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 | ||
fd08e530 PB |
1309 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
1310 | { | |
4d13d1a4 PB |
1311 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
1312 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
1313 | |
1314 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 1315 | io_req_set_refcount(req); |
4d13d1a4 PB |
1316 | __io_req_set_refcount(req->link, 2); |
1317 | return req->link; | |
fd08e530 PB |
1318 | } |
1319 | ||
1320 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
1321 | { | |
4d13d1a4 | 1322 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
1323 | return NULL; |
1324 | return __io_prep_linked_timeout(req); | |
1325 | } | |
1326 | ||
1e6fa521 JA |
1327 | static void io_prep_async_work(struct io_kiocb *req) |
1328 | { | |
1329 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
1330 | struct io_ring_ctx *ctx = req->ctx; |
1331 | ||
b8e64b53 PB |
1332 | if (!(req->flags & REQ_F_CREDS)) { |
1333 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 1334 | req->creds = get_current_cred(); |
b8e64b53 | 1335 | } |
003e8dcc | 1336 | |
e1d675df PB |
1337 | req->work.list.next = NULL; |
1338 | req->work.flags = 0; | |
feaadc4f PB |
1339 | if (req->flags & REQ_F_FORCE_ASYNC) |
1340 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
1341 | ||
1e6fa521 JA |
1342 | if (req->flags & REQ_F_ISREG) { |
1343 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1344 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 1345 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
1346 | if (def->unbound_nonreg_file) |
1347 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1348 | } | |
e1d675df PB |
1349 | |
1350 | switch (req->opcode) { | |
1351 | case IORING_OP_SPLICE: | |
1352 | case IORING_OP_TEE: | |
e1d675df PB |
1353 | if (!S_ISREG(file_inode(req->splice.file_in)->i_mode)) |
1354 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1355 | break; | |
1356 | } | |
561fb04a | 1357 | } |
cccf0ee8 | 1358 | |
cbdcb435 | 1359 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 1360 | { |
cbdcb435 | 1361 | struct io_kiocb *cur; |
54a91f3b | 1362 | |
44eff40a PB |
1363 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
1364 | struct io_ring_ctx *ctx = req->ctx; | |
1365 | ||
79ebeaee | 1366 | spin_lock(&ctx->completion_lock); |
44eff40a PB |
1367 | io_for_each_link(cur, req) |
1368 | io_prep_async_work(cur); | |
79ebeaee | 1369 | spin_unlock(&ctx->completion_lock); |
44eff40a PB |
1370 | } else { |
1371 | io_for_each_link(cur, req) | |
1372 | io_prep_async_work(cur); | |
1373 | } | |
561fb04a JA |
1374 | } |
1375 | ||
ebf93667 | 1376 | static void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1377 | { |
a197f664 | 1378 | struct io_ring_ctx *ctx = req->ctx; |
cbdcb435 | 1379 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 1380 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 1381 | |
3bfe6106 JA |
1382 | BUG_ON(!tctx); |
1383 | BUG_ON(!tctx->io_wq); | |
561fb04a | 1384 | |
cbdcb435 PB |
1385 | /* init ->work of the whole link before punting */ |
1386 | io_prep_async_link(req); | |
991468dc JA |
1387 | |
1388 | /* | |
1389 | * Not expected to happen, but if we do have a bug where this _can_ | |
1390 | * happen, catch it here and ensure the request is marked as | |
1391 | * canceled. That will make io-wq go through the usual work cancel | |
1392 | * procedure rather than attempt to run this request (or create a new | |
1393 | * worker for it). | |
1394 | */ | |
1395 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
1396 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1397 | ||
d07f1e8a PB |
1398 | trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req, |
1399 | &req->work, req->flags); | |
ebf93667 | 1400 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
1401 | if (link) |
1402 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
1403 | } |
1404 | ||
1ee4160c | 1405 | static void io_kill_timeout(struct io_kiocb *req, int status) |
8c855885 | 1406 | __must_hold(&req->ctx->completion_lock) |
89850fce | 1407 | __must_hold(&req->ctx->timeout_lock) |
5262f567 | 1408 | { |
e8c2bc1f | 1409 | struct io_timeout_data *io = req->async_data; |
5262f567 | 1410 | |
fd9c7bc5 | 1411 | if (hrtimer_try_to_cancel(&io->timer) != -1) { |
01cec8c1 PB |
1412 | atomic_set(&req->ctx->cq_timeouts, |
1413 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
135fcde8 | 1414 | list_del_init(&req->timeout.list); |
d4d19c19 | 1415 | io_cqring_fill_event(req->ctx, req->user_data, status, 0); |
91c2f697 | 1416 | io_put_req_deferred(req); |
5262f567 JA |
1417 | } |
1418 | } | |
1419 | ||
441b8a78 | 1420 | static void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 1421 | { |
441b8a78 | 1422 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
1423 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
1424 | struct io_defer_entry, list); | |
de0617e4 | 1425 | |
9cf7c104 | 1426 | if (req_need_defer(de->req, de->seq)) |
04518945 | 1427 | break; |
27dc8338 | 1428 | list_del_init(&de->list); |
907d1df3 | 1429 | io_req_task_queue(de->req); |
27dc8338 | 1430 | kfree(de); |
441b8a78 | 1431 | } |
04518945 PB |
1432 | } |
1433 | ||
360428f8 | 1434 | static void io_flush_timeouts(struct io_ring_ctx *ctx) |
89850fce | 1435 | __must_hold(&ctx->completion_lock) |
de0617e4 | 1436 | { |
441b8a78 | 1437 | u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
f010505b | 1438 | |
79ebeaee | 1439 | spin_lock_irq(&ctx->timeout_lock); |
f18ee4cf | 1440 | while (!list_empty(&ctx->timeout_list)) { |
f010505b | 1441 | u32 events_needed, events_got; |
360428f8 | 1442 | struct io_kiocb *req = list_first_entry(&ctx->timeout_list, |
135fcde8 | 1443 | struct io_kiocb, timeout.list); |
de0617e4 | 1444 | |
8eb7e2d0 | 1445 | if (io_is_timeout_noseq(req)) |
360428f8 | 1446 | break; |
f010505b MDG |
1447 | |
1448 | /* | |
1449 | * Since seq can easily wrap around over time, subtract | |
1450 | * the last seq at which timeouts were flushed before comparing. | |
1451 | * Assuming not more than 2^31-1 events have happened since, | |
1452 | * these subtractions won't have wrapped, so we can check if | |
1453 | * target is in [last_seq, current_seq] by comparing the two. | |
1454 | */ | |
1455 | events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush; | |
1456 | events_got = seq - ctx->cq_last_tm_flush; | |
1457 | if (events_got < events_needed) | |
360428f8 | 1458 | break; |
bfe68a22 | 1459 | |
135fcde8 | 1460 | list_del_init(&req->timeout.list); |
1ee4160c | 1461 | io_kill_timeout(req, 0); |
f18ee4cf | 1462 | } |
f010505b | 1463 | ctx->cq_last_tm_flush = seq; |
79ebeaee | 1464 | spin_unlock_irq(&ctx->timeout_lock); |
360428f8 | 1465 | } |
5262f567 | 1466 | |
2335f6f5 | 1467 | static void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
360428f8 | 1468 | { |
2335f6f5 PB |
1469 | if (ctx->off_timeout_used) |
1470 | io_flush_timeouts(ctx); | |
1471 | if (ctx->drain_active) | |
1472 | io_queue_deferred(ctx); | |
1473 | } | |
1474 | ||
1475 | static inline void io_commit_cqring(struct io_ring_ctx *ctx) | |
1476 | { | |
1477 | if (unlikely(ctx->off_timeout_used || ctx->drain_active)) | |
1478 | __io_commit_cqring_flush(ctx); | |
ec30e04b PB |
1479 | /* order cqe stores with ring update */ |
1480 | smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail); | |
de0617e4 JA |
1481 | } |
1482 | ||
90554200 JA |
1483 | static inline bool io_sqring_full(struct io_ring_ctx *ctx) |
1484 | { | |
1485 | struct io_rings *r = ctx->rings; | |
1486 | ||
a566c556 | 1487 | return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries; |
90554200 JA |
1488 | } |
1489 | ||
888aae2e PB |
1490 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
1491 | { | |
1492 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
1493 | } | |
1494 | ||
d068b506 | 1495 | static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) |
2b188cc1 | 1496 | { |
75b28aff | 1497 | struct io_rings *rings = ctx->rings; |
ea5ab3b5 | 1498 | unsigned tail, mask = ctx->cq_entries - 1; |
2b188cc1 | 1499 | |
115e12e5 SB |
1500 | /* |
1501 | * writes to the cq entry need to come after reading head; the | |
1502 | * control dependency is enough as we're using WRITE_ONCE to | |
1503 | * fill the cq entry | |
1504 | */ | |
a566c556 | 1505 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
2b188cc1 JA |
1506 | return NULL; |
1507 | ||
888aae2e | 1508 | tail = ctx->cached_cq_tail++; |
ea5ab3b5 | 1509 | return &rings->cqes[tail & mask]; |
2b188cc1 JA |
1510 | } |
1511 | ||
f2842ab5 JA |
1512 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1513 | { | |
44c769de | 1514 | if (likely(!ctx->cq_ev_fd)) |
f0b493e6 | 1515 | return false; |
7e55a19c SG |
1516 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
1517 | return false; | |
44c769de | 1518 | return !ctx->eventfd_async || io_wq_current_is_worker(); |
f2842ab5 JA |
1519 | } |
1520 | ||
b41e9852 | 1521 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 1522 | { |
5fd46178 JA |
1523 | /* |
1524 | * wake_up_all() may seem excessive, but io_wake_function() and | |
1525 | * io_should_wake() handle the termination of the loop and only | |
1526 | * wake as many waiters as we need to. | |
1527 | */ | |
1528 | if (wq_has_sleeper(&ctx->cq_wait)) | |
1529 | wake_up_all(&ctx->cq_wait); | |
534ca6d6 JA |
1530 | if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait)) |
1531 | wake_up(&ctx->sq_data->wait); | |
b41e9852 | 1532 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 | 1533 | eventfd_signal(ctx->cq_ev_fd, 1); |
311997b3 PB |
1534 | if (waitqueue_active(&ctx->poll_wait)) { |
1535 | wake_up_interruptible(&ctx->poll_wait); | |
4aa84f2f PB |
1536 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); |
1537 | } | |
1d7bb1d5 JA |
1538 | } |
1539 | ||
80c18e4a PB |
1540 | static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx) |
1541 | { | |
1542 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
5fd46178 JA |
1543 | if (wq_has_sleeper(&ctx->cq_wait)) |
1544 | wake_up_all(&ctx->cq_wait); | |
80c18e4a PB |
1545 | } |
1546 | if (io_should_trigger_evfd(ctx)) | |
1547 | eventfd_signal(ctx->cq_ev_fd, 1); | |
311997b3 PB |
1548 | if (waitqueue_active(&ctx->poll_wait)) { |
1549 | wake_up_interruptible(&ctx->poll_wait); | |
4aa84f2f PB |
1550 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); |
1551 | } | |
80c18e4a PB |
1552 | } |
1553 | ||
c4a2ed72 | 1554 | /* Returns true if there are no backlogged entries after the flush */ |
6c2450ae | 1555 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) |
1d7bb1d5 | 1556 | { |
b18032bb | 1557 | bool all_flushed, posted; |
1d7bb1d5 | 1558 | |
a566c556 | 1559 | if (!force && __io_cqring_events(ctx) == ctx->cq_entries) |
e23de15f | 1560 | return false; |
1d7bb1d5 | 1561 | |
b18032bb | 1562 | posted = false; |
79ebeaee | 1563 | spin_lock(&ctx->completion_lock); |
6c2450ae | 1564 | while (!list_empty(&ctx->cq_overflow_list)) { |
d068b506 | 1565 | struct io_uring_cqe *cqe = io_get_cqe(ctx); |
6c2450ae | 1566 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 1567 | |
1d7bb1d5 JA |
1568 | if (!cqe && !force) |
1569 | break; | |
6c2450ae PB |
1570 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
1571 | struct io_overflow_cqe, list); | |
1572 | if (cqe) | |
1573 | memcpy(cqe, &ocqe->cqe, sizeof(*cqe)); | |
1574 | else | |
8f6ed49a PB |
1575 | io_account_cq_overflow(ctx); |
1576 | ||
b18032bb | 1577 | posted = true; |
6c2450ae PB |
1578 | list_del(&ocqe->list); |
1579 | kfree(ocqe); | |
1d7bb1d5 JA |
1580 | } |
1581 | ||
09e88404 PB |
1582 | all_flushed = list_empty(&ctx->cq_overflow_list); |
1583 | if (all_flushed) { | |
5ed7a37d | 1584 | clear_bit(0, &ctx->check_cq_overflow); |
20c0b380 NA |
1585 | WRITE_ONCE(ctx->rings->sq_flags, |
1586 | ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW); | |
09e88404 | 1587 | } |
46930143 | 1588 | |
b18032bb JA |
1589 | if (posted) |
1590 | io_commit_cqring(ctx); | |
79ebeaee | 1591 | spin_unlock(&ctx->completion_lock); |
b18032bb JA |
1592 | if (posted) |
1593 | io_cqring_ev_posted(ctx); | |
09e88404 | 1594 | return all_flushed; |
1d7bb1d5 JA |
1595 | } |
1596 | ||
90f67366 | 1597 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 1598 | { |
ca0a2651 JA |
1599 | bool ret = true; |
1600 | ||
5ed7a37d | 1601 | if (test_bit(0, &ctx->check_cq_overflow)) { |
6c503150 PB |
1602 | /* iopoll syncs against uring_lock, not completion_lock */ |
1603 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
1604 | mutex_lock(&ctx->uring_lock); | |
90f67366 | 1605 | ret = __io_cqring_overflow_flush(ctx, false); |
6c503150 PB |
1606 | if (ctx->flags & IORING_SETUP_IOPOLL) |
1607 | mutex_unlock(&ctx->uring_lock); | |
1608 | } | |
ca0a2651 JA |
1609 | |
1610 | return ret; | |
6c503150 PB |
1611 | } |
1612 | ||
6a290a14 PB |
1613 | /* must to be called somewhat shortly after putting a request */ |
1614 | static inline void io_put_task(struct task_struct *task, int nr) | |
1615 | { | |
1616 | struct io_uring_task *tctx = task->io_uring; | |
1617 | ||
1618 | percpu_counter_sub(&tctx->inflight, nr); | |
1619 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1620 | wake_up(&tctx->wait); | |
1621 | put_task_struct_many(task, nr); | |
1622 | } | |
1623 | ||
d4d19c19 PB |
1624 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
1625 | long res, unsigned int cflags) | |
2b188cc1 | 1626 | { |
cce4b8b0 | 1627 | struct io_overflow_cqe *ocqe; |
2b188cc1 | 1628 | |
cce4b8b0 PB |
1629 | ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT); |
1630 | if (!ocqe) { | |
1631 | /* | |
1632 | * If we're in ring overflow flush mode, or in task cancel mode, | |
1633 | * or cannot allocate an overflow entry, then we need to drop it | |
1634 | * on the floor. | |
1635 | */ | |
8f6ed49a | 1636 | io_account_cq_overflow(ctx); |
cce4b8b0 | 1637 | return false; |
2b188cc1 | 1638 | } |
cce4b8b0 | 1639 | if (list_empty(&ctx->cq_overflow_list)) { |
5ed7a37d | 1640 | set_bit(0, &ctx->check_cq_overflow); |
20c0b380 NA |
1641 | WRITE_ONCE(ctx->rings->sq_flags, |
1642 | ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW); | |
1643 | ||
cce4b8b0 | 1644 | } |
d4d19c19 | 1645 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
1646 | ocqe->cqe.res = res; |
1647 | ocqe->cqe.flags = cflags; | |
1648 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); | |
1649 | return true; | |
2b188cc1 JA |
1650 | } |
1651 | ||
d4d19c19 PB |
1652 | static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1653 | long res, unsigned int cflags) | |
2b188cc1 JA |
1654 | { |
1655 | struct io_uring_cqe *cqe; | |
1656 | ||
d4d19c19 | 1657 | trace_io_uring_complete(ctx, user_data, res, cflags); |
51c3ff62 | 1658 | |
2b188cc1 JA |
1659 | /* |
1660 | * If we can't get a cq entry, userspace overflowed the | |
1661 | * submission (by quite a lot). Increment the overflow count in | |
1662 | * the ring. | |
1663 | */ | |
d068b506 | 1664 | cqe = io_get_cqe(ctx); |
1d7bb1d5 | 1665 | if (likely(cqe)) { |
d4d19c19 | 1666 | WRITE_ONCE(cqe->user_data, user_data); |
2b188cc1 | 1667 | WRITE_ONCE(cqe->res, res); |
bcda7baa | 1668 | WRITE_ONCE(cqe->flags, cflags); |
8d13326e | 1669 | return true; |
2b188cc1 | 1670 | } |
d4d19c19 | 1671 | return io_cqring_event_overflow(ctx, user_data, res, cflags); |
2b188cc1 JA |
1672 | } |
1673 | ||
8d13326e | 1674 | /* not as hot to bloat with inlining */ |
d4d19c19 PB |
1675 | static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data, |
1676 | long res, unsigned int cflags) | |
bcda7baa | 1677 | { |
d4d19c19 | 1678 | return __io_cqring_fill_event(ctx, user_data, res, cflags); |
bcda7baa JA |
1679 | } |
1680 | ||
7a612350 PB |
1681 | static void io_req_complete_post(struct io_kiocb *req, long res, |
1682 | unsigned int cflags) | |
2b188cc1 | 1683 | { |
78e19bbe | 1684 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1685 | |
79ebeaee | 1686 | spin_lock(&ctx->completion_lock); |
d4d19c19 | 1687 | __io_cqring_fill_event(ctx, req->user_data, res, cflags); |
c7dae4ba JA |
1688 | /* |
1689 | * If we're the last reference to this request, add to our locked | |
1690 | * free_list cache. | |
1691 | */ | |
de9b4cca | 1692 | if (req_ref_put_and_test(req)) { |
7a612350 | 1693 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
0756a869 | 1694 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
1695 | io_disarm_next(req); |
1696 | if (req->link) { | |
1697 | io_req_task_queue(req->link); | |
1698 | req->link = NULL; | |
1699 | } | |
1700 | } | |
c7dae4ba JA |
1701 | io_dismantle_req(req); |
1702 | io_put_task(req->task, 1); | |
bb943b82 | 1703 | list_add(&req->inflight_entry, &ctx->locked_free_list); |
d0acdee2 | 1704 | ctx->locked_free_nr++; |
180f829f PB |
1705 | } else { |
1706 | if (!percpu_ref_tryget(&ctx->refs)) | |
1707 | req = NULL; | |
1708 | } | |
7a612350 | 1709 | io_commit_cqring(ctx); |
79ebeaee | 1710 | spin_unlock(&ctx->completion_lock); |
7a612350 | 1711 | |
180f829f PB |
1712 | if (req) { |
1713 | io_cqring_ev_posted(ctx); | |
c7dae4ba | 1714 | percpu_ref_put(&ctx->refs); |
180f829f | 1715 | } |
229a7b63 JA |
1716 | } |
1717 | ||
4e3d9ff9 JA |
1718 | static inline bool io_req_needs_clean(struct io_kiocb *req) |
1719 | { | |
c854357b | 1720 | return req->flags & IO_REQ_CLEAN_FLAGS; |
4e3d9ff9 JA |
1721 | } |
1722 | ||
a38d68db | 1723 | static void io_req_complete_state(struct io_kiocb *req, long res, |
889fca73 | 1724 | unsigned int cflags) |
229a7b63 | 1725 | { |
4e3d9ff9 | 1726 | if (io_req_needs_clean(req)) |
68fb8979 | 1727 | io_clean_op(req); |
a38d68db PB |
1728 | req->result = res; |
1729 | req->compl.cflags = cflags; | |
e342c807 | 1730 | req->flags |= REQ_F_COMPLETE_INLINE; |
e1e16097 JA |
1731 | } |
1732 | ||
889fca73 PB |
1733 | static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags, |
1734 | long res, unsigned cflags) | |
bcda7baa | 1735 | { |
889fca73 PB |
1736 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) |
1737 | io_req_complete_state(req, res, cflags); | |
a38d68db | 1738 | else |
c7dae4ba | 1739 | io_req_complete_post(req, res, cflags); |
bcda7baa JA |
1740 | } |
1741 | ||
a38d68db | 1742 | static inline void io_req_complete(struct io_kiocb *req, long res) |
0ddf92e8 | 1743 | { |
889fca73 | 1744 | __io_req_complete(req, 0, res, 0); |
0ddf92e8 JA |
1745 | } |
1746 | ||
f41db273 PB |
1747 | static void io_req_complete_failed(struct io_kiocb *req, long res) |
1748 | { | |
93d2bcd2 | 1749 | req_set_fail(req); |
f41db273 PB |
1750 | io_req_complete_post(req, res, 0); |
1751 | } | |
1752 | ||
864ea921 PB |
1753 | /* |
1754 | * Don't initialise the fields below on every allocation, but do that in | |
1755 | * advance and keep them valid across allocations. | |
1756 | */ | |
1757 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1758 | { | |
1759 | req->ctx = ctx; | |
1760 | req->link = NULL; | |
1761 | req->async_data = NULL; | |
1762 | /* not necessary, but safer to zero */ | |
1763 | req->result = 0; | |
1764 | } | |
1765 | ||
dac7a098 | 1766 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1767 | struct io_submit_state *state) |
dac7a098 | 1768 | { |
79ebeaee | 1769 | spin_lock(&ctx->completion_lock); |
cd0ca2e0 | 1770 | list_splice_init(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1771 | ctx->locked_free_nr = 0; |
79ebeaee | 1772 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1773 | } |
1774 | ||
dd78f492 | 1775 | /* Returns true IFF there are requests in the cache */ |
c7dae4ba | 1776 | static bool io_flush_cached_reqs(struct io_ring_ctx *ctx) |
0ddf92e8 | 1777 | { |
c7dae4ba | 1778 | struct io_submit_state *state = &ctx->submit_state; |
dd78f492 | 1779 | int nr; |
0ddf92e8 | 1780 | |
c7dae4ba JA |
1781 | /* |
1782 | * If we have more than a batch's worth of requests in our IRQ side | |
1783 | * locked cache, grab the lock and move them over to our submission | |
1784 | * side cache. | |
1785 | */ | |
d0acdee2 | 1786 | if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH) |
cd0ca2e0 | 1787 | io_flush_cached_locked_reqs(ctx, state); |
0ddf92e8 | 1788 | |
dd78f492 | 1789 | nr = state->free_reqs; |
cd0ca2e0 PB |
1790 | while (!list_empty(&state->free_list)) { |
1791 | struct io_kiocb *req = list_first_entry(&state->free_list, | |
bb943b82 | 1792 | struct io_kiocb, inflight_entry); |
dd78f492 | 1793 | |
bb943b82 | 1794 | list_del(&req->inflight_entry); |
dd78f492 PB |
1795 | state->reqs[nr++] = req; |
1796 | if (nr == ARRAY_SIZE(state->reqs)) | |
e5d1bc0a | 1797 | break; |
1b4c351f JA |
1798 | } |
1799 | ||
dd78f492 PB |
1800 | state->free_reqs = nr; |
1801 | return nr != 0; | |
0ddf92e8 JA |
1802 | } |
1803 | ||
5d5901a3 PB |
1804 | /* |
1805 | * A request might get retired back into the request caches even before opcode | |
1806 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1807 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1808 | * and with extra caution to not get a request that is still worked on. | |
1809 | */ | |
e5d1bc0a | 1810 | static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx) |
5d5901a3 | 1811 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1812 | { |
e5d1bc0a | 1813 | struct io_submit_state *state = &ctx->submit_state; |
864ea921 PB |
1814 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
1815 | int ret, i; | |
e5d1bc0a | 1816 | |
fe7e3257 | 1817 | BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH); |
e5d1bc0a | 1818 | |
864ea921 PB |
1819 | if (likely(state->free_reqs || io_flush_cached_reqs(ctx))) |
1820 | goto got_req; | |
e5d1bc0a | 1821 | |
864ea921 PB |
1822 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH, |
1823 | state->reqs); | |
fd6fab2c | 1824 | |
864ea921 PB |
1825 | /* |
1826 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1827 | * retry single alloc to be on the safe side. | |
1828 | */ | |
1829 | if (unlikely(ret <= 0)) { | |
1830 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1831 | if (!state->reqs[0]) | |
1832 | return NULL; | |
1833 | ret = 1; | |
2b188cc1 | 1834 | } |
864ea921 PB |
1835 | |
1836 | for (i = 0; i < ret; i++) | |
1837 | io_preinit_req(state->reqs[i], ctx); | |
1838 | state->free_reqs = ret; | |
e5d1bc0a | 1839 | got_req: |
291b2821 PB |
1840 | state->free_reqs--; |
1841 | return state->reqs[state->free_reqs]; | |
2b188cc1 JA |
1842 | } |
1843 | ||
e1d767f0 | 1844 | static inline void io_put_file(struct file *file) |
8da11c19 | 1845 | { |
e1d767f0 | 1846 | if (file) |
8da11c19 PB |
1847 | fput(file); |
1848 | } | |
1849 | ||
4edf20f9 | 1850 | static void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1851 | { |
094bae49 | 1852 | unsigned int flags = req->flags; |
929a3af9 | 1853 | |
3a0a6902 PB |
1854 | if (io_req_needs_clean(req)) |
1855 | io_clean_op(req); | |
e1d767f0 PB |
1856 | if (!(flags & REQ_F_FIXED_FILE)) |
1857 | io_put_file(req->file); | |
269bbe5f BM |
1858 | if (req->fixed_rsrc_refs) |
1859 | percpu_ref_put(req->fixed_rsrc_refs); | |
99ebe4ef | 1860 | if (req->async_data) { |
094bae49 | 1861 | kfree(req->async_data); |
99ebe4ef PB |
1862 | req->async_data = NULL; |
1863 | } | |
e65ef56d JA |
1864 | } |
1865 | ||
216578e5 | 1866 | static void __io_free_req(struct io_kiocb *req) |
c6ca97b3 | 1867 | { |
51a4cc11 | 1868 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1869 | |
216578e5 | 1870 | io_dismantle_req(req); |
7c660731 | 1871 | io_put_task(req->task, 1); |
c6ca97b3 | 1872 | |
79ebeaee | 1873 | spin_lock(&ctx->completion_lock); |
bb943b82 | 1874 | list_add(&req->inflight_entry, &ctx->locked_free_list); |
c34b025f | 1875 | ctx->locked_free_nr++; |
79ebeaee | 1876 | spin_unlock(&ctx->completion_lock); |
c34b025f | 1877 | |
ecfc5177 | 1878 | percpu_ref_put(&ctx->refs); |
e65ef56d JA |
1879 | } |
1880 | ||
f2f87370 PB |
1881 | static inline void io_remove_next_linked(struct io_kiocb *req) |
1882 | { | |
1883 | struct io_kiocb *nxt = req->link; | |
1884 | ||
1885 | req->link = nxt->link; | |
1886 | nxt->link = NULL; | |
1887 | } | |
1888 | ||
33cc89a9 PB |
1889 | static bool io_kill_linked_timeout(struct io_kiocb *req) |
1890 | __must_hold(&req->ctx->completion_lock) | |
89b263f6 | 1891 | __must_hold(&req->ctx->timeout_lock) |
2665abfd | 1892 | { |
33cc89a9 | 1893 | struct io_kiocb *link = req->link; |
f2f87370 | 1894 | |
b97e736a | 1895 | if (link && link->opcode == IORING_OP_LINK_TIMEOUT) { |
c9abd7ad | 1896 | struct io_timeout_data *io = link->async_data; |
7c86ffee | 1897 | |
f2f87370 | 1898 | io_remove_next_linked(req); |
90cd7e42 | 1899 | link->timeout.head = NULL; |
fd9c7bc5 | 1900 | if (hrtimer_try_to_cancel(&io->timer) != -1) { |
d4d19c19 PB |
1901 | io_cqring_fill_event(link->ctx, link->user_data, |
1902 | -ECANCELED, 0); | |
91c2f697 | 1903 | io_put_req_deferred(link); |
d4729fbd | 1904 | return true; |
c9abd7ad PB |
1905 | } |
1906 | } | |
d4729fbd | 1907 | return false; |
7c86ffee PB |
1908 | } |
1909 | ||
d148ca4b | 1910 | static void io_fail_links(struct io_kiocb *req) |
33cc89a9 | 1911 | __must_hold(&req->ctx->completion_lock) |
9e645e11 | 1912 | { |
33cc89a9 | 1913 | struct io_kiocb *nxt, *link = req->link; |
9e645e11 | 1914 | |
f2f87370 | 1915 | req->link = NULL; |
f2f87370 PB |
1916 | while (link) { |
1917 | nxt = link->link; | |
1918 | link->link = NULL; | |
2665abfd | 1919 | |
f2f87370 | 1920 | trace_io_uring_fail_link(req, link); |
d4d19c19 | 1921 | io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0); |
91c2f697 | 1922 | io_put_req_deferred(link); |
f2f87370 | 1923 | link = nxt; |
9e645e11 | 1924 | } |
33cc89a9 | 1925 | } |
9e645e11 | 1926 | |
33cc89a9 PB |
1927 | static bool io_disarm_next(struct io_kiocb *req) |
1928 | __must_hold(&req->ctx->completion_lock) | |
1929 | { | |
1930 | bool posted = false; | |
1931 | ||
0756a869 PB |
1932 | if (req->flags & REQ_F_ARM_LTIMEOUT) { |
1933 | struct io_kiocb *link = req->link; | |
1934 | ||
1935 | if (link && link->opcode == IORING_OP_LINK_TIMEOUT) { | |
1936 | io_remove_next_linked(req); | |
1937 | io_cqring_fill_event(link->ctx, link->user_data, | |
1938 | -ECANCELED, 0); | |
1939 | io_put_req_deferred(link); | |
1940 | posted = true; | |
1941 | } | |
1942 | } else if (req->flags & REQ_F_LINK_TIMEOUT) { | |
89b263f6 JA |
1943 | struct io_ring_ctx *ctx = req->ctx; |
1944 | ||
1945 | spin_lock_irq(&ctx->timeout_lock); | |
33cc89a9 | 1946 | posted = io_kill_linked_timeout(req); |
89b263f6 JA |
1947 | spin_unlock_irq(&ctx->timeout_lock); |
1948 | } | |
93d2bcd2 | 1949 | if (unlikely((req->flags & REQ_F_FAIL) && |
e4335ed3 | 1950 | !(req->flags & REQ_F_HARDLINK))) { |
33cc89a9 PB |
1951 | posted |= (req->link != NULL); |
1952 | io_fail_links(req); | |
1953 | } | |
1954 | return posted; | |
9e645e11 JA |
1955 | } |
1956 | ||
3fa5e0f3 | 1957 | static struct io_kiocb *__io_req_find_next(struct io_kiocb *req) |
c69f8dbe | 1958 | { |
33cc89a9 | 1959 | struct io_kiocb *nxt; |
944e58bf | 1960 | |
9e645e11 JA |
1961 | /* |
1962 | * If LINK is set, we have dependent requests in this chain. If we | |
1963 | * didn't fail this request, queue the first one up, moving any other | |
1964 | * dependencies to the next request. In case of failure, fail the rest | |
1965 | * of the chain. | |
1966 | */ | |
0756a869 | 1967 | if (req->flags & IO_DISARM_MASK) { |
33cc89a9 | 1968 | struct io_ring_ctx *ctx = req->ctx; |
33cc89a9 PB |
1969 | bool posted; |
1970 | ||
79ebeaee | 1971 | spin_lock(&ctx->completion_lock); |
33cc89a9 PB |
1972 | posted = io_disarm_next(req); |
1973 | if (posted) | |
1974 | io_commit_cqring(req->ctx); | |
79ebeaee | 1975 | spin_unlock(&ctx->completion_lock); |
33cc89a9 PB |
1976 | if (posted) |
1977 | io_cqring_ev_posted(ctx); | |
f2f87370 | 1978 | } |
33cc89a9 PB |
1979 | nxt = req->link; |
1980 | req->link = NULL; | |
1981 | return nxt; | |
4d7dd462 | 1982 | } |
9e645e11 | 1983 | |
f2f87370 | 1984 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) |
3fa5e0f3 | 1985 | { |
cdbff982 | 1986 | if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK)))) |
3fa5e0f3 PB |
1987 | return NULL; |
1988 | return __io_req_find_next(req); | |
1989 | } | |
1990 | ||
2c32395d PB |
1991 | static void ctx_flush_and_put(struct io_ring_ctx *ctx) |
1992 | { | |
1993 | if (!ctx) | |
1994 | return; | |
cd0ca2e0 | 1995 | if (ctx->submit_state.compl_nr) { |
2c32395d | 1996 | mutex_lock(&ctx->uring_lock); |
2a2758f2 | 1997 | io_submit_flush_completions(ctx); |
2c32395d PB |
1998 | mutex_unlock(&ctx->uring_lock); |
1999 | } | |
2000 | percpu_ref_put(&ctx->refs); | |
2001 | } | |
2002 | ||
7cbf1722 | 2003 | static void tctx_task_work(struct callback_head *cb) |
c40f6379 | 2004 | { |
ebd0df2e | 2005 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
2006 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
2007 | task_work); | |
c40f6379 | 2008 | |
16f72070 | 2009 | while (1) { |
3f18407d PB |
2010 | struct io_wq_work_node *node; |
2011 | ||
2012 | spin_lock_irq(&tctx->task_lock); | |
c6538be9 | 2013 | node = tctx->task_list.first; |
3f18407d | 2014 | INIT_WQ_LIST(&tctx->task_list); |
6294f368 PB |
2015 | if (!node) |
2016 | tctx->task_running = false; | |
3f18407d | 2017 | spin_unlock_irq(&tctx->task_lock); |
6294f368 PB |
2018 | if (!node) |
2019 | break; | |
3f18407d | 2020 | |
6294f368 | 2021 | do { |
3f18407d PB |
2022 | struct io_wq_work_node *next = node->next; |
2023 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
2024 | io_task_work.node); | |
2025 | ||
2026 | if (req->ctx != ctx) { | |
2027 | ctx_flush_and_put(ctx); | |
2028 | ctx = req->ctx; | |
2029 | percpu_ref_get(&ctx->refs); | |
2030 | } | |
5b0a6acc | 2031 | req->io_task_work.func(req); |
3f18407d | 2032 | node = next; |
6294f368 PB |
2033 | } while (node); |
2034 | ||
7cbf1722 | 2035 | cond_resched(); |
3f18407d | 2036 | } |
ebd0df2e PB |
2037 | |
2038 | ctx_flush_and_put(ctx); | |
7cbf1722 JA |
2039 | } |
2040 | ||
e09ee510 | 2041 | static void io_req_task_work_add(struct io_kiocb *req) |
7cbf1722 | 2042 | { |
c15b79de | 2043 | struct task_struct *tsk = req->task; |
7cbf1722 | 2044 | struct io_uring_task *tctx = tsk->io_uring; |
c15b79de | 2045 | enum task_work_notify_mode notify; |
e09ee510 | 2046 | struct io_wq_work_node *node; |
0b81e80c | 2047 | unsigned long flags; |
6294f368 | 2048 | bool running; |
7cbf1722 JA |
2049 | |
2050 | WARN_ON_ONCE(!tctx); | |
2051 | ||
0b81e80c | 2052 | spin_lock_irqsave(&tctx->task_lock, flags); |
7cbf1722 | 2053 | wq_list_add_tail(&req->io_task_work.node, &tctx->task_list); |
6294f368 PB |
2054 | running = tctx->task_running; |
2055 | if (!running) | |
2056 | tctx->task_running = true; | |
0b81e80c | 2057 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 JA |
2058 | |
2059 | /* task_work already pending, we're done */ | |
6294f368 | 2060 | if (running) |
e09ee510 | 2061 | return; |
7cbf1722 | 2062 | |
c15b79de PB |
2063 | /* |
2064 | * SQPOLL kernel thread doesn't need notification, just a wakeup. For | |
2065 | * all other cases, use TWA_SIGNAL unconditionally to ensure we're | |
2066 | * processing task_work. There's no reliable way to tell if TWA_RESUME | |
2067 | * will do the job. | |
2068 | */ | |
2069 | notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL; | |
c15b79de PB |
2070 | if (!task_work_add(tsk, &tctx->task_work, notify)) { |
2071 | wake_up_process(tsk); | |
e09ee510 | 2072 | return; |
c15b79de | 2073 | } |
2215bed9 | 2074 | |
0b81e80c | 2075 | spin_lock_irqsave(&tctx->task_lock, flags); |
6294f368 | 2076 | tctx->task_running = false; |
e09ee510 PB |
2077 | node = tctx->task_list.first; |
2078 | INIT_WQ_LIST(&tctx->task_list); | |
0b81e80c | 2079 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 | 2080 | |
e09ee510 PB |
2081 | while (node) { |
2082 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
2083 | node = node->next; | |
2084 | if (llist_add(&req->io_task_work.fallback_node, | |
2085 | &req->ctx->fallback_llist)) | |
2086 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
2087 | } | |
eab30c4d PB |
2088 | } |
2089 | ||
5b0a6acc | 2090 | static void io_req_task_cancel(struct io_kiocb *req) |
c40f6379 | 2091 | { |
87ceb6a6 | 2092 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 | 2093 | |
e83acd7d | 2094 | /* ctx is guaranteed to stay alive while we hold uring_lock */ |
792bb6eb | 2095 | mutex_lock(&ctx->uring_lock); |
2593553a | 2096 | io_req_complete_failed(req, req->result); |
792bb6eb | 2097 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2098 | } |
2099 | ||
5b0a6acc | 2100 | static void io_req_task_submit(struct io_kiocb *req) |
c40f6379 JA |
2101 | { |
2102 | struct io_ring_ctx *ctx = req->ctx; | |
2103 | ||
04fc6c80 | 2104 | /* ctx stays valid until unlock, even if we drop all ours ctx->refs */ |
81b6d05c | 2105 | mutex_lock(&ctx->uring_lock); |
af066f31 | 2106 | if (likely(!(req->task->flags & PF_EXITING))) |
