, hash_node) { |
3dd0c97a | 5311 | if (io_match_task(req, tsk, cancel_all)) |
f3606e3a JA |
5312 | posted += io_poll_remove_one(req); |
5313 | } | |
221c5eb2 | 5314 | } |
79ebeaee | 5315 | spin_unlock(&ctx->completion_lock); |
b41e9852 | 5316 | |
8e2e1faf JA |
5317 | if (posted) |
5318 | io_cqring_ev_posted(ctx); | |
76e1b642 JA |
5319 | |
5320 | return posted != 0; | |
221c5eb2 JA |
5321 | } |
5322 | ||
9ba5fac8 PB |
5323 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr, |
5324 | bool poll_only) | |
e07785b0 | 5325 | __must_hold(&ctx->completion_lock) |
47f46768 | 5326 | { |
78076bb6 | 5327 | struct hlist_head *list; |
47f46768 JA |
5328 | struct io_kiocb *req; |
5329 | ||
78076bb6 JA |
5330 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
5331 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
5332 | if (sqe_addr != req->user_data) |
5333 | continue; | |
9ba5fac8 PB |
5334 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) |
5335 | continue; | |
b2cb805f | 5336 | return req; |
47f46768 | 5337 | } |
b2cb805f JA |
5338 | return NULL; |
5339 | } | |
5340 | ||
9ba5fac8 PB |
5341 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr, |
5342 | bool poll_only) | |
e07785b0 | 5343 | __must_hold(&ctx->completion_lock) |
b2cb805f JA |
5344 | { |
5345 | struct io_kiocb *req; | |
5346 | ||
9ba5fac8 | 5347 | req = io_poll_find(ctx, sqe_addr, poll_only); |
b2cb805f JA |
5348 | if (!req) |
5349 | return -ENOENT; | |
5350 | if (io_poll_remove_one(req)) | |
5351 | return 0; | |
5352 | ||
5353 | return -EALREADY; | |
47f46768 JA |
5354 | } |
5355 | ||
9096af3e PB |
5356 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
5357 | unsigned int flags) | |
5358 | { | |
5359 | u32 events; | |
47f46768 | 5360 | |
9096af3e PB |
5361 | events = READ_ONCE(sqe->poll32_events); |
5362 | #ifdef __BIG_ENDIAN | |
5363 | events = swahw32(events); | |
5364 | #endif | |
5365 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
5366 | events |= EPOLLONESHOT; | |
5367 | return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT)); | |
47f46768 JA |
5368 | } |
5369 | ||
c5de0036 | 5370 | static int io_poll_update_prep(struct io_kiocb *req, |
3529d8c2 | 5371 | const struct io_uring_sqe *sqe) |
0969e783 | 5372 | { |
c5de0036 PB |
5373 | struct io_poll_update *upd = &req->poll_update; |
5374 | u32 flags; | |
5375 | ||
0969e783 JA |
5376 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5377 | return -EINVAL; | |
c5de0036 PB |
5378 | if (sqe->ioprio || sqe->buf_index) |
5379 | return -EINVAL; | |
5380 | flags = READ_ONCE(sqe->len); | |
5381 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
5382 | IORING_POLL_ADD_MULTI)) | |
5383 | return -EINVAL; | |
5384 | /* meaningless without update */ | |
5385 | if (flags == IORING_POLL_ADD_MULTI) | |
0969e783 JA |
5386 | return -EINVAL; |
5387 | ||
c5de0036 PB |
5388 | upd->old_user_data = READ_ONCE(sqe->addr); |
5389 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
5390 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
221c5eb2 | 5391 | |
c5de0036 PB |
5392 | upd->new_user_data = READ_ONCE(sqe->off); |
5393 | if (!upd->update_user_data && upd->new_user_data) | |
5394 | return -EINVAL; | |
5395 | if (upd->update_events) | |
5396 | upd->events = io_poll_parse_events(sqe, flags); | |
5397 | else if (sqe->poll32_events) | |
5398 | return -EINVAL; | |
221c5eb2 | 5399 | |
221c5eb2 JA |
5400 | return 0; |
5401 | } | |
5402 | ||
221c5eb2 JA |
5403 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
5404 | void *key) | |
5405 | { | |
c2f2eb7d JA |
5406 | struct io_kiocb *req = wait->private; |
5407 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 5408 | |
d7718a9d | 5409 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
5410 | } |
5411 | ||
221c5eb2 JA |
5412 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
5413 | struct poll_table_struct *p) | |
5414 | { | |
5415 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
5416 | ||
e8c2bc1f | 5417 | __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data); |
eac406c6 JA |
5418 | } |
5419 | ||
3529d8c2 | 5420 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
5421 | { |
5422 | struct io_poll_iocb *poll = &req->poll; | |
c5de0036 | 5423 | u32 flags; |
221c5eb2 JA |
5424 | |
5425 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
5426 | return -EINVAL; | |
c5de0036 | 5427 | if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr) |
88e41cf9 JA |
5428 | return -EINVAL; |
5429 | flags = READ_ONCE(sqe->len); | |
c5de0036 | 5430 | if (flags & ~IORING_POLL_ADD_MULTI) |
221c5eb2 JA |
5431 | return -EINVAL; |
5432 | ||
48dcd38d | 5433 | io_req_set_refcount(req); |
c5de0036 | 5434 | poll->events = io_poll_parse_events(sqe, flags); |
0969e783 JA |
5435 | return 0; |
5436 | } | |
5437 | ||
61e98203 | 5438 | static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) |
0969e783 JA |
5439 | { |
5440 | struct io_poll_iocb *poll = &req->poll; | |
5441 | struct io_ring_ctx *ctx = req->ctx; | |
5442 | struct io_poll_table ipt; | |
0969e783 | 5443 | __poll_t mask; |
0969e783 | 5444 | |
d7718a9d | 5445 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 5446 | |
d7718a9d JA |
5447 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
5448 | io_poll_wake); | |
221c5eb2 | 5449 | |
8c838788 | 5450 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 5451 | ipt.error = 0; |
e27414be | 5452 | io_poll_complete(req, mask); |
221c5eb2 | 5453 | } |
79ebeaee | 5454 | spin_unlock(&ctx->completion_lock); |
221c5eb2 | 5455 | |
8c838788 JA |
5456 | if (mask) { |
5457 | io_cqring_ev_posted(ctx); | |
88e41cf9 JA |
5458 | if (poll->events & EPOLLONESHOT) |
5459 | io_put_req(req); | |
221c5eb2 | 5460 | } |
8c838788 | 5461 | return ipt.error; |
221c5eb2 JA |
5462 | } |
5463 | ||
c5de0036 | 5464 | static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags) |
b69de288 JA |
5465 | { |
5466 | struct io_ring_ctx *ctx = req->ctx; | |
5467 | struct io_kiocb *preq; | |
cb3b200e | 5468 | bool completing; |
b69de288 JA |
5469 | int ret; |
5470 | ||
79ebeaee | 5471 | spin_lock(&ctx->completion_lock); |
9ba5fac8 | 5472 | preq = io_poll_find(ctx, req->poll_update.old_user_data, true); |
b69de288 JA |
5473 | if (!preq) { |
5474 | ret = -ENOENT; | |
5475 | goto err; | |
b69de288 | 5476 | } |
cb3b200e | 5477 | |
c5de0036 PB |
5478 | if (!req->poll_update.update_events && !req->poll_update.update_user_data) { |
5479 | completing = true; | |
5480 | ret = io_poll_remove_one(preq) ? 0 : -EALREADY; | |
5481 | goto err; | |
5482 | } | |
5483 | ||
cb3b200e JA |
5484 | /* |
5485 | * Don't allow racy completion with singleshot, as we cannot safely | |
5486 | * update those. For multishot, if we're racing with completion, just | |
5487 | * let completion re-add it. | |
5488 | */ | |
5489 | completing = !__io_poll_remove_one(preq, &preq->poll, false); | |
5490 | if (completing && (preq->poll.events & EPOLLONESHOT)) { | |
5491 | ret = -EALREADY; | |
5492 | goto err; | |
b69de288 JA |
5493 | } |
5494 | /* we now have a detached poll request. reissue. */ | |
5495 | ret = 0; | |
5496 | err: | |
b69de288 | 5497 | if (ret < 0) { |
79ebeaee | 5498 | spin_unlock(&ctx->completion_lock); |
93d2bcd2 | 5499 | req_set_fail(req); |
b69de288 JA |
5500 | io_req_complete(req, ret); |
5501 | return 0; | |
5502 | } | |
5503 | /* only mask one event flags, keep behavior flags */ | |
9d805892 | 5504 | if (req->poll_update.update_events) { |
b69de288 | 5505 | preq->poll.events &= ~0xffff; |
9d805892 | 5506 | preq->poll.events |= req->poll_update.events & 0xffff; |
b69de288 JA |
5507 | preq->poll.events |= IO_POLL_UNMASK; |
5508 | } | |
9d805892 PB |
5509 | if (req->poll_update.update_user_data) |
5510 | preq->user_data = req->poll_update.new_user_data; | |
79ebeaee | 5511 | spin_unlock(&ctx->completion_lock); |
cb3b200e | 5512 | |
b69de288 JA |
5513 | /* complete update request, we're done with it */ |
5514 | io_req_complete(req, ret); | |
5515 | ||
cb3b200e | 5516 | if (!completing) { |
c5de0036 | 5517 | ret = io_poll_add(preq, issue_flags); |
cb3b200e | 5518 | if (ret < 0) { |
93d2bcd2 | 5519 | req_set_fail(preq); |
cb3b200e JA |
5520 | io_req_complete(preq, ret); |
5521 | } | |
b69de288 JA |
5522 | } |
5523 | return 0; | |
5524 | } | |
5525 | ||
89850fce JA |
5526 | static void io_req_task_timeout(struct io_kiocb *req) |
5527 | { | |
89850fce | 5528 | req_set_fail(req); |
505657bc | 5529 | io_req_complete_post(req, -ETIME, 0); |
89850fce JA |
5530 | } |
5531 | ||
5262f567 JA |
5532 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
5533 | { | |
ad8a48ac JA |
5534 | struct io_timeout_data *data = container_of(timer, |
5535 | struct io_timeout_data, timer); | |
5536 | struct io_kiocb *req = data->req; | |
5537 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
5538 | unsigned long flags; |
5539 | ||
89850fce | 5540 | spin_lock_irqsave(&ctx->timeout_lock, flags); |
a71976f3 | 5541 | list_del_init(&req->timeout.list); |
01cec8c1 PB |
5542 | atomic_set(&req->ctx->cq_timeouts, |
5543 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
89850fce | 5544 | spin_unlock_irqrestore(&ctx->timeout_lock, flags); |
01cec8c1 | 5545 | |
89850fce JA |
5546 | req->io_task_work.func = io_req_task_timeout; |
5547 | io_req_task_work_add(req); | |
5262f567 JA |
5548 | return HRTIMER_NORESTART; |
5549 | } | |
5550 | ||
fbd15848 PB |
5551 | static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx, |
5552 | __u64 user_data) | |
89850fce | 5553 | __must_hold(&ctx->timeout_lock) |
f254ac04 | 5554 | { |
fbd15848 | 5555 | struct io_timeout_data *io; |
47f46768 | 5556 | struct io_kiocb *req; |
fd9c7bc5 | 5557 | bool found = false; |
f254ac04 | 5558 | |
135fcde8 | 5559 | list_for_each_entry(req, &ctx->timeout_list, timeout.list) { |
fd9c7bc5 PB |
5560 | found = user_data == req->user_data; |
5561 | if (found) | |
47f46768 | 5562 | break; |
47f46768 | 5563 | } |
fd9c7bc5 PB |
5564 | if (!found) |
5565 | return ERR_PTR(-ENOENT); | |
fbd15848 PB |
5566 | |
5567 | io = req->async_data; | |
fd9c7bc5 | 5568 | if (hrtimer_try_to_cancel(&io->timer) == -1) |
fbd15848 | 5569 | return ERR_PTR(-EALREADY); |
a71976f3 | 5570 | list_del_init(&req->timeout.list); |
fbd15848 PB |
5571 | return req; |
5572 | } | |
47f46768 | 5573 | |
fbd15848 | 5574 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
89850fce | 5575 | __must_hold(&ctx->timeout_lock) |
fbd15848 PB |
5576 | { |
5577 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); | |
5578 | ||
5579 | if (IS_ERR(req)) | |
5580 | return PTR_ERR(req); | |
f254ac04 | 5581 | |
93d2bcd2 | 5582 | req_set_fail(req); |
d4d19c19 | 5583 | io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0); |
91c2f697 | 5584 | io_put_req_deferred(req); |
f254ac04 JA |
5585 | return 0; |
5586 | } | |
5587 | ||
9c8e11b3 PB |
5588 | static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data, |
5589 | struct timespec64 *ts, enum hrtimer_mode mode) | |
89850fce | 5590 | __must_hold(&ctx->timeout_lock) |
47f46768 | 5591 | { |
9c8e11b3 PB |
5592 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); |
5593 | struct io_timeout_data *data; | |
47f46768 | 5594 | |
9c8e11b3 PB |
5595 | if (IS_ERR(req)) |
5596 | return PTR_ERR(req); | |
47f46768 | 5597 | |
9c8e11b3 PB |
5598 | req->timeout.off = 0; /* noseq */ |
5599 | data = req->async_data; | |
5600 | list_add_tail(&req->timeout.list, &ctx->timeout_list); | |
5601 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode); | |
5602 | data->timer.function = io_timeout_fn; | |
5603 | hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode); | |
5604 | return 0; | |
47f46768 JA |
5605 | } |
5606 | ||
3529d8c2 JA |
5607 | static int io_timeout_remove_prep(struct io_kiocb *req, |
5608 | const struct io_uring_sqe *sqe) | |
b29472ee | 5609 | { |
9c8e11b3 PB |
5610 | struct io_timeout_rem *tr = &req->timeout_rem; |
5611 | ||
b29472ee JA |
5612 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5613 | return -EINVAL; | |
61710e43 DA |
5614 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5615 | return -EINVAL; | |
9c8e11b3 | 5616 | if (sqe->ioprio || sqe->buf_index || sqe->len) |
b29472ee JA |
5617 | return -EINVAL; |
5618 | ||
9c8e11b3 PB |
5619 | tr->addr = READ_ONCE(sqe->addr); |
5620 | tr->flags = READ_ONCE(sqe->timeout_flags); | |
5621 | if (tr->flags & IORING_TIMEOUT_UPDATE) { | |
5622 | if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS)) | |
5623 | return -EINVAL; | |
5624 | if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2))) | |
5625 | return -EFAULT; | |
5626 | } else if (tr->flags) { | |
5627 | /* timeout removal doesn't support flags */ | |
b29472ee | 5628 | return -EINVAL; |
9c8e11b3 | 5629 | } |
b29472ee | 5630 | |
b29472ee JA |
5631 | return 0; |
5632 | } | |
5633 | ||
8662daec PB |
5634 | static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags) |
5635 | { | |
5636 | return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS | |
5637 | : HRTIMER_MODE_REL; | |
5638 | } | |
5639 | ||
11365043 JA |
5640 | /* |
5641 | * Remove or update an existing timeout command | |
5642 | */ | |
61e98203 | 5643 | static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags) |
11365043 | 5644 | { |
9c8e11b3 | 5645 | struct io_timeout_rem *tr = &req->timeout_rem; |
11365043 | 5646 | struct io_ring_ctx *ctx = req->ctx; |
47f46768 | 5647 | int ret; |
11365043 | 5648 | |
89850fce | 5649 | spin_lock_irq(&ctx->timeout_lock); |
8662daec | 5650 | if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) |
9c8e11b3 | 5651 | ret = io_timeout_cancel(ctx, tr->addr); |
8662daec PB |
5652 | else |
5653 | ret = io_timeout_update(ctx, tr->addr, &tr->ts, | |
5654 | io_translate_timeout_mode(tr->flags)); | |
89850fce | 5655 | spin_unlock_irq(&ctx->timeout_lock); |
11365043 | 5656 | |
4e88d6e7 | 5657 | if (ret < 0) |
93d2bcd2 | 5658 | req_set_fail(req); |
505657bc | 5659 | io_req_complete_post(req, ret, 0); |
11365043 | 5660 | return 0; |
5262f567 JA |
5661 | } |
5662 | ||
3529d8c2 | 5663 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 5664 | bool is_timeout_link) |
5262f567 | 5665 | { |
ad8a48ac | 5666 | struct io_timeout_data *data; |
a41525ab | 5667 | unsigned flags; |
56080b02 | 5668 | u32 off = READ_ONCE(sqe->off); |
5262f567 | 5669 | |
ad8a48ac | 5670 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 5671 | return -EINVAL; |
ad8a48ac | 5672 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 5673 | return -EINVAL; |
56080b02 | 5674 | if (off && is_timeout_link) |
2d28390a | 5675 | return -EINVAL; |
a41525ab JA |
5676 | flags = READ_ONCE(sqe->timeout_flags); |
5677 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 5678 | return -EINVAL; |
bdf20073 | 5679 | |
bfe68a22 | 5680 | req->timeout.off = off; |
f18ee4cf PB |
5681 | if (unlikely(off && !req->ctx->off_timeout_used)) |
5682 | req->ctx->off_timeout_used = true; | |
26a61679 | 5683 | |
e8c2bc1f | 5684 | if (!req->async_data && io_alloc_async_data(req)) |
26a61679 JA |
5685 | return -ENOMEM; |
5686 | ||
e8c2bc1f | 5687 | data = req->async_data; |
ad8a48ac | 5688 | data->req = req; |
ad8a48ac JA |
5689 | |
5690 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
5691 | return -EFAULT; |
5692 | ||
8662daec | 5693 | data->mode = io_translate_timeout_mode(flags); |
ad8a48ac | 5694 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
b97e736a PB |
5695 | |
5696 | if (is_timeout_link) { | |
5697 | struct io_submit_link *link = &req->ctx->submit_state.link; | |
5698 | ||
5699 | if (!link->head) | |
5700 | return -EINVAL; | |
5701 | if (link->last->opcode == IORING_OP_LINK_TIMEOUT) | |
5702 | return -EINVAL; | |
4d13d1a4 PB |
5703 | req->timeout.head = link->last; |
5704 | link->last->flags |= REQ_F_ARM_LTIMEOUT; | |
b97e736a | 5705 | } |
ad8a48ac JA |
5706 | return 0; |
5707 | } | |
5708 | ||
61e98203 | 5709 | static int io_timeout(struct io_kiocb *req, unsigned int issue_flags) |
ad8a48ac | 5710 | { |
ad8a48ac | 5711 | struct io_ring_ctx *ctx = req->ctx; |
e8c2bc1f | 5712 | struct io_timeout_data *data = req->async_data; |
ad8a48ac | 5713 | struct list_head *entry; |
bfe68a22 | 5714 | u32 tail, off = req->timeout.off; |
ad8a48ac | 5715 | |
89850fce | 5716 | spin_lock_irq(&ctx->timeout_lock); |
93bd25bb | 5717 | |
5262f567 JA |
5718 | /* |
5719 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
5720 | * timeout event to be satisfied. If it isn't set, then this is |
5721 | * a pure timeout request, sequence isn't used. | |
5262f567 | 5722 | */ |
8eb7e2d0 | 5723 | if (io_is_timeout_noseq(req)) { |
93bd25bb JA |
5724 | entry = ctx->timeout_list.prev; |
5725 | goto add; | |
5726 | } | |
5262f567 | 5727 | |
bfe68a22 PB |
5728 | tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
5729 | req->timeout.target_seq = tail + off; | |
5262f567 | 5730 | |
f010505b MDG |
5731 | /* Update the last seq here in case io_flush_timeouts() hasn't. |
5732 | * This is safe because ->completion_lock is held, and submissions | |
5733 | * and completions are never mixed in the same ->completion_lock section. | |
5734 | */ | |
5735 | ctx->cq_last_tm_flush = tail; | |
5736 | ||
5262f567 JA |
5737 | /* |
5738 | * Insertion sort, ensuring the first entry in the list is always | |
5739 | * the one we need first. | |
5740 | */ | |
5262f567 | 5741 | list_for_each_prev(entry, &ctx->timeout_list) { |
135fcde8 PB |
5742 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, |
5743 | timeout.list); | |
5262f567 | 5744 | |
8eb7e2d0 | 5745 | if (io_is_timeout_noseq(nxt)) |
93bd25bb | 5746 | continue; |
bfe68a22 PB |
5747 | /* nxt.seq is behind @tail, otherwise would've been completed */ |
5748 | if (off >= nxt->timeout.target_seq - tail) | |
5262f567 JA |
5749 | break; |
5750 | } | |
93bd25bb | 5751 | add: |
135fcde8 | 5752 | list_add(&req->timeout.list, entry); |
ad8a48ac JA |
5753 | data->timer.function = io_timeout_fn; |
5754 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
89850fce | 5755 | spin_unlock_irq(&ctx->timeout_lock); |
5262f567 JA |
5756 | return 0; |
5757 | } | |
5262f567 | 5758 | |
f458dd84 PB |
5759 | struct io_cancel_data { |
5760 | struct io_ring_ctx *ctx; | |
5761 | u64 user_data; | |
5762 | }; | |
5763 | ||
62755e35 JA |
5764 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
5765 | { | |
5766 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
f458dd84 | 5767 | struct io_cancel_data *cd = data; |
62755e35 | 5768 | |
f458dd84 | 5769 | return req->ctx == cd->ctx && req->user_data == cd->user_data; |
62755e35 JA |
5770 | } |
5771 | ||
f458dd84 PB |
5772 | static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data, |
5773 | struct io_ring_ctx *ctx) | |
62755e35 | 5774 | { |
f458dd84 | 5775 | struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, }; |
62755e35 | 5776 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
5777 | int ret = 0; |
5778 | ||
f458dd84 | 5779 | if (!tctx || !tctx->io_wq) |
5aa75ed5 JA |
5780 | return -ENOENT; |
5781 | ||
f458dd84 | 5782 | cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false); |
62755e35 JA |
5783 | switch (cancel_ret) { |
5784 | case IO_WQ_CANCEL_OK: | |
5785 | ret = 0; | |
5786 | break; | |
5787 | case IO_WQ_CANCEL_RUNNING: | |
5788 | ret = -EALREADY; | |
5789 | break; | |
5790 | case IO_WQ_CANCEL_NOTFOUND: | |
5791 | ret = -ENOENT; | |
5792 | break; | |
5793 | } | |
5794 | ||
e977d6d3 JA |
5795 | return ret; |
5796 | } | |
5797 | ||
8cb01fac | 5798 | static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr) |
47f46768 | 5799 | { |
8cb01fac | 5800 | struct io_ring_ctx *ctx = req->ctx; |
47f46768 JA |
5801 | int ret; |
5802 | ||
8cb01fac PB |
5803 | WARN_ON_ONCE(req->task != current); |
5804 | ||
f458dd84 | 5805 | ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx); |
df9727af | 5806 | if (ret != -ENOENT) |
8cb01fac | 5807 | return ret; |
505657bc PB |
5808 | |
5809 | spin_lock(&ctx->completion_lock); | |
79ebeaee | 5810 | spin_lock_irq(&ctx->timeout_lock); |
47f46768 | 5811 | ret = io_timeout_cancel(ctx, sqe_addr); |
79ebeaee | 5812 | spin_unlock_irq(&ctx->timeout_lock); |
47f46768 | 5813 | if (ret != -ENOENT) |
505657bc PB |
5814 | goto out; |
5815 | ret = io_poll_cancel(ctx, sqe_addr, false); | |
5816 | out: | |
5817 | spin_unlock(&ctx->completion_lock); | |
5818 | return ret; | |
47f46768 JA |
5819 | } |
5820 | ||
3529d8c2 JA |
5821 | static int io_async_cancel_prep(struct io_kiocb *req, |
5822 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 5823 | { |
fbf23849 | 5824 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 | 5825 | return -EINVAL; |
61710e43 DA |
5826 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5827 | return -EINVAL; | |
5828 | if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags) | |
e977d6d3 JA |
5829 | return -EINVAL; |
5830 | ||
fbf23849 JA |
5831 | req->cancel.addr = READ_ONCE(sqe->addr); |
5832 | return 0; | |
5833 | } | |
5834 | ||
61e98203 | 5835 | static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags) |
fbf23849 JA |
5836 | { |
5837 | struct io_ring_ctx *ctx = req->ctx; | |
58f99373 PB |
5838 | u64 sqe_addr = req->cancel.addr; |
5839 | struct io_tctx_node *node; | |
5840 | int ret; | |
5841 | ||
8cb01fac | 5842 | ret = io_try_cancel_userdata(req, sqe_addr); |
58f99373 PB |
5843 | if (ret != -ENOENT) |
5844 | goto done; | |
58f99373 PB |
5845 | |
5846 | /* slow path, try all io-wq's */ | |
5847 | io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK)); | |
5848 | ret = -ENOENT; | |
5849 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
5850 | struct io_uring_task *tctx = node->task->io_uring; | |
fbf23849 | 5851 | |
58f99373 PB |
5852 | ret = io_async_cancel_one(tctx, req->cancel.addr, ctx); |
5853 | if (ret != -ENOENT) | |
5854 | break; | |
5855 | } | |
5856 | io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK)); | |
58f99373 | 5857 | done: |
58f99373 | 5858 | if (ret < 0) |
93d2bcd2 | 5859 | req_set_fail(req); |
505657bc | 5860 | io_req_complete_post(req, ret, 0); |
5262f567 JA |
5861 | return 0; |
5862 | } | |
5863 | ||
269bbe5f | 5864 | static int io_rsrc_update_prep(struct io_kiocb *req, |
05f3fb3c JA |
5865 | const struct io_uring_sqe *sqe) |
5866 | { | |
61710e43 DA |
5867 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5868 | return -EINVAL; | |
5869 | if (sqe->ioprio || sqe->rw_flags) | |
05f3fb3c JA |
5870 | return -EINVAL; |
5871 | ||
269bbe5f BM |
5872 | req->rsrc_update.offset = READ_ONCE(sqe->off); |
5873 | req->rsrc_update.nr_args = READ_ONCE(sqe->len); | |
5874 | if (!req->rsrc_update.nr_args) | |
05f3fb3c | 5875 | return -EINVAL; |
269bbe5f | 5876 | req->rsrc_update.arg = READ_ONCE(sqe->addr); |
05f3fb3c JA |
5877 | return 0; |
5878 | } | |
5879 | ||
889fca73 | 5880 | static int io_files_update(struct io_kiocb *req, unsigned int issue_flags) |
fbf23849 JA |
5881 | { |
5882 | struct io_ring_ctx *ctx = req->ctx; | |
c3bdad02 | 5883 | struct io_uring_rsrc_update2 up; |
05f3fb3c | 5884 | int ret; |
fbf23849 | 5885 | |
45d189c6 | 5886 | if (issue_flags & IO_URING_F_NONBLOCK) |
05f3fb3c | 5887 | return -EAGAIN; |
05f3fb3c | 5888 | |
269bbe5f BM |
5889 | up.offset = req->rsrc_update.offset; |
5890 | up.data = req->rsrc_update.arg; | |
c3bdad02 PB |
5891 | up.nr = 0; |
5892 | up.tags = 0; | |
615cee49 | 5893 | up.resv = 0; |
05f3fb3c JA |
5894 | |
5895 | mutex_lock(&ctx->uring_lock); | |
fdecb662 | 5896 | ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, |
98f0b3b4 | 5897 | &up, req->rsrc_update.nr_args); |
05f3fb3c JA |
5898 | mutex_unlock(&ctx->uring_lock); |
5899 | ||
5900 | if (ret < 0) | |
93d2bcd2 | 5901 | req_set_fail(req); |
889fca73 | 5902 | __io_req_complete(req, issue_flags, ret, 0); |
5262f567 JA |
5903 | return 0; |
5904 | } | |
5905 | ||
bfe76559 | 5906 | static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 5907 | { |
d625c6ee | 5908 | switch (req->opcode) { |
e781573e | 5909 | case IORING_OP_NOP: |
bfe76559 | 5910 | return 0; |
f67676d1 JA |
5911 | case IORING_OP_READV: |
5912 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 5913 | case IORING_OP_READ: |
bfe76559 | 5914 | return io_read_prep(req, sqe); |
f67676d1 JA |
5915 | case IORING_OP_WRITEV: |
5916 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 5917 | case IORING_OP_WRITE: |
bfe76559 | 5918 | return io_write_prep(req, sqe); |
0969e783 | 5919 | case IORING_OP_POLL_ADD: |
bfe76559 | 5920 | return io_poll_add_prep(req, sqe); |
0969e783 | 5921 | case IORING_OP_POLL_REMOVE: |
c5de0036 | 5922 | return io_poll_update_prep(req, sqe); |
8ed8d3c3 | 5923 | case IORING_OP_FSYNC: |
1155c76a | 5924 | return io_fsync_prep(req, sqe); |
8ed8d3c3 | 5925 | case IORING_OP_SYNC_FILE_RANGE: |
1155c76a | 5926 | return io_sfr_prep(req, sqe); |
03b1230c | 5927 | case IORING_OP_SENDMSG: |
fddaface | 5928 | case IORING_OP_SEND: |
bfe76559 | 5929 | return io_sendmsg_prep(req, sqe); |
03b1230c | 5930 | case IORING_OP_RECVMSG: |
fddaface | 5931 | case IORING_OP_RECV: |
bfe76559 | 5932 | return io_recvmsg_prep(req, sqe); |
f499a021 | 5933 | case IORING_OP_CONNECT: |
bfe76559 | 5934 | return io_connect_prep(req, sqe); |
2d28390a | 5935 | case IORING_OP_TIMEOUT: |
bfe76559 | 5936 | return io_timeout_prep(req, sqe, false); |
b29472ee | 5937 | case IORING_OP_TIMEOUT_REMOVE: |
bfe76559 | 5938 | return io_timeout_remove_prep(req, sqe); |
fbf23849 | 5939 | case IORING_OP_ASYNC_CANCEL: |
bfe76559 | 5940 | return io_async_cancel_prep(req, sqe); |
2d28390a | 5941 | case IORING_OP_LINK_TIMEOUT: |
bfe76559 | 5942 | return io_timeout_prep(req, sqe, true); |
8ed8d3c3 | 5943 | case IORING_OP_ACCEPT: |
bfe76559 | 5944 | return io_accept_prep(req, sqe); |
d63d1b5e | 5945 | case IORING_OP_FALLOCATE: |
bfe76559 | 5946 | return io_fallocate_prep(req, sqe); |
15b71abe | 5947 | case IORING_OP_OPENAT: |
bfe76559 | 5948 | return io_openat_prep(req, sqe); |
b5dba59e | 5949 | case IORING_OP_CLOSE: |
bfe76559 | 5950 | return io_close_prep(req, sqe); |
05f3fb3c | 5951 | case IORING_OP_FILES_UPDATE: |
269bbe5f | 5952 | return io_rsrc_update_prep(req, sqe); |
eddc7ef5 | 5953 | case IORING_OP_STATX: |
bfe76559 | 5954 | return io_statx_prep(req, sqe); |
4840e418 | 5955 | case IORING_OP_FADVISE: |
bfe76559 | 5956 | return io_fadvise_prep(req, sqe); |
c1ca757b | 5957 | case IORING_OP_MADVISE: |
bfe76559 | 5958 | return io_madvise_prep(req, sqe); |
cebdb986 | 5959 | case IORING_OP_OPENAT2: |
bfe76559 | 5960 | return io_openat2_prep(req, sqe); |
3e4827b0 | 5961 | case IORING_OP_EPOLL_CTL: |
bfe76559 | 5962 | return io_epoll_ctl_prep(req, sqe); |
7d67af2c | 5963 | case IORING_OP_SPLICE: |
bfe76559 | 5964 | return io_splice_prep(req, sqe); |
ddf0322d | 5965 | case IORING_OP_PROVIDE_BUFFERS: |
bfe76559 | 5966 | return io_provide_buffers_prep(req, sqe); |
067524e9 | 5967 | case IORING_OP_REMOVE_BUFFERS: |
bfe76559 | 5968 | return io_remove_buffers_prep(req, sqe); |
f2a8d5c7 | 5969 | case IORING_OP_TEE: |
bfe76559 | 5970 | return io_tee_prep(req, sqe); |
36f4fa68 JA |
5971 | case IORING_OP_SHUTDOWN: |
5972 | return io_shutdown_prep(req, sqe); | |
80a261fd JA |
5973 | case IORING_OP_RENAMEAT: |
5974 | return io_renameat_prep(req, sqe); | |
14a1143b JA |
5975 | case IORING_OP_UNLINKAT: |
5976 | return io_unlinkat_prep(req, sqe); | |
f67676d1 JA |
5977 | } |
5978 | ||
bfe76559 PB |
5979 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
5980 | req->opcode); | |
bd54b6fe | 5981 | return -EINVAL; |
bfe76559 PB |
5982 | } |
5983 | ||
93642ef8 | 5984 | static int io_req_prep_async(struct io_kiocb *req) |
bfe76559 | 5985 | { |
b7e298d2 PB |
5986 | if (!io_op_defs[req->opcode].needs_async_setup) |
5987 | return 0; | |
5988 | if (WARN_ON_ONCE(req->async_data)) | |
5989 | return -EFAULT; | |
5990 | if (io_alloc_async_data(req)) | |
5991 | return -EAGAIN; | |
5992 | ||
93642ef8 PB |
5993 | switch (req->opcode) { |
5994 | case IORING_OP_READV: | |
93642ef8 PB |
5995 | return io_rw_prep_async(req, READ); |
5996 | case IORING_OP_WRITEV: | |
93642ef8 PB |
5997 | return io_rw_prep_async(req, WRITE); |
5998 | case IORING_OP_SENDMSG: | |
93642ef8 PB |
5999 | return io_sendmsg_prep_async(req); |
6000 | case IORING_OP_RECVMSG: | |
93642ef8 PB |
6001 | return io_recvmsg_prep_async(req); |
6002 | case IORING_OP_CONNECT: | |
6003 | return io_connect_prep_async(req); | |
6004 | } | |
b7e298d2 PB |
6005 | printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n", |
6006 | req->opcode); | |
6007 | return -EFAULT; | |
f67676d1 JA |
6008 | } |
6009 | ||
9cf7c104 PB |
6010 | static u32 io_get_sequence(struct io_kiocb *req) |
6011 | { | |
a3dbdf54 | 6012 | u32 seq = req->ctx->cached_sq_head; |
9cf7c104 | 6013 | |
a3dbdf54 PB |
6014 | /* need original cached_sq_head, but it was increased for each req */ |
6015 | io_for_each_link(req, req) | |
6016 | seq--; | |
6017 | return seq; | |
9cf7c104 PB |
6018 | } |
6019 | ||
76cc33d7 | 6020 | static bool io_drain_req(struct io_kiocb *req) |
de0617e4 | 6021 | { |
3c19966d | 6022 | struct io_kiocb *pos; |
a197f664 | 6023 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 6024 | struct io_defer_entry *de; |
f67676d1 | 6025 | int ret; |
9cf7c104 | 6026 | u32 seq; |
de0617e4 | 6027 | |
3c19966d PB |
6028 | /* |
6029 | * If we need to drain a request in the middle of a link, drain the | |
6030 | * head request and the next request/link after the current link. | |
6031 | * Considering sequential execution of links, IOSQE_IO_DRAIN will be | |
6032 | * maintained for every request of our link. | |
6033 | */ | |
6034 | if (ctx->drain_next) { | |
6035 | req->flags |= REQ_F_IO_DRAIN; | |
6036 | ctx->drain_next = false; | |
6037 | } | |
6038 | /* not interested in head, start from the first linked */ | |
6039 | io_for_each_link(pos, req->link) { | |
6040 | if (pos->flags & REQ_F_IO_DRAIN) { | |
6041 | ctx->drain_next = true; | |
6042 | req->flags |= REQ_F_IO_DRAIN; | |
6043 | break; | |
6044 | } | |
6045 | } | |
6046 | ||
9d858b21 | 6047 | /* Still need defer if there is pending req in defer list. */ |
9cf7c104 | 6048 | if (likely(list_empty_careful(&ctx->defer_list) && |
10c66904 PB |
6049 | !(req->flags & REQ_F_IO_DRAIN))) { |
6050 | ctx->drain_active = false; | |
76cc33d7 | 6051 | return false; |
10c66904 | 6052 | } |
9cf7c104 PB |
6053 | |
6054 | seq = io_get_sequence(req); | |
6055 | /* Still a chance to pass the sequence check */ | |
6056 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) | |
76cc33d7 | 6057 | return false; |
de0617e4 | 6058 | |
b7e298d2 | 6059 | ret = io_req_prep_async(req); |
be7053b7 | 6060 | if (ret) |
1b48773f | 6061 | goto fail; |
cbdcb435 | 6062 | io_prep_async_link(req); |
27dc8338 | 6063 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 6064 | if (!de) { |
1b48773f PB |
6065 | ret = -ENOMEM; |
6066 | fail: | |
6067 | io_req_complete_failed(req, ret); | |
76cc33d7 PB |
6068 | return true; |
6069 | } | |
2d28390a | 6070 | |
79ebeaee | 6071 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 6072 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 6073 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 6074 | kfree(de); |
ae34817b | 6075 | io_queue_async_work(req); |
76cc33d7 | 6076 | return true; |
de0617e4 JA |
6077 | } |
6078 | ||
915967f6 | 6079 | trace_io_uring_defer(ctx, req, req->user_data); |
27dc8338 | 6080 | de->req = req; |
9cf7c104 | 6081 | de->seq = seq; |
27dc8338 | 6082 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 6083 | spin_unlock(&ctx->completion_lock); |
76cc33d7 | 6084 | return true; |
de0617e4 JA |
6085 | } |
6086 | ||
68fb8979 | 6087 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 6088 | { |
0e1b6fe3 PB |
6089 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
6090 | switch (req->opcode) { | |
6091 | case IORING_OP_READV: | |
6092 | case IORING_OP_READ_FIXED: | |
6093 | case IORING_OP_READ: | |
bcda7baa | 6094 | kfree((void *)(unsigned long)req->rw.addr); |
0e1b6fe3 PB |
6095 | break; |
6096 | case IORING_OP_RECVMSG: | |
6097 | case IORING_OP_RECV: | |
bcda7baa | 6098 | kfree(req->sr_msg.kbuf); |
0e1b6fe3 PB |
6099 | break; |
6100 | } | |
99bc4c38 PB |
6101 | } |
6102 | ||
0e1b6fe3 PB |
6103 | if (req->flags & REQ_F_NEED_CLEANUP) { |
6104 | switch (req->opcode) { | |
6105 | case IORING_OP_READV: | |
6106 | case IORING_OP_READ_FIXED: | |
6107 | case IORING_OP_READ: | |
6108 | case IORING_OP_WRITEV: | |
6109 | case IORING_OP_WRITE_FIXED: | |
e8c2bc1f JA |
6110 | case IORING_OP_WRITE: { |
6111 | struct io_async_rw *io = req->async_data; | |
1dacb4df PB |
6112 | |
6113 | kfree(io->free_iovec); | |
0e1b6fe3 | 6114 | break; |
e8c2bc1f | 6115 | } |
0e1b6fe3 | 6116 | case IORING_OP_RECVMSG: |
e8c2bc1f JA |
6117 | case IORING_OP_SENDMSG: { |
6118 | struct io_async_msghdr *io = req->async_data; | |
257e84a5 PB |
6119 | |
6120 | kfree(io->free_iov); | |
0e1b6fe3 | 6121 | break; |
e8c2bc1f | 6122 | } |
0e1b6fe3 PB |
6123 | case IORING_OP_SPLICE: |
6124 | case IORING_OP_TEE: | |
e1d767f0 PB |
6125 | if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED)) |
6126 | io_put_file(req->splice.file_in); | |
0e1b6fe3 | 6127 | break; |
f3cd4850 JA |
6128 | case IORING_OP_OPENAT: |
6129 | case IORING_OP_OPENAT2: | |
6130 | if (req->open.filename) | |
6131 | putname(req->open.filename); | |
6132 | break; | |
80a261fd JA |
6133 | case IORING_OP_RENAMEAT: |
6134 | putname(req->rename.oldpath); | |
6135 | putname(req->rename.newpath); | |
6136 | break; | |
14a1143b JA |
6137 | case IORING_OP_UNLINKAT: |
6138 | putname(req->unlink.filename); | |
6139 | break; | |
0e1b6fe3 | 6140 | } |
99bc4c38 | 6141 | } |
75652a30 JA |
6142 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
6143 | kfree(req->apoll->double_poll); | |
6144 | kfree(req->apoll); | |
6145 | req->apoll = NULL; | |
6146 | } | |
3a0a6902 PB |
6147 | if (req->flags & REQ_F_INFLIGHT) { |
6148 | struct io_uring_task *tctx = req->task->io_uring; | |
6149 | ||
6150 | atomic_dec(&tctx->inflight_tracked); | |
3a0a6902 | 6151 | } |
c854357b | 6152 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 6153 | put_cred(req->creds); |
c854357b PB |
6154 | |
6155 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
99bc4c38 PB |
6156 | } |
6157 | ||
889fca73 | 6158 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 6159 | { |
a197f664 | 6160 | struct io_ring_ctx *ctx = req->ctx; |
5730b27e | 6161 | const struct cred *creds = NULL; |
d625c6ee | 6162 | int ret; |
2b188cc1 | 6163 | |
b8e64b53 | 6164 | if ((req->flags & REQ_F_CREDS) && req->creds != current_cred()) |
c10d1f98 | 6165 | creds = override_creds(req->creds); |
5730b27e | 6166 | |
d625c6ee | 6167 | switch (req->opcode) { |
2b188cc1 | 6168 | case IORING_OP_NOP: |
889fca73 | 6169 | ret = io_nop(req, issue_flags); |
2b188cc1 JA |
6170 | break; |
6171 | case IORING_OP_READV: | |
edafccee | 6172 | case IORING_OP_READ_FIXED: |
3a6820f2 | 6173 | case IORING_OP_READ: |
889fca73 | 6174 | ret = io_read(req, issue_flags); |
edafccee | 6175 | break; |
3529d8c2 | 6176 | case IORING_OP_WRITEV: |
edafccee | 6177 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 6178 | case IORING_OP_WRITE: |
889fca73 | 6179 | ret = io_write(req, issue_flags); |
2b188cc1 | 6180 | break; |
c992fe29 | 6181 | case IORING_OP_FSYNC: |
45d189c6 | 6182 | ret = io_fsync(req, issue_flags); |
c992fe29 | 6183 | break; |
221c5eb2 | 6184 | case IORING_OP_POLL_ADD: |
61e98203 | 6185 | ret = io_poll_add(req, issue_flags); |
221c5eb2 JA |
6186 | break; |
6187 | case IORING_OP_POLL_REMOVE: | |
c5de0036 | 6188 | ret = io_poll_update(req, issue_flags); |
221c5eb2 | 6189 | break; |
5d17b4a4 | 6190 | case IORING_OP_SYNC_FILE_RANGE: |
45d189c6 | 6191 | ret = io_sync_file_range(req, issue_flags); |
5d17b4a4 | 6192 | break; |
0fa03c62 | 6193 | case IORING_OP_SENDMSG: |
889fca73 | 6194 | ret = io_sendmsg(req, issue_flags); |
062d04d7 | 6195 | break; |
fddaface | 6196 | case IORING_OP_SEND: |
889fca73 | 6197 | ret = io_send(req, issue_flags); |
0fa03c62 | 6198 | break; |
aa1fa28f | 6199 | case IORING_OP_RECVMSG: |
889fca73 | 6200 | ret = io_recvmsg(req, issue_flags); |
062d04d7 | 6201 | break; |
fddaface | 6202 | case IORING_OP_RECV: |
889fca73 | 6203 | ret = io_recv(req, issue_flags); |
aa1fa28f | 6204 | break; |
5262f567 | 6205 | case IORING_OP_TIMEOUT: |
61e98203 | 6206 | ret = io_timeout(req, issue_flags); |
5262f567 | 6207 | break; |
11365043 | 6208 | case IORING_OP_TIMEOUT_REMOVE: |
61e98203 | 6209 | ret = io_timeout_remove(req, issue_flags); |
11365043 | 6210 | break; |
17f2fe35 | 6211 | case IORING_OP_ACCEPT: |
889fca73 | 6212 | ret = io_accept(req, issue_flags); |
17f2fe35 | 6213 | break; |
f8e85cf2 | 6214 | case IORING_OP_CONNECT: |
889fca73 | 6215 | ret = io_connect(req, issue_flags); |
f8e85cf2 | 6216 | break; |
62755e35 | 6217 | case IORING_OP_ASYNC_CANCEL: |
61e98203 | 6218 | ret = io_async_cancel(req, issue_flags); |
62755e35 | 6219 | break; |
d63d1b5e | 6220 | case IORING_OP_FALLOCATE: |
45d189c6 | 6221 | ret = io_fallocate(req, issue_flags); |
d63d1b5e | 6222 | break; |
15b71abe | 6223 | case IORING_OP_OPENAT: |
45d189c6 | 6224 | ret = io_openat(req, issue_flags); |
15b71abe | 6225 | break; |
b5dba59e | 6226 | case IORING_OP_CLOSE: |
889fca73 | 6227 | ret = io_close(req, issue_flags); |
b5dba59e | 6228 | break; |
05f3fb3c | 6229 | case IORING_OP_FILES_UPDATE: |
889fca73 | 6230 | ret = io_files_update(req, issue_flags); |
05f3fb3c | 6231 | break; |
eddc7ef5 | 6232 | case IORING_OP_STATX: |
45d189c6 | 6233 | ret = io_statx(req, issue_flags); |
eddc7ef5 | 6234 | break; |
4840e418 | 6235 | case IORING_OP_FADVISE: |
45d189c6 | 6236 | ret = io_fadvise(req, issue_flags); |
4840e418 | 6237 | break; |
c1ca757b | 6238 | case IORING_OP_MADVISE: |
45d189c6 | 6239 | ret = io_madvise(req, issue_flags); |
c1ca757b | 6240 | break; |
cebdb986 | 6241 | case IORING_OP_OPENAT2: |
45d189c6 | 6242 | ret = io_openat2(req, issue_flags); |
cebdb986 | 6243 | break; |
3e4827b0 | 6244 | case IORING_OP_EPOLL_CTL: |
889fca73 | 6245 | ret = io_epoll_ctl(req, issue_flags); |
3e4827b0 | 6246 | break; |
7d67af2c | 6247 | case IORING_OP_SPLICE: |
45d189c6 | 6248 | ret = io_splice(req, issue_flags); |
7d67af2c | 6249 | break; |
ddf0322d | 6250 | case IORING_OP_PROVIDE_BUFFERS: |
889fca73 | 6251 | ret = io_provide_buffers(req, issue_flags); |
ddf0322d | 6252 | break; |
067524e9 | 6253 | case IORING_OP_REMOVE_BUFFERS: |
889fca73 | 6254 | ret = io_remove_buffers(req, issue_flags); |
3e4827b0 | 6255 | break; |
f2a8d5c7 | 6256 | case IORING_OP_TEE: |
45d189c6 | 6257 | ret = io_tee(req, issue_flags); |
f2a8d5c7 | 6258 | break; |
36f4fa68 | 6259 | case IORING_OP_SHUTDOWN: |
45d189c6 | 6260 | ret = io_shutdown(req, issue_flags); |
36f4fa68 | 6261 | break; |
80a261fd | 6262 | case IORING_OP_RENAMEAT: |
45d189c6 | 6263 | ret = io_renameat(req, issue_flags); |
80a261fd | 6264 | break; |
14a1143b | 6265 | case IORING_OP_UNLINKAT: |
45d189c6 | 6266 | ret = io_unlinkat(req, issue_flags); |
14a1143b | 6267 | break; |
2b188cc1 JA |
6268 | default: |
6269 | ret = -EINVAL; | |
6270 | break; | |
6271 | } | |
6272 | ||
5730b27e JA |
6273 | if (creds) |
6274 | revert_creds(creds); | |
def596e9 JA |
6275 | if (ret) |
6276 | return ret; | |
b532576e | 6277 | /* If the op doesn't have a file, we're not polling for it */ |
cb3d8972 PB |
6278 | if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) |
6279 | io_iopoll_req_issued(req); | |
def596e9 JA |
6280 | |
6281 | return 0; | |
2b188cc1 JA |
6282 | } |
6283 | ||
ebc11b6c PB |
6284 | static struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
6285 | { | |
6286 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6287 | ||
6288 | req = io_put_req_find_next(req); | |
6289 | return req ? &req->work : NULL; | |
6290 | } | |
6291 | ||
5280f7e5 | 6292 | static void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
6293 | { |
6294 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6df1db6b | 6295 | struct io_kiocb *timeout; |
561fb04a | 6296 | int ret = 0; |
2b188cc1 | 6297 | |
48dcd38d PB |
6298 | /* one will be dropped by ->io_free_work() after returning to io-wq */ |
6299 | if (!(req->flags & REQ_F_REFCOUNT)) | |
6300 | __io_req_set_refcount(req, 2); | |
6301 | else | |
6302 | req_ref_get(req); | |
5d5901a3 | 6303 | |
6df1db6b PB |
6304 | timeout = io_prep_linked_timeout(req); |
6305 | if (timeout) | |
6306 | io_queue_linked_timeout(timeout); | |
d4c81f38 | 6307 | |
4014d943 | 6308 | if (work->flags & IO_WQ_WORK_CANCEL) |
561fb04a | 6309 | ret = -ECANCELED; |
31b51510 | 6310 | |
561fb04a | 6311 | if (!ret) { |
561fb04a | 6312 | do { |
889fca73 | 6313 | ret = io_issue_sqe(req, 0); |
561fb04a JA |
6314 | /* |
6315 | * We can get EAGAIN for polled IO even though we're | |
6316 | * forcing a sync submission from here, since we can't | |
6317 | * wait for request slots on the block side. | |
6318 | */ | |
6319 | if (ret != -EAGAIN) | |
6320 | break; | |
6321 | cond_resched(); | |
6322 | } while (1); | |
6323 | } | |
31b51510 | 6324 | |
a3df7698 | 6325 | /* avoid locking problems by failing it from a clean context */ |
5d5901a3 | 6326 | if (ret) |
a3df7698 | 6327 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
6328 | } |
6329 | ||
aeca241b | 6330 | static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table, |
042b0d85 | 6331 | unsigned i) |
65e19f54 | 6332 | { |
042b0d85 | 6333 | return &table->files[i]; |
dafecf19 PB |
6334 | } |
6335 | ||
65e19f54 JA |
6336 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
6337 | int index) | |
6338 | { | |
aeca241b | 6339 | struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index); |
65e19f54 | 6340 | |
a04b0ac0 | 6341 | return (struct file *) (slot->file_ptr & FFS_MASK); |
65e19f54 JA |
6342 | } |
6343 | ||
a04b0ac0 | 6344 | static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file) |
9a321c98 PB |
6345 | { |
6346 | unsigned long file_ptr = (unsigned long) file; | |
6347 | ||
b191e2df | 6348 | if (__io_file_supports_nowait(file, READ)) |
9a321c98 | 6349 | file_ptr |= FFS_ASYNC_READ; |
b191e2df | 6350 | if (__io_file_supports_nowait(file, WRITE)) |
9a321c98 PB |
6351 | file_ptr |= FFS_ASYNC_WRITE; |
6352 | if (S_ISREG(file_inode(file)->i_mode)) | |
6353 | file_ptr |= FFS_ISREG; | |
a04b0ac0 | 6354 | file_slot->file_ptr = file_ptr; |
65e19f54 JA |
6355 | } |
6356 | ||
ac177053 PB |
6357 | static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx, |
6358 | struct io_kiocb *req, int fd) | |
09bb8394 | 6359 | { |
8da11c19 | 6360 | struct file *file; |
ac177053 | 6361 | unsigned long file_ptr; |
09bb8394 | 6362 | |
ac177053 PB |
6363 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
6364 | return NULL; | |
6365 | fd = array_index_nospec(fd, ctx->nr_user_files); | |
6366 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
6367 | file = (struct file *) (file_ptr & FFS_MASK); | |
6368 | file_ptr &= ~FFS_MASK; | |
6369 | /* mask in overlapping REQ_F and FFS bits */ | |
b191e2df | 6370 | req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT); |
ac177053 PB |
6371 | io_req_set_rsrc_node(req); |
6372 | return file; | |
6373 | } | |
d44f554e | 6374 | |
ac177053 | 6375 | static struct file *io_file_get_normal(struct io_ring_ctx *ctx, |
ac177053 PB |
6376 | struct io_kiocb *req, int fd) |
6377 | { | |
62906e89 | 6378 | struct file *file = fget(fd); |
ac177053 PB |
6379 | |
6380 | trace_io_uring_file_get(ctx, fd); | |
09bb8394 | 6381 | |
ac177053 PB |
6382 | /* we don't allow fixed io_uring files */ |
6383 | if (file && unlikely(file->f_op == &io_uring_fops)) | |
6384 | io_req_track_inflight(req); | |
8371adf5 | 6385 | return file; |
09bb8394 JA |
6386 | } |
6387 | ||
ac177053 | 6388 | static inline struct file *io_file_get(struct io_ring_ctx *ctx, |
ac177053 PB |
6389 | struct io_kiocb *req, int fd, bool fixed) |
6390 | { | |
6391 | if (fixed) | |
6392 | return io_file_get_fixed(ctx, req, fd); | |
6393 | else | |
62906e89 | 6394 | return io_file_get_normal(ctx, req, fd); |
ac177053 PB |
6395 | } |
6396 | ||
89b263f6 JA |
6397 | static void io_req_task_link_timeout(struct io_kiocb *req) |
6398 | { | |
6399 | struct io_kiocb *prev = req->timeout.prev; | |
8cb01fac | 6400 | int ret; |
89b263f6 JA |
6401 | |
6402 | if (prev) { | |
8cb01fac | 6403 | ret = io_try_cancel_userdata(req, prev->user_data); |
505657bc | 6404 | io_req_complete_post(req, ret ?: -ETIME, 0); |
89b263f6 | 6405 | io_put_req(prev); |
89b263f6 JA |
6406 | } else { |
6407 | io_req_complete_post(req, -ETIME, 0); | |
6408 | } | |
6409 | } | |
6410 | ||
2665abfd | 6411 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 6412 | { |
ad8a48ac JA |
6413 | struct io_timeout_data *data = container_of(timer, |
6414 | struct io_timeout_data, timer); | |
90cd7e42 | 6415 | struct io_kiocb *prev, *req = data->req; |
2665abfd | 6416 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 6417 | unsigned long flags; |
2665abfd | 6418 | |
89b263f6 | 6419 | spin_lock_irqsave(&ctx->timeout_lock, flags); |
90cd7e42 PB |
6420 | prev = req->timeout.head; |
6421 | req->timeout.head = NULL; | |
2665abfd JA |
6422 | |
6423 | /* | |
6424 | * We don't expect the list to be empty, that will only happen if we | |
6425 | * race with the completion of the linked work. | |
6426 | */ | |
447c19f3 | 6427 | if (prev) { |
f2f87370 | 6428 | io_remove_next_linked(prev); |
447c19f3 PB |
6429 | if (!req_ref_inc_not_zero(prev)) |
6430 | prev = NULL; | |
6431 | } | |
89b263f6 JA |
6432 | req->timeout.prev = prev; |
6433 | spin_unlock_irqrestore(&ctx->timeout_lock, flags); | |
2665abfd | 6434 | |
89b263f6 JA |
6435 | req->io_task_work.func = io_req_task_link_timeout; |
6436 | io_req_task_work_add(req); | |
2665abfd JA |
6437 | return HRTIMER_NORESTART; |
6438 | } | |
6439 | ||
de968c18 | 6440 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 6441 | { |
de968c18 PB |
6442 | struct io_ring_ctx *ctx = req->ctx; |
6443 | ||
89b263f6 | 6444 | spin_lock_irq(&ctx->timeout_lock); |
76a46e06 | 6445 | /* |
f2f87370 PB |
6446 | * If the back reference is NULL, then our linked request finished |
6447 | * before we got a chance to setup the timer | |
76a46e06 | 6448 | */ |
90cd7e42 | 6449 | if (req->timeout.head) { |
e8c2bc1f | 6450 | struct io_timeout_data *data = req->async_data; |
94ae5e77 | 6451 | |
ad8a48ac JA |
6452 | data->timer.function = io_link_timeout_fn; |
6453 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
6454 | data->mode); | |
2665abfd | 6455 | } |
89b263f6 | 6456 | spin_unlock_irq(&ctx->timeout_lock); |
2665abfd | 6457 | /* drop submission reference */ |
76a46e06 JA |
6458 | io_put_req(req); |
6459 | } | |
2665abfd | 6460 | |
c5eef2b9 | 6461 | static void __io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 6462 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 6463 | { |
906c6caa | 6464 | struct io_kiocb *linked_timeout; |
e0c5c576 | 6465 | int ret; |
2b188cc1 | 6466 | |
59b735ae | 6467 | issue_sqe: |
c5eef2b9 | 6468 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 6469 | |
491381ce JA |
6470 | /* |
6471 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
6472 | * doesn't support non-blocking read/write attempts | |
6473 | */ | |
1840038e | 6474 | if (likely(!ret)) { |
e342c807 | 6475 | if (req->flags & REQ_F_COMPLETE_INLINE) { |
c5eef2b9 | 6476 | struct io_ring_ctx *ctx = req->ctx; |
cd0ca2e0 | 6477 | struct io_submit_state *state = &ctx->submit_state; |
e65ef56d | 6478 | |
cd0ca2e0 PB |
6479 | state->compl_reqs[state->compl_nr++] = req; |
6480 | if (state->compl_nr == ARRAY_SIZE(state->compl_reqs)) | |
2a2758f2 | 6481 | io_submit_flush_completions(ctx); |
906c6caa | 6482 | return; |
0d63c148 | 6483 | } |
906c6caa PB |
6484 | |
6485 | linked_timeout = io_prep_linked_timeout(req); | |
6486 | if (linked_timeout) | |
6487 | io_queue_linked_timeout(linked_timeout); | |
1840038e | 6488 | } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