c5eef2b9 | 2107 | __io_queue_sqe(req); |
81b6d05c | 2108 | else |
2593553a | 2109 | io_req_complete_failed(req, -EFAULT); |
81b6d05c | 2110 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2111 | } |
2112 | ||
2c4b8eb6 | 2113 | static void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 2114 | { |
2c4b8eb6 | 2115 | req->result = ret; |
5b0a6acc | 2116 | req->io_task_work.func = io_req_task_cancel; |
e09ee510 | 2117 | io_req_task_work_add(req); |
c40f6379 JA |
2118 | } |
2119 | ||
2c4b8eb6 | 2120 | static void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 2121 | { |
5b0a6acc | 2122 | req->io_task_work.func = io_req_task_submit; |
e09ee510 | 2123 | io_req_task_work_add(req); |
a3df7698 PB |
2124 | } |
2125 | ||
773af691 JA |
2126 | static void io_req_task_queue_reissue(struct io_kiocb *req) |
2127 | { | |
2128 | req->io_task_work.func = io_queue_async_work; | |
2129 | io_req_task_work_add(req); | |
2130 | } | |
2131 | ||
f2f87370 | 2132 | static inline void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 2133 | { |
9b5f7bd9 | 2134 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
2135 | |
2136 | if (nxt) | |
906a8c3f | 2137 | io_req_task_queue(nxt); |
c69f8dbe JL |
2138 | } |
2139 | ||
c3524383 | 2140 | static void io_free_req(struct io_kiocb *req) |
7a743e22 | 2141 | { |
c3524383 PB |
2142 | io_queue_next(req); |
2143 | __io_free_req(req); | |
2144 | } | |
8766dd51 | 2145 | |
2d6500d4 | 2146 | struct req_batch { |
5af1d13e PB |
2147 | struct task_struct *task; |
2148 | int task_refs; | |
1b4c351f | 2149 | int ctx_refs; |
2d6500d4 PB |
2150 | }; |
2151 | ||
5af1d13e PB |
2152 | static inline void io_init_req_batch(struct req_batch *rb) |
2153 | { | |
5af1d13e | 2154 | rb->task_refs = 0; |
9ae72463 | 2155 | rb->ctx_refs = 0; |
5af1d13e PB |
2156 | rb->task = NULL; |
2157 | } | |
2158 | ||
2d6500d4 PB |
2159 | static void io_req_free_batch_finish(struct io_ring_ctx *ctx, |
2160 | struct req_batch *rb) | |
2161 | { | |
9ae72463 PB |
2162 | if (rb->ctx_refs) |
2163 | percpu_ref_put_many(&ctx->refs, rb->ctx_refs); | |
e9dbe221 PB |
2164 | if (rb->task == current) |
2165 | current->io_uring->cached_refs += rb->task_refs; | |
2166 | else if (rb->task) | |
2167 | io_put_task(rb->task, rb->task_refs); | |
2d6500d4 PB |
2168 | } |
2169 | ||
6ff119a6 PB |
2170 | static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req, |
2171 | struct io_submit_state *state) | |
2d6500d4 | 2172 | { |
f2f87370 | 2173 | io_queue_next(req); |
96670657 | 2174 | io_dismantle_req(req); |
2d6500d4 | 2175 | |
e3bc8e9d | 2176 | if (req->task != rb->task) { |
7c660731 PB |
2177 | if (rb->task) |
2178 | io_put_task(rb->task, rb->task_refs); | |
e3bc8e9d JA |
2179 | rb->task = req->task; |
2180 | rb->task_refs = 0; | |
5af1d13e | 2181 | } |
e3bc8e9d | 2182 | rb->task_refs++; |
9ae72463 | 2183 | rb->ctx_refs++; |
5af1d13e | 2184 | |
bd759045 | 2185 | if (state->free_reqs != ARRAY_SIZE(state->reqs)) |
6ff119a6 | 2186 | state->reqs[state->free_reqs++] = req; |
bd759045 | 2187 | else |
cd0ca2e0 | 2188 | list_add(&req->inflight_entry, &state->free_list); |
7a743e22 PB |
2189 | } |
2190 | ||
2a2758f2 | 2191 | static void io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 2192 | __must_hold(&ctx->uring_lock) |
905c172f | 2193 | { |
cd0ca2e0 PB |
2194 | struct io_submit_state *state = &ctx->submit_state; |
2195 | int i, nr = state->compl_nr; | |
905c172f PB |
2196 | struct req_batch rb; |
2197 | ||
79ebeaee | 2198 | spin_lock(&ctx->completion_lock); |
905c172f | 2199 | for (i = 0; i < nr; i++) { |
cd0ca2e0 | 2200 | struct io_kiocb *req = state->compl_reqs[i]; |
5182ed2e | 2201 | |
d4d19c19 PB |
2202 | __io_cqring_fill_event(ctx, req->user_data, req->result, |
2203 | req->compl.cflags); | |
905c172f PB |
2204 | } |
2205 | io_commit_cqring(ctx); | |
79ebeaee | 2206 | spin_unlock(&ctx->completion_lock); |
905c172f | 2207 | io_cqring_ev_posted(ctx); |
5182ed2e PB |
2208 | |
2209 | io_init_req_batch(&rb); | |
905c172f | 2210 | for (i = 0; i < nr; i++) { |
cd0ca2e0 | 2211 | struct io_kiocb *req = state->compl_reqs[i]; |
905c172f | 2212 | |
91c2f697 | 2213 | if (req_ref_put_and_test(req)) |
6ff119a6 | 2214 | io_req_free_batch(&rb, req, &ctx->submit_state); |
905c172f PB |
2215 | } |
2216 | ||
2217 | io_req_free_batch_finish(ctx, &rb); | |
cd0ca2e0 | 2218 | state->compl_nr = 0; |
7a743e22 PB |
2219 | } |
2220 | ||
ba816ad6 JA |
2221 | /* |
2222 | * Drop reference to request, return next in chain (if there is one) if this | |
2223 | * was the last reference to this request. | |
2224 | */ | |
0d85035a | 2225 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 2226 | { |
9b5f7bd9 PB |
2227 | struct io_kiocb *nxt = NULL; |
2228 | ||
de9b4cca | 2229 | if (req_ref_put_and_test(req)) { |
9b5f7bd9 | 2230 | nxt = io_req_find_next(req); |
4d7dd462 | 2231 | __io_free_req(req); |
2a44f467 | 2232 | } |
9b5f7bd9 | 2233 | return nxt; |
2b188cc1 JA |
2234 | } |
2235 | ||
0d85035a | 2236 | static inline void io_put_req(struct io_kiocb *req) |
e65ef56d | 2237 | { |
de9b4cca | 2238 | if (req_ref_put_and_test(req)) |
e65ef56d | 2239 | io_free_req(req); |
2b188cc1 JA |
2240 | } |
2241 | ||
91c2f697 | 2242 | static inline void io_put_req_deferred(struct io_kiocb *req) |
216578e5 | 2243 | { |
91c2f697 | 2244 | if (req_ref_put_and_test(req)) { |
543af3a1 PB |
2245 | req->io_task_work.func = io_free_req; |
2246 | io_req_task_work_add(req); | |
2247 | } | |
216578e5 PB |
2248 | } |
2249 | ||
6c503150 | 2250 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
2251 | { |
2252 | /* See comment at the top of this file */ | |
2253 | smp_rmb(); | |
e23de15f | 2254 | return __io_cqring_events(ctx); |
a3a0e43f JA |
2255 | } |
2256 | ||
fb5ccc98 PB |
2257 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
2258 | { | |
2259 | struct io_rings *rings = ctx->rings; | |
2260 | ||
2261 | /* make sure SQ entry isn't read before tail */ | |
2262 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
2263 | } | |
2264 | ||
8ff069bf | 2265 | static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf) |
e94f141b | 2266 | { |
8ff069bf | 2267 | unsigned int cflags; |
e94f141b | 2268 | |
bcda7baa JA |
2269 | cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT; |
2270 | cflags |= IORING_CQE_F_BUFFER; | |
0e1b6fe3 | 2271 | req->flags &= ~REQ_F_BUFFER_SELECTED; |
bcda7baa JA |
2272 | kfree(kbuf); |
2273 | return cflags; | |
e94f141b JA |
2274 | } |
2275 | ||
8ff069bf | 2276 | static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req) |
bcda7baa | 2277 | { |
4d954c25 | 2278 | struct io_buffer *kbuf; |
bcda7baa | 2279 | |
4d954c25 | 2280 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; |
8ff069bf PB |
2281 | return io_put_kbuf(req, kbuf); |
2282 | } | |
2283 | ||
4c6e277c JA |
2284 | static inline bool io_run_task_work(void) |
2285 | { | |
ef98eb04 | 2286 | if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) { |
4c6e277c | 2287 | __set_current_state(TASK_RUNNING); |
ef98eb04 | 2288 | tracehook_notify_signal(); |
4c6e277c JA |
2289 | return true; |
2290 | } | |
2291 | ||
2292 | return false; | |
bcda7baa JA |
2293 | } |
2294 | ||
def596e9 JA |
2295 | /* |
2296 | * Find and free completed poll iocbs | |
2297 | */ | |
2298 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
a8576af9 | 2299 | struct list_head *done) |
def596e9 | 2300 | { |
8237e045 | 2301 | struct req_batch rb; |
def596e9 | 2302 | struct io_kiocb *req; |
bbde017a XW |
2303 | |
2304 | /* order with ->result store in io_complete_rw_iopoll() */ | |
2305 | smp_rmb(); | |
def596e9 | 2306 | |
5af1d13e | 2307 | io_init_req_batch(&rb); |
def596e9 | 2308 | while (!list_empty(done)) { |
bcda7baa JA |
2309 | int cflags = 0; |
2310 | ||
d21ffe7e | 2311 | req = list_first_entry(done, struct io_kiocb, inflight_entry); |
f161340d PB |
2312 | list_del(&req->inflight_entry); |
2313 | ||
a8576af9 | 2314 | if (READ_ONCE(req->result) == -EAGAIN && |
8c130827 | 2315 | !(req->flags & REQ_F_DONT_REISSUE)) { |
bbde017a | 2316 | req->iopoll_completed = 0; |
773af691 | 2317 | io_req_task_queue_reissue(req); |
8c130827 | 2318 | continue; |
bbde017a | 2319 | } |
def596e9 | 2320 | |
bcda7baa | 2321 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2322 | cflags = io_put_rw_kbuf(req); |
bcda7baa | 2323 | |
d4d19c19 | 2324 | __io_cqring_fill_event(ctx, req->user_data, req->result, cflags); |
def596e9 JA |
2325 | (*nr_events)++; |
2326 | ||
de9b4cca | 2327 | if (req_ref_put_and_test(req)) |
6ff119a6 | 2328 | io_req_free_batch(&rb, req, &ctx->submit_state); |
def596e9 | 2329 | } |
def596e9 | 2330 | |
09bb8394 | 2331 | io_commit_cqring(ctx); |
80c18e4a | 2332 | io_cqring_ev_posted_iopoll(ctx); |
2d6500d4 | 2333 | io_req_free_batch_finish(ctx, &rb); |
581f9810 BM |
2334 | } |
2335 | ||
def596e9 | 2336 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, |
a8576af9 | 2337 | long min) |
def596e9 JA |
2338 | { |
2339 | struct io_kiocb *req, *tmp; | |
2340 | LIST_HEAD(done); | |
2341 | bool spin; | |
def596e9 JA |
2342 | |
2343 | /* | |
2344 | * Only spin for completions if we don't have multiple devices hanging | |
2345 | * off our complete list, and we're under the requested amount. | |
2346 | */ | |
915b3dde | 2347 | spin = !ctx->poll_multi_queue && *nr_events < min; |
def596e9 | 2348 | |
d21ffe7e | 2349 | list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) { |
9adbd45d | 2350 | struct kiocb *kiocb = &req->rw.kiocb; |
a2416e1e | 2351 | int ret; |
def596e9 JA |
2352 | |
2353 | /* | |
581f9810 BM |
2354 | * Move completed and retryable entries to our local lists. |
2355 | * If we find a request that requires polling, break out | |
2356 | * and complete those lists first, if we have entries there. | |
def596e9 | 2357 | */ |
65a6543d | 2358 | if (READ_ONCE(req->iopoll_completed)) { |
d21ffe7e | 2359 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2360 | continue; |
2361 | } | |
2362 | if (!list_empty(&done)) | |
2363 | break; | |
2364 | ||
2365 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
a2416e1e PB |
2366 | if (unlikely(ret < 0)) |
2367 | return ret; | |
2368 | else if (ret) | |
2369 | spin = false; | |
def596e9 | 2370 | |
3aadc23e PB |
2371 | /* iopoll may have completed current req */ |
2372 | if (READ_ONCE(req->iopoll_completed)) | |
d21ffe7e | 2373 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2374 | } |
2375 | ||
2376 | if (!list_empty(&done)) | |
a8576af9 | 2377 | io_iopoll_complete(ctx, nr_events, &done); |
def596e9 | 2378 | |
a2416e1e | 2379 | return 0; |
def596e9 JA |
2380 | } |
2381 | ||
def596e9 JA |
2382 | /* |
2383 | * We can't just wait for polled events to come to us, we have to actively | |
2384 | * find and complete them. | |
2385 | */ | |
b2edc0a7 | 2386 | static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
2387 | { |
2388 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
2389 | return; | |
2390 | ||
2391 | mutex_lock(&ctx->uring_lock); | |
540e32a0 | 2392 | while (!list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2393 | unsigned int nr_events = 0; |
2394 | ||
a8576af9 | 2395 | io_do_iopoll(ctx, &nr_events, 0); |
08f5439f | 2396 | |
b2edc0a7 PB |
2397 | /* let it sleep and repeat later if can't complete a request */ |
2398 | if (nr_events == 0) | |
2399 | break; | |
08f5439f JA |
2400 | /* |
2401 | * Ensure we allow local-to-the-cpu processing to take place, | |
2402 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 2403 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 2404 | */ |
3fcee5a6 PB |
2405 | if (need_resched()) { |
2406 | mutex_unlock(&ctx->uring_lock); | |
2407 | cond_resched(); | |
2408 | mutex_lock(&ctx->uring_lock); | |
2409 | } | |
def596e9 JA |
2410 | } |
2411 | mutex_unlock(&ctx->uring_lock); | |
2412 | } | |
2413 | ||
7668b92a | 2414 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 2415 | { |
7668b92a | 2416 | unsigned int nr_events = 0; |
e9979b36 | 2417 | int ret = 0; |
500f9fba | 2418 | |
c7849be9 XW |
2419 | /* |
2420 | * We disallow the app entering submit/complete with polling, but we | |
2421 | * still need to lock the ring to prevent racing with polled issue | |
2422 | * that got punted to a workqueue. | |
2423 | */ | |
2424 | mutex_lock(&ctx->uring_lock); | |
f39c8a5b PB |
2425 | /* |
2426 | * Don't enter poll loop if we already have events pending. | |
2427 | * If we do, we can potentially be spinning for commands that | |
2428 | * already triggered a CQE (eg in error). | |
2429 | */ | |
5ed7a37d | 2430 | if (test_bit(0, &ctx->check_cq_overflow)) |
f39c8a5b PB |
2431 | __io_cqring_overflow_flush(ctx, false); |
2432 | if (io_cqring_events(ctx)) | |
2433 | goto out; | |
def596e9 | 2434 | do { |
500f9fba JA |
2435 | /* |
2436 | * If a submit got punted to a workqueue, we can have the | |
2437 | * application entering polling for a command before it gets | |
2438 | * issued. That app will hold the uring_lock for the duration | |
2439 | * of the poll right here, so we need to take a breather every | |
2440 | * now and then to ensure that the issue has a chance to add | |
2441 | * the poll to the issued list. Otherwise we can spin here | |
2442 | * forever, while the workqueue is stuck trying to acquire the | |
2443 | * very same mutex. | |
2444 | */ | |
e9979b36 | 2445 | if (list_empty(&ctx->iopoll_list)) { |
8f487ef2 PB |
2446 | u32 tail = ctx->cached_cq_tail; |
2447 | ||
500f9fba | 2448 | mutex_unlock(&ctx->uring_lock); |
4c6e277c | 2449 | io_run_task_work(); |
500f9fba | 2450 | mutex_lock(&ctx->uring_lock); |
def596e9 | 2451 | |
8f487ef2 PB |
2452 | /* some requests don't go through iopoll_list */ |
2453 | if (tail != ctx->cached_cq_tail || | |
2454 | list_empty(&ctx->iopoll_list)) | |
e9979b36 | 2455 | break; |
500f9fba | 2456 | } |
a8576af9 | 2457 | ret = io_do_iopoll(ctx, &nr_events, min); |
f39c8a5b PB |
2458 | } while (!ret && nr_events < min && !need_resched()); |
2459 | out: | |
500f9fba | 2460 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
2461 | return ret; |
2462 | } | |
2463 | ||
491381ce | 2464 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 2465 | { |
491381ce JA |
2466 | /* |
2467 | * Tell lockdep we inherited freeze protection from submission | |
2468 | * thread. | |
2469 | */ | |
2470 | if (req->flags & REQ_F_ISREG) { | |
1c98679d | 2471 | struct super_block *sb = file_inode(req->file)->i_sb; |
2b188cc1 | 2472 | |
1c98679d PB |
2473 | __sb_writers_acquired(sb, SB_FREEZE_WRITE); |
2474 | sb_end_write(sb); | |
2b188cc1 JA |
2475 | } |
2476 | } | |
2477 | ||
b63534c4 | 2478 | #ifdef CONFIG_BLOCK |
dc2a6e9a | 2479 | static bool io_resubmit_prep(struct io_kiocb *req) |
b63534c4 | 2480 | { |
ab454438 | 2481 | struct io_async_rw *rw = req->async_data; |
b63534c4 | 2482 | |
ab454438 PB |
2483 | if (!rw) |
2484 | return !io_req_prep_async(req); | |
2485 | /* may have left rw->iter inconsistent on -EIOCBQUEUED */ | |
2486 | iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter)); | |
2487 | return true; | |
b63534c4 | 2488 | } |
b63534c4 | 2489 | |
3e6a0d3c | 2490 | static bool io_rw_should_reissue(struct io_kiocb *req) |
b63534c4 | 2491 | { |
355afaeb | 2492 | umode_t mode = file_inode(req->file)->i_mode; |
3e6a0d3c | 2493 | struct io_ring_ctx *ctx = req->ctx; |
b63534c4 | 2494 | |
355afaeb JA |
2495 | if (!S_ISBLK(mode) && !S_ISREG(mode)) |
2496 | return false; | |
3e6a0d3c JA |
2497 | if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() && |
2498 | !(ctx->flags & IORING_SETUP_IOPOLL))) | |
b63534c4 | 2499 | return false; |
7c977a58 JA |
2500 | /* |
2501 | * If ref is dying, we might be running poll reap from the exit work. | |
2502 | * Don't attempt to reissue from that path, just let it fail with | |
2503 | * -EAGAIN. | |
2504 | */ | |
3e6a0d3c JA |
2505 | if (percpu_ref_is_dying(&ctx->refs)) |
2506 | return false; | |
ef046888 JA |
2507 | /* |
2508 | * Play it safe and assume not safe to re-import and reissue if we're | |
2509 | * not in the original thread group (or in task context). | |
2510 | */ | |
2511 | if (!same_thread_group(req->task, current) || !in_task()) | |
2512 | return false; | |
3e6a0d3c JA |
2513 | return true; |
2514 | } | |
e82ad485 | 2515 | #else |
a1ff1e3f | 2516 | static bool io_resubmit_prep(struct io_kiocb *req) |
e82ad485 JA |
2517 | { |
2518 | return false; | |
2519 | } | |
e82ad485 | 2520 | static bool io_rw_should_reissue(struct io_kiocb *req) |
3e6a0d3c | 2521 | { |
b63534c4 JA |
2522 | return false; |
2523 | } | |
3e6a0d3c | 2524 | #endif |
b63534c4 | 2525 | |
8ef12efe | 2526 | static bool __io_complete_rw_common(struct io_kiocb *req, long res) |
a1d7c393 | 2527 | { |
b65c128f PB |
2528 | if (req->rw.kiocb.ki_flags & IOCB_WRITE) |
2529 | kiocb_end_write(req); | |
9532b99b PB |
2530 | if (res != req->result) { |
2531 | if ((res == -EAGAIN || res == -EOPNOTSUPP) && | |
2532 | io_rw_should_reissue(req)) { | |
2533 | req->flags |= REQ_F_REISSUE; | |
8ef12efe | 2534 | return true; |
9532b99b | 2535 | } |
93d2bcd2 | 2536 | req_set_fail(req); |
8ef12efe | 2537 | req->result = res; |
9532b99b | 2538 | } |
8ef12efe JA |
2539 | return false; |
2540 | } | |
2541 | ||
2542 | static void io_req_task_complete(struct io_kiocb *req) | |
2543 | { | |
2544 | int cflags = 0; | |
2545 | ||
2f8e45f1 PB |
2546 | if (req->flags & REQ_F_BUFFER_SELECTED) |
2547 | cflags = io_put_rw_kbuf(req); | |
8ef12efe JA |
2548 | __io_req_complete(req, 0, req->result, cflags); |
2549 | } | |
2550 | ||
2551 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, | |
2552 | unsigned int issue_flags) | |
2553 | { | |
2554 | if (__io_complete_rw_common(req, res)) | |
2555 | return; | |
2556 | io_req_task_complete(req); | |
ba816ad6 JA |
2557 | } |
2558 | ||
2559 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
2560 | { | |
9adbd45d | 2561 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 | 2562 | |
8ef12efe JA |
2563 | if (__io_complete_rw_common(req, res)) |
2564 | return; | |
2565 | req->result = res; | |
2566 | req->io_task_work.func = io_req_task_complete; | |
2567 | io_req_task_work_add(req); | |
2b188cc1 JA |
2568 | } |
2569 | ||
def596e9 JA |
2570 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
2571 | { | |
9adbd45d | 2572 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 2573 | |
491381ce JA |
2574 | if (kiocb->ki_flags & IOCB_WRITE) |
2575 | kiocb_end_write(req); | |
9532b99b | 2576 | if (unlikely(res != req->result)) { |
a1ff1e3f JA |
2577 | if (!(res == -EAGAIN && io_rw_should_reissue(req) && |
2578 | io_resubmit_prep(req))) { | |
93d2bcd2 | 2579 | req_set_fail(req); |
9532b99b PB |
2580 | req->flags |= REQ_F_DONT_REISSUE; |
2581 | } | |
8c130827 | 2582 | } |
bbde017a XW |
2583 | |
2584 | WRITE_ONCE(req->result, res); | |
b9b0e0d3 | 2585 | /* order with io_iopoll_complete() checking ->result */ |
cd664b0e PB |
2586 | smp_wmb(); |
2587 | WRITE_ONCE(req->iopoll_completed, 1); | |
def596e9 JA |
2588 | } |
2589 | ||
2590 | /* | |
2591 | * After the iocb has been issued, it's safe to be found on the poll list. | |
2592 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 2593 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
2594 | * accessing the kiocb cookie. |
2595 | */ | |
cb3d8972 | 2596 | static void io_iopoll_req_issued(struct io_kiocb *req) |
def596e9 JA |
2597 | { |
2598 | struct io_ring_ctx *ctx = req->ctx; | |
cb3d8972 PB |
2599 | const bool in_async = io_wq_current_is_worker(); |
2600 | ||
2601 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
2602 | if (unlikely(in_async)) | |
2603 | mutex_lock(&ctx->uring_lock); | |
def596e9 JA |
2604 | |
2605 | /* | |
2606 | * Track whether we have multiple files in our lists. This will impact | |
2607 | * how we do polling eventually, not spinning if we're on potentially | |
2608 | * different devices. | |
2609 | */ | |
540e32a0 | 2610 | if (list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
2611 | ctx->poll_multi_queue = false; |
2612 | } else if (!ctx->poll_multi_queue) { | |
def596e9 | 2613 | struct io_kiocb *list_req; |
915b3dde | 2614 | unsigned int queue_num0, queue_num1; |
def596e9 | 2615 | |
540e32a0 | 2616 | list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb, |
d21ffe7e | 2617 | inflight_entry); |
915b3dde HX |
2618 | |
2619 | if (list_req->file != req->file) { | |
2620 | ctx->poll_multi_queue = true; | |
2621 | } else { | |
2622 | queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie); | |
2623 | queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie); | |
2624 | if (queue_num0 != queue_num1) | |
2625 | ctx->poll_multi_queue = true; | |
2626 | } | |
def596e9 JA |
2627 | } |
2628 | ||
2629 | /* | |
2630 | * For fast devices, IO may have already completed. If it has, add | |
2631 | * it to the front so we find it first. | |
2632 | */ | |
65a6543d | 2633 | if (READ_ONCE(req->iopoll_completed)) |
d21ffe7e | 2634 | list_add(&req->inflight_entry, &ctx->iopoll_list); |
def596e9 | 2635 | else |
d21ffe7e | 2636 | list_add_tail(&req->inflight_entry, &ctx->iopoll_list); |
bdcd3eab | 2637 | |
cb3d8972 PB |
2638 | if (unlikely(in_async)) { |
2639 | /* | |
2640 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
2641 | * in sq thread task context or in io worker task context. If | |
2642 | * current task context is sq thread, we don't need to check | |
2643 | * whether should wake up sq thread. | |
2644 | */ | |
2645 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
2646 | wq_has_sleeper(&ctx->sq_data->wait)) | |
2647 | wake_up(&ctx->sq_data->wait); | |
2648 | ||
2649 | mutex_unlock(&ctx->uring_lock); | |
2650 | } | |
def596e9 JA |
2651 | } |
2652 | ||
4503b767 JA |
2653 | static bool io_bdev_nowait(struct block_device *bdev) |
2654 | { | |
9ba0d0c8 | 2655 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
2656 | } |
2657 | ||
2b188cc1 JA |
2658 | /* |
2659 | * If we tracked the file through the SCM inflight mechanism, we could support | |
2660 | * any file. For now, just ensure that anything potentially problematic is done | |
2661 | * inline. | |
2662 | */ | |
b191e2df | 2663 | static bool __io_file_supports_nowait(struct file *file, int rw) |
2b188cc1 JA |
2664 | { |
2665 | umode_t mode = file_inode(file)->i_mode; | |
2666 | ||
4503b767 | 2667 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
2668 | if (IS_ENABLED(CONFIG_BLOCK) && |
2669 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
2670 | return true; |
2671 | return false; | |
2672 | } | |
976517f1 | 2673 | if (S_ISSOCK(mode)) |
2b188cc1 | 2674 | return true; |
4503b767 | 2675 | if (S_ISREG(mode)) { |
4e7b5671 CH |
2676 | if (IS_ENABLED(CONFIG_BLOCK) && |
2677 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
4503b767 JA |
2678 | file->f_op != &io_uring_fops) |
2679 | return true; | |
2680 | return false; | |
2681 | } | |
2b188cc1 | 2682 | |
c5b85625 JA |
2683 | /* any ->read/write should understand O_NONBLOCK */ |
2684 | if (file->f_flags & O_NONBLOCK) | |
2685 | return true; | |
2686 | ||
af197f50 JA |
2687 | if (!(file->f_mode & FMODE_NOWAIT)) |
2688 | return false; | |
2689 | ||
2690 | if (rw == READ) | |
2691 | return file->f_op->read_iter != NULL; | |
2692 | ||
2693 | return file->f_op->write_iter != NULL; | |
2b188cc1 JA |
2694 | } |
2695 | ||
b191e2df | 2696 | static bool io_file_supports_nowait(struct io_kiocb *req, int rw) |
7b29f92d | 2697 | { |
b191e2df | 2698 | if (rw == READ && (req->flags & REQ_F_NOWAIT_READ)) |
7b29f92d | 2699 | return true; |
b191e2df | 2700 | else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE)) |
7b29f92d JA |
2701 | return true; |
2702 | ||
b191e2df | 2703 | return __io_file_supports_nowait(req->file, rw); |
7b29f92d JA |
2704 | } |
2705 | ||
a88fc400 | 2706 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 2707 | { |
def596e9 | 2708 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 2709 | struct kiocb *kiocb = &req->rw.kiocb; |
75c668cd | 2710 | struct file *file = req->file; |
09bb8394 JA |
2711 | unsigned ioprio; |
2712 | int ret; | |
2b188cc1 | 2713 | |
c97d8a0f | 2714 | if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode)) |
491381ce JA |
2715 | req->flags |= REQ_F_ISREG; |
2716 | ||
2b188cc1 | 2717 | kiocb->ki_pos = READ_ONCE(sqe->off); |
75c668cd | 2718 | if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) { |
ba04291e | 2719 | req->flags |= REQ_F_CUR_POS; |
75c668cd | 2720 | kiocb->ki_pos = file->f_pos; |
ba04291e | 2721 | } |
2b188cc1 | 2722 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
2723 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
2724 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
2725 | if (unlikely(ret)) | |
2726 | return ret; | |
2b188cc1 | 2727 | |
75c668cd PB |
2728 | /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */ |
2729 | if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK)) | |
2730 | req->flags |= REQ_F_NOWAIT; | |
2731 | ||
2b188cc1 JA |
2732 | ioprio = READ_ONCE(sqe->ioprio); |
2733 | if (ioprio) { | |
2734 | ret = ioprio_check_cap(ioprio); | |
2735 | if (ret) | |
09bb8394 | 2736 | return ret; |
2b188cc1 JA |
2737 | |
2738 | kiocb->ki_ioprio = ioprio; | |
2739 | } else | |
2740 | kiocb->ki_ioprio = get_current_ioprio(); | |
2741 | ||
def596e9 | 2742 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2743 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2744 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2745 | return -EOPNOTSUPP; |
2b188cc1 | 2746 | |
def596e9 JA |
2747 | kiocb->ki_flags |= IOCB_HIPRI; |
2748 | kiocb->ki_complete = io_complete_rw_iopoll; | |
65a6543d | 2749 | req->iopoll_completed = 0; |
def596e9 | 2750 | } else { |
09bb8394 JA |
2751 | if (kiocb->ki_flags & IOCB_HIPRI) |
2752 | return -EINVAL; | |
def596e9 JA |
2753 | kiocb->ki_complete = io_complete_rw; |
2754 | } | |
9adbd45d | 2755 | |
eae071c9 PB |
2756 | if (req->opcode == IORING_OP_READ_FIXED || |
2757 | req->opcode == IORING_OP_WRITE_FIXED) { | |
2758 | req->imu = NULL; | |
2759 | io_req_set_rsrc_node(req); | |
2760 | } | |
2761 | ||
3529d8c2 JA |
2762 | req->rw.addr = READ_ONCE(sqe->addr); |
2763 | req->rw.len = READ_ONCE(sqe->len); | |
4f4eeba8 | 2764 | req->buf_index = READ_ONCE(sqe->buf_index); |
2b188cc1 | 2765 | return 0; |
2b188cc1 JA |
2766 | } |
2767 | ||
2768 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2769 | { | |
2770 | switch (ret) { | |
2771 | case -EIOCBQUEUED: | |
2772 | break; | |
2773 | case -ERESTARTSYS: | |
2774 | case -ERESTARTNOINTR: | |
2775 | case -ERESTARTNOHAND: | |
2776 | case -ERESTART_RESTARTBLOCK: | |
2777 | /* | |
2778 | * We can't just restart the syscall, since previously | |
2779 | * submitted sqes may already be in progress. Just fail this | |
2780 | * IO with EINTR. | |
2781 | */ | |
2782 | ret = -EINTR; | |
df561f66 | 2783 | fallthrough; |
2b188cc1 JA |
2784 | default: |
2785 | kiocb->ki_complete(kiocb, ret, 0); | |
2786 | } | |
2787 | } | |
2788 | ||
a1d7c393 | 2789 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, |
889fca73 | 2790 | unsigned int issue_flags) |
ba816ad6 | 2791 | { |
ba04291e | 2792 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
e8c2bc1f | 2793 | struct io_async_rw *io = req->async_data; |
97284637 | 2794 | bool check_reissue = kiocb->ki_complete == io_complete_rw; |
ba04291e | 2795 | |
227c0c96 | 2796 | /* add previously done IO, if any */ |
e8c2bc1f | 2797 | if (io && io->bytes_done > 0) { |
227c0c96 | 2798 | if (ret < 0) |
e8c2bc1f | 2799 | ret = io->bytes_done; |
227c0c96 | 2800 | else |
e8c2bc1f | 2801 | ret += io->bytes_done; |
227c0c96 JA |
2802 | } |
2803 | ||
ba04291e JA |
2804 | if (req->flags & REQ_F_CUR_POS) |
2805 | req->file->f_pos = kiocb->ki_pos; | |
e149bd74 | 2806 | if (ret >= 0 && check_reissue) |
889fca73 | 2807 | __io_complete_rw(req, ret, 0, issue_flags); |
ba816ad6 JA |
2808 | else |
2809 | io_rw_done(kiocb, ret); | |
97284637 | 2810 | |
fe7e3257 | 2811 | if (check_reissue && (req->flags & REQ_F_REISSUE)) { |
97284637 | 2812 | req->flags &= ~REQ_F_REISSUE; |
a7be7c23 | 2813 | if (io_resubmit_prep(req)) { |
773af691 | 2814 | io_req_task_queue_reissue(req); |
8c130827 | 2815 | } else { |
97284637 PB |
2816 | int cflags = 0; |
2817 | ||
93d2bcd2 | 2818 | req_set_fail(req); |
97284637 PB |
2819 | if (req->flags & REQ_F_BUFFER_SELECTED) |
2820 | cflags = io_put_rw_kbuf(req); | |
2821 | __io_req_complete(req, issue_flags, ret, cflags); | |
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 | ||
18bceab1 | 5065 | /* already have a 2nd entry, fail a third attempt */ |
807abcb0 | 5066 | if (*poll_ptr) { |
18bceab1 JA |
5067 | pt->error = -EINVAL; |
5068 | return; | |
5069 | } | |
ea6a693d JA |
5070 | /* |
5071 | * Can't handle multishot for double wait for now, turn it | |
5072 | * into one-shot mode. | |
5073 | */ | |
7a274727 PB |
5074 | if (!(poll_one->events & EPOLLONESHOT)) |
5075 | poll_one->events |= EPOLLONESHOT; | |
1c3b3e65 | 5076 | /* double add on the same waitqueue head, ignore */ |
7a274727 | 5077 | if (poll_one->head == head) |
1c3b3e65 | 5078 | return; |
18bceab1 JA |
5079 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); |
5080 | if (!poll) { | |
5081 | pt->error = -ENOMEM; | |
5082 | return; | |
5083 | } | |
58852d4d | 5084 | io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake); |
de9b4cca | 5085 | req_ref_get(req); |
18bceab1 | 5086 | poll->wait.private = req; |
807abcb0 | 5087 | *poll_ptr = poll; |
18bceab1 JA |
5088 | } |
5089 | ||
68b11e8b | 5090 | pt->nr_entries++; |
18bceab1 | 5091 | poll->head = head; |
a31eb4a2 JX |
5092 | |
5093 | if (poll->events & EPOLLEXCLUSIVE) | |
5094 | add_wait_queue_exclusive(head, &poll->wait); | |
5095 | else | |
5096 | add_wait_queue(head, &poll->wait); | |
18bceab1 JA |
5097 | } |
5098 | ||
5099 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
5100 | struct poll_table_struct *p) | |
5101 | { | |
5102 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
807abcb0 | 5103 | struct async_poll *apoll = pt->req->apoll; |
18bceab1 | 5104 | |
807abcb0 | 5105 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
18bceab1 JA |
5106 | } |
5107 | ||
5b0a6acc | 5108 | static void io_async_task_func(struct io_kiocb *req) |
d7718a9d | 5109 | { |
d7718a9d JA |
5110 | struct async_poll *apoll = req->apoll; |
5111 | struct io_ring_ctx *ctx = req->ctx; | |
5112 | ||
236daeae | 5113 | trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data); |
d7718a9d | 5114 | |
74ce6ce4 | 5115 | if (io_poll_rewait(req, &apoll->poll)) { |
79ebeaee | 5116 | spin_unlock(&ctx->completion_lock); |
74ce6ce4 | 5117 | return; |
d7718a9d JA |
5118 | } |
5119 | ||
0ea13b44 | 5120 | hash_del(&req->hash_node); |
d4e7cd36 | 5121 | io_poll_remove_double(req); |
79ebeaee | 5122 | spin_unlock(&ctx->completion_lock); |
74ce6ce4 | 5123 | |
0be0b0e3 | 5124 | if (!READ_ONCE(apoll->poll.canceled)) |
5b0a6acc | 5125 | io_req_task_submit(req); |
0be0b0e3 | 5126 | else |
2593553a | 5127 | io_req_complete_failed(req, -ECANCELED); |
d7718a9d JA |
5128 | } |
5129 | ||
5130 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
5131 | void *key) | |
5132 | { | |
5133 | struct io_kiocb *req = wait->private; | |
5134 | struct io_poll_iocb *poll = &req->apoll->poll; | |
5135 | ||
5136 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
5137 | key_to_poll(key)); | |
5138 | ||
5139 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
5140 | } | |
5141 | ||
5142 | static void io_poll_req_insert(struct io_kiocb *req) | |
5143 | { | |
5144 | struct io_ring_ctx *ctx = req->ctx; | |
5145 | struct hlist_head *list; | |
5146 | ||
5147 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
5148 | hlist_add_head(&req->hash_node, list); | |
5149 | } | |
5150 | ||
5151 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
5152 | struct io_poll_iocb *poll, | |
5153 | struct io_poll_table *ipt, __poll_t mask, | |
5154 | wait_queue_func_t wake_func) | |
5155 | __acquires(&ctx->completion_lock) | |
5156 | { | |
5157 | struct io_ring_ctx *ctx = req->ctx; | |
5158 | bool cancel = false; | |
5159 | ||
4d52f338 | 5160 | INIT_HLIST_NODE(&req->hash_node); |
18bceab1 | 5161 | io_init_poll_iocb(poll, mask, wake_func); |
b90cd197 | 5162 | poll->file = req->file; |
18bceab1 | 5163 | poll->wait.private = req; |
d7718a9d JA |
5164 | |
5165 | ipt->pt._key = mask; | |
5166 | ipt->req = req; | |
68b11e8b PB |
5167 | ipt->error = 0; |
5168 | ipt->nr_entries = 0; | |
d7718a9d | 5169 | |
d7718a9d | 5170 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
68b11e8b PB |
5171 | if (unlikely(!ipt->nr_entries) && !ipt->error) |
5172 | ipt->error = -EINVAL; | |
d7718a9d | 5173 | |
79ebeaee | 5174 | spin_lock(&ctx->completion_lock); |
a890d01e | 5175 | if (ipt->error || (mask && (poll->events & EPOLLONESHOT))) |
46fee9ab | 5176 | io_poll_remove_double(req); |
d7718a9d | 5177 | if (likely(poll->head)) { |