906c6caa PB |
6489 | linked_timeout = io_prep_linked_timeout(req); |
6490 | ||
59b735ae OL |
6491 | switch (io_arm_poll_handler(req)) { |
6492 | case IO_APOLL_READY: | |
906c6caa PB |
6493 | if (linked_timeout) |
6494 | io_unprep_linked_timeout(req); | |
59b735ae OL |
6495 | goto issue_sqe; |
6496 | case IO_APOLL_ABORTED: | |
1840038e PB |
6497 | /* |
6498 | * Queued up for async execution, worker will release | |
6499 | * submit reference when the iocb is actually submitted. | |
6500 | */ | |
6501 | io_queue_async_work(req); | |
59b735ae | 6502 | break; |
1840038e | 6503 | } |
906c6caa PB |
6504 | |
6505 | if (linked_timeout) | |
6506 | io_queue_linked_timeout(linked_timeout); | |
0d63c148 | 6507 | } else { |
f41db273 | 6508 | io_req_complete_failed(req, ret); |
9e645e11 | 6509 | } |
2b188cc1 JA |
6510 | } |
6511 | ||
441b8a78 | 6512 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 6513 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 6514 | { |
10c66904 | 6515 | if (unlikely(req->ctx->drain_active) && io_drain_req(req)) |
76cc33d7 | 6516 | return; |
4fe2c963 | 6517 | |
76cc33d7 | 6518 | if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) { |
c5eef2b9 | 6519 | __io_queue_sqe(req); |
76cc33d7 PB |
6520 | } else { |
6521 | int ret = io_req_prep_async(req); | |
6522 | ||
6523 | if (unlikely(ret)) | |
6524 | io_req_complete_failed(req, ret); | |
6525 | else | |
6526 | io_queue_async_work(req); | |
ce35a47a | 6527 | } |
4fe2c963 JL |
6528 | } |
6529 | ||
b16fed66 PB |
6530 | /* |
6531 | * Check SQE restrictions (opcode and flags). | |
6532 | * | |
6533 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
6534 | */ | |
6535 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
6536 | struct io_kiocb *req, | |
6537 | unsigned int sqe_flags) | |
4fe2c963 | 6538 | { |
4cfb25bf | 6539 | if (likely(!ctx->restricted)) |
b16fed66 PB |
6540 | return true; |
6541 | ||
6542 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) | |
6543 | return false; | |
6544 | ||
6545 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
6546 | ctx->restrictions.sqe_flags_required) | |
6547 | return false; | |
6548 | ||
6549 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
6550 | ctx->restrictions.sqe_flags_required)) | |
6551 | return false; | |
6552 | ||
6553 | return true; | |
4fe2c963 JL |
6554 | } |
6555 | ||
b16fed66 PB |
6556 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
6557 | const struct io_uring_sqe *sqe) | |
282cdc86 | 6558 | __must_hold(&ctx->uring_lock) |
b16fed66 PB |
6559 | { |
6560 | struct io_submit_state *state; | |
6561 | unsigned int sqe_flags; | |
003e8dcc | 6562 | int personality, ret = 0; |
b16fed66 | 6563 | |
864ea921 | 6564 | /* req is partially pre-initialised, see io_preinit_req() */ |
b16fed66 PB |
6565 | req->opcode = READ_ONCE(sqe->opcode); |
6566 | /* same numerical values with corresponding REQ_F_*, safe to copy */ | |
6567 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
6568 | req->user_data = READ_ONCE(sqe->user_data); | |
b16fed66 | 6569 | req->file = NULL; |
b16fed66 | 6570 | req->fixed_rsrc_refs = NULL; |
b16fed66 | 6571 | req->task = current; |
b16fed66 PB |
6572 | |
6573 | /* enforce forwards compatibility on users */ | |
dddca226 | 6574 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) |
b16fed66 | 6575 | return -EINVAL; |
b16fed66 PB |
6576 | if (unlikely(req->opcode >= IORING_OP_LAST)) |
6577 | return -EINVAL; | |
4cfb25bf | 6578 | if (!io_check_restriction(ctx, req, sqe_flags)) |
b16fed66 PB |
6579 | return -EACCES; |
6580 | ||
6581 | if ((sqe_flags & IOSQE_BUFFER_SELECT) && | |
6582 | !io_op_defs[req->opcode].buffer_select) | |
6583 | return -EOPNOTSUPP; | |
3c19966d PB |
6584 | if (unlikely(sqe_flags & IOSQE_IO_DRAIN)) |
6585 | ctx->drain_active = true; | |
863e0560 | 6586 | |
003e8dcc JA |
6587 | personality = READ_ONCE(sqe->personality); |
6588 | if (personality) { | |
c10d1f98 PB |
6589 | req->creds = xa_load(&ctx->personalities, personality); |
6590 | if (!req->creds) | |
003e8dcc | 6591 | return -EINVAL; |
c10d1f98 | 6592 | get_cred(req->creds); |
b8e64b53 | 6593 | req->flags |= REQ_F_CREDS; |
003e8dcc | 6594 | } |
b16fed66 PB |
6595 | state = &ctx->submit_state; |
6596 | ||
6597 | /* | |
6598 | * Plug now if we have more than 1 IO left after this, and the target | |
6599 | * is potentially a read/write to block based storage. | |
6600 | */ | |
6601 | if (!state->plug_started && state->ios_left > 1 && | |
6602 | io_op_defs[req->opcode].plug) { | |
6603 | blk_start_plug(&state->plug); | |
6604 | state->plug_started = true; | |
6605 | } | |
6606 | ||
6607 | if (io_op_defs[req->opcode].needs_file) { | |
62906e89 | 6608 | req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd), |
ac177053 | 6609 | (sqe_flags & IOSQE_FIXED_FILE)); |
b16fed66 PB |
6610 | if (unlikely(!req->file)) |
6611 | ret = -EBADF; | |
6612 | } | |
6613 | ||
6614 | state->ios_left--; | |
6615 | return ret; | |
6616 | } | |
6617 | ||
a6b8cadc | 6618 | static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
a1ab7b35 | 6619 | const struct io_uring_sqe *sqe) |
282cdc86 | 6620 | __must_hold(&ctx->uring_lock) |
9e645e11 | 6621 | { |
a1ab7b35 | 6622 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 6623 | int ret; |
9e645e11 | 6624 | |
a6b8cadc PB |
6625 | ret = io_init_req(ctx, req, sqe); |
6626 | if (unlikely(ret)) { | |
6627 | fail_req: | |
de59bc10 PB |
6628 | if (link->head) { |
6629 | /* fail even hard links since we don't submit */ | |
93d2bcd2 | 6630 | req_set_fail(link->head); |
f41db273 | 6631 | io_req_complete_failed(link->head, -ECANCELED); |
de59bc10 PB |
6632 | link->head = NULL; |
6633 | } | |
f41db273 | 6634 | io_req_complete_failed(req, ret); |
a6b8cadc PB |
6635 | return ret; |
6636 | } | |
441b8a78 | 6637 | |
be7053b7 PB |
6638 | ret = io_req_prep(req, sqe); |
6639 | if (unlikely(ret)) | |
6640 | goto fail_req; | |
a6b8cadc | 6641 | |
be7053b7 | 6642 | /* don't need @sqe from now on */ |
236daeae OL |
6643 | trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data, |
6644 | req->flags, true, | |
6645 | ctx->flags & IORING_SETUP_SQPOLL); | |
a6b8cadc | 6646 | |
9e645e11 JA |
6647 | /* |
6648 | * If we already have a head request, queue this one for async | |
6649 | * submittal once the head completes. If we don't have a head but | |
6650 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
6651 | * submitted sync once the chain is complete. If none of those | |
6652 | * conditions are true (normal request), then just queue it. | |
6653 | */ | |
863e0560 PB |
6654 | if (link->head) { |
6655 | struct io_kiocb *head = link->head; | |
4e88d6e7 | 6656 | |
b7e298d2 | 6657 | ret = io_req_prep_async(req); |
cf109604 | 6658 | if (unlikely(ret)) |
a6b8cadc | 6659 | goto fail_req; |
9d76377f | 6660 | trace_io_uring_link(ctx, req, head); |
f2f87370 | 6661 | link->last->link = req; |
863e0560 | 6662 | link->last = req; |
32fe525b PB |
6663 | |
6664 | /* last request of a link, enqueue the link */ | |
ef4ff581 | 6665 | if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) { |
863e0560 | 6666 | link->head = NULL; |
5e159204 | 6667 | io_queue_sqe(head); |
32fe525b | 6668 | } |
9e645e11 | 6669 | } else { |
ef4ff581 | 6670 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
863e0560 PB |
6671 | link->head = req; |
6672 | link->last = req; | |
711be031 | 6673 | } else { |
be7053b7 | 6674 | io_queue_sqe(req); |
711be031 | 6675 | } |
9e645e11 | 6676 | } |
2e6e1fde | 6677 | |
1d4240cc | 6678 | return 0; |
9e645e11 JA |
6679 | } |
6680 | ||
9a56a232 JA |
6681 | /* |
6682 | * Batched submission is done, ensure local IO is flushed out. | |
6683 | */ | |
ba88ff11 PB |
6684 | static void io_submit_state_end(struct io_submit_state *state, |
6685 | struct io_ring_ctx *ctx) | |
9a56a232 | 6686 | { |
a1ab7b35 | 6687 | if (state->link.head) |
de59bc10 | 6688 | io_queue_sqe(state->link.head); |
cd0ca2e0 | 6689 | if (state->compl_nr) |
2a2758f2 | 6690 | io_submit_flush_completions(ctx); |
27926b68 JA |
6691 | if (state->plug_started) |
6692 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
6693 | } |
6694 | ||
6695 | /* | |
6696 | * Start submission side cache. | |
6697 | */ | |
6698 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 6699 | unsigned int max_ios) |
9a56a232 | 6700 | { |
27926b68 | 6701 | state->plug_started = false; |
9a56a232 | 6702 | state->ios_left = max_ios; |
a1ab7b35 PB |
6703 | /* set only head, no need to init link_last in advance */ |
6704 | state->link.head = NULL; | |
9a56a232 JA |
6705 | } |
6706 | ||
2b188cc1 JA |
6707 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
6708 | { | |
75b28aff | 6709 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 6710 | |
caf582c6 PB |
6711 | /* |
6712 | * Ensure any loads from the SQEs are done at this point, | |
6713 | * since once we write the new head, the application could | |
6714 | * write new data to them. | |
6715 | */ | |
6716 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
6717 | } |
6718 | ||
2b188cc1 | 6719 | /* |
dd9ae8a0 | 6720 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
6721 | * that is mapped by userspace. This means that care needs to be taken to |
6722 | * ensure that reads are stable, as we cannot rely on userspace always | |
6723 | * being a good citizen. If members of the sqe are validated and then later | |
6724 | * used, it's important that those reads are done through READ_ONCE() to | |
6725 | * prevent a re-load down the line. | |
6726 | */ | |
709b302f | 6727 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 6728 | { |
ea5ab3b5 | 6729 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 6730 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
6731 | |
6732 | /* | |
6733 | * The cached sq head (or cq tail) serves two purposes: | |
6734 | * | |
6735 | * 1) allows us to batch the cost of updating the user visible | |
6736 | * head updates. | |
6737 | * 2) allows the kernel side to track the head on its own, even | |
6738 | * though the application is the one updating it. | |
6739 | */ | |
17d3aeb3 | 6740 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
709b302f PB |
6741 | if (likely(head < ctx->sq_entries)) |
6742 | return &ctx->sq_sqes[head]; | |
2b188cc1 JA |
6743 | |
6744 | /* drop invalid entries */ | |
15641e42 PB |
6745 | ctx->cq_extra--; |
6746 | WRITE_ONCE(ctx->rings->sq_dropped, | |
6747 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
6748 | return NULL; |
6749 | } | |
6750 | ||
0f212204 | 6751 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 6752 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 6753 | { |
09899b19 | 6754 | struct io_uring_task *tctx; |
46c4e16a | 6755 | int submitted = 0; |
6c271ce2 | 6756 | |
ee7d46d9 PB |
6757 | /* make sure SQ entry isn't read before tail */ |
6758 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
2b85edfc PB |
6759 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
6760 | return -EAGAIN; | |
6c271ce2 | 6761 | |
09899b19 PB |
6762 | tctx = current->io_uring; |
6763 | tctx->cached_refs -= nr; | |
6764 | if (unlikely(tctx->cached_refs < 0)) { | |
6765 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
6766 | ||
6767 | percpu_counter_add(&tctx->inflight, refill); | |
6768 | refcount_add(refill, ¤t->usage); | |
6769 | tctx->cached_refs += refill; | |
6770 | } | |
ba88ff11 | 6771 | io_submit_state_start(&ctx->submit_state, nr); |
b14cca0c | 6772 | |
46c4e16a | 6773 | while (submitted < nr) { |
3529d8c2 | 6774 | const struct io_uring_sqe *sqe; |
196be95c | 6775 | struct io_kiocb *req; |
fb5ccc98 | 6776 | |
258b29a9 | 6777 | req = io_alloc_req(ctx); |
196be95c PB |
6778 | if (unlikely(!req)) { |
6779 | if (!submitted) | |
6780 | submitted = -EAGAIN; | |
fb5ccc98 | 6781 | break; |
196be95c | 6782 | } |
4fccfcbb PB |
6783 | sqe = io_get_sqe(ctx); |
6784 | if (unlikely(!sqe)) { | |
6785 | kmem_cache_free(req_cachep, req); | |
6786 | break; | |
6787 | } | |
d3656344 JA |
6788 | /* will complete beyond this point, count as submitted */ |
6789 | submitted++; | |
a1ab7b35 | 6790 | if (io_submit_sqe(ctx, req, sqe)) |
196be95c | 6791 | break; |
6c271ce2 JA |
6792 | } |
6793 | ||
9466f437 PB |
6794 | if (unlikely(submitted != nr)) { |
6795 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
d8a6df10 | 6796 | int unused = nr - ref_used; |
9466f437 | 6797 | |
09899b19 | 6798 | current->io_uring->cached_refs += unused; |
d8a6df10 | 6799 | percpu_ref_put_many(&ctx->refs, unused); |
9466f437 | 6800 | } |
6c271ce2 | 6801 | |
a1ab7b35 | 6802 | io_submit_state_end(&ctx->submit_state, ctx); |
ae9428ca PB |
6803 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
6804 | io_commit_sqring(ctx); | |
6805 | ||
6c271ce2 JA |
6806 | return submitted; |
6807 | } | |
6808 | ||
e4b6d902 PB |
6809 | static inline bool io_sqd_events_pending(struct io_sq_data *sqd) |
6810 | { | |
6811 | return READ_ONCE(sqd->state); | |
6812 | } | |
6813 | ||
23b3628e XW |
6814 | static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx) |
6815 | { | |
6816 | /* Tell userspace we may need a wakeup call */ | |
79ebeaee | 6817 | spin_lock(&ctx->completion_lock); |
20c0b380 NA |
6818 | WRITE_ONCE(ctx->rings->sq_flags, |
6819 | ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP); | |
79ebeaee | 6820 | spin_unlock(&ctx->completion_lock); |
23b3628e XW |
6821 | } |
6822 | ||
6823 | static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx) | |
6824 | { | |
79ebeaee | 6825 | spin_lock(&ctx->completion_lock); |
20c0b380 NA |
6826 | WRITE_ONCE(ctx->rings->sq_flags, |
6827 | ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP); | |
79ebeaee | 6828 | spin_unlock(&ctx->completion_lock); |
23b3628e XW |
6829 | } |
6830 | ||
08369246 | 6831 | static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) |
6c271ce2 | 6832 | { |
c8d1ba58 | 6833 | unsigned int to_submit; |
bdcd3eab | 6834 | int ret = 0; |
6c271ce2 | 6835 | |
c8d1ba58 | 6836 | to_submit = io_sqring_entries(ctx); |
e95eee2d | 6837 | /* if we're handling multiple rings, cap submit size for fairness */ |
4ce8ad95 OL |
6838 | if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE) |
6839 | to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE; | |
e95eee2d | 6840 | |
906a3c6f | 6841 | if (!list_empty(&ctx->iopoll_list) || to_submit) { |
c8d1ba58 | 6842 | unsigned nr_events = 0; |
948e1947 PB |
6843 | const struct cred *creds = NULL; |
6844 | ||
6845 | if (ctx->sq_creds != current_cred()) | |
6846 | creds = override_creds(ctx->sq_creds); | |
a4c0b3de | 6847 | |
c8d1ba58 | 6848 | mutex_lock(&ctx->uring_lock); |
906a3c6f | 6849 | if (!list_empty(&ctx->iopoll_list)) |
a8576af9 | 6850 | io_do_iopoll(ctx, &nr_events, 0); |
906a3c6f | 6851 | |
3b763ba1 PB |
6852 | /* |
6853 | * Don't submit if refs are dying, good for io_uring_register(), | |
6854 | * but also it is relied upon by io_ring_exit_work() | |
6855 | */ | |
0298ef96 PB |
6856 | if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) && |
6857 | !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
08369246 | 6858 | ret = io_submit_sqes(ctx, to_submit); |
c8d1ba58 | 6859 | mutex_unlock(&ctx->uring_lock); |
6c271ce2 | 6860 | |
acfb381d PB |
6861 | if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait)) |
6862 | wake_up(&ctx->sqo_sq_wait); | |
948e1947 PB |
6863 | if (creds) |
6864 | revert_creds(creds); | |
acfb381d | 6865 | } |
6c271ce2 | 6866 | |
08369246 XW |
6867 | return ret; |
6868 | } | |
6c271ce2 | 6869 | |
08369246 XW |
6870 | static void io_sqd_update_thread_idle(struct io_sq_data *sqd) |
6871 | { | |
6872 | struct io_ring_ctx *ctx; | |
6873 | unsigned sq_thread_idle = 0; | |
6c271ce2 | 6874 | |
c9dca27d PB |
6875 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
6876 | sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle); | |
08369246 | 6877 | sqd->sq_thread_idle = sq_thread_idle; |
c8d1ba58 | 6878 | } |
6c271ce2 | 6879 | |
e4b6d902 PB |
6880 | static bool io_sqd_handle_event(struct io_sq_data *sqd) |
6881 | { | |
6882 | bool did_sig = false; | |
6883 | struct ksignal ksig; | |
6884 | ||
6885 | if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) || | |
6886 | signal_pending(current)) { | |
6887 | mutex_unlock(&sqd->lock); | |
6888 | if (signal_pending(current)) | |
6889 | did_sig = get_signal(&ksig); | |
6890 | cond_resched(); | |
6891 | mutex_lock(&sqd->lock); | |
6892 | } | |
e4b6d902 PB |
6893 | return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
6894 | } | |
6895 | ||
c8d1ba58 JA |
6896 | static int io_sq_thread(void *data) |
6897 | { | |
69fb2131 JA |
6898 | struct io_sq_data *sqd = data; |
6899 | struct io_ring_ctx *ctx; | |
a0d9205f | 6900 | unsigned long timeout = 0; |
37d1e2e3 | 6901 | char buf[TASK_COMM_LEN]; |
08369246 | 6902 | DEFINE_WAIT(wait); |
6c271ce2 | 6903 | |
696ee88a | 6904 | snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid); |
37d1e2e3 | 6905 | set_task_comm(current, buf); |
37d1e2e3 JA |
6906 | |
6907 | if (sqd->sq_cpu != -1) | |
6908 | set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu)); | |
6909 | else | |
6910 | set_cpus_allowed_ptr(current, cpu_online_mask); | |
6911 | current->flags |= PF_NO_SETAFFINITY; | |
6912 | ||
09a6f4ef | 6913 | mutex_lock(&sqd->lock); |
e4b6d902 | 6914 | while (1) { |
1a924a80 | 6915 | bool cap_entries, sqt_spin = false; |
c1edbf5f | 6916 | |
e4b6d902 PB |
6917 | if (io_sqd_events_pending(sqd) || signal_pending(current)) { |
6918 | if (io_sqd_handle_event(sqd)) | |
c7d95613 | 6919 | break; |
08369246 XW |
6920 | timeout = jiffies + sqd->sq_thread_idle; |
6921 | } | |
e4b6d902 | 6922 | |
e95eee2d | 6923 | cap_entries = !list_is_singular(&sqd->ctx_list); |
69fb2131 | 6924 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
948e1947 | 6925 | int ret = __io_sq_thread(ctx, cap_entries); |
7c30f36a | 6926 | |
08369246 XW |
6927 | if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list))) |
6928 | sqt_spin = true; | |
69fb2131 | 6929 | } |
dd432ea5 PB |
6930 | if (io_run_task_work()) |
6931 | sqt_spin = true; | |
6c271ce2 | 6932 | |
08369246 | 6933 | if (sqt_spin || !time_after(jiffies, timeout)) { |
c8d1ba58 | 6934 | cond_resched(); |
08369246 XW |
6935 | if (sqt_spin) |
6936 | timeout = jiffies + sqd->sq_thread_idle; | |
6937 | continue; | |
6938 | } | |
6939 | ||
08369246 | 6940 | prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE); |
dd432ea5 | 6941 | if (!io_sqd_events_pending(sqd) && !current->task_works) { |
1a924a80 PB |
6942 | bool needs_sched = true; |
6943 | ||
724cb4f9 | 6944 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
aaa9f0f4 PB |
6945 | io_ring_set_wakeup_flag(ctx); |
6946 | ||
724cb4f9 HX |
6947 | if ((ctx->flags & IORING_SETUP_IOPOLL) && |
6948 | !list_empty_careful(&ctx->iopoll_list)) { | |
6949 | needs_sched = false; | |
6950 | break; | |
6951 | } | |
6952 | if (io_sqring_entries(ctx)) { | |
6953 | needs_sched = false; | |
6954 | break; | |
6955 | } | |
6956 | } | |
6957 | ||
6958 | if (needs_sched) { | |
6959 | mutex_unlock(&sqd->lock); | |
6960 | schedule(); | |
6961 | mutex_lock(&sqd->lock); | |
6962 | } | |
69fb2131 JA |
6963 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
6964 | io_ring_clear_wakeup_flag(ctx); | |
6c271ce2 | 6965 | } |
08369246 XW |
6966 | |
6967 | finish_wait(&sqd->wait, &wait); | |
6968 | timeout = jiffies + sqd->sq_thread_idle; | |
6c271ce2 | 6969 | } |
28cea78a | 6970 | |
78cc687b | 6971 | io_uring_cancel_generic(true, sqd); |
37d1e2e3 | 6972 | sqd->thread = NULL; |
05962f95 | 6973 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
5f3f26f9 | 6974 | io_ring_set_wakeup_flag(ctx); |
521d6a73 | 6975 | io_run_task_work(); |
734551df PB |
6976 | mutex_unlock(&sqd->lock); |
6977 | ||
37d1e2e3 JA |
6978 | complete(&sqd->exited); |
6979 | do_exit(0); | |
6c271ce2 JA |
6980 | } |
6981 | ||
bda52162 JA |
6982 | struct io_wait_queue { |
6983 | struct wait_queue_entry wq; | |
6984 | struct io_ring_ctx *ctx; | |
5fd46178 | 6985 | unsigned cq_tail; |
bda52162 JA |
6986 | unsigned nr_timeouts; |
6987 | }; | |
6988 | ||
6c503150 | 6989 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
6990 | { |
6991 | struct io_ring_ctx *ctx = iowq->ctx; | |
5fd46178 | 6992 | int dist = ctx->cached_cq_tail - (int) iowq->cq_tail; |
bda52162 JA |
6993 | |
6994 | /* | |
d195a66e | 6995 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
6996 | * started waiting. For timeouts, we always want to return to userspace, |
6997 | * regardless of event count. | |
6998 | */ | |
5fd46178 | 6999 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
7000 | } |
7001 | ||
7002 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
7003 | int wake_flags, void *key) | |
7004 | { | |
7005 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
7006 | wq); | |
7007 | ||
6c503150 PB |
7008 | /* |
7009 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
7010 | * the task, and the next invocation will do it. | |
7011 | */ | |
5ed7a37d | 7012 | if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow)) |
6c503150 PB |
7013 | return autoremove_wake_function(curr, mode, wake_flags, key); |
7014 | return -1; | |
bda52162 JA |
7015 | } |
7016 | ||
af9c1a44 JA |
7017 | static int io_run_task_work_sig(void) |
7018 | { | |
7019 | if (io_run_task_work()) | |
7020 | return 1; | |
7021 | if (!signal_pending(current)) | |
7022 | return 0; | |
0b8cfa97 | 7023 | if (test_thread_flag(TIF_NOTIFY_SIGNAL)) |
792ee0f6 | 7024 | return -ERESTARTSYS; |
af9c1a44 JA |
7025 | return -EINTR; |
7026 | } | |
7027 | ||
eeb60b9a PB |
7028 | /* when returns >0, the caller should retry */ |
7029 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
7030 | struct io_wait_queue *iowq, | |
7031 | signed long *timeout) | |
7032 | { | |
7033 | int ret; | |
7034 | ||
7035 | /* make sure we run task_work before checking for signals */ | |
7036 | ret = io_run_task_work_sig(); | |
7037 | if (ret || io_should_wake(iowq)) | |
7038 | return ret; | |
7039 | /* let the caller flush overflows, retry */ | |
5ed7a37d | 7040 | if (test_bit(0, &ctx->check_cq_overflow)) |
eeb60b9a PB |
7041 | return 1; |
7042 | ||
7043 | *timeout = schedule_timeout(*timeout); | |
7044 | return !*timeout ? -ETIME : 1; | |
7045 | } | |
7046 | ||
2b188cc1 JA |
7047 | /* |
7048 | * Wait until events become available, if we don't already have some. The | |
7049 | * application must reap them itself, as they reside on the shared cq ring. | |
7050 | */ | |
7051 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
7052 | const sigset_t __user *sig, size_t sigsz, |
7053 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 7054 | { |
90291099 | 7055 | struct io_wait_queue iowq; |
75b28aff | 7056 | struct io_rings *rings = ctx->rings; |
c1d5a224 PB |
7057 | signed long timeout = MAX_SCHEDULE_TIMEOUT; |
7058 | int ret; | |
2b188cc1 | 7059 | |
b41e9852 | 7060 | do { |
90f67366 | 7061 | io_cqring_overflow_flush(ctx); |
6c503150 | 7062 | if (io_cqring_events(ctx) >= min_events) |
b41e9852 | 7063 | return 0; |
4c6e277c | 7064 | if (!io_run_task_work()) |
b41e9852 | 7065 | break; |
b41e9852 | 7066 | } while (1); |
2b188cc1 JA |
7067 | |
7068 | if (sig) { | |
9e75ad5d AB |
7069 | #ifdef CONFIG_COMPAT |
7070 | if (in_compat_syscall()) | |
7071 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 7072 | sigsz); |
9e75ad5d AB |
7073 | else |
7074 | #endif | |
b772434b | 7075 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 7076 | |
2b188cc1 JA |
7077 | if (ret) |
7078 | return ret; | |
7079 | } | |
7080 | ||
c73ebb68 | 7081 | if (uts) { |
c1d5a224 PB |
7082 | struct timespec64 ts; |
7083 | ||
c73ebb68 HX |
7084 | if (get_timespec64(&ts, uts)) |
7085 | return -EFAULT; | |
7086 | timeout = timespec64_to_jiffies(&ts); | |
7087 | } | |
7088 | ||
90291099 PB |
7089 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
7090 | iowq.wq.private = current; | |
7091 | INIT_LIST_HEAD(&iowq.wq.entry); | |
7092 | iowq.ctx = ctx; | |
bda52162 | 7093 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 7094 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
90291099 | 7095 | |
c826bd7a | 7096 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 7097 | do { |
ca0a2651 | 7098 | /* if we can't even flush overflow, don't wait for more */ |
90f67366 | 7099 | if (!io_cqring_overflow_flush(ctx)) { |
ca0a2651 JA |
7100 | ret = -EBUSY; |
7101 | break; | |
7102 | } | |
311997b3 | 7103 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, |
bda52162 | 7104 | TASK_INTERRUPTIBLE); |
eeb60b9a | 7105 | ret = io_cqring_wait_schedule(ctx, &iowq, &timeout); |
311997b3 | 7106 | finish_wait(&ctx->cq_wait, &iowq.wq); |
ca0a2651 | 7107 | cond_resched(); |
eeb60b9a | 7108 | } while (ret > 0); |
bda52162 | 7109 | |
b7db41c9 | 7110 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 7111 | |
75b28aff | 7112 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
7113 | } |
7114 | ||
9123c8ff | 7115 | static void io_free_page_table(void **table, size_t size) |
05f3fb3c | 7116 | { |
9123c8ff | 7117 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); |
05f3fb3c | 7118 | |
846a4ef2 | 7119 | for (i = 0; i < nr_tables; i++) |
9123c8ff PB |
7120 | kfree(table[i]); |
7121 | kfree(table); | |
7122 | } | |
7123 | ||
7124 | static void **io_alloc_page_table(size_t size) | |
7125 | { | |
7126 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); | |
7127 | size_t init_size = size; | |
7128 | void **table; | |
7129 | ||
7130 | table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL); | |
7131 | if (!table) | |
7132 | return NULL; | |
7133 | ||
7134 | for (i = 0; i < nr_tables; i++) { | |
27f6b318 | 7135 | unsigned int this_size = min_t(size_t, size, PAGE_SIZE); |
9123c8ff PB |
7136 | |
7137 | table[i] = kzalloc(this_size, GFP_KERNEL); | |
7138 | if (!table[i]) { | |
7139 | io_free_page_table(table, init_size); | |
7140 | return NULL; | |
7141 | } | |
7142 | size -= this_size; | |
7143 | } | |
7144 | return table; | |
05f3fb3c JA |
7145 | } |
7146 | ||
28a9fe25 | 7147 | static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node) |
1642b445 | 7148 | { |
28a9fe25 PB |
7149 | percpu_ref_exit(&ref_node->refs); |
7150 | kfree(ref_node); | |
1642b445 PB |
7151 | } |
7152 | ||
b9bd2bea PB |
7153 | static void io_rsrc_node_ref_zero(struct percpu_ref *ref) |
7154 | { | |
7155 | struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs); | |
7156 | struct io_ring_ctx *ctx = node->rsrc_data->ctx; | |
7157 | unsigned long flags; | |
7158 | bool first_add = false; | |
7159 | ||
7160 | spin_lock_irqsave(&ctx->rsrc_ref_lock, flags); | |
7161 | node->done = true; | |
7162 | ||
7163 | while (!list_empty(&ctx->rsrc_ref_list)) { | |
7164 | node = list_first_entry(&ctx->rsrc_ref_list, | |
7165 | struct io_rsrc_node, node); | |
7166 | /* recycle ref nodes in order */ | |
7167 | if (!node->done) | |
7168 | break; | |
7169 | list_del(&node->node); | |
7170 | first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist); | |
7171 | } | |
7172 | spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags); | |
7173 | ||
7174 | if (first_add) | |
7175 | mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ); | |
7176 | } | |
7177 | ||
7178 | static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx) | |
7179 | { | |
7180 | struct io_rsrc_node *ref_node; | |
7181 | ||
7182 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); | |
7183 | if (!ref_node) | |
7184 | return NULL; | |
7185 | ||
7186 | if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero, | |
7187 | 0, GFP_KERNEL)) { | |
7188 | kfree(ref_node); | |
7189 | return NULL; | |
7190 | } | |
7191 | INIT_LIST_HEAD(&ref_node->node); | |
7192 | INIT_LIST_HEAD(&ref_node->rsrc_list); | |
7193 | ref_node->done = false; | |
7194 | return ref_node; | |
7195 | } | |
7196 | ||
a7f0ed5a PB |
7197 | static void io_rsrc_node_switch(struct io_ring_ctx *ctx, |
7198 | struct io_rsrc_data *data_to_kill) | |
6b06314c | 7199 | { |
a7f0ed5a PB |
7200 | WARN_ON_ONCE(!ctx->rsrc_backup_node); |
7201 | WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node); | |
6b06314c | 7202 | |
a7f0ed5a PB |
7203 | if (data_to_kill) { |
7204 | struct io_rsrc_node *rsrc_node = ctx->rsrc_node; | |
82fbcfa9 | 7205 | |
a7f0ed5a | 7206 | rsrc_node->rsrc_data = data_to_kill; |
4956b9ea | 7207 | spin_lock_irq(&ctx->rsrc_ref_lock); |
a7f0ed5a | 7208 | list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list); |
4956b9ea | 7209 | spin_unlock_irq(&ctx->rsrc_ref_lock); |
82fbcfa9 | 7210 | |
3e942498 | 7211 | atomic_inc(&data_to_kill->refs); |
a7f0ed5a PB |
7212 | percpu_ref_kill(&rsrc_node->refs); |
7213 | ctx->rsrc_node = NULL; | |
7214 | } | |
6b06314c | 7215 | |
a7f0ed5a PB |
7216 | if (!ctx->rsrc_node) { |
7217 | ctx->rsrc_node = ctx->rsrc_backup_node; | |
7218 | ctx->rsrc_backup_node = NULL; | |
7219 | } | |
8bad28d8 HX |
7220 | } |
7221 | ||
a7f0ed5a | 7222 | static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx) |
8dd03afe PB |
7223 | { |
7224 | if (ctx->rsrc_backup_node) | |
7225 | return 0; | |
b895c9a6 | 7226 | ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx); |
8dd03afe | 7227 | return ctx->rsrc_backup_node ? 0 : -ENOMEM; |
8bad28d8 HX |
7228 | } |
7229 | ||
40ae0ff7 | 7230 | static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx) |
8bad28d8 HX |
7231 | { |
7232 | int ret; | |
05589553 | 7233 | |
215c3902 | 7234 | /* As we may drop ->uring_lock, other task may have started quiesce */ |
8bad28d8 HX |
7235 | if (data->quiesce) |
7236 | return -ENXIO; | |
05589553 | 7237 | |
8bad28d8 | 7238 | data->quiesce = true; |
1ffc5422 | 7239 | do { |
a7f0ed5a | 7240 | ret = io_rsrc_node_switch_start(ctx); |
8dd03afe | 7241 | if (ret) |
f2303b1f | 7242 | break; |
a7f0ed5a | 7243 | io_rsrc_node_switch(ctx, data); |
f2303b1f | 7244 | |
3e942498 PB |
7245 | /* kill initial ref, already quiesced if zero */ |
7246 | if (atomic_dec_and_test(&data->refs)) | |
7247 | break; | |
c018db4a | 7248 | mutex_unlock(&ctx->uring_lock); |
8bad28d8 | 7249 | flush_delayed_work(&ctx->rsrc_put_work); |
1ffc5422 | 7250 | ret = wait_for_completion_interruptible(&data->done); |
c018db4a JA |
7251 | if (!ret) { |
7252 | mutex_lock(&ctx->uring_lock); | |
1ffc5422 | 7253 | break; |
c018db4a | 7254 | } |
8bad28d8 | 7255 | |
3e942498 PB |
7256 | atomic_inc(&data->refs); |
7257 | /* wait for all works potentially completing data->done */ | |
7258 | flush_delayed_work(&ctx->rsrc_put_work); | |
cb5e1b81 | 7259 | reinit_completion(&data->done); |
8dd03afe | 7260 | |
1ffc5422 | 7261 | ret = io_run_task_work_sig(); |
8bad28d8 | 7262 | mutex_lock(&ctx->uring_lock); |
f2303b1f | 7263 | } while (ret >= 0); |
8bad28d8 | 7264 | data->quiesce = false; |
05f3fb3c | 7265 | |
8bad28d8 | 7266 | return ret; |
d7954b2b BM |
7267 | } |
7268 | ||
2d091d62 PB |
7269 | static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx) |
7270 | { | |
7271 | unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK; | |
7272 | unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT; | |
7273 | ||
7274 | return &data->tags[table_idx][off]; | |
7275 | } | |
7276 | ||
44b31f2f | 7277 | static void io_rsrc_data_free(struct io_rsrc_data *data) |
1ad555c6 | 7278 | { |
2d091d62 PB |
7279 | size_t size = data->nr * sizeof(data->tags[0][0]); |
7280 | ||
7281 | if (data->tags) | |
7282 | io_free_page_table((void **)data->tags, size); | |
44b31f2f PB |
7283 | kfree(data); |
7284 | } | |
7285 | ||
d878c816 PB |
7286 | static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put, |
7287 | u64 __user *utags, unsigned nr, | |
7288 | struct io_rsrc_data **pdata) | |
1ad555c6 | 7289 | { |
b895c9a6 | 7290 | struct io_rsrc_data *data; |
2d091d62 | 7291 | int ret = -ENOMEM; |
d878c816 | 7292 | unsigned i; |
1ad555c6 BM |
7293 | |
7294 | data = kzalloc(sizeof(*data), GFP_KERNEL); | |
7295 | if (!data) | |
d878c816 | 7296 | return -ENOMEM; |
2d091d62 | 7297 | data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); |
b60c8dce | 7298 | if (!data->tags) { |
1ad555c6 | 7299 | kfree(data); |
d878c816 PB |
7300 | return -ENOMEM; |
7301 | } | |
2d091d62 PB |
7302 | |
7303 | data->nr = nr; | |
7304 | data->ctx = ctx; | |
7305 | data->do_put = do_put; | |
d878c816 | 7306 | if (utags) { |
2d091d62 | 7307 | ret = -EFAULT; |
d878c816 | 7308 | for (i = 0; i < nr; i++) { |
fdd1dc31 CIK |
7309 | u64 *tag_slot = io_get_tag_slot(data, i); |
7310 | ||
7311 | if (copy_from_user(tag_slot, &utags[i], | |
7312 | sizeof(*tag_slot))) | |
2d091d62 | 7313 | goto fail; |
d878c816 | 7314 | } |
1ad555c6 | 7315 | } |
b60c8dce | 7316 | |
3e942498 | 7317 | atomic_set(&data->refs, 1); |
1ad555c6 | 7318 | init_completion(&data->done); |
d878c816 PB |
7319 | *pdata = data; |
7320 | return 0; | |
2d091d62 PB |
7321 | fail: |
7322 | io_rsrc_data_free(data); | |
7323 | return ret; | |
1ad555c6 BM |
7324 | } |
7325 | ||
9123c8ff PB |
7326 | static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files) |
7327 | { | |
042b0d85 | 7328 | table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL); |
9123c8ff PB |
7329 | return !!table->files; |
7330 | } | |
7331 | ||
042b0d85 | 7332 | static void io_free_file_tables(struct io_file_table *table) |
9123c8ff | 7333 | { |
042b0d85 | 7334 | kvfree(table->files); |
9123c8ff PB |
7335 | table->files = NULL; |
7336 | } | |
7337 | ||
fff4db76 | 7338 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 7339 | { |
fff4db76 PB |
7340 | #if defined(CONFIG_UNIX) |
7341 | if (ctx->ring_sock) { | |
7342 | struct sock *sock = ctx->ring_sock->sk; | |
7343 | struct sk_buff *skb; | |
7344 | ||
7345 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
7346 | kfree_skb(skb); | |
7347 | } | |
7348 | #else | |
7349 | int i; | |
7350 | ||
7351 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7352 | struct file *file; | |
7353 | ||
7354 | file = io_file_from_index(ctx, i); | |
7355 | if (file) | |
7356 | fput(file); | |
7357 | } | |
7358 | #endif | |
042b0d85 | 7359 | io_free_file_tables(&ctx->file_table); |
44b31f2f | 7360 | io_rsrc_data_free(ctx->file_data); |
fff4db76 PB |
7361 | ctx->file_data = NULL; |
7362 | ctx->nr_user_files = 0; | |
1ad555c6 BM |
7363 | } |
7364 | ||
d7954b2b BM |
7365 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
7366 | { | |
d7954b2b BM |
7367 | int ret; |
7368 | ||
08480400 | 7369 | if (!ctx->file_data) |
d7954b2b | 7370 | return -ENXIO; |
08480400 PB |
7371 | ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); |
7372 | if (!ret) | |
7373 | __io_sqe_files_unregister(ctx); | |
7374 | return ret; | |
6b06314c JA |
7375 | } |
7376 | ||
37d1e2e3 | 7377 | static void io_sq_thread_unpark(struct io_sq_data *sqd) |
09a6f4ef | 7378 | __releases(&sqd->lock) |
37d1e2e3 | 7379 | { |
521d6a73 PB |
7380 | WARN_ON_ONCE(sqd->thread == current); |
7381 | ||
9e138a48 PB |
7382 | /* |
7383 | * Do the dance but not conditional clear_bit() because it'd race with | |
7384 | * other threads incrementing park_pending and setting the bit. | |
7385 | */ | |
37d1e2e3 | 7386 | clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
9e138a48 PB |
7387 | if (atomic_dec_return(&sqd->park_pending)) |
7388 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); | |
09a6f4ef | 7389 | mutex_unlock(&sqd->lock); |
37d1e2e3 JA |
7390 | } |
7391 | ||
86e0d676 | 7392 | static void io_sq_thread_park(struct io_sq_data *sqd) |
09a6f4ef | 7393 | __acquires(&sqd->lock) |
37d1e2e3 | 7394 | { |
521d6a73 PB |
7395 | WARN_ON_ONCE(sqd->thread == current); |
7396 | ||
9e138a48 | 7397 | atomic_inc(&sqd->park_pending); |
86e0d676 | 7398 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
09a6f4ef | 7399 | mutex_lock(&sqd->lock); |
05962f95 | 7400 | if (sqd->thread) |
86e0d676 | 7401 | wake_up_process(sqd->thread); |
37d1e2e3 JA |
7402 | } |
7403 | ||
7404 | static void io_sq_thread_stop(struct io_sq_data *sqd) | |
7405 | { | |
521d6a73 | 7406 | WARN_ON_ONCE(sqd->thread == current); |
88885f66 | 7407 | WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)); |
521d6a73 | 7408 | |
05962f95 | 7409 | set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
88885f66 | 7410 | mutex_lock(&sqd->lock); |
e8f98f24 JA |
7411 | if (sqd->thread) |
7412 | wake_up_process(sqd->thread); | |
09a6f4ef | 7413 | mutex_unlock(&sqd->lock); |
05962f95 | 7414 | wait_for_completion(&sqd->exited); |
37d1e2e3 JA |
7415 | } |
7416 | ||
534ca6d6 | 7417 | static void io_put_sq_data(struct io_sq_data *sqd) |
6c271ce2 | 7418 | { |
534ca6d6 | 7419 | if (refcount_dec_and_test(&sqd->refs)) { |
9e138a48 PB |
7420 | WARN_ON_ONCE(atomic_read(&sqd->park_pending)); |
7421 | ||
37d1e2e3 JA |
7422 | io_sq_thread_stop(sqd); |
7423 | kfree(sqd); | |
7424 | } | |
7425 | } | |
7426 | ||
7427 | static void io_sq_thread_finish(struct io_ring_ctx *ctx) | |
7428 | { | |
7429 | struct io_sq_data *sqd = ctx->sq_data; | |
7430 | ||
7431 | if (sqd) { | |
05962f95 | 7432 | io_sq_thread_park(sqd); |
521d6a73 | 7433 | list_del_init(&ctx->sqd_list); |
37d1e2e3 | 7434 | io_sqd_update_thread_idle(sqd); |
05962f95 | 7435 | io_sq_thread_unpark(sqd); |
37d1e2e3 JA |
7436 | |
7437 | io_put_sq_data(sqd); | |
7438 | ctx->sq_data = NULL; | |
534ca6d6 JA |
7439 | } |
7440 | } | |
7441 | ||
aa06165d JA |
7442 | static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p) |
7443 | { | |
7444 | struct io_ring_ctx *ctx_attach; | |
7445 | struct io_sq_data *sqd; | |
7446 | struct fd f; | |
7447 | ||
7448 | f = fdget(p->wq_fd); | |
7449 | if (!f.file) | |
7450 | return ERR_PTR(-ENXIO); | |
7451 | if (f.file->f_op != &io_uring_fops) { | |
7452 | fdput(f); | |
7453 | return ERR_PTR(-EINVAL); | |
7454 | } | |
7455 | ||
7456 | ctx_attach = f.file->private_data; | |
7457 | sqd = ctx_attach->sq_data; | |
7458 | if (!sqd) { | |
7459 | fdput(f); | |
7460 | return ERR_PTR(-EINVAL); | |
7461 | } | |
5c2469e0 JA |
7462 | if (sqd->task_tgid != current->tgid) { |
7463 | fdput(f); | |
7464 | return ERR_PTR(-EPERM); | |
7465 | } | |
aa06165d JA |
7466 | |
7467 | refcount_inc(&sqd->refs); | |
7468 | fdput(f); | |
7469 | return sqd; | |
7470 | } | |
7471 | ||
26984fbf PB |
7472 | static struct io_sq_data *io_get_sq_data(struct io_uring_params *p, |
7473 | bool *attached) | |
534ca6d6 JA |
7474 | { |
7475 | struct io_sq_data *sqd; | |
7476 | ||
26984fbf | 7477 | *attached = false; |
5c2469e0 JA |
7478 | if (p->flags & IORING_SETUP_ATTACH_WQ) { |
7479 | sqd = io_attach_sq_data(p); | |
26984fbf PB |
7480 | if (!IS_ERR(sqd)) { |
7481 | *attached = true; | |
5c2469e0 | 7482 | return sqd; |
26984fbf | 7483 | } |
5c2469e0 JA |
7484 | /* fall through for EPERM case, setup new sqd/task */ |
7485 | if (PTR_ERR(sqd) != -EPERM) | |
7486 | return sqd; | |
7487 | } | |
aa06165d | 7488 | |
534ca6d6 JA |
7489 | sqd = kzalloc(sizeof(*sqd), GFP_KERNEL); |
7490 | if (!sqd) | |
7491 | return ERR_PTR(-ENOMEM); | |
7492 | ||
9e138a48 | 7493 | atomic_set(&sqd->park_pending, 0); |
534ca6d6 | 7494 | refcount_set(&sqd->refs, 1); |
69fb2131 | 7495 | INIT_LIST_HEAD(&sqd->ctx_list); |
09a6f4ef | 7496 | mutex_init(&sqd->lock); |
534ca6d6 | 7497 | init_waitqueue_head(&sqd->wait); |
37d1e2e3 | 7498 | init_completion(&sqd->exited); |
534ca6d6 JA |
7499 | return sqd; |
7500 | } | |
7501 | ||
6b06314c | 7502 | #if defined(CONFIG_UNIX) |
6b06314c JA |
7503 | /* |
7504 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
7505 | * the io_uring can be safely unregistered on process exit, even if we have | |
7506 | * loops in the file referencing. | |
7507 | */ | |
7508 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
7509 | { | |
7510 | struct sock *sk = ctx->ring_sock->sk; | |
7511 | struct scm_fp_list *fpl; | |
7512 | struct sk_buff *skb; | |
08a45173 | 7513 | int i, nr_files; |
6b06314c | 7514 | |
6b06314c JA |
7515 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); |
7516 | if (!fpl) | |
7517 | return -ENOMEM; | |
7518 | ||
7519 | skb = alloc_skb(0, GFP_KERNEL); | |
7520 | if (!skb) { | |
7521 | kfree(fpl); | |
7522 | return -ENOMEM; | |
7523 | } | |
7524 | ||
7525 | skb->sk = sk; | |
6b06314c | 7526 | |
08a45173 | 7527 | nr_files = 0; |
62e398be | 7528 | fpl->user = get_uid(current_user()); |
6b06314c | 7529 | for (i = 0; i < nr; i++) { |
65e19f54 JA |
7530 | struct file *file = io_file_from_index(ctx, i + offset); |
7531 | ||
7532 | if (!file) | |
08a45173 | 7533 | continue; |
65e19f54 | 7534 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
7535 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
7536 | nr_files++; | |
6b06314c JA |
7537 | } |
7538 | ||
08a45173 JA |
7539 | if (nr_files) { |
7540 | fpl->max = SCM_MAX_FD; | |
7541 | fpl->count = nr_files; | |
7542 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 7543 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
7544 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
7545 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 7546 | |
08a45173 JA |
7547 | for (i = 0; i < nr_files; i++) |
7548 | fput(fpl->fp[i]); | |
7549 | } else { | |
7550 | kfree_skb(skb); | |
7551 | kfree(fpl); | |
7552 | } | |
6b06314c JA |
7553 | |
7554 | return 0; | |
7555 | } | |
7556 | ||
7557 | /* | |
7558 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
7559 | * causes regular reference counting to break down. We rely on the UNIX | |
7560 | * garbage collection to take care of this problem for us. | |
7561 | */ | |
7562 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7563 | { | |
7564 | unsigned left, total; | |
7565 | int ret = 0; | |
7566 | ||
7567 | total = 0; | |
7568 | left = ctx->nr_user_files; | |
7569 | while (left) { | |
7570 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
7571 | |
7572 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
7573 | if (ret) | |
7574 | break; | |
7575 | left -= this_files; | |
7576 | total += this_files; | |
7577 | } | |
7578 | ||
7579 | if (!ret) | |
7580 | return 0; | |
7581 | ||
7582 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
7583 | struct file *file = io_file_from_index(ctx, total); |
7584 | ||
7585 | if (file) | |
7586 | fput(file); | |
6b06314c JA |
7587 | total++; |
7588 | } | |
7589 | ||
7590 | return ret; | |
7591 | } | |
7592 | #else | |
7593 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7594 | { | |
7595 | return 0; | |
7596 | } | |
7597 | #endif | |
7598 | ||
47e90392 | 7599 | static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) |
05f3fb3c | 7600 | { |
50238531 | 7601 | struct file *file = prsrc->file; |
05f3fb3c JA |
7602 | #if defined(CONFIG_UNIX) |
7603 | struct sock *sock = ctx->ring_sock->sk; | |
7604 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
7605 | struct sk_buff *skb; | |
7606 | int i; | |
7607 | ||
7608 | __skb_queue_head_init(&list); | |
7609 | ||
7610 | /* | |
7611 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
7612 | * remove this entry and rearrange the file array. | |
7613 | */ | |
7614 | skb = skb_dequeue(head); | |
7615 | while (skb) { | |
7616 | struct scm_fp_list *fp; | |
7617 | ||
7618 | fp = UNIXCB(skb).fp; | |
7619 | for (i = 0; i < fp->count; i++) { | |
7620 | int left; | |
7621 | ||
7622 | if (fp->fp[i] != file) | |
7623 | continue; | |
7624 | ||
7625 | unix_notinflight(fp->user, fp->fp[i]); | |
7626 | left = fp->count - 1 - i; | |
7627 | if (left) { | |
7628 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
7629 | left * sizeof(struct file *)); | |
7630 | } | |
7631 | fp->count--; | |
7632 | if (!fp->count) { | |
7633 | kfree_skb(skb); | |
7634 | skb = NULL; | |
7635 | } else { | |
7636 | __skb_queue_tail(&list, skb); | |
7637 | } | |
7638 | fput(file); | |
7639 | file = NULL; | |
7640 | break; | |
7641 | } | |
7642 | ||
7643 | if (!file) | |
7644 | break; | |
7645 | ||
7646 | __skb_queue_tail(&list, skb); | |
7647 | ||
7648 | skb = skb_dequeue(head); | |
7649 | } | |
7650 | ||
7651 | if (skb_peek(&list)) { | |
7652 | spin_lock_irq(&head->lock); | |
7653 | while ((skb = __skb_dequeue(&list)) != NULL) | |
7654 | __skb_queue_tail(head, skb); | |
7655 | spin_unlock_irq(&head->lock); | |
7656 | } | |
7657 | #else | |
7658 | fput(file); | |
7659 | #endif | |
7660 | } | |
7661 | ||
b895c9a6 | 7662 | static void __io_rsrc_put_work(struct io_rsrc_node *ref_node) |
65e19f54 | 7663 | { |
b895c9a6 | 7664 | struct io_rsrc_data *rsrc_data = ref_node->rsrc_data; |
269bbe5f BM |
7665 | struct io_ring_ctx *ctx = rsrc_data->ctx; |
7666 | struct io_rsrc_put *prsrc, *tmp; | |
05589553 | 7667 | |
269bbe5f BM |
7668 | list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) { |
7669 | list_del(&prsrc->list); | |
b60c8dce PB |
7670 | |
7671 | if (prsrc->tag) { | |
7672 | bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL; | |
b60c8dce PB |
7673 | |
7674 | io_ring_submit_lock(ctx, lock_ring); | |
79ebeaee | 7675 | spin_lock(&ctx->completion_lock); |
b60c8dce | 7676 | io_cqring_fill_event(ctx, prsrc->tag, 0, 0); |
2840f710 | 7677 | ctx->cq_extra++; |
b60c8dce | 7678 | io_commit_cqring(ctx); |
79ebeaee | 7679 | spin_unlock(&ctx->completion_lock); |
b60c8dce PB |
7680 | io_cqring_ev_posted(ctx); |
7681 | io_ring_submit_unlock(ctx, lock_ring); | |
7682 | } | |
7683 | ||
40ae0ff7 | 7684 | rsrc_data->do_put(ctx, prsrc); |
269bbe5f | 7685 | kfree(prsrc); |
65e19f54 | 7686 | } |
05589553 | 7687 | |
28a9fe25 | 7688 | io_rsrc_node_destroy(ref_node); |
3e942498 PB |
7689 | if (atomic_dec_and_test(&rsrc_data->refs)) |
7690 | complete(&rsrc_data->done); | |
2faf852d | 7691 | } |
65e19f54 | 7692 | |
269bbe5f | 7693 | static void io_rsrc_put_work(struct work_struct *work) |
4a38aed2 JA |
7694 | { |
7695 | struct io_ring_ctx *ctx; | |
7696 | struct llist_node *node; | |
7697 | ||
269bbe5f BM |
7698 | ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work); |
7699 | node = llist_del_all(&ctx->rsrc_put_llist); | |
4a38aed2 JA |
7700 | |
7701 | while (node) { | |
b895c9a6 | 7702 | struct io_rsrc_node *ref_node; |
4a38aed2 JA |
7703 | struct llist_node *next = node->next; |
7704 | ||
b895c9a6 | 7705 | ref_node = llist_entry(node, struct io_rsrc_node, llist); |
269bbe5f | 7706 | __io_rsrc_put_work(ref_node); |
4a38aed2 JA |
7707 | node = next; |
7708 | } | |
7709 | } | |
7710 | ||
6b06314c | 7711 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
792e3582 | 7712 | unsigned nr_args, u64 __user *tags) |
6b06314c JA |
7713 | { |
7714 | __s32 __user *fds = (__s32 __user *) arg; | |
05f3fb3c | 7715 | struct file *file; |
f3baed39 | 7716 | int fd, ret; |
846a4ef2 | 7717 | unsigned i; |
6b06314c | 7718 | |
05f3fb3c | 7719 | if (ctx->file_data) |
6b06314c JA |
7720 | return -EBUSY; |
7721 | if (!nr_args) | |
7722 | return -EINVAL; | |
7723 | if (nr_args > IORING_MAX_FIXED_FILES) | |
7724 | return -EMFILE; | |
a7f0ed5a | 7725 | ret = io_rsrc_node_switch_start(ctx); |
f3baed39 PB |
7726 | if (ret) |
7727 | return ret; | |
d878c816 PB |
7728 | ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, |
7729 | &ctx->file_data); | |
7730 | if (ret) | |
7731 | return ret; | |
6b06314c | 7732 | |
f3baed39 | 7733 | ret = -ENOMEM; |
aeca241b | 7734 | if (!io_alloc_file_tables(&ctx->file_table, nr_args)) |
1ad555c6 | 7735 | goto out_free; |
65e19f54 | 7736 | |
08a45173 | 7737 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
d878c816 | 7738 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) { |
600cf3f8 PB |
7739 | ret = -EFAULT; |
7740 | goto out_fput; | |
7741 | } | |
08a45173 | 7742 | /* allow sparse sets */ |
792e3582 PB |
7743 | if (fd == -1) { |
7744 | ret = -EINVAL; | |
2d091d62 | 7745 | if (unlikely(*io_get_tag_slot(ctx->file_data, i))) |
792e3582 | 7746 | goto out_fput; |
08a45173 | 7747 | continue; |
792e3582 | 7748 | } |
6b06314c | 7749 | |
05f3fb3c | 7750 | file = fget(fd); |
6b06314c | 7751 | ret = -EBADF; |
792e3582 | 7752 | if (unlikely(!file)) |