79ebeaee | 5178 | spin_lock_irq(&poll->head->lock); |
d7718a9d JA |
5179 | if (unlikely(list_empty(&poll->wait.entry))) { |
5180 | if (ipt->error) | |
5181 | cancel = true; | |
5182 | ipt->error = 0; | |
5183 | mask = 0; | |
5184 | } | |
88e41cf9 | 5185 | if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error) |
d7718a9d JA |
5186 | list_del_init(&poll->wait.entry); |
5187 | else if (cancel) | |
5188 | WRITE_ONCE(poll->canceled, true); | |
5189 | else if (!poll->done) /* actually waiting for an event */ | |
5190 | io_poll_req_insert(req); | |
79ebeaee | 5191 | spin_unlock_irq(&poll->head->lock); |
d7718a9d JA |
5192 | } |
5193 | ||
5194 | return mask; | |
5195 | } | |
5196 | ||
59b735ae OL |
5197 | enum { |
5198 | IO_APOLL_OK, | |
5199 | IO_APOLL_ABORTED, | |
5200 | IO_APOLL_READY | |
5201 | }; | |
5202 | ||
5203 | static int io_arm_poll_handler(struct io_kiocb *req) | |
d7718a9d JA |
5204 | { |
5205 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
5206 | struct io_ring_ctx *ctx = req->ctx; | |
5207 | struct async_poll *apoll; | |
5208 | struct io_poll_table ipt; | |
b2d9c3da | 5209 | __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI; |
9dab14b8 | 5210 | int rw; |
d7718a9d JA |
5211 | |
5212 | if (!req->file || !file_can_poll(req->file)) | |
59b735ae | 5213 | return IO_APOLL_ABORTED; |
24c74678 | 5214 | if (req->flags & REQ_F_POLLED) |
59b735ae | 5215 | return IO_APOLL_ABORTED; |
b2d9c3da PB |
5216 | if (!def->pollin && !def->pollout) |
5217 | return IO_APOLL_ABORTED; | |
5218 | ||
5219 | if (def->pollin) { | |
9dab14b8 | 5220 | rw = READ; |
b2d9c3da PB |
5221 | mask |= POLLIN | POLLRDNORM; |
5222 | ||
5223 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
5224 | if ((req->opcode == IORING_OP_RECVMSG) && | |
5225 | (req->sr_msg.msg_flags & MSG_ERRQUEUE)) | |
5226 | mask &= ~POLLIN; | |
5227 | } else { | |
9dab14b8 | 5228 | rw = WRITE; |
b2d9c3da PB |
5229 | mask |= POLLOUT | POLLWRNORM; |
5230 | } | |
5231 | ||
9dab14b8 | 5232 | /* if we can't nonblock try, then no point in arming a poll handler */ |
b191e2df | 5233 | if (!io_file_supports_nowait(req, rw)) |
59b735ae | 5234 | return IO_APOLL_ABORTED; |
d7718a9d JA |
5235 | |
5236 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
5237 | if (unlikely(!apoll)) | |
59b735ae | 5238 | return IO_APOLL_ABORTED; |
807abcb0 | 5239 | apoll->double_poll = NULL; |
d7718a9d | 5240 | req->apoll = apoll; |
b2d9c3da | 5241 | req->flags |= REQ_F_POLLED; |
d7718a9d | 5242 | ipt.pt._qproc = io_async_queue_proc; |
48dcd38d | 5243 | io_req_set_refcount(req); |
d7718a9d JA |
5244 | |
5245 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
5246 | io_async_wake); | |
79ebeaee | 5247 | spin_unlock(&ctx->completion_lock); |
41a5169c HX |
5248 | if (ret || ipt.error) |
5249 | return ret ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
5250 | ||
236daeae OL |
5251 | trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data, |
5252 | mask, apoll->poll.events); | |
59b735ae | 5253 | return IO_APOLL_OK; |
d7718a9d JA |
5254 | } |
5255 | ||
5256 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
b2e720ac | 5257 | struct io_poll_iocb *poll, bool do_cancel) |
e07785b0 | 5258 | __must_hold(&req->ctx->completion_lock) |
221c5eb2 | 5259 | { |
b41e9852 | 5260 | bool do_complete = false; |
221c5eb2 | 5261 | |
5082620f JA |
5262 | if (!poll->head) |
5263 | return false; | |
79ebeaee | 5264 | spin_lock_irq(&poll->head->lock); |
b2e720ac JA |
5265 | if (do_cancel) |
5266 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
5267 | if (!list_empty(&poll->wait.entry)) { |
5268 | list_del_init(&poll->wait.entry); | |
b41e9852 | 5269 | do_complete = true; |
221c5eb2 | 5270 | } |
79ebeaee | 5271 | spin_unlock_irq(&poll->head->lock); |
3bfa5bcb | 5272 | hash_del(&req->hash_node); |
d7718a9d JA |
5273 | return do_complete; |
5274 | } | |
5275 | ||
5d709043 | 5276 | static bool io_poll_remove_one(struct io_kiocb *req) |
e07785b0 | 5277 | __must_hold(&req->ctx->completion_lock) |
d7718a9d JA |
5278 | { |
5279 | bool do_complete; | |
5280 | ||
d4e7cd36 | 5281 | io_poll_remove_double(req); |
e31001a3 | 5282 | do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true); |
d4e7cd36 | 5283 | |
b41e9852 | 5284 | if (do_complete) { |
d4d19c19 | 5285 | io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0); |
b41e9852 | 5286 | io_commit_cqring(req->ctx); |
93d2bcd2 | 5287 | req_set_fail(req); |
91c2f697 | 5288 | io_put_req_deferred(req); |
5d709043 | 5289 | } |
b41e9852 | 5290 | return do_complete; |
221c5eb2 JA |
5291 | } |
5292 | ||
76e1b642 JA |
5293 | /* |
5294 | * Returns true if we found and killed one or more poll requests | |
5295 | */ | |
6b81928d | 5296 | static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, |
3dd0c97a | 5297 | bool cancel_all) |
221c5eb2 | 5298 | { |
78076bb6 | 5299 | struct hlist_node *tmp; |
221c5eb2 | 5300 | struct io_kiocb *req; |
8e2e1faf | 5301 | int posted = 0, i; |
221c5eb2 | 5302 | |
79ebeaee | 5303 | spin_lock(&ctx->completion_lock); |
78076bb6 JA |
5304 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
5305 | struct hlist_head *list; | |
5306 | ||
5307 | list = &ctx->cancel_hash[i]; | |
f3606e3a | 5308 | hlist_for_each_entry_safe(req, tmp, list, hash_node) { |
3dd0c97a | 5309 | if (io_match_task(req, tsk, cancel_all)) |
f3606e3a JA |
5310 | posted += io_poll_remove_one(req); |
5311 | } | |
221c5eb2 | 5312 | } |
79ebeaee | 5313 | spin_unlock(&ctx->completion_lock); |
b41e9852 | 5314 | |
8e2e1faf JA |
5315 | if (posted) |
5316 | io_cqring_ev_posted(ctx); | |
76e1b642 JA |
5317 | |
5318 | return posted != 0; | |
221c5eb2 JA |
5319 | } |
5320 | ||
9ba5fac8 PB |
5321 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr, |
5322 | bool poll_only) | |
e07785b0 | 5323 | __must_hold(&ctx->completion_lock) |
47f46768 | 5324 | { |
78076bb6 | 5325 | struct hlist_head *list; |
47f46768 JA |
5326 | struct io_kiocb *req; |
5327 | ||
78076bb6 JA |
5328 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
5329 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
5330 | if (sqe_addr != req->user_data) |
5331 | continue; | |
9ba5fac8 PB |
5332 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) |
5333 | continue; | |
b2cb805f | 5334 | return req; |
47f46768 | 5335 | } |
b2cb805f JA |
5336 | return NULL; |
5337 | } | |
5338 | ||
9ba5fac8 PB |
5339 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr, |
5340 | bool poll_only) | |
e07785b0 | 5341 | __must_hold(&ctx->completion_lock) |
b2cb805f JA |
5342 | { |
5343 | struct io_kiocb *req; | |
5344 | ||
9ba5fac8 | 5345 | req = io_poll_find(ctx, sqe_addr, poll_only); |
b2cb805f JA |
5346 | if (!req) |
5347 | return -ENOENT; | |
5348 | if (io_poll_remove_one(req)) | |
5349 | return 0; | |
5350 | ||
5351 | return -EALREADY; | |
47f46768 JA |
5352 | } |
5353 | ||
9096af3e PB |
5354 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
5355 | unsigned int flags) | |
5356 | { | |
5357 | u32 events; | |
47f46768 | 5358 | |
9096af3e PB |
5359 | events = READ_ONCE(sqe->poll32_events); |
5360 | #ifdef __BIG_ENDIAN | |
5361 | events = swahw32(events); | |
5362 | #endif | |
5363 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
5364 | events |= EPOLLONESHOT; | |
5365 | return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT)); | |
47f46768 JA |
5366 | } |
5367 | ||
c5de0036 | 5368 | static int io_poll_update_prep(struct io_kiocb *req, |
3529d8c2 | 5369 | const struct io_uring_sqe *sqe) |
0969e783 | 5370 | { |
c5de0036 PB |
5371 | struct io_poll_update *upd = &req->poll_update; |
5372 | u32 flags; | |
5373 | ||
0969e783 JA |
5374 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5375 | return -EINVAL; | |
c5de0036 PB |
5376 | if (sqe->ioprio || sqe->buf_index) |
5377 | return -EINVAL; | |
5378 | flags = READ_ONCE(sqe->len); | |
5379 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
5380 | IORING_POLL_ADD_MULTI)) | |
5381 | return -EINVAL; | |
5382 | /* meaningless without update */ | |
5383 | if (flags == IORING_POLL_ADD_MULTI) | |
0969e783 JA |
5384 | return -EINVAL; |
5385 | ||
c5de0036 PB |
5386 | upd->old_user_data = READ_ONCE(sqe->addr); |
5387 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
5388 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
221c5eb2 | 5389 | |
c5de0036 PB |
5390 | upd->new_user_data = READ_ONCE(sqe->off); |
5391 | if (!upd->update_user_data && upd->new_user_data) | |
5392 | return -EINVAL; | |
5393 | if (upd->update_events) | |
5394 | upd->events = io_poll_parse_events(sqe, flags); | |
5395 | else if (sqe->poll32_events) | |
5396 | return -EINVAL; | |
221c5eb2 | 5397 | |
221c5eb2 JA |
5398 | return 0; |
5399 | } | |
5400 | ||
221c5eb2 JA |
5401 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
5402 | void *key) | |
5403 | { | |
c2f2eb7d JA |
5404 | struct io_kiocb *req = wait->private; |
5405 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 5406 | |
d7718a9d | 5407 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
5408 | } |
5409 | ||
221c5eb2 JA |
5410 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
5411 | struct poll_table_struct *p) | |
5412 | { | |
5413 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
5414 | ||
e8c2bc1f | 5415 | __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data); |
eac406c6 JA |
5416 | } |
5417 | ||
3529d8c2 | 5418 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
5419 | { |
5420 | struct io_poll_iocb *poll = &req->poll; | |
c5de0036 | 5421 | u32 flags; |
221c5eb2 JA |
5422 | |
5423 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
5424 | return -EINVAL; | |
c5de0036 | 5425 | if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr) |
88e41cf9 JA |
5426 | return -EINVAL; |
5427 | flags = READ_ONCE(sqe->len); | |
c5de0036 | 5428 | if (flags & ~IORING_POLL_ADD_MULTI) |
221c5eb2 JA |
5429 | return -EINVAL; |
5430 | ||
48dcd38d | 5431 | io_req_set_refcount(req); |
c5de0036 | 5432 | poll->events = io_poll_parse_events(sqe, flags); |
0969e783 JA |
5433 | return 0; |
5434 | } | |
5435 | ||
61e98203 | 5436 | static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) |
0969e783 JA |
5437 | { |
5438 | struct io_poll_iocb *poll = &req->poll; | |
5439 | struct io_ring_ctx *ctx = req->ctx; | |
5440 | struct io_poll_table ipt; | |
0969e783 | 5441 | __poll_t mask; |
0969e783 | 5442 | |
d7718a9d | 5443 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 5444 | |
d7718a9d JA |
5445 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
5446 | io_poll_wake); | |
221c5eb2 | 5447 | |
8c838788 | 5448 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 5449 | ipt.error = 0; |
e27414be | 5450 | io_poll_complete(req, mask); |
221c5eb2 | 5451 | } |
79ebeaee | 5452 | spin_unlock(&ctx->completion_lock); |
221c5eb2 | 5453 | |
8c838788 JA |
5454 | if (mask) { |
5455 | io_cqring_ev_posted(ctx); | |
88e41cf9 JA |
5456 | if (poll->events & EPOLLONESHOT) |
5457 | io_put_req(req); | |
221c5eb2 | 5458 | } |
8c838788 | 5459 | return ipt.error; |
221c5eb2 JA |
5460 | } |
5461 | ||
c5de0036 | 5462 | static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags) |
b69de288 JA |
5463 | { |
5464 | struct io_ring_ctx *ctx = req->ctx; | |
5465 | struct io_kiocb *preq; | |
cb3b200e | 5466 | bool completing; |
b69de288 JA |
5467 | int ret; |
5468 | ||
79ebeaee | 5469 | spin_lock(&ctx->completion_lock); |
9ba5fac8 | 5470 | preq = io_poll_find(ctx, req->poll_update.old_user_data, true); |
b69de288 JA |
5471 | if (!preq) { |
5472 | ret = -ENOENT; | |
5473 | goto err; | |
b69de288 | 5474 | } |
cb3b200e | 5475 | |
c5de0036 PB |
5476 | if (!req->poll_update.update_events && !req->poll_update.update_user_data) { |
5477 | completing = true; | |
5478 | ret = io_poll_remove_one(preq) ? 0 : -EALREADY; | |
5479 | goto err; | |
5480 | } | |
5481 | ||
cb3b200e JA |
5482 | /* |
5483 | * Don't allow racy completion with singleshot, as we cannot safely | |
5484 | * update those. For multishot, if we're racing with completion, just | |
5485 | * let completion re-add it. | |
5486 | */ | |
5487 | completing = !__io_poll_remove_one(preq, &preq->poll, false); | |
5488 | if (completing && (preq->poll.events & EPOLLONESHOT)) { | |
5489 | ret = -EALREADY; | |
5490 | goto err; | |
b69de288 JA |
5491 | } |
5492 | /* we now have a detached poll request. reissue. */ | |
5493 | ret = 0; | |
5494 | err: | |
b69de288 | 5495 | if (ret < 0) { |
79ebeaee | 5496 | spin_unlock(&ctx->completion_lock); |
93d2bcd2 | 5497 | req_set_fail(req); |
b69de288 JA |
5498 | io_req_complete(req, ret); |
5499 | return 0; | |
5500 | } | |
5501 | /* only mask one event flags, keep behavior flags */ | |
9d805892 | 5502 | if (req->poll_update.update_events) { |
b69de288 | 5503 | preq->poll.events &= ~0xffff; |
9d805892 | 5504 | preq->poll.events |= req->poll_update.events & 0xffff; |
b69de288 JA |
5505 | preq->poll.events |= IO_POLL_UNMASK; |
5506 | } | |
9d805892 PB |
5507 | if (req->poll_update.update_user_data) |
5508 | preq->user_data = req->poll_update.new_user_data; | |
79ebeaee | 5509 | spin_unlock(&ctx->completion_lock); |
cb3b200e | 5510 | |
b69de288 JA |
5511 | /* complete update request, we're done with it */ |
5512 | io_req_complete(req, ret); | |
5513 | ||
cb3b200e | 5514 | if (!completing) { |
c5de0036 | 5515 | ret = io_poll_add(preq, issue_flags); |
cb3b200e | 5516 | if (ret < 0) { |
93d2bcd2 | 5517 | req_set_fail(preq); |
cb3b200e JA |
5518 | io_req_complete(preq, ret); |
5519 | } | |
b69de288 JA |
5520 | } |
5521 | return 0; | |
5522 | } | |
5523 | ||
89850fce JA |
5524 | static void io_req_task_timeout(struct io_kiocb *req) |
5525 | { | |
5526 | struct io_ring_ctx *ctx = req->ctx; | |
5527 | ||
79ebeaee | 5528 | spin_lock(&ctx->completion_lock); |
89850fce JA |
5529 | io_cqring_fill_event(ctx, req->user_data, -ETIME, 0); |
5530 | io_commit_cqring(ctx); | |
79ebeaee | 5531 | spin_unlock(&ctx->completion_lock); |
89850fce JA |
5532 | |
5533 | io_cqring_ev_posted(ctx); | |
5534 | req_set_fail(req); | |
5535 | io_put_req(req); | |
5536 | } | |
5537 | ||
5262f567 JA |
5538 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
5539 | { | |
ad8a48ac JA |
5540 | struct io_timeout_data *data = container_of(timer, |
5541 | struct io_timeout_data, timer); | |
5542 | struct io_kiocb *req = data->req; | |
5543 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
5544 | unsigned long flags; |
5545 | ||
89850fce | 5546 | spin_lock_irqsave(&ctx->timeout_lock, flags); |
a71976f3 | 5547 | list_del_init(&req->timeout.list); |
01cec8c1 PB |
5548 | atomic_set(&req->ctx->cq_timeouts, |
5549 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
89850fce | 5550 | spin_unlock_irqrestore(&ctx->timeout_lock, flags); |
01cec8c1 | 5551 | |
89850fce JA |
5552 | req->io_task_work.func = io_req_task_timeout; |
5553 | io_req_task_work_add(req); | |
5262f567 JA |
5554 | return HRTIMER_NORESTART; |
5555 | } | |
5556 | ||
fbd15848 PB |
5557 | static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx, |
5558 | __u64 user_data) | |
89850fce | 5559 | __must_hold(&ctx->timeout_lock) |
f254ac04 | 5560 | { |
fbd15848 | 5561 | struct io_timeout_data *io; |
47f46768 | 5562 | struct io_kiocb *req; |
fd9c7bc5 | 5563 | bool found = false; |
f254ac04 | 5564 | |
135fcde8 | 5565 | list_for_each_entry(req, &ctx->timeout_list, timeout.list) { |
fd9c7bc5 PB |
5566 | found = user_data == req->user_data; |
5567 | if (found) | |
47f46768 | 5568 | break; |
47f46768 | 5569 | } |
fd9c7bc5 PB |
5570 | if (!found) |
5571 | return ERR_PTR(-ENOENT); | |
fbd15848 PB |
5572 | |
5573 | io = req->async_data; | |
fd9c7bc5 | 5574 | if (hrtimer_try_to_cancel(&io->timer) == -1) |
fbd15848 | 5575 | return ERR_PTR(-EALREADY); |
a71976f3 | 5576 | list_del_init(&req->timeout.list); |
fbd15848 PB |
5577 | return req; |
5578 | } | |
47f46768 | 5579 | |
fbd15848 | 5580 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
89850fce | 5581 | __must_hold(&ctx->timeout_lock) |
fbd15848 PB |
5582 | { |
5583 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); | |
5584 | ||
5585 | if (IS_ERR(req)) | |
5586 | return PTR_ERR(req); | |
f254ac04 | 5587 | |
93d2bcd2 | 5588 | req_set_fail(req); |
d4d19c19 | 5589 | io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0); |
91c2f697 | 5590 | io_put_req_deferred(req); |
f254ac04 JA |
5591 | return 0; |
5592 | } | |
5593 | ||
9c8e11b3 PB |
5594 | static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data, |
5595 | struct timespec64 *ts, enum hrtimer_mode mode) | |
89850fce | 5596 | __must_hold(&ctx->timeout_lock) |
47f46768 | 5597 | { |
9c8e11b3 PB |
5598 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); |
5599 | struct io_timeout_data *data; | |
47f46768 | 5600 | |
9c8e11b3 PB |
5601 | if (IS_ERR(req)) |
5602 | return PTR_ERR(req); | |
47f46768 | 5603 | |
9c8e11b3 PB |
5604 | req->timeout.off = 0; /* noseq */ |
5605 | data = req->async_data; | |
5606 | list_add_tail(&req->timeout.list, &ctx->timeout_list); | |
5607 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode); | |
5608 | data->timer.function = io_timeout_fn; | |
5609 | hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode); | |
5610 | return 0; | |
47f46768 JA |
5611 | } |
5612 | ||
3529d8c2 JA |
5613 | static int io_timeout_remove_prep(struct io_kiocb *req, |
5614 | const struct io_uring_sqe *sqe) | |
b29472ee | 5615 | { |
9c8e11b3 PB |
5616 | struct io_timeout_rem *tr = &req->timeout_rem; |
5617 | ||
b29472ee JA |
5618 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5619 | return -EINVAL; | |
61710e43 DA |
5620 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5621 | return -EINVAL; | |
9c8e11b3 | 5622 | if (sqe->ioprio || sqe->buf_index || sqe->len) |
b29472ee JA |
5623 | return -EINVAL; |
5624 | ||
9c8e11b3 PB |
5625 | tr->addr = READ_ONCE(sqe->addr); |
5626 | tr->flags = READ_ONCE(sqe->timeout_flags); | |
5627 | if (tr->flags & IORING_TIMEOUT_UPDATE) { | |
5628 | if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS)) | |
5629 | return -EINVAL; | |
5630 | if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2))) | |
5631 | return -EFAULT; | |
5632 | } else if (tr->flags) { | |
5633 | /* timeout removal doesn't support flags */ | |
b29472ee | 5634 | return -EINVAL; |
9c8e11b3 | 5635 | } |
b29472ee | 5636 | |
b29472ee JA |
5637 | return 0; |
5638 | } | |
5639 | ||
8662daec PB |
5640 | static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags) |
5641 | { | |
5642 | return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS | |
5643 | : HRTIMER_MODE_REL; | |
5644 | } | |
5645 | ||
11365043 JA |
5646 | /* |
5647 | * Remove or update an existing timeout command | |
5648 | */ | |
61e98203 | 5649 | static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags) |
11365043 | 5650 | { |
9c8e11b3 | 5651 | struct io_timeout_rem *tr = &req->timeout_rem; |
11365043 | 5652 | struct io_ring_ctx *ctx = req->ctx; |
47f46768 | 5653 | int ret; |
11365043 | 5654 | |
89850fce | 5655 | spin_lock_irq(&ctx->timeout_lock); |
8662daec | 5656 | if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) |
9c8e11b3 | 5657 | ret = io_timeout_cancel(ctx, tr->addr); |
8662daec PB |
5658 | else |
5659 | ret = io_timeout_update(ctx, tr->addr, &tr->ts, | |
5660 | io_translate_timeout_mode(tr->flags)); | |
89850fce | 5661 | spin_unlock_irq(&ctx->timeout_lock); |
11365043 | 5662 | |
79ebeaee | 5663 | spin_lock(&ctx->completion_lock); |
d4d19c19 | 5664 | io_cqring_fill_event(ctx, req->user_data, ret, 0); |
11365043 | 5665 | io_commit_cqring(ctx); |
79ebeaee | 5666 | spin_unlock(&ctx->completion_lock); |
5262f567 | 5667 | io_cqring_ev_posted(ctx); |
4e88d6e7 | 5668 | if (ret < 0) |
93d2bcd2 | 5669 | req_set_fail(req); |
ec9c02ad | 5670 | io_put_req(req); |
11365043 | 5671 | return 0; |
5262f567 JA |
5672 | } |
5673 | ||
3529d8c2 | 5674 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 5675 | bool is_timeout_link) |
5262f567 | 5676 | { |
ad8a48ac | 5677 | struct io_timeout_data *data; |
a41525ab | 5678 | unsigned flags; |
56080b02 | 5679 | u32 off = READ_ONCE(sqe->off); |
5262f567 | 5680 | |
ad8a48ac | 5681 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 5682 | return -EINVAL; |
ad8a48ac | 5683 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 5684 | return -EINVAL; |
56080b02 | 5685 | if (off && is_timeout_link) |
2d28390a | 5686 | return -EINVAL; |
a41525ab JA |
5687 | flags = READ_ONCE(sqe->timeout_flags); |
5688 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 5689 | return -EINVAL; |
bdf20073 | 5690 | |
bfe68a22 | 5691 | req->timeout.off = off; |
f18ee4cf PB |
5692 | if (unlikely(off && !req->ctx->off_timeout_used)) |
5693 | req->ctx->off_timeout_used = true; | |
26a61679 | 5694 | |
e8c2bc1f | 5695 | if (!req->async_data && io_alloc_async_data(req)) |
26a61679 JA |
5696 | return -ENOMEM; |
5697 | ||
e8c2bc1f | 5698 | data = req->async_data; |
ad8a48ac | 5699 | data->req = req; |
ad8a48ac JA |
5700 | |
5701 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
5702 | return -EFAULT; |
5703 | ||
8662daec | 5704 | data->mode = io_translate_timeout_mode(flags); |
ad8a48ac | 5705 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
b97e736a PB |
5706 | |
5707 | if (is_timeout_link) { | |
5708 | struct io_submit_link *link = &req->ctx->submit_state.link; | |
5709 | ||
5710 | if (!link->head) | |
5711 | return -EINVAL; | |
5712 | if (link->last->opcode == IORING_OP_LINK_TIMEOUT) | |
5713 | return -EINVAL; | |
4d13d1a4 PB |
5714 | req->timeout.head = link->last; |
5715 | link->last->flags |= REQ_F_ARM_LTIMEOUT; | |
b97e736a | 5716 | } |
ad8a48ac JA |
5717 | return 0; |
5718 | } | |
5719 | ||
61e98203 | 5720 | static int io_timeout(struct io_kiocb *req, unsigned int issue_flags) |
ad8a48ac | 5721 | { |
ad8a48ac | 5722 | struct io_ring_ctx *ctx = req->ctx; |
e8c2bc1f | 5723 | struct io_timeout_data *data = req->async_data; |
ad8a48ac | 5724 | struct list_head *entry; |
bfe68a22 | 5725 | u32 tail, off = req->timeout.off; |
ad8a48ac | 5726 | |
89850fce | 5727 | spin_lock_irq(&ctx->timeout_lock); |
93bd25bb | 5728 | |
5262f567 JA |
5729 | /* |
5730 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
5731 | * timeout event to be satisfied. If it isn't set, then this is |
5732 | * a pure timeout request, sequence isn't used. | |
5262f567 | 5733 | */ |
8eb7e2d0 | 5734 | if (io_is_timeout_noseq(req)) { |
93bd25bb JA |
5735 | entry = ctx->timeout_list.prev; |
5736 | goto add; | |
5737 | } | |
5262f567 | 5738 | |
bfe68a22 PB |
5739 | tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
5740 | req->timeout.target_seq = tail + off; | |
5262f567 | 5741 | |
f010505b MDG |
5742 | /* Update the last seq here in case io_flush_timeouts() hasn't. |
5743 | * This is safe because ->completion_lock is held, and submissions | |
5744 | * and completions are never mixed in the same ->completion_lock section. | |
5745 | */ | |
5746 | ctx->cq_last_tm_flush = tail; | |
5747 | ||
5262f567 JA |
5748 | /* |
5749 | * Insertion sort, ensuring the first entry in the list is always | |
5750 | * the one we need first. | |
5751 | */ | |
5262f567 | 5752 | list_for_each_prev(entry, &ctx->timeout_list) { |
135fcde8 PB |
5753 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, |
5754 | timeout.list); | |
5262f567 | 5755 | |
8eb7e2d0 | 5756 | if (io_is_timeout_noseq(nxt)) |
93bd25bb | 5757 | continue; |
bfe68a22 PB |
5758 | /* nxt.seq is behind @tail, otherwise would've been completed */ |
5759 | if (off >= nxt->timeout.target_seq - tail) | |
5262f567 JA |
5760 | break; |
5761 | } | |
93bd25bb | 5762 | add: |
135fcde8 | 5763 | list_add(&req->timeout.list, entry); |
ad8a48ac JA |
5764 | data->timer.function = io_timeout_fn; |
5765 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
89850fce | 5766 | spin_unlock_irq(&ctx->timeout_lock); |
5262f567 JA |
5767 | return 0; |
5768 | } | |
5262f567 | 5769 | |
f458dd84 PB |
5770 | struct io_cancel_data { |
5771 | struct io_ring_ctx *ctx; | |
5772 | u64 user_data; | |
5773 | }; | |
5774 | ||
62755e35 JA |
5775 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
5776 | { | |
5777 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
f458dd84 | 5778 | struct io_cancel_data *cd = data; |
62755e35 | 5779 | |
f458dd84 | 5780 | return req->ctx == cd->ctx && req->user_data == cd->user_data; |
62755e35 JA |
5781 | } |
5782 | ||
f458dd84 PB |
5783 | static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data, |
5784 | struct io_ring_ctx *ctx) | |
62755e35 | 5785 | { |
f458dd84 | 5786 | struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, }; |
62755e35 | 5787 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
5788 | int ret = 0; |
5789 | ||
f458dd84 | 5790 | if (!tctx || !tctx->io_wq) |
5aa75ed5 JA |
5791 | return -ENOENT; |
5792 | ||
f458dd84 | 5793 | cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false); |
62755e35 JA |
5794 | switch (cancel_ret) { |
5795 | case IO_WQ_CANCEL_OK: | |
5796 | ret = 0; | |
5797 | break; | |
5798 | case IO_WQ_CANCEL_RUNNING: | |
5799 | ret = -EALREADY; | |
5800 | break; | |
5801 | case IO_WQ_CANCEL_NOTFOUND: | |
5802 | ret = -ENOENT; | |
5803 | break; | |
5804 | } | |
5805 | ||
e977d6d3 JA |
5806 | return ret; |
5807 | } | |
5808 | ||
8cb01fac PB |
5809 | static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr) |
5810 | __acquires(&req->ctx->completion_lock) | |
47f46768 | 5811 | { |
8cb01fac | 5812 | struct io_ring_ctx *ctx = req->ctx; |
47f46768 JA |
5813 | int ret; |
5814 | ||
8cb01fac PB |
5815 | WARN_ON_ONCE(req->task != current); |
5816 | ||
f458dd84 | 5817 | ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx); |
79ebeaee | 5818 | spin_lock(&ctx->completion_lock); |
df9727af | 5819 | if (ret != -ENOENT) |
8cb01fac | 5820 | return ret; |
79ebeaee | 5821 | spin_lock_irq(&ctx->timeout_lock); |
47f46768 | 5822 | ret = io_timeout_cancel(ctx, sqe_addr); |
79ebeaee | 5823 | spin_unlock_irq(&ctx->timeout_lock); |
47f46768 | 5824 | if (ret != -ENOENT) |
8cb01fac PB |
5825 | return ret; |
5826 | return io_poll_cancel(ctx, sqe_addr, false); | |
47f46768 JA |
5827 | } |
5828 | ||
3529d8c2 JA |
5829 | static int io_async_cancel_prep(struct io_kiocb *req, |
5830 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 5831 | { |
fbf23849 | 5832 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 | 5833 | return -EINVAL; |
61710e43 DA |
5834 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5835 | return -EINVAL; | |
5836 | if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags) | |
e977d6d3 JA |
5837 | return -EINVAL; |
5838 | ||
fbf23849 JA |
5839 | req->cancel.addr = READ_ONCE(sqe->addr); |
5840 | return 0; | |
5841 | } | |
5842 | ||
61e98203 | 5843 | static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags) |
fbf23849 JA |
5844 | { |
5845 | struct io_ring_ctx *ctx = req->ctx; | |
58f99373 PB |
5846 | u64 sqe_addr = req->cancel.addr; |
5847 | struct io_tctx_node *node; | |
5848 | int ret; | |
5849 | ||
8cb01fac | 5850 | ret = io_try_cancel_userdata(req, sqe_addr); |
58f99373 PB |
5851 | if (ret != -ENOENT) |
5852 | goto done; | |
79ebeaee | 5853 | spin_unlock(&ctx->completion_lock); |
58f99373 PB |
5854 | |
5855 | /* slow path, try all io-wq's */ | |
5856 | io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK)); | |
5857 | ret = -ENOENT; | |
5858 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
5859 | struct io_uring_task *tctx = node->task->io_uring; | |
fbf23849 | 5860 | |
58f99373 PB |
5861 | ret = io_async_cancel_one(tctx, req->cancel.addr, ctx); |
5862 | if (ret != -ENOENT) | |
5863 | break; | |
5864 | } | |
5865 | io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK)); | |
5866 | ||
79ebeaee | 5867 | spin_lock(&ctx->completion_lock); |
58f99373 | 5868 | done: |
d4d19c19 | 5869 | io_cqring_fill_event(ctx, req->user_data, ret, 0); |
58f99373 | 5870 | io_commit_cqring(ctx); |
79ebeaee | 5871 | spin_unlock(&ctx->completion_lock); |
58f99373 PB |
5872 | io_cqring_ev_posted(ctx); |
5873 | ||
5874 | if (ret < 0) | |
93d2bcd2 | 5875 | req_set_fail(req); |
58f99373 | 5876 | io_put_req(req); |
5262f567 JA |
5877 | return 0; |
5878 | } | |
5879 | ||
269bbe5f | 5880 | static int io_rsrc_update_prep(struct io_kiocb *req, |
05f3fb3c JA |
5881 | const struct io_uring_sqe *sqe) |
5882 | { | |
61710e43 DA |
5883 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5884 | return -EINVAL; | |
5885 | if (sqe->ioprio || sqe->rw_flags) | |
05f3fb3c JA |
5886 | return -EINVAL; |
5887 | ||
269bbe5f BM |
5888 | req->rsrc_update.offset = READ_ONCE(sqe->off); |
5889 | req->rsrc_update.nr_args = READ_ONCE(sqe->len); | |
5890 | if (!req->rsrc_update.nr_args) | |
05f3fb3c | 5891 | return -EINVAL; |
269bbe5f | 5892 | req->rsrc_update.arg = READ_ONCE(sqe->addr); |
05f3fb3c JA |
5893 | return 0; |
5894 | } | |
5895 | ||
889fca73 | 5896 | static int io_files_update(struct io_kiocb *req, unsigned int issue_flags) |
fbf23849 JA |
5897 | { |
5898 | struct io_ring_ctx *ctx = req->ctx; | |
c3bdad02 | 5899 | struct io_uring_rsrc_update2 up; |
05f3fb3c | 5900 | int ret; |
fbf23849 | 5901 | |
45d189c6 | 5902 | if (issue_flags & IO_URING_F_NONBLOCK) |
05f3fb3c | 5903 | return -EAGAIN; |
05f3fb3c | 5904 | |
269bbe5f BM |
5905 | up.offset = req->rsrc_update.offset; |
5906 | up.data = req->rsrc_update.arg; | |
c3bdad02 PB |
5907 | up.nr = 0; |
5908 | up.tags = 0; | |
615cee49 | 5909 | up.resv = 0; |
05f3fb3c JA |
5910 | |
5911 | mutex_lock(&ctx->uring_lock); | |
fdecb662 | 5912 | ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, |
98f0b3b4 | 5913 | &up, req->rsrc_update.nr_args); |
05f3fb3c JA |
5914 | mutex_unlock(&ctx->uring_lock); |
5915 | ||
5916 | if (ret < 0) | |
93d2bcd2 | 5917 | req_set_fail(req); |
889fca73 | 5918 | __io_req_complete(req, issue_flags, ret, 0); |
5262f567 JA |
5919 | return 0; |
5920 | } | |
5921 | ||
bfe76559 | 5922 | static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 5923 | { |
d625c6ee | 5924 | switch (req->opcode) { |
e781573e | 5925 | case IORING_OP_NOP: |
bfe76559 | 5926 | return 0; |
f67676d1 JA |
5927 | case IORING_OP_READV: |
5928 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 5929 | case IORING_OP_READ: |
bfe76559 | 5930 | return io_read_prep(req, sqe); |
f67676d1 JA |
5931 | case IORING_OP_WRITEV: |
5932 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 5933 | case IORING_OP_WRITE: |
bfe76559 | 5934 | return io_write_prep(req, sqe); |
0969e783 | 5935 | case IORING_OP_POLL_ADD: |
bfe76559 | 5936 | return io_poll_add_prep(req, sqe); |
0969e783 | 5937 | case IORING_OP_POLL_REMOVE: |
c5de0036 | 5938 | return io_poll_update_prep(req, sqe); |
8ed8d3c3 | 5939 | case IORING_OP_FSYNC: |
1155c76a | 5940 | return io_fsync_prep(req, sqe); |
8ed8d3c3 | 5941 | case IORING_OP_SYNC_FILE_RANGE: |
1155c76a | 5942 | return io_sfr_prep(req, sqe); |
03b1230c | 5943 | case IORING_OP_SENDMSG: |
fddaface | 5944 | case IORING_OP_SEND: |
bfe76559 | 5945 | return io_sendmsg_prep(req, sqe); |
03b1230c | 5946 | case IORING_OP_RECVMSG: |
fddaface | 5947 | case IORING_OP_RECV: |
bfe76559 | 5948 | return io_recvmsg_prep(req, sqe); |
f499a021 | 5949 | case IORING_OP_CONNECT: |
bfe76559 | 5950 | return io_connect_prep(req, sqe); |
2d28390a | 5951 | case IORING_OP_TIMEOUT: |
bfe76559 | 5952 | return io_timeout_prep(req, sqe, false); |
b29472ee | 5953 | case IORING_OP_TIMEOUT_REMOVE: |
bfe76559 | 5954 | return io_timeout_remove_prep(req, sqe); |
fbf23849 | 5955 | case IORING_OP_ASYNC_CANCEL: |
bfe76559 | 5956 | return io_async_cancel_prep(req, sqe); |
2d28390a | 5957 | case IORING_OP_LINK_TIMEOUT: |
bfe76559 | 5958 | return io_timeout_prep(req, sqe, true); |
8ed8d3c3 | 5959 | case IORING_OP_ACCEPT: |
bfe76559 | 5960 | return io_accept_prep(req, sqe); |
d63d1b5e | 5961 | case IORING_OP_FALLOCATE: |
bfe76559 | 5962 | return io_fallocate_prep(req, sqe); |
15b71abe | 5963 | case IORING_OP_OPENAT: |
bfe76559 | 5964 | return io_openat_prep(req, sqe); |
b5dba59e | 5965 | case IORING_OP_CLOSE: |
bfe76559 | 5966 | return io_close_prep(req, sqe); |
05f3fb3c | 5967 | case IORING_OP_FILES_UPDATE: |
269bbe5f | 5968 | return io_rsrc_update_prep(req, sqe); |
eddc7ef5 | 5969 | case IORING_OP_STATX: |
bfe76559 | 5970 | return io_statx_prep(req, sqe); |
4840e418 | 5971 | case IORING_OP_FADVISE: |
bfe76559 | 5972 | return io_fadvise_prep(req, sqe); |
c1ca757b | 5973 | case IORING_OP_MADVISE: |
bfe76559 | 5974 | return io_madvise_prep(req, sqe); |
cebdb986 | 5975 | case IORING_OP_OPENAT2: |
bfe76559 | 5976 | return io_openat2_prep(req, sqe); |
3e4827b0 | 5977 | case IORING_OP_EPOLL_CTL: |
bfe76559 | 5978 | return io_epoll_ctl_prep(req, sqe); |
7d67af2c | 5979 | case IORING_OP_SPLICE: |
bfe76559 | 5980 | return io_splice_prep(req, sqe); |
ddf0322d | 5981 | case IORING_OP_PROVIDE_BUFFERS: |
bfe76559 | 5982 | return io_provide_buffers_prep(req, sqe); |
067524e9 | 5983 | case IORING_OP_REMOVE_BUFFERS: |