600cf3f8 | 7753 | goto out_fput; |
05f3fb3c | 7754 | |
6b06314c JA |
7755 | /* |
7756 | * Don't allow io_uring instances to be registered. If UNIX | |
7757 | * isn't enabled, then this causes a reference cycle and this | |
7758 | * instance can never get freed. If UNIX is enabled we'll | |
7759 | * handle it just fine, but there's still no point in allowing | |
7760 | * a ring fd as it doesn't support regular read/write anyway. | |
7761 | */ | |
05f3fb3c JA |
7762 | if (file->f_op == &io_uring_fops) { |
7763 | fput(file); | |
600cf3f8 | 7764 | goto out_fput; |
6b06314c | 7765 | } |
aeca241b | 7766 | io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file); |
6b06314c JA |
7767 | } |
7768 | ||
6b06314c | 7769 | ret = io_sqe_files_scm(ctx); |
05589553 | 7770 | if (ret) { |
08480400 | 7771 | __io_sqe_files_unregister(ctx); |
05589553 XW |
7772 | return ret; |
7773 | } | |
6b06314c | 7774 | |
a7f0ed5a | 7775 | io_rsrc_node_switch(ctx, NULL); |
6b06314c | 7776 | return ret; |
600cf3f8 PB |
7777 | out_fput: |
7778 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7779 | file = io_file_from_index(ctx, i); | |
7780 | if (file) | |
7781 | fput(file); | |
7782 | } | |
042b0d85 | 7783 | io_free_file_tables(&ctx->file_table); |
600cf3f8 | 7784 | ctx->nr_user_files = 0; |
600cf3f8 | 7785 | out_free: |
44b31f2f | 7786 | io_rsrc_data_free(ctx->file_data); |
55cbc256 | 7787 | ctx->file_data = NULL; |
6b06314c JA |
7788 | return ret; |
7789 | } | |
7790 | ||
c3a31e60 JA |
7791 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
7792 | int index) | |
7793 | { | |
7794 | #if defined(CONFIG_UNIX) | |
7795 | struct sock *sock = ctx->ring_sock->sk; | |
7796 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
7797 | struct sk_buff *skb; | |
7798 | ||
7799 | /* | |
7800 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
7801 | * file set. If there's no room, fall back to allocating a new skb | |
7802 | * and filling it in. | |
7803 | */ | |
7804 | spin_lock_irq(&head->lock); | |
7805 | skb = skb_peek(head); | |
7806 | if (skb) { | |
7807 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
7808 | ||
7809 | if (fpl->count < SCM_MAX_FD) { | |
7810 | __skb_unlink(skb, head); | |
7811 | spin_unlock_irq(&head->lock); | |
7812 | fpl->fp[fpl->count] = get_file(file); | |
7813 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
7814 | fpl->count++; | |
7815 | spin_lock_irq(&head->lock); | |
7816 | __skb_queue_head(head, skb); | |
7817 | } else { | |
7818 | skb = NULL; | |
7819 | } | |
7820 | } | |
7821 | spin_unlock_irq(&head->lock); | |
7822 | ||
7823 | if (skb) { | |
7824 | fput(file); | |
7825 | return 0; | |
7826 | } | |
7827 | ||
7828 | return __io_sqe_files_scm(ctx, 1, index); | |
7829 | #else | |
7830 | return 0; | |
7831 | #endif | |
7832 | } | |
7833 | ||
b60c8dce | 7834 | static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, |
e7c78371 | 7835 | struct io_rsrc_node *node, void *rsrc) |
05f3fb3c | 7836 | { |
269bbe5f | 7837 | struct io_rsrc_put *prsrc; |
05f3fb3c | 7838 | |
269bbe5f BM |
7839 | prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL); |
7840 | if (!prsrc) | |
a5318d3c | 7841 | return -ENOMEM; |
05f3fb3c | 7842 | |
2d091d62 | 7843 | prsrc->tag = *io_get_tag_slot(data, idx); |
50238531 | 7844 | prsrc->rsrc = rsrc; |
e7c78371 | 7845 | list_add(&prsrc->list, &node->rsrc_list); |
a5318d3c | 7846 | return 0; |
05f3fb3c JA |
7847 | } |
7848 | ||
7849 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
c3bdad02 | 7850 | struct io_uring_rsrc_update2 *up, |
05f3fb3c JA |
7851 | unsigned nr_args) |
7852 | { | |
c3bdad02 | 7853 | u64 __user *tags = u64_to_user_ptr(up->tags); |
98f0b3b4 | 7854 | __s32 __user *fds = u64_to_user_ptr(up->data); |
b895c9a6 | 7855 | struct io_rsrc_data *data = ctx->file_data; |
a04b0ac0 PB |
7856 | struct io_fixed_file *file_slot; |
7857 | struct file *file; | |
98f0b3b4 PB |
7858 | int fd, i, err = 0; |
7859 | unsigned int done; | |
05589553 | 7860 | bool needs_switch = false; |
c3a31e60 | 7861 | |
98f0b3b4 PB |
7862 | if (!ctx->file_data) |
7863 | return -ENXIO; | |
7864 | if (up->offset + nr_args > ctx->nr_user_files) | |
c3a31e60 JA |
7865 | return -EINVAL; |
7866 | ||
67973b93 | 7867 | for (done = 0; done < nr_args; done++) { |
c3bdad02 PB |
7868 | u64 tag = 0; |
7869 | ||
7870 | if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || | |
7871 | copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
c3a31e60 JA |
7872 | err = -EFAULT; |
7873 | break; | |
7874 | } | |
c3bdad02 PB |
7875 | if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { |
7876 | err = -EINVAL; | |
7877 | break; | |
7878 | } | |
4e0377a1 | 7879 | if (fd == IORING_REGISTER_FILES_SKIP) |
7880 | continue; | |
7881 | ||
67973b93 | 7882 | i = array_index_nospec(up->offset + done, ctx->nr_user_files); |
aeca241b | 7883 | file_slot = io_fixed_file_slot(&ctx->file_table, i); |
ea64ec02 | 7884 | |
a04b0ac0 PB |
7885 | if (file_slot->file_ptr) { |
7886 | file = (struct file *)(file_slot->file_ptr & FFS_MASK); | |
b60c8dce PB |
7887 | err = io_queue_rsrc_removal(data, up->offset + done, |
7888 | ctx->rsrc_node, file); | |
a5318d3c HD |
7889 | if (err) |
7890 | break; | |
a04b0ac0 | 7891 | file_slot->file_ptr = 0; |
05589553 | 7892 | needs_switch = true; |
c3a31e60 JA |
7893 | } |
7894 | if (fd != -1) { | |
c3a31e60 JA |
7895 | file = fget(fd); |
7896 | if (!file) { | |
7897 | err = -EBADF; | |
7898 | break; | |
7899 | } | |
7900 | /* | |
7901 | * Don't allow io_uring instances to be registered. If | |
7902 | * UNIX isn't enabled, then this causes a reference | |
7903 | * cycle and this instance can never get freed. If UNIX | |
7904 | * is enabled we'll handle it just fine, but there's | |
7905 | * still no point in allowing a ring fd as it doesn't | |
7906 | * support regular read/write anyway. | |
7907 | */ | |
7908 | if (file->f_op == &io_uring_fops) { | |
7909 | fput(file); | |
7910 | err = -EBADF; | |
7911 | break; | |
7912 | } | |
2d091d62 | 7913 | *io_get_tag_slot(data, up->offset + done) = tag; |
9a321c98 | 7914 | io_fixed_file_set(file_slot, file); |
c3a31e60 | 7915 | err = io_sqe_file_register(ctx, file, i); |
f3bd9dae | 7916 | if (err) { |
a04b0ac0 | 7917 | file_slot->file_ptr = 0; |
f3bd9dae | 7918 | fput(file); |
c3a31e60 | 7919 | break; |
f3bd9dae | 7920 | } |
c3a31e60 | 7921 | } |
05f3fb3c JA |
7922 | } |
7923 | ||
a7f0ed5a PB |
7924 | if (needs_switch) |
7925 | io_rsrc_node_switch(ctx, data); | |
c3a31e60 JA |
7926 | return done ? done : err; |
7927 | } | |
05589553 | 7928 | |
685fe7fe JA |
7929 | static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx, |
7930 | struct task_struct *task) | |
24369c2e | 7931 | { |
e941894e | 7932 | struct io_wq_hash *hash; |
24369c2e | 7933 | struct io_wq_data data; |
24369c2e | 7934 | unsigned int concurrency; |
24369c2e | 7935 | |
362a9e65 | 7936 | mutex_lock(&ctx->uring_lock); |
e941894e JA |
7937 | hash = ctx->hash_map; |
7938 | if (!hash) { | |
7939 | hash = kzalloc(sizeof(*hash), GFP_KERNEL); | |
362a9e65 YY |
7940 | if (!hash) { |
7941 | mutex_unlock(&ctx->uring_lock); | |
e941894e | 7942 | return ERR_PTR(-ENOMEM); |
362a9e65 | 7943 | } |
e941894e JA |
7944 | refcount_set(&hash->refs, 1); |
7945 | init_waitqueue_head(&hash->wait); | |
7946 | ctx->hash_map = hash; | |
24369c2e | 7947 | } |
362a9e65 | 7948 | mutex_unlock(&ctx->uring_lock); |
24369c2e | 7949 | |
e941894e | 7950 | data.hash = hash; |
685fe7fe | 7951 | data.task = task; |
ebc11b6c | 7952 | data.free_work = io_wq_free_work; |
f5fa38c5 | 7953 | data.do_work = io_wq_submit_work; |
24369c2e | 7954 | |
d25e3a3d JA |
7955 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
7956 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
24369c2e | 7957 | |
5aa75ed5 | 7958 | return io_wq_create(concurrency, &data); |
24369c2e PB |
7959 | } |
7960 | ||
5aa75ed5 JA |
7961 | static int io_uring_alloc_task_context(struct task_struct *task, |
7962 | struct io_ring_ctx *ctx) | |
0f212204 JA |
7963 | { |
7964 | struct io_uring_task *tctx; | |
d8a6df10 | 7965 | int ret; |
0f212204 | 7966 | |
09899b19 | 7967 | tctx = kzalloc(sizeof(*tctx), GFP_KERNEL); |
0f212204 JA |
7968 | if (unlikely(!tctx)) |
7969 | return -ENOMEM; | |
7970 | ||
d8a6df10 JA |
7971 | ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL); |
7972 | if (unlikely(ret)) { | |
7973 | kfree(tctx); | |
7974 | return ret; | |
7975 | } | |
7976 | ||
685fe7fe | 7977 | tctx->io_wq = io_init_wq_offload(ctx, task); |
5aa75ed5 JA |
7978 | if (IS_ERR(tctx->io_wq)) { |
7979 | ret = PTR_ERR(tctx->io_wq); | |
7980 | percpu_counter_destroy(&tctx->inflight); | |
7981 | kfree(tctx); | |
7982 | return ret; | |
7983 | } | |
7984 | ||
0f212204 JA |
7985 | xa_init(&tctx->xa); |
7986 | init_waitqueue_head(&tctx->wait); | |
fdaf083c | 7987 | atomic_set(&tctx->in_idle, 0); |
b303fe2e | 7988 | atomic_set(&tctx->inflight_tracked, 0); |
0f212204 | 7989 | task->io_uring = tctx; |
7cbf1722 JA |
7990 | spin_lock_init(&tctx->task_lock); |
7991 | INIT_WQ_LIST(&tctx->task_list); | |
7cbf1722 | 7992 | init_task_work(&tctx->task_work, tctx_task_work); |
0f212204 JA |
7993 | return 0; |
7994 | } | |
7995 | ||
7996 | void __io_uring_free(struct task_struct *tsk) | |
7997 | { | |
7998 | struct io_uring_task *tctx = tsk->io_uring; | |
7999 | ||
8000 | WARN_ON_ONCE(!xa_empty(&tctx->xa)); | |
ef8eaa4e | 8001 | WARN_ON_ONCE(tctx->io_wq); |
09899b19 | 8002 | WARN_ON_ONCE(tctx->cached_refs); |
ef8eaa4e | 8003 | |
d8a6df10 | 8004 | percpu_counter_destroy(&tctx->inflight); |
0f212204 JA |
8005 | kfree(tctx); |
8006 | tsk->io_uring = NULL; | |
8007 | } | |
8008 | ||
7e84e1c7 SG |
8009 | static int io_sq_offload_create(struct io_ring_ctx *ctx, |
8010 | struct io_uring_params *p) | |
2b188cc1 JA |
8011 | { |
8012 | int ret; | |
8013 | ||
d25e3a3d JA |
8014 | /* Retain compatibility with failing for an invalid attach attempt */ |
8015 | if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) == | |
8016 | IORING_SETUP_ATTACH_WQ) { | |
8017 | struct fd f; | |
8018 | ||
8019 | f = fdget(p->wq_fd); | |
8020 | if (!f.file) | |
8021 | return -ENXIO; | |
0cc936f7 JA |
8022 | if (f.file->f_op != &io_uring_fops) { |
8023 | fdput(f); | |
f2a48dd0 | 8024 | return -EINVAL; |
0cc936f7 JA |
8025 | } |
8026 | fdput(f); | |
d25e3a3d | 8027 | } |
6c271ce2 | 8028 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
46fe18b1 | 8029 | struct task_struct *tsk; |
534ca6d6 | 8030 | struct io_sq_data *sqd; |
26984fbf | 8031 | bool attached; |
534ca6d6 | 8032 | |
26984fbf | 8033 | sqd = io_get_sq_data(p, &attached); |
534ca6d6 JA |
8034 | if (IS_ERR(sqd)) { |
8035 | ret = PTR_ERR(sqd); | |
8036 | goto err; | |
8037 | } | |
69fb2131 | 8038 | |
7c30f36a | 8039 | ctx->sq_creds = get_current_cred(); |
534ca6d6 | 8040 | ctx->sq_data = sqd; |
917257da JA |
8041 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
8042 | if (!ctx->sq_thread_idle) | |
8043 | ctx->sq_thread_idle = HZ; | |
8044 | ||
78d7f6ba | 8045 | io_sq_thread_park(sqd); |
de75a3d3 PB |
8046 | list_add(&ctx->sqd_list, &sqd->ctx_list); |
8047 | io_sqd_update_thread_idle(sqd); | |
26984fbf | 8048 | /* don't attach to a dying SQPOLL thread, would be racy */ |
f2a48dd0 | 8049 | ret = (attached && !sqd->thread) ? -ENXIO : 0; |
78d7f6ba PB |
8050 | io_sq_thread_unpark(sqd); |
8051 | ||
de75a3d3 PB |
8052 | if (ret < 0) |
8053 | goto err; | |
8054 | if (attached) | |
5aa75ed5 | 8055 | return 0; |
aa06165d | 8056 | |
6c271ce2 | 8057 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 8058 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 8059 | |
917257da | 8060 | ret = -EINVAL; |
f2a48dd0 | 8061 | if (cpu >= nr_cpu_ids || !cpu_online(cpu)) |
e8f98f24 | 8062 | goto err_sqpoll; |
37d1e2e3 | 8063 | sqd->sq_cpu = cpu; |
6c271ce2 | 8064 | } else { |
37d1e2e3 | 8065 | sqd->sq_cpu = -1; |
6c271ce2 | 8066 | } |
37d1e2e3 JA |
8067 | |
8068 | sqd->task_pid = current->pid; | |
5c2469e0 | 8069 | sqd->task_tgid = current->tgid; |
46fe18b1 JA |
8070 | tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE); |
8071 | if (IS_ERR(tsk)) { | |
8072 | ret = PTR_ERR(tsk); | |
e8f98f24 | 8073 | goto err_sqpoll; |
6c271ce2 | 8074 | } |
97a73a0f | 8075 | |
46fe18b1 | 8076 | sqd->thread = tsk; |
97a73a0f | 8077 | ret = io_uring_alloc_task_context(tsk, ctx); |
46fe18b1 | 8078 | wake_up_new_task(tsk); |
0f212204 JA |
8079 | if (ret) |
8080 | goto err; | |
6c271ce2 JA |
8081 | } else if (p->flags & IORING_SETUP_SQ_AFF) { |
8082 | /* Can't have SQ_AFF without SQPOLL */ | |
8083 | ret = -EINVAL; | |
8084 | goto err; | |
8085 | } | |
8086 | ||
2b188cc1 | 8087 | return 0; |
f2a48dd0 PB |
8088 | err_sqpoll: |
8089 | complete(&ctx->sq_data->exited); | |
2b188cc1 | 8090 | err: |
37d1e2e3 | 8091 | io_sq_thread_finish(ctx); |
2b188cc1 JA |
8092 | return ret; |
8093 | } | |
8094 | ||
a087e2b5 BM |
8095 | static inline void __io_unaccount_mem(struct user_struct *user, |
8096 | unsigned long nr_pages) | |
2b188cc1 JA |
8097 | { |
8098 | atomic_long_sub(nr_pages, &user->locked_vm); | |
8099 | } | |
8100 | ||
a087e2b5 BM |
8101 | static inline int __io_account_mem(struct user_struct *user, |
8102 | unsigned long nr_pages) | |
2b188cc1 JA |
8103 | { |
8104 | unsigned long page_limit, cur_pages, new_pages; | |
8105 | ||
8106 | /* Don't allow more pages than we can safely lock */ | |
8107 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
8108 | ||
8109 | do { | |
8110 | cur_pages = atomic_long_read(&user->locked_vm); | |
8111 | new_pages = cur_pages + nr_pages; | |
8112 | if (new_pages > page_limit) | |
8113 | return -ENOMEM; | |
8114 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
8115 | new_pages) != cur_pages); | |
8116 | ||
8117 | return 0; | |
8118 | } | |
8119 | ||
26bfa89e | 8120 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) |
a087e2b5 | 8121 | { |
62e398be | 8122 | if (ctx->user) |
a087e2b5 | 8123 | __io_unaccount_mem(ctx->user, nr_pages); |
30975825 | 8124 | |
26bfa89e JA |
8125 | if (ctx->mm_account) |
8126 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); | |
a087e2b5 BM |
8127 | } |
8128 | ||
26bfa89e | 8129 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) |
a087e2b5 | 8130 | { |
30975825 BM |
8131 | int ret; |
8132 | ||
62e398be | 8133 | if (ctx->user) { |
30975825 BM |
8134 | ret = __io_account_mem(ctx->user, nr_pages); |
8135 | if (ret) | |
8136 | return ret; | |
8137 | } | |
8138 | ||
26bfa89e JA |
8139 | if (ctx->mm_account) |
8140 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); | |
a087e2b5 BM |
8141 | |
8142 | return 0; | |
8143 | } | |
8144 | ||
2b188cc1 JA |
8145 | static void io_mem_free(void *ptr) |
8146 | { | |
52e04ef4 MR |
8147 | struct page *page; |
8148 | ||
8149 | if (!ptr) | |
8150 | return; | |
2b188cc1 | 8151 | |
52e04ef4 | 8152 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
8153 | if (put_page_testzero(page)) |
8154 | free_compound_page(page); | |
8155 | } | |
8156 | ||
8157 | static void *io_mem_alloc(size_t size) | |
8158 | { | |
8159 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
26bfa89e | 8160 | __GFP_NORETRY | __GFP_ACCOUNT; |
2b188cc1 JA |
8161 | |
8162 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
8163 | } | |
8164 | ||
75b28aff HV |
8165 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
8166 | size_t *sq_offset) | |
8167 | { | |
8168 | struct io_rings *rings; | |
8169 | size_t off, sq_array_size; | |
8170 | ||
8171 | off = struct_size(rings, cqes, cq_entries); | |
8172 | if (off == SIZE_MAX) | |
8173 | return SIZE_MAX; | |
8174 | ||
8175 | #ifdef CONFIG_SMP | |
8176 | off = ALIGN(off, SMP_CACHE_BYTES); | |
8177 | if (off == 0) | |
8178 | return SIZE_MAX; | |
8179 | #endif | |
8180 | ||
b36200f5 DV |
8181 | if (sq_offset) |
8182 | *sq_offset = off; | |
8183 | ||
75b28aff HV |
8184 | sq_array_size = array_size(sizeof(u32), sq_entries); |
8185 | if (sq_array_size == SIZE_MAX) | |
8186 | return SIZE_MAX; | |
8187 | ||
8188 | if (check_add_overflow(off, sq_array_size, &off)) | |
8189 | return SIZE_MAX; | |
8190 | ||
75b28aff HV |
8191 | return off; |
8192 | } | |
8193 | ||
41edf1a5 | 8194 | static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot) |
7f61a1e9 | 8195 | { |
41edf1a5 | 8196 | struct io_mapped_ubuf *imu = *slot; |
7f61a1e9 PB |
8197 | unsigned int i; |
8198 | ||
6224843d PB |
8199 | if (imu != ctx->dummy_ubuf) { |
8200 | for (i = 0; i < imu->nr_bvecs; i++) | |
8201 | unpin_user_page(imu->bvec[i].bv_page); | |
8202 | if (imu->acct_pages) | |
8203 | io_unaccount_mem(ctx, imu->acct_pages); | |
8204 | kvfree(imu); | |
8205 | } | |
41edf1a5 | 8206 | *slot = NULL; |
7f61a1e9 PB |
8207 | } |
8208 | ||
bd54b6fe | 8209 | static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) |
edafccee | 8210 | { |
634d00df PB |
8211 | io_buffer_unmap(ctx, &prsrc->buf); |
8212 | prsrc->buf = NULL; | |
bd54b6fe | 8213 | } |
edafccee | 8214 | |
bd54b6fe BM |
8215 | static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) |
8216 | { | |
8217 | unsigned int i; | |
edafccee | 8218 | |
7f61a1e9 PB |
8219 | for (i = 0; i < ctx->nr_user_bufs; i++) |
8220 | io_buffer_unmap(ctx, &ctx->user_bufs[i]); | |
edafccee | 8221 | kfree(ctx->user_bufs); |
bb6659cc | 8222 | io_rsrc_data_free(ctx->buf_data); |
edafccee | 8223 | ctx->user_bufs = NULL; |
bd54b6fe | 8224 | ctx->buf_data = NULL; |
edafccee | 8225 | ctx->nr_user_bufs = 0; |
bd54b6fe BM |
8226 | } |
8227 | ||
0a96bbe4 | 8228 | static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) |
edafccee | 8229 | { |
bd54b6fe | 8230 | int ret; |
edafccee | 8231 | |
bd54b6fe | 8232 | if (!ctx->buf_data) |
edafccee JA |
8233 | return -ENXIO; |
8234 | ||
bd54b6fe BM |
8235 | ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); |
8236 | if (!ret) | |
8237 | __io_sqe_buffers_unregister(ctx); | |
8238 | return ret; | |
edafccee JA |
8239 | } |
8240 | ||
8241 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
8242 | void __user *arg, unsigned index) | |
8243 | { | |
8244 | struct iovec __user *src; | |
8245 | ||
8246 | #ifdef CONFIG_COMPAT | |
8247 | if (ctx->compat) { | |
8248 | struct compat_iovec __user *ciovs; | |
8249 | struct compat_iovec ciov; | |
8250 | ||
8251 | ciovs = (struct compat_iovec __user *) arg; | |
8252 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
8253 | return -EFAULT; | |
8254 | ||
d55e5f5b | 8255 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
8256 | dst->iov_len = ciov.iov_len; |
8257 | return 0; | |
8258 | } | |
8259 | #endif | |
8260 | src = (struct iovec __user *) arg; | |
8261 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
8262 | return -EFAULT; | |
8263 | return 0; | |
8264 | } | |
8265 | ||
de293938 JA |
8266 | /* |
8267 | * Not super efficient, but this is just a registration time. And we do cache | |
8268 | * the last compound head, so generally we'll only do a full search if we don't | |
8269 | * match that one. | |
8270 | * | |
8271 | * We check if the given compound head page has already been accounted, to | |
8272 | * avoid double accounting it. This allows us to account the full size of the | |
8273 | * page, not just the constituent pages of a huge page. | |
8274 | */ | |
8275 | static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, | |
8276 | int nr_pages, struct page *hpage) | |
8277 | { | |
8278 | int i, j; | |
8279 | ||
8280 | /* check current page array */ | |
8281 | for (i = 0; i < nr_pages; i++) { | |
8282 | if (!PageCompound(pages[i])) | |
8283 | continue; | |
8284 | if (compound_head(pages[i]) == hpage) | |
8285 | return true; | |
8286 | } | |
8287 | ||
8288 | /* check previously registered pages */ | |
8289 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
41edf1a5 | 8290 | struct io_mapped_ubuf *imu = ctx->user_bufs[i]; |
de293938 JA |
8291 | |
8292 | for (j = 0; j < imu->nr_bvecs; j++) { | |
8293 | if (!PageCompound(imu->bvec[j].bv_page)) | |
8294 | continue; | |
8295 | if (compound_head(imu->bvec[j].bv_page) == hpage) | |
8296 | return true; | |
8297 | } | |
8298 | } | |
8299 | ||
8300 | return false; | |
8301 | } | |
8302 | ||
8303 | static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, | |
8304 | int nr_pages, struct io_mapped_ubuf *imu, | |
8305 | struct page **last_hpage) | |
8306 | { | |
8307 | int i, ret; | |
8308 | ||
216e5835 | 8309 | imu->acct_pages = 0; |
de293938 JA |
8310 | for (i = 0; i < nr_pages; i++) { |
8311 | if (!PageCompound(pages[i])) { | |
8312 | imu->acct_pages++; | |
8313 | } else { | |
8314 | struct page *hpage; | |
8315 | ||
8316 | hpage = compound_head(pages[i]); | |
8317 | if (hpage == *last_hpage) | |
8318 | continue; | |
8319 | *last_hpage = hpage; | |
8320 | if (headpage_already_acct(ctx, pages, i, hpage)) | |
8321 | continue; | |
8322 | imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; | |
8323 | } | |
8324 | } | |
8325 | ||
8326 | if (!imu->acct_pages) | |
8327 | return 0; | |
8328 | ||
26bfa89e | 8329 | ret = io_account_mem(ctx, imu->acct_pages); |
de293938 JA |
8330 | if (ret) |
8331 | imu->acct_pages = 0; | |
8332 | return ret; | |
8333 | } | |
8334 | ||
0a96bbe4 | 8335 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, |
41edf1a5 | 8336 | struct io_mapped_ubuf **pimu, |
0a96bbe4 | 8337 | struct page **last_hpage) |
edafccee | 8338 | { |
41edf1a5 | 8339 | struct io_mapped_ubuf *imu = NULL; |
edafccee JA |
8340 | struct vm_area_struct **vmas = NULL; |
8341 | struct page **pages = NULL; | |
0a96bbe4 BM |
8342 | unsigned long off, start, end, ubuf; |
8343 | size_t size; | |
8344 | int ret, pret, nr_pages, i; | |
8345 | ||
6224843d PB |
8346 | if (!iov->iov_base) { |
8347 | *pimu = ctx->dummy_ubuf; | |
8348 | return 0; | |
8349 | } | |
8350 | ||
0a96bbe4 BM |
8351 | ubuf = (unsigned long) iov->iov_base; |
8352 | end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
8353 | start = ubuf >> PAGE_SHIFT; | |
8354 | nr_pages = end - start; | |
8355 | ||
41edf1a5 | 8356 | *pimu = NULL; |
0a96bbe4 BM |
8357 | ret = -ENOMEM; |
8358 | ||
8359 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
8360 | if (!pages) | |
8361 | goto done; | |
8362 | ||
8363 | vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), | |
8364 | GFP_KERNEL); | |
8365 | if (!vmas) | |
8366 | goto done; | |
edafccee | 8367 | |
41edf1a5 | 8368 | imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); |
a2b4198c | 8369 | if (!imu) |
0a96bbe4 BM |
8370 | goto done; |
8371 | ||
8372 | ret = 0; | |
8373 | mmap_read_lock(current->mm); | |
8374 | pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
8375 | pages, vmas); | |
8376 | if (pret == nr_pages) { | |
8377 | /* don't support file backed memory */ | |
8378 | for (i = 0; i < nr_pages; i++) { | |
8379 | struct vm_area_struct *vma = vmas[i]; | |
8380 | ||
40dad765 PB |
8381 | if (vma_is_shmem(vma)) |
8382 | continue; | |
0a96bbe4 BM |
8383 | if (vma->vm_file && |
8384 | !is_file_hugepages(vma->vm_file)) { | |
8385 | ret = -EOPNOTSUPP; | |
8386 | break; | |
8387 | } | |
8388 | } | |
8389 | } else { | |
8390 | ret = pret < 0 ? pret : -EFAULT; | |
8391 | } | |
8392 | mmap_read_unlock(current->mm); | |
8393 | if (ret) { | |
8394 | /* | |
8395 | * if we did partial map, or found file backed vmas, | |
8396 | * release any pages we did get | |
8397 | */ | |
8398 | if (pret > 0) | |
8399 | unpin_user_pages(pages, pret); | |
0a96bbe4 BM |
8400 | goto done; |
8401 | } | |
8402 | ||
8403 | ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage); | |
8404 | if (ret) { | |
8405 | unpin_user_pages(pages, pret); | |
0a96bbe4 BM |
8406 | goto done; |
8407 | } | |
8408 | ||
8409 | off = ubuf & ~PAGE_MASK; | |
8410 | size = iov->iov_len; | |
8411 | for (i = 0; i < nr_pages; i++) { | |
8412 | size_t vec_len; | |
8413 | ||
8414 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
8415 | imu->bvec[i].bv_page = pages[i]; | |
8416 | imu->bvec[i].bv_len = vec_len; | |
8417 | imu->bvec[i].bv_offset = off; | |
8418 | off = 0; | |
8419 | size -= vec_len; | |
8420 | } | |
8421 | /* store original address for later verification */ | |
8422 | imu->ubuf = ubuf; | |
4751f53d | 8423 | imu->ubuf_end = ubuf + iov->iov_len; |
0a96bbe4 | 8424 | imu->nr_bvecs = nr_pages; |
41edf1a5 | 8425 | *pimu = imu; |
0a96bbe4 BM |
8426 | ret = 0; |
8427 | done: | |
41edf1a5 PB |
8428 | if (ret) |
8429 | kvfree(imu); | |
0a96bbe4 BM |
8430 | kvfree(pages); |
8431 | kvfree(vmas); | |
8432 | return ret; | |
8433 | } | |
8434 | ||
2b358604 | 8435 | static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) |
0a96bbe4 | 8436 | { |
87094465 PB |
8437 | ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); |
8438 | return ctx->user_bufs ? 0 : -ENOMEM; | |
2b358604 | 8439 | } |
edafccee | 8440 | |
2b358604 BM |
8441 | static int io_buffer_validate(struct iovec *iov) |
8442 | { | |
50e96989 PB |
8443 | unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1); |
8444 | ||
2b358604 BM |
8445 | /* |
8446 | * Don't impose further limits on the size and buffer | |
8447 | * constraints here, we'll -EINVAL later when IO is | |
8448 | * submitted if they are wrong. | |
8449 | */ | |
6224843d PB |
8450 | if (!iov->iov_base) |
8451 | return iov->iov_len ? -EFAULT : 0; | |
8452 | if (!iov->iov_len) | |
2b358604 | 8453 | return -EFAULT; |
edafccee | 8454 | |
2b358604 BM |
8455 | /* arbitrary limit, but we need something */ |
8456 | if (iov->iov_len > SZ_1G) | |
8457 | return -EFAULT; | |
edafccee | 8458 | |
50e96989 PB |
8459 | if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp)) |
8460 | return -EOVERFLOW; | |
8461 | ||
2b358604 BM |
8462 | return 0; |
8463 | } | |
edafccee | 8464 | |
2b358604 | 8465 | static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, |
634d00df | 8466 | unsigned int nr_args, u64 __user *tags) |
2b358604 | 8467 | { |
bd54b6fe BM |
8468 | struct page *last_hpage = NULL; |
8469 | struct io_rsrc_data *data; | |
2b358604 BM |
8470 | int i, ret; |
8471 | struct iovec iov; | |
edafccee | 8472 | |
87094465 PB |
8473 | if (ctx->user_bufs) |
8474 | return -EBUSY; | |
489809e2 | 8475 | if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) |
87094465 | 8476 | return -EINVAL; |
bd54b6fe | 8477 | ret = io_rsrc_node_switch_start(ctx); |
2b358604 BM |
8478 | if (ret) |
8479 | return ret; | |
d878c816 PB |
8480 | ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); |
8481 | if (ret) | |
8482 | return ret; | |
bd54b6fe BM |
8483 | ret = io_buffers_map_alloc(ctx, nr_args); |
8484 | if (ret) { | |
bb6659cc | 8485 | io_rsrc_data_free(data); |
bd54b6fe BM |
8486 | return ret; |
8487 | } | |
edafccee | 8488 | |
87094465 | 8489 | for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { |
edafccee JA |
8490 | ret = io_copy_iov(ctx, &iov, arg, i); |
8491 | if (ret) | |
0a96bbe4 | 8492 | break; |
2b358604 BM |
8493 | ret = io_buffer_validate(&iov); |
8494 | if (ret) | |
0a96bbe4 | 8495 | break; |
2d091d62 | 8496 | if (!iov.iov_base && *io_get_tag_slot(data, i)) { |
cf3770e7 CIK |
8497 | ret = -EINVAL; |
8498 | break; | |
8499 | } | |
edafccee | 8500 | |
41edf1a5 PB |
8501 | ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], |
8502 | &last_hpage); | |
0a96bbe4 BM |
8503 | if (ret) |
8504 | break; | |
edafccee | 8505 | } |
0a96bbe4 | 8506 | |
bd54b6fe | 8507 | WARN_ON_ONCE(ctx->buf_data); |
0a96bbe4 | 8508 | |
bd54b6fe BM |
8509 | ctx->buf_data = data; |
8510 | if (ret) | |
8511 | __io_sqe_buffers_unregister(ctx); | |
8512 | else | |
8513 | io_rsrc_node_switch(ctx, NULL); | |
edafccee JA |
8514 | return ret; |
8515 | } | |
8516 | ||
634d00df PB |
8517 | static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, |
8518 | struct io_uring_rsrc_update2 *up, | |
8519 | unsigned int nr_args) | |
8520 | { | |
8521 | u64 __user *tags = u64_to_user_ptr(up->tags); | |
8522 | struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); | |
634d00df PB |
8523 | struct page *last_hpage = NULL; |
8524 | bool needs_switch = false; | |
8525 | __u32 done; | |
8526 | int i, err; | |
8527 | ||
8528 | if (!ctx->buf_data) | |
8529 | return -ENXIO; | |
8530 | if (up->offset + nr_args > ctx->nr_user_bufs) | |
8531 | return -EINVAL; | |
8532 | ||
8533 | for (done = 0; done < nr_args; done++) { | |
0b8c0e7c PB |
8534 | struct io_mapped_ubuf *imu; |
8535 | int offset = up->offset + done; | |
634d00df PB |
8536 | u64 tag = 0; |
8537 | ||
8538 | err = io_copy_iov(ctx, &iov, iovs, done); | |
8539 | if (err) | |
8540 | break; | |
8541 | if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { | |
8542 | err = -EFAULT; | |
8543 | break; | |
8544 | } | |
0b8c0e7c PB |
8545 | err = io_buffer_validate(&iov); |
8546 | if (err) | |
8547 | break; | |
cf3770e7 CIK |
8548 | if (!iov.iov_base && tag) { |
8549 | err = -EINVAL; | |
8550 | break; | |
8551 | } | |
0b8c0e7c PB |
8552 | err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); |
8553 | if (err) | |
8554 | break; | |
634d00df | 8555 | |
0b8c0e7c | 8556 | i = array_index_nospec(offset, ctx->nr_user_bufs); |
6224843d | 8557 | if (ctx->user_bufs[i] != ctx->dummy_ubuf) { |
0b8c0e7c PB |
8558 | err = io_queue_rsrc_removal(ctx->buf_data, offset, |
8559 | ctx->rsrc_node, ctx->user_bufs[i]); | |
8560 | if (unlikely(err)) { | |
8561 | io_buffer_unmap(ctx, &imu); | |
634d00df | 8562 | break; |
0b8c0e7c | 8563 | } |
634d00df PB |
8564 | ctx->user_bufs[i] = NULL; |
8565 | needs_switch = true; | |
8566 | } | |
8567 | ||
0b8c0e7c | 8568 | ctx->user_bufs[i] = imu; |
2d091d62 | 8569 | *io_get_tag_slot(ctx->buf_data, offset) = tag; |
634d00df PB |
8570 | } |
8571 | ||
8572 | if (needs_switch) | |
8573 | io_rsrc_node_switch(ctx, ctx->buf_data); | |
8574 | return done ? done : err; | |
8575 | } | |
8576 | ||
9b402849 JA |
8577 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
8578 | { | |
8579 | __s32 __user *fds = arg; | |
8580 | int fd; | |
8581 | ||
8582 | if (ctx->cq_ev_fd) | |
8583 | return -EBUSY; | |
8584 | ||
8585 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
8586 | return -EFAULT; | |
8587 | ||
8588 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
8589 | if (IS_ERR(ctx->cq_ev_fd)) { | |
8590 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
fe7e3257 | 8591 | |
9b402849 JA |
8592 | ctx->cq_ev_fd = NULL; |
8593 | return ret; | |
8594 | } | |
8595 | ||
8596 | return 0; | |
8597 | } | |
8598 | ||
8599 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
8600 | { | |
8601 | if (ctx->cq_ev_fd) { | |
8602 | eventfd_ctx_put(ctx->cq_ev_fd); | |
8603 | ctx->cq_ev_fd = NULL; | |
8604 | return 0; | |
8605 | } | |
8606 | ||
8607 | return -ENXIO; | |
8608 | } | |
8609 | ||
5a2e745d JA |
8610 | static void io_destroy_buffers(struct io_ring_ctx *ctx) |
8611 | { | |
9e15c3a0 JA |
8612 | struct io_buffer *buf; |
8613 | unsigned long index; | |
8614 | ||
8615 | xa_for_each(&ctx->io_buffers, index, buf) | |
8616 | __io_remove_buffers(ctx, buf, index, -1U); | |
5a2e745d JA |
8617 | } |
8618 | ||
7255834e | 8619 | static void io_req_cache_free(struct list_head *list) |
1b4c351f | 8620 | { |
68e68ee6 | 8621 | struct io_kiocb *req, *nxt; |
1b4c351f | 8622 | |
bb943b82 PB |
8623 | list_for_each_entry_safe(req, nxt, list, inflight_entry) { |
8624 | list_del(&req->inflight_entry); | |
1b4c351f JA |
8625 | kmem_cache_free(req_cachep, req); |
8626 | } | |
8627 | } | |
8628 | ||
4010fec4 | 8629 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 8630 | { |
cd0ca2e0 | 8631 | struct io_submit_state *state = &ctx->submit_state; |
bf019da7 | 8632 | |
9a4fdbd8 JA |
8633 | mutex_lock(&ctx->uring_lock); |
8634 | ||
cd0ca2e0 PB |
8635 | if (state->free_reqs) { |
8636 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); | |
8637 | state->free_reqs = 0; | |
8e5c66c4 | 8638 | } |
9a4fdbd8 | 8639 | |
cd0ca2e0 PB |
8640 | io_flush_cached_locked_reqs(ctx, state); |
8641 | io_req_cache_free(&state->free_list); | |
9a4fdbd8 JA |
8642 | mutex_unlock(&ctx->uring_lock); |
8643 | } | |
8644 | ||
43597aac | 8645 | static void io_wait_rsrc_data(struct io_rsrc_data *data) |
2b188cc1 | 8646 | { |
43597aac | 8647 | if (data && !atomic_dec_and_test(&data->refs)) |
bd54b6fe | 8648 | wait_for_completion(&data->done); |
bd54b6fe | 8649 | } |
04fc6c80 | 8650 | |
2b188cc1 JA |
8651 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
8652 | { | |
37d1e2e3 | 8653 | io_sq_thread_finish(ctx); |
2aede0e4 | 8654 | |
37d1e2e3 | 8655 | if (ctx->mm_account) { |
2aede0e4 JA |
8656 | mmdrop(ctx->mm_account); |
8657 | ctx->mm_account = NULL; | |
30975825 | 8658 | } |
def596e9 | 8659 | |
43597aac PB |
8660 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
8661 | io_wait_rsrc_data(ctx->buf_data); | |
8662 | io_wait_rsrc_data(ctx->file_data); | |
8663 | ||
8bad28d8 | 8664 | mutex_lock(&ctx->uring_lock); |
43597aac | 8665 | if (ctx->buf_data) |
bd54b6fe | 8666 | __io_sqe_buffers_unregister(ctx); |
43597aac | 8667 | if (ctx->file_data) |
08480400 | 8668 | __io_sqe_files_unregister(ctx); |
c4ea060e PB |
8669 | if (ctx->rings) |
8670 | __io_cqring_overflow_flush(ctx, true); | |
8bad28d8 | 8671 | mutex_unlock(&ctx->uring_lock); |
9b402849 | 8672 | io_eventfd_unregister(ctx); |
5a2e745d | 8673 | io_destroy_buffers(ctx); |
07db298a PB |
8674 | if (ctx->sq_creds) |
8675 | put_cred(ctx->sq_creds); | |
def596e9 | 8676 | |
a7f0ed5a PB |
8677 | /* there are no registered resources left, nobody uses it */ |
8678 | if (ctx->rsrc_node) | |
8679 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 8680 | if (ctx->rsrc_backup_node) |
b895c9a6 | 8681 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a PB |
8682 | flush_delayed_work(&ctx->rsrc_put_work); |
8683 | ||
8684 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
8685 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 8686 | |
2b188cc1 | 8687 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
8688 | if (ctx->ring_sock) { |
8689 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 8690 | sock_release(ctx->ring_sock); |
355e8d26 | 8691 | } |
2b188cc1 JA |
8692 | #endif |
8693 | ||
75b28aff | 8694 | io_mem_free(ctx->rings); |
2b188cc1 | 8695 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
8696 | |
8697 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 8698 | free_uid(ctx->user); |
4010fec4 | 8699 | io_req_caches_free(ctx); |
e941894e JA |
8700 | if (ctx->hash_map) |
8701 | io_wq_put_hash(ctx->hash_map); | |
78076bb6 | 8702 | kfree(ctx->cancel_hash); |
6224843d | 8703 | kfree(ctx->dummy_ubuf); |
2b188cc1 JA |
8704 | kfree(ctx); |
8705 | } | |
8706 | ||
8707 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
8708 | { | |
8709 | struct io_ring_ctx *ctx = file->private_data; | |
8710 | __poll_t mask = 0; | |
8711 | ||
311997b3 | 8712 | poll_wait(file, &ctx->poll_wait, wait); |
4f7067c3 SB |
8713 | /* |
8714 | * synchronizes with barrier from wq_has_sleeper call in | |
8715 | * io_commit_cqring | |
8716 | */ | |
2b188cc1 | 8717 | smp_rmb(); |
90554200 | 8718 | if (!io_sqring_full(ctx)) |
2b188cc1 | 8719 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
8720 | |
8721 | /* | |
8722 | * Don't flush cqring overflow list here, just do a simple check. | |
8723 | * Otherwise there could possible be ABBA deadlock: | |
8724 | * CPU0 CPU1 | |
8725 | * ---- ---- | |
8726 | * lock(&ctx->uring_lock); | |
8727 | * lock(&ep->mtx); | |
8728 | * lock(&ctx->uring_lock); | |
8729 | * lock(&ep->mtx); | |
8730 | * | |
8731 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
8732 | * pushs them to do the flush. | |
8733 | */ | |
5ed7a37d | 8734 | if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow)) |
2b188cc1 JA |
8735 | mask |= EPOLLIN | EPOLLRDNORM; |
8736 | ||
8737 | return mask; | |
8738 | } | |
8739 | ||
8740 | static int io_uring_fasync(int fd, struct file *file, int on) | |
8741 | { | |
8742 | struct io_ring_ctx *ctx = file->private_data; | |
8743 | ||
8744 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
8745 | } | |
8746 | ||
0bead8cd | 8747 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 8748 | { |
4379bf8b | 8749 | const struct cred *creds; |
071698e1 | 8750 | |
61cf9370 | 8751 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
8752 | if (creds) { |
8753 | put_cred(creds); | |
0bead8cd | 8754 | return 0; |
1e6fa521 | 8755 | } |
0bead8cd YD |
8756 | |
8757 | return -EINVAL; | |
8758 | } | |
8759 | ||
d56d938b PB |
8760 | struct io_tctx_exit { |
8761 | struct callback_head task_work; | |
8762 | struct completion completion; | |
baf186c4 | 8763 | struct io_ring_ctx *ctx; |
d56d938b PB |
8764 | }; |
8765 | ||
8766 | static void io_tctx_exit_cb(struct callback_head *cb) | |
8767 | { | |
8768 | struct io_uring_task *tctx = current->io_uring; | |
8769 | struct io_tctx_exit *work; | |
8770 | ||
8771 | work = container_of(cb, struct io_tctx_exit, task_work); | |
8772 | /* | |
8773 | * When @in_idle, we're in cancellation and it's racy to remove the | |
8774 | * node. It'll be removed by the end of cancellation, just ignore it. | |
8775 | */ | |
8776 | if (!atomic_read(&tctx->in_idle)) | |
eef51daa | 8777 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
8778 | complete(&work->completion); |
8779 | } | |
8780 | ||
28090c13 PB |
8781 | static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
8782 | { | |
8783 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
8784 | ||
8785 | return req->ctx == data; | |
8786 | } | |
8787 | ||
85faa7b8 JA |
8788 | static void io_ring_exit_work(struct work_struct *work) |
8789 | { | |
d56d938b | 8790 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 8791 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 8792 | unsigned long interval = HZ / 20; |
d56d938b PB |
8793 | struct io_tctx_exit exit; |
8794 | struct io_tctx_node *node; | |
8795 | int ret; | |
85faa7b8 | 8796 | |
56952e91 JA |
8797 | /* |
8798 | * If we're doing polled IO and end up having requests being | |
8799 | * submitted async (out-of-line), then completions can come in while | |
8800 | * we're waiting for refs to drop. We need to reap these manually, | |
8801 | * as nobody else will be looking for them. | |
8802 | */ | |
b2edc0a7 | 8803 | do { |
3dd0c97a | 8804 | io_uring_try_cancel_requests(ctx, NULL, true); |
28090c13 PB |
8805 | if (ctx->sq_data) { |
8806 | struct io_sq_data *sqd = ctx->sq_data; | |
8807 | struct task_struct *tsk; | |
8808 | ||
8809 | io_sq_thread_park(sqd); | |
8810 | tsk = sqd->thread; | |
8811 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
8812 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
8813 | io_cancel_ctx_cb, ctx, true); | |
8814 | io_sq_thread_unpark(sqd); | |
8815 | } | |
b5bb3a24 | 8816 | |
58d3be2c PB |
8817 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
8818 | /* there is little hope left, don't run it too often */ | |
8819 | interval = HZ * 60; | |
8820 | } | |
8821 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 8822 | |
7f00651a PB |
8823 | init_completion(&exit.completion); |
8824 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
8825 | exit.ctx = ctx; | |
89b5066e PB |
8826 | /* |
8827 | * Some may use context even when all refs and requests have been put, | |
8828 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 8829 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
8830 | * this lock/unlock section also waits them to finish. |
8831 | */ | |
d56d938b PB |
8832 | mutex_lock(&ctx->uring_lock); |
8833 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
8834 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
8835 | ||
d56d938b PB |
8836 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
8837 | ctx_node); | |
7f00651a PB |
8838 | /* don't spin on a single task if cancellation failed */ |
8839 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
8840 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
8841 | if (WARN_ON_ONCE(ret)) | |
8842 | continue; | |
8843 | wake_up_process(node->task); | |
8844 | ||
8845 | mutex_unlock(&ctx->uring_lock); | |
8846 | wait_for_completion(&exit.completion); | |
d56d938b PB |
8847 | mutex_lock(&ctx->uring_lock); |
8848 | } | |
8849 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
8850 | spin_lock(&ctx->completion_lock); |
8851 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 8852 | |
85faa7b8 JA |
8853 | io_ring_ctx_free(ctx); |
8854 | } | |
8855 | ||
80c4cbdb PB |
8856 | /* Returns true if we found and killed one or more timeouts */ |
8857 | static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
3dd0c97a | 8858 | bool cancel_all) |
80c4cbdb PB |
8859 | { |
8860 | struct io_kiocb *req, *tmp; | |
8861 | int canceled = 0; | |
8862 | ||
79ebeaee JA |
8863 | spin_lock(&ctx->completion_lock); |
8864 | spin_lock_irq(&ctx->timeout_lock); | |
80c4cbdb | 8865 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) { |
3dd0c97a | 8866 | if (io_match_task(req, tsk, cancel_all)) { |
80c4cbdb PB |
8867 | io_kill_timeout(req, -ECANCELED); |
8868 | canceled++; | |
8869 | } | |
8870 | } | |
79ebeaee | 8871 | spin_unlock_irq(&ctx->timeout_lock); |
51520426 PB |
8872 | if (canceled != 0) |
8873 | io_commit_cqring(ctx); | |
79ebeaee | 8874 | spin_unlock(&ctx->completion_lock); |
80c4cbdb PB |
8875 | if (canceled != 0) |
8876 | io_cqring_ev_posted(ctx); | |
8877 | return canceled != 0; | |
8878 | } | |
8879 | ||
2b188cc1 JA |
8880 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
8881 | { | |
61cf9370 MWO |
8882 | unsigned long index; |
8883 | struct creds *creds; | |
8884 | ||
2b188cc1 JA |
8885 | mutex_lock(&ctx->uring_lock); |
8886 | percpu_ref_kill(&ctx->refs); | |
634578f8 | 8887 | if (ctx->rings) |
6c2450ae | 8888 | __io_cqring_overflow_flush(ctx, true); |
61cf9370 MWO |
8889 | xa_for_each(&ctx->personalities, index, creds) |
8890 | io_unregister_personality(ctx, index); | |
2b188cc1 JA |
8891 | mutex_unlock(&ctx->uring_lock); |
8892 | ||
3dd0c97a PB |
8893 | io_kill_timeouts(ctx, NULL, true); |
8894 | io_poll_remove_all(ctx, NULL, true); | |
561fb04a | 8895 | |
15dff286 | 8896 | /* if we failed setting up the ctx, we might not have any rings */ |
b2edc0a7 | 8897 | io_iopoll_try_reap_events(ctx); |
309fc03a | 8898 | |
85faa7b8 | 8899 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
8900 | /* |
8901 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
8902 | * if we're exiting a ton of rings at the same time. It just adds | |
8903 | * noise and overhead, there's no discernable change in runtime | |
8904 | * over using system_wq. | |
8905 | */ | |
8906 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
8907 | } |
8908 | ||
8909 | static int io_uring_release(struct inode *inode, struct file *file) | |
8910 | { | |
8911 | struct io_ring_ctx *ctx = file->private_data; | |
8912 | ||
8913 | file->private_data = NULL; | |
8914 | io_ring_ctx_wait_and_kill(ctx); | |
8915 | return 0; | |
8916 | } | |
8917 | ||
f6edbabb PB |
8918 | struct io_task_cancel { |
8919 | struct task_struct *task; | |
3dd0c97a | 8920 | bool all; |
f6edbabb | 8921 | }; |
f254ac04 | 8922 | |
f6edbabb | 8923 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 8924 | { |
9a472ef7 | 8925 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 8926 | struct io_task_cancel *cancel = data; |
9a472ef7 PB |
8927 | bool ret; |
8928 | ||
3dd0c97a | 8929 | if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) { |
9a472ef7 PB |
8930 | struct io_ring_ctx *ctx = req->ctx; |
8931 | ||
8932 | /* protect against races with linked timeouts */ | |
79ebeaee | 8933 | spin_lock(&ctx->completion_lock); |
3dd0c97a | 8934 | ret = io_match_task(req, cancel->task, cancel->all); |
79ebeaee | 8935 | spin_unlock(&ctx->completion_lock); |
9a472ef7 | 8936 | } else { |
3dd0c97a | 8937 | ret = io_match_task(req, cancel->task, cancel->all); |
9a472ef7 PB |
8938 | } |
8939 | return ret; | |
b711d4ea JA |
8940 | } |
8941 | ||
e1915f76 | 8942 | static bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3dd0c97a | 8943 | struct task_struct *task, bool cancel_all) |
b7ddce3c | 8944 | { |
e1915f76 | 8945 | struct io_defer_entry *de; |
b7ddce3c PB |
8946 | LIST_HEAD(list); |
8947 | ||
79ebeaee | 8948 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 8949 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
3dd0c97a | 8950 | if (io_match_task(de->req, task, cancel_all)) { |
b7ddce3c PB |
8951 | list_cut_position(&list, &ctx->defer_list, &de->list); |
8952 | break; | |
8953 | } | |
8954 | } | |
79ebeaee | 8955 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
8956 | if (list_empty(&list)) |
8957 | return false; | |
b7ddce3c PB |
8958 | |
8959 | while (!list_empty(&list)) { | |
8960 | de = list_first_entry(&list, struct io_defer_entry, list); | |
8961 | list_del_init(&de->list); | |
f41db273 | 8962 | io_req_complete_failed(de->req, -ECANCELED); |
b7ddce3c PB |
8963 | kfree(de); |
8964 | } | |
e1915f76 | 8965 | return true; |
b7ddce3c PB |
8966 | } |
8967 | ||
1b00764f PB |
8968 | static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
8969 | { | |
8970 | struct io_tctx_node *node; | |
8971 | enum io_wq_cancel cret; | |
8972 | bool ret = false; | |
8973 | ||
8974 | mutex_lock(&ctx->uring_lock); | |
8975 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
8976 | struct io_uring_task *tctx = node->task->io_uring; | |
8977 | ||
8978 | /* | |
8979 | * io_wq will stay alive while we hold uring_lock, because it's | |
8980 | * killed after ctx nodes, which requires to take the lock. | |
8981 | */ | |
8982 | if (!tctx || !tctx->io_wq) | |
8983 | continue; | |
8984 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
8985 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
8986 | } | |
8987 | mutex_unlock(&ctx->uring_lock); | |
8988 | ||
8989 | return ret; | |
8990 | } | |
8991 | ||
9936c7c2 PB |
8992 | static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
8993 | struct task_struct *task, | |
3dd0c97a | 8994 | bool cancel_all) |
9936c7c2 | 8995 | { |
3dd0c97a | 8996 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 8997 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
9936c7c2 PB |
8998 | |
8999 | while (1) { | |
9000 | enum io_wq_cancel cret; | |
9001 | bool ret = false; | |
9002 | ||
1b00764f PB |
9003 | if (!task) { |
9004 | ret |= io_uring_try_cancel_iowq(ctx); | |
9005 | } else if (tctx && tctx->io_wq) { | |
9006 | /* | |
9007 | * Cancels requests of all rings, not only @ctx, but | |
9008 | * it's fine as the task is in exit/exec. | |
9009 | */ | |
5aa75ed5 | 9010 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, |
9936c7c2 PB |
9011 | &cancel, true); |
9012 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
9013 | } | |
9014 | ||
9015 | /* SQPOLL thread does its own polling */ | |
3dd0c97a | 9016 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || |
d052d1d6 | 9017 | (ctx->sq_data && ctx->sq_data->thread == current)) { |
9936c7c2 PB |
9018 | while (!list_empty_careful(&ctx->iopoll_list)) { |
9019 | io_iopoll_try_reap_events(ctx); | |
9020 | ret = true; | |
9021 | } | |
9022 | } | |
9023 | ||
3dd0c97a PB |
9024 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
9025 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
9026 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
e5dc480d PB |
9027 | if (task) |
9028 | ret |= io_run_task_work(); | |
9936c7c2 PB |
9029 | if (!ret) |
9030 | break; | |
9031 | cond_resched(); | |
9032 | } | |
9033 | } | |
9034 | ||
eef51daa | 9035 | static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx) |
0f212204 | 9036 | { |
236434c3 | 9037 | struct io_uring_task *tctx = current->io_uring; |
13bf43f5 | 9038 | struct io_tctx_node *node; |
a528b04e | 9039 | int ret; |
236434c3 MWO |
9040 | |
9041 | if (unlikely(!tctx)) { | |
5aa75ed5 | 9042 | ret = io_uring_alloc_task_context(current, ctx); |
0f212204 JA |
9043 | if (unlikely(ret)) |
9044 | return ret; | |
236434c3 | 9045 | tctx = current->io_uring; |
0f212204 | 9046 | } |