bfe76559 | 5984 | return io_remove_buffers_prep(req, sqe); |
f2a8d5c7 | 5985 | case IORING_OP_TEE: |
bfe76559 | 5986 | return io_tee_prep(req, sqe); |
36f4fa68 JA |
5987 | case IORING_OP_SHUTDOWN: |
5988 | return io_shutdown_prep(req, sqe); | |
80a261fd JA |
5989 | case IORING_OP_RENAMEAT: |
5990 | return io_renameat_prep(req, sqe); | |
14a1143b JA |
5991 | case IORING_OP_UNLINKAT: |
5992 | return io_unlinkat_prep(req, sqe); | |
f67676d1 JA |
5993 | } |
5994 | ||
bfe76559 PB |
5995 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
5996 | req->opcode); | |
bd54b6fe | 5997 | return -EINVAL; |
bfe76559 PB |
5998 | } |
5999 | ||
93642ef8 | 6000 | static int io_req_prep_async(struct io_kiocb *req) |
bfe76559 | 6001 | { |
b7e298d2 PB |
6002 | if (!io_op_defs[req->opcode].needs_async_setup) |
6003 | return 0; | |
6004 | if (WARN_ON_ONCE(req->async_data)) | |
6005 | return -EFAULT; | |
6006 | if (io_alloc_async_data(req)) | |
6007 | return -EAGAIN; | |
6008 | ||
93642ef8 PB |
6009 | switch (req->opcode) { |
6010 | case IORING_OP_READV: | |
93642ef8 PB |
6011 | return io_rw_prep_async(req, READ); |
6012 | case IORING_OP_WRITEV: | |
93642ef8 PB |
6013 | return io_rw_prep_async(req, WRITE); |
6014 | case IORING_OP_SENDMSG: | |
93642ef8 PB |
6015 | return io_sendmsg_prep_async(req); |
6016 | case IORING_OP_RECVMSG: | |
93642ef8 PB |
6017 | return io_recvmsg_prep_async(req); |
6018 | case IORING_OP_CONNECT: | |
6019 | return io_connect_prep_async(req); | |
6020 | } | |
b7e298d2 PB |
6021 | printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n", |
6022 | req->opcode); | |
6023 | return -EFAULT; | |
f67676d1 JA |
6024 | } |
6025 | ||
9cf7c104 PB |
6026 | static u32 io_get_sequence(struct io_kiocb *req) |
6027 | { | |
a3dbdf54 | 6028 | u32 seq = req->ctx->cached_sq_head; |
9cf7c104 | 6029 | |
a3dbdf54 PB |
6030 | /* need original cached_sq_head, but it was increased for each req */ |
6031 | io_for_each_link(req, req) | |
6032 | seq--; | |
6033 | return seq; | |
9cf7c104 PB |
6034 | } |
6035 | ||
76cc33d7 | 6036 | static bool io_drain_req(struct io_kiocb *req) |
de0617e4 | 6037 | { |
3c19966d | 6038 | struct io_kiocb *pos; |
a197f664 | 6039 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 6040 | struct io_defer_entry *de; |
f67676d1 | 6041 | int ret; |
9cf7c104 | 6042 | u32 seq; |
de0617e4 | 6043 | |
3c19966d PB |
6044 | /* |
6045 | * If we need to drain a request in the middle of a link, drain the | |
6046 | * head request and the next request/link after the current link. | |
6047 | * Considering sequential execution of links, IOSQE_IO_DRAIN will be | |
6048 | * maintained for every request of our link. | |
6049 | */ | |
6050 | if (ctx->drain_next) { | |
6051 | req->flags |= REQ_F_IO_DRAIN; | |
6052 | ctx->drain_next = false; | |
6053 | } | |
6054 | /* not interested in head, start from the first linked */ | |
6055 | io_for_each_link(pos, req->link) { | |
6056 | if (pos->flags & REQ_F_IO_DRAIN) { | |
6057 | ctx->drain_next = true; | |
6058 | req->flags |= REQ_F_IO_DRAIN; | |
6059 | break; | |
6060 | } | |
6061 | } | |
6062 | ||
9d858b21 | 6063 | /* Still need defer if there is pending req in defer list. */ |
9cf7c104 | 6064 | if (likely(list_empty_careful(&ctx->defer_list) && |
10c66904 PB |
6065 | !(req->flags & REQ_F_IO_DRAIN))) { |
6066 | ctx->drain_active = false; | |
76cc33d7 | 6067 | return false; |
10c66904 | 6068 | } |
9cf7c104 PB |
6069 | |
6070 | seq = io_get_sequence(req); | |
6071 | /* Still a chance to pass the sequence check */ | |
6072 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) | |
76cc33d7 | 6073 | return false; |
de0617e4 | 6074 | |
b7e298d2 | 6075 | ret = io_req_prep_async(req); |
be7053b7 | 6076 | if (ret) |
1b48773f | 6077 | goto fail; |
cbdcb435 | 6078 | io_prep_async_link(req); |
27dc8338 | 6079 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 6080 | if (!de) { |
1b48773f PB |
6081 | ret = -ENOMEM; |
6082 | fail: | |
6083 | io_req_complete_failed(req, ret); | |
76cc33d7 PB |
6084 | return true; |
6085 | } | |
2d28390a | 6086 | |
79ebeaee | 6087 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 6088 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 6089 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 6090 | kfree(de); |
ae34817b | 6091 | io_queue_async_work(req); |
76cc33d7 | 6092 | return true; |
de0617e4 JA |
6093 | } |
6094 | ||
915967f6 | 6095 | trace_io_uring_defer(ctx, req, req->user_data); |
27dc8338 | 6096 | de->req = req; |
9cf7c104 | 6097 | de->seq = seq; |
27dc8338 | 6098 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 6099 | spin_unlock(&ctx->completion_lock); |
76cc33d7 | 6100 | return true; |
de0617e4 JA |
6101 | } |
6102 | ||
68fb8979 | 6103 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 6104 | { |
0e1b6fe3 PB |
6105 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
6106 | switch (req->opcode) { | |
6107 | case IORING_OP_READV: | |
6108 | case IORING_OP_READ_FIXED: | |
6109 | case IORING_OP_READ: | |
bcda7baa | 6110 | kfree((void *)(unsigned long)req->rw.addr); |
0e1b6fe3 PB |
6111 | break; |
6112 | case IORING_OP_RECVMSG: | |
6113 | case IORING_OP_RECV: | |
bcda7baa | 6114 | kfree(req->sr_msg.kbuf); |
0e1b6fe3 PB |
6115 | break; |
6116 | } | |
99bc4c38 PB |
6117 | } |
6118 | ||
0e1b6fe3 PB |
6119 | if (req->flags & REQ_F_NEED_CLEANUP) { |
6120 | switch (req->opcode) { | |
6121 | case IORING_OP_READV: | |
6122 | case IORING_OP_READ_FIXED: | |
6123 | case IORING_OP_READ: | |
6124 | case IORING_OP_WRITEV: | |
6125 | case IORING_OP_WRITE_FIXED: | |
e8c2bc1f JA |
6126 | case IORING_OP_WRITE: { |
6127 | struct io_async_rw *io = req->async_data; | |
1dacb4df PB |
6128 | |
6129 | kfree(io->free_iovec); | |
0e1b6fe3 | 6130 | break; |
e8c2bc1f | 6131 | } |
0e1b6fe3 | 6132 | case IORING_OP_RECVMSG: |
e8c2bc1f JA |
6133 | case IORING_OP_SENDMSG: { |
6134 | struct io_async_msghdr *io = req->async_data; | |
257e84a5 PB |
6135 | |
6136 | kfree(io->free_iov); | |
0e1b6fe3 | 6137 | break; |
e8c2bc1f | 6138 | } |
0e1b6fe3 PB |
6139 | case IORING_OP_SPLICE: |
6140 | case IORING_OP_TEE: | |
e1d767f0 PB |
6141 | if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED)) |
6142 | io_put_file(req->splice.file_in); | |
0e1b6fe3 | 6143 | break; |
f3cd4850 JA |
6144 | case IORING_OP_OPENAT: |
6145 | case IORING_OP_OPENAT2: | |
6146 | if (req->open.filename) | |
6147 | putname(req->open.filename); | |
6148 | break; | |
80a261fd JA |
6149 | case IORING_OP_RENAMEAT: |
6150 | putname(req->rename.oldpath); | |
6151 | putname(req->rename.newpath); | |
6152 | break; | |
14a1143b JA |
6153 | case IORING_OP_UNLINKAT: |
6154 | putname(req->unlink.filename); | |
6155 | break; | |
0e1b6fe3 | 6156 | } |
99bc4c38 | 6157 | } |
75652a30 JA |
6158 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
6159 | kfree(req->apoll->double_poll); | |
6160 | kfree(req->apoll); | |
6161 | req->apoll = NULL; | |
6162 | } | |
3a0a6902 PB |
6163 | if (req->flags & REQ_F_INFLIGHT) { |
6164 | struct io_uring_task *tctx = req->task->io_uring; | |
6165 | ||
6166 | atomic_dec(&tctx->inflight_tracked); | |
3a0a6902 | 6167 | } |
c854357b | 6168 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 6169 | put_cred(req->creds); |
c854357b PB |
6170 | |
6171 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
99bc4c38 PB |
6172 | } |
6173 | ||
889fca73 | 6174 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 6175 | { |
a197f664 | 6176 | struct io_ring_ctx *ctx = req->ctx; |
5730b27e | 6177 | const struct cred *creds = NULL; |
d625c6ee | 6178 | int ret; |
2b188cc1 | 6179 | |
b8e64b53 | 6180 | if ((req->flags & REQ_F_CREDS) && req->creds != current_cred()) |
c10d1f98 | 6181 | creds = override_creds(req->creds); |
5730b27e | 6182 | |
d625c6ee | 6183 | switch (req->opcode) { |
2b188cc1 | 6184 | case IORING_OP_NOP: |
889fca73 | 6185 | ret = io_nop(req, issue_flags); |
2b188cc1 JA |
6186 | break; |
6187 | case IORING_OP_READV: | |
edafccee | 6188 | case IORING_OP_READ_FIXED: |
3a6820f2 | 6189 | case IORING_OP_READ: |
889fca73 | 6190 | ret = io_read(req, issue_flags); |
edafccee | 6191 | break; |
3529d8c2 | 6192 | case IORING_OP_WRITEV: |
edafccee | 6193 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 6194 | case IORING_OP_WRITE: |
889fca73 | 6195 | ret = io_write(req, issue_flags); |
2b188cc1 | 6196 | break; |
c992fe29 | 6197 | case IORING_OP_FSYNC: |
45d189c6 | 6198 | ret = io_fsync(req, issue_flags); |
c992fe29 | 6199 | break; |
221c5eb2 | 6200 | case IORING_OP_POLL_ADD: |
61e98203 | 6201 | ret = io_poll_add(req, issue_flags); |
221c5eb2 JA |
6202 | break; |
6203 | case IORING_OP_POLL_REMOVE: | |
c5de0036 | 6204 | ret = io_poll_update(req, issue_flags); |
221c5eb2 | 6205 | break; |
5d17b4a4 | 6206 | case IORING_OP_SYNC_FILE_RANGE: |
45d189c6 | 6207 | ret = io_sync_file_range(req, issue_flags); |
5d17b4a4 | 6208 | break; |
0fa03c62 | 6209 | case IORING_OP_SENDMSG: |
889fca73 | 6210 | ret = io_sendmsg(req, issue_flags); |
062d04d7 | 6211 | break; |
fddaface | 6212 | case IORING_OP_SEND: |
889fca73 | 6213 | ret = io_send(req, issue_flags); |
0fa03c62 | 6214 | break; |
aa1fa28f | 6215 | case IORING_OP_RECVMSG: |
889fca73 | 6216 | ret = io_recvmsg(req, issue_flags); |
062d04d7 | 6217 | break; |
fddaface | 6218 | case IORING_OP_RECV: |
889fca73 | 6219 | ret = io_recv(req, issue_flags); |
aa1fa28f | 6220 | break; |
5262f567 | 6221 | case IORING_OP_TIMEOUT: |
61e98203 | 6222 | ret = io_timeout(req, issue_flags); |
5262f567 | 6223 | break; |
11365043 | 6224 | case IORING_OP_TIMEOUT_REMOVE: |
61e98203 | 6225 | ret = io_timeout_remove(req, issue_flags); |
11365043 | 6226 | break; |
17f2fe35 | 6227 | case IORING_OP_ACCEPT: |
889fca73 | 6228 | ret = io_accept(req, issue_flags); |
17f2fe35 | 6229 | break; |
f8e85cf2 | 6230 | case IORING_OP_CONNECT: |
889fca73 | 6231 | ret = io_connect(req, issue_flags); |
f8e85cf2 | 6232 | break; |
62755e35 | 6233 | case IORING_OP_ASYNC_CANCEL: |
61e98203 | 6234 | ret = io_async_cancel(req, issue_flags); |
62755e35 | 6235 | break; |
d63d1b5e | 6236 | case IORING_OP_FALLOCATE: |
45d189c6 | 6237 | ret = io_fallocate(req, issue_flags); |
d63d1b5e | 6238 | break; |
15b71abe | 6239 | case IORING_OP_OPENAT: |
45d189c6 | 6240 | ret = io_openat(req, issue_flags); |
15b71abe | 6241 | break; |
b5dba59e | 6242 | case IORING_OP_CLOSE: |
889fca73 | 6243 | ret = io_close(req, issue_flags); |
b5dba59e | 6244 | break; |
05f3fb3c | 6245 | case IORING_OP_FILES_UPDATE: |
889fca73 | 6246 | ret = io_files_update(req, issue_flags); |
05f3fb3c | 6247 | break; |
eddc7ef5 | 6248 | case IORING_OP_STATX: |
45d189c6 | 6249 | ret = io_statx(req, issue_flags); |
eddc7ef5 | 6250 | break; |
4840e418 | 6251 | case IORING_OP_FADVISE: |
45d189c6 | 6252 | ret = io_fadvise(req, issue_flags); |
4840e418 | 6253 | break; |
c1ca757b | 6254 | case IORING_OP_MADVISE: |
45d189c6 | 6255 | ret = io_madvise(req, issue_flags); |
c1ca757b | 6256 | break; |
cebdb986 | 6257 | case IORING_OP_OPENAT2: |
45d189c6 | 6258 | ret = io_openat2(req, issue_flags); |
cebdb986 | 6259 | break; |
3e4827b0 | 6260 | case IORING_OP_EPOLL_CTL: |
889fca73 | 6261 | ret = io_epoll_ctl(req, issue_flags); |
3e4827b0 | 6262 | break; |
7d67af2c | 6263 | case IORING_OP_SPLICE: |
45d189c6 | 6264 | ret = io_splice(req, issue_flags); |
7d67af2c | 6265 | break; |
ddf0322d | 6266 | case IORING_OP_PROVIDE_BUFFERS: |
889fca73 | 6267 | ret = io_provide_buffers(req, issue_flags); |
ddf0322d | 6268 | break; |
067524e9 | 6269 | case IORING_OP_REMOVE_BUFFERS: |
889fca73 | 6270 | ret = io_remove_buffers(req, issue_flags); |
3e4827b0 | 6271 | break; |
f2a8d5c7 | 6272 | case IORING_OP_TEE: |
45d189c6 | 6273 | ret = io_tee(req, issue_flags); |
f2a8d5c7 | 6274 | break; |
36f4fa68 | 6275 | case IORING_OP_SHUTDOWN: |
45d189c6 | 6276 | ret = io_shutdown(req, issue_flags); |
36f4fa68 | 6277 | break; |
80a261fd | 6278 | case IORING_OP_RENAMEAT: |
45d189c6 | 6279 | ret = io_renameat(req, issue_flags); |
80a261fd | 6280 | break; |
14a1143b | 6281 | case IORING_OP_UNLINKAT: |
45d189c6 | 6282 | ret = io_unlinkat(req, issue_flags); |
14a1143b | 6283 | break; |
2b188cc1 JA |
6284 | default: |
6285 | ret = -EINVAL; | |
6286 | break; | |
6287 | } | |
6288 | ||
5730b27e JA |
6289 | if (creds) |
6290 | revert_creds(creds); | |
def596e9 JA |
6291 | if (ret) |
6292 | return ret; | |
b532576e | 6293 | /* If the op doesn't have a file, we're not polling for it */ |
cb3d8972 PB |
6294 | if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) |
6295 | io_iopoll_req_issued(req); | |
def596e9 JA |
6296 | |
6297 | return 0; | |
2b188cc1 JA |
6298 | } |
6299 | ||
ebc11b6c PB |
6300 | static struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
6301 | { | |
6302 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6303 | ||
6304 | req = io_put_req_find_next(req); | |
6305 | return req ? &req->work : NULL; | |
6306 | } | |
6307 | ||
5280f7e5 | 6308 | static void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
6309 | { |
6310 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6df1db6b | 6311 | struct io_kiocb *timeout; |
561fb04a | 6312 | int ret = 0; |
2b188cc1 | 6313 | |
48dcd38d PB |
6314 | /* one will be dropped by ->io_free_work() after returning to io-wq */ |
6315 | if (!(req->flags & REQ_F_REFCOUNT)) | |
6316 | __io_req_set_refcount(req, 2); | |
6317 | else | |
6318 | req_ref_get(req); | |
5d5901a3 | 6319 | |
6df1db6b PB |
6320 | timeout = io_prep_linked_timeout(req); |
6321 | if (timeout) | |
6322 | io_queue_linked_timeout(timeout); | |
d4c81f38 | 6323 | |
4014d943 | 6324 | if (work->flags & IO_WQ_WORK_CANCEL) |
561fb04a | 6325 | ret = -ECANCELED; |
31b51510 | 6326 | |
561fb04a | 6327 | if (!ret) { |
561fb04a | 6328 | do { |
889fca73 | 6329 | ret = io_issue_sqe(req, 0); |
561fb04a JA |
6330 | /* |
6331 | * We can get EAGAIN for polled IO even though we're | |
6332 | * forcing a sync submission from here, since we can't | |
6333 | * wait for request slots on the block side. | |
6334 | */ | |
6335 | if (ret != -EAGAIN) | |
6336 | break; | |
6337 | cond_resched(); | |
6338 | } while (1); | |
6339 | } | |
31b51510 | 6340 | |
a3df7698 | 6341 | /* avoid locking problems by failing it from a clean context */ |
5d5901a3 | 6342 | if (ret) |
a3df7698 | 6343 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
6344 | } |
6345 | ||
aeca241b | 6346 | static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table, |
042b0d85 | 6347 | unsigned i) |
65e19f54 | 6348 | { |
042b0d85 | 6349 | return &table->files[i]; |
dafecf19 PB |
6350 | } |
6351 | ||
65e19f54 JA |
6352 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
6353 | int index) | |
6354 | { | |
aeca241b | 6355 | struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index); |
65e19f54 | 6356 | |
a04b0ac0 | 6357 | return (struct file *) (slot->file_ptr & FFS_MASK); |
65e19f54 JA |
6358 | } |
6359 | ||
a04b0ac0 | 6360 | static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file) |
9a321c98 PB |
6361 | { |
6362 | unsigned long file_ptr = (unsigned long) file; | |
6363 | ||
b191e2df | 6364 | if (__io_file_supports_nowait(file, READ)) |
9a321c98 | 6365 | file_ptr |= FFS_ASYNC_READ; |
b191e2df | 6366 | if (__io_file_supports_nowait(file, WRITE)) |
9a321c98 PB |
6367 | file_ptr |= FFS_ASYNC_WRITE; |
6368 | if (S_ISREG(file_inode(file)->i_mode)) | |
6369 | file_ptr |= FFS_ISREG; | |
a04b0ac0 | 6370 | file_slot->file_ptr = file_ptr; |
65e19f54 JA |
6371 | } |
6372 | ||
ac177053 PB |
6373 | static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx, |
6374 | struct io_kiocb *req, int fd) | |
09bb8394 | 6375 | { |
8da11c19 | 6376 | struct file *file; |
ac177053 | 6377 | unsigned long file_ptr; |
09bb8394 | 6378 | |
ac177053 PB |
6379 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
6380 | return NULL; | |
6381 | fd = array_index_nospec(fd, ctx->nr_user_files); | |
6382 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
6383 | file = (struct file *) (file_ptr & FFS_MASK); | |
6384 | file_ptr &= ~FFS_MASK; | |
6385 | /* mask in overlapping REQ_F and FFS bits */ | |
b191e2df | 6386 | req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT); |
ac177053 PB |
6387 | io_req_set_rsrc_node(req); |
6388 | return file; | |
6389 | } | |
d44f554e | 6390 | |
ac177053 | 6391 | static struct file *io_file_get_normal(struct io_ring_ctx *ctx, |
ac177053 PB |
6392 | struct io_kiocb *req, int fd) |
6393 | { | |
62906e89 | 6394 | struct file *file = fget(fd); |
ac177053 PB |
6395 | |
6396 | trace_io_uring_file_get(ctx, fd); | |
09bb8394 | 6397 | |
ac177053 PB |
6398 | /* we don't allow fixed io_uring files */ |
6399 | if (file && unlikely(file->f_op == &io_uring_fops)) | |
6400 | io_req_track_inflight(req); | |
8371adf5 | 6401 | return file; |
09bb8394 JA |
6402 | } |
6403 | ||
ac177053 | 6404 | static inline struct file *io_file_get(struct io_ring_ctx *ctx, |
ac177053 PB |
6405 | struct io_kiocb *req, int fd, bool fixed) |
6406 | { | |
6407 | if (fixed) | |
6408 | return io_file_get_fixed(ctx, req, fd); | |
6409 | else | |
62906e89 | 6410 | return io_file_get_normal(ctx, req, fd); |
ac177053 PB |
6411 | } |
6412 | ||
89b263f6 JA |
6413 | static void io_req_task_link_timeout(struct io_kiocb *req) |
6414 | { | |
6415 | struct io_kiocb *prev = req->timeout.prev; | |
6416 | struct io_ring_ctx *ctx = req->ctx; | |
8cb01fac | 6417 | int ret; |
89b263f6 JA |
6418 | |
6419 | if (prev) { | |
8cb01fac PB |
6420 | ret = io_try_cancel_userdata(req, prev->user_data); |
6421 | if (!ret) | |
6422 | ret = -ETIME; | |
6423 | io_cqring_fill_event(ctx, req->user_data, ret, 0); | |
6424 | io_commit_cqring(ctx); | |
6425 | spin_unlock(&ctx->completion_lock); | |
6426 | io_cqring_ev_posted(ctx); | |
6427 | ||
89b263f6 JA |
6428 | io_put_req(prev); |
6429 | io_put_req(req); | |
6430 | } else { | |
6431 | io_req_complete_post(req, -ETIME, 0); | |
6432 | } | |
6433 | } | |
6434 | ||
2665abfd | 6435 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 6436 | { |
ad8a48ac JA |
6437 | struct io_timeout_data *data = container_of(timer, |
6438 | struct io_timeout_data, timer); | |
90cd7e42 | 6439 | struct io_kiocb *prev, *req = data->req; |
2665abfd | 6440 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 6441 | unsigned long flags; |
2665abfd | 6442 | |
89b263f6 | 6443 | spin_lock_irqsave(&ctx->timeout_lock, flags); |
90cd7e42 PB |
6444 | prev = req->timeout.head; |
6445 | req->timeout.head = NULL; | |
2665abfd JA |
6446 | |
6447 | /* | |
6448 | * We don't expect the list to be empty, that will only happen if we | |
6449 | * race with the completion of the linked work. | |
6450 | */ | |
447c19f3 | 6451 | if (prev) { |
f2f87370 | 6452 | io_remove_next_linked(prev); |
447c19f3 PB |
6453 | if (!req_ref_inc_not_zero(prev)) |
6454 | prev = NULL; | |
6455 | } | |
89b263f6 JA |
6456 | req->timeout.prev = prev; |
6457 | spin_unlock_irqrestore(&ctx->timeout_lock, flags); | |
2665abfd | 6458 | |
89b263f6 JA |
6459 | req->io_task_work.func = io_req_task_link_timeout; |
6460 | io_req_task_work_add(req); | |
2665abfd JA |
6461 | return HRTIMER_NORESTART; |
6462 | } | |
6463 | ||
de968c18 | 6464 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 6465 | { |
de968c18 PB |
6466 | struct io_ring_ctx *ctx = req->ctx; |
6467 | ||
89b263f6 | 6468 | spin_lock_irq(&ctx->timeout_lock); |
76a46e06 | 6469 | /* |
f2f87370 PB |
6470 | * If the back reference is NULL, then our linked request finished |
6471 | * before we got a chance to setup the timer | |
76a46e06 | 6472 | */ |
90cd7e42 | 6473 | if (req->timeout.head) { |
e8c2bc1f | 6474 | struct io_timeout_data *data = req->async_data; |
94ae5e77 | 6475 | |
ad8a48ac JA |
6476 | data->timer.function = io_link_timeout_fn; |
6477 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
6478 | data->mode); | |
2665abfd | 6479 | } |
89b263f6 | 6480 | spin_unlock_irq(&ctx->timeout_lock); |
2665abfd | 6481 | /* drop submission reference */ |
76a46e06 JA |
6482 | io_put_req(req); |
6483 | } | |
2665abfd | 6484 | |
c5eef2b9 | 6485 | static void __io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 6486 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 6487 | { |
d3d7298d | 6488 | struct io_kiocb *linked_timeout = io_prep_linked_timeout(req); |
e0c5c576 | 6489 | int ret; |
2b188cc1 | 6490 | |
59b735ae | 6491 | issue_sqe: |
c5eef2b9 | 6492 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 6493 | |
491381ce JA |
6494 | /* |
6495 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
6496 | * doesn't support non-blocking read/write attempts | |
6497 | */ | |
1840038e | 6498 | if (likely(!ret)) { |
e342c807 | 6499 | if (req->flags & REQ_F_COMPLETE_INLINE) { |
c5eef2b9 | 6500 | struct io_ring_ctx *ctx = req->ctx; |
cd0ca2e0 | 6501 | struct io_submit_state *state = &ctx->submit_state; |
e65ef56d | 6502 | |
cd0ca2e0 PB |
6503 | state->compl_reqs[state->compl_nr++] = req; |
6504 | if (state->compl_nr == ARRAY_SIZE(state->compl_reqs)) | |
2a2758f2 | 6505 | io_submit_flush_completions(ctx); |
0d63c148 | 6506 | } |
1840038e | 6507 | } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
59b735ae OL |
6508 | switch (io_arm_poll_handler(req)) { |
6509 | case IO_APOLL_READY: | |
6510 | goto issue_sqe; | |
6511 | case IO_APOLL_ABORTED: | |
1840038e PB |
6512 | /* |
6513 | * Queued up for async execution, worker will release | |
6514 | * submit reference when the iocb is actually submitted. | |
6515 | */ | |
6516 | io_queue_async_work(req); | |
59b735ae | 6517 | break; |
1840038e | 6518 | } |
0d63c148 | 6519 | } else { |
f41db273 | 6520 | io_req_complete_failed(req, ret); |
9e645e11 | 6521 | } |
d3d7298d PB |
6522 | if (linked_timeout) |
6523 | io_queue_linked_timeout(linked_timeout); | |
2b188cc1 JA |
6524 | } |
6525 | ||
441b8a78 | 6526 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 6527 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 6528 | { |
10c66904 | 6529 | if (unlikely(req->ctx->drain_active) && io_drain_req(req)) |
76cc33d7 | 6530 | return; |
4fe2c963 | 6531 | |
76cc33d7 | 6532 | if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) { |
c5eef2b9 | 6533 | __io_queue_sqe(req); |
76cc33d7 PB |
6534 | } else { |
6535 | int ret = io_req_prep_async(req); | |
6536 | ||
6537 | if (unlikely(ret)) | |
6538 | io_req_complete_failed(req, ret); | |
6539 | else | |
6540 | io_queue_async_work(req); | |
ce35a47a | 6541 | } |
4fe2c963 JL |
6542 | } |
6543 | ||
b16fed66 PB |
6544 | /* |
6545 | * Check SQE restrictions (opcode and flags). | |
6546 | * | |
6547 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
6548 | */ | |
6549 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
6550 | struct io_kiocb *req, | |
6551 | unsigned int sqe_flags) | |
4fe2c963 | 6552 | { |
4cfb25bf | 6553 | if (likely(!ctx->restricted)) |
b16fed66 PB |
6554 | return true; |
6555 | ||
6556 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) | |
6557 | return false; | |
6558 | ||
6559 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
6560 | ctx->restrictions.sqe_flags_required) | |
6561 | return false; | |
6562 | ||
6563 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
6564 | ctx->restrictions.sqe_flags_required)) | |
6565 | return false; | |
6566 | ||
6567 | return true; | |
4fe2c963 JL |
6568 | } |
6569 | ||
b16fed66 PB |
6570 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
6571 | const struct io_uring_sqe *sqe) | |
282cdc86 | 6572 | __must_hold(&ctx->uring_lock) |
b16fed66 PB |
6573 | { |
6574 | struct io_submit_state *state; | |
6575 | unsigned int sqe_flags; | |
003e8dcc | 6576 | int personality, ret = 0; |
b16fed66 | 6577 | |
864ea921 | 6578 | /* req is partially pre-initialised, see io_preinit_req() */ |
b16fed66 PB |
6579 | req->opcode = READ_ONCE(sqe->opcode); |
6580 | /* same numerical values with corresponding REQ_F_*, safe to copy */ | |
6581 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
6582 | req->user_data = READ_ONCE(sqe->user_data); | |
b16fed66 | 6583 | req->file = NULL; |
b16fed66 | 6584 | req->fixed_rsrc_refs = NULL; |
b16fed66 | 6585 | req->task = current; |
b16fed66 PB |
6586 | |
6587 | /* enforce forwards compatibility on users */ | |
dddca226 | 6588 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) |
b16fed66 | 6589 | return -EINVAL; |
b16fed66 PB |
6590 | if (unlikely(req->opcode >= IORING_OP_LAST)) |
6591 | return -EINVAL; | |
4cfb25bf | 6592 | if (!io_check_restriction(ctx, req, sqe_flags)) |
b16fed66 PB |
6593 | return -EACCES; |
6594 | ||
6595 | if ((sqe_flags & IOSQE_BUFFER_SELECT) && | |
6596 | !io_op_defs[req->opcode].buffer_select) | |
6597 | return -EOPNOTSUPP; | |
3c19966d PB |
6598 | if (unlikely(sqe_flags & IOSQE_IO_DRAIN)) |
6599 | ctx->drain_active = true; | |
863e0560 | 6600 | |
003e8dcc JA |
6601 | personality = READ_ONCE(sqe->personality); |
6602 | if (personality) { | |
c10d1f98 PB |
6603 | req->creds = xa_load(&ctx->personalities, personality); |
6604 | if (!req->creds) | |
003e8dcc | 6605 | return -EINVAL; |
c10d1f98 | 6606 | get_cred(req->creds); |
b8e64b53 | 6607 | req->flags |= REQ_F_CREDS; |
003e8dcc | 6608 | } |
b16fed66 PB |
6609 | state = &ctx->submit_state; |
6610 | ||
6611 | /* | |
6612 | * Plug now if we have more than 1 IO left after this, and the target | |
6613 | * is potentially a read/write to block based storage. | |
6614 | */ | |
6615 | if (!state->plug_started && state->ios_left > 1 && | |
6616 | io_op_defs[req->opcode].plug) { | |
6617 | blk_start_plug(&state->plug); | |
6618 | state->plug_started = true; | |
6619 | } | |
6620 | ||
6621 | if (io_op_defs[req->opcode].needs_file) { | |
62906e89 | 6622 | req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd), |
ac177053 | 6623 | (sqe_flags & IOSQE_FIXED_FILE)); |
b16fed66 PB |
6624 | if (unlikely(!req->file)) |
6625 | ret = -EBADF; | |
6626 | } | |
6627 | ||
6628 | state->ios_left--; | |
6629 | return ret; | |
6630 | } | |
6631 | ||
a6b8cadc | 6632 | static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
a1ab7b35 | 6633 | const struct io_uring_sqe *sqe) |
282cdc86 | 6634 | __must_hold(&ctx->uring_lock) |
9e645e11 | 6635 | { |
a1ab7b35 | 6636 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 6637 | int ret; |
9e645e11 | 6638 | |
a6b8cadc PB |
6639 | ret = io_init_req(ctx, req, sqe); |
6640 | if (unlikely(ret)) { | |
6641 | fail_req: | |
de59bc10 PB |
6642 | if (link->head) { |
6643 | /* fail even hard links since we don't submit */ | |
93d2bcd2 | 6644 | req_set_fail(link->head); |
f41db273 | 6645 | io_req_complete_failed(link->head, -ECANCELED); |
de59bc10 PB |
6646 | link->head = NULL; |
6647 | } | |
f41db273 | 6648 | io_req_complete_failed(req, ret); |
a6b8cadc PB |
6649 | return ret; |
6650 | } | |
441b8a78 | 6651 | |
be7053b7 PB |
6652 | ret = io_req_prep(req, sqe); |
6653 | if (unlikely(ret)) | |
6654 | goto fail_req; | |
a6b8cadc | 6655 | |
be7053b7 | 6656 | /* don't need @sqe from now on */ |
236daeae OL |
6657 | trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data, |
6658 | req->flags, true, | |
6659 | ctx->flags & IORING_SETUP_SQPOLL); | |
a6b8cadc | 6660 | |
9e645e11 JA |
6661 | /* |
6662 | * If we already have a head request, queue this one for async | |
6663 | * submittal once the head completes. If we don't have a head but | |
6664 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
6665 | * submitted sync once the chain is complete. If none of those | |
6666 | * conditions are true (normal request), then just queue it. | |
6667 | */ | |
863e0560 PB |
6668 | if (link->head) { |
6669 | struct io_kiocb *head = link->head; | |
4e88d6e7 | 6670 | |
b7e298d2 | 6671 | ret = io_req_prep_async(req); |
cf109604 | 6672 | if (unlikely(ret)) |
a6b8cadc | 6673 | goto fail_req; |
9d76377f | 6674 | trace_io_uring_link(ctx, req, head); |
f2f87370 | 6675 | link->last->link = req; |
863e0560 | 6676 | link->last = req; |
32fe525b PB |
6677 | |
6678 | /* last request of a link, enqueue the link */ | |
ef4ff581 | 6679 | if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) { |
863e0560 | 6680 | link->head = NULL; |
5e159204 | 6681 | io_queue_sqe(head); |
32fe525b | 6682 | } |
9e645e11 | 6683 | } else { |
ef4ff581 | 6684 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
863e0560 PB |
6685 | link->head = req; |
6686 | link->last = req; | |
711be031 | 6687 | } else { |
be7053b7 | 6688 | io_queue_sqe(req); |
711be031 | 6689 | } |
9e645e11 | 6690 | } |
2e6e1fde | 6691 | |
1d4240cc | 6692 | return 0; |
9e645e11 JA |
6693 | } |
6694 | ||
9a56a232 JA |
6695 | /* |
6696 | * Batched submission is done, ensure local IO is flushed out. | |
6697 | */ | |
ba88ff11 PB |
6698 | static void io_submit_state_end(struct io_submit_state *state, |
6699 | struct io_ring_ctx *ctx) | |
9a56a232 | 6700 | { |
a1ab7b35 | 6701 | if (state->link.head) |
de59bc10 | 6702 | io_queue_sqe(state->link.head); |
cd0ca2e0 | 6703 | if (state->compl_nr) |
2a2758f2 | 6704 | io_submit_flush_completions(ctx); |
27926b68 JA |
6705 | if (state->plug_started) |
6706 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
6707 | } |
6708 | ||
6709 | /* | |
6710 | * Start submission side cache. | |
6711 | */ | |
6712 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 6713 | unsigned int max_ios) |
9a56a232 | 6714 | { |
27926b68 | 6715 | state->plug_started = false; |
9a56a232 | 6716 | state->ios_left = max_ios; |
a1ab7b35 PB |
6717 | /* set only head, no need to init link_last in advance */ |
6718 | state->link.head = NULL; | |
9a56a232 JA |
6719 | } |
6720 | ||
2b188cc1 JA |
6721 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
6722 | { | |
75b28aff | 6723 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 6724 | |
caf582c6 PB |
6725 | /* |
6726 | * Ensure any loads from the SQEs are done at this point, | |
6727 | * since once we write the new head, the application could | |
6728 | * write new data to them. | |
6729 | */ | |
6730 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
6731 | } |
6732 | ||
2b188cc1 | 6733 | /* |
dd9ae8a0 | 6734 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
6735 | * that is mapped by userspace. This means that care needs to be taken to |
6736 | * ensure that reads are stable, as we cannot rely on userspace always | |
6737 | * being a good citizen. If members of the sqe are validated and then later | |
6738 | * used, it's important that those reads are done through READ_ONCE() to | |
6739 | * prevent a re-load down the line. | |
6740 | */ | |
709b302f | 6741 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 6742 | { |
ea5ab3b5 | 6743 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 6744 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
6745 | |
6746 | /* | |
6747 | * The cached sq head (or cq tail) serves two purposes: | |
6748 | * | |
6749 | * 1) allows us to batch the cost of updating the user visible | |
6750 | * head updates. | |
6751 | * 2) allows the kernel side to track the head on its own, even | |
6752 | * though the application is the one updating it. | |
6753 | */ | |
17d3aeb3 | 6754 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
709b302f PB |
6755 | if (likely(head < ctx->sq_entries)) |
6756 | return &ctx->sq_sqes[head]; | |
2b188cc1 JA |
6757 | |
6758 | /* drop invalid entries */ | |
15641e42 PB |
6759 | ctx->cq_extra--; |
6760 | WRITE_ONCE(ctx->rings->sq_dropped, | |
6761 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
6762 | return NULL; |
6763 | } | |
6764 | ||
0f212204 | 6765 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 6766 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 6767 | { |
09899b19 | 6768 | struct io_uring_task *tctx; |
46c4e16a | 6769 | int submitted = 0; |
6c271ce2 | 6770 | |
ee7d46d9 PB |
6771 | /* make sure SQ entry isn't read before tail */ |
6772 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
2b85edfc PB |
6773 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
6774 | return -EAGAIN; | |
6c271ce2 | 6775 | |
09899b19 PB |
6776 | tctx = current->io_uring; |
6777 | tctx->cached_refs -= nr; | |
6778 | if (unlikely(tctx->cached_refs < 0)) { | |
6779 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
6780 | ||
6781 | percpu_counter_add(&tctx->inflight, refill); | |
6782 | refcount_add(refill, ¤t->usage); | |
6783 | tctx->cached_refs += refill; | |
6784 | } | |
ba88ff11 | 6785 | io_submit_state_start(&ctx->submit_state, nr); |
b14cca0c | 6786 | |
46c4e16a | 6787 | while (submitted < nr) { |
3529d8c2 | 6788 | const struct io_uring_sqe *sqe; |
196be95c | 6789 | struct io_kiocb *req; |
fb5ccc98 | 6790 | |
258b29a9 | 6791 | req = io_alloc_req(ctx); |
196be95c PB |
6792 | if (unlikely(!req)) { |
6793 | if (!submitted) | |
6794 | submitted = -EAGAIN; | |
fb5ccc98 | 6795 | break; |
196be95c | 6796 | } |
4fccfcbb PB |
6797 | sqe = io_get_sqe(ctx); |
6798 | if (unlikely(!sqe)) { | |
6799 | kmem_cache_free(req_cachep, req); | |