cf27f3b1 PB |
9047 | if (!xa_load(&tctx->xa, (unsigned long)ctx)) { |
9048 | node = kmalloc(sizeof(*node), GFP_KERNEL); | |
9049 | if (!node) | |
9050 | return -ENOMEM; | |
9051 | node->ctx = ctx; | |
9052 | node->task = current; | |
13bf43f5 | 9053 | |
cf27f3b1 PB |
9054 | ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx, |
9055 | node, GFP_KERNEL)); | |
9056 | if (ret) { | |
9057 | kfree(node); | |
9058 | return ret; | |
0f212204 | 9059 | } |
cf27f3b1 PB |
9060 | |
9061 | mutex_lock(&ctx->uring_lock); | |
9062 | list_add(&node->ctx_node, &ctx->tctx_list); | |
9063 | mutex_unlock(&ctx->uring_lock); | |
0f212204 | 9064 | } |
cf27f3b1 | 9065 | tctx->last = ctx; |
0f212204 JA |
9066 | return 0; |
9067 | } | |
9068 | ||
cf27f3b1 PB |
9069 | /* |
9070 | * Note that this task has used io_uring. We use it for cancelation purposes. | |
9071 | */ | |
eef51daa | 9072 | static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx) |
cf27f3b1 PB |
9073 | { |
9074 | struct io_uring_task *tctx = current->io_uring; | |
9075 | ||
9076 | if (likely(tctx && tctx->last == ctx)) | |
9077 | return 0; | |
eef51daa | 9078 | return __io_uring_add_tctx_node(ctx); |
cf27f3b1 PB |
9079 | } |
9080 | ||
0f212204 JA |
9081 | /* |
9082 | * Remove this io_uring_file -> task mapping. | |
9083 | */ | |
eef51daa | 9084 | static void io_uring_del_tctx_node(unsigned long index) |
0f212204 JA |
9085 | { |
9086 | struct io_uring_task *tctx = current->io_uring; | |
13bf43f5 | 9087 | struct io_tctx_node *node; |
2941267b | 9088 | |
eebd2e37 PB |
9089 | if (!tctx) |
9090 | return; | |
13bf43f5 PB |
9091 | node = xa_erase(&tctx->xa, index); |
9092 | if (!node) | |
2941267b | 9093 | return; |
0f212204 | 9094 | |
13bf43f5 PB |
9095 | WARN_ON_ONCE(current != node->task); |
9096 | WARN_ON_ONCE(list_empty(&node->ctx_node)); | |
9097 | ||
9098 | mutex_lock(&node->ctx->uring_lock); | |
9099 | list_del(&node->ctx_node); | |
9100 | mutex_unlock(&node->ctx->uring_lock); | |
9101 | ||
baf186c4 | 9102 | if (tctx->last == node->ctx) |
0f212204 | 9103 | tctx->last = NULL; |
13bf43f5 | 9104 | kfree(node); |
0f212204 JA |
9105 | } |
9106 | ||
8452d4a6 | 9107 | static void io_uring_clean_tctx(struct io_uring_task *tctx) |
de7f1d9e | 9108 | { |
ba5ef6dc | 9109 | struct io_wq *wq = tctx->io_wq; |
13bf43f5 | 9110 | struct io_tctx_node *node; |
de7f1d9e PB |
9111 | unsigned long index; |
9112 | ||
13bf43f5 | 9113 | xa_for_each(&tctx->xa, index, node) |
eef51daa | 9114 | io_uring_del_tctx_node(index); |
b16ef427 ME |
9115 | if (wq) { |
9116 | /* | |
9117 | * Must be after io_uring_del_task_file() (removes nodes under | |
9118 | * uring_lock) to avoid race with io_uring_try_cancel_iowq(). | |
9119 | */ | |
9120 | tctx->io_wq = NULL; | |
ba5ef6dc | 9121 | io_wq_put_and_exit(wq); |
b16ef427 | 9122 | } |
de7f1d9e PB |
9123 | } |
9124 | ||
3f48cf18 | 9125 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 9126 | { |
3f48cf18 PB |
9127 | if (tracked) |
9128 | return atomic_read(&tctx->inflight_tracked); | |
521d6a73 PB |
9129 | return percpu_counter_sum(&tctx->inflight); |
9130 | } | |
9131 | ||
09899b19 PB |
9132 | static void io_uring_drop_tctx_refs(struct task_struct *task) |
9133 | { | |
9134 | struct io_uring_task *tctx = task->io_uring; | |
9135 | unsigned int refs = tctx->cached_refs; | |
9136 | ||
e9dbe221 PB |
9137 | if (refs) { |
9138 | tctx->cached_refs = 0; | |
9139 | percpu_counter_sub(&tctx->inflight, refs); | |
9140 | put_task_struct_many(task, refs); | |
9141 | } | |
09899b19 PB |
9142 | } |
9143 | ||
78cc687b PB |
9144 | /* |
9145 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
9146 | * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation. | |
9147 | */ | |
9148 | static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) | |
0e9ddb39 | 9149 | { |
521d6a73 | 9150 | struct io_uring_task *tctx = current->io_uring; |
734551df | 9151 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
9152 | s64 inflight; |
9153 | DEFINE_WAIT(wait); | |
fdaf083c | 9154 | |
78cc687b PB |
9155 | WARN_ON_ONCE(sqd && sqd->thread != current); |
9156 | ||
6d042ffb PO |
9157 | if (!current->io_uring) |
9158 | return; | |
17a91051 PB |
9159 | if (tctx->io_wq) |
9160 | io_wq_exit_start(tctx->io_wq); | |
9161 | ||
0e9ddb39 PB |
9162 | atomic_inc(&tctx->in_idle); |
9163 | do { | |
e9dbe221 | 9164 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 9165 | /* read completions before cancelations */ |
78cc687b | 9166 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
9167 | if (!inflight) |
9168 | break; | |
fdaf083c | 9169 | |
78cc687b PB |
9170 | if (!sqd) { |
9171 | struct io_tctx_node *node; | |
9172 | unsigned long index; | |
0f212204 | 9173 | |
78cc687b PB |
9174 | xa_for_each(&tctx->xa, index, node) { |
9175 | /* sqpoll task will cancel all its requests */ | |
9176 | if (node->ctx->sq_data) | |
9177 | continue; | |
9178 | io_uring_try_cancel_requests(node->ctx, current, | |
9179 | cancel_all); | |
9180 | } | |
9181 | } else { | |
9182 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
9183 | io_uring_try_cancel_requests(ctx, current, | |
9184 | cancel_all); | |
9185 | } | |
17a91051 | 9186 | |
0f212204 | 9187 | prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE); |
e9dbe221 | 9188 | io_uring_drop_tctx_refs(current); |
0f212204 | 9189 | /* |
a1bb3cd5 PB |
9190 | * If we've seen completions, retry without waiting. This |
9191 | * avoids a race where a completion comes in before we did | |
9192 | * prepare_to_wait(). | |
0f212204 | 9193 | */ |
3dd0c97a | 9194 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 9195 | schedule(); |
f57555ed | 9196 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 9197 | } while (1); |
fdaf083c | 9198 | atomic_dec(&tctx->in_idle); |
de7f1d9e | 9199 | |
8452d4a6 | 9200 | io_uring_clean_tctx(tctx); |
3dd0c97a | 9201 | if (cancel_all) { |
3f48cf18 PB |
9202 | /* for exec all current's requests should be gone, kill tctx */ |
9203 | __io_uring_free(current); | |
9204 | } | |
44e728b8 PB |
9205 | } |
9206 | ||
f552a27a | 9207 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 9208 | { |
f552a27a | 9209 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
9210 | } |
9211 | ||
6c5c240e RP |
9212 | static void *io_uring_validate_mmap_request(struct file *file, |
9213 | loff_t pgoff, size_t sz) | |
2b188cc1 | 9214 | { |
2b188cc1 | 9215 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 9216 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
9217 | struct page *page; |
9218 | void *ptr; | |
9219 | ||
9220 | switch (offset) { | |
9221 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
9222 | case IORING_OFF_CQ_RING: |
9223 | ptr = ctx->rings; | |
2b188cc1 JA |
9224 | break; |
9225 | case IORING_OFF_SQES: | |
9226 | ptr = ctx->sq_sqes; | |
9227 | break; | |
2b188cc1 | 9228 | default: |
6c5c240e | 9229 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
9230 | } |
9231 | ||
9232 | page = virt_to_head_page(ptr); | |
a50b854e | 9233 | if (sz > page_size(page)) |
6c5c240e RP |
9234 | return ERR_PTR(-EINVAL); |
9235 | ||
9236 | return ptr; | |
9237 | } | |
9238 | ||
9239 | #ifdef CONFIG_MMU | |
9240 | ||
9241 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9242 | { | |
9243 | size_t sz = vma->vm_end - vma->vm_start; | |
9244 | unsigned long pfn; | |
9245 | void *ptr; | |
9246 | ||
9247 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
9248 | if (IS_ERR(ptr)) | |
9249 | return PTR_ERR(ptr); | |
2b188cc1 JA |
9250 | |
9251 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
9252 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
9253 | } | |
9254 | ||
6c5c240e RP |
9255 | #else /* !CONFIG_MMU */ |
9256 | ||
9257 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9258 | { | |
9259 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
9260 | } | |
9261 | ||
9262 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
9263 | { | |
9264 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
9265 | } | |
9266 | ||
9267 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
9268 | unsigned long addr, unsigned long len, | |
9269 | unsigned long pgoff, unsigned long flags) | |
9270 | { | |
9271 | void *ptr; | |
9272 | ||
9273 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
9274 | if (IS_ERR(ptr)) | |
9275 | return PTR_ERR(ptr); | |
9276 | ||
9277 | return (unsigned long) ptr; | |
9278 | } | |
9279 | ||
9280 | #endif /* !CONFIG_MMU */ | |
9281 | ||
d9d05217 | 9282 | static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx) |
90554200 JA |
9283 | { |
9284 | DEFINE_WAIT(wait); | |
9285 | ||
9286 | do { | |
9287 | if (!io_sqring_full(ctx)) | |
9288 | break; | |
90554200 JA |
9289 | prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE); |
9290 | ||
9291 | if (!io_sqring_full(ctx)) | |
9292 | break; | |
90554200 JA |
9293 | schedule(); |
9294 | } while (!signal_pending(current)); | |
9295 | ||
9296 | finish_wait(&ctx->sqo_sq_wait, &wait); | |
5199328a | 9297 | return 0; |
90554200 JA |
9298 | } |
9299 | ||
c73ebb68 HX |
9300 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
9301 | struct __kernel_timespec __user **ts, | |
9302 | const sigset_t __user **sig) | |
9303 | { | |
9304 | struct io_uring_getevents_arg arg; | |
9305 | ||
9306 | /* | |
9307 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
9308 | * is just a pointer to the sigset_t. | |
9309 | */ | |
9310 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
9311 | *sig = (const sigset_t __user *) argp; | |
9312 | *ts = NULL; | |
9313 | return 0; | |
9314 | } | |
9315 | ||
9316 | /* | |
9317 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
9318 | * timespec and sigset_t pointers if good. | |
9319 | */ | |
9320 | if (*argsz != sizeof(arg)) | |
9321 | return -EINVAL; | |
9322 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
9323 | return -EFAULT; | |
9324 | *sig = u64_to_user_ptr(arg.sigmask); | |
9325 | *argsz = arg.sigmask_sz; | |
9326 | *ts = u64_to_user_ptr(arg.ts); | |
9327 | return 0; | |
9328 | } | |
9329 | ||
2b188cc1 | 9330 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
9331 | u32, min_complete, u32, flags, const void __user *, argp, |
9332 | size_t, argsz) | |
2b188cc1 JA |
9333 | { |
9334 | struct io_ring_ctx *ctx; | |
2b188cc1 JA |
9335 | int submitted = 0; |
9336 | struct fd f; | |
33f993da | 9337 | long ret; |
2b188cc1 | 9338 | |
4c6e277c | 9339 | io_run_task_work(); |
b41e9852 | 9340 | |
33f993da PB |
9341 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
9342 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))) | |
2b188cc1 JA |
9343 | return -EINVAL; |
9344 | ||
9345 | f = fdget(fd); | |
33f993da | 9346 | if (unlikely(!f.file)) |
2b188cc1 JA |
9347 | return -EBADF; |
9348 | ||
9349 | ret = -EOPNOTSUPP; | |
33f993da | 9350 | if (unlikely(f.file->f_op != &io_uring_fops)) |
2b188cc1 JA |
9351 | goto out_fput; |
9352 | ||
9353 | ret = -ENXIO; | |
9354 | ctx = f.file->private_data; | |
33f993da | 9355 | if (unlikely(!percpu_ref_tryget(&ctx->refs))) |
2b188cc1 JA |
9356 | goto out_fput; |
9357 | ||
7e84e1c7 | 9358 | ret = -EBADFD; |
33f993da | 9359 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
9360 | goto out; |
9361 | ||
6c271ce2 JA |
9362 | /* |
9363 | * For SQ polling, the thread will do all submissions and completions. | |
9364 | * Just return the requested submit count, and wake the thread if | |
9365 | * we were asked to. | |
9366 | */ | |
b2a9eada | 9367 | ret = 0; |
6c271ce2 | 9368 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 9369 | io_cqring_overflow_flush(ctx); |
89448c47 | 9370 | |
21f96522 JA |
9371 | if (unlikely(ctx->sq_data->thread == NULL)) { |
9372 | ret = -EOWNERDEAD; | |
04147488 | 9373 | goto out; |
21f96522 | 9374 | } |
6c271ce2 | 9375 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 9376 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
9377 | if (flags & IORING_ENTER_SQ_WAIT) { |
9378 | ret = io_sqpoll_wait_sq(ctx); | |
9379 | if (ret) | |
9380 | goto out; | |
9381 | } | |
6c271ce2 | 9382 | submitted = to_submit; |
b2a9eada | 9383 | } else if (to_submit) { |
eef51daa | 9384 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
9385 | if (unlikely(ret)) |
9386 | goto out; | |
2b188cc1 | 9387 | mutex_lock(&ctx->uring_lock); |
0f212204 | 9388 | submitted = io_submit_sqes(ctx, to_submit); |
2b188cc1 | 9389 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
9390 | |
9391 | if (submitted != to_submit) | |
9392 | goto out; | |
2b188cc1 JA |
9393 | } |
9394 | if (flags & IORING_ENTER_GETEVENTS) { | |
c73ebb68 HX |
9395 | const sigset_t __user *sig; |
9396 | struct __kernel_timespec __user *ts; | |
9397 | ||
9398 | ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); | |
9399 | if (unlikely(ret)) | |
9400 | goto out; | |
9401 | ||
2b188cc1 JA |
9402 | min_complete = min(min_complete, ctx->cq_entries); |
9403 | ||
32b2244a XW |
9404 | /* |
9405 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
9406 | * space applications don't need to do io completion events | |
9407 | * polling again, they can rely on io_sq_thread to do polling | |
9408 | * work, which can reduce cpu usage and uring_lock contention. | |
9409 | */ | |
9410 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
9411 | !(ctx->flags & IORING_SETUP_SQPOLL)) { | |
7668b92a | 9412 | ret = io_iopoll_check(ctx, min_complete); |
def596e9 | 9413 | } else { |
c73ebb68 | 9414 | ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts); |
def596e9 | 9415 | } |
2b188cc1 JA |
9416 | } |
9417 | ||
7c504e65 | 9418 | out: |
6805b32e | 9419 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
9420 | out_fput: |
9421 | fdput(f); | |
9422 | return submitted ? submitted : ret; | |
9423 | } | |
9424 | ||
bebdb65e | 9425 | #ifdef CONFIG_PROC_FS |
61cf9370 MWO |
9426 | static int io_uring_show_cred(struct seq_file *m, unsigned int id, |
9427 | const struct cred *cred) | |
87ce955b | 9428 | { |
87ce955b JA |
9429 | struct user_namespace *uns = seq_user_ns(m); |
9430 | struct group_info *gi; | |
9431 | kernel_cap_t cap; | |
9432 | unsigned __capi; | |
9433 | int g; | |
9434 | ||
9435 | seq_printf(m, "%5d\n", id); | |
9436 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
9437 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
9438 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
9439 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
9440 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
9441 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
9442 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
9443 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
9444 | seq_puts(m, "\n\tGroups:\t"); | |
9445 | gi = cred->group_info; | |
9446 | for (g = 0; g < gi->ngroups; g++) { | |
9447 | seq_put_decimal_ull(m, g ? " " : "", | |
9448 | from_kgid_munged(uns, gi->gid[g])); | |
9449 | } | |
9450 | seq_puts(m, "\n\tCapEff:\t"); | |
9451 | cap = cred->cap_effective; | |
9452 | CAP_FOR_EACH_U32(__capi) | |
9453 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
9454 | seq_putc(m, '\n'); | |
9455 | return 0; | |
9456 | } | |
9457 | ||
9458 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
9459 | { | |
dbbe9c64 | 9460 | struct io_sq_data *sq = NULL; |
fad8e0de | 9461 | bool has_lock; |
87ce955b JA |
9462 | int i; |
9463 | ||
fad8e0de JA |
9464 | /* |
9465 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, | |
9466 | * since fdinfo case grabs it in the opposite direction of normal use | |
9467 | * cases. If we fail to get the lock, we just don't iterate any | |
9468 | * structures that could be going away outside the io_uring mutex. | |
9469 | */ | |
9470 | has_lock = mutex_trylock(&ctx->uring_lock); | |
9471 | ||
5f3f26f9 | 9472 | if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) { |
dbbe9c64 | 9473 | sq = ctx->sq_data; |
5f3f26f9 JA |
9474 | if (!sq->thread) |
9475 | sq = NULL; | |
9476 | } | |
dbbe9c64 JQ |
9477 | |
9478 | seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1); | |
9479 | seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1); | |
87ce955b | 9480 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
fad8e0de | 9481 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
7b29f92d | 9482 | struct file *f = io_file_from_index(ctx, i); |
87ce955b | 9483 | |
87ce955b JA |
9484 | if (f) |
9485 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
9486 | else | |
9487 | seq_printf(m, "%5u: <none>\n", i); | |
9488 | } | |
9489 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
fad8e0de | 9490 | for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) { |
41edf1a5 | 9491 | struct io_mapped_ubuf *buf = ctx->user_bufs[i]; |
4751f53d | 9492 | unsigned int len = buf->ubuf_end - buf->ubuf; |
87ce955b | 9493 | |
4751f53d | 9494 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len); |
87ce955b | 9495 | } |
61cf9370 MWO |
9496 | if (has_lock && !xa_empty(&ctx->personalities)) { |
9497 | unsigned long index; | |
9498 | const struct cred *cred; | |
9499 | ||
87ce955b | 9500 | seq_printf(m, "Personalities:\n"); |
61cf9370 MWO |
9501 | xa_for_each(&ctx->personalities, index, cred) |
9502 | io_uring_show_cred(m, index, cred); | |
87ce955b | 9503 | } |
d7718a9d | 9504 | seq_printf(m, "PollList:\n"); |
79ebeaee | 9505 | spin_lock(&ctx->completion_lock); |
d7718a9d JA |
9506 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
9507 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
9508 | struct io_kiocb *req; | |
9509 | ||
9510 | hlist_for_each_entry(req, list, hash_node) | |
9511 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
9512 | req->task->task_works != NULL); | |
9513 | } | |
79ebeaee | 9514 | spin_unlock(&ctx->completion_lock); |
fad8e0de JA |
9515 | if (has_lock) |
9516 | mutex_unlock(&ctx->uring_lock); | |
87ce955b JA |
9517 | } |
9518 | ||
9519 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
9520 | { | |
9521 | struct io_ring_ctx *ctx = f->private_data; | |
9522 | ||
9523 | if (percpu_ref_tryget(&ctx->refs)) { | |
9524 | __io_uring_show_fdinfo(ctx, m); | |
9525 | percpu_ref_put(&ctx->refs); | |
9526 | } | |
9527 | } | |
bebdb65e | 9528 | #endif |
87ce955b | 9529 | |
2b188cc1 JA |
9530 | static const struct file_operations io_uring_fops = { |
9531 | .release = io_uring_release, | |
9532 | .mmap = io_uring_mmap, | |
6c5c240e RP |
9533 | #ifndef CONFIG_MMU |
9534 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
9535 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
9536 | #endif | |
2b188cc1 JA |
9537 | .poll = io_uring_poll, |
9538 | .fasync = io_uring_fasync, | |
bebdb65e | 9539 | #ifdef CONFIG_PROC_FS |
87ce955b | 9540 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 9541 | #endif |
2b188cc1 JA |
9542 | }; |
9543 | ||
9544 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
9545 | struct io_uring_params *p) | |
9546 | { | |
75b28aff HV |
9547 | struct io_rings *rings; |
9548 | size_t size, sq_array_offset; | |
2b188cc1 | 9549 | |
bd740481 JA |
9550 | /* make sure these are sane, as we already accounted them */ |
9551 | ctx->sq_entries = p->sq_entries; | |
9552 | ctx->cq_entries = p->cq_entries; | |
9553 | ||
75b28aff HV |
9554 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
9555 | if (size == SIZE_MAX) | |
9556 | return -EOVERFLOW; | |
9557 | ||
9558 | rings = io_mem_alloc(size); | |
9559 | if (!rings) | |
2b188cc1 JA |
9560 | return -ENOMEM; |
9561 | ||
75b28aff HV |
9562 | ctx->rings = rings; |
9563 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
9564 | rings->sq_ring_mask = p->sq_entries - 1; | |
9565 | rings->cq_ring_mask = p->cq_entries - 1; | |
9566 | rings->sq_ring_entries = p->sq_entries; | |
9567 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 JA |
9568 | |
9569 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
9570 | if (size == SIZE_MAX) { |
9571 | io_mem_free(ctx->rings); | |
9572 | ctx->rings = NULL; | |
2b188cc1 | 9573 | return -EOVERFLOW; |
eb065d30 | 9574 | } |
2b188cc1 JA |
9575 | |
9576 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
9577 | if (!ctx->sq_sqes) { |
9578 | io_mem_free(ctx->rings); | |
9579 | ctx->rings = NULL; | |
2b188cc1 | 9580 | return -ENOMEM; |
eb065d30 | 9581 | } |
2b188cc1 | 9582 | |
2b188cc1 JA |
9583 | return 0; |
9584 | } | |
9585 | ||
9faadcc8 PB |
9586 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
9587 | { | |
9588 | int ret, fd; | |
9589 | ||
9590 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
9591 | if (fd < 0) | |
9592 | return fd; | |
9593 | ||
eef51daa | 9594 | ret = io_uring_add_tctx_node(ctx); |
9faadcc8 PB |
9595 | if (ret) { |
9596 | put_unused_fd(fd); | |
9597 | return ret; | |
9598 | } | |
9599 | fd_install(fd, file); | |
9600 | return fd; | |
9601 | } | |
9602 | ||
2b188cc1 JA |
9603 | /* |
9604 | * Allocate an anonymous fd, this is what constitutes the application | |
9605 | * visible backing of an io_uring instance. The application mmaps this | |
9606 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
9607 | * we have to tie this fd to a socket for file garbage collection purposes. | |
9608 | */ | |
9faadcc8 | 9609 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
9610 | { |
9611 | struct file *file; | |
9faadcc8 | 9612 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
9613 | int ret; |
9614 | ||
2b188cc1 JA |
9615 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
9616 | &ctx->ring_sock); | |
9617 | if (ret) | |
9faadcc8 | 9618 | return ERR_PTR(ret); |
2b188cc1 JA |
9619 | #endif |
9620 | ||
2b188cc1 JA |
9621 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, |
9622 | O_RDWR | O_CLOEXEC); | |
2b188cc1 | 9623 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
9624 | if (IS_ERR(file)) { |
9625 | sock_release(ctx->ring_sock); | |
9626 | ctx->ring_sock = NULL; | |
9627 | } else { | |
9628 | ctx->ring_sock->file = file; | |
0f212204 | 9629 | } |
2b188cc1 | 9630 | #endif |
9faadcc8 | 9631 | return file; |
2b188cc1 JA |
9632 | } |
9633 | ||
7f13657d XW |
9634 | static int io_uring_create(unsigned entries, struct io_uring_params *p, |
9635 | struct io_uring_params __user *params) | |
2b188cc1 | 9636 | { |
2b188cc1 | 9637 | struct io_ring_ctx *ctx; |
9faadcc8 | 9638 | struct file *file; |
2b188cc1 JA |
9639 | int ret; |
9640 | ||
8110c1a6 | 9641 | if (!entries) |
2b188cc1 | 9642 | return -EINVAL; |
8110c1a6 JA |
9643 | if (entries > IORING_MAX_ENTRIES) { |
9644 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9645 | return -EINVAL; | |
9646 | entries = IORING_MAX_ENTRIES; | |
9647 | } | |
2b188cc1 JA |
9648 | |
9649 | /* | |
9650 | * Use twice as many entries for the CQ ring. It's possible for the | |
9651 | * application to drive a higher depth than the size of the SQ ring, | |
9652 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
9653 | * some flexibility in overcommitting a bit. If the application has |
9654 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
9655 | * of CQ ring entries manually. | |
2b188cc1 JA |
9656 | */ |
9657 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
9658 | if (p->flags & IORING_SETUP_CQSIZE) { |
9659 | /* | |
9660 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
9661 | * to a power-of-two, if it isn't already. We do NOT impose | |
9662 | * any cq vs sq ring sizing. | |
9663 | */ | |
eb2667b3 | 9664 | if (!p->cq_entries) |
33a107f0 | 9665 | return -EINVAL; |
8110c1a6 JA |
9666 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
9667 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9668 | return -EINVAL; | |
9669 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
9670 | } | |
eb2667b3 JQ |
9671 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
9672 | if (p->cq_entries < p->sq_entries) | |
9673 | return -EINVAL; | |
33a107f0 JA |
9674 | } else { |
9675 | p->cq_entries = 2 * p->sq_entries; | |
9676 | } | |
2b188cc1 | 9677 | |
2b188cc1 | 9678 | ctx = io_ring_ctx_alloc(p); |
62e398be | 9679 | if (!ctx) |
2b188cc1 | 9680 | return -ENOMEM; |
2b188cc1 | 9681 | ctx->compat = in_compat_syscall(); |
62e398be JA |
9682 | if (!capable(CAP_IPC_LOCK)) |
9683 | ctx->user = get_uid(current_user()); | |
2aede0e4 JA |
9684 | |
9685 | /* | |
9686 | * This is just grabbed for accounting purposes. When a process exits, | |
9687 | * the mm is exited and dropped before the files, hence we need to hang | |
9688 | * on to this mm purely for the purposes of being able to unaccount | |
9689 | * memory (locked/pinned vm). It's not used for anything else. | |
9690 | */ | |
6b7898eb | 9691 | mmgrab(current->mm); |
2aede0e4 | 9692 | ctx->mm_account = current->mm; |
6b7898eb | 9693 | |
2b188cc1 JA |
9694 | ret = io_allocate_scq_urings(ctx, p); |
9695 | if (ret) | |
9696 | goto err; | |
9697 | ||
7e84e1c7 | 9698 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
9699 | if (ret) |
9700 | goto err; | |
eae071c9 | 9701 | /* always set a rsrc node */ |
47b228ce PB |
9702 | ret = io_rsrc_node_switch_start(ctx); |
9703 | if (ret) | |
9704 | goto err; | |
eae071c9 | 9705 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 9706 | |
2b188cc1 | 9707 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
9708 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
9709 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
9710 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
9711 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
9712 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
9713 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
9714 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
9715 | |
9716 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
9717 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
9718 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
9719 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
9720 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
9721 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
9722 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 9723 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 9724 | |
7f13657d XW |
9725 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
9726 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 9727 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 9728 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e PB |
9729 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
9730 | IORING_FEAT_RSRC_TAGS; | |
7f13657d XW |
9731 | |
9732 | if (copy_to_user(params, p, sizeof(*p))) { | |
9733 | ret = -EFAULT; | |
9734 | goto err; | |
9735 | } | |
d1719f70 | 9736 | |
9faadcc8 PB |
9737 | file = io_uring_get_file(ctx); |
9738 | if (IS_ERR(file)) { | |
9739 | ret = PTR_ERR(file); | |
9740 | goto err; | |
9741 | } | |
9742 | ||
044c1ab3 JA |
9743 | /* |
9744 | * Install ring fd as the very last thing, so we don't risk someone | |
9745 | * having closed it before we finish setup | |
9746 | */ | |
9faadcc8 PB |
9747 | ret = io_uring_install_fd(ctx, file); |
9748 | if (ret < 0) { | |
9749 | /* fput will clean it up */ | |
9750 | fput(file); | |
9751 | return ret; | |
9752 | } | |
044c1ab3 | 9753 | |
c826bd7a | 9754 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
9755 | return ret; |
9756 | err: | |
9757 | io_ring_ctx_wait_and_kill(ctx); | |
9758 | return ret; | |
9759 | } | |
9760 | ||
9761 | /* | |
9762 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
9763 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
9764 | * params structure passed in. | |
9765 | */ | |
9766 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
9767 | { | |
9768 | struct io_uring_params p; | |
2b188cc1 JA |
9769 | int i; |
9770 | ||
9771 | if (copy_from_user(&p, params, sizeof(p))) | |
9772 | return -EFAULT; | |
9773 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
9774 | if (p.resv[i]) | |
9775 | return -EINVAL; | |
9776 | } | |
9777 | ||
6c271ce2 | 9778 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 9779 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 SG |
9780 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
9781 | IORING_SETUP_R_DISABLED)) | |
2b188cc1 JA |
9782 | return -EINVAL; |
9783 | ||
7f13657d | 9784 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
9785 | } |
9786 | ||
9787 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
9788 | struct io_uring_params __user *, params) | |
9789 | { | |
9790 | return io_uring_setup(entries, params); | |
9791 | } | |
9792 | ||
66f4af93 JA |
9793 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
9794 | { | |
9795 | struct io_uring_probe *p; | |
9796 | size_t size; | |
9797 | int i, ret; | |
9798 | ||
9799 | size = struct_size(p, ops, nr_args); | |
9800 | if (size == SIZE_MAX) | |
9801 | return -EOVERFLOW; | |
9802 | p = kzalloc(size, GFP_KERNEL); | |
9803 | if (!p) | |
9804 | return -ENOMEM; | |
9805 | ||
9806 | ret = -EFAULT; | |
9807 | if (copy_from_user(p, arg, size)) | |
9808 | goto out; | |
9809 | ret = -EINVAL; | |
9810 | if (memchr_inv(p, 0, size)) | |
9811 | goto out; | |
9812 | ||
9813 | p->last_op = IORING_OP_LAST - 1; | |
9814 | if (nr_args > IORING_OP_LAST) | |
9815 | nr_args = IORING_OP_LAST; | |
9816 | ||
9817 | for (i = 0; i < nr_args; i++) { | |
9818 | p->ops[i].op = i; | |
9819 | if (!io_op_defs[i].not_supported) | |
9820 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
9821 | } | |
9822 | p->ops_len = i; | |
9823 | ||
9824 | ret = 0; | |
9825 | if (copy_to_user(arg, p, size)) | |
9826 | ret = -EFAULT; | |
9827 | out: | |
9828 | kfree(p); | |
9829 | return ret; | |
9830 | } | |
9831 | ||
071698e1 JA |
9832 | static int io_register_personality(struct io_ring_ctx *ctx) |
9833 | { | |
4379bf8b | 9834 | const struct cred *creds; |
61cf9370 | 9835 | u32 id; |
1e6fa521 | 9836 | int ret; |
071698e1 | 9837 | |
4379bf8b | 9838 | creds = get_current_cred(); |
1e6fa521 | 9839 | |
61cf9370 MWO |
9840 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
9841 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
9842 | if (ret < 0) { |
9843 | put_cred(creds); | |
9844 | return ret; | |
9845 | } | |
9846 | return id; | |
071698e1 JA |
9847 | } |
9848 | ||
21b55dbc SG |
9849 | static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg, |
9850 | unsigned int nr_args) | |
9851 | { | |
9852 | struct io_uring_restriction *res; | |
9853 | size_t size; | |
9854 | int i, ret; | |
9855 | ||
7e84e1c7 SG |
9856 | /* Restrictions allowed only if rings started disabled */ |
9857 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9858 | return -EBADFD; | |
9859 | ||
21b55dbc | 9860 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 9861 | if (ctx->restrictions.registered) |
21b55dbc SG |
9862 | return -EBUSY; |
9863 | ||
9864 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
9865 | return -EINVAL; | |
9866 | ||
9867 | size = array_size(nr_args, sizeof(*res)); | |
9868 | if (size == SIZE_MAX) | |
9869 | return -EOVERFLOW; | |
9870 | ||
9871 | res = memdup_user(arg, size); | |
9872 | if (IS_ERR(res)) | |
9873 | return PTR_ERR(res); | |
9874 | ||
9875 | ret = 0; | |
9876 | ||
9877 | for (i = 0; i < nr_args; i++) { | |
9878 | switch (res[i].opcode) { | |
9879 | case IORING_RESTRICTION_REGISTER_OP: | |
9880 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
9881 | ret = -EINVAL; | |
9882 | goto out; | |
9883 | } | |
9884 | ||
9885 | __set_bit(res[i].register_op, | |
9886 | ctx->restrictions.register_op); | |
9887 | break; | |
9888 | case IORING_RESTRICTION_SQE_OP: | |
9889 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
9890 | ret = -EINVAL; | |
9891 | goto out; | |
9892 | } | |
9893 | ||
9894 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
9895 | break; | |
9896 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
9897 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
9898 | break; | |
9899 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
9900 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
9901 | break; | |
9902 | default: | |
9903 | ret = -EINVAL; | |
9904 | goto out; | |
9905 | } | |
9906 | } | |
9907 | ||
9908 | out: | |
9909 | /* Reset all restrictions if an error happened */ | |
9910 | if (ret != 0) | |
9911 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
9912 | else | |
7e84e1c7 | 9913 | ctx->restrictions.registered = true; |
21b55dbc SG |
9914 | |
9915 | kfree(res); | |
9916 | return ret; | |
9917 | } | |
9918 | ||
7e84e1c7 SG |
9919 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
9920 | { | |
9921 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9922 | return -EBADFD; | |
9923 | ||
9924 | if (ctx->restrictions.registered) | |
9925 | ctx->restricted = 1; | |
9926 | ||
0298ef96 PB |
9927 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
9928 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
9929 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
9930 | return 0; |
9931 | } | |
9932 | ||
fdecb662 | 9933 | static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, |
c3bdad02 | 9934 | struct io_uring_rsrc_update2 *up, |
98f0b3b4 PB |
9935 | unsigned nr_args) |
9936 | { | |
9937 | __u32 tmp; | |
9938 | int err; | |
9939 | ||
c3bdad02 PB |
9940 | if (up->resv) |
9941 | return -EINVAL; | |
98f0b3b4 PB |
9942 | if (check_add_overflow(up->offset, nr_args, &tmp)) |
9943 | return -EOVERFLOW; | |
9944 | err = io_rsrc_node_switch_start(ctx); | |
9945 | if (err) | |
9946 | return err; | |
9947 | ||
fdecb662 PB |
9948 | switch (type) { |
9949 | case IORING_RSRC_FILE: | |
98f0b3b4 | 9950 | return __io_sqe_files_update(ctx, up, nr_args); |
634d00df PB |
9951 | case IORING_RSRC_BUFFER: |
9952 | return __io_sqe_buffers_update(ctx, up, nr_args); | |
98f0b3b4 PB |
9953 | } |
9954 | return -EINVAL; | |
9955 | } | |
9956 | ||
c3bdad02 PB |
9957 | static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, |
9958 | unsigned nr_args) | |
98f0b3b4 | 9959 | { |
c3bdad02 | 9960 | struct io_uring_rsrc_update2 up; |
98f0b3b4 PB |
9961 | |
9962 | if (!nr_args) | |
9963 | return -EINVAL; | |
c3bdad02 PB |
9964 | memset(&up, 0, sizeof(up)); |
9965 | if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) | |
9966 | return -EFAULT; | |
9967 | return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); | |
9968 | } | |
9969 | ||
9970 | static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, | |
992da01a | 9971 | unsigned size, unsigned type) |
c3bdad02 PB |
9972 | { |
9973 | struct io_uring_rsrc_update2 up; | |
9974 | ||
9975 | if (size != sizeof(up)) | |
9976 | return -EINVAL; | |
98f0b3b4 PB |
9977 | if (copy_from_user(&up, arg, sizeof(up))) |
9978 | return -EFAULT; | |
992da01a | 9979 | if (!up.nr || up.resv) |
98f0b3b4 | 9980 | return -EINVAL; |
992da01a | 9981 | return __io_register_rsrc_update(ctx, type, &up, up.nr); |
98f0b3b4 PB |
9982 | } |
9983 | ||
792e3582 | 9984 | static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, |
992da01a | 9985 | unsigned int size, unsigned int type) |
792e3582 PB |
9986 | { |
9987 | struct io_uring_rsrc_register rr; | |
9988 | ||
9989 | /* keep it extendible */ | |
9990 | if (size != sizeof(rr)) | |
9991 | return -EINVAL; | |
9992 | ||
9993 | memset(&rr, 0, sizeof(rr)); | |
9994 | if (copy_from_user(&rr, arg, size)) | |
9995 | return -EFAULT; | |
992da01a | 9996 | if (!rr.nr || rr.resv || rr.resv2) |
792e3582 PB |
9997 | return -EINVAL; |
9998 | ||
992da01a | 9999 | switch (type) { |
792e3582 PB |
10000 | case IORING_RSRC_FILE: |
10001 | return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), | |
10002 | rr.nr, u64_to_user_ptr(rr.tags)); | |
634d00df PB |
10003 | case IORING_RSRC_BUFFER: |
10004 | return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), | |
10005 | rr.nr, u64_to_user_ptr(rr.tags)); | |
792e3582 PB |
10006 | } |
10007 | return -EINVAL; | |
10008 | } | |
10009 | ||
fe76421d JA |
10010 | static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg, |
10011 | unsigned len) | |
10012 | { | |
10013 | struct io_uring_task *tctx = current->io_uring; | |
10014 | cpumask_var_t new_mask; | |
10015 | int ret; | |
10016 | ||
10017 | if (!tctx || !tctx->io_wq) | |
10018 | return -EINVAL; | |
10019 | ||
10020 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
10021 | return -ENOMEM; | |
10022 | ||
10023 | cpumask_clear(new_mask); | |
10024 | if (len > cpumask_size()) | |
10025 | len = cpumask_size(); | |
10026 | ||
10027 | if (copy_from_user(new_mask, arg, len)) { | |
10028 | free_cpumask_var(new_mask); | |
10029 | return -EFAULT; | |
10030 | } | |
10031 | ||
10032 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
10033 | free_cpumask_var(new_mask); | |
10034 | return ret; | |
10035 | } | |
10036 | ||
10037 | static int io_unregister_iowq_aff(struct io_ring_ctx *ctx) | |
10038 | { | |
10039 | struct io_uring_task *tctx = current->io_uring; | |
10040 | ||
10041 | if (!tctx || !tctx->io_wq) | |
10042 | return -EINVAL; | |
10043 | ||
10044 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
10045 | } | |
10046 | ||
071698e1 JA |
10047 | static bool io_register_op_must_quiesce(int op) |
10048 | { | |
10049 | switch (op) { | |
bd54b6fe BM |
10050 | case IORING_REGISTER_BUFFERS: |
10051 | case IORING_UNREGISTER_BUFFERS: | |
f4f7d21c | 10052 | case IORING_REGISTER_FILES: |
071698e1 JA |
10053 | case IORING_UNREGISTER_FILES: |
10054 | case IORING_REGISTER_FILES_UPDATE: | |
10055 | case IORING_REGISTER_PROBE: | |
10056 | case IORING_REGISTER_PERSONALITY: | |
10057 | case IORING_UNREGISTER_PERSONALITY: | |
992da01a PB |
10058 | case IORING_REGISTER_FILES2: |
10059 | case IORING_REGISTER_FILES_UPDATE2: | |
10060 | case IORING_REGISTER_BUFFERS2: | |
10061 | case IORING_REGISTER_BUFFERS_UPDATE: | |
fe76421d JA |
10062 | case IORING_REGISTER_IOWQ_AFF: |
10063 | case IORING_UNREGISTER_IOWQ_AFF: | |
071698e1 JA |
10064 | return false; |
10065 | default: | |
10066 | return true; | |
10067 | } | |
10068 | } | |
10069 | ||
e73c5c7c PB |
10070 | static int io_ctx_quiesce(struct io_ring_ctx *ctx) |
10071 | { | |
10072 | long ret; | |
10073 | ||
10074 | percpu_ref_kill(&ctx->refs); | |
10075 | ||
10076 | /* | |
10077 | * Drop uring mutex before waiting for references to exit. If another | |
10078 | * thread is currently inside io_uring_enter() it might need to grab the | |
10079 | * uring_lock to make progress. If we hold it here across the drain | |
10080 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
10081 | * no new references will come in after we've killed the percpu ref. | |
10082 | */ | |
10083 | mutex_unlock(&ctx->uring_lock); | |
10084 | do { | |
10085 | ret = wait_for_completion_interruptible(&ctx->ref_comp); | |
10086 | if (!ret) | |
10087 | break; | |
10088 | ret = io_run_task_work_sig(); | |
10089 | } while (ret >= 0); | |
10090 | mutex_lock(&ctx->uring_lock); | |
10091 | ||
10092 | if (ret) | |
10093 | io_refs_resurrect(&ctx->refs, &ctx->ref_comp); | |
10094 | return ret; | |
10095 | } | |
10096 | ||
edafccee JA |
10097 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
10098 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
10099 | __releases(ctx->uring_lock) |
10100 | __acquires(ctx->uring_lock) | |
edafccee JA |
10101 | { |
10102 | int ret; | |
10103 | ||
35fa71a0 JA |
10104 | /* |
10105 | * We're inside the ring mutex, if the ref is already dying, then | |
10106 | * someone else killed the ctx or is already going through | |
10107 | * io_uring_register(). | |
10108 | */ | |
10109 | if (percpu_ref_is_dying(&ctx->refs)) | |
10110 | return -ENXIO; | |
10111 | ||
75c4021a PB |
10112 | if (ctx->restricted) { |
10113 | if (opcode >= IORING_REGISTER_LAST) | |
10114 | return -EINVAL; | |
10115 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); | |
10116 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
10117 | return -EACCES; | |
10118 | } | |
10119 | ||
071698e1 | 10120 | if (io_register_op_must_quiesce(opcode)) { |
e73c5c7c PB |
10121 | ret = io_ctx_quiesce(ctx); |
10122 | if (ret) | |
f70865db | 10123 | return ret; |
05f3fb3c | 10124 | } |
edafccee JA |
10125 | |
10126 | switch (opcode) { | |
10127 | case IORING_REGISTER_BUFFERS: | |
634d00df | 10128 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
10129 | break; |
10130 | case IORING_UNREGISTER_BUFFERS: | |
10131 | ret = -EINVAL; | |
10132 | if (arg || nr_args) | |
10133 | break; | |
0a96bbe4 | 10134 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 10135 | break; |
6b06314c | 10136 | case IORING_REGISTER_FILES: |
792e3582 | 10137 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
10138 | break; |
10139 | case IORING_UNREGISTER_FILES: | |
10140 | ret = -EINVAL; | |
10141 | if (arg || nr_args) | |
10142 | break; | |
10143 | ret = io_sqe_files_unregister(ctx); | |
10144 | break; | |
c3a31e60 | 10145 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 10146 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 10147 | break; |
9b402849 | 10148 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 10149 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
10150 | ret = -EINVAL; |
10151 | if (nr_args != 1) | |
10152 | break; | |
10153 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
10154 | if (ret) |
10155 | break; | |
10156 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
10157 | ctx->eventfd_async = 1; | |
10158 | else | |
10159 | ctx->eventfd_async = 0; | |
9b402849 JA |
10160 | break; |
10161 | case IORING_UNREGISTER_EVENTFD: | |
10162 | ret = -EINVAL; | |
10163 | if (arg || nr_args) | |
10164 | break; | |
10165 | ret = io_eventfd_unregister(ctx); | |
10166 | break; | |
66f4af93 JA |
10167 | case IORING_REGISTER_PROBE: |
10168 | ret = -EINVAL; | |
10169 | if (!arg || nr_args > 256) | |
10170 | break; | |
10171 | ret = io_probe(ctx, arg, nr_args); | |
10172 | break; | |
071698e1 JA |
10173 | case IORING_REGISTER_PERSONALITY: |
10174 | ret = -EINVAL; | |
10175 | if (arg || nr_args) | |
10176 | break; | |
10177 | ret = io_register_personality(ctx); | |
10178 | break; | |
10179 | case IORING_UNREGISTER_PERSONALITY: | |
10180 | ret = -EINVAL; | |
10181 | if (arg) | |
10182 | break; | |
10183 | ret = io_unregister_personality(ctx, nr_args); | |
10184 | break; | |
7e84e1c7 SG |
10185 | case IORING_REGISTER_ENABLE_RINGS: |
10186 | ret = -EINVAL; | |
10187 | if (arg || nr_args) | |
10188 | break; | |
10189 | ret = io_register_enable_rings(ctx); | |
10190 | break; | |
21b55dbc SG |
10191 | case IORING_REGISTER_RESTRICTIONS: |
10192 | ret = io_register_restrictions(ctx, arg, nr_args); | |
10193 | break; | |
992da01a PB |
10194 | case IORING_REGISTER_FILES2: |
10195 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
10196 | break; | |
10197 | case IORING_REGISTER_FILES_UPDATE2: | |
10198 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
10199 | IORING_RSRC_FILE); | |
10200 | break; | |
10201 | case IORING_REGISTER_BUFFERS2: | |
10202 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 10203 | break; |
992da01a PB |
10204 | case IORING_REGISTER_BUFFERS_UPDATE: |
10205 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
10206 | IORING_RSRC_BUFFER); | |
c3bdad02 | 10207 | break; |
fe76421d JA |
10208 | case IORING_REGISTER_IOWQ_AFF: |
10209 | ret = -EINVAL; | |
10210 | if (!arg || !nr_args) | |
10211 | break; | |
10212 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
10213 | break; | |
10214 | case IORING_UNREGISTER_IOWQ_AFF: | |
10215 | ret = -EINVAL; | |
10216 | if (arg || nr_args) | |
10217 | break; | |
10218 | ret = io_unregister_iowq_aff(ctx); | |
10219 | break; | |
edafccee JA |
10220 | default: |
10221 | ret = -EINVAL; | |
10222 | break; | |
10223 | } | |
10224 | ||
071698e1 | 10225 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 10226 | /* bring the ctx back to life */ |
05f3fb3c | 10227 | percpu_ref_reinit(&ctx->refs); |
0f158b4c | 10228 | reinit_completion(&ctx->ref_comp); |
05f3fb3c | 10229 | } |
edafccee JA |
10230 | return ret; |
10231 | } | |
10232 | ||
10233 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
10234 | void __user *, arg, unsigned int, nr_args) | |
10235 | { | |
10236 | struct io_ring_ctx *ctx; | |
10237 | long ret = -EBADF; | |
10238 | struct fd f; | |
10239 | ||
10240 | f = fdget(fd); | |
10241 | if (!f.file) | |
10242 | return -EBADF; | |
10243 | ||
10244 | ret = -EOPNOTSUPP; | |
10245 | if (f.file->f_op != &io_uring_fops) | |
10246 | goto out_fput; | |
10247 | ||
10248 | ctx = f.file->private_data; | |
10249 | ||
b6c23dd5 PB |
10250 | io_run_task_work(); |
10251 | ||
edafccee JA |
10252 | mutex_lock(&ctx->uring_lock); |
10253 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
10254 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
10255 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
10256 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
10257 | out_fput: |
10258 | fdput(f); | |
10259 | return ret; | |
10260 | } | |
10261 | ||
2b188cc1 JA |
10262 | static int __init io_uring_init(void) |
10263 | { | |
d7f62e82 SM |
10264 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
10265 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
10266 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
10267 | } while (0) | |
10268 | ||
10269 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
10270 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
10271 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
10272 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
10273 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
10274 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
10275 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
10276 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
10277 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
10278 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 10279 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
10280 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
10281 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
10282 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
10283 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
10284 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
10285 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
10286 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
10287 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
10288 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
10289 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
10290 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
10291 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
10292 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
10293 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
10294 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 10295 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
10296 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
10297 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 10298 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 10299 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 10300 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 10301 | |
b0d658ec PB |
10302 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
10303 | sizeof(struct io_uring_rsrc_update)); | |
10304 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
10305 | sizeof(struct io_uring_rsrc_update2)); | |
10306 | /* should fit into one byte */ | |
10307 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
10308 | ||
d3656344 | 10309 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
84557871 | 10310 | BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int)); |
16340eab | 10311 | |
91f245d5 JA |
10312 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
10313 | SLAB_ACCOUNT); | |
2b188cc1 JA |
10314 | return 0; |
10315 | }; | |
10316 | __initcall(io_uring_init); |