6800 | break; | |
6801 | } | |
d3656344 JA |
6802 | /* will complete beyond this point, count as submitted */ |
6803 | submitted++; | |
a1ab7b35 | 6804 | if (io_submit_sqe(ctx, req, sqe)) |
196be95c | 6805 | break; |
6c271ce2 JA |
6806 | } |
6807 | ||
9466f437 PB |
6808 | if (unlikely(submitted != nr)) { |
6809 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
d8a6df10 | 6810 | int unused = nr - ref_used; |
9466f437 | 6811 | |
09899b19 | 6812 | current->io_uring->cached_refs += unused; |
d8a6df10 | 6813 | percpu_ref_put_many(&ctx->refs, unused); |
9466f437 | 6814 | } |
6c271ce2 | 6815 | |
a1ab7b35 | 6816 | io_submit_state_end(&ctx->submit_state, ctx); |
ae9428ca PB |
6817 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
6818 | io_commit_sqring(ctx); | |
6819 | ||
6c271ce2 JA |
6820 | return submitted; |
6821 | } | |
6822 | ||
e4b6d902 PB |
6823 | static inline bool io_sqd_events_pending(struct io_sq_data *sqd) |
6824 | { | |
6825 | return READ_ONCE(sqd->state); | |
6826 | } | |
6827 | ||
23b3628e XW |
6828 | static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx) |
6829 | { | |
6830 | /* Tell userspace we may need a wakeup call */ | |
79ebeaee | 6831 | spin_lock(&ctx->completion_lock); |
20c0b380 NA |
6832 | WRITE_ONCE(ctx->rings->sq_flags, |
6833 | ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP); | |
79ebeaee | 6834 | spin_unlock(&ctx->completion_lock); |
23b3628e XW |
6835 | } |
6836 | ||
6837 | static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx) | |
6838 | { | |
79ebeaee | 6839 | spin_lock(&ctx->completion_lock); |
20c0b380 NA |
6840 | WRITE_ONCE(ctx->rings->sq_flags, |
6841 | ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP); | |
79ebeaee | 6842 | spin_unlock(&ctx->completion_lock); |
23b3628e XW |
6843 | } |
6844 | ||
08369246 | 6845 | static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) |
6c271ce2 | 6846 | { |
c8d1ba58 | 6847 | unsigned int to_submit; |
bdcd3eab | 6848 | int ret = 0; |
6c271ce2 | 6849 | |
c8d1ba58 | 6850 | to_submit = io_sqring_entries(ctx); |
e95eee2d | 6851 | /* if we're handling multiple rings, cap submit size for fairness */ |
4ce8ad95 OL |
6852 | if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE) |
6853 | to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE; | |
e95eee2d | 6854 | |
906a3c6f | 6855 | if (!list_empty(&ctx->iopoll_list) || to_submit) { |
c8d1ba58 | 6856 | unsigned nr_events = 0; |
948e1947 PB |
6857 | const struct cred *creds = NULL; |
6858 | ||
6859 | if (ctx->sq_creds != current_cred()) | |
6860 | creds = override_creds(ctx->sq_creds); | |
a4c0b3de | 6861 | |
c8d1ba58 | 6862 | mutex_lock(&ctx->uring_lock); |
906a3c6f | 6863 | if (!list_empty(&ctx->iopoll_list)) |
a8576af9 | 6864 | io_do_iopoll(ctx, &nr_events, 0); |
906a3c6f | 6865 | |
3b763ba1 PB |
6866 | /* |
6867 | * Don't submit if refs are dying, good for io_uring_register(), | |
6868 | * but also it is relied upon by io_ring_exit_work() | |
6869 | */ | |
0298ef96 PB |
6870 | if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) && |
6871 | !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
08369246 | 6872 | ret = io_submit_sqes(ctx, to_submit); |
c8d1ba58 | 6873 | mutex_unlock(&ctx->uring_lock); |
6c271ce2 | 6874 | |
acfb381d PB |
6875 | if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait)) |
6876 | wake_up(&ctx->sqo_sq_wait); | |
948e1947 PB |
6877 | if (creds) |
6878 | revert_creds(creds); | |
acfb381d | 6879 | } |
6c271ce2 | 6880 | |
08369246 XW |
6881 | return ret; |
6882 | } | |
6c271ce2 | 6883 | |
08369246 XW |
6884 | static void io_sqd_update_thread_idle(struct io_sq_data *sqd) |
6885 | { | |
6886 | struct io_ring_ctx *ctx; | |
6887 | unsigned sq_thread_idle = 0; | |
6c271ce2 | 6888 | |
c9dca27d PB |
6889 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
6890 | sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle); | |
08369246 | 6891 | sqd->sq_thread_idle = sq_thread_idle; |
c8d1ba58 | 6892 | } |
6c271ce2 | 6893 | |
e4b6d902 PB |
6894 | static bool io_sqd_handle_event(struct io_sq_data *sqd) |
6895 | { | |
6896 | bool did_sig = false; | |
6897 | struct ksignal ksig; | |
6898 | ||
6899 | if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) || | |
6900 | signal_pending(current)) { | |
6901 | mutex_unlock(&sqd->lock); | |
6902 | if (signal_pending(current)) | |
6903 | did_sig = get_signal(&ksig); | |
6904 | cond_resched(); | |
6905 | mutex_lock(&sqd->lock); | |
6906 | } | |
e4b6d902 PB |
6907 | return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
6908 | } | |
6909 | ||
c8d1ba58 JA |
6910 | static int io_sq_thread(void *data) |
6911 | { | |
69fb2131 JA |
6912 | struct io_sq_data *sqd = data; |
6913 | struct io_ring_ctx *ctx; | |
a0d9205f | 6914 | unsigned long timeout = 0; |
37d1e2e3 | 6915 | char buf[TASK_COMM_LEN]; |
08369246 | 6916 | DEFINE_WAIT(wait); |
6c271ce2 | 6917 | |
696ee88a | 6918 | snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid); |
37d1e2e3 | 6919 | set_task_comm(current, buf); |
37d1e2e3 JA |
6920 | |
6921 | if (sqd->sq_cpu != -1) | |
6922 | set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu)); | |
6923 | else | |
6924 | set_cpus_allowed_ptr(current, cpu_online_mask); | |
6925 | current->flags |= PF_NO_SETAFFINITY; | |
6926 | ||
09a6f4ef | 6927 | mutex_lock(&sqd->lock); |
e4b6d902 | 6928 | while (1) { |
1a924a80 | 6929 | bool cap_entries, sqt_spin = false; |
c1edbf5f | 6930 | |
e4b6d902 PB |
6931 | if (io_sqd_events_pending(sqd) || signal_pending(current)) { |
6932 | if (io_sqd_handle_event(sqd)) | |
c7d95613 | 6933 | break; |
08369246 XW |
6934 | timeout = jiffies + sqd->sq_thread_idle; |
6935 | } | |
e4b6d902 | 6936 | |
e95eee2d | 6937 | cap_entries = !list_is_singular(&sqd->ctx_list); |
69fb2131 | 6938 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
948e1947 | 6939 | int ret = __io_sq_thread(ctx, cap_entries); |
7c30f36a | 6940 | |
08369246 XW |
6941 | if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list))) |
6942 | sqt_spin = true; | |
69fb2131 | 6943 | } |
dd432ea5 PB |
6944 | if (io_run_task_work()) |
6945 | sqt_spin = true; | |
6c271ce2 | 6946 | |
08369246 | 6947 | if (sqt_spin || !time_after(jiffies, timeout)) { |
c8d1ba58 | 6948 | cond_resched(); |
08369246 XW |
6949 | if (sqt_spin) |
6950 | timeout = jiffies + sqd->sq_thread_idle; | |
6951 | continue; | |
6952 | } | |
6953 | ||
08369246 | 6954 | prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE); |
dd432ea5 | 6955 | if (!io_sqd_events_pending(sqd) && !current->task_works) { |
1a924a80 PB |
6956 | bool needs_sched = true; |
6957 | ||
724cb4f9 | 6958 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
aaa9f0f4 PB |
6959 | io_ring_set_wakeup_flag(ctx); |
6960 | ||
724cb4f9 HX |
6961 | if ((ctx->flags & IORING_SETUP_IOPOLL) && |
6962 | !list_empty_careful(&ctx->iopoll_list)) { | |
6963 | needs_sched = false; | |
6964 | break; | |
6965 | } | |
6966 | if (io_sqring_entries(ctx)) { | |
6967 | needs_sched = false; | |
6968 | break; | |
6969 | } | |
6970 | } | |
6971 | ||
6972 | if (needs_sched) { | |
6973 | mutex_unlock(&sqd->lock); | |
6974 | schedule(); | |
6975 | mutex_lock(&sqd->lock); | |
6976 | } | |
69fb2131 JA |
6977 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
6978 | io_ring_clear_wakeup_flag(ctx); | |
6c271ce2 | 6979 | } |
08369246 XW |
6980 | |
6981 | finish_wait(&sqd->wait, &wait); | |
6982 | timeout = jiffies + sqd->sq_thread_idle; | |
6c271ce2 | 6983 | } |
28cea78a | 6984 | |
78cc687b | 6985 | io_uring_cancel_generic(true, sqd); |
37d1e2e3 | 6986 | sqd->thread = NULL; |
05962f95 | 6987 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
5f3f26f9 | 6988 | io_ring_set_wakeup_flag(ctx); |
521d6a73 | 6989 | io_run_task_work(); |
734551df PB |
6990 | mutex_unlock(&sqd->lock); |
6991 | ||
37d1e2e3 JA |
6992 | complete(&sqd->exited); |
6993 | do_exit(0); | |
6c271ce2 JA |
6994 | } |
6995 | ||
bda52162 JA |
6996 | struct io_wait_queue { |
6997 | struct wait_queue_entry wq; | |
6998 | struct io_ring_ctx *ctx; | |
5fd46178 | 6999 | unsigned cq_tail; |
bda52162 JA |
7000 | unsigned nr_timeouts; |
7001 | }; | |
7002 | ||
6c503150 | 7003 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
7004 | { |
7005 | struct io_ring_ctx *ctx = iowq->ctx; | |
5fd46178 | 7006 | int dist = ctx->cached_cq_tail - (int) iowq->cq_tail; |
bda52162 JA |
7007 | |
7008 | /* | |
d195a66e | 7009 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
7010 | * started waiting. For timeouts, we always want to return to userspace, |
7011 | * regardless of event count. | |
7012 | */ | |
5fd46178 | 7013 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
7014 | } |
7015 | ||
7016 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
7017 | int wake_flags, void *key) | |
7018 | { | |
7019 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
7020 | wq); | |
7021 | ||
6c503150 PB |
7022 | /* |
7023 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
7024 | * the task, and the next invocation will do it. | |
7025 | */ | |
5ed7a37d | 7026 | if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow)) |
6c503150 PB |
7027 | return autoremove_wake_function(curr, mode, wake_flags, key); |
7028 | return -1; | |
bda52162 JA |
7029 | } |
7030 | ||
af9c1a44 JA |
7031 | static int io_run_task_work_sig(void) |
7032 | { | |
7033 | if (io_run_task_work()) | |
7034 | return 1; | |
7035 | if (!signal_pending(current)) | |
7036 | return 0; | |
0b8cfa97 | 7037 | if (test_thread_flag(TIF_NOTIFY_SIGNAL)) |
792ee0f6 | 7038 | return -ERESTARTSYS; |
af9c1a44 JA |
7039 | return -EINTR; |
7040 | } | |
7041 | ||
eeb60b9a PB |
7042 | /* when returns >0, the caller should retry */ |
7043 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
7044 | struct io_wait_queue *iowq, | |
7045 | signed long *timeout) | |
7046 | { | |
7047 | int ret; | |
7048 | ||
7049 | /* make sure we run task_work before checking for signals */ | |
7050 | ret = io_run_task_work_sig(); | |
7051 | if (ret || io_should_wake(iowq)) | |
7052 | return ret; | |
7053 | /* let the caller flush overflows, retry */ | |
5ed7a37d | 7054 | if (test_bit(0, &ctx->check_cq_overflow)) |
eeb60b9a PB |
7055 | return 1; |
7056 | ||
7057 | *timeout = schedule_timeout(*timeout); | |
7058 | return !*timeout ? -ETIME : 1; | |
7059 | } | |
7060 | ||
2b188cc1 JA |
7061 | /* |
7062 | * Wait until events become available, if we don't already have some. The | |
7063 | * application must reap them itself, as they reside on the shared cq ring. | |
7064 | */ | |
7065 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
7066 | const sigset_t __user *sig, size_t sigsz, |
7067 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 7068 | { |
90291099 | 7069 | struct io_wait_queue iowq; |
75b28aff | 7070 | struct io_rings *rings = ctx->rings; |
c1d5a224 PB |
7071 | signed long timeout = MAX_SCHEDULE_TIMEOUT; |
7072 | int ret; | |
2b188cc1 | 7073 | |
b41e9852 | 7074 | do { |
90f67366 | 7075 | io_cqring_overflow_flush(ctx); |
6c503150 | 7076 | if (io_cqring_events(ctx) >= min_events) |
b41e9852 | 7077 | return 0; |
4c6e277c | 7078 | if (!io_run_task_work()) |
b41e9852 | 7079 | break; |
b41e9852 | 7080 | } while (1); |
2b188cc1 JA |
7081 | |
7082 | if (sig) { | |
9e75ad5d AB |
7083 | #ifdef CONFIG_COMPAT |
7084 | if (in_compat_syscall()) | |
7085 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 7086 | sigsz); |
9e75ad5d AB |
7087 | else |
7088 | #endif | |
b772434b | 7089 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 7090 | |
2b188cc1 JA |
7091 | if (ret) |
7092 | return ret; | |
7093 | } | |
7094 | ||
c73ebb68 | 7095 | if (uts) { |
c1d5a224 PB |
7096 | struct timespec64 ts; |
7097 | ||
c73ebb68 HX |
7098 | if (get_timespec64(&ts, uts)) |
7099 | return -EFAULT; | |
7100 | timeout = timespec64_to_jiffies(&ts); | |
7101 | } | |
7102 | ||
90291099 PB |
7103 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
7104 | iowq.wq.private = current; | |
7105 | INIT_LIST_HEAD(&iowq.wq.entry); | |
7106 | iowq.ctx = ctx; | |
bda52162 | 7107 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 7108 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
90291099 | 7109 | |
c826bd7a | 7110 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 7111 | do { |
ca0a2651 | 7112 | /* if we can't even flush overflow, don't wait for more */ |
90f67366 | 7113 | if (!io_cqring_overflow_flush(ctx)) { |
ca0a2651 JA |
7114 | ret = -EBUSY; |
7115 | break; | |
7116 | } | |
311997b3 | 7117 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, |
bda52162 | 7118 | TASK_INTERRUPTIBLE); |
eeb60b9a | 7119 | ret = io_cqring_wait_schedule(ctx, &iowq, &timeout); |
311997b3 | 7120 | finish_wait(&ctx->cq_wait, &iowq.wq); |
ca0a2651 | 7121 | cond_resched(); |
eeb60b9a | 7122 | } while (ret > 0); |
bda52162 | 7123 | |
b7db41c9 | 7124 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 7125 | |
75b28aff | 7126 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
7127 | } |
7128 | ||
9123c8ff | 7129 | static void io_free_page_table(void **table, size_t size) |
05f3fb3c | 7130 | { |
9123c8ff | 7131 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); |
05f3fb3c | 7132 | |
846a4ef2 | 7133 | for (i = 0; i < nr_tables; i++) |
9123c8ff PB |
7134 | kfree(table[i]); |
7135 | kfree(table); | |
7136 | } | |
7137 | ||
7138 | static void **io_alloc_page_table(size_t size) | |
7139 | { | |
7140 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); | |
7141 | size_t init_size = size; | |
7142 | void **table; | |
7143 | ||
7144 | table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL); | |
7145 | if (!table) | |
7146 | return NULL; | |
7147 | ||
7148 | for (i = 0; i < nr_tables; i++) { | |
27f6b318 | 7149 | unsigned int this_size = min_t(size_t, size, PAGE_SIZE); |
9123c8ff PB |
7150 | |
7151 | table[i] = kzalloc(this_size, GFP_KERNEL); | |
7152 | if (!table[i]) { | |
7153 | io_free_page_table(table, init_size); | |
7154 | return NULL; | |
7155 | } | |
7156 | size -= this_size; | |
7157 | } | |
7158 | return table; | |
05f3fb3c JA |
7159 | } |
7160 | ||
28a9fe25 | 7161 | static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node) |
1642b445 | 7162 | { |
28a9fe25 PB |
7163 | percpu_ref_exit(&ref_node->refs); |
7164 | kfree(ref_node); | |
1642b445 PB |
7165 | } |
7166 | ||
b9bd2bea PB |
7167 | static void io_rsrc_node_ref_zero(struct percpu_ref *ref) |
7168 | { | |
7169 | struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs); | |
7170 | struct io_ring_ctx *ctx = node->rsrc_data->ctx; | |
7171 | unsigned long flags; | |
7172 | bool first_add = false; | |
7173 | ||
7174 | spin_lock_irqsave(&ctx->rsrc_ref_lock, flags); | |
7175 | node->done = true; | |
7176 | ||
7177 | while (!list_empty(&ctx->rsrc_ref_list)) { | |
7178 | node = list_first_entry(&ctx->rsrc_ref_list, | |
7179 | struct io_rsrc_node, node); | |
7180 | /* recycle ref nodes in order */ | |
7181 | if (!node->done) | |
7182 | break; | |
7183 | list_del(&node->node); | |
7184 | first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist); | |
7185 | } | |
7186 | spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags); | |
7187 | ||
7188 | if (first_add) | |
7189 | mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ); | |
7190 | } | |
7191 | ||
7192 | static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx) | |
7193 | { | |
7194 | struct io_rsrc_node *ref_node; | |
7195 | ||
7196 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); | |
7197 | if (!ref_node) | |
7198 | return NULL; | |
7199 | ||
7200 | if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero, | |
7201 | 0, GFP_KERNEL)) { | |
7202 | kfree(ref_node); | |
7203 | return NULL; | |
7204 | } | |
7205 | INIT_LIST_HEAD(&ref_node->node); | |
7206 | INIT_LIST_HEAD(&ref_node->rsrc_list); | |
7207 | ref_node->done = false; | |
7208 | return ref_node; | |
7209 | } | |
7210 | ||
a7f0ed5a PB |
7211 | static void io_rsrc_node_switch(struct io_ring_ctx *ctx, |
7212 | struct io_rsrc_data *data_to_kill) | |
6b06314c | 7213 | { |
a7f0ed5a PB |
7214 | WARN_ON_ONCE(!ctx->rsrc_backup_node); |
7215 | WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node); | |
6b06314c | 7216 | |
a7f0ed5a PB |
7217 | if (data_to_kill) { |
7218 | struct io_rsrc_node *rsrc_node = ctx->rsrc_node; | |
82fbcfa9 | 7219 | |
a7f0ed5a | 7220 | rsrc_node->rsrc_data = data_to_kill; |
4956b9ea | 7221 | spin_lock_irq(&ctx->rsrc_ref_lock); |
a7f0ed5a | 7222 | list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list); |
4956b9ea | 7223 | spin_unlock_irq(&ctx->rsrc_ref_lock); |
82fbcfa9 | 7224 | |
3e942498 | 7225 | atomic_inc(&data_to_kill->refs); |
a7f0ed5a PB |
7226 | percpu_ref_kill(&rsrc_node->refs); |
7227 | ctx->rsrc_node = NULL; | |
7228 | } | |
6b06314c | 7229 | |
a7f0ed5a PB |
7230 | if (!ctx->rsrc_node) { |
7231 | ctx->rsrc_node = ctx->rsrc_backup_node; | |
7232 | ctx->rsrc_backup_node = NULL; | |
7233 | } | |
8bad28d8 HX |
7234 | } |
7235 | ||
a7f0ed5a | 7236 | static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx) |
8dd03afe PB |
7237 | { |
7238 | if (ctx->rsrc_backup_node) | |
7239 | return 0; | |
b895c9a6 | 7240 | ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx); |
8dd03afe | 7241 | return ctx->rsrc_backup_node ? 0 : -ENOMEM; |
8bad28d8 HX |
7242 | } |
7243 | ||
40ae0ff7 | 7244 | static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx) |
8bad28d8 HX |
7245 | { |
7246 | int ret; | |
05589553 | 7247 | |
215c3902 | 7248 | /* As we may drop ->uring_lock, other task may have started quiesce */ |
8bad28d8 HX |
7249 | if (data->quiesce) |
7250 | return -ENXIO; | |
05589553 | 7251 | |
8bad28d8 | 7252 | data->quiesce = true; |
1ffc5422 | 7253 | do { |
a7f0ed5a | 7254 | ret = io_rsrc_node_switch_start(ctx); |
8dd03afe | 7255 | if (ret) |
f2303b1f | 7256 | break; |
a7f0ed5a | 7257 | io_rsrc_node_switch(ctx, data); |
f2303b1f | 7258 | |
3e942498 PB |
7259 | /* kill initial ref, already quiesced if zero */ |
7260 | if (atomic_dec_and_test(&data->refs)) | |
7261 | break; | |
c018db4a | 7262 | mutex_unlock(&ctx->uring_lock); |
8bad28d8 | 7263 | flush_delayed_work(&ctx->rsrc_put_work); |
1ffc5422 | 7264 | ret = wait_for_completion_interruptible(&data->done); |
c018db4a JA |
7265 | if (!ret) { |
7266 | mutex_lock(&ctx->uring_lock); | |
1ffc5422 | 7267 | break; |
c018db4a | 7268 | } |
8bad28d8 | 7269 | |
3e942498 PB |
7270 | atomic_inc(&data->refs); |
7271 | /* wait for all works potentially completing data->done */ | |
7272 | flush_delayed_work(&ctx->rsrc_put_work); | |
cb5e1b81 | 7273 | reinit_completion(&data->done); |
8dd03afe | 7274 | |
1ffc5422 | 7275 | ret = io_run_task_work_sig(); |
8bad28d8 | 7276 | mutex_lock(&ctx->uring_lock); |
f2303b1f | 7277 | } while (ret >= 0); |
8bad28d8 | 7278 | data->quiesce = false; |
05f3fb3c | 7279 | |
8bad28d8 | 7280 | return ret; |
d7954b2b BM |
7281 | } |
7282 | ||
2d091d62 PB |
7283 | static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx) |
7284 | { | |
7285 | unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK; | |
7286 | unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT; | |
7287 | ||
7288 | return &data->tags[table_idx][off]; | |
7289 | } | |
7290 | ||
44b31f2f | 7291 | static void io_rsrc_data_free(struct io_rsrc_data *data) |
1ad555c6 | 7292 | { |
2d091d62 PB |
7293 | size_t size = data->nr * sizeof(data->tags[0][0]); |
7294 | ||
7295 | if (data->tags) | |
7296 | io_free_page_table((void **)data->tags, size); | |
44b31f2f PB |
7297 | kfree(data); |
7298 | } | |
7299 | ||
d878c816 PB |
7300 | static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put, |
7301 | u64 __user *utags, unsigned nr, | |
7302 | struct io_rsrc_data **pdata) | |
1ad555c6 | 7303 | { |
b895c9a6 | 7304 | struct io_rsrc_data *data; |
2d091d62 | 7305 | int ret = -ENOMEM; |
d878c816 | 7306 | unsigned i; |
1ad555c6 BM |
7307 | |
7308 | data = kzalloc(sizeof(*data), GFP_KERNEL); | |
7309 | if (!data) | |
d878c816 | 7310 | return -ENOMEM; |
2d091d62 | 7311 | data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); |
b60c8dce | 7312 | if (!data->tags) { |
1ad555c6 | 7313 | kfree(data); |
d878c816 PB |
7314 | return -ENOMEM; |
7315 | } | |
2d091d62 PB |
7316 | |
7317 | data->nr = nr; | |
7318 | data->ctx = ctx; | |
7319 | data->do_put = do_put; | |
d878c816 | 7320 | if (utags) { |
2d091d62 | 7321 | ret = -EFAULT; |
d878c816 | 7322 | for (i = 0; i < nr; i++) { |
fdd1dc31 CIK |
7323 | u64 *tag_slot = io_get_tag_slot(data, i); |
7324 | ||
7325 | if (copy_from_user(tag_slot, &utags[i], | |
7326 | sizeof(*tag_slot))) | |
2d091d62 | 7327 | goto fail; |
d878c816 | 7328 | } |
1ad555c6 | 7329 | } |
b60c8dce | 7330 | |
3e942498 | 7331 | atomic_set(&data->refs, 1); |
1ad555c6 | 7332 | init_completion(&data->done); |
d878c816 PB |
7333 | *pdata = data; |
7334 | return 0; | |
2d091d62 PB |
7335 | fail: |
7336 | io_rsrc_data_free(data); | |
7337 | return ret; | |
1ad555c6 BM |
7338 | } |
7339 | ||
9123c8ff PB |
7340 | static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files) |
7341 | { | |
042b0d85 | 7342 | table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL); |
9123c8ff PB |
7343 | return !!table->files; |
7344 | } | |
7345 | ||
042b0d85 | 7346 | static void io_free_file_tables(struct io_file_table *table) |
9123c8ff | 7347 | { |
042b0d85 | 7348 | kvfree(table->files); |
9123c8ff PB |
7349 | table->files = NULL; |
7350 | } | |
7351 | ||
fff4db76 | 7352 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 7353 | { |
fff4db76 PB |
7354 | #if defined(CONFIG_UNIX) |
7355 | if (ctx->ring_sock) { | |
7356 | struct sock *sock = ctx->ring_sock->sk; | |
7357 | struct sk_buff *skb; | |
7358 | ||
7359 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
7360 | kfree_skb(skb); | |
7361 | } | |
7362 | #else | |
7363 | int i; | |
7364 | ||
7365 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7366 | struct file *file; | |
7367 | ||
7368 | file = io_file_from_index(ctx, i); | |
7369 | if (file) | |
7370 | fput(file); | |
7371 | } | |
7372 | #endif | |
042b0d85 | 7373 | io_free_file_tables(&ctx->file_table); |
44b31f2f | 7374 | io_rsrc_data_free(ctx->file_data); |
fff4db76 PB |
7375 | ctx->file_data = NULL; |
7376 | ctx->nr_user_files = 0; | |
1ad555c6 BM |
7377 | } |
7378 | ||
d7954b2b BM |
7379 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
7380 | { | |
d7954b2b BM |
7381 | int ret; |
7382 | ||
08480400 | 7383 | if (!ctx->file_data) |
d7954b2b | 7384 | return -ENXIO; |
08480400 PB |
7385 | ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); |
7386 | if (!ret) | |
7387 | __io_sqe_files_unregister(ctx); | |
7388 | return ret; | |
6b06314c JA |
7389 | } |
7390 | ||
37d1e2e3 | 7391 | static void io_sq_thread_unpark(struct io_sq_data *sqd) |
09a6f4ef | 7392 | __releases(&sqd->lock) |
37d1e2e3 | 7393 | { |
521d6a73 PB |
7394 | WARN_ON_ONCE(sqd->thread == current); |
7395 | ||
9e138a48 PB |
7396 | /* |
7397 | * Do the dance but not conditional clear_bit() because it'd race with | |
7398 | * other threads incrementing park_pending and setting the bit. | |
7399 | */ | |
37d1e2e3 | 7400 | clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
9e138a48 PB |
7401 | if (atomic_dec_return(&sqd->park_pending)) |
7402 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); | |
09a6f4ef | 7403 | mutex_unlock(&sqd->lock); |
37d1e2e3 JA |
7404 | } |
7405 | ||
86e0d676 | 7406 | static void io_sq_thread_park(struct io_sq_data *sqd) |
09a6f4ef | 7407 | __acquires(&sqd->lock) |
37d1e2e3 | 7408 | { |
521d6a73 PB |
7409 | WARN_ON_ONCE(sqd->thread == current); |
7410 | ||
9e138a48 | 7411 | atomic_inc(&sqd->park_pending); |
86e0d676 | 7412 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
09a6f4ef | 7413 | mutex_lock(&sqd->lock); |
05962f95 | 7414 | if (sqd->thread) |
86e0d676 | 7415 | wake_up_process(sqd->thread); |
37d1e2e3 JA |
7416 | } |
7417 | ||
7418 | static void io_sq_thread_stop(struct io_sq_data *sqd) | |
7419 | { | |
521d6a73 | 7420 | WARN_ON_ONCE(sqd->thread == current); |
88885f66 | 7421 | WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)); |
521d6a73 | 7422 | |
05962f95 | 7423 | set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
88885f66 | 7424 | mutex_lock(&sqd->lock); |
e8f98f24 JA |
7425 | if (sqd->thread) |
7426 | wake_up_process(sqd->thread); | |
09a6f4ef | 7427 | mutex_unlock(&sqd->lock); |
05962f95 | 7428 | wait_for_completion(&sqd->exited); |
37d1e2e3 JA |
7429 | } |
7430 | ||
534ca6d6 | 7431 | static void io_put_sq_data(struct io_sq_data *sqd) |
6c271ce2 | 7432 | { |
534ca6d6 | 7433 | if (refcount_dec_and_test(&sqd->refs)) { |
9e138a48 PB |
7434 | WARN_ON_ONCE(atomic_read(&sqd->park_pending)); |
7435 | ||
37d1e2e3 JA |
7436 | io_sq_thread_stop(sqd); |
7437 | kfree(sqd); | |
7438 | } | |
7439 | } | |
7440 | ||
7441 | static void io_sq_thread_finish(struct io_ring_ctx *ctx) | |
7442 | { | |
7443 | struct io_sq_data *sqd = ctx->sq_data; | |
7444 | ||
7445 | if (sqd) { | |
05962f95 | 7446 | io_sq_thread_park(sqd); |
521d6a73 | 7447 | list_del_init(&ctx->sqd_list); |
37d1e2e3 | 7448 | io_sqd_update_thread_idle(sqd); |
05962f95 | 7449 | io_sq_thread_unpark(sqd); |
37d1e2e3 JA |
7450 | |
7451 | io_put_sq_data(sqd); | |
7452 | ctx->sq_data = NULL; | |
534ca6d6 JA |
7453 | } |
7454 | } | |
7455 | ||
aa06165d JA |
7456 | static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p) |
7457 | { | |
7458 | struct io_ring_ctx *ctx_attach; | |
7459 | struct io_sq_data *sqd; | |
7460 | struct fd f; | |
7461 | ||
7462 | f = fdget(p->wq_fd); | |
7463 | if (!f.file) | |
7464 | return ERR_PTR(-ENXIO); | |
7465 | if (f.file->f_op != &io_uring_fops) { | |
7466 | fdput(f); | |
7467 | return ERR_PTR(-EINVAL); | |
7468 | } | |
7469 | ||
7470 | ctx_attach = f.file->private_data; | |
7471 | sqd = ctx_attach->sq_data; | |
7472 | if (!sqd) { | |
7473 | fdput(f); | |
7474 | return ERR_PTR(-EINVAL); | |
7475 | } | |
5c2469e0 JA |
7476 | if (sqd->task_tgid != current->tgid) { |
7477 | fdput(f); | |
7478 | return ERR_PTR(-EPERM); | |
7479 | } | |
aa06165d JA |
7480 | |
7481 | refcount_inc(&sqd->refs); | |
7482 | fdput(f); | |
7483 | return sqd; | |
7484 | } | |
7485 | ||
26984fbf PB |
7486 | static struct io_sq_data *io_get_sq_data(struct io_uring_params *p, |
7487 | bool *attached) | |
534ca6d6 JA |
7488 | { |
7489 | struct io_sq_data *sqd; | |
7490 | ||
26984fbf | 7491 | *attached = false; |
5c2469e0 JA |
7492 | if (p->flags & IORING_SETUP_ATTACH_WQ) { |
7493 | sqd = io_attach_sq_data(p); | |
26984fbf PB |
7494 | if (!IS_ERR(sqd)) { |
7495 | *attached = true; | |
5c2469e0 | 7496 | return sqd; |
26984fbf | 7497 | } |
5c2469e0 JA |
7498 | /* fall through for EPERM case, setup new sqd/task */ |
7499 | if (PTR_ERR(sqd) != -EPERM) | |
7500 | return sqd; | |
7501 | } | |
aa06165d | 7502 | |
534ca6d6 JA |
7503 | sqd = kzalloc(sizeof(*sqd), GFP_KERNEL); |
7504 | if (!sqd) | |
7505 | return ERR_PTR(-ENOMEM); | |
7506 | ||
9e138a48 | 7507 | atomic_set(&sqd->park_pending, 0); |
534ca6d6 | 7508 | refcount_set(&sqd->refs, 1); |
69fb2131 | 7509 | INIT_LIST_HEAD(&sqd->ctx_list); |
09a6f4ef | 7510 | mutex_init(&sqd->lock); |
534ca6d6 | 7511 | init_waitqueue_head(&sqd->wait); |
37d1e2e3 | 7512 | init_completion(&sqd->exited); |
534ca6d6 JA |
7513 | return sqd; |
7514 | } | |
7515 | ||
6b06314c | 7516 | #if defined(CONFIG_UNIX) |
6b06314c JA |
7517 | /* |
7518 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
7519 | * the io_uring can be safely unregistered on process exit, even if we have | |
7520 | * loops in the file referencing. | |
7521 | */ | |
7522 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
7523 | { | |
7524 | struct sock *sk = ctx->ring_sock->sk; | |
7525 | struct scm_fp_list *fpl; | |
7526 | struct sk_buff *skb; | |
08a45173 | 7527 | int i, nr_files; |
6b06314c | 7528 | |
6b06314c JA |
7529 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); |
7530 | if (!fpl) | |
7531 | return -ENOMEM; | |
7532 | ||
7533 | skb = alloc_skb(0, GFP_KERNEL); | |
7534 | if (!skb) { | |
7535 | kfree(fpl); | |
7536 | return -ENOMEM; | |
7537 | } | |
7538 | ||
7539 | skb->sk = sk; | |
6b06314c | 7540 | |
08a45173 | 7541 | nr_files = 0; |
62e398be | 7542 | fpl->user = get_uid(current_user()); |
6b06314c | 7543 | for (i = 0; i < nr; i++) { |
65e19f54 JA |
7544 | struct file *file = io_file_from_index(ctx, i + offset); |
7545 | ||
7546 | if (!file) | |
08a45173 | 7547 | continue; |
65e19f54 | 7548 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
7549 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
7550 | nr_files++; | |
6b06314c JA |
7551 | } |
7552 | ||
08a45173 JA |
7553 | if (nr_files) { |
7554 | fpl->max = SCM_MAX_FD; | |
7555 | fpl->count = nr_files; | |
7556 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 7557 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
7558 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
7559 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 7560 | |
08a45173 JA |
7561 | for (i = 0; i < nr_files; i++) |
7562 | fput(fpl->fp[i]); | |
7563 | } else { | |
7564 | kfree_skb(skb); | |
7565 | kfree(fpl); | |
7566 | } | |
6b06314c JA |
7567 | |
7568 | return 0; | |
7569 | } | |
7570 | ||
7571 | /* | |
7572 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
7573 | * causes regular reference counting to break down. We rely on the UNIX | |
7574 | * garbage collection to take care of this problem for us. | |
7575 | */ | |
7576 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7577 | { | |
7578 | unsigned left, total; | |
7579 | int ret = 0; | |
7580 | ||
7581 | total = 0; | |
7582 | left = ctx->nr_user_files; | |
7583 | while (left) { | |
7584 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
7585 | |
7586 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
7587 | if (ret) | |
7588 | break; | |
7589 | left -= this_files; | |
7590 | total += this_files; | |
7591 | } | |
7592 | ||
7593 | if (!ret) | |
7594 | return 0; | |
7595 | ||
7596 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
7597 | struct file *file = io_file_from_index(ctx, total); |
7598 | ||
7599 | if (file) | |
7600 | fput(file); | |
6b06314c JA |
7601 | total++; |
7602 | } | |
7603 | ||
7604 | return ret; | |
7605 | } | |
7606 | #else | |
7607 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7608 | { | |
7609 | return 0; | |
7610 | } | |
7611 | #endif | |
7612 | ||
47e90392 | 7613 | static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) |
05f3fb3c | 7614 | { |
50238531 | 7615 | struct file *file = prsrc->file; |
05f3fb3c JA |
7616 | #if defined(CONFIG_UNIX) |
7617 | struct sock *sock = ctx->ring_sock->sk; | |
7618 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
7619 | struct sk_buff *skb; | |
7620 | int i; | |
7621 | ||
7622 | __skb_queue_head_init(&list); | |
7623 | ||
7624 | /* | |
7625 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
7626 | * remove this entry and rearrange the file array. | |
7627 | */ | |
7628 | skb = skb_dequeue(head); | |
7629 | while (skb) { | |
7630 | struct scm_fp_list *fp; | |
7631 | ||
7632 | fp = UNIXCB(skb).fp; | |
7633 | for (i = 0; i < fp->count; i++) { | |
7634 | int left; | |
7635 | ||
7636 | if (fp->fp[i] != file) | |
7637 | continue; | |
7638 | ||
7639 | unix_notinflight(fp->user, fp->fp[i]); | |
7640 | left = fp->count - 1 - i; | |
7641 | if (left) { | |
7642 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
7643 | left * sizeof(struct file *)); | |
7644 | } | |
7645 | fp->count--; | |
7646 | if (!fp->count) { | |
7647 | kfree_skb(skb); | |
7648 | skb = NULL; | |
7649 | } else { | |
7650 | __skb_queue_tail(&list, skb); | |
7651 | } | |
7652 | fput(file); | |
7653 | file = NULL; | |
7654 | break; | |
7655 | } | |
7656 | ||
7657 | if (!file) | |
7658 | break; | |
7659 | ||
7660 | __skb_queue_tail(&list, skb); | |
7661 | ||
7662 | skb = skb_dequeue(head); | |
7663 | } | |
7664 | ||
7665 | if (skb_peek(&list)) { | |
7666 | spin_lock_irq(&head->lock); | |
7667 | while ((skb = __skb_dequeue(&list)) != NULL) | |
7668 | __skb_queue_tail(head, skb); | |
7669 | spin_unlock_irq(&head->lock); | |
7670 | } | |
7671 | #else | |
7672 | fput(file); | |
7673 | #endif | |
7674 | } | |
7675 | ||
b895c9a6 | 7676 | static void __io_rsrc_put_work(struct io_rsrc_node *ref_node) |
65e19f54 | 7677 | { |
b895c9a6 | 7678 | struct io_rsrc_data *rsrc_data = ref_node->rsrc_data; |
269bbe5f BM |
7679 | struct io_ring_ctx *ctx = rsrc_data->ctx; |
7680 | struct io_rsrc_put *prsrc, *tmp; | |
05589553 | 7681 | |
269bbe5f BM |
7682 | list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) { |
7683 | list_del(&prsrc->list); | |
b60c8dce PB |
7684 | |
7685 | if (prsrc->tag) { | |
7686 | bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL; | |
b60c8dce PB |
7687 | |
7688 | io_ring_submit_lock(ctx, lock_ring); | |
79ebeaee | 7689 | spin_lock(&ctx->completion_lock); |
b60c8dce | 7690 | io_cqring_fill_event(ctx, prsrc->tag, 0, 0); |
2840f710 | 7691 | ctx->cq_extra++; |
b60c8dce | 7692 | io_commit_cqring(ctx); |
79ebeaee | 7693 | spin_unlock(&ctx->completion_lock); |
b60c8dce PB |
7694 | io_cqring_ev_posted(ctx); |
7695 | io_ring_submit_unlock(ctx, lock_ring); | |
7696 | } | |
7697 | ||
40ae0ff7 | 7698 | rsrc_data->do_put(ctx, prsrc); |
269bbe5f | 7699 | kfree(prsrc); |
65e19f54 | 7700 | } |
05589553 | 7701 | |
28a9fe25 | 7702 | io_rsrc_node_destroy(ref_node); |
3e942498 PB |
7703 | if (atomic_dec_and_test(&rsrc_data->refs)) |
7704 | complete(&rsrc_data->done); | |
2faf852d | 7705 | } |
65e19f54 | 7706 | |
269bbe5f | 7707 | static void io_rsrc_put_work(struct work_struct *work) |
4a38aed2 JA |
7708 | { |
7709 | struct io_ring_ctx *ctx; | |
7710 | struct llist_node *node; | |
7711 | ||
269bbe5f BM |
7712 | ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work); |
7713 | node = llist_del_all(&ctx->rsrc_put_llist); | |
4a38aed2 JA |
7714 | |
7715 | while (node) { | |
b895c9a6 | 7716 | struct io_rsrc_node *ref_node; |
4a38aed2 JA |
7717 | struct llist_node *next = node->next; |
7718 | ||
b895c9a6 | 7719 | ref_node = llist_entry(node, struct io_rsrc_node, llist); |
269bbe5f | 7720 | __io_rsrc_put_work(ref_node); |
4a38aed2 JA |
7721 | node = next; |
7722 | } | |
7723 | } | |
7724 | ||
6b06314c | 7725 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
792e3582 | 7726 | unsigned nr_args, u64 __user *tags) |
6b06314c JA |
7727 | { |
7728 | __s32 __user *fds = (__s32 __user *) arg; | |
05f3fb3c | 7729 | struct file *file; |
f3baed39 | 7730 | int fd, ret; |
846a4ef2 | 7731 | unsigned i; |
6b06314c | 7732 | |
05f3fb3c | 7733 | if (ctx->file_data) |
6b06314c JA |
7734 | return -EBUSY; |
7735 | if (!nr_args) | |
7736 | return -EINVAL; | |
7737 | if (nr_args > IORING_MAX_FIXED_FILES) | |
7738 | return -EMFILE; | |
a7f0ed5a | 7739 | ret = io_rsrc_node_switch_start(ctx); |
f3baed39 PB |
7740 | if (ret) |
7741 | return ret; | |
d878c816 PB |
7742 | ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, |
7743 | &ctx->file_data); | |
7744 | if (ret) | |
7745 | return ret; | |
6b06314c | 7746 | |
f3baed39 | 7747 | ret = -ENOMEM; |
aeca241b | 7748 | if (!io_alloc_file_tables(&ctx->file_table, nr_args)) |
1ad555c6 | 7749 | goto out_free; |
65e19f54 | 7750 | |
08a45173 | 7751 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
d878c816 | 7752 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) { |
600cf3f8 PB |
7753 | ret = -EFAULT; |
7754 | goto out_fput; | |
7755 | } | |
08a45173 | 7756 | /* allow sparse sets */ |
792e3582 PB |
7757 | if (fd == -1) { |
7758 | ret = -EINVAL; | |
2d091d62 | 7759 | if (unlikely(*io_get_tag_slot(ctx->file_data, i))) |
792e3582 | 7760 | goto out_fput; |
08a45173 | 7761 | continue; |
792e3582 | 7762 | } |
6b06314c | 7763 | |
05f3fb3c | 7764 | file = fget(fd); |
6b06314c | 7765 | ret = -EBADF; |
792e3582 | 7766 | if (unlikely(!file)) |
600cf3f8 | 7767 | goto out_fput; |
05f3fb3c | 7768 | |
6b06314c JA |
7769 | /* |
7770 | * Don't allow io_uring instances to be registered. If UNIX | |
7771 | * isn't enabled, then this causes a reference cycle and this | |
7772 | * instance can never get freed. If UNIX is enabled we'll | |
7773 | * handle it just fine, but there's still no point in allowing | |
7774 | * a ring fd as it doesn't support regular read/write anyway. | |
7775 | */ | |
05f3fb3c JA |
7776 | if (file->f_op == &io_uring_fops) { |
7777 | fput(file); | |
600cf3f8 | 7778 | goto out_fput; |
6b06314c | 7779 | } |
aeca241b | 7780 | io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file); |
6b06314c JA |
7781 | } |
7782 | ||
6b06314c | 7783 | ret = io_sqe_files_scm(ctx); |
05589553 | 7784 | if (ret) { |
08480400 | 7785 | __io_sqe_files_unregister(ctx); |
05589553 XW |
7786 | return ret; |
7787 | } | |
6b06314c | 7788 | |
a7f0ed5a | 7789 | io_rsrc_node_switch(ctx, NULL); |
6b06314c | 7790 | return ret; |
600cf3f8 PB |
7791 | out_fput: |
7792 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7793 | file = io_file_from_index(ctx, i); | |
7794 | if (file) | |
7795 | fput(file); | |
7796 | } | |
042b0d85 | 7797 | io_free_file_tables(&ctx->file_table); |
600cf3f8 | 7798 | ctx->nr_user_files = 0; |
600cf3f8 | 7799 | out_free: |
44b31f2f | 7800 | io_rsrc_data_free(ctx->file_data); |
55cbc256 | 7801 | ctx->file_data = NULL; |
6b06314c JA |
7802 | return ret; |
7803 | } | |
7804 | ||
c3a31e60 JA |
7805 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
7806 | int index) | |
7807 | { | |
7808 | #if defined(CONFIG_UNIX) | |
7809 | struct sock *sock = ctx->ring_sock->sk; | |
7810 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
7811 | struct sk_buff *skb; | |
7812 | ||
7813 | /* | |
7814 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
7815 | * file set. If there's no room, fall back to allocating a new skb | |
7816 | * and filling it in. | |
7817 | */ | |
7818 | spin_lock_irq(&head->lock); | |
7819 | skb = skb_peek(head); | |
7820 | if (skb) { | |
7821 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
7822 | ||
7823 | if (fpl->count < SCM_MAX_FD) { | |
7824 | __skb_unlink(skb, head); | |
7825 | spin_unlock_irq(&head->lock); | |
7826 | fpl->fp[fpl->count] = get_file(file); | |
7827 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
7828 | fpl->count++; | |
7829 | spin_lock_irq(&head->lock); | |
7830 | __skb_queue_head(head, skb); | |
7831 | } else { | |
7832 | skb = NULL; | |
7833 | } | |
7834 | } | |
7835 | spin_unlock_irq(&head->lock); | |
7836 | ||
7837 | if (skb) { | |
7838 | fput(file); | |
7839 | return 0; | |
7840 | } | |
7841 | ||
7842 | return __io_sqe_files_scm(ctx, 1, index); | |
7843 | #else | |
7844 | return 0; | |
7845 | #endif | |
7846 | } | |
7847 | ||
b60c8dce | 7848 | static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, |
e7c78371 | 7849 | struct io_rsrc_node *node, void *rsrc) |
05f3fb3c | 7850 | { |
269bbe5f | 7851 | struct io_rsrc_put *prsrc; |
05f3fb3c | 7852 | |
269bbe5f BM |
7853 | prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL); |
7854 | if (!prsrc) | |
a5318d3c | 7855 | return -ENOMEM; |
05f3fb3c | 7856 | |
2d091d62 | 7857 | prsrc->tag = *io_get_tag_slot(data, idx); |
50238531 | 7858 | prsrc->rsrc = rsrc; |
e7c78371 | 7859 | list_add(&prsrc->list, &node->rsrc_list); |
a5318d3c | 7860 | return 0; |
05f3fb3c JA |
7861 | } |
7862 | ||
7863 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
c3bdad02 | 7864 | struct io_uring_rsrc_update2 *up, |
05f3fb3c JA |
7865 | unsigned nr_args) |
7866 | { | |
c3bdad02 | 7867 | u64 __user *tags = u64_to_user_ptr(up->tags); |
98f0b3b4 | 7868 | __s32 __user *fds = u64_to_user_ptr(up->data); |
b895c9a6 | 7869 | struct io_rsrc_data *data = ctx->file_data; |
a04b0ac0 PB |
7870 | struct io_fixed_file *file_slot; |
7871 | struct file *file; | |
98f0b3b4 PB |
7872 | int fd, i, err = 0; |
7873 | unsigned int done; | |
05589553 | 7874 | bool needs_switch = false; |
c3a31e60 | 7875 | |
98f0b3b4 PB |
7876 | if (!ctx->file_data) |
7877 | return -ENXIO; | |
7878 | if (up->offset + nr_args > ctx->nr_user_files) | |
c3a31e60 JA |
7879 | return -EINVAL; |
7880 | ||
67973b93 | 7881 | for (done = 0; done < nr_args; done++) { |
c3bdad02 PB |
7882 | u64 tag = 0; |
7883 | ||
7884 | if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || | |
7885 | copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
c3a31e60 JA |
7886 | err = -EFAULT; |
7887 | break; | |
7888 | } | |
c3bdad02 PB |
7889 | if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { |
7890 | err = -EINVAL; | |
7891 | break; | |
7892 | } | |
4e0377a1 | 7893 | if (fd == IORING_REGISTER_FILES_SKIP) |
7894 | continue; | |
7895 | ||
67973b93 | 7896 | i = array_index_nospec(up->offset + done, ctx->nr_user_files); |
aeca241b | 7897 | file_slot = io_fixed_file_slot(&ctx->file_table, i); |
ea64ec02 | 7898 | |
a04b0ac0 PB |
7899 | if (file_slot->file_ptr) { |
7900 | file = (struct file *)(file_slot->file_ptr & FFS_MASK); | |
b60c8dce PB |
7901 | err = io_queue_rsrc_removal(data, up->offset + done, |
7902 | ctx->rsrc_node, file); | |
a5318d3c HD |
7903 | if (err) |
7904 | break; | |
a04b0ac0 | 7905 | file_slot->file_ptr = 0; |
05589553 | 7906 | needs_switch = true; |
c3a31e60 JA |
7907 | } |
7908 | if (fd != -1) { | |
c3a31e60 JA |
7909 | file = fget(fd); |
7910 | if (!file) { | |
7911 | err = -EBADF; | |
7912 | break; | |
7913 | } | |
7914 | /* | |
7915 | * Don't allow io_uring instances to be registered. If | |
7916 | * UNIX isn't enabled, then this causes a reference | |
7917 | * cycle and this instance can never get freed. If UNIX | |
7918 | * is enabled we'll handle it just fine, but there's | |
7919 | * still no point in allowing a ring fd as it doesn't | |
7920 | * support regular read/write anyway. | |
7921 | */ | |
7922 | if (file->f_op == &io_uring_fops) { | |
7923 | fput(file); | |
7924 | err = -EBADF; | |
7925 | break; | |
7926 | } | |
2d091d62 | 7927 | *io_get_tag_slot(data, up->offset + done) = tag; |
9a321c98 | 7928 | io_fixed_file_set(file_slot, file); |
c3a31e60 | 7929 | err = io_sqe_file_register(ctx, file, i); |
f3bd9dae | 7930 | if (err) { |
a04b0ac0 | 7931 | file_slot->file_ptr = 0; |
f3bd9dae | 7932 | fput(file); |
c3a31e60 | 7933 | break; |
f3bd9dae | 7934 | } |
c3a31e60 | 7935 | } |
05f3fb3c JA |
7936 | } |
7937 | ||
a7f0ed5a PB |
7938 | if (needs_switch) |
7939 | io_rsrc_node_switch(ctx, data); | |
c3a31e60 JA |
7940 | return done ? done : err; |
7941 | } | |
05589553 | 7942 | |
685fe7fe JA |
7943 | static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx, |
7944 | struct task_struct *task) | |
24369c2e | 7945 | { |
e941894e | 7946 | struct io_wq_hash *hash; |
24369c2e | 7947 | struct io_wq_data data; |
24369c2e | 7948 | unsigned int concurrency; |
24369c2e | 7949 | |
362a9e65 | 7950 | mutex_lock(&ctx->uring_lock); |
e941894e JA |
7951 | hash = ctx->hash_map; |
7952 | if (!hash) { | |
7953 | hash = kzalloc(sizeof(*hash), GFP_KERNEL); | |
362a9e65 YY |
7954 | if (!hash) { |
7955 | mutex_unlock(&ctx->uring_lock); | |
e941894e | 7956 | return ERR_PTR(-ENOMEM); |
362a9e65 | 7957 | } |
e941894e JA |
7958 | refcount_set(&hash->refs, 1); |
7959 | init_waitqueue_head(&hash->wait); | |
7960 | ctx->hash_map = hash; | |
24369c2e | 7961 | } |
362a9e65 | 7962 | mutex_unlock(&ctx->uring_lock); |
24369c2e | 7963 | |
e941894e | 7964 | data.hash = hash; |
685fe7fe | 7965 | data.task = task; |
ebc11b6c | 7966 | data.free_work = io_wq_free_work; |
f5fa38c5 | 7967 | data.do_work = io_wq_submit_work; |
24369c2e | 7968 | |
d25e3a3d JA |
7969 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
7970 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
24369c2e | 7971 | |
5aa75ed5 | 7972 | return io_wq_create(concurrency, &data); |
24369c2e PB |
7973 | } |
7974 | ||
5aa75ed5 JA |
7975 | static int io_uring_alloc_task_context(struct task_struct *task, |
7976 | struct io_ring_ctx *ctx) | |
0f212204 JA |
7977 | { |
7978 | struct io_uring_task *tctx; | |
d8a6df10 | 7979 | int ret; |
0f212204 | 7980 | |
09899b19 | 7981 | tctx = kzalloc(sizeof(*tctx), GFP_KERNEL); |
0f212204 JA |
7982 | if (unlikely(!tctx)) |
7983 | return -ENOMEM; | |
7984 | ||
d8a6df10 JA |
7985 | ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL); |
7986 | if (unlikely(ret)) { | |
7987 | kfree(tctx); | |
7988 | return ret; | |
7989 | } | |
7990 | ||
685fe7fe | 7991 | tctx->io_wq = io_init_wq_offload(ctx, task); |
5aa75ed5 JA |
7992 | if (IS_ERR(tctx->io_wq)) { |
7993 | ret = PTR_ERR(tctx->io_wq); | |
7994 | percpu_counter_destroy(&tctx->inflight); | |
7995 | kfree(tctx); | |
7996 | return ret; | |
7997 | } | |
7998 | ||
0f212204 JA |
7999 | xa_init(&tctx->xa); |
8000 | init_waitqueue_head(&tctx->wait); | |
fdaf083c | 8001 | atomic_set(&tctx->in_idle, 0); |
b303fe2e | 8002 | atomic_set(&tctx->inflight_tracked, 0); |
0f212204 | 8003 | task->io_uring = tctx; |
7cbf1722 JA |
8004 | spin_lock_init(&tctx->task_lock); |
8005 | INIT_WQ_LIST(&tctx->task_list); | |
7cbf1722 | 8006 | init_task_work(&tctx->task_work, tctx_task_work); |
0f212204 JA |
8007 | return 0; |
8008 | } | |
8009 | ||
8010 | void __io_uring_free(struct task_struct *tsk) | |
8011 | { | |
8012 | struct io_uring_task *tctx = tsk->io_uring; | |
8013 | ||
8014 | WARN_ON_ONCE(!xa_empty(&tctx->xa)); | |
ef8eaa4e | 8015 | WARN_ON_ONCE(tctx->io_wq); |
09899b19 | 8016 | WARN_ON_ONCE(tctx->cached_refs); |
ef8eaa4e | 8017 | |
d8a6df10 | 8018 | percpu_counter_destroy(&tctx->inflight); |
0f212204 JA |
8019 | kfree(tctx); |
8020 | tsk->io_uring = NULL; | |
8021 | } | |
8022 | ||
7e84e1c7 SG |
8023 | static int io_sq_offload_create(struct io_ring_ctx *ctx, |
8024 | struct io_uring_params *p) | |
2b188cc1 JA |
8025 | { |
8026 | int ret; | |
8027 | ||
d25e3a3d JA |
8028 | /* Retain compatibility with failing for an invalid attach attempt */ |
8029 | if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) == | |
8030 | IORING_SETUP_ATTACH_WQ) { | |
8031 | struct fd f; | |
8032 | ||
8033 | f = fdget(p->wq_fd); | |
8034 | if (!f.file) | |
8035 | return -ENXIO; | |
0cc936f7 JA |
8036 | if (f.file->f_op != &io_uring_fops) { |
8037 | fdput(f); | |
f2a48dd0 | 8038 | return -EINVAL; |
0cc936f7 JA |
8039 | } |
8040 | fdput(f); | |
d25e3a3d | 8041 | } |
6c271ce2 | 8042 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
46fe18b1 | 8043 | struct task_struct *tsk; |
534ca6d6 | 8044 | struct io_sq_data *sqd; |
26984fbf | 8045 | bool attached; |
534ca6d6 | 8046 | |
26984fbf | 8047 | sqd = io_get_sq_data(p, &attached); |
534ca6d6 JA |
8048 | if (IS_ERR(sqd)) { |
8049 | ret = PTR_ERR(sqd); | |
8050 | goto err; | |
8051 | } | |
69fb2131 | 8052 | |
7c30f36a | 8053 | ctx->sq_creds = get_current_cred(); |
534ca6d6 | 8054 | ctx->sq_data = sqd; |
917257da JA |
8055 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
8056 | if (!ctx->sq_thread_idle) | |
8057 | ctx->sq_thread_idle = HZ; | |
8058 | ||
78d7f6ba | 8059 | io_sq_thread_park(sqd); |
de75a3d3 PB |
8060 | list_add(&ctx->sqd_list, &sqd->ctx_list); |
8061 | io_sqd_update_thread_idle(sqd); | |
26984fbf | 8062 | /* don't attach to a dying SQPOLL thread, would be racy */ |
f2a48dd0 | 8063 | ret = (attached && !sqd->thread) ? -ENXIO : 0; |
78d7f6ba PB |
8064 | io_sq_thread_unpark(sqd); |
8065 | ||
de75a3d3 PB |
8066 | if (ret < 0) |
8067 | goto err; | |
8068 | if (attached) | |
5aa75ed5 | 8069 | return 0; |
aa06165d | 8070 | |
6c271ce2 | 8071 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 8072 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 8073 | |
917257da | 8074 | ret = -EINVAL; |
f2a48dd0 | 8075 | if (cpu >= nr_cpu_ids || !cpu_online(cpu)) |
e8f98f24 | 8076 | goto err_sqpoll; |
37d1e2e3 | 8077 | sqd->sq_cpu = cpu; |
6c271ce2 | 8078 | } else { |
37d1e2e3 | 8079 | sqd->sq_cpu = -1; |
6c271ce2 | 8080 | } |
37d1e2e3 JA |
8081 | |
8082 | sqd->task_pid = current->pid; | |
5c2469e0 | 8083 | sqd->task_tgid = current->tgid; |
46fe18b1 JA |
8084 | tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE); |
8085 | if (IS_ERR(tsk)) { | |
8086 | ret = PTR_ERR(tsk); | |
e8f98f24 | 8087 | goto err_sqpoll; |
6c271ce2 | 8088 | } |
97a73a0f | 8089 | |
46fe18b1 | 8090 | sqd->thread = tsk; |
97a73a0f | 8091 | ret = io_uring_alloc_task_context(tsk, ctx); |
46fe18b1 | 8092 | wake_up_new_task(tsk); |
0f212204 JA |
8093 | if (ret) |
8094 | goto err; | |
6c271ce2 JA |
8095 | } else if (p->flags & IORING_SETUP_SQ_AFF) { |
8096 | /* Can't have SQ_AFF without SQPOLL */ | |
8097 | ret = -EINVAL; | |
8098 | goto err; | |
8099 | } | |
8100 | ||
2b188cc1 | 8101 | return 0; |
f2a48dd0 PB |
8102 | err_sqpoll: |
8103 | complete(&ctx->sq_data->exited); | |
2b188cc1 | 8104 | err: |
37d1e2e3 | 8105 | io_sq_thread_finish(ctx); |
2b188cc1 JA |
8106 | return ret; |
8107 | } | |
8108 | ||
a087e2b5 BM |
8109 | static inline void __io_unaccount_mem(struct user_struct *user, |
8110 | unsigned long nr_pages) | |
2b188cc1 JA |
8111 | { |
8112 | atomic_long_sub(nr_pages, &user->locked_vm); | |
8113 | } | |
8114 | ||
a087e2b5 BM |
8115 | static inline int __io_account_mem(struct user_struct *user, |
8116 | unsigned long nr_pages) | |
2b188cc1 JA |
8117 | { |
8118 | unsigned long page_limit, cur_pages, new_pages; | |
8119 | ||
8120 | /* Don't allow more pages than we can safely lock */ | |
8121 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
8122 | ||
8123 | do { | |
8124 | cur_pages = atomic_long_read(&user->locked_vm); | |
8125 | new_pages = cur_pages + nr_pages; | |
8126 | if (new_pages > page_limit) | |
8127 | return -ENOMEM; | |
8128 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
8129 | new_pages) != cur_pages); | |
8130 | ||
8131 | return 0; | |
8132 | } | |
8133 | ||
26bfa89e | 8134 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) |
a087e2b5 | 8135 | { |
62e398be | 8136 | if (ctx->user) |
a087e2b5 | 8137 | __io_unaccount_mem(ctx->user, nr_pages); |
30975825 | 8138 | |
26bfa89e JA |
8139 | if (ctx->mm_account) |
8140 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); | |
a087e2b5 BM |
8141 | } |
8142 | ||
26bfa89e | 8143 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) |
a087e2b5 | 8144 | { |
30975825 BM |
8145 | int ret; |
8146 | ||
62e398be | 8147 | if (ctx->user) { |
30975825 BM |
8148 | ret = __io_account_mem(ctx->user, nr_pages); |
8149 | if (ret) | |
8150 | return ret; | |
8151 | } | |
8152 | ||
26bfa89e JA |
8153 | if (ctx->mm_account) |
8154 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); | |
a087e2b5 BM |
8155 | |
8156 | return 0; | |
8157 | } | |
8158 | ||
2b188cc1 JA |
8159 | static void io_mem_free(void *ptr) |
8160 | { | |
52e04ef4 MR |
8161 | struct page *page; |
8162 | ||
8163 | if (!ptr) | |
8164 | return; | |
2b188cc1 | 8165 | |
52e04ef4 | 8166 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
8167 | if (put_page_testzero(page)) |
8168 | free_compound_page(page); | |
8169 | } | |
8170 | ||
8171 | static void *io_mem_alloc(size_t size) | |
8172 | { | |
8173 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
26bfa89e | 8174 | __GFP_NORETRY | __GFP_ACCOUNT; |
2b188cc1 JA |
8175 | |
8176 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
8177 | } | |
8178 | ||
75b28aff HV |
8179 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
8180 | size_t *sq_offset) | |
8181 | { | |
8182 | struct io_rings *rings; | |
8183 | size_t off, sq_array_size; | |
8184 | ||
8185 | off = struct_size(rings, cqes, cq_entries); | |
8186 | if (off == SIZE_MAX) | |
8187 | return SIZE_MAX; | |
8188 | ||
8189 | #ifdef CONFIG_SMP | |
8190 | off = ALIGN(off, SMP_CACHE_BYTES); | |
8191 | if (off == 0) | |
8192 | return SIZE_MAX; | |
8193 | #endif | |
8194 | ||
b36200f5 DV |
8195 | if (sq_offset) |
8196 | *sq_offset = off; | |
8197 | ||
75b28aff HV |
8198 | sq_array_size = array_size(sizeof(u32), sq_entries); |
8199 | if (sq_array_size == SIZE_MAX) | |
8200 | return SIZE_MAX; | |
8201 | ||
8202 | if (check_add_overflow(off, sq_array_size, &off)) | |
8203 | return SIZE_MAX; | |
8204 | ||
75b28aff HV |
8205 | return off; |
8206 | } | |
8207 | ||
41edf1a5 | 8208 | static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot) |
7f61a1e9 | 8209 | { |
41edf1a5 | 8210 | struct io_mapped_ubuf *imu = *slot; |
7f61a1e9 PB |
8211 | unsigned int i; |
8212 | ||
6224843d PB |
8213 | if (imu != ctx->dummy_ubuf) { |
8214 | for (i = 0; i < imu->nr_bvecs; i++) | |
8215 | unpin_user_page(imu->bvec[i].bv_page); | |
8216 | if (imu->acct_pages) | |
8217 | io_unaccount_mem(ctx, imu->acct_pages); | |
8218 | kvfree(imu); | |
8219 | } | |
41edf1a5 | 8220 | *slot = NULL; |
7f61a1e9 PB |
8221 | } |
8222 | ||
bd54b6fe | 8223 | static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) |
edafccee | 8224 | { |
634d00df PB |
8225 | io_buffer_unmap(ctx, &prsrc->buf); |
8226 | prsrc->buf = NULL; | |
bd54b6fe | 8227 | } |
edafccee | 8228 | |
bd54b6fe BM |
8229 | static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) |
8230 | { | |
8231 | unsigned int i; | |
edafccee | 8232 | |
7f61a1e9 PB |
8233 | for (i = 0; i < ctx->nr_user_bufs; i++) |
8234 | io_buffer_unmap(ctx, &ctx->user_bufs[i]); | |
edafccee | 8235 | kfree(ctx->user_bufs); |
bb6659cc | 8236 | io_rsrc_data_free(ctx->buf_data); |
edafccee | 8237 | ctx->user_bufs = NULL; |
bd54b6fe | 8238 | ctx->buf_data = NULL; |
edafccee | 8239 | ctx->nr_user_bufs = 0; |
bd54b6fe BM |
8240 | } |
8241 | ||
0a96bbe4 | 8242 | static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) |
edafccee | 8243 | { |
bd54b6fe | 8244 | int ret; |
edafccee | 8245 | |
bd54b6fe | 8246 | if (!ctx->buf_data) |
edafccee JA |
8247 | return -ENXIO; |
8248 | ||
bd54b6fe BM |
8249 | ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); |
8250 | if (!ret) | |
8251 | __io_sqe_buffers_unregister(ctx); | |
8252 | return ret; | |
edafccee JA |
8253 | } |
8254 | ||
8255 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
8256 | void __user *arg, unsigned index) | |
8257 | { | |
8258 | struct iovec __user *src; | |
8259 | ||
8260 | #ifdef CONFIG_COMPAT | |
8261 | if (ctx->compat) { | |
8262 | struct compat_iovec __user *ciovs; | |
8263 | struct compat_iovec ciov; | |
8264 | ||
8265 | ciovs = (struct compat_iovec __user *) arg; | |
8266 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
8267 | return -EFAULT; | |
8268 | ||
d55e5f5b | 8269 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
8270 | dst->iov_len = ciov.iov_len; |
8271 | return 0; | |
8272 | } | |
8273 | #endif | |
8274 | src = (struct iovec __user *) arg; | |
8275 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
8276 | return -EFAULT; | |
8277 | return 0; | |
8278 | } | |
8279 | ||
de293938 JA |
8280 | /* |
8281 | * Not super efficient, but this is just a registration time. And we do cache | |
8282 | * the last compound head, so generally we'll only do a full search if we don't | |
8283 | * match that one. | |
8284 | * | |
8285 | * We check if the given compound head page has already been accounted, to | |
8286 | * avoid double accounting it. This allows us to account the full size of the | |
8287 | * page, not just the constituent pages of a huge page. | |
8288 | */ | |
8289 | static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, | |
8290 | int nr_pages, struct page *hpage) | |
8291 | { | |
8292 | int i, j; | |
8293 | ||
8294 | /* check current page array */ | |
8295 | for (i = 0; i < nr_pages; i++) { | |
8296 | if (!PageCompound(pages[i])) | |
8297 | continue; | |
8298 | if (compound_head(pages[i]) == hpage) | |
8299 | return true; | |
8300 | } | |
8301 | ||
8302 | /* check previously registered pages */ | |
8303 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
41edf1a5 | 8304 | struct io_mapped_ubuf *imu = ctx->user_bufs[i]; |
de293938 JA |
8305 | |
8306 | for (j = 0; j < imu->nr_bvecs; j++) { | |
8307 | if (!PageCompound(imu->bvec[j].bv_page)) | |
8308 | continue; | |
8309 | if (compound_head(imu->bvec[j].bv_page) == hpage) | |
8310 | return true; | |
8311 | } | |
8312 | } | |
8313 | ||
8314 | return false; | |
8315 | } | |
8316 | ||
8317 | static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, | |
8318 | int nr_pages, struct io_mapped_ubuf *imu, | |
8319 | struct page **last_hpage) | |
8320 | { | |
8321 | int i, ret; | |
8322 | ||
216e5835 | 8323 | imu->acct_pages = 0; |
de293938 JA |
8324 | for (i = 0; i < nr_pages; i++) { |
8325 | if (!PageCompound(pages[i])) { | |
8326 | imu->acct_pages++; | |
8327 | } else { | |
8328 | struct page *hpage; | |
8329 | ||
8330 | hpage = compound_head(pages[i]); | |
8331 | if (hpage == *last_hpage) | |
8332 | continue; | |
8333 | *last_hpage = hpage; | |
8334 | if (headpage_already_acct(ctx, pages, i, hpage)) | |
8335 | continue; | |
8336 | imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; | |
8337 | } | |
8338 | } | |
8339 | ||
8340 | if (!imu->acct_pages) | |
8341 | return 0; | |
8342 | ||
26bfa89e | 8343 | ret = io_account_mem(ctx, imu->acct_pages); |
de293938 JA |
8344 | if (ret) |
8345 | imu->acct_pages = 0; | |
8346 | return ret; | |
8347 | } | |
8348 | ||
0a96bbe4 | 8349 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, |
41edf1a5 | 8350 | struct io_mapped_ubuf **pimu, |
0a96bbe4 | 8351 | struct page **last_hpage) |
edafccee | 8352 | { |
41edf1a5 | 8353 | struct io_mapped_ubuf *imu = NULL; |
edafccee JA |
8354 | struct vm_area_struct **vmas = NULL; |
8355 | struct page **pages = NULL; | |
0a96bbe4 BM |
8356 | unsigned long off, start, end, ubuf; |
8357 | size_t size; | |
8358 | int ret, pret, nr_pages, i; | |
8359 | ||
6224843d PB |
8360 | if (!iov->iov_base) { |
8361 | *pimu = ctx->dummy_ubuf; | |
8362 | return 0; | |
8363 | } | |
8364 | ||
0a96bbe4 BM |
8365 | ubuf = (unsigned long) iov->iov_base; |
8366 | end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
8367 | start = ubuf >> PAGE_SHIFT; | |
8368 | nr_pages = end - start; | |
8369 | ||
41edf1a5 | 8370 | *pimu = NULL; |
0a96bbe4 BM |
8371 | ret = -ENOMEM; |
8372 | ||
8373 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
8374 | if (!pages) | |
8375 | goto done; | |
8376 | ||
8377 | vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), | |
8378 | GFP_KERNEL); | |
8379 | if (!vmas) | |
8380 | goto done; | |
edafccee | 8381 | |
41edf1a5 | 8382 | imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); |
a2b4198c | 8383 | if (!imu) |
0a96bbe4 BM |
8384 | goto done; |
8385 | ||
8386 | ret = 0; | |
8387 | mmap_read_lock(current->mm); | |
8388 | pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
8389 | pages, vmas); | |
8390 | if (pret == nr_pages) { | |
8391 | /* don't support file backed memory */ | |
8392 | for (i = 0; i < nr_pages; i++) { | |
8393 | struct vm_area_struct *vma = vmas[i]; | |
8394 | ||
40dad765 PB |
8395 | if (vma_is_shmem(vma)) |
8396 | continue; | |
0a96bbe4 BM |
8397 | if (vma->vm_file && |
8398 | !is_file_hugepages(vma->vm_file)) { | |
8399 | ret = -EOPNOTSUPP; | |
8400 | break; | |
8401 | } | |
8402 | } | |
8403 | } else { | |
8404 | ret = pret < 0 ? pret : -EFAULT; | |
8405 | } | |
8406 | mmap_read_unlock(current->mm); | |
8407 | if (ret) { | |
8408 | /* | |
8409 | * if we did partial map, or found file backed vmas, | |
8410 | * release any pages we did get | |
8411 | */ | |
8412 | if (pret > 0) | |
8413 | unpin_user_pages(pages, pret); | |
0a96bbe4 BM |
8414 | goto done; |
8415 | } | |
8416 | ||
8417 | ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage); | |
8418 | if (ret) { | |
8419 | unpin_user_pages(pages, pret); | |
0a96bbe4 BM |
8420 | goto done; |
8421 | } | |
8422 | ||
8423 | off = ubuf & ~PAGE_MASK; | |
8424 | size = iov->iov_len; | |
8425 | for (i = 0; i < nr_pages; i++) { | |
8426 | size_t vec_len; | |
8427 | ||
8428 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
8429 | imu->bvec[i].bv_page = pages[i]; | |
8430 | imu->bvec[i].bv_len = vec_len; | |
8431 | imu->bvec[i].bv_offset = off; | |
8432 | off = 0; | |
8433 | size -= vec_len; | |
8434 | } | |
8435 | /* store original address for later verification */ | |
8436 | imu->ubuf = ubuf; | |
4751f53d | 8437 | imu->ubuf_end = ubuf + iov->iov_len; |
0a96bbe4 | 8438 | imu->nr_bvecs = nr_pages; |
41edf1a5 | 8439 | *pimu = imu; |
0a96bbe4 BM |
8440 | ret = 0; |
8441 | done: | |
41edf1a5 PB |
8442 | if (ret) |
8443 | kvfree(imu); | |
0a96bbe4 BM |
8444 | kvfree(pages); |
8445 | kvfree(vmas); | |
8446 | return ret; | |
8447 | } | |
8448 | ||
2b358604 | 8449 | static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) |
0a96bbe4 | 8450 | { |
87094465 PB |
8451 | ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); |
8452 | return ctx->user_bufs ? 0 : -ENOMEM; | |
2b358604 | 8453 | } |
edafccee | 8454 | |
2b358604 BM |
8455 | static int io_buffer_validate(struct iovec *iov) |
8456 | { | |
50e96989 PB |
8457 | unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1); |
8458 | ||
2b358604 BM |
8459 | /* |
8460 | * Don't impose further limits on the size and buffer | |
8461 | * constraints here, we'll -EINVAL later when IO is | |
8462 | * submitted if they are wrong. | |
8463 | */ | |
6224843d PB |
8464 | if (!iov->iov_base) |
8465 | return iov->iov_len ? -EFAULT : 0; | |
8466 | if (!iov->iov_len) | |
2b358604 | 8467 | return -EFAULT; |
edafccee | 8468 | |
2b358604 BM |
8469 | /* arbitrary limit, but we need something */ |
8470 | if (iov->iov_len > SZ_1G) | |
8471 | return -EFAULT; | |
edafccee | 8472 | |
50e96989 PB |
8473 | if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp)) |
8474 | return -EOVERFLOW; | |
8475 | ||
2b358604 BM |
8476 | return 0; |
8477 | } | |
edafccee | 8478 | |
2b358604 | 8479 | static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, |
634d00df | 8480 | unsigned int nr_args, u64 __user *tags) |
2b358604 | 8481 | { |
bd54b6fe BM |
8482 | struct page *last_hpage = NULL; |
8483 | struct io_rsrc_data *data; | |
2b358604 BM |
8484 | int i, ret; |
8485 | struct iovec iov; | |
edafccee | 8486 | |
87094465 PB |
8487 | if (ctx->user_bufs) |
8488 | return -EBUSY; | |
489809e2 | 8489 | if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) |
87094465 | 8490 | return -EINVAL; |
bd54b6fe | 8491 | ret = io_rsrc_node_switch_start(ctx); |
2b358604 BM |
8492 | if (ret) |
8493 | return ret; | |
d878c816 PB |
8494 | ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); |
8495 | if (ret) | |
8496 | return ret; | |
bd54b6fe BM |
8497 | ret = io_buffers_map_alloc(ctx, nr_args); |
8498 | if (ret) { | |
bb6659cc | 8499 | io_rsrc_data_free(data); |
bd54b6fe BM |
8500 | return ret; |
8501 | } | |
edafccee | 8502 | |
87094465 | 8503 | for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { |
edafccee JA |
8504 | ret = io_copy_iov(ctx, &iov, arg, i); |
8505 | if (ret) | |
0a96bbe4 | 8506 | break; |
2b358604 BM |
8507 | ret = io_buffer_validate(&iov); |
8508 | if (ret) | |
0a96bbe4 | 8509 | break; |
2d091d62 | 8510 | if (!iov.iov_base && *io_get_tag_slot(data, i)) { |
cf3770e7 CIK |
8511 | ret = -EINVAL; |
8512 | break; | |
8513 | } | |
edafccee | 8514 | |
41edf1a5 PB |
8515 | ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], |
8516 | &last_hpage); | |
0a96bbe4 BM |
8517 | if (ret) |
8518 | break; | |
edafccee | 8519 | } |
0a96bbe4 | 8520 | |
bd54b6fe | 8521 | WARN_ON_ONCE(ctx->buf_data); |
0a96bbe4 | 8522 | |
bd54b6fe BM |
8523 | ctx->buf_data = data; |
8524 | if (ret) | |
8525 | __io_sqe_buffers_unregister(ctx); | |
8526 | else | |
8527 | io_rsrc_node_switch(ctx, NULL); | |
edafccee JA |
8528 | return ret; |
8529 | } | |
8530 | ||
634d00df PB |
8531 | static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, |
8532 | struct io_uring_rsrc_update2 *up, | |
8533 | unsigned int nr_args) | |
8534 | { | |
8535 | u64 __user *tags = u64_to_user_ptr(up->tags); | |
8536 | struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); | |
634d00df PB |
8537 | struct page *last_hpage = NULL; |
8538 | bool needs_switch = false; | |
8539 | __u32 done; | |
8540 | int i, err; | |
8541 | ||
8542 | if (!ctx->buf_data) | |
8543 | return -ENXIO; | |
8544 | if (up->offset + nr_args > ctx->nr_user_bufs) | |
8545 | return -EINVAL; | |
8546 | ||
8547 | for (done = 0; done < nr_args; done++) { | |
0b8c0e7c PB |
8548 | struct io_mapped_ubuf *imu; |
8549 | int offset = up->offset + done; | |
634d00df PB |
8550 | u64 tag = 0; |
8551 | ||
8552 | err = io_copy_iov(ctx, &iov, iovs, done); | |
8553 | if (err) | |
8554 | break; | |
8555 | if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { | |
8556 | err = -EFAULT; | |
8557 | break; | |
8558 | } | |
0b8c0e7c PB |
8559 | err = io_buffer_validate(&iov); |
8560 | if (err) | |
8561 | break; | |
cf3770e7 CIK |
8562 | if (!iov.iov_base && tag) { |
8563 | err = -EINVAL; | |
8564 | break; | |
8565 | } | |
0b8c0e7c PB |
8566 | err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); |
8567 | if (err) | |
8568 | break; | |
634d00df | 8569 | |
0b8c0e7c | 8570 | i = array_index_nospec(offset, ctx->nr_user_bufs); |
6224843d | 8571 | if (ctx->user_bufs[i] != ctx->dummy_ubuf) { |
0b8c0e7c PB |
8572 | err = io_queue_rsrc_removal(ctx->buf_data, offset, |
8573 | ctx->rsrc_node, ctx->user_bufs[i]); | |
8574 | if (unlikely(err)) { | |
8575 | io_buffer_unmap(ctx, &imu); | |
634d00df | 8576 | break; |
0b8c0e7c | 8577 | } |
634d00df PB |
8578 | ctx->user_bufs[i] = NULL; |
8579 | needs_switch = true; | |
8580 | } | |
8581 | ||
0b8c0e7c | 8582 | ctx->user_bufs[i] = imu; |
2d091d62 | 8583 | *io_get_tag_slot(ctx->buf_data, offset) = tag; |
634d00df PB |
8584 | } |
8585 | ||
8586 | if (needs_switch) | |
8587 | io_rsrc_node_switch(ctx, ctx->buf_data); | |
8588 | return done ? done : err; | |
8589 | } | |
8590 | ||
9b402849 JA |
8591 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
8592 | { | |
8593 | __s32 __user *fds = arg; | |
8594 | int fd; | |
8595 | ||
8596 | if (ctx->cq_ev_fd) | |
8597 | return -EBUSY; | |
8598 | ||
8599 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
8600 | return -EFAULT; | |
8601 | ||
8602 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
8603 | if (IS_ERR(ctx->cq_ev_fd)) { | |
8604 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
fe7e3257 | 8605 | |
9b402849 JA |
8606 | ctx->cq_ev_fd = NULL; |
8607 | return ret; | |
8608 | } | |
8609 | ||
8610 | return 0; | |
8611 | } | |
8612 | ||
8613 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
8614 | { | |
8615 | if (ctx->cq_ev_fd) { | |
8616 | eventfd_ctx_put(ctx->cq_ev_fd); | |
8617 | ctx->cq_ev_fd = NULL; | |
8618 | return 0; | |
8619 | } | |
8620 | ||
8621 | return -ENXIO; | |
8622 | } | |
8623 | ||
5a2e745d JA |
8624 | static void io_destroy_buffers(struct io_ring_ctx *ctx) |
8625 | { | |
9e15c3a0 JA |
8626 | struct io_buffer *buf; |
8627 | unsigned long index; | |
8628 | ||
8629 | xa_for_each(&ctx->io_buffers, index, buf) | |
8630 | __io_remove_buffers(ctx, buf, index, -1U); | |
5a2e745d JA |
8631 | } |
8632 | ||
7255834e | 8633 | static void io_req_cache_free(struct list_head *list) |
1b4c351f | 8634 | { |
68e68ee6 | 8635 | struct io_kiocb *req, *nxt; |
1b4c351f | 8636 | |
bb943b82 PB |
8637 | list_for_each_entry_safe(req, nxt, list, inflight_entry) { |
8638 | list_del(&req->inflight_entry); | |
1b4c351f JA |
8639 | kmem_cache_free(req_cachep, req); |
8640 | } | |
8641 | } | |
8642 | ||
4010fec4 | 8643 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 8644 | { |
cd0ca2e0 | 8645 | struct io_submit_state *state = &ctx->submit_state; |
bf019da7 | 8646 | |
9a4fdbd8 JA |
8647 | mutex_lock(&ctx->uring_lock); |
8648 | ||
cd0ca2e0 PB |
8649 | if (state->free_reqs) { |
8650 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); | |
8651 | state->free_reqs = 0; | |
8e5c66c4 | 8652 | } |
9a4fdbd8 | 8653 | |
cd0ca2e0 PB |
8654 | io_flush_cached_locked_reqs(ctx, state); |
8655 | io_req_cache_free(&state->free_list); | |
9a4fdbd8 JA |
8656 | mutex_unlock(&ctx->uring_lock); |
8657 | } | |
8658 | ||
43597aac | 8659 | static void io_wait_rsrc_data(struct io_rsrc_data *data) |
2b188cc1 | 8660 | { |
43597aac | 8661 | if (data && !atomic_dec_and_test(&data->refs)) |
bd54b6fe | 8662 | wait_for_completion(&data->done); |
bd54b6fe | 8663 | } |
04fc6c80 | 8664 | |
2b188cc1 JA |
8665 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
8666 | { | |
37d1e2e3 | 8667 | io_sq_thread_finish(ctx); |
2aede0e4 | 8668 | |
37d1e2e3 | 8669 | if (ctx->mm_account) { |
2aede0e4 JA |
8670 | mmdrop(ctx->mm_account); |
8671 | ctx->mm_account = NULL; | |
30975825 | 8672 | } |
def596e9 | 8673 | |
43597aac PB |
8674 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
8675 | io_wait_rsrc_data(ctx->buf_data); | |
8676 | io_wait_rsrc_data(ctx->file_data); | |
8677 | ||
8bad28d8 | 8678 | mutex_lock(&ctx->uring_lock); |
43597aac | 8679 | if (ctx->buf_data) |
bd54b6fe | 8680 | __io_sqe_buffers_unregister(ctx); |
43597aac | 8681 | if (ctx->file_data) |
08480400 | 8682 | __io_sqe_files_unregister(ctx); |
c4ea060e PB |
8683 | if (ctx->rings) |
8684 | __io_cqring_overflow_flush(ctx, true); | |
8bad28d8 | 8685 | mutex_unlock(&ctx->uring_lock); |
9b402849 | 8686 | io_eventfd_unregister(ctx); |
5a2e745d | 8687 | io_destroy_buffers(ctx); |
07db298a PB |
8688 | if (ctx->sq_creds) |
8689 | put_cred(ctx->sq_creds); | |
def596e9 | 8690 | |
a7f0ed5a PB |
8691 | /* there are no registered resources left, nobody uses it */ |
8692 | if (ctx->rsrc_node) | |
8693 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 8694 | if (ctx->rsrc_backup_node) |
b895c9a6 | 8695 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a PB |
8696 | flush_delayed_work(&ctx->rsrc_put_work); |
8697 | ||
8698 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
8699 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 8700 | |
2b188cc1 | 8701 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
8702 | if (ctx->ring_sock) { |
8703 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 8704 | sock_release(ctx->ring_sock); |
355e8d26 | 8705 | } |
2b188cc1 JA |
8706 | #endif |
8707 | ||
75b28aff | 8708 | io_mem_free(ctx->rings); |
2b188cc1 | 8709 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
8710 | |
8711 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 8712 | free_uid(ctx->user); |
4010fec4 | 8713 | io_req_caches_free(ctx); |
e941894e JA |
8714 | if (ctx->hash_map) |
8715 | io_wq_put_hash(ctx->hash_map); | |
78076bb6 | 8716 | kfree(ctx->cancel_hash); |
6224843d | 8717 | kfree(ctx->dummy_ubuf); |
2b188cc1 JA |
8718 | kfree(ctx); |
8719 | } | |
8720 | ||
8721 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
8722 | { | |
8723 | struct io_ring_ctx *ctx = file->private_data; | |
8724 | __poll_t mask = 0; | |
8725 | ||
311997b3 | 8726 | poll_wait(file, &ctx->poll_wait, wait); |
4f7067c3 SB |
8727 | /* |
8728 | * synchronizes with barrier from wq_has_sleeper call in | |
8729 | * io_commit_cqring | |
8730 | */ | |
2b188cc1 | 8731 | smp_rmb(); |
90554200 | 8732 | if (!io_sqring_full(ctx)) |
2b188cc1 | 8733 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
8734 | |
8735 | /* | |
8736 | * Don't flush cqring overflow list here, just do a simple check. | |
8737 | * Otherwise there could possible be ABBA deadlock: | |
8738 | * CPU0 CPU1 | |
8739 | * ---- ---- | |
8740 | * lock(&ctx->uring_lock); | |
8741 | * lock(&ep->mtx); | |
8742 | * lock(&ctx->uring_lock); | |
8743 | * lock(&ep->mtx); | |
8744 | * | |
8745 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
8746 | * pushs them to do the flush. | |
8747 | */ | |
5ed7a37d | 8748 | if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow)) |
2b188cc1 JA |
8749 | mask |= EPOLLIN | EPOLLRDNORM; |
8750 | ||
8751 | return mask; | |
8752 | } | |
8753 | ||
8754 | static int io_uring_fasync(int fd, struct file *file, int on) | |
8755 | { | |
8756 | struct io_ring_ctx *ctx = file->private_data; | |
8757 | ||
8758 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
8759 | } | |
8760 | ||
0bead8cd | 8761 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 8762 | { |
4379bf8b | 8763 | const struct cred *creds; |
071698e1 | 8764 | |
61cf9370 | 8765 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
8766 | if (creds) { |
8767 | put_cred(creds); | |
0bead8cd | 8768 | return 0; |
1e6fa521 | 8769 | } |
0bead8cd YD |
8770 | |
8771 | return -EINVAL; | |
8772 | } | |
8773 | ||
d56d938b PB |
8774 | struct io_tctx_exit { |
8775 | struct callback_head task_work; | |
8776 | struct completion completion; | |
baf186c4 | 8777 | struct io_ring_ctx *ctx; |
d56d938b PB |
8778 | }; |
8779 | ||
8780 | static void io_tctx_exit_cb(struct callback_head *cb) | |
8781 | { | |
8782 | struct io_uring_task *tctx = current->io_uring; | |
8783 | struct io_tctx_exit *work; | |
8784 | ||
8785 | work = container_of(cb, struct io_tctx_exit, task_work); | |
8786 | /* | |
8787 | * When @in_idle, we're in cancellation and it's racy to remove the | |
8788 | * node. It'll be removed by the end of cancellation, just ignore it. | |
8789 | */ | |
8790 | if (!atomic_read(&tctx->in_idle)) | |
eef51daa | 8791 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
8792 | complete(&work->completion); |
8793 | } | |
8794 | ||
28090c13 PB |
8795 | static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
8796 | { | |
8797 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
8798 | ||
8799 | return req->ctx == data; | |
8800 | } | |
8801 | ||
85faa7b8 JA |
8802 | static void io_ring_exit_work(struct work_struct *work) |
8803 | { | |
d56d938b | 8804 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 8805 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 8806 | unsigned long interval = HZ / 20; |
d56d938b PB |
8807 | struct io_tctx_exit exit; |
8808 | struct io_tctx_node *node; | |
8809 | int ret; | |
85faa7b8 | 8810 | |
56952e91 JA |
8811 | /* |
8812 | * If we're doing polled IO and end up having requests being | |
8813 | * submitted async (out-of-line), then completions can come in while | |
8814 | * we're waiting for refs to drop. We need to reap these manually, | |
8815 | * as nobody else will be looking for them. | |
8816 | */ | |
b2edc0a7 | 8817 | do { |
3dd0c97a | 8818 | io_uring_try_cancel_requests(ctx, NULL, true); |
28090c13 PB |
8819 | if (ctx->sq_data) { |
8820 | struct io_sq_data *sqd = ctx->sq_data; | |
8821 | struct task_struct *tsk; | |
8822 | ||
8823 | io_sq_thread_park(sqd); | |
8824 | tsk = sqd->thread; | |
8825 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
8826 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
8827 | io_cancel_ctx_cb, ctx, true); | |
8828 | io_sq_thread_unpark(sqd); | |
8829 | } | |
b5bb3a24 | 8830 | |
58d3be2c PB |
8831 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
8832 | /* there is little hope left, don't run it too often */ | |
8833 | interval = HZ * 60; | |
8834 | } | |
8835 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 8836 | |
7f00651a PB |
8837 | init_completion(&exit.completion); |
8838 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
8839 | exit.ctx = ctx; | |
89b5066e PB |
8840 | /* |
8841 | * Some may use context even when all refs and requests have been put, | |
8842 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 8843 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
8844 | * this lock/unlock section also waits them to finish. |
8845 | */ | |
d56d938b PB |
8846 | mutex_lock(&ctx->uring_lock); |
8847 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
8848 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
8849 | ||
d56d938b PB |
8850 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
8851 | ctx_node); | |
7f00651a PB |
8852 | /* don't spin on a single task if cancellation failed */ |
8853 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
8854 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
8855 | if (WARN_ON_ONCE(ret)) | |
8856 | continue; | |
8857 | wake_up_process(node->task); | |
8858 | ||
8859 | mutex_unlock(&ctx->uring_lock); | |
8860 | wait_for_completion(&exit.completion); | |
d56d938b PB |
8861 | mutex_lock(&ctx->uring_lock); |
8862 | } | |
8863 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
8864 | spin_lock(&ctx->completion_lock); |
8865 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 8866 | |
85faa7b8 JA |
8867 | io_ring_ctx_free(ctx); |
8868 | } | |
8869 | ||
80c4cbdb PB |
8870 | /* Returns true if we found and killed one or more timeouts */ |
8871 | static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
3dd0c97a | 8872 | bool cancel_all) |
80c4cbdb PB |
8873 | { |
8874 | struct io_kiocb *req, *tmp; | |
8875 | int canceled = 0; | |
8876 | ||
79ebeaee JA |
8877 | spin_lock(&ctx->completion_lock); |
8878 | spin_lock_irq(&ctx->timeout_lock); | |
80c4cbdb | 8879 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) { |
3dd0c97a | 8880 | if (io_match_task(req, tsk, cancel_all)) { |
80c4cbdb PB |
8881 | io_kill_timeout(req, -ECANCELED); |
8882 | canceled++; | |
8883 | } | |
8884 | } | |
79ebeaee | 8885 | spin_unlock_irq(&ctx->timeout_lock); |
51520426 PB |
8886 | if (canceled != 0) |
8887 | io_commit_cqring(ctx); | |
79ebeaee | 8888 | spin_unlock(&ctx->completion_lock); |
80c4cbdb PB |
8889 | if (canceled != 0) |
8890 | io_cqring_ev_posted(ctx); | |
8891 | return canceled != 0; | |
8892 | } | |
8893 | ||
2b188cc1 JA |
8894 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
8895 | { | |
61cf9370 MWO |
8896 | unsigned long index; |
8897 | struct creds *creds; | |
8898 | ||
2b188cc1 JA |
8899 | mutex_lock(&ctx->uring_lock); |
8900 | percpu_ref_kill(&ctx->refs); | |
634578f8 | 8901 | if (ctx->rings) |
6c2450ae | 8902 | __io_cqring_overflow_flush(ctx, true); |
61cf9370 MWO |
8903 | xa_for_each(&ctx->personalities, index, creds) |
8904 | io_unregister_personality(ctx, index); | |
2b188cc1 JA |
8905 | mutex_unlock(&ctx->uring_lock); |
8906 | ||
3dd0c97a PB |
8907 | io_kill_timeouts(ctx, NULL, true); |
8908 | io_poll_remove_all(ctx, NULL, true); | |
561fb04a | 8909 | |
15dff286 | 8910 | /* if we failed setting up the ctx, we might not have any rings */ |
b2edc0a7 | 8911 | io_iopoll_try_reap_events(ctx); |
309fc03a | 8912 | |
85faa7b8 | 8913 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
8914 | /* |
8915 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
8916 | * if we're exiting a ton of rings at the same time. It just adds | |
8917 | * noise and overhead, there's no discernable change in runtime | |
8918 | * over using system_wq. | |
8919 | */ | |
8920 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
8921 | } |
8922 | ||
8923 | static int io_uring_release(struct inode *inode, struct file *file) | |
8924 | { | |
8925 | struct io_ring_ctx *ctx = file->private_data; | |
8926 | ||
8927 | file->private_data = NULL; | |
8928 | io_ring_ctx_wait_and_kill(ctx); | |
8929 | return 0; | |
8930 | } | |
8931 | ||
f6edbabb PB |
8932 | struct io_task_cancel { |
8933 | struct task_struct *task; | |
3dd0c97a | 8934 | bool all; |
f6edbabb | 8935 | }; |
f254ac04 | 8936 | |
f6edbabb | 8937 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 8938 | { |
9a472ef7 | 8939 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 8940 | struct io_task_cancel *cancel = data; |
9a472ef7 PB |
8941 | bool ret; |
8942 | ||
3dd0c97a | 8943 | if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) { |
9a472ef7 PB |
8944 | struct io_ring_ctx *ctx = req->ctx; |
8945 | ||
8946 | /* protect against races with linked timeouts */ | |
79ebeaee | 8947 | spin_lock(&ctx->completion_lock); |
3dd0c97a | 8948 | ret = io_match_task(req, cancel->task, cancel->all); |
79ebeaee | 8949 | spin_unlock(&ctx->completion_lock); |
9a472ef7 | 8950 | } else { |
3dd0c97a | 8951 | ret = io_match_task(req, cancel->task, cancel->all); |
9a472ef7 PB |
8952 | } |
8953 | return ret; | |
b711d4ea JA |
8954 | } |
8955 | ||
e1915f76 | 8956 | static bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3dd0c97a | 8957 | struct task_struct *task, bool cancel_all) |
b7ddce3c | 8958 | { |
e1915f76 | 8959 | struct io_defer_entry *de; |
b7ddce3c PB |
8960 | LIST_HEAD(list); |
8961 | ||
79ebeaee | 8962 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 8963 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
3dd0c97a | 8964 | if (io_match_task(de->req, task, cancel_all)) { |
b7ddce3c PB |
8965 | list_cut_position(&list, &ctx->defer_list, &de->list); |
8966 | break; | |
8967 | } | |
8968 | } | |
79ebeaee | 8969 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
8970 | if (list_empty(&list)) |
8971 | return false; | |
b7ddce3c PB |
8972 | |
8973 | while (!list_empty(&list)) { | |
8974 | de = list_first_entry(&list, struct io_defer_entry, list); | |
8975 | list_del_init(&de->list); | |
f41db273 | 8976 | io_req_complete_failed(de->req, -ECANCELED); |
b7ddce3c PB |
8977 | kfree(de); |
8978 | } | |
e1915f76 | 8979 | return true; |
b7ddce3c PB |
8980 | } |
8981 | ||
1b00764f PB |
8982 | static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
8983 | { | |
8984 | struct io_tctx_node *node; | |
8985 | enum io_wq_cancel cret; | |
8986 | bool ret = false; | |
8987 | ||
8988 | mutex_lock(&ctx->uring_lock); | |
8989 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
8990 | struct io_uring_task *tctx = node->task->io_uring; | |
8991 | ||
8992 | /* | |
8993 | * io_wq will stay alive while we hold uring_lock, because it's | |
8994 | * killed after ctx nodes, which requires to take the lock. | |
8995 | */ | |
8996 | if (!tctx || !tctx->io_wq) | |
8997 | continue; | |
8998 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
8999 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
9000 | } | |
9001 | mutex_unlock(&ctx->uring_lock); | |
9002 | ||
9003 | return ret; | |
9004 | } | |
9005 | ||
9936c7c2 PB |
9006 | static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9007 | struct task_struct *task, | |
3dd0c97a | 9008 | bool cancel_all) |
9936c7c2 | 9009 | { |
3dd0c97a | 9010 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 9011 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
9936c7c2 PB |
9012 | |
9013 | while (1) { | |
9014 | enum io_wq_cancel cret; | |
9015 | bool ret = false; | |
9016 | ||
1b00764f PB |
9017 | if (!task) { |
9018 | ret |= io_uring_try_cancel_iowq(ctx); | |
9019 | } else if (tctx && tctx->io_wq) { | |
9020 | /* | |
9021 | * Cancels requests of all rings, not only @ctx, but | |
9022 | * it's fine as the task is in exit/exec. | |
9023 | */ | |
5aa75ed5 | 9024 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, |
9936c7c2 PB |
9025 | &cancel, true); |
9026 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
9027 | } | |
9028 | ||
9029 | /* SQPOLL thread does its own polling */ | |
3dd0c97a | 9030 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || |
d052d1d6 | 9031 | (ctx->sq_data && ctx->sq_data->thread == current)) { |
9936c7c2 PB |
9032 | while (!list_empty_careful(&ctx->iopoll_list)) { |
9033 | io_iopoll_try_reap_events(ctx); | |
9034 | ret = true; | |
9035 | } | |
9036 | } | |
9037 | ||
3dd0c97a PB |
9038 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
9039 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
9040 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
e5dc480d PB |
9041 | if (task) |
9042 | ret |= io_run_task_work(); | |
9936c7c2 PB |
9043 | if (!ret) |
9044 | break; | |
9045 | cond_resched(); | |
9046 | } | |
9047 | } | |
9048 | ||
eef51daa | 9049 | static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx) |
0f212204 | 9050 | { |
236434c3 | 9051 | struct io_uring_task *tctx = current->io_uring; |
13bf43f5 | 9052 | struct io_tctx_node *node; |
a528b04e | 9053 | int ret; |
236434c3 MWO |
9054 | |
9055 | if (unlikely(!tctx)) { | |
5aa75ed5 | 9056 | ret = io_uring_alloc_task_context(current, ctx); |
0f212204 JA |
9057 | if (unlikely(ret)) |
9058 | return ret; | |
236434c3 | 9059 | tctx = current->io_uring; |
0f212204 | 9060 | } |
cf27f3b1 PB |
9061 | if (!xa_load(&tctx->xa, (unsigned long)ctx)) { |
9062 | node = kmalloc(sizeof(*node), GFP_KERNEL); | |
9063 | if (!node) | |
9064 | return -ENOMEM; | |
9065 | node->ctx = ctx; | |
9066 | node->task = current; | |
13bf43f5 | 9067 | |
cf27f3b1 PB |
9068 | ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx, |
9069 | node, GFP_KERNEL)); | |
9070 | if (ret) { | |
9071 | kfree(node); | |
9072 | return ret; | |
0f212204 | 9073 | } |
cf27f3b1 PB |
9074 | |
9075 | mutex_lock(&ctx->uring_lock); | |
9076 | list_add(&node->ctx_node, &ctx->tctx_list); | |
9077 | mutex_unlock(&ctx->uring_lock); | |
0f212204 | 9078 | } |
cf27f3b1 | 9079 | tctx->last = ctx; |
0f212204 JA |
9080 | return 0; |
9081 | } | |
9082 | ||
cf27f3b1 PB |
9083 | /* |
9084 | * Note that this task has used io_uring. We use it for cancelation purposes. | |
9085 | */ | |
eef51daa | 9086 | static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx) |
cf27f3b1 PB |
9087 | { |
9088 | struct io_uring_task *tctx = current->io_uring; | |
9089 | ||
9090 | if (likely(tctx && tctx->last == ctx)) | |
9091 | return 0; | |
eef51daa | 9092 | return __io_uring_add_tctx_node(ctx); |
cf27f3b1 PB |
9093 | } |
9094 | ||
0f212204 JA |
9095 | /* |
9096 | * Remove this io_uring_file -> task mapping. | |
9097 | */ | |
eef51daa | 9098 | static void io_uring_del_tctx_node(unsigned long index) |
0f212204 JA |
9099 | { |
9100 | struct io_uring_task *tctx = current->io_uring; | |
13bf43f5 | 9101 | struct io_tctx_node *node; |
2941267b | 9102 | |
eebd2e37 PB |
9103 | if (!tctx) |
9104 | return; | |
13bf43f5 PB |
9105 | node = xa_erase(&tctx->xa, index); |
9106 | if (!node) | |
2941267b | 9107 | return; |
0f212204 | 9108 | |
13bf43f5 PB |
9109 | WARN_ON_ONCE(current != node->task); |
9110 | WARN_ON_ONCE(list_empty(&node->ctx_node)); | |
9111 | ||
9112 | mutex_lock(&node->ctx->uring_lock); | |
9113 | list_del(&node->ctx_node); | |
9114 | mutex_unlock(&node->ctx->uring_lock); | |
9115 | ||
baf186c4 | 9116 | if (tctx->last == node->ctx) |
0f212204 | 9117 | tctx->last = NULL; |
13bf43f5 | 9118 | kfree(node); |
0f212204 JA |
9119 | } |
9120 | ||
8452d4a6 | 9121 | static void io_uring_clean_tctx(struct io_uring_task *tctx) |
de7f1d9e | 9122 | { |
ba5ef6dc | 9123 | struct io_wq *wq = tctx->io_wq; |
13bf43f5 | 9124 | struct io_tctx_node *node; |
de7f1d9e PB |
9125 | unsigned long index; |
9126 | ||
13bf43f5 | 9127 | xa_for_each(&tctx->xa, index, node) |
eef51daa | 9128 | io_uring_del_tctx_node(index); |
b16ef427 ME |
9129 | if (wq) { |
9130 | /* | |
9131 | * Must be after io_uring_del_task_file() (removes nodes under | |
9132 | * uring_lock) to avoid race with io_uring_try_cancel_iowq(). | |
9133 | */ | |
9134 | tctx->io_wq = NULL; | |
ba5ef6dc | 9135 | io_wq_put_and_exit(wq); |
b16ef427 | 9136 | } |
de7f1d9e PB |
9137 | } |
9138 | ||
3f48cf18 | 9139 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 9140 | { |
3f48cf18 PB |
9141 | if (tracked) |
9142 | return atomic_read(&tctx->inflight_tracked); | |
521d6a73 PB |
9143 | return percpu_counter_sum(&tctx->inflight); |
9144 | } | |
9145 | ||
09899b19 PB |
9146 | static void io_uring_drop_tctx_refs(struct task_struct *task) |
9147 | { | |
9148 | struct io_uring_task *tctx = task->io_uring; | |
9149 | unsigned int refs = tctx->cached_refs; | |
9150 | ||
e9dbe221 PB |
9151 | if (refs) { |
9152 | tctx->cached_refs = 0; | |
9153 | percpu_counter_sub(&tctx->inflight, refs); | |
9154 | put_task_struct_many(task, refs); | |
9155 | } | |
09899b19 PB |
9156 | } |
9157 | ||
78cc687b PB |
9158 | /* |
9159 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
9160 | * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation. | |
9161 | */ | |
9162 | static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) | |
0e9ddb39 | 9163 | { |
521d6a73 | 9164 | struct io_uring_task *tctx = current->io_uring; |
734551df | 9165 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
9166 | s64 inflight; |
9167 | DEFINE_WAIT(wait); | |
fdaf083c | 9168 | |
78cc687b PB |
9169 | WARN_ON_ONCE(sqd && sqd->thread != current); |
9170 | ||
6d042ffb PO |
9171 | if (!current->io_uring) |
9172 | return; | |
17a91051 PB |
9173 | if (tctx->io_wq) |
9174 | io_wq_exit_start(tctx->io_wq); | |
9175 | ||
0e9ddb39 PB |
9176 | atomic_inc(&tctx->in_idle); |
9177 | do { | |
e9dbe221 | 9178 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 9179 | /* read completions before cancelations */ |
78cc687b | 9180 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
9181 | if (!inflight) |
9182 | break; | |
fdaf083c | 9183 | |
78cc687b PB |
9184 | if (!sqd) { |
9185 | struct io_tctx_node *node; | |
9186 | unsigned long index; | |
0f212204 | 9187 | |
78cc687b PB |
9188 | xa_for_each(&tctx->xa, index, node) { |
9189 | /* sqpoll task will cancel all its requests */ | |
9190 | if (node->ctx->sq_data) | |
9191 | continue; | |
9192 | io_uring_try_cancel_requests(node->ctx, current, | |
9193 | cancel_all); | |
9194 | } | |
9195 | } else { | |
9196 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
9197 | io_uring_try_cancel_requests(ctx, current, | |
9198 | cancel_all); | |
9199 | } | |
17a91051 | 9200 | |
0f212204 | 9201 | prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE); |
e9dbe221 | 9202 | io_uring_drop_tctx_refs(current); |
0f212204 | 9203 | /* |
a1bb3cd5 PB |
9204 | * If we've seen completions, retry without waiting. This |
9205 | * avoids a race where a completion comes in before we did | |
9206 | * prepare_to_wait(). | |
0f212204 | 9207 | */ |
3dd0c97a | 9208 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 9209 | schedule(); |
f57555ed | 9210 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 9211 | } while (1); |
fdaf083c | 9212 | atomic_dec(&tctx->in_idle); |
de7f1d9e | 9213 | |
8452d4a6 | 9214 | io_uring_clean_tctx(tctx); |
3dd0c97a | 9215 | if (cancel_all) { |
3f48cf18 PB |
9216 | /* for exec all current's requests should be gone, kill tctx */ |
9217 | __io_uring_free(current); | |
9218 | } | |
44e728b8 PB |
9219 | } |
9220 | ||
f552a27a | 9221 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 9222 | { |
f552a27a | 9223 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
9224 | } |
9225 | ||
6c5c240e RP |
9226 | static void *io_uring_validate_mmap_request(struct file *file, |
9227 | loff_t pgoff, size_t sz) | |
2b188cc1 | 9228 | { |
2b188cc1 | 9229 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 9230 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
9231 | struct page *page; |
9232 | void *ptr; | |
9233 | ||
9234 | switch (offset) { | |
9235 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
9236 | case IORING_OFF_CQ_RING: |
9237 | ptr = ctx->rings; | |
2b188cc1 JA |
9238 | break; |
9239 | case IORING_OFF_SQES: | |
9240 | ptr = ctx->sq_sqes; | |
9241 | break; | |
2b188cc1 | 9242 | default: |
6c5c240e | 9243 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
9244 | } |
9245 | ||
9246 | page = virt_to_head_page(ptr); | |
a50b854e | 9247 | if (sz > page_size(page)) |
6c5c240e RP |
9248 | return ERR_PTR(-EINVAL); |
9249 | ||
9250 | return ptr; | |
9251 | } | |
9252 | ||
9253 | #ifdef CONFIG_MMU | |
9254 | ||
9255 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9256 | { | |
9257 | size_t sz = vma->vm_end - vma->vm_start; | |
9258 | unsigned long pfn; | |
9259 | void *ptr; | |
9260 | ||
9261 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
9262 | if (IS_ERR(ptr)) | |
9263 | return PTR_ERR(ptr); | |
2b188cc1 JA |
9264 | |
9265 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
9266 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
9267 | } | |
9268 | ||
6c5c240e RP |
9269 | #else /* !CONFIG_MMU */ |
9270 | ||
9271 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9272 | { | |
9273 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
9274 | } | |
9275 | ||
9276 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
9277 | { | |
9278 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
9279 | } | |
9280 | ||
9281 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
9282 | unsigned long addr, unsigned long len, | |
9283 | unsigned long pgoff, unsigned long flags) | |
9284 | { | |
9285 | void *ptr; | |
9286 | ||
9287 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
9288 | if (IS_ERR(ptr)) | |
9289 | return PTR_ERR(ptr); | |
9290 | ||
9291 | return (unsigned long) ptr; | |
9292 | } | |
9293 | ||
9294 | #endif /* !CONFIG_MMU */ | |
9295 | ||
d9d05217 | 9296 | static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx) |
90554200 JA |
9297 | { |
9298 | DEFINE_WAIT(wait); | |
9299 | ||
9300 | do { | |
9301 | if (!io_sqring_full(ctx)) | |
9302 | break; | |
90554200 JA |
9303 | prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE); |
9304 | ||
9305 | if (!io_sqring_full(ctx)) | |
9306 | break; | |
90554200 JA |
9307 | schedule(); |
9308 | } while (!signal_pending(current)); | |
9309 | ||
9310 | finish_wait(&ctx->sqo_sq_wait, &wait); | |
5199328a | 9311 | return 0; |
90554200 JA |
9312 | } |
9313 | ||
c73ebb68 HX |
9314 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
9315 | struct __kernel_timespec __user **ts, | |
9316 | const sigset_t __user **sig) | |
9317 | { | |
9318 | struct io_uring_getevents_arg arg; | |
9319 | ||
9320 | /* | |
9321 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
9322 | * is just a pointer to the sigset_t. | |
9323 | */ | |
9324 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
9325 | *sig = (const sigset_t __user *) argp; | |
9326 | *ts = NULL; | |
9327 | return 0; | |
9328 | } | |
9329 | ||
9330 | /* | |
9331 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
9332 | * timespec and sigset_t pointers if good. | |
9333 | */ | |
9334 | if (*argsz != sizeof(arg)) | |
9335 | return -EINVAL; | |
9336 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
9337 | return -EFAULT; | |
9338 | *sig = u64_to_user_ptr(arg.sigmask); | |
9339 | *argsz = arg.sigmask_sz; | |
9340 | *ts = u64_to_user_ptr(arg.ts); | |
9341 | return 0; | |
9342 | } | |
9343 | ||
2b188cc1 | 9344 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
9345 | u32, min_complete, u32, flags, const void __user *, argp, |
9346 | size_t, argsz) | |
2b188cc1 JA |
9347 | { |
9348 | struct io_ring_ctx *ctx; | |
2b188cc1 JA |
9349 | int submitted = 0; |
9350 | struct fd f; | |
33f993da | 9351 | long ret; |
2b188cc1 | 9352 | |
4c6e277c | 9353 | io_run_task_work(); |
b41e9852 | 9354 | |
33f993da PB |
9355 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
9356 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))) | |
2b188cc1 JA |
9357 | return -EINVAL; |
9358 | ||
9359 | f = fdget(fd); | |
33f993da | 9360 | if (unlikely(!f.file)) |
2b188cc1 JA |
9361 | return -EBADF; |
9362 | ||
9363 | ret = -EOPNOTSUPP; | |
33f993da | 9364 | if (unlikely(f.file->f_op != &io_uring_fops)) |
2b188cc1 JA |
9365 | goto out_fput; |
9366 | ||
9367 | ret = -ENXIO; | |
9368 | ctx = f.file->private_data; | |
33f993da | 9369 | if (unlikely(!percpu_ref_tryget(&ctx->refs))) |
2b188cc1 JA |
9370 | goto out_fput; |
9371 | ||
7e84e1c7 | 9372 | ret = -EBADFD; |
33f993da | 9373 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
9374 | goto out; |
9375 | ||
6c271ce2 JA |
9376 | /* |
9377 | * For SQ polling, the thread will do all submissions and completions. | |
9378 | * Just return the requested submit count, and wake the thread if | |
9379 | * we were asked to. | |
9380 | */ | |
b2a9eada | 9381 | ret = 0; |
6c271ce2 | 9382 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 9383 | io_cqring_overflow_flush(ctx); |
89448c47 | 9384 | |
21f96522 JA |
9385 | if (unlikely(ctx->sq_data->thread == NULL)) { |
9386 | ret = -EOWNERDEAD; | |
04147488 | 9387 | goto out; |
21f96522 | 9388 | } |
6c271ce2 | 9389 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 9390 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
9391 | if (flags & IORING_ENTER_SQ_WAIT) { |
9392 | ret = io_sqpoll_wait_sq(ctx); | |
9393 | if (ret) | |
9394 | goto out; | |
9395 | } | |
6c271ce2 | 9396 | submitted = to_submit; |
b2a9eada | 9397 | } else if (to_submit) { |
eef51daa | 9398 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
9399 | if (unlikely(ret)) |
9400 | goto out; | |
2b188cc1 | 9401 | mutex_lock(&ctx->uring_lock); |
0f212204 | 9402 | submitted = io_submit_sqes(ctx, to_submit); |
2b188cc1 | 9403 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
9404 | |
9405 | if (submitted != to_submit) | |
9406 | goto out; | |
2b188cc1 JA |
9407 | } |
9408 | if (flags & IORING_ENTER_GETEVENTS) { | |
c73ebb68 HX |
9409 | const sigset_t __user *sig; |
9410 | struct __kernel_timespec __user *ts; | |
9411 | ||
9412 | ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); | |
9413 | if (unlikely(ret)) | |
9414 | goto out; | |
9415 | ||
2b188cc1 JA |
9416 | min_complete = min(min_complete, ctx->cq_entries); |
9417 | ||
32b2244a XW |
9418 | /* |
9419 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
9420 | * space applications don't need to do io completion events | |
9421 | * polling again, they can rely on io_sq_thread to do polling | |
9422 | * work, which can reduce cpu usage and uring_lock contention. | |
9423 | */ | |
9424 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
9425 | !(ctx->flags & IORING_SETUP_SQPOLL)) { | |
7668b92a | 9426 | ret = io_iopoll_check(ctx, min_complete); |
def596e9 | 9427 | } else { |
c73ebb68 | 9428 | ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts); |
def596e9 | 9429 | } |
2b188cc1 JA |
9430 | } |
9431 | ||
7c504e65 | 9432 | out: |
6805b32e | 9433 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
9434 | out_fput: |
9435 | fdput(f); | |
9436 | return submitted ? submitted : ret; | |
9437 | } | |
9438 | ||
bebdb65e | 9439 | #ifdef CONFIG_PROC_FS |
61cf9370 MWO |
9440 | static int io_uring_show_cred(struct seq_file *m, unsigned int id, |
9441 | const struct cred *cred) | |
87ce955b | 9442 | { |
87ce955b JA |
9443 | struct user_namespace *uns = seq_user_ns(m); |
9444 | struct group_info *gi; | |
9445 | kernel_cap_t cap; | |
9446 | unsigned __capi; | |
9447 | int g; | |
9448 | ||
9449 | seq_printf(m, "%5d\n", id); | |
9450 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
9451 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
9452 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
9453 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
9454 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
9455 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
9456 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
9457 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
9458 | seq_puts(m, "\n\tGroups:\t"); | |
9459 | gi = cred->group_info; | |
9460 | for (g = 0; g < gi->ngroups; g++) { | |
9461 | seq_put_decimal_ull(m, g ? " " : "", | |
9462 | from_kgid_munged(uns, gi->gid[g])); | |
9463 | } | |
9464 | seq_puts(m, "\n\tCapEff:\t"); | |
9465 | cap = cred->cap_effective; | |
9466 | CAP_FOR_EACH_U32(__capi) | |
9467 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
9468 | seq_putc(m, '\n'); | |
9469 | return 0; | |
9470 | } | |
9471 | ||
9472 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
9473 | { | |
dbbe9c64 | 9474 | struct io_sq_data *sq = NULL; |
fad8e0de | 9475 | bool has_lock; |
87ce955b JA |
9476 | int i; |
9477 | ||
fad8e0de JA |
9478 | /* |
9479 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, | |
9480 | * since fdinfo case grabs it in the opposite direction of normal use | |
9481 | * cases. If we fail to get the lock, we just don't iterate any | |
9482 | * structures that could be going away outside the io_uring mutex. | |
9483 | */ | |
9484 | has_lock = mutex_trylock(&ctx->uring_lock); | |
9485 | ||
5f3f26f9 | 9486 | if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) { |
dbbe9c64 | 9487 | sq = ctx->sq_data; |
5f3f26f9 JA |
9488 | if (!sq->thread) |
9489 | sq = NULL; | |
9490 | } | |
dbbe9c64 JQ |
9491 | |
9492 | seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1); | |
9493 | seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1); | |
87ce955b | 9494 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
fad8e0de | 9495 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
7b29f92d | 9496 | struct file *f = io_file_from_index(ctx, i); |
87ce955b | 9497 | |
87ce955b JA |
9498 | if (f) |
9499 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
9500 | else | |
9501 | seq_printf(m, "%5u: <none>\n", i); | |
9502 | } | |
9503 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
fad8e0de | 9504 | for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) { |
41edf1a5 | 9505 | struct io_mapped_ubuf *buf = ctx->user_bufs[i]; |
4751f53d | 9506 | unsigned int len = buf->ubuf_end - buf->ubuf; |
87ce955b | 9507 | |
4751f53d | 9508 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len); |
87ce955b | 9509 | } |
61cf9370 MWO |
9510 | if (has_lock && !xa_empty(&ctx->personalities)) { |
9511 | unsigned long index; | |
9512 | const struct cred *cred; | |
9513 | ||
87ce955b | 9514 | seq_printf(m, "Personalities:\n"); |
61cf9370 MWO |
9515 | xa_for_each(&ctx->personalities, index, cred) |
9516 | io_uring_show_cred(m, index, cred); | |
87ce955b | 9517 | } |
d7718a9d | 9518 | seq_printf(m, "PollList:\n"); |
79ebeaee | 9519 | spin_lock(&ctx->completion_lock); |
d7718a9d JA |
9520 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
9521 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
9522 | struct io_kiocb *req; | |
9523 | ||
9524 | hlist_for_each_entry(req, list, hash_node) | |
9525 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
9526 | req->task->task_works != NULL); | |
9527 | } | |
79ebeaee | 9528 | spin_unlock(&ctx->completion_lock); |
fad8e0de JA |
9529 | if (has_lock) |
9530 | mutex_unlock(&ctx->uring_lock); | |
87ce955b JA |
9531 | } |
9532 | ||
9533 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
9534 | { | |
9535 | struct io_ring_ctx *ctx = f->private_data; | |
9536 | ||
9537 | if (percpu_ref_tryget(&ctx->refs)) { | |
9538 | __io_uring_show_fdinfo(ctx, m); | |
9539 | percpu_ref_put(&ctx->refs); | |
9540 | } | |
9541 | } | |
bebdb65e | 9542 | #endif |
87ce955b | 9543 | |
2b188cc1 JA |
9544 | static const struct file_operations io_uring_fops = { |
9545 | .release = io_uring_release, | |
9546 | .mmap = io_uring_mmap, | |
6c5c240e RP |
9547 | #ifndef CONFIG_MMU |
9548 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
9549 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
9550 | #endif | |
2b188cc1 JA |
9551 | .poll = io_uring_poll, |
9552 | .fasync = io_uring_fasync, | |
bebdb65e | 9553 | #ifdef CONFIG_PROC_FS |
87ce955b | 9554 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 9555 | #endif |
2b188cc1 JA |
9556 | }; |
9557 | ||
9558 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
9559 | struct io_uring_params *p) | |
9560 | { | |
75b28aff HV |
9561 | struct io_rings *rings; |
9562 | size_t size, sq_array_offset; | |
2b188cc1 | 9563 | |
bd740481 JA |
9564 | /* make sure these are sane, as we already accounted them */ |
9565 | ctx->sq_entries = p->sq_entries; | |
9566 | ctx->cq_entries = p->cq_entries; | |
9567 | ||
75b28aff HV |
9568 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
9569 | if (size == SIZE_MAX) | |
9570 | return -EOVERFLOW; | |
9571 | ||
9572 | rings = io_mem_alloc(size); | |
9573 | if (!rings) | |
2b188cc1 JA |
9574 | return -ENOMEM; |
9575 | ||
75b28aff HV |
9576 | ctx->rings = rings; |
9577 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
9578 | rings->sq_ring_mask = p->sq_entries - 1; | |
9579 | rings->cq_ring_mask = p->cq_entries - 1; | |
9580 | rings->sq_ring_entries = p->sq_entries; | |
9581 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 JA |
9582 | |
9583 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
9584 | if (size == SIZE_MAX) { |
9585 | io_mem_free(ctx->rings); | |
9586 | ctx->rings = NULL; | |
2b188cc1 | 9587 | return -EOVERFLOW; |
eb065d30 | 9588 | } |
2b188cc1 JA |
9589 | |
9590 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
9591 | if (!ctx->sq_sqes) { |
9592 | io_mem_free(ctx->rings); | |
9593 | ctx->rings = NULL; | |
2b188cc1 | 9594 | return -ENOMEM; |
eb065d30 | 9595 | } |
2b188cc1 | 9596 | |
2b188cc1 JA |
9597 | return 0; |
9598 | } | |
9599 | ||
9faadcc8 PB |
9600 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
9601 | { | |
9602 | int ret, fd; | |
9603 | ||
9604 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
9605 | if (fd < 0) | |
9606 | return fd; | |
9607 | ||
eef51daa | 9608 | ret = io_uring_add_tctx_node(ctx); |
9faadcc8 PB |
9609 | if (ret) { |
9610 | put_unused_fd(fd); | |
9611 | return ret; | |
9612 | } | |
9613 | fd_install(fd, file); | |
9614 | return fd; | |
9615 | } | |
9616 | ||
2b188cc1 JA |
9617 | /* |
9618 | * Allocate an anonymous fd, this is what constitutes the application | |
9619 | * visible backing of an io_uring instance. The application mmaps this | |
9620 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
9621 | * we have to tie this fd to a socket for file garbage collection purposes. | |
9622 | */ | |
9faadcc8 | 9623 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
9624 | { |
9625 | struct file *file; | |
9faadcc8 | 9626 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
9627 | int ret; |
9628 | ||
2b188cc1 JA |
9629 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
9630 | &ctx->ring_sock); | |
9631 | if (ret) | |
9faadcc8 | 9632 | return ERR_PTR(ret); |
2b188cc1 JA |
9633 | #endif |
9634 | ||
2b188cc1 JA |
9635 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, |
9636 | O_RDWR | O_CLOEXEC); | |
2b188cc1 | 9637 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
9638 | if (IS_ERR(file)) { |
9639 | sock_release(ctx->ring_sock); | |
9640 | ctx->ring_sock = NULL; | |
9641 | } else { | |
9642 | ctx->ring_sock->file = file; | |
0f212204 | 9643 | } |
2b188cc1 | 9644 | #endif |
9faadcc8 | 9645 | return file; |
2b188cc1 JA |
9646 | } |
9647 | ||
7f13657d XW |
9648 | static int io_uring_create(unsigned entries, struct io_uring_params *p, |
9649 | struct io_uring_params __user *params) | |
2b188cc1 | 9650 | { |
2b188cc1 | 9651 | struct io_ring_ctx *ctx; |
9faadcc8 | 9652 | struct file *file; |
2b188cc1 JA |
9653 | int ret; |
9654 | ||
8110c1a6 | 9655 | if (!entries) |
2b188cc1 | 9656 | return -EINVAL; |
8110c1a6 JA |
9657 | if (entries > IORING_MAX_ENTRIES) { |
9658 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9659 | return -EINVAL; | |
9660 | entries = IORING_MAX_ENTRIES; | |
9661 | } | |
2b188cc1 JA |
9662 | |
9663 | /* | |
9664 | * Use twice as many entries for the CQ ring. It's possible for the | |
9665 | * application to drive a higher depth than the size of the SQ ring, | |
9666 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
9667 | * some flexibility in overcommitting a bit. If the application has |
9668 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
9669 | * of CQ ring entries manually. | |
2b188cc1 JA |
9670 | */ |
9671 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
9672 | if (p->flags & IORING_SETUP_CQSIZE) { |
9673 | /* | |
9674 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
9675 | * to a power-of-two, if it isn't already. We do NOT impose | |
9676 | * any cq vs sq ring sizing. | |
9677 | */ | |
eb2667b3 | 9678 | if (!p->cq_entries) |
33a107f0 | 9679 | return -EINVAL; |
8110c1a6 JA |
9680 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
9681 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9682 | return -EINVAL; | |
9683 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
9684 | } | |
eb2667b3 JQ |
9685 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
9686 | if (p->cq_entries < p->sq_entries) | |
9687 | return -EINVAL; | |
33a107f0 JA |
9688 | } else { |
9689 | p->cq_entries = 2 * p->sq_entries; | |
9690 | } | |
2b188cc1 | 9691 | |
2b188cc1 | 9692 | ctx = io_ring_ctx_alloc(p); |
62e398be | 9693 | if (!ctx) |
2b188cc1 | 9694 | return -ENOMEM; |
2b188cc1 | 9695 | ctx->compat = in_compat_syscall(); |
62e398be JA |
9696 | if (!capable(CAP_IPC_LOCK)) |
9697 | ctx->user = get_uid(current_user()); | |
2aede0e4 JA |
9698 | |
9699 | /* | |
9700 | * This is just grabbed for accounting purposes. When a process exits, | |
9701 | * the mm is exited and dropped before the files, hence we need to hang | |
9702 | * on to this mm purely for the purposes of being able to unaccount | |
9703 | * memory (locked/pinned vm). It's not used for anything else. | |
9704 | */ | |
6b7898eb | 9705 | mmgrab(current->mm); |
2aede0e4 | 9706 | ctx->mm_account = current->mm; |
6b7898eb | 9707 | |
2b188cc1 JA |
9708 | ret = io_allocate_scq_urings(ctx, p); |
9709 | if (ret) | |
9710 | goto err; | |
9711 | ||
7e84e1c7 | 9712 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
9713 | if (ret) |
9714 | goto err; | |
eae071c9 | 9715 | /* always set a rsrc node */ |
47b228ce PB |
9716 | ret = io_rsrc_node_switch_start(ctx); |
9717 | if (ret) | |
9718 | goto err; | |
eae071c9 | 9719 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 9720 | |
2b188cc1 | 9721 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
9722 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
9723 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
9724 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
9725 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
9726 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
9727 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
9728 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
9729 | |
9730 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
9731 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
9732 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
9733 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
9734 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
9735 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
9736 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 9737 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 9738 | |
7f13657d XW |
9739 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
9740 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 9741 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 9742 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e PB |
9743 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
9744 | IORING_FEAT_RSRC_TAGS; | |
7f13657d XW |
9745 | |
9746 | if (copy_to_user(params, p, sizeof(*p))) { | |
9747 | ret = -EFAULT; | |
9748 | goto err; | |
9749 | } | |
d1719f70 | 9750 | |
9faadcc8 PB |
9751 | file = io_uring_get_file(ctx); |
9752 | if (IS_ERR(file)) { | |
9753 | ret = PTR_ERR(file); | |
9754 | goto err; | |
9755 | } | |
9756 | ||
044c1ab3 JA |
9757 | /* |
9758 | * Install ring fd as the very last thing, so we don't risk someone | |
9759 | * having closed it before we finish setup | |
9760 | */ | |
9faadcc8 PB |
9761 | ret = io_uring_install_fd(ctx, file); |
9762 | if (ret < 0) { | |
9763 | /* fput will clean it up */ | |
9764 | fput(file); | |
9765 | return ret; | |
9766 | } | |
044c1ab3 | 9767 | |
c826bd7a | 9768 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
9769 | return ret; |
9770 | err: | |
9771 | io_ring_ctx_wait_and_kill(ctx); | |
9772 | return ret; | |
9773 | } | |
9774 | ||
9775 | /* | |
9776 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
9777 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
9778 | * params structure passed in. | |
9779 | */ | |
9780 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
9781 | { | |
9782 | struct io_uring_params p; | |
2b188cc1 JA |
9783 | int i; |
9784 | ||
9785 | if (copy_from_user(&p, params, sizeof(p))) | |
9786 | return -EFAULT; | |
9787 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
9788 | if (p.resv[i]) | |
9789 | return -EINVAL; | |
9790 | } | |
9791 | ||
6c271ce2 | 9792 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 9793 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 SG |
9794 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
9795 | IORING_SETUP_R_DISABLED)) | |
2b188cc1 JA |
9796 | return -EINVAL; |
9797 | ||
7f13657d | 9798 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
9799 | } |
9800 | ||
9801 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
9802 | struct io_uring_params __user *, params) | |
9803 | { | |
9804 | return io_uring_setup(entries, params); | |
9805 | } | |
9806 | ||
66f4af93 JA |
9807 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
9808 | { | |
9809 | struct io_uring_probe *p; | |
9810 | size_t size; | |
9811 | int i, ret; | |
9812 | ||
9813 | size = struct_size(p, ops, nr_args); | |
9814 | if (size == SIZE_MAX) | |
9815 | return -EOVERFLOW; | |
9816 | p = kzalloc(size, GFP_KERNEL); | |
9817 | if (!p) | |
9818 | return -ENOMEM; | |
9819 | ||
9820 | ret = -EFAULT; | |
9821 | if (copy_from_user(p, arg, size)) | |
9822 | goto out; | |
9823 | ret = -EINVAL; | |
9824 | if (memchr_inv(p, 0, size)) | |
9825 | goto out; | |
9826 | ||
9827 | p->last_op = IORING_OP_LAST - 1; | |
9828 | if (nr_args > IORING_OP_LAST) | |
9829 | nr_args = IORING_OP_LAST; | |
9830 | ||
9831 | for (i = 0; i < nr_args; i++) { | |
9832 | p->ops[i].op = i; | |
9833 | if (!io_op_defs[i].not_supported) | |
9834 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
9835 | } | |
9836 | p->ops_len = i; | |
9837 | ||
9838 | ret = 0; | |
9839 | if (copy_to_user(arg, p, size)) | |
9840 | ret = -EFAULT; | |
9841 | out: | |
9842 | kfree(p); | |
9843 | return ret; | |
9844 | } | |
9845 | ||
071698e1 JA |
9846 | static int io_register_personality(struct io_ring_ctx *ctx) |
9847 | { | |
4379bf8b | 9848 | const struct cred *creds; |
61cf9370 | 9849 | u32 id; |
1e6fa521 | 9850 | int ret; |
071698e1 | 9851 | |
4379bf8b | 9852 | creds = get_current_cred(); |
1e6fa521 | 9853 | |
61cf9370 MWO |
9854 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
9855 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
9856 | if (ret < 0) { |
9857 | put_cred(creds); | |
9858 | return ret; | |
9859 | } | |
9860 | return id; | |
071698e1 JA |
9861 | } |
9862 | ||
21b55dbc SG |
9863 | static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg, |
9864 | unsigned int nr_args) | |
9865 | { | |
9866 | struct io_uring_restriction *res; | |
9867 | size_t size; | |
9868 | int i, ret; | |
9869 | ||
7e84e1c7 SG |
9870 | /* Restrictions allowed only if rings started disabled */ |
9871 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9872 | return -EBADFD; | |
9873 | ||
21b55dbc | 9874 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 9875 | if (ctx->restrictions.registered) |
21b55dbc SG |
9876 | return -EBUSY; |
9877 | ||
9878 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
9879 | return -EINVAL; | |
9880 | ||
9881 | size = array_size(nr_args, sizeof(*res)); | |
9882 | if (size == SIZE_MAX) | |
9883 | return -EOVERFLOW; | |
9884 | ||
9885 | res = memdup_user(arg, size); | |
9886 | if (IS_ERR(res)) | |
9887 | return PTR_ERR(res); | |
9888 | ||
9889 | ret = 0; | |
9890 | ||
9891 | for (i = 0; i < nr_args; i++) { | |
9892 | switch (res[i].opcode) { | |
9893 | case IORING_RESTRICTION_REGISTER_OP: | |
9894 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
9895 | ret = -EINVAL; | |
9896 | goto out; | |
9897 | } | |
9898 | ||
9899 | __set_bit(res[i].register_op, | |
9900 | ctx->restrictions.register_op); | |
9901 | break; | |
9902 | case IORING_RESTRICTION_SQE_OP: | |
9903 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
9904 | ret = -EINVAL; | |
9905 | goto out; | |
9906 | } | |
9907 | ||
9908 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
9909 | break; | |
9910 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
9911 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
9912 | break; | |
9913 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
9914 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
9915 | break; | |
9916 | default: | |
9917 | ret = -EINVAL; | |
9918 | goto out; | |
9919 | } | |
9920 | } | |
9921 | ||
9922 | out: | |
9923 | /* Reset all restrictions if an error happened */ | |
9924 | if (ret != 0) | |
9925 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
9926 | else | |
7e84e1c7 | 9927 | ctx->restrictions.registered = true; |
21b55dbc SG |
9928 | |
9929 | kfree(res); | |
9930 | return ret; | |
9931 | } | |
9932 | ||
7e84e1c7 SG |
9933 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
9934 | { | |
9935 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9936 | return -EBADFD; | |
9937 | ||
9938 | if (ctx->restrictions.registered) | |
9939 | ctx->restricted = 1; | |
9940 | ||
0298ef96 PB |
9941 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
9942 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
9943 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
9944 | return 0; |
9945 | } | |
9946 | ||
fdecb662 | 9947 | static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, |
c3bdad02 | 9948 | struct io_uring_rsrc_update2 *up, |
98f0b3b4 PB |
9949 | unsigned nr_args) |
9950 | { | |
9951 | __u32 tmp; | |
9952 | int err; | |
9953 | ||
c3bdad02 PB |
9954 | if (up->resv) |
9955 | return -EINVAL; | |
98f0b3b4 PB |
9956 | if (check_add_overflow(up->offset, nr_args, &tmp)) |
9957 | return -EOVERFLOW; | |
9958 | err = io_rsrc_node_switch_start(ctx); | |
9959 | if (err) | |
9960 | return err; | |
9961 | ||
fdecb662 PB |
9962 | switch (type) { |
9963 | case IORING_RSRC_FILE: | |
98f0b3b4 | 9964 | return __io_sqe_files_update(ctx, up, nr_args); |
634d00df PB |
9965 | case IORING_RSRC_BUFFER: |
9966 | return __io_sqe_buffers_update(ctx, up, nr_args); | |
98f0b3b4 PB |
9967 | } |
9968 | return -EINVAL; | |
9969 | } | |
9970 | ||
c3bdad02 PB |
9971 | static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, |
9972 | unsigned nr_args) | |
98f0b3b4 | 9973 | { |
c3bdad02 | 9974 | struct io_uring_rsrc_update2 up; |
98f0b3b4 PB |
9975 | |
9976 | if (!nr_args) | |
9977 | return -EINVAL; | |
c3bdad02 PB |
9978 | memset(&up, 0, sizeof(up)); |
9979 | if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) | |
9980 | return -EFAULT; | |
9981 | return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); | |
9982 | } | |
9983 | ||
9984 | static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, | |
992da01a | 9985 | unsigned size, unsigned type) |
c3bdad02 PB |
9986 | { |
9987 | struct io_uring_rsrc_update2 up; | |
9988 | ||
9989 | if (size != sizeof(up)) | |
9990 | return -EINVAL; | |
98f0b3b4 PB |
9991 | if (copy_from_user(&up, arg, sizeof(up))) |
9992 | return -EFAULT; | |
992da01a | 9993 | if (!up.nr || up.resv) |
98f0b3b4 | 9994 | return -EINVAL; |
992da01a | 9995 | return __io_register_rsrc_update(ctx, type, &up, up.nr); |
98f0b3b4 PB |
9996 | } |
9997 | ||
792e3582 | 9998 | static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, |
992da01a | 9999 | unsigned int size, unsigned int type) |
792e3582 PB |
10000 | { |
10001 | struct io_uring_rsrc_register rr; | |
10002 | ||
10003 | /* keep it extendible */ | |
10004 | if (size != sizeof(rr)) | |
10005 | return -EINVAL; | |
10006 | ||
10007 | memset(&rr, 0, sizeof(rr)); | |
10008 | if (copy_from_user(&rr, arg, size)) | |
10009 | return -EFAULT; | |
992da01a | 10010 | if (!rr.nr || rr.resv || rr.resv2) |
792e3582 PB |
10011 | return -EINVAL; |
10012 | ||
992da01a | 10013 | switch (type) { |
792e3582 PB |
10014 | case IORING_RSRC_FILE: |
10015 | return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), | |
10016 | rr.nr, u64_to_user_ptr(rr.tags)); | |
634d00df PB |
10017 | case IORING_RSRC_BUFFER: |
10018 | return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), | |
10019 | rr.nr, u64_to_user_ptr(rr.tags)); | |
792e3582 PB |
10020 | } |
10021 | return -EINVAL; | |
10022 | } | |
10023 | ||
fe76421d JA |
10024 | static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg, |
10025 | unsigned len) | |
10026 | { | |
10027 | struct io_uring_task *tctx = current->io_uring; | |
10028 | cpumask_var_t new_mask; | |
10029 | int ret; | |
10030 | ||
10031 | if (!tctx || !tctx->io_wq) | |
10032 | return -EINVAL; | |
10033 | ||
10034 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
10035 | return -ENOMEM; | |
10036 | ||
10037 | cpumask_clear(new_mask); | |
10038 | if (len > cpumask_size()) | |
10039 | len = cpumask_size(); | |
10040 | ||
10041 | if (copy_from_user(new_mask, arg, len)) { | |
10042 | free_cpumask_var(new_mask); | |
10043 | return -EFAULT; | |
10044 | } | |
10045 | ||
10046 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
10047 | free_cpumask_var(new_mask); | |
10048 | return ret; | |
10049 | } | |
10050 | ||
10051 | static int io_unregister_iowq_aff(struct io_ring_ctx *ctx) | |
10052 | { | |
10053 | struct io_uring_task *tctx = current->io_uring; | |
10054 | ||
10055 | if (!tctx || !tctx->io_wq) | |
10056 | return -EINVAL; | |
10057 | ||
10058 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
10059 | } | |
10060 | ||
071698e1 JA |
10061 | static bool io_register_op_must_quiesce(int op) |
10062 | { | |
10063 | switch (op) { | |
bd54b6fe BM |
10064 | case IORING_REGISTER_BUFFERS: |
10065 | case IORING_UNREGISTER_BUFFERS: | |
f4f7d21c | 10066 | case IORING_REGISTER_FILES: |
071698e1 JA |
10067 | case IORING_UNREGISTER_FILES: |
10068 | case IORING_REGISTER_FILES_UPDATE: | |
10069 | case IORING_REGISTER_PROBE: | |
10070 | case IORING_REGISTER_PERSONALITY: | |
10071 | case IORING_UNREGISTER_PERSONALITY: | |
992da01a PB |
10072 | case IORING_REGISTER_FILES2: |
10073 | case IORING_REGISTER_FILES_UPDATE2: | |
10074 | case IORING_REGISTER_BUFFERS2: | |
10075 | case IORING_REGISTER_BUFFERS_UPDATE: | |
fe76421d JA |
10076 | case IORING_REGISTER_IOWQ_AFF: |
10077 | case IORING_UNREGISTER_IOWQ_AFF: | |
071698e1 JA |
10078 | return false; |
10079 | default: | |
10080 | return true; | |
10081 | } | |
10082 | } | |
10083 | ||
e73c5c7c PB |
10084 | static int io_ctx_quiesce(struct io_ring_ctx *ctx) |
10085 | { | |
10086 | long ret; | |
10087 | ||
10088 | percpu_ref_kill(&ctx->refs); | |
10089 | ||
10090 | /* | |
10091 | * Drop uring mutex before waiting for references to exit. If another | |
10092 | * thread is currently inside io_uring_enter() it might need to grab the | |
10093 | * uring_lock to make progress. If we hold it here across the drain | |
10094 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
10095 | * no new references will come in after we've killed the percpu ref. | |
10096 | */ | |
10097 | mutex_unlock(&ctx->uring_lock); | |
10098 | do { | |
10099 | ret = wait_for_completion_interruptible(&ctx->ref_comp); | |
10100 | if (!ret) | |
10101 | break; | |
10102 | ret = io_run_task_work_sig(); | |
10103 | } while (ret >= 0); | |
10104 | mutex_lock(&ctx->uring_lock); | |
10105 | ||
10106 | if (ret) | |
10107 | io_refs_resurrect(&ctx->refs, &ctx->ref_comp); | |
10108 | return ret; | |
10109 | } | |
10110 | ||
edafccee JA |
10111 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
10112 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
10113 | __releases(ctx->uring_lock) |
10114 | __acquires(ctx->uring_lock) | |
edafccee JA |
10115 | { |
10116 | int ret; | |
10117 | ||
35fa71a0 JA |
10118 | /* |
10119 | * We're inside the ring mutex, if the ref is already dying, then | |
10120 | * someone else killed the ctx or is already going through | |
10121 | * io_uring_register(). | |
10122 | */ | |
10123 | if (percpu_ref_is_dying(&ctx->refs)) | |
10124 | return -ENXIO; | |
10125 | ||
75c4021a PB |
10126 | if (ctx->restricted) { |
10127 | if (opcode >= IORING_REGISTER_LAST) | |
10128 | return -EINVAL; | |
10129 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); | |
10130 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
10131 | return -EACCES; | |
10132 | } | |
10133 | ||
071698e1 | 10134 | if (io_register_op_must_quiesce(opcode)) { |
e73c5c7c PB |
10135 | ret = io_ctx_quiesce(ctx); |
10136 | if (ret) | |
f70865db | 10137 | return ret; |
05f3fb3c | 10138 | } |
edafccee JA |
10139 | |
10140 | switch (opcode) { | |
10141 | case IORING_REGISTER_BUFFERS: | |
634d00df | 10142 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
10143 | break; |
10144 | case IORING_UNREGISTER_BUFFERS: | |
10145 | ret = -EINVAL; | |
10146 | if (arg || nr_args) | |
10147 | break; | |
0a96bbe4 | 10148 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 10149 | break; |
6b06314c | 10150 | case IORING_REGISTER_FILES: |
792e3582 | 10151 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
10152 | break; |
10153 | case IORING_UNREGISTER_FILES: | |
10154 | ret = -EINVAL; | |
10155 | if (arg || nr_args) | |
10156 | break; | |
10157 | ret = io_sqe_files_unregister(ctx); | |
10158 | break; | |
c3a31e60 | 10159 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 10160 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 10161 | break; |
9b402849 | 10162 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 10163 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
10164 | ret = -EINVAL; |
10165 | if (nr_args != 1) | |
10166 | break; | |
10167 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
10168 | if (ret) |
10169 | break; | |
10170 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
10171 | ctx->eventfd_async = 1; | |
10172 | else | |
10173 | ctx->eventfd_async = 0; | |
9b402849 JA |
10174 | break; |
10175 | case IORING_UNREGISTER_EVENTFD: | |
10176 | ret = -EINVAL; | |
10177 | if (arg || nr_args) | |
10178 | break; | |
10179 | ret = io_eventfd_unregister(ctx); | |
10180 | break; | |
66f4af93 JA |
10181 | case IORING_REGISTER_PROBE: |
10182 | ret = -EINVAL; | |
10183 | if (!arg || nr_args > 256) | |
10184 | break; | |
10185 | ret = io_probe(ctx, arg, nr_args); | |
10186 | break; | |
071698e1 JA |
10187 | case IORING_REGISTER_PERSONALITY: |
10188 | ret = -EINVAL; | |
10189 | if (arg || nr_args) | |
10190 | break; | |
10191 | ret = io_register_personality(ctx); | |
10192 | break; | |
10193 | case IORING_UNREGISTER_PERSONALITY: | |
10194 | ret = -EINVAL; | |
10195 | if (arg) | |
10196 | break; | |
10197 | ret = io_unregister_personality(ctx, nr_args); | |
10198 | break; | |
7e84e1c7 SG |
10199 | case IORING_REGISTER_ENABLE_RINGS: |
10200 | ret = -EINVAL; | |
10201 | if (arg || nr_args) | |
10202 | break; | |
10203 | ret = io_register_enable_rings(ctx); | |
10204 | break; | |
21b55dbc SG |
10205 | case IORING_REGISTER_RESTRICTIONS: |
10206 | ret = io_register_restrictions(ctx, arg, nr_args); | |
10207 | break; | |
992da01a PB |
10208 | case IORING_REGISTER_FILES2: |
10209 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
10210 | break; | |
10211 | case IORING_REGISTER_FILES_UPDATE2: | |
10212 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
10213 | IORING_RSRC_FILE); | |
10214 | break; | |
10215 | case IORING_REGISTER_BUFFERS2: | |
10216 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 10217 | break; |
992da01a PB |
10218 | case IORING_REGISTER_BUFFERS_UPDATE: |
10219 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
10220 | IORING_RSRC_BUFFER); | |
c3bdad02 | 10221 | break; |
fe76421d JA |
10222 | case IORING_REGISTER_IOWQ_AFF: |
10223 | ret = -EINVAL; | |
10224 | if (!arg || !nr_args) | |
10225 | break; | |
10226 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
10227 | break; | |
10228 | case IORING_UNREGISTER_IOWQ_AFF: | |
10229 | ret = -EINVAL; | |
10230 | if (arg || nr_args) | |
10231 | break; | |
10232 | ret = io_unregister_iowq_aff(ctx); | |
10233 | break; | |
edafccee JA |
10234 | default: |
10235 | ret = -EINVAL; | |
10236 | break; | |
10237 | } | |
10238 | ||
071698e1 | 10239 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 10240 | /* bring the ctx back to life */ |
05f3fb3c | 10241 | percpu_ref_reinit(&ctx->refs); |
0f158b4c | 10242 | reinit_completion(&ctx->ref_comp); |
05f3fb3c | 10243 | } |
edafccee JA |
10244 | return ret; |
10245 | } | |
10246 | ||
10247 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
10248 | void __user *, arg, unsigned int, nr_args) | |
10249 | { | |
10250 | struct io_ring_ctx *ctx; | |
10251 | long ret = -EBADF; | |
10252 | struct fd f; | |
10253 | ||
10254 | f = fdget(fd); | |
10255 | if (!f.file) | |
10256 | return -EBADF; | |
10257 | ||
10258 | ret = -EOPNOTSUPP; | |
10259 | if (f.file->f_op != &io_uring_fops) | |
10260 | goto out_fput; | |
10261 | ||
10262 | ctx = f.file->private_data; | |
10263 | ||
b6c23dd5 PB |
10264 | io_run_task_work(); |
10265 | ||
edafccee JA |
10266 | mutex_lock(&ctx->uring_lock); |
10267 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
10268 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
10269 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
10270 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
10271 | out_fput: |
10272 | fdput(f); | |
10273 | return ret; | |
10274 | } | |
10275 | ||
2b188cc1 JA |
10276 | static int __init io_uring_init(void) |
10277 | { | |
d7f62e82 SM |
10278 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
10279 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
10280 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
10281 | } while (0) | |
10282 | ||
10283 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
10284 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
10285 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
10286 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
10287 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
10288 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
10289 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
10290 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
10291 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
10292 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 10293 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
10294 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
10295 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
10296 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
10297 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
10298 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
10299 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
10300 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
10301 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
10302 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
10303 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
10304 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
10305 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
10306 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
10307 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
10308 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 10309 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
10310 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
10311 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 10312 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 10313 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 10314 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 10315 | |
b0d658ec PB |
10316 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
10317 | sizeof(struct io_uring_rsrc_update)); | |
10318 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
10319 | sizeof(struct io_uring_rsrc_update2)); | |
10320 | /* should fit into one byte */ | |
10321 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
10322 | ||
d3656344 | 10323 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
84557871 | 10324 | BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int)); |
16340eab | 10325 | |
91f245d5 JA |
10326 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
10327 | SLAB_ACCOUNT); | |
2b188cc1 JA |
10328 | return 0; |
10329 | }; | |
10330 | __initcall(io_uring_init); |